{Zhipeng Xie and Shichao Sun}

Abstract

Most existing neural models for math word problems exploit Seq2Seq model to generate solutionexpressions sequentially from left to right, whoseresults are far from satisfactory due to the lackof goal-driven mechanism commonly seen in human problem solving. This paper proposes a treestructured neural model to generate expression treein a goal-driven manner. Given a math word problem, the model first identifies and encodes its goalto achieve, and then the goal gets decomposed intosub-goals combined by an operator in a top-downrecursive way. The whole process is repeated until the goal is simple enough to be realized by aknown quantity as leaf node. During the process,two-layer gated-feedforward networks are designedto implement each step of goal decomposition, anda recursive neural network is used to encode fulfilled subtrees into subtree embeddings, which provides a better representation of subtrees than thesimple goals of subtrees. Experimental results onthe dataset Math23K have shown that our treestructured model outperforms significantly severalstate-of-the-art models.