Retrieval methods for legal question answering in German

This thesis investigates the effectiveness of various Information Retrieval (IR) Systems in addressing legal questions formulated in layman’s terms within the German legal domain. The study utilizes the GerLayQA[3] dataset, comprising legal questions sourced from the popular online platform “Frag-einen-Anwalt.de” aiming to evaluate the capability of IR systems to retrieve relevant paragraphs from German legal texts in response to these queries. The research employs a diverse range of retrieval methods, including baseline approaches such as TF-IDF and BM25 and more advanced techniques based on state-of-the-art language models. A significant focus is placed on comparing the performance of pre-trained and fine-tuned Bi-Encoder models and highlighting the impact of domain-specific training on retrieval accuracy. Furthermore, the study explores the potential of a Retrieve-and-Re-Rank pipeline combining the efficiency of Bi-Encoders with the precision of Cross-Encoders. This approach is evaluated against single-stage retrieval methods to assess its viability in the legal domain. This work makes a novel contribution by developing and evaluating a Majority Vote IR model, which aggregates results from multiple retrieval methods through a majority voting mechanism and thereby tries to minimize the effect of any single IR system weakness. This ensemble approach is compared against individual models to determine its effectiveness in improving retrieval performance. The performance of these systems is rigorously evaluated using standard information retrieval metrics, including the Precision@k, Recall@k, F1@k-Score, and the Normalized Discounted Cumulative Gain (NDCG). The results are critically analyzed to identify the strengths and limitations of each approach within the context of German Legal Information Retrieval. This research aims to contribute to the growing field of legal AI by providing insights into the applicability and challenges of various IR techniques for processing German layman legal queries. The findings have implications for developing more accessible and efficient legal information systems, potentially closing the gap between complex legal language and public understanding.