/DDP_parsing

One implementation of the paper "Improving Multi-Party Dialogue Discourse Parsing via Domain Integration".

Primary LanguagePython

Introduction

  • One implementation of the paper "Improving Multi-Party Dialogue Discourse Parsing via Domain Integration".
  • The parsing follows the Segmented Discourse Representation Theory (SDRT) scheme.
  • Users can apply it to parse the input dialogue text, and get dependency-parsing structure and relation prediction.
  • This repo and the pre-trained model are only for research use.

Package Requirements

  1. pytorch==1.7.1
  2. transformers==4.8.2

Supported Languages

We trained and evaluated the model with two English dialogue discourse parsing corpora: STAC and Molweni.

Data Format

  • [Inference Input] InputSentence: The input dialogue content with the #ROOT# head token and utterance split \<utterance\>. The raw text will be tokenized and encoded by the roberta-base language backbone.

    • Raw Sequence Example:
      #ROOT# <utterance> A: Hi Tom, are you busy tomorrow’s afternoon? <utterance> B: I’m pretty sure I am. What’s up? <utterance> A: Can you go with me to the animal shelter?. <utterance> B: What do you want to do? ... ... <utterance> B: I wonder what he'll name it. <utterance> A: He said he’d name it after his hamster –Lemmy- he's a great Motorhead fan :-)

  • [Inference Output] all_sample_utter_level: The list of utterances, each utterance is taken as one EDU in the parsing process.

  • [Inference Output] all_predict_link: Predictions of the discourse links, in the form of [from_utterance_id, to_utterance_id], where the first node is the #ROOT# token.

    • Output Example: [[1, 0], [2, 1], [3, 2], [4, 3], [5, 3], [6, 5], [7, 6], [8, 7], [9, 8], [10, 8], [11, 10], [12, 10], [13, 12], [14, 13], [15, 14], [16, 14], [17, 16]]

  • [Inference Output] all_predict_relation: Predictions of the discourse relation classification result. See relation type mapping is defined in the code.

    • Output Example: [16, 12, 12, 12, 3, 0, 5, 0, 0, 0, 9, 7, 3, 0, 10, 7, 3]

How to use it for parsing

  • Put the dialogue content to the file ./data/text_for_inference.txt. See sample data in the file.
  • Run the script main_infer.py to obtain the dialogue discourse parsing result. See the script for detailed model output.
  • We recommend users to run the parser on a GPU-equipped environment.

Experimental Result (F1 Score)

Here is the result of the joint training on data from two domains (Molweni+STAC):

Training on Joint Data     Link     Link+Relation
Testing on STAC 75.5 57.2
Testing on Molweni 80.2 56.9

For cross-domain experimental results, interested readers can refer to the paper.

Citation

If the work is helpful, please cite following papers in your publications, reports, and slides.

@inproceedings{liu-chen-2021-improving,
    title = "Improving Multi-Party Dialogue Discourse Parsing via Domain Integration",
    author = "Liu, Zhengyuan and Chen, Nancy",
    booktitle = "Proceedings of the 2nd Workshop on Computational Approaches to Discourse",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic and Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.codi-main.11",
    pages = "122--127",
}