SIGNATURE WORK
CONFERENCE & EXHIBITION 2022

User reviews mining is crucial for product improvement which makes the consideration of feedback a substantial part of the business optimization. Time is the most valuable asset at our disposal. Optimizing a business’s time equates to optimizing its decisions. By making fast, reliable, and informed marketing decisions, the firms will maximize their financial revenues while improving the products reviewed. In this study, an Aspect-Based Sentiment Analysis strategy is proposed to facilitate the mining of reviews and the classification of its existing aspects. This paper will focus on addressing the problem of classifying the sentiment of known aspects from text data. The existing studies in the literature offer various methods including graph models, neural networks and transformers, for addressing the same problem. Unlike those studies, a novel approach is proposed for bridging between Transformers and recurrent networks to classify the polarity of aspects. Regarding Transformers, Bidirectional Encoder Representations from Transformer (BERT) is preferred. BERT is a well-known transformer-based machine learning model that is used to successfully extract and encode semantic textual information. The output of BERT will be fed to a recurrent neural network model, more precisely, a long short-term memory model (LSTM) block that, in turn, will decode the semantic textual information after reducing the BERT’s hidden layers by means of Principal Component Analysis(PCA). The present paper focuses on the SEMEVAL datasets of the years 2014, 2015, and 2016 and assess the results of the models with an accuracy and F1 score. The proposed system offers results on par with the current state of the arts models while providing a novel one-way bridge between the BERT and the LSTM blocks with limited computational cost, accommodated by PCA.