Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Multi-view Discourse Framework for Integrating Semantic and Syntactic Features in Dialog Agents

Akanksha Mehndiratta, Krishna Asawa

TL;DR

This paper tackles the challenge of maintaining coherence in open-domain, retrieval-based dialogue with long context. It introduces a Multi-view Discourse Framework (MVDF) that fuses semantic and syntactic information via Multiview Canonical Correlation Analysis (MVCCA) to produce utterance-level representations, and learns discourse tokens through Canonical Correlation Analysis (CCA) to capture inter-turn relations. The framework scores candidate responses by aligning them with learned discourse tokens, enabling discourse-aware retrieval. Experiments on the Ubuntu Dialog Corpus show significant improvements in automatic metrics and qualitative assessments, underscoring the potential of discourse-aware, multi-view integration for practical dialogue systems.

Abstract

Multiturn dialogue models aim to generate human-like responses by leveraging conversational context, consisting of utterances from previous exchanges. Existing methods often neglect the interactions between these utterances or treat all of them as equally significant. This paper introduces a discourse-aware framework for response selection in retrieval-based dialogue systems. The proposed model first encodes each utterance and response with contextual, positional, and syntactic features using Multi-view Canonical Correlation Analysis (MCCA). It then learns discourse tokens that capture relationships between an utterance and its surrounding turns in a shared subspace via Canonical Correlation Analysis (CCA). This two-step approach effectively integrates semantic and syntactic features to build discourse-level understanding. Experiments on the Ubuntu Dialogue Corpus demonstrate that our model achieves significant improvements in automatic evaluation metrics, highlighting its effectiveness in response selection.

A Multi-view Discourse Framework for Integrating Semantic and Syntactic Features in Dialog Agents

TL;DR

This paper tackles the challenge of maintaining coherence in open-domain, retrieval-based dialogue with long context. It introduces a Multi-view Discourse Framework (MVDF) that fuses semantic and syntactic information via Multiview Canonical Correlation Analysis (MVCCA) to produce utterance-level representations, and learns discourse tokens through Canonical Correlation Analysis (CCA) to capture inter-turn relations. The framework scores candidate responses by aligning them with learned discourse tokens, enabling discourse-aware retrieval. Experiments on the Ubuntu Dialog Corpus show significant improvements in automatic metrics and qualitative assessments, underscoring the potential of discourse-aware, multi-view integration for practical dialogue systems.

Abstract

Multiturn dialogue models aim to generate human-like responses by leveraging conversational context, consisting of utterances from previous exchanges. Existing methods often neglect the interactions between these utterances or treat all of them as equally significant. This paper introduces a discourse-aware framework for response selection in retrieval-based dialogue systems. The proposed model first encodes each utterance and response with contextual, positional, and syntactic features using Multi-view Canonical Correlation Analysis (MCCA). It then learns discourse tokens that capture relationships between an utterance and its surrounding turns in a shared subspace via Canonical Correlation Analysis (CCA). This two-step approach effectively integrates semantic and syntactic features to build discourse-level understanding. Experiments on the Ubuntu Dialogue Corpus demonstrate that our model achieves significant improvements in automatic evaluation metrics, highlighting its effectiveness in response selection.

Paper Structure

This paper contains 20 sections, 2 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Canonical Corelation Analysis
  4. Multiview Learning based Language Modeling
  5. Feature Extraction Using Hidden State Interpretation
  6. Multi-view Discourse Framework(MVDF)
  7. Model Definition
  8. Model Architecture
  9. Define The Views
  10. Multi-view Learning Based Utternace Level Representations
  11. Feature Extraction Based Discourse Framework
  12. Response Selection
  13. Evaluation metrics
  14. Experiment
  15. Dataset
  16. ...and 5 more sections

Figures (5)

  • Figure 1: Hidden state interpretation of CCA on two random variables
  • Figure 2: Word token representation in a multi-view space
  • Figure 3: Overview of the Multi-view learning based language modeling
  • Figure 4: Hidden state interpretation of CCA on two utterances to learn shared intent
  • Figure 5: Overview of the proposed discourse framework using CCA