Multi-Event Extraction Guided by Global Constraints

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 70–79, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Multi Event Extraction Guided by Global Constraints Roi Reichart Regina Barzilay Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {roiri, regina}@csail.mit.edu Abstract This paper addresses the extraction of event records from documents that describe multi- ple events. Specifically, we aim to identify the fields of information contained in a docu- ment and aggregate together those fields that describe the same event. To exploit the in- herent connections between field extraction and event identification, we propose to model them jointly. Our model is novel in that it integrates information from separate sequen- tial models, using global potentials that en- courage the extracted event records to have desired properties. While the model con- tains high-order potentials, efficient approxi- mate inference can be performed with dual- decomposition. We experiment with two data sets that consist of newspaper articles de- scribing multiple terrorism events, and show that our model substantially outperforms tra- ditional pipeline models. 1 Introduction Today, most efforts in information extraction have focused on the field extraction task, commonly for- mulated as a sequence tagging problem. When a document describes a single event, the list of ex- tracted fields provides a useful abstraction of the in- put document. In practice, however, a typical news- paper document describes multiple events, and a flat list of field values may not contain the sufficient structure required for many NLP applications. Our goal is therefore to extract event templates which ag- gregate field values for individual events. Consider, for instance, the New York Times arti- cle excerpt in Figure 1 that describes three related terrorist events. As this example illustrates, in order to populate the corresponding event templates, the model needs to identify segments that describe indi- vidual events. Such segmentation is challenging, as event boundaries are not explicitly demarcated in the text. Moreover, descriptions of different events are often intermingled, as in the above example, further complicating boundary recovery. In this paper, we consider a model that jointly performs event segmentation and field extraction. This model capitalizes on the inherent connection between the two tasks in order to reduce the ambi- guity of template-based extraction. For example, the distribution of field values in the text provides strong clues about event segmentation, such as the presence of multiple new fields strongly signaling a segment boundary. Likewise, knowledge of the boundaries enables the model to rule out mutually inconsistent predictions, such as extracting two distinct locations for the same event. We formulate our approach as a joint model that marks each word with field and event labels si- multaneously. At the sentence level, segmentation and field extraction taggers are implemented using separate sequence models operating over local fea- tures. At the document level, the model encourages global consistency via potentials that link the ex- tracted event records and their fields. Some of these potentials are limited to …

Content truncated to 500 words