Graph convolutional network with reinforced dependency graph and denoising mechanism for sarcasm detection

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Sarcasm detection is crucial for social media sentiment analysis and public opinion monitoring.
• Graph-based methods show promise but struggle with homogeneous dependency relationships and noise aggregation.
• Existing approaches often fail to model syntactic and emotional nuances effectively.

Purpose of the Study:

• To propose a novel graph convolutional network (GCN) framework for enhanced sarcasm detection.
• To address limitations in existing graph-based methods by modeling dependency heterogeneity and suppressing noise.
• To improve the accuracy and robustness of sarcasm detection in social media contexts.

Main Methods:

• Constructing dependency graphs considering node distances and dependency types.
• Utilizing a denoising autoencoder for robust feature representation and noise reduction.
• Applying a parameterized mechanism to penalize sparse edges and suppress irrelevant information.

Main Results:

• Achieved F1-scores of 84.97%, 71.11%, and 83.83% on three different datasets.
• The proposed GCN model significantly outperformed baseline models across all evaluated metrics.
• Demonstrated superior performance in handling syntactic and emotional information for sarcasm detection.

Conclusions:

• The novel GCN framework effectively models dependency heterogeneity and suppresses noise for improved sarcasm detection.
• The proposed method offers a robust solution for challenges in social media sentiment analysis.
• The model's outstanding performance highlights its potential for real-world applications in opinion monitoring.