Ling 596D Frames, Constructions Jan. 21, 2004

Heidi Harley

1.0 Fischer, Hall, Rakowitz and Gleitman

à A new experiment, pursuing Gleitman’s then decade-long investigation into the role syntactic frames play in restricting learner’s hypotheses about verb meaning

(1) Premise: while co-occurrence statistics, speaker attention and general heuristics about what’s an ‘object’ can guide a learner’s guesses about the meaning of novel nouns, that kind of general procedure is demonstrably inadequate for guesses about verb meaning.

(2) Background: Support for premise: ‘word-guessing’ experiments with adult subjects (Gillette & Gleitman).

à Subjects are played a video of a scene in which a mother is interacting with an infant/toddler, with no audio

à The points at which a certain word is uttered by the mother are marked by a ‘beep’ in the video

à Subject has to guess what word the ‘beep’ represents

à They are extremely good at this for nouns (almost at ceiling)

à They are very bad at this for verbs

(3) (Note: this goes with the general picture of ‘word-teaching’ incidents — nouns often occur in ostensive use ‘This is a ball’; verbs almost never do ‘This is hopping’. This despite the fact that full nouns are actually kind of rare in conversation, and in many languages usually a maximum of one noun per sentence is usual (Salish lgs, Gerdts, e.g.) Yet nouns are ‘taught’, while verbs aren’t. (One wonders if this is the pattern in languages other than English, particularly polysynthetic ones — I imagine so, actually).

(4) Consequence: Children learn nouns first, not many verbs until quite late (3 yrs old [HH: How true is this assertion?]). But when they start acquiring verbs, how do they go about it? The stimuli haven’t changed much...

(5) Hypothesis: ‘Sentence-to-world mapping’ rather than ‘word to world mapping’. Children use knowledge about the meanings associated with particular syntactic frames to narrow down their range of guesses about the content of a particular verb — and this syntactic-frame knowledge overrides otherwise strong non-linguistic heuristics (agent-first, etc.)

(6) Note that they still assume that knowledge *of the frames* must be learned/acquired before this can happen, indeed, that assumption is explanatory for them: ‘it takes time to acquire structural knowledge, and nouns but not verbs can be acquired efficiently in the absence of such knowledge’.

(7) Even allowing for a lot of heuristics that narrow down possible verb meanings (as for nouns), there’s still too big a hypothesis space. (Nice way of expressing the problem: ‘It is often and truly said that a picture is worth a thousand words, but that’s just the problem for verb frame acquisition.’

(8) Hypothesized procedure: identify nouns in sentence, note their grammatical positions (subject, obj, object of P). (Interesting: how do you identify ‘subject’ before you know structure? First noun you hear? Pronoun?) Verb learning can only occur within a prior learned vocabulary of noun meanings. “The structured sentence representation can help in acquiring the novel verb just because it is revealing of the argument-taking aspects of that verb’s interpretation.”

(9) Question really is, *can* verb meaning be learned at all in the absence of these structures?

à Methodology: show learner a scene of an action that is readily describable from two perspectives, each of which can be encoded as a different lexical verb in English. How does a child decide (without the movie director’s zoom lens) which facet of the event is being named by the verb?

à “If the learner considers the novel verb use within a syntactic structure, and requires an interpretation that is congruent both with the scene and the structure, there is a solution to the mapping problem for these verbs.”

à Otherwise, other heuristics would predict only one possible meaning for the verb without structure.

(10) Verb alone, frame A & frame B:

‘Look, biffing!’ ‘Look, ziking’

‘The rabbit is biffing the ball to the elephant’ ‘The rabbit is ziking the duck’

‘The elephant is biffing the ball from the rabbit’ ‘The duck is ziking’

(11) Causality bias: A scene depicting A feeding B is interpreted as involving ‘feeding’ rather than just ‘eating’. (Maybe a kind of Gricean effect here? Maximize descriptions when possible? ‘Understanding’ of an event (causally?) should be maximized?)

Materials: Videos

I: A rabbit is feeding an elephant with a spoon.

2: A rabbit comes up and pushes a monkey off a box.

3: A rabbit runs across the screen, followed by a skunk.

4: A monkey is riding piggy-back on a rabbit.

5: An elephant hands a ball to a rabbit.

6: A rabbit puts a blanket over a monkey.

Materials: Stimuli

The elephant is ---ing. (eat) 2NP

The bunny is ---ing the elephant. (feed) 1NP

The bunny is ---ing the monkey. (push) 2NP

The monkey is ---ing. (fall) 1NP

The bunny is ---ing the skunk. (flee) 2NP

The skunk is ---ing the bunny. (chase) 2NP (other order)

The monkey is ---ing the bunny. (ride) 2NP

The bunny is ---ing the monkey. (carry) 2NP (other order)

The elephant is ---ing the ball to thebunny. (give) 3NP

The bunny is ---ing the ball from the elephant. (take) 3NP (dif order of 1st and 3rd NPs

The bunny is ---ing the blanket onto the monkey. (put) 3NP

The bunny is ---ing the monkey with the blanket. (cover) 3NP (dif order of 2nd and 3rd NPs)

Findings:

a) with no syntactic frame, subjects were quite likely (1/3 time) to mention both possible meanings as interpretations for the verb (big hypothesis space!)

b) when syntactic frame provided, subjects essentially never provided both meanings as possible interpretations (syntax narrows hypothesis space)

c) kids were more likely to give appropriate meanings in appropriate contexts than otherwise, except for the last case, when they often said ‘put’ for the ‘cover’ frame—more often than they said ‘cover’. (‘blanket’ not an argument of ‘cover...’)

d) without syntactic frame, kids went with verb meaning consistent with agent-first syntax, where ‘agent’ is ‘most active is scene’ or ‘most in control of the scene’

(12) Possible objection: the kids were understanding the nonsense terms as synonyms for verbs whose meaning & syntax they already knew. If that’s true, they didn’t usually exploit it in their overt explanations of the meaning of the nonsense words. Children seemed to be using the principle of contrast, hence not simply ‘filling in the blank’, but rather devising a new meaning constrained by the situation and the syntax.

(13) Acquisition procedure:

segmentation à

world-word mapping for nouns à

‘partial structural representation’ à

annotation of transitive sentences via agency bias[1] à

extrapolation to hierarchical structure using ‘agent-is-subject’ cue

Then:

Once the phrase structure has been bootstrapped from PSR, the learner

can make this decision by inspecting the geometry of the tree to determine

which noun is sentence subject. If the plausible agent appears as subject with

the give/receive scene, then the situational and syntactic cues converge on

give. But if the plausible agent appears in nonsubject position then it is not

the agent, despite appearances.

(14) So kids decide to construct a verb meaning which is consistent with the non-prototypical agent-participant being the agent, based on the syntax of the sentence presented. (Note that the kid would have to have heard ‘enough’ good examples of agents-as-subjects to have comfortably identified ‘subject’ position without worrying about these confounding cases).

(15) Supplementary discussion:

à verb frame ranges for a given verb can also help

NP want NP (just looks transitive, could mean ‘take’, etc.)

NP want [to V...] (aha! must be mental state verb)

NP see NP (just looks transitive — could mean ‘touch’)

NP see [ if NP V...] (aha! must be mental state verb)

à note kids must be exploiting only positive evidence here...

V. of transfer V of mental state

3 NPs (give) CP complement (think)

V of mental transfer:

3NP frame AND CP complement frame (explain)

à Is the structural evidence in the input? (yes — 24 most-common verbs in infant-directed speech have unique conjunctions of syntactic frames)

à Is it used? (yes, at least, the possibility is there — infant-directed speech turned to jabberwocky speech, adults asked to divine verb meanings from frame sets (no videos or anything else); adults got 52% of the verbs right (compare 7% for the beep+video scenario, and 28% when nouns but no other syntactic clues are presented in conjunction with a video). Plus, the 48% wrong weren’t far wrong — they were semantically quite close, while the wrong guesses in the other test scenarios were way wrong.

à syntactic ambiguity not generally present in infant-directed speech

“Overall, our view is not that there are ‘verb classes’, each of which has

semantic components and (therefore) licenses certain structures. Rather we

suggest that verb frames have semantic implications (truth values), and verbs have meanings.”

à This is the crunch — one strong version of compositionality goes out the window here

2 A couple of notes on Pinker’s thoughts

In other words, inferring that tear means ‘tear’ from hearing paper and shreds is a kind of cognitive inference using knowledge of real-world contigencies, the same one that could be used to infer that tear means ‘tear’ when seeing paper being torn to shreds. It is not an example of learning a verb’s meaning from its syntactic properties, the process Gleitman is concerned with.

à that’s why Gleitman is using different frames that contain all the same nouns

Moreover, some of the information about how a verb is used in a sentence is based on universal features of semantics. For example, the sentence I am glipping apples could inform a learner that glip can’t mean ‘like’, because the progressive aspect marked on the verb is semantically incompatible with the stativity of liking. Here, too, one can learn something about a verb’s meaning from the sentence in which the verb is used, as opposed to the situation in which the verb is used, but the learning is driven by semantic information (in this example, that liking does not inherently involve changes over time), not syntactic information. [my emphasis --HH]

à This is a much more cogent point, and I think is at the core of the compositionality debate.

à What’s semantics and what’s syntax? The sort of syntactico-semantic feature that Pinker is pointing to here are things that are the objects of syntactic construction. Do ‘Tense’ or ‘Aspect’ exist independently of their semantic values? No.

à Yet Gleitman would have us believe that notions like ‘causality’ and ‘transfer’ are properties of frames, not of lexical items/syntacticosemantic features. This, though, is somewhat incoherent from the perspective of the strong compositionality hypothesis, I think.

à (amusing subnote: Pinker seems to want to accept a promissory account of inferral-from-mental-models over a fairly clearly worked out and at least provisionally supported account of inferral-from-syntactic-frame:

“But to support the alternative claim that verb subcategorization

information is crucial, it is necessary to show that no theory of inferring

communicative intent could ever be adequate, not that we currently don’t

have one that is fully worked out.”

à Shades of Piaget! Funny in an otherwise fairly well-reasoned paper...

Linking Rules: complicated versions of simple heuristics like ‘Agent first’.

The regularities that license this inference are what linguists

call linking rules (Carter 1988, Jackendoff 1987, 1990; Pinker 1989, Gropen et al. 1991a). For example, if A is a causal agent, A is the subject of a transitive verb.

à This idea is equally non-compositional — stipulative connections between syntactic form and meaning. For a theory like Jackendoff’s, semantic form is not computed from word meaning and sentence structure, but is mapped off of it by a series of heuristics, ‘linking rules’. In fact, the ‘frames’/construction grammar approach is very similar to the Jackendovian approach (presumably why he was there nodding approvingly at the construction grammar workshop at the LSA.)

The solution that I favor has to do with positing the equivalent of Tense/Aspect features at work introducing certain verbal arguments, their meanings compositionally calculated from the semantic content of those features/functions — vP, in particular.

This is not so much in the original spirit of Hale & Keyser, which I’ll get into next week, but more in the spirit of Harley, which I’ll also get into next week. :)

***Addenda****

What I think of as the central problem with constructions & linking frames can be highlighted by an email exchange I had after that workshop with Adele Goldberg:

Dear Adele ( & cc Ivan & Ray) --

Was thinking a little bit more about constructions & thought I'd drop a note to explain a bit more what I meant about “homophonous constructions”, now that I think *I* know what I meant. Ray’s examples with the various verb-particles, upon reflection, weren’t quite what I was getting at, since they tended to involve particular lexical items or particles.

If a construction is like a lexical item, an idiosyncratic form-meaning pair, then one and the same form ought to be able to link to two distinct and unrelated meanings, where in this case ‘form’ means ‘phrase structure’ or ‘constituent structure’ (probably oughta be 'homographic' rather than 'homophonous'). This should mean that you could take a set of lexical items, arrange them in a constituent structure to which the lexical items each contribute their normal semantics, and yet come out with two available different meanings. In the sort of framework I’m used to, this would be impossible, since the meanings of constituents are made up of meanings of the parts & the way they’re arranged, and so if the parts have the same meanings and are arranged in the same way, the whole schmozzle should compute the same meaning (compositionally).

That is, in construction grammar, there should be three sources for string-ambiguity: the usual lexical homophony or polysemy (“I saw the bank”) and string-vacuous structural differences (“I saw the man with the telescope”), and also constructional homophony (homography), where the ambiguity doesn’t come from the lexical items or from covertly distinct constituent structures, but from two different meanings being assigned to the same syntactic frame. (Ivan’s “Is he crazy?” vs. “Is he crazy!” strings are what made me think of it, although as I understand Ivan’s analysis of the exclamative vs. interrogative readings there they actually involve a difference in feature structures at some level, and so are not identical in the way I'm thinking.)

One kind of case that could be an example might be the two kinds of meanings associated with strings like ‘John struck the wall’, where the syntax seems to be the invariant (crudely, [NP [VP NP]]) and the individual lexical items also (none are getting an 'idiom'-ish reading), but there are both volitional and non-volitional readings available, with quite different semantics. In my universe, I have to ascribe these two readings either to an ambiguity in ‘strike’ or to different syntacticosemantic features being contributed from somewhere else, e.g. a ‘volitionality’ projection, or from ‘John’ (whether the NP is construed as volitional/intentional/animate etc.). But it seems to me that in the construction grammar universe, there’s another possibility: both the syntax and the words are the same, but the [NP [VP NP]] structure is listed twice in the inventory of constructions, once with a volitional meaning and once without. Does that make any sense?

yrs in curious construction-grammar naiveté, hh

Hi Heidi, Thanks for your note. I see now what you're asking. Interesting question.

You can get structurally homophonous constructions that are not truly (in our sense) formally identical, if stress is conventionally used to differentiate them. E.g.,

Truman DIED. vs TRUMAN died.

have different info. structure properties (the latter being "all-new" or "sentence focus"), and involve correspondingly different constructions. Or how about:

She called him a bus.

to mean she flagged him down a bus or she insulted him.

But that case does result from, I'd say, two different structures: Subj V Obj1 Obj2 OR Subj V Obj PRED.

If we're looking for cases in which the form is truly identical (in terms of grammatical relations and intonational properties and not just surface strings), then real homonymy is hard to find. How about this case though:

In Chinese, the same form "She stole him a shirt" can either (and equally well) mean she stole a shirt FOR him or she stole a shirt FROM him. I'd say we have there constructional polysemy, not real ambiguity, but the two possible interpretations do seem to come from the range of possible meanings of the construction, not from the word "stole" itself.

The case you mention is also possibility I suppose, but I'm not sure there's great motivation to say there's both an intentional transitive construction and an unintentional transitive construction instead of just saying that the construction is underspecified (and so is 'struck' so both interpretations are allowed in the case of that particular verb). It might be possible to do the same thing in the Chinese case, too, actually, assigning a "transfer" interpretation to the construction without specifying which direction the transfer goes.

It's an interesting question why the cases you ask about are so hard to find, but it seems to me that it might have something to do with how we decide to assign forms in the first place. In cases that have the right potential, we tend to ASSIGN them different formal properties, since the form is supposed to parallel the meaning (in both MP and in a lot of work in CxG). That is, I say that the second NP in the "call" sentence above is either an Obj2 or a PRED because it can be construed as either a theme argument (ditransitive) or as a predication (in the other construction).

Does that make sense?

Adele

---

[1] Polynesian VOS lgs are doable from this perspective...child would have to first notice that nouns depicting entities that (independently) look like proto-agents go at the end, then, using ‘agents are subjects’, build syntactic trees that put subjects at the end