Abstract Guhe
We outline an NLG system that receives input from a dynamically changing environment, e.g. a sports match or ticket reservationsystem. From this input the system incrementally generates utterances that describe the ongoing events using an underspecified semantics (CLLS; Egg et al., 2001). Since the environment can change while the utterance is being generated, underspecification is advantageous, because it facilitates immediate self-corrections. This enhances the naturalness of the produced language. Our system generates preverbal structures that are then passed on to a subsequent system that linguistically encodes them (Levelt 1989, Reiter & Dale 2000). The linguistic output is monitored, and a correction is triggered if the environment changes during utterance production or the linguistic encoding goes wrong, e.g. "The ball goes to a Tottenham uh no to an Arsenal player."
Our system is an improvement over former approaches, because most
current accounts are purely surface-oriented: after analysing the entire
utterance the faulty segment (i.e.~the reparandum) is completely
discarded and substituted by the alteration. These approaches run into
problems, especially when the reparandum is still needed, e.g. in
anaphoric references ("take the oranges to Elimira uh I mean take them
to Corning") or in elliptic constructions ("we could send engine E
Egg, M., A. Koller and J. Niehren (2001). The Constraint Language for
Lambda-Structures. To appear in Journal for Logic, Language, and
Information 10.
Levelt, W. (1989). Speaking: From intention to articulation. Cambridge,
MA: MIT Press.
Reiter, E. and R. Dale (2000). Building NLG Systems, Cambridge University Press.