关键词: 模式链接, 图结构学习, 预训练语言模型, Text-to-SQL, 异构图

Abstract: To optimize Text-to-SQL generation problem based on heterogeneous graph encoder, the SELSQL model was proposed. Firstly, an end-to-end learning framework was employed by the model, and the Poincaré distance metric in hyperbolic space was used instead of the Euclidean metric as a way to optimize semantically enhanced pattern link graph constructed from the pre-trained language model using probe technology. Secondly, K-head weighted cosine similarity and graph regularization methods were used to learn the similarity metric graph so that the initial pattern link graph was iteratively optimized during training. Finally, the improved Relationship Graph ATtention network (RGAT) graph encoder and multi-head attention mechanism were used to encode the joint semantic pattern link graphs of the two modules, and SQL statement decoding was solved using a grammar-based neural semantic decoder and a predefined structured language. Experimental results on Spider dataset show that when using ELECTRA-large pre-training model, the accuracy of SELSQL model is increased by 2.5 percentage points compared with the best baseline model, which has a great improvement effect on the generation of complex SQL statements, and has better robustness in more general usage scenarios.

Key words: schema linking, graph structure learning, pre-trained language model, Text-to-SQL, heterogeneous graph