Biomedical Event Extraction with Machine Learning

Abstract

<p> Biomedical natural language processing (BioNLP) is a subfield of natural<br /> language processing, an area of computational linguistics concerned<br /> with developing programs that work with natural language: written texts and<br /> speech. Biomedical relation extraction concerns the detection of<br /> semantic relations such as protein--protein interactions (PPI) from scientific<br /> texts. The aim is to enhance information retrieval by detecting relations<br /> between concepts, not just individual concepts as with a keyword search.<br /> <br /> In recent years, events have been proposed as a more detailed alternative for<br /> simple pairwise PPI relations. Events provide a systematic, structural<br /> representation for annotating the content of natural language texts. Events are<br /> characterized by annotated trigger words, directed and typed arguments and the<br /> ability to nest other events. For example, the sentence ``Protein A causes<br /> protein B to bind protein C&#39;&#39; can be annotated with the nested event structure<br /> <em>CAUSE(A, BIND(B, C))</em>. Converted to such formal representations, the<br /> information of natural language texts can be used by computational<br /> applications. Biomedical event annotations were introduced by the BioInfer and<br /> GENIA corpora, and event extraction was popularized by the BioNLP&#39;09 Shared Task<br /> on Event Extraction.<br /> <br /> In this thesis we present a method for automated event extraction, implemented<br /> as the Turku Event Extraction System (TEES). A unified graph format is defined<br /> for representing event annotations and the problem of extracting complex event<br /> structures is decomposed into a number of independent classification tasks.<br /> These classification tasks are solved using SVM and RLS classifiers, utilizing<br /> rich feature representations built from full dependency parsing.&nbsp; Building on<br /> earlier work on pairwise relation extraction and using a generalized graph<br /> representation, the resulting TEES system is capable of detecting binary<br /> relations as well as complex event structures.<br /> <br /> We show that this event extraction system has good performance,<br /> reaching the first place in the BioNLP&#39;09 Shared Task on Event Extraction. Subsequently,<br /> TEES has achieved several first ranks in the BioNLP&#39;11 and BioNLP&#39;13 Shared<br /> Tasks, as well as shown competitive performance in the binary relation Drug-Drug<br /> Interaction Extraction 2011 and 2013 shared tasks.<br /> <br /> The Turku Event Extraction System is published as a freely available open-source<br /> project, documenting the research in detail as well as making the method<br /> available for practical applications. In particular, in this thesis we<br /> describe the application of the event extraction method to PubMed-scale text<br /> mining, showing how the developed approach not only shows good performance, but<br /> is generalizable and applicable to large-scale real-world text mining projects.<br /> <br /> Finally, we discuss related literature, summarize the contributions of the work<br /> and present some thoughts on future directions for biomedical event extraction.<br /> This thesis includes and builds on six original research publications. The first<br /> of these introduces the analysis of dependency parses that leads to<br /> development of TEES. The entries in the three BioNLP Shared Tasks, as well as<br /> in the DDIExtraction 2011 task are covered in four publications, and the sixth<br /> one demonstrates the application of the system to PubMed-scale text mining.</p>