Meaning Representation for Knowledge Sharing in Practical Machine Translation

Knowledge-based machine translation can be viewed as the problem of extracting and representing the meaning of a text and generating a translation in a target language using the meaning representation. Meaning extraction requires the integration of information present explicitly in a text with common sense and domain knowledge given to the system. Thus, integrating linguistic knowledge of each language with general world knowledge is a central problem in machine translation , especially when more than two languages are involved. In this article we consider the design of a meaning representation that enables language-speciic lexicons to share knowledge with a language-independent world model. We illustrate how the underlying core meaning representation can be enhanced in three diierent ways to arrive at lexical, ontological, and text meaning representations. The meaning representations presented here have been implemented in the Mikrokos-mos machine translation system and used to represent Span-ish and Japanese lexicons in addition to a broad-coverage ontological world model. Most human knowledge is represented and communicated via natural languages in spite of relatively recent eeorts such as Cyc (Lenat and Guha, 1990) to \computerize" common sense knowledge. Extracting the meaning of a natural language text involves integrating the information present overtly in the text (which is typically incomplete) with the linguistic and world knowledge sources given to the system. The design of such a knowledge based natural language processing (NLP) system is inseparable from issues of sharing knowledge between linguistic and world knowledge sources, and, in turn, sharing knowledge representations between lexical, textual, and ontological meaning representations. Machine translation (MT), especially when it involves more than two languages, is an NLP problem that provides a particularly interesting context to study knowledge shar-0 ing between multiple knowledge bases containing diierent types of knowledge. In multilingual knowledge based machine translation (KBMT), the meaning of a source text is extracted using both linguistic knowledge of the source language and general world knowledge. The resulting meaning representation is then used to generate translations in one or more target languages. Language generation also requires both world knowledge and knowledge of the target language. The internal representation of the meaning of a text must be independent of particular languages to enable translation to or from any of a set of languages. A fundamental issue in the design of multilingual MT (and other NLP) systems is how to represent (i) linguistic knowledge for each language, (ii) general world knowledge, and …