Intermediate Entity-based Sparse Interpretable Representation Learning

TL;DR

This paper introduces ItsIRL, a novel method for learning sparse, interpretable entity representations that enhance biomedical task performance while preserving interpretability and enabling model debugging through counterfactual manipulation.

Contribution

ItsIRL improves upon prior IERs by maintaining interpretability and boosting predictive accuracy, especially in biomedical applications, and introduces techniques for global semantic analysis and counterfactual entity manipulation.

Findings

Enhanced performance on biomedical tasks compared to previous IERs.

Maintains interpretability and supports counterfactual entity type manipulation.

Provides methods for global semantic property analysis of learned classes.

Abstract

Interpretable entity representations (IERs) are sparse embeddings that are "human-readable" in that dimensions correspond to fine-grained entity types and values are predicted probabilities that a given entity is of the corresponding type. These methods perform well in zero-shot and low supervision settings. Compared to standard dense neural embeddings, such interpretable representations may permit analysis and debugging. However, while fine-tuning sparse, interpretable representations improves accuracy on downstream tasks, it destroys the semantics of the dimensions which were enforced in pre-training. Can we maintain the interpretable semantics afforded by IERs while improving predictive performance on downstream tasks? Toward this end, we propose Intermediate enTity-based Sparse Interpretable Representation Learning (ItsIRL). ItsIRL realizes improved performance over prior IERs on biomedical tasks, while maintaining "interpretability" generally and their ability to support model debugging specifically. The latter is enabled in part by the ability to perform "counterfactual" fine-grained entity type manipulation, which we explore in this work. Finally, we propose a method to construct entity type based class prototypes for revealing global semantic properties of classes learned by our model.