Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards a Generative Approach for Emotion Detection and Reasoning

Ankita Bhaumik, Tomek Strzalkowski

TL;DR

This work reframes emotion detection as a generative QA task using a two-step process: first generate domain-specific context, then produce step-by-step emotional reasoning and open-ended emotion labels. By adopting context-aware prompts and soft majority voting over multiple generated explanations, the approach yields higher accuracy and richer explanations than standard zero-shot and entailment-based baselines, demonstrated on ISEAR and #Emotional Tweets. The authors also release updated datasets with fine-grained emotion labels and top explanations to support further research in emotional reasoning. Overall, the method enhances interpretability and adaptability across domains, with implications for more explainable and empathetic NLP systems.

Abstract

Large language models (LLMs) have demonstrated impressive performance in mathematical and commonsense reasoning tasks using chain-of-thought (CoT) prompting techniques. But can they perform emotional reasoning by concatenating `Let's think step-by-step' to the input prompt? In this paper we investigate this question along with introducing a novel approach to zero-shot emotion detection and emotional reasoning using LLMs. Existing state of the art zero-shot approaches rely on textual entailment models to choose the most appropriate emotion label for an input text. We argue that this strongly restricts the model to a fixed set of labels which may not be suitable or sufficient for many applications where emotion analysis is required. Instead, we propose framing the problem of emotion analysis as a generative question-answering (QA) task. Our approach uses a two step methodology of generating relevant context or background knowledge to answer the emotion detection question step-by-step. Our paper is the first work on using a generative approach to jointly address the tasks of emotion detection and emotional reasoning for texts. We evaluate our approach on two popular emotion detection datasets and also release the fine-grained emotion labels and explanations for further training and fine-tuning of emotional reasoning systems.

Towards a Generative Approach for Emotion Detection and Reasoning

TL;DR

This work reframes emotion detection as a generative QA task using a two-step process: first generate domain-specific context, then produce step-by-step emotional reasoning and open-ended emotion labels. By adopting context-aware prompts and soft majority voting over multiple generated explanations, the approach yields higher accuracy and richer explanations than standard zero-shot and entailment-based baselines, demonstrated on ISEAR and #Emotional Tweets. The authors also release updated datasets with fine-grained emotion labels and top explanations to support further research in emotional reasoning. Overall, the method enhances interpretability and adaptability across domains, with implications for more explainable and empathetic NLP systems.

Abstract

Large language models (LLMs) have demonstrated impressive performance in mathematical and commonsense reasoning tasks using chain-of-thought (CoT) prompting techniques. But can they perform emotional reasoning by concatenating `Let's think step-by-step' to the input prompt? In this paper we investigate this question along with introducing a novel approach to zero-shot emotion detection and emotional reasoning using LLMs. Existing state of the art zero-shot approaches rely on textual entailment models to choose the most appropriate emotion label for an input text. We argue that this strongly restricts the model to a fixed set of labels which may not be suitable or sufficient for many applications where emotion analysis is required. Instead, we propose framing the problem of emotion analysis as a generative question-answering (QA) task. Our approach uses a two step methodology of generating relevant context or background knowledge to answer the emotion detection question step-by-step. Our paper is the first work on using a generative approach to jointly address the tasks of emotion detection and emotional reasoning for texts. We evaluate our approach on two popular emotion detection datasets and also release the fine-grained emotion labels and explanations for further training and fine-tuning of emotional reasoning systems.
Paper Structure (23 sections, 3 equations, 5 figures, 4 tables)

This paper contains 23 sections, 3 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Emotional Reasoning
  3. Methodology
  4. Context Generation
  5. Emotion Generation
  6. Answer Selection
  7. Experiments
  8. Datasets
  9. Implementation
  10. Baselines
  11. Results
  12. Zero-shot Emotion Detection
  13. Emotional Reasoning Dataset
  14. Significance of Context Generation
  15. Error Analysis
  16. ...and 8 more sections

Figures (5)

  • Figure 1: Example input text taken from ISEAR dataset. Our approach generates an open ended set of emotions along with an emotional reasoning for a final answer.
  • Figure 2: Overall architecture of our approach to generate: (1) an emotion label over a fixed set of labels (2) an open ended set of emotion words (3) top-k explanations
  • Figure 3: Human evaluation of top-3 generated emotional reasoning in ISEAR dataset. Plot shows the distribution of scores for questions 1-5.
  • Figure 4: Comparison of distributions of emotion words in gold ISEAR dataset vs. generated emotion labels using our approach
  • Figure 5: Example from ISEAR illustrates how different contexts result in a distinct CoT emotion reasoning.