خلاصة:
Nowadays the amount of textual information on the web is grown rapidly. The huge textual data needs more accurate classification algorithms. Sentiment analysis is a branch of text classification that is used to classify user opinions in case of market decisions, product evaluations or measuring consumer confidence. With the rise of the production rate of Persian text data in a commercial area, improvement of the efficiency of algorithms in Persian is a must. The structure of the Persian language such as word and sentence structures poses some challenges in this area. Deep learning algorithms are recently used in NLP and especially sentiment text classification for many dominant languages like Persian. The goal is to improve the performance of classification using deep learning issues. In this work, the authors proposed a hybrid method by a combination of structural correspondence learning (SCL) and convolutional neural network (CNN). The SCL method selects the most effective pivot features so the adaptation from one domain to similar ones cannot drop the efficiency drastically. The results showed that the proposed hybrid method that is learned from one domain can act efficiently in a similar domain. The result showed that applying a combination of SCL+CNN can improve the result of sentiment classification for two domains more than 10 percent.
ملخص الجهاز:
The application of Deep Learning in Persian Documents Sentiment Analysis Mohammad Bag her Dastgheib Sara Koleini Department of Designing & System Department of Information & Communication Operation, Regional Information Center for Technology, Regional Information Center for Science and Technology, RICeST, IRAN.
In this work, the authors proposed a hybrid method by a combination of structural correspondence learning (SCL) and convolutional neural network (CNN).
Keywords: Deep learning, Persian Documents, Sentiment Analysis, Convolutional Neural Network (CNN), Structural Correspondence Learning (SCL).
Sentiment analysis is the process of recognizing, extract, evaluate, and study of affective states and subjective information by using different methods such as machine learning or statistical techniques.
The sentiment analysis methods can be categorized into statistical / machine learning, knowledge /lexicon based and hybrid approaches as shown in Figure 1 (Shahnawaz & Astya, 2017).
Hybrid method uses statistical/machine learning and lexicon-based techniques to increase the performance and precision of sentiment analysis task as applied in this work.
In this research, the sentiment classification process for the Persian language is performed by deep learning method.
The inadequacy of resources and sentiment data forced Persian domain researchers to use machine learning methods.
Experiments and Results As mentioned in section 3, to classify sentiment of sentences we proposed a hybrid algorithm that uses domain adaptation by SCL and convolutional neural network for sentiment analysis classification.
The results of adapting laptop/mobile domains in sentiment analysis As shown in table 4, the proposed hybrid method (CNN+SCL) has the maximum F- measure score.
Persian sentiment analysis using domain adaptation by Structural Correspondence Learning.