Multi-label Sentiment Analysis Base on BERT with modified TF-IDF
Zeyi Jin, Xin Lai, Jingjig Cao
Abstract
With the evolution of web technologies, sentiment analysis, especially aspect-based sentiment analysis (ABSA), as a technology to identify and extract user opinions, has attracted widespread attention. According to the characteristics of user comments, we convert ABSA problem into a multi-label classification problem, and propose a classification model based on BERT with modified TF-IDF for ABSA. In the feature extraction phase, a modified TF-IDF method is proposed to better reflect the words' importance in multi-label classification problem by calculating the different weights in individual classes. And a new feature is generated by combining the BERT embedding with modified TF-IDF for classification task. Then the feature is input into a fully connected layer to fine-tune BERT model for the multi-label classification task. The effectiveness of supervised TF-IDF and proposed model is validated by experiments of multilabel classification on a restaurant reviews dataset.