[pt] A tarefa chamada brevemente de Texto-para-SQL envolve a geração de uma consulta SQL com base em um banco de dados relacional e uma pergunta em linguagem natural. Embora os rankings de benchmarks conhecidos indiquem que Modelos de Linguagem Grandes (LLMs) se destacam nessa tarefa, eles são avaliados em bancos de dados com esquemas bastante simples. Esta dissertação investiga inicialmente o desempenho de modelos Texto-para-SQL baseados em LLMs em um banco de dados disponível ao público (Mondial)com um esquema conceitual complexo e um conjunto de 100 perguntas em Linguagem Natural (NL). Executando sob GPT-3.5 e GPT-4, os resultados deste primeiro experimento mostram que as ferramentas baseadas em LLM têm desempenho significativamente inferior ao relatado nesses benchmarks e enfrentam dificuldades com a vinculação de esquemas e joins, sugerindo que o esquema relacional pode não ser adequado para LLMs. Essa dissertação propõe então o uso de visões e descrições de dados amigáveis ao LLM para melhorara precisão na tarefa Texto-para-SQL. Em um segundo experimento, usando a estratégia com melhor performance, custo e benefício do experimento anterior e outro conjunto com 100 perguntas sobre um banco de dados do mundo real, os resultados mostram que a abordagem proposta é suficiente para melhorar consideravelmente a precisão da estratégia de prompt. Esse trabalho conclui com uma discussão dos resultados obtidos e sugere abordagens adicionais para simplificar a tarefa de Texto-para-SQL. / [en] The Text-to-SQL task involves generating an SQL query based on a given relational database and a Natural Language (NL) question. While the leaderboards of well-known benchmarks indicate that Large Language Models (LLMs) excel in this task, they are evaluated on databases with simpler schemas. This dissertation first investigates the performance of LLM-based Text-to-SQL models on a complex and openly available database (Mondial) with a large schema and a set of 100 NL questions. Running under GPT-3.5 and GPT-4, the results of this first experiment show that the performance of LLM-based tools is significantly less than that reported in the benchmarks and that these tools struggle with schema linking and joins, suggesting that the relational schema may not be suitable for LLMs. This dissertation then proposes using LLM-friendly views and data descriptions for better accuracy in the Text-to-SQL task. In a second experiment, using the strategy with better performance, cost and benefit from the previous experiment and another set with 100 questions over a real-world database, the results show that the proposed approach is sufficient to considerably improve the accuracy of the prompt strategy. This work concludes with a discussion of the results obtained and suggests further approaches to simplify the Text-to-SQL task.

The objective of this thesis is to explore technologies and solutions and see if it is possible to make a logistical flow more efficient. The logistical flow consists of a database containing materiel for purchase or reparation. As of now, searches may either result in too many results, of which several are irrelevant, or no results at all. The search needs to be very specific to retrieve the exact item, which requires extensive knowledge about the database and its contents. Areas that will be explored include Natural Language Processing and Machine Learning techniques. To solve this, a literature study will be conducted to gain insights into existing work and possible solutions. Exploratory Data Analysis will be used to understand the patterns and limitations of the data.

Text-to-SQL, which involves translating natural language into Structured Query Language (SQL), is crucial for enabling broad access to structured databases without expert knowledge. However, designing models for such tasks is challenging due to numerous factors, including the presence of ’noise,’ such as ambiguous questions and syntactical errors. This thesis provides an in-depth analysis of the distribution and types of noise in the widely used BIRD-Bench benchmark and the impact of noise on models. While BIRD-Bench was created to model dirty and noisy database values, it was not created to contain noise and errors in the questions and gold queries. We found after a manual evaluation that noise in questions and gold queries are highly prevalent in the financial domain of the dataset, and a further analysis of the other domains indicate the presence of noise in other parts as well. The presence of incorrect gold SQL queries, which then generate incorrect gold answers, has a significant impact on the benchmark’s reliability. Surprisingly, when evaluating models on corrected SQL queries, zero-shot baselines surpassed the performance of state-of-the-art prompting methods. The thesis then introduces the concept of classifying noise in natural language questions, aiming to prevent the entry of noisy questions into text-to-SQL models and to annotate noise in existing datasets. Experiments using GPT-3.5 and GPT-4 on a manually annotated dataset demonstrated the viability of this approach, with classifiers achieving up to 0.81 recall and 80% accuracy. Additionally, the thesis explored the use of LLMs for automatically correcting faulty SQL queries. This showed a 100% success rate for specific query corrections, highlighting the potential for LLMs in improving dataset quality. We conclude that informative noise labels and reliable benchmarks are crucial to developing new Text-to-SQL methods that can handle varying types of noise.