Transcript EVALUATING MODELS OF PARAMETER SETTING - CUNY

EVALUATING MODELS OF
PARAMETER SETTING
Janet Dean Fodor
Graduate Center,
City University of New York
1
2
On behalf of CUNY-CoLAG
CUNY Computational Language Acquisition Group
With support from PSC-CUNY
William G. Sakas, co-director
Carrie Crowther
Atsu Inoue
Xuan-Nga Kam
Yukiko Koizumi
Eiji Nishimoto
Artur Niyazov
Iana Melnikova Pugach
Lisa Reisig-Ferrazzano
Iglika Stoyneshka-Raleva
Virginia Teller
Lidiya Tornyova
Erika Troseth
Tanya Viger
Sam Wagner
www.colag.cs.hunter.cuny.edu
3
Before we start…
Warning: I may skip some slides.
But not to hide them from you.
Every slide is at our website
www.colag.cs.hunter.cuny.edu
4
What we have done
 A factory for testing models of parameter
setting.
 UG + 13 parameter values → 3,072
languages (simplified but human-like).
 Sentences of a target language are the
input to a learning model.
 Is learning successful? How fast?
 Why?
5
Our Aims
 A psycho-computational model of
syntactic parameter setting.
 Psychologically realistic.
 Precisely specified.
 Compatible with linguistic theory.
 And… it must work!
6
Parameter setting as the solution
(1981)
 Avoids problems of rule-learning.
 Only 20 (or 200) facts to learn.
 Triggering is fast & automatic
= no linguistic computation is necessary.
 Accurate.
 BUT: This has never been modeled.
7
Parameter setting as the problem
(1990s)
R. Clark, and Gibson & Wexler have shown:
 P-setting is not labor-free, not always
successful. Because…
 The parameter interaction problem.
 The parametric ambiguity problem.
 Sentences do not tell which
parameter values generated them.
8
This evening…
Parameter setting:
 How severe are the problems?
 Why do they matter?
 How to escape them?
 Moving forward: from problems to
explorations.
9
Problem 1: Parameter interaction
 Even independent parameters interact in
derivations (Clark 1988,1992).
 Surface string reflects their combined effects.
 So one parameter may have no distinctive
isolatable effect on sentences.
= no trigger, no cue (cf. cue-based learner;
Lightfoot 1991; Dresher 1999)
 Parametric decoding is needed.
Must disentangle the interactions, to identify
which p-values a sentence requires.
10
Parametric decoding
Decoding is not instantaneous. It is hard work.
Because…
To know that a parameter value is necessary,
must test it in company of all other p-values.
So whole grammars must be tested against
the sentence. (Grammar-testing ≠ triggering!)
All grammars must be tested, to identify one
correct p-value. (exponential!)
11
Decoding
O3 Verb Subj O1[+WH] P Adv.
 This sets: no wh-movt, p-stranding, head
initial VP, V to I to C, no affix hopping, Cinitial, subj initial, no overt topic marking
 Doesn’t set: oblig topic, null subj, null topic
12
More decoding
Adv[+WH] P NOT Verb S KA.
 This sets everything except ±overt topic
marking.
Verb[+FIN].
 This sets nothing, not even +null subject.
13
Problem 2: Parametric ambiguity
 A sentence may belong to more than one
language.
 A p-ambiguous sentence doesn’t reveal the
target p-values (even if decoded).
 Learner must guess (= inaccurate)
or pass (= slow, + when? )
 How much p-ambiguity is there in natural
language? Not quantified; probably vast.
14
Scale of the problem (exponential)
 P-interaction and p-ambiguity are likely to
increase with the # of parameters.
20 parameters →
30 parameters →
40 parameters →
100 parameters →
220 grammars = over a million
230 grammars = over a billion
240 grammars = over a trillion
2100 grammars = ???
 How many parameters are there?
15
Learning models must scale up
 Testing all grammars against each input
sentence is clearly impossible.
 So research has turned to search methods:
how to sample and test the huge field of
grammars efficiently.
 Genetic algorithms (e.g., Clark 1992)
 Hill-climbing algorithms (e.g., Gibson &
Wexler’s TLA 1994)
16
Our approach
 Retain a central aspect of classic triggering:
Input sentences guide the learner
toward the p-values they need.
 Decode on-line; parsing routines do the work.
(They’re innate.)
 Parse the input sentence (just as adults
do, for comprehension) until it crashes.
 Then the parser draws on other p-values,
to find one that can patch the parse-tree.
17
Structural Triggers Learners (CUNY)
 STLs find one grammar for each sentence.
 More than that would require parallel
parsing, beyond human capacity.
 But the parser can tell on-line if there
is (possibly) more than one candidate.
 If so: guess, or pass (wait for unambig).
 Considers only real candidate grammars;
directed by what the parse-tree needs.
18
Summary so far…
 Structural triggers learners (STLs) retain an
important aspect of triggering (p-decoding).
 Compatible with current psycholinguistic
models of sentence processing.
 Hold promise of being efficient. (Home in on
target grammar, within human resource limits.)
 Now: Do they really work, in a domain
with realistic parametric ambiguity?
19
Evaluating learning models
 Do any models work?
 Reliably? Fast? Within human resources?
 Do decoding models work better than
domain-search (grammar-testing) models?
 Within decoding models, is guessing
better or worse than waiting?
20
Hope it works! If not…
 The challenge: What is UG good for?
 All that innate knowledge, only a few
facts to learn, but you can’t say how!
 Instead, one simple learning procedure:
 Adjust the weights in a neural network;
 Record statistics of co-occurrence frequencies.
 Nativist theories of human language are
vulnerable until some UG-based learner
is shown to perform well.
21
Non-UG-based learning
 Christiansen, M.H., Conway, C.M. and Curtin, S. (2000). A
connectionist single-mechanism account of rule-like behavior in
infancy. In Proceedings of 22nd Annual Conference of Cognitive
Science Society, 83-88. Mahwah, NJ: Lawrence Erlbaum.
 Culicover, P.W. and Nowak, A. (2003) A Dynamical Grammar.
Oxford, UK: Oxford University Press. Vol.Two of Foundations of
Syntax.
 Lewis, J.D. and Elman, J.L. (2002) Learnability and the statistical
structure of language: Poverty of stimulus arguments revisited. In B.
Skarabela et al. (eds) Proceedings of BUCLD 26, Somerville, Mass:
Cascadilla Press.
 Pereira, F. (2000) Formal Theory and Information theory: Together
again? Philosophical Transactions of the Royal Society, Series A
358, 1239-1253.
 Seidenberg, M.S., & MacDonald, M.C. (1999) A probabilistic
constraints approach to language acquisition and processing.
Cognitive Science 23, 569-588.
 Tomasello, M. (2003) Constructing a Language: A Usage-Based
Theory of Language Acquisition. Harvard University Press.
22
The CUNY simulation project
 We program learning algorithms proposed in the
literature. (12 so far)
 Run each one on a large domain of human-like
languages. 1,000 trials (1,000 ‘children’) each.
 Success rate: % of trials that identify target.
 Speed: average # of input sentences consumed
until learner has identified the target grammar.
 Reliability/speed: # of input sentences for 99%
of trials ( 99% of ‘children’) to attain the target.
 Subset Principle violations and one-step local
maxima excluded by fiat. (Explained below as necessary.)
23
Designing the language domain
 Realistically large, to test which models scale up well.
 As much like natural languages as possible.
 Except, input limited like child-directed speech.
 Sentences must have fully specified tree structure
(not just word strings), to test models like STL.
 Should reflect theoretically defensible linguistic
analyses (though simplified).
 Grammar format should allow rapid conversion into
the operations of an effective parsing device.
24
Language domains created
# params # langs
# sents
per lang
tree
structure
Gibson &
Wexler (1994)
3
8
12 or 18
Not fully
specified
Word order + V2
Bertolo et al.
(1997)
7
64
distinct
Many
Yes
G&W + V-raising +
degree-2
Kohl (1999)
12
2,304
Many
Partial
B et al. + scrambling
Sakas & Nishimoto (2002)
4
16
12-32
Yes
G&W + null subj/topic
3,072
168-1,420 Yes
S&N + Imp + wh-movt
+ piping + etc
Fodor, Melni- 13
kova & Troseth
(2002)
Language properties
25
Selection criteria for our domain
We have given priority to syntactic phenomena which:
 Occur in a high proportion of known natl langs;
 Occur often in speech directed to 2-3 year olds;
 Pose learning problems of theoretical interest;
 A focus of linguistic / psycholinguistic research;
 Syntactic analysis is broadly agreed on.
26
By these criteria
 Questions, imperatives.
 Negation, adverbs.
 Null subjects, verb movement.
 Prep-stranding, affix-hopping
(though not widespread!).
 Wh-movement, but no scrambling yet.
27
Not yet included
 No LF interface (cf. Villavicencio 2000)
 No ellipsis; no discourse contexts to license
fragments.
 No DP-internal structure; Case; agreement.
 No embedding (only degree-0).
 No feature checking as implementation of
movement parameters (Chomsky 1995ff.)
 No LCA / Anti-symmetry (Kayne 1994ff.)
28
Our 13 parameters (so far)
Parameter













Default
Subject Initial (SI)
Object Final (OF)
Complementizer Initial (CI)
V to I Movement (VtoI)
I to C Movement (of aux or verb) (ItoC)
Question Inversion (Qinv = I to C in questions only)
Affix Hopping (AH)
Obligatory Topic (vs. optional) (ObT)
Topic Marking (TM)
Wh-Movement obligatory (vs. none) (Wh-M)
Pied Piping (vs. preposition stranding) (PI)
Null Subject (NS)
Null Topic (NT)
yes
yes
initial
no
no
no
no
yes
no
no
piping
no
no
29
Parameters are not all independent
Constraints on P-value combinations:
 If [+ ObT] then [- NS].
(A topic-oriented language does not have null subjects.)
 If [- ObT] then [- NT].
(A subject-oriented language does not have null topics.)
 If [+ VtoI] then [- AH].
(If verbs raise to I, affix hopping does not occur.)
(This is why only 3,072 grammars, not 8,192.)
30
Input sentences
 Universal lexicon: S, Aux, O1, P, etc.
 Input is word strings only, no structure.
 Except, the learner knows all word
categories and all grammatical roles!
 Equivalent to some semantic bootstrapping; no prosodic bootstrapping (yet!)
31
Learning procedures
In all models tested (unless noted), learning is:
 Incremental = hypothesize a grammar
after each input. No memory for past input.
 Error-driven = if Gcurrent can parse the
sentence, retain it.
 Models differ in what the learner does when
Gcurrent fails = grammar change is needed.
32
The learning models: preview
 Learners that decode: STLs.
 Waiting (‘squeaky clean’)
 Guessing
 Grammar-testing learners
 Triggering Learning Algorithm (G&W)
 Variational Learner (Yang 2000)
 …plus benchmarks for comparison
 too powerful
 too weak
33
Learners that decode: STLs
 Strong STL: Parallel parse input sentence, find all
successful grammars. Adopt p-values they share.
(A useful benchmark, not a psychological model.)
 Waiting STL: Serial parse. Note any choice-point in
the parse. Set no parameters after a choice.
(Never guesses. Needs fully unambig triggers.)
(Fodor 1998a)
 Guessing STLs: Serial. At a choice-point, guess.
(Can learn from p-ambiguous input.) (Fodor 1998b)
34
Guessing STLs’ guessing principles
If there is more than one new p-value that
could patch the parse tree…
 Any Parse: Pick at random.
 Minimal Connections: Pick the p-value
that gives the simplest tree. ( MA + LC)
 Least Null Terminals: Pick the parse
with the fewest empty categories. ( MCP)
 Nearest Grammar: Pick the grammar that
differs least from Gcurrent.
35
Grammar-testing: TLA
 Error-driven random: Adopt any grammar.
(Another baseline; not a psychological model.)
 TLA (Gibson & Wexler, 1994): Change any one
parameter. Try the new grammar on the sentence.
Adopt it if the parse succeeds. Else pass.
 Non-greedy TLA (Berwick & Niyogi, 1996):
Change any one parameter. Adopt it. (No test of
new grammar against the sentence.)
 Non-SVC TLA (B&N 96): Try any grammar other
than Gcurrent. Adopt it if the parse succeeds.
36
Grammar-testing models with memory
 Variational Learner (Yang 2000,2002) has
memory for success / failure of p-values.
 A p-value is:
 rewarded if in a grammar that parsed an input;
 punished if in a grammar that failed.
 Reinforcement is approximate, because
of interaction. A good p-value in a bad
grammar is punished, and vice versa.
37
With memory: Error-driven VL
 Yang’s VL is not error-driven. It chooses pvalues with probability proportional to their
current success weights. So it occasionally
tries out unlikely p-values.
 Error-driven VL (Sakas & Nishimoto, 2002)
Like Yang’s original, but:
First, set each parameter to its currently
more successful value. Only if that fails,
pick a different grammar as above.
38
Previous simulation results

TLA is slower than error-driven random on the G&W
domain, even when it succeeds (Berwick & Niyogi 1996).

TLA sometimes performs better, e.g., in strongly smooth
domains (Sakas 2000, 2003).
TLA fails on 3 of G&W’s 8 languages, and on 95.4% of
Kohl’s 2,304 languages.




There is no default grammar that can avoid TLA learning
failures. The best starting grammar succeeds only 43%
(Kohl 1999).
Some TLA-unlearnable languages are quite natural, e.g.,
Swedish-type settings (Kohl 1999).
Waiting-STL is paralyzed by weakly equivalent grammars
(Bertolo et al. 1997).
39
Data by learning model
% failure
rate
# inputs
(99% of trials)
# inputs
(average)
0
16,663
3,589
88
16,990
961
TLA w/o Greediness
0
19,181
4,110
TLA without SVC
0
67,896
11,273
Strong STL
74
170
26
Waiting STL
75
176
28
Any parse
0
1,486
166
Minimal Connections
0
1,923
197
Least Null Terminals
0
1,412
160
80
180
30
Algorithm
Error-driven random
TLA original
Guessing STLs
Nearest Grammar
40
Summary of performance
 Not all models scale up well.
 ‘Squeaky-clean’ models (Strong / Waiting STL)
fail often. Need unambiguous triggers.
 Decoding models which guess are most efficient.
 On-line parsing strategies make good learning
strategies. (?)
 Even with decoding, conservative domain search
fails often (Nearest Grammar STL).
 Thus: Learning-by-parsing fulfills its promise.
Psychologically natural ‘triggering’ is efficient.
41
Now that we have a workable model…
 Use it to investigate questions of interest:
 Are some languages easier than others?
 Do default starting p-values help?
 Does overt morphological marking facilitate
syntax learning?
 etc…..
 Compare with psycholinguistic data, where
possible. This tests the model further, and
may offer guidelines for real-life studies. 42
Are some languages easier?
Guessing STL- MC
‘Japanese’

# inputs
(99% of trials)
# inputs
(average)
87
21
‘French’
99
22
‘German’
727
147
1,549
357
‘English’

43
What makes a language easier?
 Language difficulty is not predicted by how
many of the target p-settings are defaults.
 Probably what matters is parametric
ambiguity
 Overlap with neighboring languages
 Lack of almost-unambiguous triggers
 Are non-attested languages the difficult
ones? (Kohl, 1999: explanatory!)
44
Sensitivity to input properties
 How does the informativeness of the input
affect learning rate?
 Theoretical interest: To what extent can
UG-based p-setting be input-paced?
 If an input-pacing profile does not match
child learners, that could suggest biological
timing (e.g., maturation).
45
Some input properties
 Morphological marking of syntactic features:
 Case
 Agreement
 Finiteness
 The target language may not provide them.
Or the learner may not know them.
 Do they speed up learning?
Or just create more work?
46
Input properties, cont’d
For real children, it is likely that:
 Semantics / discourse pragmatics signals
illocutionary force:
[ILLOC DEC], [ILLOC Q] or [ILLOC IMP]
 Semantics and/or syntactic context reveals
SUBCAT (argument structure) of verbs.
 Prosody reveals some phrase boundaries
(as well as providing illocutionary cues).
47
Making finiteness audible
 [+/-FIN] distinguishes Imperatives from
Declaratives. (So does [ILLOC], but it’s inaudible.)
 Imperatives have null subject. E.g., Verb O1.
 A child who interprets an IMP input as a
DEC could mis-set [+NS] for a [-NS] lang.
 Does learning become faster / more accurate
when [+/-FIN] is audible? No. Why not?
 Because Subset Principle requires learner to
parse IMP/DEC ambiguous sentences as IMP.
48
Providing semantic info: ILLOC
 Suppose real children know whether an input
is Imperative, Declarative or Question.
 This is relevant to [+ItoC] vs. [+Qinv].
( [+Qinv]  [+ItoC] only in questions )
 Does learning become faster / more accurate
when [ILLOC] is audible? No. It’s slower!
 Because it’s just one more thing to learn.
 Without ILLOC, a learner could get all
word strings right, but their ILLOCs and pvalues all wrong – and count as successful.
49
Providing SUBCAT information
 Suppose real children can bootstrap verb
argument structure from meaning / local context.
 This can reveal when an argument is missing.
How can O1, O2 or PP be missing? Only by [+NT].
 If [+NT] then also [+ObT] and [-NS] (in our UG).
 Does learning become faster / more accurate
when learners know SUBCAT? Yes. Why?
 SP doesn’t choose between no-topic and nulltopic. Other triggers are rare. So triggers for
50
[+NT] are useful.
Enriching the input: Summary
 Richer input is good if it helps with something that
must be learned anyway (& other cues are scarce).
 It hinders if it creates a distinction that otherwise
could have been ignored. (cf. Wexler & Culicover 1980)
 Outcomes depend on properties of this domain,
but it can be tailored to the issue at hand.
 The ultimate interest is the light these data shed
on real language acquisition.
 We can provide profiles of UG-based / input(in)sensitive learning, for comparison with children.
 The outcomes are never quite as anticipated.
51
This is just the beginning…
 Next on the agenda 
52
Next steps ~ input properties
 How much damage from noisy input?
E.g., 1 sentence in 5 / 10 / 100 not from target language.
 How much facilitation from ‘starting small’?
E.g., Probability of occurrence inversely proportional to
sentence length.
 How much facilitation (or not) from the exact
mix of sentences in child-directed speech?
(cf. Newport, 1977; Yang, 2002)
53
Next steps ~ learning models
 Add connectionist and statistical learners.
 Add our favorite STL (= Parse Naturally),
with MA, MCP etc. and a p-value ‘lexicon’.
(Fodor 1998b)
 Implement the ambiguity / irrelevance
distinction, important to Waiting-STL.
 Evaluate models for realistic sequence
of setting parameters. (Time course data)
 Your request here 
54
www.colag.cs.hunter.cuny.edu
www.colag.cs.hunter.cuny.edu
The end
55
REFERENCES
Bertolo, S., Broihier, K., Gibson, E., and Wexler, K. (1997) Cue-based learners in parametric language systems:
Application of general results to a recently proposed learning algorithm based on unambiguous
'superparsing'. In M. G. Shafto and P. Langley (eds.) 19th Annual Conference of the Cognitive Science
Society, Lawrence Erlbaum Associates, Mahwah, NJ.
Berwick, R.C. and Niyogi, P. (1996) Learning from Triggers. Linguistic Inquiry, 27(2), 605-622.
Chomsky, N. (1995) The Minimalist Program. Cambridge MA: MIT Press.
Clark, R. (1988) On the relationship between the input data and parameter setting. NELS 19, 48-62.
Clark, R. (1992) The selection of syntactic knowledge, Language Acquisition 2(2), 83-149.
Dresher, E. (1999) Charting the learning path: Cues to parameter setting. Linguistic Inquiry 30.1, 27-67.
Fodor, J D. (1998a) Unambiguous triggers, Linguistic Inquiry 29.1, 1-36.
Fodor, J.D. (1998b) Parsing to learn. Journal of Psycholinguistic Research 27.3, 339-374.
Fodor, J.D., I. Melnikova and E. Troseth (2002) A structurally defined language domain for testing syntax
acquisition models, CUNY-CoLAG Working Paper #1.
Gibson, E. and Wexler, K. (1994) Triggers. Linguistic Inquiry 25, 407-454.
Kayne, R.S. (1994) The Antisymmetry of Syntax. Cambridge MA: MIT Press.
Kohl, K.T. (1999) An Analysis of Finite Parameter Learning in Linguistic Spaces. Master’s Thesis, MIT.
Lightfoot, D. (1991) How to set parameters: Arguments from Language Change. Cambridge, MA: MIT Press.
Sakas, W.G. (2000) Ambiguity and the Computational Feasibility of Syntax Acquisition, PhD Dissertation, City
University of New York.
Sakas, W.G. and Fodor, J.D. (2001). The Structural Triggers Learner. In S. Bertolo (ed.) Language Acquisition and
Learnability. Cambridge, UK: Cambridge University Press.
Sakas, W.G. and Nishimoto, E. (2002). Search, Structure or Heuristics? A comparative study of memoryless
algorithms for syntax acquisition. 24th Annual Conference of the Cognitive Science Society. Hillsdale, NJ:
Lawrence Erlbaum Associates.
Yang, C.D. (2000) Knowledge and Learning in Natural Language. Doctoral dissertation, MIT.
Yang, C.D. (2002) Knowledge and Learning in Natural Language. Oxford University Press.
Villavicencio, A. (2000) The use of default unification in a system of lexical types. Paper presented at the
Workshop on Linguistic Theory and Grammar Implementation, Birmingham,UK.
56
Wexler, K. and Culicover, P. (1980) Formal Principles of Language Acquisition. Cambridge MA: MIT Press.
Something we can’t do: production
What do learners say when they don’t know?
 Sentences in Gcurrent, but not in Gtarget.
Do these sound like baby-talk?
Me has Mary not kissed why?
(early)
Whom must not take candy from? (later)
 Sentences in Gtarget but not in Gcurrent.
Goblins Jim gives apples to.
57
CHILD-DIRECTED SPEECH
STATISTICS FROM THE CHILDES DATABASE






The current domain of 13 parameters is almost as
much as it’s feasible to work with – maybe we can
eventually push it up to 20.
Each ‘language’ in the domain has only the
properties assigned to it by the 13 parameters.
Painful decisions – what to include? what to omit?
To decide, we consult adult speech to children in
CHILDES transcripts. Child age approx 1½ to 2½
years (earliest produced syntax).
Child’s MLU very approx 2. Adults’ MLU from 2.5 to 5.
So far: English, French, German, Italian, Japanese.
(Child Language Data Exchange System, MacWhinney 1995)
58
STATISTICS ON CHILD-DIRECTED SPEECH FROM THE CHILDES DATABASE
ENGLISH
GERMAN
ITALIAN
JAPANESE
RUSSIAN
Name+Age (Y;M.D)
Eve 1;8-9.0
Nicole 1;8.15
Martina 1;8.2
Jun (2;2.5-25)
Varvara 1;6.5 -1;7.13
File Name
eve05-06.cha
nicole.cha
mart03,08.cha
jun041-044.cha
varv01-02.cha
Researcher/Childes folder name
BROWN
WAGNER
CALAMBRONE
ISHII
PROTASSOVA
Number of Adults
4,3
2
2,2
1
4
MLU Child
2.13
2.17
1.94
1.606
2.8
MLU Adults (avg. of all)
3.72
4.56
5.1
2.454
3.8
Total Utterances (incl. Frags.)
1304
Usable Utterances/Fragments
USABLES (% of all utterances)
1107
806/498
62%
1258
728/379
66%
1691
929/329
74%
1008
1113/578
66%
727/276
72%
DECLARATIVES
40%
42%
27%
25%
34%
DEICTIC DECLARATIVES
8%
6%
3%
8%
7%
MORPHO-SYNTACTIC QUESTIONS
10%
12%
0%
18%
2%
PROSODY-ONLY QUESTIONS
7%
5%
15%
14%
5%
WH-QUESTIONS
22%
8%
27%
15%
34%
IMPERATIVES
13%
27%
24%
11%
11%
EXCLAMATIONS
0%
0%
1%
3%
0%
LET'S CONSTRUCTIONS
0%
0%
2%
4%
2%
59
FRAGMENTS (% of all utterances)
38%
34%
26%
34%
27%
NP FRAGMENTS
25%
24%
37%
10%
35%
VP FRAGMENTS
8%
7%
6%
1%
8%
AP FRAGMENTS
4%
3%
16%
1%
7%
PP FRAGMENTS
9%
4%
5%
1%
3%
WH-FRAGMENTS
10%
2%
10%
2%
6%
OTHER (E.g. stock expressions yes, huh)
44%
60%
26%
85%
41%
COMPLEX NPs (not from fragments)
Total Number of Complex NPs
Approx 1 per 'n' utterances
140
55
88
58
105
6
13
11
19
7
NP with one ADJ
91
36
27
38
54
NP with two ADJ
7
1
2
0
4
NP with a PP
20
3
15
14
18
NP with possessive ADJ
22
7
0
0
4
NP modified by AdvP
0
0
31
1
6
NP with relative clause
0
8
13
5
5
60
DEGREE-n Utterances
DEGREE 0
88%
84%
81%
94%
77%
Degree 0 deictic (E.g. that's a duck)
8%
6%
2%
8%
18%
Degree 0 (all others)
92%
94%
98%
92%
82%
DEGREE 1+
12%
16%
19%
6%
33%
infinitival complement clause
36%
1%
31%
2%
30%
finite complement clause
12%
1%
40%
10%
26%
relative clause
10%
16%
12%
8%
3%
coordinating clause
30%
59%
9%
10%
41%
adverbial clause
11%
18%
7%
80%
0%
ANIMACY and CASE
% Utt. with +animate somewhere
62%
60%
37%
8%
31%
% Subjects (overt)
94%
91%
56%
63%
97%
% Objects (overt)
18%
23%
44%
23%
14%
Case-marked NPs
238
439
282
100
949
Nominative
191
283
36
45
552
Accusative
47
79
189
4
196
Dative
0
77
57
4
35
Genitive
0
0
0
14
98
Topic
0
0
0
34
0
Instrumental and Prepositional
0
0
0
0
68
Subject drop
0
26
379
740
124
Object drop
0
4
0
125
37
Negation Occurrences
62
73
43
72
71
Nominal
5
19
2
0
8
Sentential
57
54
41
72
63
61
www.colag.cs.hunter.cuny.edu
62