Aspect Level Sentiment Analysis Using Bi-Directional LSTM Encoder with the Attention Mechanism

Abstract

Aspect level sentiment analysis (ALSA) is a fine-grained task in sentiment analysis. It classifies the different polarities (positive, negative) for the specific aspects of each review. In general, document and sentence level sentiment analysis has achieved remarkable results, but they can’t make the right prediction for the specific aspects in the given review. In the past few years, many researchers have worked sentiment analysis using machine learning approaches, but they have a problem in polarity detection for each aspect in the sentence. Therefore, this paper proposes to classify the sentiment polarity for each aspect by using Bi-directional LSTM (Bi-LSTM) encoder combining with the attention mechanism. The attention mechanism pays attention to the aspect of a specific target in the given review. It is a type of deep learning model and also an extension of the LSTM network, which can learn the long sequences of text. Deep learning models take a long time to train and when the network is deep, it encounters the vanishing gradient problem, a long-standing issue in the neural network model. This paper also considers the hyperparameter tuning approach to solve the vanishing gradient problem for ALSA task. Hyperparameter may be the weight, bias, the number of epochs, batch size, and so on. Experiments are conducted on IMDB and SemEval 2014 Task 4 datasets and the results show that the accuracy of our sentiment model reaches 88.5%, which is higher than other LSTM-based methods like standard AE-LSTM, AT-LSTM, standard LSTM, TC-LSTM, and TD-LSTM.