090626cc tech-summit

DSNotify: Handling Broken Links in the Web of Data

Niko P. Popitsch Bernhard Haslhofer
niko.popitsch@univie.ac.at bernhard.haslhofer@univie.ac.at
University of Vienna, Department of Distributed and Multimedia Systems
Liebiggasse 4/3-4, 1010 Vienna, Austria

ABSTRACT the network leading to the practical unavailability of infor-
The Web of Data has emerged as a way of exposing struc- mation [15, 2, 18, 19, 22].
tured linked data on the Web. It builds on the central build- Recently, Linked Data has been proposed as an approach
ing blocks of the Web (URIs, HTTP) and benefits from its for exposing structured data by means of common Web
simplicity and wide-spread adoption. It does, however, also technologies such as dereferencable URIs, HTTP, and RDF.
inherit the unresolved issues such as the broken link prob- Links between resources play a central role in this Web of
lem. Broken links constitute a major challenge for actors Data as they connect semantically related data. Meanwhile
consuming Linked Data as they require them to deal with an estimated number of 4.7 billion RDF triples and 142 mil-
reduced accessibility of data. We believe that the broken lion RDF links (cf. [6]) are exposed on the Web by numerous
link problem is a major threat to the whole Web of Data data sources from different domains. DBpedia [7], the struc-
idea and that both Linked Data consumers and providers tured representation of Wikipedia, is the most prominent
will require solutions that deal with this problem. Since no one. Web agents can easily retrieve these data by derefer-
general solutions for fixing such links in the Web of Data encing URIs via HTTP and process the returned RDF data
have emerged, we make three contributions into this direc- in their own application context. In the example shown in
tion: first, we provide a concise definition of the broken Figure 1, an institution has linked a resource representing a
link problem and a comprehensive analysis of existing ap- band in their local data set with the corresponding resource
proaches. Second, we present DSNotify, a generic frame- in DBpedia in order to publish a combination of these data
work able to assist human and machine actors in fixing bro- on its Web portal.
ken links. It uses heuristic feature comparison and employs Source Resource Link Target Resource
a time-interval-based blocking technique for the underlying
http://example.com/bands/OliverBlack rdfs:seeAlso http://dbpedia.org/resource/Oliver_Black
instance matching problem. Third, we derived benchmark
datasets from knowledge bases such as DBpedia and eval- dbpprop: abstract
foaf:name
uated the effectiveness of our approach with respect to the Representation Representation

broken link problem. Our results show the feasibility of a Oliver Black
Oliver Black is a Canadian
rock group ...
time-interval-based blocking approach for systems that aim
at detecting and fixing broken links in the Web of Data.
Figure 1: Sample link to a DBpedia resource
Categories and Subject Descriptors
H.4.m [Information Systems]: Miscellaneous; H.3.3 [ The Linked Data approach builds on the Architecture of
Information Systems]: Information Search and Retrieval the World Wide Web [16] and inherits the technical benefits
such as simplicity and wide-spread adoption but also the
unsolved problems such as broken links. In the course of
General Terms time, resources and their representations can be removed
Algorithms, Theory, Experimentation, Measurement completely or “moved” to another URI meaning that they
are published under a different HTTP URI. In case of the
1. INTRODUCTION above example, the band eventually changed its name and
Data integrity on the Web is not given because URI ref- the title of their Wikipedia entry to “Townline”, with the
erences of links between resources are not as cool 1 as they result that the corresponding DBpedia entry moved from its
are supposed to be. Resources may be removed, moved, or previous URI to http://dbpedia.org/resource/Townline.
updated over time leading to broken links. These constitute In the Linked Data context, we informally speak of links
a major problem for consumers of Web data, both human pointing from one resource (source) to another resource (tar-
and machine actors, as they interrupt navigational paths in get). Such a link is broken when the representations of the
target cannot be accessed anymore. However, we consider
1
See http://www.w3.org/Provider/Style/URI a link also as broken when the representations of the target
Copyright is held by the International World Wide Web Conference Com-
resource were updated in such a way that they underwent a
mittee (IW3C2). Distribution of these papers is limited to classroom use, change in meaning the link-creator had not in mind.
and personal use by others. When regarding the changes in recent DBpedia releases,
WWW 2010, April 26–30, 2010, Raleigh, North Carolina, USA. the broken link problem becomes evident: analogous to [7],
ACM 978-1-60558-799-8/10/04.

we analyzed the instances of common DBpedia classes in the 2.1 Definition of Broken Links and Events
snapshots 3.2 (October 2008) and 3.3 (May 2009), identified We distinguish two types of broken links that differ in their
the events that occurred between these versions and catego- characteristics and in the way how they can be detected and
rized them into move, remove, and create events. The results fixed: structurally and semantically broken links.
in Table 1 show that within a period of about seven months
the DBpedia data space has grown and was considerably re- Structurally broken links. Formally, we define structurally
organized. Different from other data sources, DBpedia has broken (binary) links as follows:
the great advantage that it records move events in so-called
redirect links that are derived from redirection pages. These Definition 1 (Broken Link). Let R and D be the set
are automatically created in the Wikipedia when articles are of resources and resource representations respectively and
renamed. P(A) be the powerset of an arbitrary set A.
Class Ins. 3.2 Ins. 3.3 MV RM CR
Further let δt : R −→ P(D), be a dereferencing function
Person 213,016 244,621 2,841 20,561 49,325 returning the set of representations of a given resource at a
Place 247,508 318,017 2,209 2,430 70,730 given time t.
Organisation 76,343 105,827 2,020 1,242 28,706 Now we can define a (binary) link as a pair
Work 189,725 213,231 4,097 6,558 25,967
l = (rsource , rtarget ) with rsource ∧ rtarget ∈ R.
Table 1: Changes between two DBpedia releases. Such a link is called structurally broken if
Ins. 3.2 and Ins 3.3 denote the number of instances δt−∆ (rtarget ) = ∅ ∧ δt (rtarget ) = ∅.
of a certain DBpedia class in the respective release
That is, a link (as depicted for example in Figure 1) is
data sets, MV the moved, RM the removed, and
considered structurally broken if its target resource had rep-
CR the number of created resources.
resentations that are not retrievable anymore2 . In the re-
mainder of this paper, we will refer to structurally broken
If humans encounter broken links caused by a move event, links simply as broken links if not stated otherwise.
they can search the data source or the Web for the new lo-
cation of the target resource. However, for machine agents
broken links can lead to serious processing errors or misin-
Semantically broken links. Apart from structurally bro-
ken links, we also consider a link as broken if the human
terpretation of resources when they do not implement ap-
interpretation (the meaning) of the representations of its
propriate fallback mechanisms. If, for instance, the link in
target resource differs from the one intended by the link au-
Figure 1 breaks and the target resource becomes unavailable
thor. In a quantitative analysis of Wikipedia articles that
due to a remove or move event, the referenced biography
changed their meaning over time [13], the authors found out
information cannot be provided anymore. If the resource
that only a small number of articles (6 out of a test set
representations are updated and undergo a change in mean-
of 100 articles) underwent minor or major changes in their
ing, the Web portal could encounter the problem of exposing
meaning. However, we do not think that these results can
semantically invalid information.
be generalized for arbitrary data sources.
While the detection of broken links on the Web is sup-
In contrast to structurally broken links, semantically bro-
ported by a number of tools, there are only few approaches
ken links are much harder to detect and fix by machine ac-
for automatically fixing them [19]. Techniques for dealing
tors. But they are fixable by human actors that may, in a
with the broken link problem in the Web of Data do not
semi-automatic process, report such events to a system that
exist yet. The current approach is to rely on the HTTP
then forwards these events to subscribed actors. We have
404 Not Found response and assume that data-consuming
foreseen this in our system (see Section 3).
actors can deal with it. We consider this as insufficient and
propose DSNotify, which we informally introduced in [12],
as a possible solution. DSNotify is a generic change detec- Events. Having defined links and broken links we can now
tion framework for Linked Data sources that informs data- define events that occur when datasets are modified, possi-
consuming actors about the various types of events (create, bly leading to broken links:
remove, move, update) that can occur in data sources.
Our contributions are: (i) we formalize the broken-link Definition 2 (Event). Let E be the set of all events
problem in the context of the Web of Data and provide a and e ∈ E be a quadruple e = (r1 , r2 , τ, t), where r1 ∈ R and
comprehensive analysis of existing solution strategies (Sec- r2 ∈ R ∪ {∅} are resources affected by the event,
tion 2), (ii) we present DSNotify, focusing on its core algo- τ ∈ {created, removed, updated, moved} is the type of the
rithms for handling the underlying instance matching prob- event and t is the time when the event took place. Further
lem (Section 3), and (iii) we have derived benchmark data let L ⊆ E be a set of detected events.
sets from the ISLab Instance Matching Benchmark [11] and Then we can assert that, ∀r ∈ R :
from the DBpedia knowledge base and evaluate the effec- δt−∆ (r) = ∅ ∧ δt (r) = ∅ =⇒ L ←− L ∪ {(r, ∅, created, t)} .
tiveness of the DSNotify approach (Section 4). δt−∆ (r) = ∅ ∧ δt (r) = ∅ =⇒ L ←− L ∪ {(r, ∅, removed, t)} .
δt−∆ (r) = δt (r) =⇒ L ←− L ∪ {(r, ∅, updated, t)} .

2. THE BROKEN LINK PROBLEM Note the analogy between the definition of broken links
In the previous section, we informally described the bro- and removed events: whenever the representations of a link
ken link problem and its possible consequences. In this sec- 2
Note that our definitions do not consider a link as broken
tion we want to contribute a formal definition of a broken if only some of the representations of the target resource are
link in the context of Linked Data and discuss existing so- not retrievable anymore. We consider clarifications of this
lution strategies for dealing with this problem. issue as a topic for further research.

target are removed, the corresponding links are broken. Now Broken link type
Solution Strategy Class A B C
we have defined create, remove and update events, but what
Ignoring the Problem -
about the event type “moved ”? In fact, it is not possible Embedded Links p
to speak about moved resources considering only the pre- Relative References p
vious definitions. Although there is no concept of resource Indirection p
location in RDF, it exists in the Web of Data as it relies Versioned and Static Collections p
on dereferencable HTTP URIs. For this reason, we define a Regular Updates p
Redundancy p
weak equality relation between resources in the Web of Data
Dynamic Links a
based on a similarity relation between its representations Notification c
and build on that to define move events: Detect and Correct c
Manual Edit/Search c
Definition 3 (Move Event).
Let σ : P(D) × P(D) −→ [0, 1] be a similarity function
between two sets of resource representations. Further let Table 2: Solution strategies for the broken link prob-
Θ ∈ [0, 1] be a threshold value. lem. The strategies are classified according to Ash-
We define the maximum similarity of a resource man [2]: preventative methods (p) that try to avoid
rold ∈ {r ∈ R|δt (rold ) = ∅} with any other resource rnew ∈ broken links in the first place, adaptive methods (a)
R {rold } as simmax ≡ max(σ(δt−∆ (rold ), δt (rnew ))) . that create links dynamically thereby avoiding bro-
rold
Now we can assert that: ken links and corrective methods (c) that try to fix
∃!rnew ∈ R|δt−∆ (rnew ) = ∅ ∧ Θ < σ(δt−∆ (rold ), δt (rnew )) = broken links. Symbols: potentially fixes/avoids such
simmax =⇒ L ← L ∪ {(rnew , rold , moved, t)} . broken links ( ), does not fix/avoid such broken
rold
links ( ), partly fixes/avoids such broken links ( )
Thus, we consider a resource as moved from one HTTP
URI to another when the resource with the “old” URI was
removed, the resource with the “new” URI was created and and the target resource is referenced using e.g., an HTTP
the representations retrieved from the “old” URI are very URI reference. This method preserves link integrity when
similar3 to the representations retrieved from the “new” URI. the source resource of a link is moved (type A).

2.2 Solution Strategies Relative References. Relative references prevent broken
In consequence of Definition 1, we further provide a more links in some cases (e.g., when a whole resource collection
informal definition of link integrity: is moved). But neither does this method avoid broken links
due to removed resources (type C ), nor does it hinder links
Definition 4 (Link Integrity). Link integrity is a with external sources/targets (i.e., absolute references) from
qualitative measure for the reliability that a link leads to the breaking.
representations of a resource that were intended by the au-
thor of the link. Indirection5 . Introducing a layer of indirection allows con-
tent providers to keep links to their Web resources up to
Methods to preserve link integrity have a long history in
date. Aliases refer to the location of a resource and special
hypertext research. We have analyzed existing approaches
services translate between an alias and its referred location.
in detail, building to a great part on Ashman’s work [2] and
Moving or removing a resource requires an update in these
extending it. In the following, we categorize broken links by
service’s translation tables. Uniform Resource Names were
the events that caused their breaking:
proposed for such an indirection strategy, PURLs and DOIs
type A: links broken because source resources were moved4 ,
are two well known examples [1]. Permalinks use a similar
type B: links broken because target resources were moved
strategy, although the translation step is performed by the
and type C: links broken because source or target resources
content repository itself and not by a special (possibly cen-
were removed. The identified solution strategies are sum-
tral) service. Another special case on the Web is the use
marized in Table 2 and discussed in the following:
of small (“gravestone”) pages that reside at the locations of
moved or removed resources and indicate what happened to
Ignoring the Problem. Although this can hardly be called the resource (e.g., by automatically redirecting the HTTP
a “solution strategy”, it is the status-quo to simply ignore request to the new location).
the problem of broken links and shift it to higher-level ap- The main disadvantage of the indirection strategy is that
plications that process the data. As mentioned before, this it depends on notifications (see below) for updating the
strategy is even less acceptable in the Web of Data. translation tables. Furthermore it “. . . requires the cooper-
ation of link creators to refer to the alias instead of to the
Embedded Links. The embedded link model [8] is the most absolute address.” [2]. Another disadvantage is that spe-
common way how links on the Web are modeled. As in cial “translation services” (PURL server, CMS, gravestone
HTML, the link is embedded in the source representation pages) are required that introduce additional latency when
3
Note that in the case that there are multiple possible move accessing resources (e.g., two HTTP requests instead of one).
target resources with equal (maximum) similarity to the re- Nevertheless, indirection is an increasingly popular method
moved resource rold , no event is issued (∃! should be read as on the Web. This strategy prevents broken links of type A
“there exists exactly one”).
4 5
Note that in our definitions we did not consider links of The “Dereferencing (Aliasing) of End Points” and “For-
type A as we assumed an embedded link model for Linked warding Mechanisms and Gravestones” categories from [2]
Data sources. are combined in this category.

and B and can also help with type C links, e.g., removed
PURLs result in HTTP 410 status code (Gone), which al- links to
lows an application to react accordingly (e.g., by removing Dsrc Dtgt

the links to this resource). updates consumes
consumes monitors
Versioned and Static Collections. In this approach, no
modifications/deletions of the considered resources are al- Application notifies DSNotify
lowed. This prevents broken links of types A-C within a consumes notifies
static collection (e.g., an archive), links with sources/targets event choice
outside this collection can still break. Examples are HTML
links into the Internet Archive 6 . Furthermore, semantically User
broken links may be prevented when e.g., linking to a certain
(unmodifiable) revision of a Wikipedia article.
Figure 2: DSNotify Usage Scenario.
Regular Updates. This approach is based on predictable
updates to resource collections, so applications can easily Manual Edit/Search. In this category we summarize man-
update their links to the new (predictable) resource loca- ual strategies for fixing broken links or re-finding missing
tions, avoiding broken links of types A-C ). link targets. This “solution strategy” is currently arguably
among the most popular ones on the Web. First, content
Redundancy. Redundant copies of resource representations providers may manually update links in their contents (per-
are kept and a service forwards referrers to one of these haps assisted by automatic link checking software like W3C’s
copies as long as at least one copy still exists. This approach link checker). Second, users may re-find the target resources
is related to the versioning and indirection approaches. How- of broken links e.g., by exploiting search engines or by man-
ever, such services can reasonably be applied only to highly ual URI manipulations.
available, unmodifiable data (e.g., collections of scientific
documents). This method may prevent broken links of types
A-C, examples include LOCKSS [21] or RepWeb [23]. 3. DSNOTIFY
After having investigated possible solution strategies to
Dynamic Links. In this method, links are created dynami- deal with the broken link problem, we now present our own
cally when required and are not stored, avoiding broken links solution strategy. It is called DSNotify and is a generic
of types A-C. The links are created based on computations change detection framework that assists human and machine
that may reflect the current state of the involved resources actors in fixing broken links. It can be used as an add-on to
as well as other (external) parameters, i.e., such links may existing applications that want to preserve link integrity in
be context-dependent. However, the automatic generation the data sets that are under their control (detect and correct
of links is a non-trivial task and this solution strategy is not strategy, see below). It can also be used to notify subscribed
applicable to many real-world problems. applications of changes in a set of data sources (notification
strategy). Further, it can be set-up as a service that au-
Notification. Here, clients are informed about the events tomatically forwards requests to new resource locations of
that may lead to broken links and all required information moved resources (indirection strategy).
(e.g., new resource locations) is communicated to them so DSNotify is not meant to be a service that monitors the
they can fix affected links. This strategy was for example whole Linked Data space but rather as a light-weight com-
used in the Hyper-G system [17] where resource updates are ponent that can be tailored to application-specific needs and
propagated between document servers using a p-flood algo- detects modifications in selected Linked Data sources.
rithm. It is also the strategy of Silk and Triplify discussed A typical usage scenario is illustrated in Figure 2: an ap-
in Section 5. This method resolves broken links caused by plication hosts a data set Dsrc that contains links to a target
A-C but requires that the content provider can observe and data set Dtgt . The application consumes and integrates data
communicate such events. from both data sets, and provides a view on this data (such
as a Website) consumed by Web users. The application uses
DSNotify to monitor Dtgt as it has no control over this tar-
Detect and Correct. The solution for the broken link prob- get data set. DSNotify notifies the application (and possibly
lem proposed in this work falls mainly into this category
other subscribed actors) about the events occurring in Dtgt
that Ashman describes in her work as: “. . . the application
and the application can update and fix potentially broken
using the link first checks the validity of the endpoint refer-
links in the source data set.
ence against the information, perhaps matching it with an
expected value. If the validity test fails, an attempt to correct 3.1 General Approach
the link by relocating it may be made . . . ” [2].
As all other solutions in this category (cf., Section 5), we Our approach for fixing broken links is based on an in-
use heuristic methods to semi-automatically fix broken links dexing infrastructure. A monitor periodically accesses con-
of the types A-C. sidered data sources (e.g., a Linked Data source), creates
an item for each resource it encounters and extracts a fea-
6
For example the URI http://web.archive.org/web/ ture vector from this item’s representations. The feature
19971011064403/http://www.archive.org/index.html vector is stored together with the item’s URI in an item
references the 1997 version of the internet archive main index (ii). The set of extracted features, their implementa-
page. tion, and their extractors are configurable. Feature vectors

CA CB CC CD CE RA MB,F MC,G CH RF RD ME,I
are updated in the ii with every monitoring cycle resulting
in possible update events logged by DSNotify.
Items corresponding to resources that are not found any- m0 m1 h1

more are moved to another index, the removed item index A
B
E
F
B E
F A
B
G C
(rii). After some timeout period, items are moved from the C
D D A
G
D C
rii to a third index called the archived item index (aii) re- ii rii aii
sulting in a remove event (cf. Definition 2).
Items in the ii and the rii are periodically considered by a t
housekeeper thread (a “move detector”) that compares their m2 h2 m3 h3
feature vectors and tries to identify possible successors for H F B H F B H F B H F
B
I E I E I E I E
removed (real-world) items (cf. Definition 3). A plug-in G D C G
D C G
D C G CD
A A A A
heuristic is used for this comparison; in the default con-
figuration a vector space model acting on the extracted and timeout timeout

weighted feature vectors is used. The similarity measures for
the features themselves are configurable; for character string Figure 3: Example timeline illustrating the main
features one could use, for instance, exact string matching workflow of DSNotify. Ci ,Ri and Mi,j denote create,
or the Levenshtein distance. The similarity value calculated remove and move events of items i and j. mx and
by the heuristic is compared against threshold values 7 to hx denote monitoring and housekeeping operations
determine whether an item is another item’s predecessor respectively. The current index contents is shown
(resulting in a detected move event) or not (possibly re- in the grey boxes below the respective operation,
sulting in a detected create event). The predecessors of the the overall process is explained in the text.
newly indexed items are moved to the aii and linked to the
new corresponding items. This enables actors to query the
DSNotify indices for actual locations of items. the evaluation of our system (see Section 4).
The events detected by DSNotify are stored in a central
event log. This log as well as the indices can be accessed 3.3 Monitoring and Housekeeping
via various interfaces, such as a Java API, an XML-RPC The cooperation of monitor, housekeeper, and the indices
interface, and an HTTP interface. is depicted in Figure 3. To simplify matters, we assume an
empty dataset at the beginning. Then four items (A,B,C,D)
3.2 Time-interval-based Blocking are created before the initial DSNotify monitoring process
The main task of DSNotify is the efficient and accurate starts at m0 . The four items are detected and added to
matching of pairs of feature vectors representing the same the item index (ii). Then a new item E is created, item A
real-world item at different points in time. As in record is removed and the items B and C are “moved” to a new
linkage and related problems (cf. Section 5), the number location becoming items F and G respectively. At m1 the
of such pairs grows quadratically with the number of con- three items that are not found anymore by the monitor are
sidered items resulting in unacceptable computational ef- “moved” to the removed item index (rii) and the new item
fort. The reduction in the number of comparisons is called is added to the ii. When the housekeeper is started for the
blocking and various approaches have been proposed in the first time at h1 , it acts on the current indices and compares
past [25]. the recent new items (E,F ,G) with the recently removed
We have developed a time-interval-based blocking (TIBB) items (B,C,A). It does not include the “old” item D in its
mechanism for an efficient and accurate reduction of the feature vector comparisons. The housekeeper detects B as a
number of compared feature vector pairs. Our method in- predecessor of F and C as a predecessor of G, moves B and
cludes only the feature vectors derived from items that were C to the archived item index (aii) and links them to their
detected as being recently created or removed, blocking out successors. Between m1 and m2 a new item is created (H),
the vast majority of the items in our indices. Reconsider- two items (F ,D) are removed and the item E is “moved”
ing Definition 3, this means that we are allowing only small to item I. The monitor updates the indices accordingly at
values for ∆. Thus, if x is the number of feature vectors m2 and the subsequent housekeeping operation at h2 tries
stored in our system, n is the number of new items and r to find predecessors of the items H and I. But before this
is the number of recently removed items with n + r ≤ x, operation, the housekeeper recognizes that the retention pe-
then the number of comparisons in a single DSNotify house- riod of item A in rii is longer than the timeout period and
keeping operation is n · r instead of x2 . It is intuitively clear moves it to the aii. The housekeeper then detects E as a
that normally n and r are much smaller than x and therefore predecessor of I, moves it also to the aii and links it to I.
n·r x2 . The actual number of feature vector comparisons Between m2 and m3 no events take place and the indices
in a single housekeeper operation depends on the vitality of remain untouched by the monitor. At h3 the housekeeper
the monitored data source with respect to created, removed recognizes the timeout of the items F and D and moves
and moved items and on the frequency of housekeeping op- them to the aii leaving an empty rii.
erations8 . We have analyzed and confirmed this behavior in
7
3.4 Event Choices
Note that using threshold values for the determination As mentioned before, a threshold value (upperThreshold )
of non-matches, possible-matches and matches was already
proposed by Fellegi and Sunter in 1969 [10]. is used to decide whether two feature vectors are similar
8
As housekeeping and monitoring are separate operations enough to assume their corresponding items as a predeces-
in DSNotify, this number depends also on the monitoring sor/successor pair. Furthermore, DSNotify uses a second
frequency when lower than the housekeeping frequency. threshold value (lowerThreshold ) to decide whether two fea-

Data: Item indices ii, rii, aii
Result: List of detected events L updates monitors
begin Dsrc
Move timed-out items from rii to aii;
L ←− ∅; P M C ←− ∅; Simulator DSNotify
Dobs
foreach ni ∈ ii.getRecentItems() do
pmc ←− ∅;
foreach oi ∈ rii.getItems() do Dtgt
sim ←− calculateSimilarity(oi, ni);
if sim > lowerT hreshold then Analyzer
Etest Edet
pmc ←− pmc + {(oi, ni, sim)};
end
end Results
if pmc = ∅ then
L ←− L + {(ni, ∅, create)};
else
P M C ←− P M C + {pmc}; Figure 4: The DSNotify evaluation approach. A
end simulator takes two datasets (Dsrc and Dtgt ) and an
end eventset (Etest ) as input and continuously updates
foreach pmc ∈ P M C do a newly created observed dataset (Dobs ). DSNotify
if pmc = ∅ then
(oimax , nimax , simmax ) ←− monitors this dataset and creates a log of detected
getElementW ithM axSim(pmc); events (Edet ). This log is compared to the eventset
if simmax > upperT hreshold then Etest to evaluate the system’s accuracy.
L ←− L + {(oimax , nimax , move)};
move oimax to aii;
link oimax to nimax ; we rely on simple timeouts for removing old data items from
remove all elements from P M C where these structures but this method can still result in unaccept-
pmc.oi = oimax ;
able memory consumption when monitoring highly dynamic
else
Issue an eventChoice for pmc; data sources. More advanced strategies are under consider-
end ation. Note that we consider particularly the feature vector
end history as a very valuable data structure as it allows ex post
end analysis of the evolution of items w.r.t. their location in a
return L; data set and the values of the indexed features.
end
Algorithm 1: Central DSNotify housekeeping algorithm. 4. EVALUATION
In the evaluation of our system we concentrated on two
issues: first, we wanted to evaluate the system for its ap-
ture vectors are similar enough to be even considered as
plicability for real-world Linked Data sources, and second,
such a pair (“possible move candidates”, pmc). When none
we wanted to analyze the influence of the housekeeping fre-
of the feature vectors considered for a possible move op-
quency on the overall effectiveness of the system.
eration are similar enough (i.e., >upperThreshold ), DSNo-
We evaluated our system with datasets that we call event-
tify stores all considered pairs of feature vectors with sim-
sets. An eventset is a timely-ordered set of events (cf. Defini-
ilarity values >lowerThreshold in a so-called event choice
tion 2 and 3) that transforms a source into a target dataset.
object. Event choices are representations of decisions that
Thus, an eventset can be seen as the event history of a
have to be made outside of DSNotify, for example by hu-
dataset. We have developed a simulator that can re-play
man actors or by other machine actors that can resort to
such eventsets, interpreting the event timestamps with re-
additonal data/knowledge. These external actors may ac-
gard to a configurable duration of the whole simulation. Fig-
cess the log of event choice objects and send their decisions
ure 4 depicts an overview of our evaluation approach.
about what feature vector pair (if any) corresponds to a pre-
All experiments were carried out on a system using two
decessor/successor pair back. DSNotify will now update its
Intel Xeon CPUs with 2.66 Ghz each and 8 GB of RAM.
indices accordingly and send notifications to all subscribed
The used threshold values were 0.8 (upperThreshold ) and 0.3
actors. A detailed description of the overall housekeeping al-
(lowerThreshold ). We have created two types of eventsets
gorithm, the core of DSNotify, is presented in Algorithm 1.
from existing datasets for our evaluation: the iimb-eventsets
3.5 Item History and Event Log and the dbpedia-eventset 9 .
As discussed above, DSNotify incrementally constructs 4.1 The IIMB Eventsets
three central data structures during its operation: (i) an
The iimb-eventsets are derived from the ISLab Instance
event log containing all events detected by the system, (ii) a
Matching Benchmark [11] which contains one (source) data-
log containing all unresolved event choices and (iii) a linked
set containing 222 instances and 37 target datasets that vary
structure of feature vectors constituting a history of the re-
in number and type of introduced modifications to the in-
spective items. This latter structure is stored in the indices
stance data. It is the goal of instance matching tools to
maintained by DSNotify. All three data structures can be
match the resources in the source dataset with the resources
accessed in various ways by agents that make use of DSNo-
in the respective target dataset by comparing their instance
tify for fixing broken links as further described in [12].
data. The benchmark contains an alignment file describing
As these data structures may grow indefinitely, a strategy
9
for pruning them from time to time is required. Currently All data sets are available at http://dsnotify.org/.

90 100
Name Coverage H Hnorm hki=1s
tbox:cogito-Name 0.995 5.378 0.995

80
tbox:cogito-first sentence 0.991 5.354 0.991 q
hki=3s q
q

F1 − measure (%)
q

70
tbox:cogito-tag 0.986 1.084 0.201 hki=10s
q
q

60
tbox:cogito-domain 0.982 3.129 0.579 q q

50
hki=20s q
tbox:wikipedia-name 0.333 1.801 0.333

40
q
tbox:wikipedia-birthdate 0.225 1.217 0.225 q
hki=30s
q

30
q
tbox:wikipedia-location 0.185 0.992 0.184 q

20
q
q q q q
tbox:wikipedia-birthplace 0.104 0.553 0.102

10
q
Namespace prefix tbox: <http://islab.dico.unimi.it/iimb/tbox.owl# >

0
1 2 3 4 5 6 7 8 9 10
Table 3: Coverage, entropy and normalized en- iimb−eventset number
tropy of all properties in the iimb datasets with a
coverage > 10%. The selected properties are written Figure 5: Influence of the housekeeping interval
in bold. (hki) on the F1 -measure in the iimb-eventsets evalu-
ations.
what resources correspond to each other that can be used
to measure the effectiveness of such tools. We used this curacy with the increasing dataset number. This is also
alignment information to derive 10 eventsets, correspond- expected as the benchmarks introduces more value trans-
ing to the first 10 iimb target datasets, each containing 222 formations with higher dataset numbers, although there are
move events. The first 10 iimb datasets introduce increasing two outliners for the datasets 7 and 10.
numbers of value transformations like typographical errors
to the instance data. We used random timestamps for the 4.2 The DBpedia Persondata Eventset
events (as this data is not available in this benchmark) that In order to evaluate our approach with real-world data we
resulted in an equal distribution of events over the eventset have created a dbpedia-eventset that was derived from the
duration. person datasets of the DBpedia snapshots 3.2 and 3.310 . The
We have simulated these eventsets, monitored the chang- raw persondata datasets contain 20,284 (version 3.2) and
ing dataset with DSNotify and measured precision and recall 29,498 (version 3.3) subjects typed as foaf:Person each hav-
of the reported events with respect to the eventset informa- ing three properties foaf:name, foaf:surname and foaf:given-
tion. For a useful feature selection we first calculated the name. Naturally, these properties are very well suited to
entropy of the properties with a coverage > 10%, i.e., only uniquely identify persons as also confirmed by their high
properties were considered where at least 10% of the re- entropy values (cf. Table 4). For the same reasons as al-
sources had instance values. The results are summarized ready discussed for the iimb datasets an evaluation with
in Table 3. As the goal of the evaluation was not to op- only these properties would not clearly demonstrate our
timize the resulting precision/recall values but to analyze approach. Therefore we enriched both raw data sets with
our blocking approach, we consequently chose the properties four properties (see Table 4) from the respective DBpedia
tbox:cogito-tag and tbox:cogito-domain for the evaluation be- Mapping-based Infobox Extraction datasets [7] with smaller
cause they have good coverage but comparatively small en- coverage and entropy values.
tropy in this dataset. We calculated the entropy as shown We derived the dbpedia-eventset by comparing both data-
in Equation 1 and normalized it by dividing by ln(n). sets for created, removed or updated resources. We retrieved
n the creation and removal dates for the events from Wikipedia
H(p) = − pi ln(pi ) (1) as these data are not included in the DBpedia datasets. For
i=1 the update events we used random dates. Furthermore, we
used the DBpedia redirect dataset to identify and generate
DSNotify was configured to compare these properties using move events. This dataset contains redirection information
the Levenshtein distance and both properties contributed derived from Wikipedia’s redirect pages that are automat-
equally (weight = 1.0) to the corresponding feature vector ically created when a Wikipedia article is renamed. The
comparison. The simulation was configured to run for 60 dates for these events were also retrieved from Wikipedia.
seconds, thus the monitored datasets changed with an aver- The resulting eventset contained 3810 create, 230 remove,
age rate of 222 = 3.7 events/s.
60 4161 update and 179 move events, summing up to 8380
As stated before, the goal of this evaluation was to demon- events11 . The histogram of the eventset depicted in Figure 6
strate the influence of the housekeeping frequency on the shows a high peak in bin 14. About a quarter of all events
overall effectiveness of the system. For this, we repeated 10
the experiment with varying housekeeping intervals of 1s, The snapshots contain a subset of all instances of type
3s, 10s, 20s, 30s (corresponding to an average rate 3.7, 11.1, foaf:Person and can be downloaded from http://dbpedia.
org/ (filename: persondata_en.nt).
37.0, 74.0, 111.0 events/housekeeping cycle) and calculated 11
Another 5666 events were excluded from the eventset as
the F1 -measure (the harmonic mean of precision and recall) they resulted from inaccuracies in the DBpedia datasets.
for each dataset (Figure 5). For example there are some items in the 3.2 snapshot
that are not part of the 3.3 snapshot but were not re-
Results. The results clearly demonstrate the expected de- moved from Wikipedia (a prominent example is the re-
crease in accuracy when increasing the length of the house- source http://dbpedia.org/resource/Tim_Berners-Lee).
Furthermore several items from version 3.3 were not in-
keeping intervals, as this leads to more feature vector com- cluded in version 3.2 although the creation date of the corre-
parisons and therefore more possibilities to make the wrong sponding Wikipedia article is before the creation date of the
decisions. Furthermore, Figure 5 depicts the decreasing ac- 3.2 snapshot. We decided to generally exclude such items.

90 100
Name Coverage H Hnorm q
foaf:name
q
q

foaf:name (d) 1.00/1.00 9.91/10.28 1.00/1.00 q
q
q q

80
q
q
dsnotify:rdfhash
foaf:surname (d) 1.00/1.00 9.11/9.25 0.92/0.90 q
q q

F1 − measure (%)
q
q

70
foaf:givenname
foaf:givenname (d) 1.00/1.00 8.23/8.52 0.83/0.83 dbpedia:birthdate+dbpedia:birthplace
q q
q

60
dbpedia:birthdate (d) 0.60/0.60 5.84/5.96 0.59/0.58 q

50
q
q
dbpedia:birthplace (o) 0.48/0.47 4.24/4.32 0.43/0.42 dbpedia:birthplace
q q

40
q
q
q
q
dbpedia:height (d) 0.10/0.08 0.65/0.51 0.07/0.05

30
q
dbpedia:birthdate
dbpedia:draftyear (d) 0.01/0.01 0.06/0.05 0.01/0.01

20
q
dbpedia:height

10
Namespace prefix dbpedia: <http://dbpedia.org/ontology/> dbpedia:draftyear

0
Namespace prefix foaf: <http://xmlns.com/foaf/0.1/>
1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4

Log10(average events/housekeeping cycle)
Table 4: Coverage, type, entropy and normalized
entropy of all properties in the enriched dbpedia
Figure 7: Influence of the number of events per
3.2/3.3 persondata sets. The selected properties
housekeeping cycle on the F1 -measure of detected
are written in bold. Symbols: object property (o),
move events in the dbpedia-eventset evaluation.
datatype property (d).
1000 1500 2000

Results. The results, depicted in Figure 7, show a fast sat-
uration of the F1 -measure with an decreasing number of
Events

events per housekeeping cycle. This clearly confirms the
findings from our iimb evaluation. The accuracy of DSNo-
500

tify increases with increasing housekeeping frequencies or
0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 decreasing event rates. From a pragmatical viewpoint, this
Bins means a tradeoff between the costs for monitoring and house-
keeping operations (computational effort, network transmis-
Figure 6: Histogram of the distribution of events in sion costs, etc.) and accuracy. The curve for the simple rdf-
the dbpedia-eventset. A bin corresponds to a time hash function is surprisingly good, stabilizing at about 80%
interval of about 11 days. for the F1 -measure. This can be attributed mainly to the
high precision rates that are expected from such a function.
The curve for the combined properties shows maximum val-
occured within this time interval. We think that such event ues for the F1 -measure of about 60%.
peaks are not unusual in real-world data and are interested The measured precision and recall rates are depicted in
how our application deals with such situations. Figure 8. Both measures show a decrease with increasing
We re-played the eventsets, monitored the changing data- numbers of events per housekeeping cycle. For the preci-
set with DSNotify and measured precision and recall of the sion this can be observed mainly for low-entropy properties
reported events with respect to the eventset information (cf. whereas the recall measures for all properties are affected.
Figure 4). We repeated the simulation seven times varying
dsnotify:rdfhash
90 100

q
q q
q

the number of average events per housekeeping interval and
q q
q
q q q q
q q
q
q foaf:name
calculated the F1 -measure of the reported move events12 . q q
80

q
Precision (%)
70

q

For each simulation, DSNotify was configured to index q q
60

q q

only one of the six selected properties in Table 4. To calcu- dbpedia:birthplace
50

q q q
foaf:givenname
40

late the similarity between datatype properties, we used the q

dbpedia:birthdate+dbpedia:birthplace
q
30

Levensthein distance. For object properties we used a sim- dbpedia:birthdate
20

dbpedia:height
ple similarity function that counted the number of common
10

dbpedia:draftyear
0

property values (i.e., resources) in both resources that are 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4

compared and divided it by the number of total values. Log10(average events/housekeeping cycle)

Furthermore, we ran the simulations indexing only one
90 100

foaf:name
cumulative attribute, an rdf-hash. This hash function calcu- q
q q
q
q

foaf:givenname q
80

q q

lates an MD5 hashsum over all string-serialized properties
70

q
dsnotify:rdfhash
Recall (%)

q
q
q q q

of a resource and the corresponding similarity function re-
60

q
q
50

q

turns 1.0 if the hash-sums are equal or 0.0 otherwise. Thus
q
q
dbpedia:birthdate+dbpedia:birthplace
40

dbpedia:birthplace q

this rdf-hash is sensible to any modifications in a resource’s
30

q
q q q
q
q
dbpedia:birthdate
20

q

instance data. q
10

dbpedia:height
dbpedia:draftyear
Additionally we evaluated a combination of the dbpedia
0

1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4
birthdate and birthplace properties, each contributed with Log10(average events/housekeeping cycle)
equal weight to the weighted feature vector. The coverage of
resources that had values for at least one of these attributes
Figure 8: Influence of the number of events per
was 65% in the 3.2 snapshot and 62% in the 3.3 snapshot.
housekeeping cycle on the measured precision and
recall of detected move events in the dbpedia-eventset
12
We fixed the housekeeping period for this experiment to 30s evaluation.
and varied the simulation length from 3600 to 56.25s. Thus
the event rates varied between 2.3 to 149.0 events/second It is, again, important to state that our evaluation had not
or 35.2 to 2250.1 events/housekeeeping interval respectively.
For these calculations we considered only move, remove and the goal to maximize the accuracy of the system for these
create events (i.e., 4219 events) from the eventset as only particular eventsets but rather to reveal the characteristics
these influence the accuracy of the algorithm. of our time-interval-based blocking approach. It shows that

we can achieve good results even for attributes with low en- RDF. Triplify also provides a Linked Data Update Log that
tropy when choosing an appropriate housekeeping frequency. groups updates to an RDF model within a certain timespan
into nested collections accessible via HTTP. In principle, this
5. RELATED WORK solution provides a scalable approach for logging events that
occurred in a Linked Data source. However, this notification
Besides the already cited works, our research is also closely
approach requires clients to regularly poll this log and the
related to the areas of record linkage, a well-researched prob-
data source to capture all these events (if possible). Fur-
lem in the database domain, and instance matching, which
thermore, the current specification of the vocabulary16 used
is related to ontology alignment and schema matching.
for describing the update events does not contain moved or
Record linkage13 is concerned with finding pairs of data
created events but only “Update” and “Deletion”.
records (from one or multiple datasets) that refer to (de-
Peridot is a tool developed by IBM for automatically fix-
scribe) the same real-world entity [25, 9]. This informa-
ing links in the Web. It is based on the patents [4, 5] and the
tion is useful e.g., for joining different relations or for du-
basic idea is to calculate fingerprints of Web documents and
plicate detection [9]. Record linkage is trivial where entity-
repair broken links based on their similarity. The method
unique identifiers (such as ISBN numbers or OWL inverse-
differs from DSNotify in that we consider the structured
functional properties like foaf:mbox ) are available. When
nature of Linked Data and support domain-specific, config-
such additional identifiers are missing, tools often rely on
urable similarity-heuristics on a property level which allows
probabilistic distance metrics and machine learning meth-
more advanced comparisons methods. Furthermore, DSNo-
ods (e.g., HMMs, TD-IDF; SVM). A comprehensive survey
tify introduces the described time-interval-based blocking
of record linkage research can be found in [9]. The instance
approach and detects also create, remove and update events.
matching problem is closely related to record linkage but re-
In [19], Morishima et al. describe Link Integrity Man-
quires certain specific methods when dealing with structural
agement tools that focus on fixing broken links in the Web
and logical heterogeneities as pointed out in [11].
that occurred due to moved link targets (type B, cf. section
Furthermore, our work is closely related to current work
2.2). Similar to DSNotify, they have developed a tool called
in Semantic Web and in hypertext research:
PageChaser that uses a heuristic approach to find missing
In [20], Phelps and Wilensky propose so-called robust hy-
resources based on indexed information (namely URIs, page
perlinks based on URI references that are decorated with a
content and redirect information). An explorer component,
small14 lexical signature composed of terms extracted from
which makes use of search engines, redirect information,
the referenced document. When a target document is not
and so-called link authorities (Web pages containing well-
found, this lexical signature can be used to re-find the re-
maintained links) is used to find possible URIs of moved
source using regular Web search engines. A disadvantage
resources. They also provide a heuristics-based method to
of the robust hyperlink approach is that it requires existing
calculate such link authority pages. A major difference to
URI references to be changed which is not the case with
our approach is that PageChaser was built for fixing links in
our approach. Furthermore, it is unclear how to extend this
the (human) Web exploiting some of its characteristics (like
method to non-textual resources whereas our feature vector
locality or link authorities), while DSNotify aims at becom-
based approach could in principle be combined with most
ing a general framework for assisting actors in fixing links
existing multimedia feature extraction solutions.
based on domain-specific content features.
In [14], an algorithm for record linkage (object consoli-
In a recent paper, Van de Sompel et al. discuss a proto-
dation) based on inverse-functional properties is described.
col for time-based content negotiation that can be used to
The approach groups instances with matching IFP values
access archived representations (“Mementos”) of resources
together and determines canonical URIs for identification of
[22]. By this, the protocol enables a kind of time-travel
the real-world entities described by those instances. Natu-
when accessing archived resources (such archives would fall
rally, IFPs can be used efficiently for record linkage prob-
into the Versioned and Static Collections category in Ta-
lems but are unfortunately unavailable in many real-world
ble 2). DSNotify could be used to build such archives when
datasets. In DSNotify, IFPs can be exploited by simply us-
a monitor implementation was used that would store not
ing them as a single feature in a feature vector.
only a feature vector derived from a resource representation
The Silk framework [24] aims mainly at the automatic,
but also the representation data itself.
heuristics-based discovery of semantic relationships between
resources in the Web of Data. These heuristics may be
configured using an XML-based links specification language 6. CONCLUSIONS AND FUTURE WORK
(Silk-LSL). In order to react on changes in the interlinked We presented the broken link problem in the context of
datasets, the authors propose a new SOAP-based protocol the Web of Data as a special case of the instance matching
(Web of Data - Link Maintenance Protocol, WOD-LMP 15 ) problem and showed the feasibility of a time-interval-based
for synchronizing and maintaining links between LD sources. blocking approach for systems that aim at detecting and fix-
Triplify [3] is a system that exposes data from relational ing such broken links. Reconsidering the solution strategies
databases as Linked Data. It is based on mapping HTTP re- for the broken link problem discussed in Section 2.2, we can
quests to RDBMS queries and publishing the result data as state that DSNotify can actually be used in multiple ways,
13
Record linkage is also known under many other names, including: (1) to function as a detect and correct module in
such as object identification, data cleaning, entity resolution, an existing software, (2) as a standalone notification service
coreference resolution or object consolidation [25, 14, 9]. that keeps subscribed clients informed about changes in a
14
The authors found out that five terms are enough to data source, (3) as an indirection service that automatically
uniquely identify a Web resource in virtually all cases. forwards requests for moved resources to their new location.
15
http://www4.wiwiss.fu-berlin.de/bizer/silk/
16
wodlmp/ http://triplify.org/vocabulary/update

The various interfaces for accessing the data structures built [12] B. Haslhofer and N. Popitsch. DSNotify – detecting
by DSNotify should facilitate the integration with existing and fixing broken links in linked data sets. In 8th
applications. Our approach is by design a semi-automatic International Workshop on Web Semantics (WebS
solution that is capable of integrating human intelligence in 09), co-located with DEXA 2009, 2009.
the sense of human-based computation. We plan to further [13] M. Hepp, K. Siorpaes, and D. Bachlechner. Harvesting
elaborate on this issue. However, DSNotify cannot “cure” wiki consensus: Using Wikipedia entries as vocabulary
the Web of Data from broken links. It may rather be used for knowledge management. IEEE Internet
as an add-on for particular data providers that want to keep Computing, 11(5):54–65, 2007.
a high level of link integrity in their data. [14] A. Hogan, A. Harth, and S. Decker. Performing object
The flexibility of our tool is founded in its generic nature consolidation on the semantic web data graph. In
and its customizability. Consequently, the development and Proceedings of the 1st I3 : Identity, Identifiers,
evaluation of additional monitors, features and extractors, Identification Workshop, 2007.
heuristics and indices is one part of our future work. In
[15] D. Ingham, S. Caughey, and M. Little. Fixing the
particular we want to research feasible methods for auto-
“broken-link” problem: the W3Objects approach.
matic feature selection as well as for the determination of
Comput. Netw. ISDN Syst., 28(7-11):1255–1268, 1996.
optimal monitoring/housekeeping periods as these are the
[16] I. Jacobs and N. Walsh. Architecture of the World
key parameters for achieving good accuracy with our tool.
Wide Web, volume one. Technical report, W3C,
Further, we plan to analyze the applicability of DSNotify to
December 2004. Retrieved May 7, 2009.
other domains like the (document) Web or the file system.
[17] F. Kappe. A scalable architecture for maintaining
referential integrity in distributed information
7. ACKNOWLEDGMENTS systems. Journal of Universal Computer Science,
The authors would like to thank Martin Romauch for his 1(2):84–104, 1995.
valuable comments. The work has partly been supported [18] S. Lawrence, D. M. Pennock, G. W. Flake, R. Krovetz,
by the European Commission as part of the eContentplus F. M. Coetzee, E. Glover, F. A. Nielsen, A. Kruger,
program (EuropeanaConnect). and C. L. Giles. Persistence of web references in
scientific research. Computer, 34(2):26–31, 2001.
8. REFERENCES [19] A. Morishima, A. Nakamizo, T. Iida, S. Sugimoto, and
[1] W. Y. Arms. Uniform resource names: handles, purls, H. Kitagawa. Bringing your dead links back to life: a
and digital object identifiers. Commun. ACM, comprehensive approach and lessons learned. In HT
44(5):68, 2001. ’09: Proceedings of the 20th ACM conference on
[2] H. Ashman. Electronic document addressing: dealing Hypertext and hypermedia, pages 15–24, New York,
with change. ACM Comput. Surv., 32(3), 2000. NY, USA, 2009. ACM.
[3] S. Auer, S. Dietzold, J. Lehmann, S. Hellmann, and [20] T. A. Phelps and R. Wilensky. Robust hyperlinks cost
D. Aum¨ller. Triplify: light-weight linked data
u just five words each. Technical Report
publication from relational databases. In WWW ’09, UCB/CSD-00-1091, EECS Department, University of
New York, NY, USA, 2009. ACM. California, Berkeley, 2000.
[4] M. Beynon and A. Flegg. Hypertext request integrity [21] D. S. H. Rosenthal and V. Reich. Permanent web
and user experience, 2004. US Patent 0267726A1. publishing. In ATEC ’00: Proceedings of the annual
[5] M. Beynon and A. Flegg. Guaranteeing hypertext link conference on USENIX Annual Technical Conference,
integrity, 2007. US Patent 7290131 B2. pages 129–140. USENIX Association, 2000.
[6] C. Bizer, T. Heath, and T. Berners-Lee. Linked data - [22] H. Van de Sompel, M. L. Nelson, R. Sanderson, L. L.
the story so far. International Journal on Semantic Balakireva, S. Ainsworth, and H. Shankar. Memento:
Web and Information Systems (IJSWIS), 5(3), 2009. Time travel for the web. Arxiv preprint.,
[7] C. Bizer, J. Lehmann, G. Kobilarov, S. Auer, arxiv:0911.1112, November 2009.
C. Becker, R. Cyganiak, and S. Hellmann. DBpedia - [23] L. Veiga and P. Ferreira. Repweb: replicated web with
a crystallization point for the web of data. Web referential integrity. In SAC ’03: Proceedings of the
Semantics: Science, Services and Agents on the World 2003 ACM symposium on Applied computing, pages
Wide Web, July 2009. 1206–1211, New York, NY, USA, 2003. ACM.
[8] H. C. Davis. Referential integrity of links in open [24] J. Volz, C. Bizer, M. Gaedke, and G. Kobilarov.
hypermedia systems. In Proceedings of the 9th ACM Discovering and maintaining links on the web of data.
conference on Hypertext and hypermedia, pages In 8th International Semantic Web Conference, 2009.
207–216, New York, NY, USA, 1998. ACM. [25] W. E. Winkler. Overview of record linkage and
[9] A. K. Elmagarmid, P. G. Ipeirotis, and V. S. Verykios. current research directions. Technical report, U.S.
Duplicate record detection: A survey. IEEE Trans. on Bureau of the Census, 2006.
Knowl. and Data Eng., 19(1):1–16, 2007.
[10] I. P. Fellegi and A. B. Sunter. A theory for record
linkage. Journal of the American Statistical
Association, 64(328):1183–1210, 1969.
[11] A. Ferrara, D. Lorusso, S. Montanelli, and G. Varese.
Towards a benchmark for instance matching. In
Ontology Matching (OM 2008), volume 431 of CEUR
Workshop Proceedings. CEUR-WS.org, 2008.

090626cc tech-summit

More Related Content

What's hot (12)

Viewers also liked (9)

Similar to 090626cc tech-summit (20)

More from EuropeanaConnect (10)

Recently uploaded (20)

090626cc tech-summit