Enhancing Low-Resource Question-Answering Performance Through Word Seeding and Customized Refinement

The state-of-the-art approaches in Question-Answering (QA) systems necessitate extensive supervised training datasets. In low-resource languages (LRL), the scarcity of data poses a bottleneck, and the manual annotation of labeled data is a rigorous process. Addressing this challenge, some recent efforts have explored cross-lingual or multilingual QA learning by leveraging training data from resource-rich languages (RRL). However, the efficiency of such approaches relies on syntactic compatibility between languages. The paper introduces the innovative method that involves seeding LRL data into RRL to create a bilingual supervised corpus while preserving the syntactical structure of RRL. The method employs the translation and transliteration of selected parts-of-speech (POS) category words. Additionally, the paper also proposes a customized approach to fine-tune the models using bilingual data. Employing the bilingual data and the proposed fine-tuning approach, the most successful model has achieved a 75.62 F1 score on the XQuAD Hindi dataset and a 68.92 F1 score on the MLQA Hindi dataset in a zero-shot architecture. In the experiments conducted using few-shot learning setup, the highest F1 scores of 79.17 on the XQuAD Hindi dataset and 70.42 on the MLQA Hindi dataset have been achieved.

@article{Pandya2024,
  title     = {Enhancing Low-Resource Question-Answering Performance Through Word Seeding and Customized Refinement},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {3},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Hariom Pandya and Brijesh Bhatt},
  doi       = {10.14569/IJACSA.2024.01503138},
  url       = {https://doi.org/10.14569/IJACSA.2024.01503138}
}