Two Sides of a Coin
Translate while Classify Multilanguage Annotations with Domain Ontology-driven
Word Sense Disambiguation
Massimiliano Gioseffi and Angela Locoro
DIBRIS, Sede di Valle Puggia, Dipartimento di Informatica, Bioingegneria, Robotica e Ingegneria dei Sistemi,
University of Genova, Genova, Italy
Keywords:
Domain-driven Word Sense Disambiguation, Domain Ontologies, Text Classification, Natural Language
Processing, Machine Translation.
Abstract:
In this paper we present an approach for the translation and classification of short texts in one step. Our
work lays in the tradition of Domain-Driven Word Sense Disambiguation, though a major emphasis is given
to domain ontologies as the right tool for sense-tagging and topic detection of short texts which, by their
nature, are known to be reluctant to statistical treatment. We claim that in a scenario where users can annotate
knowledge items using different languages, domain ontologies can prove very suitable for driving the word
disambiguation and topic classification tasks. In this way, two tasks are gainfully collapsed in a single one.
Although this study is still in its infancy, in what follows we are able to articulate motivations, design, workflow
analysis, and concrete evolutions envisioned for our tool.
1 INTRODUCTION AND
MOTIVATION
Whenever a Web user enters the Google Image (Gen-
eralistic Search Engine) Area
1
and looks for pictures
that have been annotated with short captions like
“a player reading the score”
at least three different subject types are shown as top
10 ranked results: a music player, a football player
and a videogame player. The words player and score
have two different meanings in this case, depending
on the domain in which they are exploited. The first
two glosses of the words player and score respec-
tively, as they have been taken from WordNet On-
line
2
, are reported below:
player
1. a person who participates in or is skilled at
some game;
2. someone who plays a musical instrument.
1
https://www.google.it/imghp?hl=en&tab=wi. Last ac-
cessed on 21st October 2012.
2
http://wordnetweb.princeton.edu/perl/webwn. Last ac-
cessed on 21st October 2012.
score
1. a number or letter indicating quality perfor-
mance;
2. a written form of a musical composition.
Scenarios demanding automatic or semi-
automatic services for searching Web contents,
translating their annotations while classifying them
according to their topic are more and more reaching
the surface of the user’s needs iceberg. Users clamour
for a domain-oriented systematisation of available
online information with the less effort and the more
effectiveness.
A Web user trying to collect pictures of famous
musicians or a philharmonic institution engaged in the
enrichment of its local repository with Web contents,
or a music community wanting to exchange domain
digital artifacts with worldwide experts are all exam-
ples of subjects interested in services that should be
able to provide a selection, a translation and a classi-
fication of Web contents based on their topic of inter-
est.
In this paper we present an approach for the dis-
ambiguation of words in sentences by means of do-
main ontologies (i.e. semantic objects able to de-
scribe how entities relate, interact and should be in-
358
Gioseffi M. and Locoro A..
Two Sides of a Coin - Translate while Classify Multilanguage Annotations with Domain Ontology-driven Word Sense Disambiguation.
DOI: 10.5220/0004330203580363
In Proceedings of the 5th International Conference on Agents and Artificial Intelligence (ICAART-2013), pages 358-363
ISBN: 978-989-8565-39-6
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
terpreted in a specialised piece of reality), which is
able to frame the translation of short sentences into
their correct context, hence providing the right sense
for each word to be translated. Once all words in
a phrase have been sense-tagged with ontology con-
cepts, the domain of discourse can be extracted from
it in a straightforward way. As a consequence, a clas-
sification by topic for all the sources being annotated
with those short texts can be provided for free.
Although our study is still in its infancy, we be-
lieve that what follows is able to provide a worthy
articulation of our approach. The paper is organised
as follows: Section 2 outlines the related work on
Domain Driven Word Sense Disambiguation and the
main differences with our work, Section 3 presents
the design and the workflow of our system, whereas
Section 4 discusses the evolutions envisioned for our
approach. Section 5 concludes.
2 RELATED WORK
Domain-Driven or Domain-Oriented Word Sense
Disambiguation (Navigli, 2009) is strongly focused
on providing the most appropriate sense label for a
word that is being used in domain-specific texts. The
peculiarity of this approach with respect to classical
Word Sense Disambiguation, according to Navigli,
lays in the paradigm “shift from linguistic understand-
ing to a domain-oriented type-based vision of sense
ambiguity”. This is especially true for cross-lingual
Word Sense Disambiguation, where the domain in-
formation of a phrase may result crucial for bringing
positive chances of a close translation.
A major source of domain information for the dis-
ambiguation of words has been in recent years the
WordNet lexical database, as witnessed by several
research studies (Gliozzo et al., 2004), (Cucchiarelli
and Velardi, 1998), (Buitelaar and Sacaleanu, 2001).
In these scenarios WordNet is used as a domain se-
mantic model, especially in its version where synsets
are tagged with domain labels
3
. Based on such mod-
els, score formulas are computed to determine the pre-
dominant sense of a word in a text. However, Word-
Net is not a proper domain ontology. Moreover, most
of these techniques rely on a trained corpus (Koeling
and McCarthy, 2008) (e.g. SemCor
4
and the like) as
a knowledge source, instead of a domain ontology.
Notably, a recent study (Agirre et al., 2009) en-
forces evidences in favour of knowledge-based meth-
ods (among which we include domain ontologies)
3
http://wndomains.fbk.eu/.
4
http://multisemcor.fbk.eu/semcor.php.
for boosting the disambiguation task in domain-
specific environments. The authors claim that, when
tagging domain-specific corpora, knowledge-based
Word Sense Disambiguation is performing better than
generic supervised Word Sense Disambiguation sys-
tems trained on generalistic corpora. The test was
conducted on 41 domain-related and highly polyse-
mous words in the two domains of Sports and Fi-
nance. The algorithm used is called Personalised Page
Rank and was applied to WordNet graph in order to
rank word senses.
These researches were conducted as a monolin-
gual task. In addition, very few attempts have been
made in the direction of developing Domain-Driven
Word Sense Disambiguation to real case applications.
The Omega ontology (Philpot et al., 2010) was
conceived as a synthesis of WordNet and Mikrokos-
mos (O’Hara et al., 1998), (Mahesh, 1996), a con-
ceptual resource properly designed to support trans-
lation. Besides the core concept base, Omega was
designed to connect with a range of auxiliary knowl-
edge sources, including domain ontologies, incorpo-
rated into the basic conceptual structure and represen-
tation.
In this paper we try to extend these directions of
research by exploiting ontologies conceived by do-
main experts as our knowledge source, and short texts
annotations of domain specific digital sources as our
target of disambiguation, translation and classifica-
tion tasks.
3 SYSTEM ARCHITECTURE
In this Section we will briefly depict the main steps of
our approach, and will give more details of the disam-
biguation and classification algorithm.
3.1 System Workflow
Figure 1 shows the main components, outputs and
data support sources of our system.
The purpose of our approach is the translation and
classification of a sentence in English into a sentence
in Italian by means of a domain ontology-driven word
sense disambiguation algorithm. The classification
by topic of the target sentence is obtained thanks to
the ontology that has been acknowledged to represent
the correct domain of both the source and the target
sentences after the execution of the domain driven
disambiguation procedure. The main steps of the
algorithm are depicted in the sequel. For sake of
clarity the sentence in English
TwoSidesofaCoin-TranslatewhileClassifyMultilanguageAnnotationswithDomainOntology-drivenWordSense
Disambiguation
359
Figure 1: Workflow of a typical translation and classifica-
tion session with our algorithm.
“the violin player is reading the score”
will be used as a pointwise example throughout
the following description, that is also recalled in the
visual workflow of Figure 1:
1. Tokenisation and lemmatisation of the sentence.
The output of these two basic steps will be the
phrase form:
[the, violin, player, be, read, the, score].
2. Parsing with the Stanford Parser
5
. The output of
this phase will be a Parse Tree with words tagged
with their part-of-speech (POS), and so on.
3. Creation of our Program Tree based on collapsing
the Stanford Parse Tree POS structure, and adding
syntactic dependencies to it, as depicted in Figure
2. In the specific example the POS nodes used in
the Custom Tree are:
[the violin player]
NP NODE ← { NP }
5
http://nlp.stanford.edu/software/lex-parser.shtml.
[is reading the score]
VP NODE ← { VP }
[the violin], [the score]
NOUN NODE ← { DT, NN }
[is reading]
VERB NODE ← { VBZ, VBG }
and some of the word-pair dependencies
collected for the sentence are of the kind:
nsubj(reading-5, player-3)
aux(reading-5, is-4)
det(score-7,the-6)
and so on.
4. For each word in the sentence a draft translation
is tried by means of:
• MultiWordNet
6
English-Italian alignment.
• the data support sources, whose contents and
interdependencies are depicted in Figure 3, if
the word is not found in MultiWordNet.
5. For each noun and verb a disambiguation pro-
cedure is carried out by means of the ontology
loaded and composed of different domains (an ex-
ample of such an ontology is reported in Figure
4). The details of the disambiguation algorithm
are reported in Section 3.2.
6. Conversion from an English grammar to an Ital-
ian grammar phrase structure. This procedure in-
cludes the execution of the following tasks:
• Adjectives and verbs are correctly conjugated
for number and genre, and verb tense, respec-
tively.
• Idiomatic expressions are correctly rendered.
• The final translation is printed. In case the dis-
ambiguation has been carried out with the help
of an ontology, the domain labels of each noun
and verb of the phrase are shown, together with
the more general label of the upper domain to
which the domain labels belong (e.g. Music,
Sports, and so on).
3.2 Disambiguation and Classification
with Domain Ontologies
In case both an ontology like the one depicted in Fig-
ure 4 and a domain specific verb list are available, or
6
http://multiwordnet.fbk.eu/english/home.php.
ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence
360
Figure 2: An example of Program Tree obtained by collapsing the POS structure of the Stanford tree with fewer compacted
notations, and by extending it with word-pair dependencies.
Figure 3: The data support sources for translation optimisa-
tion, seen as MultiWordNet extensions.
one of the two is available instead, the domain dis-
ambiguation of nouns (resp. verbs) of the source sen-
tence is carried out. A specific algorithm is devoted
to the computation of the likelihood for each domain
and for each noun and verb in the sentence. Given w
as the word to be disambiguated, t as one of its trans-
lations in the set of its possible translation T
w
, c as
a concept in one of the domain ontologies sub-trees
O
i
, v as one of the domain specific verbs of the list
DV
i
, and d
i
as a specific domain label, the Algorithm
depicted in 1 is computing the domain likelihood for
each domain d
i
analysed.
Starting from each sub-tree root the algorithm
compares each concept of the ontology with each
translation of the word being disambiguated. Each
time there is an exact match between a translation of
w and an ontology concept c or a verb v in the domain
verbs list, the likelihood for the domain d
i
is incre-
mented by 1. The most probable domain is the one
with highest likelihood (hence, with the highest num-
ber of words matching domain concepts in the ontol-
ogy plus domain verbs in the verbs list, if any). The
winner domain is chosen and a translation is produced
according to such domain words.
Algorithm 1: DisambiguateWithOntology algorithm.
1: procedure DISONT(w, T
w
, O
i
, DV
i
)
2: for all w ∈ Words do . noun or verbs
3: for all t ∈ T
w
do . translations of w
4: for all c ∈ O
i
, v ∈ DV
i
do
5: if t = c, t = v then likelihood
d
i
+ 1
6: end if
7: end for
8: end for
9: end for
10: end procedure
In case more than one domain results with the
same likelihood score, the disambiguation is con-
ducted with the “translation by frequency”: the top
synset of MultiWordNet is taken as the ”lemma set”
L
s
from which the most suitable translation word
t
w
∈ L
s
is selected. The selection is done by choos-
ing the most frequent t
w
in the whole space of all its
synsets and glosses.
In our example, the result of the disambiguation
procedure for our sentence (with domain words
underlined in the English version) will be:
English: [“the violin
player is reading the score”]
Italian: [“il violinista sta leggendo lo spartito”]
(Music:3,Sport:2)
and the classification results (translated in English for
sake of clarity) are the following (in square brackets
both the super-class of each word in the sentence, as
well as the root concept of the winner domain ontol-
ogy sub-tree is set):
TwoSidesofaCoin-TranslatewhileClassifyMultilanguageAnnotationswithDomainOntology-drivenWordSense
Disambiguation
361
Figure 4: An ontology fragment with specific domain sub-trees under the Thing concept. The root of each sub-tree is labelled
with the upper domain label (e.g. Music, Sports, and do on).
“[Domain: Music] the violin [string instrument]
player [music performer] is reading the score [mu-
sic
artefact]”
4 DISCUSSION AND NEW
PROPOSALS
Our approach has potentials in the semantic treatment
of texts that are by their nature short and refer to spe-
cific subjects, objects, and topics of interest, such as
those that users exploit to annotate their personal or
professional digital archives. Statistics is not always
able to capture alone enough features when dealing
with short sentences that can be found isolated from
a document corpus (Wenyin et al., 2010). In addi-
tion, building a domain specific corpus or training a
statistic device on an existing one may result in less
slim or precise results if compared to the exploitation
of a codified knowledge source as a domain ontology,
which is in fact a tool especially adopted to give a
structure, an organisation and a semantic description
of resources in domain specific communities.
Although it is not always possible to disambiguate
with a domain ontology sentences of the form:
[“the player is reading the score”](Music:2,Sport:2)
as they would result in a fair likelihood for two
different domains, a valid counter example could be
the one where the presence of a single specialistic
word may make the difference. For example:
“the strong player was playing the bass in the
city orchestra near the sea and his performance was
good”
may bring to both the music (with concepts:
player, bass, orchestra) and the sports (with con-
cepts: player, bass, sea) domains. However, besides
the mere computation of domain words, the word
orchestra can be considered as a “domain hapax”.
This phenomenon is also reflected in the position of
specialistic words usually placed deep in the domain
ontology hierarchy, and this can be measured, as
exemplified in Table 1: the deepest the level in
the ontology hierarchy, the more chances has the
word (and hence, the sentence) to be assigned to
that specific domain. The degree of specificity of
a word could be considered as a valid criterion in
the topic interpretation of a sentence, hence the
ontology level reached during the disambiguation
procedure can be used as its measure. In the same
vein, if a set of words in a sentence belongs to more
than one domain, a selection measure could be the
“semantic relatedness” of such words in each domain
ontology, expressed as the number of connections
between pairwise concepts. The winner domain
for that sentence could be the one with the highest
semantic relatedness among such words, under the
hypothesis that a sentence tends to express stronger
relations between objects of the reality. An example
is depicted in Figure 5.
Table 1: Example of the deepest level reached when visit-
ing two domain ontology trees during the disambiguation
of the above sentence. According to this measure the Music
domain is chosen.
Domain ontology Ontology level
Music 3
Sports 2
5 CONCLUSIONS AND FUTURE
WORK
In this paper we have prospected an alternative di-
rection to existing domain-driven word sense disam-
biguation methodologies by proposing to exploit do-
main ontologies. We claimed that this is a promising
approach and we gave motivations to our hypothesis.
Our future work will address the testing of domain
ontology-driven word sense disambiguation against
different translation tools, and the extension of our
ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence
362
Figure 5: Semantic relatedness between (black) nodes in
two domain ontologies. The more the nodes are connected
(ontology O
1
), the highest the chance that a sentence be-
longs to that domain of discourse instead of belonging to
the other (ontology O
2
).
knowledge source with its inclusion, for example, into
the Omega ontology. We are also interested in the de-
velopment of ad hoc concepts relatedness measures
that can strength the hypothesis of domain disam-
biguation for a sentence. The application scenario en-
visioned for our tool is that of the translation, classi-
fication and retrieval of multilanguage annotations of
digital contents.
ACKNOWLEDGEMENTS
This work is partially supported by the “Indiana MAS
and the Digital Preservation of Rock Carvings: A
multi-agent system for drawing and natural language
understanding aimed at preserving rock carvings”
MIUR FIRB Project funded by the Italian Ministry of
Education, University and Research under fund iden-
tifier RBFR10PEIT.
REFERENCES
Agirre, E., De Lacalle, O. L., and Soroa, A. (2009).
Knowledge-based wsd on specific domains: perform-
ing better than generic supervised wsd. In Proceed-
ings of the 21st international jont conference on Ar-
tifical intelligence, IJCAI’09, pages 1501–1506, San
Francisco, CA, USA. Morgan Kaufmann Publishers
Inc.
Buitelaar, P. and Sacaleanu, B. (2001). Ranking and select-
ing synsets by domain relevance. In Proceedings of
the NAACL Workshop on WordNet and Other Lexical
Resources: Applications, Extensions and Customiza-
tions.
Cucchiarelli, A. and Velardi, P. (1998). Finding a domain-
appropriate sense inventory for semantically tagging a
corpus. Nat. Lang. Eng., 4(4):325–344.
Gliozzo, A. M., Magnini, B., and Strapparava, C. (2004).
Unsupervised domain relevance estimation for word
sense disambiguation. In EMNLP, pages 380–387.
ACL.
Koeling, R. and McCarthy, D. (2008). From predicting
predominant senses to local context for word sense
disambiguation. In Proceedings of the 2008 Con-
ference on Semantics in Text Processing, STEP ’08,
pages 129–138, Stroudsburg, PA, USA. Association
for Computational Linguistics.
Mahesh, K. (1996). Ontology development for machine
translation: Ideology and methodology.
Navigli, R. (2009). Word sense disambiguation: A survey.
ACM Comput. Surv., 41(2).
O’Hara, T., Mahesh, K., and Niremburg, S. (1998). Lexical
acquisition with wordnet and the mikrokosmos ontol-
ogy. In In Proceedings of the ACL Workshop on the
Use of WordNet in NLP, pages 94–101.
Philpot, A., Hovy, E., and Pantel, P. (2010). The Omega
ontology. Cambridge University Press.
Wenyin, L., Quan, X., Feng, M., and Qiu, B. (2010). A short
text modeling method combining semantic and statis-
tical information. Information Sciences, 180(20):4031
– 4041.
TwoSidesofaCoin-TranslatewhileClassifyMultilanguageAnnotationswithDomainOntology-drivenWordSense
Disambiguation
363
