Enhancing Amazon Product Review Analysis by Using Bidirectional LSTM Neural Networks for Deep Multi-Domain Sentiment

This paper shows deep multi-domain sentiment analysis utilizing bidirectional Long Short-Term Memory (LSTM) for extensive experimentation and application on various product reviews in the Amazon Multi-Domain Sentiment Dataset. The dataset contains a mix of different type products, including reviews on kitchen appliances to more defining items such as books, DVDs and electronics. We used the vectors to represent textual data in dense and continuous embeddings, so that our model can capture some area information from words. Lemmatization was also used to lemma words back to their base or root form, to normalize carrying forms, and ultimately improve the representation of input data. By incorporating these preprocessing steps into our model we could help the LSTM to capture more underlying sentiment patterns and provide for a wide range of product categories. The proposed study achieved an accuracy of 92% in terms of LSTM model architecture. This work advances state-of-the-art sentiment analysis approaches, especially for multi-domain datasets, and highlights the promise of deep learning models in processing and understanding sentiment from diverse textual sources.