A NOTE ON PARTIAL MATCH OF DESCRIPTIONS: CAN ONE 
SIMULTANEOUSLY QUESTION (RETRIEVE) AND INFORM (UPDATE)? I 
Aravind K. Joshi 
Department of Computer and Information Science 
The Moore School, University of Pennsylvania 
Philadelphia, Pa. 19104 
Summary: In data base query systems there is an 
~p~t assumption that descriptions in queries 
must match exactly, i.e., queries are for retrieval 
only, and not for retrieval and updating simul- 
taneously. A related assumption (or constraint) 
that in questions descriptions are used refer- 
entially only (i.e., a question cannot be used 
simultaneously for questioning and informing) seems 
to hold in ordinary conversations also, with some 
qualifications. Some issues related to the 
validity of such a constraint and its relation 
to partial matching of descriptions are briefly 
discussed in this note. 
i. In a question-answer system each description 
~n a query is used referentially i.e., for each 
description one expects to find an entity in the 
data base which serves as the unique referent for 
that description. For simplicity, hereafter we 
will consider only definite descriptions (in 
particular, definite noun phrases consisting of a 
definite article, an adjective, and a noun). Thus 
in (i) 
(i) Is the red hook on the table? 
the description the red book will serve to identify 
an entity, say, e I in the data base 2 and the 
description the table, an entity, say, e 2 . The 
question can be~--~red after verifying the 
appropriate relation between e I and e 2 . For the 
purpose of makingthe definiteness transparent and 
also for simplifying the discussion in this note, 
let us assume that there is exactly one book and 
one table in the data base. 
2. The match for the red book can succeed if e 1 
~s a color attribute with the value red. The 
match can fail either due to a mismatch or a 
partial match. A mismatch will occur if e I has a 
color value other than red, say green. A partial 
match will occur if e I has an unspecified value 
for the color attribute or if the possession of 
the color attribute itself has not been specified 
for e I . 
In the rest of the discussion, we will not be 
concerned with failure due to mismatch, although 
many of the issues raised below are quite relevant 
to this ease also. We will be concerned with 
partial matches only. A partial match really is 
a partially successful match, where a part of the 
description has matched exactly, and the 
remainder has failed to match due to the lack of 
some information, and not due to a mismatch. 
3. Let us consider the case of a partial match 
where the part of the description that matched is 
sufficient to identify the referent uniquely. In 
(2) this is trivially accomplished because of our 
assumption that there is exactly one book and one 
table in the data base. 3 Although we have a 
partial match (due to the lack of the color value 
or the color attribute itself for e ), it will be • . 1 
possible to answer the question either by yes or 
no depending on whether e I is on e 2 or not, since 
the referents e I and e 2 have been uniquely 
identified. How should we proceed in this case? 
i. If we insist that each description in 
the question must match exactly, then clearly, we 
have failed to establish a refer~iceand the 
question cannot be answered. 
2. On the other hand, we may ssume that 
whenever we have a partial match and the referents 
are uniquely identified somehow, we should answer 
the question, and treat that part of the 
description which was not accounted for as new 
information. This new information can then be 
used to update the data base. Thus for the 
question (2), if the partial match is due to the 
fact that in the data base the value for the color 
attribute for e I is not specified, then we can now 
specify it to be red. If, on the other hand, the 
partial match was due to the fact that the 
possession of the color attribute itself is not 
specified for el, then the updating would involve 
adding a new attribute called color for el, and 
then specifying a value for it, whieh in is this 
case is red. The first type of update can be 
called content update and the second type, 
structure update; in the first case we have made 
a local modification of assigning a value to an 
attribute, while in the second case a new 
structural item has been added. 4 
4. There are a number of issues involved in 
adopting a strategy for updating upon a partial 
match when the matched part uniquely identifies 
the referent. We will state only two of these 
issues here and pursue the second in some detail. 
a) The part of the description that was 
missing in the data base (and which led to a 
184 
partial match) is accepted as new information and 
used for updating. The strategy followed is that 
if an exact match fails due to the iack of some 
information then the missing information is treated 
as new and updating is done accordingly. This is 
a kind of default reasoning. 5 However, it is not 
clear whether we can allow such unconstrained 
updates. In data base query systems there is an 
implicit assumption that the descriptions in 
queries must match exactly, i.e., queries are for 
retrieval only 6 and not for retrieval and updating 
simultaneously. Can we relax this requirement 
somewhat? We can get some ideas by looking at 
questions in ordinary conversations, which is what 
we will do briefly in b) below. 
b) The hypothesis (or constraint) that in a 
question construct 7 definite descriptions are 
used referentially only (i.e., a question cannot 
be used simultaneously for asking a question and 
conveying some additional information) seems to 
hold in ordinary conversations also, with some 
qualifications. The three examples below briefly 
describe some of the problems involved. 
i) Suppose that i) there is only one 
individual in the context, 2) the speaker believes 
that he is a plumber, 3) the hearer is una~ure 
of his being a plumber, and 4) the speaker 
believes that the hearer is unaware of his being 
a plumber. Under such circumstances it would be 
inappropriate to use (3) to aske the question (4), 
and simultaneously inform the hearer that (5). 
(3) When did the plumber leave? 
(4) When did the person leave? 
(5) He is a plumber. 
If (3) is used by the speaker (possibly due to a 
mistaken belief that the hearer is aware that the 
person is a plumber), it is unlikely that the 
hearer will update his model without some 
clarification or some response such as 0hi I 
didn't know that he was a plumber, i.e., the hearer 
will not update without any interrupting responses. 
This example illustrates that the question 
construct cannot be used for questioning and 
informing simultaneously, and if it appears to 
have been so used (due to the speaker's ignorance 
of the hearer's lack of some information), the 
updating by hearer is not without an interrupting 
response, thus indirectly confirming the hypothesis. 
2) Again suppose that i)there is only one 
individual in the context, 2) the speaker regards 
him as a grouch, 3) the hearer has no such 
specific evaluation of him, and 4) the speaker 
believes that the hearer has no such evaluation. 
In this case, it seems not completely inappropri- 
ate for the speaker to use (6), in order to ask 
the question (7), and simultaneously inform the 
user that the speaker regards (8) to be the case. 
(6) When did the grouch leave? 
(7) When did the person leave? 
(8) He is a grouch. 
With evaluative information, simultaneously 
questioning and informing appears to be a bit 
more convenient. If (6) is used by the speaker, 
it appears that the hearer can update his model, 
without any interrupting responses, with the 
attribute grouchyattached to the entity, as 
speaker's eval-~ion (and the hearer's too if he 
agrees with the speaker). Even if the hearer 
asks for clarification, it is likely to be of the 
form 0hi I didn't know that you thought he was a 
grouch rather than 0hi I didn't know that he was 
a grouch (compare this to the response in the 
previous example). 
3) Finally, there is an apparent violation 
of the hypothesis in examples such as (9). 
(9) Who is sitting to the right of your lovely 
wife? 
(9) can be used by the speaker to ask the question 
and pay a compliment (a side effect) rather than 
to convey new information. Thus the hypothesis 
does not appear to be violated in these cases. 
5. Some of the issues which merit further 
--investigation are as follows, i) To what extent 
the hypothesis can be violated and what are the 
side effects. If the constraint is mutually 
understood by the speaker and the hearer, then 
any apparent violation of it will be recognized 
andmay be accompanied by a side effect 
(implicature?) in addition to the updating. 2) To 
what extent updating without interrupting 
responses depends on the shape of the description, 
the syntactic construct in which it appears 
(e.g., questions, it-clefts, deelaratives, etc.) 8, 
the role it plays in the construct (e.g., subject, 
topic, etc.), the discourse model (for the 
speaker and for the hearer) created so far, 9 etc. 
3) To what extent the 'new' information used 
for updating has to be somehow relevant to the 
'old' information, either by being inferrable 
from it or by being able to fit it into the 
discourse structure created so far, etc. I0 
Notes: 
i. This work is partially supported by 
NSF Grant MCS76-19466. I wish to thank 
Jerry Kaplan, Lorrie Levin, StanRosenschein, 
Ivan Sag, and Bonnie Webber for valuable 
discussions. 
Some of the issues raised here will be 
discussed in detail in a forthcoming paper by 
Joshi and Rosenschein (Strategies for reference 
and ascription in object centered representations). 
2. We will assume a rather simple- 
minded structure for the data base. It will 
consist of entities and attributes, and relations 
among entities. 
3. However, in general, unique reference 
maybe established due to the context, and the 
structure and content of the data base. 
4. In the data base context, updates are 
usually content updates. Structure updates are not 
185 
permitted. In a conversational context and 
discourse understanding, clearly, both types of 
updates are possible. In these contexts it is not 
clear whether~we can always tell which type of 
update has taken place. Structure updates should 
be harder than context updates, cognitively 
speaking; but this is only a conjecture at this 
time. 
5. See "0nreasoningbydefault" by Raymond 
Reiter (this volume). The closed world assumption 
discussed in this paper is also relevant to 
our discussion. See also "Fragments of a theory 
of human plausible reasoning" by Allan Collins 
(this volume), and "Inferencing onpartial 
information" by Aravind K. Joshi, in Pattern 
Directed Inference (ed. F. Hays-Roth and D. 
Waterman), Academic Press, 1978. 
6. See "Cooperative responses from a natural 
language data base query system: Preliminary 
report", by S. Jerrold Kaplan, Technical Report, 
Department of Computer and Information Science, 
University of Pennsylvania, November 1977. 
7. We will limit outselves only to wh 
questions and yes/no questions. 
8. Lorrie Levinhasmade a perliminary 
investigation of the update potential of some of 
these constructs (unpublished). 
9. Entity-oriented discourse models have 
been considered for problems of reference 
(see "A formal approach to discourse anaphora" 
by Bonnie Webber, Ph.D. Dissertation, Harvard 
University, 1978). 
i0. A detailed discussion of some of these 
issues will be included in a forthcoming paper 
by Joshi and Rosenschein (see note i). 
186 
