Cross-lingual Structure Transfer for Zero-resource Event Extraction

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1976–1981 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1976 Cross-lingual Structure Transfer for Zero-resource Event Extraction Di Lu1*, Ananya Subburathinam1, Heng Ji2, Jonathan May3, Shih-Fu Chang4, Avirup Sil5, and Clare Voss6 1Rensselaer Polytechnic Institute,2University of Illinois at Urbana-Champaign, 3University of Southern California Information Sciences Institute, 4Columbia University, 5IBM T.J. Watson Research Center, 6Army Research Laboratory 1dylanlu.dlu@gmail.com 2hengji@illinois.edu Abstract Most current cross-lingual transfer learning methods for Information Extraction (IE) have been applied to local sequence labeling tasks. To tackle more complex tasks such as event extraction, we need to transfer graph structures (event trigger linked to multiple arguments with various roles) across languages. We develop a novel share-and-transfer framework to reach this goal with three steps: (1) Convert each sentence in any language to language-universal graph structures; in this paper we explore two approaches based on universal dependency parses and fully-connected graphs, respectively. (2) Represent each node in these graph structures with a cross-lingual word embedding so that all sentences, regardless of language, can be represented within one shared semantic space. (3) Using this common semantic space, train event extractors on English training data and apply them to languages that do not have any event annotations. Experimental results on three languages (Spanish, Russian and Ukrainian) without any annotations show this framework achieves comparable performance to a state-of-the-art supervised model trained on more than 1,500 manually annotated event mentions. Keywords: Information Extraction, Less-Resourced Languages, Multilinguality 1. Introduction Event Extraction is an important task in Information Extrac- tion (IE) that aims to identify event triggers and arguments from unstructured texts and classify them into predefined categories. Compared to other IE tasks such as name tag- ging, the annotations for Event Extraction are more costly because they are structured and require a rich label space; full event structure annotation consists of its trigger span and type label as well as each of its one or more argu- ment spans and role labels. Publicly available annotations for event extraction exist for only a few languages, such as English, Spanish, Chinese, and Arabic (Doddington et al., 2004; Getman et al., 2018). We propose a novel share- and-transfer framework to project training data for English only and test data for zero-event-resource languages into one common semantic space, so that we can train an event extractor on English annotations and apply it to target lan- guages. Currently most successful cross-lingual transfer approaches for IE are limited to sequence labeling (Feng et al., 2018a; Xie et al., 2018; Zhang et al., 2018a; Lin et al., 2018). In contrast, event extraction requires transferring complex graph structures that contain triggers and arguments. For example, in Figure 1, the words fire/ fired combined with different arguments indicate different event types. A trans- fer approach to IE with a typical sequence-based Long Short Term Memory (LSTM) encoder will incorporate language- specific characteristics, such as word order, into word rep- resentations, reducing its effectiveness in …

Content truncated to 500 words