Problem-guided Neural Math Problem Solvers

The Math Word Problem (MWP) refers to mathematical problems described in natural language. Recent research has mostly employed sequence-to-sequence (Seq2Seq) or sequence-to-tree (Seq2Tree) approaches to generate computational trees or expressions that help solve the problems. Some works have also treated MWP as a complex relation extraction task. These approaches have shown empirical effectiveness, but recent studies have indicated that these neural solvers merely establish a shallow mapping between text and arithmetic expressions, failing to effectively utilize the textual information of the problem to guide the reasoning process. In this work, to address the issue of neural solvers' limited utilization of text and problem information, we propose a novel architecture for the encoder and decoder, where multiple question representations are obtained from a transformer-based encoder to enhance the guiding capability of problem text information in the decoding process. Each decision in the decoder is guided by the encoded information of the problem text. According to our experiments, our approach demonstrates competitive performance on four benchmark datasets.