Opinion mining in Persian language using a hybrid feature extraction approach based on convolutional neural network
Shima Zobeidi,Marjan Naderan,Seyyed Enayatallah Alavi
DOI: https://doi.org/10.1007/s11042-019-07993-4
IF: 2.577
2019-08-02
Multimedia Tools and Applications
Abstract:Nowadays, huge amounts of text data are generated due to the increase of communications, over various web sites and applications. Evaluation and extraction of information from these data is an important task, one way of which is named opinion mining. The purpose of this paper is sentiment analysis of users' opinions about various products. The proposed system classifies opinions at the sentence level based on emotions into two and multiple classes by deep learning methods. To this end, three main phases are taken: the first step contains sentences preparation for the input matrix which itself is accomplished in two levels: word-level and character-level. In word-level, each word in each sentence is given to the word2vec algorithm. In character-level, for each character in each sentence, the proposed method computes a numerical vector and creates a matrix. Next, the feature extraction part is executed which includes a Convolutional Neural Network (CNN). The generated matrices in the previous levels for each sentence are given to the CNN for embedding each sentence and therefore, utilizing both word2vec and CNN for extracting features. In the final step, the generated vectors are given to the Bidirectional Long Short Term Memory (Bi-LSTM) network for sentiment classification, not used in any of the previous methods. The performance of the proposed algorithm has been investigated on the Digikala Persian dataset on mobile and digital cameras. Results show that the proposed algorithm reaches an accuracy of 95% for two classes and 92% for multi-class classification which is comparable with previous algorithms.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering