Biomedical Event Extraction based on Knowledge-driven Tree-LSTM

Proceedings of NAACL-HLT 2019, pages 1421–1430 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1421 Biomedical Event Extraction Based on Knowledge-driven Tree-LSTM Diya Li†, Lifu Huang†, Heng Ji†, Jiawei Han∗ †Computer Science Department, Rensselaer Polytechnic Institute {lid18,huangl7,jih}@rpi.edu ∗Computer Science Department, University of Illinois at Urbana Champaign hanj@illinois.edu Abstract Event extraction for the biomedical domain is more challenging than that in the general news domain since it requires broader ac- quisition of domain-specific knowledge and deeper understanding of complex contexts. To better encode contextual information and external background knowledge, we propose a novel knowledge base (KB)-driven tree- structured long short-term memory networks (Tree-LSTM) framework, incorporating two new types of features: (1) dependency struc- tures to capture wide contexts; (2) entity prop- erties (types and category descriptions) from external ontologies via entity linking. We eval- uate our approach on the BioNLP shared task with Genia dataset and achieve a new state- of-the-art result. In addition, both quantita- tive and qualitative studies demonstrate the ad- vancement of the Tree-LSTM and the exter- nal knowledge representation for biomedical event extraction. 1 Introduction Biomedical information extraction is widely used to assist the biology community on knowledge ac- quisition and ontology construction. Biomedical events generally refer to a change of status, par- ticularly on proteins or genes. The goal of event extraction is to identify triggers and their argu- ments from biomedical text, and then assign an event type to each trigger and a role to each ar- gument. For example, in the sentence shown in Figure 1, it includes a gene expression and a pos- itive regulation event mention, both triggered by the word transduced. Tax is the Theme argument of the gene expression event. An event could also serve as an argument of another event, lead- ing to a nested structure. For instance, the gene expression event triggered by transduced is also a Theme argument of the positive regulation event as shown in Figure 1. Earlier studies on biomedical event extrac- tion rely on kernel classification methods like the support vector machines (SVMs) (Bj¨orne and Salakoski, 2011; Venugopal et al., 2014) using hand-crafted features, which require high engi- neering effort and domain-specific knowledge. Recent distributional representation based ap- proaches (Rao et al., 2017; Bj¨orne and Salakoski, 2018) explore deep neural networks which only require distributed semantic features. However, different from event extraction in the general news domain, biomedical event extraction requires broad acquisition of domain-specific knowledge and deep understanding of complex contexts. For example, in Genia event extracton of BioNLP shared task 2011 (Kim et al., 2011), about 80% of entity mentions are abbreviations of genes, pro- teins and diseases while more than 36% of event triggers and arguments are separated with more than 10 words. In order to efficiently capture indicative infor- mation from broad contexts, we first adopt tree structure based long short-term memory (Tree- LSTM) networks. Compared to the linear chain structured LSTM, the Tree-LSTM takes tree- structured network topology into consideration. As shown in the …

Content truncated to 500 words