Solving Math Word Problem with External Knowledge and Entailment Loss

Automatic math word problem(MWP) solving is an interesting task for NLP researchers in recent years. Over the last few years, a growing number of effective sequence-to-sequence deep learning-based model are proposed. However, these models do not efficiently consider factual errors as the sequence-to-sequence model can produce expressions that do not appear in the question. Additionally, these models neglect external knowledge information during the math word problem-solving process. To address these problems, we propose a model that can automatically solve math word problems with External Knowledge and Entailment Loss (MathEE). MathEE uses a Textual-Entailment auxiliary task to identify factual errors and introduces an entity graph based on external knowledge to model the highly relevant entity words in the question. Our experimental results on publicly available Chinese datasets Ape210K and Math23K show that MathEE achieves an accuracy rate of 74.43% and 78.7%, which is 2.08% and 1.6% higher than strong baseline models.