[Paper Review] Bias in Emotion Recognition with ChatGPT

Naoki Wake, Atsushi Kanehira+3 more|arXiv (Cornell University)|Oct 18, 2023

Topic Modeling4 citations

TL;DR

This study evaluates ChatGPT's emotion recognition capabilities across multiple datasets and emotion labels, revealing significant performance variability linked to label choice and dataset differences. While fine-tuning improves accuracy, model inconsistencies and sensitivity to label phrasing indicate inherent biases, underscoring the critical need for careful dataset and label selection in deploying LLMs for emotion analysis.

ABSTRACT

This technical report explores the ability of ChatGPT in recognizing emotions from text, which can be the basis of various applications like interactive chatbots, data annotation, and mental health analysis. While prior research has shown ChatGPT's basic ability in sentiment analysis, its performance in more nuanced emotion recognition is not yet explored. Here, we conducted experiments to evaluate its performance of emotion recognition across different datasets and emotion labels. Our findings indicate a reasonable level of reproducibility in its performance, with noticeable improvement through fine-tuning. However, the performance varies with different emotion labels and datasets, highlighting an inherent instability and possible bias. The choice of dataset and emotion labels significantly impacts ChatGPT's emotion recognition performance. This paper sheds light on the importance of dataset and label selection, and the potential of fine-tuning in enhancing ChatGPT's emotion recognition capabilities, providing a groundwork for better integration of emotion analysis in applications using ChatGPT.

Motivation & Objective

• To evaluate ChatGPT’s performance in recognizing nuanced emotions across diverse datasets and emotion labels.
• To investigate the impact of dataset selection and emotion label formulation on model reliability.
• To assess the effectiveness of prompt engineering and fine-tuning in improving emotion recognition accuracy.
• To identify biases and inconsistencies in LLM-based emotion recognition, particularly regarding less common emotion labels.
• To provide guidance for responsible deployment of ChatGPT in applications requiring accurate emotion detection.

Proposed method

• Evaluated ChatGPT using zero-shot prompting with predefined emotion labels on five datasets: IEMOCAP, MELD, EmoryNLP, DailyDialog, and a custom series.
• Applied few-shot prompting and fine-tuning using the gpt-3.5-turbo model on labeled data to improve performance.
• Used conversation history as context to predict the emotion of the current utterance, simulating real-time emotion recognition.
• Compared model outputs across different emotion label sets, including variations like 'happiness' vs. 'happy' and 'powerful' vs. 'peaceful'.
• Analyzed model consistency by comparing responses to identical inputs with different label names or contexts.
• Conducted qualitative analysis of ChatGPT’s definitions of emotion labels to understand its internal representation of emotional concepts.

Experimental results

Research questions

• RQ1 How does ChatGPT’s emotion recognition performance vary across different datasets and emotion label sets?
• RQ2 To what extent does fine-tuning improve ChatGPT’s emotion recognition accuracy and consistency?
• RQ3 How sensitive is ChatGPT’s output to minor changes in emotion label names, such as 'happiness' versus 'happy'?
• RQ4 Why do certain emotion labels like 'powerful' and 'peaceful' show consistently low performance across datasets?
• RQ5 What role do training data biases and labeling criteria play in shaping ChatGPT’s emotion recognition behavior?

Key findings

• ChatGPT demonstrated reasonable zero-shot performance in emotion recognition, consistent with prior findings on sentiment analysis.
• Performance varied significantly across emotion labels, with notably low accuracy for 'powerful' and 'peaceful', suggesting label-specific biases.
• Fine-tuning consistently improved model performance across all datasets, but did not fully eliminate inconsistencies across labels.
• The model showed high sensitivity to label name changes, such as 'happiness' vs. 'happy', indicating that label formulation heavily influences output.
• Despite the black-box nature of LLMs, ChatGPT generated definitions aligned with established emotion models like Ekman’s and Plutchik’s, suggesting structured internal representations.
• Label selection and dataset choice were found to be critical factors influencing model reliability, with inconsistent labeling criteria across datasets contributing to performance variance.