Translating natural language questions into SQL queries using Semantic Parsing and the Rational Speech Acts Model

Abstract

This work presents the application of the rational speech acts (RSA) model to create an entity linking module in a semantic parsing setting. In the theoretical part we review the rational speech acts model and explore a myriad of approaches to semantic parsing. Main ideas from the core articles from both fields are extracted and concisely summarised, which motivate the hypothesis behind this work. In the research part we first present the WIKISQL dataset which serves as the evaluation engine for the introduced models. Then, we give an overview of an existing SQL semantic parsing model, that acts as a base for the models proposed as part of this work. Subsequently, we demonstrate a novel model to solve the task on hand, that builds upon the aforementioned architecture, enriching it with an entity linking module. The final results are compared against a number of baselines from previous publications and exhibit a significant improvement over them. Such an enhancement goes to show that pragmatic reasoning and, in particular, the RSA, helps to disambiguate entities mentioned in the natural language question and can serve as a modular way of enriching existing semantic parsers.