ICT: System Description for CoNLL-2012

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 71–75, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics ICT: System Description for CoNLL-2012 Hao Xiong and Qun Liu Key Lab. of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {xionghao, liuqun}@ict.ac.cn Abstract In this paper, we present our system de- scription for the CoNLL-2012 coreference resolution task on English, Chinese and Arabic. We investigate a projection-based model in which we first translate Chinese and Arabic into English, run a publicly available coreference system, and then use a new projection algorithm to map the coreferring entities back from English in- to mention candidates detected in the Chi- nese and Arabic source. We compare to a baseline that just runs the English coref- erence system on the supplied parses for Chinese and Arabic. Because our method does not beat the baseline system on the development set, we submit outputs gen- erated by the baseline system as our final submission. 1 Introduction Modeling multilingual unrestricted coreference in the OntoNotes data is the shared task for CoNLL- 2012. This is an extension of the CoNLL- 2011 shared task and would involve automatic anaphoric mention detection and coreference res- olution across three languages – English, Chinese and Arabic – using OntoNotes v5.0 corpus, giv- en predicted information on the syntax, proposi- tion, word sense and named entity layers. Au- tomatic identification of coreferring entities and events in text has been an uphill battle for sev- eral decades, partly because it can require world knowledge which is not well-defined and partly owing to the lack of substantial annotated data. Figure 1: The overall process of our system, where we use Google Translator to translate Chinese and Arabic into English. For more details, readers can refer to (Pradhan et al., 2012). Before this year’s task, researchers proposed t- wo typical novel methods to address the prob- lem of natural language processing across multiple languages: projection and joint learning (Rahman and Ng, 2012). Specific to this year’s coreference resolution task, for projection based method, we could first develop a strong resolver or utilize a publicly available system on English, and trans- late other languages into English, eventually, we could project the coreferring entities resolved on English back into other language sides. General- ly, a projection method is easier to develop since it doesn’t need sentence alignment across multiple languages. Thus, in this year’s task, we investigate a translation based model to resolve coreference on English, Chinese and Arabic. The whole pro- cess is illustrated in figure 1, in which we first use Google Translator to translate Chinese and Ara- bic into English, and we then employ a strong En- glish coreference resolver to generate coreferring entities, after mapping entities from English into 71 Chinese and Arabic mention candidates, we could obtain coreferring entities for these languages. Intuitively, the performance of coreference re- solver on English should perform better than that on Chinese and …

Content truncated to 500 words