Research on Entity Recognition Based on Multi-criteria Fusion Model
Qiyu Yin · 2021
In order to fully and comprehensively utilize the Chinese named entity recognition corpus marked according to different labeling criteria, this paper, by applying the word-based BERT bidirectional language model, introducing multi-criteria shared connection layer and conditional random field (CRF) system, and using Microsoft Research Asia MSRA-NER corpus and Peking University’s People’s Daily part-of-speech tagging corpus (RMRB-98-1), firstly formed each Chinese named entity recognition model separately, and then mixed a multi-criteria fusion model with two corpora. Experiments show that the recognition effect of multi-criteria fusion model is better than that of each corpus independent model, reaching 94.46% F1 value and 94.32% F1 value respectively on MSRA-NER and RMRB-98-1 corpus. However, the experiment still has its limit in the scale of the corpus involved in the fusion. Later, integration of more corpora, and combination with entity recognition tasks in specific fields such as biology and military will be further explored to enhance the recognition effect.