Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{195558,
        author = {Namit Khanduja and Nishant Kumar and Arun Chauhan},
        title = {Sentiment Classification for Low-Resource Indian Languages Using Multilingual Transformers and Parameter-Efficient Fine-Tuning},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {10},
        pages = {7771-7788},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=195558},
        abstract = {—Sentiment classification is an important topic in field of Natural language processing and has gained a lot of attention with the advent of new language models. However, the task remains particularly challenging for the languages that are resource constrained. Indian languages face unique hurdles due to their low-resource nature, characterized by scarce annotated datasets, linguistic diversity, morphological complexity, and a prevalence of code-mixed text. Recent advancement has paved the way for development of multi lingual transformers, but their performance and computational efficiency for low resource language needs further investigation. This paper presents a multifaceted approach to improve sentiment classification across three linguistically distinct Indian languages Hindi, Tamil, and Bengali, by integrating multilingual transformer models with parameter efficient fine-tuning strategies, we explore a spectrum of models ranging from traditional machine learning baselines to state-of-the-art language models.
This study provides two facet solution for the challenges faced by low-resource languages, first the curation of the annotated sentiment corpora for three resource constrained languages Hindi, Bengali and Tamil. Second the experimental evaluation of the 15 modelling approaches which includes traditional machine learning approaches and a suite of transformer-based architectures such as mBERT, XLM-R, IndicBERT, DistilBERT, and MuRIL. Parameter efficient fine-tuning techniques like Low-Rank Adaptation and Adapter were tested on models to check for their performance while reducing the computation cost, the results show that transformer-based models, MuRIL with LoRA performed well with F1-score of 0.87 for Hindi, 0.84 for Tamil and 0.86 for Bengali. The experiments show that use of PEFT techniques to reduce the trainable parameters help in optimizing computational demands while maintaining accuracy of the model, which correlates with the demand of resource constrained environment by opt the trainable parameters. These results were consistent across all three languages, with transformer-based systems consistently outperforming classical methods in both precision and generalizability.},
        keywords = {Sentiment Analysis, Low-Resource Languages, Multilingual Transformers, Parameter-Efficient Fine-Tuning, Indian NLP},
        month = {March},
        }