Formative Evaluation of Generative AI Models in Kannada Sentiment Analysis

Abstract

Sentiment analysis in code-mixed Kannada-English social media comments presents unique challenges due to language mixing, informal grammar, and context-dependent expressions. This study compares human-annotated sentiment polarities with those assigned by AI models—ChatGPT, Gemini, and Copilot. The analysis highlights variations in sentiment interpretation, particularly in sarcasm, emoji usage, and indirect expressions in selected comments of Kannada Sentiment Dataset. Our findings reveal discrepancies where AIs either misinterpret local linguistic nuances or apply overly literal sentiment classifications. Understanding these divergences is essential for improving AI sentiment analysis models for Kannada-English code-mixed data.