Solving Math Word Problems Using AI Generated Expression Trees

This thesis investigates a structured deep learning approach to building equations that solve math word problems (MWPs), where token generation is constrained according to the output goals and model decisions are observable. A current state-of-the-art (SOTA) solution in the MWP space attempts to generate sub-goals for equation generation, similar to how humans approach MWPs. I extend the sub-goal algorithm and attempt to apply it to solve MWPs that require solution sets instead of singular equations. I find that my model achieves 55.0% solution accuracy on on Math23K, 10% worse when compared to the model I extend. On DRAW-1K, my model achieves 26.5% solution accuracy, significantly under-performing the 69.6% SOTA non-LLM approach. However, these results can be reasonably attributed to dataset size limitations.