Domain adaptive English aspect word extraction method based on bidirectional long and short-term memory network and multi-head attention mechanism

• Published in: Journal of Applied Science and Engineering Vol. 28; no. 12; pp. 2461 - 2469
• Main Author: Tianxiao Wang
• Format: Journal Article
• Language: English
• Published: Tamkang University Press 01.06.2025
• Subjects: bert; bidirectional long and short-term memory network; english aspect word extraction; multi-head attention mechanism
• ISSN: 2708-9967; 2708-9975
• DOI: 10.6180/jase.202512_28(12).0013

English aspect word extraction is a core task of aspect level sentiment analysis. With the continuous development of social networks, more users tend to make decisions based on comment text, and pay more attention to the details of comment text. Therefore, it is of great significance to accurately extract aspect words from massive review texts for users to make quick decisions. Because the annotation corpus is extremely time-consuming, labor-intensive and costly, there are relatively few aspect word datasets in public currently, which affects the effective training of neural network models. To solve this problem, this paper proposes a novel domain adaptive English aspect word extraction method based on bidirectional long and short-term memory network and multi-head attention mechanism. Firstly, the dual labels of aspect emotion and aspect extraction are used for labeling. Secondly, by parallel aspect extraction and aspect emotion classification task channel, it uses the BERT, bidirectional long and short term memory networks (Bi-LSTM) and multi-head self-attention to extract deeper semantic information, near and far feature information. Finally, conditional random field (CRF) classifier and Softmax classifier are used for classification. In order to verify the effectiveness of the new method, experiments are carried out on Laptop, Restaurant and Device datasets respectively. Experimental results show that the proposed method performs better than multiple baseline models.