AOED: Generating SQL with the Aggregation Operator Enhanced Decoding

Yilin Li, Xuan Pan, Dongming Zhao, Minhui Wang, Yanlong Wen

Abstract: NL2SQL is a translation task that converts natural language queries to SQL. We revisit the popular NL2SQL models and find that the accuracy of aggregation operator prediction remains a bottleneck of current NL2SQL models. We present a novel statistics-based approach called AOED, which stands for Aggregation Operator Enhanced Decoding, to help predict aggregation operator. AOED is a carefully designed mechanism that takes full advantage of the statistical information of the aggregation keywords in the natural language query to help improve the prediction accuracy of the aggregation operator. Experiments on the WikiSQL dataset show that our model outperforms the state-of-the-art model SQLova and NL2SQL-RULE by 3.4% and 0.7% on overall SQL results in the logical form accuracy and by 0.2% and 0.7% on aggregation operator result.