Multi-lingual Sentence Generation from the PIVOT interlingua

[MT Summit III, Washington, DC, USA, July 1991] Multi-lingual Sentence Generation from the PIVOT Interlingua Akitoshi OKUMURA Kazunori MURAKI Susumu AKAMINE C & C Information Technology Research Laboratories 4-1-1 Miyazaki, Miyamae-ku, Kawasaki 216, JAPAN {okumura,k-muraki,akamine}%mtl.cl.nec.co.jp@sj.nec.com Abstract This paper proposes a strategy for French and Spanish sentence generation systems, based on the English generation system. The English generation mode! consists of four procedures, conceptual wording (sentence-structure plan- ning), syntactic selection, ordering and mor- phological generation. The analysis of linguis- tic similarities and differences between English, French and Spanish reveals that a single model is applicable for Spanish and French generation by modifying the contents of the dictionary and the knowledge base. The differences can natu- rally be represented in the knowledge base and dictionary, which is difficult in the transfer ap- proach. So, it is possible to develop Spanish and French generation systems from the En- glish generation system, merely by modifying the contents of the dictionary and knowledge base. Using this strategy, the authors have suc- ceeded in developing English, Spanish, French, Korean and Japanese sentence generation sys- tems. The input for each system is a language- independent conceptual representation, which is an interlingua of the machine translation system PIVOT. The sizes for each dictionary are 70,000 English words, 5,000 Spanish words, 3,000 French words, 110,000 Japanese words and 10,000 Korean words. The systems are im- plemented on NEC’s work station. 1 Introduction Multi-lingual machine translation system (MMTS) has been gaining more and more attention in recent years, There are two different methods for use in the system, a transfer approach and an interlingua approach. The in- terlingua method is suitable for MMTS, because of the low cost of the system development and maintenance[l]. From the viewpoint of system development effectiveness, the rules and dictionary information for the established system should be exploited for another language system development. When a sentence generation system is de- veloped in one language, it is desirable to make the most use of the existing system, in order to develop another generation or analysis system more efficiently. For its purpose, linguistic phenomena of the languages should be analyzed from the semantic, syntactic and morpholog- ical view. It. is necessary to reveal their similarities and differences, A generation model should be designed to maximize the universal characteristics, but still to min- imize language-specific characteristics. The authors have comparatively analyzed subjects, sentence structures, grammatical cases, voice and mor- phological features in English, Spanish and French, and found it possible to develop Spanish and French gen- eration systems from English, merely by modifying the contents of the dictionary and knowledge base. The gen- eration model consists of four procedures; conceptual wording (sentence-structure planning), syntactic selec- tion, word ordering and morphological generation. In each procedure, linguistic differences can be treated by the change in lexical information in the dictionary and knowledge bases. Most of the generation rules can be shared in the three languages. This paper describes a strategy for developing the Spanish and French generation systems from one …

Content truncated to 500 words