LAASF: A Language-Aware Adaptive Sentiment Framework with Adapter Fusion for Low-Resource Languages

Abstract

Sentiment analysis in low-resource Indian languages such as Kannada is challenging due to the scarcity of annotated data and the presence of code-mixed and transliterated text. In this work, we propose the Language-Aware Adaptive Sentiment Framework (LAASF), which incorporates an adapter-based fusion strategy. This framework performs instance-level model selection to handle diverse input types within a single system. The framework leverages the multilingual transformer models MuRIL and XLM-RoBERTa (XLM-R), integrating language-specific and sentiment-specific adapters to effectively combine linguistic and sentiment knowledge. To address the data scarcity problem, cross-lingual alignment was performed to transfer sentiment knowledge from English to Kannada. To improve robustness across domains, a score-based adaptive ensemble method was introduced and evaluated on native, code-mixed, transliterated, and mixed Kannada datasets. Experimental results demonstrated strong performance of the proposed model, with an accuracy of 0.80 on native data and 0.77 on mixed datasets. The findings highlight that adaptive model selection combined with adapter fusion enables effective sentiment analysis in low-resource and heterogeneous multilingual settings.