=Paper=
{{Paper
|id=None
|storemode=property
|title=An Approach for the Semantic Contextualization of the Web Advertisement Process
|pdfUrl=https://ceur-ws.org/Vol-682/paper7.pdf
|volume=Vol-682
|dblpUrl=https://dblp.org/rec/conf/esws/StojanovicS10
}}
==An Approach for the Semantic Contextualization of the Web Advertisement Process==
https://ceur-ws.org/Vol-682/paper7.pdf
An approach for the semantic contextualization of the web advertisement
process
Ljiljana Stojanovic Roland Stuehmer
FZI Karlsruhe FZI Karlsruhe
Haid-und-Neu Strasse 10-14 Haid-und-Neu Strasse 10-14
76131 Karlsruhe 76131 Karlsruhe
+49 721 9654804 +49 721 9654872
Ljijana.Stojanovic@fzi.de stuehmer@fzi.de
ABSTRACT communication, the user’s behavior cannot be captured in the
The modern advertisement theory is based on the “contextual real-time (on the client side) easily and is therefore omitted from
priming effects”: the product attributes primed by the ad the traditional web advertisement process.
context may result in the formation or change of beliefs about
Modern web technologies are enabling more client-side control of
the advertised brand, thereby affecting consumers' evaluations
of the brand. Therefore, a web ad should be tailored as much as the user’s behavior and there is already work done in developing
possible to the user’s current context (interests) in order to technologies for gathering a web user’s behavior while browsing
affect the user’ attention appropriately. AJAX-based web pages [1].
In this paper we present an approach for the semantic-based
In this paper we leverage on that work in developing an approach
personalized advertising on the web.
for the dynamic and personalized web advertisement. In the
Keywords nutshell of the approach is the real-time and complex processing
Personalized advertisement, complex event processing, of the user’s behavior in a web page.
semantics
In fact, the user’s interaction with a web page is interpreted as a
set of events, which are combined in order to discover the “very
current” interest of the user. Events, simple or complex, are
1. INTRODUCTION models for things that happen e.g., when a user interacts with a
A contextual web advertising system scans the text of a website Web page. Events are consumed in some meaningful way e.g., for
for keywords and returns advertisements to the webpage based on monitoring reasons or to trigger actions such as responses.
what the user is viewing. Contextual advertising has made a Semantics is used for a better interpretation of the user’s behavior
major impact on earnings of many websites. Because the by taking into account the meta information assigned to parts of
advertisements are more targeted, they are more likely to be the web page, which the user has visited. The user’s
clicked, thus generating revenue for the owner of the website (and interest/profile generated in this way is used for the very
the server of the advertisement). However, despite being targeted, personalized ad generation.
current approaches for contextual web advertising are not
Additionally we define a model for updating ads once the current
personalized, i.e. they are not taking into account the user, but
user’s interest has been changed such that displayed ads are not
only the characteristics of the web site. On the other hand, the
any more the most relevant one. In that way our approach
modern advertisement theory is based on the “contextual priming
supports dynamic adaptation of ads ensuring their high relevance
effects”: the product attributes primed by the ad context may
for the user.
result in the formation or change of beliefs about the advertised
brand, thereby affecting consumers' evaluations of the brand. In this paper we present the whole approach for the personalized
Therefore, an ad should be tailored as much as possible to the and dynamic web advertisement, including the technical
user’s interests in order to affect the user appropriately. architecture for detecting and composing (semantic) events in
Consequently, this implies a need for real-time tracking a web Web clients, that is, as explained above, the basic mechanism for
user’s behavior in order to detect her/his current interests, by discovering and updating real-time profile of a web user (i.e.
assuming that her/his current interest will correlate to the visited her/his interests). Additionally we demonstrate the validity of the
elements in a web page. Moreover, due to the different contexts approach in two evaluation studies.
that can be found in a web page, such personalized ads should be
The paper is structured as follows: In Section 2 we describe
dynamically changed, according to the changes in the user’s
methods for tracking a user’s behavior in a web page (as a set of
interests. However, due to the request/response style of web
semantically enriched events), whereas in Section 3 the
methods for complex processing of these events are given
as an approach for discovering the current interest of the
42
web user. Section 4 presents the architecture for generating page which can be dynamically filled by an ad provider as a
personalized and dynamic web ads based on detecting response to an event the client sends. In our approach ad content
“unusual” user’s behavior, whereas Section 5 contains is created based on a current user’s attention. In order to
some implementation details and in Section 6 we present some accomplish this we need as much (meta-) information as possible
evaluation results. We will discuss related work and conclude the about the content of the Web page. Therefore, we assume
paper in the last remaining sections. semantically enriched Web content such that context extraction is
easier and more precise. Additionally, every page is split up in a
number of Semantic Web Widgets (SWW). We introduce
2. TRACKING A WEB USER’S Semantic Web Widgets as self-contained components annotated
with semantic data and displayed in a Web page. Semantic Web
BEHAVIOUR Widgets give a high-level description of the content, and provide
The main issue in enabling personalization of the web usage is to the basic context of data contained in the widgets. For instance on
enable capturing of actions or changes in Web documents. These a news portal incorporating semantic advertising one widget
can be treated as events, which an event-driven system will react could be used for listing all news belonging to one subcategory,
to. For our use case of advertising we will focus on events created e.g., politics, another one for arts, etc. In Figure 1 we show an
from a user’s interaction with Web documents. After having RDFa example of the semantic description for an arts event listed
extracted events from a Web document, they must be processed in a widget related to musicals. The code snippet presents an
in order to interpret them semantically, to be able to react on them event named “Mary Poppins Show” described using RDF
appropriately. The following two subsections describe our Schemata for Dublin Core, vCard and iCal vocabularies.
approach for these two issues: generation and processing of Web Information such as categories, start and duration of the musical
events. are provided together with contact information, location and so
on.
2.1 Simple Event generation
A simple event in Web clients is characterized by two
dimensions; the type of event (e.g. click, mouseover) and the part
of the Web page, where the event occurred (e.g. a node of the
Document Object Model of the Web document). This node is,
however, just a syntactical artifact of the document as it is
presented in a Web browser. Adding this node or parts of it to the
event body will not significantly add meaning to the event and
not ease the understanding of the event for the recipient of the
event.
We therefore propose to add semantic information to the event
which pertains to the actual domain knowledge that the Web page
is about. In order to enable this, the first step is to represent the
content of a Web page in a form that can be used for generating
meaningful events. To do so without having to manually annotate
every Web document, we envision a mechanism, which ensures
the relevance of the annotations. This can be done in many (semi-
Figure 1. An example for a musical listed in a Semantic Web
) automatic ways, e.g. by providing Web forms (page templates),
Widget.
which for a given user’s input, automatically adds the proper
semantic relationships between the form fields. In this way all 2.2 Event enrichment
user generated content will be annotated. The Web forms are
created based on supported vocabularies for a particular Web site. In this subsection we focus on enriching simple events with
Our particular focus is on widely spread vocabularies such as semantics from the context of the Web page in which the event
Dublin Core, Creative Commons, FOAF, GeoRSS and occurred.
OpenCalais. Regarding the format of structured data, RDFa [2], A simple event in Web clients is characterized by two
eRDF and Microformats are all good candidates for this purpose. dimensions; the type of event (e.g. click, mouseover) and the part
They support semantics embedded within actual Web page data of the Web page, where the event occurred (e.g. a node in the
and allow reusable semantic markup inside of Web pages. In our Document Object Model of the Web document). Subscribing to
implementation we use RDFa, since in comparison to eRDF it is a simple events of these types therefore requires the specification of
more encouraged candidate by the W3C. Comparing it further to type and the specification of the node or nodes where the events
Microformats, RDFa is more flexible in mixing different existing may originate. Both dimensions are retained in an event instance
vocabularies. by using the attributes jsEventType and cssSelector (see Figure 2
In the remaining part of this section we give an example for more explanations).
demonstrating the generation of events in the context of a
Semantic Advertising scenario. The ad space is a part of the Web
43
In order to better understand these events and make sense of what these rules are created they are pulled by the next client request
happened we must enrich the content of events when they are and loaded into the rule engine for the run time.
produced. The jsEventType tells us what a user has done and the
For the run time we have developed a client-side event-condition-
cssSelector tells us where on the Web page the user did it.
action (ECA) rule engine. It uses a lightweight rule language
However, the latter is a purely presentation-dependent measure.
which supports ECA rules described in more detail in [3].
There is no semantics which has any meaning beyond the
context of a specific Web page structure. We propose to extract Very briefly, we JSON-Rules, our client-side rule language that
presentation-independent semantic information from the Web resembles a lightweight reaction rule language tailored to the
page if present. Instead of creating events from interaction with needs of Rich Internet Applications, specifically applications that
purely syntactic items of a Web document, we create events about profit from or require Complex Event Processing, condition
interaction with semantic concepts which the document stands evaluation on a working memory, and running rule actions
for. As an example, an event should not represent e.g., a click on written in JavaScript. As a representation for our rules we use
a certain headline element of a Web document but rather a user’s JSON, because it is natively usable within JavaScript. JSON can
interaction with an article talking about politics and certain specify objects, arrays and primitives. Rule objects in our JSON-
persons mentioned within. Rules language contain the three attributes event, condition and
action. The event part consists of patterns in the event pattern
To annotate a Web page with semantic data such as the topics of
language Snoop [4]. Snoop contains a fairly comprehensive list of
an article, we use RDFa. Defined in [2] RDFa is a means of
Boolean and temporal operators. They are modeled in our
adding RDF data to existing Web pages by using inline XHTML
ontology. What is missing in Snoop are operators which inspect
attributes.
the contents of input events such as attributes other than
After detecting an event which happened in the context of a timestamps and type. Therefore, we added a FilterEvent as an
certain DOM node of a Web document, we collect all semantic example of what is needed to filter events by their content.
information in the Web page about the thing that is reported in
The condition part consists of conjunctive predicates over
that given DOM node. We currently achieve this by employing
variables from a working memory. The action part in turn
the client-side RDFa library ubiquity (http://ubiquity-
contains one or more JavaScript code blocks to gain a maximum
rdfa.googlecode.com/). The lifting of context is achieved in a
degree of versatility for the rule author. Alternatively for rule
two-phase process. In the first phase we collect the list of RDF
actions we offer to trigger certain desired events as well as
subjects of possible triples. This is done close to where the event
manipulations of the working memory. The latter types of action
happened in the document to provide accurate context. In the
offer greater declarativity while formulating rules. This increase
second phase we collect every triple with these subjects from the
is, however, bought at the price of some flexibility. Thus, we still
overall document in order to provide a very rich context. To find
offer all three kinds of rule actions which can be freely mixed.
valid subjects the first phase traverses the node where the event
happened and its complete subtree. If the given main node does Figure 2. Example of a single Rule
not contain a subject, the immediate dominator node containing a
subject is added to the list. This serves two purposes,
guaranteeing a single root subject for orphan properties and
objects in the subtree and guaranteeing a non-empty result set.
In the second phase all triples with the given subjects are
collected from the entire document tree. The gathered triples are
then reified and appended as a bag to the event payload. Even if
the event itself becomes part of more complex events during the
process of correlating and aggregating events, this basic data is
retained as part of the simple event.
3. UNDERSTANDING THE USER’S
INTEREST – COMPLEX EVENT
The rules on the client serve to detect users exhibiting
PROCESSING
interesting behavior as learned from the average usage patterns.
Simple events extracted from Web documents must be combined The user causes events to occur by interacting with the Web page,
in order to detect complex situations which might be interpreted detected by the event processor and rule engine. Rules are
as a user’s interest. This is the task of Complex Event Processing. triggered which create intermediate events in a hierarchy of event
Detecting the behavior of Web users according to our proposal is abstraction. These events are subsequently accumulated until
divided into design time and run time. The design time consists of sufficient interest according to the ad provider is recorded
(i) semantically enhancing the Web page and then (ii) recording (threshold achieved) and actions can be taken by further rules.
average viewing statistics of the annotated elements, e.g. from log
The distinction between run time and design time in this
files. From the statistical data we generate client-side rules. Once
section is not a strict temporal distinction as the names would
44
suggest. Rather, because new users will inevitably alter our This process can be done automatically. A simple sequence along
knowledge of what is interesting there is a loop in the process, with its confidence might be “politics” followed by “flowers"
feeding back from the run time into the design time to evolve new with a low confidence of 2%. This means that from previous
rules for future users. users only a fraction of 2% have looked at a politics widget
followed by looking at a flowers widget.
Figure 2 shows an example rule. It can be automatically
created from analyzing histories of interesting behavior. The only This pattern in the users behavior can be treated as unusual, i.e.
his/her interests for "politics" and "flowers" are distinguished
requirement is knowledge, that e.g. states that only two percent of
from the interest of others, so that this can be used for developing
users look at a politics item followed by a science item. The
a very personalized ad. In fact, we argue that more information
actual rule consists of an event part starting at line 5 and an action content (for generating ads) is stored in the exceptional behavior,
part starting at line 20. The rule resembles an event-condition- than in the usual/expected one. A simple explanation is that
action rule where the condition is left blank, i.e. is always true. expected behavior is too general to detect what is specific in the
behavior of the customers (cf. example from the brick and mortar
environment from Introduction). Once when enough “unusual
4. GENERATING THE PERSONALIZED behavior” is accounted for a user a new ad should be issued.
ADS DYNAMICALLY Such an ad will very likely attract the attention of the user, since
it directly corresponds to his short-term profile. Further
Figure 3 shows a rough architecture of our approach: Part b) on processing of e.g. the time interval within the two participating
the right hand side of the figure depicts the components of our events could be envisioned.
client-side rule engine. Multiple event sources provide input for
Each complex event expression is embedded in an event-
the event detection, creating complex events. Also, a working
condition-action rule with the probability as the consequence. The
memory submits its changes to a Rete network, evaluating rule
consequence forms another event which is processed further by
conditions. The logic for both the event detection and condition higher-level rules.
evaluation is supplied by rules from a repository, generated from
past user activities. Part a) on the left hand side places the
client-side components above the protocol boundary dividing
client and server. Below on the server or several distributed
servers hold the Web content as well as the advertising content.
The Web content is annotated, providing semantic relations to
the advertisements. Short-term user models provide a temporal
model of how a user interacts with the Web content. The ad
provider analyses user models to provide up-to-date and
personalized advertisements.
On the other hand we anticipate annotations to be mostly used
on elements at, or not far below the level of single widgets or
paragraphs. Reasons for this are of practical nature, in keeping
the number of events manageable. Handling too much detail
might have further adverse effects at this point, creating a large Figure 3: Architecture: a) Logical Architecture b) Client-side
number of event types which are almost never used (created or User Behavior Analysis.
consumed). There might, for example, be no measurable
interaction of the user with a certain word in a Web page, In order to enable such a processing, we extended the set of
whereas the surrounding paragraph might encounter detectable traditional event processing operators with two additional ones
mouse clicks or mouse hovering/movement. Filter(E1; condition) and Thres(E1; threshold) as follows:
As mentioned in the introduction, web ads should be continuously Filter is modeled like event masks. The Filter enforces a
updated to the web user’s interests, which implies a need for the condition on each occurrence of event E1. This allows e.g. for
automatic triggering of a new ad, once that user’s interests has fine-grained content-based filtering/masking of events.
been dramatically changed. In this work we use the notion of Thres is another content-based operator which we need to extend
unusual behavior as the criteria for generating a new ad. In the the Snoop algebra with. Thres(E1; threshold) accumulates the
following we describe that principle shortly. events of type E1 until the boolean function “threshold” returns
In order to form complex event expressions, the RFDa true, releasing all accumulated events as a complex event and
annotations are combined with a temporal model. Such starting accumulation anew.
expressions group the user's atomic actions into temporal contexts
like e.g. sequences of clicks. Determining sequences of interest is
based on analyzing historical (log) data statistically. By using 5. IMPLEMENTATION: CLIENT-SIDE
data mining algorithms for click streams such as [5], historical
data is transformed into knowledge about unusual sequences of EVENT-ENABLED RULE ENGINE
interaction such as clicks. Subsequently, the corresponding For our implementation we chose JavaScript from the available
complex event expressions can be created. Web programming languages, for reasons of widespread
45
availability. The data structures and program logic we overall page. The results are very encouraging: in the average
implemented are roughly divided into the following areas: 85% of keywords generated in our approach were described as
adapters for the rule language and remote event sources, the “very relevant” and 98% as “relevant” (very similar results across
working memory, condition representation and evaluation as well all three domains).
as complex event detection.
The traditional approach achieved 65% success for “very
For Complex Event Processing we are using a graph based relevant” and 85% success for “relevant” ad-keywords. This
approach as proposed in [4]. Initially the graph is a tree with result demonstrates the advantages of our approach for generating
nested complex events being parents of their less deeply nested very relevant ads.
sub-events, down to the leaves being simple events. However,
In comparison, Web Usage Mining (e.g., [5]) is used on log files
common subtrees may be shared by more than one parent. This
which are analyzed on the server side at certain intervals or
saves space and time compared to detecting the same sub-events
possibly in a continuous fashion. It is important, however, to
multiple times, and renders the former tree a directed acyclic
stress that our approach detected all events on the client. Events
graph.
occurred purely by folding and unfolding widgets as parts of the
When using the term event, the distinction must be drawn page. No communication with the server took place and hence no
between event occurrences (i.e. instances) and event types, artifacts are visible in server log files. Thus, our approach extends
usually done implicitly. In the detection graph the nodes are event clickstream analysis to regions which were previously invisible to
types, they exist before there are any instances. Event instances server-based mining techniques.
exist after simple instances arrive and are fed into the graph at the
Moreover, our approach is a truly event-driven application,
leaves. Complex instances are then formed at the parent nodes,
meaning that we detect events in real-time, as soon as they
which in turn propagate their results upwards. Every complex
happen. In contrast, traditional mining techniques function in a
event occurrence carries pointers to the set of its constituent event
query-driven manner where results are only created at intervals,
occurrences, so that the events and their parameters can be
such as daily analyses of the log files.
accessed later. Once an occurrence is computed at a node which
is attached to a rule, the state of the associated Rete node is
started and actions are triggered.
7. RELATED WORK
In Web advertising there are essentially two main approaches,
6. EVALUATION contextual advertising and behavioral advertising. Contextual
advertising [6] is driven by the user’s context, represented usually
To evaluate the return of targeted advertisements we created a
in the form of keywords that are extracted from the Web page
demo Web page with some news articles. Each news article is
content, are related to the user’s geographical location, time and
contained in a separate part of the page, termed Semantic Web
other contextual factors. An ad provider (ad serving service)
Widget (cf. Section 2.1). Each widget is annotated using RDFa
utilizes these meta data to deliver relevant ads. Similarly, a users’
using basic keywords and concepts pertaining to the article. For a
search words can also be used to deliver related advertisement in
user entering our demo, each widget is at first partially concealed.
search engine results page, Google’s second pillar in online
This is done to solicit an action from the user when “unfolding”
advertising. However, contextual advertising, although exploited
the widget. Thereby the user expresses interest. This creates
today by major advertising players (e.g., GoogleAdsense, Yahoo!
explicit events which can then be processed by our engine. Our
Publisher Network, Microsoft adCenter, Ad-in-Motion etc.),
initial evaluation of the ad quality was performed as follows:
shows serious weaknesses. Very often the automatically detected
1. We selected three different news domains (politics, culture, context is wrong, and hence ads delivered within that context are
sports) in order to prove the domain-independence of the irrelevant. For instance, a banner ad offering a travel deal to
approach and pull into the demo Web page, as separate evaluation Florida can possibly be seen side-by-side to a story of a tornado
sessions. tearing through Florida. This is happening because the context
was determined using purely keywords such as “Florida, “shore”
2. We selected five users (PhD students from the Institute)
etc (i.e., without taking keyword semantics into account). While
with different cultural backgrounds.
there are improvements in contextual advertising (e.g., language-
3. The users should browse the demo Web page and judge independent proximity pattern matching algorithm [7]), this
about the relevance of generated ad-keywords in the case of a) the approach still often leads companies to investments that are
keywords generated statistically from the Web page (Google wasting their advertising budgets, brand promotion and
approach) and b) keywords generated by using the event-driven sentiment. In contrast, our approach utilizes semantics to cure
approach described in this paper. In order to ensure a fair major drawbacks of today’s contextual advertising. Semantic
comparison, the users did not know which list of ad-keywords Web technologies can be used to improve analysis of the meaning
was produced by which method. of a Web page, and accordingly to ensure that the Web page
contains the most appropriate advertising.
We ask the users to rate the gathered keywords in terms of
relevance to what they had been doing in the news portal and to The second approach to Web advertising is based on the user’s
compare this with a static list of keywords extracted from the behavior, collected through the user’s Web browsing history (i.e.,
46
behavioral targeted advertising). The behavior model for each connecting the Internet of Things with the Internet of Services,
user is established by a persistent cookie. For example, Web sites two basic elements of the Future Internet.
for online shopping utilize cookies to record the user’s past
activities and thereby gain knowledge about the user or a cluster
of users. There are several reasons why behavioral targeted 9. REFERENCES
advertisement via cookies is not a definitive answer to all
[1] Schmidt, K.U., Stojanovic, L., Stojanovic, N., Thomas, S.:
advertisement problems. First, if a user, after browsing the
On enriching ajax with semantics: The web personalization
information about an item purchases that item, he or she will not use case. In: Proceedings of the EuropeanSemanticWeb
be interested in that particular good afterwards. Therefore, all ads Conference, ESWC2007. Volume 4519 of Lecture Notes in
and “special deals” offered to the user later while browsing that Computer Science., Springer-Verlag (July 2007)
Web site are useless. Also, the short-term user interest should be
detected more quickly (i.e., during the current user session).
[2] Adida, B., Birbeck, M., McCarron, S., Pemberton, S.: Rdfa in
Displayed ads need to reflect current moods or transient user
xhtml: Syntax and processing. Online Resource.
interest. For example, a user looking hastily to buy a gift of
http://www.w3.org/TR/rdfa-syntax/ (October 2008)
flowers is not interested in ads related to his/her long-term
profile, created during previous purchases unrelated good or
services. Further on, there are problems with cookies. Computers [3] Schmidt, K.U., Stühmer, R., Stojanovic, L.: From business
are sometimes shared and users get to see ads governed by other rules to application rules in rich internet applications.
user’s cookies. Finally, given the European Union’s Directive and Scalable Computing: Practice and Experience 9(4) (2008)
US legislation concerned with restricted use of cookies, 329–340
behavioral targeted advertisement based on cookies is not a
promising direction for Web advertising. [4] Chakravarthy, S., Krishnaprasad, V., Anwar, E., Kim, S.K.:
Composite events for active databases: Semantics, contexts
We believe that short-term profiling (in contrast to long-term
and detection. In Bocca, J.B., Jarke, M., Zaniolo, C., eds.:
profiles created by cookies) is a valid and possibly augmenting
20th International Conference on Very Large Data Bases,
approach in terms of personalization and identification of the
September 12–15, 1994, Santiago, Chile proceedings, Los
user’s interest. We realize a short-term profiling using client-side
Altos, CA 94022, USA, Morgan Kaufmann Publishers
Complex Event Processing techniques (cf. Section 2.2), and
(1994) 606–617
background semantics (cf. Section 2). Such profiles are
automatically detected, are always up-to-date and fully
personalized. [5] Liu, B.: Web Data Mining. Data-Centric Systems and
Applications. Springer Berlin Heidelberg (2007)
The work from [8] describes event processing for Web clients.
Events are observed on the client; however, complex events are
[6] Kenny, D., Marshall, J.: Contextual marketing–the real
not detected in the client. All simple events are propagated to the business of the Internet. Harvard Business Review 78(6)
server for detection of patterns. This incurs latency and reduced (2000) 119–25
locality for the processing of events, so the advantages of client-
side event processing are lost.
[7] Schonfeld, E.: Proximic signs deals with yahoo and ebay to
turn product listings into contextual ads; taking on adsense.
Online Article.
8. CONCLUSION http://www.techcrunch.com/2008/01/15/proximic-signs-
deals-with-yahoo-andebay-to-turn-product-listings-into-
In this paper we present a novel approach for the personalized contextual-ads-taking-on-adsense/ (January 2008) Last
and dynamic ad delivery on the web. The approach is based on visited: August 2009.
the complex processing of the semantically enriched events
generated out of the user’s interaction with web content.
[8] Carughi, G.T., Comai, S., Bozzon, A., Fraternali, P.:
Additionally, the approach introduce the notion of “unusual Modeling distributed events in data-intensive rich internet
behavior” as a criteria for determining the dynamics of the new applications. In Benatallah, B., Casati, F., Georgakopoulos,
ads delivery for a particular user. This work goes beyond the D., Bartolini, C., Sadiq, W., Godart, C., eds.: WISE. Volume
web advertisement use case – in fact it opens possibilities to build 4831 of Lecture Notes in Computer Science., Springer
event-driven applications for the (Semantic) Web. We envision (2007) 593–602
the future of the (Semantic) Web as a huge, decentralized event
repository (the so-called Event Cloud in Event processing
terminology), which will contain information about the real-time
activities of different Web users. Such an event repository will
enable different kinds of processing of the real-time information,
making the Semantic Web really active, i.e. the environment can
react and adapt itself on the signals sensed from the environment,
47