Global Locality in Biomedical Relation and Event Extraction

TL;DR

This paper introduces a novel approach for simultaneous relation and event extraction in biomedical texts, leveraging an adapted transformer architecture with multi-head attention to improve dependency modeling across long sentences, outperforming existing methods.

Contribution

It presents a new joint extraction model using multi-head attention and convolutions, addressing the limitations of previous methods that focus on single entity pairs in short spans.

Findings

Outperforms state-of-the-art on multiple biomedical benchmarks

Effectively models long-range dependencies in biomedical sentences

Demonstrates the benefit of a transformer-based architecture in biomedical relation extraction

Abstract

Due to the exponential growth of biomedical literature, event and relation extraction are important tasks in biomedical text mining. Most work only focus on relation extraction, and detect a single entity pair mention on a short span of text, which is not ideal due to long sentences that appear in biomedical contexts. We propose an approach to both relation and event extraction, for simultaneously predicting relationships between all mention pairs in a text. We also perform an empirical study to discuss different network setups for this purpose. The best performing model includes a set of multi-head attentions and convolutions, an adaptation of the transformer architecture, which offers self-attention the ability to strengthen dependencies among related elements, and models the interaction between features extracted by multiple attention heads. Experiment results demonstrate that our approach outperforms the state of the art on a set of benchmark biomedical corpora including BioNLP 2009, 2011, 2013 and BioCreative 2017 shared tasks.