Event Linking with Sentential Features from Convolutional Neural Networks

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 239–249, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Event Linking with Sentential Features from Convolutional Neural Networks Sebastian Krause, Feiyu Xu, Hans Uszkoreit, Dirk Weissenborn Language Technology Lab German Research Center for Artificial Intelligence (DFKI) Alt-Moabit 91c, 10559 Berlin, Germany {skrause,feiyu,uszkoreit,diwe01}@dfki.de Abstract Coreference resolution for event men- tions enables extraction systems to process document-level information. Current sys- tems in this area base their decisions on rich semantic features from various knowledge bases, thus restricting them to domains where such external sources are available. We propose a model for this task which does not rely on such features but instead utilizes sentential features coming from convolutional neural networks. Two such networks first process coreference candi- dates and their respective context, thereby generating latent-feature representations which are tuned towards event aspects rel- evant for a linking decision. These rep- resentations are augmented with lexical- level and pairwise features, and serve as input to a trainable similarity function pro- ducing a coreference score. Our model achieves state-of-the-art performance on two datasets, one of which is publicly avail- able. An error analysis points out direc- tions for further research. 1 Introduction Event extraction aims at detecting mentions of real- world events and their arguments in text documents of different domains, e.g., news articles. The sub- sequent task of event linking is concerned with resolving coreferences between recognized event mentions in a document, and is the focus of this paper. Several studies investigate event linking and re- lated problems such as relation mentions spanning multiple sentences. Swampillai and Stevenson (2010) find that 28.5 % of binary relation mentions in the MUC 6 dataset are affected, as are 9.4 % of relation mentions in the ACE corpus from 2003. Ji and Grishman (2011) estimate that 15 % of slot fills in the training data for the “TAC 2010 KBP Slot Filling” task require cross-sentential inference. To confirm these numbers, we analyzed the event annotation of the ACE 2005 corpus and found that approximately 23 % of the event mentions lack ar- guments which are present in other mentions of the same event instance in the respective document. These numbers suggest that event linking is an im- portant task. Previous approaches for modeling event men- tions in context of coreference resolution (Bejan and Harabagiu, 2010; Sangeetha and Arock, 2012; Liu et al., 2014) make either use of external feature sources with limited cross-domain availability like WordNet (Fellbaum, 1998) and FrameNet (Baker et al., 1998), or show low performance. At the same time, recent literature proposes a new kind of feature class for modeling events (and relations) in order to detect mentions and extract their argu- ments, i.e., sentential features from event-/relation- mention representations that have been created by taking the full extent and surrounding sentence of a mention into account (Zeng et al., 2014; Nguyen and Grishman, 2015; Chen et al., 2015; dos Santos et al., 2015; Zeng et al., 2015). …

Content truncated to 500 words