Specification of the Linked Media Layer

Deliverable D2.5 Speciﬁcation of the Linked Media Layer
Jan Bouchner, Lukáš Pikora, Barbora ˇCervenková, Tomáš Kliegr, Jaroslav Kuchaˇr, Ivo Lašek/ UEP
Mathilde Sahuguet, Benoit Huet, José-Luis Redondo Garcia, Raphaël Troncy/ EURECOM
Jan Thomsen, Ali Sarioglu / CONDAT
Lyndon Nixon / STI
Evlampios Apostolidis, Vasileios Mezaris / CERTH
Daniel Stein, Stefan Eickeler / Fraunhofer IAIS
11/10/2013
Work Package 2: Linking hypervideo to Web content
LinkedTV
Television Linked To The Web
Integrated Project (IP)
FP7-ICT-2011-7. Information and Communication Technologies
Grant Agreement Number 287911

Specification of the Linked Media Layer D2.5
Dissemination level PU
Contractual date of delivery 30/09/2013
Actual date of delivery 11/10/2013
Deliverable number D2.5
Deliverable name Specification of the Linked Media Layer
File D2.5_Specification*of*the*Linked*Media*Layer.tex
Nature Report
Status & version Reviewed & v1.0
Number of pages 51
WP contributing to the deliver-
able
2
Task responsible UEP
Other contributors EURECOM, CONDAT, STI, CERTH, FRAUNHOFER IAIS
Author(s) Jan Bouchner, Lukáš Pikora, Barbora ˇCervenková, Tomáš Kliegr,
Jaroslav Kuchaˇr, Ivo Lašek/ UEP
Mathilde Sahuguet, Benoit Huet, José-Luis Redondo Garcia,
Raphaël Troncy/ EURECOM
Jan Thomsen, Ali Sarioglu / CONDAT
Lyndon Nixon / STI
Evlampios Apostolidis, Vasileios Mezaris / CERTH
Daniel Stein, Stefan Eickeler / Fraunhofer IAIS
Reviewer Dorothea Tsatsou
EC Project Officer Thomas Kuepper
Keywords Video Enrichment, Linked Media, Crawling, Information Retrieval
Abstract (for dissemination) This deliverable describes the process of providing a set of prospec-
tive enrichment resources for seed video content in the LinkedTV
project. Additionally, we participate in the Search and Hyperlinking
task of the MediaEval 2013 benchmarking initiative and we report
the approaches and results we obtained.
© LinkedTV Consortium, 2013 2/51

History
Table 1: History of the document
Date Version Name Comment
09/07/2013 v0.1 Kliegr, UEP deliverable structure
22/08/2013 v0.2 Kliegr, UEP unstructured search chapter
27/08/2013 v0.3 Sahuguet, EURECOM MediaCollector chapter
03/09/2013 v0.4 Thomsen, CONDAT Linked Media Architecture chapter
19/09/2013 v0.5 Sahuguet, EURECOM MediaEval chapter
26/09/2013 v0.6 Nixon, STI LSI chapter and introduction
26/09/2013 v0.65 Kliegr, UEP revise introduction
27/09/2013 v0.7 Kliegr, UEP add statistics of the index
08/10/2013 v0.8 Kliegr, UEP address QA comments
09/10/2013 v0.9 Troncy, EURECOM complete MediaCollector chapter
09/10/2013 v0.95 Troncy, EURECOM general proof read
10/10/2013 v0.97 Troncy, EURECOM address second QA comments
11/10/2013 v1.0 Kliegr, UEP address comments of the scientiﬁc coordinator

0 Table of Content
0 Table of Content 4
1 Introduction 6
2 Linked Media Layer Architecture 8
2.1 Linked Media Workﬂow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Linked Media Layer Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Repository Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.2 Media Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.3 Metadata Repository . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.4 Integration Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.5 Service Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.6 Administration REST API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.7 RDF Repository REST API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Unstructured Search Module 13
3.1 Frameworks Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.1 Apache Nutch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.2 Apache Solr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.3 Apache HBase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Workﬂow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.1 White-list . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.2 Crawling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.2.1 Inject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2.2 Generate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2.3 Fetch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2.4 Parse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2.5 Updatedb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2.6 Solrindex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.3 Refreshing index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Developed Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1 Multimedia Indexing Plugin for Nutch . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.2 Rendering-Based Multimedia Extraction - HtmlUnit and Dom managers . . . . . . 16
3.3.3 Nutch Annnotated Crawling List Extension . . . . . . . . . . . . . . . . . . . . . . . 18
3.3.4 Nutch-Solr bridge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4 Index structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5 API Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5.1 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5.2 API response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.6 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6.1 Sample Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6.2 RBB scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6.3 S&V scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.6.4 Index statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.7 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.7.1 Dashboard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.7.2 Metadata Extraction Service . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.7.3 Entity enrichment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4 Media Collector 29
4.1 Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3 Query and Response Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.4.1 RBB scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.4.2 Sound & Vision scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5 Searching Media Resources with LSI 35
5.1 Technology used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2 Use of the REST API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2.1 Find Media Resources. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.3 Use in LinkedTV Media Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6 MediaEval task 39
6.1 The challenge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.1 Search task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.2 Hyperlinking task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.2 Pre-processing step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.2.1 Concepts Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.2.2 Scene Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2.3 Keywords Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2.4 Optical Character Recognition (OCR) . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.2.5 Face Detection and Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.2.6 Named Entities Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3 From visual cues to detected concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.4 Lucene indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.4.1 Video index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.4.2 Scene index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.4.3 Shot index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.4.4 Speech segments index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.4.5 Sliding windows index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.5 Search task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.6 Hyperlinking task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.6.1 Hyperlinking using the search component . . . . . . . . . . . . . . . . . . . . . . . 45
6.6.2 MoreLikeThis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.7.1 Search task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.7.2 Hyperlinking task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.7.3 Some ﬁgures on visual concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.8 Lessons learned and future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7 Conclusion 48

1 Introduction
The vision of “Linked Media” in the LinkedTV project is a vision of a potentially Web scale layer of
structured and semantic metadata about media items, connected to the Web of Data, so that agents
can determine significant links between different online media resources.
Today’s Web is increasingly non-textual in content, yet Web systems are still widely tuned to the pro-
cessing of textual information from Web pages, even when considering the widely used search engines
or in fact the hyperlinking mechanism that so fundamentally underlies the concept of the Web. Integrat-
ing all of this, rapidly expanding in scale, non-textual content into the existing Web infrastructure is a
significant challenge which is still being addressed in work on media analysis, annotation, management,
search and retrieval.
As Television and the Web converge, this disjunction between the traditional end user consumption of
audiovisual streams and the interaction with online media resources within web sites becomes very clear.
LinkedTV aims to seamlessly integrate content from both sources into a single interactive experience,
which requires that cross-network media shares a common description and an agreed means to access
and process that description, and that a LinkedTV system is able to compare media item descriptions
in order to generate links between media that can be used in enriching a TV program or enabling end
users to browse within a collection of media items from different sources.
We have proposed a set of principles to underlie Linked Media [19]:
1. Web media descriptions need a common representation of media structure
2. Web media descriptions need a common representation of media content
3. Web media descriptions need to use a media ontology which supports description of both the
structure and content of media
4. The descriptions of media in terms of common representations of structure and content are the
basis for deriving links across media on the Web
In the same presentation, a set of specifications are proposed that can fulfil the requirements of each
principle1:
1. Media structure can be represented using W3C Media Fragment URIs
2. Media content can be represented using the Linked Data URI concept space
3. Media descriptions can be based on the W3C Ontology for Media Resources, the W3C Open
Annotation Model (core) and some extensions
4. Media hyperlinking can be derived from the semantic analysis of media descriptions, e.g. concept
similarity or named entity detection
LinkedTV contributes to the vision of Linked Media, since the generation of links between media
resources is a fundamental part of the LinkedTV experience of enriched TV programming. It has taken
up the use of the specifications mentioned above, and proposes a new media ontology based on re-use
of existing media ontologies where appropriate and extended with the requisite additional information
needed for linking media (the LinkedTV Core ontology). Finally, it is researching how media hyperlinks
can be derived on the basis of these shared representations of media structure and content, part of
which is to generate relevant semantic descriptions of existing online media resources (whether from
online media platforms such as YouTube, social networks such as Twitter and Facebook, or white-listed
web sites from the scenarios themselves as specified in [2]) to permit their comparison with the seed
media fragments.
While previous deliverables have addressed the common representations of media structure and
content for LinkedTV and introduced the LinkedTV ontology, this deliverable focuses on the work done
to build upon this and provide Web services for exposing online media resources to the “Linked Media
layer”2 and systems for processing the media metadata in that layer, in order to derive representative
links to relevant and conceptually related media for enriching a seed media fragment.
1Slides available at http://www.slideshare.net/linkedtv/www-linked-media-keynote
2We define the “Linked Media Layer” as a decentralised web-based platform based on the distribution of web media annotations
and associated conceptual models over the web. It extends the Linked Data model of web-scale interlinking of concepts and
metadata about concepts, with media annotations which tie online media with Linked Data concepts. See also http://www.link
edtv.eu/development/linked-media-layer/.

According to the speciﬁcation provided in the preceding deliverable D2.3[1], a set of technologies for
retrieving the enrichment content from the web was developed and integrated via the LinkedTV Linked
Media Layer, which is covered in (Section 2). The two fundamental resources for obtaining this content
are regular web sites and a web service access to popular social networks and media sharing platforms
such as Flickr, Instagram, Twitter, Facebook, Google+ or YouTube. To exploit these resources, LinkedTV
developed a so-called Unstructured Search Module presented in Section 3 which crawls a list of web
sites provided by the LinkedTV content partners and creates a full-text index. This system was crafted
to include also multimedia content (images, podcasts, video) available on those web sites.
Social networks and media sharing platforms that expose web APIs are harnessed through the so-
called Media Collector described in Section 4 and the complementary Linked Services Infrastruc-
ture described in Section 5. Since supporting a new web site or a web API requires some amount of
customization and consumes computational and storage resources, the set of resources covered was
prioritized to include those relevant to the two LinkedTV scenarios – RBB and Sound & Vision. To this
end, the respective chapters include also illustrative examples from these domains. The scientiﬁcally
relevant algorithms developed for or already incorporated into the LinkedTV services have been put to
test within the MediaEval benchmarking initiative (Section 6). The span of this section overreaches WP2
as also the approaches to video and speech processing developed within WP1 are discussed. A brief
summary of the overall state of the WP2 integration activities is provided in the conclusion.

2 Linked Media Layer Architecture
The Linked Media Layer is an integrated collection of components, metadata and services for managing,
adding, updating and querying of enriched annotations of media resources and media fragments. This
chapter describes the architecture, which underlies the realization of the Linked Media Layer as the cen-
tral component of the LinkedTV Platform. The Deliverable D5.4 [5] having a “Confidential” dissemination
level, the overview presented in this chapter provides the minimal description between the Linked Media
Layer, the individual “enrichment services” and the rest of the LinkedTV workflow.
2.1 Linked Media Workflow
The Linked Media Layer has to support the Linked Media Workflow as defined and realized in LinkedTV.
In itself, it does not contain any components which generate data, but it rather provides the means to
store, access, search, link, update or manage that data. Figure 1 shows the basic Linked Media Work-
flow from ingestion, analysis, to serialization and enrichment. This workflow generates and aggregates
metadata upon which personalization can be performed. The playout filters and adapts the linked media
data to context and personal profiles. Each of these steps consists in itself of multiple sub-processes.
The Linked Media Layer also allows interaction with these sub-processes on an individual basis. A more
detailed description of the integration of these different process steps within the Linked Media Layer is
provided in the Deliverable D5.4 [5].
Figure 1: Linked Media Workflow
2.2 Linked Media Layer Architecture
The Linked Media Layer provides all services for accessing the persistency components such as storage
for video content, annotations, ontologies and user data as well as functions for retrieval, data access
and management. It also provides the central integration communication infrastructure. Thus, the Linked
Media Layer consists itself of three main sub-layers (Figure 1):
– Repository Layer: includes all persistency storage facilities for videos, metadata, ontologies,
personalization and context data, as well as for application specific data.
– Integration Layer: includes the central integration and communication component called the
Linked Media Service Bus.
– Service Layer: provides API services for retrieval, accessing and manipulating data stored in the
repositories.
All layers are conceptual layers but not physical layers. The different components can be (and actually
are) distributed over various different locations, hosted by different partners. They could also be hosted
by cloud services.
2.2.1 Repository Layer
The Repository Layer is usually located on the server side of the platform. However, some smaller data
sets, such as for the user profile or history can also be stored on the client side. For the Metadata
Repository, the ontologies, user and social data, an RDF repository (or Triple Store) is provided which
also offers other types of storage (SQL, XML, Documents, Free Text). The triple store used in LinkedTV
is OpenLink Virtuoso3.
3http://virtuoso.openlinksw.com/

Figure 2: Linked Media Layer
2.2.2 Media Storage
In general, the Linked Media Layer in LinkedTV does not require that media resources which are be-
ing analyzed and annotated are stored within the platform itself, but can rather be accessed either
through URIs over the Internet or by connecting to media archives (e.g. of a broadcaster). However, the
LinkedTV platform does support the storage of media resources which is needed in particular for the
analysis part.
2.2.3 Metadata Repository
The Metadata Repository is the central part of the LinkedTV platform, as it contains the metadata ag-
gregated through the different processes for analysis, fragment detection and annotation. At present, it
contains the following data:
– General Metadata: Metadata which applies to the base resource as a whole, such as title, dura-
tion, owner, author, format, genre, etc. The General Metadata is a subset of the EBUCore Meta-
data Set4. Table 2.2.3 lists the attributes used. This metadata is not in RDF format as it is collected
before the RDF-ization steps. In the LinkedTV platform, it is stored within an SQL Database, how-
ever NoSQL Databases would be also appropriate. Access to the general metadata is provided
through http://api.linkedtv.eu/mediaresource/.
– Annotated Media Fragments Metadata: The annotated and enriched Media Fragments are all
those data which are generated by the Metadata Conversion Service (tv2rdf) [4] and the Enrich-
ment Service (TVEnrichment) (Section 4 of this deliverable). They are provided in RDF format
and stored within the RDF Repository. Additionally, annotations can be added and managed via
a so-called Editor Tool [3]. The RDF Repository is accessible at http://data.linkedtv.eu while
the SPARQL interface is at http://data.linkedtv.eu/sparql.
2.2.4 Integration Layer
The Integration Layer provides the components and services for the integration of the different LinkedTV
components and the communication between them. The main purpose is to ensure an open, extendable,
distributed scalable integration backbone for LinkedTV in order to support the Linked Media Workflow
but also different workflows and business models. Since the analysis process, data provision and ap-
plications can be located at different physical places, it is also possible to flexibly integrate third party
content or service providers.
4http://tech.ebu.ch/docs/tech/tech3293v1_4.pdf

Attribute Description
Identifier The unique id (key) of the Media Resource (this is a UUID)
TitleName String. The title of the Media Resource (if not available, by default the file-
name)
Language URI | String
Locator The logical address at which the resource can be accessed (e.g. a URL or a
DVB URI).
DateInserted DateTime of import into the media repository
Collection The collection where the media resource belongs to (in our case the original
ftp directory)
Publisher URI, String (e.g. ’rbb’, http://www.rbb-online.de, ’dbpedia:rbb’)
FrameSizeWidth Decimal (e.g. 720)
FrameSizeHeight Decimal (e.g. 480)
Compression URI, String (e.g. ogg, mpeg4)
Format URI, String; MIME Type
Duration Decimal (in seconds)
hasRelations MediaResourceRelation. Handles the relations to other related resources, like
EXMARaLDA file, playout versions in different formats or metadata files
LinkedTVStatus Not part of Ontology for Media Resource, used to describe the workflow status
of this media resource in LinkedTV (LINKEDTV_STATUS)
Table 2: Object Model of the AdminDB of the LinkedTV Platform
The main component of the Integration Layer is the Linked Media Service Bus (LMSB). The Linked
Media Service Bus is based on general Enterprise Service Bus (ESB) technology and provides the
following features, to name just the most relevant ones in the Linked Media context:
– support of Enterprise Integration Patterns (EIP)5
– message routing
– publish/subscribe messaging
– automatic type conversion (custom converters could be added)
– numerous different types of endpoints (e.g. file system, ftp, email, RSS, REST services)
– logging and monitoring
The LinkedTV LMSB is realized on top of the Apache Camel6.
2.2.5 Service Layer
The Service Layer of the Linked Media Layer includes the general LinkedTV services which are provided
for the different components of the LinkedTV Platform for analysis, annotation, enrichment, or personal-
ization, as well as external clients such as the Editor Tool or the LinkedTV Player. The Service Layer is
completely realized through a set of REST services.
The two main repositories, the Administration Database and the RDF Repository, are exposed via
two different REST Base URLs at http://api.linkedtv.eu and http://data.linkedtv.eu. The first
one gives access only to complete media resources, whereas the second one gives access to the RDF
data of complete media resources as well as media fragments, annotations, and more. Both REST APIs
ensure that the LinkedTV Platform is an open platform which can be used by other clients as well as
integrated with other IPTV Environments.
2.2.6 Administration REST API
The base URL for the REST API URIs is http://api.linkedtv.eu. Table 3 lists the REST API functions
for getting and setting complete media resources. This also includes the management of files related to
a media resource, such as analysis result files, RDF files, etc.
5http://www.enterpriseintegrationpatterns.com/
6http://camel.apache.org/

Operation HTTP Method URI Pattern
Get all media resources, Filter by sta-
tus, Filter by publisher
GET /mediaresource ?status ?publisher
Create a media resource POST /mediaresource
Get a media resource GET /mediaresource/{id}
Replace a media resource PUT /mediaresource/{id}
Delete a media resource DELETE /mediaresource/{id}
Get all relations for a media resource GET /mediaresource/{id}/relation
Create a media resource relation POST /mediaresource/{id}/relation
Get a media resource relation GET /mediaresource/{id}/relation/{id}
Replace a media resource relation PUT /mediaresource/{id}/relation/{id}
Delete a media resource relation DELETE /mediaresource/{id}/relation/{id}
Table 3: Administration REST API functions of the Linked Media Layer
A JSON representation of the response can be retrieved by adding the extension .json to the re-
quest (e.g. http://api.linkedtv.eu/mediaresource.json). The deliverable D5.4 provides a detailed
description of the REST API functions [5].
2.2.7 RDF Repository REST API
The base URL for the REST API URIs is http://data.linkedtv.eu. Table 4 lists the collection of REST
API calls to access the RDF Repository.
Operation HTTP Method URI Pattern
Get all media resources GET /mediaresource
Get a specific media re-
source
GET /mediaresource/{id}
Get all media fragments
of a media resource
GET /mediaresource/{id}/mediafragment OR /mediafrag-
ment?isFragmentOf={mediaresourceURI}]
Get all media fragments
of a media resource for a
specific time interval
GET http://data.linkedtv.eu/mediafragment?isFragmentOf=
{mediaresourceURI}&min-
temporalStart={startTime}&max-
temporalEnd={endTime}
Get all shots of a media
resource
GET http://data.linkedtv.eu/mediaresource/{ID}/shot
Get all annotations of me-
dia resource Full view
GET /mediaresource/{id}/annotation?_view=full
Get all media fragments GET /mediafragment
Get all shots GET /shot
Get all annotations GET /annotation
Get spatial objects GET /spatialobject
Add an annotation to a
media fragment
PUT /mediaresource/{id}/annotation
Table 4: Triple Store REST API functions of the Linked Media Layer
Besides these specific REST calls (more will be added in future releases) the general SPARQL
endpoint at http://data.linkedtv.eu/sparql allows to retrieve arbitrary data from the Triple Store via
SPARQL queries. The response can be requested in various formats through indicating the content
type. The following content types are supported:
– text/javascript, application/javascript: JSONP representations (on some browsers)
– text/plain: Plain text representation
– application/rdf+xml, text/turtle: RDF/XML or Turtle representation
– application/json: JSON representation

– text/xml, application/xml: XML representation
– text/html: model rendered by stylesheet applied to XML rendering
The REST API for the RDF triple graph is implemented with Elda7, a JAVA implementation of the
Linked Data API speciﬁcation8. The Linked Data API provides a conﬁgurable way to access RDF data
using simple RESTful URLs that are translated into queries to the SPARQL endpoint. For a more detailed
description of the RDF Repository API, see the deliverable D5.4 [5].
7http://www.epimorphics.com/web/tools/elda.html
8http://code.google.com/p/linked-data-api/

3 Unstructured Search Module
This section describes the progress made on the “Unstructured Search Module”, which serves within
the LinkedTV project for retrieval of additional content from a curated list of web sites. The functionality,
dependencies, and the main technologies used for the implementation described here were specified in
the Deliverable D2.3 [1] (Section 3.1). The Unstructured search module performs crawling, a methodical,
automated manner to collect data from Web resources. It also provides indexing of the crawled data,
so that these can be quickly retrieved in response to a query. The query interface is exposed via REST
web services.
This section is organized as follows: section 3.1 focuses on a detailed description of the technological
background. The interconnection of the individual components into a complete workflow is covered in
section 3.2. The extensions and new modules developed within LinkedTV are described in section 3.3.
The index structure is described in section 3.4, and the API of the Unstructured Search Module in
section 3.5. We present some benchmark results in section 3.6 and future work in section 3.7.
3.1 Frameworks Used
The Unstructured Search Module is developed on top of the Apache Nutch, Apache Solr and Apache
HBase frameworks. These frameworks are used to gather information from the white-listed web sites
(Apache Nutch), store them (Apache HBase) and index them for fast retrieval (Apache Solr). All these
frameworks are developed and maintained by the non-profit Apache Foundation and released under the
open-source Apache License9.
The selected frameworks represent the de-facto standard in the open-source community, which pro-
vides a rich set of features as well as the availability of free support via the mailing lists. Nevertheless, it
was necessary to perform some customization to cater for the specific LinkedTV needs, mainly related
to the indexing of multimedia content (videos, podcasts). The implemented extensions are described in
section 3.3 and the foreseen extensions in section 3.7.
3.1.1 Apache Nutch
Within the Unstructured Search Module, Apache Nutch (nutch.apache.org) is used as a highly extensi-
ble and scalable web crawler. Apache Nutch can run on a single machine, but it can also run in a cluster
(Apache Hadoop). Apache Hadoop allows for the distributed processing of large data sets across clus-
ters of computers. Hadoop, Nutch core and all their modules and plugins are written in Java. Plugin is a
set of extensions which correspond to the extension points defined by Nutch. Some of extension points
are e.g.:
– IndexingFilter
– Parser
– HtmlParseFilter
– URLFilter
Apache Nutch (in version 2.x) implements an abstract storage layer, which enables to use a data
store of choice. This abstraction is handled by Apache Gora (gora.apache.org), a new source open
source framework providing an in-memory data model and persistence for big data.
3.1.2 Apache Solr
Apache Solr (solr.apache.org) is an advanced search platform based on Apache Lucene (lucene.apa
che.org). Major features are powerful full-text search, faceted search, near real-time indexing, database
integration and more. Additionally, Solr provides comprehensive web-based administration interfaces.
Apache Lucene (lucene.apache.org) is a high-performance text search engine. It is provided as a
library written in Java. Lucene is used by Solr at its core for full-text indexing and search. Solr is also
written in Java and runs as a standalone full-text search server within a servlet container, for example in
Apache Tomcat (tomcat.apache.org) .
Solr provides a REST API, which can be called over HTTP from any programming language or
platform.
9http://www.apache.org/licenses/LICENSE-2.0.html

3.1.3 Apache HBase
HBase (http://hbase.apache.org/) is non-relational, distributed, column-oriented database for big
data running on top of HDFS10 and providing BigTable-like capabilities for Hadoop. It is useful for
random, realtime read/write access to Big Data because BigTable can be used both as an input source
and as an output target for MapReduce jobs.
3.2 Workflow
Section 3.1 described the frameworks on which the Unstructured Search Module is based on. This
section gives a step-by-step account of the crawling process.
3.2.1 White-list
To start indexing data, Nutch requires a seed file. This seed file contains the urls of web pages which
should be crawled. The seed is an ordinary text file with a list of webpages, one url per line, e.g.:
http://www.kulturradio.de/
http://www.fritz.de/
http://www.funkhauseuropa.de/
After the seed file has been specified, it is possible to restrict the crawling to some domains, or
more generally, to url patterns. This is done by defining regular expressions in a configuration file
regex-urlfilter.txt. For example, if crawling is limited to the domains defined above, the specification
is as follows:
+^http://([a-z0-9]*.)*kulturradio.de/
+^http://([a-z0-9]*.)*fritz.de/
+^http://([a-z0-9]*.)*funkhauseuropa.de/
The lines above will instruct Nutch to include all url in these three domains (respecting robots.txt file).
Example of robots.txt disallowing Nutch to crawl the website:
User-agent: Nutch
Disallow: /
3.2.2 Crawling
Once the seed and regex-urlfilter configuration files have been specified, the crawling can start. All of
the steps described below (see schema 3) are provided by Java classes within the packages:
– org.apache.nutch.crawl
– org.apache.nutch.fetcher
– org.apache.nutch.parse
The workflow runs within Nutch crawl script called crawl with four input parameters
crawl <seedDir> <crawlID> <solrURL> <numberOfRounds>
where:
– seedDir path to the directory with seed files,
– crawlID the id to prefix storage schemas to operate on,
– solrURL url where Solr index is located,
– numberOfRounds iterations of generate - fetch - parse - update cycle (earlier versions used depth
parameter).
Example:
crawl /opt/ir/nutch/urls webpage http://127.0.0.1/solr/ 10
10Hadoop Distributed Filesystem

CRAWL LIST
urls (seed)
CRAWL LIST
urls (seed)
1. INJECT
CRAWL DB
Known or
fetched urls
LINK DB
URLs and links
PLUGIN:
MediaExtractor
RENDER PAGERENDER PAGE
EXTRACT MEDIAEXTRACT MEDIA
SAVE TO PAGE
METADATA
SAVE TO PAGE
METADATA
SOLR
INDEX
2. GENERATE
SEGMENTS
201307081225
crawl generate – set of URLs
crawl_fetch – status of each URL
content – raw content
parse_text – text
parse_data – outlinks + metadata
crawl_parse – outlinks for update db
3. FETCH2. GENERATE
fetch list
4. PARSE
5. UPDATE DB
6. INVERT LINKS
7. INDEX
MediaSolrIndexer
EXTRACT MEDIA
FROM PAGE METADATA
EXTRACT MEDIA
FROM PAGE METADATA
CREATE NEW
NUTCHDOCUMENTS
CREATE NEW
NUTCHDOCUMENTS
WRITE IN CONTEXTWRITE IN CONTEXT
NUTCH
Figure 3: Crawling workflow.
3.2.2.1 Inject This step takes the urls from the seed file and adds them to a crawlDB (backend
datastore). In the Unstructured Search Module, an HBase table is used as the backend datastore.
These urls are stored with information about fetch schedule, fetch status and metadata.
3.2.2.2 Generate This step generates a subset of data from a crawlDB, which should be fetched.
The main purpose is to generate a segment which tells Nutch which and how to fetch the urls.
Segment is a set of web pages that are fetched and indexed as a unit. It holds all information used
for generate/fetch/update crawling cycle (for example fetchlist which is a file that lists the set of pages
to be fetched). There is a number of parameters that can be specified for this step. Empirically, one of
the most important parameters is the fetch list size – number of top urls to be selected. The result is a
fetchlist (list of urls to be fetched) which is stored in the segment directory.
3.2.2.3 Fetch This step fetches the urls generated in the previous step. Nutch fetches the resources
using a defined protocol (for example http) and places the data back to a segment directory under the
timestamp when the url was created. One of the key parameters of this step is the number of fetching
threads per task. This step usually is the most time-consuming one.
3.2.2.4 Parse The fetched data are parsed within this step. The parser processes the textual content
of each web page and extracts the links. The output of the parser is the parsed text, metadata, and
the newly acquired urls. This step was extended with our own solution for crawling multimedia content
(section 3.3)
3.2.2.5 Updatedb Within this step, the list of urls to be fetched is updated with the newly acquired urls
(outlinks) during the parsing step. After this, the crawling process continues to cycle through generate,
fetch, parse and updatedb steps. A pre-specified number of iterations can be made.
3.2.2.6 Solrindex At the end of the workflow described above, the Solrindex step is initiated to save
crawled data to an index. Nutch delegates all data created in the parsing step to the IndexingFilter
extension, which generates the data to be indexed. The output of the IndexingFilter extension is a
NutchDocument. This document is again delegated to Nutch.
Nutch then decides which data should be indexed based on the mapping file. The mapping file (see
section 3.3.4) defines which NutchDocument fields will be mapped to which SolrDocument fields. The
implementation of the mapping file in Nutch was too restricted for LinkedTV purposes, leading to the
need to develop an extension/replacement for this step as described in section 3.3.4.
3.2.3 Refreshing index
One of the main requirements in the LinkedTV scenarios is keeping the index in sync with the white-
listed web sites. For this purpose, Nutch implements the Adaptive Fetch Schedule. This feature allows

to adapt the revisit frequency for a particular page to the rhythm of changes in the page content.
When a new url is added to the crawlDB, it is initially set to be re-fetched at the default interval. The
next time the page is visited, the Adaptive Fetch Schedule will increase the interval if the page has not
changed and decrease it if the page has changed. The maximum and a minimum revisit interval are
configuration parameters. Over the time, the pages that change often will tend to be re-indexed more
frequently than the ones that do not change frequently.
3.3 Developed Extensions
3.3.1 Multimedia Indexing Plugin for Nutch
Nutch, by default, does not offer features for multimedia parsing. To the best of our knowledge, there is
no plugin that would allow an easy integration and retrieval of multimedia content. As a consequence, the
corresponding extension had to be designed for LinkedTV purposes. This lead to substantial changes
in the index structure, because Nutch follows a single web page - single index entry paradigm, which
no longer holds if images, podcasts, and videos featured on a page become objects of standalone
importance.
The main class of the developed plugin is MediaParseFilter. This class overrides the default Nutch
extension point ParseFilter. This can be considered as entry point for every web page to be parsed. In
its overridden method called “filter”, this class searches for multimedia on the webpage. Additionally, it
provides additional meta data about the web page.
At the beginning of every parse step for a single web page, lists of Document Object Model (DOM)
managers and taggers are created. DOM manager is implementing a common interface that we dubbed
as DomManager. The interface has only one method matchAndParseMetadata() declared. Its purpose
is to:
– parse a web page,
– search for multimedia objects,
– identify additional metadata.
After the list is initialized, every instance of the DOM manager from the list is run. The input param-
eters of the matchAndParseMetadata() method are a representation of HTML page, url of the page and
the set of known ids (id is an url of the media found) to prevent from storing twice the same multimedia
item from one page by more dom managers11. The page content is represented by the W3C Document
Object Model, which allows the managers to use XPath to query for elements and attributes. Regular
expressions are also frequently used.
The result of the method is a set of multimedia objects. Each of these objects is saved as a separate
document to the index (during indexing phase), and assigned with a document representing the HTML
page within which the document was located. This results in every occurrence of a multimedia object
having a unique index document (as opposed to every unique object). This behaviour is required to
create a custom solution for the Solrindex step (see section 3.2.2.6). The metadata and the time of
parsing are also saved.
This plugin performs also a first filtering of some multimedia objects such as logos and banners. The
current implementation uses a blacklist of names, which, if present in the multimedia object’s file name,
cause the object to be skipped. This simple, yet effective technique can be extended in multiple ways:
1. blacklist based on object dimensions,
2. number of occurrences of the object on the web site,
3. content-based classification.
3.3.2 Rendering-Based Multimedia Extraction - HtmlUnit and Dom managers
HtmlUnit (http://htmlunit.sourceforge.net/) is an open source Java library. It creates HTTP calls
which imitate the browser functionality from a code. It proved to be useful to use this library to access
and process the rendered source code, in addition to using the unprocessed source code as directly
11This could happen, for example, with videos represented by both <object> and <embed> html elements, which are often
repeated within the same page multiple times (with small syntactical variations), but referencing the same multimedia object.

returned from the server over the HTTP protocol. The motivation for using the rendered code is the
fact that many pages from the whitelist specified by the content partners contain multimedia objects
embedded in such a way that they cannot be found in the unprocessed source code.
Illustrative example. Consider the web page http://www.rbb-online.de/brandenburgaktuell/
index.html:
Multimedia representation before rendering by HtmlUnit:
function setupLb_586015052(){
jwplayer("lb_586015052").setup({
’modes’: [
{type: ’flash’,
src: ’/cqbase/jwplayer/5_9/player.swf’,
config: {
’file’:
’/rbb/aktuell/dossier/aktuell_20130418_Zeitzeugentrailer_17._Juni_m_16_9_512x288.mp4’,
’provider’: ’rtmp’,
streamer’: ’rtmp://ondemand.rbb-online.de/ondemand’
With HtmlUnit:
<object type="application/x-shockwave-flash" data="/cqbase/jwplayer/5_9/player.swf"
width="100%" height="100%" bgcolor="#000000" id="1817906585" name="1817906585"
tabindex="0">
<param name="allowfullscreen" value="true">
<param name="allowscriptaccess" value="always">
<param name="seamlesstabbing" value="true">
<param name="wmode" value="transparent">
<param name="flashvars" value="netstreambasepath=http%3A%2F%2Fwww.rbb-online.de%2F
brandenburgaktuell%2Findex.html&id=1817906585&image=%2Fetc%2Fmedialib%2Frbb%2Frbb%2F
aktuell%2Ffans_fiebern_championsleage.file.512.288.jpg&skin=%2Fstatic%2Frbb%2F
jwplayer%2Fskins%2Fclassic%2Fclassic.xml&screencolor=0x000000&bufferlength=5
&smoothing=true&idlehide=false&stretching=uniform&wmode=transparent&bgcolor=0x000000&
autostart=false&file=%2Frbb%2Faktuell%2Faktuell_20130525_fan_m_16_9_512x288.mp4&
provider=rtmp&streamer=rtmp%3A%2F%2Fondemand.rbb-online.de%2Fondemand&
controlbar.position=bottom&display.icons=true&dock=false">
</object>
The object element above is generated by JavaScript. But additionally, the script checks whether
Flash plugin is installed before rendering the object. Therefore, custom BrowserVersion (objects of this
class represent one specific version of a given browser) that supports Flash should be used. After that,
the web page is represented in the code like an HtmlPage object.
HtmlPage is a class that represents a single web page along with the client data (HTML, JavaScript,
CSS etc.) and provides different methods to access the page content (html elements) like getElement-
ById or getByXPath. In order to increase the number of identified videos per web page on the whitelist,
a survey of coding practices for embedding multimedia objects12 into a web page was performed. Con-
sequently, several classes called dom managers were designed (non-exhaustive list):
– ObjectTagDomManagerImpl
– VideoTagDomManagerImpl
– JavascriptContentDomManagerImpl
– IFrameDomManagerImpl
– YoutubeObjectTagDomManagerImpl
12Videos and podcasts, images do not generally pose a processing problem.

3.3.3 Nutch Annnotated Crawling List Extension
For the purposes of keeping urls from the two whitelists (RBB and Sound&Vision) apart within one
physical index, it was necessary to create an additional configuration file for Nutch. This apparent simple
task required a greater amount of changes because Nutch does not support custom configuration files.
Nutch source code was extended with one class called DomainTaggerImpl.
This class is used to load the annotated filter configuration file, which contains a flag indicating to
which whitelist the url belongs. After detection of a domain source, the url is assigned to the appro-
priate field called domain_source. To make Nutch aware of this configuration file, a new property to
configuration file nutch-default.xml needs to be added:
<property>
<name>urlfilter.whitelist.file</name>
<value>whitelist-urlfilter.txt</value>
<description>
Name of a file on CLASSPATH containing urls of whitelisted pages
used by media-extractor (DomainTaggerImpl) plugin
</description>
</property>
A format of the configuration file (whitelist-urlfilter.txt):
# LinkedTV configuration file for white-listed urls.
# RBB White List - http://linkedtv.eu/wiki/index.php/RBB_White_List
# SV White List - http://linkedtv.eu/wiki/index.php/Sound_and_Vision_White_List
#
# If you want to add resource you can do it in this way:
#
# Format:
# Set prefix "R_" for urls which you want to have indexed as "rbb" in the Solr field
# called "domain_source"
# or set prefix "S_" for urls which you want to have indexed as "sv" in the Solr
# field called "domain_source"
#
# then add url (without http or www)
# example: R_rbb-online.de
R_antennebrandenburg.de/
R_inforadio.de/
S_amsterdammuseum.nl/
S_imdb.com/
S_linkedmdb.org
3.3.4 Nutch-Solr bridge
The development of extensions for handling multimedia objects described in sections 3.3.1 and 3.3.3
triggered the need to implement a replacement for the SolrIndex class (section 3.2.2.6), which is unable
to work with the resulting custom data structure.
Mapping of fields created by Nutch to fields defined (and expected) by Solr (schema.xml) is set in a
configuration file called solrindex-mapping.xml. This file contains the setting used by the Solrindex step
at the end of the script crawling process. For the LinkedTV purposes, this file cannot be used because it
does not offer the flexibility required to deal with the customizations resulting from multimedia processing.
Instead, the process of placing fields in the index is performed by the new class MediaSolrIndexerJob.
This class extends the default Solr IndexerJob class and contains two inner classes:
SolrMediaIndexerJobMapper
SolrMediaIndexerJobReducer
These classes contain overriden methods to map, reduce, load and setup configuration (in addition to
other methods). The main purpose of the map method is to create Nutch documents for all media items
found in a web page including the web page itself. Nutch takes all the created Nutch documents, maps

them on Solr documents and writes them to the Solr index. After the documents are written, they are
available in Solr for querying.
3.4 Index structure
One document in the index can be represented as an XML element <doc>. The children elements of
<doc> correspond to fields. There are four types of documents in the index: webpage, video, podcast
and image.
Illustrative Example. Document of image type.
<doc>
<str name="media_type">image</str>
<str name="media_addition"/>
<str name="media_url">
http://www.zdf.de/ZDFmediathek/contentblob/1938364/timg298x168blob/8171749
</str>
<str name="digest">5fc5d474b5b6383079780d955ddc6432</str>
<date name="tstamp">2013-08-07T22:06:14.237Z</date>
<str name="media_title">Bild zeigt den Berliner Reichstag bei Nacht</str>
<str name="url">
http://www.zdf.de/Berlin-direkt/Team-Berlin-direkt-22674890.html
</str>
<str name="source_code_rendered">
</str>
<str name="domain_source">rbb</str>
<str name="id">
http://www.zdf.de/ZDFmediathek/contentblob/1938364/timg298x168blob/8171749
</str>
<str name="source_code">
</str>
<str name="content">
HERE IS A CONTENT OF A HTML PAGE WHERE THE IMAGE IS LOCATED BUT WITHOUT HTML
ELEMENTS. JUST TEXT REPRESENTATION. IT IS OMITTED HERE FOR SHORTHAND.
</str>
<str name="media_solver_class">ImgMetaManagerImpl</str>
<str name="title"/>
<long name="parse_time_unix_timestamp">1374718537</long>
<date name="parse_time">2013-07-25T04:15:37.99Z</date>
<str name="media_description">Bild zeigt den Berliner Reichstag bei Nacht</str>
<long name="_version_">1441505267507265536</long></doc>
<doc>
This entry shows how the document is represented in Solr. Below, we will briefly describe some of
the fields:
– media_type: the document type (e.g. an image)
– media_url: the location of the source document (e.g. the URL of the image)
– id: a unique key for every document in the index, media_url is used for the key
The semantics of fields source_code and source_code_rendered differs depending on the me-
dia_type. For multimedia objects (media_type different from web page), the source code contains the id
of the document (web page) containing the source code:
<str name="source_code">
</str>
With this information, it is possible to retrieve the original document, which will be a web page:

<str name="media_type">webpage</str>
The ﬁelds source_code and source_code_rendered (see section 3.3.2) store complete source code
of a web page with html elements and javascript.
<str name="source_code_rendered"><?xml version="1.0" encoding="UTF-8"?>
<html lang="de" class=" js n
..... etc.
<str name="source_code"><!DOCTYPE html> query
Example: q=media_url:*
row --> number of items to retrive
Default: 10
Example: row=100
3.5.1 Syntax
Wildcards and Operators. Wildcards ? and * replace one or more characters. The common boolean
operators AND, OR, NOT are also available.

Range Queries. It is possible to use [ from TO to ] syntax to form range queries. The symbol * may
be used for either or both endpoints to specify an open-ended range query.
– field:[* TO 100] matches all field values less than or equal to 100
– field:[100 TO *] matches all field values greater than or equal to 100
– field:[* TO *] matches all documents with the field present
In the LinkedTV setup, range queries are intended for the timestamp* fields. The complete ISO 8601
date syntax that fields support, or the DateMath Syntax to get relative dates. For example to get all
documents crawled since August 22, 2013:
parse_time:[2013-08-22T07:33:21.699Z TO *]
Alternatively, the range query can be executed against the parse_time_unix_timestamp field, which is
stored as data type long:
parse_time_unix_timestamp:[1377149601 TO *]
Proximity. Words should be within a specific distance away.
– content:"Berlin parliament"~2 finds documents, where words “Berlin” and “parliament” are within 2
words of each other
Fuzzy Search. Finding similar terms (Levenshtein13 or edit distance) in spelling.
– media_title:meine~0.7 finds documents with title similar to “meine” with similarity 0.7
Priorities (boost). It is possible to set priorities for individual fields. Default priority is 1. Higher number
means higher priority.
– (media_title:*Berlin*)^5 (media_description:*Parlament*)^0.5
3.5.2 API response
The mapping relationship between the values of the Solr fields (section 3.4) and the response fields
(section 4) is as follows:
– url -> micropostUrl
– parse_time_unix_timestamp -> timestamp
– parse_time -> publicationDate
– media_type -> type
query: page | image (also photo is supported) | video | podcast (also audio is supported) | webpage
response: photo | video | audio | webpage
– media_url -> mediaUrl
if media_type equals "image" then media_url -> posterUrl
– media_description -> micropost.html
– media_title -> micropost.plainText
The content of the publicationDate field is set to the timestamp when the page was parsed by Nutch, i.e.
it corresponds to the retrieved date, rather than to a publication date as for example contained in meta
name="DC.Date" from a source code of a page.
Document uri authority (DNS host name) is used as source id (e.g. www.ard.de). Source id is an
unique identifier containing items with properties like mediaUrl, type, publicationDate.
13http://en.wikipedia.org/wiki/Levenshtein_distance

www.ard.de: [
{
micropostUrl: "http://www.ard.de/",
micropost: {
html: "Woller und Schwester Hanna (Bild: ARD/Barbara Bauriedl)",
plainText: "Woller und Schwester Hanna (Bild: ARD/Barbara Bauriedl)"
},
mediaUrl:
"http://static.daserste.de/cmspix/tvtipp/thumb_128_72_05022013187955.jpg",
type: "photo",
timestamp: 1398772232014,
publicationDate: "2013-08-29T13:50:32Z"
}
]
3.6 Evaluation
To demonstrate the functionality of the Unstructured Search Module, the section 3.6.1 presents several
sample queries on the LinkedTV content and section 3.6.4 presents some statistics about the current
index.
3.6.1 Sample Queries
This subsection lists several queries based on the LinkedTV scenarios. The queries were executed
against the index populated with web pages indexed from web sites on the whitelist provided by the
LinkedTV content partners. The queries were manually constructed based on entities appearing in the
LinkedTV scenarios. All queries returned result as listed at the time when they were designed. Due to
changes on the web sites (especially RBB content is made unavailable after a certain time period), the
listed response of the queries is valid only as of time of writing.
The search was performed mainly over the content, title and media_title fields, which hold the plain
text content of the web page (or web page holding the multimedia object), value of HTML title tag and
value of the img tag’s title attribute. Higher precision can be obtained by involving the foreseen Metadata
Extraction Service (MES) service (section 3.7.2). The list of result for individual queries listed below are
abridged to contain only the most relevant results (unless noted otherwise).
3.6.2 RBB scenarios
The following queries address some of the information requirements related to a person called Nina from
the RBB scenarios [6].
At the beginning of the scenario, Nina starts to watch TV news. The news are mainly about the
US president Barack Obama and his visit to Berlin, but there are also other news items. Below is a
demonstration of results returned by several specific queries.
Query RBB 1. The news starts with the president’s departure. The prominent entities shown include
Air Force One14 or Airport Tegel, where the aircraft landed.
– Object: Air Force One
– Query to API: q=title:"Air Force One" AND (media_type:video OR media_type:webpage) AND
domain_source:rbb
– Relevant media content returned from the API:
◦ http://www.ardmediathek.de/rbb-fernsehen/theodor/am-steuerknueppel-von-honec
kers-air-force-one?documentId=16752626
◦ http://www.whitehousemuseum.org/special/AF1/
14The official air traffic control call sign of a United States Air Force aircraft carrying the President of the United States

Figure 4: Airport Tegel. Source: http://www.rbb-online.de/politik/thema/Flughafen-BER/BER-A
ktuelles/hintergrund/beitraege/txl/_jcr_content/image.img.jpg/rendition=original/size=
708x398.jpg
This query found relevant information on the entity Air Force One. The ﬁrst one is a video about
Air Force One pilots from media library and the second result is a web page about history of the
aircraft with a lot of photos and other information.
– Object: Airport Tegel
– Query to API: q=(title:"Tegel" OR media_title:*Tegel*) AND domain_source:"rbb" AND (media_type:
"podcast" OR media_type:"image")
◦ http://www.rbb-online.de/politik/thema/Flughafen-BER/BER-Aktuelles/hintergrund
/beitraege/berliner_flughaefen/_jcr_content/pictureList/13748/image.img.jpg/re
ndition=original/size=708x398.jpg
◦ http://www.rbb-online.de/politik/thema/Flughafen-BER/BER-Aktuelles/hintergrund
/beitraege/txl/_jcr_content/image.img.jpg/rendition=original/size=708x398.jpg
◦ www.inforadio.de//cqbase/flash/jwplay.swf?streamer=rtmp://stream5.rbb-online.de
/inf&autostart=false&image=/etc/medialib/rbb/inf/bild/politiker__prominente/ma
thias_koehne__spd.file.220.165.jpg&file=/programm/boe/201306/28/NP079362.mp3
◦ www.inforadio.de//cqbase/flash/jwplay.swf?streamer=rtmp://stream5.rbb-online.de
/inf&autostart=false&image=/etc/medialib/rbb/inf/bild/verkehr/luftverkehr/flug
hafen_berlin_tegel0.file.220.165.jpg&file=/programm/boe/201212/10/NP071961.mp3
This query found relevant information about the entity Airport Tegel. The results include images
(see Fig. 4) of the airport and podcasts with content relevant to the query entity.
Query RBB 2. According to the scenario, Nina is curious about the meeting at the Brandenburg Gate.
She wants to hear what Obama said. She also wants to see the image gallery of this big event.
– Object: Barack Obama at Brandenburg Gate
– Query to API: q=(title:"Barack Obama Brandenburg Gate"~5 OR (title:"Barack Obama Branden-
burger Tor"~5)) OR (content: "Barack Obama Brandenburg Gate"~5 OR (content:"Barack Obama
Brandenburger Tor"~5)) AND domain_source:"rbb"
◦ http://www.radioeins.de/cqbase/jwplayer/5_9/player.swf?file=http%3A%2F%2Fdownlo
ad.radioeins.de%2Fmp3%2F_programm%2F8%2F20130619_1620_obama.mp3&autostart=true

Figure 5: Barack Obama at Brandenburg Gate. Source: http://www.radioeins.de/etc/medialib/rb
b/rad/bilder0/201306/us_praesident_brack.file.66656.460.259.jpg
◦ http://www.dw.de//js/jwplayer/player.swf?streamer=rtmp://tv-od.dw.de/flash/&pro
vider=rtmp&DWQualityPlugin.isold=false&DWQualityPlugin.initialbandwidth=&DWQua
lityPlugin.pluginmode=FLASH&controlbar.position=over&controlbar.idlehide=true&
file=je/je20130619_obamaspeech_sd_avc.mp4
◦ http://dw.de/us-president-barack-obama-speaks-at-brandenburg-gate/a-16894277
◦ http://www.radioeins.de/etc/medialib/rbb/rad/bilder0/201306/us_praesident_brac
k.file.66656.460.259.jpg
◦ http://fritz.de/etc/medialib/rbb/fri/bilder/beitraege/neues_wort/aktuell/2013/
130618_obama_berlin.file.jpg
This query found relevant information on Obama’s speech at the Brandenburg Gate including
images (consider e.g. Figure 5), a podcast, a video and a webpage.
Query RBB 3. Nina is interested in the speeches of Klaus Wowereit (Governing Mayor of Berlin).
– Object: Klaus Wowereit and his speech at Brandenburg Gate item Query to API: q=(title:"Wowereit
Obamas"~50) OR (content:"Wowereit Obama"~50) AND domain_source: "rbb" AND (media_type:"podcast"
OR media_type:"video" OR media_type:"webpage")
◦ http://www.berlin.de/landespressestelle/archiv/20130619.1340.386279.html
This query found relevant information about the text of the Governing Mayor’s speech (Klaus Wo-
wereit) during Obama’s visit.
Query RBB 4. The voice in the news mentions Barack Obama’s half-sister Auma. Nina would like to
get some additional information about Auma Obama.
– Object: Auma Obama
– Query to API: q=title:"Auma Obama" AND domain_source:"rbb"
◦ http://de.wikipedia.org/wiki/Auma_Obama
◦ http://mp3-download.swr.de/swr1/bw/leute/auma-obama-entwicklungshelferin-in-ken
ia.6444m.mp3
◦ http://www.dw.de/die-kenianische-menschenrechtlerin-auma-obama/a-16703706
This query found relevant information about Auma Obama including a podcast.

Figure 6: Tim Raue, the chef at Schloss Charlottenburg. Source: http://www.radioeins.de/etc/med
ialib/rbb/rad/personen/r/tim_raue___dpa_bildfunk.file.12367.460.259.jpg
Query RBB 5. The next chapter of news is about the official dinner at Schloss Charlottenburg. Nina is
very interested in it and tries to find out about the chef in charge, Tim Raue (see Figure 6).
– Object: Schloss Charlottenburg
– Query to API: q=(title: "Schloss Charlottenburg" OR media_title:*Charlottenburg*)
AND domain_source:"rbb"
◦ http://www.berlin.de/aktuelles/obama-besuch-2013/3099566-3095135.gallery.html?p
age=1
– Object: Tim Raue
– Query to API: q=(title:"Tim Raue" OR media_title:*Raue*) AND domain_source:"rbb"
◦ http://www.radioeins.de/etc/medialib/rbb/rad/personen/r/tim_raue___dpa_bildfunk
.file.12367.460.259.jpg
◦ http://www.radioeins.de/cqbase/jwplayer/5_9/player.swf?file=http%3A%2F%2Fdownlo
ad.radioeins.de%2Fmp3%2F_programm%2F8%2F20121215_arena_raue.mp3&autostart=true
◦ http://www.berlin.de/restaurants/1640924-1622830-restaurant-tim-raue.html
Query RBB 6. The last spot in the video is about politicians from Brandenburg. One of them is Matthias
Platzeck. Nina wants to know more about him. She wants to see his profile and get more information.
– Object: Matthias Platzeck
– Query to API: q=(title:"Matthias Platzek" OR media_title:*Platzeck*) AND domain_source:"rbb"
◦ http://www.antennebrandenburg.de/programm/dossiers_programm/2013/matthias_platz
eck.html
◦ http://www.antennebrandenburg.de/etc/medialib/rbb/ant/bilder/politiker/matthias
_platzeck0.file.200.150.jpg
◦ http://www.inforadio.de/etc/medialib/rbb/inf/bild/politiker__prominente/matthia
s_platzeck22.file.28953.402.171.jpg
◦ http://www.berlin.de/aktuelles/berlin/3137562-958092-platzeck-wird-nach-auszeit
-in-seiem-buer.html

◦ http://www.inforadio.de/programm/schema/sendungen/int/201301/16/platzeck_und_w
owereit.html
The results presents a wealth of information about the query entity. Nina can view Platzeck’s
profile, relevant articles. The results also include pictures featuring Platzeck.
3.6.3 S&V scenarios
The queries below address some of the information requirements defined for the two selected S&V
scenarios [6]. There is a person called Rita defined in the first scenario, and persons Bert and Anne in
the second scenario.
Query S&V 1. Rita watches the show and she is interested in finding out more about the show’s host
Nelleke van der Krogt.
– Object: Nelleke van der Krogt
– Query to API: q=(title:"Nelleke van der Krogt" OR content:"Nelleke van der Krogt") AND do-
main_source:"sv"
◦ http://www.rmo.nl/actueel/nieuws/2011/tussen-kunst-en-kitsch
◦ http://cultuurgids.avro.nl/front/archiefkunst.html?trefwoord=NellekevanderKrogt
Query S&V 2. Rita also sees that the show has been recorded in the Hermitage Museum in Am-
sterdam. She has always wanted to visit the museum, and to find out what the link is between the
Amsterdam Hermitage and the Hermitage in St. Petersburg.
– Object 1: Hermitage Museum in Amsterdam
– Object 2: Hermitage in St. Petersburg
– Query to API 1: q=(title:"Hermitage" AND content:"Hermitage Amsterdam history"~5) AND do-
main_source:"sv"
– Query to API 2: q=(title:"Hermitage" AND content:"Hermitage Petersburg"~5) AND domain_source:"sv"
◦ http://www.hermitage.nl/en/hermitage_amsterdam/
◦ http://www.artcyclopedia.com/featuredmuseums-feb99.html
◦ http://www.hermitage.nl/en/st-petersburg_en_rusland/de_stad_st-petersburg.htm
◦ http://www.the-athenaeum.org/sources/detail.php?id=141
Results include web pages with articles about both Hermitages.
Query S&V 3. The show is coming to an end and the final person on the show has brought in a painting
that has the signature Jan Sluijters. It is a famous Dutch painter. Rita wants to learn more about this
painter (see Figure 7) and the art styles he represents.
– Object: Jan Sluijters
– Query to API: q=title:"Jan Sluijters" OR content:"Jan Sluijters"
◦ http://de.wikipedia.org/wiki/Jan_Sluijters
◦ http://www.artcyclopedia.com/artists/sluijters_jan.html
◦ http://www.avro.nl/Images/jan_sluijters_tcm8-261885.jpg
◦ http://staticextern.avro.nl/gfx/museumtv/D01280_1_tcm4-65412.jpg
Results include images with paintings and also web pages about the painter and his life.

Figure 7: Jan Sluijters’s painting. Source: http://staticextern.avro.nl/gfx/museumtv/D01280_1_
tcm4-65412.jpg
Query S&V 4. The next scenario is about Bert and Anne. They also watch TV. Anne sees a chapter
about Greek mythology (gods Hebe and Karitas) and she wants to brush up on her knowledge about
these gods and the Greek mythology in general.
– Object 1: Hebe
– Object 1: Karitas
– Query to API 1: q=(title:"Hebe" AND content:"Griechische Gottheit")
– Query to API 2: q=title:"Karitas"
– Query to API 3: q=title:"mythologie" AND content:*Griech*
◦ http://www.the-athenaeum.org/art/detail.php?ID=92990
◦ http://www.the-athenaeum.org/art/detail.php?ID=87341
◦ http://de.wikipedia.org/wiki/Karitas
◦ http://de.wikipedia.org/wiki/Dia_(Mythologie)
◦ http://de.wikipedia.org/wiki/Griechische_Mythologie
The results contain multiple relevant web pages related to the queries. The ﬁrst two items are
paintings with titles containing the word “Hebe”. However, these are false positives.
Query S&V 5. There is also one chapter in the programme about a gold watch by the famous watch
maker Breguet. Anne likes the watch and she notices that a clock system is made like a Jacquemart15.
– Object 1: Famous watch maker Breguet
– Object 1: Jacquemart
– Query to API 1: q=title:"Breguet" AND (content:"watch" OR content:"clock" OR content:"Uhr")
– Query to API 2: q=title:"Jacquemart"
◦ http://www.getty.edu/news/press/center/breguet.html
◦ http://de.wikipedia.org/wiki/Abraham_Louis_Breguet
◦ http://de.wikipedia.org/wiki/Jacquemart_(Uhr)
Those results could satisfy Anne’s information need.
15System that produces sound by having one or more human ﬁgures strike the bells.

3.6.4 Index statistics
The statistics listed below illustrate the growth of the index of the Unstructured Search Module from
an empty initial state (new software version released) over the course of two weeks and one week
respectively.
Old statistics (31/05/2013 - 14/06/2013).
– Video: 3917
– Podcast: 30
– Image: 14904
– Webpage: web pages were not stored in the index
Current statistics (21/09/2013 - 30/09/2013).
– Video: 5584
– Podcast: 2693
– Image: 79996
– Webpage: 73251
The increase in the number of crawled resources shows increasing effectiveness of the Unstructured
Search Module in extracting multimedia objects from the whitelists, the lists of web sites provided by the
LinkedTV content partners16.
3.7 Extensions
We already foresee a number of extensions to the Unstructured Search Module. A dashboard will be
contributed in year 3 of the project. For the remaining two extensions (Metadata Extraction Service
and Entity enrichment), a feasibility study will be performed, and based on its results the extension or a
proof-of-concept will be developed.
3.7.1 Dashboard
Dashboard a web application that will serve index statistics through a friendly graphical interface. The
dashboard will complement the existing reporting facilities in Solr, which e.g. does not distinguish be-
tween the types of documents stored (webpage, image, video, podcast).
3.7.2 Metadata Extraction Service
The purpose of the Metadata Extraction Service is to identify text associated with the multimedia object.
The extracted text will serve as metadata about the object, and will be stored into separate fields of the
index, allowing for higher precision queries.
3.7.3 Entity enrichment
A requirement of WP4 is to have enrichment content annotated with entity analysis. A foreseen exten-
sion of the current indexing process is to use an entity classification service to annotate the indexed
textual content and store the result into the index. Due to high time-complexity of the entity-classification
process, it is questionable whether such processing is feasible for all documents in the index.
16It should be noted that the list of web sites on these whitelists also grew in the respective period.

4 Media Collector
The widespread availability of mobile phones with higher resolution cameras has transformed citizens
into media publishers and witnesses who are used to comment and share event-related media on social
networks. Some examples with global impact include the shootings in Utøya, which first appeared on
Twitter, the capture and arrest of Muammar Gaddafi, which first appeared on YouTube, or the emergency
ditching of a plane in the Hudson river, which first appeared on Twitpic. Some news agencies17 have
even specialized in aggregating and brokering this user-generated content. In this section, we propose
a new approach for retrieving all those event-related media items that are being published by users on
several social networks [22, 17, 32]. They will provide a new source of LinkedTV enrichments that will
prove to be particularly relevant in the context of a news program enrichment [6].
We have developed a so-called Media Collector composed of media item extractors for all the media
sharing networks listed in Table 5. Furthermore, the Media Collector also integrates the Unstructured
Search Module described in the section 3 via the Search API that this module exposes. Consequently,
the Media Collector is the central module for the enrichment process that provides content coming from
social networks and white-listed web sites, and that can be later on validated by an editor and/or further
filtered by the LinkedTV personalization component.
The remainder of this section is organized as follows: in Section 4.1, we provide a classification
of social networks according to the media support they provide. In Section 4.2, we detail the process
of extraction and reconciliation of media items from different sources while we describe the query and
response formats of the Media Collector in section 4.3. Finally, we provide examples of results given by
the Media Collector for the LinkedTV scenarios in the section 4.4.
4.1 Social Networks
A social network is an online service or media platform that focuses on building and reflecting social re-
lationships among people who share interests and/or activities. The boundary between social networks
and media platforms is rather blurry. Several media sharing platforms, such as YouTube, enable people
to upload content and optionally allow other people to react to this content in the form of comments,
likes or dislikes. On other social networks (e.g., Facebook), users can update their status, post links
to stories, upload media content and also give readers the option to react. Finally, there are hybrid
clients (e.g., TweetDeck for Twitter using Twitpic) where social networks integrate with media platforms
typically via third party applications. Therefore, we consider three types of support of media items with
social networks:
– First-order support: The social network is centered on media items and posting requires the inclu-
sion of a media item (e.g. YouTube, Flickr);
– Second-order support: The social network lets users upload media items but it is also possible to
post only textual messages (e.g. Facebook);
– Third-order support: The social network has no direct support for media items but relies on third
party application to host media items, which are linked to the status update (e.g. Twitter before the
introduction of native photo support).
We consider 12 different social networks that all have powerful and stable APIs and, together, rep-
resent the majority of the market. The criteria for including media sharing platforms follow a study
performed by the company Sysomos, specialized in social media monitoring and analytics [12]. Table 5
lists these platforms according to the categorization defined above.
4.2 Implementation
Twitter and its ecosystem (TwitPic, TwitterNative, MobyPicture, Lockerz or yfrog), GooglePlus and YouTube,
Facebook and Instagram, Flickr and FlickrVideos, MySpace, all offer search APIs over the content they
host. Those search functions, however, provide results that vary according to the time the query has
been triggered, covering a window of time that ranges from only the recent past to many years ago. In
addition, they offer different parameters that enable to customize search queries (e.g. filtering by loca-
tion). The Media Collector is composed of media item extractors for these 12 media sharing platforms.
17e.g. Citizenside (http://www.citizenside.com)

Social Network URL Category Comment
Google+ http://google.com/+ second-order Links to media items are returned via the
Google+ API.
MySpace http://myspace.com second-order Links to media items are returned via the
MySpace API.
Facebook http://facebook.com second-order Links to media items are returned via the
Facebook API.
Twitter http://twitter.com second-/third-order In second order mode, links to media items
are returned via the Twitter API. In third or-
der mode, Web scraping or media platform
API usage are necessary to retrieve links
to media items. Many people use Twitter
in third order mode with other media plat-
forms.
Instagram http://instagram.com first-order Links to media items are returned via the
Instagram API.
YouTube http://youtube.com first-order Links to media items are returned via the
YouTube API.
Flickr http://flickr.com first-order Links to media items are returned via the
Flickr API.
MobyPicture http://mobypicture.com first-order Media platform for Twitter. Links to media
items are returned via the MobyPicture API.
Twitpic http://twitpic.com first-order Media platform for Twitter. Links to media
items must be retrieved via Web scraping.
img.ly http://img.ly first-order Media platform for Twitter. Links to media
Lockerz https://lockerz.com/ first-order Media platform for Twitter. Links to media
yfrog http://yfrog.com first-order Media platform for Twitter. Links to media
Table 5: Social networks with different support levels for media items and techniques needed to retrieve
them

Figure 8: The media collector architecture: it proposes a hybrid approach for the media item extraction
process using a combination of API access and Web scraping.
It takes as input a search term and a parallel key-search is then performed to these social networks.
Each platform has a 30 second timeout window to deliver its results. When the timeout has expired, or
when all social networks have responded, a uniﬁed output is delivered [22, 17, 32]. Figure 8 depicts the
overall architecture of the media collector.
The metadata attached to the microposts retrieved vary in terms of schemas, data types and serial-
ization formats. We harmonize these results and project them to a common schema (section 4.3). This
component performs also a cleansing process, discarding items which are older than seven days ago,
in order to keep only fresh media items. Besides this abstraction layer on top of the native data formats
of those social networks, we perform a similar task for the social interactions (Table 6 [32]).
Likes Shares Comments Views
Facebook Like
Google+ +1
Instagram Like
Flickr Favorite
YouTube Like
YouTube Favorite
Twitter Favorite
Facebook Share
Google+ Share
Twitter native ReTweet
Facebook Comments
Google+ Comments
Instagram Comments
Twitter manual RT, @Replies
Twitpic Comments
MobyPicture Comments
Flickr Comments
YouTube Views
Flickr Views
Twitpic Views
MobyPicture Views
Table 6: Abstract social network interaction paradigms and their underlying native counterparts [32]
Media Collector provides not only a way to capture a snapshot at a particular instant of what has been
shared in social media platforms, but enables also to monitor the results of a search query over a longer
period of time, by automatically re-issuing the same query at a regular frequency and by cumulating the
results [16].
The Media Collector was originally developed by Thomas Steiner at https://github.com/tomay
ac/media-server and forked twice for the purpose of LinkedTV at https://github.com/vuknje/medi
a-server (in order to enable temporal search across social media platforms) and later one at https:
//github.com/MathildeS/media-server (in order to integrate the Unstructured Search Module). It is
based on NodeJS18.
4.3 Query and Response Format
Different settings are possible depending on the type of search performed:
– a S&V or RBB focused search can be made based on the white lists they provided, mainly using
the unstructured search module (section 3);
– fresh media items can be searched on social networks using their provided APIs.
The query pattern is therefore: http://linkedtv.eurecom.fr/api/mediacollector/search/TYPE
/TERM where:
18http://nodejs.org/

– TYPE is one of [RBB, SV, freshMedia, combined] and
– TERM is the search term.
The resulting behavior is:
– RBB: returns results from the RBB white list. The unstructured search module provides media
items crawled from numerous white-listed web sites, while we also query some APIs offering white-
listed content namely the YouTube API (on specific channels) and the Arte replay programs API
(arte+7).
– SV: returns results from the Sound and Vision white list. Similarly as RBB, we query the unstruc-
tured search module and YouTube white-listed channels.
– freshMedia: returns fresh media items from the 12 social platforms.
– combined: combines results from all possible sources.
The Media Collector produces results into a unified JSON structure that first contains sources and
then items for each source:
{
www.ard.de: [],
Arte+7: [],
TwitterNative: [],
Instagram: [
{...},
{...},
{...},
{...},
{...},
{...}
{...},
],
YouTube: [],
FlickrVideos: [],
Flickr: [],
MobyPicture: [],
TwitPic: [],
Lockerz: [ ],
}
An item is described using the following attributes:
– mediaUrl: deep link to the media item itself.
– posterUrl: URL of a thumbnail (miniature) of the media item.
– micropostUrl: URL of the web document where the item is embedded (permalink), e.g. a YouTube
page.
– micropost:
◦ html: text description of the micropost in the original marked-up format.
◦ plainText: cleaned text description of the micropost where the markup and some characters
are removed.
– userProfileUrl: URL of the user that has published the micropost,
– type: type of the media item (photo or video).
– timestamp: reference time when the micropost was authored or the media item was uploaded
(date in UNIX timestamp format).
– publicationDate: publication date using human-friendly syntax ("yyyy-MM-dd’T’HH:mm:ss’Z’").

Figure 9: Screenshot of the web page at http://www.s-bahn-berlin.de/unternehmen/firmenprofi
l/historie.htm, relevant to the named entity S-Bahn in the RBB scenario
– socialInteractions:
◦ likes: total number of like, +1, heart for this media item.
◦ shares: total number of re-share, re-post of this media item.
◦ comments: total number of comments made about this media item.
◦ views: number of times that this item has been viewed
4.4 Examples
We present in this section some example queries and content returned by the MediaCollector. Similar
to the section 3.6.1, those queries are based on the LinkedTV scenarios. The search terms used
correspond to the most frequent entities detected in media fragment annotations.
We observe that results coming from white-listed web sites are really relevant for the Sound and
Vision scenario dedicated to cultural heritage material while recent media shared on social networks
make more sense in the RBB scenario focused on news content.
4.4.1 RBB scenario.
– Named entity: S-Bahn
– Query: http://linkedtv.eurecom.fr/api/mediacollector/search/RBB/S-Bahn
– Media content returned from API:
◦ source: YouTube
type: video
url : http://www.youtube.com/watch?v=jHdpP2X9tE8
◦ source: www.s-bahn-berlin.de
type: webpage
url: http://www.s-bahn-berlin.de/unternehmen/firmenprofil/historie.html
◦ source: www.mdr.de
type: photo
url: http://www.mdr.de/sachsen-anhalt/halle/halleipzig100_v-standard43_zc-698fff
06.jpg?version=54569

Figure 10: Painting of Johan and Mies Drabbe by Jan Toorop, 1898 (collection H.F. Elout), relevant to
the named entity Toorop in the Sound and Vision scenario. Source: http://www.geschiedeniszeelan
d.nl/topics/kunstindomburg_22.jpg
4.4.2 Sound & Vision scenario.
– Named entity: Toorop
– Query: http://linkedtv.eurecom.fr/api/mediacollector/search/SV/Toorop
– Media content returned from API:
◦ source: YouTube
type: video
url : http://www.youtube.com/watch?v=csVdymOUyNA
◦ source: www.geschiedeniszeeland.nl
type: photo
url: http://www.geschiedeniszeeland.nl/topics/kunstindomburg_18.jpg
◦ source: www.geschiedeniszeeland.nl
type: photo
url: http://www.geschiedeniszeeland.nl/topics/kunstindomburg_22.jpg
◦ source: www.boijmans.nl
type: photo
url: http://www.boijmans.nl/images/pages/slides/72/2090_MK_De_Theems_bij_Londen_
van_Jan_Toorop_web.jpg

5 Searching Media Resources with LSI
The Linked Services Infrastructure (LSI) is another component, provided as a REST service, which can
be integrated into the Media Collector and that provide relevant media resources (images and video)
from online media repositories for a given Linked Data (DBPedia) concept.
In the case of LSI, we focus on making use of web APIs for online media platforms such as Flickr or
YouTube, defining mapping rules between the semantic query (in terms of a Linked Data resource) and
the structural API query to the non-semantic Web API. The process of transforming a semantic query to
a structural query is called “lowering” while the transformation of the structured resource (usually JSON
or XML) to a semantic result (RDF) is called “lifting”. The use of mapping rules means that - provided the
Web API does not change - media can be retrieved from that source repeatedly with the actual mapping
only needing to be defined once. Media resource matches from different APIs is collected in parallel,
while a local store of metadata of media resources relevant to a known, expected concept has been
added to improve retrieval speed. LSI returns a list of matching media resources in RDF, with additional
metadata for each media resource which could be used in subsequent ranking and selection of “most
relevant” matches.
The rest of this section provides more details on the LSI implementation and usage. A standalone
LSI instance for testing can be found at http://production.sti2.org/lsi.
5.1 Technology used
Linked Services Infrastructure makes use of Semantic Web Service technology developed in the Open
University. The iServe19 platform acts a repository for the descriptions of the Web APIs. These descrip-
tions are themselves semantic, using the specifications hRESTs and MSM20 to generate a RDF based
machine processable description of the Web APIs inputs and outputs. Lifting and lowering rulesets (in
XSLT extended by SPARQL based inline queries) define how a Linked Data URI may be mapped into a
request to the Web API, and how the Web API response can be mapped to the LSI’s RDF model (based
on the W3C Ontology for Media Resources) for a consistent and aggregated response of relevant media
resource descriptions to the client.
The REST API allows a single service to be queried (e.g. based on its description, such as only
services who have content for a particular type of concept), a concept to be mapped to media resources
by a single service call, or a concept to be mapped to media resources aggregated from all services
called (based on their relevance to the concept in the query). Parameters allow that only image or video
content is returned, a limit placed on the amount of media contents in the response, and whether Named
Entity Recognition is used in the media resource titles and descriptions (this can reduce the latency in
waiting on a response from LSI). Named Entity Recognition is used with an optional additional “context”
parameter: the idea is that aggregating media content on the basis of a single concept can be very
general (e.g. “Berlin”), while the provision of an additional contextual concept (e.g. “Obama”) enables
more focused media content responses from the LSI.
To determine sets of concepts related to a media item, we extract named entities as Linked Data
URIs from the free text metadata of those media items such as their titles, descriptions, or user-provided
tags. We work on using the Linked Data graph to map contextual concepts to related concept sets, for
example “US Presidents” as the context to a query on “Berlin” should return images and videos related
to US presidents and Berlin, e.g. a video annotated with the concepts “Berlin + Obama” will match a LSI
query for “Berlin + US Presidents”, based on the semantic closeness of Obama to the category of US
presidents within the DBPedia graph.
5.2 Use of the REST API
The LSI API currently supports the invocation of the following APIs:
– Foursquare Venues API (https://developer.foursquare.com/overview/venues)
– Ookaboo API (about.ookaboo.com/a/ookaboo_api)
– Flickrwrappr API (wifo5-03.informatik.uni-mannheim.de/flickrwrappr)
– YouTube data API (https://developers.google.com/youtube)
19http://iserve.kmi.open.ac.uk/
20http://iserve.kmi.open.ac.uk/wiki/index.php/IServe_vocabulary

Figure 11: Search interface for ’Berlin + US Presidents’
– Instagram API (instagram.com/developer/)
– Flickr APIs (www.flickr.com/services/api/)
The Linked Service Infrastructure (LSI) provides a simple RESTful API for search and invocation of Web
services. The primary REST method usable for extracting online media resources for a Linked Data URI
is “Find Media Resources”.
5.2.1 Find Media Resources.
Request URL : http://production.sti2.org/lsi/api/invoke
HTTP Method : GET or POST
HTTP Request Header : Use the Accept header to specify the required response format. Adding
“Accept: application/rdf+xml” to the header of the request yields responses in the RDF/XML format;
Adding “Accept: text/html” to the header of the request yields responses in the format of HTML snippets.
Name Type Example
lod URI http://dbpedia.org/resource/Berlin
mediaType String all, video, image
limit Integer 10
ner Boolean yes or no
context String http://dbpedia.org/resource/US_Presidents
Table 7: Example usage of LSI for enriching content based on a search term and a context
Supported LOD URIs : The current implementation is tuned to DBPedia concepts while it will make a
best effort with any Linked Data URI (any URI from which RDF based metadata can be found and used).
– Instances of dbp:Place use geolocation information. Two properties, i.e. latitude (wgs84:lat) and
longitude (wgs84:long), will be used to assemble the actual requests to be sent to the endpoints
of services
– All the instances of rdf:resource can be used in a service call using its value of rdfs:label.
Note: for multi-lingual labels, LSI will choose the one in English.

Figure 12: Video results for the query “Berlin + US Presidents”
Example : Here is an example for invoking the LSI API via HTTP GET.
Request:
http://production.sti2.org/lsi/api/invoke?lod=http://dbpedia.org/resource/Berlin&mediaType=all&li
mit=10&ner=yes&context=http://dbpedia.org/resource/US_Presidents
Response:
The response will be in the format of RDF and conform to the W3C Ontology for Media Resource. It
contains a set of Media Resource instances with some properties and values according to what can be
extracted from the original Web API. There is a mix of media characteristics (duration of video, dimen-
sions of images) and conceptual relationships via the property hasKeyword (generated from textual titles
and descriptions processed via Named Entity Recognition tools). For example, the RDF description of
a YouTube video can look like this:
<rdf:Description rdf:about="http://production.sti2.org/lsi/media/2684e43caae442ce87e009f6c09214f2">
<ma:hasKeyword rdf:resource="http://dbpedia.org/resource/Ronald_Reagan"/>
<ma:hasKeyword rdf:resource="http://dbpedia.org/resource/Robert_Tear"/>
<ma:hasKeyword rdf:resource="http://rdf.freebase.com/ns/en/berlin"/>
<ma:hasKeyword rdf:resource="http://dbpedia.org/resource/State_Opening_of_Parliament"/>
<ma:hasKeyword rdf:resource="http://dbpedia.org/resource/Berlin"/>
<ma:hasKeyword rdf:resource="http://rdf.freebase.com/ns/en/ronald_reagan"/>
<ma:hasKeyword rdf:resource="http://dbpedia.org/resource/Ronald_Reagan_Presidential_Library"/>
<ma:title>Reagan - Tear Down This Wall</ma:title>
<ma:publisher rdf:resource="http://www.youtube.com"/>
<ma:locator rdf:datatype="http://www.w3.org/2001/XMLSchema#anyURI">
https://www.youtube.com/watch?v=WjWDrTXMgF8&feature=youtube_gdata_player</ma:locator>
<rdf:type rdf:resource="http://www.w3.org/ns/ma-ont#MediaResource"/>
<ma:duration rdf:datatype="http://www.w3.org/2001/XMLSchema#integer">240</ma:duration>
<ma:releaseDate rdf:datatype="http://www.w3.org/2001/XMLSchema#dateTime">
2007-07-04T03:17:00</ma:releaseDate>
<ma:description>Reagan’s Berlin address.</ma:description>
<foaf:thumbnail rdf:resource="https://i.ytimg.com/vi/WjWDrTXMgF8/default.jpg"/>
<ma:contributor rdf:resource="https://gdata.youtube.com/feeds/api/users/JohnJ2427"/>
</rdf:Description>
5.3 Use in LinkedTV Media Layer
LSI will be added as an additional content source for the LinkedTV Media Collector (section 4) providing
image and video resources matching the Linked Data concept with which the seed video fragment
has been annotated. A ﬁrst step in this integration will be the adaptation of the LSI response format
to be conform with the JSON syntax expected by the Media Collector, which can be extended to call
additionally the LSI REST API with the media annotation concept.
In due course, we want to explore which additional media resource properties are useful for en-
richment selection and can be provided with the needed consistency and correctness (Web APIs vary
in what information they provide about the media resources and how trustworthy those values are for
automated processing).

Finally, further research needs to consider how contextual factors can be increasingly used to re-
trieve, in the ﬁrst instance, the most relevant media resources from large scale media collections like
Flickr or YouTube, and how these contextual factors can be brought into the initial query construction to
LSI / Media Collector (e.g. Enrichments for a concept like “Chinese restaurants” could be geo-located to
be close to the viewer’s current location).

6 MediaEval task
MediaEval is a benchmarking initiative for multimedia evaluation. It offers several tasks, among which
the Search and Hyperlinking task we participated in. The MediaEval 2013 workshop will take place in
Barcelona, Catalunya, Spain, on Friday-Saturday 18-19 October 201321, just before ACM Multimedia
201322.
In this section, we will describe what the goals of this challenge are and how it connects to LinkedTV
goals, especially to the task of linking hypervideos to Web content. The following partners contributed to
the LinkedTV submission to the challenge: CERTH, EURECOM, Fraunhofer and UEP (in alphabetical
order).
6.1 The challenge
The MediaEval Search and Hyperlinking task tackles the issue of information seeking in a video dataset [7].
The scenario proposed is that of a user looking for a known video segment, and who could be interested
in watching related content proposed by the system. Hence, such a task is at the heart of LinkedTV: it
is a way to test the LinkedTV framework for hyperlinking videos together on a TV dataset, in real-world
conditions. After evaluation, such a work can be extended to hyperlinking of videos to web content.
The dataset offered for this task contains 2323 videos from the BBC, amounting to 1697 hours
of television content of all sort: news shows, talk shows, series, documentaries, etc. The collection
contains not only the videos and audio tracks, but also some additional information:
– subtitles (manually transcribed)
– two types of ASR transcripts (LIMSI [11] and LIUM/Vocapia [28])
– metadata giving show title, description, date of airing, format, etc
– additional metadata: synopsis and cast from the BBC web site)
– shot boundaries and keyframes
– face detection and face similarity information
– concept detection
6.1.1 Search task
The goal of the Search task is to search for a known video segment in the dataset, using a textual query
provided by a user. The query is constituted of two parts: the ﬁrst part gives information for a text search
while the other part gives cues on visual information in the searched segments, using words. 29 users
deﬁned 50 search queries related to video segments watched among this dataset.
Here are two examples of query:
<top>
<itemId>item_1</itemId>
<queryText>Little Britian Fat Fighters problem of gypsies in the area</queryText>
<visualQueues>fat club comedfy</visualQueues>
</top>
<top>
<itemId>item_2</itemId>
<queryText>(Two pints of lager) Sandwich Maker Dirty Shirt My girlfriend is a
housekeeper </queryText>
<visualQueues>The guy with the dirty shirt In a Pub Having a Chat and making fun of
his girlfriend </visualQueues>
</top>
The evaluation of the search task is based on the following measures:
21http://www.multimediaeval.org/mediaeval2013/
22http://acmmm13.org/

– the Mean Reciprocal Rank (MRR) assesses the ranks of the relevant segment returned for the
queries. It averages the multiplicative inverse of the ranks of the correct answers (within a given
time windows, here 60s).
– the Mean Generalized Average Precision (mGAP) is a variation of the previous that takes into
account the distance to the actual relevant jump-in point. Hence, this measure also takes into
account the start time of the segment returned.
– the Mean Average Segment Precision (MASP) assesses of the search in term of both precision
of the retrieved segments and the length of the segments that should be watched before reaching
the relevant content [9]. It takes into account the length of overlap between the returned segments
and the relevant segment. It hence favors segments whose boundaries are close to the expected
ones.
A window size of 10, 30 and 60 seconds was originally planned to report results, but in the final results
delivered, only the 60s time window was used.
6.1.2 Hyperlinking task
The Hyperlinking task aimed at offering to the viewer content from the dataset related to what (s)he is
watching. The user defines an anchor (a video segment, identified by the video name, the begin time
and end time): this anchor is the basis for seeking related content in the collection. Condition LA requires
to use only this segment for the hyperlinking, while condition LC also allowed to use the context of this
segment (i.e. a broader segment from the video).
Examples of hyperlinking queries:
<anchor>
<anchorId>anchor_1</anchorId>
<startTime>13.07</startTime>
<endTime>13.22</endTime>
<item>
<fileName>v20080511_203000_bbcthree_little_britain</fileName>
</item>
</anchor>
<anchor>
<anchorId>anchor_2</anchorId>
<item>
<fileName>v20080510_002000_bbcthree_two_pints_of_lager_and</fileName>
</item>
</anchor>
A manual evaluation by users will follow by performing crowd-sourcing evaluation through Mechanical
Turk. It was carried out for the top 10 ranks of the runs. Mean Average Precision will be reported, as
well as precision at different ranks (i.e. how many relevant targets were retrieved in the top n ranks).
6.2 Pre-processing step
First, we worked on further processing the dataset, in order to have as much information as possible.
We summarized the pre-processing performances in table 8: the processing time is fairly important due
to the size of the dataset (1697 hours of video). The data extracted is described in the next paragraphs.
6.2.1 Concepts Detection
For visual concept detection, we follow the approach presented in [29], using a sub-set of 10 base (key-
frame) detectors. The algorithm is applied on the key-frames of the video, aiming to detect 151 semantic

Table 8: Performances of the different analysis techniques on the whole dataset
Concepts Detection 20 days on 25 4-cores computers
Scene Segmentation 2 days on 6 cores
Keywords Extraction 5 hours
OCR 1 days on 10 cores
Face Detection and Tracking 4 days on 40 4-cores computers
Named Entities Extraction 4 days
concepts, both static and dynamic ones, selected from the list of concepts defined in the TRECVID
2012 SIN task [21]. The 10 used classification modules are derived from different combinations of the
employed interest point detector, descriptor and visual word assignment method. Specifically, the con-
sidered interest point detectors are the Harris-Laplace corner detector [10] and a dense pixel sampling
strategy, while the employed descriptors are the well known SIFT [15] and two colored variations of
it, named RGB-SIFT and Opponent-SIFT [34]. Then, the low-level descriptors are assigned to visual
words from two vocabularies that were created off-line through K-means clustering, employing hard-
and soft-assignment [35], respectively.
For each one of the employed classification modules, one vector per key-frame is finally extracted and
used as the actual input to the utilized SVM classifier. In order to increase the computational efficiency,
linear SVM classifiers are employed instead of kernel SVMs that are typically used for this task, while
another boost in time performance is obtained by using only one range file for the classification of the
overall set of concepts (not one range file per concept, as when the algorithm runs for a small collection
of videos). The latter results in a slightly lower detection accuracy, however reducing by 145 times the
needed processing time, which is crucial when the algorithm is applied on large collections of videos,
such as the MediaEval dataset. The output of each of the employed classifiers is a Degree of Confidence
(DoC) score for the corresponding concept, which expresses the classifier’s confidence in this concept
being suitable for annotating the current shot. This process is iterated for each considered concept and
for all used modules, and the extracted DoC scores are averaged to generate the final concept detection
score. Finally, a vector of such scores, where each element of the vector corresponds to a different
concept, is the system’s output.
6.2.2 Scene Segmentation
The employed technique for scene segmentation is based on the algorithm introduced in [30]. This
method groups the shots of the video (either automatically detected, or predefined as in the MediaEval
2013 Search and Hyperlinking task) into sets that correspond to individual scenes of the video, based
on the visual similarity and the temporal consistency among them. Specifically, one representative key-
frame is extracted from each shot of the video and the visual similarity between pairs of key-frames
is estimated via HSV histogram comparison. The grouping of shots into scenes is then performed, by
utilizing the two proposed extensions of the well-known Scene Transition Graph (STG) method [38],
which clusters shots into scenes by examining whether a link, between two shots, exists.
The first extension, called Fast STG, reduces the computational cost of shot grouping, by considering
shot linking transitivity and the fact that scenes are by definition convex sets of shots, thus limiting the
number of shot pairs whose possible linking needs to be evaluated. The latter allows for faster detec-
tion of the scene boundaries, while maintaining the same performance with the original STG algorithm.
The second extension, called Generalized STG, builds on the former in order to construct a proba-
bilistic framework, towards multiple STGs combination, alleviating the need for manual STG parameter
selection. As described in [30], this probabilistic framework can also be used for the realization of a
multi-modal approach for scene segmentation, allowing the fusion of STGs built by considering different
forms of information extracted from the video, such as low level audio or visual features, visual concepts
and audio events.
6.2.3 Keywords Extraction
Keyword extraction is based on the algorithm as presented at [33]. It is mainly based on the inverse
document frequency (TF-IDF, see [27]) paradigm and employs Snowball23 as stemmer. The stop-word
list was manually curated based on the subtitles.
23http://snowball.tartarus.org/

6.2.4 Optical Character Recognition (OCR)
For text localization, we employ the algorithm presented at [31]. The detection is based on color seg-
mentation, using statistical region merging [20]. For a refined text separation, a Gaussian model is
computed based on uniform colored connected components. For Optical Character Recognition (OCR),
we employ the widely used tesseract24 engine.
6.2.5 Face Detection and Tracking
Face analysis starts with face detection, where all frames of the videos are processed to extract faces.
For this task, we use the well-known Viola and Jones’ cascade face detector [36], or more precisely
its implementation in the C++ openCV library as improved by Lienhart and Maydt [13]. Detection is
combined with a skin color detector [18] for filtering out detections that are not likely to be faces.
The tested framework performs well on images. However, we had to adapt it to videos and to create
face tracks. We use spatio-temporal information in order to smooth the results. We link detected faces
through shots using a spatio-temporal matching of faces: if two faces in adjacent frames are in a similar
position, we assume we can match them. Linear interpolation of missing faces also relies on matching
similar bounding boxes in close but none adjacent frames through a shot. This process enables to
smooth the tracking results and to reject some false positive (when a track is too short, it is considered
as a false alarm).
6.2.6 Named Entities Extraction
For every video subtitle in the collection, we perform a named-entity recognition process using the
NERD framework [23, 24, 26, 25]. Due to the huge size of the corpora, the extraction operation has
rely exclusively in the service Textrazor25, which provides low response times and a bigger processing
quota than other available alternatives (thanks to a special academic agreement). For TV shows with
long duration, the transcripts have been divided into chunks and submitted separately in order to fit with
the maximum number of bytes supported by Textrazor. The output result is a collection of entities per
subtitle file, aligned to the NERD Ontology v0.526. In most of the cases, entities are disambiguated to
resources in the Web of Data where extra information about them can be retrieved. This information
was aimed at enriching the data. In order to use the detected concepts, we needed to be able to add
semantics on top of them.
6.3 From visual cues to detected concepts
Linking visual concepts to semantics was a key challenge of this work. Taking as input the visual queues
per search query, we have run keyword extraction operations using Alchemy API27. We have then
aligned every spotted keyword with some LSCOM concepts (currently, from a subset of more than one
hundred items currently considered in our automatic concept detection algorithms) by using a semantic
word distance based on Wordnet synsets [14]. The output includes two indexes: the confidence score,
which helps to decide how close a visual concept is related to a keyword, and the relevance, which gives
an idea of the importance of a certain concept inside the scope of the text submitted. Manual evaluations
lead us to discard concepts when the confidence score was below 0.7.
In the next section, we will see the next steps of this work: designing a framework for the search and
hyperlinking tasks.
6.4 Lucene indexing
We have used the search platform Solr (see section 3.1.2). We indexed all available data in a Lucene
index at different granularities: video level, scene level, shot level, subtitle block level, speech segments
from transcripts. Hence, searches can be performed at a chosen granularity, the next step being to
design an appropriate query.
Different indexes were created with different information, as described below. Documents were rep-
resented by both textual fields (for a text search) and floating point fields (with a float value). Float
24http://code.google.com/p/tesseract-ocr/
25http://www.textrazor.com/
26http://nerd.eurecom.fr/ontology/nerd-v0.5.n3
27http://www.alchemyapi.com/

fields are used to represent concepts: search will be performed using a range query: e.g., query for
documents where the concept Animal is between 0.63 and 1.
6.4.1 Video index
Videos were defined using their name. Different text fields enabled to store and index diverse information
taken from the metadata: title of the series the show is part of, episode title, channel the video was aired
on, synopsis, short synopsis, description, cast. We also indexed subtitle and keywords extracted, plus
the number of shot in each video(for the purpose of some search algorithm described in the next section).
6.4.2 Scene index
Scenes (created using scene segmentation, see 6.2.2) were defined using the id of the video they are
part of, with begin time and end time. Each scene was aligned with subtitles. Hence, we indexed each
scene with associated textual information: subtitle and various metadata (title, cast, synopsis, etc). Field
representing concepts were also introduced. For each concept, we stored the value of the highest score
of the concept across shots that constitute the scene.
6.4.3 Shot index
Similarly to scenes, shots were defined using the id of the video they are part of, with begin time and
end time. We indexed subtitles (aligned to each shot) as well as OCR when detected. The number of
faces per shot was also indexed but not used in the search and hyperlinking algorithms.
6.4.4 Speech segments index
We create speech segments by merging adjacent speech segments when they were uttered by the
same speaker. Speech segments don’t overlap and are connected to each other. It was prepared only
for LIMSI transcripts using information about speakers. Hence, speech segment are defined using video
id, begin time and end time. The indexed text was the corresponding transcript. We also made one
attempt to use synonyms at index time.
This algorithm produces short video segments, whose length can vary from 1 to 45 seconds. This al-
gorithm is not suitable for videos with one person speaking, because it will produce an extreme situation
with a single long segment. An improvement that can be done is to take into account dialog/trialog/mul-
tidialog based segments or more speech segment from different speakers in one video segment. In the
final runs, we didn’t submit any results based on speech segments because of poor results.
6.4.5 Sliding windows index
Sliding window algorithm [8] was used with different parameters and different schema in Apache Solr.
The index was created as described below.
First, each document is divided into sentence segments. Because of different format of data, the
sentence segments have to be prepared differently. For LIUM transcripts, a fix length “sentence” of
about 17 words was used, as it is considered as the average length of English sentence [37]. For LIMSI
transcripts, splitting is done on punctuation. Finally, for subtitles, each line in <span> tag was extracted
as one sentence.
Sentences are processed using POS (part-of-speech) tagging (for each sentence segment) to obtain
information about the number of open words they contain. Open words are those words that carry the
content or the meaning of a sentence - verbs, nouns, adjectives, adverbs and interjections. For open
class words recognition, MaxtenTagger28 was used.
Last, we index sentences as follows: we insert each sentence into sliding window and count open
words for the whole actual window (a window can contain many sentences). When the sliding window
is full, i.e. when the count of open class words is bigger than the parameter specified at startup, we
merge sentences into one text and index it as one segment (start time is start time of first sentence
and end time is end time of last sentence). We then remove the first sentence and repeat this process
until we reach the end of the document. The last segment in the document can have a smaller count of
open class words than other. This algorithm produces many overlapping segments and the density of
coverage document is high.
28http://nlp.stanford.edu/software/tagger.shtml

6.5 Search task
The search task required to divide videos into smaller segments, between 1 and 15 minutes. We submit-
ted 9 different runs by adopting two different strategies: we either used pre-constructed segments that
were indexed in the Lucene engine, or we performed queries creating segments on the fly, by merging
video segments based on their score.
For most of the search, we concatenated textual and visual query (content of queryText and visu-
alQueues) to perform the text query, as it yields better results than using the text from the textual query
only (when otherwise using the same settings).
After indexing the entire collection of videos, we found out that performing a text query on the video
index returned the accurate video in the top of the list. Hence, for some runs, we could restrict the pool
of videos that are going to be searched to a small number. The query then has to be made in two steps:
first, we query the overall video index, extracting the 20 first videos, and then we perform additional
queries for smaller segments restricted to this smaller dataset.
Search with visual cues. Text search is straightforward with Lucene/Solr: the search engine provides
a default text search based on TF-IDF values. For this reason, incorporating visual information in the
search task requires to design a new query that combines the provided visual cues and the output of
the visual concept detection algorithm. Therefore, we initially created a mapping between keywords
(i.e. terms) extracted from the visual cues query via text analysis 6.3, and the visual concepts that
are considered by the applied concept detection algorithm. Based on this mapping, we were able to
integrate the output of this algorithm (i.e. presence of visual concepts in the searched segment) into the
search task, resulting in an enriched query that includes both textual and visual information.
For each one of the considered visual concepts we defined two values: its highest score across the
entire group of key-frames/shots of the MediaEval data-set and a “valid detection” rate, calculated by
examining its presence/absence within the key-frames/ shots of the MediaEval dataset that correspond
to the top-100 highest scores for this concept. Built on this, we performed a filtering step using the
“valid detection” rate, that occurred from the top-100 highest scores, as a new confidence score about
this concept and we then discarded the concepts with a rate lower than 0.5. Afterwards, we linearly
normalized these rates between 0 and 1, by mapping the highest score among them to 1.
Finally, when a visual concept was detected from the visual cues query and if it hadn’t been rejected
by the previous filtering, we also added a range query to the Lucene query, beside the textual query:
in order for the concept to be present, its normalized score had to lie above a threshold value, i.e.
between this threshold and 1. We empirically set the threshold at 0.7. Hence, we considered that all the
normalized scores above this threshold truly contain the given concept. The range query was as follows:
Animal:[0.7 TO 1]
We submitted 9 runs in total:
– scenes-C: Scene search using textual and visual cues. No filtering by video.
– scenes-noC: Scene search using textual cues only. For comparison purposes, this run has the
same settings as the previous one, except that it makes no use of the visual concepts.
– part-scenes-C: Partial scenes search from shot boundary using textual and visual cues. This
search was made in three steps: first, we filtered the list of videos as explained earlier; then we
queried for shots inside each video and ordered them by score. As a shot is a unit that is too small
to be returned to a viewer, we completed the segment using the scenes boundaries: for each shot,
we created a segment that begins with the shot and ends at the end of the scene this shot is part
of.
– part-scenes-noC: Partial scenes search from shot boundary using textual cues only. Same settings
as the previous except from the visual concepts.
– clustering10-C: Temporal clustering shots within a video using text and visual cues. The clustering
was done as following: First, we filtered out the set of videos to search as explained earlier.
Second, we computed scores of every shot in the video, and clustered together shots that were
closer than 10 seconds apart. We added the score of shots together to form the score of the
segment.

– clustering10-noC: Temporal clustering shots within a video using text search only. Same settings
as the previous except from the visual concepts.
– scenes-S or scenes-U or scenes-I: Scene search using only textual cue from transcript (scenes-I
for LIMSI, scenes-U for LIUM) or subtitle (scenes-S). No metadata was used.
– slidingWindows-60-I or slidingWindows-60-S: Search over segments created by the sliding window
algorithm (section 6.4.5) for LIMSI transcripts (slidingWindows-60-I) and subtitles (slidingWindows-
60-S). The size of the sliding windows is 60.
– slidingWindows-40-U: The same as above for LIUM transcript with sliding window size of 40.
6.6 Hyperlinking task
For the hyperlinking task, we performed two styles of query, either using the search component or using
a feature of the Solr component.
6.6.1 Hyperlinking using the search component
Our first approach amounts to use the search component to perform the hyperlinking task. The challenge
was to create the query from the given anchor. For the first condition (LA), the text query was made by
extracting keywords from the subtitle (aligned at start time and end time), using Alchemy API29 default
keyword extraction. The visual cues were extracted from the shots contained in the anchor. If the
anchor was constituted by more than one shot, we took for each concepts the highest score over all
shots. Indeed, we considered that if a concepts appeared in at least one shot, it should be taken into
account. Then, we performed queries using the scene approach and the shot clustering approach.
For the second setting (LC), we used the following: for the textual query, we use both keywords
extracted from subtitle aligned to the anchor and to the context, and gave a higher weight to words
coming from the anchor than to words from the context. For the visual query, we took the highest score
of all concepts across the anchor and its context.
6.6.2 MoreLikeThis
Our second approach amount to use the MoreLikeThis feature of the Solr component combined with
THD (Targeted Hypernym Discovery) annotation. It is based on the premise that similar video segments
are those that have the same topic. If a user is watching a video, e.g. on the theme “drugs and
legalisation of cannabis”, we suppose that he is also interested in other videos with this subject. We
experimented with THD and MoreLikeThis components.
THD : this is an unsupervised entity discovery and classification system. Recognized entities are
enriched with links from Wikipedia (resp. DBpedia) and types from DBpedia and YAGO knowledge
bases providing high semantic interoperability30. For example, for the text “Vltava” spotted in one video
segment, THD returns annotations such as “Vltava, Rivers of the Central Bohemian Region, River,
BodyOfWater, Place,” etc.
MoreLikeThis : this feature from Apache Solr is aimed at finding documents similar to the provided
one. It compares documents based on specific fields. More information about the functionality of More-
LikeThis can be found in the official documentation of Solr31.
The hyperlinking framework has 5 main steps:
– make segments from the LIMSI transcript with sliding window algorithm
– annotate text with THD annotations and index them
– create, annotate and index temporary document from the anchor for query
– ask with MorelikeThis for similar documents
– delete temporary document (to exclude contamination of index)
29http://www.alchemyapi.com/
30https://ner.vse.cz/thd/about/
31http://wiki.apache.org/solr/MoreLikeThis

6.7 Results
In this section, we report the results of our approaches given the measures described in 6.1.1 and 6.1.2.
We do not have a baseline to compare to, nor the evaluation of other participants’ runs. Those results
will be disclosed during the day of the MediaEval workshop, which takes place in Barcelona, Catalunya,
Spain, on Friday-Saturday 18-19 October 2013. Note that all these results are still being validated by
the organizers of the benchmarking task. Bugs in the evaluation script have been found at the time of
writing this deliverable and it is likely that all scores reported in this section will change. We expect that
the updated results will be soon available at the challenge web site at http://www.multimediaeval.o
rg/mediaeval2013/hyper2013/.
6.7.1 Search task
The results of the search task are listed in table 9. Overall, runs using scenes and the sliding window
with the size 60 approach have the best performances. First, we notice than under the same conditions,
subtitles perform significantly better than any of the transcripts. This was expected since subtitles are
human-generated and are thus more accurate than transcripts. Hence, the runs that got the best results
for each measure were generated using subtitles. If the subtitles were not available, the best option
would be to use the scenes with LIMSI transcripts (scenes-I) approach as it yields the higher results
over all measures when using transcripts.
It is also interesting to note that using the visual concepts in the query slightly increases the results
for all measures (e.g., clustering10-C vs clustering10-noC). This could even be improved when more
visual concepts will be taken into account (we used a subset of 151 concepts).
Table 9: Results of the Search task
Run MRR mGAP MASP
scenes-C 0.30946346604 0.176992668643 0.195100503655
scenes-noC 0.309135738072 0.176714986483 0.194691217552
scenes-S 0.315224269655 0.163526098095 0.202065496585
scenes-I 0.261337609154 0.144401090929 0.158159589898
scenes-U 0.245763345432 0.134440892192 0.152835325636
part-scenes-C 0.22839916764 0.12414594745 0.102360579399
part-scenes-noC 0.228115789255 0.12399354915 0.102092742224
clustering10-C 0.292888150402 0.152518978931 0.181357372131
clustering10-noC 0.284924316767 0.147870395447 0.171263250592
slidingWindows-60-S 0.283311232393 0.192486941209 0.202706179669
slidingWindows-60-I 0.196479921556 0.120561880191 0.120417180025
slidingWindows-40-U 0.236816510313 0.134195367345 0.150069623509
In the following, we further look into the measures and compare the three runs that have the best per-
formances: scenes-C, scenes-S and slidingWindows-60-S. MRR favors the approaches using scenes,
scenes-S having slightly higher results. Hence, retrieving segments as scenes returns a better list of
results in terms of “corresponding” segments, within a time window of 60s. When looking at mGAP, the
sliding window approach is better: as this measure takes into account the distance of the jump-in point
to the actual start of the video, it shows that using sliding windows is more precise when it comes to a
more fine-grained analysis. Retrieved segments start points are closer to the actual ones, to an extent
that enables them to catch up with other approaches with better MRR ranking. Last, MASP measures
are very similar for scenes-S and slidingWindows approaches, the latter being higher by very little. This
measure takes into account the boundaries of the segment watched: once again, such results hint that
the sliding windows approach produces segments closer to the expected ones in term of start time and
end time. A next step of this work could then be to refine the scenes in order to preserve the ranking but
redefine the start and end times.
We can question to which extent it is possible to retrieve a segment within a very small time window,
let’s say 10 seconds. To this effect, it would be interesting to compare the degree of precision (in term of
segment boundaries) offered by the sliding windows approach with what could be produced by a human.

Table 10: Results of the Hyperlinking task
Run MAP P-5 P-10 P-20
LA clustering10 0.2764 0.3733 0.2967 0.2200
LA THDMoreLikeThis 0.4760 0.4667 0.4533 0.3533
LA scenes 0.5848 0.7467 0.6633 0.5233
LC clustering10 0.3392 0.4467 0.3733 0.2417
LC THDMoreLikeThis 0.5457 0.5733 0.5433 0.4483
LC scenes 0.7042 0.8533 0.8000 0.6683
6.7.2 Hyperlinking task
For the hyperlinking task, there were originally 98 anchors. Due to time and financial constraints, the
organizers have chosen a representative subset of 30 anchors to be evaluated via crowdsourcing. The
results can be seen in table 10: MAP is the mean average precision, while P-5, P-10 and P-20 are the
precisions at rank 5, 10 and 20.
For both LA and LC conditions, runs using scenes outperform other runs for all metrics. The THD/-
MoreLikeThis approach comes second, and the least performant run is the one using the clustering
search method. As expected, using the context increases the precision when hyperlinking video seg-
ments. It is also notable that the precision at rank n decreases with n. From this observation, we can
conclude that those approaches will suit a scenario of a search engine, where mainly the top results will
raise the user’s attention.
6.7.3 Some figures on visual concepts
As we described earlier in this document, keywords spotted in the text query through a mapping were
linked to visual concepts. We performed a filtering over concepts at two levels: first, we discarded visual
concepts where the “valid detection” score among the top 100 scores was under 0.5. Second, we also
applied a threshold of 0.7 on the mapping between text and concepts on the 50 search queries.
Table 11: Number of concepts at various stages
Filtering level # concepts before filtering # concepts after filtering # concepts final
visual level 151 47 41
mapping level 976 144 over 50 queries
As can be seen in table 11, our filtering stages drastically reduces the number of concepts that are
actually taken into account in the queries. 976 concepts were discovered in the 50 queries, and we
eventually only take 41 of them into account in the final queries: those 41 concepts are spread over
23 queries, the remaining 27 queries having no visual concept attached after filtering. However, search
results indicate that using concepts globally provides an improvement over textual query. This finding is
very promising and should lead to greater research work in that direction.
6.8 Lessons learned and future work
This challenge enabled us to test our approaches with an entire dataset of diverse shows of a TV
provider, which is very close to the LinkedTV conditions. Real users evaluated the results in the hyper-
linking task, which is of great interest for our work.
The results evidence that approaches using scenes outperform others. In the hyperlinking task, using
scenes along with context of the anchor returns results with a high precision (0.85 and 0.80) at rank 5
and 10: hence, proposing the first items to a viewer is very likely to reach the expected goal. Hence,
improving on scene segmentation, by using features such as speaker clustering, faces or semantics
may be a future direction to study.
We also observe that using visual concepts in the queries improve the results: we filtered both
concepts detected in the videos and concepts mapped from the query, and still got better results than
without using those visual cues. Those outcomes show that there is room for improvement and the use
of visual concepts is another direction that should be better explored. We see two main processes that
worth being looked at: first, at the text query level, the mapping between text and concepts; second, at
the video analysis level, the confidence in visual concepts from the video.

7 Conclusion
In this deliverable, we already described the ﬁrst software results for retrieval of enrichment for seed
media fragments. The consortium has developed a diverse set of tools for acquiring enrichment content
from various web sources including common media web services. The Unstructured search module
provides a consortium-hosted search engine, which crawls content on a set of white-listed web sites and
makes it available via full-text search exposed via a REST API. The Media Collector provides a single
point for placing a query over a range of popular web resources for media content, including Flickr and
YouTube. At the time of writing, these services have been already integrated with the LinkedTV platform,
which provides a central point for orchestrating the analysis of seed content and its enrichment. The
LSI service, which provides access to a list of API’s complementary to those interfaced by the Media
Collector, is yet to be integrated. This deliverable gave a concise speciﬁcation of the APIs exposed by
the platform32.
The validity of the approach taken has been underpinned by the participation of the LinkedTV con-
sortium in the search and hyperlinking tasks of the MediaEval contest. This task is particularly relevant
to the LinkedTV use case, since it tested the ability of the tools and techniques available within the con-
sortium to handle the problem of retrieving multimedia content relevant to a textual query and to another
video segment. This deliverable gave a brief account of the approaches taken and the results obtained.
32A more detailed description of the platform is also available in the Deliverable D5.4 [5]

References
[1] LinkedTV consortium. Deliverable 2.3: Specification of web mining process for hypervideo concept
identification, 2012.
[2] LinkedTV consortium. Deliverable 6.1: Scenario descriptions, 2012.
[3] LinkedTV consortium. Deliverable 1.3: Linkedtv annotation tool - first release, 2013.
[4] LinkedTV consortium. Deliverable 2.4: Annotation and retrieval module of media fragments, 2013.
[5] LinkedTV consortium. Deliverable 5.4: Final linkedtv integrating platform, 2013.
[6] LinkedTV consortium. Deliverable 6.2: Scenario demonstrators, 2013.
[7] M. Eskevich, J Gareth J.F., S. Chen, R. Aly, and R. Ordelman. The Search and Hyperlinking Task
at MediaEval 2013. In MediaEval 2013 Workshop, Barcelona, Spain, October 18-19 2013.
[8] M. Eskevich, G.J.F. Jones, C. Wartena, M. Larson, R. Aly, T. Verschoor, and R. Ordelman. Compar-
ing retrieval effectiveness of alternative content segmentation methods for Internet video search.
In Content-Based Multimedia Indexing (CBMI), 2012 10th International Workshop on, pages 1–6,
2012.
[9] Maria Eskevich, Walid Magdy, and Gareth J. F. Jones. New metrics for meaningful evaluation of in-
formally structured speech retrieval. In Proceedings of the 34th European conference on Advances
in Information Retrieval, ECIR’12, pages 170–181, Berlin, Heidelberg, 2012. Springer-Verlag.
[10] Chris Harris and Mike Stephens. A combined corner and edge detector. In Proceedings of 4th
Alvey Vision Conference, pages 147–151, 1988.
[11] Lori Lamel and Jean-Luc Gauvain. Speech processing for audio indexing. In Advances in Natural
Language Processing (LNCS 5221), pages 4–15. Springer, 2008.
[12] Sheldon Levine. How People Currently Share Pictures On Twitter, 2011. http://blog.sysomos.c
om/2011/06/02/.
[13] R. Lienhart and J. Maydt. An extended set of Haar-like features for rapid object detection. In Image
Processing. 2002. Proceedings. 2002 International Conference on, volume 1, pages I–900 – I–903
vol.1, 2002.
[14] Dekang Lin. An Information-Theoretic Definition of Similarity. In Proceedings of the Fifteenth In-
ternational Conference on Machine Learning, ICML ’98, pages 296–304, San Francisco, CA, USA,
1998. Morgan Kaufmann Publishers Inc.
[15] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. International Journal
of Computer Vision, 60(2):91–110, November 2004.
[16] Vuk Milicic, José Luis Redondo García, Giuseppe Rizzo, and Raphaël Troncy. Tracking and Ana-
lyzing The 2013 Italian Election. In 10th Extended Semantic Web Conference (ESWC’13), Demo
Session, Montpellier, France, 2013.
[17] Vuk Milicic, Giuseppe Rizzo, José Luis Redondo García, Raphaël Troncy, and Thomas Steiner. Live
Topic Generation from Event Streams. In 22nd World Wide Web Conference (WWW’13), Demo
Session, Rio de Janeiro, Brazil, 2013.
[18] J. See N. A. Abdul Rahim, C. W. Kit. RGB-H-CbCr Skin Colour Model for Human Face Detec-
tion. In MMU International Symposium on Information & Communications Technologies (M2USIC),
Petaling Jaya, Malaysia, 2006.
[19] Lyndon Nixon. The importance of Linked Media to the Future Web: a proposal for the linked media
research agenda. In 22nd International World Wide Web Conference companion (WWW’13), pages
455–456, Rio de Janeiro, Brazil, 2013.

[20] R. Nock and F. Nielsen. Statistical Region Merging. IEEE Transactions on Pattern Analysis and
Machine Intelligence (TPAMI), 26(11):1452–1458, November 2004.
[21] Paul Over, George Awad, Martial Michel, Jonathan Fiscus, Greg Sanders, Barbara Shaw, Wessel
Kraaij, Alan F. Smeaton, and Georges Quéenot. TRECVID 2012 – An Overview of the Goals,
Tasks, Data, Evaluation Mechanisms and Metrics. In Proceedings of TRECVID 2012. NIST, USA,
2012.
[22] Giuseppe Rizzo, Thomas Steiner, Raphaël Troncy, Ruben Verborgh, José Luis Redondo García,
and Rik Van de Walle. What Fresh Media Are You Looking For? Retrieving Media Items from
Multiple Social Networks. In International Workshop on Socially-aware multimedia (SAM’12), Nara,
Japan, 2012.
[23] Giuseppe Rizzo and Raphaël Troncy. NERD: A Framework for Evaluating Named Entity Recogni-
tion Tools in the Web of Data. In 10th International Semantic Web Conference (ISWC’11), Demo
Session, pages 1–4, Bonn, Germany, 2011.
[24] Giuseppe Rizzo and Raphaël Troncy. NERD: Evaluating Named Entity Recognition Tools in the
Web of Data. In Workshop on Web Scale Knowledge Extraction (WEKEX’11), pages 1–16, Bonn,
Germany, 2011.
[25] Giuseppe Rizzo and Raphaël Troncy. NERD: A Framework for Unifying Named Entity Recognition
and Disambiguation Extraction Tools. In 13th Conference of the European Chapter of the Associa-
tion for computational Linguistics (EACL’12), 2012.
[26] Giuseppe Rizzo, Raphaël Troncy, Sebastian Hellmann, and Martin Bruemmer. NERD meets NIF:
Lifting NLP Extraction Results to the Linked Data Cloud. In 5th International Workshop on Linked
Data on the Web (LDOW’12), Lyon, France, 2012.
[27] S. E. Robertson and S. Walker. Some simple effective approximations to the 2-Poisson model
for probabilistic weighted retrieval. In Proceedings of the 17th annual international ACM SIGIR
conference on Research and development in information retrieval, SIGIR ’94, pages 232–241, New
York, NY, USA, 1994. Springer-Verlag New York, Inc.
[28] A. Rousseau, F. Bougares, P. Deléglise, H. Schwenk, and Y. Estèv. LIUM’s systems for the IWSLT
2011 Speech Translation Tasks. In Proceedings of IWSLT 2011, San Francisco, USA, 2011.
[29] Panagiotis Sidiropoulos, Vasileios Mezaris, and Ioannis Kompatsiaris. Enhancing Video concept
detection with the use of tomographs. In Proceedings of the IEEE International Conference on
Image Processing (ICIP), 2013.
[30] Panagiotis Sidiropoulos, Vasileios Mezaris, Ioannis Kompatsiaris, Hugo Meinedo, Miguel Bugalho,
and Isabel Trancoso. Temporal Video Segmentation to Scenes Using High-Level Audiovisual Fea-
tures. IEEE Transactions on Circuits and Systems for Video Technology, 21(8):1163–1177, August
2011.
[31] Daniel Stein, Stefan Eickeler, Rolf Bardeli, Evlampios Apostolidis, Vasileios Mezaris, and Meinard
Müller. Think Before You Link – Meeting Content Constraints when Linking Television to the Web.
In Proc. NEM Summit, Nantes, France, October 2013. to appear.
[32] Thomas Steiner. A Meteoroid on Steroids: Ranking Media Items Stemming from Multiple Social
Networks. In 22nd World Wide Web Conference (WWW’13), Demo Session, Rio de Janeiro, Brazil,
2013.
[33] S. Tschöpel and D. Schneider. A lightweight keyword and tag-cloud retrieval algorithm for automatic
speech recognition transcripts. In Proceedings of the 11th Annual Conference of the International
Speech Communication Association ISCA (INTERSPEECH), 2010.
[34] Koen van de Sande, Theo Gevers, and Cees Snoek. Evaluating Color Descriptors for Object and
Scene Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(9):1582–
1596, September 2010.
[35] Jan C. van Gemert, Cor J. Veenman, Arnold W. M. Smeulders, and Jan-Mark Geusebroek. Visual
Word Ambiguity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(7):1271–
1283, July 2010.

[36] P. Viola and M. Jones. Rapid object detection using a boosted cascade of simple features. In Com-
puter Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer
Society Conference on, volume 1, pages I–511 – I–518 vol.1, 2001.
[37] G.C. Whitworth. Indian English: an examination of the errors of idiom made by Indians in writing
English. Bahri Publications, 1982.
[38] Minerva Yeung, Boon-Lock Yeo, and Bede Liu. Segmentation of video by clustering and graph
analysis. Computer Vision and Image Understanding, 71(1):94–109, July 1998.

Specification of the Linked Media Layer

More Related Content

What's hot (12)

Similar to Specification of the Linked Media Layer (20)

More from LinkedTV (20)

Recently uploaded (20)

Specification of the Linked Media Layer