The model that we use for this approach basically consists of three components:
The reversible system introduced in section 5.2 (see figure 5.1) can be represented more detailed as shown in figure 5.5.
Figure 5.5: The Architecture of the new reversible system.
The linguistic core system consists of a reversible grammar (plus lexicon) and the uniform algorithm used for parsing or generation. Both together constitute the competence base of a natural language system. Thus, we are following an extended competence description as known from cognitive linguistics (see e.g., [Schwarz1992]), such that we do not only consider the grammatical knowledge as the only source of language competence (cf. [Chomsky1986]), but also those processes that are able to determine all possible well-formed grammatical structures of a language, at least potentially. However, during run-time of a system we assume that only a subset of possible solutions are computed. This makes sense at least in the case of generation because in practice only one utterance should be produced for a given semantic representation. In our view, the methods or techniques that are used to determine most adequate readings or paraphrases in some specific situation, or those that are used to perform for instance some kind of robustness or monitoring, belong to the performance issue of language use.
The derivations obtained during parsing and generation are stored in a common memory which we will call common knowledge pool for derivation trees, abbreviated as CKP.
The editor is that part of the system that compares structures obtained during monitoring and stored in CKP and eventually revises previously computed solutions. We assume that the editor has a functional nature consisting of a set of specific functions for performing particular tasks (e.g., generation of unambiguous utterances or paraphrases). These functions are selected according to the major flow of control (either understanding or generation) and in dependence of the specified goals, e.g., `be as unambiguous as possible', `re-solve ambiguities explicitly'.
Monitoring is not performed by a separate module. Instead the understander or generator fulfill this part (either during generation or understanding). The basic task of monitoring is to gain information about processing which is not necessarily obvious, i.e., a device is called for which this information can be available to the speaker or the hearer. In has often been argued in cognitive psychology (cf. [Berg1986, Levelt1989]) that it is highly desirable to find a mechanism that is an integral and independently motivated part of the whole system and one that performs the monitoring function by its own nature. Clearly, if it is possible to use devices that are needed anyway for natural language processing, this fulfills our effort to avoid redundancies and to obtained a certain degree of economy.
Whether monitoring and editing should take place is decided by a central planning unit which is part of the conceptual component. We assume that the degree of monitoring and editing that takes place during processing depends on the degree of attention a system has in ongoing communications. This depends on the relevance of the dialog contribution and on the specific dialog situation. For example, if monitoring is switched on by the central planner then it depends on the relevance of the information to be uttered if editing should take place. If the system is under time pressure then the planner might decide to avoid editing because it would take to much time in order to utter the information immediately. Or if the editor cannot find an un-ambiguous reading then the planner has to decide whether to recompute the communication goal that possibly can be mapped to an un-ambiguous utterance (actually the planner would delegate the task of re-computation of the original plan to the text planner; this is indicated by the dashed line in figure 5.5.). The central planning unit will not be further investigated in this thesis because we are mainly concerned with the realization aspects of monitoring and editing on the grammatical level. However, as mentioned above we assume the existing of this unit as the primarily activator for monitoring.
The integrated approach developed in this thesis, is basically used for revision of previously produced utterances. These utterances are assumed to be grammatically well-formed. The integrated approach is used in order to reduce the risk of being misunderstood in the case of the utterance's ambiguity. However, monitoring can also be very useful for detection and correction of speech errors. There is psychological evidence that monitoring mechanisms are also involved in the case of self-repair of lexical errors (using wrong lexemes for a known concept), errors forced by trouble during phonological encoding (wrong spelling or prosody), slip of the tongue phenomena or errors on morphological and grammatical level [Levelt1989]. In this thesis we do not consider self-repair of utterances containing errors. In principle, it could be questionable, whether a computational model of such error handling is of interest not only for cognitive psychology but also for the development of NLS, in general. However, the detailed discussion of this question is beyond the scope of this work.