Cross-Sentence N-ary Relation Extraction with Graph LSTMs

Transactions of the Association for Computational Linguistics (TACL), 2017 |

Past work in relation extraction has focused on binary relations in single sentences. Recent NLP inroads in high-value domains have sparked interest in the more general setting of extracting n-ary relations that span multiple sentences. In this paper, we explore a general relation extraction framework based on graph long short-term memory networks (graph LSTMs) that can be easily extended to cross-sentence n-ary relation extraction. The graph formulation provides a unified way of exploring different LSTM approaches and incorporating various intra-sentential and intersentential dependencies, such as sequential, syntactic, and discourse relations. A robust contextual representation is learned for the entities, which serves as input to the relation classifier. This simplifies handling of relations with arbitrary arity, and enables multi-task learning with related relations. We evaluate this framework in two important precision medicine settings, demonstrating its effectiveness with both conventional supervised learning and distant supervision. Cross-sentence extraction produced larger knowledge bases. and multi-task learning significantly improved extraction accuracy. A thorough analysis of various LSTM approaches yielded useful insight the impact of linguistic analysis on extraction accuracy.

论文与出版物下载

Cross-Sentence N-ary Relation Extraction with Graph LSTMs

29 11 月, 2018

Implementation of TACL 2017 paper: "Cross-Sentence N-ary Relation Extraction with Graph LSTMs." Nanyun Peng, Hoifung Poon, Chris Quirk, Kristina Toutanova and Wen-tau Yih.