<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0821">
  <Title>Improved Language Modeling for Statistical Machine Translation</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Statistical machine translation (SMT) makes use of a noisy channel model where a sentence -e in the desired language can be conceived of as originating as a sentence -f in a source language. The goal is to find, for every input utterance -f, the best hypothesis</Paragraph>
    <Paragraph position="2"> bilistic constraints on the association of source and target strings. P(-e) is a language model specifying the probability of target language strings. Usually, a standard word trigram model of the form</Paragraph>
    <Paragraph position="4"> is used, where -e = e1,...,el. Each word is predicted based on a history of two preceding words.</Paragraph>
    <Paragraph position="5"> Most work in SMT has concentrated on developing better translation models, decoding algorithms, or minimum error rate training for SMT. Comparatively little effort has been spent on language modeling for machine translation. In other fields, particularly in automatic speech recognition (ASR), there exists a large body of work on statistical language modeling, addressing e.g. the use of word classes, language model adaptation, or alternative probability estimation techniques. The goal of this study was to use some of the language modeling techniques that have proved beneficial for ASR in the past and to investigate whether they transfer to statistical machine translation. In particular, this includes language models that make use of morphological and part-of-speech information, so-called factored language models.</Paragraph>
  </Section>
class="xml-element"></Paper>