Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Think Twice: A Human-like Two-stage Conversational Agent for Emotional Response Generation

Yushan Qian, Bo Wang, Shangzhao Ma, Wu Bin, Shuo Zhang, Dongming Zhao, Kun Huang, Yuexian Hou

TL;DR

A two-stage conversational agent is proposed for the generation of emotional dialogue that outperforms the compared models in the emotion generation and maintains the semantic performance in the automatic and human evaluations.

Abstract

Towards human-like dialogue systems, current emotional dialogue approaches jointly model emotion and semantics with a unified neural network. This strategy tends to generate safe responses due to the mutual restriction between emotion and semantics, and requires rare emotion-annotated large-scale dialogue corpus. Inspired by the "think twice" behavior in human dialogue, we propose a two-stage conversational agent for the generation of emotional dialogue. Firstly, a dialogue model trained without the emotion-annotated dialogue corpus generates a prototype response that meets the contextual semantics. Secondly, the first-stage prototype is modified by a controllable emotion refiner with the empathy hypothesis. Experimental results on the DailyDialog and EmpatheticDialogues datasets demonstrate that the proposed conversational outperforms the comparison models in emotion generation and maintains the semantic performance in automatic and human evaluations.

Think Twice: A Human-like Two-stage Conversational Agent for Emotional Response Generation

TL;DR

A two-stage conversational agent is proposed for the generation of emotional dialogue that outperforms the compared models in the emotion generation and maintains the semantic performance in the automatic and human evaluations.

Abstract

Towards human-like dialogue systems, current emotional dialogue approaches jointly model emotion and semantics with a unified neural network. This strategy tends to generate safe responses due to the mutual restriction between emotion and semantics, and requires rare emotion-annotated large-scale dialogue corpus. Inspired by the "think twice" behavior in human dialogue, we propose a two-stage conversational agent for the generation of emotional dialogue. Firstly, a dialogue model trained without the emotion-annotated dialogue corpus generates a prototype response that meets the contextual semantics. Secondly, the first-stage prototype is modified by a controllable emotion refiner with the empathy hypothesis. Experimental results on the DailyDialog and EmpatheticDialogues datasets demonstrate that the proposed conversational outperforms the comparison models in emotion generation and maintains the semantic performance in automatic and human evaluations.
Paper Structure (19 sections, 8 equations, 3 figures, 7 tables)

This paper contains 19 sections, 8 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Preliminaries
  4. Prototype Utterance Generator
  5. Dialogue Emotion Detector
  6. Controllable Emotion Refiner
  7. Experiments
  8. Datasets
  9. Compared Models
  10. Implementation Details
  11. Evaluation Metrics
  12. Automatic Evaluation
  13. Human Evaluation
  14. Main Results
  15. Human A/B Test
  16. ...and 4 more sections

Figures (3)

  • Figure 1: The real-life examples of emotion adjustment in the human dialogue. The "think twice" strategy can be observed in human intelligent behavior and effectively improves the quality of emotional responses by rewriting expressions or adding extra information. Bold tokens are the rewritten or added part.
  • Figure 2: The overall architecture of the proposed two-stage conversational agent. The first stage includes Prototype Utterance Generator and Dialogue Emotion Detector, which generates prototype response $U_m$ and detects the contextual emotion state as the expected emotion $e_{n+1}$ of the final response, respectively. The second stage includes Rewrite Module, Add Module, and Selector. Rewrite Module and Add Module refine $U_m$ according to $e_{n+1}$, and the Selector selects the final response from the outputs of Rewrite Module and Add Module based on the GLEU score.
  • Figure 3: Compare the prototype response and refined response with respect to the correctness and significance of the emotion. The left and right columns indicate prototype and refined responses, respectively. The red and blue columns indicate correct (i.e., coherent to the contextual emotion) and incorrect emotions, respectively. The length of the columns indicates the significance of the emotion.