Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ECR-Chain: Advancing Generative Language Models to Better Emotion-Cause Reasoners through Reasoning Chains

Zhaopei Huang, Jinming Zhao, Qin Jin

TL;DR

The paper tackles Causal Emotion Entailment (CEE) and the lack of explainable insight into the emotion-generation process. It introduces the Emotion-Cause Reasoning Chain (ECR-Chain), a step-by-step reasoning framework modeled as theme -> reaction -> appraisal -> stimulus, and leverages Chain-of-Thought prompting to guide large language models. A four-stage automated pipeline constructs an ECR-Chain set from LLMs to supervise smaller models via multi-task learning, enabling both direct prediction and explainable reasoning paths. Empirical results on the RECCON-DD dataset show improved CEE performance for both few-shot LLM prompting and supervised Vicuna-7B training, along with demonstrable explainability of the generated rationales, advancing both accuracy and interpretability in emotion-cause reasoning.

Abstract

Understanding the process of emotion generation is crucial for analyzing the causes behind emotions. Causal Emotion Entailment (CEE), an emotion-understanding task, aims to identify the causal utterances in a conversation that stimulate the emotions expressed in a target utterance. However, current works in CEE mainly focus on modeling semantic and emotional interactions in conversations, neglecting the exploration of the emotion-generation process. This hinders the models from deeply understanding emotions, restricting their ability to produce explainable predictions. In this work, inspired by the emotion generation process of "stimulus-appraisal-emotion" in the cognitive appraisal theory, we introduce a step-by-step reasoning method, Emotion-Cause Reasoning Chain (ECR-Chain), to infer the stimulus from the target emotional expressions in conversations. Specifically, we first introduce the ECR-Chain to ChatGPT via few-shot prompting, which significantly improves its performance on the CEE task. We further propose an automated construction process to utilize ChatGPT in building an ECR-Chain set, which can enhance the reasoning abilities of smaller models through supervised training and assist the Vicuna-7B model in achieving state-of-the-art CEE performance. Moreover, our methods can enable these generative language models to effectively perform emotion-cause reasoning in an explainable manner. Our code, data and more details are at https://github.com/hzp3517/ECR-Chain.

ECR-Chain: Advancing Generative Language Models to Better Emotion-Cause Reasoners through Reasoning Chains

TL;DR

The paper tackles Causal Emotion Entailment (CEE) and the lack of explainable insight into the emotion-generation process. It introduces the Emotion-Cause Reasoning Chain (ECR-Chain), a step-by-step reasoning framework modeled as theme -> reaction -> appraisal -> stimulus, and leverages Chain-of-Thought prompting to guide large language models. A four-stage automated pipeline constructs an ECR-Chain set from LLMs to supervise smaller models via multi-task learning, enabling both direct prediction and explainable reasoning paths. Empirical results on the RECCON-DD dataset show improved CEE performance for both few-shot LLM prompting and supervised Vicuna-7B training, along with demonstrable explainability of the generated rationales, advancing both accuracy and interpretability in emotion-cause reasoning.

Abstract

Understanding the process of emotion generation is crucial for analyzing the causes behind emotions. Causal Emotion Entailment (CEE), an emotion-understanding task, aims to identify the causal utterances in a conversation that stimulate the emotions expressed in a target utterance. However, current works in CEE mainly focus on modeling semantic and emotional interactions in conversations, neglecting the exploration of the emotion-generation process. This hinders the models from deeply understanding emotions, restricting their ability to produce explainable predictions. In this work, inspired by the emotion generation process of "stimulus-appraisal-emotion" in the cognitive appraisal theory, we introduce a step-by-step reasoning method, Emotion-Cause Reasoning Chain (ECR-Chain), to infer the stimulus from the target emotional expressions in conversations. Specifically, we first introduce the ECR-Chain to ChatGPT via few-shot prompting, which significantly improves its performance on the CEE task. We further propose an automated construction process to utilize ChatGPT in building an ECR-Chain set, which can enhance the reasoning abilities of smaller models through supervised training and assist the Vicuna-7B model in achieving state-of-the-art CEE performance. Moreover, our methods can enable these generative language models to effectively perform emotion-cause reasoning in an explainable manner. Our code, data and more details are at https://github.com/hzp3517/ECR-Chain.
Paper Structure (18 sections, 1 equation, 4 figures, 5 tables)

This paper contains 18 sections, 1 equation, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Causal Emotion Entailment
  4. Chain-of-Thought for Reasoning
  5. Method
  6. Task Definition
  7. Overview
  8. Few-shot Prompting with ECR-Chain
  9. Supervised Learning with ECR-Chain
  10. Automated Construction of ECR-Chain Set
  11. Multi-task Training with ECR-Chain Set
  12. Experiment
  13. Dataset
  14. Implementation Details
  15. CEE with Few-shot Learning
  16. ...and 3 more sections

Figures (4)

  • Figure 1: An example of emotion-cause reasoning in a conversation. (a): The basic form of the CEE task, requires identifying the causal utterances (#2 & #3) that stimulate the speaker's emotion in the target utterance (#5). (b): Our proposed reasoning method, ECR-Chain, begins by summarizing the conversation theme, then describes the reactions and infers the appraisals of the target speaker, and finally deduces emotion-causes.
  • Figure 2: Illustration of our designed few-shot prompt for reasoning along the ECR-Chain. Detailed prompt examples are presented in the Appendix.
  • Figure 3: An example of the automated construction process of ECR-Chain set. 1) Reasoning: Instruct the LLM with a < reasoning> prompt to generate a raw rationale without providing labels; 2) Filtering: Delete chains which lead to incorrect answers (#4); 3) Rationalization: For each golden answers not predicted (#1 & #3), supplement an additional chain by instructing the LLM with a < rationalization> prompt; 4) Consolidation: Merge all retained and supplemented chains and consolidate semantically similar items by instructing the LLM with a < consolidation> prompt.
  • Figure 4: Case Study on Explainable ECR. (a) shows the question and the labeled causal utterances. (b), (c), and (d) show rationales generated by three models. According to the analysis of GPT-4, the output of our multi-task model is considered to be "more concise and straightforward". The other two models, however, are thought to "contain inaccuracies or assumptions", as indicated by the orange background content in (b) and (d). Nonetheless, GPT-4 states that ChatGPT "provides a more detailed analysis of the stimuli" and "exhibits a more comprehensive understanding of the situation", as indicated by the green background content in (d). This is the reason for the higher score awarded to ChatGPT.