Using Ontological Knowledge and Large Language Model Vector Similarities to Extract Relevant Concepts in VAT-Related Legal Judgments

In this paper, we present OntoVAT, a multilingual ontology designed for extracting knowledge in legal judgments related to VAT (Value-Added Tax). This is, to our knowledge, the first extensive ontol- ogy in the VAT domain. OntoVAT aims to encapsulate critical concepts in the European VAT area and offers a scalable and reusable knowledge structure to support the automatic identification of VAT-specific con- cepts in legal texts. Additionally, OntoVAT supports various Artificial Intelligence and Law (AI&Law) tasks, such as extracting legal knowledge, identifying keywords, modeling topics, and extracting semantic relations. Developed using OWL with SKOS lexicalization, OntoVAT’s initial ver- sion includes ontological patterns and relations. It is available in three languages, marking a collaborative effort between computer scientists and subject matter experts. In this work, we also present an application scenario where the knowledge encoded within OntoVAT is leveraged in combination with several recent Large Language Models (LLMs). For this application, for which we used the most powerful open source LLMs avail- able today (both generative and non-generative, including legal LLMs), we show the system’s design and some preliminary results.