Aiding Pronoun Translation with Co-Reference Resolution

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 252–261, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Aiding Pronoun Translation with Co-Reference Resolution Ronan Le Nagard and Philipp Koehn University of Edinburgh Edinburgh, United Kingdom s0678231@sms.ed.ac.uk, pkoehn@inf.ed.ac.uk Abstract We propose a method to improve the trans- lation of pronouns by resolving their co- reference to prior mentions. We report re- sults using two different co-reference res- olution methods and point to remaining challenges. 1 Introduction While machine translation research has made great progress over the last years, including the in- creasing exploitation of linguistic annotation, the problems are mainly framed as the translation of isolated sentences. This restriction of the task ig- nores several discourse-level problems, such as the translation of pronouns. Pronouns typically refer to earlier mention of entities, and the nature of these entities may matter for translation. A glaring case is the translation of the English it and they into languages with gram- matical gender (as for instance, most European languages). If it refers to an object that has a male grammatical gender in the target language, then its translation is a male pronoun (e.g., il in French), while referring to a female object requires a fe- male pronoun (e.g., elle in French). Figure 1 illustrates the problem. Given a pair of sentence such as The window is open. It is blue. the translation of it cannot be determined given only the sentence it occurs in. It is essential that we connect it to the entity the window in the pre- vious sentence. Making such a connection between references to the same entity is called co-reference resolu- tion, or anaphora resolution.1 While this problem 1In the context of pronouns, anaphora resolution and co- reference resolution are identical, but they differ in other con- texts. has motivated significant research in the field of natural language processing, the integration of co- reference resolution methods into machine transla- tion has been lacking. The recent wave of work on statistical machine translation has essentially not moved beyond sentence-level and has not touched co-reference resolution. Our approach to aiding pronoun translation with co-reference resolution can be outlined as follows. On both training and test data, we identify the anaphoric noun of each occurrence of it and they on the source side (English). We then identify the noun’s translation into the target language (in our experiments, French), and identify the target noun’s grammatical gender. Based on that gender, we replace it with it-masculine, it-feminine or it- neutral (ditto for they). We train a statistical ma- chine translation system with a thusly annotated corpus and apply it to the annotated test sentences. Our experiments show some degree of suc- cess of the method, but also highlight that current co-reference resolution methods (we implemented Hobbs and Lappin/Laess) have not yet achieved sufficient performance to significantly reduce the number of errors in pronoun translation. 2 Related Work 2.1 Co-Reference and Machine Translation The problem of anaphora resolution applied to …

Content truncated to 500 words