Findings of ACL

2023

QUD dependency parsing

Discourse Analysis via Questions and Answers: Parsing Dependency Structures of Questions Under Discussion

Wei-Jen Ko, Yating Wu, Cutter Dalton, Dananjay Srinivas, Greg Durrett, Junyi Jessy Li

Findings of the Association for Computational Linguistics (ACL) · 2023

paperCodeBibTeX

Overview

The parser treats each sentence as an answer to a question raised earlier in the document. It predicts an earlier anchor and generates the question that connects them. Training uses DCQA, the crowdsourced dataset introduced by Ko et al. (2022).

We evaluate the questions with human readers, compare the resulting dependencies with Rhetorical Structure Theory trees, and use the parses to examine how simplified news articles add and reorganize information.

Paper

From an aid package to a funding gap

Sentence 4 answers a question raised by sentence 3. That question labels the dependency between them.

Sentence 3 · Anchor

The House approves a smaller relief package than California had requested.

Question on the dependency edge

What would the shortfall be?

Sentence 4 · Answer

California still faces a funding gap of roughly $2 billion to $4 billion.

Source

Figure 1(a), page 2. Annotated DCQA example in Figure 1(a). Question quoted; anchor and answer paraphrased.

Run the parser

The Colab notebook runs the pipeline from numbered sentences to questions and their anchors.

  1. Prepare the text

    Put one sentence per line in inputa/, with a sentence number, a tab, and the sentence.

  2. Find an anchor

    For each answer sentence, predict an earlier sentence that raises the question.

  3. Generate questions

    Generate candidate questions from the predicted anchor, then rank them.

  4. Read the parse

    Each question labels an edge between the answer sentence and its anchor.

The original setup uses the bundled Transformers code and separate model downloads. Follow the repository’s environment instructions.

Source

Input format and pipeline

Parser code
Anchor prediction, question generation, and reranking.
Models and setup
Checkpoints and the original environment.
Colab
Run the pipeline in a notebook.

Citation

Download BibTeX
View BibTeX
@inproceedings{qudparsing2023,
  title={{Discourse Analysis via Questions and Answers: Parsing Dependency Structures of Questions Under Discussion}},
  author={Ko, Wei-Jen and Wu, Yating and Dalton, Cutter and Srinivas, Dananjay and Durrett, Greg and Li, Junyi Jessy},
  year={2023},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2023},
  doi={10.18653/v1/2023.findings-acl.710},
  pages={11181--11195},
  url={https://aclanthology.org/2023.findings-acl.710/},
  publisher={Association for Computational Linguistics}
}