Training Classifiers with Natural Language Explanations.

Area of Science:

• Machine Learning
• Natural Language Processing
• Data Science

Background:

• Training accurate classifiers typically requires a large number of labeled data points.
• Each individual label provides limited information, especially in binary classification tasks.
• Existing methods often face bottlenecks in the data labeling process.

Purpose of the Study:

• To introduce BabbleLabble, a novel framework for training classifiers.
• To enable annotators to provide natural language explanations for labeling decisions.
• To accelerate the development of accurate classifiers through efficient data annotation.

Main Methods:

• Annotators provide natural language explanations for each data point's label.
• A semantic parser converts these explanations into programmatic labeling functions.
• These functions generate noisy labels for unlabeled data, which is then used for classifier training.

Main Results:

• Classifiers trained using BabbleLabble achieved comparable F1 scores to traditional methods.
• The framework enabled training 5-100x faster across three relation extraction tasks.
• A simple rule-based semantic parser proved sufficient despite inherent imperfections in labeling functions.

Conclusions:

• BabbleLabble offers a significantly more efficient approach to training classifiers.
• Natural language explanations can effectively replace direct labeling for faster model development.
• The framework demonstrates the potential of semantic parsing in streamlining machine learning workflows.