Learning finite-state models for machine translation

Abstract

In formal language theory, finite-state transducers are well-know models for simple "input-output" mappings between two languages. Even if more powerful, recursive models can be used to account for more complex mappings, it has been argued that the input-output relations underlying most usual natural language pairs can essentially be modeled by finitestate devices. Moreover, the relative simplicity of these mappings has recently led to the development of techniques for learning finite-state transducers from a training set of inputoutput sentence pairs of the languages considered. In the last years, these techniques have lead to the development of a number of machine translation systems. Under the statistical statement of machine translation, we overview here how modeling, learning and search problems can be solved by using stochastic finite-state transducers. We also review the results achieved by the systems we have developed under this paradigm. As a main conclusion of this review we argue that, as task complexity and training data scarcity increase, those systems which rely more on statistical techniques tend produce the best results.

GREAT: open source software for statistical machine translation

“…Only a few techniques to learn SFSTs can be found in the literature (Bangalore and Riccardi, 2002;Knight and Al-Onaizan, 1998;Casacuberta and Vidal, 2007). However, a relation between regular translations generated by SFSTs and regular languages over some alphabet of string pairs was established through morphisms (Berstel, 1979).…”

“…This property was used to propose a method of inference of SFSTs based on the inference of stochastic finite-state automata (SFSAs) (Casacuberta and Vidal, 2004). This method, which has been widely used in SMT applications (Casacuberta and Vidal, 2007;Pérez et al, 2008;González and Casacuberta, 2009), is known as GIATI and is the training framework of GREAT.…”

Joining linguistic and statistical methods for Spanish-to-Basque speech translation

“…Finite-state transducers are versatile models that count on thoroughly studied efficient implementations for training (Casacuberta and Vidal, 2007) and decoding (Mehryar Mohri and Riley, 2003). Definition and layout for probabilistic finite-state machines (automata and transducers) were comprehensively described in (Vidal et al, 2005a,b), and so we are going to follow that formalism.…”

Statistical approaches for natural language modelling and monotone statistical machine translation

“…SFSTs also permit a simple integration with other information sources, which makes it easy to apply SFSTs to more difficult tasks such as speech translation [Casacuberta et al, 2004]. SFSTs and the corresponding training and search techniques have been studied by several authors, in many cases explicitly motivated by MT applications [E. Vidal and Segarra, 1989, Oncina et al, 1993, Knight and Al-Onaizan, 1998, Mäkinen, 1999, Amengual et al, 2000, Alshawi et al, 2000a, Casacuberta, 2000a, Vilar, 2000, Vogel and Ney, 2000, Picó and Casacuberta, 2001, Bangalore and Riccardi, 2003, Kumar and Byrne, 2003, Casacuberta and Vidal, 2004, Tsukada and Nagata, 2004, Casacuberta et al, 2005, Kumar et al, 2006, Casacuberta and Vidal, 2007, Mariòo et al, 2006. There are other statistical models for MT that are based on alignments between words (statistical word-alignment models) or between word sequences (phrase-based models or alignment templates) [Och andNey, 2004, Zens, 2008].…”

“…The GIATI technique [Casacuberta and Vidal, 2007] has been applied to machine translation [Casacuberta and Vidal, 2004], speech translation [Casacuberta et al, 2004] and computed-assisted translation [Barrachina et al]. The results obtained using GIATI suggest that, among all the SFST learning techniques tested, GIATI is the only one that can cope with translation tasks under real conditions of vocabulary sizes and amounts of training data available.…”