Optimization Of Computational Formalisms For Solving Math Word Problems In Regional Languages

The automation of math word problem (MWP) solving has seen significant progress for high-resource languages like English, yet remains underdeveloped for regional languages due to linguistic complexities and data scarcity. This paper presents a comprehensive comparative analysis of computational formalisms for solving MWPs in Hindi as a case study for regional languages, evaluating rule-based, statistical, neural, and large language model (LLM) approaches. We introduce a novel Hindi MWP corpus of 2,500 annotated problems and develop a knowledge-based solver combining verb-operation mappings with constraint logic. Our experiments reveal that while LLMs (GPT-3.5) achieve the highest exact match accuracy (82.5%), rule-based methods offer superior interpretability and speed (78.2% accuracy at 120ms), and hybrid approaches demonstrate a 6.4% performance improvement over standalone models. The study identifies key challenges in regional language MWP solving, including morphological variability (22% error rate in neural models) and implicit operation resolution, while proposing future directions in neuro-symbolic integration and low-resource adaptation. These findings establish critical benchmarks for developing equitable, multilingual educational AI systems, balancing accuracy with pedagogical practicality for underserved language communities.