Advances in
Word Sense Disambiguation
Tutorial at ACL 2005
June 25, 2005
Ted Pedersen
University of Minnesota, Duluth
http://www.d.umn.edu/~tpederse
Rada Mihalcea
University of North Texas
http://www.cs.unt.edu/~rada
Goal of the Tutorial
• Introduce the problem of word sense disambiguation
(WSD), focusing on the range of formulations and
approaches currently practiced.
• Accessible to anyone with an interest in NLP.
• Persuade you to work on word sense disambiguation
– It’s an interesting problem
– Lots of good work already done, still more to do
– There is infrastructure to help you get started
• Persuade you to use word sense disambiguation in your
text applications.
2
Outline of Tutorial
•
•
•
•
•
•
•
•
3
Introduction (Ted)
Methodolodgy (Rada)
Knowledge Intensive Methods (Rada)
Supervised Approaches (Ted)
Minimally Supervised Approaches (Rada) / BREAK
Unsupervised Learning (Ted)
How to Get Started (Rada)
Conclusion (Ted)
Part 1:
Introduction
Outline
•
•
•
•
•
5
Definitions
Ambiguity for Humans and Computers
Very Brief Historical Overview
Theoretical Connections
Practical Applications
Definitions
• Word sense disambiguation is the problem of selecting a
sense for a word from a set of predefined possibilities.
– Sense Inventory usually comes from a dictionary or thesaurus.
– Knowledge intensive methods, supervised learning, and
(sometimes) bootstrapping approaches
• Word sense discrimination is the problem of dividing the
usages of a word into different meanings, without regard to
any particular existing sense inventory.
– Unsupervised techniques
6
Outline
•
•
•
•
•
7
Definitions
Ambiguity for Humans and Computers
Very Brief Historical Overview
Theoretical Connections
Practical Applications
Computers versus Humans
• Polysemy – most words have many possible meanings.
• A computer program has no basis for knowing which one
is appropriate, even if it is obvious to a human…
• Ambiguity is rarely a problem for humans in their day to
day communication, except in extreme cases…
8
Ambiguity for Humans - Newspaper Headlines!
•
•
•
•
•
•
•
•
•
9
DRUNK GETS NINE YEARS IN VIOLIN CASE
FARMER BILL DIES IN HOUSE
PROSTITUTES APPEAL TO POPE
STOLEN PAINTING FOUND BY TREE
RED TAPE HOLDS UP NEW BRIDGE
DEER KILL 300,000
RESIDENTS CAN DROP OFF TREES
INCLUDE CHILDREN WHEN BAKING COOKIES
MINERS REFUSE TO WORK AFTER DEATH
Ambiguity for a Computer
• The fisherman jumped off the bank and into the water.
• The bank down the street was robbed!
• Back in the day, we had an entire bank of computers
devoted to this problem.
• The bank in that road is entirely too steep and is really
dangerous.
• The plane took a bank to the left, and then headed off
towards the mountains.
10
Outline
•
•
•
•
•
11
Definitions
Ambiguity for Humans and Computers
Very Brief Historical Overview
Theoretical Connections
Practical Applications
Early Days of WSD
• Noted as problem for Machine Translation (Weaver, 1949)
– A word can often only be translated if you know the specific sense
intended (A bill in English could be a pico or a cuenta in Spanish)
• Bar-Hillel (1960) posed the following:
– Little John was looking for his toy box. Finally, he found it. The
box was in the pen. John was very happy.
– Is “pen” a writing instrument or an enclosure where children play?
…declared it unsolvable, left the field of MT!
12
Since then…
• 1970s - 1980s
– Rule based systems
– Rely on hand crafted knowledge sources
• 1990s
– Corpus based approaches
– Dependence on sense tagged text
– (Ide and Veronis, 1998) overview history from early days to 1998.
• 2000s
– Hybrid Systems
– Minimizing or eliminating use of sense tagged text
– Taking advantage of the Web
13
Outline
•
•
•
•
•
14
Definitions
Ambiguity for Humans and Computers
Very Brief Historical Overview
Interdisciplinary Connections
Practical Applications
Interdisciplinary Connections
• Cognitive Science & Psychology
– Quillian (1968), Collins and Loftus (1975) : spreading activation
• Hirst (1987) developed marker passing model
• Linguistics
– Fodor & Katz (1963) : selectional preferences
• Resnik (1993) pursued statistically
• Philosophy of Language
– Wittgenstein (1958): meaning as use
– “For a large class of cases-though not for all-in which we employ
the word "meaning" it can be defined thus: the meaning of a word
is its use in the language.”
15
Outline
•
•
•
•
•
16
Definitions
Ambiguity for Humans and Computers
Very Brief Historical Overview
Theoretical Connections
Practical Applications
Practical Applications
• Machine Translation
– Translate “bill” from English to Spanish
• Is it a “pico” or a “cuenta”?
• Is it a bird jaw or an invoice?
• Information Retrieval
– Find all Web Pages about “cricket”
• The sport or the insect?
• Question Answering
– What is George Miller’s position on gun control?
• The psychologist or US congressman?
• Knowledge Acquisition
– Add to KB: Herb Bergson is the mayor of Duluth.
• Minnesota or Georgia?
17
References
•
•
•
•
•
•
•
•
•
18
(Bar-Hillel, 1960) The Present Status of Automatic Translations of Languages. In
Advances in Computers. Volume 1. Alt, F. (editor). Academic Press, New York, NY. pp
91-163.
(Collins and Loftus, 1975) A Spreading Activation Theory of Semantic Memory.
Psychological Review, (82) pp. 407-428.
(Fodor and Katz, 1963) The structure of semantic theory. Language (39). pp 170-210.
(Hirst, 1987) Semantic Interpretation and the Resolution of Ambiguity. Cambridge
University Press.
(Ide and Véronis, 1998)Word Sense Disambiguation: The State of the Art..
Computational Linguistics (24) pp 1-40.
(Quillian, 1968) Semantic Memory. In Semantic Information Processing. Minsky, M.
(editor). The MIT Press, Cambridge, MA. pp. 227-270.
(Resnik, 1993) Selection and Information: A Class-Based Approach to Lexical
Relationships. Ph.D. Dissertation. University of Pennsylvania.
(Weaver, 1949): Translation. In Machine Translation of Languages: fourteen essays.
Locke, W.N. and Booth, A.D. (editors) The MIT Press, Cambridge, Mass. pp. 15-23.
(Wittgenstein, 1958) Philosophical Investigations, 3rd edition. Translated by G.E.M.
Anscombe. Macmillan Publishing Co., New York.
Part 2:
Methodology
Outline
•
•
•
•
•
20
General considerations
All-words disambiguation
Targeted-words disambiguation
Word sense discrimination, sense discovery
Evaluation (granularity, scoring)
Overview of the Problem
• Many words have several meanings (homonymy / polysemy)
–Ex: “chair” – furniture or person
–Ex: “child” – young person or human offspring
• Determine which sense of a word is used in a specific sentence
• Note:
– often, the different senses of a word are closely related
• Ex: title
- right of legal ownership
- document that is evidence of the legal ownership,
– sometimes, several senses can be “activated” in a single context
(co-activation)
• Ex: “This could bring competition to the trade”
competition: - the act of competing
- the people who are competing
21
Word Senses
• The meaning of a word in a given context
• Word sense representations
– With respect to a dictionary
chair = a seat for one person, with a support for the back; "he put his coat
over the back of the chair and sat down"
chair = the position of professor; "he was awarded an endowed chair in
economics"
– With respect to the translation in a second language
chair = chaise
chair = directeur
– With respect to the context where it occurs (discrimination)
“Sit on a chair” “Take a seat on this chair”
“The chair of the Math Department” “The chair of the meeting”
22
Approaches to Word Sense Disambiguation
• Knowledge-Based Disambiguation
– use of external lexical resources such as dictionaries and thesauri
– discourse properties
• Supervised Disambiguation
– based on a labeled training set
– the learning system has:
• a training set of feature-encoded inputs AND
• their appropriate sense label (category)
• Unsupervised Disambiguation
– based on unlabeled corpora
– The learning system has:
• a training set of feature-encoded inputs BUT
• NOT their appropriate sense label (category)
23
All Words Word Sense Disambiguation
• Attempt to disambiguate all open-class words in a text
“He put his suit over the back of the chair”
• Knowledge-based approaches
• Use information from dictionaries
– Definitions / Examples for each meaning
• Find similarity between definitions and current context
• Position in a semantic network
• Find that “table” is closer to “chair/furniture” than to “chair/person”
• Use discourse properties
• A word exhibits the same sense in a discourse / in a collocation
24
All Words Word Sense Disambiguation
• Minimally supervised approaches
– Learn to disambiguate words using small annotated corpora
– E.g. SemCor – corpus where all open class words are
disambiguated
• 200,000 running words
• Most frequent sense
25
Targeted Word Sense Disambiguation
• Disambiguate one target word
“Take a seat on this chair”
“The chair of the Math Department”
• WSD is viewed as a typical classification problem
– use machine learning techniques to train a system
• Training:
– Corpus of occurrences of the target word, each occurrence
annotated with appropriate sense
– Build feature vectors:
• a vector of relevant linguistic features that represents the context (ex:
a window of words around the target word)
• Disambiguation:
– Disambiguate the target word in new unseen text
26
Targeted Word Sense Disambiguation
•
•
Take a window of n word around the target word
Encode information about the words around the target word
– typical features include: words, root forms, POS tags, frequency, …
• An electric guitar and bass player stand off to one side, not really part of
the scene, just as a sort of nod to gringo expectations perhaps.
• Surrounding context (local features)
– [ (guitar, NN1), (and, CJC), (player, NN1), (stand, VVB) ]
• Frequent co-occurring words (topical features)
– [fishing, big, sound, player, fly, rod, pound, double, runs, playing, guitar, band]
– [0,0,0,1,0,0,0,0,0,0,1,0]
• Other features:
– [followed by "player", contains "show" in the sentence,…]
– [yes, no, … ]
27
Unsupervised Disambiguation
• Disambiguate word senses:
– without supporting tools such as dictionaries and thesauri
– without a labeled training text
• Without such resources, word senses are not labeled
– We cannot say “chair/furniture” or “chair/person”
• We can:
– Cluster/group the contexts of an ambiguous word into a number
of groups
– Discriminate between these groups without actually labeling
them
28
Unsupervised Disambiguation
• Hypothesis: same senses of words will have similar neighboring
words
• Disambiguation algorithm
– Identify context vectors corresponding to all occurrences of a particular
word
– Partition them into regions of high density
– Assign a sense to each such region
“Sit on a chair”
“Take a seat on this chair”
“The chair of the Math Department”
“The chair of the meeting”
29
Evaluating Word Sense Disambiguation
• Metrics:
– Precision = percentage of words that are tagged correctly, out of the
words addressed by the system
– Recall = percentage of words that are tagged correctly, out of all words
in the test set
– Example
• Test set of 100 words
• System attempts 75 words
• Words correctly disambiguated 50
Precision = 50 / 75 = 0.66
Recall = 50 / 100 = 0.50
• Special tags are possible:
– Unknown
– Proper noun
– Multiple senses
• Compare to a gold standard
– SEMCOR corpus, SENSEVAL corpus, …
30
Evaluating Word Sense Disambiguation
• Difficulty in evaluation:
– Nature of the senses to distinguish has a huge impact on results
• Coarse versus fine-grained sense distinction
chair = a seat for one person, with a support for the back; "he put his coat
over the back of the chair and sat down“
chair = the position of professor; "he was awarded an endowed chair in
economics“
bank = a financial institution that accepts deposits and channels the money
into lending activities; "he cashed a check at the bank"; "that bank holds
the mortgage on my home"
bank = a building in which commercial banking is transacted; "the bank is
on the corner of Nassau and Witherspoon“
• Sense maps
– Cluster similar senses
– Allow for both fine-grained and coarse-grained evaluation
31
Bounds on Performance
• Upper and Lower Bounds on Performance:
– Measure of how well an algorithm performs relative to the difficulty of
the task.
• Upper Bound:
– Human performance
– Around 97%-99% with few and clearly distinct senses
– Inter-judge agreement:
• With words with clear & distinct senses – 95% and up
• With polysemous words with related senses – 65% – 70%
• Lower Bound (or baseline):
– The assignment of a random sense / the most frequent sense
• 90% is excellent for a word with 2 equiprobable senses
• 90% is trivial for a word with 2 senses with probability ratios of 9 to 1
32
References
33
•
(Gale, Church and Yarowsky 1992) Gale, W., Church, K., and Yarowsky, D. Estimating
upper and lower bounds on the performance of word-sense disambiguation programs
ACL 1992.
•
(Miller et. al., 1994) Miller, G., Chodorow, M., Landes, S., Leacock, C., and Thomas, R.
Using a semantic concordance for sense identification. ARPA Workshop 1994.
•
(Miller, 1995) Miller, G. Wordnet: A lexical database. ACM, 38(11) 1995.
•
(Senseval) Senseval evaluation exercises http://www.senseval.org
Part 3:
Knowledge-based Methods for
Word Sense Disambiguation
Outline
• Task definition
– Machine Readable Dictionaries
•
•
•
•
35
Algorithms based on Machine Readable Dictionaries
Selectional Restrictions
Measures of Semantic Similarity
Heuristic-based Methods
Task Definition
• Knowledge-based WSD = class of WSD methods relying
(mainly) on knowledge drawn from dictionaries and/or raw
text
• Resources
– Yes
• Machine Readable Dictionaries
• Raw corpora
– No
• Manually annotated corpora
• Scope
– All open-class words
36
Machine Readable Dictionaries
• In recent years, most dictionaries made available in
Machine Readable format (MRD)
– Oxford English Dictionary
– Collins
– Longman Dictionary of Ordinary Contemporary English (LDOCE)
• Thesauruses – add synonymy information
– Roget Thesaurus
• Semantic networks – add more semantic relations
– WordNet
– EuroWordNet
37
MRD – A Resource for Knowledge-based WSD
• For each word in the language vocabulary, an MRD
provides:
– A list of meanings
– Definitions (for all word meanings)
– Typical usage examples (for most word meanings)
WordNet definitions/examples for the noun plant
1.
2.
3.
4.
38
buildings for carrying on industrial labor; "they built a large plant to
manufacture automobiles“
a living organism lacking the power of locomotion
something planted secretly for discovery by another; "the police used a plant to
trick the thieves"; "he claimed that the evidence against him was a plant"
an actor situated in the audience whose acting is rehearsed but seems
spontaneous to the audience
MRD – A Resource for Knowledge-based WSD
• A thesaurus adds:
– An explicit synonymy relation between word meanings
WordNet synsets for the noun “plant”
1. plant, works, industrial plant
2. plant, flora, plant life
• A semantic network adds:
– Hypernymy/hyponymy (IS-A), meronymy/holonymy (PART-OF),
antonymy, entailnment, etc.
WordNet related concepts for the meaning “plant life”
{plant, flora, plant life}
hypernym: {organism, being}
hypomym: {house plant}, {fungus}, …
meronym: {plant tissue}, {plant part}
holonym: {Plantae, kingdom Plantae, plant kingdom}
39
Outline
• Task definition
– Machine Readable Dictionaries
•
•
•
•
40
Algorithms based on Machine Readable Dictionaries
Selectional Restrictions
Measures of Semantic Similarity
Heuristic-based Methods
Lesk Algorithm
•
(Michael Lesk 1986): Identify senses of words in context
using definition overlap
Algorithm:
1. Retrieve from MRD all sense definitions of the words to be
disambiguated
2. Determine the definition overlap for all possible sense combinations
3. Choose senses that lead to highest overlap
Example: disambiguate PINE CONE
• PINE
1. kinds of evergreen tree with needle-shaped leaves
2. waste away through sorrow or illness
• CONE
1. solid body which narrows to a point
2. something of this shape whether solid or hollow
3. fruit of certain evergreen trees
41
Pine#1  Cone#1 = 0
Pine#2  Cone#1 = 0
Pine#1  Cone#2 = 1
Pine#2  Cone#2 = 0
Pine#1  Cone#3 = 2
Pine#2  Cone#3 = 0
Lesk Algorithm for More than Two Words?
• I saw a man who is 98 years old and can still walk and tell jokes
– nine open class words: see(26), man(11), year(4), old(8), can(5), still(4),
walk(10), tell(8), joke(3)
• 43,929,600 sense combinations! How to find the optimal sense
combination?
• Simulated annealing (Cowie, Guthrie, Guthrie 1992)
– Define a function E = combination of word senses in a given text.
– Find the combination of senses that leads to highest definition overlap
(redundancy)
1. Start with E = the most frequent sense for each word
2. At each iteration, replace the sense of a random word in the set with a
different sense, and measure E
3. Stop iterating when there is no change in the configuration of senses
42
Lesk Algorithm: A Simplified Version
•
Original Lesk definition: measure overlap between sense definitions for
all words in context
–
•
Simplified Lesk (Kilgarriff & Rosensweig 2000): measure overlap
between sense definitions of a word and current context
–
•
43
Identify simultaneously the correct senses for all words in context
Identify the correct sense for one word at a time
Search space significantly reduced
Lesk Algorithm: A Simplified Version
• Algorithm for simplified Lesk:
1.Retrieve from MRD all sense definitions of the word to be
disambiguated
2.Determine the overlap between each sense definition and the
current context
3.Choose the sense that leads to highest overlap
Example: disambiguate PINE in
“Pine cones hanging in a tree”
• PINE
1. kinds of evergreen tree with needle-shaped leaves
2. waste away through sorrow or illness
44
Pine#1  Sentence = 1
Pine#2  Sentence = 0
Evaluations of Lesk Algorithm
• Initial evaluation by M. Lesk
– 50-70% on short samples of text manually annotated set, with respect
to Oxford Advanced Learner’s Dictionary
• Simulated annealing
– 47% on 50 manually annotated sentences
• Evaluation on Senseval-2 all-words data, with back-off to
random sense (Mihalcea & Tarau 2004)
– Original Lesk: 35%
– Simplified Lesk: 47%
• Evaluation on Senseval-2 all-words data, with back-off to
most frequent sense (Vasilescu, Langlais, Lapalme 2004)
– Original Lesk: 42%
– Simplified Lesk: 58%
45
Outline
• Task definition
– Machine Readable Dictionaries
•
•
•
•
46
Algorithms based on Machine Readable Dictionaries
Selectional Preferences
Measures of Semantic Similarity
Heuristic-based Methods
Selectional Preferences
• A way to constrain the possible meanings of words in a
given context
• E.g. “Wash a dish” vs. “Cook a dish”
– WASH-OBJECT vs. COOK-FOOD
• Capture information about possible relations between
semantic classes
– Common sense knowledge
• Alternative terminology
– Selectional Restrictions
– Selectional Preferences
– Selectional Constraints
47
Acquiring Selectional Preferences
• From annotated corpora
– Circular relationship with the WSD problem
• Need WSD to build the annotated corpus
• Need selectional preferences to derive WSD
• From raw corpora
– Frequency counts
– Information theory measures
– Class-to-class relations
48
Preliminaries: Learning Word-to-Word Relations
• An indication of the semantic fit between two words
1. Frequency counts
– Pairs of words connected by a syntactic relations
Count (W 1 , W 2 , R )
2. Conditional probabilities
– Condition on one of the words
P (W 1 | W 2 , R ) 
49
Count (W 1 , W 2 , R )
Count (W 2 , R )
Learning Selectional Preferences (1)
• Word-to-class relations (Resnik 1993)
– Quantify the contribution of a semantic class using all the concepts
subsumed by that class
P ( C 2 | W 1 , R ) log
A (W 1 , C 2 , R ) 

P (C 2 | W1 , R )
P ( C 2 | W 1 , R ) log
P (C 2 )
P (C 2 | W1 , R )
C2
P (C 2 )
– where
P (C 2 | W1 , R ) 
Count (W 1 , C 2 , R )
Count (W 1 , R )
Count (W 1 , C 2 , R ) 

W 2 C 2
50
Count (W 1 , W 2 , R )
Count (W 2 )
Learning Selectional Preferences (2)
• Determine the contribution of a word sense based on the assumption of
equal sense distributions:
– e.g. “plant” has two senses  50% occurences are sense 1, 50% are sense 2
• Example: learning restrictions for the verb “to drink”
– Find high-scoring verb-object pairs
Co-occ score
11.75
11.75
11.75
10.53
10.2
9.34
Verb
drink
drink
drink
drink
drink
drink
Object
tea
Pepsi
champagne
liquid
beer
wine
– Find “prototypical” object classes (high association score)
51
A(v,c)
Object class
3.58 (beverage, [drink, …])
2.05 (alcoholic_beverage, [intoxicant, …])
Learning Selectional Preferences (3)
• Other algorithms
– Learn class-to-class relations (Agirre and Martinez, 2002)
• E.g.: “ingest food” is a class-to-class relation for “eat chicken”
– Bayesian networks (Ciaramita and Johnson, 2000)
– Tree cut model (Li and Abe, 1998)
52
Using Selectional Preferences for WSD
Algorithm:
1. Learn a large set of selectional preferences for a given syntactic
relation R
2. Given a pair of words W1– W2 connected by a relation R
3. Find all selectional preferences W1– C (word-to-class) or C1– C2
(class-to-class) that apply
4. Select the meanings of W1 and W2 based on the selected semantic
class
Example: disambiguate coffee in “drink coffee”
1. (beverage) a beverage consisting of an infusion of ground coffee beans
2. (tree) any of several small trees native to the tropical Old World
3. (color) a medium to dark brown color
Given the selectional preference “DRINK BEVERAGE” : coffee#1
53
Evaluation of Selectional Preferences for WSD
• Data set
– mainly on verb-object, subject-verb relations extracted from
SemCor
• Compare against random baseline
• Results (Agirre and Martinez, 2000)
– Average results on 8 nouns
– Similar figures reported in (Resnik 1997)
Object
Precision
Random
19.2
Word-to-word
95.9
Word-to-class
66.9
Class-to-class
66.6
54
Recall
19.2
24.9
58.0
64.8
Subject
Precision Recall
19.2
19.2
74.2
18.0
56.2
46.8
54.0
53.7
Outline
• Task definition
– Machine Readable Dictionaries
•
•
•
•
55
Algorithms based on Machine Readable Dictionaries
Selectional Restrictions
Measures of Semantic Similarity
Heuristic-based Methods
Semantic Similarity
• Words in a discourse must be related in meaning, for the
discourse to be coherent (Haliday and Hassan, 1976)
• Use this property for WSD – Identify related meanings for
words that share a common context
• Context span:
1. Local context: semantic similarity between pairs of words
2. Global context: lexical chains
56
Semantic Similarity in a Local Context
• Similarity determined between pairs of concepts, or
between a word and its surrounding context
• Relies on similarity metrics on semantic networks
– (Rada et al. 1989)
carnivore
fissiped mamal, fissiped
wolf
dingo
canine, canid
wild dog
dog
hyena dog
feline, felid
hyena
hunting dog
dachshund
57
terrier
bear
Semantic Similarity Metrics (1)
• Input: two concepts (same part of speech)
• Output: similarity measure
•
(Leacock and Chodorow 1998)
 Path ( C 1 , C 2 ) 
Similarity ( C 1 , C 2 )   log 

2
D


, D is the taxonomy depth
– E.g. Similarity(wolf,dog) = 0.60 Similarity(wolf,bear) = 0.42
•
(Resnik 1995)
– Define information content, P(C) = probability of seeing a concept of type
C in a large corpus
IC ( C )   log( P ( C ))
– Probability of seeing a concept = probability of seeing instances of that
concept
– Determine the contribution of a word sense based on the assumption of
equal sense distributions:
• e.g. “plant” has two senses  50% occurrences are sense 1, 50% are sense 2
58
Semantic Similarity Metrics (2)
• Similarity using information content
– (Resnik 1995) Define similarity between two concepts (LCS = Least
Common Subsumer)
Similarity ( C 1 , C 2 )  IC ( LCS ( C 1 , C 2 ))
– Alternatives (Jiang and Conrath 1997)
Similarity ( C 1 , C 2 )  2  IC ( LCS ( C 1 , C 2 )) 
• Other metrics:
( IC ( C 1 )  IC ( C 2 ))
– Similarity using information content (Lin 1998)
– Similarity using gloss-based paths across different hierarchies (Mihalcea
and Moldovan 1999)
– Conceptual density measure between noun semantic hierarchies and
current context (Agirre and Rigau 1995)
– Adapted Lesk algorithm (Banerjee and Pedersen 2002)
59
Semantic Similarity Metrics for WSD
• Disambiguate target words based on similarity with one
word to the left and one word to the right
– (Patwardhan, Banerjee, Pedersen 2002)
Example: disambiguate PLANT in “plant with flowers”
PLANT
1.
2.
plant, works, industrial plant
plant, flora, plant life
Similarity (plant#1, flower) = 0.2
Similarity (plant#2, flower) = 1.5
: plant#2
• Evaluation:
60
– 1,723 ambiguous nouns from Senseval-2
– Among 5 similarity metrics, (Jiang and Conrath 1997) provide the
best precision (39%)
Semantic Similarity in a Global Context
•
•
Lexical chains (Hirst and St-Onge 1988), (Haliday and Hassan 1976)
“A lexical chain is a sequence of semantically related words, which
creates a context and contributes to the continuity of meaning and the
coherence of a discourse”
Algorithm for finding lexical chains:
1. Select the candidate words from the text. These are words for which we can
compute similarity measures, and therefore most of the time they have the
same part of speech.
2. For each such candidate word, and for each meaning for this word, find a
chain to receive the candidate word sense, based on a semantic relatedness
measure between the concepts that are already in the chain, and the candidate
word meaning.
3. If such a chain is found, insert the word in this chain; otherwise, create a new
chain.
61
Semantic Similarity of a Global Context
A very long train traveling along the rails with a constant velocity v in a
certain direction …
train
#1: public transport
#2: order set of things
#3: piece of cloth
travel
#2: undergo transportation
rail
#1: a barrier
#3: a small bird
62
#1 change location
# 2: a bar of steel for trains
Lexical Chains for WSD
• Identify lexical chains in a text
– Usually target one part of speech at a time
• Identify the meaning of words based on their membership
to a lexical chain
• Evaluation:
– (Galley and McKeown 2003) lexical chains on 74 SemCor texts
give 62.09%
– (Mihalcea and Moldovan 2000) on five SemCor texts give 90%
with 60% recall
• lexical chains “anchored” on monosemous words
– (Okumura and Honda 1994) lexical chains on five Japanese texts
give 63.4%
63
Outline
• Task definition
– Machine Readable Dictionaries
•
•
•
•
64
Algorithms based on Machine Readable Dictionaries
Selectional Restrictions
Measures of Semantic Similarity
Heuristic-based Methods
Most Frequent Sense (1)
• Identify the most often used meaning and use this meaning
by default
Example: “plant/flora” is used more often than “plant/factory”
- annotate any instance of PLANT as “plant/flora”
• Word meanings exhibit a Zipfian distribution
– E.g. distribution of word senses in SemCor
0.9
0.8
Frequency
0.7
0.6
Noun
0.5
Verb
0.4
Adj
0.3
Adv
0.2
0.1
0
1
2
3
4
5
6
7
Sense number
65
8
9
10
Most Frequent Sense (2)
•
•
Method 1: Find the most frequent sense in an annotated
corpus
Method 2: Find the most frequent sense using a method
based on distributional similarity (McCarthy et al. 2004)
1. Given a word w, find the top k distributionally similar words
Nw = {n1, n2, …, nk}, with associated similarity scores {dss(w,n1),
dss(w,n2), … dss(w,nk)}
2. For each sense wsi of w, identify the similarity with the words nj,
using the sense of nj that maximizes this score
3. Rank senses wsi of w based on the total similarity score
Score ( ws i ) 

dss ( w , n j )
n jN w
wnss ( ws i , n j )
 wnss ( ws
ws i ' senses ( w )
where
wnss ( ws i , n j ) 
66
max
ns x  senses ( n j )
( wnss ( ws i , ns x ))
i
', n j )
,
Most Frequent Sense(3)
• Word senses
– pipe #1 = tobacco pipe
– pipe #2 = tube of metal or plastic
• Distributional similar words
– N = {tube, cable, wire, tank, hole, cylinder, fitting, tap, …}
• For each word in N, find similarity with pipe#i (using the sense that
maximizes the similarity)
– pipe#1 – tube (#3) = 0.3
– pipe#2 – tube (#1) = 0.6
• Compute score for each sense pipe#i
– score (pipe#1) = 0.25
– score (pipe#2) = 0.73
Note: results depend on the corpus used to find distributionally
similar words => can find domain specific predominant senses
67
One Sense Per Discourse
• A word tends to preserve its meaning across all its occurrences in a
given discourse (Gale, Church, Yarowksy 1992)
• What does this mean?
E.g. The ambiguous word PLANT occurs 10 times in a discourse
all instances of “plant” carry the same meaning
• Evaluation:
– 8 words with two-way ambiguity, e.g. plant, crane, etc.
– 98% of the two-word occurrences in the same discourse carry the same
meaning
• The grain of salt: Performance depends on granularity
– (Krovetz 1998) experiments with words with more than two senses
– Performance of “one sense per discourse” measured on SemCor is approx.
70%
68
One Sense per Collocation
• A word tends to preserver its meaning when used in the same
collocation (Yarowsky 1993)
– Strong for adjacent collocations
– Weaker as the distance between words increases
• An example
The ambiguous word PLANT preserves its meaning in all its
occurrences within the collocation “industrial plant”, regardless
of the context where this collocation occurs
• Evaluation:
– 97% precision on words with two-way ambiguity
• Finer granularity:
– (Martinez and Agirre 2000) tested the “one sense per collocation”
hypothesis on text annotated with WordNet senses
– 70% precision on SemCor words
69
References
•
•
•
•
•
•
•
•
•
•
•
70 •
(Agirre and Rigau, 1995) Agirre, E. and Rigau, G. A proposal for word sense disambiguation
using conceptual distance. RANLP 1995.
(Agirre and Martinez 2001) Agirre, E. and Martinez, D. Learning class-to-class selectional
preferences. CONLL 2001.
(Banerjee and Pedersen 2002) Banerjee, S. and Pedersen, T. An adapted Lesk algorithm for
word sense disambiguation using WordNet. CICLING 2002.
(Cowie, Guthrie and Guthrie 1992), Cowie, L. and Guthrie, J. A. and Guthrie, L.: Lexical
disambiguation using simulated annealing. COLING 2002.
(Gale, Church and Yarowsky 1992) Gale, W., Church, K., and Yarowsky, D. One sense per
discourse. DARPA workshop 1992.
(Halliday and Hasan 1976) Halliday, M. and Hasan, R., (1976). Cohesion in English.
Longman.
(Galley and McKeown 2003) Galley, M. and McKeown, K. (2003) Improving word sense
disambiguation in lexical chaining. IJCAI 2003
(Hirst and St-Onge 1998) Hirst, G. and St-Onge, D. Lexical chains as representations of
context in the detection and correction of malaproprisms. WordNet: An electronic lexical
database, MIT Press.
(Jiang and Conrath 1997) Jiang, J. and Conrath, D. Semantic similarity based on corpus
statistics and lexical taxonomy. COLING 1997.
(Krovetz, 1998) Krovetz, R. More than one sense per discourse. ACL-SIGLEX 1998.
(Lesk, 1986) Lesk, M. Automatic sense disambiguation using machine readable dictionaries:
How to tell a pine cone from an ice cream cone. SIGDOC 1986.
(Lin 1998) Lin, D An information theoretic definition of similarity. ICML 1998.
References
•
•
•
•
•
•
•
•
•
•
•
71
•
(Martinez and Agirre 2000) Martinez, D. and Agirre, E. One sense per collocation and
genre/topic variations. EMNLP 2000.
(Miller et. al., 1994) Miller, G., Chodorow, M., Landes, S., Leacock, C., and Thomas, R.
Using a semantic concordance for sense identification. ARPA Workshop 1994.
(Miller, 1995) Miller, G. Wordnet: A lexical database. ACM, 38(11) 1995.
(Mihalcea and Moldovan, 1999) Mihalcea, R. and Moldovan, D. A method for word
sense disambiguation of unrestricted text. ACL 1999.
(Mihalcea and Moldovan 2000) Mihalcea, R. and Moldovan, D. An iterative approach
to word sense disambiguation. FLAIRS 2000.
(Mihalcea, Tarau, Figa 2004) R. Mihalcea, P. Tarau, E. Figa PageRank on Semantic
Networks with Application to Word Sense Disambiguation, COLING 2004.
(Patwardhan, Banerjee, and Pedersen 2003) Patwardhan, S. and Banerjee, S. and
Pedersen, T. Using Measures of Semantic Relatedeness for Word Sense Disambiguation.
CICLING 2003.
(Rada et al 1989) Rada, R. and Mili, H. and Bicknell, E. and Blettner, M. Development
and application of a metric on semantic nets. IEEE Transactions on Systems, Man, and
Cybernetics, 19(1) 1989.
(Resnik 1993) Resnik, P. Selection and Information: A Class-Based Approach to Lexical
Relationships. University of Pennsylvania 1993.
(Resnik 1995) Resnik, P. Using information content to evaluate semantic similarity.
IJCAI 1995.
(Vasilescu, Langlais, Lapalme 2004) F. Vasilescu, P. Langlais, G. Lapalme "Evaluating
variants of the Lesk approach for disambiguating words”, LREC 2004.
(Yarowsky, 1993) Yarowsky, D. One sense per collocation. ARPA Workshop 1993.
Part 4:
Supervised Methods of Word
Sense Disambiguation
Outline
• What is Supervised Learning?
• Task Definition
• Single Classifiers
– Naïve Bayesian Classifiers
– Decision Lists and Trees
• Ensembles of Classifiers
73
What is Supervised Learning?
• Collect a set of examples that illustrate the various possible
classifications or outcomes of an event.
• Identify patterns in the examples associated with each
particular class of the event.
• Generalize those patterns into rules.
• Apply the rules to classify a new event.
74
Learn from these examples :
“when do I go to the store?”
Day
75
CLASS
F1
Go to Store? Hot Outside?
F2
Slept Well?
F3
Ate Well?
1
YES
YES
NO
NO
2
NO
YES
NO
YES
3
YES
NO
NO
NO
4
NO
NO
NO
YES
Learn from these examples :
“when do I go to the store?”
Day
76
CLASS
F1
Go to Store? Hot Outside?
F2
Slept Well?
F3
Ate Well?
1
YES
YES
NO
NO
2
NO
YES
NO
YES
3
YES
NO
NO
NO
4
NO
NO
NO
YES
Outline
• What is Supervised Learning?
• Task Definition
• Single Classifiers
– Naïve Bayesian Classifiers
– Decision Lists and Trees
• Ensembles of Classifiers
77
Task Definition
• Supervised WSD: Class of methods that induces a classifier from
manually sense-tagged text using machine learning techniques.
• Resources
– Sense Tagged Text
– Dictionary (implicit source of sense inventory)
– Syntactic Analysis (POS tagger, Chunker, Parser, …)
• Scope
– Typically one target word per context
– Part of speech of target word resolved
– Lends itself to “targeted word” formulation
• Reduces WSD to a classification problem where a target word is
assigned the most appropriate sense from a given set of possibilities
based on the context in which it occurs
78
Sense Tagged Text
Bonnie and Clyde are two really famous criminals, I think
they were bank/1 robbers
My bank/1 charges too much for an overdraft.
I went to the bank/1 to deposit my check and get a new ATM
card.
The University of Minnesota has an East and a West Bank/2
campus right on the Mississippi River.
My grandfather planted his pole in the bank/2 and got a great
big catfish!
The bank/2 is pretty muddy, I can’t walk there.
79
Two Bags of Words
(Co-occurrences in the “window of context”)
FINANCIAL_BANK_BAG:
a an and are ATM Bonnie card charges check Clyde
criminals deposit famous for get I much My new overdraft
really robbers the they think to too two went were
RIVER_BANK_BAG:
a an and big campus cant catfish East got grandfather great
has his I in is Minnesota Mississippi muddy My of on planted
pole pretty right River The the there University walk West
80
Simple Supervised Approach
Given a sentence S containing “bank”:
For each word Wi in S
If Wi is in FINANCIAL_BANK_BAG then
Sense_1 = Sense_1 + 1;
If Wi is in RIVER_BANK_BAG then
Sense_2 = Sense_2 + 1;
If Sense_1 > Sense_2 then print “Financial”
else if Sense_2 > Sense_1 then print “River”
else print “Can’t Decide”;
81
Supervised Methodology
• Create a sample of training data where a given target word
is manually annotated with a sense from a predetermined
set of possibilities.
– One tagged word per instance/lexical sample disambiguation
• Select a set of features with which to represent context.
– co-occurrences, collocations, POS tags, verb-obj relations, etc...
• Convert sense-tagged training instances to feature vectors.
• Apply a machine learning algorithm to induce a classifier.
– Form – structure or relation among features
– Parameters – strength of feature interactions
• Convert a held out sample of test data into feature vectors.
– “correct” sense tags are known but not used
• Apply classifier to test instances to assign a sense tag.
82
From Text to Feature Vectors
• My/pronoun grandfather/noun used/verb to/prep fish/verb
along/adv the/det banks/SHORE of/prep the/det
Mississippi/noun River/noun. (S1)
• The/det bank/FINANCE issued/verb a/det check/noun
for/prep the/det amount/noun of/prep interest/noun. (S2)
S1
S2
83
P-2
P-1
P+1 P+2 fish check river interest SENSE TAG
adv
det
prep
det
Y
N
Y
N
SHORE
det
verb
det
N
Y
N
Y
FINANCE
Supervised Learning Algorithms
• Once data is converted to feature vector form, any
supervised learning algorithm can be used. Many have
been applied to WSD with good results:
–
–
–
–
–
–
–
–
–
84
Support Vector Machines
Nearest Neighbor Classifiers
Decision Trees
Decision Lists
Naïve Bayesian Classifiers
Perceptrons
Neural Networks
Graphical Models
Log Linear Models
Outline
•
•
•
•
•
85
What is Supervised Learning?
Task Definition
Naïve Bayesian Classifier
Decision Lists and Trees
Ensembles of Classifiers
Naïve Bayesian Classifier
• Naïve Bayesian Classifier well known in Machine
Learning community for good performance across a range
of tasks (e.g., Domingos and Pazzani, 1997)
…Word Sense Disambiguation is no exception
• Assumes conditional independence among features, given
the sense of a word.
– The form of the model is assumed, but parameters are estimated
from training instances
• When applied to WSD, features are often “a bag of words”
that come from the training data
– Usually thousands of binary features that indicate if a word is
present in the context of the target word (or not)
86
Bayesian Inference
p ( S | F 1 , F 2 , F 3 ,..., Fn )
•
•
•
•
87
p ( F 1 , F 2 , F 3 ,..., Fn | S )* p ( S )
p ( F 1, F 2 , F 3 ,..., Fn )
Given observed features, what is most likely sense?
Estimate probability of observed features given sense
Estimate unconditional probability of sense
Unconditional probability of features is a normalizing
term, doesn’t affect sense classification
Naïve Bayesian Model
S
F1
F2
F3
F4
Fn
P ( F 1, F 2 ,..., Fn | S )  p ( F 1 | S ) * p ( F 2 | S ) * ... * p ( Fn | S )
88
The Naïve Bayesian Classifier
sense  argmax
p ( F 1 | S ) * ... * p ( Fn | S ) * p ( S )
sense  S
– Given 2,000 instances of “bank”, 1,500 for bank/1 (financial sense)
and 500 for bank/2 (river sense)
• P(S=1) = 1,500/2000 = .75
• P(S=2) = 500/2,000 = .25
– Given “credit” occurs 200 times with bank/1 and 4 times with bank/2.
• P(F1=“credit”) = 204/2000 = .102
• P(F1=“credit”|S=1) = 200/1,500 = .133
• P(F1=“credit”|S=2) = 4/500 = .008
– Given a test instance that has one feature “credit”
• P(S=1|F1=“credit”) = .133*.75/.102 = .978
• P(S=2|F1=“credit”) = .008*.25/.102 = .020
89
Comparative Results
• (Leacock, et. al. 1993) compared Naïve Bayes with a
Neural Network and a Context Vector approach when
disambiguating six senses of line…
• (Mooney, 1996) compared Naïve Bayes with a Neural
Network, Decision Tree/List Learners, Disjunctive and
Conjunctive Normal Form learners, and a perceptron when
disambiguating six senses of line…
• (Pedersen, 1998) compared Naïve Bayes with Decision
Tree, Rule Based Learner, Probabilistic Model, etc. when
disambiguating line and 12 other words…
• …All found that Naïve Bayesian Classifier performed as
well as any of the other methods!
90
Outline
•
•
•
•
•
91
What is Supervised Learning?
Task Definition
Naïve Bayesian Classifiers
Decision Lists and Trees
Ensembles of Classifiers
Decision Lists and Trees
• Very widely used in Machine Learning.
• Decision trees used very early for WSD research (e.g.,
Kelly and Stone, 1975; Black, 1988).
• Represent disambiguation problem as a series of questions
(presence of feature) that reveal the sense of a word.
– List decides between two senses after one positive answer
– Tree allows for decision among multiple senses after a series of
answers
• Uses a smaller, more refined set of features than “bag of
words” and Naïve Bayes.
– More descriptive and easier to interpret.
92
Decision List for WSD (Yarowsky, 1994)
• Identify collocational features from sense tagged data.
• Word immediately to the left or right of target :
– I have my bank/1 statement.
– The river bank/2 is muddy.
• Pair of words to immediate left or right of target :
– The world’s richest bank/1 is here in New York.
– The river bank/2 is muddy.
• Words found within k positions to left or right of target,
where k is often 10-50 :
– My credit is just horrible because my bank/1 has made several
mistakes with my account and the balance is very low.
93
Building the Decision List
• Sort order of collocation tests using log of conditional
probabilities.
• Words most indicative of one sense (and not the other)
will be ranked highly.
p ( S  1| F  Collocatio n )
i
i
Abs (log
)
p ( S  2 | F  Collocatio n )
i
i
94
Computing DL score
– Given 2,000 instances of “bank”, 1,500 for bank/1
(financial sense) and 500 for bank/2 (river sense)
• P(S=1) = 1,500/2,000 = .75
• P(S=2) = 500/2,000 = .25
– Given “credit” occurs 200 times with bank/1 and 4
times with bank/2.
• P(F1=“credit”) = 204/2,000 = .102
• P(F1=“credit”|S=1) = 200/1,500 = .133
• P(F1=“credit”|S=2) = 4/500 = .008
– From Bayes Rule…
• P(S=1|F1=“credit”) = .133*.75/.102 = .978
• P(S=2|F1=“credit”) = .008*.25/.102 = .020
– DL Score = abs (log (.978/.020)) = 3.89
95
Using the Decision List
• Sort DL-score, go through test instance looking for
matching feature. First match reveals sense…
96
DL-score
Feature
Sense
3.89
credit within bank
Bank/1 financial
2.20
bank is muddy
Bank/2 river
1.09
pole within bank
Bank/2 river
0.00
of the bank
N/A
Using the Decision List
CREDIT?
BANK/1 FINANCIAL
IS MUDDY?
BANK/2 RIVER
POLE?
BANK/2 RIVER
97
Learning a Decision Tree
• Identify the feature that most “cleanly” divides the training
data into the known senses.
– “Cleanly” measured by information gain or gain ratio.
– Create subsets of training data according to feature values.
• Find another feature that most cleanly divides a subset of
the training data.
• Continue until each subset of training data is “pure” or as
clean as possible.
• Well known decision tree learning algorithms include ID3
and C4.5 (Quillian, 1986, 1993)
• In Senseval-1, a modified decision list (which supported
some conditional branching) was most accurate for English
Lexical Sample task (Yarowsky, 2000)
98
Supervised WSD with Individual Classifiers
• Many supervised Machine Learning algorithms have been
applied to Word Sense Disambiguation, most work
reasonably well.
– (Witten and Frank, 2000) is a great intro. to supervised learning.
• Features tend to differentiate among methods more than
the learning algorithms.
• Good sets of features tend to include:
–
–
–
–
–
Co-occurrences or keywords (global)
Collocations (local)
Bigrams (local and global)
Part of speech (local)
Predicate-argument relations
• Verb-object, subject-verb,
– Heads of Noun and Verb Phrases
99
Convergence of Results
• Accuracy of different systems applied to the same data
tends to converge on a particular value, no one system
shockingly better than another.
– Senseval-1, a number of systems in range of 74-78% accuracy for
English Lexical Sample task.
– Senseval-2, a number of systems in range of 61-64% accuracy for
English Lexical Sample task.
– Senseval-3, a number of systems in range of 70-73% accuracy for
English Lexical Sample task…
• What to do next?
100
Outline
•
•
•
•
•
101
What is Supervised Learning?
Task Definition
Naïve Bayesian Classifiers
Decision Lists and Trees
Ensembles of Classifiers
Ensembles of Classifiers
• Classifier error has two components (Bias and Variance)
– Some algorithms (e.g., decision trees) try and build a
representation of the training data – Low Bias/High Variance
– Others (e.g., Naïve Bayes) assume a parametric form and don’t
represent the training data – High Bias/Low Variance
• Combining classifiers with different bias variance
characteristics can lead to improved overall accuracy
• “Bagging” a decision tree can smooth out the effect of
small variations in the training data (Breiman, 1996)
– Sample with replacement from the training data to learn multiple
decision trees.
– Outliers in training data will tend to be obscured/eliminated.
102
Ensemble Considerations
• Must choose different learning algorithms with
significantly different bias/variance characteristics.
– Naïve Bayesian Classifier versus Decision Tree
• Must choose feature representations that yield significantly
different (independent?) views of the training data.
– Lexical versus syntactic features
• Must choose how to combine classifiers.
– Simple Majority Voting
– Averaging of probabilities across multiple classifier output
– Maximum Entropy combination (e.g., Klein, et. al., 2002)
103
Ensemble Results
• (Pedersen, 2000) achieved state of art for interest and line
data using ensemble of Naïve Bayesian Classifiers.
– Many Naïve Bayesian Classifiers trained on varying sized
windows of context / bags of words.
– Classifiers combined by a weighted vote
• (Florian and Yarowsky, 2002) achieved state of the art for
Senseval-1 and Senseval-2 data using combination of six
classifiers.
– Rich set of collocational and syntactic features.
– Combined via linear combination of top three classifiers.
• Many Senseval-2 and Senseval-3 systems employed
ensemble methods.
104
References
•
•
•
•
•
•
•
•
•
105
(Black, 1988) An experiment in computational discrimination of English word senses.
IBM Journal of Research and Development (32) pg. 185-194.
(Breiman, 1996) The heuristics of instability in model selection. Annals of Statistics (24)
pg. 2350-2383.
(Domingos and Pazzani, 1997) On the Optimality of the Simple Bayesian Classifier
under Zero-One Loss, Machine Learning (29) pg. 103-130.
(Domingos, 2000) A Unified Bias Variance Decomposition for Zero-One and Squared
Loss. In Proceedings of AAAI. Pg. 564-569.
(Florian an dYarowsky, 2002) Modeling Consensus: Classifier Combination for Word
Sense Disambiguation. In Proceedings of EMNLP, pp 25-32.
(Kelly and Stone, 1975). Computer Recognition of English Word Senses, North Holland
Publishing Co., Amsterdam.
(Klein, et. al., 2002) Combining Heterogeneous Classifiers for Word-Sense
Disambiguation, Proceedings of Senseval-2. pg. 87-89.
(Leacock, et. al. 1993) Corpus based statistical sense resolution. In Proceedings of the
ARPA Workshop on Human Language Technology. pg. 260-265.
(Mooney, 1996) Comparative experiments on disambiguating word senses: An
illustration of the role of bias in machine learning. Proceedings of EMNLP. pg. 82-91.
References
•
•
•
•
•
•
•
106
(Pedersen, 1998) Learning Probabilistic Models of Word Sense Disambiguation. Ph.D.
Dissertation. Southern Methodist University.
(Pedersen, 2000) A simple approach to building ensembles of Naive Bayesian classifiers
for word sense disambiguation. In Proceedings of NAACL.
(Quillian, 1986). Induction of Decision Trees. Machine Learning (1). pg. 81-106.
(Quillian, 1993). C4.5 Programs for Machine Learning. San Francisco, Morgan
Kaufmann.
(Witten and Frank, 2000). Data Mining – Practical Machine Learning Tools and
Techniques with Java Implementations. Morgan-Kaufmann. San Francisco.
(Yarowsky, 1994) Decision lists for lexical ambiguity resolution: Application to accent
restoration in Spanish and French. In Proceedings of ACL. pp. 88-95.
(Yarowsky, 2000) Hierarchical decision lists for word sense disambiguation. Computers
and the Humanities, 34.
Part 5:
Minimally Supervised Methods
for Word Sense Disambiguation
Outline
• Task definition
– What does “minimally” supervised mean?
• Bootstrapping algorithms
– Co-training
– Self-training
– Yarowsky algorithm
• Using the Web for Word Sense Disambiguation
– Web as a corpus
– Web as collective mind
108
Task Definition
• Supervised WSD = learning sense classifiers starting with
annotated data
• Minimally supervised WSD = learning sense classifiers
from annotated data, with minimal human supervision
• Examples
– Automatically bootstrap a corpus starting with a few human
annotated examples
– Use monosemous relatives / dictionary definitions to automatically
construct sense tagged data
– Rely on Web-users + active learning for corpus annotation
109
Outline
• Task definition
– What does “minimally” supervised mean?
• Bootstrapping algorithms
– Co-training
– Self-training
– Yarowsky algorithm
• Using the Web for Word Sense Disambiguation
– Web as a corpus
– Web as collective mind
110
Bootstrapping WSD Classifiers
• Build sense classifiers with little training data
– Expand applicability of supervised WSD
• Bootstrapping approaches
– Co-training
– Self-training
– Yarowsky algorithm
111
Bootstrapping Recipe
• Ingredients
– (Some) labeled data
– (Large amounts of) unlabeled data
– (One or more) basic classifiers
• Output
– Classifier that improves over the basic classifiers
112
… plant#1 growth is retarded …
… a nuclear power plant#2 …
… plants#1 and animals …
… industry plant#2 …
Classifier 1
Classifier 2
113
… building the only atomic plant …
… plant growth is retarded …
… a herb or flowering plant …
… a nuclear power plant …
… building a new vehicle plant …
… the animal and plant life …
… the passion-fruit plant …
Co-training / Self-training
– A set L of labeled training examples
– A set U of unlabeled examples
– Classifiers Ci
• 1. Create a pool of examples U'
– choose P random examples from U
• 2. Loop for I iterations
– Train Ci on L and label U'
– Select G most confident examples and add to L
• maintain distribution in L
– Refill U' with examples from U
• keep U' at constant size P
114
Co-training
• (Blum and Mitchell 1998)
• Two classifiers
– independent views
– [independence condition can be relaxed]
• Co-training in Natural Language Learning
–
–
–
–
115
Statistical parsing (Sarkar 2001)
Co-reference resolution (Ng and Cardie 2003)
Part of speech tagging (Clark, Curran and Osborne 2003)
...
Self-training
• (Nigam and Ghani 2000)
• One single classifier
• Retrain on its own output
• Self-training for Natural Language Learning
– Part of speech tagging (Clark, Curran and Osborne 2003)
– Co-reference resolution (Ng and Cardie 2003)
• several classifiers through bagging
116
Parameter Setting for Co-training/Self-training
• 1. Create a pool of examples U'
P o o l s ize
– choose P random examples from U
• 2. Loop for I iterations
– Train Ci on L and label U'
– Select G most confident examples and add to L
I te r a tio n s
• maintain distribution in L
– Refill U' with examples from U
• keep U' at constant size P
G ro w th s ize
•A major drawback of bootstrapping
117
–“No principled method for selecting optimal values for these
parameters” (Ng and Cardie 2003)
Experiments with Co-training / Self-training
for WSD
• Training / Test data
– Senseval-2 nouns (29 ambiguous nouns)
– Average corpus size: 95 training examples, 48 test examples
• Raw data
– British National Corpus
– Average corpus size: 7,085 examples
• Co-training
– Two classifiers: local and topical classifiers
• Self-training
– One classifier: global classifier
•
118
(Mihalcea 2004)
Parameter Settings
• Parameter ranges
– P = {1, 100, 500, 1000, 1500, 2000, 5000}
– G = {1, 10, 20, 30, 40, 50, 100, 150, 200}
– I = {1, ..., 40}
• 29 nouns → 120,000 runs
• Upper bound in co-training/self-training performance
–
–
–
–
–
Optimised on test set
Basic classifier: 53.84%
Optimal self-training: 65.61%
Optimal co-training: 65.75%
~25% error reduction
• Per-word parameter setting:
– Co-training = 51.73%
– Self-training = 52.88%
• Global parameter setting
119
– Co-training = 55.67%
– Self-training = 54.16%
• Example: lady
– basic = 61.53%
– self-training = 84.61%
[20/100/39]
– co-training = 82.05%
[1/1000/3]
Yarowsky Algorithm
• (Yarowsky 1995)
• Similar to co-training
• Differs in the basic assumption (Abney 2002)
– “view independence” (co-training) vs. “precision independence”
(Yarowsky algorithm)
• Relies on two heuristics and a decision list
– One sense per collocation :
• Nearby words provide strong and consistent clues as to the sense of a
target word
– One sense per discourse :
• The sense of a target word is highly consistent within a single
document
120
Learning Algorithm
•
A decision list is used to classify instances of target word :
“the loss of animal and plant species through extinction …”
• Classification is based on the highest ranking rule that
matches the target context
121
LogL
Collocation
Sense
…
…
…
9.31
flower (within +/- k words)
 A (living)
9.24
job (within +/- k words)
 B (factory)
9.03
fruit (within +/- k words)
 A (living)
9.02
plant species
 A (living)
...
...
…
Bootstrapping Algorithm
Sense-A: life
Sense-B: factory
• All occurrences of the target word are identified
• A small training set of seed data is tagged with word sense
122
Bootstrapping Algorithm
Seed set grows and residual set shrinks ….
123
Bootstrapping Algorithm
Convergence: Stop when residual set stabilizes
124
Bootstrapping Algorithm
• Iterative procedure:
– Train decision list algorithm on seed set
– Classify residual data with decision list
– Create new seed set by identifying samples that are tagged with a
probability above a certain threshold
– Retrain classifier on new seed set
• Selecting training seeds
– Initial training set should accurately distinguish among possible
senses
– Strategies:
125
• Select a single, defining seed collocation for each possible sense.
Ex: “life” and “manufacturing” for target plant
• Use words from dictionary definitions
• Hand-label most frequent collocates
Evaluation
• Test corpus: extracted from 460 million word corpus of multiple
sources (news articles, transcripts, novels, etc.)
• Performance of multiple models compared with:
– supervised decision lists
– unsupervised learning algorithm of Schütze (1992), based on
alignment of clusters with word senses
126
Word
Senses
Supervised
Unsupervised
Schütze
Unsupervised
Bootstrapping
plant
living/factory
97.7
92
98.6
space
volume/outer
93.9
90
93.6
tank
vehicle/container
97.1
95
96.5
motion
legal/physical
98.0
92
97.9
…
…
…
-
…
Avg.
-
96.1
92.2
96.5
Outline
• Task definition
– What does “minimally” supervised mean?
• Bootstrapping algorithms
– Co-training
– Self-training
– Yarowsky algorithm
• Using the Web for Word Sense Disambiguation
– Web as a corpus
– Web as collective mind
127
The Web as a Corpus
• Use the Web as a large textual corpus
– Build annotated corpora using monosemous relatives
– Bootstrap annotated corpora starting with few seeds
• Similar to (Yarowsky 1995)
• Use the (semi)automatically tagged data to train WSD
classifiers
128
Monosemous Relatives
•
Idea: determine a phrase (SP) which uniquely identifies the
sense of a word (W#i)
1. Determine one or more Search Phrases from a machine readable
dictionary using several heuristics
2. Search the Web using the Search Phrases from step 1.
3. Replace the Search Phrases in the examples gathered at 2 with W#i.
–
Output: sense annotated corpus for the word sense W#i
As a pastime, she enjoyed reading.
Evaluate the interestingness of the website.
As an interest, she enjoyed reading.
Evaluate the interest of the website.
129
Heuristics to Identify Monosemous Relatives
• Synonyms
– Determine a monosemous synonym
– remember#1 has recollect as monosemous synonym  SP=recollect
• Dictionary definitions (1)
– Parse the gloss and determine the set of single phrase definitions
– produce#5 has the definition “bring onto the market or release”  2
definitions: “bring onto the market” and “release”
eliminate “release” as being ambiguous  SP=bring onto the market
• Dictionary defintions (2)
– Parse the gloss and determine the set of single phrase definitions
– Replace the stop words with the NEAR operator
– Strengthen the query: concatenate the words from the current synset using
the AND operator
– produce#6 has the synset {grow, raise, farm, produce} and the definition
“cultivate by growing” 
SP=cultivate NEAR growing AND (grow OR raise OR farm OR produce)
130
Heuristics to Identify Monosemous Relatives
• Dictionary definitions (3)
– Parse the gloss and determine the set of single phrase definitions
– Keep only the head phrase
– Strengthen the query: concatenate the words from the current synset using
the AND operator
– company#5 has the synset {party,company} and the definition “band of
people associated in some activity” 
SP=band of people AND (company OR party)
131
Example
• Building annotated corpora for the noun interest.
#
1
2
3
4
5
6
7
S ynset
{ interest# 1 , in vo lve m e nt}
D efinitio n
sense o f co ncern w ith a nd curio sity ab o ut
so m eo ne o r so m ething
{ interest# 2 ,interestin g ness} the p o w er o f attractin g o r ho ld ing o ne ’s interest
{ sake, interest# 3 }
reaso n fo r w a ntin g so m e thin g d o ne
{ interest# 4 }
fixed charge fo r b o rro w in g m o ne y; u sually a
p ercentage o f the a m o unt b o rro w ed
{ p astim e,interest# 5 }
a sub ject o r p ursuit that o ccup ies o ne’s tim e and
tho u gh ts
{ interest# 6 , stake}
a right o r legal share o f so m eth in g; financial
invo lve m en t w ith so m ething
{ interest# 7 , interest gro up } a so cial gro up w ho se m e m b ers co n tro l so m e field
o f activity a nd w ho ha ve co m m o n aim s
S ense #
1
2
3
4
5
6
132
7
S earch p hrase
sense o f co ncern A N D (interest O R invo lve m e nt)
interestig ness
reaso n fo r w a ntin g A N D (in terest O R sake)
fixed charge A N D interest
p ercentage o f a m o u nt A N D interest
p astim e
right share A N D (interest O R sta ke)
legal share A N D (intere st O R stake)
financial in vo lve m e nt A N D (in terest O R sta ke)
interest gro up
Example
• Gather 5,404 examples
• Check the first 70 examples  67 correct; 95.7% accuracy.
1. I appreciate the genuine interest#1 which motivated you to write your
message.
2. The webmaster of this site warrants neither accuracy, nor interest#2.
3. He forgives us not only for our interest#3, but for his own.
4. Interest#4 coverage, including rents, was 3.6x
5. As an interest#5, she enjoyed gardening and taking part into church
activities.
6. Voted on issues, they should have abstained because of direct and indirect
personal interests#6 in the matters of hand.
7. The Adam Smith Society is a new interest#7 organized within the APA.
133
Experimental Evaluation
• Tests on 20 words
– 7 nouns, 7 verbs, 3 adjectives, 3 adverbs (120 word meanings)
– manually check the first 10 examples of each sense of a word
=> 91% accuracy (Mihalcea 1999)
134
W o rd
P o lysem y
co u n t
E xam p les
in S em C o r
T o tal exam p les acq u ired
E xam p les m a- C o rrect
n u ally ch ecked exam p les
interest
rep o rt
co m p any
scho o l
p ro d uce
rem em b er
w rite
sp eak
sm all
clearly
TOTAL
(2 0 w o rd s)
7
7
9
7
7
8
8
4
14
4
120
139
71
90
146
148
166
285
147
192
48
2582
5404
4196
6292
2490
4982
3573
2914
4279
10954
4031
80741
70
70
80
59
67
67
69
40
107
29
1080
67
63
77
54
60
57
67
39
92
28
978
Web-based Bootstrapping
• Similar to Yarowsky algorithm
• Relies on data gathered from the Web
1. Create a set of seeds (phrases) consisting of:
–
–
–
•
Sense tagged examples in SemCor
Sense tagged examples from WordNet
Additional sense tagged examples, if available
Phrase?
–
–
At least two open class words;
Words involved in a semantic relation (e.g. noun phrase, verb-object,
verb-subject, etc.)
2. Search the Web using queries formed with the seed expressions found
at Step 1
–
–
135
Add to the generated corpus of maximum of N text passages
Results competitive with manually tagged corpora (Mihalcea 2002)
The Web as Collective Mind
• Two different views of the Web:
– collection of Web pages
– very large group of Web users
• Millions of Web users can contribute their knowledge to a
data repository
• Open Mind Word Expert (Chklovski and Mihalcea, 2002)
• Fast growing rate:
– Started in April 2002
– Currently more than 100,000 examples of noun senses in several
languages
136
OMWE
online
137
http://teach-computers.org
Open Mind Word Expert: Quantity and Quality
• Data
– A mix of different corpora: Treebank, Open Mind Common Sense, Los
Angeles Times, British National Corpus
• Word senses
– Based on WordNet definitions
• Active learning to select the most informative examples for learning
– Use two classifiers trained on existing annotated data
– Select items where the two classifiers disagree for human annotation
• Quality:
– Two tags per item
– One tag per item per contributor
• Evaluations:
– Agreement rates of about 65% - comparable to the agreements
rates obtained when collecting data for Senseval-2 with trained
lexicographers
– Replicability: tests on 1,600 examples of “interest” led to 90%+
replicability
138
References
•
•
•
•
•
•
•
•
•
•
•
139
(Abney 2002) Abney, S. Bootstrapping. Proceedings of ACL 2002.
(Blum and Mitchell 1998) Blum, A. and Mitchell, T. Combining labeled and unlabeled
data with co-training. Proceedings of COLT 1998.
(Chklovski and Mihalcea 2002) Chklovski, T. and Mihalcea, R. Building a sense tagged
corpus with Open Mind Word Expert. Proceedings of ACL 2002 workshop on WSD.
(Clark, Curran and Osborne 2003) Clark, S. and Curran, J.R. and Osborne, M.
Bootstrapping POS taggers using unlabelled data. Proceedings of CoNLL 2003.
(Mihalcea 1999) Mihalcea, R. An automatic method for generating sense tagged
corpora. Proceedings of AAAI 1999.
(Mihalcea 2002) Mihalcea, R. Bootstrapping large sense tagged corpora. Proceedings
of LREC 2002.
(Mihalcea 2004) Mihalcea, R. Co-training and Self-training for Word Sense
Disambiguation. Proceedings of CoNLL 2004.
(Ng and Cardie 2003) Ng, V. and Cardie, C. Weakly supervised natural language
learning without redundant views. Proceedings of HLT-NAACL 2003.
(Nigam and Ghani 2000) Nigam, K. and Ghani, R. Analyzing the effectiveness and
applicability of co-training. Proceedings of CIKM 2000.
(Sarkar 2001) Sarkar, A. Applying cotraining methods to statistical parsing. Proceedings
of NAACL 2001.
(Yarowsky 1995) Yarowsky, D. Unsupervised word sense disambiguation rivaling
supervised methods. Proceedings of ACL 1995.
Part 6:
Unsupervised Methods of Word
Sense Disambiguation
Outline
•
•
•
•
•
141
What is Unsupervised Learning?
Task Definition
Agglomerative Clustering
LSI/LSA
Sense Discrimination Using Parallel Texts
What is Unsupervised Learning?
• Unsupervised learning identifies patterns in a large sample
of data, without the benefit of any manually labeled
examples or external knowledge sources
• These patterns are used to divide the data into clusters,
where each member of a cluster has more in common with
the other members of its own cluster than any other
• Note! If you remove manual labels from supervised data
and cluster, you may not discover the same classes as in
supervised learning
– Supervised Classification identifies features that trigger a sense tag
– Unsupervised Clustering finds similarity between contexts
142
Cluster this Data!
Facts about my day…
143
Day
F1
Hot Outside?
F2
Slept Well?
F3
Ate Well?
1
YES
NO
NO
2
YES
NO
YES
3
NO
NO
NO
4
NO
NO
YES
Cluster this Data!
Facts about my day…
144
Day
F1
Hot Outside?
F2
Slept Well?
F3
Ate Well?
1
YES
NO
NO
2
YES
NO
YES
3
NO
NO
NO
4
NO
NO
YES
Cluster this Data!
145
Day
F1
Hot Outside?
F2
Slept Well?
F3
Ate Well?
1
YES
NO
NO
2
YES
NO
YES
3
NO
NO
NO
4
NO
NO
YES
Outline
•
•
•
•
•
146
What is Unsupervised Learning?
Task Definition
Agglomerative Clustering
LSI/LSA
Sense Discrimination Using Parallel Texts
Task Definition
• Unsupervised Word Sense Discrimination: A class of
methods that cluster words based on similarity of context
• Strong Contextual Hypothesis
– (Miller and Charles, 1991): Words with similar meanings tend to
occur in similar contexts
– (Firth, 1957): “You shall know a word by the company it keeps.”
• …words that keep the same company tend to have similar meanings
• Only use the information available in raw text, do not use
outside knowledge sources or manual annotations
• No knowledge of existing sense inventories, so clusters are
not labeled with senses
147
Task Definition
• Resources:
– Large Corpora
• Scope:
– Typically one targeted word per context to be discriminated
– Equivalently, measure similarity among contexts
– Features may be identified in separate “training” data, or in the
data to be clustered
– Does not assign senses or labels to clusters
• Word Sense Discrimination reduces to the problem of
finding the targeted words that occur in the most similar
contexts and placing them in a cluster
148
Outline
•
•
•
•
•
149
What is Unsupervised Learning?
Task Definition
Agglomerative Clustering
LSI/LSA
Sense Discrimination Using Parallel Texts
Agglomerative Clustering
• Create a similarity matrix of instances to be discriminated
– Results in a symmetric “instance by instance” matrix, where each
cell contains the similarity score between a pair of instances
– Typically a first order representation, where similarity is based on
the features observed in the pair of instances
• Apply Agglomerative Clustering algorithm to matrix
– To start, each instance is its own cluster
– Form a cluster from the most similar pair of instances
– Repeat until the desired number of clusters is obtained
( X
 Y )
( X
 Y )
150
• Advantages : high quality clustering
• Disadvantages – computationally expensive, must carry
out exhaustive pair wise comparisons
Measuring Similarity
• Integer Values
– Matching Coefficient
– Jaccard Coefficient
– Dice Coefficient
X Y
X Y
X Y
2 X Y
X  Y
• Real Values
– Cosine
151


X  Y
X
Y
Instances to be Clustered
S1
P-2
P-1
P+1
P+2
fish
check
river
interest
adv
det
prep
det
Y
N
Y
N
det
prep
det
N
Y
N
Y
verb
det
Y
N
N
N
N
N
N
N
S2
S3
det
adj
S4
det
noun
noun noun
S1
S1
152
S2
S3
S4
3
4
2
2
0
S2
3
S3
4
2
S4
2
0
1
1
Average Link Clustering
aka McQuitty’s Similarity Analysis
S1
S1
S2
3
S3
4
S4
2
S2
S3
S4
3
4
2
2
0
2
1
0
S1S3
S2
3 2
S1S3
 2 .5
2
1
S2
3 2
S1S3S2
1 .5  1 .5
S1S3S2
S4
2
1 .5  1 .5
2
153
S4
 1 .5
 1 .5
2 1
2
 1 .5
2 1
 1 .5
2
0
 2 .5
2
S4
S4
0
Evaluation of Unsupervised Methods
• If Sense tagged text is available, can be used for evaluation
– But don’t use sense tags for clustering or feature selection!
• Assume that sense tags represent “true” clusters, and
compare these to discovered clusters
– Find mapping of clusters to senses that attains maximum accuracy
• Pseudo words are especially useful, since it is hard to find
data that is discriminated
– Pick two words or names from a corpus, and conflate them into
one name. Then see how well you can discriminate.
– http://www.d.umn.edu/~kulka020/kanaghaName.html
• Baseline Algorithm– group all instances into one cluster,
this will reach “accuracy” equal to majority classifier
154
Baseline Performance
S1
S2
S3
Totals
S3
S2
S1
Totals
C1
0
0
0
0
C1
0
0
0
0
C2
0
0
0
0
C2
0
0
0
0
C3
80
35
55
170
C3
55
35
80
170
Totals
80
35
55
170
Totals
55
35
80
170
(0+0+55)/170 = .32
if C3 is S3
155
(0+0+80)/170 = .47
if C3 is S1
Evaluation
• Suppose that C1 is labeled S1, C2 as S2, and C3 as S3
• Accuracy = (10 + 0 + 10) / 170 = 12%
• Diagonal shows how many members of the cluster actually belong to
the sense given on the column
• Can the “columns” be rearranged to improve the overall accuracy?
– Optimally assign clusters to senses
156
S1
S2
S3
Totals
C1
10
30
5
45
C2
20
0
40
60
C3
50
5
10
65
Totals
80
35
55
170
Evaluation
• The assignment of C1 to S2,
C2 to S3, and C3 to S1
results in 120/170 = 71%
• Find the ordering of the
columns in the matrix that
maximizes the sum of the
diagonal.
• This is an instance of the
Assignment Problem from
Operations Research, or
finding the Maximal
Matching of a Bipartite
Graph from Graph Theory.
157
S2
S3
S1
Totals
C1
30
5
10
45
C2
0
40
20
60
C3
5
10
50
65
Totals
35
55
80
170
Agglomerative Approach
• (Pedersen and Bruce, 1997) explore discrimination with a
small number (approx 30) of features near target word.
–
–
–
–
Morphological form of target word (1)
Part of Speech two words to left and right of target word (4)
Co-occurrences (3) most frequent content words in context
Unrestricted collocations (19) most frequent words located one
position to left or right of target, OR
– Content collocations (19) most frequent content words located one
position to left or right of target
• Features identified in the instances be clustered
• Similarity measured by matching coefficient
• Clustered with McQuitty’s Similarity Analysis, Ward’s
Method, and the EM Algorithm
– Found that McQuitty’s method was the most accurate
158
Experimental Evaluation
•
Adjectives
–
–
–
–
•
•
–
–
–
–
159
•
Agree, 74% majority (1109)
Close, 77% majority (1354)
Help, 78% majority (1267)
Include, 91% majority (1526)
•
Chief, 86%
Common, 80%
Last, 79%
Public, 63%
Nouns
–
–
–
–
–
Bill, 68% majority (1341)
Concern, 64% majority (1235)
Drug, 57% majority (1127)
Interest, 59% majority (2113)
Line, 37% majority (1149)
Verbs
Adjectives
–
–
–
–
Chief, 86% majority (1048)
Common, 84% majority (1060)
Last, 94% majority (3004)
Public, 68% majority (715)
Nouns
–
–
–
–
–
•
Bill, 75%
Concern, 68%
Drug, 65%
Interest, 65%
Line, 42%
Verbs
–
–
–
–
Agree, 69%
Close, 72%
Help, 70%
Include, 77%
Analysis
• Unsupervised methods may not discover clusters
equivalent to the classes learned in supervised learning
• Evaluation based on assuming that sense tags represent the
“true” cluster are likely a bit harsh. Alternatives?
– Humans could look at the members of each cluster and determine
the nature of the relationship or meaning that they all share
– Use the contents of the cluster to generate a descriptive label that
could be inspected by a human
• First order feature sets may be problematic with smaller
amounts of data since these features must occur exactly in
the test instances in order to be “matched”
160
Outline
•
•
•
•
•
161
What is Unsupervised Learning?
Task Definition
Agglomerative Clustering
LSI/LSA
Sense Discrimination Using Parallel Texts
Latent Semantic Indexing/Analysis
• Adapted by (Schütze, 1998) to word sense discrimination
• Represent training data as word co-occurrence matrix
• Reduce the dimensionality of the co-occurrence matrix via
Singular Value Decomposition (SVD)
– Significant dimensions are associated with concepts
• Represent the instances of a target word to be clustered by
taking the average of all the vectors associated with all the
words in that context
– Context represented by an averaged vector
• Measure the similarity amongst instances via cosine and
record in similarity matrix, or cluster the vectors directly
162
Co-occurrence matrix
163
apple
blood
cells
ibm
data
pc
2
0
0
1
3
body
0
3
0
0
disk
1
0
0
petri
0
2
lab
0
sales
box
tissue
graphics
memory
organ
plasma
1
0
0
0
0
0
0
0
2
0
0
2
1
2
0
3
0
1
2
0
0
1
0
0
0
2
0
1
0
1
0
3
0
2
0
2
0
2
1
3
0
0
0
2
3
0
0
1
2
0
0
linux
2
0
0
1
3
2
0
1
1
0
0
debt
0
0
0
2
3
4
0
2
0
0
0
Singular Value Decomposition
A=UDV’
164
U
.35
.09
-.2
.02
.63
.20
-.00
-.02
.08 -.09 -.44
-.04
-.6
-.02
-.01
.41 -.22
.20
-.39
.00
.03
.09
.83
.05
-.26
-.01
.00
.29 -.68 -.45 -.34 -.31 .02 -.21
.01
.43
-.02
-.07
.37 -.01 -.31 .09
.03
.31
-.00
.08
.05
.08
.08
-.00
-.01
.30 -.07 -.49 -.52 .14
-.3
-.30
.00
-.07
.05 -.49 .59
.35
.13
.52 -.09 .40
.44
.39 -.60 .31
.08 -.45 .25 -.02 .17
165
.72 -.48 -.04
.46
.11 -.08 .24 -.01 .39
.56
.25
D
9.19
6.36
3.99
3.25
2.52
2.30
1.26
0.66
0.00
0.00
0.00
166
V
167
.21
.08
-.04
.28
.04
.86
-.05
-.05
-.31
-.12
.03
.04
-.37
.57
.39
.23
-.04
.26
-.02
.03
.25
.44
.11
-.39
-.27
-.32
-.30
.06
.17
.15
-.41
.58
.07
.37
.15
.12
-.12
.39
-.17
-.13
.71
-.31
-.12
.03
.63
-.01
-.45
.52
-.09
-.26
.08
-.06
.21
.08
-.02
.49
.27
.50
-.32
-.45
.13
.02
-.01
.31
.12
-.03
.09
-.51
.20
.05
-.05
.02
.29
.08
-.04
-.31
-.71
.25
.11
.15
-.12
.02
-.32
.05
-.59
-.62
-.23
.07
.28
-.23
-.14
-.45
.64
.17
-.04
-.32
.31
.12
-.03
.04
-.26
.19
.17
-.06
-.07
-.87
-.10
-.07
.22
-.20
.11
-.47
-.12
-.18
-.27
.03
-.18
.09
.12
-.58
.50
Co-occurrence matrix after SVD
168
apple
blood
pc
.73
.00
body
.00
disk
cells
ibm
data
tissue
graphics
memory
organ
plasma
.11 1.3 2.0
.01
.86
.77
.00
.09
1.2
1.3 .00 .33
1.6
.00
.85
.84
1.5
.76
.00
.01 1.3 2.1
.00
.91
.72
.00
.00
germ
.00
1.1
1.2 .00 .49
1.5
.00
.86
.77
1.4
lab
.21
1.7
2.0 .35 1.7
2.5
.18
1.7
1.2
2.3
sales
.73
.15
.39 1.3 2.2
.35
.85
.98
.17
.41
linux
.96
.00
.16 1.7 2.7
.03
1.1
1.0
.00
.13
debt
1.2
.00
.00 2.1 3.2
.00
1.5
1.1
.00
.00
Effect of SVD
• SVD reduces a matrix to a given number of dimensions
This may convert a word level space into a semantic or
conceptual space
– If “dog” and “collie” and “wolf” are dimensions/columns in the
word co-occurrence matrix, after SVD they may be a single
dimension that represents “canines”
• The dimensions are the principle components that may
(hopefully) represent the meaning of concepts
• SVD has effect of smoothing a very sparse matrix, so that
there are very few 0 valued cells
169
Context Representation
• Represent each instance of the target word to be clustered
by averaging the word vectors associated with its context
– This creates a “second order” representation of the context
• The context is represented not only by the words that occur
therein, but also the words that occur with the words in the
context elsewhere in the training corpus
170
Second Order Context Representation
• I got a new disk today!
• What do you think of linux?
apple
blood
cells
disk
.76
.00
.01
linux
.96
.00
.16
ibm
data
tissue
1.3
2.1
.00
1.7
2.7
.03
graphics
memory
organ
Plasma
.91
.72
.00
.00
1.1
1.0
.00
.13
• These two contexts share no words in common, yet they
are similar! disk and linux both occur with “Apple”,
“IBM”, “data”, “graphics”, and “memory”
• The two contexts are similar because they share many
second order co-occurrences
171
Second Order Context Representation
• The bank of the Mississippi River was washed away.
172
First vs. Second Order Representations
• Comparison made by (Purandare and Pedersen, 2004)
• Build word co-occurrence matrix using log-likelihood ratio
– Reduce via SVD
– Cluster in vector or similarity space
– Evaluate relative to manually created sense tags
• Experiments conducted with Senseval-2 data
– 24 words, 50-200 training and test examples
– Second order representation resulted in significantly better
performance than first order, probably due to modest size of data.
• Experiments conducted with line, hard, serve
– 4000-5000 instances, divided into 60-40 training-test split
– First order representation resulted in better performance than
second order, probably due to larger amount of data
173
Analysis
• Agglomerative methods based on direct (first order)
features require large amounts of data in order to obtain a
reliable set of features
• Large amounts of data are problematic for agglomerative
clustering (due to exhaustive comparisons)
• Second order representations allow you to make due with
smaller amounts of data, and still get a rich (non-sparse)
representation of context
• http://senseclusters.sourceforge.net is a complete system
for performing unsupervised discrimination using first or
second order context vectors in similarity or vector space,
and includes support for SVD, clustering and evaluation
174
Outline
•
•
•
•
•
175
What is Unsupervised Learning?
Task Definition
Agglomerative Clustering
LSI/LSA
Sense Discrimination Using Parallel Texts
Sense Discrimination Using Parallel Texts
• There is controversy as to what exactly is a “word sense”
(e.g., Kilgarriff, 1997)
• It is sometimes unclear how fine grained sense distinctions
need to be to be useful in practice.
• Parallel text may present a solution to both problems!
– Text in one language and its translation into another
• Resnik and Yarowsky (1997) suggest that word sense
disambiguation concern itself with sense distinctions that
manifest themselves across languages.
– A “bill” in English may be a “pico” (bird jaw) in or a “cuenta”
(invoice) in Spanish.
176
Parallel Text
• Parallel Text can be found on the Web and there are several
large corpora available (e.g., UN Parallel Text, Canadian
Hansards)
• Manual annotation of sense tags is not required! However,
text must be word aligned (translations identified between
the two languages).
– http://www.cs.unt.edu/~rada/wpt/
Workshop on Parallel Text, NAACL 2003
• Given word aligned parallel text, sense distinctions can be
discovered. (e.g., Li and and Li, 2002, Diab, 2002)
177
References
•
•
•
•
•
•
•
•
•
•
178
(Diab, 2002) Diab, Mona and Philip Resnik, An Unsupervised Method for Word Sense
Tagging using Parallel Corpora, Proceedings of ACL, 2002.
(Firth, 1957) A Synopsis of Linguistic Theory 1930-1955. In Studies in Linguistic
Analysis, Oxford University Press, Oxford.
(Kilgarriff, 1997) “I don’t believe in word senses”, Computers and the Humanities (31)
pp. 91-113.
(Li and Li, 2002) Word Translation Disambiguation Using Bilingual Bootstrapping.
Proceedings of ACL. Pp. 343-351.
(McQuitty, 1966) Similarity Analysis by Reciprocal Pairs for Discrete and Continuous
Data. Educational and Psychological Measurement (26) pp. 825-831.
(Miller and Charles, 1991) Contextual correlates of semantic similarity. Language and
Cognitive Processes, 6 (1) pp. 1 - 28.
(Pedersen and Bruce, 1997) Distinguishing Word Sense in Untagged Text. In
Proceedings of EMNLP2. pp 197-207.
(Purandare and Pedersen, 2004) Word Sense Discrimination by Clustering Contexts in
Vector and Similarity Spaces. Proceedings of the Conference on Natural Language and
Learning. pp. 41-48.
(Resnik and Yarowsky, 1997) A Perspective on Word Sense Disambiguation Methods
and their Evaluation. The ACL-SIGLEX Workshop Tagging Text with Lexical
Semantics. pp. 79-86.
(Schutze, 1998) Automatic Word Sense Discrimination. Computational Linguistics, 24
(1) pp. 97-123.
Part 7:
How to Get Started in
Word Sense Disambiguation
Research
Outline
• Where to get the required ingredients?
–
–
–
–
Machine Readable Dictionaries
Machine Learning Algorithms
Sense Annotated Data
Raw Data
• Where to get WSD software?
• How to get your algorithms tested?
– Senseval
180
Machine Readable Dictionaries
• Machine Readable format (MRD)
– Oxford English Dictionary
– Collins
– Longman Dictionary of Ordinary Contemporary English (LDOCE)
• Thesauri – add synonymy information
– Roget Thesaurus http://www.thesaurus.com
• Semantic networks – add more semantic relations
– WordNet http://www.cogsci.princeton.edu/~wn/
• Dictionary files, source code
– EuroWordNet http://www.illc.uva.nl/EuroWordNet/
• Seven European languages
181
Machine Learning Algorithms
• Many implementations available online
• Weka: Java package of many learning algorithms
– http://www.cs.waikato.ac.nz/ml/weka/
– Includes decision trees, decision lists, neural networks, naïve
bayes, instance based learning, etc.
• C4.5: C implementation of decision trees
– http://www.cse.unsw.edu.au/~quinlan/
• Timbl: Fast optimized implementation of instance based
learning algorithms
– http://ilk.kub.nl/software.html
• SVM Light: efficient implementation of Support Vector
Machines
– http://svmlight.joachims.org
182
Sense Tagged Data
• A lot of annotated data available through Senseval
– http://www.senseval.org
• Data for lexical sample
– English (with respect to Hector, WordNet, Wordsmyth)
– Basque, Catalan, Chinese, Czech, Romanian, Spanish, etc.
– Data produced within Open Mind Word Expert project http://teachcomputers.org
• Data for all words
– English, Italian, Czech (Senseval-2 and Senseval-3)
– SemCor (200,000 running words)
http://www.cs.unt.edu/~rada/downloads.html
• Pointers to additional data available from
– http://www.senseval.org/data.html
183
Sense Tagged Data – Lexical Sample
<instance id="art.40008" docsrc="bnc_ANF_855">
<answer instance="art.40008" senseid="art%1:06:00::"/>
<context>
The evening ended in a brawl between the different factions in Cubism, but it brought a
moment of splendour into the blackouts and bombings of war. [/p] [p]
Yet Modigliani was too much a part of the life of Montparnasse, too involved with the
individuals leading the " new art " , to remain completely aloof.
In 1914 he had met Hans Arp, the French painter who was to become prominent in the
new Dada movement, at the artists' canteen in the Avenue du Maine.
Two years later Arp was living in Zurich, a member of a group of talented emigrant
artists who had left their own countries because of the war.
Through casual meetings at cafes, the artists drew together to form a movement in
protest against the waste of war, against nationalism and against everything
pompous, conventional or boring in the <head>art</head>of the Western world.
</context>
</instance>
184
Sense Tagged Data – SemCor
<p pnum=1>
<s snum=1>
<wf cmd=ignore pos=DT>The</wf>
<wf cmd=done rdf=group pos=NNP lemma=group wnsn=1 lexsn=1:03:00::
pn=group>Fulton_County_Grand_Jury</wf>
<wf cmd=done pos=VB lemma=say wnsn=1 lexsn=2:32:00::>said</wf>
<wf cmd=done pos=NN lemma=friday wnsn=1 lexsn=1:28:00::>Friday</wf>
<wf cmd=ignore pos=DT>an</wf>
<wf cmd=done pos=NN lemma=investigation wnsn=1
lexsn=1:09:00::>investigation</wf>
<wf cmd=ignore pos=IN>of</wf>
<wf cmd=done pos=NN lemma=atlanta wnsn=1 lexsn=1:15:00::>Atlanta</wf>
<wf cmd=ignore pos=POS>'s</wf>
<wf cmd=done pos=JJ lemma=recent wnsn=2 lexsn=5:00:00:past:00>recent</wf>
<wf cmd=done pos=NN lemma=primary_election wnsn=1
lexsn=1:04:00::>primary_election</wf>
<wf cmd=done pos=VB lemma=produce wnsn=4 lexsn=2:39:01::>produced</wf>
…
185
Raw Data
• For use with
– Bootstrapping algorithms
– Word sense discrimination algorithms
• British National Corpus
– 100 million words covering a variety of genres, styles
– http://www.natcorp.ox.ac.uk/
• TREC (Text Retrieval Conference) data
– Los Angeles Times, Wall Street Journal, and more
– 5 gigabytes of text
– http://trec.nist.gov/
• The Web
186
Outline
• Where to get the required ingredients?
–
–
–
–
Machine Readable Dictionaries
Machine Learning Algorithms
Sense Annotated Data
Raw Data
• Where to get WSD software?
• How to get your algorithms tested?
– Senseval
187
WSD Software – Lexical Sample
•
Duluth Senseval-2 systems
– Lexical decision tree systems that participated in Senseval-2 and 3
– http://www.d.umn.edu/~tpederse/senseval2.html
•
SyntaLex
– Enhance Duluth Senseval-2 with syntactic features, participated in Senseval-3
– http://www.d.umn.edu/~tpederse/syntalex.html
•
WSDShell
– Shell for running Weka experiments with wide range of options
– http://www.d.umn.edu/~tpederse/wsdshell.html
•
SenseTools
– For easy implementation of supervised WSD, used by the above 3 systems
– Transforms Senseval-formatted data into the files required by Weka
– http://www.d.umn.edu/~tpederse/sensetools.html
•
SenseRelate::TargetWord
– Identifies the sense of a word based on the semantic relation with its neighbors
– http://search.cpan.org/dist/WordNet-SenseRelate-TargetWord
– Uses WordNet::Similarity – measures of similarity based on WordNet
• http://search.cpan.org/dist/WordNet-Similarity
188
WSD Software – All Words
• SenseLearner
–
–
–
–
A minimally supervised approach for all open class words
Extension of a system participating in Senseval-3
http://lit.csci.unt.edu/~senselearner
Demo on Sunday, June 26 (1:30-3:30)
• SenseRelate::AllWords
– Identifies the sense of a word based on the semantic relation with
its neighbors
– http://search.cpan.org/dist/WordNet-SenseRelate-AllWords
– Demo on Sunday, June 26 (1:30-3:30)
189
WSD Software – Unsupervised
• Clustering by Committee
– http://www.cs.ualberta.ca/~lindek/demos/wordcluster.htm
• InfoMap
– Represent the meanings of words in vector space
– http://infomap-nlp.sourceforge.net
• SenseClusters
– Finds clusters of words that occur in similar context
– http://senseclusters.sourceforge.net
– Demo Sunday, June 26 (4:00-6:00)
190
Outline
• Where to get the required ingredients?
–
–
–
–
Machine Readable Dictionaries
Machine Learning Algorithms
Sense Annotated Data
Raw Data
• Where to get WSD software?
• How to get your algorithms tested?
– Senseval
191
Senseval
•
•
•
•
•
Evaluation of WSD systems http://www.senseval.org
Senseval 1: 1999 – about 10 teams
Senseval 2: 2001 – about 30 teams
Senseval 3: 2004 – about 55 teams
Senseval 4: 2007(?)
• Provides sense annotated data for many languages, for
several tasks
– Languages: English, Romanian, Chinese, Basque, Spanish, etc.
– Tasks: Lexical Sample, All words, etc.
192
• Provides evaluation software
• Provides results of other participating systems
Senseval
S ens ev al E va luatio ns
160
140
120
100
T as k s
T e am s
80
S y s te m s
60
40
20
0
S e n se va l 1
193
S e n se va l 2
S e n se va l 3
Part 8:
Conclusions
Outline
•
•
•
•
195
The Web and WSD
Multilingual WSD
The Next Five Years (2005-2010)
Concluding Remarks
The Web and WSD
• The Web has become a source of data for NLP in general,
and word sense disambiguation is no exception.
• Can find hundreds/thousands(?) of instances of a particular
target word just by searching.
• Search Engines :
– Alta Vista – allows scraping, at a modest rate. Insert 5 second
delays on your queries to Alta-Vista so as to not overwhelm the
system. No API provided, but Perl::LWP works nicely.
http://search.cpan.org/dist/libwww-perl/
– Google – does not allow scraping, but provides an API to access
search engine. However, the API limits you to 1,000 queries per
day.
http://www.google.com/apis/
196
The Web and WSD
• The Web can search as a good source of information for
selecting or verifying collocations and other kinds of
association.
– “strong tea” : 13,000 hits
– “powerful tea” : 428 hits
– “sparkling tea” : 376 hits
197
The Web and WSD
• You can find sets of related words from the Web.
– http://labs.google.com/sets
– Give Google Sets two or three words, it will return a set of words it
believes are related
– Could be the basis of extending features sets for WSD, since many
times the words are related in meaning
• Google Sets Input: bank, credit
• Google Sets Output: bank, credit, stock, full, investment, invoicing,
overheads, cash low, administration, produce service, grants, overdue
notices
– Great source of info about names or current events
• Google Sets Input: Nixon, Carter
• Google Sets Output: Carter, Nixon, Reagan, Ford, Bush, Eisenhower,
Kennedy, Johnson
198
A Natural Problem for the Web and WSD
• Organize Search Results by concepts, not just names.
– Separate the Irish Republican Army (IRA) from the Individual
Retirement Account (IRA).
• http://clusty.com is an example of a web site that attempts
to cluster content.
– Finds a set of pages, and labels them with some descriptive term.
– Very similar to problem in word sense discrimination, where
cluster is not associated with a known sense.
199
The Web and WSD, not all good news
• Lots and lots of junk to filter through. Lots of misleading
and malicious content on web pages.
• Counts as reported by search engines for hits are
approximations and vary sometime from query to query.
Over time they will change, so it’s very hard to reproduce
experimental results over time.
• Search engines could close down API, prohibit scraping,
etc. – there are no promises made.
• Can be slow to get data from the Web.
200
Outline
•
•
•
•
201
The Web and WSD
Multilingual WSD
The Next Five Years (2005-2010)
Concluding Remarks
Multilingual WSD
• Parallel text is a potential meeting ground between raw
untagged text (like unsupervised methods use) and sense
tagged text (like the supervised methods need)
• A source language word that is translated into various
different target language forms may be polysemous in the
source language
202
A Clever Way to Sense Tag
• Expertise of native speakers can be used to create sense
tagged text, without having to refer to dictionaries!
• Have a bilingual native speaker pick the proper translation
for a word in a given context.
http://www.teach-computers.org/word-expert/english-hindi/
http://www.teach-computers.org/word-expert/english-french/
• This is a much more intuitive way to sense tag text, and
depends only on the native speakers expertise, not a set of
senses as found in a particular dictionary.
203
Outline
•
•
•
•
204
The Web and WSD
Multilingual WSD
The Next Five Years (2005-2010)
Concluding Remarks
The Next Five Years
• Applications, applications, applications, and applications.
Where are the applications? WSD needs to demonstrate an
impact on applications in the next five years.
• Word Sense Disambiguation will be deployed in an
increasing number of applications over the next five years.
– However, not in Machine Translation. Too difficult to integrate
WSD into current statistical systems, and this won’t change soon.
– Most likely applications include web search tools and email
organizers and search tools (like gmail).
• If you are writing papers, “bake off” evaluations will meet
with more rejection that acceptance
• If you have a potential application for Word Sense
Disambiguation in any of its forms, tell us!! Please!
205
Outline
•
•
•
•
206
The Web and WSD
Multilingual WSD
The Next Five Years (2005-2010)
Concluding Remarks
Concluding Remarks
• Word Sense Disambiguation has something for everyone!
–
–
–
–
–
–
–
Statistical Methods
Knowledge Based systems
Supervised Machine Learning
Unsupervised Learning
Semi-Supervised
Bootstrapping and Co-training
Human Annotation of Data
• The impact of high quality WSD will be huge. NLP
consumers have become accustomed to systems that only
make coarse grained distinctions between concepts, or who
might not make any at all.
• Real Understanding? Real AI?
207
Thank You!
• Rada Mihalcea ([email protected])
– http://www.cs.unt.edu/~rada
• Ted Pedersen ([email protected])
– http://www.d.umn.edu/~tpederse
208
Descargar

Notebook - University of Minnesota Duluth Welcomes You