Leveraging Multilingual Training for Limited Resource Event Extraction

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1201–1210, Osaka, Japan, December 11-17 2016. Leveraging Multilingual Training for Limited Resource Event Extraction Andrew Hsi Yiming Yang Jaime Carbonell Ruochen Xu Carnegie Mellon University Pittsburgh, PA 15213 USA {ahsi, yiming, jgc, ruochenx}@cs.cmu.edu Abstract Event extraction has become one of the most important topics in information extraction, but to date, there is very limited work on leveraging cross-lingual training to boost performance. We propose a new event extraction approach that trains on multiple languages using a combination of both language-dependent and language-independent features, with particular focus on the case where target domain training data is of very limited size. We show empirically that multilin- gual training can boost performance for the tasks of event trigger extraction and event argument extraction on the Chinese ACE 2005 dataset. 1 Introduction Traditionally, event extraction has focused on monolingual training – typically English (Grishman et al., 2005; Ji and Grishman, 2008; Gupta and Ji, 2009; Liao and Grishman, 2010; Liao and Grishman, 2011; Li et al., 2013; Bronstein et al., 2015), and occasionally Chinese or other languages (Chen and Ji, 2009b; Piskorski et al., 2011; Li et al., 2012; Chen and Ng, 2012; Chen and Ng, 2014). However, apart from a few isolated studies (Chen and Ji, 2009a; Piskorski et al., 2011), to date there is very little work leveraging cross-lingual information for event extraction. Cross-lingual approaches have proven useful for many other tasks in natural language processing (NLP), including part-of-speech (POS) tagging (Snyder et al., 2009; Cohen et al., 2011), dependency parsing (Zeman and Resnik, 2008; Cohen et al., 2011; McDonald et al., 2011; Ammar et al., 2016), and named entity recognition (Richman and Schone, 2008). An important issue in event extraction is that the amount of available training data is often insufficient or unbalanced across domains and/or languages. Event extraction training datasets typically contain merely a few hundreds of documents, owing to the complexity and high costs of human annotation. This issue is even more severe for new event types in new languages. This provides strong motivation to leverage existing language resources for event extraction in new languages. This problem is closely related to low-resource NLP, which has been gathering increased interest among researchers (Garrette et al., 2013; Miao et al., 2013; Duong et al., 2014; Duong et al., 2015). In this paper, we propose a novel approach for cross-lingual event extraction, which trains on multiple languages and leverages both language-dependent and language-independent features in order to boost performance. Using such a system we aim to jointly leverage available multilingual resources (annotated data and induced features) to overcome the annotation-scarcity issue in the target language of interest. Empirically we show that our approach can substantially improve the performance of monolingual sys- tems for the task of Chinese event argument extraction. Our approach is novel compared to existing work in that we have no reliance on using either high quality machine translation or manually aligned documents, which may be …

Content truncated to 500 words