Exophoric Pronoun Resolution in Dialogues with Topic Regularization

Introduction

This is the data and the source code for the EMNLP 2021 paper "Exophoric Pronoun Resolution in Dialogues with Topic Regularization."[PAPER][VIDEO][SLIDES][POSTER]

Abstract

Resolving pronouns to their referents has long been studied as a fundamental natural language understanding problem. Previous works on pronoun coreference resolution (PCR) mostly focus on resolving pronouns to mentions in text while ignoring the exophoric scenario. Exophoric pronouns are common in daily communications, where speakers may directly use pronouns to refer to some objects present in the environment without introducing the objects first. Although such objects are not mentioned in the dialogue text, they can often be disambiguated by the general topics of the dialogue. Motivated by this, we propose to jointly leverage the local context and global topics of dialogues to solve the out-of-text PCR problem. Extensive experiments demonstrate the effectiveness of adding topic regularization for resolving exophoric pronouns.

An example of resolving exophoric pronouns in daily dialogues with and without the help of dialogue topics.

Definition of Exophora

In daily dialogues, speakers may often use exophoric pronouns to refer to objects in the environment that all speakers and listeners are aware of without introducing them in the first place. Such use of pronouns is called exophora.

Task

• Dataset: VisPro
  • It annotates in-text and out-of-text reference for pronouns in visual dialogues.
  • Even without the image, human can resolve 96% of exophoric pronouns in VisPro with only dialogue texts.
• Task definition:
  • Given a pronoun in a dialogue text, find its in-text reference in the dialogue as well as out-of-text reference in an object pool (containing 384 common objects in VisPro).
  • If a pronoun has any in-text reference, such as "them", it is endophoric.
  • If a pronoun only has out-of-text reference, such as “he”, it is exophoric.

• How do dialogue topics help？
  • Objects more relevant to the topics are more likely to appear in the environment, such as "baseball player" under this dialogue topic.
  • Our model assigns higher scores to out-of-text objects which are more relevant to the dialogue topics.

Usage of Model

Getting Started

• Install python 3.6+ and the following requirements: pip install -r requirements.txt.
• Build custom kernels by running setup_all.sh.
  • There are 3 platform-dependent ways to build custom TensorFlow kernels. Please comment/uncomment the appropriate lines in the script.

Preprocessing data

The training, validation, and test split of VisPro contains 4,000, 500, and 500 dialogs, respectively.

The data preprocessing is described in Sec. 5.4 of the paper. To enable straight reproduction, we upload the preprocessed data files in the data folder.

Evaluation only

• Run python get_lda.py to generate LDA topic labels.
• Download model checkpoints of topic_bert_base from checkpoint to the logs/topic_bert_base directory.
• Prediction: python predict.py topic_bert_base. Outputs are saved in the output directory.
• Evaluation: python evaluate.py topic_bert_base.

Training from scratch

• Download pretrained coreference BERT models with ./download_pretrained.sh <model_name> (bert_base, spanbert_base, bert_large, spanbert_large). This downloads BERT/SpanBERT models finetuned on OntoNotes as the baseline models to models folder.
• Run python get_lda.py to generate LDA topic labels.
• Experiment configurations are found in experiments.conf.
• Choose an experiment that you would like to run, e.g. topic_bert_base.
• For training and prediction, set the GPU environment variable, which the code treats as shorthand for CUDA_VISIBLE_DEVICES.
• Training: python train.py <experiment>
• Results are stored in the logs directory and can be viewed via TensorBoard.
• Prediction: python predict.py <experiment>. Outputs are saved in the output directory.
• Evaluation: python evaluate.py <experiment>.

Acknowledgment

We built the training framework based on the original BERT and SpanBERT for Coreference Resolution.

Others

If you have questions about the data or the code, you are welcome to open an issue or send me an email, I will respond to that as soon as possible.

The readers are welcome to star/fork this repository and use it to train your own model, reproduce our experiment, and follow our future work. Please kindly cite our paper:

@inproceedings{DBLP:conf/emnlp/YuZSZX021,
  author    = {Xintong Yu and
               Hongming Zhang and
               Yangqiu Song and
               Changshui Zhang and
               Kun Xu and
               Dong Yu},
  title     = {Exophoric Pronoun Resolution in Dialogues with Topic Regularization},
  booktitle = {Proceedings of the 2021 Conference on Empirical Methods in Natural
               Language Processing, {EMNLP} 2021, Virtual Event / Punta Cana, Dominican
               Republic, 7-11 November, 2021},
  pages     = {3832--3845},
  publisher = {Association for Computational Linguistics},
  year      = {2021},
  url       = {https://aclanthology.org/2021.emnlp-main.311},
}