The ISL EDTRL system

The ISL EDTRL System Juergen Reichert, Alex Waibel Interactive Systems Laboratories, Carnegie Mellon University and Karlsruhe University juergen@ira.uka.de, ahw@cs.cmu.edu Abstract For the translation of text and speech, statistical methods on one side and interlingua based methods on the other have been used successfully. However, the former requires programming grammars for each language, plus the design of an interlingua, while the latter requires the collection of a large parallel corpus for every language pair. To alleviate these problems, we propose an approach that combines the advantages from both worlds. The proposed approach makes use of English or enriched English as an interlingua and can cascade data-driven translation systems into and from this interlingua. We show that enriching English with linguistic information that is automatically derived i only on English data performs better than pure cascaded systems. 1. Introduction In recent years, a number of translation approaches have been proposed to provide for reliable meaningful translation of text and of speech from one language to another. These include direct approaches (statistical, example-based), transfer and interlingua based approaches. Translation performance is usually the one (if not the most important) consideration for the evaluation of these systems, but has to be balanced by considerations of robustness and portability as additional important dimensions to the translation problem. The translation of speech, in particular, is faced with both of these additional challenges: the input speech and its recognition is fragmentary, ill-formed and errorful, and speech translation systems are frequently required to handle multiple language pairs and language directions to allow for successful cross-lingual dialogs between humans. To accommodate these additional constraints a popular approach has been the interlingua approach to translation. Here, an intermediate representation of meaning is chosen to express the key idea or intent of the speaker. An input sentence is parsed in terms of its key semantic content, represented in an interlingua structure, and from there an equivalent sentence is generated in another language. The use of an interlingua has several advantages. First, adding a new languages to an existing system is simplified, since a new language has to translate only into and out of the interlingua, and we do not require separate translators for every other language pair the system supports. Second, the translation step extracts only the key intentions from a speaker's utterance, thereby handling colloquial expressions, and reducing the sensitivity to redundancies, and disfluencies in spoken language. Third, the system can generate paraphrases from the interlingua back into one's own language to provide meaningful feedback and verification of the translation, before it is delivered into another language. The advantages, however, come at a price: The extraction of the key content from a disfluent input sentence requires the development of semantic grammars that extract key information into frames, concepts and slots. Both, the design of a suitable, unambiguous and language independent interlingua as well as the development of grammars that map sentences to meaning are domain-dependent and have to be repeated for each topic or domain. Their development is labor intensive …

Content truncated to 500 words