SentMixSent: A Comparative Study of Sentiment Analysis Techniques for Code-Mixed Text

Understanding human emotions in multilingual digital conversations is challenging due to frequent language switching, non-standard spellings, and the lack of structured linguistic resources. This study introduces SentMixSent, a novel comparative framework for sentiment analysis in EnglishKonkani, Hindi-English, Marathi-English, and English-MarathiKonkani code-mixed text. We systematically examine three methodologies: (1) Lexicon-Driven Sentiment Analysis, utilizing sentiment lexicons and polarity propagation; (2) Rule-Based Sentiment Alignment, incorporating transliteration-aware sentiment mapping and heuristic enhancements; and (3) Machine Learning-Based Sentiment Detection, employing supervised classifiers such as Logistic Regression, SVM, and BiLSTM with hybrid feature engineering. Our findings reveal that rule-based enhancements improve lexicon-driven sentiment identification, while machine learning techniques outperform traditional approaches when trained on diverse labeled datasets. The hybrid integration of linguistic rules with deep learning-backed sentiment classification yields a 5.1 % increase in F1-score over conventional lexicon-based methods. Beyond sentiment analysis, SentMixSent serves as a scalable, cross-lingual intelligence system with applications in social media analytics, multilingual customer sentiment tracking, and context-aware emotion recognition in code-mixed digital discourse.