Echo questions #390

nschneid opened this Issue Dec 30, 2016 · 10 comments


None yet

4 participants


Mark Steedman writes:

There is a class of "echo questions" discussed by Huddleston and Pullum, as in B's responses below:

A: I bought a Picasso!

B: You bought a PICASSO?
B': You bought a WHAT?
B'': You bought WHAT?
B''': You WHAT?
B'''': WHAT?

They are grammatically (and intonationally) distinct from real questions, like "did you buy a Picasso?" and "What did you buy?" and "What did you do?".

What relations should we use for "WHAT" in the above examples? Should we treat B' like a normal noun, B'' like a normal NP, B''' like a normal VP, and B'''' like a normal sentence—or is there a more elegant solution?

jnivre commented Dec 30, 2016

Syntactically, WHAT is "obj" in both B' and B'' and "root" in B''' and B''''. Morphologically, it is PRON as usual in all cases.


Mark again:

I don't see how "what" can be SYNTACTICALLY object in B': You bought a WHAT?

You can't extract it. You can't passivize it.

*A what did you buy?
*What did you buy a?
*Picasso was bought a (by Mary)

amir-zeldes commented Dec 30, 2016 edited

@nschneid I think the phrase is an object either way; those arguments about extraction and passivization without the article apply to common nouns too:

*Dog did you buy the?
The dog, did you buy? (probably only possible using Yiddish movement)
*Dog was bought a (by Mary)
A dog was bought (by Mary)

The object is the entire phrase [a what], which is headed by what (and the head has the function 'obj' as pointed out by @jnivre )

I think you can move the entire phrase, just like most NPs, and transform using a cleft etc.:

It was [a WHAT] that you bought?
The thing you bought was [a WHAT]?

The only open question I see is whether it's still to be tagged as PRON. In PTB it would still be WP, so I would be for PRON too. It really depends on your research question or application when searching for stuff like this, but my expectation is that when I look for PRON, I get all pronouns, which would include this case (it is a pronoun in so far as its denotation is completely contextual, because it stands in for a nonu). I could imagine a context in which I don't want it (e.g. I'm looking for all nominal phrase contexts that can/cannot license an article), but I think a robotic what -> PRON decision is more predictable and easier to enforce and validate.


Oh and for the non-NP cases I'd do:

What? root(what)
You what? root(what), nsubj(what,you)
You bought what? obj(bought,what)
You bought a what? obj(bought,what), det(what,a)


those arguments about extraction and passivization without the article apply to common nouns too

Yes, I think the more relevant observation is that it's odd to passivize the entire NP:

  • non-echo:
    A Picasso was bought (by Mary)
    The Picasso was bought (by Mary)
  • echo:
    ?A WHAT was bought (by Mary)
    ?The WHAT was bought (by Mary)

"A WHAT was bought?!" certainly sounds more marginal than "She bought a WHAT?!" But I don't know if the marginality reflects the syntactic relation or the semantic/discourse properties of echo questions.

In other words: I think echo questions are a distinct construction (with special semantics/pragmatics and intonation), but I don't know that it needs a special syntactic relation in the UD context.

nschneid commented Dec 31, 2016 edited

I cannot think of other constructions where a pronoun is preceeded by an article: *the it, *a something, *the who (rock band names excepted). This gives me pause about det(what, a). According to this query, there are no instances of who, whom, or what taking a determiner in the English UD treebank. The relationship between "a" and "what" almost feels like parataxis to me.


OK, the pronoun "one" can follow an article: definite determiners as in this query, and the expression "not a one". But it is not a WH-pronoun, and it is actually tagged as a noun in the treebank.


Aha—Huddleston & Pullum, p. 891: "In You need to buy a new what?, the echo-question word what is a common noun with a determiner and adjectival modifier as dependents: interrogative what, by contrast, is a pronoun, unable to take such dependents." I'd feel better about det(what, a) if "what" was tagged as a noun.


In UD v2 we do not change the UPOSTAG each time the function of a word changes. So what can stay PRON in UD even if Huddleston & Pullum call it a common noun in a specific construction.

jnivre commented Jan 1, 2017

I did not mean that "WHAT" by itself is the object in "You bought a WHAT". I only meant that the incoming dependency is labeled "obj". Thus:

nsubj(bought, You)
obj(bought, WHAT)
det(what, a)

But I think Amir made this clear in his comments.

@dan-zeman dan-zeman added this to the lg-specific v2 milestone Jan 4, 2017
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment