Transformer-Based Model for the Natural Language Textual Sequences to SQL-Queries Conversion

The concept of using natural language instead of SQL queries formed the basis of a new type of data processing called Natural Language Interface to Database (NLIDB). This paper considers a neural network (NN) approach for the NLIDB systems development. Its essence is the transformation of the input text sequence specified by the user in natural language into commands in SQL, which is inherent in text-to-SQL and sequence-to-sequence tasks. The Transformer model architecture of the NN was justified in the research. This architecture allows the NN to convert an input sequence (such as a natural language text query) into an output sequence (translated into SQL format text query). WikiSQL dataset was used for NN learning and testing. Numerous experiments were conducted in order to determine the optimal parameters of the NN model were conducted. Bilingual Evaluation Understanding (BLEU) metric was used as a metric for evaluating the quality of converting an input text sequence to SQL queries. The BLEU metric for the neural network model trained on the WikiSQL sample is 56%.