Global Reasoning over Database Structures for Text-to-SQL Parsing

摘要: State-of-the-art semantic parsers rely on auto-regressive decoding, emitting one symbol at a time. When tested against complex databases that are unobserved training time (zero-shot), the parser often struggles to select correct set of database constants in new database, due local nature decoding. In this work, we propose globally reasons about structure output query make more contextually-informed selection constants. We use message-passing through graph neural network softly subset for query, conditioned question. Moreover, train model rank queries based global alignment question words. apply our techniques current state-of-the-art Spider, zero-shot parsing dataset with databases, increasing accuracy from 39.4% 47.4%.