Cross-lingual Structure Transfer for Relation and Event Extraction

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 313–325, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 313 Cross-lingual Structure Transfer for Relation and Event Extraction Ananya Subburathinam1*, Di Lu1* , Heng Ji2, Jonathan May3, Shih-Fu Chang4, Avirup Sil5, Clare Voss6 1Computer Science Department, Rensselaer Polytechnic Institute 2Computer Science Department, University of Illinois at Urbana-Champaign 3University of Southern California Information Sciences Institute 4Electrical Engineering Department, Columbia University 5IBM T.J. Watson Research Center 6Army Research Laboratory hengji@illinois.edu Abstract The identification of complex semantic struc- tures such as events and entity relations, al- ready a challenging Information Extraction task, is doubly difficult from sources written in under-resourced and under-annotated lan- guages. We investigate the suitability of cross- lingual structure transfer techniques for these tasks. We exploit relation- and event-relevant language-universal features, leveraging both symbolic (including part-of-speech and depen- dency path) and distributional (including type representation and contextualized representa- tion) information. By representing all entity mentions, event triggers, and contexts into this complex and structured multilingual common space, using graph convolutional networks, we can train a relation or event extractor from source language annotations and apply it to the target language. Extensive experiments on cross-lingual relation and event transfer among English, Chinese, and Arabic demon- strate that our approach achieves performance comparable to state-of-the-art supervised mod- els trained on up to 3,000 manually annotated mentions: up to 62.6% F-score for Relation Extraction, and 63.1% F-score for Event Ar- gument Role Labeling. The event argument role labeling model transferred from English to Chinese achieves similar performance as the model trained from Chinese. We thus find that language-universal symbolic and distribu- tional representations are complementary for cross-lingual structure transfer. 1 Introduction Advanced Information Extraction (IE) tasks entail predicting structures, such as relations between en- tities, and events involving entities. Given a pair of entity mentions, Relation Extraction aims to iden- tify the relation between the mentions and clas- sify it by predefined type. Event Extraction aims to identify event triggers and their arguments in Figure 1: Information Network example constructed by Event (blue) and Relation (green dashed) Extraction. unstructured texts and classify them respectively by predefined types and roles. As Figure 1 il- lustrates, both tasks entail predicting an informa- tion network (Li et al., 2014) for each sentence, where the entity mentions and event triggers are nodes, and the relations and event-argument links are edges labeled with their relation and argument roles, respectively. There are certain relations and events that are of primary interest to a given community and so are reported predominantly in the low-resource language data sources available to that commu- nity. For example, though English language news will occasionally discuss the Person Aung San Suu Kyi, the vast majority of Physical-Located relations and Meeting events involving her are only reported locally in Burmese news, and thus, without Burmese relation and event extraction, a knowledge graph of this person will lack this available information. …

Content truncated to 500 words