A Topic-Aware Data Generation Framework for Math Word Problems

In educational practice, math word problems (MWPs) are important material to train pupils’ to solve real-world math problems. In this paper, our focus is on the learning-based topic-aware MWP generation problem. To summarize, there are several challenges with respect to the problem. The first one is how to design the generation model considering the structural characteristics of math equations as well as the topic information. Second, how to generate multiple MWPs covering diverse topics is another challenge. To address the above challenges, we propose a novel framework, GenMWP, to automatically generate high-quality MWPs. First, we use the Tree-LSTM model to capture the structure of the equation. Second, we design a topic embedding model to consider the topic information, allowing the user to input any words or sentences to describe the topic. Third, given only an equation, we also support generating multiple MWPs covering different topics. The automatic evaluation result shows that GenMWP outperforms other baseline methods, and the human evaluation result shows that the MWPs generated by GenMWP have a higher rating in fluency, consistency, equation relevance, and topic relevance.