Bottom-up Graph Representation Learning for Symbolic Expression Solving

Symbolic expression solving aims to find semantically equivalent solutions for a given input, posing higher capabilities on semantic understanding for deep neural networks(NNs). In this paper, we propose adopting graph as an efficient representation for expressions, enhance the semantic understanding applicability of NNs for more complex scenarios. We explicitly reformulate the graphs as NNs inputs rich in semantic information, instead of relying on serialized expression string. To fully leverage semantic information in graphs, we present a bottom-up graph representation learning NN model, and provide comprehensive empirical evidence to show that our approach significantly improves ability to capture semantic information. Rigorously evaluated across three mathematical tasks and four baseline experimental scenarios, the results confirm that our method excels at uncovering hidden structural information and advancing symbolic expression solving.