Civil law jurisdictions share historical and doctrinal and philosophical roots, and are characterized by broadly similar features that have not been capitalized to date in research. These include an emphasis on legislative norms (as opposed to customary law and precedent); a mix of inquisitive and adversarial proceedings; a strong role reserved for the judge (as opposed, e.g., to jury-based fact determination); a highly-formalized style of judgment centred around syllogisms (as opposed to common law narratives); and a language and formulae specific to legal pleadings and concepts. As a result, civil court judgments and approaches to the law tend to be similar, regardless of the court’s seat or the language of the proceedings.

In this context, we examine the possibility of developing multilingual AI methods that can be transferred and generalized across civil jurisdictions to leverage the advances achieved in data-rich environments to data-poor ones. Recent research has evidenced that language models, especially when fine-tuned to be domain and/or language specific, can achieve strong performances on a variety of tasks that were once believed to be the sole remit of legal professionals. Such models, however, often rely on substantial amounts of data, and are thus presumably ill-fit for jurisdictions where the data is either unavailable and/or the available dataset is too small, and not expected to grow significantly (e.g., paper-based developing countries). What these jurisdictions have in common, however, is often a language with another, data-rich jurisdiction, and roots in a legal family, such as civil law. For instance, Liechtenstein’s relatively-small jurisprudence is all in German, and of the civil-law type; while training a language model on the Principality’s legal dataset would be challenging, it should be possible to rely on the (broader) case law of its German-speaking European neighbors.

In this light, the project starts with an exploratory study on a multilingual dataset from Belgian court decisions to examine the possibility of transferring and generalizing models developed in that jurisdiction to other jurisdictions sharing the same languages and civil law tradition. The models will be tested, and if needed retrained, with datasets made up of court decisions from Germany, the Netherlands and France. Belgian courts decisions represent an optimal dataset, as the state’s jurisprudence is both broad and multi-lingual: the resulting models, based on Dutch, French, and German court judgments, will then more easily be checked and benchmarked against that of its neighbors, namely the Netherlands, France, and Germany.

To assess the promises (and limitations) of this approach, we explore four main tasks, which have been at the forefront of the use of artificial intelligence in the judicial domain: anonymization, argument mining, predictive models, and legal explainability (i.e., justification of decision by legal reasoning). These tasks are characterized by a high reliance on written, text data, in a way that is meant to associate textual symbols (e.g., the court’s reasoning) to an outcome – making them ripe for computer-assisted analysis. These tasks are also, to some extent interrelated: detecting arguments helps in understanding roles in a judgment, which in turn informs anonymization (e.g., where only claimants need to be anonymized, but not judges), predictive justice (to identify the determinants of a given case), and legal explainability.

As such, it will be essential to identify what kind of model (fine-tuned to a particular task or generalist) performs the best, and in what conditions. In this light, the research proposal will focus on testing the model developed in several contexts, and checked against several benchmarks, be they off-the-shelf or designed specifically for this research endeavor.