Abstract:
Intelligent interface, to enhance e cient interactions between user and databases, is the need of
many commercial applications. Oftentimes, users are not familiar with how to frame a structured
query as they may not be aware of structure of the database and it is also not expected that
the users are required to learn SQL or other query languages to access the database. Hence to
simplify task of accessing the database, text-to-SQL models attempt to translate a user's natural
language question to corresponding SQL query. Converting natural language to SQL, the model
needs to have the ability to create an accurate mapping between the natural language keywords
to SQL keywords along with their corresponding tables and columns. Recently, lots of generative
text-to-SQL models have been developed. Some of them are using greedy search in their decoder.
Hence we choose one of such model[1] and implemented beam search on that. Apart from this we
tried explore a discriminative approach for text-to-SQL generation task. A discriminative re-ranker
has been proposed on the top of a generative text-to-SQL model for improvement of the accuracy
by extracting the best SQL query from a set of beam search predicted candidates. We proposed
a schema agnostic discriminative re-ranker built using XLNet ne-tuned classi er for calculating
similarity score between natural language and predicted SQL. We used that score to re-rank the
beam candidates in a perfect order.
