ATHENA: Mathematical Reasoning with Thought Expansion

Anonymous

Abstract: Solving math word problems depends on how to articulate the problems, the lens through which models view human linguistic expressions. Real-world settings count on such a method even more due to their lexical sophistication on the same mathematical operations. Earlier works constrain available thinking processes by repeatedly training the patterns or relations between quantities without considering their validity in the context of the problems. We tackle the above challenges and propose Attention-based THought Expansion Network Architecture (ATHENA) to learn mathematics so that it can be practical enough in real-world settings. We introduce thought expansion that maximizes feasible reasoning pathways by mimicking human thinking mechanisms. Thought expansion generates candidate thoughts carrying consistent representation for each mathematical expression and yields reasonable thoughts, filtered by solidly updated reasoning vectors. Our experiments show that ATHENA achieves a new state-of-the-art stage toward the ideal method that is compelling in variant questions even when the informativeness in training examples is restricted.

Paper Type: long

Research Area: NLP Applications