Aspect Term Extraction Based on BiLSTM-CRF Model

Abstract

Aspect term extraction is an important subtask in aspect sentiment analysis, and it is a necessary step to complete other subtasks. Existing studies focus on complex and changeable models and only use single dataset for training, which is not conducive to the research and has no good to the application of aspect term extraction task. Therefore, this paper seeks for a simple and effective model to complete the task. We transformed the aspect term extraction task into sequence tagging task, and applied the BiLSTM-CRF model to extract the aspect terms. Experiment results and case studies showed that the F1 score of proposed model on the laptop dataset is 80.13, which is 6.35 higher than the best baseline model. On the restaurant dataset, the F1 score reached 85.2, 1.19 higher than the best baseline model. It proved that the BiLSTM-CRF has better performance than the baseline, and had greater advantages on multiple aspect words recognition. In addition, we applied BiLSTM-CRF model to our practical task, and constructed an aspect-level Yelp dataset in a semi-supervised method. The parameter setting of the method was discussed.