Learning Seq2Seq Model with Dynamic Schema Linking for NL2SQL

Abstract

AbstractNL2SQL (Natural Language to SQL) is a cutting-edge problem in the field of semantic parsing and TableQA. “CCKS2022: Financial NL2SQL Evaluation” raises a challenging scenario for building NL2SQL systems in the financial domain. To deal with the problem of small-scale data and the requirement of adapting to financial scenarios, we propose an NL2SQL approach to automatically converts natural language questions into SQL queries to achieve accurate table question answering. We use cross-validation and schema linking method that fuses table-column-value information to make full use of all the training data. Then we train a Transformer-based Seq2Seq semantic parsing model with T5 as pre-training model to understand common questions in the financial field and parse the database’s tables, attributes, foreign keys and other complex relationships, and finally generate SQL queries. Experiments are conducted to test the effectiveness of our strategy. Our best ensemble model achieves EM score of 0.287 on testing set and ranks 2nd in the competition.KeywordsSemantic parsingNatural language to SQLSchema linkingSeq2Seq model