RAIN: A Relation-based Arithmetic model with Implicit Numbers

Arithmetic question is one type of the frequently asked questions in Machine Reading Comprehension. The key to answering an arithmetic question is constructing an accurate arithmetic expression by selecting the proper numbers from a candidate set and the operations for selected numbers. Previous work pays little attention on building candidate set, and usually constructs the arithmetic expression depending on each individual number, which ignores the relations between numbers. This paper therefore introduces a Relation-based Arithmetic model with Implicit Numbers (RAIN) for representing and utilizing the numbers (including both numbers explicitly occurring in passages and numbers implicitly inferred from passages) and their relations. Specifically, a Number Relation-based Attention (NRA) mechanism is proposed to mine the relations between numbers. Experimental results on a publicly available dataset DROP show that the RAIN achieves 63.97/67.32 EM/F1 and outperforms several strong baselines. Specifically, it yields overwhelming 73.84/73.86 EM/F1 on arithmetic question. The code for our model is available at https://github.com/MyPaperCode/RAIN.