Improving Question Answering System for Low resource Languages

Abstract

The state-of-the-art approaches using transformer based architectures have shown notable performance enhancements in various Natural Language Processing (NLP) applications. Usage of such architecture necessitates extensive supervised training datasets. In Low Resource Languages (LRLs), the scarcity of data poses a bottleneck, and the manual annotation of labeled data is a rigorous process. The multilingual pre-trained models make it possible to develop applications for LRLs using the model of Resource-Rich Languages (RRLs). However,the structural characteristics of the target languages can impact task specific learning and the performance of such models on the LRLs is not compatible with of RRLs. This study focuses on addressing the data scarcity issue in LRLs within the Question Answering (QA) domain. We propose three approaches to enhance the performance of LRLs using multilingual transformer architecture. Firstly, we investigated the influence of structural diversities of languages on performance. Our customized approach leverages task-specific data from LRL families through transfer learning from RRLs, combining supervised and unsupervised data. Experimental results using XLM-R, mBERT, and IndicBERT transformer models on Indic languages (Hindi, Bengali, and Telugu) demon strate substantial improvements in few-shot learning scenarios. On the XQUAD-Hindi dataset, the few-shot learning using Bengali improves the benchmark mBERT (F1/EM) score by +(10.86/7.87) and XLM-R score by +(3.84/4.42). Few-shot learning using Tel ugu has also improved the mBERT score by +(10.42/7.36) and +(3.04/2.72) for XLM-R. In addition, our model has demonstrated benchmark-compatible performance in a zero shot setup with single-epoch task learning. Secondly, we introduce a method that involves incorporating LRL data into RRL to create a bilingual supervised corpus while preserving syntactical structures. This approach utilizes translation and transliteration techniques for selected Parts-of-Speech (POS) cat i egory words.