Our system SAVANT3 functions according to the surprise-based explanation principle. It tries to make an incompatibility explicit, and thus offers the student the opportunity to explain it, i.e. to invalidate it directly or indirectly.
[interaction with SAVANT3]
context: this dialog is proposed as an introduction to a digital communication course. Some concepts like digital symbols were presented, and then a printed page is displayed. The student may consider the page as a succession of characters (teletext-like coding) or as a succession of pixels (Fax-like coding). In any case, problems may arise.
S1- Can we consider a printed text as a digital message ?
E1- sure
S2- According to what I understood: this typed page can be considered as a digital message. We have to consider symbols allowing a complete description of the page
E2- characters
S3- yes... mmm... According to what I understood: the points (pixels) in the page are not taken as symbols. What else in the page could we code, if we except printed characters?
E3- you mean if there are drawings?
S4- But there may be a problem ... If we consider that one wants to code for figures (schemas, photos, signature)...
E4- but this is a printed page
S5- yes... According to what I understood: one chooses a TELETEXT-like coding. Printed characters are not the only thing to be coded...
E5- no
S6- Could you say more?...
E6- margins, paragraphs
S7- yes... mmm... But wait... According to what you said: one wants to code for positions (paragraphs, etc.). I must say I'm confused !
E7- you use special codes
S8- well, well!... Okay...
The functioning of SAVANT3 is quite simple, and reminds of the entrapment strategy used in WHY ([Collins 1976]). The knowledge given to SAVANT3 on a specific subject consists of a set of incompatibilities:
[ p11 & p12 & . . . & p1n ] ==> F
[ p21 & p22 & . . . & p2m ] ==> F
. . .
This knowledge is thus structured as if it contained only consistency production rules (as those of the COVADIS system described in [Ayel 1990]). Some of these clauses may become invalid at a given point of the conversation with the student (when at least one term is false). SAVANT3 looks for a valid clause (e.g. clause j if all pji are known to be true or are still unknown), it utters known terms and pretends to be surprised (as in S7). The student then has the possibility to deny a mentioned term (direct invalidation, as in E4) or an non-mentioned one (indirect invalidation, as in E7).
The functioning of SAVANT3 is illustrated below (cf. [dessalles 1991] for greater detail). At each point during the interaction, the system makes utterances or feigns surprise in order to verify the most promising valid clause (i.e. the clause with highest proportion of known terms). If enough terms in this clause are known, then the program utters surprise. Otherwise it utters a "canned" sentence associated with the next unknown term in the clause (as in S5). This clause may change several times during interaction. Notice that the possibility for SAVANT3 to recognize indirect invalidations comes from the fact that some premises are intentionally "forgotten" when surprised is uttered, as was the case in S7.
S4-
|
But
there may be a problem ... If we consider that
|
standard
text
|
contradiction:
figures without
|
| one
wants to code for figures (schemas, photos, signature)...
|
wording
of a predicate
|
pixels
| |
| E4-
|
but
this is a printed page
|
keyword
detection
|
direct
|
| S5-
|
yes...
According to what I understood:
|
standard
text
|
invalidation
|
| one
chooses a TELETEXT-like coding.
|
wording
of a predicate
|
||
| Printed
characters are not the only thing to be coded...
|
"canned"
sentence associated with a predicate
|
contradiction:
teletext &
| |
| E5-
|
no
|
keyword
detection
|
positions
|
| S6-
|
Could
you say more?...
|
standard
text
|
without
codes
|
| E6-
|
margins,
paragraphs
|
keyword
detection
|
SAVANT3 is able to utter surprise and to recognize relevant explanations given by the student as invalidations, but it is unable to recognize surprise in the student's utterances. Our program PARADISE (see [dessalles 1990]), designed to reconstruct conversations, is able to recognize that some utterances are intended to express a logical incompatibility (as in [ex_canteen]), and is able to suggest explanations as solutions to these incompatibilities. To achieve this, PARADISE considers predicates in the interlocutor's utterance[*] and looks for a clause in its knowledge that contains most of these predicates. This clause, if found, expresses an incompatibility, and may be considered as the intended meaning of the interlocutor's utterance.
[*] PARADISE has no syntactic capabilities and must be given utterances in a succession of subject / verb / complement inputs.