An investigation into the deep learning approach in sentimental analysis using graph-based theories.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Sentiment analysis models, particularly deep learning, often lack transparency due to complex, high-dimensional feature processing.
• Graph-based approaches offer traceability and interpretability in complex system modeling.

Purpose of the Study:

• To develop a transparent and efficient deep neural network for sentiment analysis by integrating Credit Assignment Paths theory.
• To improve the interpretability of deep learning models in natural language analytics.

Main Methods:

• Proposed a novel algorithm utilizing deterministic edge/node embeddings with attention scores for feature extraction and neuron importance attribution.
• Applied the model to the Twitter Health News dataset, extending it for various analytical approximations (e.g., tweet/aspect level, source, frequency).

Main Results:

• The model demonstrated transparency and traceability in its sentiment analysis process.
• Achieved rapid convergence with an overall accuracy of approximately 83% and identified 94% of true positive sentiments.
• Inferred features were conditioned by user preferences and activation derivatives, allowing for feature rejection if not sufficiently scored.

Conclusions:

• The proposed method offers a favorable trade-off between transparency and efficiency in deep neural networks for sentiment analysis.
• The approach enables ideal weight assignment to specific active features, enhancing model interpretability.
• Future work includes incorporating dynamic graph structures and testing on diverse datasets like SemEval and COVID-19 tweets.