GraphMemDialog: Optimizing End-to-End Task-Oriented Dialog Systems Using Graph Memory Networks

Jie Wu, Ian G. Harris, Hongzhi ZhaoAAAI 2022 · DOI

Introduction

Task-oriented dialog systems (TDSs) are designed to assist users in completing specific tasks, such as restaurant reservations or navigation, by retrieving information from external Knowledge Bases (KBs). However, current models often struggle with two critical challenges: static knowledge representations that fail to incorporate conversational context, and a reliance on sequential processing that ignores the structural grammatical relationships within a dialog.

GraphMemDialog addresses these issues by introducing a novel end-to-end framework that utilizes a Graph Memory Network (GMN) for context-aware knowledge retrieval and a dual-path encoder to capture both sequential flow and structural dependencies.

Methodology

The GraphMemDialog architecture is composed of three primary modules designed to bridge the gap between dialog history and external knowledge.

The GraphMemDialog architecture, illustrating the Context Graph Encoder, Knowledge Encoder (GMN), and Response Decoder.
  1. Context Graph Encoder: This module processes dialog history through two parallel paths. A Hierarchical Attention Encoder (BiGRU) captures the sequential narrative, while a modified Graph Attention Network (GAT) processes a dependency parsing graph of the sentences. This dual approach ensures the model understands both the "story" of the chat and the underlying grammatical structure.
  2. Knowledge Encoder (GMN): At the core of the system is a Graph Memory Network that represents the KB. Unlike traditional static memory networks, this GMN employs a multi-hop reasoning mechanism and a learnable controller to "write" dialog context into KB entity memories. This makes the KB representation dynamic and context-aware.
  3. Response Decoder: The system uses a "sketch" GRU to first generate a response template (e.g., "@poi is @distance away") and then populates these tags by querying the context-aware KB representations.

Experimental Evaluation

The model was rigorously evaluated on three major public benchmarks: CamRest (restaurant reservations), In-Car Assistant (weather, navigation, and scheduling), and Multi-WOZ 2.1 (a complex multi-domain dataset).

Key Results

GraphMemDialog achieved state-of-the-art (SOTA) performance across all datasets, significantly outperforming strong baselines such as GLMP and Fg2Seq.

Performance comparison of GraphMemDialog against various baselines across three benchmark datasets.
  • CamRest: Achieved a 22.3 BLEU score, representing a ~10% improvement over previous best models.
  • In-Car Assistant: Demonstrated a substantial jump in Entity F1 (64.5), indicating superior accuracy in information retrieval.
  • Multi-WOZ 2.1: Outperformed baselines by a large margin, particularly in specialized domains like "Attraction" and "Restaurant."

Discussion and Insights

The success of GraphMemDialog highlights several key findings:

  • Dynamic KB Retrieval: Treating KB entities as dynamic vectors that incorporate dialog history allows for more effective reasoning.
  • Structural Synergy: Ablation studies confirmed that while sequential encoding is fundamental, incorporating structural dependency graphs via the GAT significantly reduces response generation errors.
  • Multi-hop Reasoning: The GMN's ability to perform multiple iterations of updates allows it to capture deeper semantic relationships between the dialog and the KB than single-pass models.

Conclusion and Future Work

GraphMemDialog sets a new standard for task-oriented dialog systems by effectively fusing context into knowledge retrieval. While the model relies on external dependency parsing tools which may introduce errors, and involves higher computational complexity due to multi-hop reasoning, its performance gains are substantial. Future research may focus on extending this GMN framework to even more complex, multi-domain reasoning tasks where inter-entity relationships are highly dynamic.

Summary updated

BibTeX
@inproceedings{wu2022graphmemdialog,
  author = {Jie Wu and Ian G. Harris and Hongzhi Zhao},
  title = {{GraphMemDialog: Optimizing End-to-End Task-Oriented Dialog Systems Using Graph Memory Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i10.21403},
}
Figure