Math2Sym: A System for Solving Elementary Problems via Large Language Models and Symbolic Solvers

Traditional models for solving math word problems (MWPs) often struggle to capture both linguistic context and arithmetic reasoning. We propose Math2Sym, a novel approach integrating large language models (LLMs) with symbolic solvers. This method leverages LLMs' language comprehension and symbolic computation's precision to efficiently convert MWPs into solvable symbolic form. We introduce the EMSF dataset for training models to formalize math problems across various complexities. On our defined test set benchmark, fine-tuned models outperform GPT-3.5 by 17% in few-shot tasks and perform comparably to GPT-4-mini on elementary math problems.