Generalizing Math Word Problem Solvers via Solution Diversification

Current math word problem (MWP) solvers are usually Seq2Seq models trained by the (one-problem; one-solution) pairs, each of which is made a description and solution showing reasoning flow to get correct answer. However, one MWP naturally has multiple equations. The training an solver with pairs excludes other solutions, thus limits generalizability solver. One feasible this limitation augment solutions given problem. it difficult collect diverse accurate through human efforts. In paper, we design new framework for introducing buffer discriminator. includes generated encourage data diversity. discriminator controls quality buffered participate in training. Our flexibly applicable wide setting fully, semi-weakly weakly supervised all solvers. We conduct extensive experiments on benchmark dataset Math23k named Weak12k, show that our improves performance various under different settings generating solutions.