A feature integration method for Chinese chunking
Chen Lyu, Xuejing Xu, Jiangping Huang · 2023
Chunking is a crucial task in natural language processing. The task aims to divide a text into syntactically correlated non-overlapping chunks. The discrete model and neural model can be applied to Chinese chunking. The feature representations of them are different and both can achieve excellent performance. In this paper, we build both of the models and make a detailed comparison between the two models. Furthermore, we present the feature integration method to integrate the advantages of the two models. Experiments show the effectiveness of the feature integration method. Since the neural model utilizes pre-trained word embeddings, it can be regarded as a semi-supervised learning method. In order to make a fairer comparison between the discrete model and neural model, we incorporate word clusters into these models. Experimental results show that the word cluster information does not significantly improve chunking performance and the feature integration method still improves the performance of both the discrete model and neural model.