Efficient recognition of dynamic user emotions based on deep neural networks

Area of Science:

• Natural Language Processing
• Machine Learning
• Sentiment Analysis

Background:

• Extracting user sentiment from web text is challenging.
• Existing methods often overlook syntactic and semantic information in emotional text.

Purpose of the Study:

• To develop a dynamic user text sentiment analysis technique.
• To improve the extraction of emotional text features by incorporating syntactic and semantic information.

Main Methods:

• Utilized a writing feature abstraction process based on Orthogonal Recurrent Neural Network-Long Short-Term Memory (ON-LSTM) and attention mechanisms.
• Introduced Att-ON-LSTM to capture semantic and hierarchical information.
• Developed ON-LSTM-LS model incorporating tag semantics for enhanced feature extraction.

Main Results:

• Att-ON-LSTM improved micro-average F1 by 2.27% and macro-average F1 by 1.7% over Bi-LSTM.
• ON-LSTM-LS with tag semantics increased accuracy by 0.78% and macro-averaged F1 by 1.04% compared to Att-ON-LSTM.

Conclusions:

• Deep learning models, particularly ON-LSTM-based approaches, demonstrate high capability in dynamic user sentiment analysis.
• Incorporating syntactic and semantic information significantly enhances sentiment classification performance.