Math word problem (MWP) solving faces a dilemma in number representation learning. In order to avoid the number representation issue and reduce the search space of feasible solutions, existing works striving for MWP solving usually replace real numbers with symbolic placeholders to focus on logic reasoning. However, different from common symbolic reasoning tasks like program synthesis and knowledge graph reasoning, MWP solving has extra requirements in numerical reasoning. In other words, instead of the number value itself, it is the reusable numerical property that matters more in numerical reasoning. Therefore, we argue that injecting numerical properties into symbolic placeholders with contextualized representation learning schema can provide a way out of the dilemma in the number representation issue here. In this work, we introduce this idea to the popular pre-training language model (PLM) techniques and build MWP-BERT, an effective contextual number representation PLM. We demonstrate the effectiveness of our MWP-BERT on MWP solving and several MWP-specific understanding tasks on both English and Chinese benchmarks.

Math word problem (MWP) solving is the task of transforming a sequence of natural language problem descriptions to executable math equations. An MWP solver not only needs to understand complex scenarios described in the problem texts, but also identify the key mathematical variables and associate text descriptions with math equation logic. Although recent sequence modeling MWP solvers have gained credits on the math-text contextual understanding, pre-trained language models (PLM) have not been explored for solving MWP, considering that PLM trained over free-form texts is limited in representing text references to mathematical logic. In this work, we introduce MWP-BERT to obtain pre-trained token representations that capture the alignment between text description and mathematical logic. Additionally, we introduce a keywordbased prompt matching method to address the MWPs requiring common-sense knowledge. On a benchmark Math23K dataset and a new Ape210k dataset, we show that MWP-BERT outperforms the strongest baseline model by 5-10% improvement on accuracy.