Advanced techniques for cross-language annotation projection in legal texts

Nowadays, the majority of the services we benefit from, are provided online and their use is regulated by the acceptance to the terms of service by the users. All our data are handled accordingly with the clauses of such document and all our behaviours must comply with it. Given so, it would be very useful to find automated techniques to ensure fairness of the document or inform the users about possible threats. The focus of this work, is to create resources aimed to the development of such tools in languages other than English, which may lack in linguistic resources and annotated corpus. The enormous breakthroughs of the last years in Natural Language Processing techniques made it possible the creation of such tools through automated and unsupervised process. One of the means to achieve that is through the annotation projection between two parallel corpora. The difficulties and costs of creating ad hoc resource for every language has brought the need to find another way for achieving the goal.\\ This work investigates the cross language annotation projection technique based on sentence embedding and similarity metrics to find matches between sentences. Several combination of methods and algorithms are compared, among which there are monolingual and multilingual embedding neural models. The experiments are conducted on two datasets, where the reference language is always English and the projection are evaluated on Italian, German and Polish. The results obtained provide a robust and reliable technique for the task and a good starting point to build multilingual tools.