BART goes multilingual: The UniTN / Essex submission to the CoNLL-2012 Shared Task

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 122–128, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics BART goes multilingual: The UniTN / Essex submission to the CoNLL-2012 Shared Task Olga Uryupina‡ Alessandro Moschitti‡ Massimo Poesio‡† ‡University of Trento † University of Essex uryupina@gmail.com, moschitti@disi.unitn.it, massimo.poesio@unitn.it Abstract This paper describes the UniTN/Essex sub- mission to the CoNLL-2012 Shared Task on the Multilingual Coreference Resolution. We have extended our CoNLL-2011 submission, based on BART, to cover two additional lan- guages, Arabic and Chinese. This paper fo- cuses on adapting BART to new languages, discussing the problems we have encountered and the solutions adopted. In particular, we propose a novel entity-mention detection algo- rithm that might help identify nominal men- tions in an unknown language. We also dis- cuss the impact of basic linguistic information on the overall performance level of our coref- erence resolution system. 1 Introduction A number of high-performance coreference resolu- tion (CR) systems have been created for English in the past decades, implementing both rule-based and statistical approaches. For other languages, how- ever, the situation is far less optimistic. For Ro- mance and German languages, several systems have been developed and evaluated, in particular, at the SemEval-2010 track 1 on Multilingual Coreference Resolution (Recasens et al., 2010). For other lan- guages, individual approaches have been proposed, covering specific subparts of the task, most com- monly pronominal anaphors (cf., for example, (Iida and Poesio, 2011; Arregi et al., 2010) and many oth- ers). Two new languages, Arabic and Chinese, have been proposed for the CoNLL-2012 shared task (Pradhan et al., 2012). They present a challeng- ing problem: the systems are required to pro- vide entity mention detection (EMD) and design a proper coreference resolver for both languages. At UniTN/Essex, we have focused on these parts of the task, relying on a modified version of our last-year submission for English. Most state-of-the-art full-scale coreference reso- lution systems rely on hand-written rules for the mention detection subtask.1 For English, such rules may vary from corpus to corpus, reflecting specifics of particular guidelines (e.g. whether nominal pre- modifiers can be mentions, as in MUC, or not, as in most other corpora). However, for each corpus, such heuristics can be adjusted in a straightforward way. Creating a robust rule-based EMD module for a new language, on the contrary, is a challenging issue that requires substantial linguistic knowledge. In this paper, we advocate a novel approach, re- casting parse-based EMD as a statistical problem. We consider a node-filtering model that does not rely on any linguistic expertise in a given language. In- stead, we use tree kernels (Moschitti, 2008; Mos- chitti, 2006) to induce a classifier for mention NP- nodes automatically from the data. Another issue to be solved when designing a coreference resolution system for a new language is a possible lack of relevant linguistic information. Most state-of-the-art CR algorithms rely on rela- tively advanced linguistic representations of men- tions. This can be seen as a …

Content truncated to 500 words