To appear in Journal of Linguistics
Stephen C. Levinson, Presumptive meanings: the theory of generalized
conversational implicature. Cambridge MA: MIT Press, 2000. Pp. xxiii + 480.
Reviewed by ROBYN CARSTON, University College London
The basic thesis of this book is that there is a level of utterance-type meaning,
which is distinct from, and intermediate between, sentence-type meaning and
utterance-token meaning. That is, it is more than encoded linguistic meaning
but generally less than the full interpretation of an utterance. Here are some
examples, where (a) is a sentence and (b) is its utterance-type meaning in each
Some of the children passed the test.
but not all of the children passed the test.
Mary looked at John and he smiled.
Mary looked at John and then he=John smiled.
Nick was instrumental in lighting the fire.
something odd in the way Nick lit the fire.
Can you pass the salt?
I request that you pass the salt.
The highlighted elements in each of the (b) representations are not derived by
linguistic decoding but are pragmatically inferred.
In the Gricean pragmatics tradition, pragmatically inferred meaning is
usually closely associated with context-dependence and with maxims or
principles which are geared to the recovery of the speaker’s intended meaning.
However, while Levinson agrees that this is the right way to view the
processes of full interpretation of an utterance token, he takes a quite different
stance on the pragmatics of utterance-type meaning, which is a matter of
preferred or default (or ‘presumptive’) interpretations, ‘which are carried by
the structure of utterances, given the structure of the language, and not by
virtue of the particular contexts of utterances’ (1). And while these default
interpretations are licensed by certain pragmatic principles or heuristics, they
are ‘based not on direct computations about speaker-intention but rather on
general expectations about how language is normally used’ (22). That is, they
are generated automatically by default usage rules associated with certain
linguistic expressions and structures. So, for instance, the quantitative term
some in (1a) carries a default rule licensing the inference to ‘not all’ and the
conjunction and in (2a) carries a default rule to the effect that the event
described in the first conjunct preceded that described in the second conjunct.
Since these are default inferences, hence defeasible, their results can be
overridden and this is where context does play a role: if the default output is
inconsistent with the context it is dropped. In the case of (1a), for instance, if
there is a contextual assumption to the effect that all of the children passed the
test then this will defeat the default inference given in (1b).
Levinson mentions a number of pragmatic phenomena, including
illocutionary force (as in (4) above), conversational routines and
presuppositions, which contribute to the level of utterance-type meaning, but
the focus of the book is on a class of conversational implicatures, exemplified
in (1b)-(3b) above. He makes a sharp distinction between these generalized
conversational implicatures (GCIs) and conversational implicatures of a
particularized sort (PCIs):
Did the children’s summer camp go well?
Some of them got stomach ‘flu.
GCI: Not all of the children got stomach ‘flu.
The summer camp didn’t go as well as hoped.
While the PCI of B’s utterance depends on the context provided by A’s
question and would not arise in a different context (e.g. a context in which the
issue is whether all the children were able to sit their exams), the GCI would
arise quite generally across contexts. These two domains of pragmatic
inference work in totally distinct ways: PCIs depend on some (unspecified)
maxim of relevance which is responsive to particular contextual assumptions,
while GCIs are underpinned by three informativeness principles (based
roughly on Grice’s quantity and manner maxims), each of which licenses the
hearer to employ a corresponding heuristic:
(6) Q-HEURISTIC: What isn’t said to be the case is not the case.
I-HEURISTIC: What is said in a simple (unmarked) way represents a
M-HEURISTIC: What is said in an abnormal (marked) way represents an
The Q-heuristic has to be relativized to a relevant scale of lexical alternates,
e.g. <all, some> for (1) and (5) above. The I-heuristic and the M-heuristic are
responsible for the implicatures in (2b) and (3b) respectively. As Levinson
acknowledges, this system is similar to that of Horn (1984), with the effects of
the I- and M-principles reflecting his ‘division of pragmatic labour’: two
coextensive expressions differing in formal markedness tend to become
associated with complementary subsets of the original extension (e.g. kill and
cause to die).
These, then, are the core ideas explored in the book, which is organised
into a short introduction, four long chapters and a short epilogue. The first big
chapter sets out to make the case that GCIs comprise a distinct domain within
pragmatics. It traces the Gricean background within which the distinction
between generalized and particularized conversational implicature arose and
argues that an approach like Relevance Theory (RT) (Sperber & Wilson
1986/95) which does not give the distinction any theoretical weight and
employs the same communicative principle and comprehension procedure in
the derivation of all conversational implicatures cannot do justice to the nature
of these generalized inferences. The case for GCIs is given empirical support
by the observation, again from Horn, that languages do not lexicalize the
meanings ‘not all’, ‘not always’, ‘not both’ (as opposed to ‘none’, ‘never’,
‘nor’). The idea is that this is because each of these meanings is inferred by
default from the words some, sometimes, and or, respectively.
The second chapter explores the three species of GCI in considerable
detail. Levinson provides a wealth of examples of each kind and candidly
acknowledges that some of them raise problems for his account. For example,
the scales at issue in the generation of scalar Q-implicatures may be context-
dependent (e.g. a scale consisting of celebrities ordered in terms of their
popularity) rather than a matter of semantic entailment (as in the cases of
all/some, and/or and the number terms), so that this kind of Q-inference
crosscuts the generalized/particularized distinction. A quite disparate range of
phenomena fall in the class of I-based inferences, including conjunction
buttressing, bridging inferences, some cases of pronominal reference
resolution as in (2) above, lexical narrowings, possessive interpretations, and
several of these can have more than one outcome, so don’t seem to be cases
which have a default/preferred interpretation after all. The chapter ends with a
discussion of the potential conflicts among the three principles and resolves
the problem by imposing an order of priority on them: first Q-inferences, then
M-inferences and finally I-inferences.
The third chapter is, to my mind at least, the most interesting, as it is
here that Levinson confronts the role of pragmatic inference in determining
the truth-conditional content of an utterance. That pragmatics plays this role is
widely acknowledged nowadays by pragmatists across various frameworks,
but it tends still to be resisted by advocates of a truth-conditional semantics for
natural language, as it causes obvious problems for a compositionality
principle conceived in truth-conditional terms and calls into question the
traditional semantics/pragmatics distinction. Although he doubts that it will
ultimately work, Levinson would like to ‘limit the damage’ with the
hypothesis that it is just his chosen domain of pragmatic inferences, GCIs, that
can affect truth conditions. They can do this in a range of ways, including
playing a role in processes of disambiguation and reference resolution, but
most significantly, there are certain situations in which their own content is
actually composed into the truth conditions of the utterance. This occurs in
the class of what he calls ‘intrusive’ constructions (which include negations,
conditionals, disjunctions, comparatives). He calls them intrusive because
they have the property that ‘the truth conditions of the whole expression
depend on the implicatures of some of its constituent parts’ (213-14):
If both teams got three goals the game was a draw.
If both teams got exactly three goals the game was a draw.
It’s better to drive home and drink a bottle of wine than to drink
a bottle of wine and drive home.
It’s better to drive home and then drink a bottle of wine than to
drink a bottle of wine and then drive home.
For (7a), the GCI of the embedded sentence both teams got three goals,
namely ‘at most three goals’, is composed with the encoded semantics ‘at least
three goals’ to give the truth conditions in (7b); similarly, mutatis mutandis,
Thus, what is a non-truth-conditional element (an implicature) of the
simple sentence becomes part of the truth conditions of the more complex
sentence in which the simple one is embedded. This seems barely coherent
and leads to the prediction that the intuitively valid argument in (9) is invalid,
since the truth conditions of premise 2 don’t match those of the antecedent of
the conditional in premise 1:
If both teams got three goals then the game was a draw.
Both teams got three goals.
The game was a draw.
Relevance theorists, on the other hand, predict the intuitive validity of (9),
since they take the view that utterances of the complex sentences in (7a) and
(8a) AND utterances of the simple sentences on their own are equally likely to
be pragmatically enriched; this is not a matter of implicature in either case but
of pragmatic development of the schematic encoded logical form of the
utterance (see Carston forthcoming). For a recent bid to save the traditional
semantic picture by limiting the truth-conditional effects of pragmatics to the
saturation of linguistically given variables, see King & Stanley (in press).
The fourth chapter, which relies heavily on two papers Levinson
published in this journal in 1987 and 1991, argues for the very interesting
hypothesis that the three Binding Conditions of generative grammar can be
reduced to a single grammatical condition with the effects of the other two
being secured by default pragmatic inferences of the Q and M variety.
There is no space here for detailed assessment of Levinson’s important
project, which challenges much received thinking. (For a recent thoughtful
critique, see Bezuidenhout 2002). However, as a relevance theorist I am
bound to issue the following caveat: readers not well acquainted with
relevance theory will get a rather skewed view of it from this book. Levinson
repeatedly claims that, since RT is a theory of context-sensitive inference, it is
inherently incapable of accounting for generalized inferences such as those
above – he gives NO argument to substantiate this serious allegation. He
makes other claims about RT: ‘… according to [Sperber & Wilson] all
inference involved in implicature derivation is deductive, hence the inferences
must be monotonic’ (56); ‘Relevance theorists propose that there is a special
kind of implicature, an explicature, that embellishes logical forms in limited
ways’ (238); ‘Wilson and Sperber … have argued that pragmatics amounts to
nothing more than central reasoning processes applied to linguistic stimuli’
(371). The first claim here is false, the second a distortion, and the third,
which did appear in an early RT paper, has long since been superseded (see
any RT publication since 1994, in particular Sperber & Wilson 1995, Carston
2002, Wilson & Sperber forthcoming).
The issue of whether or not default inferences of the sort that Levinson
proposes are in fact carried out in the on-line process of utterance
interpretation is currently one of the main foci of work in the newly
developing field of experimental pragmatics (see, in particular, Bott & Noveck
2003, Katsos et al. 2003). Bott & Noveck asked adult subjects to respond with
‘true’ or ‘false’ to utterances of underinformative sentences such as ‘Some
robins are birds’ or ‘Some elephants are mammals’. Subjects who respond on
the basis of linguistic meaning alone will say ‘true’ while those who have
performed the pragmatic scalar inference, giving ‘some but not all robins are
birds’, etc. will say ‘false’. Responses were given under one of two
conditions: (a) with a short time lag (900 msecs) between presentation of the
sentence and subjects’ response, and (b) with a longer time lag (3 seconds).
The point of this was to control for the amount of processing effort subjects
could expend before giving their response. The default inference account
predicts that the inference is drawn automatically and only subsequently
cancelled when checked against context (general knowledge that all robins are
birds, etc), so that one would expect fewer ‘true’ responses in the short time
condition than in the longer time condition. The reverse is predicted by RT,
which does not assume any automatic default pragmatic inferences: the
pragmatically enriched interpretation (prompting the response of ‘false’)
should take longer than the encoded logical response. The results were
statistically significant: 72% of the subjects responded ‘true’ in the short time
lag condition, while only 56% responded ‘true’ in the longer time lag
condition. This is at odds with the view that the pragmatic interpretation
arises from an automatic default inference which is only subsequently
cancelled. The authors conclude that there is no evidence that some has a
default interpretation of ‘some but not all’. Needless to say, much more
empirical testing of the predictions of different pragmatic theories is needed
before final judgement is made, but the GCI theorist cannot take heart from
the results so far.
Finally, although much of the material in this book has been around in
some form or other for well over a decade, it is very useful to have it all
collected together in one volume. There are many interesting and provocative
lateral thoughts to be found in the notes to the chapters, and the short epilogue
sets out issues which will be debated in pragmatics for many years to come.
Bezuidenhout, A. (2002). Generalized conversational implicatures and default
pragmatic inferences. In Campbell, J., O’Rourke, M. & Shier, D.
(eds.) Meaning and truth: investigations in philosophical semantics.
New York: Seven Bridges Press. 257-283.
Bott, L. & Noveck, I. (2003). Some utterances are underinformative: the onset
and time course of scalar inferences. Ms. Institute of Cognitive
Carston, R. (2002). Thoughts and utterances: the pragmatics of explicit
communication. Oxford: Blackwell.
Carston, R. (forthcoming). Truth-conditional content and conversational
implicature. In Bianchi, C. (ed.) Proceedings of the Genoa workshop
on context. Stanford, CA: CSLI Publications.
Horn, L. (1984). Toward a new taxonomy for pragmatic inference: Q- and R-
based implicature. In Schiffrin, D. (ed.) Meaning, form and use in
context. Washington DC: Georgetown University Press. 11-42.
Katsos, N., Breheny, R., Williams, J. & Lee, M.W. (2003). Are generalised
conversational implicature generated on-line by default? Ms.,
RCEAL, University of Cambridge, Cambridge.
King, J. & Stanley, J. (in press). Semantics, pragmatics, and the role of
semantic content. In Szabo, Z. (ed.) Semantics vs. pragmatics.
Oxford: Oxford University Press.
Sperber, D. & Wilson, D. (1986/95). Relevance: communication and
cognition. Oxford: Blackwell. Second edition with new postface 1995.
Wilson, D. & Sperber, D. (forthcoming). Relevance theory. In Horn, L. &
Ward, G. (eds.) Handbook of pragmatics. Oxford: Blackwell.
Department of Phonetics and Linguistics,