Computational
Linguistics
CSC 485 Summer 2020
5B
5b. Resolution of ambiguity
Gerald Penn
Department of Computer Science, University of Toronto
Copyright © 2017 Suzanne Stevenson, Graeme Hirst and Gerald Penn. All rights reserved.
Ambiguity resolution
• Problem of chart parsing: Structural ambiguity:
Time flies like an arrow.
… paint the office in the building near the research
center by the gym …
• Parser finds all possible parses.
2
Ambiguity resolution
• Chart parsing is founded on idea of exploring large space of ambiguities.
• Slow? Not that slow, given a typical median sentence length of < 40.
• There are simply too many parses on average – it’s proven too hard to write grammars for all and only the right readings.
• Too much work for semantics.
• Have to narrow down this potential.
• Possible solution: stop at first parse. • Problems?
3
Ambiguities and parsing
• Questions:
• Are structural ambiguities really a problem?
• If so, what kinds of ambiguities?
• Some real text:
In a general way such speculation is epistemologically relevant, as suggesting how organisms maturing and evolving in the physical environment we know might conceivably end up discoursing of abstract objects as we do. — Quine
W.V. Quine. “Speaking of objects.” Proceedings and Addresses of the American Philosophical Association, Vol. 31 (1957–1958), pp. 5–22. Quoted in: Steven Abney, “Statistical methods and linguistics.” In: Judith Klavans and Philip Resnik (eds.), The Balancing Act: Combining Symbolic and Statistical Approaches to Language. The MIT Press, Cambridge, MA. 1996.
4
Ambiguities and parsing: Example
S
PP S
P NP NP VP S
in Art Adj N Art N V AdjP Conj VP a gen.way suchspec. is Adv Adj as V
epistemologically relevant suggesting
S
Comp NP VP
how
organisms...
The sentence is unambiguous and we have found the parse. What is the problem?
5
Another example
In the usual way such people think, blithely ignorant as bleating sheep, politicians
fulminating and bloviating on their oversized TVs, Americans ignore evidence credibly
presented pointing out the results of their choices.
S
PP AdjP Abs
S
NP VP
P NP
AdvAdj P
In NP RCblithely
P politicians ...TVs
Art Adj N
the usual
ignorant
as bleating sheep
such people think
Americans ignore evidence ...
way
Adapted from Steve Abney’s example
6
Another example
In a general way such speculation is epistemologically relevant, as suggesting how organisms maturing and evolving in the physical environment we know might
conceivably end up ...
P P
P NP
in NP
Art Adj N
a way
S
AdjP
Abs
organisms maturing ... environ- ment
S
NP VP
we know might ...
AdvAdj P P
R
C epist’ly
relevant
general
ulation is
such spec-
as suggesting how
Adapted from Steve Abney’s example
7
Combinatorial explosions of parses
• Ordinary sentences can have hundreds of different parses due to combinatorial explosion (Church and Patil).
• That combinatorial explosion arises to a great extent from the fact that syntactic categories do not – arguably cannot – incorporate all of the real-world knowledge that we bring to bear on this problem.
10
Combinatorial explosions of parses
• More than 300 parses for 2% of sentences in corpus.
• E.g., 692 parses for:
For each plant give the ratio of 1973 to 1972 figures for each type of production cost and overhead cost.
11
Find the structural ambiguities
The 168-year-old Sunday tabloid will cease to exist after this week, Murdoch said today in an announce- ment to staff e-mailed to news organizations. ... Such has been the outcry over the phone hacking of everyday people during times of emotional turmoil that David Cameron’s government on Thursday postponed a decision on News Corp’s bid to purchase full control of BSkyB until September.
Bloomberg, 8 July 2011
12
At once too many and too few readings
Listen....
Bloomberg, 8 July 2011
13
At once too many and too few readings
Listen....
OK, robustness is important, but enough is enough.
Bloomberg, 8 July 2011
14
Global and local ambiguity
• Global ambiguity: A sentence has multiple interpretations.
I saw the man with the telescope. Time flies.
• See which interpretation(s) people prefer.
• Local ambiguity: Resolved by later input.
• The horse raced...
Mary expected the woman...
15
Syntactic sources of ambiguity 1
• Derived from PoS ambiguity: Time flies.
• Attachment of one phrase to another: examined the fingerprint with the microscope
the horse in the barn that the vet examined
learned that Nadia arrived on Sunday
He brought the car back {undamaged|undismayed}.
• Gap ambiguities:
the boys that the police debated about fighting
16
Syntactic sources of ambiguity 2
• Internal structure of a phrase:
winter boot sale
airport long term car park courtesy vehicle pickup point
• Alternative semantic role of subconstituent: The tourists objected to the guide that they couldn’t hear. I want the music box on the table.
17
What do people do? 1
• Look at human behaviour:
• Expected / preferred interpretations.
• Clues for successfully pruning parses.
• Some human strategies: ...
18
What do people do? 2
• Minimal attachment:
Prefer the simplest structure.
Karen knew the schedule ...
❶ [S [NP [PN Karen]] [VP [V knew [NP the schedule ...
❷ [S [NP [PN Karen]] [VP [V knew [S [NP the schedule
...
Karen knew the schedule {by heart | was wrong}.
W.D.Marslen-Wilson et al. Prosodic effects in minimal attachment. Quarterly Journal of Experimental Psychology, 45A(1), 73–87, 1992.
Fits ❶.
Requires ❷; hence need to back up; longer processing time.
20
What do people do? 3
• Recency (local/right association):
Associate new input with most recent part of the parse tree.
Karen met the mother of a singer who ...
❶ [NP the mother [PP [P of] [NP a singer [S who ... ❷ [NP the mother [PP [P of] [NP a singer]] [S who ...
• Notice that this might contradict minimal attachment. When?
21
What do people do? 4
• Lexical preferences:
Words (especially verbs) may have defaults for their containing nearby structures.
The tourists {objected | signalled} to the guide that they {couldn’t hear | didn’t like}.
❶ Prefer: AGENT object to PATIENT
(but AGENT object to PATIENT MESSAGE is also possible).
❷ Prefer: AGENT signal to PATIENT MESSAGE (but AGENT signal to PATIENT is also possible).
• Might contradict minimal attachment or recency.
22
PP attachment ambiguity
• Prepositional phrase attachment.
• An example problem that is a focus of much work
in disambiguation.
• A common ambiguity.
• A specific example of a very general type (modification ambiguity).
• Representative of properties of many types of ambiguities.
24
Why is PP attachment hard?
• Sometimes seems to require complex knowledge of the world:
Optical anisotropy of the copolyester melts can be determined by examination of the materials with the use of an optical microscope.
This is the first examination of the material with the impurity CVL in the region of deeply core shells.
The kinetic advantage arising upon using the NaH/Al
mixture to prepare the doped hydride was well reproduced in our examination of the materials variable dopant amounts and preparation conditions.
???
with
(1) Brewbaker, James L. and Marshall, William B. Liquid crystalline copolyesters of 4-hydroxybenzoic acid and substituted 4-hydroxybenzoic acids. U.S Patent 5268443. (2) V. B. Mikhailik. XEOL studies of impurity core-valence luminescence in mixed rubidium caesium chloride crystals. Journal of Physical Studies, 9 (2005) 182–184.
(3) P. Wang, X.D. Kanga, H.M. Cheng. Dependence of H-storage performance on preparation conditions in TiF3 doped NaAlH4. J of Alloys & Compounds, 2006, 217–22.
25
When is PP attachment easy (-ier)? 1
• Many unambiguous cases:
• The man with the telescope saw me.
The signals were analyzed with the oscilloscope.
• Sometimes syntax really should be able to say ‘no.’.
26
When is PP attachment easy (-ier)? 2
•
More often, syntax can say ‘probably not’:
• The preposition of rarely attaches to a transitive verb.
• Strong constraints on attaching PPs to pronouns and proper names.
He examined it with a microscope.
She examined John with a stethoscope.
But: I saw {John | him} with a hat. *{John | He} with a hat saw me.
Functioning
as an AdjP, not restrictive
27
Lexical preferences again
• Lexical preferences: Words (especially verbs) may have defaults for their containing or nearby structures — i.e., preferred disambiguation.
• Examples for PP attachment:
– Preposition p prefers to be attached to a verb.
– Verb v prefers PPs with preps p1 or p2 or nouns n1 or n2, but dislikes PPs with prep p3 or noun n3.
– When it’s the head of an NP in a PP, noun n1 prefers the PP to be attached to noun n2 or n3, or verb v1 or v2, if one of these is available.
28
Limitations of lexical preferences
• Preferences are only preferences:
• Might not be satisfiable.
• Might conflict.
• Might be overridden by coherence, plausibility.
• A given attachment problem might have no applicable preferences.
29
How to use lexical preferences?
• If a word w had some preferences ...
• How would we know what they are?
• How would we apply them in a parser?
30
Corpus-based attachment disambiguation
• Gather statistics for lexical usages from a corpus.
• That means the corpus gives the preferred structure
• That means that the corpus may have to be manually annotated by people (expensive)
• Use statistics to numerically estimate the parameters of a model.
• Apply model to new cases.
31
Corpora
• Corpus (pl. corpora): A large collection of text (or similar material).
• General or specialized content;
e.g., news, blog, technical, ESL, errors, ...
• May be (manually or automatically) annotated; e.g., with parse, meaning, correction, ...
32
Some important corpora
• Brown Corpus (1M words);
British National Corpus (100M words).
• Tagged with part of speech of each word.
• Wall Street Journal Corpus 1987–92 (80M
words).
• English Gigaword Corpus (~6B words).
• Penn Treebank (1.6M sentences of WSJ).
• Each with complete human-created parse tree.
• Canadian Hansard aligned French–English corpus.
33
Corpus statistics
• Can count linguistic phenomena in corpus. • E.g., count how many times a with-PP is noun-
attached or verb-attached in Penn Treebank.
• Problems:
• Sparse data — even with large corpora.
• Required information may not be explicit in corpus.
34
Corpora with Grammars
• Conventional view: Use these counts to estimate numerical parameters of prior, otherwise discrete-looking grammars.
• Avant-garde view: treat corpora themselves as a means of specifying the grammars.
• the phrase-structure rules are grounded in actual data.
• the phrase-structure rules are specified in context.
35
Statistical pattern recognition algorithms 1
• Use corpus statistics to train an algorithm — i.e., set parameters of an underlying model.
• Typically output is classification of input.
• E.g., classify (examine, the materials, with the microscope) as a V-attachment or NP-attachment situation.
• Given input = (V, NP, PP), should PP attach to V or NP?
36
Statistical pattern recognition algorithms 2
• Types of training:
• Supervised: Learn from data with known answers: From set of pairs {input, output}, learn to classify new inputs.
• Unsupervised: Given only inputs and maybe possible outcomes.
• In between:
Bootstrapping, minimally supervised.
37
A three-way partition of corpus data
• Training data.
• Development (validation, verification) data.
• To test successive versions of algorithm under development, to guide adjustments to approach.
• Test data.
• For testing of final version of algorithm. (No more tweaking allowed!)
38