When language is to be generated from an underlying knowledge base, lexical items need to be linked somehow to the representational units in the KB. In the following, we assume a KL-ONE style representation (LOOM], in this case) and draw in particular on two distinctions made in this family of languages: the division between concepts and relations (or roles) holding among concepts, and the division between terminological knowledge (concept and relation definitions) and assertional knowledge (instances of concepts and relations, representing entities of the world).
The typical one-to-one mapping from KB units to lexical items that language generators employ, evades the problem of genuine lexical choice, which means selecting from similar words and thus requires these to be adequately represented as synonyms or near-synonyms [DiMarco, Hirst, Stede 1993]. While in principle one could do so by creating an individual concept for each of them in the KB, this ``solution'' leads to a vast proliferation of concepts that need not be distinguished for the purposes of reasoning with the KB. Moreover, since we are interested in multilingual generation, a one-to-one correspondence between concepts and lexical items would require the presence of language-specific concepts in the KB, which is equally undesirable from the knowledge engineer's perspective. Even when the target languages are closely related, we have to deal with incongruities like this one from a bilingual automobile manual: Disconnect the spark plug wire and ...---Das Zundkabel abziehen und .... The closest translation of disconnect is trennen, and that of abziehen is pull off. Hence, the German sentence describes the nature of the physical movement and some property of the connection between the two parts, whereas the English version focuses on the effect that the action will have on the connection of the parts. Either way of using literal translations would be awkward in the given context. We will examine this example more closely.
" }