Reference and Relevance
Reference and Relevance
DEIRDRE WILSON
Abstract
In this paper, I compare three accounts of reference assignment which have the
same overall structure but differ in the pragmatic criteria they propose for
evaluating alternative interpretations. I argue that the criteria proposed by
Lewis (1979) and Clark (1977) are inadequate to deal with a certain range of
examples which are more insightfully handled by the relevance-theoretic
criterion proposed by Sperber & Wilson (1986).
1 Introduction
The assignment of reference to definite noun-phrases such as "the man’, ‘the
door’, has been approached from several perspectives: semantic, pragmatic and
psycholinguistic. In this paper, I will take a pragmatic perspective, considering
reference assignment as part of an overall process of utterance interpretation
whose goal is to recognise what propositions and propositional attitudes the
speaker intended to convey. My theoretical framework will be broadly
Gricean: I will assume without argument that utterance interpretation involves
a substantial element of non-demonstrative inference, and that the inference
process consists in a search, among candidate interpretations, for one that
satisfies a pragmatic criterion which it is the task of pragmatic theory to
describe. Grice's proposed criterion was based, of course, on his Co-Operative
Principle and associated maxims of truthfulness, informativeness, relevance
and clarity. In a number of recent works, Dan Sperber and I have argued for
an alternative criterion of consistency with a principle of optimal relevance
(see for example Sperber & Wilson 1986, 1987). The aim of this paper is to
consider a range of broadly Gricean approaches to reference assignment,
examine their proposed criteria for evaluating interpretations, and show that
the relevance-theoretic criterion is preferable on both empirical and theoretical
grounds.
168 Deirdre Wilson
2 The problem
The pragmatic literature on reference assignment has concentrated on two
main problems:
(a) what happens when there is more than one previously established
candidate referent and the hearer has to make a choice?
(b) what happens when there is no previously mentioned candidate and an
appropriate referent has to be inferred?
These problems are illustrated in (1) and (2):
(1) a. The room had three doors, one of which was open. I closed the
door.
b. Sean Penn attacked a photographer. The man was quite badly hurt.
(2) a. 1 walked into the room, Both windows were open.
b. Jane dropped linguistics. The lectures were boring.
In (la), though the NP ‘the door’ could in principle refer to any of the three
doors previously mentioned, the natural assumption is that it was intended to
refer to the open one; in (1b), though the NP ‘the man’ could refer to either
Sean Penn or the photographer, the natural assumption is that it was intended
to refer to the photographer. The question raised by utterances like (1) is how
the choice between the two explicitly mentioned candidates is made.
In (2) the problem is slightly different. Instead of having several
candidate referents to choose from, the hearer has to go through an inference
process to establish even one. In (2a), for example, it is nowhere explicitly
stated that the room the speaker went into had two windows: in order to assign
reference to the definite NP 'both windows’, however, the hearer has to
assume that the room had two windows and that these were the open ones.
Similarly, in assigning reference to the definite NP ‘the lectures’ in (2b), the
hearer has to assume that studying linguistics involves going to lectures and
that the linguistics lectures were the boring ones.
Examples of this type have been investigated by Herb Clark and his
associates under the name ‘bridging’. Intuitively, bridging involves the
retrieval or construction of a contextual assumption introducing a candidate
referent, which is then used to assign reference to a subsequent definite NP.
I will call the referent-introducing assumption the bridging assumption, or
bridge. Thus, for (2a) the bridging assumption would be (3a), and retrieval of
this assumption would make (2a) essentially equivalent to (3b):
Reference and Relevance 169
(3) a. The room had two windows.
b. I walked into the room. The room had two windows. Both
windows were open.
I will also be looking at a third set of examples, which has been
relatively little studied, and which presents the hearer with a combination of
the problems raised by (1) and (2): that is, it involves cases of bridging where
there is not merely one possible bridging assumption but two. Consider (4a)
and (4b):
(4) a. 1 switched from linguistics to geography. The lectures were less
boring.
b. Tran from the classroom into the playground. The children were
making too much noise,
In (4a) there are two potential bridges, (5a) and (5b), yielding the
interpretations in (6a) and (6b), respectively:
(5) a. Studying linguistics involves going to lectures.
b. Studying geography involves going to lectures.
(6) a. The linguistics lectures were less boring than the geography ones.
b. The geography lectures were less boring than the linguistics ones.
Notice that whichever interpretation is chosen, both bridges will be used; the
issue is, which of them will be used to assign reference to the NP "the
lectures’ in (4a)? 1 take it that in normal circumstances the preferred
interpretation would be the one based on (5b) and (6b).
In (4b) there are again two potential bridges and two potential
interpretations, as in (7) and (8):
(7) a. There were children in the classroom.
b. There were children in the playground.
(8) a. The children in the classroom were making too much noise.
b. The children in the playground were making too much noise.
In this case, unlike the previous one, there is no need to use both bridges. I
take it that in this case, without further information, either interpretation would
be possible, though there may be a preference for the one based on (7b) and
(8b).
170 Deirdre Wilson
The general problem that interests me, then, is this: given a definite
description such as ‘the man’, ‘the door’, how does the hearer choose the
intended referent from a range of candidate referents whose existence may not
have been explicitly stipulated, and may therefore have to be inferred?
3 Some previous accounts
Examples such as (1) and (2) refute one traditional semantic analysis of
definite descriptions, on which a definite NP such as "the door” is taken to
refer to the unique object in the universe of discourse which meets the
description of being a door. In fact, the inadequacy of this analysis was
already recognised in David Lewis's "General Semantics’ (1970/83), a classic
sketch of a formal semantics for English. In the appendix to that paper, Lewis
wrote:
‘consider the sentence ‘The door is open’. This does not mean
that the one and only door that now exists is open; nor does it
mean that the one and only door near the place of utterance,
or pointed at, or mentioned in previous discourse, is open.
Rather, it means that the one and only door among the objects
that are somehow prominent on the occasion is open.’
(Lewis 1970/83: 227)
With the introduction of the notion of prominence, cognitive factors explicitly
entered the domain of semantics. An object might be made prominent, said
Lewis, by being nearby, or pointed at, or mentioned, but the crucial
consideration was more general than these. Contextual prominence, he said,
was determined by "mental factors such as the speaker's expectations
regarding the things he is likely to bring to the attention of his audience’
(Lewis 1970/83: 227-8). Thus the role of cognitive factors in reference
assignment was explicitly acknowledged.
This insight was developed in Lewis's paper 'Scorekeeping in a
language game’ (1979/83). Here, the earlier notion of prominence was
replaced by a notion of salience, where an object could become salient, for
example, by being easily noticeable, or by being mentioned in discourse.
Lewis then proposed a general rule for assigning reference to definite
descriptions: a definite description refers to the most salient object in the
domain of discourse meeting that description. Consider (9), for example;
Reference and Relevance 171
(9) The dog got in a fight with another dog.
What this means, says Lewis, is that the most salient dog got in a fight with
some less salient dog.
Finally, Lewis invoked a notion of pragmatic acceptability, and argued
that if the rule of reference assignment yielded pragmatically unacceptable
results, a reinterpretation must take place. To motivate this addition, he used
the following scenario. At a certain point in a conversation, there are two
salient cats: mine, which is with us in the room, and yours, which you've left
at home but have just been telling me about. At a point when your cat is the
most salient one, mine gets under your feet, and I say:
(10) Watch out, you're about to tread on the cat.
Question: which cat should you take me to have meant? Although your cat is
the most salient one, it can’t be the one I meant to refer to, says Lewis,
because on that interpretation my utterance will be pragmatically unacceptable.
Accordingly, he proposed a ‘rule of accommodation for comparative salience’,
which runs as follows:
Rule of accommodation for comparative salience
If at time t something is said that requires, if it is to be
acceptable, that x be more salient than y; and if, just before t,
Xx is no more salient than y; then - ceteris paribus and within
certain limits - at t, x becomes more salient than vy,
In other words, for pragmatic reasons, the relative salience of objects can be
shifted so that the output of the reference assignment rule is pragmatically
acceptable,
Lewis's remarks on the notion of pragmatic acceptability are extremely
brief. He merely mentions three contributory factors. First, truth: an utterance
is pragmatically unacceptable if it is untrue; second, non-triviality, by which
he means, essentially, informativeness: an utterance is pragmatically
unacceptable if it is uninformative; and third, what he calls warranted
assertion, by which he means, essentially, that an utterance is pragmatically
unacceptable if it is unevidenced. These factors have obvious similarities to
Gricean maxims, though no maxim of relevance is mentioned. His claim is,
then, that reference is assigned to referring expressions depending, on the one
hand, on the comparative salience of candidate referents, and on the other
172 Deirdre Wilson
hand, on the desire to obtain an overall interpretation that satisfies a pragmatic
criterion requiring an utterance to be true, informative and evidenced.
Herb Clark’s analysis of the bridging inferences illustrated in (2) and
(4) above has the same overall structure. In his best-known paper, "Bridging”
(1977), Clark discusses examples such as (11):
(11) Tlooked into the room. The ceiling was high.
Taking a resolutely cognitive approach, he argues that reference assignment
involves not so much the location of an appropriate object as the retrieval of
an appropriate mental representation from memory. Sometimes, as in (11), an
appropriate representation will have to be constructed ad hoc, and it is this ad
hoc construction process that he calls bridging.
Which bridge should the hearer construct? Why is it that in (11), for
example, the appropriate bridge is the one that links the NP "the ceiling’ to the
ceiling in the room the speaker looked into, rather than, say, the ceiling in the
speaker's bedroom at home? Clark's answer is that the appropriate bridge will
be the shortest plausible one, i.e. the shortest bridge which contains only
assumptions that in the circumstances are likely to be true.
To get an intuitive understanding of how bridges may vary in length,
consider the following examples:
(12) a. John and Mary went out last night. The three witches were
excellent.
b. John and Mary went to the theatre last night. The three witches
were excellent.
c. John and Mary went to a play last night. The three witches were
excellent.
д. John and Mary went to Macbeth last night. The three witches were
excellent.
The examples are given in increasing order of acceptability, with (12a) being
the least acceptable and (12d) the most. These differences in acceptability
might be explained in terms of the length of the bridge needed to introduce an
appropriate referent for the NP 'the three witches’. In (12a), for example, in
accessing an appropriate referent, the hearer has to construct or retrieve the
following assumptions:
Reference and Relevance 173
(13) a. John and Mary went to the theatre.
b. At the theatre they saw a play.
¢. The play they saw was Macbeth.
d. Macbeth has parts for three witches.
In (12d), by contrast, he has to retrieve only the information in (13d). Notice
that, like Lewis's notion of salience, Clark’s notion of length of the bridge is
a factor affecting the accessibility of referents: the shorter the bridge, the more
accessible the referent, and in this case the more acceptable the overall result.
On pragmatic acceptability, Clark is no less vague than Lewis. He
claims, first, that the bridge must be short, and, second, that the bridge must
be plausible: it is the one, he says, "that requires the fewest assumptions, yet
whose assumptions are all plausible given the listener's knowledge of the
speaker, the situation, and facts about the world" (Clark 1977: 420). Finally,
the resulting interpretation must satisfy a pragmatic criterion called the
Given-New Contract, which is introduced as an extension of Grice's
Co-Operative Principle (see Clark & Haviland 1977), and applies to all
utterances containing referring expressions. To satisfy the Given-New
Contract, an utterance must meet two conditions: (a) the intended referent must
be uniquely identifiable from memory, and (b) the utterance must make some
informative assertion about that referent. In other words, for a bridging
utterance such as (2), (4) or (11) to be pragmatically acceptable, the bridge
must be short, the bridge must be plausible, the referent must be uniquely
identifiable, and the resulting overall interpretation must be informative.
Clearly, the proposals of Clark and Lewis have a common structure, and
clearly this structure is essentially correct. Reference assignment is affected,
on the one hand, by the accessibility of potential referents, and on the other
hand, by the pragmatic acceptability of the resulting overall interpretation.
What is curious, though, is how cursory the analysis of pragmatic acceptability
is on these accounts. Neither Clark nor Lewis makes any reference 10
relevance;' both appeal 10 commonsense notions of truth, degree of evidence,
and informativeness. It might seem, indeed, that though commonsensical, the
accounts are so vague as to be unfalsifiable. I will argue that, vague though
they are, they make definite predictions, and that these predictions are false.
Before doing so, however, | want to sketch a relevance-theoretic account of
reference assignment and to show that it has much the same structure as those
Clark & Haviland (1977) introduce the Given-New Contract as an extension of Grice’s
Co-Operative Principle. Clark may therefore intend to invoke the Co-Operative Principle and
maxims as additional constraints on interpretation - a possibility that is considered below.
174 Deirdre Wilson
of Clark and Lewis, differing only in the criterion of pragmatic acceptability
proposed.
4 Relevance theory
Relevance theory, unlike Grice’s pragmatic theory, is not a rule-based or
maxim-based system. It is based on a few very simple ideas. First, that every
utterance has a variety of possible interpretations, all compatible with the
information that is linguistically encoded. Second, that not all these
interpretations are equally accessible on a given occasion. Third, that hearers
are equipped with a single, very general criterion for evaluating interpretations
as they occur to them. And fourth, that this criterion is strong enough to
exclude all but at most a single possible interpretation, so that the hearer is
entitled to assume that the first interpretation that satisfies the criterion is the
only one.
The criterion that Dan Sperber and I proposed in our book Relevance
develops from a basic assumption about human cognition. The assumption is
that human cognition is relevance-oriented: we pay attention to information
that seems relevant to us. Now every act of communication starts out as a
request for attention. As a consequence, it creates an expectation of relevance.
It 1s around that expectation of relevance that our criterion is built.
Relevance is defined in terms of contextual effect and processing effort,
Contextual effects are achieved when newly presented information interacts
with a context of existing assumptions in one of three ways: by strengthening
an existing assumption, by contradicting and eliminating an existing
assumption, or by combining with an existing assumption to yield a contextual
implication: that is, a logical implication derivable neither from the new
information alone, nor from the context alone, but from the new information
and the context combined. We claim that newly-presented information is
rclevant in a context when and only when it achieves contextual effects in that
context, and the greater the contextual effects, the greater the relevance.
Contextual effects, however, do not come free: they cost some mental
effort to derive. The effort needed to compute the contextual effects of an
utterance depends on three main factors: first, the linguistic complexity of the
utterance; second, the accessibility of the context; and third, the inferential
effort needed to compute the contextual effects of the utterance in the chosen
context. Each increment in processing effort detracts from overall relevance.
So: the greater the contextual effects, the greater the relevance; and the smaller
the effort needed to achieve those effects, the greater the relevance.
Reference and Relevance — 175
Every utterance, we said, creates expectations of relevance, and around
that expectation our criterion for evaluating possible interpretations is built. In
our book, we define a notion of optimal relevance which is meant to spell out
what the hearer is looking for in terms of effort and effect:
Optimal relevance
An utterance, on a given interpretation, is optimally relevant
iff:
(a) it achieves enough effects to be worth the hearer's
attention;
(b) it puts the hearer to no unjustifiable effort in achieving
those effects.
A word of explanation about each of these clauses.
On the effect side, what the hearer is entitled to look for is enough
effects to make the utterance worth his while to process. In general, what that
means is that he’s entitled to expect more effects than he would have got from
any other information that he could have been processing at the time. How
much that is depends on what is going on elsewhere in his cognitive
environment. For example, suppose that someone walks into an important
lecture and says (14):
(14) Ladies and gentlemen, [ have to tell you that the building's on fire.
"The building’ is a referential expression, and different assignments of
reference will lead to different levels of contextual effect. In the
circumstances, the first hypothesis to come to the audience's mind will be that
‘the building’ means the building where the lecture is taking place. Clearly the
utterance, on this interpretation, would have enough effects to be worth the
audience's attention: their minds would be immediately filled with thoughts
of how to get out. Given that at a formal lecture the audience is supposed to
be entirely absorbed in what the lecturer is saying, it is hard to see what other
interpretation would have enough effects to justify the interruption; and in
these circumstances, the interpretation just suggested is basically the only
possible one.
It might be thought that in other circumstances the intended
interpretation would be harder to pin down. Surely there might be several
radically different combinations of content and context, all of which would
yield enough contextual effects to make the utterance worth the audience's
attention? That is where clause (b) of the definition of optimal relevance
176 Deirdre Wilson
comes in. Clearly, a speaker who wants to avoid any. risk of misunderstanding
should make sure that there is no interpretation more accessible to the hearer
than the intended one, which has enough effects to be worth the hearer’s
attention, and which might therefore lead him astray. Clause (b) of the
definition of optimal relevance, which excludes the unjustifiable expenditure
of effort in recovering the intended effects, covers this type of case: that is, it
excludes the possibility that the hearer will be expected to recover, process and
accept the wrong interpretation before lighting on the intended one. From
clause (b), it follows that a speaker aiming at optimal relevance should try to
formulate her utterance in such a way that the first acceptable interpretation
to occur to the hearer is the one she intended to convey. From the hearer's
point of view, this clause has an immediate practical consequence. Having
found an interpretation which satisfies his expectations of relevance in a way
the speaker might manifestly have foreseen, he need look no further, The first
such interpretation is the only such interpretation, and is the one the hearer
should choose.’
I should nole here in passing that in order to be acceptable and
comprehensible, an utterance does not actually have to be optimally relevant.
Suppose that, as you are about to leave the house, I warn you that it is raining;
as it happens, this is something you already know. In the circumstances, the
proposition 1 have expressed will have no contextual effects and hence be
irrelevant to you; nonetheless, my utterance will be both comprehensible and
acceptable as long as you can see how I might reasonably have expected it to
be relevant. The actual pragmatic criterion we propose, then, is a criterion of
consistency with the principle of relevance which takes account of this type
of case:
Criterion of consistency with the principle of relevance
An utterance, on a given interpretation, is consistent with the
principle of relevance if and only if the speaker might
rationally have expected it to be optimally relevant to the
hearer on that interpretation.
Vague as this criterion may sound, it makes one important prediction
that I have not seen matched elsewhere. This prediction follows from clause
*We are not claiming, of course, that the interpretation selected by this criterion is
invariably correct; merely that it is the only one the hearer is rationally justified in choosing.
In a broadly Gricean framework, utterance interpretation is a non-demonsirative inference
process which takes place at a risk; there is no guarantee that its outcome will be correct.
Reference and Relevance 177
(b) of the definition of optimal relevance and its consequence that the first
interpretation tested and found consistent with the principle of relevance is 1he
only interpretation consistent with the principle of relevance. Let us assume
that, in interpreting an utterance, the hearer starts with a small initial context
left over, say, from his processing of the previous utterance; he computes the
contextual effects of the utterance in that initial context; if these are not
enough to make the utterance worth his attention, he expands the context,
obtaining further effects, and repeats the process until he has enough effects
to make the utterance optimally relevant in a way the speaker could manifestly
have foreseen. At that point, he has an interpretation consistent with the
principle of relevance, and it follows that he should stop; or, at least, he is
entitled to go on on his own account but is not entitled to assume that the
speaker intended to communicate anything more. In other words, all the hearer
is entitled to impute as part of the intended interpretation is the minimal
context and set of contextual effects that would be enough to make the
utterance worth his attention. I will make use of that fact below.
5 Relevance theory and reference
How, within this framework, might the hearer identify the intended referent
of a definite NP such as the door’? In Relevance, we assume, as does Herb
Clark, that reference assignment involves the retrieval or construction of an
appropriate conceptual representation, one that uniquely identifies the intended
referent, We assume, furthermore, that this representation is incorporated into
a representation of the proposition expressed by the utterance, so that in
interpreting (15a), for example, the hearer might construct a conceptual
representation such as (15b), which will act as input to further inference
processes:
(15) a. The door is open.
b. Door dis open at time t.
These are not, incidentally, assumptions that a semanticist like Lewis would
necessarily share, Lewis is interested in assigning the correct truth conditions
to an utterance such as (15a). On his analysis, the utterance would simply be
evaluated against a package of contextual information containing, among other
things, a salience ranking of the candidate referents. There is no suggestion
that a representation of the selected candidate should be integrated into a
conceptual representation such as (15b).
178 Deirdre Wilson
Any cxamination of the pragmatics of reference assignment is
complicated by the fact that tentative assignments are apparently made while
the utterance is still in progress. What strategy should hearers pursue in
making these tentative choices? There is clear experimental evidence about
what they actually do. In this, as in every other aspect of utterance
interpretation, they follow the path of least effort; or rather, they follow this
path unless it already seems likely to them that the resulting interpretation will
be pragmatically unacceptable,
On the face of it, this behaviour is rather surprising. What right do
hearers have to assume that the interpretation they find easiest to construct will
be the one the speaker intended? Relevance theory suggests an explanation. As
we have seen, a speaker aiming at optimal relevance must put her hearer to no
unjustifiable effort in achieving the intended effects. Clearly, a speaker who
allows her hearer, even momentarily, to recover, process and accept the wrong
interpretation would put him to considerable unjustifiable effort. Relevance
theory thus sheds light on the least-effort strategy automatically adopted by
hearers in utterance interpretation. I know of no other pragmatic theory which
does the same.
Returning now to our original examples, consider how (la) and (2a)
might be analysed within this framework. On hearing the definite NP ‘the
door” in (la), the hearer will have immediately accessible (from his processing
of the first part of the utterance) the information that the room the speaker was
talking about had three doors, one of which was open. Suppose, then, that the
open door (strictly speaking, the hearer’s mental representation of the open
door) is the most accessible one, and therefore the first to be tested for
pragmatic acceptability. It is easy to see how the utterance might be acceptable
on this interpretation. We can all think of circumstances in which a speaker
might find it relevant to mention that she had closed an open door, and of
contextual effects that the utterance, on this interpretation, might be intended
to achieve: it might imply, for instance, that the speaker wanted to shut off
noise from outside, to signal that the lecture was about to begin, to deter
latecomers, and so on. By the criterion of consistency with the principle of
relevance, this first acceptable interpretation will be the only acceptable
interpretation, and is the one the hearer should choose.
Clark & Haviland (1977), for example, build into their analysis of utterance
interpretation the idea that the first acceptable candidate should be selected, but their
framework offers no explanation of why this should be so. Perhaps an explanation could be
constructed around Grice's Manner maxim 'Avoid obscurity’; but I know of no serious
attempt within the Gricean framework 10 develop such an account.
Reference and Relevance — 179
Suppose, alternatively - a hypothesis that is implausible in this case, but
might be worth considering in others - that representations of all three doors
are equally accessible. Then, in order to choose between them, the hearer must
look for a context which combines with one of them to yield an interpretation
consistent with the principle of relevance. As noted above, it is easy to think
of effects that a speaker might have intended to achieve by mentioning that
she had closed an open door; it is much harder to think of effects that she
might have intended to achieve by mentioning that she had closed a door that
was already shut. By the criterion of consistency with the principle of
relevance, the first accessible context to yield an acceptable overall
interpretation is the only acceptable context, and is the one the hearer should
choose.
Whichever supposition we make about the relative accessibility of
candidate referents, the result will be the same. Either the open door will be
the most accessible candidate, and will combine with an easily accessible
context to yield an overall interpretation consistent with the principle of
relevance: or all three doors will be equally accessible, and the choice among
them will be determined by the accessibility of potential contexts with which
they might combine to yield an overall interpretation consistent with the
principle of relevance; here again the choice will fall on the open door. These
two factors - accessibility of referents and accessibility of contexts - are
crucial to the relevance-theoretic account of reference assignment.
Notice, incidentally, that we are now in a position to say something
about the stylistic infelicity of examples such as (16):
(16) The room had three doors, all but one of which were closed. I closed
the door.
Although in (16), as in (la), there is only a single open door, which the
speaker must be presumed to have closed, (16), on this interpretation, could
not be optimally relevant. It could not be optimally relevant because, although
it might achieve adequate effects, it would put the hearer to some un justifiable
effort in deriving them. In the first place, the existence of an open door has
to be inferred in (16), whereas it was explicitly stipulated in (1a); yet this extra
inferential effort is not offset by a corresponding decrease in linguistic
complexity. In the second place, intuitively at least, (16) focuses the hearer's
attention on the doors which were closed, thus making these the first to be
tested and adding to the hearer's effort. More generally, it follows from the
definition of optimal relevance that an utterance, on a given interpretation, can
not be optimally relevant if some other utterance that would have been equally
180 Deirdre Wilson
easy for the speaker to produce would manifestly have achieved the intended
effects more economically. This sheds some light on the stylistic infelicity of
examples (12a)-(12c) above.‘
In analysing the bridging example (2a) above, the only additional factor
that has to be taken into account is that there is no ready-made referent
waiting to be retrieved and the hearer must engage in some form of inference
to derive one. As always, the most accessible (here, easily inferable) referent
will be the first to be tested, and the result will be accepted if it meets the
overall criterion of consistency with the principle of relevance.
Here we must make some tentative assumptions about how inferred
referents are accessed. In Relevance, as mentioned above, we assume that the
hearer starts out with an initial context consisting of the information he has
most recently processed, and that this can be expanded in a variety of
directions: for instance, by adding encyclopaedic information about objects
mentioned in the utterance or the initial context. Thus, in (2a), the hearer
might expand the context by adding to it the information that rooms may have
one or more windows; he could then make the hypothesis that the room the
speaker went into had two windows, and use these in assigning reference to
the definite noun-phrase both windows’. It is easy to think of effects that a
speaker might expect her utterance, on this interpretation, to achieve: for
example, that the room was cold, that the windows should be closed; that the
room was well-aired and comfortable; that the occupants of the house had not
gone out, and so on. Hence, inferred referents work in just the same way as
explicitly mentioned referents, and are affected by the same two factors:
accessibility of referents and accessibility of potential contexts with which
these might interact to yield an overall interpretation consistent with the
principle of relevance.
‘What we are claiming goes beyond the idea that there is some absolute limit on the
amount of effort the interpretation of an utterance may require. To satisfy himself that an
utterance, on a given interpretation, is consistent with the principle of relevance, the hearer
has take into account, not whether some absolute threshold of effort has been passed, but
whether some other utterance could have achieved the same effects more economically.
Consider (i), for example:
(i) John and Mary went to see the Scottish play last night.
The three witches were excellent.
Whether a hearer finds this utterance acceptable may depend on his familiarity with the
superstition in certain acting circles that bans the use of the name Macbeth on pain of bad
luck. In such circles, (i) may well be the most economical way available to the speaker of
achieving the intended effects.
Reference and Relevance — 181
6 A problem with Clark and Lewis’s criterion
A detailed comparison of these three approaches to reference assignment is
beyond the scope of this paper. Such a comparison might explore the relations
between Lewis's notion of salience (a property of objects) and Sperber &
Wilson's notion of accessibility (a property of mental representations), or
between Clark's contract-based approach to pragmatics and Sperber &
Wilson's principle-based approach. Here my ambition is more limited. As we
have seen, the three approaches to reference assignment have a common
overall structure, differing only in the pragmatic criteria they propose for
evaluating alternative interpretations. I want to compare these pragmatic
criteria by looking at the predictions they make across a certain range of
examples, and to argue that the relevance-theoretic criterion sheds light on
these examples in a way that Clark and Lewis's criteria do not.
Comparisons are not easy, since the proposed analyses are quite
complex. In particular, although all are sensitive to the accessibility of
candidate referents, none of them claims straightforwardly that the most
accessible referent is invariably selected. That claim could be straightforwardly
refuted by comparing examples such as (1b) above and (17):
(1) b. Sean Penn attacked a photographer. The man was quite badly hurt.
(17) Sean Penn attacked a photographer. The man must be deranged.
In each of these examples there are two candidate referents for the NP “the
man’: in (1b) the intended referent is the photographer; in (17) it is Sean Penn.
An account based on accessibility of referents alone can not explain these
results. It might deal with (1b) by assuming that the most recently mentioned
man is the most salient/accessible one; but this would make the wrong
predictions about (17). It might deal with (17) by assuming that the referent
of a preceding subject NP is the most salient/accessible one; but this would
make the wrong predictions about (1b). Since the first parts of these two
utterances are identical, it seems that whatever man is made most accessible
by one of them must be made most accessible by the other, and an account
based on accessibility of referents alone would make the wrong predictions
about at least one of these examples.
Lewis could, however, appeal to his rule of accommodation for
comparative salience to explain these results. According to Lewis, although
5See Kleiber (1990, forthcoming) for further important arguments against the view that
accessibility of referents is the only factor affecting reference assignment.
182 Deirdre Wilson
reference is invariably assigned to the most salient candidate, initial salience
rankings must be adjusted so that their output is pragmatically acceptable. By
Lewis's criterion, to be pragmatically acceptable, an utterance must be true,
informative and evidenced. As we have seen, in (1b) "the man’ should be
interpreted as referring to the photographer, but this is also the interpretation
on which (1b) is most likely to be true. Similarly, in (17) "the man’ should be
interpreted as referring to Sean Penn, but this is also the interpretation on
which (17) is most likely to be true. Hence, the rule of accommodation,
interacting with a criterion of overall pragmatic acceptability, should have no
trouble with examples such as (1b) and (17).
Or consider (4a) above, where there are two candidate referents for the
NP ’the lectures’ and the intended referent is the geography lectures rather
than the linguistics ones:
(4) a. I switched from linguistics to geography. The lectures were less
boring.
Here the potential bridges are (5a) and (Sb), with resulting interpretations (6a)
and (6b), respectively, of which (6b) will be preferred:
(5) a. Studying linguistics involves going to lectures.
b. Studying geography involves going to lectures.
(6) a. The linguistics lectures were less boring than the geography
lectures.
b. The geography lecturcs were less boring than the linguistics
lectures.
As it stands, Clark's analysis fails to explain why (6b) is preferred. He claims,
on the one hand, that the selected bridge must be both short and plausible: this
would not distinguish between (5a) and (5b), which are equally short and
equally likely to be true. He also claims that the resulting overall interpretation
must be informative; but this would not distinguish (6a) from (6b), which
would both, quite clearly, be informative.
Let us assume, then, that Clark, like Lewis, would be prepared to
evaluate not only the bridge but also the resulting overall interpretation for
likelihood of being true. This would enable him to explain the preference for
(6b). Suppose you were told that Mary had switched from linguistics to
geography, and then asked whether it was more likely that Mary found the
linguistics lectures less boring than the geography lectures, or the geography
lectures less boring than the linguistics lectures. You would no doubt reply
Reference and Relevance — 183
that it is more likely that she found the geography lectures less boring than the
linguistics lectures. Thus Clark, like Lewis, could claim that the choice of an
appropriate referent is affected by the resulting interpretation’s likelihood of
being true.
It is more plausible to suppose, however, that in line with his general
commitment to a Gricean approach, Clark intends to supplement his
Given-New Contract with some version of Grice’s maxims, in which a
criterion of truthfulness would replace Lewis's criterion of truth. On this
approach, overall interpretations would be evaluated not for likelihood of
being true but for likelihood that the speaker believes that they are true. An
approach based on such a criterion would have no difficulty in distinguishing
interpretations (6a) and (6b). Suppose you were told that Mary had switched
from linguistics to geography and then asked whether it was more likely that
Mary believed that the linguistics lectures were less boring than the geography
lectures, or that the geography lectures were less boring than the linguistics
lectures. You would no doubt reply that it is more likely that she believed that
the geography lectures were less boring than the linguistics lectures. Thus
Clark could claim that the choice of an appropriate referent is affected by the
resulting interpretation’s likelihood of being, if not true, at least truthful, i.e.
believed by the speaker to be true.
To refute these accounts, we need an utterance with the following
structure: first, it has two possible overall interpretations, which are equally
informative, equally evidenced, and equally likely to be true (or truthful);
second, hearers systematically prefer one interpretation to the other; and third,
this choice cannot be explained by appealing to differences in the inherent
accessibility of referents. If Clark and Lewis are right, utterances of this type
should not exist. If they do, something is missing from their accounts.
Example (18) seems to me a case in point:
(18) I switched from linguistics to geography. The lectures were just as
boring.
Here there are two potential bridges, (19a) and (19b), and two possible overall
interpretations, (20a) and (20b):
(19) a. Studying linguistics involves going to lectures.
b. Studying geography involves going to lectures.
184 Deirdre Wilson
(20) a. The linguistics lectures were just as boring as the geography
lectures.
b. The geography lectures were just as boring as the linguistics
lectures.
The natural interpretation of (18), 1 take it, is (20b), with bridging assumption
(19b) being used to assign reference to the NP ‘the lectures’. How might this
be explained on Clark's (or Lewis's) account? The choice is not determined
by the length or plausibility of the bridge, since the two potential bridges are
equally short and equally plausible. lt must therefore depend on the resulting
interpretation's likelihood of being true (or truthful), informative or evidenced.
But as far as I can tell, interpretations (20a) and (20b) are logically equivalent,
and hence equally likely to be true (or truthful), equally informative, and
equally evidenced. Thus, the pragmatic criteria proposed by Clark (and Lewis)
can not explain the preference for (20b) over (20a).
On Clark and Lewis's approaches, the choice of (20b) would therefore
have to be explained by appeal to variations in the accessibility of candidate
referents. It might be claimed, for example, that though both bridges are
equally short and plausible, one of them is inherently more accessible than the
other. This is a reasonable claim, since bridging assumptions are contextual
assumptions, which quite generally vary in accessibility. Let us suppose that
the inherent accessibility of a bridge is determined by the syntactic position
of its antecedent NP. We might then maintain that the bridge based on the
most recently mentioned antecedent NP is the most accessible one, and thus
explain why the bridge in (19b), based on the most recently mentioned NP,
‘geography’, is preferred to the bridge in (19a).
This hypothesis will not do. If it did, then by changing the order of the
NPs in (18) we should be able to change the interpretation. But compare (18)
and (21):
(18) I switched from linguistics to geography. The lectures were just as
boring.
(21) 1 switched 10 geography from linguistics. The lectures were just as
boring.
Clearly, the change in the order of NPs does not lead to a change in
interpretation. In both these utterances, the preferred bridge for the NP ‘the
lectures’ is (19b), giving rise to the overall interpretation (20b).
Indeed, it is hard to conceive of any accessibility ranking based on
surface syntactic position alone that would explain the full range of choices
Reference and Relevance — 185
that hearers apparently make in examples of this type. In each of (22a-e), for
instance, there is a clear preference, in interpreting the NP ‘the lectures’, for
the bridge based on the antecedent NP geography’, despite the variations in
its surface position:
(22) a. Geography followed linguistics. The lectures were just as boring.
b. Linguistics was followed by geography. The lectures were just as
boring.
c. Linguistics gave way to geography. The lectures were just as
boring.
d. After linguistics came geography. The lectures were just as boring.
e. Geography came after linguistics. The lectures were just as boring.
Nor can these choices be explained by invoking a rule of accommodation
which would adjust inherent salience rankings to meet some overall criterion
of pragmatic acceptability. As we have seen, the resulting overall
interpretations of (18) and (21) are logically equivalent, and hence equally
likely to be true (or truthful), informative and evidenced; hence the pragmatic
criteria proposed by Clark and Lewis do not choose between them.
To save these criteria, one would therefore have to find an alternative
accessibility ranking, based on something other than surface syntactic position,
that would explain the preference for the bridge in (19b) across the full range
of examples (18), (21) and (22). One might appeal, for instance, as do Erku
& Gundel (1987), to a notion of topic, and claim that the most accessible
bridge is the one based on the topic NP; one might appeal to an accessibility
hierarchy based on underlying thematic role (see Keenan & Comrie 1977); or
one might explore the notion of focus used in some of the recent Al literature
(see for example Asher & Wada 1988). Let me emphasise that I am not
against attempts to explore these possibilities. Clearly, reference assignment
is affected by accessibility of referents, and it may turn out that surface
syntactic position of the antecedent NP is not the only factor affecting
accessibility of referents in this type of example. My point here is that Clark
and Lewis's criteria force them to appeal to variations in accessibility of
referents in analysing these examples; that in the absence of any clear account
of how accessibility is determined, the resulting analyses cannot but be
unilluminating; and that the relevance-theoretic criterion, as | will shorily
argue, offers some immediate insight into how these examples work.
There is, however, a more serious objection to Clark and Lewis’s
criteria. As we have seen, they claim that overall interpretations are evaluated
for truth (or truthfulness), informativeness and degree of evidence. In fact,
186 Deirdre Wilson
there are two conceptions of how the evaluation procedure might work. On the
first, and weaker, conception, an interpretation would be acceptable as long as
it was true (or truthful), informative and evidenced; on the second, and
stronger. conception (the one 1 have tacitly used above), the only acceptable
interpretation would be the one that was most likely to be true (or truthful),
most informative and most strongly evidenced. I consider each conception in
turn.
Suppose we adopt the weaker conception. Then the proposed pragmatic
criteria will be unable to distinguish between the possible interpretations of
(1b) and (17) above. If Sean Penn attacked a photographer it may be true that
the photographer got hurt, but it may also be true that Sean Penn got hurt, If
Sean Penn attacked a photographer, it may be true that Sean Penn is deranged,
but it may also be truc that the photographer is deranged. Although in each
case onc of these interpretations is more likely to be true, neither can be ruled
out on grounds of simple truth (or truthfulness) alone. The same point applies
to evaluations based on simple informativeness and degree of evidence: both
utterances would be informative and evidenced on either interpretation.
By the same token, an evaluation procedure based on simple truth (or
truthfulness), informativeness and degree of evidence would be unable to
choose between the two possible interpretations of (4a) above. If Mary
switches from linguistics to geography, it may be true that she finds the
geography lectures less boring, but it would be quite possible for her to
discover, after making the change, that the linguistics lectures were less boring
after all; similarly, the utterance might be informative and evidenced on either
interpretation.
In other words, the proposed pragmatic criteria, understood in this way,
are not powerful enough to choose between the candidate interpretations of
(1b), (4a) and (17). But these were just the examples used above to support
the view that reference assignment does not depend on accessibility of
referents alone, but is also influenced by the need to obtain an overall
interpretation that satisfies some criterion of pragmatic acceptability. A
criterion as weak as the one we are now considering would be unable to
distinguish among the candidate interpretations of (1b), (4a) and (17). If their
pragmatic criteria are understood in this way, the whole weight of Clark and
Lewis's account will be forced back onto considerations of accessibility of
referents, and most of their insights will be lost.
This weak conception of the pragmatic criterion can be further
undermined by comparing (4a) with (23):
Reference and Relevance 187
(4) a. I switched from linguistics to geography. The lectures were less
boring.
(23) I switched from linguistics to geography. The lectures were too
boring.
In (da), as we have seen, the correct bridge for the NP ‘the lectures’ is based
on the NP ‘geography’; in (23), it is based on the NP "linguistics”. As we have
also seen, neither choice appears to follow from simple considerations of truth
(or truthfulness), informativeness or degree of evidence. To account for the
interpretation of these examples, given this weak conception of the pragmatic
criterion, one would therefore have to appeal to considerations of accessibility
of referents. However since the first parts of these utterances are identical, it
is hard to see how *geography' could be the most accessible candidate in (4a)
and linguistics’ in (23). It seems clear that in order to understand these
examples, we need a pragmatic criterion that does more work than Clark and
Lewis’s would on the weak interpretation we are now considering.
In looking above at how Clark and Lewis might deal with the examples
(1b), (4a) and (17), I tacitly used a stronger version of their criteria. I took it
that these criteria were designed to select not merely an interpretation on
which the utterance is true (truthful), informative and evidenced, but the
interpretation on which the utterance is most likely to be true (truthful), most
informative and most strongly evidenced. Such a criterion, as I have shown
above, is powerful enough to distinguish the various candidate interpretations
of (1b), (4a) and (17). However, it tco encounters a serious problem. Though
it may be able to choose among alternative interpretations of a given utierance,
the only way it can do so is by considering them all. Recall that what we are
looking for is not just a criterion for evaluating alternative reference
assignments, but one for evaluating overall interpretations in general.
Evaluating overall interpretations means not just considering the two possible
assignments of reference | have mentioned for examples like (4a): it means
considering all possible parsings, disambiguations, assignments of reference,
restorations of ellipsed material and resolutions of vagueness, and all potential
sets of contextual assumptions that might have a bearing on these choices.
Clearly, an attempt to take all these possibilities into account would create a
combinatorial explosion of gigantic proportions, and be quite unlike anything
that hearers actually do.
What seems to me one of the virtues of relevance theory is that its
pragmatic criterion is explicitly designed to offer a way out of this dilemma.
On the relevance-theoretic account, although the hearer may quite often have
to make choices, the pragmatic criterion itself does not force him to consider
188 Deirdre Wilson
all possible interpretations as a necessary part of the comprehension process.
As we have seen, on a relevance-theoretic account, the hearer is justified in
accepting the first accessible interpretation that is pragmatically acceptable: he
has a criterion, in other words, that enables him to recognise an acceptable
interpretation as soon as he sees it. This seems to be a virtue that other
pragmatic theories might consider building into their accounts.
7 A relevance-theoretic solution
So what has relevance theory to say about the examples we have been
discussing? Consider (4a). Intuitively, there are two ways in which the second
part of this utterance can achieve relevance in a context created by the earlier
part: first, as an explanation of why [ switched from linguistics to geography,
and second, as a description of what 1 found after making the switch. On the
preferred interpretation (6b), where ‘the lectures’ refers to the geography
lectures, and against an easily accessible set of contextual assumptions about
what might make one switch subjects, the utterance will simultaneously
achieve relevance in both these ways. By contrast, on the alternative
interpretation (6a), where ‘the lectures’ refers to the linguistics lectures, the
utterance will achieve relevance in neither of these ways. Assuming that
hearers are alert to these quite typical ways of achieving relevance, they will
easily access interpretation (6b), which will be relevant enough to be worth
their attention, making it unnecessary to look further.
Consider (18) and (21):
(18) I switched from linguistics to geography. The lectures were just as
boring.
(21) I switched to geography from linguistics. The lectures were just as
boring.
Here, one of the two possible lines of interpretation mentioned above will be
ruled out: it is hard to see how either the fact that the geography lectures were
just as boring as the linguistics lectures, or the fact that the linguistics lectures
were just as boring as the geography lectures, could be an adequate
explanation of the switch from one to the other. By contrast, on the
assumption that the boring lectures were the geography ones, then both (18)
and (21) can achieve relevance as descriptions of what 1 found after making
the switch. Again assuming that hearers are alert to these lines of
Reference and Relevance 189
interpretation, the intended analysis will be easily discovered and all others
accordingly disconfirmed.
Now consider (4b):
(4) b. Iran from the classroom into the playground. The children were
making too much noise.
Suppose the hearer decides that the second part of the utterance is intended as
an explanation of the action described in the first. There are still two possible
lines of interpretation: I might have run out of the classroom because the
children in the classroom were making too much noise; or | might have run
into the playground because the children in the playground were making too
much noise. Either interpretation could yield enough effects to make the
utterance worth the hearer's attention. The choice between them must therefore
depend on the accessibility of contexts capable of combining with one of them
to yield adequate effects for no unjustifiable effort in a way the speaker could
manifestly have foreseen. In real life, assumptions about the character of the
speaker would affect the accessibility of potential contexts: is she the sort of
person likely to be frightened by children, is she a stickler for discipline, and
so on? In a questionnaire situation, which favours the use of stereotypical
assumptions, it would no doubt be easier to imagine her as a teacher running
into the playground to stop the noise rather than a teacher fleeing the
classroom because of the noise; hence this interpretation should be preferred.
This account, though sketchy, seems to me to give some genuine insight
into the examples we have been discussing. It saves Clark and Lewis's claim
that while the accessibility of candidate referents plays a role in reference
assignment, this role is not decisive - a claim that, as we have seen, is not
necessarily compatible with their own pragmatic criteria. lt sheds light on the
least-effort strategy pursued by hearers - a strategy that is rather surprising on
Clark and Lewis's accounts. It maintains, as do Clark and Lewis, that what is
decisive in reference assignment is the discovery of an overall interpretation
which meets a criterion of pragmatic acceptability. It differs, however, as to
the proposed criterion. The criterion proposed by relevance theory is one of
consistency with the principle of relevance: to discover an interpretation that
meets the criterion, the hearer must be able to access, not only an appropriate
referent, but also a context in which the speaker might reasonably have
expected her utterance, on that assignment of reference, to be optimally
relevant.
190 Deirdre Wilson
Finally, this account sheds light on some stylistic questions that have
been to some extent neglected in the philosophical literature on reference
assignment. Here is a representative remark from Lewis:
"It is not good conversational practice to rely too heavily on
rules of accommodation. [In a monologue where] ‘the cat’
denotes first Bruce then Albert, then Bruce again, what I say
is to some extent confusing and hard to follow. But even if my
monologue is not perfectly acceptable, its flaws are much less
serious than the flaws that are averted by shifts of salience in
accordance with our rule of accommodation. Confusing shifts
of salience and reference are not as bad as falsity, trivial truth,
or unwarranted assertion.’
[Lewis1979/1983: 243]
That, 1 think, sums up the current philosophical state of the art. The criteria
proposed by philosophers for evaluating alternative interpretations tend to
focus on their informational content, and to set aside stylistic questions,
questions of effort. If relevance theory is right, economy of effort is just as
central to the evaluation process as adequacy of effect. Philosophers have
recently begun to wonder about the differential acceptability of examples such
as (1a) and (16) above. I would suggest that unless we begin to take seriously
the role of processing effort in evaluating overall interpretations, these
examples will continue to puzzle us.
Reference and Relevance 191
References
Asher, N. & H. Wada 1988. A computational account of syntactic, semantic
and discourse principles for anaphora resolution. Journal of Semantics
6: 309-44.
Clark, H. 1977. Bridging. In P. Johnson-Laird & P. Wason (eds) Thinking:
Readings in Cognitive Science. CUP, Cambridge:411-20,
Clark, H. & S. Haviland 1977. Comprehension and the Given-New Contract.
In R. Freedle (ed.) Discourse Production and Comprehension. Ablex,
Norwood, N.J.: 1-40.
Erku, F. & J. Gundel 1987. The pragmatics of indirect anaphors. In J.
Verschueren & M. Bertuccelli-Papi (eds) The Pragmatic Perspective.
Benjamins, Amsterdam: 533-46.
Keenan, E.L. & В. Comrie 1977. NP accessibility and universal grammar.
Linguistic Inquiry 8: 63-100.
Kleiber, G. 1990. Marqueurs referentiels et processus interpretatifs. Cahiers
de linguistique francaise 11.
Kleiber, G. forthcoming. Reference pronominale: comment analyser le pronom
il. To appear in B. Munch (ed.) Proceedings of the International
Semiotics Conference on Reference and Indexicalisation.
Lewis, D. 1970/83. General semantics. Synthese 22: 18-67. Reprinted in D.
Lewis 1983 Philosophical Papers. Oxford University Press, Oxford.
Lewis, D. 1979/83. Scorekeeping in a language game. Reprinted in D. Lewis
1983 Philosophical Papers. OUP, Oxford.
Sperber, D. & D. Wilson 1986. Relevance: Communication and Cognition.
Blackwell, Oxford.
Sperber, D. & D. Wilson 1987. A precis of Relevance, Presumptions of
relevance. Behavioral and Brain Sciences 10: 697-754,
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement