From: AAAI Technical Report SS-92-02. Compilation copyright © 1992, AAAI (www.aaai.org). All rights reserved.

Hyperproof:
Logical Reasoning with Diagrams

Jon Barwise and John Etchemendy

Brief overview an order of magnitude shorter and more natural thau

Hyperproof is a Macintosh program designed to help corresponding proofs in first-order logic which eschew
students learn howto reason logically, using either sen- diagrams. Understanding why this is so can only be of
tenccs of first-order logic, or diagrams, or both. 1 An help in automating reasoning.
ilhistration of a typical Hyperproof screen is shownin Second, for certain kinds of problems, the algorithms
Figure 1. used for checking diagrammatic proofs can be turned
Hyperproofis a descendant of our earlier courseware on their head to discover one-step solutions that people
Tarski’s World [2]. Tarski’s World uses diagrams to cannot even follow, must less find. This suggests that
help teach the language of first-order logic, and has these algorithms, or ones like them, could be put to
been very successful. Hyperproof extends the ideas good use in automating certain reasoning tasks.
embodied in Tarski’s World into tile realm of reason- Finally, HyperTroof has a much more liberal notion
ing. It is intended to be part of a course aimed at of what counts as an inference task than does tradi-
teaching students to solve analytical reasoning puzzles tional logic. For example, in this system, one is able to
and logic problems by whatever means of representa- prove that a purported conclusion does not follow from
tion are most appropriate to the problem at hand. On the premises, by constructing a diagram which depicts
our view, an important part of solving such a problem a situation in which the information in the premises
is figuring out what representational scheme to use, obtains but the conclusion fails. Thus we can pose
whether sentences, diagrams of some sort, a combina- questions neutrally: does the conclusion follow fi’om
tion of both, or something else altogether. the premises or not? One and the same set of tech-
Before we say more about what Hyperproof is, let’s niques and tools can be used for providing the solution,
say a bit about what it isn’t, just to forestall any mis- whether it is positive or negative. Wethink that much
conceptions. Hyperproof is not an automatic theorem ordinary reasoning is like this: visualizing a situation
prover. Nor is it an AI system. It is rather a compu- where the premises hold but some undesirable possible
tational environment which provides a setting: consequence fails, and then working toward realizing
that situation.
1. for teachers to pose interesting inference problems
using pictures and first-order sentences,
Historical introduction
2. for the student to give mathematically rigorous
proofs, again using some combination of pictures and Over the past twenty years, some workers in AI have
first-order sentences, and explored uses of diagrams in automating reasoning. By
contrast, logicians have had little truck with diagralns
3. for checking these proofs to makesure that they are and other forms of non-linguistic representation. ~- The
correct and fillfill the original goal set by the prob- traditional attitude is all too well captured by the fol-
]ei1].
lowing quotation:
While Ilyperproof is not all automated reasoner or [The diagram] is only an heuristic to prompt cer-
an AI system, we believe that it is of someinterest to tain trains of inference; ... it is dispensable as a
these fields. First, proofs in Hyperproof are typically proof-theoretic device; indeed, ... it has no proper
1While the ideas for Hyperproo/areours, the program place in the proof as such. For the proof is a syn-
itself is the crcation of GerardAllwein, MarkGreaves, and tactic object consisting only of sentences arranged
MikeLcnz with help from Alan Bush, DougFelt, and Chris in a finite and inspectable array. (Neil Tcunant,
Phoenix. Weexpress our appreciation of and admiration in [6])
for their work here. The term "Hyperproof" is suggested
by the notion of hypertext, where information is presented 2There have been a few notable exceptions, of course,
in a host of non-linear, often non-textual means. like Venn,Euler, and Peirce.

77
 It-

I,
’ ~!’HBBetween(,

GJveaGivel~]
~- FrontOf(o,c)
¯ 3x Left0r(x, d) <--> Lerge(b)
Given
I|
3x :ly (~Tet(x) ~Tet(y) A Between(b, x,y))
Glven
Given
lea
Iil
¯ Left0f(d, f) Given
II
Dodec(e) Given
Iill
-~3gLeft0f(xo e) V Large(e) ~ven
~il

Figure 1: Problem: Which objects are named a, b, ... ?

One aim of our work, as explained in [1], is to demon- this is probably an understatement. Figuring out how
strate that this dogmais quite misguided. Webelieve to represent the information at hand is often the most
that many of the problems people have putting their important part of the solution. What we seek to teach
knowledgeof logic to work, whether in machines or in our students, then, is to use the most appropriate form
their own lives, stems from the logocentricity which of representation for the reasoning task at hand. As
has pervaded its study for the past hundred years. long as the purported proof really does clearly demon-
Weapproach logic from an informational perspec- strate that the information represented by the conclu-
tive. Whereverthere is structure, there is information. sion is implicit in the information represented by the
But in order for agents (animals, people, or comput- premises, the purported proof is valid.
ers) to traffic in information, the information must, in Whyare logicians so suspicious of diagrams and
some way or other, be presented to or represented by other forms of non-textual representation? The an-
the agent. Typically, a given representation, or family swer goes back to the tradition in geometry, where
of representations, will represent certain information diagrams were viewed with suspicion by the ancients.
explicitly, while other information will be implicit in Certain mistaken proofs were seen to result from being
the information explicitly represented. Inference, as we led astray by a misleading diagram that accompanied
understand the term, is the task of extracting informa- it. So, the tradition went, the diagram should, in the-
tion implicit in someexplicitly presented information. ory, be eliminable from the proof. The textual part of
This informational perspective is part of the modern, the proof should stand on its own two feet.
semantic approach to logic associated with names like This argument is itself a nonsequitur. If we threw
G6del, Tarski, l~obinson, and the like. On this view, out every form of reasoning that could be misapplied
a purported rule of inference is valid or not depending by the careless, we would have very little left. Math-
on whcthcr it in fact guarantees that the information ematical induction, for example, would go. No, the
represented by the conclusion is implicit in the infor- correct response is not to throw out methods of proof
mation represented by the premises. But when one that have been misapplied in the past, but rather to
takes this informational perspective seriously, the lo- give a careful analysis of such methods with the aim of
gician’s disdain for nontextual representations seems understanding exactly when they are valid and when
like an oversight, a case of dogmathat desperately they are not.
needs reexamination. The most casual survey of the A nice case study along these lines has been car-
ways people actually represent information shows an ried out by Sun-Joo Shin in [3] and reported in [4]. In
enormous variety of representational devices that go introductory logic, many people teach the method of
beyond simple text. Venn diagrams. But often they also tell their students
In carrying out a reasoning task, part of the solution that Venn diagrams are only a heuristic aid to giving
lies in figuring out how to represent the problem. In proofs. A "real" proof has to be given in first-order
problem solving, well begun really is half done. Indeed, logic. Shin shows that this is a mistake. She gives a

78
 [] Given
¯ VX(TeL(x) -.) Small(x)) Given
¯ Vx (Cube(x) 4 Hedtum(x)) Given
¯ VX(Dodec(x) ~, Large(x)) Given
j= VVVU (Semeslze(u,v)<-> SemeCol(u, Given
I" Vw -~Dodec(w) Glvee
I" Vx Vy ((Tet(x) ^ Tet(y)) -’, Given
¯ 3W Tet(w) Given

Figure 2: Another Hypevproof Problem

rigorous analysis of Venndiagrams as a formal system, both.

with its ownsyntax, semantics, and notion of logical Hyperproofhas all the usual rules of first-order logic,
consequence. She shows that the usual rules of infer- plus a number of rules for dealing with diagrams and
ence are sound with respect to this notion of logical the interplay between diagrams and sentences. In this
consequence. Furthermore, she shows that, with some short abstract, we discuss two of these rules, the rules
trivial additional rules that had been overlooked, one of Observe and Cases Exhausted (and a special case
caq give a completeness proof for the deductive system. of the latter, Apply). The rule Observe allows us
An even earlier study is given by John Sowain [5]. to extract sentential information from diagrammatic,
In the early chapters of that book, Sowagives a careful while the latter allows us to go the other way around.
formal treatment of Peirce’s existential graphs, includ- In Hyperproof every diagram represents partial in-
ing their syntax, semantics, and proof theory. This is a formation about some class of blocks worlds, the kind
formalism that is expressively equivalent to first-order of partial information that can be modelled by a par-
logic, but it is given in purely diagrammatic terms. tial first-order structure. Observe allows the user to
The importance of these results is this. They show infer any sentence which is definitely true in the par-
that there is no principled distinction between infer- tial structure that corresponds to the diagram--that
ence formalisms that use text and those that use dia- is, true according to the truth evaluation schema em-
grams. One can have rigorous, logically sound (and bodied in the (weak) Kleene three-valued logic. The
complete) formal systems based on diagrams. And three values are true, false, and unknown.
while Venn diagrams may seem very simple, they are
The truth behind the old cliche referred to earlier
probably uscd more often in real life than first-order is evident from the uses of this rule. For any given
formalisms. diagram there are literally an infinite number of sen-
tences that can be observed to hold on the basis of the
Hyperproof diagram. And typically, if one wanted to completely
Hyperproof is a working program that we are nowusing specify a given diagram with sentences, it would take
iu teaching elementary logic courses. Wewill demon- an enormous number. Indeed, it is not always possible
strate thc program at the meeting. Here we would to do it at all, since there are elementary equivale,t.
like to draw attention to one very important part of diagrams (i.e., diagrams satisfying the same first-order
Hyperproof. sentences) that represent distinct possibilities.
One of the differences between diagrams and text is But the most important features of the Hypcrproof
suggested by the old Chinese saying "A picture is worth system are the techniques for transfering information
ten thousand words." The truth behind this is that di- from sentences into a range of diagrams. Thus, we
agrams and pictures are extremely good a presenting a are typically give a diagram D and some sentences
wealth of specific, conjunctive information. It is much S1,...,S, describing the target world. Based on one
harder to use them to present indefinite information, of these sentences, say S, we see that there is a range of
ncgative information, or disjunctive information. For possibilities, each of which can be embodiedby flesh-
these, sentences are often better. That is why we fa- ing out the diagram D in some way. Thus we create
vor a heterogcncous system, one that allows us to use some new diagrams D1,..., Dk, each aa amplification

79
 I Problem3 Counterexamplo

Figure 3: Hyperproofs solution

of D (in that it carries more information than D), and diagram D* which is incompatible with D’ and which
say that these exhaust all the possibilities, given the settles all issues raised in D~, we call this a possible
sentence S. This step is valid if any way the world counterexample, and give it to the user as a possibility
could be, consistent with D and S, is also consistent that needs to be considered explicitly. In this way, the
with at least one of these new diagrams. This is the search space is greatly reduced and the rule is sound,
rule of Cases Exhaustive. in that it never ratifies an invalid use of Apply. The
A special instance of Cases Exhaustive is where algorithm is not complete, in that it will not be able
k = l, that is, where there is just one new diagram. to recognize some valid uses of Apply, but in such
Applying the rule in this case means that the informa- cases it returns a possible counterexample for the Nser
tion added to D in the new diagram D1 is implicit in to consider.
tile sentence S. Wecall this special instance Apply, This rule, and the more general version of Cases
since you can think of it as applying the information Exhaustive, give proofs in Hyperproof quite a differ-
in S to add new information to the diagram. ent character than we expected. It is very difficult
Rather than describe the general Cases Exhaus- to find problems where the best solution relies on the
tive algorithm, let us discuss the special case of Ap- standard sentential rules of inference. It is typically
ply. The interested reader can readily think through muchmore efficient to break into several diagrammal~le
the extension to the more general rule. So we imagine cases and then work on these diagrams.
ourselvcs in tile following situation. The user has a di-
agram D and a sentence S. She wants to apply S to The role of the computer
I.
amplify tile diagram to D
As we said earlier, Hyperproof is not an AI program. It;
Ideally, what one would like would be an algorithm
is, rather, a platform for teaching students somebasic
which did the following: If this Apply is legitimate,
then say so. If not, then produce a counterexample, principles of valid reasoning. The role of the compllter
is two-fold:
by displaying a diagram D* that amplifies D in a way
I.
where S is true but which is incompatible with D ¯ It is an environmentfor creating proofs: it provides a
This specification is not in general realizable in a convenient environment for generating visually per-
computationally feasible manner. The reason is that it spicuous representations of the situations the stLi-
can happen that such a diagram would have to settle dents are asked to reason about.
too many issues, issues which themselves can be re- * It is a proof checker: it checks the students work to
solved in a variety of ways. Thus one gets into a huge see if each step is correct, and if the proof as a whole
scarch space. satisfies the goal of the problem.
In Hyperproofwesolve this problem in two ways, re-
lying on partial information. First, we only search for This said, though, it must be said that the program
countercxamplcs amongthose diagrams that settle just does have some smarts. It is able to perform some
issues raised in DS.3 If we can find a such a diagram reasoning tasks that outstrip our original iutentioas
in which S is true, then we will have found a coun- for it. While we do not want to make any AI claims
tcrexample. But we do not insist that the sentence for our program, we would be gratified if the features
S come out as true in D*, only that it not come out that make it smarter than we intended were of use il~
as false. If it comesout as neither true nor false in a AI.
Most noticeable of these are the rules of Apply and
3 For example,if D’ settles the size of a single block b, Cases Exhaustive. By repeated uses of these rules,
and that is all, then the only issue we consider is the size one can basically have the computer figure out the
of that block. All other amplifications of D are ignored. cases for a case analysis. All one does is start it off

8O
with one case to activate the issues you want consid- 6. Neil Tennant, "The Withering Away of Formal Se-
ered in your case analysis. mantics," Mind and Language, Vol. 1, No. 4, 1986,
Using this trick, the user can prod the program to pp. 302-318.
performfeats of inference that are very difficult for the Jon Barwise
user herself to follow. For example, it may produce Department of Computer Science
a set of exhaustive cases which she cannot readily see Indiana University
to be exhaustive. For our pedagogical purposes, this Bloomington, IN 47405
poses a bit of a problem, since we want to teach the
student that proofs are composed out of self-evident E-mail: [email protected]
steps. But if you think of using Hyperproof-like tech-
nology for automated or computer-aided inference in
other settings, this feature could be a big win. John Etchemendy
Weclose with a simple example of this phenomenon. Center for the Study of Language and Information
Figure 2 shows a sample problem given in Hyperproof. Ventura Hall
The goal of the problem is to determine the sizes and Stanford University
shapes of the objects depicted. (In Hyperproof, we as- Stanford, CA94305
sumethat there are only three sizes of objects, small,
mediumand large, and three shapes, cubes, tetrahe- E-mail: [email protected]
dra, and dodecahedra, so the sizes and shapes of the
objects depicted are, in fact, all determined by the sen-
tences in the given.) This is not an entirely trivial
chore. However,using a trick, and Hyperproof, she can
make short work of it. If she makes a random guess
at the sizes and shapes, and asks the program to ver-
ify that her guess follows from the given information,
Hyperproofquickly comes up with the counterexample
shown in Figure 3. This counterexample is, of course,
4the correct answer to the problem.
This is a fairly simple example, but the program can
solve much more complicated problems in the same
way. For example, try to figure out which objects de-
picted in Figure 1 have which of the names used in the
sentences displayed. The trick just described makes
short work of it.

References
1. Jon Barwise and John Etchemendy, "Valid inference
and visual representation," in Visualization in Math-
emalics, ed by Zimmermanand Cunningham, Math-
ematical Association of America, 1990
2. Jon Barwise and John Etchemendy, Tarski’s World
3.0, CSLI Lecture Notes, University of Chicago Press
1991
3. Sun-Joo Shin, Vaild Reasoning and Visual Represen-
tation, Dissertation, Stanford University, 1991
4. Sun-Joo Shin, "An Information-Theoretic Analysis
of Valid Reasoning with Venn Diagrams," in Sit-
nation theory and its applications, Part 2, ed by
Barwise et. al., CSLI Lecture Notes, University of
Chicago Press, 1991
5. John Sowa, Conceptual Structures: Information
Processing in Mind and Machine, Addison-Wesley,
1984
Interestingly, it takes a lot longer for the programto de-
termine that the counterexampleis the only solution than
that it is a solution.

81