Modelling sentiments based on objectivity and subjectivity with self-attention mechanisms.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Digital commerce growth necessitates understanding customer feedback through sentiment analysis.
• Sentiment analysis accuracy can be improved by classifying review subjectivity and objectivity.
• Attention mechanisms offer a way to enhance aspect-level sentiment analysis.

Purpose of the Study:

• To enhance sentiment analysis by incorporating subjectivity and objectivity classification with attention mechanisms.
• To evaluate the performance of attention-based models against traditional classification models.
• To assess the impact of data augmentation on sentiment analysis accuracy.

Main Methods:

• Utilized three corpora: ShopeeRD (subjectivity), Wiki-en and IMDb (objectivity).
• Employed Word2Vec for word embeddings and a bidirectional LSTM with an attention layer (LSTM-ATT).
• Benchmarked LSTM-ATT against Logistics Regression (LR) and Linear SVC (L-SVC), incorporating data augmentation with AUG-BERT.

Main Results:

• LSTM-ATT achieved 69.0% accuracy on subjective reviews, outperforming L-SVC.
• With AUG-BERT, LSTM-ATT reached 70.0% accuracy on subjective and 60.0% on objective reviews.
• Augmented models consistently showed improved performance over non-augmented counterparts.

Conclusions:

• Attention mechanisms combined with subjectivity/objectivity classification significantly improve sentiment analysis accuracy.
• The LSTM-ATT model demonstrates superior performance, especially when augmented with AUG-BERT.
• Findings highlight the potential of nuanced sentiment analysis for understanding customer reviews in e-commerce.