Detection of Missed Links in Large Legal Corpora

Legal texts have a large number of explicit and implicit links to other documents. Extraction of those links would help to systematize and analyze lawmaking activity better. Legal documents are often multi-topic and fragmented, making standard topical similarity search methods useless to reveal those links. In this paper, we propose Siamese network-based approaches, which can tackle that problem. Our approaches incorporate SentenceBERT and Doc2Vec models and generate document-level embeddings, which can be helpful to build a large-scale legal information retrieval system. The experiments on Russian and multilingual legal corpora confirm the applicability of the proposed methods. Namely, SentenceBERT-based models show the best performance, although they have to be fine-tuned on a multilingual labeled corpus.