In its most general form, structured data represents information arranged according to predefined schemas that make it machine-readable and systematically interpretable. It can support reasoning, grounding, and it can be easily automatically processed. With the rise of large language models, these properties are becoming increasingly valuable. Structured data offers a way to make model behavior more reliable, interpretable, and efficient. As a result, the ability to work effectively with structured data, both to use it and to generate it, has become an increasingly important direction in modern natural language processing.
This thesis explores how structured data can be used to enhance natural language processing (NLP) tasks, and how language models can be improved to generate structured data more accurately. Throughout, a strong emphasis is placed on efficiency and scalability, as real-world applications depend on systems that are not only capable but also often small, specialized, and resource-aware.
The work is divided into three main technical parts. The first part investigates how existing structured resources in the Wikipedia ecosystem can be exploited to support applications with wide-reaching community impact. By enriching textual input with structured and cross-lingual information, we demonstrate how these methods can improve content generation quality and accessibility, particularly for low-resource languages, thereby contributing to narrowing knowledge gaps across language editions. The second part focuses on improving the generation of structured data itself. It addresses key challenges such as the lack of high-quality datasets for training and evaluation and the need to handle dynamic constraints in evolving knowledge bases. The proposed methods aim to make structured NLP systems more accurate, flexible, and efficient for practical tasks. The third part addresses the rising inference costs associated with the modern (large) language models. In this part we propose a model routing strategy that can be applied to a broad range of NLP tasks, including structured ones, that maintains a high performance while lowering the costs significantly.
Together, these contributions highlight how structured data can act as both an input and an output signal by guiding, constraining, and grounding generative models while benefiting from them in return. This offers a foundation for the next generation of NLP systems: models that are efficient, interpretable, and capable of integrating structured knowledge seamlessly into real-world applications.
