Relieving the Computational Bottleneck: Joint Inference for Event Extraction with High-Dimensional Features

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 831–843, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Relieving the Computational Bottleneck: Joint Inference for Event Extraction with High-Dimensional Features Deepak Venugopal and Chen Chen and Vibhav Gogate and Vincent Ng Department of Computer Science and Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 dxv021000@utdallas.edu, {yzcchen,vgogate,vince}@hlt.utdallas.edu Abstract Several state-of-the-art event extraction sys- tems employ models based on Support Vec- tor Machines (SVMs) in a pipeline architec- ture, which fails to exploit the joint depen- dencies that typically exist among events and arguments. While there have been at- tempts to overcome this limitation using Markov Logic Networks (MLNs), it re- mains challenging to perform joint infer- ence in MLNs when the model encodes many high-dimensional sophisticated fea- tures such as those essential for event ex- traction. In this paper, we propose a new model for event extraction that combines the power of MLNs and SVMs, dwarfing their limitations. The key idea is to reli- ably learn and process high-dimensional features using SVMs; encode the output of SVMs as low-dimensional, soft formu- las in MLNs; and use the superior joint in- ferencing power of MLNs to enforce joint consistency constraints over the soft for- mulas. We evaluate our approach for the task of extracting biomedical events on the BioNLP 2013, 2011 and 2009 Genia shared task datasets. Our approach yields the best F1 score to date on the BioNLP’13 (53.61) and BioNLP’11 (58.07) datasets and the second-best F1 score to date on the BioNLP’09 dataset (58.16). 1 Introduction Event extraction is the task of extracting and la- beling all instances in a text document that corre- spond to a pre-defined event type. This task is quite challenging for a multitude of reasons: events are often nested, recursive and have several arguments; there is no clear distinction between arguments and events; etc. For instance, consider the BioNLP Ge- nia event extraction shared task (N´edellec et al., 2013). In this task, participants are asked to extract instances of a pre-defined set of biomedical events from text. An event is identified by a keyword called the trigger and can have an arbitrary number of arguments that correspond to pre-defined argu- ment types. The task is complicated by the fact that an event may serve as an argument of another event (nested events). An example of the task is shown in Figure 1. As we can see, event E13 takes as arguments two events, E14 and E12, which in turn has E11 as one of its arguments. A standard method that has been frequently em- ployed to perform this shared task uses a pipeline architecture with three steps: (1) detect if a token is a trigger and assign a trigger type label to it; (2) for every detected trigger, determine all its argu- ments and assign types to each detected argument; and (3) combine the extracted triggers and argu- ments to obtain events. Though …

Content truncated to 500 words