Research on Named Entity Recognition in Judicial Field Based on ERNIE-Gram

Named Entity Recognition is a key and fundamental task in natural language processing and can benefit many downstream tasks such as knowledge graph construction, question answering system, machine reading, etc. In view of the contradiction between the rapidly increasing of judgement documents and the low efficiency of manual analysis in the judicial field, we propose a NER model by using a combination of ERNIE-Gram, BiGRU, and CRF. ERNIE-Gram adopts multi-granularity n-gram language learning mechanism to learn the semantic of n-grams more adequately. Thus, we first employ the ERNIE-Gram to capture rich language representation, then we feed them into the BiGRU to obtain more important text features, finally, we use the CRF to decode and output optimal labeling sequence. We conduct experiments on an open dataset of the 2021 “Challenge of AI in Law” information extraction subtask and compare our model with currently familiar models for NER task. Experimental results demonstrate that our proposed model achieves a F1-score of 87.01%, and outperforms all the baseline models.