Cross-Sentence N-ary Relation Extraction with Graph LSTMs

Cross-Sentence N-ary Relation Extraction with Graph LSTMs Nanyun Peng1∗ Hoifung Poon2 Chris Quirk2 Kristina Toutanova3∗ Wen-tau Yih2 1 Center for Language and Speech Processing, Computer Science Department Johns Hopkins University, Baltimore, MD, USA 2 Microsoft Research, Redmond, WA, USA 3 Google Research, Seattle, WA, USA npeng1@jhu.edu, kristout@google.com {hoifung,chrisq,scottyih}@microsoft.com Abstract Past work in relation extraction has focused on binary relations in single sentences. Re- cent NLP inroads in high-value domains have sparked interest in the more general setting of extracting n-ary relations that span mul- tiple 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 in- corporating various intra-sentential and inter- sentential dependencies, such as sequential, syntactic, and discourse relations. A robust contextual representation is learned for the en- tities, which serves as input to the relation clas- sifier. This simplifies handling of relations with arbitrary arity, and enables multi-task learning with related relations. We evaluate this frame- work in two important precision medicine set- tings, demonstrating its effectiveness with both conventional supervised learning and distant supervision. Cross-sentence extraction pro- duced larger knowledge bases. and multi-task learning significantly improved extraction ac- curacy. A thorough analysis of various LSTM approaches yielded useful insight the impact of linguistic analysis on extraction accuracy. 1 Introduction Relation extraction has made great strides in newswire and Web domains. Recently, there has ∗ This research was conducted when the authors were at Microsoft Research. been increasing interest in applying relation extrac- tion to high-value domains such as biomedicine. The advent of $1000 human genome1 heralds the dawn of precision medicine, but progress in personalized can- cer treatment has been hindered by the arduous task of interpreting genomic data using prior knowledge. For example, given a tumor sequence, a molecular tumor board needs to determine which genes and mu- tations are important, and what drugs are available to treat them. Already the research literature has a wealth of relevant knowledge, and it is growing at an astonishing rate. PubMed2, the online repository of biomedical articles, adds two new papers per minute, or one million each year. It is thus imperative to advance relation extraction for machine reading. In the vast literature on relation extraction, past work focused primarily on binary relations in single sentences, limiting the available information. Con- sider the following example: “The deletion mutation on exon-19 of EGFR gene was present in 16 patients, while the L858E point mutation on exon-21 was noted in 10. All patients were treated with gefitinib and showed a partial response.”. Collectively, the two sentences convey the fact that there is a ternary interaction between the three entities in bold, which is not expressed in either sentence alone. Namely, tumors with L858E mutation in EGFR gene can be treated with gefitinib. Extracting such knowledge clearly requires moving beyond binary relations and single sentences. N-ary relations and cross-sentence …

Content truncated to 500 words