Automatically Labeled Data Generation for Large Scale Event Extraction

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 409–419 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1038 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 409–419 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1038 Automatically Labeled Data Generation for Large Scale Event Extraction Yubo Chen1,2, Shulin Liu1,2, Xiang Zhang1, Kang Liu1 and Jun Zhao1,2 1 National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China 2 University of Chinese Academy of Sciences, Beijing, 100049, China {yubo.chen, shulin.liu, xiang.zhang, kliu, jzhao}@nlpr.ia.ac.cn Abstract Modern models of event extraction for tasks like ACE are based on supervised learning of events from small hand-labeled data. However, hand-labeled training da- ta is expensive to produce, in low cov- erage of event types, and limited in size, which makes supervised methods hard to extract large scale of events for knowledge base population. To solve the data label- ing problem, we propose to automatically label training data for event extraction vi- a world knowledge and linguistic knowl- edge, which can detect key arguments and trigger words for each event type and em- ploy them to label events in texts auto- matically. The experimental results show that the quality of our large scale automat- ically labeled data is competitive with e- laborately human-labeled data. And our automatically labeled data can incorporate with human-labeled data, then improve the performance of models learned from these data. 1 Introduction Event Extraction (EE), a challenging task in In- formation Extraction, aims at detecting and typ- ing events (Event Detection), and extracting ar- guments with different roles (Argument Identifi- cation) from natural-language texts. For exam- ple, in the sentence shown in Figure 1, an EE system is expected to identify an Attack even- t triggered by threw and extract the correspond- ing five augments with different roles: Yesterday (Role=Time), demonstrators (Role=Attacker), s- tones (Role=Instrument), soldiers (Role=Target), and Israeli (Role=Place). To this end, so far most methods (Nguyen et al., Michelle Obama and Barack Obama were on October 3, 1992. Marry Person Person Time Figure 1: This sentence expresses an Attack event triggered by threw and containing five arguments. 2016; Chen et al., 2015; Li et al., 2014; Hong et al., 2011; Ji and Grishman, 2008) usually adopted su- pervised learning paradigm which relies on elab- orate human-annotated data, such as ACE 20051, to train extractors. Although this paradigm was widely studied, existing approaches still suffer from high costs for manually labeling training da- ta and low coverage of predefined event types. In ACE 2005, all 33 event types are manually pre- defined and the corresponding event information (including triggers, event types, arguments and their roles) are manually annotated only in 599 English documents since the annotation process is extremely expensive. As Figure 2 shown, n- early 60% of event types in ACE 2005 have less than 100 labeled samples and there are even three event types …

Content truncated to 500 words