Detecting sarcasm in user-generated content integrating transformers and gated graph neural networks

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Social media presents challenges for sentiment analysis due to sarcasm.
• Sarcasm detection is complex, often using positive words for negative emotions.
• Automated systems struggle with nuanced language in user-generated content.

Purpose of the Study:

• To develop a novel model for accurate sarcasm detection.
• To enhance the understanding of ironic cues in online text.
• To improve automated sentiment analysis for social media.

Main Methods:

• A hybrid model combining Bidirectional Encoder Representations from Transformers (BERT) and Gated Graph Neural Networks (GGNN).
• Integration of a self-attention mechanism to capture subtle ironic expressions.
• Utilizing dependency and emotion graphs with GGNN for global semantic understanding.

Main Results:

• The BERT-GGNN model achieved 92.00% accuracy and 91.51% F1 score on the Headlines dataset.
• The model attained 86.49% accuracy and 86.59% F1 score on the Riloff dataset.
• Superior performance compared to conventional BERT-GCN models, validated by ablation studies.

Conclusions:

• The proposed BERT-GGNN model significantly improves sarcasm detection accuracy.
• The integration of GGNN is crucial for handling complex ironic expressions in social media.
• This approach offers a more robust solution for sentiment analysis in user-generated content.