SENTIMENT ANALYSIS OF PRODUCT REVIEWS WITH DEEP LEARNING MODELS

Abstract

This article researches the usage of modern deep learning models for sentiment analysis of product reviews gathered from online commercial platforms. For companies that launch and sell products, feedback analysis is valuable for gathering intel on their customers' opinions. Reviews allow companies to understand exactly what customers like or dislike about their products. This information helps businesses identify their products' strengths and weaknesses and use this knowledge to improve product quality, introduce new features, or solve problems that may arise. With the rapid growth of consumer review data available on the internet, opinion mining has become a complex task that can benefit from automation based on machine learning techniques.

This article describes the hands-on approach of review analysis using modern NLP models based on transformers, such as BERT and GPT-3. During previous work analysis, it was concluded that those models are trending in NLP nowadays.

Models were trained using pre-processed Amazon Product Reviews dataset. During the research, a few minor tasks were accomplished, such as GPT-3 prompt generation and selection of main metrics for accuracy evaluation. Essential steps of development were documented and described. Research results show that both models can accurately analyze text sentiment and are eligible for practical use, but GPT-3 based model shows higher balanced accuracy on provided dataset. The caveats of each approach were described and concluded.

The research results will help obtain more accurate and comprehensive answers about customers' attitudes to products and contribute to improving interaction in the aspect of company-client in this matter. The developed models can be applied to automate the analysis of feedback from those data sources where there is no scoring functionality, for example, social networks.