II~PERENCING AND SEARCH FOR AN ANSWER IN TIBAQ 
Petr Oirk~, Jan HaJi~ 
Center of Biomathematics, Czechoslovak Academy of Sciences, 
142 20 Praha 4, Vtde~sk~ 1083, Czechoslovakia 
and 
Faculty of Mathematics and Physics, Charles University, 
118 00 Praha 1, Malostransk~ n. 25, Czechoslovakia 
The objective of this contribution is to describe two 
procedures employed in the question answering method TIBAQ 
(~ext-and Inference-Based Answering of ~uestior ~, an overall 
outline of which is given in Sgall's paper (COLING 82 - Pro- 
ceed~n6s) o 
(i) The procedure of inferencing is activated after the 
input sentence has been recognized as a question. However, it 
would be useless to search through the whole data base. Thus 
only a small part of the data base SQ, viz. the set of re- 
presentations of sentences relevant with respect to a given 
question Q, is activated, which consists of those sentences 
that contain at least one term sem~utically equivalent with 
an element actually occurring in Q, i.e. in the question to 
be answered. This set is selected from the data base for the 
inference rules to operate on it. Let us denote by Cn(SQ) the 
set of all consequences of SQ. Thus Cn(SQ) is the theoretical 
set of statements relevant for the build-up of the answer. 
Such a set of consequences would grow beyond any limits, so 
that it is necessary to formulate a strategy which controls 
the whole process. Two different strategies will be discussed. 
In the first experimen%e with TIBAQ, we represent the 
meaning of the sentence by a dependency tree. Inferences are 
- 139 - 
performed by using rules for nondestructive conditional re- 
writing of such trees. Prom a formal point of view a set of 
inference rules can be considered as a set of productions 
over oriented projective labelled trees. Labels are used for 
representing elementary terms of knowledge (meaning), while 
pro~ectivity permits a treatment of the contextual role of 
the words in the sentence. For linguistically motivated rules, 
which are emphasized in our paper, we can use a bottom-up 
strategy carried out here by using Colmerauer°s Q-systems. 
Such a strategy is not fully satisfactory for an inference 
process controlled by logical patterns. The logical deductive 
ability of the system should rather be connected with back- 
ward chaining strategies and thus more advanced devices, 
using e.g. backtracking mechanism, are needed. 
(ii) The procedure of the search for an answer to the 
question Q operates in the set Cn(SQ). A consequence cECn(SQ) 
must fulfil the following conditions to be chosen as a (full 
or partial) answer to Q: 
(a) the root of c must be either identical with the 
root of Q (identity means coincidence in all parts of the 
complex label of the node, where "coincidence" is defined as 
allowing for certain specific differences such as that of 
singular vs. plural under certain conditions, of an adverbial 
of Manner vs. adverbial of Regard, etc.), or, if the lexical 
part of the label of the root of Q equals "d~lat" (English do 
in the meaning of a full verb), then the lexical part of the 
label of the root of c may have any shape provided it includes 
and index denoting the feature "Activity" ! 
(b) c must comprise a path that is identical with the 
path in Q that leads to the node labelled by the representat- 
ion of the question word (WH), except that the lexical part 
of the label of the counterpart of WH in c consists in a con- 
crete lexical word, possibly accompanied by words dependent on 
it. 
- 140 - 
(c) £ and Q must have at least one more node in common 
(with a coinciding and lexically specified label), dependent 
on an identical node on the path to WH (in c). 
If there is no c matching the conditions (a) to (c), 
th~ system answers I don°t know. The distinction between a 
full and a partial answer is determined as follows: if the 
dependency trees of Q and £ match the conditions (a) to (c) 
and Q does not comprise any nods not having a counterpart in 
c, then _o is printed as a full answer, otherwise it is print- 
ed as a partial answer, prefixed by "I know that ...". 
This procedure makes it possible to respect among other 
relevant issues also the dichotomy of topic and focus, so that 
e.g. if the set of statements contains the assertion .Arithmet- 
ical operations are carried out by the device D" (rather than 
"The device D carries out arithmetical operations"), then the 
questions "What is carried out by the device D?" will be 
answered "I know that arithmetical operations are carried out 
by the device D", which points out the possibility that 
arithmetical operations constitute only an unimportant part 
of the set of processes carried out by D. 
- 141 - 
