Authors:
Vasily Derbentsev
1
;
Vitalii Bezkorovainyi
1
;
Andriy Matviychuk
1
;
Oksana Pomazun
1
;
Andrii Hrabariev
1
and
Alexey Hostryk
2
Affiliations:
1
Kyiv National Economic University named after Vadym Hetman, 54/1 Peremogy Ave., Kyiv, 03680, Ukraine
;
2
Odessa National Economic University, 8 Preobrazhenskaya Str., Odessa, 65082, Ukraine
Keyword(s):
Sentiment Analysis, Social Media, Deep Learning, Convolutional Neural Networks, Long Short-Term Memory, Word Embeddings.
Abstract:
This paper describes Deep Learning approach of sentiment analyses which is an active research subject in the domain of Natural Language Processing. For this purpose we have developed three models based on Deep Neural Networks (DNNs): Convolutional Neural Network (CNN), and two models that combine convolutional and recurrent layers based on Long-Short-Term Memory (LSTM), such as CNN-LSTM and Bi-Directional LSTM-CNN (BiLSTM-CNN). As vector representations of words were used GloVe and Word2vec word embeddings. To evaluate the performance of the models, were used IMDb Movie Reviews and Twitter Sentiment 140 datasets, and as a baseline classifier was used Logistic Regression. The best result for IMDb dataset was obtained using CNN model (accuracy 90.1%), and for Sentiment 140 the model based on BiLSTM-CNN showed the highest accuracy (82.1%) correspondinly. The accuracy of the proposed models is a quite acceptable for practical use and comparable to state of the art models.