Comparing deep learning architectures for sentiment analysis on drug reviews.

Area of Science:

• Natural Language Processing
• Machine Learning
• Pharmacovigilance

Background:

• Sentiment analysis of online user reviews is crucial for understanding drug effectiveness and side effects.
• Deep learning techniques have advanced NLP tasks, yet their application to drug review sentiment analysis is underexplored.

Purpose of the Study:

• To benchmark various deep learning architectures for sentiment analysis of drug reviews.
• To evaluate the performance of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Bidirectional Encoder Representations from Transformers (BERT) models.

Main Methods:

• Compared CNN and LSTM recurrent neural networks, including model combinations and pre-trained word embeddings.
• Explored BERT with Bi-LSTM for drug review sentiment analysis, assessing performance and training time.

Main Results:

• BERT models achieved the highest accuracy in sentiment analysis of drug reviews.
• CNN models provided acceptable results with substantially lower training times compared to BERT.

Conclusions:

• BERT is effective for drug review sentiment analysis but computationally intensive.
• CNNs present a practical alternative for sentiment analysis when training time is a constraint.