The University of Maryland translation system for IWSLT 2007

A comparison of linguistically and statistically enhanced models for speech-to-speech machine translation Alicia P´erez(1), V´ıctor Guijarrubia(1), Raquel Justo(1), M. In´es Torres(1), Francisco Casacuberta(2) (1)Dep. Electricity and Electronics. University of the Basque Country manes@we.lc.ehu.es (2)Dep. of Information Systems and Computation. Technical University of Valencia fcn@dsic.upv.es Abstract The goal of this work is to improve current translation mod- els by taking into account additional knowledge sources such as semantically motivated segmentation or statistical catego- rization. Specifically, two different approaches are discussed. On the one hand, phrase-based approach, and on the other hand, categorization. For both approaches, both statistical and linguistic alternatives are explored. As for translation framework, finite-state transducers are considered. These are versatile models that can be easily integrated on-the-fly with acoustic models for speech translation purposes. In what the experimental framework concerns, all the models pre- sented were evaluated and compared taking confidence in- tervals into account. 1. Introduction Stochastic finite-state transducers (SFST) [1] consist of a sub-set of probabilistic translation models [2]. They are ver- satile models that count on efficient algorithms for inference from training samples [3], composition with other finite-state models [4, 5] that allows for a hierarchical structure of sev- eral knowledge sources, minimization and decoding [6]. The main goal of this paper is to carry out a compara- tive study on the benefits of additional knowledge sources within finite-state framework constrained to GIATI method- ology [7]. We aim at comparing the classical GIATI ap- proach, where the source sentence to be translated is ana- lyzed word-by-word, with category-based and phrase-based approaches, where either the analysis of the source sentence or the generation of target sentence is driven by other kind of tokens. We intend to find out wether it is worth or not to create more complex models such as class-based or phrase- based models within this finite-state framework. In previous works we studied such models taking just linguistic approach into account. The contribution of this paper is to tackle those models under purely statistical framework and to compare with previous models, which joined linguistic knowledge sources within the statistical framework. The experimental layout entails both text and speech translation. On the other hand, this work is an attempt to quanti- tatively determine the contribution of linguistic knowledge sources in contrast to statistical ones for the same task and under exactly the same conditions. Nevertheless, when the discrepancy in score of two systems is marginal we may wonder whether it is possible to assert that one system out- performs the other. The arising question is what a marginal difference stands for or what might be considered as a sig- nificant difference. In this context, the scores are presented along with their confidence interval. In addition, whenever we say that a system outperforms another, the results are ac- companied by the probability of improvement. The organization of this paper is as follows: first of all, an overview of both the baseline (in this paper also referred to as word-based), category-based and phrase-based finite-state transducers is presented …

Content truncated to 500 words