Analysis of Chinese Named Entity Recognition Results Based on Different Paradigm

Abstract

Named-entity recognition (NER) is a sub-task of information extraction. Currently, there are various approaches to achieve the NER task and the results are getting better. However, the internal structure of the models is becoming more and more complex, making them much less comparable. In view of the above reasons, we categorize NER methods and implement three different NER paradigms (classification-based model, sequence annotation-based model, and pointer network-based model) using the simplest structure for three Chinese NER datasets (PeopleDailyNER, FinanceSinaNER, and CLUENER). We find that the pointer network-based model works well for datasets with short sentence lengths and large data sizes; the classification-based model is good for datasets with large data sizes and numerous entity categories; in the vast majority of cases, the sequence annotation-based model has better entity recognition performance than the other two models.