Paraphrase-Sense-Tagged Sentences
Anne Cocos and Chris Callison-Burch
Department of Computer and Information Science
University of Pennsylvania
odonnell.anne@gmail.com, ccb@cis.upenn.edu
Astratto
Many natural language processing tasks re-
quire discriminating the particular meaning of
a word in context, but building corpora for
developing sense-aware models can be a chal-
lenge. We present a large resource of example
usages for words having a particular mean-
ing, called Paraphrase-Sense-Tagged Sentences
(PSTS). Built on the premise that a word’s
paraphrases instantiate its fine-grained mean-
ing (cioè., bug has different meanings corre-
sponding to its paraphrases fly and microbe)
the resource contains up to 10,000 sentences
for each of 3 million target-paraphrase pairs
where the target word takes on the meaning
of the paraphrase. We describe an automatic
method based on bilingual pivoting used to
enumerate sentences for PSTS, and present
two models for ranking PSTS sentences based
on their quality. Finalmente, we demonstrate the
utility of PSTS by using it to build a dataset
for the task of hypernym prediction in con-
testo. Training a model on this automatically
generated dataset produces accuracy that is
competitive with a model trained on smaller
datasets crafted with some manual effort.
1 introduzione
Word meaning is context-dependent. Whereas lex-
ical semantic tasks like relation prediction have
been studied extensively in a non-contextual set-
ting, applying such models to a downstream task
like textual inference or question answering re-
quires taking the full context into account. For
esempio, it may be true that rotavirus is a type of
bug, but rotavirus is not within the realm of pos-
sible answers to the question ‘‘Which bug caused
the server outage?’’
Many tasks in natural
language processing
require discerning the meaning of polysemous
words within a particular context. It can be a chal-
lenge to develop corpora for training or evaluating
sense-aware models, because particular attention
must be paid to making sure the distribution of
instances for a given word reflects its various
introduces Paraphrase-
meanings. This paper
Sense-Tagged Sentences (PSTS),1 a large resource
of example usages of English words having a
particular meaning. Rather than assume a rigid
inventory of possible senses for each word, PSTS
is grounded in the idea that the many fine-grained
meanings of a word are instantiated by its para-
frasi. Per esempio, the word bug has different
meanings corresponding to its paraphrases fly,
error, and microbe, and PSTS includes sentences
where bug takes on each of these meanings
(Figura 1). Overall, the resource contains up to
10,000 sentences for each of roughly 3 million
English lexical and phrasal paraphrases from
the Paraphrase Database (PPDB) (Bannard and
Callison-Burch, 2005; Ganitkevitch et al., 2013;
Pavlick et al., 2015).
PSTS was compiled by automatically extracting
sentences from the English side of bilingual paral-
lel corpora using a technique inspired by bilingual
pivoting (Bannard and Callison-Burch, 2005). For
instance, to find a sentence containing bug where
it means fly, we select English sentences where
bug is translated to the French mouche, Spanish
mosca, or one of the other foreign words that
bug shares as a translation with fly. Qualitative
analysis of the sentences in PSTS indicates that
this is a noisy process, so we implement and
compare two methods for ranking sentences by
the degree to which they are ‘‘characteristic’’ of
their associated paraphrase meaning. When used
to rank PSTS sentences, a supervised regression
model trained to correlate with human judgments
of sentence quality, and an unsupervised lexical
substitution model (Melamud et al., 2016) lead to,
1http://psts.io.
714
Operazioni dell'Associazione per la Linguistica Computazionale, vol. 7, pag. 714–728, 2019. https://doi.org/10.1162/tacl a 00295
Redattore di azioni: Sebastian Pad´o. Lotto di invio: 5/2019; Lotto di revisione: 9/2019; Pubblicato 12/2019.
C(cid:2) 2019 Associazione per la Linguistica Computazionale. Distribuito sotto CC-BY 4.0 licenza.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
Manually sense-tagged corpora, such as SemCor
(Miller et al., 1994) or OntoNotes (Weischedel
et al., 2013), can then be used to train supervised
word sense disambiguation (WSD) classifiers to
predict sense labels on untagged text (Ando, 2006;
Zhong and Ng, 2010; Rothe and Sch¨utze, 2015).
Top-performing supervised WSD systems achieve
roughly 74% accuracy in assigning WordNet
sense labels to word instances (Ando, 2006;
Rothe and Sch¨utze, 2015). In shared task set-
tings, supervised classifiers typically out-perform
unsupervised WSD systems (Mihalcea et al.,
2004).
Within the set of unsupervised methods, one
long-standing idea is to use foreign translations
as proxies for sense labels of polysemous words
(Brown et al., 1991; Dagan, 1991). This is based
on the assumption that a polysemous English
word e will often have different translations into a
target language, depending on the sense of e that
si usa. To borrow an example from Gale et al.
(1992), if the English word sentence is translated
to the French peine (judicial sentence) in one
context and the French phrase (syntactic sentence)
in another, then the two instances in English can
be tagged with appropriate sense labels based
on a mapping from the French translations to
the English sense inventory. This technique has
been frequently applied to automatically generate
sense-tagged corpora, in order to overcome the
costliness of manual sense annotation (Gale et al.,
1992; Dagan and Itai, 1994; Diab and Resnik,
2002; Ng et al., 2003; Chan and Ng, 2005;
Apidianaki, 2009; Lefever et al., 2011). Our ap-
proach to unsupervised sense tagging in this
paper is related, but different. Like the translation
proxy approach, our method relies on having
bilingual parallel corpora. But in our case, IL
sense labels are grounded in English paraphrases,
rather than in foreign translations. This means
that our method does not require any manual
mapping from foreign translations to an English
sense inventory. It also enables us to generate
sense-tagged examples using bitext over multiple
pivot languages, without having to resolve sense
mapping between languages.
There is a close relationship between sense
tagging and paraphrasing. Some research efforts
assume that words have a discrete sense inventory,
and they represent each word sense as a set or
cluster of paraphrases (Mugnaio, 1995; Cocos and
Callison-Burch, 2016). Other work (Melamud
Figura 1: We assume that the fine-grained mean-
ings of the noun bug are instantiated by its para-
frasi. Example usages of bug pertaining to
each paraphrase are extracted automatically via a
method inspired by bilingual pivoting (Bannard
and Callison-Burch, 2005).
rispettivamente, 89% E 96% precision within the
top-10 sentences.
In Section 5 we demonstrate a use of PSTS by
automatically constructing a training set for the
task of hypernym prediction in context (Shwartz
and Dagan, 2016; Vyas and Carpuat, 2017). In this
task, a system is presented with a pair of words
and sentence-level contexts for each, and must
predict whether a hypernym relation holds for
that word pair in the given contexts. We auto-
matically generate training data for this task
from PSTS, creating a training set with 5 E
30 times more training instances than the two
existing datasets for this task—both of which
rely on manually generated resources. We train
a contextual hypernym prediction model on the
PSTS-derived dataset, and show that it leads to
prediction accuracy that is competitive with or
better than than the same model trained on the
smaller training sets.
2 Related Work
Generalmente, there are three basic categories of
techniques for generating sense-tagged corpora:
manual annotation, application of supervised mod-
els for word sense disambiguation, and unsuper-
vised methods. Manual annotation asks humans to
hand-label word instances with a sense tag, assum-
ing that the word’s senses are enumerated in an
underlying sense inventory (typically WordNet
[Mugnaio, 1995]) (Edmonds and Cotton, 2001;
Mihalcea et al., 2004; Petrolito and Bond, 2014).
715
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
et al., 2015UN), including in lexical substitution
(McCarthy and Navigli, 2007, 2009), represents
the contextualized meaning of a word instance by
the set of paraphrases that could be substituted
for it. This paper takes the view that assuming
a discrete underlying sense inventory can be too
rigid for many applications; humans have notori-
ously low agreement in manual sense-tagging
compiti (Cinkov´a et al., 2012), and the appropriate
sense granularity varies by setting. Invece, we
assume a ‘‘one paraphrase per fine-grained mean-
ing’’ model in this paper as a generalizable ap-
proach to word sense modeling. In PSTS, a word
type has as many meanings as it has paraphrases,
but its paraphrase-sense-tagged instances can be
grouped based on a coarser sense inventory if so
desired.
3 Constructing PSTS
For a paraphrase pair like coach↔trainer, PSTS
includes a set of sentences Scoach,trainer con-
taining coach in its trainer sense (per esempio., My coach
cancelled the workout), and a set of sentences
Scoach,trainer containing trainer in its coach sense
(per esempio., It’s just a sprain, according to her trainer).
This section describes the method for enumerating
sentences corresponding to a particular paraphrase
pair for inclusion in PSTS.
3.1 Sentence Extraction
Our method for extracting sentences for PSTS
is inspired by bilingual pivoting (Bannard and
Callison-Burch, 2005), which discovers same-
language paraphrases by ‘‘pivoting’’ over bilin-
gual parallel corpora. Specifically, if the English
phrases coach and trainer are each translated to the
same Slovenian phrase trener in some contexts,
this is taken as evidence that coach and trainer
have approximately similar meaning. We apply
this idea in reverse: to find English sentences
where coach means trainer (as opposed to bus
or railcar), we extract sentences from English-
Slovenian parallel corpora where coach has been
aligned to their shared translation trener.
The starting point for extracting PSTS is the
PPDB (Ganitkevitch et al., 2013; Pavlick et al.,
2015), a collection of over 80M lexical (one-word)
and phrasal English paraphrase pairs.2 Because
2Note that although the term paraphrase is generally
used to denote different words or phrases with approximately
716
Figura 2: Extracting sentences containing the noun
x = bug in its y = virus sense for PSTS set Sxy.
In Step 1, the set F xy of translations shared by bug
and virus is enumerated. In Step 2, the translations
f ∈ F xy are ranked by P M I(sì, F ), in order to
prioritize bug’s translations most ‘characteristic’
of its meaning in the virus sense. In Step 3,
sentences where bug has been aligned to the
French translation f = virus are extracted from
bitext corpora and added to the set Sxy.
PPDB was built using the pivot method, it follows
that each paraphrase pair x↔y in PPDB has at least
one shared foreign translation. The paraphrases for
a target word x are used as proxy labels for x’s
fine-grained senses.
The process for extracting PSTS sentences
Sx,y for x↔y consists of three steps: (1) trovare
a set F xy of shared translations for x and y,
(2) prioritizing translations that are most ‘‘char-
acteristic’’ of x’s shared meaning with y, E
(3) extracting sentences from bilingual parallel
corpora. The process is illustrated in Figure 2, E
described in further detail below.
Step 1: Finding Shared Translations.
In order
to find sentences containing the English term x
where it takes on its meaning as a paraphrase
of y, we begin by finding the sets of foreign
the same meaning, the noisy bilingual pivoting process can
produce paraphrase pairs that are more loosely semantically
related (cioè., meronyms, holonyms, or even antonyms). Here
we take a broader definition of paraphrase to mean any pair
derived from bilingual pivoting.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
(x ↔ y)
F
log p(F |sì)
log p(F )
PMI(sì, F )
Sentence segment
hot ↔ warm
hot ↔ spicy
hot ↔ popular
c´alida (es)
ciepłego (pl)
chaudes (fr)
(zh)
−1.96
−3.92
−3.30
−4.41
−1.61
−1.92
−8.19
tr`es vogue (fr)
tr`es demande (fr) −9.11
−3.61
´epic´e (fr)
(zh)
(zh)
−12.75
−14.34
−12.63
−17.75
−14.32
−12.98
−17.40
−17.47
−11.77
10.79
10.42
9.33
13.34
12.72
11.06
9.21
8.36
8.17
With the end of the hot season last year, …
I think that a hot cup of milk…would be welcome.
Avoid getting your feet too close to hot surfaces…
People with digestion issues should shun hot dishes.
Hot jambalaya!
…a manufacturer of soy sauce, hot pepper paste…
…skin aging – a hot topic in the cosmetic industry.
This area of technology is hot.
Now the town is a hot spot for weekend outings.
Tavolo 1: Example PSTS sentence segments for the adjective x = hot as a paraphrase of
y ∈ {warm, spicy, popular}. For each example, the pivot translation f is given along with its
translation probability p(F |sì), foreign word probability p(F ), and PMI(sì, F ).
translations for x and y, F x and F y respectively.
These translations are enumerated by processing
the phrase-based alignments induced between
English sentences and their translations within
a large, amalgamated set of English-to-foreign
bitext corpora. Once the translation sets F x and
F y are extracted for the individual terms, we take
their intersection as the set of shared translations,
F xy.
Step 2: Prioritizing Characteristic Transla-
zioni. Our goal is to build Sxy such that its
sentences containing x are ‘‘highly characteristic’’
of x’s shared meaning with y, and vice versa. How-
ever, not all pivot translations f ∈ F xy produce
equally characteristic sentences. Per esempio,
consider the paraphrase pair bug ↔ worm. Their
shared translation set, F bug,worm, includes the
French terms ver (worm) and esp`ece (species),
(insect). In selecting
and the Chinese term
sentences for Sbug,worm, PSTS should prioritize
English sentences where bug has been trans-
lated to the most characteristic translation for
worm—ver—over the more general
or esp`ece.
We propose using pointwise mutual informa-
zione (PMI) as a measure to quantify the degree to
which a foreign translation is ‘‘characteristic’’ of
an English term. To avoid unwanted biases that
might arise from the uneven distribution of lan-
guages present in our bitext corpora, we treat PMI
as language-specific and use shorthand notation
fl to indicate that f comes from language l. IL
PMI of English term e with foreign word fl can be
computed based on the statistics of their alignment
in bitext corpora:
PMI(e, fl) =
P(e, fl)
P(e) · p(fl)
=
P(fl|e)
P(fl)
(1)
The term in the numerator of the rightmost ex-
pression is the translation probability p(fl|e), Quale
indicates the likelihood that English word e is
aligned to foreign term fl in an English-l parallel
corpus. Maximizing this term promotes the most
frequent foreign translations for e. The term in
the denominator is the likelihood of the foreign
word, P(fl). Dividing by this term down-weights
the emphasis on frequent foreign words. This is
especially helpful for mitigating errors due to mis-
alignments of English words with foreign stop
words or punctuation. Both p(fl|e) and p(fl) are
estimated using maximum likelihood estimates
from an automatically aligned English-l parallel
corpus.
Step 3: Extracting Sentences. To extract
Sxy, we first order the shared translations for
paraphrase pair x↔y, f ∈ F xy, by decreasing
P M I(sì, F ). Then, for each translation f in order,
we extract up to 2500 sentences from the bitext
corpora where x is translated to f . This process
continues until Sxy reaches a maximum size of 10k
sentences. Tavolo 1 gives examples of sentences
extracted for various paraphrases of the adjective
hot, ordered by decreasing PMI.
PSTS is extracted from the same English-to-
foreign bitext corpora used to generate English
PPDB (Ganitkevitch et al., 2013), consisting of
Sopra 106 million sentence pairs, and spanning 22
pivot languages. Sentences are extracted for all
3 thresh-
paraphrases with a minimum PPDBSCORE
old of at least 2.0. The threshold value serves
to produce a resource corresponding to the
highest-quality paraphrases in PPDB, and elim-
inates considerable noise. In total, sentences were
3The PPDBSCORE is a supervised metric trained to correlate
with human judgments of paraphrase quality (Pavlick et al.,
2015).
717
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
POS
Paraphrase pairs Mean |Sxy| Median |Sxy|
N*
V*
R*
J*
Total
1.8M
1.1M
0.1M
0.3M
3.3M
856
972
1385
972
918
75
54
115
72
68
Tavolo 2: Number of paraphrase pairs and sen-
tences in PSTS by macro-level part of speech
(POS). The number of sentences per pair is
capped at 10k in each direction.
extracted for over 3.3M paraphrase pairs covering
nouns, verbs, adverbs, and adjectives (21 part-of-
speech tags total). Tavolo 2 gives the total number
of paraphrase pairs covered and average number
of sentences per pair in each direction. Results
are given by macro-level part-of-speech, Dove,
Per esempio, N* covers part-of-speech tags NN,
NNS, NNP, and NNPS, and phrasal constituent
tag NP.
4 PSTS Validation and Ranking
Bilingual pivoting is a noisy process (Bannard and
Callison-Burch, 2005; Chan et al., 2011; Pavlick
et al., 2015). Although shared translations for
each paraphrase pair were carefully selected using
PMI in an attempt to mitigate noise in PSTS,
the analysis of PSTS sentences that follows in
this section indicates that their quality varies.
Therefore, we follow the qualitative analysis by
proposing and evaluating two metrics for ranking
target word instances to promote those most char-
acteristic of the associated paraphrase meaning.
4.1 Qualitative Evaluation of PSTS
Our primary question is whether automatically
extracted PSTS sentences for a paraphrase pair
truly reflect the paraphrase meaning. Specifically,
for sentences like sbug where sbug ∈ Sbug,virus,
does the meaning of the word bug in sbug actually
reflect its shared meaning with virus?
We used human judgments to investigate this
question. For a pair like bug↔insect, annotators
were presented with a sentence containing bug
from Sbug,insect, and asked whether bug means
roughly the same thing as insect in the sentence.
The annotators chose from responses yes (IL
meanings are roughly similar), NO (the meanings
are different), unclear (there is not enough con-
textual information to tell), or never (these phrases
never have similar meaning). We instructed anno-
tators to ignore grammaticality in their responses,
and concentrate specifically on the semantics
of the paraphrase pair.
Human annotation was run in two rounds,
with the first round of annotation completed by
NLP researchers, and the second (much larger)
round completed by crowd workers via Amazon
Mechanical Turk (MTurk). In the first round (done
by NLP researchers), a batch of 240 sentence-
paraphrase instances (covering lexical and phrasal
noun, verb, adjective, and adverb paraphrases)
corresponding to 40 hand-selected polysemous
target words was presented to a group of 10
annotators, split into five teams of two. To encour-
age consistency, each pair of annotators worked
together to annotate each instance. For redun-
dancy, we also ensured that each instance was an-
notated separately by two pairs of researchers.
In this first round, the annotators had inter-pair
agreement of 0.41 Fleiss’ kappa (after mapping
all never and unclear answers to no), indicating
weak agreement (Fleiss, 1971).
In the second round we generated 1000 sentence-
paraphrase instances, and each instance was eval-
uated individually by seven workers on MTurk.
In each MTurk assignment, we also included an
instance from the first round that was annotated as
unanimously yes or unanimously no by the NLP
researchers in order to gauge agreement between
rounds. The crowd annotators had inter-annotator
agreement of 0.34 Fleiss’ kappa (after mapping all
never and unclear answers to no)—slightly lower
than that of the NLP researchers in round 1. IL
crowd workers had 75% absolute agreement with
the ‘‘control’’ instances inserted from the pre-
vious round.
There was weak inter-annotator agreement
in both annotation rounds. To determine why,
we manually examined 100 randomly selected
instances that received an even or nearly even split
of yes and no responses. Most of the time (71%),
annotators disagreed on the boundary between
‘‘roughly similar’’ and ‘‘different’’ meanings. For
esempio, in ‘‘An American cannot rent a car in
Canada, drive it to the USA and then return it
to Canada.’’, annotators were closely split on
whether the target word drive had roughly similar
meaning to its paraphrase guide. Another common
reason for disagreement was ambiguity of the
target word within the given context (13%), as in
718
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
the instance ‘‘I think some bug may have gotten
in the clean room.’’ (paraphrase virus). Further
disagreements occurred when the target word and
paraphrase were morphologically different forms
of the same lemma (6%) (‘‘…a matter which
is very close to our hearts…’’ with paraphrase
closely). The remaining 10% of closely split
instances are generally cases where annotators
did not consider all possible senses of the target
word and paraphrase. Per esempio, in ‘‘It does not
look good for the intelligence agency chief’’, only
four of seven crowd workers said that service was
an appropriate paraphrase for its synonym agency.
4.1.1 Human Annotation Results
To quantify the overall quality of sentences in
PSTS, we calculate the average human rating
for each annotated instance, where no (32.1%
of all annotations), never (3.9%), and unclear
(2.8%) answers are mapped to the value 0, and yes
answers are mapped to the value 1. The combined
results of this calculation from both rounds are
given in Figure 3. Overall, the average rating
È 0.61, indicating that more sentence-paraphrase
instances from PSTS are judged by humans to
have similar meaning than dissimilar meaning. In
general, adjectives produce higher-quality PSTS
sentences than the other parts of speech. For
nouns and adjectives, phrasal paraphrase pairs
are judged to have higher quality than lexical
paraphrase pairs. For verbs and adverbs, the results
are reversed.
To understand why some sentences are of poor
quality, we manually examine 100 randomly se-
lected instances with average human rating below
0.3. On close inspection, we disagreed with the
low rating for 25% of the sentences (which mirrors
the finding of 75% absolute agreement between
expert- and crowd-annotated control instances in
the second round of annotation). In those cases,
either the meaning of the target in context is a
rare sense of the target or paraphrase (per esempio., ‘‘the
appropriation is intended to cover expenses’’
with paraphrase capture), or the target word is
ambiguous in its context but could be construed
to match the paraphrase meaning (per esempio., ‘‘We’re
going to treat you as a victim in the field.’’ with
paraphrase discuss).
For the truly poor-quality sentences, in roughly
one third of cases the suggested PPDB paraphrase
for the target word is of poor quality due to
misspellings (per esempio., manage↔mange) or other
719
Figura 3: Human evaluation of the degree to
which a PSTS sentence from Sxy containing term
x reflects x’s shared meaning with its paraphrase
sì (range 0 A 1; higher scores are better).
noise in the bilingual pivoting process. One
common source of noise was mis-tagging of the
target word in context, leading to a suggested
paraphrase pertaining to the wrong part of speech.
Per esempio, in the sentence ‘‘Increase in volume
was accompanied by a change to an ovaloid
or elongate shape’’, the target elongate, Quale
appears as an adjective, was mis-tagged as a verb,
yielding the suggested but erroneous paraphrase
lie.
The remaining poor-quality sentences (roughly
50 del 100 examined) were cases where the
target word simply did not take on its shared
meaning with the suggested paraphrase. Most
of these occurred due to polysemous foreign
translations. Per esempio, PSTS wrongly suggests
the sentence ‘‘…to become a part of Zimbabwe’s
bright and positive history’’ as an example of
bright
taking on the meaning of high-gloss.
This error happens because the shared Spanish
translation, brillante, can be used with both the
literal and figurative senses of bright, but high-
gloss only matches the literal sense.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
4.2 Sentence Quality Ranking
Given the amount of variation in PSTS sentence
quality, it would be useful to have a numeric
quality estimate.
In the formation of PSTS
(Sezione 3) we used P M I(sì, F ) of the English
paraphrase y with the shared foreign translation
f to estimate how characteristic a sentence
containing English target word x is of its shared
sense with y. Ma
the Spearman correlation
between PMI and the average human ratings
for the annotated sentence-paraphrase instances
È 0.23 (P < 0.01), indicating only weak positive
correlation. Therefore, in order to enable selection
within PSTS of the most characteristic sentences
for each paraphrase pair for downstream tasks,
we propose and evaluate two models to re-rank
PSTS sentences in a way that better corresponds
to human quality judgments.
4.2.1 Supervised Regression Model
The first ranking model is a supervised regression,
trained to correlate with human quality judgments.
Concretely, given a target word x, its paraphrase
y, and a sentence sx ∈ Sx,y,
the model
predicts a score whose magnitude indicates how
characteristic sx is of x’s shared meaning with y.
This task is formulated as ordinary least squares
linear regression, where the dependent variable is
the average human quality rating for a sentence-
paraphrase instance, and the features are computed
based on the input sentence and paraphrase pair.
There are four groups, or types, of features
used in the model that are computed for each
paraphrase-sentence instance, (x↔y, sx ∈ Sx,y):
PPDB Features. Seven features from PPDB 2.0
for paraphrase pair x↔y are used as input to the
model. These include the pair’s PPDBSCORE, and
translation and paraphrase probabilities.
Contextual Features. Three contextual features
are designed to measure the distributional sim-
ilarity between the target x and paraphrase y, as
well as the substitutability of paraphrase y for
the target x in the given sentence. They include
the mean cosine similarity between word embed-
dings4 for paraphrase y and tokens within a two-
word context window of x in sentence sx; the
cosine similarity between context-masked embed-
4For computing all contextual features, we used 300-
dimensional skip-gram embeddings (Mikolov et al., 2013)
trained on the Annotated Gigaword corpus (Napoles et al.,
2012).
720
Mean contextual similarity
(cid:2)
w∈W cos(vy,vw)
f (y, sx) =
|W |
AddCos (Melamud et al., 2015b)
|W |·cos(vx,vy )+
(cid:2)
w∈W cos(vy,vw)
f (x, y, sx) =
2·|W |
Context-masked embedding similarity (Vyas and
Carpuat, 2017)
f (x, y, sx) = cos(vx,mask, vy,mask)
vx,mask = [vx (cid:5)vWmin; vx (cid:5)vWmax; vx (cid:5)vWmean ]
Table 3: Contextual features used for sentence
quality prediction, given paraphrase pair x↔y and
sentence sx ∈ Sx,y. W contains words within
a two-token context window of x in sx. vx
is the word embedding for x. vW(cid:2) are vectors
composed of the column-wise min/max/mean of
embeddings for w ∈ W . The (cid:5) symbol denotes
element-wise multiplication.
dings for x and y in sx (Vyas and Carpuat, 2017),
and the AddCos lexical substitution metric where
y is the substitute, x is the target, and the con-
text is extracted from sx (Melamud et al., 2015b)
(Table 3).
Syntactic Features. Five binary features indicate
the coarse part-of-speech label assigned to para-
phrase x ↔ y (NN, VB, RB, or JJ), and whether
x ↔ y is a lexical or phrasal paraphrase.
PMI. The final feature is simply P M I(y, f ).
The features used as input to the model training
process are the 16 listed above, as well as their
interactions as modeled by degree-2 polynomial
combinations (153 features total). During training
and validation, we apply feature selection using
recursive feature elimination in cross-validation
(RFECV) (Guyon et al., 2002).
We train the model on the 1227 sentence-
paraphrase instances that were annotated in one or
both rounds of human evaluation, after ignoring
instances marked as ‘‘unclear’’ by two or more
workers. The quality rating for each instance is
taken as the average annotator score, where no,
never, and unclear answers are mapped to the
value 0, and yes answers are mapped to the value 1.
We refer to the predicted quality scores produced
by this model as the REG(ression) score.
4.2.2 Unsupervised LexSub Model
Lexical substitution (hereafter LexSub) is the
task of identifying meaning-preserving substitutes
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
t
l
a
c
_
a
_
0
0
2
9
5
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3
target words in context
for
(McCarthy and
Navigli, 2007, 2009). For example, finding valid
substitutes for bug in There are plenty of places to
plant a bug in her office might include microphone
or listening device but not glitch. The tasks of sense
tagging and LexSub are closely related, since
valid substitutes for a polysemous word must
adhere to the correct meaning in each instance.
Indeed, early LexSub systems explicitly included
sense disambiguation as part of their pipeline
(McCarthy and Navigli, 2007), and later studies
have shown that performing sense disambiguation
can improve the results of LexSub models and
vice versa (Cocos et al., 2017; Alagi´c et al., 2018).
We adopt an off-the-shelf LexSub model called
CONTEXT2VEC (Melamud et al., 2016) as an unsu-
pervised sentence ranking model. CONTEXT2VEC
learns word and context embeddings using a bidi-
rectional long short-term memory model such that
words and their appropriate contexts have high
cosine similarity. In order to apply CONTEXT2VEC
to ranking sentence-paraphrase instances, we
calculate the cosine similarity between the para-
phrase’s CONTEXT2VEC word embedding and the
context of the target word in the sentence, using a
pre-trained model.5 The resulting score is hereafter
referred to as the C2V score.
4.2.3 Ranking Model Comparison
We compare the PSTS REG and C2V scoring models
under two evaluation settings. First, we measure
the correlation between predicted sentence scores
under each model, and the average human rating
for annotated sentences. Second, we compare the
precision of the top-10 ranked sentences under
each model based on human judgments. In the
latter experiment, we also compare with a baseline
LexSub-based sentence selection and ranking
model in order to validate bilingual pivoting as a
worthwhile sentence selection approach.
To calculate correlation between C2V model
rankings and human judgments, we simply
generate a C2V score for each of the 1227 human-
annotated sentence-paraphrase instances. For the
REG model, because the same instances were used
for training, we use 5-fold cross-validation to
estimate model correlation. In each fold, we first
run RFECV on the training portion, then train
a regression model on the selected features and
predict ratings for the test portion. The predicted
5http://u.cs.biu.ac.il/∼nlp/resources/
downloads/context2vec/.
LexSub
(baseline)
PSTS+REG
PSTS+C2V
ρ
P@1
P@5
P@10
–
0.91
0.93
0.92
0.40
0.85
0.88
0.89
0.34
0.98
0.97
0.96
Table 4: Correlation (ρ) of REG and C2V scores
with human ratings for 1227 PSTS sentence-
paraphrase instances, and precision of top-1/5/10
ranked sentences as evaluated by humans.
ratings on held-out portions from each fold are
compared to the mean annotator ratings, and
Spearman correlation is calculated on the com-
bined set of all instances.
We calculate precision under each model by
soliciting human judgments, via the same crowd-
sourcing interface used to gather sentence anno-
tations in Section 4.1. Specifically, for each of
40 hand-picked polysemous target words t (10
each nouns, verbs, adjectives, and adverbs), we
select two paraphrases p and ask workers to judge
whether t takes on the meaning of p in the top-10
PSTS sentences from St,p as ranked by REG or
C2V.
We also use top-10 precision to see how
our bilingual pivoting approach for enumerating
meaning-specific sentences compares to a system
that enumerates sentences using a LexSub model
alone, without bilingual pivoting. The baseline
LexSub model selects sentences containing coach
in its trainer sense by scoring trainer as a sub-
stitute for coach in a large set of candidate sen-
tences using CONTEXT2VEC, and ranking them. We
consider the union of all PSTS sentence sets con-
taining coach, Scoach,∗, as candidates. The top-10
scoring sentences are evaluated by humans for
precision, and compared to the ranked sets of top-
10 PSTS sentences under the REG and C2V models.
Results are given in Table 4.
The supervised REG model produces a higher
correlation (0.40) between model scores and hu-
man ratings than does the unsupervised C2V model
(0.34) or the PMI metric (0.23), indicating that
REG may be preferable to use in cases where
sentence quality estimation for a wide quality
range is needed. Although a correlation of 0.40
is not very high, it is important to note that the
correlation between each individual annotator and
721
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
t
l
a
c
_
a
_
0
0
2
9
5
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3
the mean of other annotators over all
target
sentence-paraphrase instances was only 0.36.
Thus the model predicts the mean annotator rating
with roughly the same reliability as individual
annotators.
For applications where it is necessary to choose
only the highest-quality examples of target words
with a specific paraphrase-aligned meaning, the
C2V ranking of PSTS sentences is best. We found
that 96% of top-10 ranked sentences under this
model were evaluated by humans to be good
examples of target words with the specified mean-
ing, versus 89% for the REG model and 92% for
the LexSub baseline. This indicates that the dif-
ferent methods for enumerating example sentences—
bilingual pivoting (PSTS) and LexSub score—are
complementary, and that combining the two pro-
duces the best results.
5 Hypernym Prediction in Context
Finally, we aim to demonstrate that PSTS can be
used to automatically construct a training dataset
for the task of predicting hypernymy in context,
without relying on manually annotated resources
or a pre-trained word sense disambiguation model.
Most work on hypernym prediction has been
done out of context: The input to the task is a
pair of terms like (table, furniture), and the model
predicts whether the second term is a hypernym
of the first (in this case, it is). However, both
Shwartz and Dagan (2016) and Vyas and Carpuat
(2017) point out that hypernymy between two
terms depends on their context. For example, the
table mentioned in ‘‘He set the glass down on the
table’’ is indeed a type of furniture, but in ‘‘Results
are reported in table 3.1’’ it is not. This is the
motivation for studying the task of predicting
hypernymy within a given context, where the
input to the problem is a pair of sentences each
containing a target word, and the task is to predict
whether a hypernym relationship holds between
the two targets. Example task instances are in
Table 5.
Previous work on this task has relied on either
human annotation, or the existence of a man-
ually constructed lexical semantic resource (i.e.,
WordNet), to generate training data. In the case of
Shwartz and Dagan (2016), who examined fine-
grained semantic relations in context, a dataset
of 3,750 sentence pairs was compiled by auto-
matically extracting sentences from Wikipedia
containing target words of interest, and asking
crowd workers to manually label sentence pairs
with the appropriate fine-grained semantic re-
lation.6 Subsequently, Vyas and Carpuat (2017)
studied hypernym prediction in context. They
generated a larger dataset of 22k sentence pairs
which used example sentences from WordNet as
contexts, and WordNet’s ontological structure to
find sentence pairs where the presence or absence
of a hypernym relationship could be inferred. This
section builds on both previous works, in that
we generate an even larger dataset of over 84k
sentence pairs for studying hypernymy in context,
and use the existing test sets for evaluation. How-
ever, unlike the previous methods, our dataset
is constructed without any manual annotation or
reliance on WordNet for contextual examples.
Instead, we leverage the sense-specific contexts in
PSTS to generate training instances automatically.
5.1 Producing a Training Set
Because PSTS can be used to query sentences
containing target words with a particular fine-
grained sense, our hypothesis is that, given a set
of term pairs having known potential semantic
relations, we can use PSTS to automatically pro-
duce a large training set of sentence pairs for
contextual hypernym prediction. More specifi-
cally, our goal is to generate training instances of
the form:
(t, w, ct, cw, l)
where t is a target term, w is a possibly related
term, ct and cw are contexts, or sentences, con-
taining t and w respectively, and l is a binary
label indicating whether t and w are a hyponym-
hypernym pair in the senses as they are expressed
in contexts ct and cw. The proposed method for
generating such instances from PSTS relies on
WordNet (or another lexical semantic resource)
only insofar as we use it to enumerate term pairs
(t, w) with known semantic relation; the contexts
(ct, cw) in which these relations hold or do not are
generated automatically from PSTS.
6In this study, which included the relations equivalence,
forward and reverse entailment, negation/alternation, other-
related, and independence, hyponym–hypernym pairs were
labeled as forward entailment and hypernym–hyponym pairs
labeled as reverse entailment instances.
722
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
t
l
a
c
_
a
_
0
0
2
9
5
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3
Ex.
Target
Word (t)
Related
Word (w)
(a)
tuxedo
dress
(b)
defendant
plaintiff
(c)
bug
microphone
Contexts
Hypernym (l)
ct: People believe my folderol because I wear a black tuxedo.
cw: The back is crudely constructed and is probably an addition for fancy dress.
ct: The plaintiff had sued the defendant for defamation.
cw: The court found that the plaintiff had made sufficiently full disclosure.
ct: An address error usually indicates a software bug.
cw: You have to bring the microphone to my apartment.
Yes
No
No
Table 5: Example instances for contextual hypernym prediction, selected from the PSTS-derived dataset.
The training set is deliberately constructed to
include instances of the following types:
(a) Positive instances, where (t, w) hold a hyper-
nym relationship in contexts ct and cw (l = 1)
(Table 5, example a).
(b) Negative instances, where (t, w) hold some
semantic relation other than hypernymy (such
as meronymy or antonymy) in contexts ct and
cw (l = 0). This will encourage the model to
discriminate true hypernym pairs from other
semantically related pairs (Table 5, example
b shows an antonym pair in context).
(c) Negative instances, where (t, w) hold a known
semantic relation, including possibly hyper-
nymy, in some sense, but the contexts ct and
cw are not indicative of this relation (l = 0).
This will encourage the model to take con-
text into account when making a prediction
(Table 5, example c).
Beginning with a target word t, the procedure
for generating training instances of each type from
PSTS is as follows:
Find related terms The first step is to find
related terms w such that the pair (t, w) are related
in WordNet with relation type r (which could be
one of synonym, antonym, hypernym, hyponym,
meronym, or holonym), and t ↔ w is a paraphrase
pair present in PSTS. The related terms are not
constrained to be hypernyms, in order to enable
generation of instances of type (b) above.
Generate contextually related instances
(types
(a) and (b) above). Given term pair (t, w) with
known relation r, generate sentence pairs where
this relation is assumed to hold as follows. First,
order PSTS sentences in Stw (containing target t)
and Stw (containing related term w in its sense as a
paraphrase of t) by decreasing quality score. Next,
choose the top-k sentences from each ordered list,
and select sentence pairs (ct, cw) ∈ Stw × Stw
where both sentences are in their respective top-k
lists. Add each sentence pair to the dataset as a
positive instance (l = 1) if r = hypernym, or as a
negative instance (l = 0) if r is something other
than the hypernym relation.
Generate contextually unrelated instances
(type
(c) above). Given term pair (t, w) with known
relation r, generate sentence pairs where this
relation is assumed not to hold as follows. First,
pick a confounding term t(cid:8) that is a paraphrase
of t (i.e., t ↔ t(cid:8) is in PPDB), but unrelated to w
in PPDB. This confounding term is designed to
represent an alternative sense of t. For example,
a confounding term corresponding to the term
pair (t, w) =(bug, microphone) could be glitch
because it represents a sense of bug that is different
from bug’s shared meaning with microphone.
the top-k/2 sentences containing
Next, select
related term w in its sense as w(cid:8) from Sw,w(cid:8)
in terms of quality score. Choose sentence pairs
(ct, cw) ∈ St,w ×Sw,w(cid:8)
to form negative instances.
To form the PSTS-derived contextual hyper-
nym prediction dataset, this process is carried
out for a set of 3,558 target nouns drawn from the
Shwartz and Dagan (2016) and Vyas and Carpuat
(2017) datasets. For each target noun, all PPDB
paraphrases that are hypernyms, hyponyms, syn-
onyms, antonyms, co-hyponyms, or meronyms
from WordNet were selected as related terms.
There were k = 3 sentences selected for each
target/related term pair, where the PSTS sen-
tences were ranked by the C2V model. This process
resulted in a dataset of over 84k instances, of which
32% are positive contextual hypernym pairs (type
(a)). The 68% of negative pairs are made up of
38% instances where t and w hold some relation
723
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
t
l
a
c
_
a
_
0
0
2
9
5
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3
other than hypernymy in context (type (b)), and
30% instances where t and w are unrelated in the
given context (type (c)).
5.2 Baseline IMS Training Set
In order to compare the quality of the PSTS-
derived contextual hypernym dataset
to one
produced using sentences sense-tagged by a super-
vised WSD model, we generate a baseline training
set using word instances with senses tagged by the
English all-words WSD model It Makes Sense
(IMS) (Zhong and Ng, 2010). IMS is a supervised
sense tagger that uses a SVM classifier operating
over syntactic and contextual features.
We begin by extracting an inventory of sen-
tences pertaining to WordNet senses using IMS.
Specifically, a pre-trained, off-the-shelf version of
IMS7 is used to predict WordNet 3.0 sense labels
for instances of the same target nouns present in
the PSTS-derived training set. The instances are
drawn from the English side of the same English-
foreign bitext used to extract PSTS, so the source
corpora for the PSTS-derived and IMS contextual
hypernym datasets are the same. We select the top
sentences for each sense of each target noun, as
ranked by IMS model confidence, as a sentence
inventory for each sense.
Next,we extract training instances (t, w, ct, cw, l)
using the same procedure outlined in Section 5.1.
Term pairs (t, w) are selected such that t and w
have related senses in WordNet, and both t and
w are within the set of target nouns. Related
instances are generated from the top-3 IMS-ranked
sentences for the related senses of t and w, and
unrelated sentences are chosen by selecting an
un-related WordNet sense of t to pair with the
original sense of w, and vice versa. Finally, we
truncate the resulting set of training instances
to match the PSTS-derived dataset in size and
instance type distribution: 84k instances total, with
32% positive (contextual hypernym) pairs, 38%
contextually related non-hypernym pairs, and 30%
contextually unrelated pairs.
5.3 Contextual Hypernym Prediction Model
Having automatically generated a dataset from
PSTS for studying hypernymy in context, the next
steps are to adopt a contextual hypernym pre-
diction model to train on the dataset, and then
7https://www.comp.nus.edu.sg/∼nlp.
724
Figure 4: The contextual hypernym prediction
model is based on BERT (Devlin et al., 2019).
Input sentences ct and cw are tokenized, prepended
with a [CLS] token, and separated by a [SEP]
token. The target word t in the first sentence, ct,
and the related word w in the second sentence,
cw, are surrounded by < and > gettoni. The class
label (hypernym or not) is predicted by feeding the
output representation of the [CLS] token through
fully-connected and softmax layers.
to evaluate its performance on existing hypernym
prediction test sets.
The model adopted for predicting hypernymy
in context is a fine-tuned version of the BERT
pre-trained transformer model (Devlin et al.,
2019) (Figura 4). Specifically, we use BERT in
its configuration for sentence pair classification
compiti, where the input consists of two tokenized
sentences (ct and cw), preceded by a [CLS]
token and separated by a [SEP] token. In order
to highlight the target t and related term w in
each respective sentence, we surround them with
left and right bracket tokens ‘‘<’’ and ‘‘>’’. IL
model predicts whether the sentence pair contains
contextualized hypernyms or not by processing the
input through a transformer encoder, and feeding
the output representation of the [CLS] token
through fully connected and softmax layers.
5.4 Experiments
To test our hypothesis that PSTS can be used to
generate a large, high-quality dataset for training
a contextualized hypernym prediction model,
we perform experiments that compare the per-
formance of
the BERT hypernym prediction
model on existing test sets after training on our
PSTS dataset, versus training on on datasets built
using manual resources or a supervised WSD
modello.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
We use two existing test sets for contextual
hypernym prediction in our experiments. The first,
abbreviated S&D-binary, is a binarized version of
the fine-grained semantic relation dataset from
Shwartz and Dagan (2016). The original dataset
contained five relation types, but we convert all
forward entailment and flipped reverse entailment
instances to positive (hypernym) instances, E
the rest to negative instances. The resulting data-
set has 3750 instances (18% positive and 82%
negative), split
into train/dev/test portions of
2630/190/930 instances, rispettivamente. The second
dataset used in our experiments is WordNet
Hypernyms in Context (WHiC) from Vyas and
Carpuat (2017). It contains 22,781 instances (23%
positive and 77% negative), split
into train/
dev/test portions of 15716/1704/5361 instances,
rispettivamente. There are two primary differences
between the WHiC and S&D-binary datasets.
Primo, S&D-binary contains negative instances
where the word pair has a semantic relation
other than hypernymy in the given contexts (cioè.,
type (B) from Table 5) whereas WHiC does
non. Secondo, because its sentences are extracted
from Wikipedia, S&D-binary contains some
instances where the meaning of a word in context
is ambiguous; WHiC sentences selected from
WordNet are unambiguous. Our PSTS-derived
contextual hypernym prediction dataset, Quale
contains semantically related negative instances
and has some ambiguous contexts (as noted
in Section 4.1.1) is more similar in nature to
S&D-binary.
For both the S&D-binary and WHiC datasets,
we compare results of the BERT sentence pair
classification model on the test portions after fine-
tuning on the PSTS dataset, the supervised IMS
baseline dataset, the original training set, or a
combination of the PSTS dataset with the original
training set. In order to gauge how different the
datasets are from one another, we also experiment
with training on S&D-binary and testing on
WHiC, and vice versa. In each case we use the
dataset’s original dev portion for tuning the BERT
model parameters (batch size, number of epochs,
and learning rate). Results are reported in terms
of weighted average F-Score over the positive and
negative classes, and given in Table 6.
In the case of S&D-binary, we find that training
on the 85k-instance PSTS dataset leads to a modest
improvement in test set performance of 0.6% Sopra
training on the original 2.6k-instance manually
Training Set
S&D-binary
WHiC
IMS
PSTS
PSTS+WHiC
PSTS+S&D-binary
Test Set
WHiC
S&D-binary
68.6
78.7
69.8
73.4
78.5
79.2
71.7
81.4
79.7
82.5
Tavolo 6: Performance of the BERT fine-tuned
contextual hypernym prediction model on two
existing test sets, segmented by training set. Tutto
results are reported in terms of weighted average
F1.
annotated training set. Combining the PSTS and
original training sets leads to a 4.2% relative
performance improvement over training on the
original dataset alone, and outperforms the IMS
baseline built using a supervised WSD system.
Tuttavia, on the WHiC dataset, it turns out that
training on the PSTS dataset as opposed to the
original 15.7k-instance WHiC training set leads to
a relative 6.7% drop in performance. But training
the model on the PSTS training data leads to better
performance on WHiC than training on instances
produced using the output of the supervised IMS
WSD system, or from training on S&D-binary. It
is not surprising that the PSTS-derived training set
performs better on the S&D-binary test set than it
does on the WHiC test set, given the more similar
composition between PSTS and S&D-binary.
6 Conclusione
We present PSTS, a resource of up to 10k English
sentence-level contexts for each of over 3M pa-
raphrase pairs. The sentences were enumerated
using a variation of bilingual pivoting (Bannard
and Callison-Burch, 2005), which assumes that an
English word like bug takes on the meaning of its
paraphrase fly in sentences where it is translated
to a shared foreign translation like mouche (fr).
Human assessment of the resource shows that
sentences produced by this automated process
have varying quality, so we propose two methods
to rank sentences by how well they reflect the
meaning of the associated paraphrase pair. UN
supervised regression model has higher overall
correlation (0.4) with human sentence quality
725
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
judgments, whereas an unsupervised ranking
method based on lexical substitution produces
highest precision (96%) for the top-10 ranked
sentences.
We leveraged PSTS to automatically produce
a contextualized hypernym prediction training
set, without
the need for a supervised sense
tagging model or existing hand-crafted lexical
semantic resources. To evaluate this training set,
we adopted a hypernym prediction model based
on the BERT transformer (Devlin et al., 2019).
We showed that this model, when trained on the
large PSTS training set, achieves a slight gain of
0.6% accuracy relative to training on a smaller,
manually annotated training set, without the need
for manual annotations. This suggests that it is
worth exploring the use of PSTS to generate
sense-specific datasets for other contextualized
compiti.
Ringraziamenti
We are grateful for support from the Allen
Institute for Artificial Intelligence (AI2) Key
Scientific Challenges program and the Google
Ph.D. Fellowship program. This work was also
supported by DARPA under the LORELEI pro-
gram (HR0011-15-C-0115). The views and con-
clusions contained in this publication are those
of the authors and should not be interpreted as
representing official policies or endorsements of
DARPA and the U.S. Government.
We especially thank our anonymous reviewers
for their thoughtful, substantive, and constructive
comments.
Riferimenti
Domagoj Alagi´c, Jan ˇSnajder, and Sebastian
Pad´o. 2018. Leveraging lexical substitutes for
unsupervised word sense induction. Nel professionista-
ceedings of the Thirty-Second AAAI Conference
on Artificial Intelligence, pages 5004–5011,
New Orleans, LA.
Rie Kubota Ando. 2006. Applying alternating
structure optimization to word sense disam-
biguation. In Proceedings of the Tenth Con-
ference on Computational Natural Language
Apprendimento (CoNLL), pages 77–84, New York,
NY.
Marianna Apidianaki. 2009. Data-driven semantic
analysis for multilingual WSD and lexical
selection in translation. Negli Atti del
12th Conference of the European Chapter of
the Association for Computational Linguistics
(EACL), pages 77–85, Athens.
Colin Bannard and Chris Callison-Burch. 2005.
Paraphrasing with bilingual parallel corpora.
In Proceedings of the 43rd Annual Meeting
on Association for Computational Linguistics
(ACL), pages 597–604, Ann Arbor, MI.
Peter F. Brown, Stephen A. Della Pietra, Vincent
J. Della Pietra, and Robert L. Mercer. 1991.
Word-sense disambiguation using statistical
metodi. In Proceedings of the 29th Annual
Riunione dell'Associazione per il Computazionale
Linguistica (ACL), pages 264–270, Berkeley,
CA.
Tsz Ping Chan, Chris Callison-Burch, E
Benjamin Van Durme. 2011, Luglio. Reranking
bilingually extracted paraphrases using mono-
lingual distributional similarity. Negli Atti
of the GEMS 2011 Workshop on GEometri-
cal Models of Natural Language Semantics,
pages 33–42, Edinburgh.
Yee Seng Chan and Hwee Tou Ng. 2005. Scal-
ing up word sense disambiguation via parallel
texts. In Proceedings of the Twentieth National
Conference on Artificial Intelligence (AAAI),
pages 1037–1042, Pittsburgh, PAPÀ.
Silvie Cinkov´a, Martin Holub, and Vincent
Kr´ıˇz. 2012. Managing uncertainty in semantic
tagging. In Proceedings of the 13th Conference
Di
the Associa-
tion for Computational Linguistics (EACL),
pages 840–850, Avignon.
the European Chapter of
Anne Cocos, Marianna Apidianaki, and Chris
Callison-Burch. 2017. Word sense filtering im-
proves embedding-based lexical substitution.
In Proceedings of the 1st Workshop on Sense,
Concept and Entity Representations and their
Applications, pages 110–119, Valencia.
Anne Cocos and Chris Callison-Burch. 2016.
Clustering paraphrases by word sense.
In
Proceedings of the 15th Annual Conference
Di
IL
the North American Chapter of
Associazione per la Linguistica Computazionale:
726
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
Tecnologie del linguaggio umano
HLT), pages 1463–1472, San Diego, CA.
(NAACL-
Ido Dagan. 1991. Lexical disambiguation: fonti
of information and their statistical realization.
In Proceedings of the 29th Annual Meeting of
the Association for Computational Linguistics
(ACL), pages 341–342, Berkeley, CA.
Ido Dagan and Alon Itai. 1994. Word sense
disambiguation using a second language mo-
nolingual corpus. Linguistica computazionale,
20(4):563–596.
Jacob Devlin, Ming-Wei Chang, Kenton Lee,
and Kristina Toutanova. 2019. BERT: Pre-
training of deep bidirectional transformers for
language understanding. Negli Atti del
2019 Annual Conference of the North American
Capitolo dell'Associazione per il calcolo
Linguistica (NAACL), Minneapolis, MN.
Mona Diab and Philip Resnik. 2002. An unsu-
pervised method for word sense tagging us-
ing parallel corpora. In Proceedings of 40th
Annual Meeting of the Association for Com-
Linguistica putazionale (ACL), pages 255–262,
Philadelphia, PAPÀ.
Philip Edmonds
and Scott Cotton.
2001.
SENSEVAL-2: overview. Negli Atti di
SENSEVAL-2 Second International Workshop
on Evaluating Word Sense Disambiguation
Sistemi, pages 1–5, Toulouse.
Joseph L. Fleiss. 1971. Measuring nominal scale
agreement among many raters. Psicologico
Bulletin, 76(5):378.
William A. Gale, Kenneth W. Church, and David
Yarowsky. 1992. Using bilingual materials to
develop word sense disambiguation methods.
Negli Atti di
the Fourth International
Conference on Theoretical and Methodological
Issues in Machine Translation, pages 101–112,
Montr´eal.
Juri Ganitkevitch, Benjamin Van Durme, E
Chris Callison-Burch. 2013. PPDB: The Para-
phrase Database. Negli Atti del 2013
Conference of the North American Chapter
of the Association for Computational Linguis-
tic: Tecnologie del linguaggio umano (NAACL-
HLT), pages 758–764, Atlanta, GA.
727
Isabelle Guyon, Jason Weston, Stephen Barnhill,
and Vladimir Vapnik. 2002. Gene selection for
cancer classification using support vector ma-
chines. Apprendimento automatico, 46(1-3):389–422.
Els Lefever, V´eronique Hoste, and Martine
De Cock. 2011. ParaSense or how to use par-
allel corpora for word sense disambiguation.
In Proceedings of the 49th Annual Meeting of
the Association for Computational Linguistics:
Tecnologie del linguaggio umano (ACL): Corto
Papers-Volume 2, pages 317–322, Portland,
OR.
Diana McCarthy and Roberto Navigli. 2007.
SemEval-2007 Task 10: English lexical sub-
stitution task. In Proceedings of the 4th Inter-
national Workshop on Semantic Evaluations
(SemEval-2007), pages 48–53, Prague.
Diana McCarthy and Roberto Navigli. 2009. IL
English lexical substitution task. Language
Resources and Evaluation Special Issue on
Computational Semantic Analysis of Language:
SemEval-2007 and Beyond, 43(2):139–159.
Oren Melamud, Ido Dagan, and Jacob Goldberger.
2015UN. Modeling word meaning in context with
substitute vectors. Negli Atti del 2015
Conference of the North American Chapter of
the Association for Computational Linguistics:
Tecnologie del linguaggio umano (NAACL-HLT),
pages 472–482.
Oren Melamud, Jacob Goldberger, and Ido Dagan.
2016. context2vec: Learning generic con-
text embedding with bidirectional LSTM. In
Proceedings of The 20th SIGNLL Conference
on Computational Natural Language Learning
(CONLL), pages 51–61, Berlin.
lexical substitution.
Oren Melamud, Omer Levy, and Ido Dagan.
2015B. A simple word embedding model
Negli Atti di
for
the 1st Workshop on Vector Space Modeling
for Natural Language Processing, pages 1–7,
Denver, CO.
Rada Mihalcea, Timothy Chklovski, and Adam
Kilgarriff. 2004. The SENSEVAL-3 english
Negli Atti di
lexical
SENSEVAL-3: Third International Workshop
on the Evaluation of Systems for the Semantic
Analysis of Text, pages 25–28, Barcelona.
sample
task.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
Tomás Mikolov, Ilya Sutskever, Kai Chen, Greg
Corrado, and Jeffrey Dean. 2013. Distributed
representations of words and phrases and their
compositionality. In Advances in Neural Informa-
tion Processing Systems 26, Lake Tahoe, NV.
Giorgio A. Mugnaio. 1995. WordNet: A lexical
database for English. Communications of the
ACM, 38(11):39–41.
Giorgio A. Mugnaio, Martin Chodorow, Shari
Landes, Claudia Leacock, and Robert G. Thomas.
1994. Using a semantic concordance for sense
identification. In Human Language Technol-
ogy: Proceedings of a Workshop, pages 240–243,
Plainsboro, NJ.
Courtney Napoles, Matthew Gormley,
E
Benjamin Van Durme. 2012. Annotated Giga-
word. In Proceedings of the Joint Workshop
on Automatic Knowledge Base Construction
and Web-scale Knowledge Extraction (AKBC-
WEKEX), pages 95–100, Montr´eal.
Hwee Tou Ng, Bin Wang, and Yee Seng
Chan. 2003. Exploiting parallel texts for word
sense disambiguation: An empirical study. In
Proceedings of the 41st Annual Meeting of
the Association for Computational Linguistics
(ACL), pages 455–462, Sapporo.
Ellie Pavlick,Pushpendre Rastogi, Juri Ganitkevitch,
Benjamin Van Durme, and Chris Callison-
Burch. 2015. PPDB 2.0: Better paraphrase
ranking, fine-grained entailment relations, word
embeddings, and style classification. Nel professionista-
ceedings of the 53rd Annual Meeting of the As-
sociation for Computational Linguistics and the
7th International Joint Conference on Natural
Language Processing (ACL) (Volume 2: Corto
Carte), pages 425–430, Beijing.
Tommaso Petrolito and Francis Bond. 2014.
A survey of WordNet annotated corpora. In
Proceedings of the Seventh Global WordNet
Conferenza, pages 236–245, Tartu.
Sascha Rothe and Hinrich Sch¨utze. 2015. Auto-
extend: Extending word embeddings to embed-
dings for synsets and lexemes. Negli Atti
of the 53rd Annual Meeting of the Association
for Computational Linguistics and the 7th
International Joint Conference on Natural Lan-
guage Processing (ACL) – Volume 1: Lungo
Carte, pages 1793–1803, Beijing.
Vered Shwartz and Ido Dagan. 2016. Adding
context to semantic data-driven paraphrasing.
In Proceedings of the Fifth Joint Conference
on Lexical and Computational Semantics,
pages 108–113, Berlin.
Yogarshi Vyas and Marine Carpuat. 2017. Detect-
ing asymmetric semantic relations in con-
testo: A case study on hypernymy detection.
In Proceedings of the 6th Joint Conference on
Lexical and Computational Semantics (*SEM),
pages 33–43, Vancouver.
Ralph Weischedel, Martha Palmer, Mitchell
Marcus, Eduard Hovy, Sameer Pradhan, Lance
Ramshaw, Nianwen Xue, Ann Taylor, Jeff
Kaufman, Michelle Franchini, Mohammed
El-Bachouti, Robert Belvin, and Ann Houston.
2013. OntoNotes release 5.0 LDC2013T19.
Linguistic Data Consortium, Philadelphia, PAPÀ.
Zhi Zhong and Hwee Tou Ng. 2010. It Makes
Sense: A wide-coverage word sense disam-
biguation system for free text. In Procedi-
ings of the ACL 2010 System Demonstrations,
pages 78–83, Uppsala.
728
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
9
5
1
9
2
3
5
1
4
/
/
T
l
UN
C
_
UN
_
0
0
2
9
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
