Querying Databases with Natural Language: The use of Large Language Models for Text-to-SQL tasks

The Text-to-SQL task involves generating SQL queries based on a given relational database and a Natural Language (NL) question. Although Large Language Models (LLMs) show good performance on well-known benchmarks, they are evaluated on databases with simpler schemas. This dissertation first evaluates their effectiveness on a complex and openly available database (Mondial) using GPT-3.5 and GPT-4. The results indicate that LLM-based models perform poorly and struggle with schema linking and joins. To improve accuracy, this work proposes the use of LLM-friendly views and data descriptions. A second experiment on a real-world database confirms that this approach enhances the accuracy of the Text-to-SQL task.