Corresponding Author

Liyi Zhang, Wuhan University, China,

Document Type



Online product reviews contain a lot of valuable information regarding product problems, which are very useful for producers to find product pain points and improve product quality. However, many studies focus only on the sentiment polarity of the product aspect, ignoring specific product problem information in online reviews. In this paper, product aspects and specific problem information are extracted from online reviews to help producers find the specific pain points of products. We call this task Review Problem Mining (RPM). At the same time, existing methods of review information extraction depend heavily on manually constructed features or large-scale data. To address these limitations, we proposed a new joint model BERT-CRF which integrates the popular pre-trained language model BERT and conditional random fields (CRF). The proposed method introduces external knowledge through BERT to reduce the model’s dependence on training data and uses CRF to model the dependencies among tags. To verify the validity of our method, we constructed a dataset from and carried out the experiments. Experimental results show that the proposed method is highly effective.