Cross-lingual Coreference Resolution of Pronouns

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 14–24, Dublin, Ireland, August 23-29 2014. Cross-lingual Coreference Resolution of Pronouns Michal Nov´ak and Zdenˇek ˇZabokrtsk´y Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, CZ-11800 {mnovak,zabokrtsky}@ufal.mff.cuni.cz Abstract This work is, to our knowledge, a first attempt at a machine learning approach to cross-lingual coreference resolution, i.e. coreference resolution (CR) performed on a bitext. Focusing on CR of English pronouns, we leverage language differences and enrich the feature set of a standard monolingual CR system for English with features extracted from the Czech side of the bitext. Our work also includes a supervised pronoun aligner that outperforms a GIZA++ baseline in terms of both intrinsic evaluation and evaluation on CR. The final cross-lingual CR system has successfully outperformed both a monolingual CR and a cross-lingual projection system. 1 Introduction Coreference resolution (CR) is a well-established task in the field of Natural Language Processing (NLP). The majority of papers published so far has focused on the monolingual CR, mostly experimenting on the English data. An important step towards multilingual CR was the CoNLL-2012 Shared Task in Modeling Multilingual Unrestricted Coreference in OntoNotes, where the participants were asked to build a CR system that could be applied on three typologically different languages contained in the OntoNotes corpus (Hovy et al., 2006): English, Chinese, and Arabic. Same just as in other NLP tasks such as part-of-speech tagging or parsing, recent years have witnessed a rising interest in cross-lingual projection techniques, mostly aiming at under-resourced languages. However, little attention is paid to leveraging cross-lingual information for CR in two resource-rich languages. This is probably due to lack of bilingual resources annotated with coreference since such techniques would require rich linguistic annotation on both sides of the bitext. Moreover, to solve this issue using a supervised learner, one needs the gold standard of coreference at least on the target side of the bitext. On the other hand, given such data, the typological differences in languages can be exploited to aid a CR system to perform better than if CR is performed independently for each language. The motivation for solving this task is threefold. Firstly, even though Statistical Machine Translation (SMT) has been attracting interest of the community for years, most systems do not take information be- yond the sentence boundary into account, leaving the issues of discourse coherence unresolved. Having a better-quality bitext with coreference resolved could drive research in discourse-aware SMT forward. Secondly, although inter-sentential relations are neglected in SMT, current phrase-based system uninten- tionally resolve some of the coreference links within the sentence, using just the power of phrases. This might be leveraged by using the SMT output instead of a human-translated output in a cross-lingual CR scenario. Finally, even monolingual CR may be improved by applying semi-supervised learning methods in a smart way on a large bilingual corpus with automatic rich annotations, such as CzEng 1.0 (Bojar et al., …

Content truncated to 500 words