Attentive deep neural networks for legal document retrieval

Legal text retrieval serves as a key component in wide range of legal processing tasks such question answering, case entailment, and statute law retrieval. The performance depends, to large extent, on the representation text, both query documents. Based good representations, model can effectively match its relevant Because documents often contain long articles only some parts are queries, it is quite challenge for existing models represent In this paper, we study use attentive neural network-based document We propose general approach using deep networks with attention mechanisms. it, develop two hierarchical architectures sparse sentences articles, name them Attentive CNN Paraformer. methods evaluated datasets different sizes characteristics English, Japanese, Vietnamese. Experimental results show that: (i) substantially outperform non-neural terms across languages; (ii) Pretrained transformer-based achieve better accuracy small at cost high computational complexity while lighter weight achieves datasets; (iii) Our proposed Paraformer outperforms state-of-the-art COLIEE dataset, achieving highest recall F2 scores top-N task.