A Distance-Aware Multi-Task Framework for Conversational Discourse Parsing

Yaxin Fan, Peifeng Li, Fang Kong, Qiaoming ZhuCOLING 2022 · OpenReview

Abstract

Conversational discourse parsing aims to construct an implicit utterance dependency tree to reflect the turn-taking in a multi-party conversation. Existing works are generally divided into two lines: graph-based and transition-based paradigms, which perform well for short-distance and long-distance dependency links, respectively. However, there is no study to consider the advantages of both paradigms to facilitate conversational discourse parsing. As a result, we propose a distance-aware multi-task framework DAMT that incorporates the strengths of transition-based paradigm to facilitate the graph-based paradigm from the encoding and decoding process. To promote multi-task learning on two paradigms, we first introduce an Encoding Interactive Module (EIM) to enhance the flow of semantic information between both two paradigms during the encoding step. And then we apply a Distance-Aware Graph Convolutional Network (DAGCN) in the decoding process, which can incorporate the different-distance dependency links predicted by the transition-based paradigm to facilitate the decoding of the graph-based paradigm. The experimental results on the datasets STAC and Molweni show that our method can significantly improve the performance of the SOTA graph-based paradigm on long-distance dependency links.

Introduction

In multi-party conversations, identifying the structure of replies—known as Conversational Discourse Parsing—is a complex task. It involves building a dependency tree where nodes represent utterances and edges denote relationships like "Question" or "Acknowledgement." Current approaches typically follow one of two paradigms:

  1. Graph-based models: These optimize the global structure by evaluating all possible utterance pairs. While effective for adjacent messages (distance 1), they struggle with long-distance dependencies.
  2. Transition-based models: These build the tree incrementally. By maintaining a conversational state, they are better at capturing long-distance links but often lack a global perspective.

Research indicates a "distance gap": graph models dominate at short distances, while transition models excel as distance increases. The DAMT framework seeks to bridge this gap by combining the strengths of both paradigms through explicit multi-task interaction.

The DAMT Framework

The proposed Distance-Aware Multi-Task (DAMT) framework integrates graph-based and transition-based paradigms through two primary innovations:

Encoding Interactive Module (EIM)

Unlike standard multi-task learning that relies on simple shared backbones, the EIM facilitates explicit interaction. It utilizes Self-Attention (SA) and Unidirectional Cross-Attention (UCA) to allow the graph-based and transition-based views to exchange semantic information. This ensures that the initial representations are informed by the strengths of both architectures.

Distance-Aware Graph Convolutional Network (DAGCN)

The DAGCN acts as a decoding bridge. The transition-based module first predicts a preliminary structure, which is then processed by the DAGCN. By using trainable weights based on the distance between utterances, the graph-based decoder can leverage the transition-based model's expertise for long-distance links while maintaining global optimization.

The DAMT framework architecture, illustrating the interaction between the Encoding Interactive Module (EIM) and the Distance-Aware Graph Convolutional Network (DAGCN).

Experimental Evaluation

The model was evaluated on two benchmarks for Segmented Discourse Representation Theory (SDRT):

  • STAC: Strategic dialogues from an online version of the game The Settlers of Catan.
  • Molweni: A large-scale dataset based on the Ubuntu Chat corpus, focusing on technical support.

Both datasets feature complex "crossing dependencies" where multiple conversation threads occur simultaneously, making long-distance link detection critical.

Key Results

DAMT achieved state-of-the-art (SOTA) performance on both benchmarks. On the Molweni dataset, it reached an F1 score of 82.25% for link prediction, a significant improvement over previous models.

The framework's most significant contribution is its mastery of long-distance links (spanning 4 or more utterances). While baseline models typically see a sharp decline in accuracy as distance increases, DAMT maintains a significantly higher performance floor, effectively closing the distance gap identified in previous research.

Accuracy comparison across dependency link distances, demonstrating DAMT's superior performance on long-distance links compared to SSAM and PN baselines.

Conclusion

The DAMT framework demonstrates that explicit multi-task interaction between graph-based and transition-based paradigms significantly enhances conversational discourse parsing. Future research will focus on improving relation type classification—distinguishing between types like "Comment" and "Result"—and exploring more granular distance embeddings to further refine performance in extremely long-running conversation threads.

Summary updated

Abstract

Conversational discourse parsing aims to construct an implicit utterance dependency tree to reflect the turn-taking in a multi-party conversation. Existing works are generally divided into two lines: graph-based and transition-based paradigms, which perform well for short-distance and long-distance dependency links, respectively. However, there is no study to consider the advantages of both paradigms to facilitate conversational discourse parsing. As a result, we propose a distance-aware multi-task framework DAMT that incorporates the strengths of transition-based paradigm to facilitate the graph-based paradigm from the encoding and decoding process. To promote multi-task learning on two paradigms, we first introduce an Encoding Interactive Module (EIM) to enhance the flow of semantic information between both two paradigms during the encoding step. And then we apply a Distance-Aware Graph Convolutional Network (DAGCN) in the decoding process, which can incorporate the different-distance dependency links predicted by the transition-based paradigm to facilitate the decoding of the graph-based paradigm. The experimental results on the datasets STAC and Molweni show that our method can significantly improve the performance of the SOTA graph-based paradigm on long-distance dependency links.
BibTeX
@inproceedings{fan2022distance,
  author = {Yaxin Fan and Peifeng Li and Fang Kong and Qiaoming Zhu},
  title = {{A Distance-Aware Multi-Task Framework for Conversational Discourse Parsing}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2022},
  url = {https://openreview.net/forum?id=JRzHrfwvWy},
}
Figure