=Paper=
{{Paper
|id=None
|storemode=property
|title=SLINT: a schema-independent linked data interlinking system
|pdfUrl=https://ceur-ws.org/Vol-946/om2012_Tpaper1.pdf
|volume=Vol-946
|dblpUrl=https://dblp.org/rec/conf/semweb/NguyenIL12
}}
==SLINT: a schema-independent linked data interlinking system==
https://ceur-ws.org/Vol-946/om2012_Tpaper1.pdf
SLINT: A Schema-Independent Linked Data
Interlinking System
Khai Nguyen1 , Ryutaro Ichise2 , and Bac Le1
1
University of Science, Ho Chi Minh, Vietnam
{nhkhai,lhbac}@fit.hcmus.edu.vn
2
National Institute of Informatics, Tokyo, Japan
ichise@nii.ac.jp
Abstract. Linked data interlinking is the discovery of all instances that
represent the same real-world object and locate in different data sources.
Since different data publishers frequently use different schemas for stor-
ing resources, we aim at developing a schema-independent interlinking
system. Our system automatically selects important predicates and use-
ful predicate alignments, which are used as the key for blocking and in-
stance matching. The key distinction of our system is the use of weighted
co-occurrence and adaptive filtering in blocking and instance matching.
Experimental results show that the system highly improves the precision
and recall over some recent ones. The performance of the system and the
efficiency of main steps are also discussed.
Keywords: linked data, schema-independent, blocking, interlinking.
1 Introduction
Years of effort in building linked data has brought a huge amount of data in the
LOD. However, maximizing the efficiency of linked data in the development of
semantic web is still facing many difficulties. One of the current challenges is to
integrate the individual data sources for building a common knowledge system.
When different data source may contain heterogeneous instances, which co-refer
to the same real-world objects, data integration process requires the detection of
such objects to ensure the integrity and consistency of data. Detecting all iden-
tities between data sources is the mission of data interlinking. Data interlinking
consist of two main steps, blocking and instance matching. While blocking aims
at pruning the number of comparison, instance matching is to determine the
matching status of two interested instances.
Current interlinking methods can be categorized into two main groups: schema-
dependent [2,7,10] and schema-independent [1,3,4,9]. The former requires the
knowledge about meaning of RDF predicates (e.g. predicate #preLabel declares
the label of object) and the predicate alignments (e.g. predicate #preLabel
matched with predicate #name). In contrast, the latter does not need these
information, therefore it does not rely on human knowledge about the schema.
Because a linked data instance is a set of many RDF triples (subject, predicate,
object), the schema of a data source refers to the list of all used predicates,
which are closely related to vocabulary and ontology. The schemas are usually
different for each data sources, even in the same data source but different do-
mains. Clearly, schema-independent methods are more applicable when it can
2 K. Nguyen, R. Ichise, B. Le
work on every kind of source or domain without any human’s instruction. Be-
sides, manual specifications of interlinking rules frequently ignore the hidden
useful predicate alignments.
We present SLINT system, which use a new approach for schema-independent
linked data interlinking. SLINT automatically selects important RDF predicates
using the coverage and discriminability. The selected predicates are combined to
construct the predicate alignments in conciliation of data type. We estimate the
confidence of predicate alignments to collect the most appropriate alignments
for blocking and interlinking. By this way, the collective information of instance
is frequently leveraged. Blocking is therefore more complete, compact, and sup-
portive for interlinking. Also, we apply adaptive filtering techniques for blocking
and instance matching. In experiment, we compare SLINT with three systems,
which participated OAEI 2011 instance matching campaign, and report the high
improvement on both precision and recall. Experiments on the performance of
SLINT and the efficiency of blocking step are also reported.
The paper is organized as follow: the next section is the overview of previous
work. Section 3 describes the detail of SLINT system. Section 4 reports our
experimental evaluation. Section 5 closes the paper with conclusion and outlook.
2 Related work
Data interlinking is an early studied area, however, this problem in linked data
has just been recently attended. Silk [10], a well-known framework, provides a
declarative interface for user to define the predicate alignments as well as the sim-
ilarity metrics for matching. Silk was used as a main component of the LDIF [8],
a multiple linked data sources integration framework. Recently, Isele and Bizer
have improved their Silk by applying an automatic linkage rules generation using
genetic algorithm [3]. The work is very interesting at the modeling of appropri-
ate fitness function and specific transformations for genetic programming in the
context of interlinking. This work makes Silk to be schema-independent. With
the similar objective, RAVEN [4] minimize human curation effort using active
learning, while Nikolov et al. also use genetic algorithm with the research tar-
get is an unsupervised learning process [6]. Also with schema-independent goal,
Nguyen et al. suggest using decision tree classifier for determining the matching
status of two instances [5].
Zhishi.Links [7] is one of the current state-of-the-art matchers. This system
adopt the idea of Silk’s pre-matching step, by using label of objects such as
skos:prefLabel or scheme:label, to group similar instances. Afterward, a more
complex semantic similarity is utilized for matching. This system ranks first at
the OAEI 2011 instance matching, while the second best is SERIMI [1], a schema-
independent system. SERIMI selects RDF predicates and predicate alignments
using entropy and RDF object similarity, respectively. AgreementMaker [2] is an
ontology matching and instance matching system. It firstly generates candidates
set by comparing the labels of instances. These candidates are then divided into
smaller subsets, in which every pair is matched to produce the final alignments.
Most of previous interlinking systems do not deeply investigate on blocking
step, which generates potential identity pairs of instances. Song and Heffin focus
SLINT: A Schema-Independent Linked Data Interlinking System 3
Fig. 1. Summary of data interlinking process
on blocking scheme for linked data interlinking for parallel independent work [9].
It is a very interesting idea when the authors propose an unsupervised learning
for maximizing the usefulness of blocking keys, which are the combinations of
RDF predicates. The authors conduct experiments on some large datasets, which
also proof for the scalability.
In general, the schema-dependent approaches compare two instances by spec-
ified properties. That is, they can detect almost right identity pairs but the preci-
sion may be low on highly ambiguous data sources. The reason is that some useful
information can be ignored since the manual predicate alignment frequently is
not an optimal solution. In contrast, schema-independent approaches reconcile
precision and recall because of elaborate analysis on the data. Although these
approaches need to collect predicate alignments, the matching is more effective
when collective information is frequently used. Comparing SLINT with previous
interlinking systems, the prominent differences are the predicate selection, pred-
icate alignment, and adaptive filtering for blocking and interlinking. In the next
section, we describe these elements as the details of SLINT.
3 Schema-independent linked data interlinking system
This section describes the SLINT system. The overview of the interlinking pro-
cess for source data Ds and target data Dt is shown in Figure 1. In this figure,
the small circles and triangles respectively stand for instances and theirs RDF
predicates. The referred circles of each step are the output of that step. The
SLINT system consists of four steps. The interlinking process begins with pred-
icate selection, which collects the important predicates from all predicates of
each data sources. In the second step, predicate alignment, selected predicates
are combined in accordance with their data type to construct the raw predicate
alignments. We estimate the confidence of every raw alignment to measure its
appropriateness. A raw alignment will be called a key alignment if its confidence
4 K. Nguyen, R. Ichise, B. Le
satisfies a filtering condition. These key alignments provide much useful informa-
tion in blocking and instance matching steps. Next, blocking step is designed to
reduce the number of comparison by producing identity candidates of instances.
The instance matching afterward only need to verifies the retrieved candidates
for discovering the identity pairs. The followings are the detail of each step.
3.1 Predicate selection
The mission of this step is to find the important predicates from the schema,
which consists of all predicates appearing in interested data source. We use two
criteria for determining the importance level of predicate p: coverage cov(p, D)
and discriminability dis(p, D). Eq.1 and Eq. 2 are the explanations of these
criteria when considering predicate p of data source D.
|{x|∃ < s, p, o >∈ x, x ∈ D}|
cov(p, D) = . (1)
|D|
|{o|∃x ∈ D, < s, p, o >∈ x}|
dis(p, D) = . (2)
|{< s, p, o > |∃x ∈ D, < s, p, o >∈ x}|
In these equation, x represents an instance and is a set of RDF triple <
s, p, o > (subject, predicate, object). D is the interested data source and is a
set of instances. From each input source, we collect the predicates having high
score of coverage and discriminability. A predicate p is selected if it satisfies the
condition in Eq.3, which inherits from the idea of [9].
(cov(p, D) ≥ α) ∧ (dis(p, D) ≥ β) ∧ (HM ean(cov(p, D), dis(p, D)) ≥ γ). (3)
The α and β imply the minimum standard of an important predicate, whereas
γ, the condition for harmonic mean of dis(p, D) and cov(p, D), is the main
requirement. Therefore, we set small values for α and β and larger value for γ.
Song and Heffin focus on learning blocking key by iteratively maximize the
coverage and discriminability of the set of predicates [9]. In our system, we use
the same discriminability function with theirs and slightly different function for
coverage. For the numerator of Eq. 1, while they use the number of RDF subjects,
we use the number of instances, because we aim at finding the frequency of
predicate over instances, not over RDF subjects.
Important predicates are expected to be used for declaring the common prop-
erties and distinct information of objects. Since coverage and discriminability
respectively express the former and latter, the combination of them is therefore
appropriate for the objective of predicate selection. If a predicate has a high
coverage but a low discriminability or otherwise, it will not be important. An
example for this kind of predicate is rdf:type. This predicate is frequently used
but it usually describes a limit range of various RDF objects when observing the
instances in the same domain.
3.2 Predicate alignment
In this step, we find the appropriate alignments of predicates between the source
data and target data. An alignment of two predicates is considered to be appro-
priate if the interested predicates describe the similar properties of instances.
SLINT: A Schema-Independent Linked Data Interlinking System 5
From selected predicates of source data and target data, we connect every type-
matched pair and select the alignments whose confidence is higher than threshold
δ. Selected predicate alignments are called key alignments. The confidence of an
alignment is the Dice coefficient between the representatives of RDF objects de-
scribed by its formed predicates. Eq. 4 is the equation of confidence conf (ps , pt )
for the alignment between predicate ps in source data Ds and predicate pt in
target data Dt .
2 × |R(Os ) ∩ R(Ot )|
conf (ps , pt ) = , Ok = {o|∃x ∈ Dk , < s, pk , o >∈ x}. (4)
|R(Os )| + |R(Ot )|
In above equation, R is the function that returns the representative elements
of RDF objects. The return values of R depend on the type of predicates. We
divide the predicates into five different types: string, URI, decimal, integer, and
date. This separation is based on the variety of data types in the real world and
covers most of current types of linked data. For string, we extract the word token
of RDF objects. For URI, we omit the domain part and use the same manner as
for string, with the assumption that slash ‘/’ is the token separator. For decimal,
we take the 2-decimal digits rounded values. For integer and date, we do not
transform the RDF objects and use the original values. For determining type of
a predicate, we use the major type of RDF objects declared by this predicate.
For example, if 51% appearance times of p is to describe decimal values, the
data type of p will be decimal. Currently, we detect the type of RDF objects
without the consideration about the difference in their metric (e.g. the units of
time, distance, area).
The confidence of an alignment represents the similarity of RDF objects be-
tween two data sources. The predicates having the same meaning frequently
describe the similar information. Therefore, alignments of matched predicates
usually have higher confidence than the others. It means that a predicate satis-
fying the requirements of an important predicate is verified again, by considering
the confidence of all alignments in which it appears. For example, the predicate
rdfs:comment has possibility to be important but the confidences of its align-
ments are usually low because the denominator of Eq.3 is very high in this case.
A common limiting point of almost previous systems is the use of string mea-
surement for every type of RDF objects. Clearly, this approach is not sufficient to
cover the meaning of RDF objects, thus, does not well estimate the similarity of
non-string values. We discriminate data types not only in combining predicates,
but also in blocking and instance matching.
It is not easy for a person to detect all useful predicate alignments, this step
is therefore very meaningful, and in accompaniment with predicate selection, it
tackles the schema-independent goals. The next steps are the use of selected key
alignments and their formed predicates.
3.3 Blocking
As we introduced, the blocking aims at retrieving the candidates for instance
matching step by grouping similar instances into the same block. A candidate
is a pair of two instances, one belongs to source data and one belongs to target
data. The blocking can be divided into three sub phases. The first phase indexes
6 K. Nguyen, R. Ichise, B. Le
Algorithm 1: Generating candidates set
Input: Ds , Dt , P rs , P rt , ζ,
Output: Candidate set C
1 H ←∅
2 M [|Ds |, |Dt |] ← {0}
3 C ←∅
4 foreach < D, P >∈ {< Ds , P rs >, < Dt , P rt >} do
5 foreach x ∈ D do
6 foreach pi ∈ P do P
7 sumConf ← pj ∈{P rs ,P rt }\P conf (pi , pj )
8 foreach r ∈ Rp(O), O = {o| < s, pi , o >∈ x} do
9 if not H.ContainsKey(r.Label) then
10 H.AddKey(r.Label)
11 H.AddValue(r.Label, D, < x, r.V alue × sumConf >)
12 foreach key ∈ H.AllKeys() do
13 foreach < xs , vs >∈ H.GetV alues(key, Ds ) do
14 foreach < xt , vt >∈ H.GetV alues(key, Dt ) do
15 M [xs , xt ] ← M [xs , xt ] + vs × vt
16 foreach xs ∈ Ds do
17 foreach xt ∈ Dt do
18 maxs ← Max(M [xs , xj ]), ∀xj ∈ Dt
19 maxt ← Max(M [xi , xt ]), ∀xi ∈ Ds
20 max ← HMean(maxs , maxt )
21 if M [xs , xt ] ≥ ζ and Mmax
[xs ,xt ]
≥ then
22 C ← C∪ < xs , xt >
23 return C
every instance in each data source by the value extracted from its RDF objects.
The second phase traverses the index table and builds a weighted co-occurrence
matrix. The final phase uses this matrix as the input information when it applies
a filtering technique to select candidates. Algorithm 1 is the pseudo-code of the
whole blocking process. In this algorithm, P rs and P rt represent the list of
predicates that form the key alignments, where P rk belongs to Dk . H, M , C,
Rp represent the inverted-index table, weighted co-occurrence matrix, candidates
set, and representative extraction method, respectively.
The lines 4-11 perform the invert-indexing, a well-known indexing technique.
By once traversing each data source, we extract the representatives of RDF
objects and use them as the keys of invert-index table. An element r in the
representatives set of RDF objects consists of two fields: the label r.Label and
value r.V alue. While r.Label is the return value of representative extraction
SLINT: A Schema-Independent Linked Data Interlinking System 7
method R as in predicate alignment step, r.V alue is computed in accordance
with the data type of predicate pi . If pi is string or URI, we set the value to
TF-IDF score of the token. If pi is either decimal, integer, or date, we assign the
value to a fixed number, which is 1.
After constructing the invert-index table, we compute weighted co-occurrence
matrix M as the lines 12-15, by accumulating the value for each matrix element.
The lines 16-22 are the process of adaptive filtering. An instance pair <
xs , xt > will be considered as a candidate if its weighted co-occurrence value
M [xs , xt ] satisfies the threshold , after divided for the harmonic mean of maxi-
mum weighted co-occurrences of xs and xt . In addition, we use ζ, a small thresh-
old, to avoid the surjection assumption. The identities frequently have the high
weighted co-occurrences; however, these values are variable for different pairs.
Choosing a fixed threshold for selecting candidates is not good in this situation
and is a tedious task. Therefore, we use the coefficient of M [xs , xt ] and max,
which is a data driven element and expresses the adaptive filtering idea.
Blocking is very important because it reduces the number of comparison in
instance matching. However, it seems not to have been sufficiently attended when
most of previous systems use quite simple method for blocking. In comparison
with blocking step in previous interlinking systems, the key difference of our
method is the weighted co-occurrence matrix and the adaptive filtering. While
previous systems compare the pairs of RDF objects, we aggregate the product
of the weight of their matched representatives. For candidate selection, Silk [10]
and Zhishi.Links [7] use top-k strategy, which selects k candidates for each in-
stance. The approach is very good for controlling the number of candidates, but
determining the value of k is not easy. Song and Heffin use thresholding selection
[9], which is also similar with SERIMI [1]. Our method also use thresholding ap-
proach as the availability of ζ. However, the key idea of our selection method
is the adaptive filtering because the impact of ζ is not high. Frequently, there
are many of non-identity pairs between two data sources, ζ is therefore usually
configured with a low value.
Next, we input the set of candidates C and the key alignments A into the
final step, instance matching.
3.4 Instance matching
The instance matching verifies the selected candidates to determine their identity
state. We compute the matching score for every candidate and choose the ones
that have high score as the identity pairs. For each element in A, we compute the
similarity of RDF objects, which declared by the involved predicates of interested
key alignment. The final score of two instances is the weighted average value of
all these similarities, and the weights are the confidences of key alignments. Eq.5
is the computation of matching score between instance xs ∈ Ds and xt ∈ Dt .
1 X
score(xs , xt ) = conf (ps , pt ) × sim(R(Os ), R(Ot )),
W
∈A
W here Ok = {o|∃x ∈ Dk , < s, pk , o >∈ x} (5)
X
W = conf (ps , pt ).
∈A
8 K. Nguyen, R. Ichise, B. Le
In this equation, R stands for the representative extraction methods, which are
similar to those in predicate alignment step.
Categorizing five data types, we implement three different versions of sim
function in accordance with the type of predicates. For decimal and integer, we
take the variance of the values to remedy the slight difference of data representa-
tions. For date, the sim function yields 1 or 0 when the values are equal or not,
respectively. A date is usually important if it is a property of the object (e.g.
birthday, decease date, release date of a movie). Therefore, the exact matching is
an appropriate selection for dates comparison. For string and URI, we compute
the TF-IDF modified cosine similarity, as given in Eq.6. TF-IDF is used because
its advantage in disambiguating the instances sharing common tokens. TF-IDF
also minimizes the weight for the stop-words, which are usually useless.
P
q∈Qs ∩Qt T F IDF (q, Qs )T F IDF (q, Qt )
sim(Qs , Qt ) = qP P . (6)
2 2
q∈Qs T F IDF (q, Qs ) × q∈Qt T F IDF (q, Qt )
Similar with blocking step, we do not use fixed single threshold for filtering the
candidates. Two instances will be considered as an identity pair if their score is
higher than the maximum score of the candidates in which either of instances
appears. The final identities set I is formalized in Eq.7.
I = {< xs , xt > |score(xs , xt ) ≥ η∧
score(xs , xt ) (7)
≥ θ}.
max∀∈C,xm ≡xs ∨xn ≡xt score(xm , xn )
An identity pair is expected to be the highest correlated candidate of each in-
stance. However, it usually is not true because the ambiguity of instances. A
thresholding method that relies on the highest score would be better in this
situation. While true identity pair and its impostors have the similar score, θ is
assigned with a quite large value. On the other hand, η is additionally config-
ured as the minimum requirement of an identity. Like in Algorithm 1, η ensures
there is no assumption about the surjection of given data sources.
The key distinctions of our approach in comparison with the previous are the
use of weighted average and adaptive filtering. Previous systems do not have the
data driven information like confidence of key alignments. Silk [10] provides a
manual weighting method; however, a good weight usually depends much on the
human knowledge about the data. For identity selection, Zhishi.Links [7] and
AgreementMaker[2] eventually select the best correlated candidates, while Silk
[10] and SERIMI [1] use threshold-based selection. We compare the interlinking
result of our system with those of Zhishi.Links, SERIMI, and AgreementMaker
in our experiments, which are reported in the next section.
4 Experiment
4.1 Experiment setup
We evaluate the efficiency of blocking step and the whole interlinking process
of SLINT. We also compare SLINT with Zhishi.Links [7], SERIMI [1], and
SLINT: A Schema-Independent Linked Data Interlinking System 9
AgreementMaker [2], which recently participated OAEI 2011 instance match-
ing campaign. Discussion on predicate selection and predicate alignment are
also included. For every test in our experiment, we use the same value for each
threshold. We set α, β ,γ (Eq. 3), δ (Eq. 4), ζ, (Algorithm 1), η, and θ (Eq. 7)
to 0.25, 0.25, 0.5, 0.25, 0.1, 0.5, 0.25, and 0.95, respectively. The fixed values of
α, β, γ and δ express the schema-independent capability of SLINT.
Like previous studies, for blocking, we use two evaluation metrics: pair com-
pleteness (PC) and reduction ratio (RR); For interlinking, we use recall (Rec),
precision (Prec), and F1 score, the harmonic mean of recall and precision. Eq.8
and Eq.9, Eq.10, and Eq.11 are the computations of used metrics.
N umber of correct candidates
PC = . (8)
N umber of actual identity pairs
N umber of candidates
RR = 1 − . (9)
N umber of all pairs
N umber of correct identity pairs
Rec = . (10)
N umber of actual identity pairs
N umber of correct identity pairs
P rec = . (11)
N umber of discovered pairs
The performance of an interlinking system is also very important. We report the
execution times of SLINT when running on a desktop machine equipped with
2.66Ghz quad-core CPU and 4GB of memory.
4.2 Datasets and discussion on predicate selection & predicate
alignment
We use 9 datasets in experiment. The first 7 datasets are IM@OAEI2011 datasets,
the ones used in instance matching campaign at the OAEI 20113 . Concretely, we
use the datasets of interlinking New York Times track, which asks participants
to detect identity pairs from NYTimes to DBpedia, Freebase, and Geonames.
These datasets belong to three domains: locations, organizations, and people.
The IM@OAEI2011 datasets are quite small. Therefore, for evaluating the com-
putational performance of the system, we select two larger datasets. The first
one, a medium dataset, contains 13758 pairs in film domain between DBpedia
and LinkedMDB4 . The second one is a quite large dataset, which contains 86456
pairs in locations domain between DBpedia and Geonames5 . All datasets are
downloaded by dereferencing URI and stored in external memory in advance.
We remove triples having owl:sameAs predicates and rdf:seeAlso predicates of
course. Table 1 gives the overview of these datasets. In this table, IM@OAEI2011
datasets are from D1 to D7, and the last two datasets are D8 and D9. We also
include the number of predicates and predicate alignments in this table. Denotes
that s and t are the source and target data, respectively; P rd and P fd are the
3
http://oaei.ontologymatching.org/2011/instance/
4
http://downloads.dbpedia.org/3.7/links/linkedmdb links.nt.bz2
5
http://downloads.dbpedia.org/3.7/links/geonames links.nt.bz2
10 K. Nguyen, R. Ichise, B. Le
Table 1. Number of predicates and predicate alignments
ID Source Target Domain Pairs P rs P rt P fs P ft A K
D1 NYTimes DBpedia Locations 1920 12 2859 6 24 26 7
D2 NYTimes DBpedia Organizations 1949 10 1735 6 14 21 5
D3 NYTimes DBpedia People 4977 10 1941 5 20 33 8
D4 NYTimes Freebase Locations 1920 12 775 6 18 20 4
D5 NYTimes Freebase Organizations 3044 10 1492 5 18 33 3
D6 NYTimes Freebase People 4979 10 1844 5 18 32 5
D7 NYTimes Geonames Locations 1789 12 32 6 13 14 4
D8 DBpedia LinkdMDB Movie 13758 1343 54 16 7 31 14
D9 DBpedia Geonames Locations 86456 8194 17 11 7 27 8
number of predicates in data source d before and after selected by predicate
selection step, respectively; A and K are the number of all predicate alignments
and only key alignments, respectively.
In general, excepts in NYTimes, the number of available predicates in the
schema of each data source is very large, but the important predicates occupy a
very small percent. As our observation, the predicates declaring the label or the
name of objects are always aligned with a very high confidence. The non-string
type predicates also frequently construct the key alignments. For example, in lo-
cations domain, the key alignments always contain the right combination of pred-
icates declaring latitudes and longitudes. The predicate releaseDate of DBpedia
is successfully combined with predicate date and predicate initial relase date of
LinkedMDB. An interesting key alignment in dataset D6 is the high confidence
combination of core#preLabel of NYTimes and user.mikeshwe.default domain.vi-
deosurf card.videosurf link text of Freebase. The latter predicate may be difficult
for manual selection since the meaning of the predicate name does not imply the
label. Clearly, it is not easy for human to detect every compatible predicates.
When manually doing this task, we may lose to leverage all useful information.
4.3 Blocking and interlinking result
This section reports the result of blocking and the whole interlinking process.
Concretely, we report the pair completeness and reduction ratio of blocking, and
precision, recall, F1 score, and the runtime of the system. Table 2 shows these
metrics on each dataset. According to this table, although we cannot retain all
identity pairs, the lowest PC is still very high at 0.94. Besides, the high RRs
reveal that the numbers of retrieved candidates are very small if compared with
the numbers of total pairs. For all the evidences of PC and RR, the aim of
blocking is successfully achieved.
For interlinking, the precison and recall are very competitive. The recall,
which is not much lower than pair completeness, implies that the instance match-
ing performs a good work. The high precision implies that our system has a ef-
ficient disambiguation capability on tested datasets. It seems easy for SLINT to
interlink people domain, whereas in locations domain, SLINT achieves the best
result on IM@OAEI2011 datasets involving with Geonames.
The execution time of SLINT is very good in overview. Because we use co-
occurrence structure in blocking, the memory on tested machine cannot satisfy
SLINT: A Schema-Independent Linked Data Interlinking System 11
Table 2. Number of candidates, PC, RR, Rec, Prec, F1, and execution time
Dataset Blocking Interlinking
Candidates PC RR Prec Rec F1 Runtime
D1 4102 0.9901 0.9989 0.9636 0.9651 0.9644 3.55
D2 3457 0.9831 0.9970 0.9768 0.9487 0.9625 4.29
D3 9628 0.9950 0.9972 0.9883 0.9841 0.9862 12.74
D4 3580 0.9849 0.9990 0.9486 0.9521 0.9504 3.78
D5 7744 0.9823 0.9992 0.9610 0.9560 0.9585 6.71
D6 10333 0.9938 0.9996 0.9944 0.9904 0.9924 18.25
D7 2473 0.9961 0.9959 0.9883 0.9888 0.9885 1.63
D8 33926 0.9948 0.9998 0.9317 0.9868 0.9584 67.76
D9 418592 0.9468 0.9999 0.9782 0.9418 0.9596 2465.38
a very large dataset. In our context, interlinking dataset D9 has such issue.
We temporarily implement a parallel program for re-computing every element
of the co-occurrence matrix. The interlinking on this dataset is therefore takes
much time because the repeat of data traversing and high computational cost.
However, the high speeds on other datasets are really promising for scaling-up
our system.
The advantage of blocking is very high if we compare the time of interlinking
with and without this step. For example, the time for instance matching step to
match 33926 candidates of dataset D8 is 12.2 seconds. It means that the time
for matching all available pairs will be nearly 17 hours, whereas this number is
only 67.76 seconds in total if we implement the blocking step. Blocking averagely
occupies 58% total runtime of interlinking process on the nine tested datasets.
Although this number is over a half, the advantage of blocking is still very
considerable.
4.4 Comparison with previous interlinking systems
As mentioned, we compare our system with AgreementMaker [2], SERIMI [1],
and Zhishi.Links [7]. Because these systems recently participated instance match-
ing campaign of the OAEI 2011, we use the results on IM@OAEI2011 datasets
for comparison. Table 3 shows the interlinking result of SLINT and others. As
showed in this table, it is clear that our system totally outperforms the others
on both precision and recall. AgreementMaker has a competitive precision with
SLINT on dataset D3 but this system is much lower in recall. Zhishi.Links re-
sults on dataset D3 are very high, but the F1 score of SLINT is still 0.05 higher
in overall.
The prominent differences of SLINT and these systems are that we use the
confidence of alignment as the weight for blocking and instance matching, and
discriminate data types with the use of TF-IDF for the token of string and URI.
Generally, SLINT is verified as the best accurate one among compared systems.
5 Conclusion
In this paper, we present SLINT, an efficient schema-independent linked data
interlinking system. We select important predicates by predicate’s coverage and
discriminability. The predicate alignments are constructed and filtered for ob-
taining key alignments. We implement an adaptive filtering technique to produce
12 K. Nguyen, R. Ichise, B. Le
Table 3. Comparison with previous interlinking systems.
Dataset SLINT Agree.Maker SERIMI Zhishi.Links
Prec Rec F1 Prec Rec F1 Prec Rec F1 Prec Rec F1
D1 0.96 0.97 0.96 0.79 0.61 0.69 0.69 0.67 0.68 0.92 0.91 0.92
D2 0.98 0.95 0.96 0.84 0.67 0.74 0.89 0.87 0.88 0.90 0.93 0.91
D3 0.99 0.98 0.99 0.98 0.80 0.88 0.94 0.94 0.94 0.97 0.97 0.97
D4 0.95 0.95 0.95 0.88 0.81 0.85 0.92 0.90 0.91 0.90 0.86 0.88
D5 0.96 0.96 0.96 0.87 0.74 0.80 5.92 0.89 0.91 0.89 0.85 0.87
D6 0.99 0.99 0.99 0.97 0.95 0.96 0.93 0.91 0.92 0.93 0.92 0.93
D7 0.99 0.99 0.99 0.90 0.80 0.85 0.79 0.81 0.80 0.94 0.88 0.91
H-mean. 0.97 0.97 0.97 0.92 0.80 0.85 0.89 0.88 0.89 0.93 0.92 0.92
candidates and identities. Compare with the most recent systems, SLINT highly
outperforms the precision and recall in interlinking. The performance of SLINT
is also very high when it takes around 1 minute to detect more than 13,000
identity pairs.
Although SLINT has good result on tested datasets, it is not sufficient to
evaluate the scalability of our system, which we consider as the current limiting
point because of the used of weighted co-occurrence matrix. We will investigate
about a solution for this issue in our next work. Besides, we also interested
in automatic configuration for every threshold used in SLINT and improving
SLINT into a novel cross-domain interlinking system.
References
1. S. Araujo, D. Tran, A. de Vries, J. Hidders, and D. Schwabe. SERIMI: Class-based
disambiguation for effective instance matching over heterogeneous web data. In
SIGMOD’12 15th Workshop on Web and Database, pages 19–25, 2012.
2. I. F. Cruz, F. P. Antonelli, and C. Stroe. AgreementMaker: efficient matching for
large real-world schemas and ontologies. VLDB Endow., 2:1586–1589, 2009.
3. R. Isele and C. Bizer. Learning linkage rules using genetic programming. In ISWC’
11 6th Workshop on Ontology Matching, pages 13–24, 2011.
4. A. Ngomo, J. Lehmann, S. Auer, and K. Höffner. RAVEN - Active learning of link
specifications. In ISWC’ 11 6th Workshop on Ontology Matching, pages 25–36,
2011.
5. K. Nguyen, R. Ichise, and B. Le. Learning approach for domain-independent linked
data instance matching. In KDD’12 2nd Workshop on Minning Data Semantic,
pages 7:1–7:8, 2012.
6. A. Nikolov, M. d’Aquin, and E. Motta. Unsupervised learning of link discovery
configuration. In ESCW’12, pages 119–133, 2012.
7. X. Niu, S. Rong, Y. Zhang, and H. Wang. Zhishi.links results for OAEI 2011. In
ISWC’ 11 6th Workshop on Ontology Matching, pages 220–227, 2011.
8. A. Schultz, A. Matteini, R. Isele, C. Bizer, and C. Becker. LDIF - Linked data
integration framework. In ISWC’ 11 2nd Workshop on Consuming Linked Data,
2011.
9. D. Song and J. Heflin. Automatically generating data linkages using a domain-
independent candidate selection approach. In ISWC’ 11, pages 649–664, 2011.
10. J. Volz, C. Bizez, M. Gaedke, and G. Kobilarov. Discovering and maintaining links
on the web of data. In ISWC’ 09, pages 650–665, 2009.