Publikation

Lexical Choice and Knowledge Representation

Stephan Busemann

In: Jochen Heinsohn , Bernhard Hollunder (Hrsg.). DFKI Workshop on Taxonomic Reasoning. DFKI Workshops February 26 Saarbrücken Germany Seiten 33-39 DFKI Documents (D) 92-08 DFKI 1992.

Abstrakt

Recently the problem of choosing communicatively adequate lexemes has attracted much interest in the NL generation community, in general, the task amounts to deciding for a given representation of an intended meaning, which words will most appropriately convey that meaning to the addressee. Whether lexical choice must be exact in the sense that all and only the intended meaning is verbalized, depends on the respective communication situation. In a multimodal discourse, where language is supplemented by gestures or graphics, the linguistic device need not convey everything to the partner. In written discourse without a predefined context, as in DISCO, exact verbalization seems much more in order. In all theories of lexical choice, the convergence problem has to be solved: there is always a decision for exactly one lexical item. We may distinguish the following subtasks of lexical choice:
  • Definite reference, proforms: Events and objects must often be described using words that allow for an unambiguous identification of the referent. The problem subdivides in finding appropriate words for the referents and in describing the relations between them, as deictic and intrinsic readings of "The ball is infront of the car" suggest.
  • Social jugdement: Some words carry social jugdements with them. German Putzfrau and Raumpflegerin mean both cleaning woman, but only the latter is now used officially.1 The former has a pejorative connotation. See [6].
  • Collocations: There are different kinds of cooccurrence restrictions between lexemes. Some words cannot be used together with others, some tend to be used together with others and some yield a different meaning when used with certain others (idioms).
  • Choice of open class words: Given a conceptual representation of the intended meaning, an appropriate word for each concept must be identified.
In this paper, we assume that for lexical selection the following kinds of knowledge are necessary:
  • the concepts of the meaning representation language
  • lexical entries (lemmata and/or phrasal items including semantic and syntactic in formation, among other things )
  • knowledge about the reader (including the reader's goals ans beliefs)
  • knowledge about the linguistic, situational, and social context
We will show that lexical choice requires a domain model based on linguistic considerations, and that standard KL-ONE techniques are insufficient for parts of the task at hand.

Busemann_1992_LCAKR.pdf (pdf, 261 KB )

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence