Flexible finite-state lexical selection for rule-based machine translation

Flexible finite-state lexical selection for rule-based machine translation Francis M. Tyers, Felipe Sánchez-Martínez, Mikel L. Forcada Departament de Llenguatges i Sistemes Informàtics Universitat d’Alacant E-03071 Alacant {ftyers,fsanchez,mlf}@dlsi.ua.es Abstract In this paper we describe a module (rule for- malism, rule compiler and rule processor) designed to provide flexible support for lex- ical selection in rule-based machine trans- lation. The motivation and implementation for the system is outlined and an efficient algorithm to compute the best coverage of lexical-selection rules over an ambiguous input sentence is described. We provide a demonstration of the module by learning rules for it on a typical training corpus and evaluating against other possible lexical- selection strategies. The inclusion of the module, along with rules learnt from the parallel corpus provides a small, but con- sistent and statistically-significant improve- ment over either using the highest-scoring translation according to a target-language model or using the most frequent aligned translation in the parallel corpus which is also found in the system’s bilingual dictio- naries. 1 Introduction This paper presents a module for lexical selection to be used in rule-based machine translation (RBMT). The module consists of an XML-based formalism for specifying lexical-selection rules in the form of constraints, a compiler which converts the rules written in this format to a finite-state transducer, and a processor which applies the rule transducer to ambiguous input sentences. The paper also presents a method of learning lexical-selection rules from a parallel corpus. Lexical selection is the task of choosing, given several source-language (SL) translations with the c⃝ 2012 European Association for Machine Translation. same part-of-speech (POS), the most adequate translation among them in the target language (TL). The task is related to the task of word-sense disam- biguation (Ide and Véronis, 1998). The difference is that its aim is to find the most adequate trans- lation, not the most adequate sense. Thus, it is not necessary to choose between a series of fine- grained senses if all these senses result in the same final translation. The dominant approach to MT for language pairs with sufficient training data is phrase-based statis- tical machine translation; in this approach, lexical selection is performed by a combination of coocur- rence in the phrase table, and score from the target- language model (Koehn, 2010). There have how- ever been attempts to improve on this by looking at global lexical selection over the whole sentence, see e.g. (Venkatapathy and Bangalore, 2007; Carpuat and Wu, 2007). In order to test different approaches to lexical se- lection for RBMT, we use the Apertium (Forcada et al., 2011) platform. This free/open-source platform includes 30 language pairs (as of February 2012). Sánchez-Martínez et al. (2007) describe a method to perform lexical selection in Apertium based on training a source-language bag-of-words model using TL cooccurrence statistics. This ap- proach was tested, but abandoned as it produced less adequate translations than using the transla- tion marked as default by a linguist in the bilingual dictionary. Other possible solutions would be to generate all possible combinations …

Content truncated to 500 words