Neg-Raising — 1
There is one particular aspect to the question of scope that deserves special attention. This is the question of how to account for the worldwide phenomenon that, with certain verbs, a negation that is interpreted as part of the subordinate clause is found grammatically in the main clause, thereby giving the impression that it takes semantic scope over the whole sentence. I mean cases such as I don’t think she’ll win, with not in the position for negation over the whole sentence with think as main verb, whereas the sentence is normally understood as saying ‘I think she won’t win’, where not has scope only over the embedded complement clause. I will use the term Neg-Raising (NR) to denote this phenomenon, regardless of the fact that this term originated as one of the names (along with Negative Transportation, Negation Absorption and others) of a putative transformation in the framework of early transformational grammar. Traditional prescriptive grammars used to dictate that this way of speaking or writing should be avoided, as it was ‘illogical’. Yet the phenomenon is part of human language and thus needs an explanation.
What I will argue for in this post series is that NR is indeed a rule of syntax. This does not exclude a pragmatic motivation for the widespread use of NR constructions, conceivably based on the wish to speak in a modest manner (I think she won’t win is blunter than I don’t think she’ll win), but the tendency to let negation take the largest possible scope when the message has an overall negative purport is, in my view, a universal and functional trait of language, built into the human language system as it links up with cognition—regardless of the frequent, politeness-induced use of NR in sentences of the type I don’t think that … . In the light of the way linguistic theory has developed over the past half century, I also feel that if the concerted efforts, over the past forty years, by pragmaticists, by Chomskyan grammarians and even by formal semanticists to write NR out of the grammar-cum-semantics script and lodge it in pragmatics had been backed by more careful and more evenhanded arguments, it would have become clear much earlier that they were bound to fail. One cannot help but think that the efforts in question were motivated by the generally prevailing wish to debunk Generative Semantics, the most profound and most abstract of all linguistic theories.
After a few decades of relative rest, the question of how to account for NR phenomena has again come to the fore in a recent lengthy paper by Chris Collins and Paul Postal, ‘Classical NEG Raising’, which has been announced as a Linguistic Inquiry Monograph, to appear in May 2014. There is also a paper by Wim Klooster, ‘Negative Raising revisited’, in Germania et alia. A linguistic webschrift for Hans den Besten, edited by Jan Koster and Henk van Riemsdijk, 2003, available on the internet.
But let us start with the beginning. In early transformational grammar (TG), NR was taken to be a cyclic rule ‘promoting’ the negation standing over an object or subject complement clause to the position reserved for sentence negation, without, however, that becoming the ‘new’ interpretation of the sentence. As such, NR was a prototypical rule of Generative Semantics (Semantic Syntax or SeSyn), although it was already part of the repertory of rules in earlier forms of TG. A load of arguments was adduced to show the reality of NR as a syntactic rule.
An important and still widely quoted argument is found in Robin Lakoff’s 1969 paper ‘A syntactic argument for Negative Transporation’ (Papers of the Fifth Regional Meeting of the Chicago Linguistic Society, 1969, pp. 140–7, reprinted in P. A. M. Seuren ed., Semantic Syntax, OUP, 1974, pp. 175–82). Robin Lakoff observed was that a restricted class of negative polarity elements (NPIs) can occur in clauses under an NR verb such as think, from which the negation has been raised but not in clauses under a verb that does not allow for NR, like claim (the NPIs are italicised):
(1) a I don’t think that Jim will arrive until tomorrow.
— b. *I don’t claim that Jim will arrive until tomorrow.
(2) a. I don’t think Jim is in the least interested.
— b. *I don’t claim Jim is in the least interested.
Since in general NPIs like until or in the least require the negation to be in the same clause, the fact that the negation is, exceptionally, in a higher clause in (1a) and (2a) is evidence for NR. However, Robin Lakoff herself cast doubt on this argument by pointing at grammatical sentences like No-one thought that John would leave until tomorrow, which is semantically equivalent to Everyone thought that John would not leave until tomorrow. Her problem was that it was unclear how NR could transform the latter into the former.
I tried to resolve this question in my 1974 paper ‘Negative’s travels’ (in my edited little volume Semantic Syntax, OUP, 1974, pp. 183–208). The solution I proposed there was, for me, a dramatic confirmation of the Semantic Syntax approach to grammar and language. I observed that NR also occurs under a number of ‘universal’ predicates like modal verbs of necessity/obligation, the propositional operator and, the quantifier all, and the predicate cause, often but not always with a lexical switch from the ‘universal’ to its ‘particular’ counterpart, entailed by a subaltern entailment. A clear example without lexical switch is the French sentence (3a), where the negation ne…pas takes grammatical scope over the verb doit (must) but in whose semantic representation the negation takes scope only over the embedded clause, as overtly expressed in the less felicitous (3b):
(3) a. Ça ne doit pas être gai. (that mustn’t [be very nice])
— b. Ça doit ne pas être gai. (that must [not be very nice])
The position of ne…pas in (3a) shows that it is grammatically in construction with doit, manifesting an underlying structure where the negation takes scope over the verb devoir (must), whereas in (3b) it is grammatically in construction with être gai. It is hard to deny that this is a case of NR: the negation has been ‘promoted’ from the subordinate infinitival clause to the main finite clause. (The same is true for English mustn’t in the translation of (3a). Unsophisticated grammarians tend to say that must here takes grammatical scope over not, but this is false, since the suffixated form n’t results from a transposition of a higher not to suffix position, as in can’t or cannot, written as one word, which means ‘not possible’ and not ‘possible not’; see my Semantic Syntax, pp. 111–6.) (3a) is a case of NR without lexical switch. Cases with lexical switch are, for example, (4a–c):
(4) a. That can’t be true. from: [must [not [that be true]]]
— b. No-one left. from: [everybody x [not [x leave]]]
— c. He didn’t allow me to go. (from: [he cause [not [I go]]]
The inference that, at least in some cases, the negated particulars can’t, no-one and not allow derive from their nonnegated universal counterparts standing over a lower negation is based on the fact that the nonnegated versions That must not be true, Everybody did not leave and He made me not go are clearly less favoured than their NR-induced versions in (4). Apparently, ‘universal’ operators that have an entailed ‘particular’ counterpart have a tendency to induce NR.
This is also the solution I proposed in my 1974 article to Robin Lakoff’s problem with sentences like No-one thought that John would leave until tomorrow. Its derivation is now clear. From an underlying (5a) cyclic NR first gives (5b); from there, by repeated cyclic NR, (5c); from there, again by cyclic NR but with lexical switch from everyone to someone, we get (5d), and, finally, through the cyclic and postcyclic rules of English grammar, (5e):
(5) a. [everyone x [x think [until tomorrow [not [John leave]]]]]
— b. [everyone x [x think [not [until tomorrow [John leave]]]]]
— c. [everyone x [not [x think [until tomorrow [John leave]]]]]
— d. [not [someone x [x think [until tomorrow [John leave]]]]]
— e. No-one thought that John would leave until tomorrow.
The negation is thus seen to ‘travel’ upwards from being the lowest to being the highest operator in the SA of sentence (5e). This solution is only viable, of course, if prepositions, quantifiers and other operators are treated as semantic predicates, as proposed by McCawley in the late 1960s. Yet the problem with this solution, neat though it may be, is that for this to take place the normal cyclic procedures of the grammar must be suspended till NR has done its work cyclically through the whole structure. In particular, the negation not must be kept from being lowered into its argument-S as long as this Neg-Raising process is going on. This is one of the reasons why I later (in my Semantic Syntax of 1996, p. 114, note 4) proposed that grammars contain a precycle, reserved for certain ‘privileged’ cyclic rules that receive VIP treatment in that they are allowed to precede the normal syntactic cycle. Their function is to fashion SAs into a shape fit for further syntactic treatment. NR is then best taken to be a ‘VIP’ cyclic rule (possibly along with one of the forms of Conjunction Reduction) in that sense.
The treatment sketched above implies that until likewise induces NR, so that semantic until–not becomes syntactic not–until. And analogously for even: not even means ‘even not’ (Laurence R. Horn, A Natural History of Negation, Chicago University Press, 1989, p. 151). Both not until and not even can be preposed, but only with AUX-Inversion, as in:
(6) a. Not until five did he leave.
— b. Not even then did he leave.
This shows that AUX-Inversion is sensitive to syntactic negativity of the preposed adverbial, whereby the fact that in (6a,b) the negation semantically does not apply to until or even but to the main verb leave no longer counts. Moreover, one will note that English not even has counterparts in other languages with the two words inverted: même pas in French, sogar nicht in German, zelfs niet in Dutch, which means that in the other languages the counterpart of English even does not induce NR. (No amount of pragmatic reasoning will be able to account for such phenomena.)
A further corollary of this analysis is that the negation that is up for Raising must always be the syntactically highest operator of the embedded S in the precyclic cycle. It may also be semantically the highest, as in (7a), which can have the meaning of (7b) but not of (7c):
(7) a. I don’t suppose many members agree.
— b. ≈ I suppose not many members agree.
— c. ≉ I suppose many members do not agree.
This explains why (8a) lacks an NR reading and must mean literally ‘it is not the case that I think that she may make it on time’, since in She may not make it on time the highest operator is may, with not below may. By contrast, (8b) can be read as ‘I think that it is not possible that she makes it on time’, since can’t in She can’t make it on time represents ‘not possible’, with ‘not’ taking scope over ‘possible’:
(8) a. I don’t think she may make it on time.
— b. I don’t think she can make it on time.
A wealth of further evidence for NR, taken from a variety of languages, is provided in Larry Horn’s book-length article ‘Remarks on Neg-Raising’ in Pragmatics (= Syntax and Semantics, vol. 9), 1978, edited by Peter Cole, pp. 129–220. Ironically, however, Horn shows a preference for dismissing all this carefully gathered evidence in favour of NR and argues visibly towards a pragmatic explanation (about which more in the next post), only to conclude at the end that NR “lies at the heart of the intersection of syntax, semantics, and communicative intent” (p. 214).
I will, in the end, subscribe to this conclusion (though with a few qualifications), but without sharing Horn’s aversion to a syntactic, and his preference for a pragmatic, account. But before I come to that, I want to focus on Horn’s argument, presented on pp. 174-5 of his 1978 article, against my explanation of the grammaticality of (9c) below, with the plural verb form were, presented in my ‘Negative’s travels’ of 1974, pp. 204–5. My explanation rests on the observation that (9c) is grammatical and can only be interpreted as (9a). In my analysis, (9a) is first reduced by NR (lifting a unified ‘not’ over ‘and’, which is replaced with ‘or’, i.e. one half of De Morgan’s Laws) to (9b) and then, after repeated NR, to what ultimately becomes (9c). The crucial point is the plural were in (9c), which is remarkable since (9d) is ungrammatical. This plural is taken to derive from the semantic plurality inherent in (9a), where the predicate be late is assigned to two subject referents, Harry and Fred: we say Harry and Fred were late but Harry or Fred was late. This plural assignment, however, must be optional, since (9e) can also have the NR reading of (9a), besides the non-NR reading of (9f), where the negation stands over believe:
(9) a. I believe [and [[not [Harry be late], [not [Fred be late]]]]]
— b. ==> I believe [not [or [[either Harry be late], [Fred be late]]]]
— c. ==> I don’t believe that either Harry or Fred were late.
— d. *Either Harry or Fred were late.
— e. I don’t believe that either Harry or Fred was late.
— f. not [I believe [or [[Harry be late], [Fred be late]]]]
I’m afraid Horn misrepresented my argument, saying that, according to me, (9c) in the sense of (9a) “must take plural verb agreement” (Horn p. 174), whereas all I say is “Any syntax of English will have to explain why [9c] is grammatical although [9d] is not” (my p. 204). He then rejected my argument on the grounds that “I, for one, can interpret [9e] as corresponding to either [9a] or [9f], although [9c] does favor the former reading” (Horn p. 175). But that is precisely what I said in my article! Therefore, Horn’s conclusion (p. 175) that the data of (9c,d,e) “do not refute Seuren’s claim, based on the more restrictive dialect, […] they merely muddy the water” is not only ungenerous, it is also baseless. What “muddied the water” was Horn’s misrepresentation of my analysis and his not so relevant invocation of ‘dialectal’ differences. But perhaps we should put this down to the prevailing atmosphere of the day, where debunking Generative Semantics was the thing to do. … to be continued