Theses

Despite being a fast-paced research field, text-to-SQL systems face critical challenges. The datasets used for the training and evaluation of these systems play a vital role in de- termining their performance as well as the progress in the field. In this work, we introduce a methodology for text-to-SQL dataset analysis, and we perform an in-depth analysis of several text-to-SQL datasets, providing valuable insights into their capabilities and limita- tions and how they affect training and evaluation of text-to-SQL systems. We investigate existing evaluation methods, and propose an informative system evaluation based on er- ror analysis. We show how our dataset analysis can help explain the behavior of a system on different datasets. Using our error analysis, we further show how we can pinpoint the sources of errors of a text-to-SQL system for a particular dataset and reveal opportunities for system improvements.

The most popular category of recommender systems is Collaborative Filtering (CF) recommenders, which serve as specialized information retrieval systems that generate personalized suggestions by leveraging user feedback. However, these systems have some intrinsic limitations. Firstly, they face the item cold-start problem, where new items lacking historical interactions are difficult to recommend. Secondly, these algorithms are notably sensitive to popularity bias, disproportionately favoring well-known items over less popular but, possibly more relevant, alternatives. To counter these challenges, the research community frequently extends CF algorithms by incorporating side information, resulting in the formation of hybrid recommender systems. Although this strategy partially resolves the aforementioned issues, it introduces an additional problem: the algorithms have a tendency to redundantly emphasize specific item relationships that are present in both user feedback and side information. To address this multifold problem, we introduce KoalaRec, a novel neighborhood learning-based recommendation model. KoalaRec incorporates an orthogonal regularization term, formulated on the principles of multiview subspace clustering techniques, to encourage diverse information extraction from both user feedback data and side information. This unique approach enables the algorithm to learn distinct item-item relationships from each data source, thereby mitigating the risk of redundant emphasis on particular item connections. We experimentally show that KoalaRec effectively addresses the issue of popularity bias, as the model outperforms other evaluated algorithms in terms of catalog coverage and the Gini coefficient, albeit at the cost of a minor reduction in recommendation accuracy in some cases.

While data guides and influences many human activities, the barriers posed by the tools that are needed to retrieve it, such as Structured Query Language (SQL), make data inaccessible for many users. To lift these barriers, researchers have been working on creating natural language interfaces that would allow users to access databases solely through natural language. Natural language interfaces employ Text-to-SQL systems that can translate a natural language question from the user to an SQL query that can retrieve the data they need. Recently, novel Text-to-SQL systems are adopting deep learning methods with very promising results. At the same time, several challenges remain open, making this area an active and flourishing field of research and development. To make real progress in building Text-to-SQL systems, we need to de-mystify what has been done, understand how and when each approach can be used, and, finally, identify the research challenges ahead of us. We present a detailed taxonomy of neural Text-to-SQL systems that will enable a deeper study of all the parts of such a system. This taxonomy will allow us to make a better comparison between different approaches, as well as highlight specific challenges in each step of the process, thus enabling researchers to better strategize their quest towards the “holy grail" of database accessibility. However, how can the user verify that the generated SQL query matches their intent if they are not familiar with SQL? To tackle this problem, a system that can translate the SQL query back to natural language is needed (also known as an SQL-to-Text system). We explore the SQL-to-Text problem, we examine its challenges and peculiarities, and present a Transformer-based model that can generate fluent query explanations. Additionally, we look into the difficulties of automatically evaluating the performance of such a system and we examine how different metrics behave in the SQL-to-Text setting.

Query optimization is a well-studied problem in the database community. Recently, deep learning methods have been applied either to assist the query optimizer on measures like cardinality estimation, computational cost prediction, and query execution time prediction or to implement neural-based optimizers from scratch. Despite the promising results, very few tackle multiple aspects of the optimizer at the same time or combine both the underlying data and a query workload. QPSeeker takes a step towards a neural database planner, encoding first the information provided from the workload and second the underlying tables, using the power of specially designed language models for tabular data. Next, it applies a special form of attention to combine these two sources to approximate the distributions of cardinalities, costs, and execution times of possible query plans. At inference, when a query is submitted to the database, QPSeeker uses its learned cost model and traverses the query plan space using Monte Carlo Tree Search to provide the best execution plan for the query.

Database democratization focuses on making databases accessible to non-expert users that are not familiar with database query languages like SQL. In this direction, a lot of effort has already been put into two problems: Text-to-SQL, which focuses on translating a natural language query to SQL, and SQL-to-Text, which is the inverse problem. However, work has lagged behind in explaining query results in natural language. We first define the Query Results-to-Text problem as: given the results of a query, produce a natural language verbalisation describing these results. Then we attempt solving Query Results-to-Text by defining a model, namely QR2T. We propose pretraining QR2T using real-world table datasets focusing on table understanding. We use a preprocessing step that transforms the query so that the query results, which are the input of our model, include additional information, which QR2T can utilize leading to a more informative verbalisation. Finally, we create two Query Results-to-Text benchmarks, which are the first datasets that contain query result verbalisations for both fine-tuning and evaluation.

SQL is the most popular and easy to use language for querying and analyzing large volumes of data in databases. Nowadays, databases are complex, containing several tables and attributes, making data exploration tasks difficult even for expert users. Our aim is to assist users in discovering the most interesting tables and attributes, while they are forming their queries. For that purpose, a recommendation system is proposed in the form of auto-completion. The proposed system uses Multi-Armed Bandits, a category of algorithms that balance exploration and exploitation, which are very suitable for interactive online learning. Furthermore, user information from past interactions can be used to create a context, making the recommendations personalized. Our system provides recommendations that speed up the process of writing SQL statements while learning from user feedback.

Databases contain a vast amount of data, used to support a range of operations, from business operations, scientific experiments to activities in our everyday lives. However they are still inaccessible for non-technical users, without knowledge of Structured Query Language (SQL). Natural language interfaces to databases lift these obstacles for such users and they have recently bloomed. In this thesis, we will start by presenting the NL2SQL problem (translating Natural Language to Structured Query Language), its most important aspects and the anatomy of a NL2SQL system. We will compare some systems and see how each one of them chooses to tackle the problem. In the main part of this work, we will focus on the SQLNet system which uses deep learning methods to tackle the NL2SQL problem. We will also test our own implementation of the system, investigate possible improvements and test how well it works on various cases.