Graph Alignment for Cross-Domain Text-to-SQL
Yadong Liu, Ya-Hong Hu, Zhen Li, Zhengdong Zhu · 2022 7th International Conference on Intelligent Computing and Signal Processing (ICSP) · 2022
Text-to-SQL, the task of translating the natural language utterance into SQL, has attracted much attention recently. Under the cross-domain setting, the traditional semantic parse model is difficult to adapt to the invisible database schema. The key to being able to better handle cross-domain issues lies in the encoding method for modeling the natural language utterance and the database schema and establishing alignment between them. We propose a Graph Alignment for cross-domain Text-to-SQL (GASQL) to provide a method that unified encodes the natural language utterance and the database schema. Following the unified encoding method, we propose a well-designed graph alignment module to further learn the alignment between the natural language utterance and the database schema. We conducted experiments on the challenging Spider benchmark, and the results proved that our model can align the natural language utterance and database schema well, and achieved good results.