SMARAGD: Synthesized sMatch for Accurate and Rapid AMR Graph Distance

The semantic similarity of graph-based meaning representations, such as Abstract Meaning Representation (AMR), is typically assessed using graph matching algorithms, such as SMATCH (Cai and Knight, 2013). However, SMATCH suffers from NP-completeness, making its large-scale application, e.g., for AMR clustering or semantic search, infeasible. To mitigate this issue, we propose SMARAGD (Synthesized sMatch for accurate and rapid AMR graph distance). We show the potential of neural networks to approximate the SMATCH scores and graph alignments, i) in linear time using a machine translation framework to predict the alignments, or ii) in constant time using a Siamese CNN to directly predict SMATCH scores. We show that the approximation error can be substantially reduced by applying data augmentation and AMR graph anonymization.