<?xml version="1.0" standalone="yes"?>
<Paper uid="J04-2004">
  <Title>c(c) 2004 Association for Computational Linguistics Machine Translation with Inferred Stochastic Finite-State Transducers</Title>
  <Section position="2" start_page="0" end_page="206" type="abstr">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> Formal transducers give rise to an important framework in syntactic-pattern recognition (Fu 1982; Vidal, Casacuberta, and Garc'ia 1995) and in language processing (Mohri 1997). Many tasks in automatic speech recognition can be viewed as simple translations from acoustic sequences to sublexical or lexical sequences (acoustic-phonetic decoding) or from acoustic or lexical sequences to query strings (for database access) or (robot control) commands (semantic decoding) (Vidal, Casacuberta, and Garc'ia 1995; Vidal 1997; Bangalore and Ricardi 2000a, 2000b; Hazen, Hetherington, and Park 2001; Mou, Seneff, and Zue 2001; Segarra et al. 2001; Seward 2001).</Paragraph>
    <Paragraph position="1"> Another similar application is the recognition of continuous hand-written characters (Gonz'alez et al. 2000). Yet a more complex application of formal transducers is language translation, in which input and output can be text, speech, (continuous) handwritten text, etc. (Mohri 1997; Vidal 1997; Bangalore and Ricardi 2000b, 2001; Amengual et al. 2000).</Paragraph>
    <Paragraph position="2"> Rational transductions (Berstel 1979) constitute an important class within the field of formal translation. These transductions are realized by the so-called finite-state transducers. Even though other, more powerful transduction models exist, finite-state transducers generally entail much more affordable computational costs, thereby making these simpler models more interesting in practice.</Paragraph>
    <Paragraph position="3"> One of the main reasons for the interest in finite-state machines for language translation comes from the fact that these machines can be learned automatically from examples (Vidal, Casacuberta, and Garc'ia 1995). Nowadays, only a few techniques exist for inferring finite-state transducers (Vidal, Garc'ia, and Segarra 1989; Oncina, [?] Departamento de Sistemas Inform'aticos y Computaci'on, Instituto Tecnol'ogico de Inform'atica, 46071 Valencia, Spain. E-mail:{fcn, evidal}@iti.upv.es.</Paragraph>
    <Paragraph position="4">  Computational Linguistics Volume 30, Number 2 Garc'ia, and Vidal 1993; M&amp;quot;akinen 1999; Knight and Al-Onaizan 1998; Bangalore and Ricardi 2000b; Casacuberta 2000; Vilar 2000). Nevertheless, there are many techniques for inferring regular grammars from finite sets of learning strings which have been used successfully in a number of fields, including automatic speech recognition (Vidal, Casacuberta, and Garc'ia 1995). Some of these techniques are based on results from formal language theory. In particular, complex regular grammars can be built by inferring simple grammars that recognize local languages (Garc'ia, Vidal, and Casacuberta 1987).</Paragraph>
    <Paragraph position="5"> Here we explore this idea further and propose methods that use (simple) finite-state grammar learning techniques, such as n-gram modeling, to infer rational transducers which prove adequate for language translation.</Paragraph>
    <Paragraph position="6"> The organization of the article is as follows. Sections 2 and 3 give the basic definitions of a finite-state transducer and the corresponding stochastic extension, presented within the statistical framework of language translation. In Section 4, the proposed method for inferring stochastic finite-state transducers is presented. The experiments are described in Section 5. Finally, Section 6 is devoted to general discussion and conclusions.</Paragraph>
  </Section>
class="xml-element"></Paper>