QuExEnt: Improved Zero-Shot Classification from Explanations Through Quantifier Modeling and Curriculum Learning

Sayan Ghosh, Rakesh R Menon, Shashank Srivastava

Abstract: A hallmark of human intelligence is the ability to learn new concepts purely from language. While recent advances in training machine learning models via natural language explanations show promise, these approaches still fall short on modeling the the intricacies of natural language (such as quantifiers) or in mimicking human behavior in learning a suite a tasks with varying difficulty. In this work, we present QuExEnt, to learn better zero-shot classifiers from explanations by using three strategies - (1) model the semantics of quantifiers present in explanations (including exploiting ordinal strength relationships, such as 'always' > 'likely'), (2) aggregating information from multiple explanations using an attention-based mechanism, and (3) model training via curriculum learning from tasks with simple explanations to tasks with complex explanations. With these strategies, QuExEnt outperforms prior work showing an absolute gain of up to 7% on the recently proposed CLUES benchmark in generalizing to unseen classification tasks.

Track: Non-Archival (will not appear in proceedings)