Social Sciences · Political Science and International Relations
MWP-T5: Fine-tuning with Numeracy for Math Word Problem Generation
See on the mapAbstract
Language models have been fine-tuned to generate math word problems (MWPs) from a provided context representing a premise. Approaches have also been made to control the difficulty of the generated MWPs by controlling the number and type of underlying equations of the given MWPs. Additionally, attempts have also been made to generate only univariate MWPs since multivariate MWPs represent complex relationships between various entities and thus can prove to be a challenging task. However, existing approaches are prone to generating semantically impossible or mathematically invalid MWPs with unsatisfactory language quality. They also tend to create MWPs that do not represent the expected equations either by having an incorrect number of equations, missing operands, or a completely incorrect equation. This leads to a loss of control over the MWP generation due to a lack of a guiding metric that allows the language model to learn mathematically valid MWPs that represent equations accurately. In our work, we propose a novel conditional fine-tuning strategy to inject numeracy and MWP solvability into a Text-To-Text Transfer Transformer (T5) to create MWP-T5, a practical model capable of generating mathematically valid MWPs offering greater control over difficulty without compromising language quality. We perform our experiments on real-world MWP datasets and achieve state-of-the-art status by significantly outperforming all existing quantitative and qualitative baselines.