A New Feature Fusion Method Based on Pre-Training Model for Sequence Labeling
Shichuan Yu, Yan Yang · 2023
To fuse vocabulary features into the pre-training model is the mainstream data feature processing method for sequence labelling tasks. In general, the feature fusion methods that have been proposed at present are direct fusion outside the pre-training model or fusion of lexical features using attention mechanism. However, the study found that this way of vocabulary enhancement does not conform to the word formation rules of modern Chinese. In the Chinese language, it is easy to fuse irrelevant or even incorrect lexical features into the sequence using the above feature processing methods, which is bad for the experimental results of the Chinese sequence labelling task. To solve these problems, we propose to use Cosine Similarity Adapter to process lexical features in Chinese sequence labelling tasks. CSBERT is a hybrid model using this structure based on BERT, which conforms to the word formation rules of modern Chinese to a certain extent. It can fuse the features of the word into the character or eliminate the features of the word in the character according to the cosine similarity between the character vector and a word vector. The experimental results show that CSBERT has better ability to label Chinese sequences than the benchmark model. CSBERT has achieved the best experimental results such as F1-Score on 7 open datasets and the best ability of multi-label classification, which proves that the model has good practical value.