Improving Distantly-Supervised Relation Extraction through BERT-based Label & Instance Embeddings

TL;DR

This paper introduces REDSandT, a transformer-based method leveraging BERT to improve distantly-supervised relation extraction by capturing a wider range of relations and reducing noise through informative embeddings and attention mechanisms.

Contribution

The paper presents a novel BERT-based RE model that uses structured input and label-instance embeddings to better handle noisy data and recognize long-tail relations.

Findings

Achieves state-of-the-art AUC of 0.424 on NYT-10 dataset.

Captures a broader set of relations with higher confidence.

Effectively reduces noise in distantly-supervised relation extraction.

Abstract

Distantly-supervised relation extraction (RE) is an effective method to scale RE to large corpora but suffers from noisy labels. Existing approaches try to alleviate noise through multi-instance learning and by providing additional information, but manage to recognize mainly the top frequent relations, neglecting those in the long-tail. We propose REDSandT (Relation Extraction with Distant Supervision and Transformers), a novel distantly-supervised transformer-based RE method, that manages to capture a wider set of relations through highly informative instance and label embeddings for RE, by exploiting BERT's pre-trained model, and the relationship between labels and entities, respectively. We guide REDSandT to focus solely on relational tokens by fine-tuning BERT on a structured input, including the sub-tree connecting an entity pair and the entities' types. Using the extracted informative vectors, we shape label embeddings, which we also use as attention mechanism over instances to further reduce noise. Finally, we represent sentences by concatenating relation and instance embeddings. Experiments in the NYT-10 dataset show that REDSandT captures a broader set of relations with higher confidence, achieving state-of-the-art AUC (0.424).