Using Misunderstanding and Discussion in Dialog as a Knowledge Acquisition or Enhancement Procecss - Distributed Artificial Intelligence, Agent Technology, and Collaborative Applications - page 84

Λ = ( Q i +1 → Q i +2 ), …, ( Q n −1 →¬( P n ))).

If the 'teacher' gives the knowledge ( Q i → Q i +1 ) then a conflict arises as from ∆ as we can deduce
P n from P 1 and now from Θ, ( Q i → Q i +1 ) and Λ as we can deduce ¬( P n )) from P 1 .

These two types of conflicts are hard to detect as we have to locate the two or three series of implica-
tion being involved. The conflict elimination is done in the same way as the previous case: The 'student'
will have to check with the 'teacher' the validity of each implication from ∆ and Θ until having found
one which is wrong in order to remove it.
Fact conflicts. This case appears when the 'student' thinks a fact true ( p ( A )) and the 'teacher' tells
it the fact is false (¬( p ( A ))) or the 'teacher' gives it a knowledge which allows the 'student' to deduce
that its fact is false. The conflict management is close to the implication conflict one and several cases
can occur (the 'student' here thinks ( p ( A ))):

• If the 'teacher' tells it (¬( p ( A ))), then the 'student' removes ( p ( A )) from its KB;

• If the 'student' thinks, ( p ( X )→ p 2 ( X )), ( p 2 ( X )→ p 3 ( X )), …, ( p n −1 ( X ) → p n ( X )) and the 'teacher' tells

it (¬( p n ( A ))) the a conflict arises as the 'student' can deduce ( p n ( A )) from ( p ( A )) and the series
of implications, which is contradictory with (¬( p n ( A ))) . The 'student' will have then to remove
( p ( A )) or one of the previous implications. It will then perform a series of information requests to
the 'teacher'.

Connexity variations in conflict resolution. We can notice a connexity increasing when the 'student'
attempts to learn new data. In the example where the 'student' has to know the predicate ( rectangle ( x )) in
order to learn the new formula, it was already knowing the three predicates ( parallelogram ( x )), ( DITM ( x ))
and ( DLE ( x )), and by learning the new predicate, the three old ones have been added to themselves a
new link between them. The connexity is then increased.
However, when handling conflicts, an implication can be questioned then removed, leading to a
decrease of the KB connexity level if the implication would be located inside an implication chain and
would be a cut-vertex. In order to minimize the risks of a connexity decrease, the 'student' can use an
heuristic for selecting implications in an optimal order: By starting with end of chain of implications,
their removal would not reduce the KB connexity level.

Symmetry Between Dialog And KB Management In Discussions With Issues . When an issue
(misunderstanding or KB conflict) arises within the discussion, it breaks the symmetry of the current
discussion and opens a new one, bound to solve the issue. This new discussion will install a new sym-
metry with a repair process headed by one of the agents, depending on the type of the issue.
Tables 1 and 2 describe the different steps of our repair common scheme, for respectively a misun-
derstanding and a conflict situations. 'S' means a 'student's' intervention and 'T' a 'teacher's' one.
In Table 1, a say - dissatisfaction message informs the 'teacher' of the 'student's' misunderstanding.
Let us assume the 'student' don't understand a predicate P . The 'teacher' will head the dialog and
use a misunderstanding strategy to solve the issue. Each give - information () message conduces the
'student' to check into its KB the presence of the provided implication. Once the 'teacher' deems the
student have the knowledge required to understand P , it will assert the knowledge link by giving an
explanation, the formula F . The 'student' ends this repair discussion by informing its interlocutor with
a say - satisfaction message.