Modeling verbal inflection for English to German SMT

Proceedings of the First Conference on Machine Translation, Volume 1: Research Papers, pages 21–31, Berlin, Germany, August 11-12, 2016. c⃝2016 Association for Computational Linguistics Modeling verbal inflection for English to German SMT Anita Ramm IMS University of Stuttgart, Germany ramm@ims-uni.stuttgart.de Alexander Fraser CIS University of Munich, Germany fraser@cis.uni-muenchen.de Abstract German verbal inflection is frequently wrong in standard statistical machine translation approaches. German verbs agree with subjects in person and num- ber, and they bear information about mood and tense. For subject–verb agreement, we parse German MT output to iden- tify subject–verb pairs and ensure that the verb agrees with the subject. We show that this approach improves subject-verb agreement. We model tense/mood transla- tion from English to German by means of a statistical classification model. Although our model shows good results on well- formed data, it does not systematically improve tense and mood in MT output. Reasons include the need for discourse knowledge, dependency on the domain, and stylistic variety in how tense/mood is translated. We present a thorough analysis of these problems. 1 Introduction Statistical machine translation of English into Ger- man faces two main problems involving verbs: (i) correct placement of the verbs, and (ii) generation of the appropriate inflection for the verb. The position of verbs in German and English differs greatly and often large-range reorderings are needed to place the German verbs in the cor- rect positions. Gojun and Fraser (2012) showed that the preordering approach applied on English– to–German SMT overcomes large problems with both missing and misplaced verbs. Fraser et al. (2012) proposed an approach for handling inflectional problems in English to Ger- man SMT, focusing on the problems of sparsity caused by nominal inflection. However, they do not handle the verbs, ensuring neither that verbs appear in the correct position (which is a problem due to the highly divergent word order of English and German), nor that verbs are correctly inflected (problematic due to the richer system of verbal in- flection in German). In many cases, verbs do not match their subjects (in person and number) which makes understanding of translations difficult. In addition to person and number, the German verbal inflection also includes information about tense and mood. If these are wrong (i.e. do not cor- respond to the tense/mood in the source), very important information, such as point of time and modality of an action/state expressed by the verb, is incorrect. This can lead to false understanding of the overall sentence. In this paper, we reimplement the nominal in- flection modeling for translation to German pre- sented by Fraser et al. (2012) and combine it with the reordering of the source data (Gojun and Fraser, 2012). In a novel extension, we present a method for correction of the agreement errors, and an approach for modeling the translation of tense and mood from English into German. While the subject-verb agreement problems are dealt with successfully, modeling of tense/mood translation is problematic due to many reasons which we will analyze in detail. In …

Content truncated to 500 words