BERT-based implicit aspect extraction

The traditional implicit aspect extraction methods require too much manual feature engineering, which are inefficient when processing large-scale data. To address the problem, BERT-based model with different kinds of classifiers is proposed. In this paper, the text are inputted into BERT to get the embeddings. Then the embeddings of the final hidden layer are combined with different classifiers to obtain implicit aspect. Five classific text classifiers are used for comparison. The experimental results show that the method in this paper outperform state-of-the-art works. The accuracy and the macro-F1 score on the SemEval implicit extraction data set can reach 78.11% and 73.19% respectively.