Robust Text-to-SQL Generation with Execution-Guided Decoding

• DOI: 10.48550/arxiv.1807.03100
• Publication Date: 2018-01-01
• Authors: Chenglong Wang; Kedar Tatwawadi; Marc Brockschmidt; Po-Sen Huang; Yi Mao; Oleksandr Polozov; Rishabh Singh
• Keywords: Executable; Leverage (statistics)

We consider the problem of neural semantic parsing, which translates natural language questions into executable SQL queries. introduce a new mechanism, execution guidance, to leverage semantics SQL. It detects and excludes faulty programs during decoding procedure by conditioning on partially generated program. The mechanism can be used with any autoregressive generative model, we demonstrate four state-of-the-art recurrent or template-based parsing models. that guidance universally improves model performance various text-to-SQL datasets different scales query complexity: WikiSQL, ATIS, GeoQuery. As result, achieve accuracy 83.8% WikiSQL.