A noisy label and negative sample robust loss function for DNN-based distant supervised relation extraction.

Area of Science:

• Natural Language Processing
• Machine Learning

Background:

• Distantly supervised relation extraction (DSRE) is a key NLP method.
• DSRE faces challenges from noisy labels and class imbalance, common in NLP.
• Existing research has not simultaneously addressed both issues in DSRE.

Purpose of the Study:

• To propose a novel loss function for DSRE.
• To create a loss function robust to noisy labels and efficient for imbalanced datasets.
• To be the first to simultaneously address noisy labels and class imbalance in DSRE.

Main Methods:

• Quantifying the negative impacts of noisy labels and class imbalance.
• Constructing a loss function using linear programming to minimize these impacts.
• Evaluating the loss function on various datasets, including Docred and CoNLL 2003.

Main Results:

• The proposed loss function demonstrates robustness to noisy labels.
• The loss function is efficient for imbalanced class datasets.
• Deep Neural Network (DNN) models with the new loss function outperform state-of-the-art methods.

Conclusions:

• The developed loss function effectively addresses simultaneous noisy label and class imbalance problems in DSRE.
• This represents a significant advancement in handling data imperfections in NLP.
• The approach shows superior performance compared to existing robust loss functions.