In the area of natural language systems a similar method that also integrates understanding and generation like Levelt's model is known under the term of anticipation feedback loop (AFL) which has been motivated as a special case of exploitation of a user model [Wahlster and Kobsa1986]. The basic idea of the AFL model is the use of the system's natural language understanding part to anticipate the prefered user's interpretation of an utterance which the system plans to realize. User-modelling is necessary to answer a question like (cf. [Wahlster1991])
If I had to analyze this communication act relative to the assumed knowledge of the user, then what would be the effect on me?
To answer such question a produced utterance is fed back to the system's NLU part under the consideration of the user model. If the result of the understanding process does not match the system's intention in planning, it has to re-plan its utterance. Figure 5.4 shows the schematic structure of a system which incorporates an anticipation feedback loop.
Figure 5.4: Schematic structure of an Anticipation Feedback Loop, based on
Wahlster and Kobsa [1986].
A possible utterance S generated from a semantic representation SR is fed back to the understanding component of the system. The result SR has been computed under consideration of the user model. If this result does not match the original goal SR the system interprets this as a possible source of misunderstandings. Therefore SR is iteratively revised until the analysis of the produced utterances matches the system's intention.
In [Jameson and Wahlster1982] this method was used for generating elliptical utterances in the HAM-ANS system. A local anticipation feedback loop is used to ensure that the system's generated ellipses are not so brief as to be ambiguous or misleading. Suppose for example the user as entered the question:
If the room has in fact three beds and three desks, an appropriate answer of the system would be
instead of one of the possible other answers like
because they are undetermined with respect to correct interpretation. Therefore, the semantic representation SR of the user's question is taken into account during the process of ellipsis generation in the following way. After constructing the complete semantic representation SR of the system's answer, SR is compared from the top downward with SR in order to determine the set of essentially identical subtrees of both representations. The top-down approach orders them automatically with respect to their size. The smallest identical subtree is chosen as a possible candidate for an elliptic utterance. Before that partial semantic representation SR is actually passed to the surface transformation process, SR is fed back to the system to check whether SR will possibly be understood by the user according to the system's intention. Using the ellipsis reconstruction component of the system's understanding part, SR is compared with SR to be able to determine whether SR is actually a part of SR (which must be the case because otherwise the system itself is in an inconsistent state). If SR occurs only in one subtree of SR then it is chosen as a candidate and passed to the verbalization component. If two or more subtrees exist in SR that match SR , SR is rejected and the next larger subtree of SR is chosen from the list of possible candidates and the same method is applied again. In summary, before verbalizing an elliptic utterance immediately, the system attempts to reconstruct its semantic representation as the user would, i.e. by determining how it fits into the structure of the original question SR .