COLING 82, ,IT. Horeck2) (ed.) 
North.Holland Publishing Company 
© Academi~ 1982 
Taking the Initiative in Natural Language 
Data Base Interactions: Justifying Why 
Bonnie Webber and Aravind Joshl 
Department of Computer and Information Science 
University of Pennsylvania 
Philadelphia PA 19104 
I. Introduction 
In answering a factual database query, one often has the option of providing 
more than just the answer explicitly requested. As part of our research on 
Natural Language interactions with databases~ we have been looking at three ways 
in which the system could so "take the initiative" in constructing a response: 
(i) pointing out incorrect presuppositions reflected in the user's query \[4,5\]; 
(2) offering to "monitor" for the requested information or additional relevant 
information as the system learns of it \[6,7\]; and (3) providing grounds for the 
system's response i.e., "justifying why". The following responses illustrate 
"presupposition correctlon"~ "monitor offers" and "justification", respectively. 
This paper describes our research on producing justifications. ("U" refers to the 
user, "S" to the system.) 
U: Is John taking four courses? 
SI: No. John can't take any courses: he's not a student. 
$2: NoD three. Shall I let you know if he registers for a fourth? 
$3: No, three - CIS531, CIS679 and Linguistics 650. 
Database systems are growing more complex in both their domain models and 
reasoning capabilities. For many reasons, it is important that such systems be 
able to point the user to what responses follow from. In this regard, 
Justification differs from argument, whose role is to convince the listener. The 
role of justification is rather to give the listener an informed basis for either 
accepting the system's response or quest\[onlng its premises or reasoning. 
There are at least three issues involved in producing Justifications: 
(I) what Justification does a query call for? 
(2) what justification can the system give? 
(3) What constitutes a clear and understandab\]e justification? 
The first depends prlmarily on whether or not the user's perceived 
expectations have been fulfilled by the system's answer. For example, the use of 
"still" in a query indicates that the user expects that the potentially changing 
situation described in hls/her query has not in fact changed. 
U: Does anyone still have an Incomplete in CSE1107 
$I: Yes, John does. 
$2: No. The last remaining Incomplete was made up on 1 December. 
If the user's expectation is correct (i.e., not all Ineompletes have been nkade 
up), then facts corroborating that expectation may constitute appropriate 
justification as in the "yes" answer above. If it isn't correct (as ~n the "no" 
case), then specifying the event that enabled the change may constitue a more 
appropriate justification. 
413 
414 B. WEBBER and A. JOSHI 
More often than not, the user's expectations will not be as clearly 
identiflable as in the "still" case above. The system will have to formulate and 
keep a growing model of the user, as its basis for determining what situation the 
user believes to hold and hence what s/he expects to learn from his or her 
queries. This is a slgnlficant area of research, which we are pursulng in 
parallel wlth the work report'ed on in this paper. 
The second ~ssue in Justifying a dlrect answer is what justification the 
system can give. In the simple case, the system may know no more than the 
partlcular facts in its data base. In that case, the only justification the 
system can provide is those facts themselves (or some description of them). For 
example, suppose the system knows for each student-course palr only whether the 
student h~d passed (P) or failed (F) the course. Then only the following 
justflcatlon is possible: 
U: Did anyone fail CIS531? 
S: Yes, John and Martha. 
On the other hand, some reasoning may have been involved in deriving the answer, 
as in the case where the system knows (e.g. has an axiom to the effect) that 
"falling" a graduate course follows from a receiving a grade of C or worse. In 
this case, speclfyi~g the particular grade each falling student recelved could be 
used to justify the system's assertion that they had failed - i.e., 
U: Did anyone fall CIS531? 
S: Yes, John recelved a C and Martha, a D. 
Or the system could give some indlcatlon of its reasoning as well, as in 
U: Did anyone fail CIS531? 
S: Yes. With a graduate course, everyone fails who gets a C or below. In 
CIS531, John received a C and Martha, a D. 
We shall clalm that posslble Justificatlons are related but not equivalent to 
how the system comes to know an answer. The reason for this reservation has to do 
with the third issue mentioned earlier in justifying a direct answer and the one 
we will address further In this paper: that is, what constitutes a clear and 
understandable justlficatlon. 
II° Towards Clear and Informative Justlfications 
The issue of clear and understandable justifications is important everywhere 
that reasoning is involved in deriving an answer, be it database systems, expert 
systems or elsewhere. While our theme is extended database questlon-answerlng, we 
believe our points are of general applicablllty. Where reasoning is involved in 
deriving an answer, clarity and ease of understanding translate into two goals: 
(i) succinct justlflcatlons that do not say more than Is necessary (lest they lose 
the user) and (2) conceptually well-structured justifications that the user can 
understand and evaluate. To meet the first goal, we are drawing upon an oft-made 
observation \[2,|0, II\], first made in Arlstotle's Rhe___ toric, that in exposltlon, 
"abbreviated proofs" ~ what Sadoek has called "modus breves" - are both sufflelent 
and desirable. Our approach to the second goal draws on the strong similarity we 
see between reasoning and hierarchical planning. Again, while our examples will 
be drawn prlmarlly from the database domain, our approach, discussed in more 
deta~l below, should be of general interest. 
JUSTIFYING WIlY 415 
A. Succinct Justifications - "Modus Brevis" 
As a simple illustration of "modus brevis" and its use in forming succinct 
justifications, consider a modus ponens deduction, possibly used in deriving an 
assertion and now to be used in justifying it. It has been observed that in 
presenting such reasoning one need not make explicit all three parts of the proof 
- the major premise (A -> B), the mlnor premise (A), and the conclusion (B). 
Rather it is sufficient to state the conclusion, with either the major premise or 
minor premise (but not both) as support. So suppose in response to the query "Did 
John fall physics?", the system makes the following modus ponens deduction 
Anyone who gets below a C fails physics. 
(e major premise *) 
John got below a C. (* minor premise *) 
John failed physics. (* conclusion *) 
The system can then justify its "yes" answer in either of the following ways, 
relying on the user's ability to recognize the underlying deduction. 
S: Yes. Everyone failed physlcs who got below a C. 
S: Yes. He got below a C. 
"Modus brevls" forms can be used in justifying other types of reasoning as 
well) both deductive and non-monotonlc. However, the speaker must be able to 
assume that the listener can, on the bas~s of what is essentially a clue to an 
argument, reconstruct that argument. On the other hand, whether the listener is 
convinced by the argument s/he deduces - i.e., whether s/he accepts the inferred 
premise - is a separate issue: the listener can always initiate a subsequent 
interaction to confirm that s/he has inferred what the speaker has intended or to 
~uestion it. For example, 
U: Did John fail physics? 
S: Yes. He got a B. 
U: Is the failing grade really B or below? 
Since the successful use of' "modus brevis" in justifications depends 
essentially on the listener's ability to reconstruct an argument from a single 
clue, it is only used in place of very short reasoning chains. On the other hand, 
the reasoning we may want to Justify may grow quite large and complex. Thus we 
expect to "modus brevis" forms to reduce the bulk of substructures, rather than 
for Justifying entire arguments. Currently we are cataloguing "modus brevls" 
forms of various argument types and notlng context-dependencies in their use. 
These schemas wlll then be used as tools for generating succinct justlfications. 
B. Clear Justifications - Hierarchical Reasoning 
The other goal of our research into producing justifications involves 
creating text structures which convey appropriate justifications in an 
understandable way. We have two claims to make here. The first is that just as 
actions have a hierarchlcal conceptual organlzatlon, so does reasonln~ - which is 
essentlally the act of supporting or denying a propesltion. The former 
organization can be used in formln~ plans, revising plans, or describing them to 
another person \[3,9\]. Siml\]ar\]y, the hierarchical organization of reasoning can 
be used both in constructln 8 a proof and in ~a result. Our second claim 
Is that the computatlonally efflclent reasoning strategies used to r~ a 
proposltien (~.e., respond to a query) are not necessarily the best ones to use in 
Justifying a result. What one wants rather is the ability to use the system's 
reasoning to suggest and instantiate conceptually more accessible strategles fnr 
organlz~ng and presenting justifications. Both claims will be discussed in this 
section. 
" 
416 B. WEBBER ~nd A. JOSH1 
Many researchers have already observed that explanations have a tree-llke 
structure. This observation reflects a view of each supported assertion as a 
non-terminal node in a tree, with the sub-tree under it corresponding to the 
reasons given in its support \[2,11\]. Since a statement acting as a "reason" may 
in turn be supported by other statements/reasons, explanations have a recursive 
structure. 
While the above is true, it masks what we see as a more significant recursive 
organization - one that reflects the inherently recursive strategies that people 
use in reasoning (i.e., in supporting or denying propositions). These strategies 
are recurslve because they contain subtasks that call in turn for other 
propositions to be supported or denied. One way to accomplish this is to chose 
and invoke another strategy. The kinds of strategies we have in mind are things 
like: 
o Simple Backward Chaining - to show that Q is true, find a set of 
propositions P1,...,Pk from whose simultaneous satisfaction Q follows. For 
each Pi, show that it follows. Hence Q must be true. 
o Simple Case Analysis - to show that Q is true, find some proposition P from 
which Q follows, ~ndependent of P's truth value. Assume P and show that Q 
follows. Assume ~P and show the same. Since either P or ~P must be true, Q 
must be true. (Alternatively, to show Q ~s false, find some P from which ~Q 
follows, independent of P's truth value. Assume P and show ~Q follows. Do 
the same for -P. Since P or ~P, ~Q must be true - hence Q is false.) 
o General Case Analysis - to show that Q is true, find some assertion P that 
is partltionable into P1,...,Pk. Assume each Pi in turn and show that Q 
follows from Pi. Since some Pi must be true given P is, Q must be true. 
(This has the obvious complementary strategy for showing Q false.) 
o Reduction ad Absurdum - to show that Q is false, find some proposition P 
whose both assertion and negation follow from Q. Assume Q and show that P 
follows. Show that ~P follows. Since Q leads to both P and ~P, Q must be 
false. 
(Other strategies are noted in the full version of this paper.) Wherever a 
strategy calls for showing "P follows" or "~P follows", there another strategy may 
be chosen and invoked in support. That such strategies are used in reasoning is 
well-known. What is significant where explanation and Justification are concerned 
is that where the strategy is clear from the text, the explanation or 
justification is that much easier to follow. 
To see this, consider the following tale, whose humor follows in part from 
the recursive use of simple case analysis in support of successive alternatives. 
What is there to be frightened of__~? 
War was on the horizon. Two students in the Yeshiva were discussing the 
situation. 
"I hope I'm not called," said one. "I'm not the type for war. I have the 
courage of the spirit, but nevertheless I shrink from it." 
"But what is there to be frightened about?" asked the other. "Let's analyze 
it After all, there are two possibilities: either war will break out or it 
won't. If it doesn't, there's no cause for alarm. If it does, there are 
two possibilities: either they take you or they don't take you. If they 
don't, alarm is needless. And even if they do, there are two possibilities: 
either you're given combat duty, or non-combatant duty. If non-combatant, 
JUSTIFYING WHY 417 
what is there to be worried about? And if combat duty, there are two 
possibilities: you'll be wounded, or you won't. Now if you're not wounded, 
you can forget your fears. But even if you are wounded, there are two 
possibilities: either you're wonded gravely or you're wounded slightly. If 
you're wounded slightly, your fear is nonsensical, and if you're wounded 
gravely, there are still two possibilitles: either you succumb and die, or 
you don't succumb and you 1lye. If you don't die, things are fine, and even 
if you do die, there are two possibilities: either you will be buried in a 
Jewish cemetery or you won't. Now if you're buried in a Jewish cemetery, 
what Is there to worry about, and even if you are not ... but why be 
afraid? There may not be any war at all!" \[I\] p.63 
In this example, "there's no call for worry" is the Q meant to be proven. The 
initial P being used to support Q independent of its truth value is "war will 
break out". Assuming "P (i.e., war won't break out), then Q follows because 
"obviously" -P -> Q. On the other hand, to show Q follows from assuming P, the 
speaker invokes a simple case analysis strategy again, this time finding P'- "they 
take you \[into the army\]" - meant to support Q independent of its truth value, and 
so forth. 
Our second claim is that the reasoning strategy used to prove some 
proposition (i.e., respond to some query) is not necessarily the best one to use 
in justifying the result to the user. What one wants is to be able to use proofs 
to suggest an appropriate organization of supportable strategies that c~n be 
Instantlated to form the basis for an understandable justification. Moore's 
"Blocks World" example \[8\] provides a good case in point. In this example, there 
are three blocks A,B and C. A is on B (On A B) and B is on C (On B C)'. A is 
green (Green A), C is blue (Blue C) and B's color is not known. It is also the 
case that whatever is blue is not green and vice versa (ALL x . Green x => -Blue 
x). The question is 
"Is there a green block on a non-green block?" 
(EXIST x,y . Green x AND "Green y AND On x,y) 
Resolutlon Is the only slmple machine reasoning method that can find the 
correct answer "yes" to this problem. (Simple backward-chalnlng or 
forward-deductlon systems require an additional procedure called "restricted 
goal/fact resolution".) Converting the above facts and axioms to clause form and 
using resolution, one proof goes as follows: 
(1) -Green x OR Green y OR "On x,y 
(2) Green A 
(3) Green y OR "On A,y 
(4) On A,B 
(5) Green B 
(6) Green y OR -On B,y 
(7) On B,C 
(8) Green C 
(9) -Green z OR -Blue z 
(i0) -Blue C 
(II) Blue C 
(12) NIL 
\[negation of theorem\] 
\[axiom\] 
\[resolving 1 and 2\] 
\[axiom\] 
\[resolving 3 and 4\] 
\[resolving I and 5\] 
\[axiom\] 
\[resolving 6 and 7\] 
\[axiom\] 
\[resolving 8 and 9\] 
\[axiom\] 
\[resolving I0 and 11\] 
What this proof does not make obvious is that the answer follows by 
considering all colors that B can take (that is, green and non-green). Neither 
does the proof make obvious that the answer follows in either case, even though a 
different situation holds. That these are the elements of what people give in 
what they think of as understandable justifications can be seen in protoeals that 
we have collected of people justifying their answers to Moore's problem: most of 
them do so using a ~ case analysis strategy. For example, 
418 B. WEBBER and A. JOSHI 
"Yes - it doesn't matter what color B is. If it's green, then it is the 
green block on top of a nos-green block C. If it's not green, then A is the 
green block on top of a non-green block B." 
Our point is that while resolution theorem provers may be appropriate 
reasoning engines for data base systems, their proofs do not form a good basis for 
understandable Justifications. Thus at least part of our research is aimed at 
discovering whether one could recognize in the proof tree of a theorem which of 
the above understandable reasoning strategies could be in justifying the result 
and then construct an appropriate valid justification in terms of those 
strategies. 
III. Conclusion 
This paper has reported briefly on our research on justification. It is an 
abbreviated version of our technical report CIS-82-1, which can be obtained by 
writing to the authors. 
The authors would llke to thank Barbara Grosz for her helpful comments 
on previous drafts. 

REFERENCES 

\[1\] Ausubel, N. (ed.) A Treasur~ of Jewish Folklore. New York: Crown 
Publishers, 1948. (Abridged edition published by Bantam Books, 1980.) 

\[2\] Cohen, R. "Investigation of Processing Strategies for the Structural Analysls 
of Arguments". Proc. 19th Annual ACL Meeting. Stanford CA, June 1981. 

\[3\] Grosz, B. Th____eeRepresentatlon an___ddUs__ee of Foc____us In. Di~ Understanding. 
Technical Report 151. SRI International, Menlo Park CA, 1977. 

\[4\] Kaplan, S.J. Cooperative Responses from a Portable Natural Lan~ Data Base 
Que~ System. Ph.D. thesis. Department of Computer and Information Science, 
University of Pennsylvania. 

\[5\] Mays, E. "Correcting Misconceptions about Data Base Structure". Proc. 
3-CSCSI. Victoria, B.C., May 1980. 

\[6\] Flays, E., Lanka, S., Joshl, A.K. and Webber, B.L. "Natural Language 
Interaction with Dynamic Knowledge Bases: Monitoring as Response". Proc. 
8-1JCAI. Vancouver, B.C., August 1981. 

\[7\] Mays, E., Joshi, A.K. and Webber, B.L. "Taking the Initiative in Natural 
Language Data Base Interactions: Monitoring as Response". Proc. 1982 
European Conference on Artificial Intelligence, Orsay, France, July 1982. 

\[8\] Moore, R. Reasoning from Incomplete Knowledge in a Procedural Deduction 
S s~. Technical Report 347. MIT Artlflcial Intelligence Lab. Cambridge MA, 
1975. 

\[9\] Sacerdoti, E. A Structure for Plans and Behavior. New York: Elsevler, 1977 

\[i0\] Sadock, J. '~odus Brevis: The truncated argument". Papers from the 13th 
Chicago Linguistics Society Meeting, Chicago IL, 1977. 

\[II\] Welner, J. "BLAH, A System which Explains its Reasoning." Artificial 
Intelllgence, 15, 19-48. 
