Entropy-guided Script-aware Mixture-of-Experts to Tri-lingual Sentiment Analysis of Marathi Hindi English communication

Multilingualism in the Indian online platforms are a problem that involves script-switching and transliteration as it is a complex phenomenon that presents a challenge to sentiment analysis systems. Indicatively, both Devanagari and Latin scripts can be used in one discourse unit in Marathi, Hindi and English. By representational interference and distributional mismatch, the performance of traditional multilingual transformer models is reduced when exposed to romanized data with mixed scripts. This paper provides Tri-Indic MoE-Sent, an early entropy-based script-aware mixture-of-experts (MoE) model of tri-lingual sentiment analysis. The framework is comprised of three neural experts of English, Devanagari (Marathi/Hindi), and romanised/code-mixed scripts and a lightweight gating network, which is operated by the ratio of scripts and entropy of the data. Also, there is a lexicon-based, symbolic, sentiment-analytical branch, which explains the effects of linguistic negation, emotional intensity through negation, and emojis. In the case of Hindi, Marathi, English and Hinglish, the findings reveal that performance is improved in Hindi and mixed script strength is increased in comparison to the multilingual baselines (XLM-R [12] and IndicBERT [5]). The suggested framework and the findings were based on the theory that multilingual sentiment could be addressed as a combination of the various distribution of different regimes of script, which could be used to conditional computation. Ablation and calibration analyses demonstrate that expert specialization regime-sensitive is more efficient and stable than monolithic multilingual encoders.