Machine Translation 15: 43–74, 2000.
© 2001 Kluwer Academic Publishers. Printed in the Netherlands.
43
Statistical Translation of Text and Speech: First
Results with the RWTH System
CHRISTOPH TILLMANN, STEPHAN VOGEL, HERMANN NEY and
HASSAN SAWAF
Lehrstuhl für Informatik VI, Rheinisch-Westfälische Technische Hochschule Aachen, Ahornstraße
55, 52056 Aachen, Germany
(E-mail: {tillmann,vogel,ney,sawaf}@informatik.rwth-aachen.de)
Abstract. In this paper, we describe a first version of a system for statistical translation and present
experimental results. The statistical translation approach uses two types of information: a translation
model and a language model. The language model used is a standard bigram model. The translation
model is decomposed into lexical and alignment models. After presenting the details of the alignment
model, we describe the search problem and present a dynamic programming-based solution for the
special case of monotone alignments. So far, the system has been tested on two limited-domain
tasks for which a bilingual corpus is available: the EuTrans traveller task (Spanish–English, 500-
word vocabulary) and the Verbmobil task (German–English, 3000-word vocabulary). We present
experimental results on these tasks. In addition to the translation of text input, we also address the
problem of speech translation and suitable integration of the acoustic recognition process and the
translation process.
Key words: bilingual alignment, dynamic programming, hidden Markov models, learning from
bilingual corpora, statistical machine translation
1. Introduction
In this paper we describe the present status (as of August 1998) of the MT sys-
tem developed at RWTH Aachen and report experimental results obtained for the
EuTrans task and Verbmobil task. Although the ultimate goal of these tasks is
speech translation as opposed to text translation, the experimental tests reported
are limited to text input.
The approach is based on the same statistical principles as used in speech
recognition (Jelinek, 1976) and in the Candide system by IBM (Berger et al., 1994).
Nevertheless, there are a number of characteristic features addressed in our system
and in this paper:
- The alignment models used in our system are based on a Hidden Markov
Model (HMM) as in speech recognition.
- For the search, i.e. the generation of the unknown target sentence, we present
a dynamic programming (DP) approach. To this purpose, we propose align-