Research on Chinese Keyword Extraction Based on the ALBERT Pre-Training Model
Xiangcheng Zhang, Jianfeng Cui, Xu Zhang · 2022
The classic word embedding-based Chinese keyword extraction techniques have a number of drawbacks, including poor word segmentation and an inability to recognize word ambiguity. An efficient ALBERT-BiGRU-CRF-based Chinese keyword extraction model has been presented to address the aforementioned issues. The model first obtains dynamic Chinese character vector representation based on ALBERT pre-training model, fuses lexical information by lexical enhancement method, then extracts contextual long-range semantic features using bidirectional gated recurrent units, and finally introduces conditional random fields to calculate the global optimal solution to obtain the final keywords. The experimental findings indicate that the model’s Fl value is 81.36%, which is higher than that of the conventional keyword extraction models and demonstrates that the suggested technique can successfully enhance Chinese keyword extraction performance.