roberta-temporal-predictor

A RoBERTa-base model that is fine-tuned on the The New York Times Annotated Corpus to predict temporal precedence of two events. This is used as the ``temporality prediction'' component in our ROCK framework for reasoning about commonsense causality. See our paper for more details.

Usage

You can directly use this model for filling-mask tasks, as shown in the example widget. However, for better temporal inference, it is recommended to symmetrize the outputs as $P(E_1 \prec E_2) = \frac{1}{2} (f(E_1,E_2) + f(E_2,E_1))$ where f(E_1,E_2) denotes the predicted probability for E_1 to occur preceding E_2. For simplicity, we implement the following TempPredictor class that incorporate this symmetrization automatically. Below is an example usage for the TempPredictor class:

from transformers import (RobertaForMaskedLM, RobertaTokenizer)
from src.temp_predictor import TempPredictor

TORCH_DEV = "cuda:0" # change as needed

tp_roberta_ft = src.TempPredictor(
    model=RobertaForMaskedLM.from_pretrained("CogComp/roberta-temporal-predictor"),
    tokenizer=RobertaTokenizer.from_pretrained("CogComp/roberta-temporal-predictor"),
    device=TORCH_DEV
)

E1 = "The man turned on the faucet."
E2 = "Water flows out."
t12 = tp_roberta_ft(E1, E2, top_k=5)
print(f"P('{E1}' before '{E2}'): {t12}")

BibTeX entry and citation info

@misc{zhang2022causal,
      title={Causal Inference Principles for Reasoning about Commonsense Causality}, 
      author={Jiayao Zhang and Hongming Zhang and Dan Roth and Weijie J. Su},
      year={2022},
      eprint={2202.00436},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}