EAGS: An extracting auxiliary knowledge graph model in multi-turn dialogue generation.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Multi-turn dialogue generation is crucial for question answering systems but current methods often miss subjective information by focusing on latent aspects or external knowledge.
• Limited availability of relevant external knowledge and complete entity-linked graphs hinders the effectiveness of existing approaches.
• Dependency trees offer explicit sentence information, which is often overlooked in dialogue generation.

Purpose of the Study:

• To propose the EAGS model for knowledge graph-enabled multi-turn dialogue generation.
• To combine explicit dependency tree information with implicit semantic information for improved dialogue quality.
• To develop a model that generates dialogues without requiring external knowledge sources.

Main Methods:

• The EAGS model extracts and builds a dependency knowledge graph from input sentences.
• It utilizes node representations shared with Bi-GRU at each time step for semantic-level processing.
• A multi-task training approach is employed to balance local and global feature extraction.

Main Results:

• The EAGS model successfully integrates explicit structural information (dependency trees) with implicit semantic information.
• It demonstrates the ability to build and utilize internal knowledge graphs from existing sentences.
• Experiments on large-scale datasets show superior performance compared to baseline models in both automatic and human evaluations.

Conclusions:

• The EAGS model offers a novel approach to multi-turn dialogue generation by leveraging explicit dependency structures.
• The model effectively captures subjective key information often lost in latent-based methods.
• The EAGS model provides a robust and self-contained solution for generating high-quality dialogues.