This article offers a long short-term memory (LSTM) based structured prediction model taking into account existing approaches to sequence tagging tasks and allowing for extraction of user opinions from reviews. We propose a model configuration and state transition rules which allow us to use past predictions of the model alongside sentence features. We create a body of annotated user reviews about mobile phones from Amazon for model training and evaluation. The model trained on reviews corpus with recommended hyperparameter values. Experiment shows that the proposed model has a 4.51% increase in the F1 score for aspects detection and a 5.44% increase for aspect descriptions compared to the conditional random field (CRF) model with the use of LSTM when F1 spans are matched strictly.
      The extraction of user opinions on mobile phones from reviews outside of the collected corpus was conducted as practical confirmation of the proposed model. In addition, opinions from other product categories like skin care products, TVs and tablets were extracted. The examples show that the model can successfully extract user opinions from different kinds of reviews. The results obtained can be useful for computational linguists and machine learning professionals, heads and managers of online stores for consumer preference determination, product recommendations and for providing rich catalog searching tools.

This study was conducted under government order of the Ministry of Education and Science of Russia, project No. 8.8184.2017/8.9