Abstract:
"Over the years, relational database management systems (RDBMs) have been in use among both technical and non-technical professionals. However, it is apparent that the query languages that are used to extract data from RDBMs require extensive knowledge. Hence, ample research has been conducted on natural language interfaces for query languages such as SQL. Nevertheless, previous research contains some limitations including limited support for unseen databases, limited accuracy for complex natural language expressions and complexity of generating constants in SQL. It is evident that the limitation of parsing complex natural language expressions to SQL queries has rarely been addressed in the existing literature.
A rule-based approach is applied in this research to overcome the limited accuracy of translating complex natural language expressions to SQL. Thereby, a unique set of rules is defined to identify table names, column names, conditions, table joins, aggregate operators, GROUP BY and ORDER BY scenarios that are associated with a given natural language expression. Furthermore, the rules are derived by applying dependency parsing and regular expressions.
The accuracy of QueryLift was evaluated using a self-composed dataset which has yielded 84% accuracy. Spider dataset was used to conduct competitive benchmarking and ensure cross domain adaptability. The system, QueryLift was compared with NaLIR and Templar which are state of the art rule-based systems. Thereby, QueryLift has revealed 4.1% accuracy which is higher than the accuracy of both NaLIR and Templar."
