Is (President, ) a Correct Sense Pair?
Linking and Creating Bilingual Sense Correspondences
Fumiyo Fukumoto, Yoshimi Suzuki and Attaporn Wangpoonsarp
Graduate Faculty of Interdisciplinary Research, University of Yamanashi, Japan
Keywords:
Bilingual Sense Correspondence, Domain Specific Senses.
Abstract:
This paper presents a method of linking and creating bilingual sense correspondences between English and
Japanese noun word dictionaries. Locally, we extracted bilingual noun words using sentence-based similarity.
Globally, for each monolingual dictionary, we identified domain-specific senses using a textual corpus with
category information. We incorporated these, i.e., we assigned a sense to each noun word of the extracted
bilingual words keeping domain (category) consistency. Evaluation on the WordNet 3.0 and EDR Japanese
dictionaries using Reuters and Mainichi Japanese newspaper corpora showed 23.1% improvement of bilingual
noun word extraction over the baseline with local data view only. Moreover, we found that the extracted
bilingual noun senses can be used as a lexical resource for the machine translation.
1 INTRODUCTION
There has long been a great deal of interest in retrieval
of bilingual lexicons with the availability of a number
of large-scale corpora. The extracted bilingual lexi-
con is widely used for cross-lingual NLP applications,
such as machine translation (MT), cross-lingual infor-
mation retrieval (CLIR), multilingual topic tracking,
text classification and summarization. Much of the
previous work on finding bilingual lexicons has fo-
cused on word correspondence, rather than meaning.
Interpretation of whether the extracted bilingual lex-
icon is correct relies on human intervention. For ex-
ample, in most cases, we can accept a pair of English
word “president” and Japanese word “ ” as a
bilingual lexicon, while the noun “president” has at
least six different senses in WordNet, and “ ”
has two senses in the EDR dictionary.
In this paper, we propose a method for bilingual
noun word sense correspondence. We focused on
manually constructed monolingual dictionaries in dif-
ferent languages because dictionaries such as Word-
Net (Miller et al., 1990; Miller, 1995), ACQUILEX
(Briscoe, 1991), COMLEX (Grishman et al., 1994),
and EDR Japanese dictionaries
1
are fine-grained and
they are successfully utilized not only for NLP but
also for NLP applications. We used them to make
up the deficit of corpus statistics obtained from com-
1
www2.nict.go.jp/ipp/EDR/ENG/indexTop.html
parable corpora. We identified bilingual noun word
senses based on local and global data views. Here,
local data view indicates relevance between English
and Japanese sentences. We extracted bilingual noun
words by applying sentence-based similarity.
Global data view refers to domain/category infor-
mation. The assumption behind this is that predom-
inant sense of a word can depend on the domain or
source of a document (Gale et al., 1992; McCarthy
et al., 2007). We identified domain-specific senses us-
ing a corpus with category information. This method
first identifies each sense of a word in the dictionary to
its corresponding category. For each category, we cre-
ated a graph, where each node refers to each sense of
a word, and edges between nodes indicate similarity
between two nodes. We applied a Markov Random
Walk (MRW) model to the graph and ranked scores
for each sense. Finally, we incorporated the local
and global information, i.e., we assigned a domain-
specific sense to each noun word of the extracted
bilingual words maintaining domain/category consis-
tency.
The rest of the paper is organized as follows. The
next section describes our approach, especially local
and global data views. Section 3 reports some exper-
iments with a discussion of evaluation. Finally, we
summarize existing related work and conclude in Sec-
tion 5.
Fukumoto F., Suzuki Y. and Wangpoonsarp A.
Is (President,åd’
˘
gç¸t
´
sé
˘
aŸ) a Correct Sense Pair? - Linking and Creating Bilingual Sense Correspondences.
DOI: 10.5220/0006489900390048
In Proceedings of the 9th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (KEOD 2017), pages 39-48
ISBN: 978-989-758-272-1
Copyright
c
 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

2 SYSTEM DESIGN
The method consists of three steps: (1) bilingual noun
word extraction, (2) identification of domain-specific
senses, and (3) bilingual sense correspondence. Here-
after, we describe bilingual noun words as BN words
and bilingual noun word senses as BNSs.
2.1 Local Data View
The first step is to extract BN words from the corpora.
This process consists of two sub-steps: (i) retrieval
of relevant documents, and (ii) BN word extraction
based on sentence level similarity.
2.1.1 Retrieval of Relevant Documents
We used Reuters’96 and the Mainichi Japanese news-
paper documents. Let d
J
i
(1 ≤ i ≤ s) and d
E
j
(1 ≤ j ≤
t) be a Mainichi document and a Reuters document,
respectively. Each Reuters document d
E
j
is trans-
lated into a Japanese document d
E mt
j
using English-
Japanese MT software. We calculated similarity be-
tween two documents using BM25 (Robertson and
Walker, 1994), which is widely used in information
retrieval studies
2
. BM25 is given by:
BM25(d
J
i
,d
E mt
j
) =
∑
w∈d
E mt
j
w
(1)
(k
1
+ 1)t f
i
K +t f
i
(k
3
+ 1)qt f
j
k
3
+ qt f
j
, (1)
where w is a word within d
E mt
j
; w
(1)
= log
(t−n+0.5)
(n+0.5)
is the weight of w; t is the number of Mainichi doc-
uments; n is the number of documents containing w;
K refers to k
1
((1 − b) + b
dl
i
avdl
); k
1
, b, and k
3
are pa-
rameters set to 1, 1, and 1,000, respectively; dl
i
is
the document length of d
J
i
; avdl is the average doc-
ument length in words; and t f
i
and qt f
j
are the fre-
quencies of occurrence of the w (∈ d
E mt
j
) in d
J
i
, and
d
E mt
j
, respectively. If the similarity value between d
J
i
and d
E mt
j
is higher than the lower bound L
θ
, these are
regarded as relevant documents, and a document pair
is created.
2.1.2 Bilingual Word Extraction
We assume that BN word correspondences obtained
using relevant documents are unreliable as many noun
words appear in a pair of relevant documents. There-
fore, we applied sentence level retrieval. First, we ap-
plied simple χ
2
statistics to the extracted pairs of rel-
evant documents and extracted Mainichi noun word
w
J
and Reuters noun word w
E
pairs with χ
2
values
2
http://trec.nist.gov/
Table 1: Variables of χ
2
statistics.
w
E
¬w
E
w
J
f (w
J
,w
E
) = a f (w
J
,¬ w
E
) = b
¬w
J
f (¬ w
J
,w
E
) = c f (¬ w
J
, ¬ w
E
) = d
greater than zero. The χ
2
statistic measures the lack
of independence between w
J
and w
E
, and can be com-
pared to the χ
2
distribution with one degree of free-
dom to judge extremeness (Yang and Pedersen, 1997).
It is defined by:
χ
2
(w
J
,w
E
) =
(ad − bc)
2
(a +b)(a + c)(b + d)(c + d)
. (2)
a, b, c, and d in Eq. (2) are shown in Table 1. f (x, y)
in Table 1 refers to the co-occurrence frequencies of
x and y on the Japanese and English sides, respec-
tively. Next, for each w
J
and w
E
pair, we apply sen-
tence level similarity given by:
S sim(w
J
,w
E
) =
max
S w
J
∈Set
J
,S w
E
∈Set
E
sim(S w
J
,S w
E
),
where
sim(S w
J
,S w
E
) =
| S w
J
∩ S
mt
w
E
|
| S w
J
| + | S
mt
w
E
| −2× | S w
J
∩ S
mt
w
E
| +2
. (3)
Set
J
and Set
E
are sets of sentences that include w
J
and
w
E
, respectively; |X| is the number of noun words in a
sentence X ; | S w
J
∩ S
mt
w
E
| refers to the number of
noun words that appear in both S w
J
and S
mt
w
E
; and
S
mt
w
E
is the translation result of S w
E
. The larger
value of sim(S w
J
,S w
E
) indicates the more similar
these two sentences S w
J
and S w
E
. As shown in Eq.
(4), we retrieved w
J
and w
E
as the BN word such that
the similarity between w
J
and w
0
E
∈ BP(w
J
) are the
largest value. Here, BP(w
J
) is a set of BN word pairs,
each of which includes w
J
on the Japanese side.
hw
J
,w
E
i = arg max
hw
J
,w
0
E
i∈BP(w
J
)
S
sim(w
J
,w
0
E
). (4)
2.2 Global Data View
The second step is to identify domain-specific senses
for each noun word. Globally, we used bilingual cate-
gory/domain correspondences, Reuters and Mainichi
categories. The process is applied independently to
the English (WordNet) and Japanese (EDR) dictionar-
ies. We used the MRW model to ranking the senses.
For each Reuters category c
E
and Mainichi category
c
J
such as “sports” and “economy” assigned to the
Reuters (Mainichi) documents, we created a Graph G

= (S, E) that reflects the relationships between senses
in a set S. S refers to a set of noun word senses in the
Reuters (Mainichi) documents assigned to the cate-
gory c
E
(c
J
). Each sense s
i
∈ S is represented by a
vector. Each dimension of a vector corresponds to
each word appearing in syn ∪ gloss, where syn in-
dicates a synset and gloss refers to a gloss text in
a dictionary. Each element of a dimension is a fre-
quency count of the word in syn ∪ gloss. E is a set
of edges, which is a subset of S ∗ S. Each edge e
i j
∈
E is associated with an affinity weight aw(i → j). We
used two affinity types between i and j. Directed re-
lation is a hyponym relation, and undirected relation
is a synonym relation. The weight is computed using
the standard cosine measure between two senses. The
transition probability from s
i
to s
j
is then defined by
normalizing the corresponding affinity weight:
p(i → j) =









aw(i→ j)
|S|
∑
k=1
aw(i→k)
, if Σaw 6= 0
0 , otherwise.
(5)
We used the row-normalized matrix U =
(U
i j
)
|S|∗|S|
to describe G with each sense correspond-
ing to the transition probability, where U
i j
= p(i → j).
To make U a stochastic matrix, the rows with all zero
elements are replaced by a smoothing vector with all
elements set to
1
|S|
. The matrix form of the saliency
score Score(s
i
) can be formulated in a recursive form
as in the MRW model:
~
λ = µU
T
~
λ +
(1−µ)
|S|
~e, where
~
λ
= [Score(s
i
)]
|S|∗1
is a vector of saliency score for the
senses. ~e is a column vector with all elements equal
to 1. µ is a damping factor. We set µ to 0.85, as in the
PageRank (Brin and Pagee, 1998). The final asym-
metric transition matrix M is given by:
M = µU
T
+
(1 − µ)
| S |
~e~e
T
(6)
Each sense score in a specific category is obtained
by the principal eigenvector of the new transition ma-
trix M. We applied the algorithm for each category.
3
For implementation, we used the Eigen library.
4
We
chose a vector with the largest eigenvalues. We nor-
malized a vector, and obtained rank scores of senses.
2.3 Bilingual Sense Correspondence
The final step is to retrieve BNSs using both results
obtained by local and global data views. Let hw
J
,w
E
i
3
The principal eigenvector is obtained by the power
method and inverse iteration method.
4
http://eigen.tuxfamily.org/index.php?title=Main Page
∈ Set
{w
J
,w
E
}
be a pair of nouns obtained by the BN
word extraction procedure where w
J
has the number
of m senses w
J s
i
(1 ≤ i ≤ m) and w
E
has the n senses
w
E s
j
(1 ≤ j ≤ n), respectively. We retrieved BNS
pair w
J s
i
and w
E s
j
:
hw
J s
i
,w
E s
j
i
(c
J
,c
E
)
s.t. w
J s
i
∈ Set
c
J
,
w
E s
j
∈ Set
c
E
,
hw
J
,w
E
i satisfies RNN. (7)
We recall that the corpora used to identify domain-
specific senses are Reuters and Mainichi documents,
each of which has different categories. Therefore,
we estimated the category correspondence accord-
ing to the χ
2
(c
J
,c
E
) statistics shown in Eq. (2).
In Eq. (2), we simply replaced w
J
by c
J
and w
E
by c
E
as each document of relevant document pairs
has category information. The subscript (c
J
,c
E
) of
hw
J s
i
,w
E s
j
i
(c
J
,c
E
)
in formula (7) refers to the re-
sult of category correspondence between Reuters and
Mainichi obtained by Eq. (2). Set
c
J
and Set
c
E
show
ranked sense lists for c
J
, and c
E
, respectively, ob-
tained by the MRW model. RNN is the so-called Re-
ciprocal Nearest Neighbors in that two noun senses
are each other’s most similar noun (Hindle, 1990).
3 EXPERIMENTS
We evaluated each of the three procedures in the ex-
periments, (1) BN extraction by using the results of
relevant documents retrieval, (2) retrieving domain
specific senses, and (3) BNS correspondences.
3.1 Local Data View
In this subsection, we report the results of BN extrac-
tion.
3.1.1 Experimental Setup
We used Reuters’96 and Mainichi Japanese newspa-
per corpora from 20 August 1996 to 19 August 1997.
The Reuters’96 corpus consists of 806,792 documents
organized into three types of coarse-grained cate-
gories, i.e., topic, industry, and region. Each consists
of 126, 870, and 366 categories. The Mainichi cor-
pus consists of 119,051 documents organized into 16
categories. The difference in dates between Reuters
and Mainichi documents is less than ±3 days, e.g.,
when the date of the Reuters document is 27 August,
the corresponding Mainichi data is from 24 to 30 Au-
gust. We set a small date difference because if some

Table 2: Bilingual noun extraction.
Pairs Eng Jap # of bilingual # of nouns IRS
nouns (top 1,000) (top 1,000)
Docs
196,368 133,854 20,882
172,895 162 2.32
Docs & Sent 115,918 329 4.83
event occurred at some specific time period, the press
of all countries report it on one of these days. We used
English-Japanese MT software (Internet Honyakuno-
Ousama for Linux, Ver.5, IBM Corp.). Each Reuters
document was translated into a Japanese document,
and BM25 was applied.
The training data for choosing the lower bound L
θ
were Reuters 20 August 1996 and Mainichi from 17
to 23 August 1996. The total number of English and
Japanese documents collected were 2,586 and 2,137,
respectively, and the number of relevant documents
collected manually for evaluation was 157. The clas-
sification was determined to be correct if the two hu-
man judges agreed on the evaluation. We used the
F-score for evaluation of relevant document retrieval,
which is a measure that balances precision (Prec) and
recall (Rec). Let cSet be a set of correct document
pairs. The definitions of Prec and Rec are given by:
Prec =
| {(d
J
,d
E
) | (d
J
,d
E
) ∈ cSet, BM25(d
J
,d
E
) ≥ L
θ
} |
| {(d
J
,d
E
) | BM25(d
J
,d
E
) ≥ L
θ
|
Rec =
| {(d
J
,d
E
) | (d
J
,d
E
) ∈ cSet, BM25(d
J
,d
E
) ≥ L
θ
} |
| {(d
J
,d
E
) | (d
J
,d
E
) ∈ cSet |
The best performance of F-score was 0.564 when the
L
θ
value was 150. We used the value (L
θ
= 150) to
extract BN word pairs. We used one year of Reuters
and Mainichi documents except for the training data
that were used to estimate L
θ
values. The difference
in dates between them was less than ±3 days.
3.1.2 Results
The results of BN words extraction are shown in Table
2. We compared the results obtained by our method,
“Docs & Sent” with the results obtained by only ap-
plying χ
2
statistics to the results of relevant docu-
ments, “Docs” to examine how the sentence-based
retrieval influences the performance. “Pairs” in Ta-
ble 2 shows the number of bilingual document pairs.
“Eng” and “Jap” show the numbers of Reuters and
Mainichi documents within the pairs satisfying the
similarity lower bound L
θ
= 150, respectively. “# of
nouns” shows the number of correct BN in the top-
most 1,000 according to the χ
2
statistics (Docs) and
sentence-based similarity (Docs & Sent).
As shown in Table 2, sentence-based retrieval con-
tributed to a reduction in the number of useless BN
words without a decrease in accuracy, as about 67%
(115,918/172,895) of the size obtained by “Docs” was
retrieved, while about 2.03 (329/162) times the num-
ber of correct BN words were obtained in the topmost
1,000 nouns. “IRS” (Inverse Rank Score) is a mea-
sure of system performance by considering the rank
of correct BN words within the candidate nouns; it is
the sum of the inverse rank of each matching noun,
e.g., correct BN words by manual evaluation matches
at ranks 2 and 4 give an IRS of
1
2
+
1
4
= 0.75. A higher
IRS value indicates better system performance. Table
2 shows that sentence-based retrieval also contributes
to ranking performance compared with the results ob-
tained by applying χ
2
statistics only.
3.2 The Global Data View
Secondly, we report the results of retrieving domain
specific senses from the WordNet and EDR dictionar-
ies.
3.2.1 WordNet 3.0
We assigned the Reuters categories to each sense of
words in WordNet. We selected 38 Reuters cate-
gories, each of which is assigned to more than 80,000
documents. For each category, we collected noun
words with frequencies ≥ 5 within a document from
the one-year Reuters corpus. There are no existing
sense-tagged data for Reuters categories that could be
used for evaluation. Therefore, we selected a limited
number of words and evaluated these words qualita-
tively. To do this, we used SFC resources (Magnini
and Cavaglia, 2000), which annotate WordNet 2.0
synsets with domain labels. We selected 4 cate-
gories that are easy to manually identify correspond-
ing Reuters and SFC categories. Statistics of data
and the results are shown in Table 3. “Words” shows
the number of different words with frequency ≥ 5
within a document assigned to the category shown
in “Reu,” and “Sense” shows their total number of
senses. “SFC” refers to the number of senses appear-
ing in the SFC resource. “DSS” (Domain-Specific
Senses) is the number of senses assigned by our
method. “DSS of the # of Correct S” shows the num-
ber of correct senses in the topmost 1,000, and “D&S”
refers to the number of correct senses appearing in
both DSS and SFC.

Table 3: The results of sense assignments (WordNet).
Reu / SFC Words Senses SFC DSS
# of Correct S
IRS
DSS D&S
Economics / Economy 10,284 32,550 1,032 6,355 623 120 2.31
Sports / Sports 7,437 26,478 339 4,457 573 158 2.01
War / Military 7,681 26,831 913 4,366 510 68 1.41
Politics / Politics 10,349 32,536 793 5,981 612 103 2.21
Average 8,938 29,598 769 5,290 580 112 1.99
Table 4: DSS against the FS heuristic (WordNet).
Cat Sense DSS FS
Economics 5.6 0.73 0.68
Sports 4.5 0.71 0.69
War 4.9 0.52 0.30
Politics 4.2 0.75 0.68
Average 4.8 0.68 0.59
As shown in Table 3, the numbers of correct
senses obtained by “DSS” and “D&S” did not exactly
match. This is not surprising because our method
used the Reuters corpus, while the SFC resource con-
sists of 96% of the WordNet synsets, each of which
is manually annotated using 115 different SFC. The
IRS depends on the category, and the average IRS
was 1.99. It is interesting to note that some senses of
words that were obtained correctly by our approach
did not appear in the SFC resource. For example, the
words “shot” and “strike” in the Sport category, and
“liberty” and “military hospital” in the war/military
category were obtained by our approach but did not
appear in the SFC. This is because we used WordNet
3.0, while SFC was based on WordNet 2.0. These
observations clearly support the usefulness of our au-
tomated method.
In the WSD task, a first sense (FS) heuristic is of-
ten applied because of its power and lack of a require-
ment for expensive hand-annotated data sets (Cotton
et al., 1998; McCarthy et al., 2007). We compared the
results obtained by DSS to those obtained by the FS
heuristic. For each category, we randomly selected 10
words from the senses assigned by DSS, and selected
10 sentences from the documents belonging to each
corresponding category. As a result, we tested 100
sentences for each category. Table 4 shows the results.
“Sense” refers to the number of average senses per
word. The average precision (Avg) of our method was
0.68, while the result obtained by the FS was 0.59.
3.2.2 EDR Dictionary
We assigned categories from Mainichi Japanese doc-
uments to each sense of words in the EDR Japanese
Table 5: The results of sense assignments (EDR).
Cat Words Senses Cor IRS
Economics 15,906 30,869 740 6.39
Sports 17,556 33,595 559 2.77
International 13,906 27,239 451 5.63
Average 15,789 30,568 583 4.93
word dictionary
5
. We selected 11 out of 16 categories,
each of which had a sufficient number of documents.
All documents were tagged by the morphological an-
alyzer Chasen (Matsumoto et al., 2000), and noun
words were extracted. We choose three categories for
which it is easy to manually create correct data. Ta-
ble 5 shows the statistics of the data and the results
of assignment. “Cor” shows the number of senses as-
signed by our approach correctly in the topmost 1,000
senses. The average IRS obtained by the EDR was
4.93 and Reuters was 1.99. This is reasonable as
the assignment task using EDR is easy compared to
WordNet, i.e., the average number of senses per word
of the former is 1.93, while that of the latter is 3.31.
Table 6: DSS against the FS heuristic (EDR).
Cat Sense DSS FS
Economics 2.793 0.79 0.60
Sports 2.873 0.65 0.52
International 2.873 0.68 0.58
Average 2.846 0.70 0.57
In the WSD task, we randomly selected 30 words
from the senses assigned by DSS. For each word, we
selected 10 sentences from the documents belonging
to each corresponding category. The FS in the EDR is
determined based on the EDR corpus. Table 6 shows
the results. As can be seen in Table 6, DSS was also
better than the FS heuristics in Japanese data. The
overall performance for the FS (0.57) was not better,
similar to the case for the English data (0.59), while
the number of senses per word in the Japanese re-
source was smaller than that in WordNet. There were
many senses that did not occur in the EDR corpus, i.e.,
5
www2.nict.go.jp/out-promotion/techtransfer/EDR/
index.html?

Table 7: The results of bilingual sense correspondences.
Approach Candidates # of Correct ext.(1,000) RNN (1,071) IRS
BN BNS BN BNS BN BNS BN BNS
Local 171,895 —– 329 —– 431 —– 2.32 —–
Local & Global 171,895 115,918 437 312 679 580 4.10 4.35
Table 8: Examples of bilingual noun senses.
Cat pair # of Pairs # of Correct Examples
(Mai, Reu) Cand RNN senses(%) Sense id (gloss text)
(International, 269 10 9 (90.0) EDR: 01 (an irregular group of soldiers
given to sneak attacks)
Government) WordNet: guerrilla 01 (a member of an irregular
armed force)
(Economics, 4,964 58 44 (75.9) EDR: 09 (stock, share)
Economics) WordNet: stock 01 (the capital raised by a corpora-
tion through the issue of shares)
(Sports,
Sports)
6,560 68 48 (70.6) EDR: 02 (in golf, the average number of
strokes for playing around a course)
WordNet: par 01 (the number of strokes set for each
hole on a golf course)
(Local news, 2,903 34 23 (67.6) EDR: 02 (the act of killing someone)
Crime) WordNet: killer 01 (someone who causes the death of
a person)
62,460 nouns appeared in both EDR and Mainichi
newspapers (from 1991 to 2000), 164,761 senses in
all. Of these, 114,267 senses did not appear in the
EDR corpus. This also demonstrates that automatic
DSS works well compared to the frequency-based FS
heuristics.
3.3 Bilingual Sense Correspondence
Finally, we evaluated the performance of BNS corre-
spondences.
3.3.1 Experimental Setup
The data for BNS correspondence were the Reuters
and Mainichi corpora from the same period, i.e., 20
August 1996 to 19 August 1997. The total num-
bers of documents were 806,791, and 119,822, re-
spectively. Locally, we extracted BN words us-
ing sentence-based similarity. We retrieved cross-
lingually relevant Japanese documents with English
documents. The difference in dates between them
was less than ±3 days. We used these documents
to extract BN words. Globally, we assigned domain-
specific senses for each of the 38 categories for Word-
Net 3.0 and 11 categories for EDR. We estimated cat-
egory correspondences, and retrieved BNSs accord-
ing to their correspondences. We obtained 92 cate-
gory correspondences in all with χ
2
values larger than
zero. From these data, we extracted BNSs.
3.3.2 Results
Table 7 shows the results of BNS correspondences.
“Local” indicates the results using only sentence-
based similarity, and “Local & Global” shows the
results obtained by our method. “# of Correct ext.”
refers to the number of correct extractions within the
topmost 1,000, and “RNN” shows the number of cor-
rect extractions by applying RNN. “BN” refers to the
number of BN words, and “BNS” is their senses.
As can be clearly seen in Table 7, the results with
integration of local and global data views improved
overall performance of BN word extraction compared
to local data only because 10.8% (437−329)/1,000
improvement within the topmost 1,000, and 23.1%
(679−431)/1,071 improvement with RNN. We ob-
tained 312 BNSs within the topmost 1,000, although
bilingual sense correspondence is a difficult task.
Moreover, RNN is effective for BNS correspon-
dences. We obtained a total of 1,071 BNSs, which
580 were correct, while that without RNN was 312
within the topmost 1,000.
Table 8 lists examples of BNSs for each category
correspondence. (x, y) of the category pair refers to
the Mainichi and Reuters category correspondence.
“Cand” denotes the number of extracted word pairs,
and “RNN” shows word pairs obtained by RNN.
“Sense id” shows the sense and its order of appear-
ance in each dictionary, WordNet and EDR. Table 8

Table 9: Constituents of bilingual sense of words.
Eng authority, board, budget, business, case
coast, company deficit, finance, group, issue
money, opposition people, power, president
seat, space, state, trial
Jap , , , , ,
, , , , , , ,
, , , , , , ,
, , , , , ,
, ,
Table 10: Examples of bilingual senses.
Cat pair Sense pair
(Mai, Reu) (Mai, Reu)
(top news, industrial) ( 01, authority 01)
(international, international) ( 02, authority 05)
(top news, election) ( 02, case 01)
(top news, violence) ( 01, case 02)
(top news, funding) ( 01, deficit 01)
(top news, politics) ( 02, deficit 02)
shows that the first three senses of words, “guerrilla,”
“stock,” and “par” corresponded correctly. However,
“the act of killing someone” in the EDR was in-
correctly identified by “killer” (person) in WordNet.
Our approach for identifying BNSs is based on term-
based corpus statistics. It will be necessary to inves-
tigate other types of lexicon, such as verb and subjec-
tive/objective noun collocations, for further improve-
ment. Our method for category correspondence is
very simple as it is based on the χ
2
statistics. Much of
the previous work on text categorization indicated that
hierarchical structure (e.g., MeSH, Yahoo!, LookS-
mart) improves the accuracy of text categorization,
and it is worth attempting to use a hierarchical struc-
ture to determine corresponding categories with high
accuracy.
Finally, we compared the results of bilingual sense
correspondences with a machine readable bilingual
dictionary and two English-Japanese MT systems.
We randomly choose 80 BNSs from 580 correct bilin-
gual senses obtained by RNN. There were 40 different
category pairs, and as shown in Table 9, the number
of different English and Japanese words were 20, and
32, respectively. Table 10 shows examples of BNSs
and its corresponding category pairs.
We choose titles of documents as a translation test
data. Because they are short sentences, and hard to be
affected by a syntactic analyzer in MT systems which
enable to approximate a fair comparison. For each of
the 80 BNS pairs, we created translation test data.
For example, in the word, “authority” of BNS pair
( 01, authority 01) in Table 10, we randomly se-
lected 10 titles sentences including “authority” from
the Reuters documents assigned to the category “in-
Table 11: Comparison against a dictionary and MT systems.
# of Correct senses (%)
Eijirou 329 / 800 (41.2)
Honyakuno-Ousama 521 / 800 (65.1)
SYSTRANet 550 / 800 (68.8)
RNN 562 / 800 (70.3)
dustrial”. We translated these test data by using two
MT systems
6
, and compared the translation result of
“authority”. We also compared the results obtained by
our method with English-Japanese dictionary, Eijirou
on the WEB
7
. In the English-Japanese dictionary, we
used the first-sense heuristic (choosing the first sense
of a word). The comparative results are shown in Ta-
ble 11.
As we can see from Table 11 that the results ob-
tained by RNN was 70.3%, and it was better to the
results obtained by bilingual dictionary (41.2%), and
slightly better to the results obtained by SYSTRANet
(68.8%). Moreover, it works well compared to the
Honyakuno-Ousama (MT) that we used in the process
of BN word extraction as local data view. It can be ob-
served from these results that the extracted BNSs can
be used as a lexical resource for MT system.
4 RELATED WORK
Our approach to link bilingual word senses in dictio-
naries can be regarded as a type of ontology align-
ment. The earliest such attempts were Chimaera
(McGuinness et al., 2000) and PROMPT (Noy and
Musen, 2000). Suchanek et al. developed an ontol-
ogy alignment system called PARIS which relies on
instance overlap-based cues to align instances, cate-
gories, and relations from two knowledge bases. They
reported that the method is effective, although it re-
mains insufficient to align ontologies that share few
or no data entries in common. Wijaya et al. focused
on the problem, and presented a method of aligning
ontologies called PIDGIN that employs a very large
natural language text as interlingua and graph-based
self-supervised learning (Wijaya et al., 2013). The
use of corpora is similar to our method, although the
target of integration is quite different, i.e., PIDGIN
aimed at relation and category alignment, while our
method aligned noun word senses.
In the context of bilingual lexicon extraction,
much of the previous work used comparable cor-
pora. One attempt involved directly retrieving bilin-
6
We used Internet Honyakuno-Ousama, and SYSTRANet
English-Japanese MT software. www.systranet.com/translate
7
www.alc.co.jp

gual lexicons from corpora (Fung and Cheung, 2004;
Gaussier et al., 2004). The alternative approach con-
sists of two steps: first, cross-lingual relevant doc-
uments are retrieved from comparable corpora, and
then bilingual term correspondences within these rel-
evant documents are estimated. Much of the previ-
ous work in finding relevant documents used MT sys-
tems or existing bilingual lexicons to translate one
language into another (Utsuro et al., 2003). Docu-
ment pairs are then retrieved based on document sim-
ilarity. Another approach to retrieving relevant doc-
uments involves the collection of relevant document
URLs from the WWW (Resnik and Smith, 2003).
Munteanu et al. proposed a method for extracting par-
allel sub-sentential fragments from very non-parallel
bilingual corpora (Munteanu and Marcu, 2006). All
of these methods successfully extracted bilingual lex-
icons but they ignored the meanings of words. One at-
tempt to deal with the meaning of words is Kusner et
al.’s method (Kusner et al., 2015). They presented the
Word Mover’s Distance (WMD) between text doc-
uments by utilizing word2vec embeddings (Mikolov
et al., 2013a; Mikolov et al., 2013b). Word2vec learns
a vector representation for each word using a neural
network architecture consisting of three layers, i.e. an
input layer, a projecting layer, and an output layer to
predict nearby words.
There also has been a lot of work where bilin-
gual word vectors are induced using parallel corpora
(Brown et al., 1993; Haghighi et al., 2008; Dyer et al.,
2013; Kocisky et al., 2014). Dyer et al. presented
an alignment model called FASTALIGN model which
uses an alignment distribution defined by a single pa-
rameter that measures how close the alignment is to
the diagonal (Dyer et al., 2013). Blunsem et al. ex-
tended Dyer et al.’s model for learning bilingual word
representations. They marginalize out the alignments
which enable to capture more bilingual semantic con-
text (Kocisky et al., 2014). Gouws et al. proposed
a simple and computationally efficient model called
BioBOWA (Bilingual Bag-of-Words without Align-
ments) for learning bilingual distributed representa-
tions of words which can scale to large monolingual
datasets, and does not require word-aligned parallel
training data (Gouws et al., 2015). The method re-
quires monolingual data which trains directly, and ex-
tracts a bilingual signal from a smaller set of raw-text
sentence-aligned data. They evaluated the induced
cross-lingual embeddings on the two tasks, i.e. doc-
ument classification and lexical translation task, and
the results showed that the method outperforms cur-
rent state-of-the-art methods, especially it contributes
to reduce computational cost.
In the context of domain-specific senses of a
word, Magnini et al. presented a lexical resource
where WordNet 2.0 synsets were annotated with Sub-
ject Field Codes (SFC) by a procedure that exploits
the WordNet structure (Magnini and Cavaglia, 2000;
Bentivogli et al., 2004). 96% of the WordNet synsets
of the noun hierarchy could have been annotated us-
ing 115 different SFC, while identification of the do-
main labels for senses required a considerable amount
of hand-labeling. McCarthy et al. presented a method
to find domain-specific predominant noun senses au-
tomatically using a thesaurus acquired from raw tex-
tual corpora and the WordNet similarity package (Mc-
Carthy et al., 2007). They tested two domains,
“Sports” and “Finance.”
indent To our knowledge, there have been only a few
previous work on bilingual sense extraction (Plous
and Ji, 2003). One approach is word translation dis-
ambiguation presented by Li et al. (Li and Li, 2004).
Their method is based on one-to-many sense map-
ping. They used a machine learning technique that
repeatedly constructs classifiers in the two languages
in parallel. They reported that the approach sig-
nificantly outperformed existing methods using two
nouns (Pedersen, 2000), and seven of the twelve En-
glish words studied in WSD research by (Yarowsky,
1995). Their method requires a small number of
sense-tagged training data in both of the two lan-
guages, while our method requires documents as-
signed to categories and a dictionary with gloss text
which unfortunately hinders a direct and fair compar-
ison between their system and ours.
Our method has three novel aspects. First, we pro-
pose a method to integrate different data views to im-
prove the quality of bilingual sense correspondence.
Second, from the perspective of existing knowledge-
based integration, we propose a method for corre-
sponding senses between two monolingual dictionar-
ies via many-to-many sense mapping. Finally, from
the perspective of robustness, the method is auto-
mated, and requires only documents assigned to do-
mains/categories and a dictionary with gloss text. It
can be applied easily to a new domain, sense in-
ventory, or different languages given sufficient doc-
uments.
5 CONCLUSION
We have developed an approach for linking and cre-
ating bilingual sense correspondences by combining
local and global data views. Future work will in-
clude: (i) investigating other types of lexicon for fur-
ther improvement, (ii) extending the method to use hi-
erarchical structures of categories for category corre-

spondence, (iii) retrieving other parts of speech word
senses, and (iv) evaluation of the method using dic-
tionaries other than English WordNet and Japanese
EDR.
ACKNOWLEDGEMENTS
The authors would like to thank Bernardo Magnini for
providing SFC resources, and the anonymous review-
ers for their comments and suggestions. This work
was supported by the Telecommunications Advance-
ment Foundation, and Support Center for Advanced
Telecommunications Technology Research, Founda-
tion.
REFERENCES
Bentivogli, L., Forner, P., Magnini, B., and Pianta, E.
(2004). Revising the wordnet domains hierarchy: Se-
mantics, coverage and balancing. In Proceedings of
the Workshop on Multilingual Linguistic Ressources,
pages 101–108.
Brin, S. and Pagee, L. (1998). Lexical Issues in Natural
Language Processing. In Proc. of the 7th International
Conference on World Wide Web 7, pages 107–117.
Briscoe, E. J. (1991). Lexical Issues in Natural Language
Processing. In Natural Language and Speech. Pro-
ceedings of the Symposium on Natural Language and
Speech, pages 39–68.
Brown, P. F., Pietra, V. J. D., Pietra, S. A. D., and Mercer,
R. L. (1993). The Mathematics of Statistical Machine
Translation: Parameter Estimation. 19(2):263–311.
Cotton, S., Edmonds, P., Kilgarriff, A., and
Palmer, M. (1998). SENSEVAL-2,
http://www.sle.sharp.co.uk/senseval2/.
Dyer, C., Chahuneau, V., and Smith, N. A. (2013). A Sim-
ple, Fast, and Effective Reparameterization of IBM
Model 2. In Proc. of the 2013 Conference of the
North American Chapter of the Association for Com-
putational Linguistics: Human Language Technolo-
gies, pages 644–648.
Fung, P. and Cheung, P. (2004). Mining Very Non-Parallel
Corpora: Parallel Sentence and Lexicon Extraction
via Bootstrapping and EM. In Proc. of 2004 Con-
ference on Empirical Methods in Natural Language
Processing, pages 57–63.
Gale, W. A., Church, K. W., and Yarowsky, D. (1992). One
Sense Per Discourse. In Proc. of Speech and Natural
Language Workshop 1992, pages 233–237.
Gaussier, E., Renders, H.-M., Matveeva, I., Goutte, C., and
D
´
ejean, H. (2004). A Geometric View on Bilingual
Lexicon Extraction from Comparable Corpora. In
Proc. of the 42nd Annual Meeting of the Association
for Computational Linguistics, pages 527–534.
Gouws, S., Bengio, Y., and Corrado, G. (2015). Bil-
BOWA:Fast Bilingual Distributed Representations
without Word Alignments. In Proc. of the 32nd In-
ternational Conference on Machine Learning, pages
748–756.
Grishman, R., MacLeod, C., and Meyers, A. (1994). COM-
LEX Syntax: Building a Computational Lexicon. In
Proc. of the 15th International Conference on Compu-
tational Linguistics, pages 268–272.
Haghighi, A., Liang, P., Kirkpatrick, T. B., and Klein, D.
(2008). Learning Bilingual Lexicons from Monolin-
gual Corpora. In Proc. of 46th Annual Meeting of
the Association for Computational Linguistics: Hu-
man Language Technologies, pages 771–779.
Hindle, D. (1990). Noun Classification from Predicate-
Argument Structures. In Proc. of the 28th Annual
Meeting of the Association for Computational Lin-
guistics, pages 268–275.
Kocisky, T., Hermann, K. M., and Blunsom, P.
(2014). Learning Bilingual Word Representations by
Marginalizing Alignments. In Proc. of the 52nd An-
nual Meeting of the Association for Computational
Linguistics, pages 224–229.
Kusner, M. J., Sun, Y., Kolkin, N. I., and Weinberger, K. Q.
(2015). From Word Embeddings to Document Dis-
tances. In Proc. of the 32nd International Conference
on Machine Learning, pages 957–966.
Li, H. and Li, C. (2004). Word Translation Disambigua-
tion Using Bilingual Bootstrapping. In Computational
Linguistics, pages 1–22.
Magnini, B. and Cavaglia, G. (2000). Integrating Subject
Field Codes into WordNet. In Proc. of LREC-2000.
Matsumoto, Y., Kitauchi, A., Yamashita, T., Hirano, Y.,
Matsuda, Y., Takaoka, K., and Asahara, M. (2000).
Japanese Morphological Analysis System ChaSen
Version 2.2.1. In NAIST Technical Report NAIST.
McCarthy, D., Koeling, R., Weeds, J., and Carroll, J.
(2007). Unsupervised Acquisition of Predominant
Word Senses. In Computational Linguistics, volume
33(4), pages 553–590.
McGuinness, D. L., Fikes, R., Rice, J., and Wilder, S.
(2000). An Environment for Merging and Testing
Large Ontologies. In Proc. of the Conference on Prin-
ciples of Knowledge Representation and Reasoning,
pages 483–493.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013a).
Efficient estimation of word representations in vector
space. CoRR, abs/1301.3781.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., and
Dean, J. (2013b). Distributed Representations of
Words and Phrases and Their Compositionality. In
Proc. of NIPS, pages 3111–3119.
Miller, G. A. (1995). Wordnet: A lexical database for en-
glish. Commun. ACM, 38:39–41.
Miller, G. A., Beckwith, R. T., Fellbaum, C. D., Gross, D.,
and Miller, K. J. (1990). WordNet: An on-line Lexical
Database. In International Journal of Lexicography,
volume 3(4), pages 235–244.
Munteanu, D. S. and Marcu, D. (2006). Extracting Parallel
Sub-Sentential Fragments from Non-Parallel Corpora.

In Proc. of the 21st International Conference on Com-
putational Linguistics and 44th Annual Meeting of the
Association for Computational Linguistics, pages 81–
88.
Noy, N. F. and Musen, M. A. (2000). Prompt: Algo-
rithm and Tool for Automated Ontology Merging and
Alignment. In Proc. of the 17th National Conference
on Artificial Intelligence, pages 450–455.
Pedersen, T. (2000). A Simple Approach to Building En-
sembles of Naive Bayesian Classifiers for Word Sense
Disambiguation. In Proc. of the 2nd Conference on
Empirical Methods in Natural Language Processing,
pages 197–207.
Plous, S. and Ji, H. (2003). A Model for Matching Se-
mantic Maps between Language (French/English, En-
glish/French). In Computational Linguistics, pages
155–178.
Resnik, P. and Smith, N. A. (2003). The Web as a Parallel
Corpus. In Computational Linguistics, volume 29(3),
pages 349–380.
Robertson, S. E. and Walker, S. (1994). Some Simple Ef-
fective Approximations to the 2-Poisson Model for
Probabilistic Weighted Retrieval. In Proc. of the
17th Annual International ACM SIGIR Conference on
Research and Development in Information Retrieval,
pages 232–241.
Utsuro, T., Horiuchi, T., Hamamoto, T., Hino, K., and
Nakayama, T. (2003). Effect of Cross-Language IR in
Bilingual Lexicon Acquisition from Comparable Cor-
pora. In Proc. of the 10th Conference of the Euro-
pean Chapter of the Association for Computational
Linguistics, pages 355–362.
Wijaya, D., Talukdar, P. P., and Mitchell, T. (2013). PID-
GIN: Ontology Alignment using Web Text as Inter-
lingua. In Proc. of ACM Conference on Information
and Knowledge Management, pages 589–598.
Yang, Y. and Pedersen, J. O. (1997). A Comparative
Study on Feature Selection in Text Categorization. In
Proc. of the 14th International Conference on Ma-
chine Learning, pages 412–420.
Yarowsky, D. (1995). Unsupervised Word Sense Disam-
biguation Rivaling Supervised Methods. In Proc. of
the 33rd Annual Meeting of the Association for Com-
putational Linguistics, pages 189–196.