Sentiment Analysis of Code-Mixed Indian Languages Using Context-Engineered Prompts

Abstract

This study investigates the application of Large Language Models (LLMs) on sentiment analysis tasks on code-mixed Indian languages, wherein English and Indian languages are combined. This study also points out challenges faced by this technical community, such as uncertainty associated with transliteration and cultural nuances, and holds that classical models require substantial fine-tuning to perform effectively. This study proposes the application of context-engineered prompts, wherein linguistic cues and appropriate sentiment judgments from a cultural standpoint, could provide improvements on sentiment type categorization, as opposed to proposals on developing another model, and demonstrates more consistent behavior than simpler prompting strategies when handling mixed polarity and culturally implicit sentiment. This requires no extra training data and thus pertains to efficiency as well as valuable observations on improving LLM usability on diverse linguistic settings.