Translation-Based Projection for Multilingual Coreference Resolution

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 720–730, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Translation-Based Projection for Multilingual Coreference Resolution Altaf Rahman and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {altaf,vince}@hlt.utdallas.edu Abstract To build a coreference resolver for a new language, the typical approach is to first coreference-annotate documents from this tar- get language and then train a resolver on these annotated documents using supervised learn- ing techniques. However, the high cost asso- ciated with manually coreference-annotating documents needed by a supervised approach makes it difficult to deploy coreference tech- nologies across a large number of natural lan- guages. To alleviate this corpus annotation bottleneck, we examine a translation-based projection approach to multilingual corefer- ence resolution. Experimental results on two target languages demonstrate the promise of our approach. 1 Introduction Noun phrase (NP) coreference resolution is the task of determining which NPs (or mentions) refer to each real-world entity in a document. Recent years have witnessed a surge of interest in multilin- gual coreference resolution. For instance, the ACE 2004/2005 evaluations and SemEval-2010 Shared Task 1 have both involved coreference resolution in multiple languages. As evidenced by the partici- pants in these evaluations, the most common ap- proach to building a resolver for a new language is supervised, which involves training a resolver on coreference-annotated documents from the tar- get language. Although supervised approaches work reasonably well, they present a challenge to deploy- ing coreference technologies across a large number of natural languages. Specifically, for each new lan- guage of interest, one has to hire native speakers of the language to go through the labor-intensive, time- consuming process of hand-annotating a potentially large number of documents with coreference anno- tation before a supervised resolver can be trained. One may argue that a potential solution to this corpus annotation bottleneck is to employ an unsu- pervised or heuristic approach to coreference resolu- tion, especially in light of the fact that they have re- cently started to rival their supervised counterparts. However, by adopting these approaches, we are sim- ply replacing the corpus annotation bottleneck by another, possibly equally serious, bottleneck, the knowledge acquisition bottleneck. Specifically, in these approaches, one has to employ knowledge of the target language to design coreference rules (e.g., Mitkov (1999), Poon and Domingos (2008), Raghu- nathan et al. (2010)) or sophisticated generative models (e.g., Haghighi and Klein (2007,2010), Ng (2008)) to combine the available knowledge sources. One could argue that designing coreference rules and generative models may not be as time- consuming as annotating a large coreference corpus. This may be true for a well-studied language like English, where we can easily compose a rule that disallows coreference between two mentions if they disagree in number and gender, for instance. How- ever, computing these features may not be as simple as we hope for a language like Chinese: the lack of morphology complicates the determination …

Content truncated to 500 words