I suspect that everything I say here is well known, but I hadn’t realised any of it until recently, and so it might be news to some readers as well. Most of the good ideas are from conversations with Will Starr and Ishani Maitra. The subject is sentences where ‘sure’ is followed by an embedded question. Here are three interesting properties about such constructions.
‘Sure + wh-‘ is a very strong NPI
What is your reaction to the following purported sentences?
(1) John is sure who killed JFK.
(2) John isn’t sure who killed JFK.
(3) Is John sure who killed JFK?
(4) Anyone who is sure who killed JFK is excessively credulous.
If you’re like me, (1) is clearly defective, (2) is perfectly well-formed, and (3) and (4) are fairly questionable. (Right now I think (3) isn’t quite as bad as (4), though I go back and forth on that.) If that’s right, ‘sure+wh-‘ is not only a negative polarity item, it is a very strong one, requiring something like overt negation to be fully licenced. Back in the day, there used to be several idiomatic constructions with this property. But over the years those phrases have either fallen out of usage (e.g., ‘give a red cent’) or have ceased to be NPIs (e.g., ‘give a damn’). I was surprised to discover that contemporary English has an NPI so strong as to prefer overt negation, and that it is not an idiom.
This came up because Will Starr has been doing some really fascinating work on the relationship between conditionals and questions. (I suspect some of that work will start getting a lot of attention soon, since it’s got the potential to revolutionise the way we think about conditionals; but that’s for another post.) One aspect of that connection is that ‘if’-clauses can be embedded questions. For instance, (5) means roughly the same thing as (6).
(5) Does Bob know if John will be at the party?
(6) Does Bob know whether John will be at the party?
But these two don’t quite mean the same thing, because ‘knows whether’ is not an NPI, while ‘knows if’ is. (Or at least it is in the idiolects of a few people I’ve spoken to.) Compare (7), (8), (9) and (10)
(7) Bob knows if John will be at the party.
(8) Bob knows whether John will be at the party.
(9) Bob doesn’t know if John will be at the party.
(10) Bob doesn’t know whether John will be at the party.
I think (7) is, every so slightly, worse than (8) through (10). So I think ‘knows if’ is a, very weak, NPI.
I said earlier that I was surprised that contemporary English contains as strong an NPI as ‘sure+wh-‘. I knew such strong NPIs used to exist, but I thought they had gone extinct in recent decades. But I don’t think I ever knew there was an NPI as weak as ‘knows if’. If I’ve classified it correctly, it’s a very strange creature. But I’m not sure whether I have classified it correctly.
Attitudes, Factivity and Embedding
One constraint on theories of the semantics of attitude verbs and embedded questions is that they have to explain why (11) is well-formed, while (12) isn’t.
(11) John knows who killed JFK.
(12) John believes who killed JFK.
In the past I’ve been most impressed by theories that explained this asymmetry in terms of the fact that ‘knows’ is factive while ‘believes’ is not. But that can’t be what’s central, since ‘sure’ isn’t factive, but ‘sure’ takes embedded questions as complements. Admittedly, ‘sure’ only takes embedded questions as complements in negated contexts, but still, if you’re classifying sentences, it seems (13) has to go with (14), and not with (15).
(13) John isn’t sure who killed JFK.
(14) John doesn’t know who killed JFK.
(15) John doesn’t believe who killed JFK.
Whatever explains why (15) doesn’t work, it can’t be (or at least it can’t just be) the non-factivity of ‘believes’. There must be something else going on, and I’m not sure what it could be.
Denotation of Embedded Questions
When we think about ‘knows+wh-‘ constructions, it is tempting to say that the denotation of an embedded question is the true answer to that question. After all, (11) is true iff there is a _p_ such that _p_ is the true answer to the question “Who killed JFK?” and John knows _p_. The simplest way to get that to turn out correct is that “who killed JFK” in (11) simply denotes the true answer to the question “Who killed JFK?”.
Obviously this theory has to be qualified to some extent to explain the fact that (12) and (15) are not well-formed. Thinking about ‘sure’ shows that it needs to be qualified even further.
Suppose that Bill is sure Fidel Castro killed JFK. And suppose that he’s wrong; it has Lee Harvey Oswald. Then (16) is intuitively false.
(16) Bill isn’t sure who killed JFK.
Bill is sure; he’s just wrong. But on the view in question, “who killed JFK” denotes the proposition that Lee Harvey Oswald killed JFK. Then (16) should be equivalent to (17).
(17) Bill isn’t sure that Lee Harvey Oswald killed JFK.
But (17) is true, Bill isn’t sure of this. To be sure, (17) is misleading, since it implicates that Bill takes Lee Harvey Oswald’s guilt to be a live option. But I think once we distinguish truth from assertability, it is clear that (17) is true. So the simple hypothesis about the denotation of “who killed JFK” can’t be right.
It seems like what we need to say is that the sentence _S Vs Q_, where _Vs_ is an attitude verb, and _Q_ is a question, is true iff for some _p_, _p_ is a (possibly false) answer to the question _Q_, and _S Vs p_. That’s why it’s false to say (16). There is an answer to the question “Who killed JFK?” that Bill is sure of. It’s a false answer, but it’s an answer. The reason that _S knows Q_ requires knowing the true answer isn’t something that follows from the semantics of embedded questions, but something that follows from the factivity of ‘knows’. But I’m not sure what kind of compositional theory could deliver these truth conditions.