GLGR: Question-aware Global-to-Local Graph Reasoning for Multi-party Dialogue Reading Comprehension

Yanling Li, Bowei Zou, Yifan Fan, Xibo Li, Ai Ti Aw, Yu HongEMNLP 2023 · DOI

Abstract

Multi-party Dialogue Reading Comprehension (MDRC) is a complex task where information is often scattered across multiple speakers and non-linear discourse structures. Traditional models frequently struggle with irrelevant dialogue noise and fail to capture the fine-grained semantic relationships within individual utterances. This paper introduces GLGR (Global-to-Local Graph Reasoning), a two-stage framework designed to address these issues. GLGR employs a Question-aware Interlocutor-Utterance Graph (QIUG) to filter global context and a Local Semantic Role Graph (LSRG) to pinpoint precise semantic clues. Evaluated on the Molweni and FriendsQA benchmarks, GLGR achieves state-of-the-art performance, demonstrating the effectiveness of its hierarchical, question-aware reasoning approach.

Introduction: The Challenge of Multi-party Dialogue

Multi-party Dialogue Reading Comprehension (MDRC) is significantly more challenging than standard machine reading comprehension. Unlike typical documents where information flows linearly, multi-party chats—such as technical support threads or TV show scripts—feature clues scattered across different speakers. These conversations are often disordered, and speakers frequently rely on pronouns or shorthand, making it difficult for models to track the flow of information.

Existing models generally suffer from two major limitations:

  1. Noise Sensitivity: Many models build dialogue graphs without considering the specific question being asked. This leads them to process irrelevant parts of the conversation, introducing noise that degrades performance.
  2. Local Semantic Neglect: While models may capture high-level structure (who spoke to whom), they often ignore the internal semantic structure of sentences—the "who did what to whom" relationships. These fine-grained details are often where the critical clues for answering a question reside.

The GLGR Framework

To overcome these challenges, the authors propose GLGR, a two-stage "Global-to-Local" reasoning approach. This method first identifies the relevant context within the entire dialogue and then zooms in on the semantic details of specific utterances.

Overview of the MDRC task and the proposed Global-to-Local graph structures (QIUG and LSRG).

1. Global Reasoning: QIUG

The Question-aware Interlocutor-Utterance Graph (QIUG) expands traditional dialogue graphs by introducing a dedicated "Question" node. This node connects to every utterance, acting as a filter that allows the model to prioritize utterances relevant to the query. The QIUG models six distinct relation types, including "querying," "replying," and "issuing," to capture the complex dynamics of multi-party interaction.

2. Local Reasoning: LSRG

The Local Semantic Role Graph (LSRG) focuses on the internal structure of sentences. Using Semantic Role Labeling (SRL), the model identifies predicates (actions) and their associated arguments (actors and objects). To bridge information across different sentences, GLGR utilizes coreference resolution (e.g., linking "it" to a previously mentioned "device") and synonym identification to merge nodes across the local graphs of various utterances. This allows the model to "reason" that a speaker's answer about "restarting" relates to a previous question about "installing," even without exact word matches.

Two-Stage Encoder Architecture

The GLGR architecture follows a progressive pipeline designed to refine information from the broad context down to specific answer spans.

The complete GLGR architecture, showing the progressive encoding from global to local graphs.
  • Stage 1 (Global): A Graph Attention Network (GAT) processes the QIUG to determine which parts of the conversation are most relevant to the question.
  • Stage 2 (Local): A Graph Transformer processes the LSRG, refined by the global context from the first stage, to pinpoint exact semantic clues.
  • Extraction: The final representations are fed into a pointer network to extract the precise answer span from the dialogue.

Experimental Evaluation

The model was evaluated on two major MDRC benchmarks: Molweni, a large-scale technical discussion dataset, and FriendsQA, derived from transcripts of the TV show Friends.

Key Results

GLGR achieved significant performance gains over strong baselines like BERT and ELECTRA:

  • Molweni: Reached 73.6% F1, outperforming the previous state-of-the-art (ESA).
  • FriendsQA: Reached 77.2% F1.
  • Robustness: The model showed consistent improvements regardless of the underlying backbone (BERT-base, BERT-large, or ELECTRA), proving the architectural design's stability.
  • Ablation Studies: Removing either the global (QIUG) or local (LSRG) components led to a performance drop, confirming that both "zoom levels" are necessary for effective reasoning.

Conclusion and Future Work

GLGR demonstrates that making the reasoning process question-aware and hierarchical—moving from global context to local semantics—is highly effective for multi-party dialogue. However, the model's reliance on external preprocessing tools (SRL, coreference resolution) means that upstream errors can propagate. Additionally, performance begins to degrade in very long conversations (beyond 30 utterances). Future work may involve exploring more efficient "Heterogeneous Graph Transformers" to reduce computational overhead and better handle long-range dependencies.

Summary updated

BibTeX
@inproceedings{li2023glgr,
  author = {Yanling Li and Bowei Zou and Yifan Fan and Xibo Li and Ai Ti Aw and Yu Hong},
  title = {{GLGR: Question-aware Global-to-Local Graph Reasoning for Multi-party Dialogue Reading Comprehension}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  doi = {10.18653/v1/2023.findings-emnlp.122},
}
Figure