10th International Congress on Information and Communication Technology

Authors - Amir Ince, Saurav Keshari Aryal, Howard Prioleau
Abstract - With the rise of social media, vast amounts of text, including code-switching, are being generated, presenting unique linguistic challenges for sentiment analysis. This study explores how existing models perform without fine-tuning to understand the challenges of analyzing code-switched data. We propose a prompt tuning approach based on generated versus human-labeled code-switched dataset. Our results show that the Few-shot technique and the Prompt Optimization Framework with Dataset Examples offer the most consistent performance, highlighting the importance of real-world examples and language-specific data in improving multilingual sentiment analysis. However, the studied models and technique do no exhibit the ability to significant triage sentiments for Hindi and Dravidian languages.