You are on page 1of 17

Grice for graphics: 1 Introduction Author's address

Human Communication
pragmatic Research Centre
University of Edinburgh
Petre and Green (1992) observe intriguing
implicature in phenomena which occur when real designers
2 Buccleuch Place
Edinburgh, EH8 9LW
network diagrams use complex graphical representations. They Scotland

also offer a promising account for these phe­ Jon Oberlander is EPSRC Advanced
Jon Oberlander nomena. This paper explores an alternative Fellow at the Human Communication
Research Centre. The support of
account, which relies on the notion of graphical the Economic and Social Research
Intriguing phenomena occur implicature. Whether or not the current Council for HCRC is gratefully
when experts use computer- acknowledged.
assisted design tools in elec­
account is useful has broader significance,
tronics. It can be seen that because if it is useful, it helps supply the first This paper is based on a talk pre­
tools must support information premise for the following argument: (i) there sented at the Summer School organ­
ised by the Joint Councils' Initiative
access, 'escape from formal­ are certain parallels between pragmatic phe­ in Cognitive Science and Human-
ism', 'secondary notation', nomena in natural language and graphical rep­ Computer Interaction, held at the
and differences between University of Edinburgh in September
individual users. This paper
resentation; (ii) formal techniques have been 1993. Thanks to Corin Gurr, John
explores a new account for the developed for modelling some of the natural Lee, Ehud Reiter, Keith Stenning,
Paul Stiff and my three referees for
data, relying on the idea of language phenomena; and thus (iii) once the helpful comments; I am particularly
graphical implicature, gener­ graphical data are better-understood, it may be grateful to Thomas Green for his
alised from Grice's conversa­ possible to treat them with formal techniques encouragement, without which this
tional implicature. All com­ paper would never have been start­
from natural language pragmatics. ed, let alone finished. My thanks
municative artefacts carry
also to Marian Petre for her help in
implicatures, significance The paper has the following structure. First,
providing figures 1 , 5, and 6; and my
beyond their literal meaning. Petre and Green's phenomena are introduced, acknowledgements to P.U.F.
The important thing is to con­ and their account summarised. Then the next Publishers for permission to repro­
trol them systematically, so duce them here from Le Travail
section introduces some key concepts from lin­ Humain, 55.
that a graphic avoids unwant­
ed implicatures, and carries
guistic pragmatics, focussing on Grice's (1975)
the desired ones. Whether or theory of implicature. Section 4 indicates how
not the current account is use­ Grice's ideas have been applied in the graphical
ful has broader significance. If domain, by Marks and Reiter (1990). The final
it is useful, it may prove possi­ part of the paper shows how the ideas apply to
ble to predict and explain the
Petre and Green's observations, and relates the
properties of complex graphi­
cal representations, by borrow­ two accounts.
ing formal techniques from
natural language pragmatics.

2 Network diagrams

Petre and Green (1992) report on expert users'


use of graphical notations for computer-assisted
design in electronics (CAD-Ε). They inter­
viewed five engineers, all with over ten years'
experience in digital electronic design, and reg­
ular users of CAD tools. The engineers demon­
strated their tools and activities, and discussed

163

Information Design Journal 8:2 (1996), 163-179. DOI 10.1075/idj.8.2.05obe


ISSN 0142-5471 / Ε-ISSN 1569-979X © John Benjamins Publishing Company
Jon Oberlander ■ Grice for graphics

their practices. T h e key representation for such main points are reproduced nearly verbatim
users is the schematic drawing, an example of here:
which appears in figure I. Such drawings are Overviews The designers found graphics
largely graphical, and the CAD tools permit better (than text) for overviews (p.51).
multiple layering, allowing zooming in and out. Zooming Designers emphasize the need for
Figure 1 T h e drawings are also supported by textual improved detail on the overall gestalt-like views
A schematic annotations and appendices. obtained by changing scale (zooming) (p.54).
drawing.
Neighbourhood: adjacency and locality The
Reproduced
from Petre and 2.1 Petre and Green's observations nature of graphics allows connectedness to be
Green's figure Petre and Green (1992: 51-65) report seven represented by a line, and thus kept separate
1 (p.52). themes that emerged in their interviews. Their from relatedness (represented by adjacency).

Information design journal 8/2 (1996) 163-179

164
Jon Oberlander ■ Grice for graphics

Hence, adjacency or position is always, even information should be presented in a perceptual


locally, a secondary cue, a manipulable variable, form; redundant recoding should present
available for reinforcing associations, for sug­ important information in both textual and dia­
gesting structure, or for giving hints to the grammatic forms; restrictions should be placed
reader. Text, however, must either use adjacen­ upon users, so that only easily understood
cy to indicate both relationships, or introduce objects are presented; underlying mechanisms
symbolic links (e.g., cross-references or vari­ should be directly revealed, and respond to
ables) (p. 55). manipulation; revision of the diagram should
Shifts from graphics to text Text has its uses; be easy.
even after reasonable evolution, electronics Green uses the notion of cognitive dimen­
schematics remain a multi-level system with sions to characterise notations in the context
alternative representations in text (p.57). of their support environment, and their users'
Viscosity Massive changes to graphic repre­ tasks. Viscosity is a measure of resistance to
sentations are hard work: if a group of compo­ local change; hidden dependencies are frequent
nents is moved, connections have to be re­ when structural relationships are not made
established. Similarly, the cases where moving explicit in the notation; role-expressiveness
a piece of text is difficult are where connectivity measures the ease of seeing the function of sub­
has been established explicitly; in other cases, structures in the notation; premature commit­
blocks of text may be moved freely without dif­ ment occurs when users are forced into making
ficulty (p. 61). decisions too early in a process.
Search trails Searching is well supported in Petre and Green (1992: 50) observe that two
CAD-E in the domain of connectivity, but not other predictions flow from these models. First,
in the domain of functionality (p.63). that users need to escape from formalism, so
Vocabulary and space consumption Although that if a task requires information to be stored,
it is commonly believed that graphics is more and that information cannot be captured in the
compact than text, the comparison is not chosen notation, then users will switch to
straightforward, particularly because most another notation to achieve their goal. Second­
graphical notations rely to some extent on text ly, users will adopt secondary perceptual nota­
to keep their vocabularies manageable. Com­ tions to capture or emphasize certain informa­
plete system descriptions tend to be massive tion. That is, they will go beyond the 'official'
whether in graphics or text (p.64). interpretation of the diagrammatic symbols,
and - for instance - use spatial arrangement to
2.2 An explanation in terms of cognitive encode useful information. In fact, they later
dimensions concede (p66) that the Fitter and Green model
Their primary claim is that these observations does not really predict the use of secondary
are to be expected on the basis of Fitter and notation, because it focusses exclusively on the
Green's (1979) criteria for information accessi­ intended interpretation of the notation; and nor
bility; or from Green's (1989) cognitive dimen­ does the model deal with the escape from for­
sions, some of which refine the Fitter and Green malism, because it focusses on the properties
criteria. of single notations, rather than seeing them as
Fitter and Green state five criteria, whose sat­ forming elements in a set of tools. In addition,
isfaction contributes to information accessibility constraints on vocabulary size are not really
in a given diagrammatic notation: relevant accounted for by the model, and it does not

Information design journal 8/2 (1996) 163-179

165
Jon Oberlander ■ Grice for graphics

address differences among individuals' skill on language use. Gazdar (1979) proposed the
with the notations. influential equation 'Pragmatics = Meaning -
They argue that cognitive dimensions Truth Conditions' (p = Μ - τ C). On this basis,
provide a more useful framework, since it offers pragmatics is devoted to the study of all the
a mechanical model of the mental operations aspects of meaning which lie outside the mere
people go through as they use a notation, in a conditions of truth for sentences. For Gazdar,
rich context, embracing both tasks and tools. one of the most important aspects of an utter­
ance's meaning is the way it changes the context
2.3 An explanation in terms of Gricean in which it is uttered.
pragmatics Over time, differing pragmatic traditions have
Cognitive dimensions undoubtedly offer a use­ emerged; both Gazdar and Levinson (1983)
ful starting point for explanations; the alterna­ offer useful surveys of the varying schools
tive approach explored here is intended to of thought. For instance, followers of Austin
complement it, rather than replace it. T h e new (1962) emphasize that language use is a form of
approach, exploiting the notion of graphical action, and investigate the conceptual relations
implicature, makes for stronger explanations in among speech acts, their preconditions and
some areas than cognitive dimensions, but is effects (cf. for example Searle (1969)). By con­
correspondingly weaker elsewhere. trast, work in the Gricean tradition emphasizes
In the long run, graphical implicatures can the view that communication takes place by
be discovered and analysed in all kinds of every­ virtue of a set of assumptions about rational co­
day graphic objects, such as the London Under­ operation (Grice (1975,1978)) - p e o p l e provi­
ground diagram, weather maps, income sionally guess that other people will act in
support forms, or timetables. T h e next sections co-operative ways, and these guesses can be
provide some background on linguistic prag­ confirmed or upset as a conversation unfolds.
matics, and a description of previous work on A key Gricean idea is that a participant in a
graphical implicature, before returning to dis­ discourse will understand utterances directed
cuss how the Gricean approach can help explain at her in part because she recognises that the
Petre and Green's data. utterances were constructed with the intention
that she so recognise them. Griceans have thus
extensively investigated the linguistic role of
3 Pragmatics and the Gricean theory of defeasible (cancellable) inference, or implica­
implicature ture, particularly where this concerns mental
states.
According to Morris (1938), the study of sys­
tems of communication can be divided into 3.1 Co-operation and implicature
three parts. Syntax is devoted to 'the formal First, consider two concrete examples of what
relation of signs to one another'; semantics to Grice takes to be implicature, the dialogue in
'the relations of signs to the objects to which the (1), and the alternative texts in (2): 1
signs are applicable'; and pragmatics studies (1) a. A: Can you tell me the time?
'the relations of signs to interpreters'. Although b. B: Well, the milkman has come.
the notion of 'signs' can be read very widely,
most work in pragmatics has followed in the 1. The linguistic examples, (1) - (5), are all taken directly
footsteps of syntax and semantics, and focussed from Levinson (1983).

Information design journal 8/2 (1996) 163-179

166
Jon Oberlander· Grice for graphics

(2) a. T h e lone ranger jumped on his horse 2. do not make your contribution more infor­
and rode into the sunset. mative than is required.
b. T h e lone ranger rode into the sunset The maxim of Relation Make your contribu­
and jumped on his horse. tions relevant.
(Ia) has the implicature that A would like  to The maxim of Manner Be perspicuous, and
tell her the current time. (lb) carries the impli­ specifically:
cature that  does not know the precise time, 1. avoid obscurity;
but has some information which may permit A 2. avoid ambiguity;
to compute it. (2a)'s implicature is that the 3. be brief;
jump happened first, and was followed by rid­ 4. be orderly.
ing. By contrast, (2b)'s implicature is that the The subcomponents of a maxim - such as the
riding preceded the jumping. In both (1) and four parts of the maxim of manner - can be
(2), implicatures go beyond the literal truth termed submaxims.
conditional meaning. For instance, all that mat­ How does an individual Η in a conversation
ters for the truth of a complex sentence of the actually retrieve the implicatures q carried by an
form A and  is that both A and  be true: the utterance by a speaker S? Grice suggests that the
order of mention of the components is irrele­ following pattern of inference is typical. Η starts
vant. As Ρ =  -   suggests, implicatures help from the premise that S has said that p, and
to bridge the gap between truth conditions and notes that there's no reason to think S is not
'real' meaning. observing the maxims, or at least the Co-opera­
But how do implicatures actually function in tive Principle. H then reasons that in order for S
communication? Grice's account is deceptively to say that ρ and indeed be observing the max­
simple. His claim is that conversational partici­ ims (or the Co-operative Principle), S must
pants adhere to a Co-operative Principle. This think that q. S must also know that it is mutual
can be phrased as an injunction: make your knowledge that q must be supposed if S is to be
contribution such as is required, at the stage at taken to be co-operating. Now, S has done
which it occurs, by the accepted purpose or nothing to stop H, the addressee, thinking that
direction of the talk exchange in which you are q. So H concludes that S intends H to think that
engaged (Grice (1975: 26)). 2 In itself, this is q, and in saying that ρ has implicated q. There
very general, and Grice suggests four maxims, is, in principle, no reason to think that individu­
which can be thought of as guiding usual com­ als actually pass through this cycle of inference
munication: Quality, Quantity, Relation and every time they work out what a speaker means.
Manner. These are spelt out in Grice (1975: But, for the kind of implicatures Grice calls
26-7) as follows: conversational, the pattern can be reconstruct­
The maxim of Quality Try to make your con­ ed if necessary.
tribution one that is true, specifically:
1. do not say what you believe to be false; 3.2 Observing, violating and flouting
2. do not say that for which you lack adequate Conversation proceeds smoothly so long as
evidence. people observe the maxims. Things go wrong
The maxim of Quantity 1. make your contri­ when people violate them. However, Grice is
bution as informative as is required; well aware that people don't simply follow the
conversational maxims in a flat-footed way.
2. Page references to Grice are to versions in Grice (1989). People are perfectly capable of exploiting them

Information design journal 8/2 (1996) 163-179

167
Jon Oberlander- Grice for graphics

for effect: they can disobey, or flout a maxim or operating will have to find a reason why not.
set of maxims. Take the maxim of quantity, and One possibility is that (5b) holds, so that (by the
consider (3) and (4). maxim of quality), the speaker could not have
(3) a. Nigel has fourteen children. uttered a briefer description.
b. H e has exactly fourteen children.
(4) a. Either John will come, or he won't.
b. There's no point in worrying about it. 4 Implicatures in automatic
A straightforward utterance of (3 a) implicates network diagram design
(3b) : saying that someone has fourteen children
Marks and Reiter (1990) explored the possibili­
when they actually have fifteen would be true,
ty of applying the Gricean theory of implicature
but would violate the injunction to be as infor­
to graphical representations. In particular, they
mative as is required. By contrast, an utterance
addressed the automatic design of network dia­
(4a) appears to flout the maxim, since it is com­
grams, and developed their system, ANDD, to
pletely uninformative, and merely states a tau­
avoid unwanted conversational implicatures.
tology. Someone who hears (4a), and believes
Here, I summarise ANDD's task and structure,
that its utterer is still observing the Co-opera­
and then note the types of graphical implica­
tive Principle, will have to work out how voicing
tures which Marks and Reiter maintain should
this tautology could add to the conversation.
be avoided. The section concludes with a dis­
One solution is to infer that the speaker means
cussion of their arguments for three additional
that there is a fact of the matter, but that this is
conversational maxims.
inaccessible to the conversational participants;
hence (4b) is a natural implicature. 4.1 ANDD: an automated network diagram
T h e distinction between observing and flout­ designer
ing can also be illustrated by considering the ANDD's task is to design network diagrams
maxim of manner. Take (2) and (5): which accurately portray input network models.
(2) a. The lone ranger jumped on his horse T h e domain of the models could be, for
and rode into the sunset. instance, configurations of clients and servers
b. First he jumped, and then he rode. on a computer network; or they could be com­
(5) a. Miss S produced a series of sounds cor­ ponents in an electronic circuit. ANDD's net­
responding closely to the score of an aria from work diagrams have both syntax and semantics.
Rigoletto. T h e morphological elements are node symbols,
b. She wasn't singing in the conventional link symbols, text labels, and diacritical sym­
sense. bols. Each class of symbols has a set of graphical
An utterance of (2a) usually implicates (2b), properties: for instance, node symbols have
because speakers observe the injunction to be shape, size, pen colour, and fill colour. T h e syn­
orderly, and describe events in the same order tax also specifies a set of perceptual organisation
in which they happened. By contrast, the utter­ relations, comprising: sequential layout (top -
ance (5a) flouts the injunction to be brief. In bottom; left - right), proximity grouping, align­
interpreting the utterance, a hearer would nor­ ment, symmetry, similarity, and ordering.
mally conclude that the speaker is describing Semantically, an expressive mapping relates the
Miss S as singing. Since this much shorter network-model to the syntax of the network-
description was not used, the hearer who diagram: for instance, vertices and edges are
believes that the Co-operative Principle is still related to node and link symbols in figure 2.

Information design journal 8/2 (1996) 163-179

168
Jon Oberlander- Grice for graphics

ANDD generates a diagram by using three rule- Figure 2 A net


based systems to create an expressive mapping, work diagram
of a comput­
choose suitable values for graphical properties, er's disk sub­
and choose suitable locations for the symbols. system. This
T h e last stage is computationally hard, so ANDD diagram has
uses syntactic constraints together with existing no unwanted
local layout to heuristically generate a full lay­ implicatures;
the relation­
out.
ships between
In fact, Marks and Reiter's claim is that many input, disks
possible layouts are ruled out because they carry and output are
unwanted graphical implicatures, and that if the clear. Based
system can detect these, avoiding them will help on Marks and
further constrain the space of full layouts. Reiter's figure
l ( p . 456).
4.2 Avoiding unwanted implicatures
Assuming a particular network model, one can Figure 3
stipulate that the network diagram in figure 2 A variant of
figure 2, pos­
conveys precisely the intended information. sessing
Given this, Marks and Reiter maintain that the unwanted
network diagrams in figures 3 and 4 carry implicatures.
unwanted conversational implicatures. Based on
They suggest that figure 3 is misleading for at Marks and
Reiter's figure
least four reasons. First, the pen-width used for 2 (p. 456).
the channel-facility queue differs from pen-
width for all other queues. T h e implicature is
that this queue must be different in some
respect, since otherwise the same pen-width
would have been used. Secondly, disk symbols
are perceptually grouped into two gestalts. T h e
implicature is that there is some basis in the
Figure 4
domain for this grouping, since otherwise ele­
Another variant
ments would have been placed regularly. of figure 2,
Thirdly, device-queue symbols are laid out possessing dif­
irregularly. T h e implicature is that one of them ferent unwant­
is uniquely different, since otherwise elements ed implica­
tures. Based
would have been placed regularly. Finally, a dif­
on Marks and
ferent font is used for the channel-facility's text Reiter's figure
label. T h e implicature is that it must have dif­ 3(p.456).
ferent sub-system status, since otherwise the
same font would have been used.
Note that, just like the diagram in figure 2,
the one in figure 3 faithfully portrays all the fea­
tures of the assumed network model. T h e

Information design journal 8/2 (1996) 163-179

169
Jon Oberlander- Grice for graphics

difficulty with the second diagram is not that it sure to convey sub-system affiliation.
is incorrect, but that various aspects of it - not Returning to the faults in figure 3, they
specified in the semantics for the diagram - are observe that the use of spurious graphical-
misleading. All four of Marks and Reiter's prob­ property values may do their damage by violat­
lems arise because there is a general expectation ing the maxim of relation. What is wrong with
of regularity, and any departures from this are some of the departures from regularity, such as
assumed to be meaningful: the diagram con­ the distinct pen-width for the channel-facility
tains irregularities where the domain has none. queue, is that irrelevant properties are intro­
In one case, involving the grouping of disk sym­ duced. As Marks and Reiter observe, such
bols, an irregularity gives rise to a further regu­ problems may equally well be attributable to
larity, so the diagram also contains a regularity violations of the maxim of manner. Certainly,
where the domain has none. Figure 4 provides a other departures from regularity are more easily
further example of an 'over-regular' diagram. explained in this way: for example, the irregular
Its node symbols are ordered by size, and the layout of the device-queue symbols violates the
implicature is that there must be a similar injunction to be orderly.
ordering relation among the vertices in the net­
work model. Assuming that the network model 4.3 Elaborating the maxims
is as before, this implicature is again misleading. Marks and Reiter propose that the submaxims
So, it is claimed that figures 3 and 4 are defec­ of manner must be augmented with additional
tive for pragmatic, rather than semantic rea­ concepts to deal with the graphical domain. In
sons. That is: they are correct with respect to particular, a diagram generator must strive for
the expressive mapping, but give rise to spuri­ appropriate perceptual organisation, and
ous implicatures. But if a network diagram's observe human perceptual limits. Spurious per­
conversational implicatures are the culprit, it ceptual organisations that are orthogonal to the
must be possible to sketch a Gricean account diagram's intended information should be
which cites the conversational maxims which avoided. Graphical-property values should be
are being violated. limited in number, so that they are easily distin­
Marks and Reiter attempt to do just this. For guished, and perceptually dissimilar.
instance, they suggest that unwanted implica­ These are useful notions, although it is
tures arise if two graphical properties (such as possible to argue that they overlap to some
shape and colour) are used to communicate one extent with Grice's original submaxims of man­
network attribute, since the independence of ner. Spurious perceptual organisations take two
the properties will spuriously suggest that other forms: diagrams with too little regularity with
shape/colour combinations might exist. Such respect to the domain; and diagrams with too
redundancy runs counter to the maxim of quan­ much regularity. Grice's injunction to be order­
tity, since in some sense the diagram's contribu­ ly is violated by the first form. The additional
tion is more informative than is required. concept that is required is essentially a dual -
However, they also note that redundancy - or don't be too orderly. T h e orderliness submaxim
more generally, extra information - is actually could thus be broken in two, in the manner of
often required. For example, figure 2 contains the maxims of quality and quantity:
background context to allow the role of the Submaxim of Order
overloaded servers to be understood; it also uses 1. make your contribution as orderly as is
perceptual grouping as well as diacritical enclo­ required;

Information design journal 8/2 (1996) 163-179

170
Jon Oberlander- Grice for graphics

2. do not make your contribution more orderly implicature is responsible? As well as the con­
than is required. versational implicatures already discussed,
Perceptual limits are perhaps more contro­ there is a type of implicature which Grice calls
versial. Nonetheless, it might be argued that the conventional. These implicatures are inferences
injunction to observe perceptual limits is, in attached to individual words, and are not
fact, another way of expressing the injunction licensed directly by conversational maxims.
to avoid ambiguity. If I find it difficult to distin­ Some have seen the distinction as important,
guish two graphical-property values, a given and within the linguistic domain, Grice pro­
value may be regarded as ambiguous, denoting posed a number of tests to isolate conversation­
one or other network attribute. al implicature. T h e precise set has been the
Marks and Reiter also follow Hirschberg subject of argument, but Levinson's (1983:
(1991) in suggesting that the failure to use 119-20) six properties are a good summary; the
basic-level predicates (after Rosch (1978)) can first four are from Grice, the fifth from Sadock
be a further source of unwanted implicatures. (1978), and the last is Levinson's own:
In natural language, a predicate such as dog is Cancellability (or defeasibility) If extra informa­
basic-level, while animal is superordinate, and tion is added to the context, an implicature that
poodle is subordinate. Departures from basic- normally holds can be over-ridden.
level are obviously permitted, but always carry Non-detachability T h e precise surface form of
additional implicatures. Marks and Reiter sug­ the expression is irrelevant, so the implicature
gest that the preference for basic-level expres­ will go through if another expression with the
sions carries over to diagrams. For example, it is same meaning is substituted for it.
assumed that the basic-level, preferred status of Calculability T h e implicature can, in principle,
symbols in a computer display is static and non- by computed from first principles, via the Co­
blinking. So if a diagram designer chooses operative Principle and background knowledge.
blinking symbols for the whole display, the user Non-conventionality Implicatures are not part of
will interpret the choice of non-basic level the conventional meaning of expressions.
attributes as meaningful. Again, this is a useful Reinforceability The implicature can be made
suggestion, although Fodor (1983: 96) argues explicit without causing anomalous redundan­
that the use of basic-level expressions is in fact cy.
already dictated by Grice's maxim of quantity Universality If the implicature is generated via
and submaxim of brevity. Basic-level predicates an assumption of co-operation, all languages
provide a crucial layer in language, where brevi­ should share it.
ty and informativeness are most effectively None of these is entirely uncontroversial,
combined. So unprincipled choice of non-basic but they can all be defended against the most
symbols violates either the injunction to be obvious criticisms. For current purposes, three
brief, or the injunction to be informative. points should be noted. First, both defeasibility
and non-conventionality help to distinguish
4.4 Tests for implicatures implicatures from truth-conditional meaning.
Marks and Reiter's discussion provides useful Secondly, there is a trade-off between calcula­
examples of graphical implicatures. But is there bility and conventionality: something which is
a general way of telling when part of a diagram's calculable will usually be non-conventional;
meaning has been generated by an implicature? something which is conventional (such as the
And is there a way of telling which kind of contrastive sense to but) will usually not be cal-

Information design journal 8/2 (1996) 163-179

171
Jon Oberlander ■ Grice for graphics

culable. Finally, the principle of non-detacha- it could be argued that some other graphical
bility does not govern implicatures generated by implicatures are actually conventional (and not
the maxim of manner, since this is clearly con­ calculable). For example, some complex graph­
cerned with the surface form of conversational ical entities form patterns which are highly
contributions. recognisable; the symbols associated with the
So, are the graphical implicatures discussed bi-stable flip-flop discussed in section 5.1 are
by Marks and Reiter genuinely conversational of this kind. Practised users of a notation will
in nature? interpret such symbols 'directly', without hav­
They are certainly defeasible in general: tex­ ing to build up their interpretation from the
tual annotation can generally over-ride any meanings of their parts. These complex graphi­
graphical implicature. For instance, the impli­ cal entities thus resemble idiomatic phrases in
catures associated with font choice can be natural language. However, although their
defeated by adding text to the diagram's legend, meaning can be read off directly, it need not be.
such as Choice of fonts is not meaningful. In net­ Thus, their interpretations, and associated
work diagrams, relative size of symbols does not implicatures, are still calculable.
carry implicatures concerning the size of The re-inforceability of graphical implica­
domain objects, but in other domains, a fre­ tures seems relatively uncontroversial; indeed,
quent textual annotation is Not to scale, which some of Marks and Reiter's problems with
cancels the relevant implicatures. Graphical redundancy may be traced to reinforcement.
implicatures can also be over-ridden directly by When the diagram in figure 2 uses perceptual
graphical means. In network diagrams, the sub- grouping as well as diacritical enclosure to con­
maxim of order leads to related components vey sub-system affiliation, the former works via
being grouped together, and unrelated compo­ implicature, but is reinforced via the semanti-
nents being separated. However, sometimes cally-specified enclosure symbolism.
related components must be separated for inci­ Universality may seem more difficult: it might
dental reasons. In these cases, the explicit lines be argued that the extraordinary range of differ­
connecting them defeat implicatures concern­ ent graphical notations means that there are few
ing unrelatedness. T h e defeasibility of graphical 'universal' implicatures, carried by a symbol (or
implicatures helps to show that they are conver­ ordering of symbols) wherever they appear. But
sational, and not conventional, since the latter this is not necessary. What matters is that if a
cannot be over-ridden. configuration has an implicature in one graphi­
However, like conventional implicatures in cal notation, the implicature should be retained
language, many graphical implicatures are by the configuration's translation in any other
detachable; this is, however, because they are notation of equivalent expressiveness. T h e basic
traceable to the maxim of manner and its sub- point is just that if the implicature arises from
maxims. Since these concern the surface form principles of co-operation, it should be repro­
of representations, variations in the form can ducible in any equivalent notation used co­
cause implicatures to vanish. T h e calculability operatively. The universality property thus
of the implicatures has already been demon­ helps distinguish conventional implicatures
strated, in the sketch of the reasoning behind from conversational. The real problem with
the implicatures relating to regularity and over- using universality as a test will reside in the
regularity. In theory, then, the non-convention­ difficulty of establishing which notational sys­
ality of these implicatures is assured. However, tems are genuinely equivalent.

Information design journal 8/2 (1996) 163-179

172
Jon Oberlander ■ Grice for graphics

5 Implicatures in CAD-E 5.1 Design errors and unwanted implicatures


Looking at the schematic drawings, some
These concepts can be used to re-examine Petre aspects of the network diagram are seen to be
and Green's data on computer-assisted design fixed by the expressive mapping, but others are
Figure 5
A competent in electronics. This section discusses a simple not. Only the former are part of what Petre and
electronic example in terms of graphical implicatures, and Green term 'the official notation' (p.67). So
schematic then assesses the extent to which an implicature there is a clear distinction between the seman­
drawing. account can explain the themes which arose tics and pragmatics of diagrams - between the
Reproduced literal and the 'full' meaning of the graphics.
from Petre and
from Petre and Green's interviews. Finally, the
implicature account is compared with Petre Compare the diagram in figure 1 (repeated here
Green's figure
Kp.52). and Green's proposal. as figure 5) with that in figure 6.

Information design journal 8/2 (1996) 163-179

173
Jon Oberlander- Grice for graphics

According to Petre and Green, the two


machines depicted are equivalent, but the one
in figure 6 fails to use what they term 'manipu-
lable cues'. T h e diagram suffers from 'neglect of Figure 6 An incompetent electronic schematic drawing,
adjacency cues: things are grouped visually representing the same circuit as in figure 5. The circles
which are not related logically' (p.57); compo­ enclose significant mis-cues:
nents are dispersed which ought to be related (1) symmetry misleadingly relates unrelated components;
(for instance, u8 and UIO); the signal flow is (2) extra bends in a wire have been left by past edits -
there is also convoluted wiring at the left of the diagram;
badly confused, via convoluted wiring; and two
(3) input 1 to the gate has redundant entry routes, and the
gates have been rendered to look like a bi-stable angle of entry is odd; and
flip-flop, which they label 'a striking mis-use of (4) the gates misleadingly resemble a bi-stable flip-flop.
a strong perceptual cue' (p. 61). Reproduced from Petre and Green's figure 4 (p.56).

Information design journal 8/2 (1996) 163-179

174
Jon Oberlander ■ Grice for graphics

In current terms, the diagrams denote the vocabulary and discriminability. This issue is
same network model, and use the same expres­ revisited towards the end of the next subsection.
sive mapping. Where they differ is in their prag­
matic status. The connections between the 5.2 Petre and Green's observations
components are basically the same in the two So, the implicature-based account can help
diagrams, but the use of space is very different: characterise specific design errors. But can it
the connections are specified in the network say anything about the emergence of the seven
model, but the actual layout is a pragmatic mat­ themes which characterise Petre and Green's
ter. The diagram in figure 5 observes the interviews with AD-E users? Consider each
Gricean maxims, whereas that in figure 6 in turn.
violates them. Overviews 'The designers found graphics better
T h e specific faults in figure 6 can be traced to for overviews.' A circuit graphic collects togeth­
the violation of particular conversational max­ er many components in a single, synoptic repre­
ims. Since they are largely layout problems, sentation. By contrast, to convey the same
they are best thought of as violations of the information through text requires the assembly
maxim of manner. When things are grouped of multiple lists and representations. The
visually which are functionally or logically dis­ unified representation therefore juxtaposes
tinct, the diagram violates the second part of the many more components, and hence the scope
new submaxim of order. It does so because the for graphical implicatures is greater than the
diagram immediately juxtaposes various sym­ scope for textual implicatures in the alternative
bols, and thereby generates the unwanted representations. It follows that the schematic
implicature that the symbols' denotations are in drawing is a richer medium for pragmatically
some way related. When related components encoding information, such as functional relat-
are dispersed, the diagram violates the first part edness. This may well be one reason why graph­
of the submaxim of order. By failing to juxta­ ics often seem to say more than texts, even
pose symbols which should be related, the dia­ where these are known to be equivalent at the
gram fails to be as orderly as it should. Convo­ semantic level.
luted wiring may be thought of as a violation of Zooming 'Designers emphasize the need for
the submaxim of brevity, since connections are improved detail... by changing scale.' A synop­
made longer and more complex than is tic representation restricts the granularity of
justifiable. 3 The fake bi-stable flip-flop can be representation. Zooming in on details via in-
thought of as violating the submaxim of ambi­ place expansion is particularly good because it
guity, because the gates' rendering leads natu­ retains graphical context, and therefore pre­
rally to an additional, unintended interpreta­ serves the background's implicatures, which
tion. Since this interpretation is also incorrect, specify implicit relationships.
the diagram may even appear to violate the Neighbourhood: adjacency and locality In
maxim of quality - although it is strictly correct. schematic drawings, 'adjacency or position is
T h e bi-stable flip-flop is particularly interesting, always, even locally, a secondary cue.' T h e
because it also raises the issue of diagram drawings separate out connectivity from relat-
edness, whereas text conflates them, or has to
introduce symbolic linking. In fact, the main
3. Alternately, it may be maintained that unnecessary loops
and angles are irrelevant contributions, and hence violate difference resides in the dimensionality of the
the maxim of relation. modality: diagrams use two dimensions, and

Information design journal 8/2 (1996) 163-179

175
Jon Oberlander ■ Grice for graphics

unformatted text uses only one. But both dia­ guity'; moreover, a single physical component
grams and texts carry implicatures relating to can have multiple graphical representations,
order. The 'perceptual grouping' conventions giving differing views of its function. So viola­
governing the layout of diagram elements in tions of either brevity or informativeness are
figure 5 play a similar role to the descriptive also a danger; as with natural language, any
conventions governing the ordering of clauses choice of a non-standard representation will
in example (2). Textual ordering generalises to appear meaningful, as a departure from the
diagrammatic juxtaposition; but in both cases, expected, basic-level choice. And the fake bi­
extra information - beyond the semantic - is stable flip-flop from figure 6 shows that a con­
carried by pragmatic means. structed representation can accidentally
Shifts from graphics to text 'Text has its uses.' It is resemble a basic-level entity. 4 Petre and Green
the general purpose tool of last resort. As Petre emphasize that, as with adjacency, correct
and Green observe, even the best-evolved nota­ choice of representations is an acquired skill,
tion will sometimes fail to express a given and that individual differences along this skill
proposition; there must always be an escape dimension tend to distinguish novice designers
from formalism, and the place of refuge usually from experts. This situation is very similar to
lies in natural language. It has already been that confronting second language learners:
noted that the cancellation of graphical implica­ pragmatic aspects of a language are often much
tures can often be achieved via textual annota­ harder to master than its semantics.
tion. It is plausible to suggest that such Vocabulary and space consumption 'Most graphi­
cancellation is one of the major reasons why cal notations rely to some extent on text to keep
shifts from graphics to text should be necessary. their vocabularies manageable.' The problem of
Another relates to vocabulary size, considered graphical ambiguity just noted is solved by
below. using the additional modality of text: 50 to 65
Viscosity 'Massive changes to graphic represen­ component symbols are augmented with textual
tations are hard work ... moving a piece of text annotation to represent 100 to 500 physical part
is difficult... where connectivity has been estab­ types. But why are graphical vocabularies kept
lished explicitly.' This suggests that, at least in so small? T h e answer lies in the domain of per­
some cases, it is preferable to encode informa­ ceptual discriminability. It might not be impos­
tion pragmatically, rather than semantically, sible to design 500 discriminable symbols, but
where there is a choice. If relationships can be uncertainties over symbol interpretation would
left implicit in the graphic, then less work has to become more likely. It is better to say some­
be done when elements are shifted around. This thing general than to say something more pre­
could be one reason why the diacritical enclo­ cise, if the latter is destined to be misunder­
sures for sub-systems considered by Marks and stood. T h e over-generality of the symbols is
Reiter are not used in the real world schematics manageable because it is known that their full
studied by Petre and Green. interpretation depends on the division of labour
Search trails 'Searching is well supported in between graphics and language. In this way,
CAD-Ε in the domain of connectivity, but not in intended imprecision prevents unintended
the domain of functionality.' Petre and Green ambiguity.
observe that symbol shape conveys a compo­ 4. This may occur because basic-level entities can actually
nent's function, but not the values it operates be graphically complex, resembling idiomatic phrases in
over. Such abstraction inevitably leads to 'ambi­ natural language, rather than unanalysable words.

Information design journal 8/2 (1996) 163-179

176
Jon Oberlander- Grice for graphics

5.3 Comparison with cognitive dimensions effectively than others. Cognitive dimensions
Petre and Green (1992: 50) initially argue that offer no direct explanation for limits on the
the themes which emerged in their interviews graphical vocabulary; however, given the limits,
are predictable from Fitter and Green's (1979) cognitive dimensions do predict the necessity of
criteria, or from Green's (1989) cognitive a division of labour - some work carried out
dimensions framework. They ultimately suggest beyond the constraints of the graphical formal­
(pp. 65-7) that four aspects of the data are most ism. Similarly, various trade-offs can be predict­
important: ed: premature commitment is acceptable if the
Information access Notations and their support overall system has low viscosity; or poor role-
tools should make the information conveyed as expressiveness in the main notation is permitted
obvious as possible. if support tools can reveal the roles using alter­
Escape from formalism Where tasks require infor­ nate notations.
mation to be encoded that cannot be fitted into How, then, does the Gricean account com­
the main notational scheme, other mechanisms pare? Information access is at the core of the
must be provided. approach, and hence Fitter and Green's is the
Secondary notation Users will exploit various closer model. Levinson (1983: 121) noted that
techniques - such as informative layout - to go pragmatics should cover non-linguistic commu­
beyond the semantics of the official notation. nication: 'If the maxims are derivable from con­
Individual differences Experts and novices differ siderations of rational co-operation, we should
particularly in their ability to use secondary expect them to be universal in application, at
notation effectively. least in co-operative kinds of interaction.'
Fitter and Green's approach most squarely Indeed, it seems that network diagram design­
addresses the themes related to information ers do observe the Co-operative Principle.
access. There, it goes some way towards Their contributions are tailored to the purposes
explaining why, for instance, graphics are good of their communications, and can be thought of
for overviews. However, because it is concerned as being carried out in accordance with Grice's
with the usability and semantics of individual maxims of conversation. Errors which frustrate
notations, the Fitter-Green approach does not information access can be seen as arising from
touch on either the escape from formalism - violations of the maxims, which generate
possible only with multiple notations - or the unwanted implicatures. T h e Gricean account
pervasive use of secondary notation - possible has greater reach than Fitter and Green, howev­
only by going beyond the basic semantics of the er. First, the escape from formalism is predicted
notation. Nor were individual skill differences a to flow from the need to cancel unwanted impli­
matter of concern. catures of various kinds. Secondly, secondary
Green's cognitive dimensions, however, have notation is the very stuff of graphical pragmatics
a broader applicability. Because notations are - meaningful structures which go beyond the
seen in the context of task and support environ­ plain semantics of the system. Finally, individu­
ment, the need to support the escape from for­ al differences arise exactly because the mastery
malism is explained. Secondary notation is of pragmatic implicatures is one of the hardest
useful because it reveals dependencies and things to achieve with a new representational
encourages role-expressiveness. But because system.
secondary notation is not automatically sup­ In this sense, the Gricean account goes
ported, some individuals will exploit it more beyond Fitter and Green, and covers territory

Information design journal 8/2 (1996) 163-179

177
Jon Oberlander ■ Grice for graphics

that Petre and Green argue falls within the 6 Conclusions


domain of cognitive dimensions. There is no
question, however, of its replacing cognitive It was argued that there are certain parallels
dimensions. The latter obviously has wider between pragmatic phenomena in natural lan­
scope, taking into account tasks and support guage and graphical representation. Elsewhere,
environments; unlike the implicature story, it formal techniques have been developed for
can support claims concerning trade-offs modelling some of the natural language phe­
between various system properties. For this rea­ nomena. Thus, it is to be hoped that once the
son, it is more likely to provide a useful guide to graphical data are better-understood, it may be
the design of human-computer interaction. possible to treat them with formal techniques
However, it can be argued that the Gricean from natural language pragmatics.
account complements cognitive dimensions, It is highly unlikely that a single set of princi­
and is at least compatible with it. This paper has ples will be found which apply to all graphical
concentrated on pragmatic aspects of the net­ objects. As Reiter (1995) has recently argued,
work notation, but pragmatic studies in general there is a strong 'sublanguage' component to
have always emphasized the need to look at graphical implicature: some rules apply only in
notations in their context of use, and the crucial certain genres. For example, salience rules, like
role of the notations' users. All discussion of the most important object is the one in the middle,
pragmatic implicatures must ultimately be are strongly genre-dependent. Grouping rules
grounded in models of human inference, which may be somewhat more general, but even here
show how people combine various goals and it seems that using horizontal/vertical alignment
beliefs in context, to derive interpretations and to show grouping is acceptable in some genres
actions. From this perspective, the Gricean but not in others. So even without universal
account has many assumptions in common with principles, the tools of pragmatics are already
cognitive dimensions. There can and should be beginning to help understand the many and var­
constructive interaction between the two ied conventions of graphical design.
approaches. At this stage, some might baulk at the idea of
examining graphics from a perspective which
was pioneered in linguistics, precisely because
they attach value to real data, rather than to the
toy examples so beloved of linguists. However,
Levinson (1983: 25) rightly observed that 'For
many linguists, one of the major contributions
of pragmatics has been to direct attention once
again to actual language usage.' The discussion
here, of Petre and Green's examples and obser­
vations, is intended to be only an initial step
towards the pragmatic study of actual graphical
usage.

Information design journal 8/2 (1996) 163-179

178
Jon Oberlander- Grice for graphics

References
Austin, J. L. (1962) How to do things with words. Oxford:
Clarendon Press
Fitter, M. and Green, T. R. G. (1979) When do diagrams
make good computer languages? International Journal of
Man-Machine Studies, 1 1 : 2 3 5 - 2 6 1
Fodor, J. A. (1983) Modularity of mind. Cambridge, MA: MIT
Press
Gazdar, G. (1979) Pragmatics: Implicature, presupposition
and logical form. New York: Academic Press
Green, T. R. G. (1989) Cognitive dimensions of notations. In
A. Sutcliffe and L. Macaulay (eds.) People and computers
5. Cambridge: Cambridge University Press
Grice, H. P. (1975) Logic and conversation. In P. Cole and
J. L. Morgan (eds.) Syntax and semantics 3: Speech acts,
pp 4 1 - 5 8 . New York: Academic Press. (Reprinted as
Essay 2 in Grice 1989)
Grice, H. P. (1978) Further notes on logic and conversation.
In P. Cole (ed.) Syntax and semantics 9: Pragmatics,
pp 113-128. New York: Academic Press. (Reprinted as
Essay 3 in Grice 1989)
Grice, H. P. (1989) Studies in the way of words. Cambridge,
MA: Harvard University Press
Hirschberg, J. (1991) A theory of scalar implicature. New
York: Garland Publishing (Published version of 1985 the­
sis, University of Pennsylvania)
Levinson, S. C. (1983) Pragmatics. Cambridge: Cambridge
University Press
Marks, J. and Reiter, E. (1990) Avoiding unwanted conversa­
tional implicatures in text and graphics. In Proceedings of
the Eighth National Conference on Artificial Intelligence,
pp 450-456. Boston, 29 July-3 August, 1990
Morris,  W. (1938) Foundations of the theory of signs. In
0. Neurath, R. Carnap and  Morris (eds.) international
encyclopedia of unified science, pp 77-138. Chicago:
Chicago University Press
Petre, M. and Green, T. R. G. (1992) Requirements of graph­
ical notations for professional users: electronics CAD sys­
tems as a case study. Le travail humain, 55: 4 7 - 7 0
Reiter, E. (1995) Sublanguages in text and graphics. Paper
presented at the first International Workshop on
Intelligence and Multimodality in Multimedia Interfaces,
Edinburgh, July, 1995
Rosch, E. (1978) Principles of categorization. In E. Rosch
and B. Lloyd (eds.) Cognition and categorization,
pp 27-48. Hillsdale, New Jersey: Lawrence Erlbaum
Associates
Sadock, J. M. (1978) On testing for conversational implica­
ture. In P. Cole (ed.) Syntax and semantics 9: Pragmatics,
pp 281-298. New York: Academic Press
Searle, J. (1969) Speech acts: An essay in the philosophy of
language. Cambridge: Cambridge University Press

Information design journal 8/2 (1996) 163-179

179