Math Word Problem Solving with Dual Solving Thinking Structure

Math Word Problem, abbreviated as MWP, constitute a class of elementary mathematical application scenarios, commonly encountered within primary secondary educational curricula. Such problems are characterized by their presentation in natural language, eschewing the utilization of intricate mathematical symbols or formulas. The focal task of model geared towards addressing Math Word Problem is twofold: first, to discern and interpret the mathematical concepts encapsulated within the textual descriptions provided, and second, to formulate an appropriate mathematical expression or equation reflective of the problem's essence. Subsequently, these models proceed to compute the solution corresponding to the derived expression, thereby completing the problem-solving process. Currently, advancements in this domain primarily revolve around improvements to encoders and decoders within models, with a relatively narrow scope of thinking processes during model solving. Consequently, the efficacy of expression solving for complex semantic Math Word Problem remains to be enhanced. In contrast to the prevalent sequence-to-tree approach, humans often employ a top-down thinking process to solve Math Word Problem. We introduce a Math Word Problem solving model featuring a dual-thinking structure, comprising a top-down solver, a step-by-step anthropomorphic thinking solver, and a consistent comparison learning module. By aligning expressions derived from two distinct solvers at multiple granularities to maintain consistency, we aim to enhance reasoning correctness. Experimental results across multiple datasets demonstrate the superiority of our approach over existing baselines, particularly in achieving notable improvements in solving accuracy for complex problems.