This repo contains the code and data of the following paper:

A Dataset for Document Grounded Conversations. Kangyan Zhou, Shrimai Prabhumoye, Alan W Black. EMNLP 2018. arXiv

This paper introduces a document grounded dataset for text conversations. We define "Document Grounded Conversations" as conversations that are about the contents of a specified document. In this dataset the specified documents were Wikipedia articles about popular movies. The dataset contains 4112 conversations with an average of 21.43 turns per conversation. This positions this dataset to not only provide a relevant chat history while generating responses but also provide a source of information that the models could use. We describe two neural architectures that provide benchmark performance on the task of generating the next response. We also evaluate our models for engagement and fluency, and find that the information from the document helps in generating more engaging and fluent responses.

• Total number of conversations(having at least 1 turns): 4112

• Total number of distinct users participating the conversations: 4929

• Rating

  • The whole data is splited into 3 parts, from rate 1(lower) to 3(better).
  • Rate 1: Low BLEU score(< 0.1), or the number of turns lower than 10
  • Rate 2: All the conversations that do not fit in rate 1 or 3
  • Rate 3: Conversations with more than 12 turns, with BLEU score over one std

• Overall stats based on ratings

  Stats\Ratings	Rate 1	Rate 2	Rate 3
  Total Number of conversations	1443	2142	527
  Total Number of utterances	28536	80104	21360
  Mean/Std utterances per conversations	19.77 ± 13.68	35.39 ± 8.48	40.53 ± 12.92
  Mean/Std length of utterances	7.51 ± 50.19	10.56 ± 8.51	16.57 ± 15.23

• All the conversation data files are in the 'Conversations' folder, with train/validation/test split.
• Each conversation file is a json file. Fields in the file includes
  • 'date': the time the file is created.
  • 'history': a list of json object containing the whole chat history. Fields in the json objects in the list includes
    • 'text': the text of each utterance.
    • 'utcTimestamp': the server utc timestamp of this utterance.
    • 'uid': the user id of this utterance.
    • 'docIdx': the current section index of the wiki document when the utterance is said. There are in total 4 sections for each document.
  • 'rating': A number from 1 or 2 or 3. A larger number means the quality of the conversation is better.
  • 'uid1LogInTime', 'uid1LogOutTime', 'uid2LogInTime', 'uid2LogOutTime': The log in and log out time of user1 and user2. Any of the field could be missing due to some technical issue.
  • 'uid1response', 'uid2response': a json object contains the status and response of user after finishing the conversation. Fields in the object includes
    • 'type': should be one of ['finish', 'abandon','abandonWithouAnsweringFeedbackQuestion']. 'finish' means the user successfully finishes the conversation, either by completing 12 or 15 turns or in the way that the other user leaves the conversation first. 'abandon' means the user abandons the conversation in the middle, but entering the feedback page. 'abandonWithouAnsweringFeedbackQuestion' means the user just disconnects or closes the web page without providing the feedback.
    • 'response': the answer to the post-conversation questions. The worker can choose multiple of them. The options presented to the user are as follows:
      • For type 'finish'
        • 1: The conversation is understandable.
        • 2: The other user is actively responding me.
        • 3: The conversation goes smoothly.
      • For type 'abandon'
        • 1: The other user is too rude.
        • 2: I don't know how to proceed with the conversation.
        • 3: The other user is not responding to me.
      • For users given the document
        • 4: I have watched the movie before.
        • 5: I have not watched the movie before.
      • For the users without the document
        • 4: I will watch the movie after the other user's introduction.
        • 5: I will not watch the movie after the other user's introduction.
    • 'feedback': the user-entered feedback after the conversation.
  • 'user1_id', 'user2_id': the generated user id of user1 and user2.
  • 'status': 0 means the conversation finished unexceptedly (at least one user abandoned the conversations). 1 means the conversation finished correctly.
  • 'whoSawDoc': Should be one of ['user1'], ['user2'], ['user1', 'user2']. Indicating which user read the document.
  • 'wikiDocumentIdx': the index of the wiki document.

• All the wiki document data files are in the 'WikiData' folder
• An example of the data is included in the repository.
• Each file is a json file. Fields in the file includes
  • '0', '1', '2', '3': the introductions and key scenes in the movie. The number corresponds to the 'docIdx' field in the 'history' in the conversation data file. '0' corresponds to the introductions, and '1', '2', and '3' are the index of key scenes.
  • The meaning of each field in '0':
    • 'cast': the cast of the movie, possibly with a brief introduction of the role.
    • 'critical_response': a list of selected critical responses of the movie. Some of them are positive reviews, while some are negatives.
    • 'director': the director of the movie.
    • 'genre': the genre of the movie.
    • 'introduction': the introduction of the movie, with some basic information.
    • 'movieName': the name of the movie.
    • 'rating': a list of ratings from major review websites for this movie. The ratings usually includes some of the Rotten Tomatoes, IMDB, CinemaScore, or Metacritic.
    • 'year': the year the movie came out.
  • 'wikiDocumentIdx': the index of this wiki document, corresponding to the 'wikiDocumentIdx' field in conversation data. You can use this key to map the this document to the conversation file.

@inproceedings{cmu_dog_emnlp18,
title={A Dataset for Document Grounded Conversations},
author={Zhou, Kangyan and Prabhumoye, Shrimai and Black, Alan W},
year={2018},
booktitle={Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing}
}