Lightly-supervised Representation Learning with Global Interpretability

TL;DR

This paper introduces a lightly-supervised method for named entity classification that combines representation learning with interpretability, outperforming existing bootstrapping methods and producing a globally interpretable decision list.

Contribution

It presents a novel approach that learns custom embeddings for entities and patterns from limited annotations, enabling both improved extraction accuracy and interpretability.

Findings

Outperforms three state-of-the-art bootstrapping methods on CoNLL-2003 and OntoNotes datasets.

Produces a globally-interpretable decision list with minimal performance loss.

Demonstrates the effectiveness of representation learning for interpretable information extraction.

Abstract

We propose a lightly-supervised approach for information extraction, in particular named entity classification, which combines the benefits of traditional bootstrapping, i.e., use of limited annotations and interpretability of extraction patterns, with the robust learning approaches proposed in representation learning. Our algorithm iteratively learns custom embeddings for both the multi-word entities to be extracted and the patterns that match them from a few example entities per category. We demonstrate that this representation-based approach outperforms three other state-of-the-art bootstrapping approaches on two datasets: CoNLL-2003 and OntoNotes. Additionally, using these embeddings, our approach outputs a globally-interpretable model consisting of a decision list, by ranking patterns based on their proximity to the average entity embedding in a given class. We show that this interpretable model performs close to our complete bootstrapping model, proving that representation learning can be used to produce interpretable models with small loss in performance.