A finite-state framework for log-linear models in machine translation

A finite-state framework for log-linear models in Machine Translation1 Jorge Gonz´alez and Francisco Casacuberta Departamento de Sistemas Inform´aticos y Computaci´on Instituto Tecnol´ogico de Inform´atica Universidad Polit´ecnica de Valencia {jgonzalez,fcn}@dsic.upv.es Abstract. Log-linear models represent nowadays the state-of-the-art in statistical machine translation. There, several models are combined al- together into a whole statistical approach. Finite-state transducers cons- titute a special type of statistical translation model whose interest has been proved in different translation tasks. The goal of this work is to introduce a finite-state framework for a log-linear modelling approach in statistical machine translation. Results for a French-English technical translation task show the convenience of the proposed methods. 1 Introduction Statistical machine translation is a fascinating field of natural language processing where models can be learned automatically from training bilingual text using ef- ficient estimation procedures. Initial, word-based alignment models were quickly overcome by the use of the so called phrase-based alignment models [1]. However, the performance of machine translation systems increases significantly when ad- ditional models are used in a log-linear modelling approach [2]. These models are typically: phrase-based translation models, language models, sentence-length models, statistical dictionaries, etc. This approach represents the state-of-the art in statistical machine translation [3]. Given a source sentence s, the statistical machine translation problem can be expressed by means of a log-linear approach in this manner: ˆt = argmax t Pr(t|s) = argmax t � i λifi(s, t) Finite-state transducers constitute a special type of statistical translation model whose interest in statistical machine translation has been proved in different translation tasks [4, 5]. However, the use of additional models as in the log- linear modelling approach is not effortless in a finite-state framework. The goal of this work is to introduce a finite-state framework for a log-linear modelling approach in statistical machine translation by means of the combina- tion and integration of different phrase-based stochastic finite-state transducers. 1 This work has been partially supported by the EC (FEDER) and the Spanish projects TIN2006-15694-C02-01 and the Consolider Ingenio 2010 CSD2007-00018 12th EAMT conference, 22-23 September 2008, Hamburg, Germany 41 2 Finite state models A stochastic finite-state transducer is a tuple A = (Σ, ∆, Q, i, f, P), where Σ is an alphabet of input symbols, ∆ is an alphabet of output symbols, Q is a finite set of states, functions i : Q → [0, 1] and f : Q → [0, 1] refer to the probability of each state to be, respectively, initial and final, and partial function P : Q × {Σ⋆ × ∆⋆} × Q → [0, 1] defines a set of transitions between pairs of states in such a way that each transition is labelled with an input and an output string, and is assigned a probability. Moreover, consistency properties have to be respected for functions i, f and P in order to be able to define a distribution of probabilities on the free monoid. 2.1 Phrase-based models The derivation of phrase-based models stems from the concept of bilingual seg- mentation, i.e. …

Content truncated to 500 words