Topic sentiment analysis based on deep neural network using document embedding technique

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Sentiment Analysis (SA) is challenging due to domain-specific language, limiting traditional models and lexicons.
• Sequential Topic Modeling (TM) and SA lack accuracy; simultaneous approaches often fail with domain-specific terms.

Purpose of the Study:

• To propose a novel supervised hybrid Topic Sentiment Analysis (TSA) approach, ETSANet, for improved domain-specific sentiment analysis.
• To address limitations of existing methods in accurately capturing domain-specific sentiment polarity.

Main Methods:

• Developed Semantically Topic-Related Documents Finder (STRDF) to identify contextually relevant training documents using Semantic Topic Vectors.
• Employed a hybrid CNN-GRU deep neural network trained on semantically related documents.
• Utilized a hybrid metaheuristic optimization (Grey Wolf + Whale Optimization) for hyperparameter tuning.

Main Results:

• ETSANet effectively extracts semantic relationships between topics and training data.
• The hybrid CNN-GRU model achieved superior performance compared to existing methods.
• Demonstrated a 1.92% increase in accuracy over state-of-the-art techniques.

Conclusions:

• ETSANet offers a more accurate and robust approach to domain-specific sentiment analysis.
• The novel STRDF and hybrid deep learning architecture are key to the improved performance.