DADgraph: A Discourse-aware Dialogue Graph Neural Network for Multiparty Dialogue Machine Reading Comprehension

Jiaqi Li, Ming Liu, Zihao Zheng, Heng Zhang, Bing Qin, Min-Yen Kan, Ting Liu

Abstract

Multiparty dialogues in platforms like Slack or IRC are often non-linear and entangled, posing significant challenges for traditional Machine Reading Comprehension (MRC) models designed for linear prose. This paper introduces DADgraph, a Discourse-aware Dialogue Graph Neural Network that explicitly models the discourse structure of conversations. By representing utterances as nodes and functional relationships (e.g., Question-Answer, Elaboration) as edges, DADgraph enables efficient information flow across long-distance dependencies. Evaluated on the Molweni dataset, DADgraph outperforms strong baselines in both accuracy and computational efficiency.

Introduction

Traditional MRC models, such as those trained on SQuAD, typically assume a linear, logical progression of text. However, multiparty dialogues are inherently "entangled." For instance, an utterance might respond to a comment made several turns prior, while intervening messages belong to a separate side-conversation. Standard NLP models treating text as a flat sequence struggle to navigate these non-linear dependencies because they lack an understanding of the discourse structure—the web of functional relationships connecting disparate parts of a conversation.

Discourse graph example showing non-linear relations between utterances.

Methods: The DADgraph Architecture

The DADgraph framework employs a three-stage pipeline to process dialogue:

  1. Sequential Context Encoding: Initial representations of each utterance are generated using a pre-trained BERT model, followed by a Bi-directional GRU (Bi-GRU) to capture the basic temporal flow.
  2. Discourse Graph Modeling: This core component constructs a directed graph where nodes represent utterances and edges represent one of 16 specific discourse relations (e.g., Acknowledgement, Clarification, Elaboration, Question-Answer Pair). A Graph Convolutional Network (GCN) processes this graph, allowing information to "jump" directly between related utterances, bypassing irrelevant noise.
  3. MRC Module: Graph-aware representations are combined with word-level embeddings. An attention mechanism facilitates interaction between the question and the dialogue to predict the answer span.
The three-stage DADgraph architecture: Sequential Encoding, Discourse Graph Modeling, and MRC.

Datasets and Evaluation

The primary evaluation was conducted on the Molweni corpus, which contains over 10,000 dialogues and 30,000 questions derived from the Ubuntu Chat Corpus. Molweni is uniquely valuable because it is hand-annotated with discourse structures, providing the ground truth links required for DADgraph's internal mapping. The model was also benchmarked against FriendsQA.

Key Results and Insights

DADgraph demonstrated superior performance and efficiency compared to several strong baselines:

  • Performance: On the Molweni dataset, DADgraph achieved an Exact Match (EM) of 46.5 and an F1-score of 61.5, surpassing standard BERT (58.0 F1) and specialized models like DialogueGCN (61.0 F1).
  • Efficiency: Due to the sparsity of discourse graphs (averaging ~1.06 relations per utterance), DADgraph required significantly less GPU memory (approx. 1/4th) and training time (approx. 1/2) compared to fully-connected graph models.
  • Discourse Sensitivity: Ablation studies revealed that the type of discourse relation (e.g., "Answer" vs. "Acknowledgement") is more critical for performance than the mere existence of a link.

Conclusion and Future Work

DADgraph effectively addresses the entanglement problem in multiparty dialogues by leveraging discourse relations. However, the model currently relies on human-annotated discourse structures. Future work involves integrating automated discourse parsers to enable real-world deployment, as well as exploring more sophisticated interaction layers between questions and dialogue context. Despite these advances, a significant gap remains between model performance (61.5 F1) and human performance (80.2 F1), highlighting multiparty dialogue as a continuing challenge in natural language understanding.

Summary updated

BibTeX
@inproceedings{li2021dadgraph,
  author = {Jiaqi Li and Ming Liu and Zihao Zheng and Heng Zhang and Bing Qin and Min-Yen Kan and Ting Liu},
  title = {{DADgraph: A Discourse-aware Dialogue Graph Neural Network for Multiparty Dialogue Machine Reading Comprehension}},
  booktitle = {IJCNN},
  year = {2021},
  url = {https://openreview.net/forum?id=2tp-PyNpus2},
}
Figure