Searching for legal documents is a specialized Information Retrieval task that is relevant for expert users (lawyers and their assistants) and for non-expert users. By searching previous court decisions (cases), a user can better prepare the legal reasoning of a new case. Being able to search using a natural language text snippet instead of a more artificial query could help to prevent query formulation issues. Also, if semantic similarity could be modeled beyond exact lexical matches, more relevant results can be found even if the query terms don’t match exactly. For this domain, we formulated a task to compare different ways of modeling semantic similarity at paragraph level, using neural and non-neural systems. We compared systems that encode the query and the search collection paragraphs as vectors, enabling the use of cosine similarity for results ranking. After building a German dataset for cases and statutes from Switzerland, and extracting citations from cases to statutes, we developed an algorithm for estimating semantic similarity at paragraph level, using a link-based similarity method. When evaluating different systems in this way, we find that semantic similarity modeling by neural systems can be boosted with an extended attention mask that quenches noise in the inputs.

CoSimLex is a dataset that can be used to evaluate the ability of context-dependent word embed- dings for modeling subtle, graded changes of meaning, as perceived by humans during reading. At SemEval-2020, task 3, subtask 1 is about ”predicting the (graded) effect of context in word similarity”, using CoSimLex to quantify such a change of similarity for a pair of words, from one context to another. Here, a meaning shift is composed of two aspects, a) discrete changes observed between different word senses, and b) more subtle changes of meaning representation that are not captured in those discrete changes. Therefore, this SemEval task was designed to allow the evaluation of systems that can deal with a mix of both situations of semantic shift, as they occur in the human perception of meaning. The described system was developed to improve the BERT baseline provided with the task, by reducing distortions in the BERT semantic space, compared to the human semantic space. To this end, complementarity between 768- and 1024-dimensional BERT embeddings, and average word sense vectors were used. With this system, after some fine-tuning, the baseline performance of 0.705 (uncentered Pearson correlation with human semantic shift data from 27 annotators) was enhanced by more than 6%, to 0.7645. We hope that this work can make a contribution to further our understanding of the semantic vector space of human perception, as it can be modeled with context-dependent word embeddings in natural language processing systems.