A Neuro-Symbolic Framework for Conflict Detection in the Swedish Legal Domain : A Knowledge-Driven Approach Combining Natural Language Processing and Logic Programming

Sammanfattning

The manual process of identifying and interpreting conflicts within the complex structures of legal documents is a time-consuming and laborintensive task. This thesis addresses this need by proposing an automated solution for conflict detection that assists to streamline regulatory management. Additionally the thesis presents a neuro-symbolic AI framework designed to tackle this problem, with a focus on its application to non-universal languages like Swedish. A proof of concept (PoC) was developed to detect contradictions in rudimentary text. Due to a lack of available data, Large LanguageModels (LLMs) were utilized as intermediary Named Entity Recognition (NER) model to facilitate the creation of a usable dataset. The proposed framework employs Natural Language Processing (NLP), including NER, to identify key textual elements such as parties, obligations, and alternative options. A symbolic reasoning engine, powered by Prolog, is then used to detect contradictions and inconsistencies based on a set of logical rules. While this initial implementation revealed overfitting and a high rule translation error rate, the thesis explores these issues and their potential solutions. The work provides a foundational pipeline for future research to build upon, paving the way for more sophisticated automated legal analysis.