Systematic review and comparative synthesis of cross lingual question answering systems for low resource Indic languages
TL;DR
A systematic review of 137 shortlisted research articles reveals that the performance of a cross-lingual QA system is structurally influenced by the interaction between linguistic typology, model design, and not exclusively by model scale or architecture.
Abstract
Recent advancements in Question Answering (QA) methods on resource-rich languages have significantly improved answer accuracy and contextual understanding performance. However, extending these capabilities to low-resource Indic languages remains a persistent challenge. This study carries out a systematic review of 137 shortlisted research articles, and a detailed comparative synthesis of 78 core articles is further conducted by considering dataset diversity, modelling paradigms, alignment strategies, and evaluation frameworks across extractive, generative, and conversational QA systems. To improve analytical rigour and follow a systematic review process, four Research Questions (RQs) are defined to verify dataset suitability, cross-lingual alignment strategies, adaptation mechanisms, and evaluation outcomes. The review additionally explores recent advancements and methods used in building QA systems with a primary focus on learning multilingual datasets, pre-trained language models, cross-lingual transfer methods, and evaluation metrics specifically tailored for Indic languages. Even though language-specific models utilize fine-grained linguistic features, a comparative analysis shows that the multilingual models’ broader representation and transfer capabilities help them perform better than language-specific frameworks. In general, the review findings indicate that the performance variations are more context-specific rather than consistent across different datasets, task formulations, and alignment strategies. In contrast to existing QA surveys, the proposed review offers an Indic-centred synthesis that explicitly relates to dataset design, alignment paradigms, and evaluation practices to other structural linguistic variables such as morphology richness, script diversity, and typological divergence. The review findings reveal that the performance of a cross-lingual QA system is structurally influenced by the interaction between linguistic typology, model design, and not exclusively by model scale or architecture.