Cross-Domain Dutch Coreference Resolution

Proceedings of Recent Advances in Natural Language Processing, pages 186–193, Hissar, Bulgaria, 12-14 September 2011. Cross-Domain Dutch Coreference Resolution Orph´ee De Clercq, V´eronique Hoste LT3, Language and Translation Technology Team University College Ghent Groot-Brittannielaan 45 B - 9000 Gent, Belgium orphee.declercq@hogent.be veronique.hoste@hogent.be Iris Hendrickx Centre of Linguistics University of Lisbon Av. Prof. Gama Pinto, 2 1649-003 Lisbon, Portugal iris@clul.ul.pt Abstract This article explores the portability of a coreference resolver across a variety of eight text genres. Besides newspaper text, we also include administrative texts, au- tocues, texts used for external communi- cation, instructive texts, wikipedia texts, medical texts and unedited new media texts. Three sets of experiments were conducted. First, we investigated each text genre individually, and studied the ef- fect of larger training set sizes and in- cluding genre-specific training material. Then, we explored the predictive power of each genre for the other genres conducting cross-domain experiments. In a final step, we investigated whether excluding gen- res with less predictive power increases overall performance. For all experiments we use an existing Dutch mention-pair re- solver and report on our experimental re- sults using four metrics: MUC, B-cubed, CEAF and BLANC. We show that resolv- ing out-of-domain genres works best when enough training data is included. This ef- fect is further intensified by including a small amount of genre-specific text. As far as the cross-domain performance is con- cerned we see that especially genres of a very specific nature tend to have less gen- eralization power. 1 Introduction Coreference resolution is the task of automatically recognizing which words or expressions refer to the same discourse entity in a particular text or di- alogue.1 In the last decade considerable efforts 1In this article we only discuss nominal coreference, i.e. which coreferential relations exist between noun phrases (common and proper nouns, pronouns). have been put in annotating corpora with coref- erential relations. Not only a widespread lan- guage such as English (e.g. ACE-2 (Doddington et al., 2004), ARRAU (Poesio and Artstein, 2008), OntoNotes 3.0 (Weischedel et al., 2009)), but also Czech (PDT 2.0 (Kuˇcov´a and Hajiˇcov´a, 2004)), Catalan (AnCora-Ca (Recasens and Mart´ı, 2010)) and Italian (I-CAB (Magnini et al., 2006))2 can now rely on substantial resources for coreference research. One of the challenges in many current NLP tasks is to test their portability across different do- mains and languages. This portability to other languages was the main objective of the SemEval 2010 Task on Coreference Resolution in Multi- ple Languages (Recasens et al., 2010). The is- sue of domain portability was the focus of the ACL 2010 Workshop on Domain Adaptation for NLP (Daum´e III et al., 2010). In this paper we investigate the performance of an existing mention-pair coreference resolver for Dutch (Hoste, 2005; Hendrickx et al., 2008b) across various text genres. More specifically we want to know whether training on out-of-domain data can be done without performance loss. The above-mentioned corpora designed for corefer- ence resolution consist almost exclusively of text from the same genre, i.e. newspaper texts, and as …

Content truncated to 500 words