Disentangled Memory Retrieval Towards Math Word Problem Generation

Anonymous

Abstract: The task of math word problem (MWP) generation, which generates a MWP given an equation and relevant topic words, has increasingly attracted researchers' attention. In this work, we propose a seq2seq model with a disentangled memory retrieval module to better take advantage of the logical description and scenario description within a MWP and more relevant training data to improve the generation quality. We first disentangle the training MWPs into logical descriptions and scenario description and then record them in respective memory modules. Later, we use the given equation and topic words as queries to retrieve the most relevant logical descriptions and scenario description from the corresponding memory modules respectively. The retrieved results are then used to complement the process of the MWP generation. Extensive experiments verify the superior performance and effectiveness of our method.

Paper Type: long