Datasets:

jamescalam

/

ai-arxiv-chunked

like 32

1707.06347

3

form ˆg=ˆEt/bracketleftBig ∇θlogπθ(at|st)ˆAt/bracketrightBig (1) whereπθis a stochastic policy and ˆAtis an estimator of the advantage function at timestep t. Here, the expectation ˆEt[...] indicates the empirical average over a finite batch of samples, in an algorithm that alternates between sampling and optimization. Implementations that use automatic differentiation software work by constructing an objective function whose gradient is the policy gradient estimator; the estimator ˆ gis obtained by differentiating the objective LPG(θ) =ˆEt/bracketleftBig logπθ(at|st)ˆAt/bracketrightBig . (2) While it is appealing to perform multiple steps of optimization on this loss LPGusing the same trajectory, doing so is not well-justified, and empirically it often leads to destructively large policy updates (see Section 6.1; results are not shown but were similar or worse than the “no clipping or penalty” setting). 2.2 Trust Region Methods

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

4

updates (see Section 6.1; results are not shown but were similar or worse than the “no clipping or penalty” setting). 2.2 Trust Region Methods In TRPO [Sch+15b], an objective function (the “surrogate” objective) is maximized subject to a constraint on the size of the policy update. Specifically, maximize θˆEt/bracketleftbiggπθ(at|st) πθold(at|st)ˆAt/bracketrightbigg (3) subject to ˆEt[KL[πθold(·|st),πθ(·|st)]]≤δ. (4) Here,θoldis the vector of policy parameters before the update. This problem can efficiently be approximately solved using the conjugate gradient algorithm, after making a linear approximation to the objective and a quadratic approximation to the constraint. The theory justifying TRPO actually suggests using a penalty instead of a constraint, i.e., solving the unconstrained optimization problem maximize θˆEt/bracketleftbiggπθ(at|st)

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

5

solving the unconstrained optimization problem maximize θˆEt/bracketleftbiggπθ(at|st) πθold(at|st)ˆAt−βKL[πθold(·|st),πθ(·|st)]/bracketrightbigg (5) for some coefficient β. This follows from the fact that a certain surrogate objective (which computes the max KL over states instead of the mean) forms a lower bound (i.e., a pessimistic bound) on the performance of the policy π. TRPO uses a hard constraint rather than a penalty because it is hard to choose a single value of βthat performs well across different problems—or even within a single problem, where the the characteristics change over the course of learning. Hence, to achieve our goal of a first-order algorithm that emulates the monotonic improvement of TRPO, experiments show that it is not sufficient to simply choose a fixed penalty coefficient βand optimize the penalized objective Equation (5) with SGD; additional modifications are required. 2 3 Clipped Surrogate Objective

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

6

objective Equation (5) with SGD; additional modifications are required. 2 3 Clipped Surrogate Objective Letrt(θ) denote the probability ratio rt(θ) =πθ(at|st) πθold(at|st), sor(θold) = 1. TRPO maximizes a “surrogate” objective LCPI(θ) =ˆEt/bracketleftbiggπθ(at|st) πθold(at|st)ˆAt/bracketrightbigg =ˆEt/bracketleftBig rt(θ)ˆAt/bracketrightBig . (6) The superscript CPI refers to conservative policy iteration [KL02], where this objective was proposed. Without a constraint, maximization of LCPIwould lead to an excessively large policy update; hence, we now consider how to modify the objective, to penalize changes to the policy that movert(θ) away from 1. The main objective we propose is the following: LCLIP(θ) =ˆEt/bracketleftBig

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

7

movert(θ) away from 1. The main objective we propose is the following: LCLIP(θ) =ˆEt/bracketleftBig min(rt(θ)ˆAt,clip(rt(θ),1−/epsilon1,1 +/epsilon1)ˆAt)/bracketrightBig (7) where epsilon is a hyperparameter, say, /epsilon1= 0.2. The motivation for this objective is as follows. The first term inside the min is LCPI. The second term, clip( rt(θ),1−/epsilon1,1+/epsilon1)ˆAt, modifies the surrogate objective by clipping the probability ratio, which removes the incentive for moving rtoutside of the interval [1−/epsilon1,1 +/epsilon1]. Finally, we take the minimum of the clipped and unclipped objective, so the final objective is a lower bound (i.e., a pessimistic bound) on the unclipped objective. With this scheme, we only ignore the change in probability ratio when it would make the objective improve,

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

8

scheme, we only ignore the change in probability ratio when it would make the objective improve, and we include it when it makes the objective worse. Note that LCLIP(θ) =LCPI(θ) to first order aroundθold(i.e., where r= 1), however, they become different as θmoves away from θold. Figure 1 plots a single term (i.e., a single t) inLCLIP; note that the probability ratio ris clipped at 1−/epsilon1 or 1 +/epsilon1depending on whether the advantage is positive or negative. rLCLIP 0 11 +/epsilon1A> 0 r LCLIP0 1 1−/epsilon1A< 0 Figure 1: Plots showing one term (i.e., a single timestep) of the surrogate function LCLIPas a function of the probability ratio r, for positive advantages (left) and negative advantages (right). The red circle on each plot shows the starting point for the optimization, i.e., r= 1. Note that LCLIPsums many of these terms. Figure 2 provides another source of intuition about the surrogate objective LCLIP. It shows how

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

9

Figure 2 provides another source of intuition about the surrogate objective LCLIP. It shows how several objectives vary as we interpolate along the policy update direction, obtained by proximal policy optimization (the algorithm we will introduce shortly) on a continuous control problem. We can see that LCLIPis a lower bound on LCPI, with a penalty for having too large of a policy update. 3 0 1 Linear interpolation factor0.02 0.000.020.040.060.080.100.12 Et[KLt] LCPI=Et[rtAt] Et[clip(rt,1 ,1+)At] LCLIP=Et[min(rtAt,clip(rt,1 ,1+)At)] Figure 2: Surrogate objectives, as we interpolate between the initial policy parameter θold, and the updated policy parameter, which we compute after one iteration of PPO. The updated policy has a KL divergence of about 0.02 from the initial policy, and this is the point at which LCLIPis maximal. This plot corresponds to the first policy update on the Hopper-v1 problem, using hyperparameters provided in Section 6.1. 4 Adaptive KL Penalty Coefficient Another approach, which can be used as an alternative to the clipped surrogate objective, or in

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

10

4 Adaptive KL Penalty Coefficient Another approach, which can be used as an alternative to the clipped surrogate objective, or in addition to it, is to use a penalty on KL divergence, and to adapt the penalty coefficient so that we achieve some target value of the KL divergence dtargeach policy update. In our experiments, we found that the KL penalty performed worse than the clipped surrogate objective, however, we’ve included it here because it’s an important baseline. In the simplest instantiation of this algorithm, we perform the following steps in each policy update: •Using several epochs of minibatch SGD, optimize the KL-penalized objective LKLPEN(θ) =ˆEt/bracketleftbiggπθ(at|st) πθold(at|st)ˆAt−βKL[πθold(·|st),πθ(·|st)]/bracketrightbigg (8) •Computed=ˆEt[KL[πθold(·|st),πθ(·|st)]] –Ifd<d targ/1.5,β←β/2 –Ifd>d targ×1.5,β←β×2

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

11

–Ifd<d targ/1.5,β←β/2 –Ifd>d targ×1.5,β←β×2 The updated βis used for the next policy update. With this scheme, we occasionally see policy updates where the KL divergence is significantly different from dtarg, however, these are rare, and βquickly adjusts. The parameters 1 .5 and 2 above are chosen heuristically, but the algorithm is not very sensitive to them. The initial value of βis a another hyperparameter but is not important in practice because the algorithm quickly adjusts it. 5 Algorithm The surrogate losses from the previous sections can be computed and differentiated with a minor change to a typical policy gradient implementation. For implementations that use automatic differentation, one simply constructs the loss LCLIPorLKLPENinstead ofLPG, and one performs multiple steps of stochastic gradient ascent on this objective. Most techniques for computing variance-reduced advantage-function estimators make use a learned state-value function V(s); for example, generalized advantage estimation [Sch+15a], or the 4 finite-horizon estimators in [Mni+16]. If using a neural network architecture that shares parameters

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

12

4 finite-horizon estimators in [Mni+16]. If using a neural network architecture that shares parameters between the policy and value function, we must use a loss function that combines the policy surrogate and a value function error term. This objective can further be augmented by adding an entropy bonus to ensure sufficient exploration, as suggested in past work [Wil92; Mni+16]. Combining these terms, we obtain the following objective, which is (approximately) maximized each iteration: LCLIP +VF+S t (θ) =ˆEt/bracketleftbig LCLIP t (θ)−c1LVF t(θ) +c2S[πθ](st)/bracketrightbig , (9) wherec1,c2are coefficients, and Sdenotes an entropy bonus, and LVF tis a squared-error loss (Vθ(st)−Vtarg t)2. One style of policy gradient implementation, popularized in [Mni+16] and well-suited for use with recurrent neural networks, runs the policy for Ttimesteps (where Tis much less than the episode length), and uses the collected samples for an update. This style requires an advantage

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

13

with recurrent neural networks, runs the policy for Ttimesteps (where Tis much less than the episode length), and uses the collected samples for an update. This style requires an advantage estimator that does not look beyond timestep T. The estimator used by [Mni+16] is ˆAt=−V(st) +rt+γrt+1+···+γT−t+1rT−1+γT−tV(sT) (10) wheretspecifies the time index in [0 ,T], within a given length- Ttrajectory segment. Generalizing this choice, we can use a truncated version of generalized advantage estimation, which reduces to Equation (10) when λ= 1: ˆAt=δt+ (γλ)δt+1+···+···+ (γλ)T−t+1δT−1, (11) whereδt=rt+γV(st+1)−V(st) (12) A proximal policy optimization (PPO) algorithm that uses fixed-length trajectory segments is shown below. Each iteration, each of N(parallel) actors collect Ttimesteps of data. Then we

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

14

shown below. Each iteration, each of N(parallel) actors collect Ttimesteps of data. Then we construct the surrogate loss on these NTtimesteps of data, and optimize it with minibatch SGD (or usually for better performance, Adam [KB14]), for Kepochs. Algorithm 1 PPO, Actor-Critic Style foriteration=1 ,2,...do foractor=1,2,...,N do Run policy πθoldin environment for Ttimesteps Compute advantage estimates ˆA1,..., ˆAT end for Optimize surrogate Lwrtθ, withKepochs and minibatch size M≤NT θold←θ end for 6 Experiments 6.1 Comparison of Surrogate Objectives First, we compare several different surrogate objectives under different hyperparameters. Here, we compare the surrogate objective LCLIPto several natural variations and ablated versions. No clipping or penalty: Lt(θ) =rt(θ)ˆAt Clipping: Lt(θ) = min(rt(θ)ˆAt,clip(rt(θ)),1−/epsilon1,1 +/epsilon1)ˆAt

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

15

KL penalty (fixed or adaptive) Lt(θ) =rt(θ)ˆAt−βKL[πθold,πθ] 5 For the KL penalty, one can either use a fixed penalty coefficient βor an adaptive coefficient as described in Section 4 using target KL value dtarg. Note that we also tried clipping in log space, but found the performance to be no better. Because we are searching over hyperparameters for each algorithm variant, we chose a computationally cheap benchmark to test the algorithms on. Namely, we used 7 simulated robotics tasks2 implemented in OpenAI Gym [Bro+16], which use the MuJoCo [TET12] physics engine. We do one million timesteps of training on each one. Besides the hyperparameters used for clipping ( /epsilon1) and the KL penalty ( β,dtarg), which we search over, the other hyperparameters are provided in in Table 3. To represent the policy, we used a fully-connected MLP with two hidden layers of 64 units, and tanh nonlinearities, outputting the mean of a Gaussian distribution, with variable standard

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

16

and tanh nonlinearities, outputting the mean of a Gaussian distribution, with variable standard deviations, following [Sch+15b; Dua+16]. We don’t share parameters between the policy and value function (so coefficient c1is irrelevant), and we don’t use an entropy bonus. Each algorithm was run on all 7 environments, with 3 random seeds on each. We scored each run of the algorithm by computing the average total reward of the last 100 episodes. We shifted and scaled the scores for each environment so that the random policy gave a score of 0 and the best result was set to 1, and averaged over 21 runs to produce a single scalar for each algorithm setting. The results are shown in Table 1. Note that the score is negative for the setting without clipping or penalties, because for one environment (half cheetah) it leads to a very negative score, which is worse than the initial random policy. algorithm avg. normalized score No clipping or penalty -0.39 Clipping,/epsilon1= 0.1 0.76 Clipping, /epsilon1= 0.2 0.82 Clipping,/epsilon1= 0.3 0.70 Adaptive KL dtarg= 0.003 0.68

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

17

Clipping,/epsilon1= 0.3 0.70 Adaptive KL dtarg= 0.003 0.68 Adaptive KL dtarg= 0.01 0.74 Adaptive KL dtarg= 0.03 0.71 Fixed KL,β= 0.3 0.62 Fixed KL,β= 1. 0.71 Fixed KL,β= 3. 0.72 Fixed KL,β= 10. 0.69 Table 1: Results from continuous control benchmark. Average normalized scores (over 21 runs of the algorithm, on 7 environments) for each algorithm / hyperparameter setting . βwas initialized at 1. 6.2 Comparison to Other Algorithms in the Continuous Domain Next, we compare PPO (with the “clipped” surrogate objective from Section 3) to several other methods from the literature, which are considered to be effective for continuous problems. We compared against tuned implementations of the following algorithms: trust region policy optimization [Sch+15b], cross-entropy method (CEM) [SL06], vanilla policy gradient with adaptive stepsize3, 2HalfCheetah, Hopper, InvertedDoublePendulum, InvertedPendulum, Reacher, Swimmer, and Walker2d, all “-v1”

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

18

3After each batch of data, the Adam stepsize is adjusted based on the KL divergence of the original and updated policy, using a rule similar to the one shown in Section 4. An implementation is available at https://github.com/ berkeleydeeprlcourse/homework/tree/master/hw4 . 6 A2C [Mni+16], A2C with trust region [Wan+16]. A2C stands for advantage actor critic, and is a synchronous version of A3C, which we found to have the same or better performance than the asynchronous version. For PPO, we used the hyperparameters from the previous section, with /epsilon1= 0.2. We see that PPO outperforms the previous methods on almost all the continuous control environments. 0 1000000500 0500100015002000HalfCheetah-v1 0 100000005001000150020002500Hopper-v1 0 100000002000400060008000InvertedDoublePendulum-v1 0 100000002004006008001000InvertedPendulum-v1 0 1000000120 100 80 60 40 20 Reacher-v1 0 1000000020406080100120Swimmer-v1 0 10000000100020003000Walker2d-v1 A2C

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

19

20 Reacher-v1 0 1000000020406080100120Swimmer-v1 0 10000000100020003000Walker2d-v1 A2C A2C + Trust Region CEM PPO (Clip) Vanilla PG, Adaptive TRPO Figure 3: Comparison of several algorithms on several MuJoCo environments, training for one million timesteps. 6.3 Showcase in the Continuous Domain: Humanoid Running and Steering To showcase the performance of PPO on high-dimensional continuous control problems, we train on a set of problems involving a 3D humanoid, where the robot must run, steer, and get up off the ground, possibly while being pelted by cubes. The three tasks we test on are (1) RoboschoolHumanoid: forward locomotion only, (2) RoboschoolHumanoidFlagrun: position of target is randomly varied every 200 timesteps or whenever the goal is reached, (3) RoboschoolHumanoidFlagrunHarder, where the robot is pelted by cubes and needs to get up off the ground. See Figure 5 for still frames of a learned policy, and Figure 4 for learning curves on the three tasks. Hyperparameters are provided in Table 4. In concurrent work, Heess et al. [Hee+17] used the adaptive KL

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

20

variant of PPO (Section 4) to learn locomotion policies for 3D robots. 0 50M Timestep01000200030004000RoboschoolHumanoid-v0 0 100M Timestep05001000150020002500RoboschoolHumanoidFlagrun-v0 0 100M Timestep0100020003000RoboschoolHumanoidFlagrunHarder-v0 Figure 4: Learning curves from PPO on 3D humanoid control tasks, using Roboschool. 7 Figure 5: Still frames of the policy learned from RoboschoolHumanoidFlagrun. In the first six frames, the robot runs towards a target. Then the position is randomly changed, and the robot turns and runs toward the new target. 6.4 Comparison to Other Algorithms on the Atari Domain We also ran PPO on the Arcade Learning Environment [Bel+15] benchmark and compared against well-tuned implementations of A2C [Mni+16] and ACER [Wan+16]. For all three algorithms, we used the same policy network architechture as used in [Mni+16]. The hyperparameters for PPO are provided in Table 5. For the other two algorithms, we used hyperparameters that were tuned to maximize performance on this benchmark.

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

21

are provided in Table 5. For the other two algorithms, we used hyperparameters that were tuned to maximize performance on this benchmark. A table of results and learning curves for all 49 games is provided in Appendix B. We consider the following two scoring metrics: (1) average reward per episode over entire training period (which favors fast learning), and (2) average reward per episode over last 100 episodes of training (which favors final performance). Table 2 shows the number of games “won” by each algorithm, where we compute the victor by averaging the scoring metric across three trials. A2C ACER PPO Tie (1) avg. episode reward over all of training 1 18 30 0 (2) avg. episode reward over last 100 episodes 1 28 19 1 Table 2: Number of games “won” by each algorithm, where the scoring metric is averaged across three trials. 7 Conclusion We have introduced proximal policy optimization, a family of policy optimization methods that use multiple epochs of stochastic gradient ascent to perform each policy update. These methods have the stability and reliability of trust-region methods but are much simpler to implement, requiring only few lines of code change to a vanilla policy gradient implementation, applicable in more general settings (for example, when using a joint architecture for the policy and value function), and have

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

22

only few lines of code change to a vanilla policy gradient implementation, applicable in more general settings (for example, when using a joint architecture for the policy and value function), and have better overall performance. 8 Acknowledgements Thanks to Rocky Duan, Peter Chen, and others at OpenAI for insightful comments. 8 References [Bel+15] M. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. “The arcade learning environment: An evaluation platform for general agents”. In: Twenty-Fourth International Joint Conference on Artificial Intelligence . 2015. [Bro+16] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba. “OpenAI Gym”. In: arXiv preprint arXiv:1606.01540 (2016). [Dua+16] Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel. “Benchmarking Deep Reinforcement Learning for Continuous Control”. In: arXiv preprint arXiv:1604.06778 (2016).

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

23

Reinforcement Learning for Continuous Control”. In: arXiv preprint arXiv:1604.06778 (2016). [Hee+17] N. Heess, S. Sriram, J. Lemmon, J. Merel, G. Wayne, Y. Tassa, T. Erez, Z. Wang, A. Eslami, M. Riedmiller, et al. “Emergence of Locomotion Behaviours in Rich Environments”. In: arXiv preprint arXiv:1707.02286 (2017). [KL02] S. Kakade and J. Langford. “Approximately optimal approximate reinforcement learning”. In: ICML . Vol. 2. 2002, pp. 267–274. [KB14] D. Kingma and J. Ba. “Adam: A method for stochastic optimization”. In: arXiv preprint arXiv:1412.6980 (2014). [Mni+15] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves,

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

24

M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. “Human-level control through deep reinforcement learning”. In: Nature 518.7540 (2015), pp. 529–533. [Mni+16] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. P. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu. “Asynchronous methods for deep reinforcement learning”. In: arXiv preprint arXiv:1602.01783 (2016). [Sch+15a] J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel. “High-dimensional continuous control using generalized advantage estimation”. In: arXiv preprint arXiv:1506.02438 (2015). [Sch+15b] J. Schulman, S. Levine, P. Moritz, M. I. Jordan, and P. Abbeel. “Trust region policy optimization”. In: CoRR, abs/1502.05477 (2015).

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

25

optimization”. In: CoRR, abs/1502.05477 (2015). [SL06] I. Szita and A. L¨ orincz. “Learning Tetris using the noisy cross-entropy method”. In: Neural computation 18.12 (2006), pp. 2936–2941. [TET12] E. Todorov, T. Erez, and Y. Tassa. “MuJoCo: A physics engine for model-based control”. In: Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on . IEEE. 2012, pp. 5026–5033. [Wan+16] Z. Wang, V. Bapst, N. Heess, V. Mnih, R. Munos, K. Kavukcuoglu, and N. de Freitas. “Sample Efficient Actor-Critic with Experience Replay”. In: arXiv preprint arXiv:1611.01224 (2016). [Wil92] R. J. Williams. “Simple statistical gradient-following algorithms for connectionist reinforcement learning”. In: Machine learning 8.3-4 (1992), pp. 229–256. 9 A Hyperparameters Hyperparameter Value Horizon (T) 2048

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

26

9 A Hyperparameters Hyperparameter Value Horizon (T) 2048 Adam stepsize 3×10−4 Num. epochs 10 Minibatch size 64 Discount (γ) 0.99 GAE parameter ( λ)0.95 Table 3: PPO hyperparameters used for the Mujoco 1 million timestep benchmark. Hyperparameter Value Horizon (T) 512 Adam stepsize ∗ Num. epochs 15 Minibatch size 4096 Discount (γ) 0.99 GAE parameter ( λ) 0.95 Number of actors 32 (locomotion), 128 (flagrun) Log stdev. of action distribution LinearAnneal(−0.7,−1.6) Table 4: PPO hyperparameters used for the Roboschool experiments. Adam stepsize was adjusted based on the target value of the KL divergence. Hyperparameter Value Horizon (T) 128 Adam stepsize 2.5×10−4×α Num. epochs 3 Minibatch size 32×8 Discount (γ) 0.99 GAE parameter ( λ)0.95 Number of actors 8 Clipping parameter /epsilon10.1×α VF coeff.c1(9) 1

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

27

Number of actors 8 Clipping parameter /epsilon10.1×α VF coeff.c1(9) 1 Entropy coeff. c2(9) 0.01 Table 5: PPO hyperparameters used in Atari experiments. αis linearly annealed from 1 to 0 over the course of learning. B Performance on More Atari Games Here we include a comparison of PPO against A2C on a larger collection of 49 Atari games. Figure 6 shows the learning curves of each of three random seeds, while Table 6 shows the mean performance. 10 10002000Alien 0250500750Amidar 020004000Assault 0250050007500Asterix 150020002500Asteroids 0100000020000003000000Atlantis 05001000BankHeist 5000100001500020000BattleZone 1000200030004000BeamRider 304050Bowling 050100Boxing 0200400Breakout 500010000Centipede 200040006000ChopperCommand 50000100000CrazyClimber 02000040000DemonAttack 17.5 15.0 12.5 10.0 DoubleDunk 0250500750Enduro 100 50 0FishingDerby 0102030Freeway 100200300Frostbite

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

28

10.0 DoubleDunk 0250500750Enduro 100 50 0FishingDerby 0102030Freeway 100200300Frostbite 02000040000Gopher 250500750Gravitar 10 8 6 4 IceHockey 0200400600Jamesbond 0500010000Kangaroo 2000400060008000Krull 02000040000KungFuMaster 050100MontezumaRevenge 100020003000MsPacman 25005000750010000NameThisGame 100 0Pitfall 20 020Pong 0500PrivateEye 050001000015000Qbert 25005000750010000Riverraid 02000040000RoadRunner 246Robotank 050010001500Seaquest 5001000SpaceInvaders 02000040000StarGunner 20 15 10 Tennis 30004000TimePilot 0100200300Tutankham 0100000200000UpNDown 0 40M Frames0510Venture 0 40M Frames50000100000150000VideoPinball 0 40M Frames20004000WizardOfWor 0 40M Frames0200040006000Zaxxon A2C ACER

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

29

0 40M Frames20004000WizardOfWor 0 40M Frames0200040006000Zaxxon A2C ACER PPOFigure 6: Comparison of PPO and A2C on all 49 ATARI games included in OpenAI Gym at the time of publication. 11 A2C ACER PPO Alien 1141.7 1655.4 1850.3 Amidar 380.8 827.6 674.6 Assault 1562.9 4653.8 4971.9 Asterix 3176.3 6801.2 4532.5 Asteroids 1653.3 2389.3 2097.5 Atlantis 729265.3 1841376.0 2311815.0 BankHeist 1095.3 1177.5 1280.6 BattleZone 3080.0 8983.3 17366.7 BeamRider 3031.7 3863.3 1590.0 Bowling 30.1 33.3 40.1 Boxing 17.7 98.9 94.6 Breakout 303.0 456.4 274.8 Centipede 3496.5 8904.8 4386.4 ChopperCommand 1171.7 5287.7 3516.3

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

30

Centipede 3496.5 8904.8 4386.4 ChopperCommand 1171.7 5287.7 3516.3 CrazyClimber 107770.0 132461.0 110202.0 DemonAttack 6639.1 38808.3 11378.4 DoubleDunk -16.2 -13.2 -14.9 Enduro 0.0 0.0 758.3 FishingDerby 20.6 34.7 17.8 Freeway 0.0 0.0 32.5 Frostbite 261.8 285.6 314.2 Gopher 1500.9 37802.3 2932.9 Gravitar 194.0 225.3 737.2 IceHockey -6.4 -5.9 -4.2 Jamesbond 52.3 261.8 560.7 Kangaroo 45.3 50.0 9928.7 Krull 8367.4 7268.4 7942.3 KungFuMaster 24900.3 27599.3 23310.3 MontezumaRevenge 0.0 0.3 42.0 MsPacman 1626.9 2718.5 2096.5 NameThisGame 5961.2 8488.0 6254.9

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

31

MsPacman 1626.9 2718.5 2096.5 NameThisGame 5961.2 8488.0 6254.9 Pitfall -55.0 -16.9 -32.9 Pong 19.7 20.7 20.7 PrivateEye 91.3 182.0 69.5 Qbert 10065.7 15316.6 14293.3 Riverraid 7653.5 9125.1 8393.6 RoadRunner 32810.0 35466.0 25076.0 Robotank 2.2 2.5 5.5 Seaquest 1714.3 1739.5 1204.5 SpaceInvaders 744.5 1213.9 942.5 StarGunner 26204.0 49817.7 32689.0 Tennis -22.2 -17.6 -14.8 TimePilot 2898.0 4175.7 4342.0 Tutankham 206.8 280.8 254.4 UpNDown 17369.8 145051.4 95445.0 Venture 0.0 0.0 0.0 VideoPinball 19735.9 156225.6 37389.0 WizardOfWor 859.0 2308.3 4185.3

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

1707.06347

32

VideoPinball 19735.9 156225.6 37389.0 WizardOfWor 859.0 2308.3 4185.3 Zaxxon 16.3 29.0 5008.7 Table 6: Mean final scores (last 100 episodes) of PPO and A2C on Atari games after 40M game frames (10M timesteps). 12

1707.06347

Proximal Policy Optimization Algorithms

We propose a new family of policy gradient methods for reinforcement learning, which alternate between sampling data through interaction with the environment, and optimizing a "surrogate" objective function using stochastic gradient ascent. Whereas standard policy gradient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the benefits of trust region policy optimization (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time.

http://arxiv.org/pdf/1707.06347

cs.LG

20170720

20170828

2009.14395

0

Can Automatic Post-Editing Improve NMT? Shamil Chollampatt Rakuten, Inc. shamil.chollampatt@rakuten.comRaymond Hendy Susanto Rakuten, Inc. raymhs91@gmail.com Liling Tan Rakuten, Inc. liling.tan@rakuten.comEwa Szymanska Rakuten, Inc. ewa.szymanska@rakuten.com Abstract Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

1

of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https:// github.com/shamilcm/pedra . 1 Introduction Automatic Post-Editing (APE) aims to reduce manual post-editing effort by automatically fixing errors in the machine-translated output. Knight and Chander (1994) first proposed APE to cope with systematic errors in selecting appropriate articles for Japanese to English translation. Earlier application of statistical phrase-based models for APE treated it as a monolingual re-writing task without considering the source sentence (Simard et al., 2007; B ´echara et al., 2011). Modern APE models take the source text and machine-translated text as input and output the post-edited text in the target

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

2

2007; B ´echara et al., 2011). Modern APE models take the source text and machine-translated text as input and output the post-edited text in the target language (see Figure 1). Source text (English): Will he send the gifts to the house? Machine translated text (German): Die Geschenke in mein Haus schicken?        (The gifts)                    (to my)     (house)        (send) Post-edited text (German): Wird er die Geschenke ins Haus schicken? (Will he)               (the gifts)            (to the) (house)        (send)Figure 1: An example of post-editing given the source text in English and the translated text in German. APE models are usually trained and evaluated in ablack-box scenario where the underlying MT model and the decoding process are inaccessible making it difficult to improve the MT system directly. APE can be effective in this case to improve the MT output or to adapt its style or domain. Recent advancement of APE has shown remarkable success on statistical machine translation (SMT) outputs (Junczys-Dowmunt and Grundkiewicz, 2018; Correia and Martins, 2019) even

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

3

when trained with limited number of post-edited training instances (generally “triplets” consisting ofsource ,translated , and post-edited segments), with or without additional large-scale artificial data (Junczys-Dowmunt and Grundkiewicz, 2016; Negri et al., 2018). Substantial improvements have been reported especially on English-German (ENDE) WMT APE shared tasks on SMT (Bojar et al., 2017; Chatterjee et al., 2018), when models were trained with fewer than 25,000 human post-edited triplets. However, on NMT, strong APE models have failed to show any notable improvement (Chatterjee et al., 2018, 2019; Ive et al., 2020) when trained on similar-sized human post-edited data. This has led to questions regarding the usefulness of APE with current NMT systems that produce improved translations compared to SMT.arXiv:2009.14395v1 [cs.CL] 30 Sep 2020 Junczys-Dowmunt and Grundkiewicz (2018) concluded that the results of the WMT’18 APE (NMT) task “might constitute the end of neural automatic post-editing for strong neural in-domain systems”

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

4

task “might constitute the end of neural automatic post-editing for strong neural in-domain systems” and that “neural-on-neural APE might not actually be useful”. Contrary to this belief, we hypothesize that a competitive neural APE model still has potential to further improve strong state-of-the-art in-domain NMT systems when trained on adequate human post-edited data. We compile a new large post-edited corpus, SubEdits , which consists of actual human postedits of translations of drama and movie subtitles produced by a strong in-domain proprietary NMT system. We use this corpus to train a state-of-theart neural APE model (Correia and Martins, 2019), with the goal of answering the following three research questions to better assess the relevance of APE going forward: Can APE substantially improve in-domain NMT with adequate data size? How much does artificial APE data help? How significant is domain shift for APE? Spoilers: Through automatic and human evaluation, we confirm our hypothesis that, in order to notably improve over the original NMT output (“donothing” baseline), state-of-the-art APE models

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

5

need to be trained on a larger number of human post-edits, unlike the case with SMT. Training on datasets of sizes in the scale of those from the WMT APE tasks, even with large-scale in-domain artificial APE corpora, leads to underperformance. Our experimental results also highlight that APE models are highly sensitive to domain differences. To effectively exploit out-of-domain post-edited corpora such as SubEdits in other domains, it has to be carefully mixed with available in-domain data. 2 SubEdits Corpus Human post-edited corpora of NMT outputs from previous WMT APE shared tasks usually consist of fewer than 25,000 instances. Large-scale artificial corpora such as eSCAPE (Negri et al., 2018), do not adequately cater to the primary APE objective of correcting systematic errors of the MT outputs since the pseudo “post-edits” are independent human-translated references often differing greatly from the MT output. Table 1 lists the real and artificial APE corpora on NMT outputs. Due toLang. Size Domain Human post-edited corpora QT21EN-LV 21KLife (Specia et al., 2017) Sciences

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

6

Human post-edited corpora QT21EN-LV 21KLife (Specia et al., 2017) Sciences WMT’18 & ’19 APEEN-DE 15K IT(Chatterjee et al., 2018) WMT’19 APEEN-RU 17K IT(Chatterjee et al., 2019) APE-QUEST EN-NL 11K Legal (Ive et al., 2020) EN-FR 10K EN-PT 10K SubEdits ( this work ) EN-DE 161K Subtitles Artificial corpora eSCAPE EN-DE 7.2M Mixed (Negri et al., 2018) EN-IT 3.3M EN-RU 7.7M Table 1: APE corpora on NMT outputs and their sizes in terms of number of post-edited triplets. the paucity of larger human post-edited corpora on NMT outputs, a study of APE performance under sufficient supervised training data conditions was not possible previously. To enable such a study, we introduce the SubEdits EN-DE post-editing corpus with over 161K triplets of source sentences, NMT translations, and human post-edits of NMT translations. 2.1 Corpus Collection

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

7

NMT translations, and human post-edits of NMT translations. 2.1 Corpus Collection SubEdits corpus is collected from a database of subtitles of a popular video streaming platform, Rakuten Viki (https://www.viki.com/) Every subtitle segment had been originally manually transcribed and translated to English before translating it to German using a proprietary NMT system employed by the platform and specialized at translating subtitles. Viki community1members who volunteer as subtitle translators would then post-edit the machine-translated subtitles to further improve it, if necessary. 2.2 Corpus Filtering We use an adaptation of Gale-Church filtering (Tan and Pal, 2014) used for machine translation for filtering the triplets. The global character mean ratio rcis computed as the ratio between the number of characters in the source and machine translated portions of the entire corpus. We remove triplets (src,mt,pe) from the corpus where the ratio between the number of characters of source ( src) and post-edit ( pe) does not lie within a threshold range of(1t)rcand(1 +t)rcwitht= 0:2. We nor1https://contribute.viki.com/ No. of No. of tokens triplets src mt pe

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

8

No. of No. of tokens triplets src mt pe Train 141,413 1,432,247 1,395,211 1,423,257 Dev 10,000 101,330 98,581 100,795 Test 10,000 101,709 99,032 101,112 Table 2: Statistics of the SubEdits corpus malize punctuation2and remove duplicate triplets. Among the triplets that share the same srcandmt segments, we choose only the one with the longest pe. Finally, we remove triplets that are not correctly identified with the respective source and target language using a language identification tool3(Lui and Baldwin, 2012). We set aside 10,000 triplets as development set and 10,000 triplets as test set. The final statistics are shown in Table 2. 3 BERT Encoder-Decoder APE Model BERT Encoder-Decoder APE (Correia and Martins, 2019) is a state-of-the-art neural APE model based on a Transformer model (Vaswani et al., 2017) with the encoder and decoder initialized with pre-trained multilingual BERT (Devlin et al., 2019) weights and fine-tuned on post-editing data.

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

9

multilingual BERT (Devlin et al., 2019) weights and fine-tuned on post-editing data. A single encoder is used to encode both the source text and the machine-translated text by concatenating them with the separator token [SEP] . The encoder component of the model is identical to the original Transformer encoder initialized with pre-trained weights from the multilingual BERT. For the decoder, Correia and Martins (2019) initialized the context attention weights with the corresponding BERT self-attention weights. Also, the weights of the self-attention layers of the encoder and decoder are tied. All other weights are initialized with corresponding weights from the same multilingual BERT model as well. BERT Encoder-Decoder APE was shown to outperform other state-of-the-art APE models (Tebbifakhr et al., 2018; Junczys-Dowmunt and Grundkiewicz, 2018) on SMT outputs even in the absence of additional large-scale artificial data that competing models have used. An improved variant of this model with additional in-domain artificial data, despite being the winning submission of the recent

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

10

model with additional in-domain artificial data, despite being the winning submission of the recent WMT’19 APE EN-DE (NMT) task (Lopes et al., 2019), only performed marginally better than the baseline NMT output. For the purpose of this study, we base our experiments on the BERT Encoder2Using Moses normalize-punctuation.perl script. 3https://github.com/saffsd/langid.pyDecoder APE architecture (Correia and Martins, 2019). 4 Experimental Setup 4.1 Model Hyperparameters For the BERT Encoder-Decoder model ( BERT Enc-Dec ), we use the implementation4and model hyperparameters used by Correia and Martins (2019) and initialize the encoder and decoder with cased multilingual BERT (base) from Transformers5library (Wolf et al., 2019). The encoder and decoder follow the architecture of BERT (base) with 12 layers and 12 attention heads, an embedding size of 768, and a feed-forward layer size of 3072. We set the effective batch size to 4096 tokens for parameter updates. We train BERT Enc-Dec on a single NVIDIA Quadro RTX6000 GPU. Training on our SubEdits corpus took approximately 5 hours

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

11

for parameter updates. We train BERT Enc-Dec on a single NVIDIA Quadro RTX6000 GPU. Training on our SubEdits corpus took approximately 5 hours to converge. We validate and save checkpoints at every 2000 steps and use early-stopping (patience of 4 checkpoints) to select the model based on best perplexity. We use a decoding beam size of 5. As a control measure, we compare BERT EncDec against two vanilla Transformer APE models using automatic metrics. The Transformer APE models use BERT vocabularies and tokenization, and employ a single encoder to encode the concatenationsrcandmt, but they are not initialized with pre-trained weights. The following are the descriptions of the two Transformer APE baselines: TF (base) A Transformer (base) (Vaswani et al., 2017) model with 6 hidden layers implemented in OpenNMT-py.6The embedding size is 512 with 2048 feed-forward units. We use default learning parameters in OpenNMT-py: Adam optimizer with a learning rate of 2 and Noam scheduler. TF (BERT size.) A bigger Transformer with the same number of layers, attention heads, embedding dimensions, hidden, and feed-forward dimensions as BERT Enc-Dec, but without any pre-training

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

12

same number of layers, attention heads, embedding dimensions, hidden, and feed-forward dimensions as BERT Enc-Dec, but without any pre-training and tying of self-attention layers. All learning hyperparameters follow that of TF (base) model. 4.2 Pre-processing and Post-processing SubEdits corpus contains HTML tags such as line breaks ( <br> ) and italic tags ( <i>), and symbols denoting musical notes ( ,) and segments often 4https://github.com/deep-spin/OpenNMT-APE 5https://github.com/huggingface/transformers 6https://github.com/OpenNMT/OpenNMT-py BLEU"ChrF"TER# Proprietary NMT 46.83 63.81 37.20 Google Translate 40.96 59.20 41.91 Microsoft Translator 38.78 57.68 43.72 SYSTRAN 38.06 56.74 44.37 Table 3: Comparison of the proprietary NMT to leading commercial MT systems on an in-domain test set. begin with hyphens (-). We applied several processing steps to make the data as close as possible to natural sentences on which BERT has been pretrained on. The triplets with multi-line src,mt,

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

13

to natural sentences on which BERT has been pretrained on. The triplets with multi-line src,mt, andpecontaining <br> tags are split into separate training instances7and we remove italics and other HTML tags, musical note symbols, and leading hyphens. Thereafter, the input is tokenized with the BERT tokenization and word-piece segmentation in the Transformers library. During test time, we keep track of the changes made to input such as deletion of leading hyphens, music symbols, and italics tags, and splitting at <br> tags. After decoding, the outputs are detokenized and post-processed to re-introduce the tracked changes and evaluated. 4.3 Evaluation We evaluate the models using three different automatic metrics: BLEU (Papineni et al., 2002), ChrF (Popovi ´c, 2015), and TER (Snover et al., 2006). For our evaluation on SubEdits test set, differing from WMT APE task evaluation, we post-process and detokenize the outputs and use SacreBLEU8(Post, 2018) to evaluate BLEU and ChrF, and TERCOM9 to compute TER with normalization. Significance test is done by bootstrap re-sampling on BLEU

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

14

to compute TER with normalization. Significance test is done by bootstrap re-sampling on BLEU with 1000 samples (Koehn, 2004). Additionally, we conduct human evaluation to ascertain the improvement of the BERT Enc-Dec APE model and to determine the human upper-bound performance for the SubEdits benchmark (see Section 5.3). We also compare the APE model on the canonical WMT APE dataset (Section 5.6 and Table 7). We follow their evaluation method and use the released tokenized post-edited reference to compute BLEU, ChrF, and TER on the tokenized output.5 Results and Discussion 5.1 Proprietary In-domain NMT We first assess the quality of an proprietary indomain NMT system that is used for compiling the SubEdits corpus. We use it as a black-box system and use the evaluation results from Table 3 to demonstrate that it is a strong baseline for studying APE performance on NMT outputs. We compare the proprietary NMT system to three leading commercial EN-DE NMT systems: Google Translate, Microsoft Translator, and SYSTRAN, on a separate in-domain EN-DE test set of 5,136 subtitle segments with independent reference translations (i.e., not post-edits of any system)

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

15

of 5,136 subtitle segments with independent reference translations (i.e., not post-edits of any system) fetched from the same video streaming platform as the SubEdits corpus. The results (as of May 2020) are summarized in Table 3. Unsurprisingly, the proprietary NMT system specialized at translating drama subtitles substantially outperforms other general MT systems. 5.2 APE Performance on SubEdits Table 4 reports the performance of vanilla transformer and BERT Enc-Dec APE models and compares it the do-nothing NMT baseline (the output produced by the proprietary in-domain NMT system). TF (base) APE improves over the do-nothing NMT baseline output ( p <0:05), particularly on TER scores. However, TF ( BERT size) APE shows a smaller improvement on ChrF and TER scores and a drop in BLEU. Even with the SubEdits corpus, large networks such as TF ( BERT size) tends to overfit. However, with pre-trained BERT initialization, BERT Enc-Dec APE shows substantial improvement across all metrics. Unlike previous studies that report marginal improvements (Chatterjee et al., 2018, 2019), our results show that a strong APE model trained on large human postedits can significantly outperform ( p <0:001) a

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

16

strong APE model trained on large human postedits can significantly outperform ( p <0:001) a strong in-domain NMT system. 5.3 Human Evaluation To validate the improvement in automatic evaluation scores and to estimate the human upper-bound performance on SubEdits, we conducted human evaluation. We hired five German native freelance translators who are also proficient in English and 7We only separate at <br> when the src,mt, and pecontains same number of <br> symbols. 8https://github.com/mjpost/sacreBLEU 9http://www.cs.umd.edu/˜snover/tercom/ No. of Dev Test Params BLEU" ChrF" TER# BLEU" ChrF" TER# do-nothing NMT 62.07 71.66 27.68 61.88 71.33 28.06 w/ TF (Base) APE 105.5M 62.47 72.26 25.65 62.26 71.97 25.94 w/ TF ( BERT size.) APE 290.4M 62.04 72.04 25.73 61.62 71.65 26.14

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

17

w/ TF ( BERT size.) APE 290.4M 62.04 72.04 25.73 61.62 71.65 26.14 w/ BERT Enc-Dec APE 262.4M 64.88 74.94 23.29 64.53 74.71 23.72 Table 4: Performance of APE models on the SubEdits test set. Figure 2: Interface used to rate the translations. had prior experience with English/German translation. Given the original English text, the annotators were asked to rate the adequacy (from 1 to 5) for three German translations: (1) the do-nothing baseline output (NMT), (2) BERT Enc-Dec APE output (APE), and (3) the human post-edited text (Human). Figure 2 shows the interface presented to the annotators for rating the translations. The three translations are presented on the same screen in random order and the annotators are unaware of their origin. Following recent WMT APE tasks (Bojar et al., 2017; Chatterjee et al., 2018, 2019), our human

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

18

their origin. Following recent WMT APE tasks (Bojar et al., 2017; Chatterjee et al., 2018, 2019), our human evaluation is also based solely on adequacy assessments. Previous studies reported a high correlation of fluency judgments with adequacy (CallisonBurch et al., 2007) making the fluency annotationsAnnotator NMT APE Human # Eval. A 3.7 4.2 4.5 593 / 603 B 3.5 4.0 4.4 594 / 603 C 3.7 4.3 4.4 603 / 603 D 2.8 3.4 3.8 587 / 603 E 3.3 3.8 4.3 602 / 603 A-E 3.4 3.9 4.3 2979 / 3015 Table 5: Average adequacy scores (1-5) rated by annotators (A to E). Overall average is shown in the last row (A-E). superfluous (Przybocki et al., 2009). Unlike the recent WMT APE tasks, we did not opt for direct assessments (Graham et al., 2013) since we wanted to evaluate the degradation or improvement in the quality of the NMT output due to APE and human

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

19

assessments (Graham et al., 2013) since we wanted to evaluate the degradation or improvement in the quality of the NMT output due to APE and human post-edits on the same English source segments. We elicit judgments for all test set instances where the APE model modified the NMT output beyond simple edits on punctuation, HTML tags, spacing, or casing. 2,815 out of the 10,000 instances in our test set contains non-simple edits. A set of 50 instances out of 2,815 was evaluated by all annotators to compute inter-annotator agreement.10 After evaluation, we filtered out the instances where the annotator was unable to decide a score for any of the three translations. The average scores by each annotator (A to E) and the overall average scores are shown in Table 5. The numerator of the “# Eval.” column indicates the number of evaluations used for the average score computation after filtering out the “ I can’t decide ” annotations. The results of our human evaluation (Table 5) show that all five annotators rate the APE output better than baseline NMT output by at least +0:5on average, reaching an overall score of 3.9. All the

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

20

than baseline NMT output by at least +0:5on average, reaching an overall score of 3.9. All the five annotators rated the human post-edited output substantially better than the NMT output and the APE output, which indicates that quality of the post-edits in the SubEdits corpus is high. Human post-edits received an overall average score of 4.3. Using the repeated set of 46 instances,11we com10Each annotator scored 603 test instances. 11We removed 4 instances out of the 50, where one or more annotators chose the “I can’t decide” option. pute inter-annotator agreement using average pairwise Cohen’s Kappa (Cohen, 1960) to be 0.27 which is considered to be fair (Landis and Koch, 1977) and similar to that observed for adequacy judgments in WMT tasks (Callison-Burch et al., 2007). However, the ranges of scores used by the annotators differ considerably (especially, annotator ‘D’). Hence, measures such as a weighted Kappaw(Cohen, 1968), which assigns partial credit to smaller disagreements and works better with ordinal data (such as our adequacy judgments), is more suitable. We compute the average pairwise

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

21

credit to smaller disagreements and works better with ordinal data (such as our adequacy judgments), is more suitable. We compute the average pairwise quadratically weighted Kappa wto be 0:50, and consider their agreement to be moderate. 5.4 Can APE substantially improve in-domain NMT with adequate data size? To analyze the effect of training data size with respect to APE performance, we train BERT Enc-Dec APE with varying sizes of training data from the SubEdits corpus and evaluated the models on the SubEdits development set. For each training data size, ranging from 6,250 to 125,000, we train three models on three random samples of the respective size from the SubEdits training set. Each point in Figure 3 denotes the mean score of the three models (the vertical error bars at each point denote the minimum andmaximum scores). The do-nothing NMT baseline score is represented by a horizontaldotted line. As a reference, we mark the size equivalent to that of WMT’18 APE EN-DE (NMT) training set (13,441 triplets) with the vertical dotted line. The rightmost point on each graph represents the score if the full training corpus is used. Although the sizes of WMT APE dataset and the SubEdits corpus are not directly comparable,

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

22

the score if the full training corpus is used. Although the sizes of WMT APE dataset and the SubEdits corpus are not directly comparable, we see that size does matter for better APE performance. When the APE model was trained on a subset of SubEdits corpus that is of the same size as the WMT’18 APE training data, it performs worse than the baseline in terms of BLEU score and only marginally improves in ChrF and TER scores (see intersection points of the vertical and horizontal lines in Figure 3). Interestingly, doubling the amount of training data from 12,500 to 25,000 provides slight BLEU gains above the do-nothing baseline and increasing the data size to 50,000 training instances improves the model further by +1 BLEU. The curves continue to show an increasing trend. After 100,000 training instances, the data size effect on score im-provement slows down. This experiment shows the possibility that previous work on APE for NMT outputs might have reached a plateau simply due to the lack of human post-edited data rather than the limited usefulness of APE models. 5.5 How much does artificial APE data help? Previous work using strong neural APE models

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

23

the limited usefulness of APE models. 5.5 How much does artificial APE data help? Previous work using strong neural APE models (Junczys-Dowmunt and Grundkiewicz, 2018; Tebbifakhr et al., 2018) relied predominantly on artificial corpora such as that released by JunczysDowmunt and Grundkiewicz (2016) and the eSCAPE corpora (Negri et al., 2018). However, artificial post-edits are either generated from monolingual corpora or independent reference translations and they do not directly address the errors made by the MT system that is to be fixed by APE. We compare the APE model performance when trained on large-scale in-domain and out-of-domain artificial data (in the order of millions of triplets) to training on the human post-edited SubEdits corpus (over 141K human post-edits). As out-of-domain artificial data, we use the eSCAPE EN-DE NMT corpus and filter sentences that have between 0 and 200 characters resulting in 5.3 million triplets. As

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

24

corpus and filter sentences that have between 0 and 200 characters resulting in 5.3 million triplets. As in-domain artificial data, we generated an artificial APE corpus using the same approach used to create the eSCAPE corpus by decoding the source sentences from the OpenSubtitles2016 parallel corpus (Lison and Tiedemann, 2016), which is also from the subtitle domain12using the same proprietary NMT system we use to create the SubEdits corpus; the corresponding references translations become the artificial post-edits. We use the same filtering criteria and pre-processing methods for SubEdits (Section 2.2 and 4.2) resulting in 5.6 million artificial triplets. We set aside 10,000 triplets from each artificial corpus and use it as a development set when training solely on the corresponding corpus. We refer to this artificial corpus as SubEscape. We compare the performance of the BERT EncDec APE trained on SubEdits corpus to that when trained on the artificial corpora in Table 6. We

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

25

We compare the performance of the BERT EncDec APE trained on SubEdits corpus to that when trained on the artificial corpora in Table 6. We find that training on artificial corpora alone, irrespective of their domain, cannot improve over the do-nothing baseline and in fact, degrades the performance substantially. However, when we combine SubEscape with up-sampled (10 ) SubEdits cor12Although both SubEdits and SubEscape are from the subtitle domain, the translations in SubEscape are from www.opensubtitles.org/ whereas the SubEdits post-edits are compiled from Rakuten Viki. 625012500 25000 50000100000 125000 141413 Number of training instances606162636465BLEU “do-nothing” Baseline 625012500 25000 50000100000 125000 141413 Number of training instances707172737475ChrF “do-nothing” Baseline 625012500 25000 50000100000 125000 141413 Number of training instances232425262728TER “do-nothing” BaselineFigure 3: Performance of BERT Enc-Dec APE model with varying training data size in terms of BLEU, ChrF, and

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

26

“do-nothing” BaselineFigure 3: Performance of BERT Enc-Dec APE model with varying training data size in terms of BLEU, ChrF, and TER metrics on the SubEdits dev set. The vertical dotted line in each figure shows the data size used for WMT APE EN-DE (NMT) task (13,441 triplets) and the horizontal dotted line shows the NMT Baseline results. BLEU"ChrF"TER# do-nothing NMT 61.88 71.33 28.06 w/ BERT Enc-Dec APE trained on: SubEdits (R) 64.53 74.71 23.72 eSCAPE (A) 52.35 65.65 31.95 SubEscape (A) 50.51 65.89 32.78 +SubEdits 10(A+R) 64.59 75.09 23.41 Table 6: APE performance on SubEdits test set when trained with real (R) and artificial (A) training corpora. pus, we get a small improvement, particularly in terms of ChrF and TER. 5.6 How significant is domain shift for APE?

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

27

pus, we get a small improvement, particularly in terms of ChrF and TER. 5.6 How significant is domain shift for APE? While NMT performance has been known to be particularly domain-dependant (Chu and Wang, 2018), domain shift between NMT and APE training has not been investigated previously. To assess this, we evaluate BERT Enc-Dec APE on the canonical WMT’18 APE EN-DE (NMT) dataset.13. The baseline NMT system and datasets used for the WMT’18 task is from the Information Technology (IT) domain and is notably different from the domain of SubEdits. We experiment with different methods of combining SubEdits (out-domain) with the WMT APE training data (in-domain). For all experiments, we use 1,000 instances held out from the WMT’18 APE training data as the validation set. The results are reported in Table 7. When trained on SubEdits alone, despite its size, we see that there is a drastic drop in performance compared to training the much smaller WMT APE data alone. When we combine SubEdits with 10  upsampled WMT APE training data, we observe

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

28

compared to training the much smaller WMT APE data alone. When we combine SubEdits with 10  upsampled WMT APE training data, we observe 13WMT’19 APE task also used the same dataset for benchmarking EN-DE APE systemsBLEU"ChrF"TER# do-nothing NMT 74.73 85.89 16.84 w/ BERT Enc-Dec APE trained on: WMT’18 APE (I) 75.08 85.81 16.88 SubEdits (O) 49.05 69.48 39.30 +WMT’18 APE (O+I) 74.93 85.90 16.92 +WMT’18 APE 10(O+I) 75.27 86.08 16.62 Table 7: APE performance with in (I) and out-ofdomain (O) training data on WMT APE NMT test set. some improvement, particularly in terms of BLEU (p < 0:05), over training with WMT APE data alone. These results show that in-domain training data is crucial to training APE models to improve in-domain NMT. 6 Analysis 6.1 Impact of APE with varying NMT quality To study the impact of APE with varying quality

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

29

in-domain NMT. 6 Analysis 6.1 Impact of APE with varying NMT quality To study the impact of APE with varying quality of NMT output, we conduct analysis on subsets of our development set with varying translation qualities (Figure 4). We split the SubEdits development set into 10 subsets by aggregating those triplets with the NMT output scoring >90TER (lowest quality), 9081TER,:::,2011TER, and10(highest quality). They are ordered from left to right in the x-axis in Figure 4 according to increasing MT quality. y-axis denotes the difference (
) between the TER score of APE output and NMT output for each subset. The more negative
TER indicates a larger improvement due to APE. We find that on the lower quality subsets, APE improves over NMT substantially. This improvement margin reduces with improving NMT quality and can deteriorate the NMT output when NMT quality is at the highest. This experiment MT Quality 40 30 20 10 0TERAPENMT Figure 4: Translation quality difference due to APE (y-axis) shown by the
TER APENMT with increasing MT quality ( x-axis). Negative
TER indicates improvement in performance.

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

30

(y-axis) shown by the
TER APENMT with increasing MT quality ( x-axis). Negative
TER indicates improvement in performance. shows that APE contributes to improving overall NMT performance by predominantly fixing poorer quality NMT outputs. The APE model’s error will dominate and APE can become counter-productive when NMT output is nearly perfect (i.e., when there are very few or no post-edits done on them as indicated by sentence-level TER scores of <10). APE task remains relevant until NMT systems achieve this state, which is still not the case even for strong in-domain NMT systems as indicated by our experiments. 6.2 Qualitative Analysis We qualitatively analyze the output produced by APE on the SubEdits development set to better understand the improvements and errors made by the APE model. Table 8 shows three example outputs produced by the APE model along with the original English text (SRC), the do-nothing baseline output (NMT), and the human post-edits (Human). APE is able to fix incorrect named-entity translations made by the NMT system. Example 1 demonstrates an example (“Zhongyuan Palast” !“Palast Zhongcui”) where the incorrect entity is corrected

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

31

Zhongcui”) where the incorrect entity is corrected by the APE model to match the human post-edits. NMT often under-translates and misses phrases and the APE models usually can patch these undertranslations, e.g. Example 2 where the prepositional phrase “to the resort” !“zum Resort” was missing in the MT outputs and the APE model was able to mend the translation. As much as sentence-level APE works well empirically, the lack of context results in erroneousExample 1: Incorrect named entities SRC Go to Zhongcui Palace ! NMT Geh zum Zhongyuan Palast ! APE Geh zum Palast Zhongcui ! Human Geht zum Palast Zhongcui ! Example 2: Missing phrases SRC Let’s go back to the resort and we’ll talk it out. NMT Geh zurck und wir werden reden. APE Geh zurck zum Resort und wir werden reden. Human Lass uns zurck zum Resort gehen und darber reden. Example 3: Requires more context SRC Before coming, City Master negotiated with me.

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

32

Human Lass uns zurck zum Resort gehen und darber reden. Example 3: Requires more context SRC Before coming, City Master negotiated with me. NMT Bevor ergekommen ist, hat der Stadtmeister ml cit mir verhandelt. APE Bevor wirkommen, hat die Stadtmeisterin mit mir verhandelt. Human Bevor ichkam, hat die Stadtmeisterin mit mir verhandelt. Table 8: Examples where the APE model proposes changes to the NMT output on the SubEdits test set. The original sentence in English (SRC) and the human post-edit (Human) is also shown. translation by the NMT system where it tries to infer a wrong pronoun and the APE model attempts to assume yet another wrong pronoun, e.g. translating a pronoun-dropped source text in Example 3. Often, the prior or future context from video, audio, or other subtitle instances is necessary to fill these contextual gaps. Sentence-level APE cannot address these issues robustly, which calls for further research on multimodal (Deena et al., 2017;

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

33

these contextual gaps. Sentence-level APE cannot address these issues robustly, which calls for further research on multimodal (Deena et al., 2017; Caglayan et al., 2019) and document-level (Hardmeier et al., 2015; V oita et al., 2019) translation and post-editing, especially for subtitles. 7 Related Work Until 2018, APE models were benchmarked on SMT outputs through various WMT APE tasks (Bojar et al., 2015, 2016, 2017). The scale of postedited data provided by these tasks was in the order of 10,000 to 25,000 triplets. The largest collection of human post-edits, released by Zhechev (2012), however, was on SMT and consisted of 30,000 to 410,000 triplets across 12 language pairs. On SMT output, participating systems showed impressive gains even with small training datasets from WMT APE tasks (Junczys-Dowmunt and JunczysDowmunt, 2017; Tebbifakhr et al., 2018). The results of subsequent APE (NMT) tasks were not as promising with only marginal improvements on English-German and no improvement on EnglishRussian (Chatterjee et al., 2019).

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

34

results of subsequent APE (NMT) tasks were not as promising with only marginal improvements on English-German and no improvement on EnglishRussian (Chatterjee et al., 2019). Previously, there was no study to assess the necessity of larger human post-edited training data on APE performance on NMT outputs which we address in this paper. APE models were predominantly trained on large-scale artificial data combined with a few thousand human post-edits. Junczys-Dowmunt and Grundkiewicz (2016) proposed generation of large-scale artificial APE training data via round-trip translation approach inspired from back-translation (Sennrich et al., 2016). They combined artificial training data with real data provided by WMT APE tasks to train their model. Using a similar approach of generating artificial APE data, Freitag et al. (2019) trained a monolingual re-writing APE model trained on the generated artificial training data alone. Contrary to the roundtrip translation approach, large-scale artificial APE data was generated by simply translating source sentences using NMT and SMT systems and using the reference translations as the “pseudo” post-edits

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

35

data was generated by simply translating source sentences using NMT and SMT systems and using the reference translations as the “pseudo” post-edits to create eSCAPE corpus (Negri et al., 2018). Using the eSCAPE English-Italian APE corpus, Negri et al. (2017) assessed the performance of an online APE model in a simulated environment where the APE model is updated at test time with new user inputs. They found that their online APE models trained on eSCAPE found it difficult to improve specialized in-domain NMT systems. Such an analysis by training on artificial corpora may not adequately assess the actual potential of APE since these corpora do not fully cater to the task and can be noisy. The “synthetic” post-edits are independent or loosely coupled with the MT outputs, and are often drastically different from the MT output. This makes analyzing APE performance over competitive NMT systems on actual post-edited data an important step in understanding the potential of APE research. Contrary to previous conclusions, our analysis shows that a competitive in-domain NMT system can be markedly improved by a strong neural APE model when trained on sufficient human post-edited training data. 8 Conclusion

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

36

in-domain NMT system can be markedly improved by a strong neural APE model when trained on sufficient human post-edited training data. 8 Conclusion APE has been an effective option to fix systematic MT errors and improve translations from black-box MT services. However, on NMT outputs, APE has shown hardly any improvement since training has been done on limited human post-edited data. The newly collected SubEdits corpus is the largest corpus of NMT human post-edits collected so far. We reassessed the usefulness of APE on NMT usingthis corpus. We showed that with a larger human post-edited corpus, a strong neural APE model can substantially improve a strong in-domain NMT system. While artificial APE corpora help, we showed that the APE model performs better when trained on adequate human post-edited data (SubEdits) compared to large-scale artificial corpora. Finally, our experiments comparing in and out-domain APE show that domain-specificity of training affects APE performance drastically and a combination of in and out-of-domain data with certain upscaling alleviates the domain-shift problem for APE.

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

37

APE performance drastically and a combination of in and out-of-domain data with certain upscaling alleviates the domain-shift problem for APE. We find that APE mostly contributes to improving NMT performance by fixing the poorer-quality outputs that still exist with strong in-domain NMT systems. We release the post-editing datasets used in this paper (SubEscape and SubEdits) along with pre/post-processing scipts at PEDRa GitHub repository ( https://github.com/shamilcm/pedra ) Acknowledgements We thank the anonymous reviewers for their useful comments. We also thank Rakuten Viki community members who had contributed subtitle post-edits that helped building the SubEdits dataset. References Hanna B ´echara, Yanjun Ma, and Josef van Genabith. 2011. Statistical post-editing for a statistical MT system. In Proceedings of the 13th Machine Translation Summit . Ondˇrej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Shujian Huang, Matthias Huck, Philipp Koehn, Qun Liu, Varvara Logacheva, Christof Monz, Matteo Negri, Matt Post,

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

38

Matthias Huck, Philipp Koehn, Qun Liu, Varvara Logacheva, Christof Monz, Matteo Negri, Matt Post, Raphael Rubino, Lucia Specia, and Marco Turchi. 2017. Findings of the 2017 conference on machine translation (WMT17). In Proceedings of the Second Conference on Machine Translation . Ondˇrej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Aur ´elie N´ev´eol, Mariana Neves, Martin Popel, Matt Post, Raphael Rubino, Carolina Scarton, Lucia Specia, Marco Turchi, Karin Verspoor, and Marcos Zampieri. 2016. Findings of the 2016 conference on machine translation. In Proceedings of the First Conference on Machine Translation: Volume 2, Shared Task Papers . Ondˇrej Bojar, Rajen Chatterjee, Christian Federmann, Barry Haddow, Matthias Huck, Chris Hokamp, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Matt Post, Carolina Scarton, Lucia

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

39

Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Matt Post, Carolina Scarton, Lucia Specia, and Marco Turchi. 2015. Findings of the 2015 Workshop on Statistical Machine Translation. InProceedings of the 10th Workshop on Statistical Machine Translation . Ozan Caglayan, Pranava Madhyastha, Lucia Specia, and Lo ¨ıc Barrault. 2019. Probing the need for visual context in multimodal machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies . Chris Callison-Burch, Cameron Fordyce, Philipp Koehn, Christof Monz, and Josh Schroeder. 2007. (meta-) evaluation of machine translation. In Proceedings of the Second Workshop on Statistical Machine Translation . Rajen Chatterjee, Christian Federmann, Matteo Negri, and Marco Turchi. 2019. Findings of the WMT 2019 shared task on automatic post-editing. In Proceedings of the Fourth Conference on Machine Translation: Shared Task Papers . Rajen Chatterjee, Matteo Negri, Raphael Rubino, and Marco Turchi. 2018. Findings of the WMT 2018

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

40

Translation: Shared Task Papers . Rajen Chatterjee, Matteo Negri, Raphael Rubino, and Marco Turchi. 2018. Findings of the WMT 2018 shared task on automatic post-editing. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers . Chenhui Chu and Rui Wang. 2018. A survey of domain adaptation for neural machine translation. In Proceedings of the 27th International Conference on Computational Linguistics . Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement , 20(1):37–46. Jacob Cohen. 1968. Weighted Kappa: Nominal scale agreement provision for scaled disagreement or partial credit. Psychological bulletin , 70(4):213. Gonc ¸alo M. Correia and Andr ´e F. T. Martins. 2019. A simple and effective approach to automatic postediting with transfer learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics . Salil Deena, Raymond WM Ng, Pranava Madhyastha, Lucia Specia, and Thomas Hain. 2017. Exploring the use of acoustic embeddings in neural machine translation. In Proceedings of the 2017 IEEE Automatic Speech Recognition and Understanding Workshop , pages 450–457.

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

41

the use of acoustic embeddings in neural machine translation. In Proceedings of the 2017 IEEE Automatic Speech Recognition and Understanding Workshop , pages 450–457. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) .Markus Freitag, Isaac Caswell, and Scott Roy. 2019. APE at scale and its implications on MT evaluation biases. In Proceedings of the Fourth Conference on Machine Translation (Volume 1: Research Papers) . Yvette Graham, Timothy Baldwin, Alistair Moffat, and Justin Zobel. 2013. Continuous measurement scales in human evaluation of machine translation. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse . Christian Hardmeier, Preslav Nakov, Sara Stymne, J ¨org Tiedemann, Yannick Versley, and Mauro Cettolo. 2015. Pronoun-focused mt and cross-lingual pronoun prediction: Findings of the 2015 discomt shared task on pronoun translation. In Proceedings of the Second Workshop on Discourse in Machine

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

42

shared task on pronoun translation. In Proceedings of the Second Workshop on Discourse in Machine Translation (DiscoMT) . Julia Ive, Lucia Specia, Sara Szoc, Tom Vanallemeersch, Joachim Van den Bogaert, Eduardo Farah, Christine Maroti, Artur Ventura, and Maxim Khalilov. 2020. A post-editing dataset in the legal domain: Do we underestimate neural machine translation quality? In Proceedings of The 12th Language Resources and Evaluation Conference . Marcin Junczys-Dowmunt and Roman Grundkiewicz. 2016. Log-linear combinations of monolingual and bilingual neural machine translation models for automatic post-editing. In Proceedings of the First Conference on Machine Translation: Volume 2, Shared Task Papers . Marcin Junczys-Dowmunt and Roman Grundkiewicz. 2018. MS-UEdin submission to the WMT2018 APE shared task: Dual-source transformer for automatic post-editing. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers . Marcin Junczys-Dowmunt and Marcin JunczysDowmunt. 2017. The AMU-UEdin submission to the WMT 2017 shared task on automatic postediting. In Proceedings of the Second Conference on Machine Translation .

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

43

to the WMT 2017 shared task on automatic postediting. In Proceedings of the Second Conference on Machine Translation . Kevin Knight and Ishwar Chander. 1994. Automated postediting of documents. In Proceedings of the 12th AAAI National Conference on Artificial Intelligence . Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing . J. Richard Landis and Gary G. Koch. 1977. The measurement of observer agreement for categorical dat. Biometrics , 33(1):159–174. Pierre Lison and J ¨org Tiedemann. 2016. OpenSubtitles2016: Extracting large parallel corpora from movie and TV subtitles. In Proceedings of the 10th International Conference on Language Resources and Evaluation . Ant´onio V . Lopes, M. Amin Farajian, Gonc ¸alo M. Correia, Jonay Tr ´enous, and Andr ´e F. T. Martins. 2019. Unbabel’s submission to the WMT2019 APE shared task: BERT-based encoder-decoder for automatic post-editing. In Proceedings of the Fourth Conference on Machine Translation (Volume 3: Shared Task Papers) .

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

44

task: BERT-based encoder-decoder for automatic post-editing. In Proceedings of the Fourth Conference on Machine Translation (Volume 3: Shared Task Papers) . Marco Lui and Timothy Baldwin. 2012. langid.py: An off-the-shelf language identification tool. In Proceedings of the ACL 2012 System Demonstrations . Matteo Negri, Marco Turchi, Nicola Bertoldi, and Marcello Federico. 2017. Online neural automatic postediting for neural machine translation. In Proceedings of the Fifth Italian Conference on Computational Linguistics . Matteo Negri, Marco Turchi, Rajen Chatterjee, and Nicola Bertoldi. 2018. eSCAPE: a large-scale synthetic corpus for automatic post-editing. In Proceedings of the 11th International Conference on Language Resources and Evaluation . Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics . Maja Popovi ´c. 2015. chrF: character n-gram F-score for automatic MT evaluation. In Proceedings of the 10th Workshop on Statistical Machine Translation . Association for Computational Linguistics.

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

45

for automatic MT evaluation. In Proceedings of the 10th Workshop on Statistical Machine Translation . Association for Computational Linguistics. Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers . Mark Przybocki, Kay Peterson, S ´ebastien Bronsart, and Gregory Sanders. 2009. The NIST 2008 Metrics for machine translation challenge overview, methodology, metrics, and results. Machine Translation , 23(2-3):71–103. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . Michel Simard, Cyril Goutte, and Pierre Isabelle. 2007. Statistical phrase-based post-editing. In Proceedings of Human Language Technologies: The 2007 Annual Conference of the North American Chapter of the Association for Computational Linguistics . Matthew Snover, Bonnie Dorr, Richard Shwartz, Linnea Micciulla, and John Makhoul. 2006. A study of translation edit rate with targeted human annotation. InProceedings of the Seventh Conference of the Association for Machine Translation in the Americas .

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

46

translation edit rate with targeted human annotation. InProceedings of the Seventh Conference of the Association for Machine Translation in the Americas . Lucia Specia, Kim Harris, Frdric Blain, Aljoscha Burchardt, Viviven Macketanz, Inguna Skadia, MatteoNegri, , and Marco Turchi. 2017. Translation quality and productivity: A study on rich morphology languages. In Proceedings of Machine Translation Summit XVI . Liling Tan and Santanu Pal. 2014. Manawi: Using multi-word expressions and named entities to improve machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation . Amirhossein Tebbifakhr, Ruchit Agrawal, Matteo Negri, and Marco Turchi. 2018. Multi-source transformer with combined losses for automatic post editing. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers . Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30 . Elena V oita, Rico Sennrich, and Ivan Titov. 2019.

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

2009.14395

47

you need. In Advances in Neural Information Processing Systems 30 . Elena V oita, Rico Sennrich, and Ivan Titov. 2019. Context-aware monolingual repair for neural machine translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the Ninth International Joint Conference on Natural Language Processing . Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R’emi Louf, Morgan Funtowicz, and Jamie Brew. 2019. HuggingFace’s Transformers: State-of-the-art natural language processing. arXiv preprint , arXiv:1910.03771. Ventsislav Zhechev. 2012. Machine translation infrastructure and post-editing performance at Autodesk. InProceedings of the AMTA 2012 Workshop on PostEditing Technology and Practice .

2009.14395

Can Automatic Post-Editing Improve NMT?

Automatic post-editing (APE) aims to improve machine translations, thereby reducing human post-editing effort. APE has had notable success when used with statistical machine translation (SMT) systems but has not been as successful over neural machine translation (NMT) systems. This has raised questions on the relevance of APE task in the current scenario. However, the training of APE models has been heavily reliant on large-scale artificial corpora combined with only limited human post-edited data. We hypothesize that APE models have been underperforming in improving NMT translations due to the lack of adequate supervision. To ascertain our hypothesis, we compile a larger corpus of human post-edits of English to German NMT. We empirically show that a state-of-art neural APE model trained on this corpus can significantly improve a strong in-domain NMT system, challenging the current understanding in the field. We further investigate the effects of varying training data sizes, using artificial training data, and domain specificity for the APE task. We release this new corpus under CC BY-NC-SA 4.0 license at https://github.com/shamilcm/pedra.

http://arxiv.org/pdf/2009.14395

In EMNLP 2020

cs.CL

20200930

20200930

1706.05125

0

arXiv:1706.05125v1 [cs.AI] 16 Jun 2017Deal or No Deal? End-to-End Learning for Negotiation Dialog ues Mike Lewis1, Denis Yarats1, Yann N. Dauphin1, Devi Parikh2,1and Dhruv Batra2,1 1Facebook AI Research2Georgia Institute of Technology {mikelewis,denisy,ynd }@fb.com{parikh,dbatra }@gatech.edu Abstract Much of human dialogue occurs in semicooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other’s reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts , in which the model plans ahead by simulating possible complete continuations of the conversation, and find that

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

1

rollouts , in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available.1 1 Introduction Intelligent agents often need to cooperate with others who have different goals, and typically use natural language to agree on decisions. Negotiation is simultaneously a linguistic and a reasoning problem, in which an intent must be formulated and then verbally realised. Such dialogues contain both cooperative and adversarial elements, and require agents to understand, plan, and generate utterances to achieve their goals (Traum et al., 2008; Asher et al., 2012). We collect the first large dataset of natural language negotiations between two people, and show 1https://github.com/facebookresearch/end-to-end-nego tiatorthat end-to-end neural models can be trained to negotiate by maximizing the likelihood of human actions. This approach is scalable and domainindependent, but does not model the strategic skills required for negotiating well. We further show that models can be improved by training and decoding to maximize reward instead of likelihood—by training with self-play reinforcement learning, and using rollouts to estimate the expected reward of utterances during decoding. To study semi-cooperative dialogue, we gather a dataset of 5808 dialogues between humans on a

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

2

expected reward of utterances during decoding. To study semi-cooperative dialogue, we gather a dataset of 5808 dialogues between humans on a negotiation task. Users were shown a set of items with a value for each, and asked to agree how to divide the items with another user who has a different, unseen, value function (Figure 1). We first train recurrent neural networks to imitate human actions. We find that models trained to maximise the likelihood of human utterances can generate fluent language, but make comparatively poor negotiators, which are overly willing to compromise. We therefore explore two methods for improving the model’s strategic reasoning skills— both of which attempt to optimise for the agent’s goals, rather than simply imitating humans: Firstly, instead of training to optimise likelihood, we show that our agents can be considerably improved using self play , in which pre-trained models practice negotiating with each other in order to optimise performance. To avoid the models diverging from human language, we interleave reinforcement learning updates with supervised updates. For the first time, we show that end-toend dialogue agents trained using reinforcement learning outperform their supervised counterparts in negotiations with humans. Secondly, we introduce a new form of planning for dialogue called dialogue rollouts , in which an

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

3

learning outperform their supervised counterparts in negotiations with humans. Secondly, we introduce a new form of planning for dialogue called dialogue rollouts , in which an agent simulates complete dialogues during decoding to estimate the reward of utterances. We show Figure 1: A dialogue in our Mechanical Turk interface, which we used to collect a negotiation dataset. that decoding to maximise the reward function (rather than likelihood) significantly improves performance against both humans and machines. Analysing the performance of our agents, we find evidence of sophisticated negotiation strategies. For example, we find instances of the model feigning interest in a valueless issue, so that it can later ‘compromise’ by conceding it. Deceit is a complex skill that requires hypothesising the other agent’s beliefs, and is learnt relatively late in child development (Talwar and Lee, 2002). Our agents have learnt to deceive without any explicit human design, simply by trying to achieve their goals. The rest of the paper proceeds as follows: §2 describes the collection of a large dataset of humanhuman negotiation dialogues. §3 describes a baseline supervised model, which we then show can be improved by goal-based training ( §4) and decoding (§5).§6 measures the performance of our

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

4

be improved by goal-based training ( §4) and decoding (§5).§6 measures the performance of our models and humans on this task, and §7 gives a detailed analysis and suggests future directions. 2 Data Collection 2.1 Overview To enable end-to-end training of negotiation agents, we first develop a novel negotiation task and curate a dataset of human-human dialogues for this task. This task and dataset follow our proposed general framework for studying semicooperative dialogue. Initially, each agent is shown an input specifying a space of possible actions and a reward function which will score the outcome of the negotiation. Agents then sequentially take turns of either sending natural language messages, or selecting that a final decision has been reached. When one agent selects that anagreement has been made, both agents independently output what they think the agreed decision was. If conflicting decisions are made, both agents are given zero reward. 2.2 Task Our task is an instance of multi issue bargaining (Fershtman, 1990), and is based on DeVault et al. (2015). Two agents are both shown the same collection of items, and instructed to divide them so that each item assigned to one agent. Each agent is given a different randomly generated value function, which gives a non-negative

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

5

that each item assigned to one agent. Each agent is given a different randomly generated value function, which gives a non-negative value for each item. The value functions are constrained so that: (1) the total value for a user of all items is 10; (2) each item has non-zero value to at least one user; and (3) some items have nonzero value to both users. These constraints enforce that it is not possible for both agents to receive a maximum score, and that no item is worthless to both agents, so the negotiation will be competitive. After 10 turns, we allow agents the option to complete the negotiation with no agreement, which is worth 0 points to both users. We use 3 item types (books ,hats,balls ), and between 5 and 7 total items in the pool. Figure 1 shows our interface. 2.3 Data Collection We collected a set of human-human dialogues using Amazon Mechanical Turk. Workers were paid $0.15 per dialogue, with a $0.05 bonus for maximal scores. We only used workers based in the United States with a 95% approval rating and at least 5000 previous HITs. Our data collection interface was adapted from that of Das et al. (2016). We collected a total of 5808 dialogues, based on 2236 unique scenarios (where a scenario is the Crowd Sourced Dialogue

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

6

We collected a total of 5808 dialogues, based on 2236 unique scenarios (where a scenario is the Crowd Sourced Dialogue Agent 1 Input 3xbook value =1 2xhat value =3 1xball value =1Agent 2 Input 3xbook value =2 2xhat value =1 1xball value =2 Dialogue Agent 1: I want the books and the hats, you get the ball Agent 2: Give me a book too and we have a deal Agent 1: Ok, deal Agent 2: <choose> Agent 1 Output 2xbook 2xhatAgent 2 Output 1xbook 1xballPerspective: Agent 1 Perspective: Agent 2Input 3xbook value =1 2xhat value =3 1xball value =1 Output 2xbook 2xhatDialogue write: I want the books and the hats, you get the ball read: Give me a book too and we have a deal write: Ok, deal read:<choose> Input 3xbook value =2 2xhat value =1 1xball value =2Dialogue read: I want the books and the hats, you get the ball write: Give me a book too and we have a deal read: Ok, deal

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

7

read: I want the books and the hats, you get the ball write: Give me a book too and we have a deal read: Ok, deal write:<choose>Output 1xbook 1xball Figure 2: Converting a crowd-sourced dialogue (left) into t wo training examples (right), from the perspective of each user. The perspectives differ on their inpu t goals, output choice, and in special tokens marking whether a statement was read or written. We train con ditional language models to predict the dialogue given the input, and additional models to predict t he output given the dialogue. available items and values for the two users). We held out a test set of 252 scenarios (526 dialogues). Holding out test scenarios means that models must generalise to new situations. 3 Likelihood Model We propose a simple but effective baseline model for the conversational agent, in which a sequenceto-sequence model is trained to produce the complete dialogue, conditioned on an agent’s input. 3.1 Data Representation Each dialogue is converted into two training examples, showing the complete conversation from the perspective of each agent. The examples differ on their input goals, output choice, and whether utterances were read or written. Training examples contain an input goal g, specifying the available items and their values, a

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

8

on their input goals, output choice, and whether utterances were read or written. Training examples contain an input goal g, specifying the available items and their values, a dialogue x, and an output decision ospecifying which items each agent will receive. Specifically, we represent gas a list of six integers corresponding to the count and value of each of the three item types. Dialogue xis a list of tokens x0..Tcontaining the turns of each agent interleaved with symbols marking whether a turn was written by the agent or their partner, terminating in a special token indicating one agent has marked that an agree-ment has been made. Output ois six integers describing how many of each of the three item types are assigned to each agent. See Figure 2. 3.2 Supervised Learning We train a sequence-to-sequence network to generate an agent’s perspective of the dialogue conditioned on the agent’s input goals (Figure 3a). The model uses 4 recurrent neural networks, implemented as GRUs (Cho et al., 2014): GRU w, GRUg, GRU− →o, and GRU← −o. The agent’s input goals gare encoded using GRUg. We refer to the final hidden state as hg.

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

9

The agent’s input goals gare encoded using GRUg. We refer to the final hidden state as hg. The model then predicts each token xtfrom left to right, conditioned on the previous tokens and hg. At each time step t, GRUwtakes as input the previous hidden state ht−1, previous token xt−1(embedded with a matrix E), and input encoding hg. Conditioning on the input at each time step helps the model learn dependencies between language and goals. ht=GRUw(ht−1,[Ext−1,hg]) (1) The token at each time step is predicted with a softmax, which uses weight tying with the embedding matrix E(Mao et al., 2015): pθ(xt|x0..t−1,g)∝exp(ETht) (2) Input Encoder Output Decoder write: Take one hat read: Ineed two write: deal . . . (a) Supervised TrainingInput Encoder Output Decoder write: Take one hat read: Ineed two write: deal . . . (b) Decoding, and Reinforcement Learning Figure 3: Our model: tokens are predicted conditioned on pre vious words and the input, then the output

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

10

(b) Decoding, and Reinforcement Learning Figure 3: Our model: tokens are predicted conditioned on pre vious words and the input, then the output is predicted using attention over the complete dialogue. In supervised training (3a), we train the model to predict the tokens of both agents. During decoding and reinforcement learning (3b) so me tokens are sampled from the model, but some are generated by the other ag ent and are only encoded by the model. Note that the model predicts both agent’s words, enabling its use as a forward model in Section 5. At the end of the dialogue, the agent outputs a set of tokens orepresenting the decision. We generate each output conditionally independently, using a separate classifier for each. The classifiers share bidirectional GRU oand attention mechanism (Bahdanau et al., 2014) over the dialogue, and additionally conditions on the input goals. h− →o t=GRU− →o(h− →o t−1,[Ext,ht]) (3) h← −o t=GRU← −o(h← −o t+1,[Ext,ht]) (4) ho t= [h← −o t,h− →o t] (5) ha

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

11

t+1,[Ext,ht]) (4) ho t= [h← −o t,h− →o t] (5) ha t=W[tanh(W′ho t)] (6) αt=exp(w·ha t)/summationtext t′exp(w·ha t′)(7) hs= tanh(Ws[hg,/summationdisplay tαtht]) (8) The output tokens are predicted using softmax: pθ(oi|x0..t,g)∝exp(Woihs) (9) The model is trained to minimize the negative log likelihood of the token sequence x0..Tconditioned on the input goals g, and of the outputs oconditioned on xandg. The two terms are weighted with a hyperparameter α. L(θ) =−/summationdisplay x,g/summationdisplay tlogpθ(xt|x0..t−1,g) /bracehtipupleft /bracehtipdownright/bracehtipdownleft /bracehtipupright Token prediction loss −α/summationdisplay x,g,o/summationdisplay

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

12

Token prediction loss −α/summationdisplay x,g,o/summationdisplay jlogpθ(oj|x0..T,g) /bracehtipupleft /bracehtipdownright/bracehtipdownleft /bracehtipupright Output choice prediction loss(10) Unlike the Neural Conversational Model (Vinyals and Le, 2015), our approach shares all parameters for reading and generating tokens.3.3 Decoding During decoding, the model must generate an output token xtconditioned on dialogue history x0..t−1and input goals g, by sampling from pθ: xt∼pθ(xt|x0..t−1,g) (11) If the model generates a special end-of-turn token, it then encodes a series of tokens output by the other agent, until its next turn (Figure 3b). The dialogue ends when either agent outputs a special end-of-dialogue token. The model then outputs a set of choices o. We choose each item independently, but enforce consistency by checking the solution is in a feasible set O: o∗= argmax o∈O/productdisplay ipθ(oi|x0..T,g) (12)

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

13

o∗= argmax o∈O/productdisplay ipθ(oi|x0..T,g) (12) In our task, a solution is feasible if each item is assigned to exactly one agent. The space of solutions is small enough to be tractably enumerated. 4 Goal-based Training Supervised learning aims to imitate the actions of human users, but does not explicitly attempt to maximise an agent’s goals. Instead, we explore pre-training with supervised learning, and then fine-tuning against the evaluation metric using reinforcement learning. Similar two-stage learning strategies have been used previously (e.g. Li et al. (2016); Das et al. (2017)). During reinforcement learning, an agent Aattempts to improve its parameters from conversations with another agent B. While the other agent Bcould be a human, in our experiments we used our fixed supervised model that was trained to imitate humans. The second model is fixed as we found that updating the parameters of both agents led to divergence from human language. In effect, read: You get one book and I’ll take everything else.write: Great deal, thanks! write: No way, I

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

14

read: You get one book and I’ll take everything else.write: Great deal, thanks! write: No way, I need all 3 hatsread: Ok, fineread: I’ll give you 2read: No problemread: Any time choose: 3xhatchoose: 2xhatchoose: 1xbookchoose: 1xbook 9611 Dialogue history Candidate responses Simulation of rest of dialogue Score Figure 4: Decoding through rollouts: The model first generat es a small set of candidate responses. For each candidate it simulates the future conversation by samp ling, and estimates the expected future reward by averaging the scores. The system outputs the candidate wi th the highest expected reward. agentAlearns to improve by simulating conversations with the help of a surrogate forward model. AgentAreads its goals gand then generates tokensx0..nby sampling from pθ. Whenxgenerates an end-of-turn marker, it then reads in tokens xn+1..mgenerated by agent B. These turns alternate until one agent emits a token ending the dialogue. Both agents then output a decision oand collect a reward from the environment (which will be 0 if they output different decisions). We denote

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

15

collect a reward from the environment (which will be 0 if they output different decisions). We denote the subset of tokens generated by AasXA(e.g. tokens with incoming arrows in Figure 3b). After a complete dialogue has been generated, we update agent A’s parameters based on the outcome of the negotiation. Let rAbe the score agent Aachieved in the completed dialogue, Tbe the length of the dialogue, γbe a discount factor that rewards actions at the end of the dialogue more strongly, and µbe a running average of completed dialogue rewards so far2. We define the future rewardRfor an action xt∈XAas follows: R(xt) =/summationdisplay xt∈XAγT−t(rA(o)−µ) (13) We then optimise the expected reward of each actionxt∈XA: LRL θ=Ext∼pθ(xt|x0..t−1,g)[R(xt)] (14) The gradient of LRL θis calculated as in REINFORCE (Williams, 1992): ∇θLRL θ=/summationdisplay

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

16

The gradient of LRL θis calculated as in REINFORCE (Williams, 1992): ∇θLRL θ=/summationdisplay xt∈XAExt[R(xt)∇θlog(pθ(xt|x0..t−1,g))] (15) 2As all rewards are non-negative, we instead re-scale them by subtracting the mean reward found during self play. Shift ing in this way can reduce the variance of our estimator.Algorithm 1 Dialogue Rollouts algorithm. 1:procedure ROLLOUT (x0..i,g) 2:u∗←∅ 3: forc∈{1..C}do⊲ Ccandidate moves 4:j←i 5: do ⊲Rollout to end of turn 6: j←j+1 7: xj∼pθ(xj|x0..j−1,g) 8: whilexk/∈{read: ,choose:} 9:u←xi+1..xj⊲uis candidate move 10: fors∈{1..S}do⊲Ssamples per move 11: k←j ⊲ Start rollout from end of u

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

17

10: fors∈{1..S}do⊲Ssamples per move 11: k←j ⊲ Start rollout from end of u 12: whilexk∝ne}ationslash=choose: do ⊲Rollout to end of dialogue 13: k←k+1 14: xk∼pθ(xk|x0..k−1,g) ⊲Calculate rollout output and reward 15: o←argmaxo′∈Op(o′|x0..k,g) 16: R(u)←R(u)+r(o)p(o′|x0..k,g) 17: ifR(u)> R(u∗)then 18: u∗←u 19: returnu∗⊲Return best move 5 Goal-based Decoding Likelihood-based decoding ( §3.3) may not be optimal. For instance, an agent may be choosing between accepting an offer, or making a counter offer. The former will often have a higher likelihood under our model, as there are fewer ways to agree than to make another offer, but the latter may lead

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

18

under our model, as there are fewer ways to agree than to make another offer, but the latter may lead to a better outcome. Goal-based decoding also allows more complex dialogue strategies. For example, a deceptive utterance is likely to have a low model score (as users were generally honest in the supervised data), but may achieve high reward. We instead explore decoding by maximising expected reward. We achieve this by using pθas a forward model for the complete dialogue, and then deterministically computing the reward. Rewards for an utterance are averaged over samples to calculate expected future reward (Figure 4). We use a two stage process: First, we generateccandidate utterances U=u0..c, representing possible complete turns that the agent could make, which are generated by sampling from pθ until the end-of-turn token is reached. Let x0..n−1 be current dialogue history. We then calculate the expected reward R(u)of candidate utterance u=xn,n+kby repeatedly sampling xn+k+1,T frompθ, then choosing the best output ousing Equation 12, and finally deterministically computing the reward r(o). The reward is scaled by the probability of the output given the dialogue, because if the agents select different outputs then

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

19

probability of the output given the dialogue, because if the agents select different outputs then they both receive 0 reward. R(xn..n+k) =Ex(n+k+1..T;o)∼pθ[r(o)pθ(o|x0..T)] (16) We then return the utterance maximizing R. u∗= argmax u∈UR(u) (17) We use 5 rollouts for each of 10 candidate turns. 6 Experiments 6.1 Training Details We implement our models using PyTorch. All hyper-parameters were chosen on a development dataset. The input tokens are embedded into a 64-dimensional space, while the dialogue tokens are embedded with 256-dimensional embeddings (with no pre-training). The input GRUghas a hidden layer of size 64 and the dialogue GRUw is of size 128. The output GRU − →oand GRU← −o both have a hidden state of size 256, the size of hsis 256 as well. During supervised training, we optimise using stochastic gradient descent with a minibatch size of 16, an initial learning rate of 1.0, Nesterov momentum with µ=0.1 (Nesterov,

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

20

minibatch size of 16, an initial learning rate of 1.0, Nesterov momentum with µ=0.1 (Nesterov, 1983), and clipping gradients whose L2norm exceeds 0.5. We train the model for 30 epochs and pick the snapshot of the model with the best validation perplexity. We then annealed the learning rate by a factor of 5 each epoch. We weight the terms in the loss function (Equation 10) using α=0.5. We do not train against output decisions where humans selected different agreements. Tokens occurring fewer than 20 times are replaced with an ‘unknown’ token.During reinforcement learning, we use a learning rate of 0.1, clip gradients above 1.0, and use a discount factor of γ=0.95. After every 4 reinforcement learning updates, we make a supervised update with mini-batch size 16 and learning rate 0.5, and we clip gradients at 1.0. We used 4086 simulated conversations. When sampling words from pθ, we reduce the variance by doubling the values of logits (i.e. using temperature of 0.5). 6.2 Comparison Systems We compare the performance of the following:

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616

1706.05125

21

variance by doubling the values of logits (i.e. using temperature of 0.5). 6.2 Comparison Systems We compare the performance of the following: LIKELIHOOD uses supervised training and decoding (§3), RLis fine-tuned with goal-based selfplay (§4), ROLLOUTS uses supervised training combined with goal-based decoding using rollouts (§5), and RL+ROLLOUTS uses rollouts with a base model trained with reinforcement learning. 6.3 Intrinsic Evaluation For development, we use measured the perplexity of user generated utterances, conditioned on the input and previous dialogue. Results are shown in Table 3, and show that the simple LIKELIHOOD model produces the most human-like responses, and the alternative training and decoding strategies cause a divergence from human language. Note however, that this divergence may not necessarily correspond to lower quality language—it may also indicate different strategic decisions about what to say. Results in §6.4 show all models could converse with humans. 6.4 End-to-End Evaluation We measure end-to-end performance in dialogues both with the likelihood-based agent and with humans on Mechanical Turk, on held out scenarios. Humans were told that they were interacting

1706.05125

Deal or No Deal? End-to-End Learning for Negotiation Dialogues

Much of human dialogue occurs in semi-cooperative settings, where agents with different goals attempt to agree on common decisions. Negotiations require complex communication and reasoning skills, but success is easy to measure, making this an interesting task for AI. We gather a large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other's reward functions must reach an agreement (or a deal) via natural language dialogue. For the first time, we show it is possible to train end-to-end models for negotiation, which must learn both linguistic and reasoning skills with no annotated dialogue states. We also introduce dialogue rollouts, in which the model plans ahead by simulating possible complete continuations of the conversation, and find that this technique dramatically improves performance. Our code and dataset are publicly available (https://github.com/facebookresearch/end-to-end-negotiator).

http://arxiv.org/pdf/1706.05125

cs.AI

20170616

20170616