Análisis del ambiente externo

A corresponding integration of the vertical and horizontal dimensions of explaining spatial cognition is exemplified by the system SOCCER of the project VITRA [ANDRÉ

ET AL. 1988]: in this case, the exemplary radio sports reporter from the beginning is considered. The explanation of the visual perception, which is part of the foundation by

reference semantics of the utterances of the radio reporter, follows the exemplary line given in section 4.3.2 up to the field of spatial concepts. Simplified versions of the concepts underlying static spatial relations, like »being in«, »- at«, »- near«, »- to the left« holding between a reduced version of sortal objects are determined. The concepts of spatial events, like »doing a double pass with«, are additionally defined as a temporal sequence of phases during which certain spatial relations hold. From the resulting sets of spatial propositions, some are finally chosen to be communicated and transformed into a corresponding verbal manifestation (Fig. 123):

S1: Miller, the defender, stands just left to the penalty spot. S2: Miller gets the ball and runs with it close to the centre circle.

As was mentioned in sections 3.5 and 4.4.2, any adequate theory of communication explaining the behavior of a speaker also has to consider the audience in a particular way: the speaker has to be conceived of as somebody who also sets himself in the posi- tion of his audience. He has to play anticipatorily its role in the language game in order to really communicate. In VITRA, this demand is answered by means of the listener model ANTLIMA: we focus here only on the static spatial relations, as in sentence S1, although spatial events as in S2 are dealt with accordingly, as well. The understanding of the audience is modeled with the three steps described above:

First, the proposition of the (planned) utterance is projected to the lower levels im- plementing the spatial field: i.e., restrictions of the spatial interaction with other objects

are transferred mainly into restrictions of the locations of the objects (plus the part-whole as- pects of the objects involved); this transforma- tion – the schema of the corresponding spatial concept – is encoded in ANTLIMA by means of functions called ‘TyPoFs’,98

which are al- ready applied for recognizing spatial relations: they can easily be viewed as the characteristic functions of the fuzzy sets of situations to be described by the corresponding relation (cf. again Fig’s 66 & 67, Sect. 4.3.2.3).

Second, the context of the planned utterance is revised on the lower level, i.e., as a mental image: the locations of the objects are chosen by means of a hill-climbing algorithm ruled by the TyPoFs and depending on the contextual positions. Figure 124 illustrates the influence of three different geometrical contexts (starting positions) on the location selected, namely ‘to be in front of the penalty area’. The hill- climbing algorithm determines maximally typical positions for all objects localized with respect to the geometric restrictions given by the predication. Therefore, the image con- struction concretizes the consequences of an additional proposition to the given contex- tual image – an implicit type of spatial reasoning. If an image can be constructed with highly typical positions for all restrictions, the utterance under consideration must be rated plausible in the given context.99

Third, the schemata of the spatial concepts (object models, TyPoFs, and definitions of spatial events) are applied to (re)construct the context on the level of the spatial field: this finally renders explicit the implicatures included. Another set of spatial propositions is the result.

That set modeling the anticipated understanding of the audience has to be compared in the listener model with the understanding intended by the speaker, i.e., what has been actually perceived: the differing propositions are used in an anticipation feedback loop for an enhancement of the propositions to be effectively uttered (cf. again Fig. 123, and Fig. 84, Sect. 4.4.2.2).

Note, that the image constructed – i.e., the image the speaker anticipates the listeners can construct when told the proposition in question – cannot directly be compared to the set of propositions describing what the speaker has observed. A first guess might be to use the percept instead – after all, the audience should have a mental image correspond- ing to the speaker's percept. Percept and mental image are assumed to be of the same type, so that the comparison can be done syntactically. Unfortunately, such a solution is not exactly plausible. That conception does not take into account that the speaker's communicative intentions are – even in the case of an objective description – not identi-

98

‘TyPof’ is a speaking acronym for ‘Typicality Potential Field’, alluding to its use in a gradient search: it tips off the maximally typical positions falling under the spatial concept in question.

99

The resulting image is later used as the starting point for constructing the image for the utterances planned next, and also to check whether a noun phrase to be employed in that utterance denotes uniquely an object in that imaginary visual field of the listener solving the question of reference.

Figure 124: Visualization of the TyPoF for a player being in front of a penalty area, and approximation paths for several contexts

cal to the speaker's “raw percept”: it is the set of (spatial) propositions re- flecting what the speaker has recog- nized in the percept, which has to be considered. Even if we assume that the comparison between the percept and the mental image could be used – for example by means of the distance between the two incarnations of an object in the two images – we still have a serious difficulty: are all dif- ferences really equivalent? Imagine a soccer field with two balls – a black one representing the position per- ceived by the speaker, and a white one representing the position antici- pated by the image construction. Let

us assume furthermore that the two balls are in one case about one foot apart some- where in the middle of the field away from any landmark, and in another case – with the very same distance between each other – on different sides of the outside line (Fig. 125). It should be obvious that in the first case, the difference is not considered essential, and correspondingly should not trigger a reaction in the listener model. How- ever in the second case, the two positions are different: if the white ball is the one out- side the field, the listener model has in fact predicted that the audience falsely under- stands that the ball is outside of the game: a correction is then highly recommended.100

The recognition component of the speaker model classifies exactly percepts with es- sential differences; it generates the same sets of propositions if two percepts do not dif- fer essentially. Therefore in the listener model with mental images, the very same “cog- nitive abilities” are employed with respect to the mental image in order to generate a propositional description of what the audience (at least presumably) is able to recognize in its mental image. That set of propositions can easily be compared to the analogous set of the speaker providing the means for dealing with the problem of adequacy mentioned above. Thus, the sequence of recognition and secondary selection based on the antici- pated mental image reflects exactly the speaker’s own activities with respect to his per- cept: recognition and primary selection. The analogy of the »seeing by one’s mind’s eye« and the »seeing by the physical eyes« becomes even more plausible: as was said before, it is believed that the listeners can “see” the consequences of integrating a new proposition in the contextual knowledge in the mental image.

As is demonstrated in Figure 126, the spatial restrictions holding for an object simul- taneously (e.g., during an event phase) can be easily combined on the level of TyPoF’s. Only if the combination is consistent, the resulting typicality field has a maximum close to the ultimate value. Furthermore, the context-sensitivity of the algorithm for finding the maximum of the typicality distribution as demonstrated in Figure 123 adds another advantage when considering spatial events: the positions of the objects at consecutive

100

See also again Figure 85 (Sect. 4.4.2.2): the argument used here is also valid for the comparison step in the listener’s anticipation feedback loop: the pictures can only be compared with respect to a particular “reading”, not as such.

moments of an event phase are developed in a cinematographic procedure taking the position at t-1 as starting position for the gradient search for the position at t.101

The further uses of the difference propositions in the listener model are not directly related to the data type »image«. They have been described elsewhere in detail [SCHIRRA 1997].