Towards Interpretable Math Word Problem Solving with Grounded Linguistic Logic Reasoning

Anonymous

Abstract: Automatically math word problem (MWP) solving is a challenging artificial intelligence task since a machine should be able to not only understand problem comprehensively on linguistics but also the grounded math logic entailed in problem. Recently, lots of deep learning models have made great progress in MWP solving on answer accuracy, they rely on shallow heuristics to achieve high performance, lacking of grounded math logic reasoning, which makes them uninterpretable. To address this issue and push the research boundary of MWPs to interpretable MWP solving, we construct a large-scale and high-quality MWP dataset named InterMWP which consists of 11507 MWP data and annotates interpretable algebraic knowledge formulas as the grounded linguistic logic of each solving equation and asks for a solver to output the formulas when it decides current predicted node is a inner-node (operator) during expression reasoning. We further propose a strong baseline called InterSolver to show the effectiveness of our constructed dataset and show how to harvest these logic knowledge by fusing logic knowledge with semantic representation to improve problem solving and make a step towards providing interpretability. Experimental results shows that our InterSolver has strong logical formula-based interpretability while achieving high answer accuracy simultaneously.