Multi-classfication Sentiment Analysis Based on the Fused Model
Yingbin Xue, Xiaoye Wang, Zan Gao · 2019
The traditional methods of sentiment classification usually see the data has positive and negative two kinds of attitudes only. But actually, the real data has multi-category sentiment, is positive, negative, neutral and not mentioned four classes. Therefore, when using common classifying methods to analyzing the data sentiment, if the number of few class data is too scarce, it is difficult to learn useful information from them and the final classifying result will tend to most classes. In order to obtain accurate classification results, this paper proposes a multi-classification method based on the combination of Bert (Bidirectional Encoder Representation from Transformers) model and Liblinear (A Library for Large Linear Classification) model (It is abbreviated as B-Liblinear). Due to the Bert model's breakthrough in data preprocessing, this paper prepressed training data set, and obtained the word vector and sentence vectors from data. Next, combined with attribute label and sentiment tendency data, the unstructured data was converted into a structured training data set. It was as the standard input data of Liblinear model to construct a classification model. This model's classification mechanism is "one vs. rest", it can effectively solve the heavy class imbalance problem of massive data in multiple classification tasks. In this paper, the classification result of B-Liblinear model was compared with several classical multi-classification methods. And the experimental results show that the combination of Bert model and Liblinear of dealing with the text multi-classification problem is more accurate.