Exploiting Document Level Information to Improve Event Detection via Recurrent Neural Networks

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 352–361, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Exploiting Document Level Information to Improve Event Detection via Recurrent Neural Networks Shaoyang Duan1,2 and Ruifang He1,2 and Wenli Zhao1,2 1School of Computer Science and Technology, Tianjin University, Tianjin, China 2Tianjin Key Laboratory of Cognitive Computing and Applications, Tianjin, China {syduan,rfhe}@tju.edu.cn, wlzhao@gmail.com Abstract This paper tackles the task of event de- tection, which involves identifying and categorizing events. The previous work mainly exists two problems: (1) the tradi- tional feature-based methods apply cross- sentence information, yet need taking a large amount of human effort to de- sign complicated feature sets and infer- ence rules; (2) the representation-based methods though overcome the problem of manually extracting features, while just depend on local sentence representation. Considering local sentence context is in- sufficient to resolve ambiguities in identi- fying particular event types, therefore, we propose a novel document level Recurrent Neural Networks (DLRNN) model, which can automatically extract cross-sentence clues to improve sentence level event de- tection without designing complex reason- ing rules. Experiment results show that our approach outperforms other state-of- the-art methods on ACE 2005 dataset nei- ther the external knowledge base nor the event arguments are used explicitly. 1 Introduction Event detection is a crucial subtask of event ex- traction, which aims to extract event triggers (most often a single verb or noun) and classify them in- to specific types in text. For instance, accord- ing to the ACE 2005 annotation guideline1, in the sentence “central command says troops were in- volved in a gun battle yesterday”, an event de- tection system should be able to detect an Attack event with the trigger word “battle”. However, this 1https://www.ldc.upenn.edu/sites/www.ldc.upenn. edu/files/english-events-guidelines-v5.4.3.pdf. task is very challenging, as the same event might appear with various trigger words and a trigger ex- pression might evoke different event types in dif- ferent context. Most of the existing methods either employed feature-based models with cross-sentence level information (Ji and Grishman, 2008)(Liao and Grishman, 2010)(Hong et al., 2011)(Huang and Riloff, 2012) or followed representation-based architectures with sentence level context (Chen et al., 2015)(Nguyen and Grishman, 2015)(Li- u et al., 2016)(Nguyen and Grishman, 2016)(N- guyen et al., 2016)(Liu et al., 2017)(Chen et al., 2017). Both models have some inherent flaws: (1) feature-based approaches not only need to e- laborately design rich features and often suffer er- ror propagation from the existing natural language processing tools (i.e part of speech tags and de- pendency), but also the cross-sentence clues are embodied by devising complex inference rules, which is difficult to cover all the semantic laws; (2) though representation-based models can ef- fectively alleviate the problem of manually ex- tract features, local sentence context information may be insufficient for event detection models or even humans to classify events from isolated sen- tences. For example, consider the following sen- tences from ACE2005 dataset: S1: Saba hasn’t delivered yet.2 S2: I knew it was time to leave.3 It is …

Content truncated to 500 words