An exponential translation model for target language morphology

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 230–238, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics An exponential translation model for target language morphology Michael Subotin Paxfire, Inc. Department of Linguistics & UMIACS, University of Maryland msubotin@gmail.com Abstract This paper presents an exponential model for translation into highly inflected languages which can be scaled to very large datasets. As in other recent proposals, it predicts target- side phrases and can be conditioned on source- side context. However, crucially for the task of modeling morphological generalizations, it estimates feature parameters from the entire training set rather than as a collection of sepa- rate classifiers. We apply it to English-Czech translation, using a variety of features captur- ing potential predictors for case, number, and gender, and one of the largest publicly avail- able parallel data sets. We also describe gen- eration and modeling of inflected forms un- observed in training data and decoding proce- dures for a model with non-local target-side feature dependencies. 1 Introduction Translation into languages with rich morphology presents special challenges for phrase-based meth- ods. Thus, Birch et al (2008) find that transla- tion quality achieved by a popular phrase-based sys- tem correlates significantly with a measure of target- side, but not source-side morphological complexity. Recently, several studies (Bojar, 2007; Avramidis and Koehn, 2009; Ramanathan et al., 2009; Yen- iterzi and Oflazer, 2010) proposed modeling target- side morphology in a phrase-based factored mod- els framework (Koehn and Hoang, 2007). Under this approach linguistic annotation of source sen- tences is analyzed using heuristics to identify rel- evant structural phenomena, whose occurrences are in turn used to compute additional relative frequency (maximum likelihood) estimates predicting target- side inflections. This approach makes it difficult to handle the complex interplay between different predictors for inflections. For example, the ac- cusative case is usually preserved in translation, so that nouns appearing in the direct object position of English clauses tend to be translated to words with accusative case markings in languages with richer morphology, and vice versa. However, there are exceptions. For example, some verbs that place their object in the accusative case in Czech may be rendered as prepositional constructions in English (Naughton, 2005): David was looking for Jana David hledal Janu David searched Jana-ACC Conversely, direct objects of some English verbs can be translated by nouns with genitive case markings in Czech: David asked Jana where Karel was David zeptal se Jany kde je Karel David asked SELF Jana-GEN where is Karel Furthermore, English noun modifiers are often rendered by Czech possessive adjectives and a ver- bal complement in one language is commonly trans- lated by a nominalizing complement in another lan- guage, so that the part of speech (POS) of its head need not be preserved. These complications make it difficult to model morphological phenomena using 230 closed-form estimates. This paper presents an alter- native approach based on exponential phrase mod- els, which can straightforwardly handle feature sets with arbitrarily elaborate source-side dependencies. …

Content truncated to 500 words