Bootstrapping Events and Relations from Text

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 296–305, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Bootstrapping Events and Relations from Text Ting Liu ILS, University at Albany, USA tliu@albany.edu Tomek Strzalkowski ILS, University at Albany, USA Polish Academy of Sciences tomek@albany.edu Abstract In this paper, we describe a new approach to semi-supervised adaptive learning of event extraction from text. Given a set of exam- ples and an un-annotated text corpus, the BEAR system (Bootstrapping Events And Relations) will automatically learn how to recognize and understand descriptions of complex semantic relationships in text, such as events involving multiple entities and their roles. For example, given a series of descriptions of bombing and shooting inci- dents (e.g., in newswire) the system will learn to extract, with a high degree of accu- racy, other attack-type events mentioned elsewhere in text, irrespective of the form of description. A series of evaluations using the ACE data and event set show a signifi- cant performance improvement over our baseline system. 1 Introduction We constructed a semi-supervised machine learning process that effectively exploits statisti- cal and structural properties of natural language discourse in order to rapidly acquire rules to de- tect mentions of events and other complex rela- tionships in text, extract their key attributes, and construct template-like representations. The learning process exploits descriptive and struc- tural redundancy, which is common in language; it is often critical for achieving successful com- munication despite distractions, different con- texts, or incompatible semantic models between a speaker/writer and a hearer/reader. We also take advantage of the high degree of referential consistency in discourse (e.g., as observed in word sense distribution by (Gale, et al. 1992), and arguably applicable to larger linguistic units), which enables the reader to efficiently correlate different forms of description across coherent spans of text. The method we describe here consists of two steps: (1) supervised acquisition of initial extrac- tion rules from an annotated training corpus, and (2) self-adapting unsupervised multi-pass boot- strapping by which the system learns new rules as it reads un-annotated text using the rules learnt in the first step and in the subsequent learning passes. When a sufficient quantity and quality of text material is supplied, the system will learn many ways in which a specific class of events can be described. This includes the capability to detect individual event mentions using a system of context-sensitive triggers and to isolate perti- nent attributes such as agent, object, instrument, time, place, etc., as may be specific for each type of event. This method produces an accurate and highly adaptable event extraction that significant- ly outperforms current information extraction techniques both in terms of accuracy and robust- ness, as well as in deployment cost. 2 Learning by bootstrapping As a semi-supervised machine learning method, bootstrapping can start either with a set of prede- fined rules or patterns, or with a collection of training examples (seeds) annotated by a domain expert …

Content truncated to 500 words