Transcript Attachment Ambiguity and Phrase Ordering

Corpus Studies of Constituent Ordering
Tom Wasow
An example, from Steven Pinker’s
The Language Instinct, p. 131:
In my laboratory we use it as an easily studied instance
of mental grammar, allowing us to document
 in great detail
 the psychology of linguistic rules
 from infancy to old age
 in both normal and neurologically impaired people,
 in much the same way that biologists focus on the fruit fly
Drosophila to study the machinery of the genes.
One of the other 119 possible orders:
In my laboratory we use it as an easily studied instance
of mental grammar, allowing us to document
 the psychology of linguistic rules
 in great detail
 in both normal and neurologically impaired people,
 from infancy to old age
 in much the same way that biologists focus on the fruit fly
Drosophila to study the machinery of the genes.
And another order
?? In my laboratory we use it as an easily studied
instance of mental grammar, allowing us to document
 in much the same way that biologists focus on the fruit fly
Drosophila to study the machinery of the genes
 in both normal and neurologically impaired people,
 in great detail
 the psychology of linguistic rules
 from infancy to old age
What makes some orders
sound more natural than others?
 The
answer might shed light on the
psychological processes underlying
language use.
 It might also have practical applications:
– for on-line style checkers
– for machine translation
– for other applications requiring robust
generation
The Alternations I Studied



Heavy Noun Phrase Shift:
– We take too many dubious idealizations for granted.
– We take for granted too many dubious idealizations.
The Verb-Particle Construction:
– We figured out the problem.
– We figured the problem out.
Dative Alternation:
– Kim handed a toy to the baby.
– Kim handed the baby a toy.
Factors I Looked At
 Structural
complexity (or “weight”)
 Discourse status (or “newness”)
 Semantic connectedness of verb and
following constituents
 Lexical biases of verbs
 Ambiguity avoidance
Grammatical Weight
 Behaghel’s
“Gesetz der Wachsenden Glieder”:
“Von zwei Gliedern von verschiedenem Umfang
steht das umfangreichere nach.”
Translation
Law of Growing Constituents:
Of two constituents of different size, the larger one
follows the smaller one
 In
other words:
Simple phrases precede complex ones.
Many Proposals to Make Behaghel’s
Generalization Precise
 Some
absolute, others relative
 Some categorical, others graded
 Corpus data support relative, graded
definition
 Various proposed measures are so highly
correlated that they can’t be distinguished
Categorial Weight Definitions
• An NP is heavy if it "dominates S”
[Ross (1967, rule 3.26)]
• "the condition on complex NP shift is that the NP dominate an S or a PP"
[Emonds (1976; 112)]
• "Counting a nominal group as heavy means either that two or more
nominal groups...are coordinated...., or that the head noun of a nominal
group is postmodified by a phrase or clause"
[Erdmann (1988; 328), emphasis in original]
• "the dislocated NP [in HNPS] is licensed when it contains at least two
phonological phrases"
[Zec and Inkelas (1990; 377)]
• "it is possible to formalize the intuition of 'heaviness' in terms of an aspect
of the meaning of the constituents involved, namely their givenness in the
discourse"
[Niv (1992; 3)]
Graded Weight Definitions
• Number of words dominated
[Hawkins (1990)]
• Number of nodes dominated
[Hawkins (1994)]
• Number of phrasal nodes (i.e. maximal
projections) dominated
[Rickford, et al (1995; 111)]
Numbers of Examples
V DO X
V X DO
TOTAL
HNPS
10,592
694
11,286
DAV-Prt
426
496
615
1,205
1,041
1,701
Testing Adequacy of Categorial
Definitions using HNPS
Testing Categorical Definitions as
Relative Criteria using HNPS
Weight Effects Increase Smoothly
Weights of Both Constituents Matter in HNPS
Weights of Both Constituents Matter in DA
The Overlap of the Weight Measures
More on Overlap of Weight Measures
More on Overlap of Weight Measures
Still More on Overlap of Weight Measures
Correlation Coefficients for
3 Weight Measures
HNPS
Words & Nodes
.94
Words & Phrasal Nodes .96
Nodes & Phrasal Nodes .94
DA
.96
.97
.96
V-Prt
.99
.95
.98
HNPS and Collocations
Two Verb Classes and HNPS
 Vt (for "transitive verbs") require NP objects in all their
subcategorizations: bring, carry, make, place, put, set,
take.
 Vp (for "prepositional verbs") can occur with NP
objects but also have uses with an immediately
following PP and no NP object: add, build, call, draw,
give, hold, leave, see, show, write.
Predictions
SPEAKER'S PERSPECTIVE
PERSPECTIVE
Vt
HNPS rare
HNPS relatively common
Vp
HNPS very rare
HNPS relatively common
LISTENER'S
Results from Brown Corpus
Results from Switchboard Corpus
Two Verb Classes and DA
Vs(for ”sentential verbs") may be followed by
an NP and that-clause or infinitval VP: offer,
show, teach, tell, write
 Vn (for ”non-sentential verbs") may not be
followed by an NP and that-clause or infinitval
VP :
assign, bring, give, hand, pay, send, take
Predictions
SPEAKER'S PERSPECTIVE
Vs double object relatively common
Vn double object relatively rare
common
LISTENER'S PERSPECTIVE
double object relatively rare
double object relatively
Corpus Results for DA Verb Classes
End of material on weight
Newness
 The
“Given-Before-New Principle”, as
formulated by Clark & Clark:
“Given information should appear before new
information.”
 Many variants in the literature.
Are weight and newness distinct effects?
 New
information requires more more words
to convey than old information (e.g.,
descriptions vs. pronouns)
 Is one of these factors just a side-effect of
the other?
 Surprisingly, nobody asked this question
until a few years ago.
Weight and newness are distinct.
 With
my students, I conducted corpus
analyses and a production experiment to
tease weight and newness apart.
 Both methods showed the two factors were
not reducible to one.
Weight vs. Newness in
Heavy NP Shift Corpus Study
Weight & Newness
Aren’t the Whole Story
“On this side of the Atlantic, the Lancaster-Oslo/Bergen
corpus was designed to replicate as closely as possible
the Brown corpus, the only difference being that this
corpus contains British rather than American English
texts.”
Judith Klavans, “Computational Linguistics,” in W.
O’Grady, M. Dobrovolsky, & F. Katamba,
Contemporary Linguistics: An Introduction
Another Factor: Semantic Connectedness
Behaghel again:
“das geistig eng Zusammengehörige auch eng
zusammengestellt wird”
Translation
What belongs together mentally is also
placed close together
Collocations and Idioms
 Idioms
(semantically opaque collocations):
– …bring pressure to bear
 Semantically
transparent collocations:
– …bring the meeting to an end
 Non-collocations:
– ...bring a pencil to the meeting
Heavy NP Shift and Semantic Connectedness
Dependent vs. Independent Particles
 Dependent:
They ate the cookies up.
– The meaning of “up” is dependent on the
meaning of “ate”, since the cookies don’t go up.
 Independent:
They picked the cookies up.
– The meaning of “up” is independent of the
meaning of “ate”, since the cookies go up.
Particle Position and
Semantic Connectedness
75.0%
V - OBJ - PRT
V - PRT - OBJ
50.0%
25.0%
0.0%
Independent Prt
Dependent Prt
Another Factor: Verb Bias
Possible Explanations for Factors
Influencing Order Variation
 Short
before long is easier to process,
because hard tasks are postponed.
 Given before new facilitates efficient
communication by establishing common
ground.
Possible Explanations for Factors
Influencing Order Variation (continued)
 Long
phrases and new information are hard
to produce and thus get postponed.
 Choices in word order allows speakers
flexibility in production.
 Our memory for words includes information
about what constructions they occur in and
how frequently.
Another Possible Factor:
Ambiguity Avoidance
 Global
ambiguity:
– I saw a man wearing an odd hat with a telescope.
– I saw with a telescope a man wearing an odd hat.
 Local
ambiguity:
– They gave Grant’s letters to Lincoln to a museum.
– They gave a museum Grant’s letters to Lincoln.
Corpus Study of Global Ambiguity and
Heavy NP Shifting
Corpus Search for Local Ambiguity
 Few
ambiguities of the relevant form (3)
– The company gave the U.S. rights to the drug to
the Population Council…
 More
unambiguous word orders (56)
– Giuliani gave the commissioner the ceremonial
key to the city…
 But
all unambiguous cases are also cases of
short-before-long.
Experimental Method
1. Speaker
silently reads a
sentence:
A museum received
Grant's letters to
Lincoln from the
foundation.
LISTENER
SPEAKER
Experimental Method
What did the
foundation do?
2. Sentence
disappears
from screen.
Listener
reads
question
from list.
LISTENER
SPEAKER
Experimental Method
3. Speaker answers
the listener’s
question.
The foundation gave .... the
museum, um, Grant's letter's
to Lincoln.
Listener chooses
the correct response
on list (from two
choices).
LISTENER
SPEAKER
Experimental Results on Local Ambiguity
100%
90%
80%
V-NP-NP
70%
60%
V-NP-PP bias
V-NP-NP bias
50%
40%
30%
20%
10%
0%
No potential
local ambiguity
Potential
local ambiguity
Implications of Experiment
 Phrase
ordering is not driven by an
ambiguity avoidance strategy
 Instead, it’s influenced by lexical
preferences and structural complexity
 There is a reverse ambiguity effect that
needs to be explained.
Why the reverse ambiguity effect?
A Conjecture
 Lexical/
Structural Priming -- the ambiguous
stimuli have the same PP in the NP as is
needed in the PP response:
– They received Grant’s letters to Lincoln from a
foundation.
– They received Grant’s letters about Lincoln from
a foundation.
Why don’t speakers use phrase ordering
to avoid ambiguity?
 Listeners
have a very high tolerance for
ambiguity
 Pragmatics rules out many ambiguities
 Prosody rules out others
 Still others just don’t matter
Conclusions
 A variety
of syntactic, semantic, and
discourse factors influence ordering.
 Some things you might think would
influence ordering don’t.
Questions for Future Research
 How
strong is each of the various factors
influencing ordering?
 Do ordering preferences in speech and
writing differ?
 Why are people so tolerant of ambiguity?