Research on Text Sentiment Analysis Method Based on Features Combined with Dual Channel Convolutional Neural Network

Abstract

E-commerce platform hold vast, valuable and unstructured comment text. Merchants can use
this type of review text to improve goods and services, etc. However, only a deep understanding
of natural language can better analyze this type of data. Taking into account that traditional
word vectors as input convolutional neural network cannot utilize the unique information of
emotional feature in the sentiment analysis task, and it’s difficult to recognize the same word as
different parts of different sentences. In this paper, a sentiment analysis model for Chinese
product reviews is proposed, which is based on the features of speech and viewpoint, it’s
combined with convolutional neural network. First of all, we use part-of-speech, dependency
parsing and semantic dependency parsing to formulate rules for extracting viewpoint features.
Secondly, based on the representation of word vectors, the part-of-speech feature and viewpoint
feature can be introduced by vector stitching. Finally, the word vectors and the extended feature
vectors are used as the two input channels of sentiment analysis for convolutional neural
network. This method uses Tan Songbo’s public hotel review data set for evaluation, and shows
that it has better classification performance based on the comparison between traditional text
convolutional neural networks and other classifiers.