Math Word Problem Solving with Guided-Context Tree-Structured Model

Math Word Problems (MWP) are a complex task that per- form very poorly when using traditional transformers like the T5 (Text- to-Text Transfer Transformer). While SEQ2Tree models have been put forth, most existing iterations of such models for Math Word Problem Solvers tend to perform rather poorly when compared to its counterparts such as SEQ2SEQ models. But Expression Trees tend to have upsides such as providing greater relational reasoning between quantities. Hence, we propose a model which improve on current iterations of Math Word Problem Solvers by making use of Expression trees with the help of added context to help guide the model to generate the intended equations. It presents the solutions with explainable deductive reasoning steps to iteratively construct the target expressions, involving a previous operation over two objects. We infer from the results presented in this paper how our model performs considerably well. There is an increase in expression accuracy of 30% and an increase in answer accuracy by 21% when comparing our model to the Graph-to-Tree Structured (GTS) model. Experimental results also suggest that the difficulty level of problems plays an important role in determining performance.