Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Emotion Classification to Emotional Reasoning: Enhancing Emotional Intelligence in Large Language Models

Arjhun Sreedar, Rohan Pillay, Laukik Patade

TL;DR

This paper addresses the limitation of emotion understanding as flat label prediction in LLMs by introducing synthetic emotional reasoning data to train smaller open LLMs to infer, explain, and reason about emotions through structured chain-of-thought. It proposes a dual-generation pipeline (MADS and a concise version) to create therapy-style dialogues and convert them into Emotion MCQs with explanations, and fine-tunes 7B-scale models using LoRA for emotional reasoning tasks evaluated on EmoBench. The key contributions are a dataset of 3,000+ EU/EA reasoning items, three baselines and rigorous EmoBench-style evaluation showing substantial improvements in $EU$ and $EA$, and a demonstration that emotional reasoning can emerge without architectural changes. The results highlight significant gains for several models (notably Mistral-7B) and point to practical implications for emotionally intelligent agents, while identifying remaining challenges in ambiguous emotions and social-context reasoning.

Abstract

This work investigates whether synthetic emotional chain-of-thought data can improve the emotional reasoning abilities of smaller open large language models (LLMs). We design a multi-agent generation pipeline that produces therapy-style conversations and converts them into structured emotion multiple-choice questions (MCQs) with explanations. We propose that fine-tuning a variety of 7B models on this dataset should yield substantial gains in emotional understanding and emotional awareness on EmoBench-style evaluations, suggesting that emotional reasoning can be induced without architectural changes. Our results demonstrate that fine-tuned Mistral 7B achieves EU improvements from 10.5 to 20.5 and EA improvements from 40.5 to 60.0, validating the effectiveness of synthetic emotional reasoning data for enhancing model capabilities in nuanced emotional tasks.

From Emotion Classification to Emotional Reasoning: Enhancing Emotional Intelligence in Large Language Models

TL;DR

This paper addresses the limitation of emotion understanding as flat label prediction in LLMs by introducing synthetic emotional reasoning data to train smaller open LLMs to infer, explain, and reason about emotions through structured chain-of-thought. It proposes a dual-generation pipeline (MADS and a concise version) to create therapy-style dialogues and convert them into Emotion MCQs with explanations, and fine-tunes 7B-scale models using LoRA for emotional reasoning tasks evaluated on EmoBench. The key contributions are a dataset of 3,000+ EU/EA reasoning items, three baselines and rigorous EmoBench-style evaluation showing substantial improvements in and , and a demonstration that emotional reasoning can emerge without architectural changes. The results highlight significant gains for several models (notably Mistral-7B) and point to practical implications for emotionally intelligent agents, while identifying remaining challenges in ambiguous emotions and social-context reasoning.

Abstract

This work investigates whether synthetic emotional chain-of-thought data can improve the emotional reasoning abilities of smaller open large language models (LLMs). We design a multi-agent generation pipeline that produces therapy-style conversations and converts them into structured emotion multiple-choice questions (MCQs) with explanations. We propose that fine-tuning a variety of 7B models on this dataset should yield substantial gains in emotional understanding and emotional awareness on EmoBench-style evaluations, suggesting that emotional reasoning can be induced without architectural changes. Our results demonstrate that fine-tuned Mistral 7B achieves EU improvements from 10.5 to 20.5 and EA improvements from 40.5 to 60.0, validating the effectiveness of synthetic emotional reasoning data for enhancing model capabilities in nuanced emotional tasks.
Paper Structure (24 sections, 1 figure, 1 table, 2 algorithms)

This paper contains 24 sections, 1 figure, 1 table, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Task and data source
  5. Synthetic data generation pipeline
  6. Multi-Agent Dialogue System (MADS) for EU/EA generation.
  7. Concise single-conversation pipeline.
  8. Persona and attribute diversity.
  9. Data preprocessing
  10. Data annotation
  11. Baselines
  12. Our approach
  13. Comparison to EmoBench data.
  14. Parameter-Efficient Fine-Tuning with LoRA.
  15. Training objective and instruction format.
  16. ...and 9 more sections

Figures (1)

  • Figure 1: Overall methodology pipeline for synthetic emotional reasoning data generation, model fine-tuning, and EmoBench-style evaluation. The pipeline begins with persona and theme selection, proceeds through multi-agent dialogue generation with quality supervision, extracts structured EU/EA items, fine-tunes models using PEFT, and evaluates on held-out EmoBench metrics.