Generating Structured Database Queries Using Deeply-Bidirectional Natural Language Encodings

In the era of Information and Technology, databases have emerged as an irreplaceable solution to store world's data and effectively use it. They make interaction with data more efficient but on the other hand, it requires the user to understand a query language to access the data. A system that could effectively translate simple Natural Language to database queries would solve the problem of the steep learning curve required to learn database interface query language. The accepted paradigm to solve the problem is to model it as a sequence-to-sequence problem that can in essence translate Natural Language into database queries. Existing state-of-the-art techniques uses bidirectional LSTMs to solve this problem. We propose a deeply bidirectional model that uses parallelized transformer encoders to encode, which can impart bidirectionality to the model which aids in efficient learning. Our proposed model gives a logical form accuracy of 52.2% and an execution accuracy of 67.9%.