Exploring Emotion Recognition Capability and Emotional Fairness of Large Language Models
Demidova, Anastasiia
Demidova, Anastasiia
Author
Supervisor
Department
Natural Language Processing
Embargo End Date
30/05/2025
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This thesis investigates whether improving emotion recognition capabilities in LLMs affects their emotional biases. To address this question, we enhance emotion recognition through various techniques and then evaluate biases in both vanilla and modified models. This work presents both discriminative and generative LLMs, leveraging approaches such as fine-tuning, incontext learning, and feature fusion with selfattention. Due to the lack of multilabelled multi-modal data, we also introduce a synthetic audio generation pipeline using emotional text-to-speech (TTS) synthesis to study the impact of audio features on emotion classification. Furthermore, we propose an emotional bias evaluation pipeline that measures biases of LLMs across various domains using situation-based emotional prompts. To quantitatively assess the relationship between emotion recognition improvements and emotional bias, we introduce a Pearson correlation analysis between F1scores and bias rate.
Citation
Anastasiia Demidova, “Exploring Emotion Recognition Capability and Emotional Fairness of Large Language Models,” Master of Science thesis, Natural Language Processing, MBZUAI, 2025.
Source
Conference
Keywords
Large Language Model (LLM), Bias, Emotion Recognition, Discriminative Pre-trained Model, Generative Pre-trained Model, Audio Synthesis