Suvach -- Generated Hindi QA benchmark

TL;DR

This work introduces Suvach, a Hindi extractive QA benchmark generated by large language models to avoid machine translation biases and better reflect Indic-language capabilities. It presents an end-to-end workflow that uses Hindi Wikipedia as context, performs content chunking with retrieval, and employs one-shot prompts to generate high-quality QA items. A validation layer checks context relevance, question relevance, answer accuracy, and clarity, yielding a dataset with over 100k MCQ-style QA pairs. The methodology is designed to generalize to other tasks and Indic languages, offering a scalable path toward more reliable evaluation of Hindi EQA models.

Abstract

Current evaluation benchmarks for question answering (QA) in Indic languages often rely on machine translation of existing English datasets. This approach suffers from bias and inaccuracies inherent in machine translation, leading to datasets that may not reflect the true capabilities of EQA models for Indic languages. This paper proposes a new benchmark specifically designed for evaluating Hindi EQA models and discusses the methodology to do the same for any task. This method leverages large language models (LLMs) to generate a high-quality dataset in an extractive setting, ensuring its relevance for the target language. We believe this new resource will foster advancements in Hindi NLP research by providing a more accurate and reliable evaluation tool.

Figures (4)

• Figure 1: End to End workflow used for question generation and validation
• Figure 2: Samples of chunks and the generated questions
• Figure 3: The prompt used for generation of question
• Figure 4: The prompt used for validation of the generated question