Transcript ppt

Unambiguous + Unlimited = Unsupervised
Using the Web for
Natural Language Processing Problems
Marti Hearst
School of Information, UC Berkeley
Joint work with Preslav Nakov
BYU CS Colloquium, Dec 6, 2007
This research supported in part by NSF DBI-0317510
Natural Language Processing
 The ultimate goal: write programs that read and
understand stories and conversations.
 This is too hard! Instead we tackle sub-problems.
 There have been notable successes lately:
 Machine translation is vastly improved
 Speech recognition is decent in limited circumstances
 Text categorization works with some accuracy
Marti Hearst, BYU CS 2007
How can a machine understand these differences?
Get the cat with the gloves.
Marti Hearst, BYU CS 2007
How can a machine understand these differences?
Get the sock
from the cat
with the gloves.
Get the glove
from the cat
with the socks.
Marti Hearst, BYU CS 2007
How can a machine understand these
differences?




Decorate the cake with the frosting.
Decorate the cake with the kids.
Throw out the cake with the frosting.
Throw out the cake with the kids.
Marti Hearst, BYU CS 2007
Why is this difficult?
 Same syntactic structure, different meanings.
 Natural language processing algorithms have to
deal with the specifics of individual words.
 Enormous vocabulary sizes.
 The average English speaker’s vocabulary is around
50,000 words,
 Many of these can be combined with many others,
 And they mean different things when they do!
Marti Hearst, BYU CS 2007
How to tackle this problem?
 The field was stuck for quite some time.
 Hand-enter all semantic concepts and relations
 A new approach started around 1990
 Get large text collections
 Compute statistics over the words in those collections
 There are many different algorithms.
Marti Hearst, BYU CS 2007
Size Matters
Recent realization: bigger is better than smarter!
Banko and Brill ’01: “Scaling to Very, Very Large
Corpora for Natural Language Disambiguation”, ACL
Marti Hearst, BYU CS 2007
Example Problem
 Grammar checker example:
Which word to use?
<principal> <principle>
 Solution: use well-edited text and look at which
words surround each use:
 I am in my third year as the principal of Anamosa High
School.
 School-principal transfers caused some upset.
 This is a simple formulation of the quantum mechanical
uncertainty principle.
 Power without principle is barren, but principle without
power is futile. (Tony Blair)
Marti Hearst, BYU CS 2007
Using Very, Very Large Corpora
 Keep track of which words are the neighbors of each
spelling in well-edited text, e.g.:
 Principal: “high school”
 Principle: “rule”
 At grammar-check time, choose the spelling best
predicted by the surrounding words.
 Surprising results:
 Log-linear improvement even to a billion words!
 Getting more data is better than fine-tuning algorithms!
Marti Hearst, BYU CS 2007
The Effects of LARGE Datasets
 From Banko & Brill ‘01
Marti Hearst, BYU CS 2007
How to Extend this Idea?
 This is an exciting result …
 BUT relies on having huge amounts of text
that has been appropriately annotated!
Marti Hearst, BYU CS 2007
How to Avoid Manual Labeling?
 “Web as a baseline” (Lapata & Keller 04,05)
 Main idea: apply web-determined counts to
every problem imaginable.
 Example: for t in {<principal> <principle>}
 Compute f(w-1, t, w+1)
 The largest count wins
Marti Hearst, BYU CS 2007
Web as a Baseline
 Works very well in some cases





machine translation candidate selection
article generation
noun compound interpretation
noun compound bracketing
adjective ordering
Significantly better than the
best supervised algorithm.
Not significantly different
from the best supervised.
 But lacking in others
 spelling correction
 countability detection
 prepositional phrase attachment
 How to push this idea further?
Marti Hearst, BYU CS 2007
Using Unambiguous Cases
 The trick: look for unambiguous cases to start
 Use these to improve the results beyond what cooccurrence statistics indicate.
 An Early Example:
 Hindle and Rooth, “Structural Ambiguity and Lexical
Relations”, ACL ’90, Comp Ling’93
 Problem: Prepositional Phrase attachment
 I eat/v spaghetti/n1 with/p a fork/n2.
 I eat/v spaghetti/n1 with/p sauce/n2.
 Question: does n2 attach to v or to n1?
Marti Hearst, BYU CS 2007
Using Unambiguous Cases
 How to do this with unlabeled data?
 First try:
 Parse some text into phrase structure
 Then compute certain co-occurrences
f(v, n1, p) f(n1, p)
f(v, n1)
 Problem: results not accurate enough
 The trick: look for unambiguous cases:
 Spaghetti with sauce is delicious. (pre-verbal)
 I eat with a fork.
(no direct object)
 Use these to improve the results beyond what cooccurrence statistics indicate.
Marti Hearst, BYU CS 2007
Unambiguous + Unlimited = Unsupervised
 Apply the Unambiguous Case Idea to the Very, Very Large
Corpora idea
 The potential of these approaches are not fully realized
 Our work (with Preslav Nakov):
 Structural Ambiguity Decisions



PP-attachment
Noun compound bracketing
Coordination grouping
 Semantic Relation Acquisition


Hypernym (ISA) relations
Verbal relations between nouns

SAT Analogy problems
Marti Hearst, BYU CS 2007
Applying U + U = U to Structural Ambiguity
 We introduce the use of (nearly) unambiguous
features:
 Surface features
 Paraphrases
 Combined with ngrams
 From very, very large corpora
 Achieve state-of-the-art results without labeled
examples.
Marti Hearst, BYU CS 2007
Noun Compound Bracketing
(a)
(b)
[ [ liver cell ] antibody ]
[ liver [cell line] ]
(left bracketing)
(right bracketing)
In (a), the antibody targets the liver cell.
In (b), the cell line is derived from the liver.
Marti Hearst, BYU CS 2007
Dependency Model
 right bracketing: [w1[w2w3] ]
 w2w3 is a compound (modified by w1)
 home health care
 w1 and w2 independently modify w3
 adult male rat
w1
w2
w3
w1
w2
w3
 left bracketing : [ [w1w2 ]w3]
 only 1 modificational choice possible
 law enforcement officer
Marti Hearst, BYU CS 2007
Our U + U + U Algorithm




Compute bigram estimates
Compute estimates from surface features
Compute estimates from paraphrases
Combine these scores with a voting algorithm to
choose left or right bracketing.
 We use the same general approach for two other
structural ambiguity problems.
Marti Hearst, BYU CS 2007
Using n-grams to make predictions
 Say trying to distinguish:
[home health] care
home [health care]
 Main idea: compare these co-occurrence
probabilities
 “home health” vs
 “health care”
Marti Hearst, BYU CS 2007
Computing Bigram Statistics
 Dependency Model, Frequencies
Compare #(w1,w2) to #(w1,w3)
 Dependency model, Probabilities
Pr(left) = Pr(w1w2|w2)Pr(w2w3|w3)
Pr(right) = Pr(w1w3|w3)Pr(w2w3|w3)
right
w1
w2
w3
left
 So we compare Pr(w1w2|w2) to Pr(w1w3|w3)
Marti Hearst, BYU CS 2007
Using ngrams to estimate probabilities
 Using page hits as a proxy for n-gram counts
 Pr(w1w2|w2) = #(w1,w2) / #(w2)
 #(w2)
 #(w1,w2)
word frequency; query for “w2”
bigram frequency; query for “w1 w2”
 smoothed by 0.5
 Use 2 to determine if w1 is associated with w2
(thus indicating left bracketing), and same for w1
with w3
Marti Hearst, BYU CS 2007
Our U + U + U Algorithm




Compute bigram estimates
Compute estimates from surface features
Compute estimates from paraphrases
Combine these scores with a voting algorithm to
choose left or right bracketing.
Marti Hearst, BYU CS 2007
Web-derived Surface Features
 Authors often disambiguate noun compounds using
surface markers, e.g.:
 amino-acid sequence  left
 brain stem’s cell  left
 brain’s stem cell  right
 The enormous size of the Web makes these
frequent enough to be useful.
Marti Hearst, BYU CS 2007
Web-derived Surface Features:
Dash (hyphen)
 Left dash
 cell-cycle analysis  left
 Right dash
 donor T-cell  right
 Double dash
 T-cell-depletion  unusable…
Marti Hearst, BYU CS 2007
Web-derived Surface Features:
Possessive Marker
 Attached to the first word
 brain’s stem cell  right
 Attached to the second word
 brain stem’s cell  left
 Combined features
 brain’s stem-cell  right
Marti Hearst, BYU CS 2007
Web-derived Surface Features:
Capitalization
 anycase – lowercase – uppercase
 Plasmodium vivax Malaria  left
 plasmodium vivax Malaria  left
 lowercase – uppercase – anycase
 brain Stem cell  right
 brain Stem Cell  right
 Disable this on:
 Roman digits
 Single-letter words: e.g. vitamin D deficiency
Marti Hearst, BYU CS 2007
Web-derived Surface Features:
Embedded Slash
 Left embedded slash
 leukemia/lymphoma cell  right
Marti Hearst, BYU CS 2007
Web-derived Surface Features:
Parentheses
 Single-word
 growth factor (beta)  left
 (brain) stem cell  right
 Two-word
 (growth factor) beta  left
 brain (stem cell)  right
Marti Hearst, BYU CS 2007
Web-derived Surface Features:
Comma, dot, semi-colon
 Following the first word
 home. health care  right
 adult, male rat  right
 Following the second word
 health care, provider  left
 lung cancer: patients  left
Marti Hearst, BYU CS 2007
Web-derived Surface Features:
Dash to External Word
 External word to the left
 mouse-brain stem cell  right
 External word to the right
 tumor necrosis factor-alpha  left
Marti Hearst, BYU CS 2007
Other Web-derived Features:
Abbreviation
 After the second word
 tumor necrosis factor (NF)  right
 After the third word
 tumor necrosis (TN) factor  right
 We query for, e.g., “tumor necrosis tn factor”
 Problems:
 Roman digits: IV, VI
 States: CA
 Short words: me
Marti Hearst, BYU CS 2007
Other Web-derived Features:
Concatenation
 Consider health care reform
 healthcare : 79,500,000
 carereform : 269
 healthreform: 812
 Adjacency model
 healthcare vs. carereform
 Dependency model
 healthcare vs. healthreform
 Triples
 “healthcare reform” vs. “health carereform”
Marti Hearst, BYU CS 2007
Other Web-derived Features:
Reorder
 Reorders for “health care reform”
 “care reform health”  right
 “reform health care”  left
Marti Hearst, BYU CS 2007
Other Web-derived Features:
Internal Inflection Variability
 Vary inflection of second word
 tyrosine kinase activation
 tyrosine kinases activation
Marti Hearst, BYU CS 2007
Other Web-derived Features:
Switch The First Two Words
 Predict right, if we can reorder
 adult male rat
 male adult rat
as
Marti Hearst, BYU CS 2007
Our U + U + U Algorithm




Compute bigram estimates
Compute estimates from surface features
Compute estimates from paraphrases
Combine these scores with a voting algorithm to
choose left or right bracketing.
Marti Hearst, BYU CS 2007
Paraphrases
 The semantics of a noun compound is often made
overt by a paraphrase (Warren,1978)
 Prepositional
 stem cells in the brain  right
 cells from the brain stem  left
 Verbal
 virus causing human immunodeficiency  left
 Copula
 office building that is a skyscraper  right
Marti Hearst, BYU CS 2007
Paraphrases
 prepositional paraphrases:
 We use: ~150 prepositions
 verbal paraphrases:
 We use: associated with, caused by, contained in, derived
from, focusing on, found in, involved in, located at/in,
made of, performed by, preventing, related to and used
by/in/for.
 copula paraphrases:
 We use: is/was and that/which/who
 optional elements:
 articles: a, an, the
 quantifiers: some, every, etc.
 pronouns: this, these, etc.
Marti Hearst, BYU CS 2007
Paraphrases: pattern (1)
(1)v n1 p n2  v n2 n1

Can we turn “n1 p n2” into a noun compound “n2 n1”?



meet/v demands/n1 from/p customers/n2 
meet/v the customer/n2 demands/n1
Problem: ditransitive verbs like give



(noun)
gave/v an apple/n1 to/p him/n2 
gave/v him/n2 an apple/n1
Solution:



no determiner before n1
determiner before n2 is required
the preposition cannot be to
Marti Hearst, BYU CS 2007
Paraphrases: pattern (2)
(2)v n1 p n2  v p n2 n1

(verb)
If “p n2” is an indirect object of v, then it could
be switched with the direct object n1.


had/v a program/n1 in/p place/n2 
had/v in/p place/n2 a program/n1
Determiner before n1 is required to prevent
“n2 n1” from forming a noun compound.
Marti Hearst, BYU CS 2007
Paraphrases: pattern (3)
(3)v n1 p n2  p n2 * v n1
(verb)

“*” indicates a wildcard position (up to
three intervening words are allowed)

Looks for appositions, where the PP has
moved in front of the verb, e.g.


I gave/v an apple/n1 to/p him/n2 
to/p him/n2 I gave/v an apple/n1
Marti Hearst, BYU CS 2007
Paraphrases: pattern (4)
(4)v n1 p n2  n1 p n2 v

(noun)
Looks for appositions, where “n1 p n2” has
moved in front of v


shaken/v confidence/n1 in/p markets/n2 
confidence/n1 in/p markets/n2 shaken/v
Marti Hearst, BYU CS 2007
Paraphrases: pattern (5)
(5)v n1 p n2  v PRONOUN p n2
(verb)

n1 is a pronoun  verb (Hindle&Rooth, 93)

Pattern (5) substitutes n1 with a dative pronoun
(him or her), e.g.


put/v a client/n1 at/p odds/n2 
put/v him at/p odds/n2
Marti Hearst, BYU CS 2007
Paraphrases: pattern (6)
(6)v n1 p n2  BE n1 p n2
(noun)

BE is typically used with a noun attachment

Pattern (6) substitutes v with a form of to be (is
or are), e.g.


eat/v spaghetti/n1 with/p sauce/n2 
is spaghetti/n1 with/p sauce/n2
Marti Hearst, BYU CS 2007
Our U + U + U Algorithm




Compute bigram estimates
Compute estimates from surface features
Compute estimates from paraphrases
Combine these scores with a voting algorithm to
choose left or right bracketing.
Marti Hearst, BYU CS 2007
Evaluation: Datasets
 Lauer Set
 244 noun compounds (NCs)
 from Grolier’s encyclopedia
 inter-annotator agreement: 81.5%
 Biomedical Set
 430 NCs
 from MEDLINE
 inter-annotator agreement: 88% ( =.606)
Marti Hearst, BYU CS 2007
Co-occurrence Statistics
 Lauer set
 Bio set
Marti Hearst, BYU CS 2007
Paraphrase and Surface Features Performance
 Lauer Set
 Biomedical Set
Marti Hearst, BYU CS 2007
Individual Surface Features Performance: Bio
Marti Hearst, BYU CS 2007
Individual Surface Features Performance: Bio
Marti Hearst, BYU CS 2007
Results Lauer
Marti Hearst, BYU CS 2007
Results: Comparing with Others
Marti Hearst, BYU CS 2007
Results Bio
Marti Hearst, BYU CS 2007
Results for Noun Compound Bracketing
 Introduced search engine statistics that go
beyond the n-gram (applicable to other
tasks)
 surface features
 paraphrases
 Obtained new state-of-the-art results on NC
bracketing
 more robust than Lauer (1995)
 more accurate than Keller&Lapata (2004)
Marti Hearst, BYU CS 2007
Prepositional Phrase Attachment
Problem:
(a) Peter spent millions of dollars.
(b) Peter spent time with his family.
Which attachment for quadruple:
(noun attach)
(verb attach)
(v, n1, p, n2)
Results:
Much simpler than other algorithms
As good as or better than best unsupervised,
and better than some supervised approaches
Marti Hearst, BYU CS 2007
Noun Phrase Coordination
 (Modified) real sentence:
 The Department of Chronic Diseases and Health
Promotion leads and strengthens global efforts to
prevent and control chronic diseases or disabilities
and to promote health and quality of life.
Marti Hearst, BYU CS 2007
NC coordination: ellipsis
 Ellipsis
 car and truck production
 means car production and truck production
 No ellipsis
 president and chief executive
 All-way coordination
 Securities and Exchange Commission
Marti Hearst, BYU CS 2007
Results
428 examples from Penn TB
Marti Hearst, BYU CS 2007
Semantic Relation Detection
 Goal: automatically augment a lexical database
 Many potential relation types:
 ISA (hypernymy/hyponymy)
 Part-Of (meronymy)
 Idea: find unambiguous contexts which (nearly)
always indicate the relation of interest
Marti Hearst, BYU CS 2007
Lexico-Syntactic Patterns
Marti Hearst, BYU CS 2007
Lexico-Syntactic Patterns
Marti Hearst, BYU CS 2007
Adding a New Relation
Marti Hearst, BYU CS 2007
Semantic Relation Detection
 Lexico-syntactic Patterns:
 Should occur frequently in text
 Should (nearly) always suggest the relation of interest
 Should be recognizable with little pre-encoded
knowledge.
 These patterns have been used extensively by
other researchers.
Marti Hearst, BYU CS 2007
Semantic Relation Detection
 What relationship holds between two nouns?
 olive oil – oil comes from olives
 machine oil – oil used on machines
 Assigning the meaning relations between these
terms has been seen as a very difficult solution
 Our solution:
 Use clever queries against the web to figure out the
relations.
Marti Hearst, BYU CS 2007
Queries for Semantic Relations
 Convert the noun-noun compound into a query of the form:

noun2 that * noun1

“oil that * olive(s)”
 This returns search result snippets containing interesting
verbs.
 In this case:





Come from
Be obtained from
Be extracted from
Made from
…
Marti Hearst, BYU CS 2007
Uncovering Semantic Relations
 More examples:
 Migraine drug -> treat, be used for, reduce, prevent
 Wrinkle drug -> treat, be used for, reduce, smooth
 Printer tray -> hold, come with, be folded, fit under,
be inserted into
 Student protest -> be led by, be sponsored by, pit, be,
be organized by
Marti Hearst, BYU CS 2007
Conclusions
 Unambiguous + Unlimited = Unsupervised
 The enormous size of the web opens new opportunities for
text analysis
 There are many words, but they are more likely to appear together
in a huge dataset
 This allows us to do word-specific analysis
 To counter the labeled-data roadblock, we start with
unambiguous features that we can find naturally.
 We’ve applied this to structural and semantic language problems.
 These are stepping stones towards sophisticated language
understanding.
Marti Hearst, BYU CS 2007
Thank you!
http://biotext.berkeley.edu
Supported in part by NSF DBI-0317510