Text-to-SQL for Fact Extraction and Verification Over Tabular Data

This thesis presents an integrated approach for extracting and validating facts from tabular data sourced from the Feverous dataset, which comprises Wikipedia pages. Beyond translating natural language queries into SQL statements optimized for dataset structures, this methodology integrates a retriever component to efficiently identify relevant data entries. The study evaluates the efficacy of this integrated approach in accurately extracting and validating facts from tabular data. Furthermore, the research explores the application of retrieval models, followed by sequence-to-sequence (seq2seq) models and language model (LLM) prompts for constructing knowledge bases, extracting information, and implementing query answering systems.