Robust Commonsense Reasoning Against Noisy Labels Using Adaptive Correction.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Commonsense reasoning using knowledge graphs (KGs) is crucial for AI but is hindered by unrealistic assumptions of clean training data.
• Existing methods lack robustness to label noise, a common issue in real-world datasets, limiting practical applications.
• The challenge of commonsense reasoning with mislabeled training samples remains largely unexplored.

Purpose of the Study:

• To develop a robust method for commonsense reasoning that effectively handles mislabeled training samples.
• To enhance the practical applicability of commonsense reasoning models by improving their resilience to label noise.
• To provide a generally applicable framework for boosting the robustness of existing commonsense reasoning approaches.

Main Methods:

• Constructed diverse knowledge and model augmentations.
• Developed a multiple-choice alignment method to categorize training samples into clean, semi-clean, and unclean sets.
• Designed adaptive label correction techniques for semi-clean and unclean samples to leverage noisy information.

Main Results:

• The proposed method significantly enhances robustness against label noise in commonsense reasoning.
• Overall performance on commonsense reasoning benchmarks (CommonsenseQA, OpenbookQA) was substantially improved.
• The approach demonstrated general applicability across multiple existing commonsense reasoning frameworks.

Conclusions:

• The developed method effectively addresses the critical issue of noisy labels in commonsense reasoning.
• This work offers a practical solution for building more reliable AI systems capable of understanding and reasoning with real-world data.
• The findings pave the way for more robust and performant AI in complex reasoning tasks.