Detecting Errors with Zero-Shot Learning.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Traditional error detection in large datasets is costly and rule-dependent.
• Deep learning models struggle with low error rates and insufficient negative samples.
• Existing methods often use data augmentation, leading to biased feature learning.

Purpose of the Study:

• To propose a novel deep learning model for efficient error detection in relational datasets.
• To address the challenge of limited negative samples in real-world data.
• To develop a model that identifies data inconsistencies without predefined rules or constraints.

Main Methods:

• Developed SAT-GAN (Self-Attention Generative Adversarial Network), an Auto-Encoder Generative Adversarial Network-based model.
• Integrated self-attention mechanisms with pre-trained language models to capture semantic features and functional dependencies.
• Employed zero-shot learning to train the model effectively despite the scarcity of negative samples.

Main Results:

• SAT-GAN achieved an average F1-score of 0.95 across five diverse datasets.
• Demonstrated significant improvement over rule-based methods, with at least a 46.2% F1-score increase.
• Outperformed state-of-the-art deep learning approaches in scenarios lacking rules and negative samples.

Conclusions:

• SAT-GAN offers a robust and efficient solution for error detection in relational datasets.
• The model's zero-shot learning approach effectively handles class imbalance and data scarcity.
• SAT-GAN's ability to learn semantic features eliminates the need for manual rules or constraints.