=Paper=
{{Paper
|id=None
|storemode=property
|title=Scenario-Driven Selection and Exploitation of Semantic Data for Optimal Named Entity Disambiguation
|pdfUrl=https://ceur-ws.org/Vol-925/paper_6.pdf
|volume=Vol-925
|dblpUrl=https://dblp.org/rec/conf/ekaw/AlexopoulosRG12
}}
==Scenario-Driven Selection and Exploitation of Semantic Data for Optimal Named Entity Disambiguation==
https://ceur-ws.org/Vol-925/paper_6.pdf
Scenario-Driven Selection and Exploitation of
Semantic Data for Optimal Named Entity
Disambiguation
Panos Alexopoulos, Carlos Ruiz, and José-Manuel Gómez-Pérez
iSOCO, Avda del Partenon 16-18, 28042, Madrid, Spain,
{palexopoulos,cruiz,jmgomez}@isoco.com
Abstract. The rapidly increasing use of large-scale data on the Web has
made named entity disambiguation a key research challenge in Informa-
tion Extraction (IE) and development of the Semantic Web. In this paper
we propose a novel disambiguation framework that utilizes background
semantic information, typically in the form of Linked Data, to accurately
determine the intended meaning of detected semantic entity references
within texts. The novelty of our approach lies in the definition of a struc-
tured semi-automatic process that enables the custom selection and use
of the semantic data that is optimal for the disambiguation scenario at
hand. This process allows our framework to adapt to the particular char-
acteristics of different domains and scenarios and, as experiments show,
to be more effective than approaches primarily designed to work in open
domain and unconstrained situations.
1 Introduction
Information Extraction (IE) involves the automatic extraction of structured in-
formation from texts, such as entities and their relations, in an effort to make
the information of these texts more amenable to applications related to Question
Answering, Information Access and the Semantic Web. In turn, named entity
resolution is an IE subtask that involves detecting mentions of named entities
(e.g. people, organizations or locations) within texts and mapping them to their
corresponding entities in a given knowledge source. The typical problem in this
task is ambiguity, i.e. the situation that arises when a term may refer to mul-
tiple different entities. For example, “Tripoli” may refer, among others, to the
capital of Libya or to the city of Tripoli in Greece. Deciding which reference is
the correct one in a given text is a challenging task which a significant number
of approaches have been trying to facilitate for a long time now [2] [3] [6] [7] [5]
[8].
The majority of these approaches rely on the strong contextual hypothesis
of Miller and Charles [9] according to which terms with similar meanings are
often used in similar contexts. The role of these contexts, which practically serve
as disambiguation evidence, is typically played by already annotated documents
(e.g. wikipedia articles) which are used to train term classifiers. These classifiers
2
link a term to its correct meaning entity, based on the similarity between the
term’s textual context and the contexts of its potential entities [8] [10].
An alternative kind of disambiguation evidence that has recently begun to
be used are semantic structures like ontologies and Linked Data [7] [6] [12]. The
respective approaches typically employ graph-related measures to determine the
similarity between the graph formed by the entities found within the ambiguous
term’s textual context and the graphs formed by each candidate entity’s “neigh-
bor” entities in the ontology. The candidate entity with the best matching graph
is assumed to be the correct one.
An obvious limitation of this is the need for comprehensive semantic infor-
mation as input to the system; nevertheless the increasing availability of such
information on the Web, typically in the form of Linked Data, can help over-
come this problem to a significant degree. Still, however, effectiveness of these
approaches is highly dependent on the degree of alignment between the content
of the texts to be disambiguated and the semantic data to be used. This means
that the ontology’s elements (concepts, instances and relations) should cover the
domain(s) of the texts to be disambiguated but should not contain other addi-
tional elements that i) do not belong to the domain or ii) do belong to it but do
not appear in the texts.
To show why this is important, consider an excerpt from a contemporary foot-
ball match description saying that “Ronaldo scored two goals for Real Madrid”.
To disambiguate the term “Ronaldo” in this text using an ontology, the only
contextual evidence that can be used is the entity “Real Madrid”, yet there are
two players with that name that are semantically related to it, namely Cristiano
Ronaldo (current player) and Ronaldo Luis Nazario de Lima (former player).
Thus, if both relations are considered then the term will not be disambiguated.
Yet, the fact that the text describes a contemporary football match suggests
that, in general, the relation between a team and its former players is not ex-
pected to appear in it. Thus, for such texts, it would make sense to ignore this
relation in order to achieve more accurate disambiguation.
Unfortunately, current approaches do not facilitate such a fine-grained control
over which parts of a given ontology should be used for disambiguation in a given
scenario and which not. Some of them allow the constraining of the concepts to
which the potential entities may belong [6] [8], but they do not do the same for
relations nor do they provide any structured process and guidelines for better
execution of this task. That is because their goal is to build scenario and domain
independent disambiguation systems where a priori knowledge about what enti-
ties and relations are expected to be present in the text is usually unavailable.
Indeed, this is the case in scenarios involving news articles, blog posts, tweets
and generally texts whose exact content cannot really be predicted. Yet there
can be also specialized scenarios where such predictions can be safely made.
One such scenario is the one above about football match descriptions. This
was in the context of the project BuscaMedia1 and involved the disambiguation
of football related entities within texts describing highlights of football matches.
1
http://www.cenitbuscamedia.es/
3
The nature of these texts made safe the assumption that the entities expected
to be found in them were players, coaches and teams and that the relations
implied between them were the ones of current membership (i.e. players and
coaches related to their current team). A similarly specialized scenario was in
the project GLOCAL2 , involving the disambiguation of location entities within
historical texts describing military conflicts. Again, the nature of these texts
allowed us to expect to find in them, among others, military conflicts, locations
where these conflicts took place and people and groups that participated in them.
Given that, in this paper we define a novel ontology-based disambiguation
framework that is particularly applicable to similar to the above scenarios where
knowledge about what entities and relations are expected to be present in the
texts is available. Through a structured semi-automatic process the framework
enables i) the exploitation of this a priori knowledge for the selection of the
subset of domain semantic information that is optimal for the disambiguation
scenario at hand, ii) the use of this subset for the generation of disambigua-
tion evidence and iii) the use of this evidence for the disambiguation of entities
within the scenario’s texts. As we will show in the rest of the paper, this process
allows our system to be more effective in such constrained scenarios than other
disambiguation approaches designed to work in unconstrained ones.
The rest of the paper is as follows. Section 2 presents related work while
section 3 describes in detail our proposed framework. Section 4 presents exper-
imental results regarding the framework’s effectiveness in the two application
scenarios mentioned above. Finally, in sections 5 and 6 we make a critical dis-
cussion of our work, we summarize its key aspects and we outline the potential
directions it could take in the future.
2 Related Work
A recent ontology-based entity disambiguation approach is described in [7] where
an algorithm for entity reference resolution via Spreading Activation on RDF
Graphs is proposed. The algorithm takes as input a set of terms associated with
one or more ontology elements and uses the ontology graph and spreading acti-
vation in order to compute Steiner graphs, namely graphs that contain at least
one ontology element for each entity. These graphs are then ranked according to
some quality measures and the highest ranking graph is expected to contain the
elements that correctly correspond to the entities.
Another approach is that of [4] where the application of restricted relation-
ship graphs (RDF) and statistical NLP techniques to improve named entity
annotation in challenging Informal English domains is explored. The applied re-
strictions are i) domain ones where various entities are a priori ruled out and ii)
real world ones that can be identified using the metadata about entities as they
appear in a particular post (e.g. that an artist has released only one album, or
has a career spanning more than two decades).
2
http://glocal-project.eu/
4
In [5] Hassel et al. propose an approach based on the DBLP-ontology which
disambiguates authors occurring in mails published in the DBLP-mailing list.
They use ontology relations of length one or two, in particular the co-authorship
and the areas of interest. Also, in [12] the authors take into account the semantic
data’s structure, which is based on the relations between the resources and,
where available, the human-readable description of a resource. Based on these
characteristics, they adapt and apply two text annotation algorithms: a structure
based one (Page Rank) and a content-based one.
Several approaches utilize Wikipedia as a highly structured knowledge source
that combines annotated text information (articles) and semantic knowledge
(through the DBPedia3 [1] and YAGO [13] ontologies). For example, DBPe-
dia Spotlight [8] is a tool for automatically annotating mentions of DBPedia
resources in text by using i) a lexicon that associates multiples resources to an
ambiguous label and which is constructed from the graph of labels, redirects and
disambiguations that DBPedia ontology has and ii) a set of textual references to
DBPedia resources in the form of Wikilinks. These references are used to gather
textual contexts for the candidate entities from wikipedia articles and use them
as disambiguation evidence.
A similar approach that uses the YAGO ontology is the AIDA system [6]
which combines three entity disambiguation measures: the prior probability of
an entity being mentioned, the similarity between the contexts of a mention and
a candidate entity, and the semantic coherence among candidate entities for all
mentions together. The latter is calculated based on the distance between two
entities in terms of type and subclassOf edges as well as the number of incoming
links that their Wikipedia articles share.
The difference between the above approaches and our framework is detected
in the way they treat the available semantic data. For example, Spotlight uses
the DBPedia ontology only as an entity lexicon without really utilizing any
of its relations, apart from the redirect and disambiguation ones. Thus, it’s
more text-based than ontology-based. On the other hand, AIDA builds an entity
relation graph by considering only the type and subclassOf relations as well
as “assumed” relations inferred by the links within the articles. The problem
with this approach is that important semantic relations that are available in the
ontology are not utilized and, of course, there is no control over which edges of
the derived ontology graph should be utilized in the given scenario. Such control
is not provided either in [7] or any of the rest aforementioned approaches except
for that of [5] which, however, is specific for the scientific publications domain.
3 Proposed Disambiguation Framework
Our framework targets the task of entity disambiguation based on the intuition
that a given ontological entity is more likely to represent the meaning of an
ambiguous term when there are many ontologically related to it entities in the
3
http://dbpedia.org
5
text. These related entities can be seen as evidence whose quantitative and
qualitative characteristics can be used to determine the most probable meaning
of the term. For example, consider a historical text containing the term “Tripoli”.
If this term is collocated with terms like “Siege of Tripolitsa” and “Theodoros
Kolokotronis” (the commander of the Greeks in this siege) then it is fair to
assume that this term refers to the city of Tripoli in Greece rather than the
capital of Libya.
Nevertheless, as we already showed in the introduction, which entities and to
what extent should serve as evidence in a given scenario depends on the domain
and expected content of the texts that are to be analyzed. For that, the key
ability our framework provides to its users is to construct, in a semi-automatic
manner, semantic evidence models for specific disambiguation scenarios and use
them to perform entity disambiguation within them. In particular, our frame-
work comprises the following components:
– A Disambiguation Evidence Model that contains, for a given scenario,
the entities that may serve as disambiguation evidence for the scenario’s
target entities (i.e. entities we want to disambiguate). Each pair of a target
entity and an evidential one is accompanied by a degree that quantifies the
latter’s evidential power for the given target entity.
– A Disambiguation Evidence Model Construction Process that builds,
in a semi-automatic manner, a disambiguation evidence model for a given
scenario.
– An Entity Disambiguation Process that uses the evidence model to de-
tect and extract from a given text terms that refer to the scenario’s target
entities. Each term is linked to one or more possible entity uris along with
a confidence score calculated for each of them. The entity with the highest
confidence should be the one the term actually refers to.
In the following paragraphs we elaborate on each of the above components.
3.1 Disambiguation Evidence Model and its Construction
For the purposes of this paper we define an ontology as a tuple O = {C, R, I, iC , iR }
where
– C is a set of concepts.
– I is a set of instances.
– R is a set of binary relations that may link pairs of concept instances.
– iC is a concept instantiation function C → I.
– iR is a relation instantiation function R → I × I.
The Disambiguation Evidence Model defines for each ontology instance
which other instances and to what extent should be used as evidence towards
its correct meaning interpretation. More formally, given a domain ontology O, a
disambiguation evidence model is defined as a function dem : I × I → [0, 1]. If
6
i1 , i2 ∈ I then dem(i1 , i2 ) is the degree to which the existence, within the text,
of i2 should be considered an indication that i1 is the correct meaning of any
text term that has i1 within its possible interpretations.
To construct the optimal evidence model for a given disambiguation scenario
we proceed as follows: First, based on the scenario, we determine the concepts
the instances of which we wish to disambiguate (e.g. players, teams and man-
agers for the football match scenario). Then, for each of these concepts, we
determine the related to them concepts whose instances may serve as contextual
disambiguation evidence. The result of the above analysis should be a disam-
biguation evidence concept mapping function evC : C → C × Rn which given
a target concept ct ∈ C returns the concepts which may act as evidence for
it along with the ontological relations whose composition links this concept to
the target one. Table 1 contains an example of such a function for the football
match descriptions scenario where, for instance, soccer players provide evidence
for other soccer players that play in the same team. This mapping, shown in the
second row of the table, is facilitated by the composition of the relations dbp-
prop:currentclub (that relates players to their current teams) and its inverse
one is dbpprop:currentclub of (that relates teams to their current players).
Table 2 illustrates a similar mapping for the military conflict texts scenario.
Table 1. Sample Disambiguation Evidence Concept Mapping for Football Match De-
scriptions
Target Concept Evidence Concept Relation(s) linking Evi-
dence to Target
dbpedia-owl:SoccerPlayer dbpedia-owl:SoccerClub is dbpprop:currentclub of
dbpedia-owl:SoccerPlayer dbpedia-owl:SoccerPlayer dbpprop:currentclub, is dbp-
prop:currentclub of
dbpedia-owl:SoccerClub dbpedia-owl:SoccerPlayer dbpprop:currentclub
dbpedia-owl:SoccerClub dbpedia- dbpedia-owl:managerClub
owl:SoccerManager
dbpedia- dbpedia-owl:SoccerClub is dbpedia-owl:managerClub of
owl:SoccerManager
Using the disambiguation evidence concept mapping, we can then automat-
ically derive the disambiguation evidence model dem as follows: Given a tar-
get concept ct ∈ C and an evidence concept ce ∈ C then for each instance
it ∈ iC (ct ) and ie ∈ iC (ce ) that are related to each other through the com-
position of relations {r1 , r2 , ..., rn } ∈ evC (ct ) we derive the set of instances
It ⊆ I which share common names with it and are also related to ie through
{r1 , r2 , ..., rn } ∈ evC (ct ). Then the value of dem for this pair of instances is
computed as follows:
1
dem(it , ie ) = (1)
|It |
7
Table 2. Sample Disambiguation Evidence Concept Mapping for Military Conflict
Texts
Target Concept Evidence Concept Relation(s) linking Evi-
dence to Target
dbpedia- dbpedia- dbpprop:place
owl:PopulatedPlace owl:MilitaryConflict
dbpedia- dbpedia- dbpprop:place, dbpedia-
owl:PopulatedPlace owl:MilitaryConflict owl:isPartOf
dbpedia- dbpedia- is dbpprop:commander of, dbp-
owl:PopulatedPlace owl:MilitaryPerson prop:place
dbpedia- dbpedia- dbpedia-owl:isPartOf
owl:PopulatedPlace owl:PopulatedPlace
dbpedia- dbpedia- dbpprop:commander
owl:MilitaryPerson owl:MilitaryConflict
The intuition behind this formula is that the evidential power of a given entity
is inversely proportional to the number of different target entities it provides
evidence for. If, for example, a given military person has fought in many different
locations with the same name, then its evidential power for this name is low.
3.2 Entity Disambiguation Process
The entity reference resolution process for a given text document and a disam-
biguation evidence model starts by extracting from the text the set of terms
T that match to some instance belonging to a target or an evidence concept,
that is some i ∈ iC (c), c ∈ Ct ∪ Ce . Along with that we derive a term-meaning
mapping function m : T → I that returns for a given term t ∈ T the instances
it may refer to. We also consider Itext to be the superset of these instances.
t
Then we consider the set of potential target instances found within the Itext ⊆
t
Itext and for each it ∈ Itext we derive all the instances ie from Itext for which
dem(it , ie ) > 0. Subsequently, by combining the evidence model dem with the
term meaning function m we are able to derive an entity-term support function
t t
sup : Itext × T → [0, 1] that returns for a target entity it ∈ Itext and a term
t ∈ T the degree to which t supports it :
1 !
sup(it , t) = dem(it , ie ) (2)
|m(t)|
ie ∈m(t)
Using this function we are able to calculate for a given term in the text the
confidence that it refers to the entity it ∈ m(t) as follows:
"
!
t∈T K(it , t)
conf (it ) = " " " ∗ sup(it , t) (3)
i! ∈m(t)
t t∈T K(it , t) t∈T
where K(it , t) = 1 if sup(it , t) > 0 and 0 otherwise. In other words, the overall
support score for a given candidate target entity is equal to the sum of the
8
entity’s partial supports (i.e. function sup) weighted by the relative number of
terms that support it. It should be noted that in the above process we adopt the
one referent per discourse approach which assumes one and only one meaning
for a term in a discourse.
4 Framework Application and Evaluation
To evaluate the effectiveness of our framework we applied it in the two scenarios
we mentioned in the introduction, the one involving disambiguation in football
match descriptions and the other in texts describing military conflicts. In both
cases we used DBPedia as a source of semantic information and we i) defined
a disambiguation evidence model for each scenario and ii) used these models to
perform entity disambiguation in a representative set of texts. Then we measured
the precision and recall of the process. Precision was determined by the fraction
of correctly interpreted terms (i.e. terms for which the interpretation with the
highest confidence was the correct one) to the total number of interpreted terms
(i.e. terms with at least one interpretation). Recall was determined by the frac-
tion of correctly interpreted terms to the total number of annotated terms in
the input texts. It should be noted that all target terms for disambiguation in
the input texts were known to the knowledge base (i.e. DBPedia).
Finally, the results of the above evaluation process were compared to those
achieved by two publicly available semantic annotation and disambiguation sys-
tems, namely DBPedia Spotlight 4 [8], AIDA5 [6]. The two systems were chosen
for comparison because i) they also use DBPedia as a knowledge source and ii)
they provide some basic mechanisms for constraining the types of entities to be
disambiguated, though not in the same methodical way as our framework does.
Practically, the two systems merely provide the users the capability to select the
classes whose instances are to be included in the process. In all cases, it should be
made clear that the goal of this comparison was not to disprove the effectiveness
and value of these systems as tools for open domain and unconstrained situa-
tions but rather to verify our claim that our approach is more appropriate for
disambiguation in “controlled” scenarios, i.e. scenarios in which a priori knowl-
edge about what entities and relations are expected to be present in the text is
available. A useful evaluation of popular semantic entity recognition systems for
open scenarios may be found at [11].
4.1 Football Match Descriptions Scenario
In this scenario we had to semantically annotate a set of textual descriptions of
football match highlights like the following: “It’s the 70th minute of the game
and after a magnificent pass by Pedro, Messi managed to beat Claudio Bravo.
Barcelona now leads 1-0 against Real.”. These descriptions were used as meta-
data of videos showing these highlights and our goal was to determine, in an
4
http://dbpedia-spotlight.github.com/demo/index.html
5
https://d5gate.ag5.mpi-sb.mpg.de/webaida/
9
unambiguous way, which were the participants (players, coaches and teams) in
each video. The annotated descriptions were then to be used as part of a se-
mantic search application where users could retrieve videos that showed their
favorite player or team, with much higher accuracy.
To achieve this goal, we applied our framework and we built a disambiguation
evidence model, based on DBPedia, that had as an evidence mapping function
that of table 1. This function was subsequently used to automatically calculate
(through equation 1) the function dem for all pairs of target and evidence en-
tities. Table 3 shows a small sample of these pairs where, for example, Getafe
acts as evidence for the disambiguation of Pedro Leon because the latter is a
current player of it. Its evidential power, however, for that player is 0.5, since
in the same team there is another player with the same name (i.e. Pedro Rios
Maestre).
Table 3. Examples of Target-Evidential Entity Pairs for the Football Scenario
Target Entity Evidential Entity dem
dbpedia:Real Sociedad dbpedia:Claudio Bravo (footballer) 1.0
dbpedia:Pedro Rodriguez Ledesma dbpedia:FC Barcelona 1.0
dbpedia:Pedro Leon dbpedia:Getafe CF 0.5
dbpedia:Pedro Rios Maestre dbpedia:Getafe CF 0.5
dbpedia:Lionel Messi dbpedia:FC Barcelona 1.0
Using this model, we applied our disambiguation process in 50 of the above
texts, all containing ambiguous entity references. The overall number of ref-
erences was 126 with about 90% of them being ambiguous. In average, each
ambiguous entity reference had 3 possible interpretations with player names be-
ing the most ambiguous. Table 4 shows the results achieved by our approach as
well as by DBPedia Spotlight and AIDA. It should be noted that when using
the latter systems, we used their concept selection facilities in order to constrain
the space of possible interpretations. Still, as one can see from the table data,
the constraining of the semantic data that our custom disambiguation evidence
model facilitated (e.g. the consideration of only the current membership relation
between players and teams) was more effective and managed to yield significantly
better results.
Table 4. Entity Disambiguation Evaluation Results in the Football Scenario
System/Approach Precision Recall F1 Measure
Proposed Approach 84% 81% 82%
AIDA 62% 56% 59%
DBPedia Spotlight 85% 26% 40%
10
4.2 Military Conflict Texts Scenario
In this scenario our task was to disambiguate location references within a set of
textual descriptions of military conflicts like the following: “The Siege of Augusta
was a significant battle of the American Revolution. Fought for control of Fort
Cornwallis, a British fort near Augusta, the battle was a major victory for the
Patriot forces of Lighthorse Harry Lee and a stunning reverse to the British and
Loyalist forces in the South”. For that we used again DBPedia and we defined the
disambiguation evidence mapping function of table 2 which, in turn, produced
the evidence model that is (partially) depicted in table 5.
Table 5. Examples of Target-Evidential Entity Pairs for the Miltary Conflict Scenario
Location Evidential Entity dem
dbpedia:Columbus, Georgia James H. Wilson 1.0
dbpedia:Columbus, New Mexico dbpedia:Pancho Villa 1.0
dbpedia:Beaufort County, South Carolina dbpedia:Raid at Combahee Ferry 1.0
dbpedia:Beaufort County, South Carolina dbpedia:James Montgomery (colonel) 1.0
dbpedia:Beaufort County, North Carolina dbpedia:Battle Of Washington 1.0
dbpedia:Beaufort County, North Carolina dbpedia:John G. Foster 1.0
Using this model, we applied, as in the football scenario, our disambiguation
process in a set of 50 military conflict texts, targeting the locations mentioned in
them. The average reference ambiguity of this set was 5 in a total of 55 locations.
Table 6 shows the achieved results which verify the ability of our framework to
improve disambiguation effectiveness.
Table 6. Entity Disambiguation Evaluation Results in the Military Conflict Scenario
System/Approach Precision Recall F1 Measure
Proposed Approach 88% 83% 85%
DBPedia Spotlight 71% 69% 70%
AIDA 44% 40% 42%
5 Discussion
It should have been made clear from the previous sections that our framework is
not independent of the content or domain of the input texts but rather adaptable
to them. That’s exactly its main differentiating feature as our purpose was not
to build another generic disambiguation system but rather a reusable framework
that can i) be relatively easily adapted to the particular characteristics of the
domain and application scenario at hand and ii) exploit these characteristics in
11
order to increase the effectiveness of the disambiguation process. Our motivation
for that was that, as the comparative evaluation of the previous section showed,
the scenario adaptation capabilities of existing generic disambiguation systems
can be inadequate in certain scenarios (like the ones described in this paper),
thus limiting their applicability and effectiveness.
Of course, the usability and effectiveness of our approach is directly pro-
portional to the content specificity of the texts to be disambiguated and the
availability of a priori knowledge about their content. The greater these two
parameters are, the more applicable is our approach and the more effective the
disambiguation is expected to be. The opposite is true as the texts become more
generic and the information we have out about them more scarce. A method that
could a priori assess how suitable is our framework for a given scenario would be
useful, but it falls outside the scope of this paper. Also, the framework’s approach
is not completely automatic as it requires some knowledge engineer or domain
expert to manually define the scenario’s disambiguation evidence mapping func-
tion. Nevertheless, this function is defined at the schema level thus making the
number of required mappings for most scenarios rather small and manageable.
Finally, although we haven’t formally evaluated the scalability of our ap-
proach, the fact that our framework is based on the constraining of the semantic
data to be used makes us expect that it will perform faster than traditional ap-
proaches that use the whole amount of data. Furthermore, as the disambiguation
evidence model may be constructed offline and stored in some index, the most
probable bottleneck of the process will be the phase of determining the candidate
entities for the extracted terms rather than the resolution process. Nevertheless,
a more rigorous scalability study will have to be made as part of future work.
6 Conclusions and Future Work
In this paper we proposed a novel framework for optimizing named entity dis-
ambiguation in well-defined and adequately constrained scenarios through the
customized selection and exploitation of semantic data. First we described how,
given a priori knowledge about the domain(s) and expected content of the texts
that are to be analyzed, one can use the semantic data and define an evidence
model that determines which and to what extent semantic entities should be
used as contextual evidence for the disambiguation task at hand. Then we de-
scribed the process through which such a model can be actually used for this
task. The overall framework was experimentally evaluated in two specific sce-
narios and the results verified its superiority over existing approaches that are
designed to work in open domains and unconstrained scenarios.
Future work will focus on the further automation of the disambiguation evi-
dence model construction by means of data mining and machine learning tech-
niques. Moreover, an online tool to enable users to dynamically build such models
out of existing semantic data and use them for disambiguation purposes, will be
developed.
12
Acknowledgements
This work was supported by the Spanish project CENIT-2009-1026 BuscaMedia
and by the European Commission under contract FP7- 248984 GLOCAL.
References
1. Auer, S., Bizer, C., Kobilarov, G., Lehmann J., Cyganiak, R., Ives, Z.G.: DBpedia:
A Nucleus for a Web of Open Data. In Proceedings of the 6th International Semantic
Web Conference, pages 722-735, 2007.
2. Fader, A., Soderland, S., Etzioni, O.: Scaling wikipedia-based named entity disam-
biguation to arbitrary web text. In Proceedings of the WikiAI 09 - IJCAI Work-
shop: User Contributed Knowledge and Artificial Intelligence: An Evolving Synergy,
Pasadena, CA, USA, July 2009.
3. Ferragina, P., Scaiella, U.: TAGME: on-the-fly annotation of short text fragments
(by wikipedia entities). In Proceedings of the 19th ACM international conference
on Information and knowledge management, ACM, New York, SA, 1625-1628.
4. Gruhl, D., Nagarajan, M., Pieper, J., Robson, C., Sheth A.P.: Context and domain
knowledge enhanced entity spotting in informal text. In Proceedings of the 8th
International Semantic Web Conference, pages 260-276, 2009.
5. Hassell, J., Aleman-Meza, B., Arpinar, I.: Ontology-driven automatic entity disam-
biguation in unstructured text. In Proceedings of the 3rd European Semantic Web
Conference, pages 44-57, Springer Berlin, Heidelberg, 2006.
6. Hoffart, J., Yosef, M.A., Bordino, I., Frstenau, H, Pinkal, M., Spaniol, M., Taneva,
B., Thater, S., Weikum, G.: Robust disambiguation of named entities in text. In
Proceedings of the Conference on Empirical Methods in Natural Language Process-
ing, Association for Computational Linguistics, Stroudsburg, PA, USA, 782-792.
7. Kleb, J., Abecker, A.: Entity Reference Resolution via Spreading Activation on
RDF-Graphs. In Proceedings of the 7th European Semantic Web Conference, pages
152-166, Springer Berlin, Heidelberg, 2006.
8. Mendes, P.N., Jakob, M., Garcia-Silva, A., Bizer, C.: DBpedia spotlight: shedding
light on the web of documents. In Proceedings of the 7th International Conference
on Semantic Systems, ACM, New York, USA, 1-8, 2011.
9. Miller, G., Charles, W.: Contextual correlates of semantic similarity. Language and
Cognitive Processes, 6(1):128.
10. Pilz, A., Paass, G.: Named entity resolution using automatically extracted seman-
tic information. Workshop on Knowledge Discovery, Data Mining, and Machine
Learning, page 84-91, 2009.
11. Rizzo G., Troncy, R.: NERD: A Framework for Evaluating Named Entity Recog-
nition Tools in the Web of Data. In 10th International Semantic Web Conference,
Demo Session, pages 1-4, Bonn, Germany, 2011.
12. Rusu, D., Fortuna, B., Mladenic, D.: Automatically Annotating Text with Linked
Open Data. In 4th Linked Data on the Web Workshop (LDOW 2011), 20th World
Wide Web Conference, Hyderabad, India, 2011.
13. Suchanek, F.M., Kasneci, G., Weikum, G.: YAGO: A Core of Semantic Knowledge.
In 16th World Wide Web Conference, 2007.