A hybrid model for low-resource language text classification and comparative analysis
Context: The growing digital content in many languages helps users share diverse information. However, classifying user reviews is time-consuming and biased. Transformers like BERT excel in NLP, but low-resource languages still face challenges due to limited data, computational resources, and lingui...
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025
|
| Online Access: | http://psasir.upm.edu.my/id/eprint/119285/ http://psasir.upm.edu.my/id/eprint/119285/1/119285.pdf |
| Summary: | Context: The growing digital content in many languages helps users share diverse information. However, classifying user reviews is time-consuming and biased. Transformers like BERT excel in NLP, but low-resource languages still face challenges due to limited data, computational resources, and linguistic tools. Objective: The objective of this paper is twofold: (1) to evaluate and compare existing text classification methods using a newly annotated dataset for Malay, a low-resource language; and (2) to propose a new hybrid model for classifying low-resource languages that combines rule-based linguistic features with transfer learning approaches. Methods: For this analysis, five tools—LangDetect, spaCy, FastText, XLM-RoBERTa and LLaMA were applied. The study compares these tools against a low-resource dataset (Malay) to identify gaps and limitations in performance. The research focuses on several main areas: (i) Challenges in Low-Resource Languages, (ii) Comparative Analysis, (iii) Proposed Model, and (iv) Empirical Evaluation. The dataset includes 74,931 user reviews from Google Play Store apps (MyBayar, PDRM, MyJPJ, and MySejahtera). A subset of 2621 reviews was selected and annotated by two independent coders, and Fleiss’ Kappa was used to ensure reliable agreement for a ground-truth dataset. Results: The proposed hybrid model demonstrated statistically significant improvements in classification performance, achieving an accuracy of 84 %. Paired t-tests further confirm these improvements, showing significant differences in F1-score compared to baseline methods (p < 0.05). Conclusion: Findings emphasize the need for tailored NLP approaches for underrepresented languages, showing the importance of custom models to handle language diversity and further development in low-resource language. |
|---|