Context-Dependent Text-to-SQL Generation with Intermediate Representation

Context-Dependent Text-to-SQL Generation with Intermediate Representation

연구 분야: Databases

학회: Pacific Rim International Conference on Artificial Intelligence

In recent years, the Text-to-SQL task has become a research hotspot in semantic analysis. Among them, context-dependent Text-to-SQL task has received more and more attention as it meets the needs of actual scenarios. The core of the problem is how to use historical interaction information and database schema to understand the context. Most existing research ignores the structure of SQL queries and introduces low-level information such as variable names and parameters, and the mismatch problem between intents expressed in utterance and the implementation details in SQL still exists. In this paper, SemQL is applied to serve as an intermediate representation between utterance and SQL, meanwhile, the Coarse-to-Fine neural architecture is adopted to decompose decoding process of SemQL into two stages. We validated the performance of our model on SParC and CoSQL datasets, which outperforms the existing ones and achieves excellent results on both datasets.