sergioburdisso/dialog2flow-dataset

Dialog2Flow Training Corpus

This page hosts the dataset introduced in the paper "Dialog2Flow: Pre-training Soft-Contrastive Action-Driven Sentence Embeddings for Automatic Dialog Flow Extraction" published in the EMNLP 2024 main conference. Here we are not only making available the dataset but also each one of the 20 (standardized) task-oriented dialogue datasets used to build it.

The corpus consists of 3.4 million utterances/sentences annotated with dialog act and slot labels across 52 different domains. Domain names and dialog act labels were manually standardized across the 20 datasets.

Load Training Datasets

From this corpus, in the paper, 3 datasets were created for training the sentence encoders, one for the single target (D2F_single) training containing the subset with both dialog act and slots annotation; and other two for the joint target (DFD_joint), one containing the subset with dialog acts and another with slots only. To use them, you can use one of the following names, respectively:

1. "dialog-acts+slots": (utterance, action label) pairs.
2. "dialog-acts": (utterance, dialog act label) pairs.
3. "slots": (utterance, slots label) pairs.

For instance, below is one example to load the "dialog-acts+slots" dataset:

from datasets import load_dataset

dataset = load_dataset('sergioburdisso/dialog2flow-dataset', 'dialog-acts+slots', trust_remote_code=True)

print(dataset)

Output:

DatasetDict({
    train: Dataset({
        features: ['utterance', 'label'],
        num_rows: 1577184
    })
    validation: Dataset({
        features: ['utterance', 'label'],
        num_rows: 4695
    })
})

Load (Individual) Task-Oriented Dialog Datasets

We also provide access to each one of the 20 task-oriented dialogue datasets standardizing annotation and format used to build the corpus. To load each dataset we can simply use its name as given in the following table with information about License and number of dialogues in each dataset:

Dataset Name	Train	Validation	Test	Total	License
ABCD	8034	1004	1004	10042	MIT License
BiTOD	2952	295	442	3689	Apache License 2.0
Disambiguation	8433	999	1000	10432	MiT License
DSTC2-Clean	1612	506	1117	3235	GNU GENERAL PUBLIC LICENSE Version 3, 29 June 2007
FRAMES	1329	-	40	1369	GNU GENERAL PUBLIC LICENSE Version 3, 29 June 2007
GECOR	676	-	-	676	CC BY 4.0
HDSA-Dialog	8438	1000	1000	10438	MIT License
KETOD	4247	545	532	5324	MiT License
MS-DC	10000	-	-	10000	MICROSOFT RESEARCH LICENSE TERMS
MulDoGO	59939	1150	2319	63408	Community Data License Agreement – Permissive – Version 1.0
MultiWOZ_2.1	8434	999	1000	10433	MiT License
MULTIWOZ2_2	8437	1000	1000	10437	Mit License
SGD	16142	2482	4201	22825	CC BY-SA 4.0
SimJointGEN	100000	10000	10000	120000	No license
SimJointMovie	384	120	264	768	No license
SimJointRestaurant	1116	349	775	2240	No license
Taskmaster1	6170	769	769	7708	Attribution 4.0 International (CC BY 4.0)
Taskmaster2	17304	-	-	17304	Creative Commons Attribution 4.0 License (CC BY 4.0)
Taskmaster3	22724	17019	17903	57646	Creative Commons Attribution 4.0 License (CC BY 4.0)
WOZ2_0	600	200	400	1200	Apache License 2.0

For instance, below is one example to load the "WOZ2_0" dataset:

from datasets import load_dataset

dataset = load_dataset('sergioburdisso/dialog2flow-dataset', 'WOZ2_0', trust_remote_code=True)

print(dataset)

Output:

DatasetDict({
    test: Dataset({
        features: ['dialog'],
        num_rows: 400
    })
    train: Dataset({
        features: ['dialog'],
        num_rows: 600
    })
    validation: Dataset({
        features: ['dialog'],
        num_rows: 200
    })
})

Note that, unlike previous datasets that contained utterance-label pairs, these individual datasets consist of only one feature "dialog" since their a collection of dialogs (not utterances). Each dialog in turn has the JSON structure described in Appendix A of the paper. For instance, let's get the first dialog of the train split:

print(dataset["train"][0]["dialog"])

Output:

[
   {
      "speaker":"user",
      "text":"Are there any eritrean restaurants in town?",
      "domains":[
         "restaurant"
      ],
      "labels":{
         "dialog_acts":{
            "acts":[
               "inform"
            ],
            "main_acts":[
               "inform"
            ],
            "original_acts":[
               "inform"
            ]
         },
         "slots":[
            "food"
         ],
         "intents":"None"
      }
   },
    ...
   {
      "speaker":"system",
      "text":"There is a wide variety of Chinese restaurants, do you have an area preference or a price preference to narrow it down?",
      "domains":[
         "restaurant"
      ],
      "labels":{
         "dialog_acts":{
            "acts":[
               "request"
            ],
            "main_acts":[
               "request"
            ],
            "original_acts":[
               "request"
            ]
         },
         "slots":[
            "area"
         ],
         "intents":"None"
      }
   },
  ...
]

Corpus Details

Stats

• Utterances: 3.4M
• Domains: 52
• Dialogs: 369,174
• Labels:
  • Dialog acts: 18
  • Slots: 524
  • Actions (dialog act + slots): 3,982

Full List of Dialog Acts

List of the final 18 dialog act labels along with their proportion in the corpus:

inform (64.66%) · request (12.62%) · offer (6.62%) · inform_success (3.07%) · good_bye (2.67%) · agreement (2.45%) · thank_you (2.25%) · confirm (2.10%) · disagreement (1.60%) · request_more (1.06%) · request_alternative (0.90%) · recommendation (0.70%) · inform_failure (0.64%) · greeting (0.31%) · confirm_answer (0.18%) · confirm_question (0.17%) · request_update (0.02%) · request_compare (0.01%)

Full List of Domains

List of the final 52 domain names along with their proportion in the corpus:

movie (32.98%) · restaurant (13.48%) · hotel (10.15%) · train (4.52%) · flight (4.30%) · event (3.56%) · attraction (3.50%) · service (2.44%) · bus (2.28%) · taxi (2.21%) · rentalcars (2.20%) · travel (2.16%) · music (1.81%) · medium (1.66%) · ridesharing (1.30%) · booking (1.21%) · home (1.01%) · finance (0.79%) · airline (0.69%) · calendar (0.69%) · fastfood (0.68%) · insurance (0.61%) · weather (0.58%) · bank (0.47%) · hkmtr (0.36%) · mlb (0.35%) · ml (0.31%) · food (0.30%) · epl (0.30%) · pizza (0.25%) · coffee (0.24%) · uber (0.24%) · software (0.23%) · auto (0.21%) · nba (0.20%) · product_defect (0.17%) · shipping_issue (0.16%) · alarm (0.13%) · order_issue (0.13%) · messaging (0.13%) · hospital (0.11%) · subscription_inquiry (0.11%) · account_access (0.11%) · payment (0.10%) · purchase_dispute (0.10%) · nfl (0.09%) · chat (0.08%) · police (0.07%) · single_item_query (0.06%) · storewide_query (0.06%) · troubleshoot_site (0.06%) · manage_account (0.06%)

More details about the corpus can be found in Section 4 and Appendix A of the original paper.

Citation

If you found the paper and/or this repository useful, please consider citing our work :)

EMNLP paper: here.

@inproceedings{burdisso-etal-2024-dialog2flow,
    title = "{D}ialog2{F}low: Pre-training Soft-Contrastive Action-Driven Sentence Embeddings for Automatic Dialog Flow Extraction",
    author = "Burdisso, Sergio  and
      Madikeri, Srikanth  and
      Motlicek, Petr",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.emnlp-main.310",
    pages = "5421--5440",
}

License

Individual datasets were originally loaded from DialogStudio and therefore, this project follows their licensing structure. For detailed licensing information, please refer to the specific licenses accompanying the datasets provided in the table above.

All extra content purely authored by us is released under the MIT license:

Copyright (c) 2024 Idiap Research Institute.

MIT License.