Stephen C. Levinson (2000)

Presumptive meanings:
The theory of generalized conversational implicature,

MIT Press, Cambridge, Massachusetts & London, England, 480p.

Reviewed by: Bert Bultinck

Seventeen years after his ground-breaking introduction to pragmatics (Levinson 1983), Stephen Levinson presents another voluminous book on the subject, this time limiting himself to only one of the central topics in pragmatics: the so-called "Gricean implicatures". More precisely, Levinson addresses the problem of "generalized conversational implicatures" (as opposed to "particularized conversational implicatures"). According to Levinson, only the former are truly linguistic in that only they do not rely on "specific contextual assumptions" (p. 16): "the point being emphasized is that a theory of GCIs has to be supplemented with a theory of PCIs that will have at least as much, and possibly considerably more, importance to a general theory of communication. It is just to a linguistic theory that GCIs have an unparalleled import" (p. 22, Levinson's emphasis).

Presumptive meanings contains four major chapters, a succinct introduction and a short epilogue. For the background that students might need in order to comprehend the issues discussed in this book, Levinson refers back to his 1983 introduction. It is clear that Levinson addresses himself mainly to people who are at least familiar with the Gricean canon, even though he includes a fairly extensive "note to students". I will first summarize the four essential components, before embarking on a critical evaluation of his analyses.

In the introduction and in the first chapter, Levinson claims that the existence of "preferred interpretations" is self-evident. In what follows he will concentrate on an analysis of how these preferred interpretations remain relatively, and structurally, invariant over changes in context. Levinson's insights rest on the acceptance of a general human capacity of "inference", but he also refers to two more specific ideas. The first is "an argument from design": due to our anatomic make-up, there seems to be a bottle-neck in our information system: the "transmission rate of human speech" (the actual time we need in order to form phonologically discrete sounds) is about four times slower than other aspects of human speech production (prearticulation processes, parsing and comprehension procedures). This loss of speed, Levinson assumes, has to be made up by inferencing processes: if we can express more by using less words, we gain in speed. Another, related basic intuition concerns "games of pure coordination", referring back to game-theoretic experiments from the sixties, in which people appeared to be able to coordinate their actions, solely on the basis of default inferences and assumptions about what the other would be doing. These two ideas (Levinson mentions them in passing, but he keeps referring back to them throughout the rest of the book) and the corresponding domain of information theory constitute the background for his discussion of the three "heuristics" that guide generalized conversational implicatures. He exemplifies his principles with one example sentence: "There's a blue pyramid on the red cube". The first heuristic can be summed up in the formula "what isn't said, isn't" (labeled Q) and crucially relies on salient lexical contrasts to rule out a whole number of states-of-affairs. In his example, this principle licenses, e.g., the inference that "there is not a cone on the red cube". The second heuristic says that "what is simply described is stereotypically exemplified" (I). This principle licenses the inference that, e.g., there is direct contact between the pyramid and the cube - if the contact was indirect, the speaker should have said so, because the stereotypical situation described by an "on"-relationship is one involving direct contact. The third principle is the counterpart of the second: "What's said in an abnormal way, isn't normal; or Marked message indicates marked situation" (M). If, for instance, one should say: "The blue cuboid block is supported by the red cube", one can infer that the block in question is not a stereotypical one. The marked form "cuboid" licenses the inference that the block is not a canonical cube. Levinson also discusses how his three new principles correspond to the Gricean maxims and other, neo-Gricean proposals, distinguishing them by means of diagnostic principles, as well as relating them to more general "reasoning systems" (e.g., "default logics"). Also, the so-called projection problem is treated by introducing "Gazdar's bucket". First entailments, then clausal Q GCIs, then scalar ones, then M GCIs and finally I GCIs (with I and M implicatures the distinction between clausal and scalar ones is irrelevant) are added to the context (the bucket), provided the later additions are compatible with the previous ones and with the general context. At this point, Levinson launches his first extensive attack on Relevance Theory, the framework that will remain his sparring-partner throughout the rest of the book. His main critique resides in the observation that Sperber and Wilson and their followers do not allow for an intermediate level of generalized conversational implicatures, in between "literal meaning" (semantics) and once-off ("nonce") inferences. Finally, Levinson tackles another rivaling account (which works with the concept of "accommodation") and demonstrates how GCIs can account for lexicalization patterns.

The second chapter surveys "the phenomena'": Levinson lists and discusses famous and lesser-known applications of each of his communicative heuristics. For each heuristic, the author delves into the history of the concept (loyally turning to Grice's views at the beginning of each discussion), reiterates the criticism that has been voiced against the concept and the subsequent analyses and finishes by enumerating a fairly extensive list of examples. Obviously, the highly productive Q-implicatures receive the most attention: various views on scalar and clausal implicatures are discussed (with an interesting elaboration on negative scalar items, and Levinson's claim that "negation reverses scales" (p. 80; cf. also Atlas and Levinson (1981)).

The somewhat longer third chapter tries to compute the consequences of the theory of GCIs for the semantics-pragmatics interface, in what Levinson calls "the architecture [of an overall theory of meaning]" (p.166). He explicitly attacks the Chomskyan modular perspective with its well-known consequences in terms of the alleged autonomy of the different modules and the hierarchies based on "logical priorities". Because GCIs have truth-conditional implications, semantics (traditionally defined as truth-conditional meaning) can no longer be thought of as an independent level. With the aid of a number of examples, Levinson proceeds to show how semantics and implicatures interpenetrate. What he calls "semantics" seems to encompass the traditional semantic representations (even though Levinson advocates a much more abstract "meaning representation", and as such embraces the trend toward "underdetermined" meanings, cf. Atlas 1989, Carston 1998, Pustejovsky 1995, Sperber and Wilson 1986) and a second, truth-conditional level that enriches the semantic structure of the first component (taking into account implicatures, presuppositions and other pragmatic meaning contributions). The domains that Levinson refers to in order to demonstrate his claim of "pragmatic intrusion" are deictic and other classic semantico-pragmatic problems: reference resolution, generality narrowing, scope assignments, ellipses. I will give only two examples, as I will come back to this discussion in the critical evaluation. Consider the utterances Take these three drinks to the three people over there and take these four to the four people over there, in the situation in which there are two sets of glasses (one with three, one with four). Levinson argues that the numerals in question have an 'at least'-semantics, and hence the expression these three drinks is ambiguous - it may refer to the set with three or to the set with four glasses. Because of the well-known scalar implicatures (resulting in 'exact'-readings of the numerals) the expression is disambiguated. The second example concerns the ellipsis in an answer like John (the question being for example Who came?). The semantics of the answer can be enriched to 'John came' by means of an I-inference and thus acquires a truth-conditional content. After an extensive counter-screening procedure, in which Levinson investigates various "architectures" that would seem to be able to save the integrity of semantics and pragmatics, he concludes that semantics and pragmatics must indeed interleave. In passing, he draws up a very handy table of most of the terminologies that have been used to refer to the domain in between "what is said" and "what is implicated". The chapter closes with more evidence for the bi-directional informational relation between semantics and pragmatics.

In the final chapter Levinson confronts the generative paradigm head-on by developing an alternative account of so-called binding phenomena, in this case the binding of anaphora. He starts this chapter, which relies heavily on two of his earlier articles (Levinson 1987, 1991), by emphasizing the importance of implicatures for grammaticalization processes and then summarizes the generative approaches of the problem of anaphora, relying on binding conditions and (technically defined) domains. Levinson works with the assumption that semantic generality is the essential property for an anaphoric expression: "Anaphoric potential is, on the whole, a property not of expressions but of uses of expressions" (p. 270), since "just about any semantically general expression can function anaphorically" (p.268). He lists a few examples with so-called "discourse anaphora" to demonstrate this and then focuses on intrasentential phenomena.

The generative account of anaphora is based on three Binding Conditions. The first Binding Condition states (loosely paraphrased) that an Anaphor (e.g., a reflexive pronoun) is bound in its governing category. The second Binding Condition holds that nonreflexive pronouns must be free in their governing category. The third states that a lexical NP must be free everywhere. In the first version of his analysis Levinson reduces the second and the third of the three Binding Conditions to a pragmatic rule. Levinson claims that the latter two Conditions are superfluous: the phenomena they are meant to cover can more generally be explained by the concept of scalar implicatures. The pronoun in John likes himself receives a coreferential interpretation (by grammatical stipulation), but the pronoun in John likes him usually does not get a coreferential reading, not because of a grammatical rule but as a default implicature. The choice for him in the scale Q-implicates that the coreferential reading was not intended (if it was, the speaker should have said so). The default coreferential reading of nonreflexive pronouns in subordinate clauses (not accounted for in the generative account) is explained by the fact that the speaker does not have the same choice as in the former examples: in John told her that he gave her a valentine, he is most naturally taken to refer to John because himself is ungrammatical (a simple I-inference takes care of the coreference). Along similar lines, also the third Binding Condition ("lexical NPs must be free everywhere") is reduced to a pragmatic inference: if a reflexive pronoun might have been used, but instead a lexical NP is used, a Q-inference will induce a noncoreferential interpretation. If a reflexive could not have been used, however (in which case a pronoun would be the default option), the use of a lexical NP M-implicates disjoint reference.

Levinson also proposes a second version, in which only the second Binding Condition is accepted as a grammatical rule (or even only as a pragmatic rule called the "Disjoint Reference Presumption") and the other two are reduced to pragmatic inferences. Eventually, each account is associated with different languages and a synthesis between the two accounts is formulated, with a diachronic progression from so-called B-first languages (languages that are compatible with the second account) to A-first languages (languages compatible with the first account) is presented.


Levinson's book is an insightful survey of the various issues, problems and puzzles that are connected with one of the most famous concepts in linguistic pragmatics, the Gricean implicatures. The author not only critically evaluates different perspectives on the theory of implicatures, but also emphasizes one important aspect of Grice's original theory that has never been discussed so extensively: the generalized conversational implicatures. In constant, often implicit, but sometimes vehement discussion with rivaling accounts (mostly Relevance Theory), which deny the existence of an intermediate level between "once-off" inferences and "literal" or "conventional" meaning, he offers data that suggest the existence of such a level, thereby crucially relying on the difference between entailment and defeasible inference. The book is a bit repetitive at times, but at the same time it is an almost encyclopedic source of information for researchers of Gricean implicatures and the structure of linguistic communication in general. Also, it stimulates the discussion on the semantics-pragmatics interface, is cognizant of new evolutions in generative theories of meaning and formal semantics, and offers a few interesting suggestions as to how the Gricean tradition could be associated with recent work in information theory. Even though it relies heavily on the work of (especially) Larry Horn and others (including some earlier articles by the author himself), the book is original and thought-provoking for anyone interested in the theory of implicatures and the principles of linguistic communication in general.

Nevertheless, there are a number of problems with his perspective on generalized conversational implicatures . I will focus on what seems to me the most important one. In broad terms, the main problem with Levinson's account is that it relies too much on a compartmentalized view of meaning. Levinson criticizes rigidly modular approaches (Fodor 1983) a number of times in his book, but as far as the semantics/pragmatics interface is concerned, he still holds on to a rather rigid division between the two: "the distinction between semantics and pragmatics is one of a number of essential distinctions in the study of meaning [...] nothing is gained by lumping in our attempt to understand human communication" (p.7). Even though part of his book is an explicit account of how truth conditions of some utterances depend partly on conversational implicatures, Levinson contends that semantics and pragmatics are distinct domains, which operate on specific, domain-exclusive principles, even though they are "interleaved" (p.187). The problem with this view is that the status of these "semantic" meanings is unclear; moreover, the relationship between semantics and pragmatics rests on a few specific, but on the whole rather marginal linguistic phenomena.

In the third chapter, Levinson argues that implicatures contribute essentially to truth-conditions. If that is so, then "semantics" can no longer be equivalent with "truth-conditional meaning", or, at the very least, the domain of semantics will no longer be sufficient to determine the truth conditions. In his discussion of the position of "semantic retreat" (which consists of leaving all sorts of interpretation to pragmatics, except for the computation of so-called logical forms on the basis of the lexicon and the syntax), Levinson argues for "a new, strange level populated by semantic wraiths - a level of fragmentary structures, underspecification and half-information, even archi-sememes" (p.241). For the domain of truth conditions Levinson claims that "we need to find another level at which these insights hold, and it will be a level to which pragmatic information has contributed" (p.241), but although a number of schematic proposals figure in the book, it is not clear which model Levinson wants to defend. "Another level", seems to imply the existence of a third level, next to the semantic and the pragmatic level. But this statement amounts to no more than a suggestion, and the architecture is never fully worked out. This is very unfortunate, because in a certain respect it is the central question of the book.

On the other hand, in his critique of "pragmantics" (a position held by a number of Cognitive Linguists) he programmatically claims: "Pragmantics can be avoided if we can find a way of accounting for pragmatic intrusion into truth conditions while maintaining the modularity of a distinct pragmatics (built on nonmonotonic principles) and semantics (built on monotonic principles)" (p.243). Reasoning systems based on non-monotonic principles are defeasible, reasoning systems based on monotonic principles (as, e.g., deductive systems) are not. Here, the hypothetical third level seems be absent again, and instead the term "semantics" seems to receive yet another meaning. Semantics "might continue to operate on the basis of logical principles" (p.244). What could these principles be, if they cannot be relations like entailment and the traditional sense-relations (because these hold at a different, yet to be discovered level)? In other words, if semantics interleaves with pragmatic principles so that truth-conditions cannot be established without pragmatic intrusion, how can the semantic level "continue to operate on the basis of logical principles"? Or, alternatively, if the semantic level consists of underspecified half-information, how can it at the same time be based on truth-conditional principles? In a work that addresses a problem like the semantics/pragmatics interface, a clearer picture of what is meant by the terms semantics and pragmatics would be welcome.

But more important than these potentially confusing if not contradictory statements is the question why monotonicity should be so central that it necessitates the postulation of two distinct domains in the first place? More specifically, why should the defeasibility of so-called cancellation phrases be more relevant than the meanings of lexical items (e.g., quantifiers and other scalars) as they are actually used in contexts? The linguistic phenomenon that hides behind the term "non-monotonicity" is the possibility of nuancing or correcting statements that were made earlier by means of a number of specific phrases, among which the famous cancellation phrases (The man has three, and even four children) and suspension phrases (some, if not all people left the building). True, these corrections seem to involve specific lexical items (often characterized by their so-called "scalar" nature) and are only applicable under some special conditions (the suspension phrase is normally only used when the correction is in terms of an item higher up the scale - the utterance the man has three, if not two children is odd). But does the existence of these phenomena really justify a division of meaning into two separate parts, with all the confusion that this distinction brings along? Especially in a book that is partly devoted to showing how pragmatic principles, in this case Gricean implicatures, contribute to truth conditions, it is decidedly odd that semantics and pragmatics are still viewed as distinct levels.

From an empirical point of view, meanings can be more or less conventional, or more or less context-independent, not semantic or pragmatic. The so-called defeasibility of the alleged Gricean implicatures merely amounts to a specific way of nuancing the meaning of scalar items. The special conditions associated with these nuancing phrases are indeed caused by the relationship between certain items (scalarity), but there is no reason to assume that this relationship is more basic in any possible sense than the meanings that can be found in corpora. The fact that there is an understanding with the speakers of English that anything that is hot is also 'warm' gives rise to these if not-phrases. But that does not mean that the meaning of warm is 'at least warm, if not hot': meanings are not constructs that are retroactively devised on the basis of special properties that can be isolated in truth-conditional tests. Meanings, in any meaningful sense of the word (Putnam 1975) can be found in the usage of words. These empirical phenomena have to be the starting-point of any semantic theory: look at one thousand instances of three and interpret their meaning. In a second phase, the special properties (those that can be revealed in special constructions and tests) must then be integrated in a unified account. In fact, semantic theories that postulate the inverted hierarchy and start with these "special properties" can not even start a semantic analysis, due to lack of data. 'At least'-meanings of numerals can only be found in logical systems, hardly ever in ordinary language.

The cardinal numerals are actually the best example to demonstrate the terminological and conceptual-architectural confusion that is unavoidable if empirical data are disregarded to the advantage of a largely theoretical and highly debatable distinction between truth-conditional semantics and Gricean pragmatics. Larry Horn has actually admitted that his classical Gricean treatment of the cardinal numbers (Horn 1972, 1989) runs into trouble, and subsequently restricted his neo-Gricean approach to other scalar items (Horn 1992). Also Levinson argues that "the number words [are] not the correct testbed for the whole theory of scalar implicature" (p.90). Nevertheless, the numerals figure prominently in the evidence that Levinson presents for a number of claims. In his discussion of indexical resolution (p.178), ellipsis unpacking (p.183), the conditional (p.206), metalinguistic negation (p.213), negation in general (p.255), and Sag's model for pragmatic intrusion (p.247) the numerals are called upon to demonstrate certain allegedly Gricean phenomena. In this line of thinking, numerals like three have 'at least three' as literal, "semantic" meaning and the 'exactly three' interpretation is the result of a Gricean Q1-implicature.

This concrete example makes the question that was posed in general terms above even more acute: where does this 'at least'-meaning come from? Corpus data reveal that "bare numerals" (as in John has three children) always have 'exactly'-interpretations. It is not economical to explain this meaning via a scalar implicature that transforms the "actual" 'at least'-meaning into an 'exactly'-meaning, if this supposedly happens every time. If generalized conversational implicatures are so general that they occur every time, it is simpler to accept the 'exactly'-reading simply as the meaning of numerals. The data are overwhelmingly against 'at least'-meanings, to the extent that this 'at least'-meaning can only be expressed when the language user explicitly modifies the numeral by using the phrases at least three, three if not more etc. The monotonicity phenomena that are deemed so central can easily be explained as nuancing or correction phrases (Bultinck forthcoming).

Presumptive meanings is, I repeat, an excellent survey of (neo-)Gricean phenomena, analyses and theoretical assumptions. As a general hypothesis of how human communication works, it provides at the very least a plausible picture. Grice's conversational principles and maxims offer an intuitively appealing, if hardly falsifiable, account of how language users interpret and produce non- or less conventional utterances. But the distinction between semantics and pragmatics as two distinct levels of meaning is confusing and leads Griceans to postulate meanings that are not empirically verifiable, or only very marginally so. The postulation of these logic-inspired meanings leads to an uneconomical analysis. A unified perspective on meaning should replace the modular distinction between semantics and pragmatics: a meaning can be more or less conventionalized, but not semantic or pragmatic.



Bert Bultinck, PhD-candidate in linguistics, working with the Belgian Fund
For Scientific Research (Flanders) and affiliated to the University of
Antwerp, working on the concept of literal meaning in semantics and
pragmatics, Grice's legacy, and the semantics and pragmatics of numerals.