Modeling Selectional Preferences of Verbs and Nouns in String-to-Tree Machine Translation

Proceedings of the First Conference on Machine Translation, Volume 1: Research Papers, pages 32–42, Berlin, Germany, August 11-12, 2016. c⃝2016 Association for Computational Linguistics Modeling Selectional Preferences of Verbs and Nouns in String-to-Tree Machine Translation Maria N˘adejde School of Informatics University of Edinburgh m.nadejde@sms.ed.ac.uk Alexandra Birch School of Informatics University of Edinburgh a.birch@ed.ac.uk Philipp Koehn Department of Computer Science Johns Hopkins University phi@jhu.edu Abstract We address the problem of mistranslated predicate-argument structures in syntax- based machine translation. This paper ex- plores whether knowledge about semantic affinities between the target predicates and their argument fillers is useful for translat- ing ambiguous predicates and arguments. We propose a selectional preference fea- ture based on the selectional association measure of Resnik (1996) and integrate it in a string-to-tree decoder. The feature models selectional preferences of verbs for their core and prepositional arguments as well as selectional preferences of nouns for their prepositional arguments. We compare our features with a variant of the neural relational dependency language model (RDLM) (Sennrich, 2015) and find that neither of the features improves au- tomatic evaluation metrics. We conclude that mistranslated verbs, errors in the tar- get syntactic trees produced by the de- coder and underspecified syntactic rela- tions are negatively impacting these fea- tures. 1 Introduction Syntax-based machine translation systems have had some success when applied to language pairs with major structural differences such as German- English or Chinese-English. Modeling the target side syntactic structure is important in order to produce grammatical, fluent translations and could be an intermediate step on which to build a se- mantic representation of the target sentence. How- ever these systems still suffer from errors such as scrambled or mis-translated predicate-argument structures. We give a few examples of such er- rors in Table 1. In example a) the baseline system MT1 mistranslates the verb besichtigt as viewed. The system MT2 which uses information about the semantic affinity between the verb and its argu- ment produces the correct translation visited. The semantic affinity score , shown on the right, for the verb viewed and argument trip in the syntac- tic relation prep on is indicating a stronger affinity than for the baseline translation. In example b) the baseline system MT1 mistranslates the noun Aufnahmen as recordings while the system MT2 produces the correct translation images which is a better fit for the prepositional modifier from the telescope. Syntax-based MT systems handle long distance reordering with synchronous translation rules such as: root → ⟨RB∼0V BZ∼1sich nsubj∼2prep∼3, RB∼0nsubj∼2V BZ∼1prep∼3⟩ This rule is useful for reordering the verb and its arguments according to the target side word or- der. However the rule does not contain the lexical head for the verb, the subject and the prepositional modifier. Therefore the entire predicate argument structure is translated by subsequent independent rules. The language model context will capture at most the verb and one main argument. Due to the lack of a larger source or target context the re- sulting predicate-argument structures are often not semantically coherent. This paper …

Content truncated to 500 words