A multilevel framework for incremental development of MT systems

MT Summit VII Sept. 1999 A Multilevel Framework for Incremental Development of MT Systems Rémi Zajac Computing Research Laboratory, New Mexico State University Box 30001 Dept. 3CRL, Las Cruces NM 88003 zajac@crl.nmsu.edu Abstract We describe a Machine Translation framework aimed at the rapid development of large scale robust machine translation systems for assimilation purposes, where the MT system is incorporated as one of the tools in an analyst’s workstation. The multilevel architecture of the system is designed to enable early delivery of functional translation capabilities and incremental improvement of quality. A crucial aspect of the framework is a careful articulation of a software architecture, a linguistic architecture and an incremental development process of linguistic knowledge. 1 Introduction The machine translation systems that are being developed at CRL are designed for assimilation purposes and are targeted at a large variety of source texts, including news articles, Web pages, newsgroups articles and email traffic. Thus, coverage and robustness are emphasized over depth of analysis, and accuracy over stylistic fluidity. Moreover, these systems are for the most part developed under severe resource constraints. Some of the new languages which are or will be covered are so-called ‘low-density languages’: languages for which there are little or no electronic resources, comparatively little expertise and few descriptive linguistic works published. An example of such a language under development at CRL is Persian. The lack of electronic resources, including bilingual corpora or even monolingual corpora rules out statistical and learning- based approaches to machine translation. As a consequence, language resources are carefully structured and the organized to support rapid and large scale acquisition of resources (computational dictionaries and grammars). Robustness is also a fundamental issue, and the architecture of the machine translation system itself is designed to produce translations even with incomplete resources (although breadth of lexical coverage is a minimum requirement). One of the desiderata of the MT design is the ability to produce translations after a very short period of development: the incremental addition of linguistic knowledge in the system improves the translation quality without the need to restructure the MT software or already acquired knowledge. In this paper, we illustrate how the chosen structuration of the language resources supports on the one hand rapid and incremental acquisition of resources and enables robust processing on the other hand. 1.1 Past experience: the Temple project One of the results of the Temple project at CRL, a three year effort in building a set of MT systems translating from Arabic, Japanese, Russian and Spanish to English with low amount of resources (Vanni & Zajac 97), is that a carefully designed MT architecture is crucial for developing MT systems with a minimal amount of effort, and that the quality of the software contributes significantly to the quality of the final result. The various Temple MT systems were built reusing existing components and resources whenever they existed, even if the quality was low. This experience taught us some important lessons on the construction of robust machine translation systems. In …

Content truncated to 500 words