Joint Reference and Relation Extraction from Legal Documents with Enhanced Decoder Input

Abstract This paper deals with an important task in legal text processing, namely reference and relation extraction from documents, which includes two subtasks: 1) extraction; 2) determination. Motivated by the fact that subtasks are related share common information, we propose a joint learning model solves simultaneously both subtasks. Our employs Transformer-based encoder-decoder architecture non-autoregressive decoding allows relaxing sequentiality of traditional seq2seq models extracting references relations one inference step. We also method to enrich decoder input learnable meaningful information therefore, improve accuracy. Experimental results on dataset consisting 5031 documents Vietnamese 61,446 show our proposed performs better than several strong baselines achieves F 1 score 99.4% for task.