Phrase-level alignment generation using a smoothed loglinear phrase-based statistical alignment model

Phrase-level alignment generation using a smoothed loglinear phrase-based statistical alignment model Daniel Ortiz-Mart´ınez1, Ismael Garc´ıa-Varea2, and Francisco Casacuberta1 1 Dpto. de Sist. Inf. y Comp., Univ. Polit´ecnica de Valencia, 46071 Valencia, Spain dortiz@dsic.upv.es fcn@dsic.upv.es 2 Dpto. de Sist. Inf., Univ. de Castilla-La Mancha, 02071 Albacete, Spain ivarea@info-ab.uclm.es Abstract. We present a phrase-based statistical alignment model togheter with a set of different smoothing techniques to be applied when the best phrase-to-phrase alignment for a pair of sentences is to be computed. We follow a loglinear approach, which allows us to introduce different scoring functions to control specific aspects of phrase-level alignments. Experimental results for a well-known shared task on word alignment evaluation are reported, showing the great importance of smoothing in the generation of alignments. As a step forward, we also discuss the adap- tation of the proposed model for its use in a CAT (Computer Assisted Translation) system. 1 Introduction Statistical Machine translation (SMT) is an area of great interest in the NLP community that deals with the transformation of text or speech from a source language into a target language. From a purely statistical point of view, the translation process can be for- mulated as follows: A source language string f is to be translated into a target language string e. Every target string is regarded as a possible translation for the source language string with maximum a-posteriori probability Pr(e|f). Ac- cording to Bayes’ theorem, the target string ˆe that maximizes the product of both the target language model Pr(e) and the string translation model Pr(f|e) must be chosen. The equation that models this process is: ˆeI 1 = arg max e � Pr(e) · Pr(f|e) � (1) State-of-the-art statistical translation systems follow a phrase-based approach, that is, the structural relations between source and target sentences are captured by means of phrases instead of isolated words. In this paper we tackle the problem of generating alignments at phrase level by means of smoothed phrase-based statistical alignment models. As far as we know the problem of finding the best alignment at phrase level has not been ex- tensively addressed in the literature. For example, in [1] three different techniques 12th EAMT conference, 22-23 September 2008, Hamburg, Germany 160 for obtaining phrase-level alignments are compared, but there is no mention at all of the phrase-level aligment coverage problems that arise when real tasks and applications are used. Different applications can benefit from the techniques proposed here, ranging from phrase-based SMT systems to machine-aided NLP tools, as for example CAT [2]. Under the CAT framework, we are given a source sentence and a prefix of the target sentence, and the goal is to obtain the best suffix that constitutes a complete translation. Then the first problem to be solved is how to align the given prefix with the corresponding portion of the source sentence. The techniques proposed here can be easily adapted to deal with this problem. 2 Phrase-based SMT Different translation models have been proposed depending on how the relation between …

Content truncated to 500 words