Interpretable Multi-Head Self-Attention Architecture for Sarcasm Detection in Social Media

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Social media is crucial for businesses, requiring analysis of customer feedback.
• Sentiment analysis of social media text is hindered by the ambiguity of sarcasm.
• Sarcasm detection is challenging due to its linguistic nature, often conveying the opposite of literal meaning.

Purpose of the Study:

• To develop an interpretable deep learning model for accurate sarcasm detection in social media conversations.
• To enhance the performance of sentiment analysis by effectively identifying and classifying sarcastic expressions.
• To provide insights into the linguistic cues that contribute to sarcasm in text.

Main Methods:

• Developed an interpretable deep learning model incorporating multi-head self-attention and gated recurrent units.
• Multi-head self-attention identifies key sarcastic cue-words.
• Gated recurrent units capture long-range dependencies between cue-words for classification.

Main Results:

• Achieved state-of-the-art results on multiple social media and online media datasets.
• The model demonstrates high effectiveness in detecting sarcasm.
• The approach yields interpretable models that highlight contributing sarcastic cues.

Conclusions:

• The proposed deep learning model effectively detects sarcasm in social media text.
• Interpretability of the model allows for identification of crucial sarcastic cues.
• This approach significantly improves sentiment analysis by addressing the challenge of sarcasm.