OpExHAN: opinion extraction using hierarchical attention network from unstructured reviews

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning
• Artificial Intelligence

Background:

• Online marketplaces generate vast amounts of customer reviews.
• Analyzing these reviews is crucial for consumers to make purchasing decisions.
• Existing methods face challenges in effectively processing and understanding review sentiment.

Purpose of the Study:

• To develop a robust framework for analyzing and classifying customer opinions from online product reviews.
• To address the challenge of information overload in massive review datasets.
• To improve the accuracy and efficiency of sentiment analysis in e-commerce.

Main Methods:

• A hierarchical attention network (HAN) framework was developed.
• Amazon's Smartphone review dataset was utilized and preprocessed using NLP techniques.
• GloVe embedding was applied for word vector representation, followed by a Bi-GRU model for encoding.
• The model classifies reviews into five opinion classes: extremely positive, positive, neutral, negative, and extremely negative.

Main Results:

• The proposed OpExHAN model demonstrated superior performance compared to baseline methods.
• Achieved an accuracy of 94.6% and precision and recall of 91%.
• The framework effectively captures contextual information at word and sentence levels.

Conclusions:

• The hierarchical attention network-based framework provides an effective solution for large-scale review analysis.
• The OpExHAN model significantly enhances sentiment classification accuracy and reliability.
• This approach aids consumers in making informed purchasing decisions by summarizing product opinions.