A Bottom-Up DAG Structure Extraction Model for Math Word Problems

Research on automatically solving mathematical word problems (MWP) has a long history. Most recent works adopt Seq2Seq approach to predict the result equations as sequence of quantities and operators. Although can be written sequence, it is essentially structure. More precisely, Direct Acyclic Graph (DAG) whose leaf nodes are quantities, internal root arithmetic or comparison In this paper, we propose novel Seq2DAG extract equation set directly DAG It extracted in bottom-up fashion by aggregating sub-expressions layer iteratively. The advantages our three-fold: intrinsically suitable solve multivariate problems, always outputs valid structure, its computation satisfies commutative law for +, x =. Experimental results Math23K DRAW1K demonstrate that model outperforms state-of-the-art deep learning methods. We also conduct detailed analysis show strengths limitations approach.