Bootstrapping Distributional Feature
Vector Quality
Maayan Zhitomirsky-Geffet∗
Bar-Ilan University
Ido Dagan∗∗
Bar-Ilan University
This article presents a novel bootstrapping approach for improving the quality of feature vector
weighting in distributional word similarity. The method was motivated by attempts to utilize
distributional similarity for identifying the concrete semantic relationship of lexical entailment.
Our analysis revealed that a major reason for the rather loose semantic similarity obtained by
distributional similarity methods is insufficient quality of the word feature vectors, caused by
deficient feature weighting. This observation led to the definition of a bootstrapping scheme
which yields improved feature weights, and hence higher quality feature vectors. The under-
lying idea of our approach is that features which are common to similar words are also most
characteristic for their meanings, and thus should be promoted. This idea is realized via a
bootstrapping step applied to an initial standard approximation of the similarity space. Le
superior performance of the bootstrapping method was assessed in two different experiments,
one based on direct human gold-standard annotation and the other based on an automatically
created disambiguation dataset. These results are further supported by applying a novel quanti-
tative measurement of the quality of feature weighting functions. Improved feature weighting
also allows massive feature reduction, which indicates that the most characteristic features
for a word are indeed concentrated at the top ranks of its vector. Enfin, experiments with
three prominent similarity measures and two feature weighting functions showed that the
bootstrapping scheme is robust and is independent of the original functions over which it is
appliqué.
1. Introduction
1.1 Motivation
Distributional word similarity has long been an active research area (Hindle 1990; Ruge
1992; Grefenstette 1994; Lee 1997; Lin 1998; Dagan, Lee, and Pereira 1999; Weeds and
∗ Department of Information Science, Bar-Ilan University, Ramat-Gan, Israel.
E-mail: zhitomim@mail.biu.ac.il.
∗∗ Department of Computer Science, Bar-Ilan University, Ramat-Gan, Israel. E-mail: dagan@cs.biu.ac.il.
Submission received: 6 Décembre 2006; revised submission received: 9 Juillet 2008; accepted for publication:
21 Novembre 2008.
© 2009 Association for Computational Linguistics
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
Weir 2005). This paradigm is inspired by Harris’s distributional hypothesis (Harris
1968), which states that semantically similar words tend to appear in similar contexts.
In a computational realization, each word is characterized by a weighted feature vector,
where features typically correspond to other words that co-occur with the characterized
word in the context. Distributional similarity measures quantify the degree of similarity
between a pair of such feature vectors. It is then assumed that two words that occur
within similar contexts, as measured by similarity of their context vectors, are indeed
semantically similar.
The distributional word similarity measures were often applied for two types of
inferences. The first type is making similarity-based generalizations for smoothing
word co-occurrence probabilities, in applications such as language modeling and dis-
ambiguation. Par exemple, assume that we need to estimate the likelihood of the verb–
object co-occurrence pair visit–country, although it did not appear in our sample corpus.
Co-occurrences of the verb visit with words that are distributionally similar to country,
such as state, city, and region, cependant, do appear in the corpus. Par conséquent, nous
may infer that visit–country is also a plausible expression, using some mathematical
scheme of similarity-based generalization (Essen and Steinbiss 1992; Dagan, Marcus,
and Markovitch 1995; Karov and Edelman 1996; Ng and Lee 1996; Ng 1997; Dagan,
Lee, and Pereira 1999; Lee 1999; Weeds and Weir 2005). The rationale behind this
inference is that if two words are distributionally similar then the occurrence of one
word in some contexts indicates that the other word is also likely to occur in such
contexts.
A second type of semantic inference, which primarily motivated our own research,
is meaning-preserving lexical substitution. Many NLP applications, such as question
answering, information retrieval, information extraction, et (multi-document) sum-
marization, need to recognize that one word can be substituted by another one in a
given context while preserving, or entailing the original meaning. Naturellement, recogniz-
ing such substitutable lexical entailments is a prominent component within the textual
entailment recognition paradigm, which models semantic inference as an application-
independent task (Dagan, Glickman, and Magnini 2006). Accordingly, several textual
entailment systems did utilize the output of distributional similarity measures to model
entailing lexical substitutions (Jijkoun and de Rijke 2005; Adams 2006; Ferrandez et al.
2006; Nicholson, Stokes, and Baldwin 2006; Vanderwende, Menezes, and Snow 2006).
In some of these papers the distributional information typically complements man-
ual lexical resources in textual entailment systems, most notably WordNet (Fellbaum
1998).
Lexical substitution typically requires that the meaning of one word entails
the meaning of the other. Par exemple, in question answering, the word company
in a question can be substituted in an answer text by firm, automaker, or subsidiary,
whose meanings entail the meaning of company. Cependant, as it turns out, traditional
distributional similarity measures do not capture well such lexical substitution
relationships, but rather capture a somewhat broader (and looser) notion of semantic
similarité. Par exemple, quite distant co-hyponyms such as party and company
also come out as distributionally similar to country, due to a partial overlap of
their semantic properties. Clairement, the meanings of these words do not entail each
other.
Motivated by these observations, our long-term goal is to investigate whether the
distributional similarity scheme may be improved to yield tighter semantic similarities,
and eventually better approximation of lexical entailments. This article presents one
component of this research plan, which focuses on improving the underlying semantic
436
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
quality of distributional word feature vectors. The article describes the methodology,
definitions, and analysis of our investigation and the resulting bootstrapping scheme
for feature weighting which yielded improved empirical performance.
1.2 Main Contributions and Outline
As a starting point for our investigation, an operational definition was needed for
evaluating the correctness of candidate pairs of similar words. Following the lexical
substitution motivation, in Section 3 we formulate the substitutable lexical entailment
relation (or lexical entailment, for brevity), refining earlier definitions in Geffet and
Dagan (2004, 2005). Generally speaking, this relation holds for a pair of words if a
possible meaning of one word entails a meaning of the other, and the entailing word can
substitute the entailed one in some typical contexts. Lexical entailment overlaps partly
with traditional lexical semantic relationships, while capturing more generally the
lexical substitution needs of applications. Empirically, high inter-annotator agreement
was obtained when judging the output of distributional similarity measures for lexical
entailment.
Suivant, we analyzed the typical behavior of existing word similarity measures relative
to the lexical entailment criterion. Choosing the commonly used measure of Lin (1998)
as a representative case, the analysis shows that quite noisy feature vectors are a major
cause for generating rather “loose” semantic similarities. On the other hand, one may
expect that features which seem to be most characteristic for a word’s meaning should
receive the highest feature weights. This does not seem to be the case, cependant, pour
common feature weighting functions, such as Point-wise Mutual Information (Church
and Patrick 1990; Hindle 1990).
Following these observations, we developed a bootstrapping formula that improves
the original feature weights (Section 4), leading to better feature vectors and better
similarity predictions. The general idea is to promote the weights of features that are
common for semantically similar words, since these features are likely to be most char-
acteristic for the word’s meaning. This idea is implemented by a bootstrapping scheme,
where the initial (and cruder) similarity measure provides an initial approximation for
semantic word similarity. The bootstrapping method yields a high concentration of
semantically characteristic features among the top-ranked features of the vector, lequel
also allows aggressive feature reduction.
The bootstrapping scheme was evaluated in two experimental settings, which cor-
respond to the two types of applications for distributional similarity. D'abord, it achieved
significant improvements in predicting lexical entailment as assessed by human judg-
ments, when applied over several base similarity measures (Section 5). Additional
analysis relative to the lexical entailment dataset revealed cleaner and more charac-
teristic feature vectors for the bootstrapping method. To obtain a quantitative analysis
of this behavior, we defined a measure called average common-feature rank ratio.
This measure captures the idea that a prominent feature for a word is expected to be
prominent also for semantically similar words, while being less prominent for unrelated
words. To the best of our knowledge this is the first proposed measure for direct analysis
of the quality of feature weighting functions, without the need to employ them within
some vector similarity measure.
As a second evaluation, we applied the bootstrapping scheme for similarity-based
prediction of co-occurrence likelihood within a typical pseudo-word sense disambigua-
tion experiment, obtaining substantial error reductions (Section 7). Section 8 concludes
437
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
this article, suggesting the relevance of our analysis and bootstrapping scheme for the
general use of distributional feature vectors.1
2. Background: Distributional Similarity Models
This section reviews the components of the distributional similarity approach and
specifies the measures and functions that were utilized by our work.
The Distributional Hypothesis assumes that semantically similar words appear in
similar contexts, suggesting that semantic similarity can be detected by comparing
contexts of words. This is the underlying principle of the vector-based distributional
similarity model, which comprises two phases. D'abord, context features for each word are
constructed and assigned weights; alors, the weighted feature vectors of pairs of words
are compared by a vector similarity measure. The following two subsections review
typical methods for each phase.
2.1 Features and Weighting Functions
In the typical computational setting, word contexts are represented by feature vectors.
A feature represents another word (or term) w(cid:1) with which w co-occurs, and possibly
specifies also the syntactic relationship between the two words, as in Grefenstette (1994),
Lin (1998), and Weeds and Weir (2005). Ainsi, a word (or term) w is represented by
a feature vector, where each entry in the vector corresponds to a feature f . Pado and
Lapata (2007) demonstrate that using syntactic dependency-based features helps to
distinguish among classes of lexical relations, which seems to be more difficult when
using “bag of words” features that are based on co-occurrence in a text window.
A syntactic-based feature f for a word w is defined as a triple:
(cid:2) fw, syn rel, f role(cid:3)
where fw is a context word (or term) that co-occurs with w under the syntactic depen-
dency relation syn rel. The feature role ( f role) corresponds to the role of the feature word
fw in the syntactic dependency, being either the head (denoted h) or the modifier (de-
noted m) of the relation. Par exemple, given the word company, the feature (cid:2)earnings, gen,
h(cid:3) corresponds to the genitive relationship company’s earnings, et (cid:2)investor, pcomp of, m(cid:3)
corresponds to the prepositional complement relationship the company of the investor.2
Throughout this article we use syntactic dependency relationships generated by the
Minipar dependency parser (Lin 1993). Tableau 1 lists common Minipar dependency
relations involving nouns. Minipar also identifies multi-word expressions, which is
1 A preliminary version of the bootstrapping method was presented in Geffet and Dagan (2004). That
paper presented initial results for the bootstrapping scheme, when applied only over Lin’s measure and
tested by the manually judged dataset of lexical entailment. The current research extends our initial
results in many respects. It refines the definition of lexical entailment; utilizes a revised test set of larger
scope and higher quality, annotated by three assessors; extends the experiments to two additional
similarity measures; provides comparative qualitative and quantitative analysis of the bootstrapped
vectors, while employing our proposed average common-feature rank ratio; and presents an additional
evaluation based on a pseudo-WSD task.
2 Following a common practice, we consider the relationship between a head noun (company in the
example) and the nominal complement of a modifying prepositional phrase (investor) as a single direct
dependency relationship. The preposition itself is encoded in the dependency relation name, with a
distinct relation for each preposition.
438
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
Tableau 1
Common grammatical relations of Minipar involving nouns.
Relation
appo
comp1
det
gen
mod
pnmod
pcomp
post
vrel
obj
obj2
subj
s
Description
apposition
first complement
determiner
genitive marker
the relationship between a word and its adjunct modifier
post nominal modifier
nominal complement of prepositions
post determiner
passive verb modifier of nouns
object of verbs
second object of ditransitive verbs
subject of verbs
surface subject
advantageous for detecting distributional similarity for such terms. Par exemple,
Curran (2004) reports that multi-word expressions make up between 14–25% of the
synonyms in a gold-standard thesaurus.
in the corpus. F
Ainsi, in our representation the corpus is first transformed to a set S of dependency
relationship instances of the form (cid:2)w,F (cid:3), where each pair corresponds to a single co-
occurrence of w and f
is termed as a feature of w. Alors, a word
w is represented by a feature vector, where each entry in the vector corresponds to
one feature f . The value of the entry is determined by a feature weighting function
weight(w, F ), which quantifies the degree of statistical association between w and f in the
set S. Par exemple, some feature weighting functions are based on the logarithm of the
word–feature co-occurrence frequency (Ruge 1992), or on the conditional probability of
the feature given the word (Pereira, Tishby, and Lee 1993; Dagan, Lee, and Pereira 1999;
Lee 1999).
Probably the most widely used feature weighting function is (point-wise) Mutual
Information (MI) (Church and Patrick 1990; Hindle 1990; Luk 1995; Lin 1998; Gauch,
Wang, and Rachakonda 1999; Dagan 2000; Baroni and Vegnaduzzo 2004; Chklovski
and Pantel 2004; Pantel and Ravichandran 2004; Pantel, Ravichandran, and Hovy 2004;
Weeds, Weir, and McCarthy 2004), defined by:
weightMI(w, F ) = log2
P.(w,F )
P.(w)P.( F )
(1)
We calculate the MI weights by the following statistics in the space of co-occurrence
instances S:
weightMI(w, F ) = log2
count(w, F ) · nrels
count(w) · count( F )
(2)
where count(w, F ) is the frequency of the co-occurrence pair (cid:2)w,F (cid:3) in S, count(w) et
count( F ) are the independent frequencies of w and f in S, and nrels is the size of S. High
MI weights are assumed to correspond to strong word–feature associations.
439
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
Curran and Moens (2002) argue that, generally, informative features are statis-
tically correlated with their corresponding headword. Ainsi, they suggest that any
statistical test used for collocations is a good starting point for improving feature-
weight functions. In their experiments the t-test-based metric yielded the best empirical
performance.
Cependant, a known weakness of MI and most of the other statistical weighting
functions used for collocation extraction, including t-test and χ2, is their tendency to
inflate the weights for rare features (Dunning 1993). En outre, a major property of
lexical collocations is their “non-substitutability”, as termed in Manning and Schutze
(1999). C'est, typically neither a headword nor a modifier in the collocation can be
substituted by their synonyms or other related terms. This implies that using modifiers
within strong collocations as features for a head word would provide a rather small
amount of common features for semantically similar words. Ainsi, these functions
seem less suitable for learning broader substitutability relationships, such as lexical
entailment.
Similarity measures that utilize MI weights showed good performance, cependant.
En particulier, a common practice is to filter out features by minimal frequency and
weight thresholds. Alors, a word’s vector is constructed from the remaining (not filtered)
features that are strongly associated with the word. These features are denoted here as
active features.
In the current work we use MI for data analysis, and for the evaluations of vector
quality and word similarity performance.
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
2.2 Vector Similarity Measures
Once feature vectors have been constructed the similarity between two words is de-
fined by some vector similarity measure. Similarity measures which have been used
in the cited papers include weighted Jaccard (Grefenstette 1994), Cosine (Ruge 1992),
and various information theoretic measures, as introduced and reviewed in Lee (1997,
1999). In the current work we experiment with the following three popular similarity
measures.
1.
The basic Jaccard measure compares the number of common features with
the overall number of features for a pair of words. One of the weighted
generalizations of this scheme to non-binary values replaces intersection
with minimum weight, union with maximum weight, and set cardinality
with summation. This measure is commonly referred to as weighted Jaccard
(WJ) (Grefenstette 1994; Dagan, Marcus, and Markovitch 1995; Dagan
2000; Gasperin and Vieira 2004), defined as follows:
simWJ(w, v) =
(cid:1)
(cid:1)
f ∈F(w)∩F(v) min(weight(w,F ),weight(v,F ))
f ∈F(w)∪F(v) maximum(weight(w,F ),weight(v,F ))
(3)
where F(w) and F(v) are the sets of active features of the two words w
and v. The appealing property of this measure is that it considers the
association weights rather than just the number of common features.
2.
The standard Cosine measure (COS), which is popularly employed for
information retrieval (Salton and McGill 1983) and also utilized for
440
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
learning distributionally similar words (Ruge 1992; Caraballo 1999; Gauch,
Wang, and Rachakonda 1999; Pantel and Ravichandran 2004), is defined as
follows:
simCOS(w, v) =
(cid:1)
√(cid:1)
F (weight(w,F )·weight(v,F ))
√(cid:1)
F (weight(w,F ))2·
F (weight(v,F ))2
(4)
This measure computes the cosine of the angle between the two feature
vectors, which normalizes the vector lengths and thus avoids inflated
discrimination between vectors of significantly different lengths.
3.
A popular state of the art measure has been developed by Lin (1998),
motivated by Information Theory principles. This measure behaves quite
similarly to the weighted Jaccard measure (Weeds, Weir, and McCarthy
2004), and is defined as follows:
simLIN(w, v) =
(cid:1)
(cid:1)
f ∈F(w)∩F(v)(weightMI (w,F )+weightMI (v,F ))
f ∈F(w) weightMI (w,F )+
f ∈F(v) weightMI (v,F )
(cid:1)
(5)
where F(w) and F(v) are the active features of the two words. The weight
function used originally by Lin is MI (Équation 1).
It is interesting to note that a relatively recent work by Weeds and Weir (2005) inves-
tigates a more generic similarity framework. Within their framework, the similarity of
two nouns is viewed as the ability to predict the distribution of one of them based on
that of the other. Their proposed formula combines the precision and recall of a potential
“retrieval” of similar words based on the features of the target word. The precision of
w’s prediction of v’s feature distribution indicates how many of the features of the word
w co-occurred with the word v. The recall of w’s prediction of v’s features indicates
how many of the features of v co-occurred with w. Words with both high precision
and high recall can be obtained by computing their harmonic mean, mh (or F-score),
and a weighted arithmetic mean. Cependant, after empirical tuning of weights for the
arithmetic mean, Weeds and Weir’s formula practically reduces to Lin’s measure, comme
was anticipated by their own analysis (in Section 4 of their paper).
Par conséquent, we choose the Lin measure (Équation 5) (henceforth denoted as LIN)
as representative for the state of the art and utilize it for data analysis and as a starting
point for improvement. To further explore and evaluate our new weighting scheme,
independently of a single similarity measure, we conduct evaluations also with the
other two similarity measures of weighted Jaccard and Cosine.
3. Substitutable Lexical Entailment
As mentioned in the Introduction, the long term research goal which inspired our work
is modeling meaning–entailing lexical substitution. Motivated by this goal, nous
proposed in earlier work (Geffet and Dagan 2004, 2005) a new type of lexical
relationship which aims to capture such lexical substitution needs. Here we adopt that
approach and formulate a refined definition for this relationship, termed substitutable
lexical entailment. In the context of the current article, utilizing a concrete target notion
of word similarity enabled us to apply direct human judgment for the “correctness”
(relative to the defined notion) of candidate word pairs suggested by distributional
similarité. Utilizing these judgments we could analyze the behavior of alternative
441
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
distributional vector representations and, in particular, conduct error analysis for word
pair candidates that were judged negatively.
The discussion in the Introduction suggested that multiple text understanding
applications need to identify term pairs whose meanings are both entailing and sub-
stitutable. Such pairs seem to be most appropriate for lexical substitution in a meaning
preserving scenario. To model this goal we present an operational definition for a lexical
semantic relationship that integrates the two aspects of entailment and substitutability,3
which is termed substitutable lexical entailment (or lexical entailment, for brevity).
This relationship holds for a given directional pair of terms (w, v), saying that w entails
v, if the following two conditions are fulfilled:
1. Word meaning entailment: the meaning of a possible sense of w implies a
possible sense of v;
2.
Substitutability: w can substitute for v in some naturally occurring sentence,
such that the meaning of the modified sentence would entail the meaning
of the original one.
To operationally assess the first condition (by annotators) we propose considering
the meaning of terms by existential statements of the form “there exists an instance of
the meaning of the term w in some context” (notice that, unlike propositions, it is not
intuitive for annotators to assign truth values to terms). Par exemple, the word company
would correspond to the existential statement “there exists an instance of the concept
company in some context.” Thus, if in some context “there is a company” (in the sense
of “commercial organization”) then necessarily “there is a firm” in that context (dans le
corresponding sense). Donc, we conclude that the meaning of company implies the
meaning of firm. On the other hand, “there is an organization” does not necessarily imply
the existence of company, since organization might stand for some non-profit association,
aussi. Donc, we conclude that organization does not entail company.
To assess the second condition, the annotators need to identify some natural con-
text in which the lexical substitution would satisfy entailment between the modified
sentence and the original one. Practically, in our experiments presented in Section 5 le
human assessors could consult external lexical resources and the entire Web to obtain all
the senses of the words and possible sentences for substitution. We note that the task of
identifying the common sense of two given words is quite easy since they mutually dis-
ambiguate each other, and once the common sense is known it naturally helps finding a
corresponding common context. We note that this condition is important, in particular,
in order to eliminate cases of anaphora and co-reference in contexts, where two words
quite different in their meaning can sometimes appear in the same contexts only due
to the text pragmatics in a particular situation. Par exemple, in some situations worker
and demonstrator could be used interchangeably in text, but clearly it is a discourse co-
reference rather than common meaning that makes the substitution possible. Plutôt,
we are interested in identifying word pairs in which one word’s meaning provides
a reference to the entailed word’s meaning. This purpose is exactly captured by the
existential propositions of the first criterion above.
3 The WordNet definition of the lexical entailment relation is specified only for verbs and, donc, is not
felicitous for general purposes: A verb X entails Y if X cannot be done unless Y is, or has been, fait (par exemple.,
snore and sleep).
442
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
As reported further in Section 5.1, we observed that assessing these two conditions
for candidate word similarity pairs was quite intuitive for annotators, and yielded good
cross-annotator agreement. Dans l'ensemble, substitutable lexical entailment captures directly
the typical lexical substitution scenario in text understanding applications, as well as in
generic textual entailment modeling. En fait, this relation partially overlaps with several
traditional lexical semantic relations that are known as relevant for lexical substitution,
such as synonymy, hyponymy, and some cases of meronymy. Par exemple, we say
that the meaning of company is lexically entailed by the meaning of firm (synonym)
or automaker (hyponym), while the word government entails minister (meronym) as The
government voted for the new law entails A minister in the government voted for the new law.
On the other hand, lexical entailment is not just a superset of other known relations,
but it is rather designed to select those sub-cases of other lexical relations that are needed
for applied entailment inference. Par exemple, lexical entailment does not cover all cases
of meronyms (par exemple., division does not entail company), but only some sub-cases of part-
whole relationship mentioned herein. En outre, some other relations are also covered
by lexical entailment, like ocean and water and murder and death, which do not seem to
directly correspond to meronymy or hyponymy relations.
Notice also that whereas lexical entailment is a directional relation that specifies
which word of the pair entails the other, the relation may hold in both directions
for a pair of words, as is the case for synonyms. More detailed motivations for the
substitutable lexical entailment relation and analysis of its relationship to traditional
lexical semantic relations appear in Geffet (2006) and Geffet and Dagan (2004, 2005).
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
4. Bootstrapping Feature Weights
To gain a better understanding of distributional similarity behavior we first analyzed
the output of the LIN measure, as a representative case for the state of the art, et
regarding lexical entailment as a reference evaluation criterion. We judge as correct,
with respect to lexical entailment, those candidate pairs of the distributional similarity
method for which entailment holds at least in one direction.
Par exemple, the word area is entailed by country, since the existence of country
entails the existence of area, and the sentence There is no rain in subtropical countries during
the summer period entails the sentence There is no rain in subtropical areas during the summer
period. As another example, democracy is a type of country in the political sense, thus the
existence entailment holds and also the sentence Israel is a democracy in the Middle East
entails Israel is a country in the Middle East.
On the other hand, our analysis revealed that many candidate word similarity pairs
suggested by distributional similarity measures do not correspond to “tight” semantic
relationships. En particulier, many word pairs suggested by the LIN measure do not
satisfy the lexical entailment relation, as demonstrated in Table 2.
A deeper look at the corresponding word feature vectors reveals typical reasons
for these lexical entailment prediction errors. Most relevant for the scope of the cur-
rent article, in many cases highly ranked features in a word vector (when sorting the
features by their weight) do not seem very characteristic for the word meaning. Ce
is demonstrated in Table 3, which shows the top 10 features in the vector for country.
As can be seen, some of the top features are either too specific (landlocked, airspace),
and are thus less reliable, or too general (destination, ambition), thus not indicative and
may co-occur with many different types of words. On the other hand, intuitively more
characteristic features of country, like population and governor, occur further down the
443
Computational Linguistics
Volume 35, Nombre 3
Tableau 2
The top 20 most similar words for country (and their ranks) in the similarity list of LIN, followed
by the next four words in the similarity list that were judged as entailing at least in one direction.
nation
region
state
*monde
*island
*province
1
2
3
4
5
6
*city
territoire
area
*town
republic
*north
7
8
9
10
11
12
economy
*neighbor
*sector
*member
*party
government
13
14
15
16
17
18
*entreprise
*industry
kingdom
place
colony
democracy
19
20
30
35
41
82
Twelve out of 20 top similarities (60%) were judged as mutually non-entailing and are marked
with an asterisk. The similarity data was produced as described in Section 5.
Tableau 3
The top 10 ranked features for country produced by MI, the weighting function employed in the
LIN method.
Feature
weightMI
Commercial bank, gen, h
Destination, pcomp of, m
Airspace, pcomp of, h
Landlocked, mod, m
Trade balance, gen, h
Sovereignty, pcomp of, h
Ambition, nn, h
Bourse, gen, h
Politician, gen, h
Border, pcomp of, h
8.08
7.97
7.83
7.79
7.78
7.78
7.77
7.72
7.54
7.53
sorted feature list, at positions 461 et 832. Dans l'ensemble, features that seem to characterize
the word meaning well are scattered across the ranked feature list, while many non-
indicative features receive high weights. This behavior often yields high similarity
scores for word pairs whose semantic similarity is rather loose while missing some
much tighter similarities.
En outre, we observed that characteristic features for a word w, which should
receive higher weights, are expected to be common for w and other words that are
semantically similar to it. This observation suggests a computational scheme which
would promote the weights of features that are common for semantically similar words.
Bien sûr, there is an inherent circularity in such a scheme: to determine which features
should receive high weights we need to know which words are semantically similar,
while computing distributional semantic similarity already requires pre-determined
feature weights.
This kind of circularity can be approached by a bootstrapping scheme. We first
compute initial distributional similarity values, based on an initial feature weighting
fonction. Alors, to learn more accurate feature weights for a word w, we promote
features that characterize other words that are initially known to be similar to w. Par
the same rationale, features that do not characterize many words that are sufficiently
similar to w are demoted. Even if such features happen to have a strong direct statis-
tical association with w they would not be considered reliable, because they are not
supported by additional words that have a similar meaning to that of w.
444
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
4.1 Bootstrapped Feature Weight Definition
The bootstrapped feature weight is defined as follows. D'abord, some standard word
similarity measure sim is computed to obtain an initial approximation of the similarity
espace. Alors, we define the word set of a feature f , denoted by WS( F ), as the set of words
for which f is an active feature. Recall from Section 2.2 that an active feature is a feature
that is strongly associated with the word, c'est, its (initial) weight is higher than an
empirically predefined threshold, θweight. The semantic neighborhood of w, denoted by
N(w), is defined as the set of all words v which are considered sufficiently similar to
w, satisfying sim(w, v) > θsim, where θsim is a second empirically determined threshold.
The bootstrapped feature weight, denoted weightB, is then defined by:
weightB(w, F ) =
(cid:1)
v∈WS( F )∩N(w) sim(w, v)
(6)
C'est, we identify all words v that are in the semantic neighborhood of w and are also
characterized by f , and then sum the values of their similarities to w.
Intuitively, summing these similarity values captures simultaneously a desired
balance between feature specificity and generality, addressing the observations in the
beginning of this section. Some features might characterize just a single word that is
very similar to w, but then the sum of similarities will include a single element, yielding
a relatively low weight. This is why the sum of similarities is used rather than an
average value, which might become too high by chance when computed over just a
single element (or very few elements). Relatively generic features, which occur with
many words and are thus less indicative, may characterize more words within N(w)
but then on average the similarity values of these words with w is likely to be lower,
contributing smaller values to the sum. To receive a high overall weight a reliable feature
has to characterize multiple words that are highly similar to w.
We note that the bootstrapped weight is a sum of word similarity values rather
than a direct function of word–feature association values, which is the more common
approche. It thus does not depend on the exact statistical co-occurrence level between
w and f . Plutôt, it depends on a more global assessment of the association between
f and the semantic vicinity of w. We notice that the bootstrapped weight is deter-
mined separately relative to each individual word. This differs from measures that are
global word-independent functions of the feature, such as the feature entropy used in
Grefenstette (1994) and the feature term strength relative to a predefined class as em-
ployed in Pekar, Krkoska, and Staab (2004) for supervised word classification.
4.2 Feature Reduction and Similarity Re-Computation
Once the bootstrapped weights have been computed, their accuracy is sufficient to allow
for aggressive feature reduction. As shown in the following section, in our experiments
it sufficed to use only the top 100 features for each word in order to obtain optimal
word similarity results, because the most informative features now receive the highest
weights.
Enfin, similarity between words is re-computed over the reduced vectors using the
sim function with weightB replacing the original feature weights. The resulting similarity
measure is further referred to as simB.
445
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
5. Evaluation by Lexical Entailment
To test the effectiveness of the bootstrapped weighting scheme, we first evaluated
whether it contributes to better prediction of lexical entailment. This evaluation was
based on gold-standard annotations determined by human judgments of the substi-
tutable lexical entailment relation, as defined in Section 3. The new similarity scheme,
simB, based on the bootstrapped weights, was first computed using the standard LIN
method as the initial similarity measure. The resulting similarity lists of simLIN (le
original LIN method) and simB
LIN (Bootstrapped LIN) schemes were evaluated for a sam-
ple of nouns (Section 5.2). Alors, the evaluation was extended (Section 5.3) to apply the
bootstrapping scheme over the two additional similarity measures that were presented
in Section 2.2, simWJ (weighted Jaccard) and simCOS (Cosine). Along with these lexical
entailment evaluations we also analyzed directly the quality of the bootstrapped fea-
ture vectors, according to the average common-feature rank ratio measure, ce qui était
defined in Section 6.
5.1 Experimental Setting
Our experiments were conducted using statistics from an 18 million token subset of the
Reuters RCV1 corpus (known as Reuters Corpus, Volume 1, English Language, 1996-
08-20 à 1997-08-19), parsed by Lin’s Minipar dependency parser (Lin 1993).
The test set of candidate word similarity pairs was constructed for a sample of
30 randomly selected nouns whose corpus frequency exceeds 500. In our primary exper-
iment we computed the top 40 most similar words for each noun by the simLIN and by
simB
LIN measures, yielding 1,200 pairs for each method, et 2,400 pairs altogether. About
800 of these pairs were common for the two methods, therefore leaving approximately
1,600 distinct candidate word similarity pairs. Because the lexical entailment relation is
directional, each candidate pair was duplicated to create two directional pairs, yielding
a test set of 3,200 pairs. Ainsi, for each pair of words, w and v, the two ordered pairs (w, v)
et (v, w) were created to be judged separately for entailment in the specified direction
(whether the first word entails the other). Par conséquent, a non-directional candidate
similarity pair w, v is considered as a correct entailment if it was assessed as an entailing
pair at least in one direction.
The assessors were only provided with a list of word pairs without any contextual
information and could consult any available dictionary, WordNet, and the Web.
The judgment criterion follows the criterion presented in Section 3. En particulier,
the judges were asked to apply the two operational conditions, existence and sub-
stitutability in context, to each given pair. Prior to performing the final test of the
annotation experiment, the judges were presented with an annotated set of entailing
and non-entailing pairs along with the existential statements and sample sentences for
substitution, demonstrating how the two conditions could be applied in different cases
of entailment. En outre, they had to judge a training set of several dozen pairs and
then discuss their judgment decisions with each other to gain a better understanding of
the two criteria.
The following example illustrates this process. Given a non-directional pair
{entreprise, organization} two directional pairs are created: (entreprise, organization) et
(organization, entreprise). The former pair is judged as a correct entailment: the existence
of a company entails the existence of an organization, and the meaning of the sentence:
John works for a large company entails the meaning of the sentence with substitution:
John works for a large organization. Ainsi, company lexically entails organization, but not
446
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
vice versa (as shown in Section 3.3), therefore the second pair is judged as not entailing.
Eventually, the non-directional pair {entreprise, organization} is considered as a correct
entailment.
Enfin, the test set of 3,200 pairs was split into three disjoint subsets that were
judged by three native English speaking assessors, each of whom possessed a Bach-
elors degree in English Linguistics. For each subset a different pair of assessors was
assigned, each person judging the entire subset. The judges were grouped into three
different pairs (c'est à dire., JudgeI+JudgeII, JudgeII+JudgeIII, and JudgeI+JudgeIII). Each pair
was assigned initially to judge all the word similarities in each subset, and the third
assessor was employed in cases of disagreement between the first two. The majority
vote was taken as the final decision. Ainsi, each assessor had to fully annotate two
thirds of the data and for a third subset she only had to judge the pairs for which there
was disagreement between the other two judges. This was done in order to measure the
agreement achieved for different pairs of annotators.
The output pairs from both methods were mixed so the assessors could not associate
a pair with the method that proposed it. We note that this evaluation methodology,
in which human assessors judge the correctness of candidate pairs by some semantic
substitutability criterion, is similar to common evaluation methodologies used for para-
phrase acquisition (Barzilay and McKeown 2001; Lin and Pantel 2001; Szpektor et al.
2004).
Measuring human agreement level for this task, the proportions of matching de-
cisions were 93.5% between Judge I and Judge II, 90% for Judge I and Judge III, et
91.2% for Judge II and Judge III. The corresponding kappa values are 0.83, 0.80, et 0.80,
which is regarded as “very good agreement” (Landis and Koch 1997). It is interesting
to note that after some discussion most of the disagreements were settled, and the few
remaining mismatches were due to different understandings of word meanings. These
findings seem to have a similar flavor to the human agreement findings reported for the
Recognizing Textual Entailment challenges (Bar-Haim et al. 2006; Dagan, Glickman, et
Magnini 2006), in which entailment was judged for pairs of sentences. En fait, the kappa
values obtained in our evaluation are substantially higher than reported for sentence-
level textual entailment, which suggests that it is easier to make entailment judgments
at the lexical level than at the full sentence level.
The parameter values of the algorithms were tuned using a development set of
similarity pairs generated for 10 additional nouns, distinct from the 30 nouns used for
the test set. The parameters were optimized by running the algorithm systematically
with various values across the parameter scales and judging a sample subset of the
résultats. weightMI = 4 was found as the optimal MI threshold for active feature weights
(features included in the feature vectors), yielding a 10% precision increase of simLIN and
removing over 50% of the data relative to no feature filtering. Accordingly, this value
also serves as the θweight threshold in the bootstrapping scheme (Section 4). As for the
θsim parameter, the best results on the development set were obtained for θsim = 0.04,
θsim = 0.02, and θsim = 0.01 when bootstrapping over the initial similarity measures
LIN, WJ, and COS, respectivement.
5.2 Evaluation Results for simB
LIN
We measured the contribution of the improved feature vectors to the resulting preci-
sion of simLIN and simB
LIN in predicting lexical entailment. The results are presented in
Tableau 4, where precision and error reduction values were computed for the top 20,
30, et 40 word similarity pairs produced by each method. It can be seen that the
447
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
Tableau 4
Lexical entailment precision values for top-n similar words by the Bootstrapped LIN and the
original LIN method.
Top-n
Words
Correct
Error Rate
Entailments (%) Reduction (%)
simLIN
simB
LIN
Top 20
Top 30
Top 40
52.0
48.2
41.0
57.9
56.2
49.7
12.3
15.4
14.7
Bootstrapped LIN method outperformed the original LIN approach by 6–9 precision
points at all top-n levels. As expected, the precision for the shorter top 20 list is higher
for both methods, thus leaving a bit less room for improvement.
Dans l'ensemble, the Bootstrapped LIN method extracted 104 (21%) more correct similarity
pairs than the other measure and reduced the number of errors by almost 15%. We also
computed the relative recall, which shows the percentage of correct word similarities
found by each method relative to the joint set of similarities that were extracted by
both methods. The overall relative recall of the Bootstrapped LIN was quite high (94%),
exceeding LIN’s relative recall (de 78%) par 16 percentage points. We found that the
bootstrapped method covers over 90% of the correct similarities learned by the original
method, while also identifying many additional correct pairs.
It should be noted at this point that the current limited precision levels are deter-
mined not just by the quality of the feature vectors but significantly by the nature of the
vector comparison measure itself (c'est à dire., the LIN formula, as well weighted Jaccard and
Cosine as reported in Section 5.3). It was observed in other work (Geffet and Dagan
2005) that these common types of vector comparison schemes exhibit certain flaws
in predicting lexical entailment. Our present work thus shows that the bootstrapping
method yields a significant improvement in feature vector quality, but future research
is needed to investigate improved vector comparison schemes.
An additional indication of the improved vector quality is the massive feature
reduction allowed by having the most characteristic features concentrated at the top
ranks of the vectors. The vectors of active features of LIN, as constructed after standard
feature filtering (Section 5.1), could be further reduced by the bootstrapped weighting
to about one third of their size. As illustrated in Figure 1, changing the vector size
significantly affects the similarity results. In simB
LIN the best result was obtained with
the top 100 features per word, while using less than 100 or more than 150 features
caused a 5–10% decrease in performance. On the other hand, an attempt to cut off the
lower ranked features of the MI weighting always resulted in a noticeable decrease in
precision. These results show that for MI weighting many important features appear
further down in the ranked vectors, while for the bootstrapped weighting adding too
many features adds mostly noise, since most characteristic features are concentrated
at the top ranks. Ainsi, in addition to better feature weighting, the bootstrapping step
provides effective feature reduction, which improves vector quality and consequently
the similarity results.
We note that the optimal vector size we obtained conforms to previous results—
Par exemple, by Widdows (2003), Curran (2004), and Curran and Moens (2002)—who
also used reduced vectors of up to 100 features as optimal for learning hyponymy and
synonymy, respectivement. In Widdows the known SVD method for dimension reduction
448
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
Chiffre 1
Percentage of correct entailments within the top 40 candidate pairs of each of the methods,
LIN and Bootstrapped LIN (denoted as LINB in the figure), when using varying numbers of
top-ranked features in the feature vector. The value of “All” corresponds to the full size of
vectors and is typically in the range of 300–400 features.
of LSA-based vectors is applied, whereas in Curran, and Curran and Moens, only
the strongly associated verbs (direct and indirect objects of the noun) are selected as
“canonical features” that are expected to be shared by true synonyms.
Enfin, we tried executing an additional bootstrapping iteration of weightB calcula-
tion over the similarity results of simB
LIN. The resulting increase in precision was much
smaller, of about 2%, showing that most of the potential benefit is exploited in the
first bootstrapping iteration (which is not uncommon for natural language data). Sur
the other hand, computing the bootstrapping weight twice increases computation time
significantly, which led us to suggest a single bootstrapping iteration as a reasonable
cost-effectiveness tradeoff for our data.
5.3 Evaluation for simB
WJ and simB
COS
To further validate the behavior of the bootstrapping scheme we experimented with two
additional similarity measures, weighted Jaccard (simWJ) and Cosine (simCOS) (described
in Section 2.2). For each of the additional measures the experiment repeats the main
three steps described in Section 4: Initially, the basic similarity lists are calculated for
each of the measures using MI weighting; alors, the bootstrapped weighting, weightB, est
computed based on the initial similarities, yielding new word feature vectors; finally,
the similarity values are recomputed by the same vector similarity measure using the
new feature vectors.
To assess the effectiveness of weightB we computed the four alternative output
similarity lists, using the simWJ and simCOS similarity measures, each with the weightMI
449
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
Tableau 5
Comparative precision values for the top 20 similarity lists of the three selected similarity
measures, with MI and Bootstrapped feature weighting for each.
Measure
LIN–LINB WJ–WJB
COS–COSB
Correct Similarities (%)
52.0–57.9
51.0–54.8
46.1–50.9
and weightB weighting functions. The four lists were judged for lexical entailment by
three assessors, according to the same procedure described in Section 5.1. To make the
additional manual evaluation affordable we judged the top 20 similar words in each list
for each of the 30 target nouns of Section 5.1.
Tableau 5 summarizes the precision values achieved by LIN, WJ, and COS with
both weightMI and weightB. As shown in the table, bootstrapped weighting consistently
contributed between 4–6 points to the accuracy of each method in the top 20 similarité
list. We view the results as quite positive, considering that improving over top 20
similarities is a much more challenging task than improving over longer similarity lists,
while the improvement was achieved only by modifying the feature vectors without
changing the similarity measure itself (as hinted in Section 5.2). Our results are also
compatible with previous findings in the literature (Dagan, Lee, and Pereira 1999;
Weeds, Weir, and McCarthy 2004) that found LIN and WJ to be more accurate for
similarity acquisition than COS. Dans l'ensemble, the results demonstrate that the bootstrapped
weighting scheme consistently produces improved results.
An interesting behavior of the bootstrapping process is that the most prominent
features for a given target word converge across the different initial similarity measures,
as exemplified in Table 6. En particulier, although the initial similarity lists overlap only
partly,4 the overlap of the top 30 features for our 30-word sample was ranging between
88% et 100%. This provides additional evidence that the quality of the bootstrapped
weighting is quite similar for various initial similarity measures.
6. Analyzing the Bootstrapped Feature Vector Quality
In this section we provide an in-depth analysis of the bootstrapping feature weighting
quality compared to the state-of-the-art MI weighting function.
6.1 Qualitative Observations
The problematic feature ranking noticed at the beginning of Section 4 can be revealed
more objectively by examining the common features which contribute most to the word
similarity scores. To that end, we examine the common features of the two given words
and sort them by the sum of their weights in both word vectors. Tableau 7 shows the top
10 common features by this sorting for a pair of truly similar (lexically entailing) words
(country–state), and for a pair of non-entailing words (country–party). For each common
feature the table shows its two corresponding ranks in the feature vectors of the two
words.
4 Overlap rate was about 40% between COS and WJ or LIN, et 70% between WJ and LIN. The overlap
was computed following the procedure of Weeds, Weir, and McCarthy (2004), disregarding the order of
the similar words in the lists. Fait intéressant, they obtained roughly similar figures, de 28% overlap for COS
and WJ, 32% overlap for COS and LIN, et 81% overlap between LIN and WJ.
450
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
Tableau 6
Top 30 features of town by bootstrapped weighting based on LIN, WJ, and COS as initial
similarities. The three sets of words are almost identical, with relatively minor ranking
differences.
LINB
WJB
COSB
southern
northern
office
eastern
remote
official
troop
northeastern
people
coastal
attack
based
populated
northwestern
base
home
south
west
western
neighboring
resident
plant
police
held
locate
trip
city
site
camp
surrounding
southern
northern
office
official
coastal
eastern
northeastern
remote
troop
people
based
populated
attack
home
northwestern
south
western
west
resident
neighboring
maison
city
base
trip
camp
held
north
locate
surrounding
police
northern
southern
remote
eastern
official
based
northeastern
office
coastal
northwestern
people
attack
troop
home
south
western
city
populated
base
resident
north
west
neighboring
trip
surrounding
police
held
locate
maison
camp
It can be observed in Table 7 that for both word pairs the common features are
scattered across the pair of feature vectors, making it difficult to distinguish between
the truly similar and the non-similar pairs. We suggest, on the other hand, that the
desired behavior of effective feature weighting is that the common features of truly
similar words would be concentrated at the top ranks of both word vectors. In other
words, if the two words are semantically similar then we expect them to share their
most characteristic features, which are in turn expected to appear at the higher ranks
of each feature vector. The common features for non-similar words are expected to be
scattered all across each of the vectors. En fait, these expectations correspond exactly to
the rationale behind distributional similarity measures: Such measures are designed to
assign higher similarity scores for vector pairs that share highly weighted features.
Comparatively, we illustrate the behavior of the Bootstrapped LIN method relative to
the observations regarding the original LIN method, using the same running example.
Tableau 8 shows the top 10 features of country. We observe that the list now contains
features that are intuitively quite indicative and reliable, while many too specific or
idiomatic features, and too general ones, were demoted (compare with Table 3). Tableau 9
shows that most of the top 10 common features for country–state are now ranked highly
451
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
for both words. On the other hand, there are only two common features (among the
top 100 features) for the incorrect pair country–party, both with quite low ranks (compare
with Table 7), while the rest of the common features for this pair did not pass the top 100
cutoff.
Par conséquent, Tableau 10 demonstrates a much more accurate similarity list for coun-
try, where many incorrect (non-entailing) word similarities, like party and company, étaient
demoted. Plutôt, additional correct similarities, like kingdom and land, were promoted
(compare with Table 2). In this particular case all the remaining errors correspond to
words that are related quite closely to country, denoting geographic concepts. Many of
these errors are context dependent entailments which might be substitutable in some
cases, but they violate the word meaning entailment condition (par exemple., country–neighbor,
country–port). Apparently, these words tend to occur in contexts that are typical for
country in the Reuters corpus. Some errors violating the substitutability condition of
lexical entailment were identified as well, such as industry–product. These cases are
quite hard to differentiate from correct entailments, since the two words are usually
closely related to each other and also share highly ranked features, because they often
appear in similar characteristic contexts. It may therefore be difficult to filter out such
Tableau 7
LIN (MI) weighting: The top 10 common features for country–state and country–party, along with
their corresponding ranks in each of the two feature vectors. The features are sorted by the sum
of their feature weights with both words.
Country–State
Ranks
Country–Party
Ranks
Broadcast, pcomp in, h
Goods, mod, h
Civil servant, gen, h
Bloc, gen, h
Nonaligned, mod, m
Neighboring, mod, m
Statistic, pcomp on, h
Border, pcomp of, h
Northwest, mod, h
Trip, pcomp to, h
24
140
64
30
55
15
165
10
41
105
Brass, nn, h
50
64
16 Concluding, pcomp of, h
73
54 Representation, pcomp of, h
82
Patriarch, pcomp of, h
77
128
Friendly, mod, m
60
58
Expel, pcomp from, h
59
165
43 Heartland, pcomp of, h
102
Surprising, pcomp of, h
247
114
Issue, pcomp between, h
103
174
34 Contravention, pcomp in, m 129
22
20
27
28
83
30
23
38
51
43
Tableau 8
Top 10 features of country by the Bootstrapped feature weighting.
Feature
WeightB
Industry, gen, h
Airport, gen, h
Visit, pcomp to, h
Neighboring, mod, m
Loi, gen, h
Économie, gen, h
Population, gen, h
Stock market, gen, h
Governor, pcomp of, h
Parliament, gen, h
1.21
1.16
1.06
1.04
1.02
1.02
0.93
0.92
0.92
0.91
452
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
Tableau 9
Bootstrapped weighting: top 10 common features for country–state and country–party along with
their corresponding ranks in the two (sorted) feature vectors.
Country–State
Ranks
Country–Party
Neighboring, mod, m
Industry, gen, h
Impoverished, mod, m
Governor, pcomp of, h
Population, gen, h
City, gen, h
Économie, gen, h
Parliament, gen, h
Citizen, pcomp of, h
Loi, gen, h
3
1
8
10
6
17
5
10
14
4
1 Relation, pcomp with, h
11 Minister, pcomp from, h
8
9
16
18
15
22
25
33
Ranks
12
77
26
49
Tableau 10
Top 20 most similar words for country and their ranks in the similarity list by the Bootstrapped
LIN measure.
nation
state
*island
region
area
1
2
3
4
5
territoire
*neighbor
colony
*port
republic
6
7
8
9
10
11
*province
12
*city
*town
13
kingdom 14
15
*district
zone
atterrir
place
economy
*monde
16
17
18
19
20
Note that four of the incorrect similarities from Table 2 were replaced with correct entailments
resulting in a 20% increase of precision (reaching 60%).
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
non-substitutable similarities merely by the standard distributional similarity scheme,
suggesting that additional mechanisms and data types would be required.
6.2 The Average Common-Feature Rank Ratio
It should be noted at this point that these observations regarding feature weight be-
havior are based on subjective intuition of how characteristic features are for a word
meaning, which is quite difficult to assess systematically. Donc, we next propose a
quantitative measure for analyzing the quality of feature vector weights.
More formally, given a pair of feature vectors for words w and v we first define
their average common-feature rank with respect to the top-n common features, denoted
acfrn, as follows:
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
(cid:1)
acfrn(w, v) =
1
n
f ∈top−n(F(w)∩F(v))
1
2 [rank(w, F ) + rank(v, F )]
(7)
where rank(w, F ) is the rank of feature f in the vector of the word w when features are
sorted by their weight, and F(w) is the set of features in w’s vector. top-n is the set of
top n common features to consider, where common features are sorted by the sum of
their weights in the two word vectors (the same sorting as in Table 7). Autrement dit,
acfrn(w, v) is the average rank in the two feature vectors of their top n common features.
453
Computational Linguistics
Volume 35, Nombre 3
Using this measure, we expect that a good feature weighting function would
typically yield lower values of acfrn for truly similar words (as low ranking values
correspond to higher positions in the vectors) than for non-similar words. Ainsi, given
a pre-judged test set of pairs of similar and non-similar words, we define the ratio,
acfr-ratio, between the average acfrn of the set of all the non-similar words, denoted as
Non-Sim, and the average acfrn of the set of all the known pairs of similar words, Sim, à
be an objective measure for feature weighting quality, as follows:
acfrn − ratio =
1
|Non−Sim|
1
|Sim|
(cid:1)
(cid:1)
w,v∈Non−Sim acfrn(w,v)
w,v∈Sim acfrn(w,v)
(8)
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
As an illustration, the two word pairs in Table 7 yielded acfr10(country, state) = 78
and acfr10(country, party) = 64. Both values are quite high, showing no principal differ-
ence between the tighter lexically entailing similarity versus a pair of non-similar (ou
rather loosely related) words. This behavior indicates the deficiency of the MI feature
weighting function in this case. On the other hand, the corresponding values for the
two pairs produced by the Bootstrapped LIN method (for the features in Table 9) sont
acfr10(country, state) = 12 and acfr10(country, party) = 41. These figures clearly reflect the
desired distinction between similar and non-similar words, showing that the common
features of the similar words are indeed concentrated at much higher ranks in the
vectors than the common features of the non-similar words.
In recent work on distributional similarity (Curran 2004; Weeds and Weir 2005) un
variety of alternative weighting functions were compared. Cependant, the quality of these
weighting functions was evaluated only through their impact on the performance of
a particular word similarity measure, as we did in Section 5. Our acfr-ratio measure
provides the first attempt to analyze the quality of weighting functions directly, relative
to a pre-judged word similarity set, without reference to a concrete similarity measure.
6.3 An Empirical Assessment of the acfr-ratio
In this subsection we report an empirical comparison of the acfr-ratio obtained for the MI
and BootstrappedLIN weighting functions. To that end, we have run the Minipar system
on the full Reuters RCV1 corpus, which contains 2.5 GB of English news stories, et
then calculated the MI-weighted feature vectors. The optimized threshold on the feature
weights, θweight, was set to 0.2. Plus loin, to compute the Bootstrapped LIN feature weights
a θsim of 0.02 was applied to the LIN similarity values. In this experiment we employed
the full bootstrapped vectors (c'est à dire., without applying feature reduction by the top 100
cutoff). This was done to avoid the effect of the feature vector size on the acfrn metric,
which tends to naturally assign higher scores to shorter vectors.
As computing the acfr-ratio requires a pre-judged sample of candidate word simi-
larity pairs, we utilized the annotated test sample of candidate pairs of word similarities
described in Section 5, which contains both entailing and non-entailing pairs.
D'abord, we computed the average common-feature rank scores (acfrn) (with varying
values of n) for weightMI and for weightB over all the pairs in the test sample. Fait intéressant,
the mean acfrn scores for weightB range within 110–264 for n = 10. . . 100, tandis que le
corresponding range for weightMI is by an order of magnitude higher: 780–1,254, despite
the insignificant differences in vector sizes. Donc, we conclude that the common
features that are relevant to establishing distributional similarity in general (regardless
of entailment) are much more scattered across the vectors by MI weighting, while with
bootstrapping they tend to appear at higher positions in the vectors. These figures
454
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
reflect a desired behavior of the bootstrapping function which concentrates most of the
prominent common features for all the distributionally similar words (whether entailing
ou non) at the lower ranks of their vectors. En particulier, this explains the ability of our
method to perform a massive feature reduction as demonstrated in Section 5, and to
produce more informative vectors, while demoting and eliminating much of the noise
in the original vectors.
Suivant, we aim to measure the discriminative power of the compared methods to
distinguish between entailing and non-entailing pairs. To this end we calculated the
acfr-ratio, which captures the difference in the average common feature ranks between
entailing vs. non-entailing pairs, for both the MI-based and bootstrapped vectors.
The obtained results are presented in Figure 2. As can be seen the acfr-ratio values
are consistently higher for Bootstrapped LIN than for MI. C'est, the bootstrapping
method assigns much higher acfrn scores to entailing words than to non-entailing ones,
whereas for MI the corresponding acfrn scores for entailing and non-entailing pairs are
roughly equal. En particulier, we notice that the largest gaps in acfr-ratio occur for lower
numbers of top common features, whose weights are indeed the most important and
influential in distributional similarity measures. Ainsi, these findings suggest a direct
indication of an improved quality of the bootstrapped feature vectors.
7. A Pseudo-Word Sense Disambiguation Evaluation
The lexical entailment evaluation reported herein corresponds to the lexical substitution
application of distributional similarity. The other type of application, as reviewed in the
Introduction, is similarity-based prediction of word co-occurrence likelihood, needed
for disambiguation applications. Comparative evaluations of distributional similarity
methods for this type of application were commonly conducted using a pseudo-word
sense disambiguation scheme, which is replicated here. In the next subsections we first
describe how distributional similarity can help improve word sense disambiguation
(WSD). Then we describe how the pseudo-word sense disambiguation task, lequel
Chiffre 2
Comparison between the acfr-ratio for MI and Bootstrapped LIN methods, when using varying
numbers of common top-ranked features in the words’ feature vectors.
455
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
corresponds to the general WSD setting, was used to evaluate the co-occurrence like-
lihood predictions obtained by alternative similarity methods.
7.1 Similarity Modeling for Word Sense Disambiguation
WSD methods need to identify the correct sense of an ambiguous word in a given
contexte. Par exemple, a test instance for the verb save might be presented in the con-
text saving Private Ryan. The disambiguation method must decide whether save in this
particular context means rescue, preserve, keep, lay aside, or some other alternative.
Sense recognition is typically based on context features collected from a sense-
annotated training corpus. Par exemple, the system might learn from the annotated
training data that the word soldier is a typical object for the rescuing sense of save, as in:
They saved the soldier. In this setting, distributional similarity is used to reduce the data
sparseness problem via similarity-based generalization. The general idea is to predict
the likelihood of unobserved word co-occurrences based on observed co-occurrences
of distributionally similar words. Par exemple, assume that the noun private did not
occur as a direct object of save in the training data. Encore, some of the words that are
distributionally similar to private, like soldier or sergeant, might have occurred with save.
Ainsi, a WSD system may infer that the co-occurrence save private is more likely for the
rescuing sense of save because private is distributionally similar to soldier, which did co-
occur with this sense of save in the annotated training corpus. In general terms, the WSD
method estimates the co-occurrence likelihood for the target sense and a given context
word based on training data for words that are distributionally similar to the context
word.
This idea of similarity-based estimation of co-occurrence likelihood was applied
in Dagan, Marcus, and Markovitch (1995) to enhance WSD performance in machine
translation and recently in Gliozzo, Giuliano, and Strapparava (2005), who employed a
Latent Semantic Analysis (LSA)-based kernel function as a similarity-based represen-
tation for WSD. Other works employed the same idea for pseudo-word sense dis-
ambiguation, as explained in the next subsection.
7.2 The Pseudo-Word Sense Disambiguation Setting
Sense disambiguation typically requires annotated training data, created with consid-
erable human effort. Yarowsky (1992) suggested that when using WSD as a test bed
for comparative algorithmic evaluation it is possible to set up a pseudo-word sense
disambiguation scheme. This scheme was later adopted in several experiments, et
was popular for comparative evaluations of similarity-based co-occurrence likelihood
estimation (Dagan, Lee, and Pereira 1999; Lee 1999; Weeds and Weir 2005). We followed
closely the same experimental scheme, as described subsequently.
D'abord, a list of pseudo-words is constructed by “merging” pairs of words into a single
pseudo word. In our experiment each pseudo-word constitutes a pair of randomly
chosen verbs, (v, v(cid:1)), where each verb represents an alternative “sense” of the pseudo-
word. The two verbs are chosen to have almost identical probability of occurrence,
which avoids a word frequency bias on the co-occurrence likelihood predictions.
Suivant, we consider occurrences of pairs of the form (cid:2)n, (v, v’)(cid:3) , où (v, v(cid:1)) is a
pseudo-word and n is a noun representing the object of the pseudo-word. Such pairs
are constructed from all co-occurrences of either v or v(cid:1) with the object n in the corpus.
Par exemple, given the pseudo-word (rescue, keep) and the verb–object co-occurrence in
the corpus rescue–private we construct the pair (cid:2)private, (rescue, keep)(cid:3). Given such a test
456
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
pair, the disambiguation task is to decide which of the two verbs is more likely to co-
occur with the given object noun, aiming to recover the original verb from which this
pair was constructed. In this example we would like to predict that rescue is more likely
to co-occur with private as an object than keep.
In our experiment 80% of the constructed pairs were used for training, providing
the co-occurrence statistics for the original known verb in each pair (c'est à dire., either (cid:2)n, v(cid:3)
ou (cid:2)n, v’(cid:3)). From the remaining 20% of the pairs those occurring in the training corpus
were discarded, leaving as a test set only pairs which do not appear in the training part.
Ainsi, predicting the co-occurrence likelihood of the noun with each of the two verbs
cannot rely on direct frequency estimation for the co-occurrences, but rather only on
similarity-based information.
To make the similarity-based predictions we first compute the distributional sim-
ilarity scores for all pairs of nouns based on the training set statistics, where the co-
occurring verbs serve as the features in the distributional vectors of the nouns. Alors,
given a test pair (cid:2)(v,v’), n(cid:3) our task is to predict which of the two verbs is more likely
to co-occur with n. This verb is thus predicted as being the original verb from which
the pair was constructed. To this end, the noun n is substituted in turn with each of its
k distributionally most similar nouns, ni, and then both of the obtained “similar” pairs
(cid:2)ni, v(cid:3) et (cid:2)ni, v(cid:1)(cid:3) are sought in the training set.
Suivant, we would like to predict that the more likely co-occurrence between (cid:2)n, v(cid:3) et
(cid:2)n, v’(cid:3) is the one for which more pairs of similar words were found in the training set.
Several approaches were used in the literature to quantify this decision procedure and
we have followed the most recent one from Weeds and Weir (2005). Each similar noun
ni is given a vote, which is equal to the difference between the frequencies of the two
co-occurrences (ni, v) et (ni, v(cid:1)), and which it casts to the verb with which it co-occurs
more frequently. The votes for each of the two verbs are summed over all k similar
nouns ni and the one with most votes wins. The winning verb is considered correct if it
is indeed the original verb from which the pair was constructed, and a tie is recorded
if the votes for both verbs are equal. Enfin, the overall performance of the prediction
method is calculated by its error rate:
error = 1
T
(#of incorrect choices +
#of ties
2
)
(9)
where T is the number of test instances.
In the experiment, we used the 1,000 most frequent nouns in our subset of the
Reuters corpus (of Section 5.1). The training and test data were created as described
herein, using the Minipar parser (Lin 1993) to produce verb–object co-occurrence pairs.
The k = 40 most similar nouns for each test noun were computed by each of the three
examined similarity measures LIN, WJ, and COS (as in Section 5), with and without
bootstrapping. The six similarity lists were utilized in turn for the pseudo-word sense
disambiguation task, calculating the corresponding error rate.
7.3 Results
Tableau 11 shows the error rate improvements after applying the bootstrapped weighting
for each of the three similarity measures. The largest error reduction, by over 15%, était
obtained for the LIN method, with quite similar results for WJ. This result is better than
the one reported by Weeds and Weir (2005), who achieved about 6% error reduction
compared to LIN.
457
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
Tableau 11
The comparative error rates of the pseudo-disambiguation task for the three examined similarity
measures, with and without applying the bootstrapped weighting for each of them.
Measure
LIN–LINB
WJ–WJB
COS–COSB
Error rate
0.157–0.133
0.150–0.132
0.155–0.145
This experiment shows that learning tighter semantic similarities, based on the im-
proved bootstrapped feature vectors, correlates also with better similarity-based infer-
ence for co-occurrence likelihood prediction. En outre, we have seen once again that
the bootstrapping scheme does not depend on a specific similarity measure, reducing
the error rates for all three measures.
8. Conclusions
The primary contribution of this article is the proposal of a bootstrapping method
that substantially improves the quality of distributional feature vectors, as needed for
statistical word similarity. The main idea is that features which are common for similar
words are also most characteristic for their meanings and thus should be promoted. Dans
fact, beyond its intuitive appeal, this idea corresponds to the underlying rationale of
the distributional similarity scheme: Semantically similar words are expected to share
exactly those context features which are most characteristic for their meaning.
The superior empirical performance of the resulting vectors was assessed in the
context of the two primary applications of distributional word similarity. The first is
lexical substitution, which was represented in our work by a human gold standard
for the substitutable lexical entailment relation. The second is co-occurrence likelihood
prediction, which was assessed by the automatically computed scores of the common
pseudo-word sense disambiguation evaluation. An additional outcome of the improved
feature weighting is massive feature reduction.
Experimenting with three prominent similarity measures showed that the boot-
strapping scheme is robust and performs well when applied over different measures.
Notably, our experiments show that the underlying assumption behind the boot-
strapping scheme is valid, c'est, available similarity metrics do provide a reason-
able approximation of the semantic similarity space which can be then exploited via
bootstrapping.
The methodology of our investigation has yielded several additional contributions:
Utilizing a refined definition of substitutable lexical entailment both as an
end goal and as an analysis vehicle for distributional similarity. It was
shown that the refined definition can be judged directly by human subjects
with very good agreement. Dans l'ensemble, lexical entailment is suggested as a
useful model for lexical substitution needs in semantic-oriented
applications.
A thorough error analysis of state of the art distributional similarity
performance was conducted. The main observation was deficient quality
of the feature vectors, which reduces the eventual quality of similarity
measures.
1.
2.
458
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
3.
Inspired by the qualitative analysis, we proposed a new analytic measure
for feature vector quality, namely average common-feature rank ratio
(acfr-ratio), which is based on the common ranks of the features for pairs of
words. This measure estimates the ability of a feature weighting method to
distinguish between pairs of similar vs. non-similar words. To the best of
our knowledge this is the first proposed measure for direct analysis of the
quality of feature weighting functions, without the need to employ them
within some vector similarity measure.
The ability to identify the most characteristic features of words can have additional ben-
efits, beyond their impact on traditional word similarity measures (as evaluated in this
article). A demonstration of such potential appears in Geffet and Dagan (2005), lequel
presents a novel feature inclusion scheme for vector comparison. That scheme utilizes
our bootstrapping method to identify the most characteristic features of a word and
then tests whether these particular features co-occur also with a hypothesized entailed
word. The empirical success reported in that paper provides additional evidence for the
utility of the bootstrapping method.
Plus généralement, our motivation and methodology can be extended in several di-
rections by future work on acquiring lexical entailment or other lexical-semantic rela-
tion. One direction is to explore better vector comparison methods that will utilize
the improved feature weighting, as shown in Geffet and Dagan (2005). Another direc-
tion is to integrate distributional similarity and pattern-based acquisition approaches,
which were shown to provide largely complementary information (Mirkin, Dagan, et
Geffet 2006). An additional potential is to integrate automatically acquired relationships
with the information found in WordNet, which seems to suffer from several serious
limitations (Curran 2005), and typically overlaps to a rather limited extent with the
output of automatic acquisition methods. As a parallel direction, future research should
explore in detail the impact of different lexical-semantic acquisition methods on text
understanding applications.
Enfin, our proposed bootstrapping scheme seems to have a general appeal for
improving feature vector quality in additional unsupervised settings. We thus hope that
this idea will be explored further in other NLP and machine learning contexts.
Les références
Adams, Rod. 2006. Textual entailment
through extended lexical overlap. Dans
Proceedings of the Second PASCAL Challenges
Workshop on Recognising Textual Entailment,
pages 68–73, Venice.
Bar-Haim, Roy, Ido Dagan, Bill Dolan,
Lisa Ferro, Danilo Giampiccolo, Bernardo
Magnini, and Idan Szpektor. 2006. Le
second PASCAL recognising textual
entailment challenge. In Proceedings of the
Second PASCAL Challenges Workshop on
Recognising Textual Entailment, pages 1–9,
Venice.
Baroni, Marco and S. Vegnaduzzo. 2004.
Identifying subjective adjectives through
web-based mutual information. Dans
Proceedings of KONVENS–04, pages 17–24,
Vienna.
Barzilay, Regina and Kathleen McKeown.
2001. Extracting paraphrases from a
parallel corpus. In Proceedings of ACL /
EACL–01, pages 50–57, Toulouse.
Caraballo, Sharon A. 1999. Automatic
construction of a hypernym-labeled noun
hierarchy from text. In Proceedings of
ACL–99, pages 120–126, College Park, MARYLAND.
Chklovski, Timothy and Patrick Pantel.
2004. VerbOcean: Mining the web for
fine-grained Semantic Verb Relations. Dans
Proceedings of EMNLP–04, pages 33–40,
Barcelona.
Church, Kenneth W. and Hanks Patrick.
1990. Word association norms, mutual
information, and lexicography.
Computational Linguistics, 16(1):22–29.
Curran, James R. 2004. From Distributional
to Semantic Similarity. Ph.D. Thesis,
459
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Computational Linguistics
Volume 35, Nombre 3
School of Informatics of the University
of Edinburgh, Écosse.
Curran, James R. 2005. Supersense tagging of
unknown nouns using semantic similarity.
In Proceedings of ACL–2005, pages 26–33,
Ann-Arbor, MI.
Curran, James R. and Marc Moens. 2002.
Improvements in automatic thesaurus
extraction. In Proceedings of the Workshop
on Unsupervised Lexical Acquisition,
pages 59–67, Philadelphia, Pennsylvanie.
Dagan, Ido. 2000. Contextual Word
Similarity. In Rob Dale, Hermann Moisl,
and Harold Somers, editors, Handbook
of Natural Language Processing. Marcel
Dekker Inc, Chapter 19, pages 459–476,
New York, New York.
Dagan, Ido, Oren Glickman, and Bernardo
Magnini. 2006. The PASCAL recognising
textual entailment challenge. Lecture Notes
in Computer Science, 3944:177–190.
Dagan, Ido, Lillian Lee, and Fernando
Pereira. 1999. Similarity-based models of
co-occurrence probabilities. Machine
Apprentissage, 34(1-3):43–69.
Dagan, Ido, Shaul Marcus, and Shaul
Markovitch. 1995. Contextual word
similarity and estimation from sparse
data. Computer, Speech and Language,
9:123–152.
Dunning, Ted E. 1993. Accurate methods for
the statistics of surprise and coincidence.
Computational Linguistics, 19(1):61–74.
Essen, U., and V. Steinbiss. 1992.
Co-occurrence smoothing for stochastic
language modeling. In ICASSP–92,
1:161–164, Piscataway, New Jersey.
Fellbaum, Christiane, editor. 1998. WordNet:
An Electronic Lexical Database. AVEC Presse,
Cambridge, MA.
Ferrandez, O., R.. M.. Terol, R.. Munoz, P..
Martinez-Barco, and M. Palomar. 2006.
An approach based on logic forms and
WordNet relationships to textual
entailment performance. In Proceedings
of the Second PASCAL Challenges
Workshop on Recognising Textual Entailment,
pages 22–26, Venice.
Gasperin, Caroline and Renata Vieira. 2004.
Using word similarity lists for resolving
indirect anaphora. In Proceedings of
ACL–04 Workshop on Reference Resolution,
pages 40–46, Barcelona.
Gauch, Susan, J.. Wang, et S. Mahesh
Rachakonda. 1999. A corpus analysis
approach for automatic query expansion
and its extension to multiple databases.
ACM Transactions on Information Systems
(TOIS), 17(3):250–269.
460
Geffet, Maayan. 2006. Refining the
Distributional Similarity Scheme for Lexical
Entailment. Ph.D. Thesis. School of
Computer Science and Engineering,
Hebrew University, Jerusalem, Israel.
Geffet, Maayan and Ido Dagan. 2004. Feature
vector quality and distributional similarity.
In Proceedings of COLING–04, Article
number: 247, Genève.
Geffet, Maayan and Ido Dagan. 2005. Le
distributional inclusion hypotheses and
lexical entailment. In Proceedings of
ACL–05, pages 107–114, Ann-Arbor, MI.
Gliozzo, Alfio, Claudio Giuliano, and Carlo
Strapparava. 2005. Domain kernels for
word sense disambiguation. In Proceedings
of ACL–05, pages 403–410, Ann-Arbor, MI.
Grefenstette, Gregory. 1994. Exploration in
Automatic Thesaurus Discovery. Kluwer
Academic Publishers, Norwell, MA.
Harris, Zelig S. 1968. Mathematical structures
of language. Wiley, New Jersey.
Hindle, D. 1990. Noun classification from
predicate-argument structures. Dans
Proceedings of ACL–90, pages 268–275,
Pittsburgh, Pennsylvanie.
Jijkoun, Valentin and Maarten de Rijke. 2005.
Recognizing textual entailment: Is word
similarity enough? In Joaquin Quinonero
Candela, Ido Dagan, Bernardo Magnini,
and Florence d’Alche-Buc, editors, Machine
Learning Challenges, Evaluating Predictive
Uncertainty, Visual Object Classification
and Recognizing Textual Entailment, D'abord
PASCAL Machine Learning Challenges
Workshop, MLCW 2005, Southampton, ROYAUME-UNI,
Lecture Notes in Computer Science 3944,
pages 449–460, Springer, New York, New York.
Karov, Oui. et S. Edelman. 1996.
Learning similarity-based word sense
disambiguation from sparse data.
In E. Ejerhed and I. Dagan, editors,
Fourth Workshop on Very Large Corpora.
Association for Computational Linguistics,
Somerset, New Jersey, pages 42–55.
Landis, J.. R.. and G. G. Koch. 1997. Le
measurements of observer agreement for
categorical data. Biometrics, 33:159–174.
Lee, Lillian. 1997. Similarity-Based Approaches
to Natural Language Processing. Ph.D. thesis,
Université Harvard, Cambridge, MA.
Lee, Lillian.1999. Measures of distributional
similarité. In Proceedings of ACL–99,
pages 25–32, College Park, MARYLAND.
Lin, Dekang. 1993. Principle-based parsing
without overgeneration. In Proceedings of
ACL–93, pages 112–120, Columbus, OH.
Lin, Dekang. 1998. Automatic retrieval and
clustering of similar words. In Proceedings
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Zhitomirsky-Geffet and Dagan
Bootstrapping Distributional Feature Vector Quality
of COLING/ACL–98, pages 768–774,
Montréal.
Lin, Dekang and Patrick Pantel. 2001.
Discovery of inference rules for question
answering. Natural Language Engineering,
7(4):343–360.
Luk, Alpha K. 1995. Statistical sense
disambiguation with relatively small
corpora using dictionary definitions.
In Proceedings of ACL–95, pages 181–188,
Cambridge, MA.
Manning, Christopher D. and Hinrich
Schutze. 1999. Foundations of Statistical
Natural Language Processing. AVEC Presse,
Cambridge, MA.
Mirkin, Shachar, Ido Dagan, and Maayan
Geffet. 2006. Integrating pattern-based
and distributional similarity methods
for lexical entailment acquisition. Dans
Proceedings of the COLING/ACL–06 Main
Conference Poster Sessions, pages 579–586,
Sydney.
Ng, H. T. 1997. Exemplar-based word
sense disambiguation: Some recent
improvements. In Proceedings of
EMNLP– 97, pages 208–213,
Providence, RI.
Ng, H. T. and H. B. Lee. 1996. Integrating
multiple knowledge sources to
disambiguate word sense: Un
exemplar-based approach. Dans
Proceedings of ACL–1996, pages 40–47,
Santa Cruz, Californie.
Nicholson, Jeremy, Nicola Stokes, et
Timothy Baldwin. 2006. Detecting
entailment using an extended
implementation of the basic elements
overlap metric. In Proceedings of the
Second PASCAL Challenges Workshop
on Recognising Textual Entailment,
pages 122–127, Venice.
Pado, Sebastian and Mirella Lapata. 2007.
Dependency-based construction of
semantic space models. Informatique
Linguistics, 33(2):161–199.
Pantel, Patrick and Deepak Ravichandran.
2004. Automatically labeling semantic
classes. In Proceedings of HLT/NAACL–04,
pages 321–328, Boston, MA.
Pantel, Patrick, D. Ravichandran, et
E. Hovy. 2004. Towards terascale
knowledge acquisition. In Proceedings of
COLING–04, Article number: 771, Genève.
Pekar, Viktor, M.. Krkoska, et S. Staab.
2004. Feature weighting for
co-occurrence-based classification
of Words. In Proceedings of COLING–04,
Article number: 799, Genève.
Pereira, Fernando, Naftali Tishby, et
Lillian Lee. 1993. Distributional
clustering of English words. Dans
Proceedings of ACL–93, pages 183–190,
Colombus, OH.
Ruge, Gerda. 1992. Experiments on
linguistically-based term associations.
Information Processing & Management,
28(3):317–332.
Salton, G. and M. J.. McGill. 1983.
Introduction to Modern Information
Retrieval. McGraw-Hill, New York, New York.
Szpektor, Idan, H. Tanev, Ido Dagan, et
B. Coppola. 2004. Scaling web-based
acquisition of entailment relations. Dans
Proceedings of EMNLP–04, pages 41–48,
Barcelona.
Vanderwende, Lucy, Arul Menezes, and Rion
Snow. 2006. Microsoft research at RTE-2:
Syntactic contributions in the entailment
task: An implementation. In Proceedings
of the Second PASCAL Challenges Workshop
on Recognising Textual Entailment,
pages 27–32, Venice.
Weeds, Julie and David Weir. 2005.
Co-occurrence retrieval: A flexible
framework for lexical distributional
similarité. Computational Linguistics,
31(4):439–476.
Weeds, Julie, D. Weir, and D. McCarthy.
2004. Characterizing measures of lexical
distributional similarity. In Proceedings
of COLING–04, pages 1015–1021,
Suisse.
Widdows, D. 2003. Unsupervised methods
for developing taxonomies by combining
syntactic and statistical information.
In Proceedings of HLT/NAACL 2003,
pages 197–204, Edmonton.
Yarowsky, D. 1992. Word-sense
disambiguation using statistical models
of Roget’s categories trained on large
corpora. In Proceedings of COLING–92,
pages 454–460, Nantes.
461
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
je
D
o
w
n
o
un
d
e
d
F
r
o
m
h
t
t
p
:
/
/
d
je
r
e
c
t
.
m
je
t
.
e
d
toi
/
c
o
je
je
/
je
un
r
t
je
c
e
–
p
d
F
/
/
/
/
3
5
3
4
3
5
1
7
9
8
6
5
1
/
c
o
je
je
.
0
8
–
0
3
2
–
r
1
–
0
6
–
9
6
p
d
.
F
b
oui
g
toi
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Télécharger le PDF