Criticality Prediction and Information Retrieval in Swiss Legal Data

We introduce two novel tasks for legal Natural Language Processing (NLP), aiming to add diversity to the field. As certain cases may have a more significant impact on jurisdiction than others, legal experts are interested in predicting the level of controversy surrounding a case. Additionally, being able to retrieve relevant laws or Leading Decision (BGE) a case depends on could greatly reduce court costs. The potential use cases for both tasks are numerous and not only include reducing court delays but also prioritizing cases for better judgments. To our knowledge, there have been no previous attempts to predict the criticality of a legal case or the retrieve relevant laws and BGE in Switzerland.
To address this, we publicly release two multilingual datasets containing cases from the Federal Supreme Court of Switzerland. The Swiss-Criticality-Prediction dataset features two approaches for labeling cases as critical, while the Swiss-Doc2doc-IR dataset includes links for each case to cited laws and BGE. In the Criticality task, we assess the performance of multilingual BERT-based models. For the Information Retrieval task, we evaluate existing models from the BEIR benchmark. We observe that all models fall short in their performance. Nonetheless, we find that domain- specific pre-training proves to be advantageous for both tasks.