Teaching Machines Through Human Explanations

Explore an explanation-based learning framework for natural language processing that improves label efficiency and model reliability in this 45-minute conference talk. Discover how human explanations can be leveraged to create more effective NLP systems with fewer training examples compared to traditional deep learning approaches. Learn about techniques like neural rule grounding and soft rule matching that allow models to generalize from explanations. Examine case studies demonstrating improved performance on tasks like relation extraction and hate speech classification using this framework. Gain insights into making NLP model development more accessible and less reliant on large labeled datasets or machine learning expertise.

Intro
A Surprisingly "Simple" Recipe for Modern NLP
Cost of data labeling: relation extraction
Cost of data labeling: more complex task
Workaround for (less) data labeling?
How "labels" alone could make things wrong
From "labels" to "explanations of labels" One explanation generalizes to many examples
Learning from Human Explanation
Our Focus: Natural Language Explanations
Learning with Human Explanations
Explanations to "labeling rules"
Generalizing explanations Matching labeling rules to create pseudo labeled data
Challenge: Language Variations
Neural Rule Grounding for rule generalization
A Learable, Soft Rule Matching Function
Neural Execution Tree (NEXT) for Soft Matching
Study on Label Efficiency (TACRED)
Results: Hate Speech (Binary) Classification
Take-aways . "One explanation generalizes to many examples" - better label efficiency vs. conventional supervision