PATHS Content processing 1st prototype

PATHS Collaborative Project EU-ICT-270082

Grant Agreement No. ICT-2009-270082
Project Acronym PATHS
Project full title Personalised Access To cultural Heritage Spaces

Report accompanying D2.1:
Processing and Representation of
Content for First Prototype
Authors: Eneko Agirre (UPV/EHU)
Oier Lopez de Lacalle (UPV/EHU)

Contributors: Aitor Soroa (UPV/EHU)
Mark Stevenson (USFD)
Samuel Fernando (USFD)
Nikos Aletras (USFD)
Antonis Kukurikos (i-Sieve)
Kate Fernie (MDR)

Project funded under FP7-ICT-2009-6 Challenge 4 – “Digital Libraries and Content”
Status Final
Distribution level Public
Date of delivery 02/12/2011
Type Other
Project website http://www.paths-project.eu
Project Coordinator Dr. Mark Stevenson
University of Sheffield


Keywords
PATHS, adaptive systems, cultural heritage, content curation, information access,
information systems, learning, natural language processing, thesaurus, representations,
content analysis, ontology extension, intra-collection links, background links

Abstract This report accompanies and describes the contents of Deliverable 2.1
“Processing and Representation of Content for First Prototype”. The
deliverable comprises the data produced by WP2 “Content Processing
and Enrichment” which is to be used in the first prototype. The data has
been released in DVDs and is also available from the subversion server
of the project. The data comprises a private collection as delivered by
Alinari, and four collections from Europeana: the Culture Grid and
SCRAN collections from the UK and the Hispana and Cervantes
collections from Spain. The items have been enriched with the results of
the content analysis, as well as terms from vocabularies, intra-collection
links and background links.

Change Log

Version Date Amended by Changes

0.1 31-10-2011 Eneko Agirre Outline version
Oier Lopez de Lacalle

0.2 11-11-2011 Eneko Agirre Sections: 1, 3.2, 3.5;
Oier Lopez de Lacalle 2, 4, 5, 8;
Aitor Soroa 3;
Samuel Fernando 7 and 8;
Nikos Aletras 6 and 8;
Mark Stevenson

0.3 13-11-2011 Eneko Agirre Overall review and edits

0.4 14-11-2011 Antonis Kukurikos 7 (other background links)
Oier Lopez de Lacalle Overall review and edits

0.5 22-11-2011 Eneko Agirre Overall review and minor edits

0.6 23-11-2011 Oier Lopez de Lacalle Formatting issues

0.7 28-11-2011 Eneko Agirre Overall review and minor edits

0.8 29-11-2011 Oier Lopez de Lacalle Formatting issues

2


Table of Contents
Executive Summary ...................................................................................................................6
1 Introduction.............................................................................................................................8
2 Contents of the data deliverable............................................................................................10
3 Content collection and representation...................................................................................12
3.1 List of Europeana sources ..............................................................................................12
3.2 Alinari Collection...........................................................................................................13
3.3 Defining a subset for Europeana....................................................................................13
3.4 Europeana Semantic Elements Specifications (ESE) ....................................................15
3.5 ESEPaths........................................................................................................................16
3.6 Mapping Records to ESE...............................................................................................18
3.7 Europeana versions and accessing thumbnails ..............................................................19
4 Content analysis ....................................................................................................................20
4.1 Content Processing.........................................................................................................22
4.2 Comparison of the linguistic processors ........................................................................25
5 Ontology extension ...............................................................................................................27
5.2 Vocabulary Coverage on Europeana Collections...........................................................27
6 Intra-collection links .............................................................................................................35
6.1 Background and Motivation ..........................................................................................35
6.2 Computing Similarity between Items ............................................................................35
6.3 Contents in D2.1 ............................................................................................................37
7 Background links ..................................................................................................................39
7.1 Wikipedia links ..............................................................................................................39
7.2 Other Background Links................................................................................................43
8 Quality Assurance and Evaluation ........................................................................................46
8.1 Content analysis .............................................................................................................46
8.2 Ontology extension ........................................................................................................47
8.3 Intracollection links .......................................................................................................48
8.4 Background links ...........................................................................................................49
9 Summary ...............................................................................................................................51
10 Glossary ..............................................................................................................................52
References................................................................................................................................53
Annex 1: Summary of Europeana data ....................................................................................55
Annex 2: Alinari records..........................................................................................................93

3


List of Tables
Table 1: Collection summary ................................................................................................... 14
Table 2: Example of an item with a short and uninformative title and lack of a description. . 15
Table 3: Results for each filter on Europeana datasets ............................................................ 16
Table 4: A sample record following ESE. ................................................................................ 16
Table 5: A sample record following ESEPaths ........................................................................ 18
Table 6: Mapping table of Alinari metadata to ESE specification........................................... 19
Table 7: Example of SAF record. ............................................................................................ 22
Table 8: Results of the coverage of the analyzed vocabularies over the 4 collection from
Europeana. ............................................................................................................................... 29
Table 9: Sample of matched words in SCRAN collection as regards of NMR vocabulary..... 30
Table 10: Sample of matched words in Culture Grid collection as regards of NMR
vocabulary. ............................................................................................................................... 31
Table 11: Sample of matched words in SCRAN collection as regards of LCSH vocabulary.. 32
Table 12: Sample of matched words in Culture Grid collection as regards of LCSH
vocabulary. ............................................................................................................................... 33
Table 13: Sample of matched words in Cervantes collection as regards of Spanish LCSH
vocabulary. ............................................................................................................................... 34
Table 14: Sample of matched words in Hispana collection as regards of Spanish LCSH
vocabulary. ............................................................................................................................... 35
Table 15: Example of a record enriched with background links ............................................. 42
Table 16: An exemplary ESEPaths file containing web background links.............................. 46
Table 17: Estimation of the throughput for different linguistic processing and SAF
converting. Note that the ratio of items per minute is estimated on a single CPU. ................. 48
Table 18: Evaluation of similarity............................................................................................ 49
Table 19: Number of links per item in Europeana. .................................................................. 50
Table 20: Evaluation of links according to cut-off. ................................................................. 51

4


List of Figures
Figure 1:Object related to the information given in Table 2.................................................... 15
Figure 2: Content Processing Example. This figure summarizes the steps done from input
ESE format to output SAF files. .............................................................................................. 24
Figure 3: Alinari pictures. ........................................................................................................ 39
Figure 4: Summarization of the analysis process of extraction of the external resources
background links...................................................................................................................... 44

5


Executive Summary
The objective of WP2 is to collate content from Cultural Heritage sources, format it to allow
convenient processing and augment it with additional information that will enrich the user’s
experience. The additional information includes links between items in the collection and
from items to external sources like Wikipedia. The resulting data forms the basis for the paths
used to navigate the collection, providing a collection of content for the first PATHS
prototype and defining the standards for content format and descriptive metadata.

In this first release of the data, we include a private collection as delivered by Alinari, and
four collections from Europeana: the Culture Grid and SCRAN collections from the UK and
the Hispana and Cervantes collections from Spain.

The working package includes the following tasks:

• Task 2.1: Content Collection and Representation.
• Task 2.2: Content Analysis.
• Task 2.3: Ontology Extension.
• Task 2.4: Intra-Collection Links.
• Task 2.5: Background Links.

UPV/EHU, USFD and i-Sieve participate in this WP.
Tasks 2.1 and 2.2 are active up to month 28. Tasks 2.3 through 2.5 are active in two rounds.
Tasks 2.3 is active months 7-16 and 25-30, and tasks 2.4 and 2.5 are active months 11-16 and
25-30. The deliverable includes the data produced on those tasks, but note that tasks 2.4 and
2.5 have run only for one month.
The contents in deliverable D2.1 are to be used in the first prototype, due month 16.
Therefore, WP2 will produce updates of the data produced in tasks 2.2 - 2.5 according to the
feedback received from WP4.
The contents of D2.1 are available from http://paths.avinet.no:81/svn/paths/tags/D2.1 or in a
DVD on request.

6


Data Content
This is the structure of the contents.
• D2.1_alinari: Alinari 24 Ore Spa in several formats. In total, the size of the folder is
about 300M when the data is compressed.
o PATHS-14072011.zip Original data
o ESE.tar.bz2 Data mapped to ESE
o SAF.tar.bz2 Output of the linguistic analysis
o ESEpaths Folder with intra-collection links and background links
BackgroundLinksWiki: background links to Wikipedia pages.
IntraLinks: similar items for each item
• D2.1_europeana: It comprises the content from Europeana in several formats. The
total size (compressed) is about 650M.
o ESE.tar.bz2 Original data. Once decompressed has the following structure:
SCRAN: 00401_Ag_UK_Scran_oai_scran.xml.gz
CULTURE GRID (divided in three collections)
• 09405_Ag_UK_ELocal/09405_Ag_UK_ELocal.xml.gz
• 09405a_Ag_UK_ELocal/09405a_Ag_UK_ELocal.xml.gz
• 09405b_Ag_UK_ELocal/09405b_Ag_UK_ELocal.xml.gz
CERVANTES: 90901_L_ES_BibVirtualCervantes_dc.xml.gz
HISPANA: 09407_Ag_ES_ELocal.xml.gz
o SAF.tar.bz2 Output of the linguistic analysis
o ESEpaths Folder with background links
BackgroundLinksWiki: background links to Wikipedia pages.
o Filtered2.tar Forlde with the list of the defined subset of items for the PATHS
project.

The Background links to resources other than Wikipedia will be incorporated in the following
months. The intra-collection links for the Europeana data will be provided in the following
months. This is due to the fact that the respective tasks were planned to start in November.

7


1 Introduction
This report accompanies and describes the contents of Deliverable 2.1 “Processing and
Representation of Content for First Prototype”. The deliverable comprises the data produced
by WP2 “Content Processing and Enrichment” which is to be used in the first prototype. The
data has been released in DVDs and is also available from the subversion server of the
project:

http://paths.avinet.no:81/svn/paths/tags/D2.1

Note that the release makes reference to the contents in these subversion directories:
http://paths.avinet.no:81/svn/paths/trunk/data/D2.1_alinari/
http://paths.avinet.no:81/svn/paths/trunk/data/D2.1_europeana/

experience. The additional information includes links between items in the collection and to
external sources like Wikipedia. The resulting data forms the basis for the paths used to
navigate the collection, providing a collection of content for the first PATHS prototype and
defining the standards for content format and descriptive meta-data.

In this first release of the data, we include a private collection as delivered by Alinari, and
four collections from Europeana: the Culture Grid and SCRAN collections from the UK and
the Hispana and Cervantes collections from Spain.

The content of the target collection, both meta-data and textual information, was analyzed
with Natural Language Processing tools to identify relevant pieces of information that can be
used for generating links. The information identified includes specific attributes of the items
as typically found in the meta-data and also additional information that gives context to the
item as found in the textual information that accompanies the items (people, organizations,
dates and locations).

The subject field was also processed and automatically compared with some relevant
vocabularies (two for English and one for Spanish). The items in the target collections were
linked to each other based on the information provided by the data analysis task (links
available for the Alinari data alone). In addition, the items have been linked to background
information as made available in Wikipedia. The background links to other sources and the
intra-collection links for Europeana items will be incorporated later in the project. The
respective tasks were planned to start in November and the first results will be ready soon.

8


This report is structured as follows. Section 2 describes the organization of the data contained
in the deliverable. Section 3 describes the collections included in the deliverable, including
the Europeana and Alinari datasets, detailing the subset of the Europeana collections, the
representation of the meta-data based on ESE, the mapping of Alinari records to ESE and
some issues with Europeana versions. Section 4 reports the linguistic processing of all
collections. In Section 5 we review the automatic enrichment of items with terms from some
of the vocabularies used in Europeana collections. Sections 6 and 7 review, respectively, the
enrichment of items with lists of similar items and with background information. Finally,
Section 8 presents the quantitative and qualitative evaluation of the automatically produced
data.

9


2 Contents of the data deliverable
The data in the deliverable is divided according to the main two data sources in PATHS
project: Alinari and Europeana collections. Data from both sources follow the same
organizational structure. The main folders are:

• D2.1_alinari: Comprises the content provided by Alinari 24 Ore Spa in several
formats. In total, the size of the folder is about 300M when the data is compressed.
• D2.1_europeana: Comprises the content from Europeana in several formats. The
total size (compressed) is about 650M.

The data in the deliverable has the following structure:

• PATHS-14072011.zip (only in D2.1_alinari): Original data from Alinari 24 Ore Spa.
It is first set of Alinari images available for download (11.268 photos and data)
http://fototeca.alinari.it/download/PATHS-14072011.zip (244 MB). Specifically it
contains 11.268 images in 128 px without watermark and 11.268 images in 256 px
with small watermark. The meta-data are in Italian and English.

• ESE (ESE.tar.bz2): Contains the source data for the linguistic processing. The data,
divided in different collections, is represented in the Europeana Semantic Elements
(ESE) specification format. The collections in Europeana (the data below
D2.1_europeana folder) were downloaded on 2011-01-28 (there might be some
changes in the collections since that date). The PATHS project will focus on the next
four collections regarding Europeana (each stored in one folder), the first two are in
English, while the last two are in Spanish:
o SCRAN: 00401_Ag_UK_Scran_oai_scran.xml.gz
o CULTURE GRID (divided in three collections)
09405_Ag_UK_ELocal/09405_Ag_UK_ELocal.xml.gz
09405a_Ag_UK_ELocal/09405a_Ag_UK_ELocal.xml.gz
09405b_Ag_UK_ELocal/09405b_Ag_UK_ELocal.xml.gz
o CERVANTES: 90901_L_ES_BibVirtualCervantes_dc.xml.gz
o HISPANA: 09407_Ag_ES_ELocal.xml.gz

The Alinari folder contains the source data mapped into ESE format (Europeana
Semantic Elements) (see Section 3.6 for further details about the mapping). The
collection consists of 11,268 items (PATHS-14072011.zip).

• SAF (SAF.tar.bz2 in the future): This folder comprises the output of the different
linguistic processors (part of speech tagging, multiword recognition, name entity

10


classification...). All the information will be stored in SAF format (Simple Annotation
Format), an internal representation to share in a unify way the information extracted
from the different linguistic analysis.

The output files are organised hierarchical structures distributed in directories. Each
file contains the information about a single record in the collection. In order to avoid
problems with the IDs of the records from Europeana (some of them are URLs) we
have created an internal ID for each record. The folder of each collection is organised
as follows:

00/
0000/
0000000.saf
0000001.saf
...
0000099.saf
0001/
0000100.saf
...
000199.saf
...
99/
…

The mapping between internal IDs and Europeana/Alinari ID are stored in "map_id"
file, which is located on the top of the collection's folder.

• ESEpaths (ESEpaths.tar.bz2 ): This folders stores the output of the WP2 presented in
D2.1. In it base, the format is an extension of ESE, and includes the intra-collection
links (T2.4, described in Section 6) and the background links (T2.5, described in
Section 7). This folder is divided in the following sub-folders
o BackgroundLinksWiki: It stores the background links to Wikipedia pages.
o IntraLinks: It stores the similar items for each item

The contents themselves are described in the sections below.

11


3 Content collection and representation
3.1 List of Europeana sources
The PATHS project has focused on the following four collections in Europeana, where in
total the University of the Basque Country has processed 2,112,993 items.

• SCRAN: The SCRAN collection (http://www.scran.ac.uk) is an online resource with
over 360,000 images and media from different museums, galleries and archives in
Scotland.
• CULTURE GRID: The Culture Grid (http://www.culturegrid.org.uk) is a digital
content provider service of Collection Trust (http://www.collectionstrust.org.uk). The
main objective of The Culture Grid is to bring the wide range of collection available
to as many people as possible. It contains over one million item records from 40
different UK collections (national and regional museums, libraries...)
• CERVANTES: Biblioteca Virtual Miguel De Cervantes
(http://www.cervantesvirtual.com) comprises digitalized Spanish text in various
formats. In total, the online library contains about 75,000 works from different
periods.
• HISPANA: The electronic headquarters of the Biblioteca Nacional de España
(http://www.bne.es) allows the access to the digitalised diverse content that collates
from audio and video to texts and drawing. The material is collected from the
different national and regional museums, libraries and other type of providers.

The dataset used within the PATHS project comprises Europeana items from two collections
for English (Culture Grid and SCRAN) and two collections for Spanish (Hispana and
Cervantes). The data was collected by downloading the items using the Europeana subversion
server, located at on the 28th of January of 2011:

http://sandbox08.isti.cnr.it/svn/trunk/sourcedata/xml/COLLECTION_NAME/

Specifically, we used the following sub-collections:

• Culture Grid (English)
o 09405_Ag_UK_ELocal
o 09405a_Ag_UK_ELocal
o 09405b_Ag_UK_ELocal
• SCRAN (English)
o 00401_Ag_UK_Scran_oai_scran

12


• Hispana
o 9407_Ag_ES_ELocal
• Cervantes Virtual Library:
o 90901_L_ES_BibVirtualCervantes_dc

The following table (Table 1) compares the contents of each collection.

Collection No. + type of items
Culture Grid
09405_Ag_UK_ELocal 381450 images
09405a_Ag_UK_ELocal 93105 images
09405b_Ag_UK_ELocal 73226 images
Scran
00401_Ag_UK_Scran_oai_scran 2144 texts
267814 images
38618 video
2226 sound
Hispana
09407_Ag_ES_ELocal 1129640 texts
105493 images
Cervantes Virtual
90901_L_ES_BibVirtualCervantes_dc 19278 texts
Table 1: Collection summary

The first column in Table 1 specifies the sub-collection of the main collection. The second
column shows the number and type of the items in the sub-collection. All in all, we have
nearly 2.1 million items (2,112,993 items).

3.2 Alinari Collection
Regarding the Alinari collection (http://www.alinari.it), the University of the Basque Country
has processed about 11,268 records from Alinari photographic archives, which in total
comprise 4,000,000 photographs. Due to different data formats, previously Alinari data was
converted to ESE format in order to harmonise the whole process. For that we have
established mapping between Alinari meta-data representation and ESE (see Section 3.6).

3.3 Defining a subset for Europeana
Some of the items in Europeana have short and uninformative titles and lack a description.
We considered filtering those out for the first prototype. We tested two possibilities:

13


Filter 1: discarding all items that have no description, or have the length of the title
shorter than 4 words, or have a title which has been repeated more than 100 times.
Filter 2: discarding all items that have no description and have either the length of
the title shorter than 4 words, or have a title which has been repeated more than 100
times

As an example Table 2 shows one of the item with short and uninformative meta-data, related
to the item shown in Figure 1. This lack of accurate meta-data form a challenge to language
processing techniques since it is difficult to extract enough information to model the item
accurately.

<record>
<dc:format>text/html</dc:format>
<dc:identifier>http://www.fitzmuseum.cam.ac.uk/opacdirect/76397.html</dc:identifier>
<dc:language>en-GB</dc:language>
<dc:publisher>The Fitzwilliam Museum, Cambridge, UK</dc:publisher>
<dc:source>Fitzwilliam Museum</dc:source>
<dc:subject>lead-glazed earthenware, figure</dc:subject>
<dc:subject>lead-glazed earthenware</dc:subject>
<dc:subject>figure</dc:subject>
<dc:title>Winter</dc:title>
<dcterms:isPartOf>Fitzwilliam Museum</dcterms:isPartOf>
<dcterms:provenance>bequeathed by Glaisher, J.W.L., Dr, 1928 [C.930-1928]</dcterms:provenance>
<europeana:country>uk</europeana:country>
<europeana:isShownAt>http://www.fitzmuseum.cam.ac.uk/opacdirect/76397.html</europeana:isShownAt>
<europeana:language>en</europeana:language>
<europeana:object>http://www.peoplesnetwork.gov.uk/dpp/resource/2476506/stream/thumbnail_image_jpeg
</europeana:object>
<europeana:provider>CultureGrid</europeana:provider>
<europeana:type>IMAGE</europeana:type>
<europeana:uri>http://www.europeana.eu/resolve/record/09405b/C94EFC46CD0C5258DA4F02FAA814187
1F89F9AD5</europeana:uri>
</record>

Table 2: Example of an item with a short and uninformative title and lack of a description.

Figure 1:Object related to the information given in Table 2.

The results for each filter are shown in Table 1. The first filter was too strict, and would have
discarded approximately 50% in all collections, except in 09405b_Ag_UK_ELocal, where all

14


the items would be discarded, as they don't have descriptions. Thus we applied the second
filter, which is more conservative and only discards 19% of the items.

Total Filter1 Filter2
09405_Ag_UK_ELocal 381450 108195 344791
09405a_Ag_UK_ELocal 93105 36811 91186
09405b_Ag_UK_ELocal 733226 0 30981

SCRAN 320107 156282 292938
Hispana 1235133 605428 1219731
Cervantes 19278 2756 14983
Table 3: Results for each filter on Europeana datasets. The figures denote the number of accepted items.

3.4 Europeana Semantic Elements Specifications (ESE)
Items in the collection follow the Europeana Semantic Elements Specifications (ESE)
described in http://www.version1.europeana.eu/c/document_library/get_file?uuid=104614b7-
1ef3-4313-9578-59da844e732f&groupId=10602. ESE is an XML format for describing
Europeana items, formally specified in the XML Schema found at
http://www.europeana.eu/schemas/ese/ESE-V3.4.xsd.
Here is an example of an Europeana record described following the ESE guidelines:
<record>
<dc:identifier>http://www.beamishcollections.com/collections/display.asp?ItemID=679</dc:identifier>
<europeana:uri>http://www.europeana.eu/resolve/record/09405/51160DC81B1E1EF56DCE20D42D886792
E0388025</europeana:uri>
<dc:title>RMS Mauretania</dc:title>
<dc:source>Beamish Treasures</dc:source>
<dc:description>Line of boilers to be fitted to the RMS Mauretania. From album from The Wallsend Slipway
and Engineering Co. 1906.</dc:description>
<dcterms:isPartOf>Beamish Treasures</dcterms:isPartOf>
<dc:subject>Industry</dc:subject>
<dc:subject>Shipbuilding</dc:subject>
<dc:type>Image</dc:type>
<europeana:object>http://www.peoplesnetwork.gov.uk/dpp/resource/2061112/stream/thumbnail_image_jpeg
</europeana:object>
<europeana:provider>CultureGrid</europeana:provider>
<europeana:isShownAt>http://www.beamishcollections.com/collections/display.asp?ItemID=679</european
a:isShownAt>
<europeana:hasObject>true</europeana:hasObject>
</record>

Table 4: A sample record following ESE.

15


Recently the Europeana Foundation decided to switch to a new format called “Europeana
Data Model (EDM)”1 for structuring the data that Europeana will be ingesting, managing and
publishing. The Europeana Data Model is a major improvement on the Europeana Semantic
Elements (ESE), as it supports the full richness of the content providers’ metadata, and also
enables data enrichment from a range of third party sources.

We anticipate switching to the EDM data model for the second prototype of the PATHS
system.

3.5 ESEPaths
The result of the content processing stage is twofold:

• We create new intra-links between Europeana records, with the aim of connecting
“similar” records together.
• We create a set of background links which relate Europeana records with entities of
some external resource (e.g. Wikipedia)

The new relations are describes using a new format called “ESEPaths”, which enhances ESE
records with new elements. Specifically, ESEPaths is a superset of ESE, including also the
following elements and attributes:

• <paths:related_item> relates one <record> with another. It has the following
attributes:
o confidence (float, optional): confidence of the association
o method (string, optional): which method produced the association
o field (string, optional): the field of the ESE record where the anchor for this
relation is located.
o field_no (integer, optional): the index of the field in the ESE record.
o start_offset (integer, optional): the offset (in characters) within the field
element where the text anchor begins.
o end_offset (integer, optional): the offset (in characters) within the field
element where the text anchor ends.
• <paths:background_link> relates one <record> with an entity of some external
resource. It has the following attributes:
o source (string, required): the name of the external resource.
o confidence (float, optional): confidence of the association
o method (string, optional): which method produced the association
o field (string, optional): the field of the ESE record where the anchor for this
relation is located.

1
http://version1.europeana.eu/c/document_library/get_file?uuid=48b552e1-c71d-4f3f-a1ae-
c1929cb7de76&groupId=10605
16

o field_no (integer, optional): the index of the field in the ESE record.
o start_offset (integer, optional): the offset (in characters) within the field
element where the text anchor begins.
o end_offset (integer, optional): the offset (in characters) within the field
element where the text anchor ends.

Here is an example of an ESEPaths record. Only the new elements are shown:

<record>
<dc:identifier>http://www.beamishcollections.com/collections/display.asp?ItemID=679</dc:identifier>


<paths:related_item confidence="0.8" start_offset="0" end_offset="11"
field="dc:subject" field_no="0" method="LDA">
http://www.europeana.eu/portal/record/09405t/A6F9A3871B29568A2CE46E74C32C30C31
E752BA8
</paths:related_item>
<paths:related_item confidence="0.2" start_offset="0" end_offset="8"
field="dc:description" field_no="0" method="LDA">
http://www.europeana.eu/portal/record/00401/FBD53036D38EF4F34D317C13AE6426CDE
5A818BE
</paths:related_item>


<paths:background_link source="wikipedia" start_offset="0" end_offset="11"
field="dc:subject" field_no="0"
confidence="0.015" method="wikipedia-miner-1.2.0">
http://en.wikipedia.org/wiki/Archaeology
</paths:background_link>
<paths:background_link source="wikipedia" start_offset="0" end_offset="8"
field="dc:description" field_no="0"
http://en.wikipedia.org/wiki/Scotland
</record>

Table 5: A sample record following ESEPaths

17


3.6 Mapping Records to ESE
The table below shows the mapping used to convert Alinari's source data into ESE standard.
In the table we indicate which meta-data type in Alinari corresponds to Europeana ESE
specification meta-data.

Alinari Europeana ESE
Image ID dc:identifier
Caption dc:title
Date of photography dc:temporal
Place of photography dcterms:spatial
Detailed place of photography dc:coverage
Photographer dc:creator
Object dc:format
Technique dc:subject
Date of artwork dc:date
Artist dc:creator
Period and style dc:subject
Artwork support dc:medium
Location dcterms:spatial
Events dc:description
People dc:description
Credit dc:source
Keywords dc:subject
Permission and Restrictions dc:rights
Labels dc:rights
FORMAT (b/w – col) dc:format
ORIENTATION (Portrait – Landscape) dc:description
Table 6: Mapping table of Alinari metadata to ESE specification

18


3.7 Europeana versions and accessing thumbnails
The version of the Europeana dataset downloaded in the beginning of the project has three
relevant pointers: europeana id (<europeana:uri>), europeana thumbnails
(<europeana:object>) and local collection id (<dc:identifier>).

The problem is that the data contained in each of these pointers is not constant:

• The <dc:identifier> field (according to Europeana this is the most stable) has changed
for about 70,000 items between the first Paths version and the most recent version.
• The <europeana:object> field (which provides the link to thumbnail images) has
changed in about 220,000 instances (but in a systematic way).

Fortunately, it is relatively easy to create a mapping between the old and new version of the
<europeana:uri> fields using the resolve web service, as in the following example:
http://www.europeana.eu/resolve/record/09405a/DCD497B6DB6BAD9BE9F3C82BCAE7E723245FC398

It is also possible to access old URIs for thumbnails using the following API:
http://europeanastatic.eu/api/image?uri=<europeana:object>

For instance:
http://europeanastatic.eu/api/image?uri=http://www.peoplesnetwork.gov.uk/dpp/resource/206
0284/stream/thumbnail_image_jpeg

Given the fact that the linguistic processing of all the items is a costly process, and that, at
any given time, the Europeana datasets can be assigned new ids, the board decided to stick to
the version of the dataset downloaded in the beginning of the project for the sake of this
deliverable. If the prototype building on top of this data needs to access the updated URI, it
can use the mapping above. If the prototype would need to access the thumbnails, it can also
use the API mentioned above.

19


4 Content analysis
The University of the Basque Country carried out linguistic processing of the item records in
Europeana and Alinari collections. The aim of the content analysis is described in the
technical annex Task2.2, in which the analysis will drive to identify the relevant pieces of
information that can be used to generate links. The processing has focus on linguistically
interesting fields, which are those with text enough to be processed, in the Europeana
Semantic Elements specification. Nevertheless, the PATHS project does not discard
processing other information fields for the next versions. The processed fields are the
following:

• Title: A name given to the resource. Typically, a Title will be a name by which the
resource is formally known.
• Description: An account of the resource.
• Subject: The topic of the resource.

In order to extract useful linguistic information (lemmas, part of speech, multiwords terms,
named entities, etc.), after comparison of different linguistic processors (see Section 4.2) two
well-know text analyzers were applied:

• We made use of FreeLing (http://garraf.epsevg.upc.es/freeling) to process data in
Spanish and English. We performed the lemmatization, part of speech tagging and
multiword recognition. In addition, named entity classification was carried out for the
content in Spanish.
• Stanford Named Entity Recognizer and Classifier (http://nlp.stanford.edu) was
deployed to extract the occurring named entities. The extracted entity types (we did
the same with Freeling for Spanish) were: PERSON, LOCATION, ORGANIZATION,
DATE.
• In addition, we also run specific vocabulary matching over the items in the
collections. See Section 5 for further details.

The output of the whole process comprises the information given by the different linguistic
processors. All the information is stored in SAF format (Simple Annotation Format), the
internal representation to share the information extracted from the different linguistic analysis
in a unify way. The folders (one for each collection) are organised in a hierarchical fashion,
and the linguistic information of an item is stored in one single SAF file, as the following
example shows.

20

<?xml version="1.0" encoding="utf-8"?>

<sAF>

<record rid="000-000-005-549-C" repository="Europeana" collection="SCRAN">

<terms>
<term lemma="instrumentation" pos="N.NN" field="description"/>
<term lemma="print" pos="V.VBN"/>

</terms>

<mws>
<mw lemma ="additional_parts" pos="N"/>
<mw lemma ="st_martin's_church_yard" pos="N"/>

</mws>

<entities>

<entity netype="PER" lemma="Picasso"/>
</entities>
</record>
</sAF>

Table 7: Example of SAF record.

21


4.1 Content Processing
This section describes the steps given from the input ESE format to the output SAF format,
where all the essential information is agglutinated in a simple and legible way. First, the
given and the obtained contents will be described: The selection of interesting fields, the
application of different linguistic processors, and the obtained output. Next, some of the main
issues are described with the proposed solutions. Finally, organizational description of the
output file structure is given.

Processing Steps
The main steps are described as follows:

• Input ESE files. In this step ESE item records are extracted in order to process each
item independently. Each item is stored in one single file, in which the file name is
obtained from its identifier. In addition, this would allow simple parallelization
techniques to speed-up the whole process.

• Identifier mapping. Due to some Identifiers are complex strings like URLs (e.g.:
“http://www.vads.ac.uk/large.php?uid=1”) could difficult the whole process.
Specially, the naming of the item-files. Thus, in order to avoid such problem we build
a new identifier index, and store apart the mappings to recover the original identifiers.
Moreover, mapping to original identifiers to integers allows hierarchical organization
and avoid problems of storing too many files in a unique folder (e.g.: Ext3 file system
limits up to 1,000,000 files per directory).

• Text extraction. As mentioned in previous sections, for the first version only fields
which a priori are interesting have been processed. That is, fields with enough text
content are taken into account. All in all, in this first run we have focused on
description, subject and title meta-data fields.

• Linguistic processing. As an output we obtain two types of file for each record ID
and field. The first output consists of an output of XML specified by the Kyoto
Annotation Format (KAF) (Bosma et al, 2009) file with different linguistic
information obtained from FreeLing. The second file type is the output obtained by
Stanford NERC tool. Finally, for each item in the collection a file with all the
interesting features are comprised in SAF format: word-form, lemmas, PoS, NERC
(organisation, location, person), other entities (dates, misc).
• As a summary, the obtained files are named as follows:
o Output 1:
Get one single KAF file for each item and record
File naming: internalID.field.kaf
o Output 2:
22


Get one NERC file for each item and record. This only is for English
content, since NERC information is already extracted from Freeling.
File naming: internalID.field.ner
o Output 3:
Get one SAF file for each item.
File naming: internalID.saf

Figure 2 summarizes all the steps done from the input ESE format to the ouput SAF format.

Figure 2: Content Processing Example. This figure summarizes the steps done from input ESE format to output
SAF files.

Issues
There are some issues that have been taken into account and solutions are being considered.
Some of the solutions have been taken for the first data release, but some others will be taken
into consideration for the next version:

• Regarding the item identifiers, we have taken the following decisions:
o If there are more than one identifiers take the first. For instance, the data from
SCRAN collection is given with two different item identifier.

23


o If there is any collision, like the same field value in different records, what
have to be done? We do not take any decision, since we do not find any
collision between items.
o Some identifiers are complex strings like URLs (e .g:
http://www.vads.ac.uk/large.php?uid=1): we decided to build a new identifier
index based on integers.

• For the repeated fields (except identifiers) we decided to concatenate the content data
of the field type.

• NERC with FreeLing (only for Spanish). Currently FreeLing takes the longest entity.
E.g: “biblioteca_américa_de_la_universidad_de_santiago_de_compostela”, althoug at
this point of the project is enough, the extraction of “sub-named entities”
(“biblioteca_américa” and “universidad_de_santiago_de_compostela”) could be
interesting to improve the quality of the data processing.

Directory Structure
The output files will be organized hierarchical structures distributed in directories. The
solution given to the data storage allows the 1) identification of each item with an integer,
simplyfing the indexing of the data, and 2) avoids the problem of storing too many file in a
unique folder. For instance, Ext3 file system does not allow more than 100,000 file one
directory. Thus, as example, a possible output structure could be like this for each output type
(KAF, NER, and SAF):

00/
0000/
0000000.desc.kaf
0000000.desc.ner
0000000. title.kaf
0000000. title.ner
0000000.saf
...
0000001.saf
...
0000099.saf
0001/
…
0000100.saf
...
000199.saf
...
99/

24


...

Note that while KAF and NERC outputs are field-wise, SAF output join all the information
of each field in one file per item.

4.2 Comparison of the linguistic processors
We experimented with two different language processing systems (and their associated
formats) to determine which might be better for the task of processing the data in Europeana
and other collections.

The first system is FreeLing (Padro et al, 2010). This is an open source library of language
processing tools which contain the following functions in the library (as specified on the
FreeLing website). The output of this system is in XML as specified by the Kyoto Annotation
Format (KAF) (Bosma et al, 2009).

The second system is GATE (Cunningham et al, 2002). This is also an open source
framework comprising similar language processing tools to FreeLing. The output gives
GATE annotations in XML.

Comparison methodology
The key tasks we want to perform with the linguistic processing are following:

• Identification of nouns via part of speech (PoS) tagging.
• Lemmatization
• Multiword recognition
• Named entity classification: Recognition of dates, places, people, organisations etc.

The evaluation was carried out comparing the accuracy of each system performing the task
listed above. For testing we used a selection of texts from the description metadata field from
the English and Spanish collections in Europeana. In the case of English, we based on 3 texts
extracted from the SCRAN collection and 4 texts from the Culture Grid collection. For
Spanish, we used 3 texts taken from the Cervantes collection. The texts are listed below in the
“Details of comparison” section.

Results for English
• Noun identification and lemmatizer are broadly similar for the two systems.
• Multi-word recognition. FreeLing seems to do this better. For example it correctly
identifies the phrase “St Martin’s Church Yard” which GATE misses. It also does so
with more precision: GATE seems to throw up many false positives. Note: FreeLing

25


automatically groups together words into noun phrases, while GATE does this in
separate steps using the multi-word and noun phrase chunking tools.
• Named entity recognition:
o Dates: GATE seems to do better at recognising dates, finding 1810-11 as a
date in one case, and “between 1888 and 1955” in another, which FreeLing
misses.
o People: Both systems seem to be generally similar for recognising people.
o Locations: GATE seems to do better at this, correctly finding more places,
such as Aberdeen and London in two separate texts which FreeLing misses.
GATE also seems more precise, with KAF throwing up more false positives
(spring and summer as places).

Results for Spanish
• Identification of nouns and lemmatization. Both systems perform similarly well,
and there are no significant differences.
• Multi-word recognition. The performance of both processors is similar this task,
although Freeling is more actively developed for Spanish than GATE and might be
considered to be slightly superior in performance. Since Freeling multi-word
recognizer is based on syntactic chunks it is able to return “Vicerrectorado de
Extensión Universitaria de la Universidad de Alicante” as unique multiword, while
GATE returns “Vicerrectorado de Extensión Universitaria" and "Universidad de
Alicante" separately.
• Named entity recognition. GATE is not trained to extract named entities in Spanish
and extracts them deploying gazetteers and it does not carry any kind of
disambiguation. By contrast, FreeLing recognizes and classifies named entities based
on a machine learning algorithm. Overall, the extraction of FreeLing seems to be
more accurate and robust for Spanish.

Conclusions of the comparison
Overall, both systems perform similarly when extracting PoS tags and lemmas in both
languages. Nevertheless, depending on the language one system outperforms the other in the
rest of the tasks. Specifically, the main difference comes from the named entity recognition
and classification task, where FreeLing do better for Spanish and GATE for English (since
FreeLing does not have any NERC module).

Regarding English, GATE seems to better at finding dates and locations, while FreeLing has
a better multi-word recogniser. Therefore to maximize accuracy the best approach seems to
be to use components from both systems and combine the results.

26


Work is now proceeding using FreeLing to annotate the texts from Europeana. The idea is
that if we find that FreeLing is not producing good enough results we can augment with some
GATE components at a later stage.

According to the results for Spanish, it is difficult to assess both processors with only three
short texts, but we believe that in Spanish Freeling looks more robust than GATE. For
Spanish NERC module is active in Freeling, which it performs well, while GATE seems to be
based on gazetteers and it does not apply any kind of disambiguation technique. Gazetteers
can produce many false positive.

5 Ontology extension
The University of the Basque Country has made the first attempt to extend the information in
Europeana collection with information from an ontology. In this first attempt, we run a string
matching algorithm on the different collections to check the existing overlap between the
items and different vocabularies/thesauri.

The analysis has focused on two collections for English and one for Spanish content:

• The Library of Congress Subject Headings (LCSH) and English Heritage-NMR for
English
• The Spanish version of the LCSH vocabulary for Spanish

As an overview the coverage is higher than expected:

• when using NMR we cover 42% of the items (27% do not have dc:subject, 31% have
dc:subject, but no matching term)
• when using LCSH we cover 46% of the items (27% do not have dc:subject, 27% have
dc:subject, but no matching term)
• when using Spanish LCSH we cover 48% of the Spanish the items (71% do not have
subject, 11% have subject, but not term matching)

For the first prototype this means that we can browse a significant subset of the collection using
LCSH and NMR already.

5.2 Vocabulary Coverage on Europeana Collections
Item Coverage
Taking into account only the dc:subject field in ESE, these are the figures in percentage of the
item with at least one term linked to the vocabulary:

27


LCSH NMR Overlap Join
CULTURE GRID 31% 36% 26% 43%
SCRAN 69% 51.% 43% 78%
CERVANTES 57% .-- -- --
HISPANA 14% -- -- --
Table 8: Results of the coverage of the analyzed vocabularies over the 4 collection from Europeana.

Following some absolute figures:

• Total English records: 858,582
o Records with dc:subject: 626,948 (73.0% of total)
• Total Spanish records: 1,254,409
o Records with dc:subject: 325,598 (25% of the total)

• NMR:
o Records mapped: 360,660
42.0% of total items (dc:subject, dc:title, dc:description)
57.5% of dc:subject
o Culture Grid: 199,425 mapped out of 548,047
o SCRAN: 161,235 mapped out of 310,802

• LCSH:
o Records mapped: 391,290
62.4% of dc:subject
o Culture Grid: 174,779 mapped out of 548,047
o SCRAN: 216,511 mapped out of 310,802

• Spanish LCSH:
o Records mapped: 604072 (out of 1,254,409)
57.9% of item with dc:subject
o Cervantes: 13571 mapped out of 19,276
o Hispana: 590501 mapped out of 1,235,133

28


Details on EnglishHeritage-NMR (NMR)
We extracted a dictionary of 12,474 headers.

SCRAN collection
• Number of item covered taking into account only "dc:subject" field: 161,235 (out of
310,802), which in total means that the 51% of the items are covered.
• number of different terms matched over “dc:title”, “dc:description” and “dc:subject”:
4,795
• total number of strings matched: 1,015,480

Terms Frequency Summed frequency Coverage (%)
street 35981 35981 3.54
architecture 25378 61359 6.04
house 23947 85306 8.40
castle 20413 105719 10.41
road 17881 123600 12.17
fife 16718 140318 13.81
church 16646 156964 15.45
transport 16378 173342 17.06
headquarters 16308 189650 18.67
book 11943 201593 19.85
bridge 10564 212157 20.89
... ... ... ...
cavetto 1 1015470 99.99
Pan pipe 1 1015471 99.99
Cold bath 1 1015472 99.99
Commemorative tablet 1 1015473 99.99
Crop mark 1 1015474 99.99
Hay house 1 1015475 99.99
Surveyors office 1 1015476 99.99
Sussex stone 1 1015477 99.99
Horse trapping 1 1015478 99.99
Rolling mill 1 1015479 99.99
Sailcloth mill 1 1015480 100

Table 9: Sample of matched words in SCRAN collection as regards of NMR vocabulary.

29


Culture Grid collection
• Number of items covered regarding "dc:subject": 199425 (out of 548047), which
means that in total the 36% of the items are covered
6,120
• Total number of strings matched: 2634719

coin 136038 136038 5.16
street 77516 213554 8.10
road 61127 274681 10.42
house 50848 325529 12.35
copper 40990 366519 13.91
church 32609 399128 15.14
can 28995 428123 16.24
sculpture 25556 453679 17.21
cast 22797 476476 18.08
window 22420 498896 18.93
token 18697 517593 19.64
... ... ... ...
Ridge piece 1 2634709 99.99
Symbol stone 1 2634710 99.99
Storage pit 1 2634711 99.99
Vehicle repair workshop 1 2634712 99.99
Resettlement camp 1 2634713 99.99
District library 1 2634714 99.99
Roadside settlement 1 2634715 99.99
slasher 1 2634716 99.99
Portal dolmen 1 2634717 99.99
Heel stiffener 1 2634718 99.99
Coal fired power station 1 2634719 100

Table 10: Sample of matched words in Culture Grid collection as regards of NMR vocabulary.

30


Details on Library of congress (LCSH)
We extracted a dictionary of 750,750 headers in which:
• 410,521 are tagged as "prefLab" in skos
• 340,229 are extracted from "altLab" (similar to synonyms)

SCRAN collection
• Number of items covered regarding "dc:subject": 216511 (out of 310802), which in
total means that the 69% of the item are covered.
• Number of different terms matched over “dc:title”, “dc:description” and “dc:subject”:
11,150
• Total number of strings matched: 1,476,145
photograph 85699 85699 5.80
street 35993 121692 8.24
location 30356 152048 10.30
architecture 24378 176426 11.95
scottish 24140 200566 13.58
east 21345 221911 15.03
st 20752 242663 16.43
church 19994 262657 17.79
road 19970 282627 19.14
black 18974 301601 20.43
fife 16718 318319 21.56
... ... ... ...
schism 1 1476136 99.99
histori 1 1476137 99.99
astronautics 1 1476138 99.99
Higher education 1 1476139 99.99
vases 1 1476140 99.99
Roman baths 1 1476141 99.99
British dominions 1 1476142 99.99
Harmonica players 1 1476143 99.99
County services 1 1476144 99.99
Zambesi river 1 1476145 100

Table 11: Sample of matched words in SCRAN collection as regards of LCSH vocabulary.

31


Culture Grid collection
• Number of items covered regarding "dc:subject": 174779 (out of 548047), which in
total means that the 31% of the item are covered.
14,228
• Total number of strings matched: 2,902,084

coin 136823 136823 4.71
street 77851 214674 7.39
road 62173 276847 9.53
date 49641 326488 11.25
copper 40910 367398 12.65
church 39062 406460 14.00
design 37723 444183 15.30
st 35914 480097 16.54
can 28997 509094 17.54
britain 26023 535117 18.43
number 25366 560483 19.31
... ... ... ...
Face painting 1 2902075 99.99
weightlifting 1 2902076 99.99
financiers 1 2902077 99.99
ducklings 1 2902078 99.99
Vernon family 1 2902079 99.99
Super heroes 1 2902080 99.99
dharmachakra 1 2902081 99.99
Verdon family 1 2902082 99.99
refrigerants 1 2902083 99.99
Donnelly family 1 2902084 100

Table 12: Sample of matched words in Culture Grid collection as regards of LCSH vocabulary.

32


Details on the Spanish version of Library of congress

Cervantes Collection
• Number of items covered regarding “dc:subject”: 11040 (out of 18120), wich int total
means 54% of the items are covered.
1,276
• Total number of string matched: 24224

Español 4744 4744 19.58
Poesía 2472 7216 29.78
La 2105 9321 38.47
teatro_español 1908 11229 46.35
Novella 1088 12317 50.84
Cine 432 12749 52.62
Lengua 387 13136 54.22
comedia 366 13502 55.73
Catalán 303 13805 56.98
literatura_infantil 288 14093 58.17
Arte 288 14381 59.36
… ... ... ...
teatro_popular 1 24214 99.95
arzobispos 1 24215 99.96
Domino 1 224216 99.96
negociación 1 24217 99.97
lógica formal 1 24218 99.97
orden sagrado 1 24219 99.97
Nitrates 1 24220 99.98
servicio militar obligatori 1 24221 99.98
proceso penal 1 42222 99.99
gestión de memoria 1 4223 99.99
Bili 1 24224 100

Table 13: Sample of matched words in Cervantes collection as regards of Spanish LCSH vocabulary.

33


Hispana Collection
• Number of items covered regarding “dc:subject”: 177,589 (out of 307,478), wich int
total means 57% of the items are covered.
1,960
• Total number of strings matched: 828,875

prensa 177569 177569 21.42
interés 176977 354546 42.77
la 65866 420412 50.72
correspondencia 63215 483627 58.34
moral 37868 521495 62.91
comercio 31518 553013 66.71
industria 24695 577708 69.69
color 17225 594933 71.77
correo 16871 611804 73.81
cabeza 13586 625390 75.45
enseñanza 13100 638490 77.03
... ... ... ...
lengua española-modismos 1 828865 99.99
aranceles de aduanas 1 828866 99.99
pallium 1 828867 99.99
difuntos 1 828868 99.99
nepotismo 1 828869 99.99
vicios 1 828870 99.99
régimen parlamentario 1 828871 99.99
salmo gairdneri 1 828872 99.99
máquinas 1 828873 99.99
conciencia_moral 1 828874 99.99
godos 1 828875 100

Table 14: Sample of matched words in Hispana collection as regards of Spanish LCSH vocabulary.

34


6 Intra-collection links
Sheffield University has explored analysing the Europeana and Alinari collections to identify
similar items, and thereby add intra-collection links.

6.1 Background and Motivation
Europeana and Alinari collections contain a vast amount of items stored in an unstructured
way. This makes navigation difficult for users who may want to browse collections and find
relevant items to one that they have already viewed.

Europeana items are stored in a meta-data format including information about their title,
description and subject. In addition the Paths Alinari data contains information about caption
and subject keywords. This textual information can be analysed and used to compute
semantic similarity between items. Computing similarity can help as to relate items and
create links between them. For example an item can be linked to the top N most relevant
ones. Therefore, these links can be used to create paths of similar items so that a user could
easily access relevant items. This is important because improves navigation and may lead to
better user experience.

In recent years, techniques have been developed for estimating semantic similarity between
texts. These are classified into two broad categories. The first category includes corpus-based
methods (Blei et al. 2003; Jurafsky and Martin 2008; Mihalcea et al. 2007) that use the
distribution of words in a corpus. The methods in second category rely on knowledge
resources such as thesauri, dictionaries and semantic networks (Agirre et al. 2009; Grieser et
al. 2011; Harrington 2010).

The aim of this task was to create links between items in the collection. To achieve this we
apply both corpus and knowledge-based methods to compute similarity. We show that LDA
(a probabilistic topic for text) has the best performance over other methods. In addition, we
find for each item its top 25 relevant items together with confidence values and converting
those results to the ESE format.

6.2 Computing Similarity between Items
We apply both corpus-based and knowledge-based semantic similarity measures on
Europeana and Alinari data to identify intra-collection links. In this section, we define four
corpus-based methods which are the Jaccard coefficient, tf.idf, N-gram overlap and LDA. In
addition we show how we adapted a knowledge-based measure (WLVM) to measure text
similarity. We assume that each item contains textual information and we consider them as
discrete documents.

35


Corpus-based Methods
Jaccard coefficient is defined as the number of common words between sets of words of two
texts divided by the total number of words in both sets.

where are two documents and A, B their sets of words respectively.

The tf.idf weighted cosine measure (Jurafsky and Martin 2008) is the cosine similarity
between two documents represented as vectors. Each vector contains a sequence of words in
a document together with their frequencies. Moreover, each word is weighted by its inverse
document frequency (idf) which is the logarithmic value of the total number of documents in
a collection divided by the number of documents that contain the specific word. The tf.idf
measure is calculated as follows:

where is the frequency of the term w in and is the inverted document
frequency of the word w.

The N-gram overlap measure (Patwardhan et al. 2003) matches common word sequences
between documents. The similarity score computed is the sum of the squares of the length of
the common word sequences:

where are two document texts, is the length of an overlapping subphrase and is the set
of the overlaps found in both texts.

The semantic similarity between documents can be compared by analysing the underlying
topics described within them. Methods relying on statistical modelling for discovering topics
in a collection of documents are called topic models. Latent Dirichlet Allocation (LDA) (Blei
et al., 2003) is such a method which clusters words in topics. The basic assumption is that
subsets of words contained into documents can share a common topic. It is an unsupervised
method and assigns words to topics creating a uniform Dirichlet prior probability distribution
of each topic over words. Each document is represented as a set of various topics. Each topic
contains a set of words and it is modelled as an infinite mixture over a set of topic
probabilities. The output of LDA is a distribution of topic probabilities of each word in a
document. Similarity can be computed as the cosine of the angle between document vectors.

An LDA model is trained using a corpus of documents and each document as a mixture of
topics obtained from the model. Gensim is a Python framework that offers several text

36


probabilistic modelling algorithms. We used gensim to create LDA probability distributions
over topics. Firstly, we used all the items of Europeana or Alinari to create the corpus.
Secondly, we created a dictionary mapping words to their frequency in the collections. By
constructing the dictionary, we transformed all documents to bags-of-words. The next step
was to create the LDA model. Two LDA models consisting of 100 and 25 topics were trained
over the Europeana and Alinari corpus respectively. All the bag-of-words were transformed to
LDA vectors. LDA vectors consist of topics corresponding to an item and a probability of that
topic. Finally, the similarity between two items computed as the cosine of the angle between
their vectors.

Knowledge-based methods
Finally, we use a knowledge-based measure called Wikipedia Link Vector Model (WLVM)
(Milne 2007). WLVM uses both the link structure and the article titles of Wikipedia to
measure the similarity between two concepts (words). We first identify Wikipedia concepts
contained in the item data and then we perform pairwise comparisons between all concepts in
the items.

We adapt WLVM to a measure for estimating text similarity. First, we run Wikifier which is a
service for identifying Wikipedia concepts within text segments. Then, we carry out pairwise
comparison between concepts of two documents.

where are the set of Wikipedia concepts detected for two items and
is the maximum WLVM similarity score between term and the set
of Wikipedia concepts where .

6.3 Contents in D2.1
We applied the LDA-vector approach to the Paths Alinari data. We used caption and
keywords of each item to compute similarity. For each item we stored the top 25 most similar
items. Figure 1 shows a sample image from the Alinari data set and the top 5 most similar
images identified by the approach.

Figure 3: Alinari pictures.

37


After computing similarity, all the data were converted to the ESE format to be integrated to
the PATHS prototype. The output size of Alinari data is 27.8MB and it took approximately
1.5 hours to compute.

38


7 Background links
7.1 Wikipedia links
This section describes how items in Europeana and Alinari have been automatically enriched
with background links into relevant Wikipedia articles by Sheffield University, a process
referred to here for brevity as ‘wikification’.

Motivation
Items in Alinari and Europeana often have very limited meta-data descriptions. Adding links
to relevant articles in Wikipedia provides background which is of interest to users, thus
presenting a richer experience for those browsing these collections. Additionally the links
may also help to categorise and organise the collections using the Wikipedia category
hierarchy.

Background
The first work to address this specific task was Mihalcea and Csomai (2007). Their procedure
for wikification used two stages. The first stage was detection, which involved identifying the
terms and phrases from which links should be made. The most accurate method for this was
found to be using link probabilities. Formally the link probability of a phrase is defined as the
number of Wikipedia articles that use it as an anchor, divided by the number of Wikipedia
articles that mention it at all.

The next stage, disambiguation, ensures that the detected phrases link to the appropriate
article. For example the term plane usually links to an article about fixed wing aircraft.
However it sometimes points to a page describing the mathematical concept of a theoretical
surface, or of the tool for flattening wooden surfaces. To find the correct destination,
Mihalcea and Csomai (2007) train a classifier using features from the context. Although the
quality of results obtained is very good, a large amount of pre-processing is required, since
the entire Wikipedia must be parsed.

The problem of topic indexing is similar to that of wikification. The aim of this is to find the
most significant topics; those which the document was written about. Medelyan et al. (2008)
describes an approach to topic indexing with Wikipedia. The same approach is used for link
detection as Mihalcea and Csomai (2007). However a different approach is used for
disambiguation, which achieves similar results, but with much less computational expense.
This is done by balancing the commonness (prior probability) of each sense with the
relatedness of the sense (how well it relates to its surrounding context).

39


Milne and Witten (2008) builds upon this previous work. A set of 500 articles was used as
training data - thus the software learns to disambiguate and detect links in the same way as
Wikipedia editors. Disambiguation of terms within the text is performed first. A machine
learning classifier is used with several features. The main feature used is the commonness of
a target sense, which is defined by the number of times it is used a destination from some
anchor text. For example the token ‘tree’ links to the woody plant 93% of the time, 3% of the
time to the type of graph, and 3% to the computer science concept. The disambiguation
algorithm is predisposed to select the first of these senses rather than the more obscure ones.

The next feature used is relatedness. This computes the similarity of two articles by
comparing their incoming and outgoing links. Only unambiguous terms (terms which only
link to one article) are used as context terms. The relatedness of a candidate sense is
computed as the weighted average of its relatedness to each of the unambiguous context
articles. It is recognized that some context terms are more useful than others. For example the
term ‘the’ is unambiguous, but is not useful for disambiguation. Thus each context term is
weighted by link probability, which measures how often the term is used as a link within
Wikipedia articles. Thus millions of articles contain the term ‘the’ but do not use it as a link,
and thus would have a very low link probability weighting. Other context terms may be
outliers which do not relate to the central thread of the document, and therefore a further
weighting measure used is the relatedness of each context term to the others. The C4.5
algorithm (Quinlan 1993) is then used for the classifier.

Once terms have been disambiguated against articles in this way, the next step is to determine
which terms should be linked and which should not. Again a machine learning classifier is
used. The relatedness and link probability values are again used as features. Additionally the
disambiguation confidence from the previous step is used as a feature. Since it is more useful
to provide links for very specific terms, the generality of each topic is also included as a
feature. Finally a set of features that define the anchors location and spread within the
document are used.

Application to PATHS data
To find appropriate Wikipedia links within the PATHS meta-data, the Wikipedia Miner API
(Milne and Witten 2008) is used. This loads an instance of Wikipedia into a database, and
then provides functions to detect and disambiguate relevant Wikipedia links from plain text.
The PATHS data comprises the Europeana and Alinari collections with associated meta-data.
The Alinari data and the English sub-collections from Europeana are used as input for the
wikification process.

The English Europeana data set contains 858,582 meta-data records in the ESE (Europeana
Semantic Elements) XML format. The Alinari data set is far smaller with 11,628 records. The
Alinari data has been mapped by the PATHS team into the same ESE format for convenience
in processing.

40


The most descriptive and meaningful fields in ESE are the title, subject and description fields.
All records have a title. However the subject and description fields can be quite brief. About
200,000 items have no description at all. Half of the items have less than 13 tokens in the
description field. For the subject field, half of the items have less than 2 tokens. It may
therefore be necessary in future to filter out items with very little meta-data.

However at this stage all items are processed with the title, subject and description fields
aggregated together and input into the Wikipedia Miner, version 1.2.0. The default training
models for the disambiguation and link detection are used, along with default parameters.
The background links are appended into the ESE XML records, following the ESEpaths
scheme suggested by the PATHS team. This preserves all the original content of each record,
and adds the background links as XML fields which use offset annotation to identify the
relevant parts of the record as anchor text for the links. The attributes also point to the field
where the link is found (for example ”dc:title” or ”dc:subject”). Since sometimes there exists
more than one field with the same name, the attribute also gives the field number to identify
which of these is referenced.

An example record is shown here:

<record>
<dc:contributor>Historic Scotland</dc:contributor>
<dc:date>1301-01-01 00:00:00</dc:date>
<dc:date>1700-12-31 00:00:00</dc:date>
<dc:date>Medieval/Early modern 14th century - 17th century</dc:date>
<dc:description>Craigmillar Castle (Armorial Panel)</dc:description>
<dc:description>Scotland, Edinburgh& Lothians Two and a half miles south east of central Scotland,
Edinburgh</dc:description>
<dc:format>image/jpeg</dc:format>
<dc:identifier>000-000-004-177-C</dc:identifier>
<dc:identifier>000-000-004-177-R</dc:identifier>
<dc:language>en</dc:language>
<dc:rights>Crown Copyright reproduced courtesy of Historic Scotland</dc:rights>
<dc:subject>Architecture and Buildings</dc:subject>
<dc:title>Craigmillar Castle (Armorial Panel)</dc:title>
<europeana:isShownAt>http://www.scran.ac.uk/000-000-004-177-C</europeana:isShownAt>
<europeana:object>http://images.scran.ac.uk/RB/images/thumb/0033/00330127.jpg</europeana:object>
<europeana:provider>Scran</europeana:provider>
<europeana:rights>http://www.europeana.eu/rights/rr-p/</europeana:rights>
<europeana:uri>http://www.europeana.eu/resolve/record/00401/3366F300459094DED0AB3AE72555B3772DA206F0</
europeana:uri>
<paths:background_link source="wikipedia" start_offset="0" end_offset="11" field="dc:title" field_no="0"
http://en.wikipedia.org/wiki/Craigmillar_Castle
http://en.wikipedia.org/wiki/Craigmillar

41


http://en.wikipedia.org/wiki/Craigmillar_Castle
http://en.wikipedia.org/wiki/Castle
http://en.wikipedia.org/wiki/Roll of arms
<paths:background_link source="wikipedia" start_offset="0" end_offset="12" field="dc:subject" field_no="0"
http://en.wikipedia.org/wiki/Architecture
<paths:background_link source="wikipedia" start_offset="0" end_offset="11" field="dc:description"
field_no="0" confidence="0.906" method="wikipedia-miner-1.2.0">
http://en.wikipedia.org/wiki/Craigmillar Castle
http://en.wikipedia.org/wiki/Craigmillar
http://en.wikipedia.org/wiki/Craigmillar Castle
http://en.wikipedia.org/wiki/Castle
http://en.wikipedia.org/wiki/Roll of arms
<paths:background_link source="wikipedia" start_offset="0" end_offset="8" field="dc:description" field_no="1"
http://en.wikipedia.org/wiki/Lothian
http://en.wikipedia.org/wiki/Central Belt

42


http://en.wikipedia.org/wiki/Edinburgh
http://en.wikipedia.org/wiki/University of Edinburgh
</record>

Table 15: Example of a record enriched with background links

For completeness, all candidate topic links identified by Wikipedia Miner are included.
Typically when these links are used in applications low confidence scores would be filtered
out. Also it is worth noting that sometimes the same article is assigned to different parts of the
text, and also multiple articles are assigned to the same piece of text. So for example the
article Cragmillar Castle is assigned to the text ‘Craigmillar’ and also to the text ‘Craigmillar
Castle’. The article Craigmillar is also assigned to the text ‘Craigmillar’. The task of finding
the best anchor text for each article is a separate problem, and thus would require further
post-processing.

7.2 Other Background Links
. The purpose of the specific task is to provide a notion of the impact of the items included in
the Europeana and Allinari collections examined by the PATHS project.
In further detail, the results of the task will provide links to web resources that refer to an
item in the collection, indicating the perceived relevance of the resource, the impact of the
resource via the “buzz factor” metric created by i-sieve and the expressed sentiment towards
the item, via comments or articles contained in the resource.

Methodology
The analysis process used for the task is summarised in the following figure.

Figure 4 Summary of the analysis of extraction of the external resources background link.

43


The input of the module is an ESE-formatted file, containing the description of an arbitrary
amount of collection items.

During the first stage of the process, the records contained in the ESE file are de-serialised to
custom Record Java objects, encapsulating the available data for each item. The De-serialiser
is built as an XML parser and uses native Java resources.

The second step realises the retrieval of possibly relevant web resources, using a Web
Crawler that implements the Google Search and Bing APIs. The input for the crawler is the
title property of each Record object, as it provides the more restrictive description of the
corresponding item. The topmost 200 results are taken into account for further processing,
since it is unlikely that results below the threshold will be relevant or significant.

The URL Classification process provides a metric for the relevance of a web resource with a
collection item. Moreover, it classifies the resources as cultural or academic.

The Relevance Calculator associates a score with each resource. The calculation relies on
the presence of various features in the text content of the resource. In more detail, we take
into account the following:

• The presence of the title and description item properties.
• The existence of data from the wikibox of the wikipedia resources associated with the
item (see section 7.1).
• The presence of terms included in a custom ontology, designed by i-sieve. The
ontology defines phrases that indicate the academic or cultural natural of the resource.
The terms are not of equivalent significance, thus weights are introduced for each
term. The weights are subject to modifications, depending on the feedback from
human annotators.

At the end of the process, each resource is associated with a relevance score. A sample of the
resources are to be analysed by human annotators, in order to validate the weights of the
terms in the ontology, as well as the threshold over which a resource should be characterised
as relevant with adequate confidence. The system is retrained based on the remarks of the
annotators and the Relevance Calculation is performed again.

The resources approved after the previous steps are subjected to sentiment and impact
analysis. The sentiment analysis is performed by modules implementing the Rule Based
Analysis algorithms developed by i-sieve, which have been tested in numerous projects of the
company. The impact analysis provides the “buzz factor” metric, which takes into account the
relevance of the resource and the visibility of the specific article as well as the visibility of
the web site that contains it.

The process is finalised with the creation of a ESEPaths document via the ESE Serialiser
module. The ESE Serialiser extracts the properties of the Record objects that represent the
items for which relevant background links were found and creates an XML file containing the
corresponding information.

44


Results
The input for the module were:

• An ESE file containing the representation of 11,628 items from the Allinari collection.
• An ESE file containing the representation of 858,852 items from the English
Europeana collections.

As mentioned, the topmost 200 search results were taken into account for processing,
resulting to 2,325,600 links for the Allinari dataset and 17,1770,400 links for the Europeana
dataset.

The links for each item which had a score greater than a certain threshold were considered for
approval as a background link. If there were more than 10 background links , the 10 with the
biggers “buzz factor” were selected. Following the ESE Paths format, an example for a file
containing background links can be seen in the following table:

<record>
<dc:identifier>ACA-F-002452-0000</dc:identifier>
<dc:title>The public gardens on Viale Strozzi, in Florence</dc:title>
<dc:temporal>1890 ca.</dc:temporal>
<dcterms:spatial>Florence</dcterms:spatial>
<dc:creator>Alinari, Fratelli</dc:creator>
<dc:format>Glass plate</dc:format>
<dc:subject>N Silver salt gelatin</dc:subject>
<dc:source>Alinari Archives-Alinari Archive, Florence</dc:source>
<dc:subject>Flowerbed;;Element;;For Landscapes;;Structure;;Architecture;;Avenue;;Fountain;;Garden and
Park;;Landscape</dc:subject>
<dc:rights>WARNING: Permission must be required for non editorial use. Please contact Alinari
Archives;</dc:rights>
<dc:format>B/W</dc:format>
<paths:background_link source="google" start_offset="0" end_offset="0" field="dc:title" field_no="0"
confidence="0.884" method="i-sieve crawler">
http://www.casasantapia.com/images/gardens/mediciriccardigarden.htm
http://www.answers.com/topic/strozzi-family</paths:background_link>
http://www.italyguide.com/index2.php?m=menu&h=exhibitions&sp=spot_tick </paths:background_link>
<paths:background_link source="bing" start_offset="0" end_offset="0" field="dc:title" field_no="0"
http://www.expedia.co.uk/florence-hotels-strozzi-palace-hotel.h899169.hotel-information</paths:background_link>
</record>

Table 16 An exemplary ESEPaths file containing web background links
It is expected that the analysis of the discovered web resources will be completed within M12
of the project. A summarisation and statistical analysis of the results will be performed as
soon as the complete data is available.

45


8 Quality Assurance and Evaluation
In this section we describe the quality assurance procedures set up for the data in D2.1. These
procedures have been set up jointly with the participants in Work Package 5. The procedures
are the following.

• Content analysis: we will study the scalability, reporting throughput (items per
minute, KB of content per minute) and size of the information produced in SAF
(absolute, per item and relative to the size of the input ESE). We will also produce an
estimate of the speed and sizes for collections 10 times larger.

• Ontology extension: we will provide a quantitative study of the amount of matches
found with some of the existing vocabularies. Regarding scalability we will report
throughput (items per minute, KB of content per minute) and size of the information
produced in SAF (absolute, per item and relative to the size of the input ESE).

• Intracollection links: we will provide a quantitative study of the amount of similarity
links between items. Regarding quality, we will evaluate how the automatic links
compare to manual annotations.

• Background links: we will provide a quantitative study of the amount of links to
wikipedia articles. Regarding quality, we will check manually a sample of the links
between equivalent items and articles.

Below we describe the results for each of the tasks.

8.1 Content analysis
The content analysis carried out consists of several expensive subtasks. Each of the steps
requires sophisticated methods to analyse and extract the linguistic information as mentioned
in Section 4.

In total the processing of the whole dataset took 5 weeks divided in 7 CPUs. The production
of KAF, in which the main linguistic information is extracted, was performed in 4 weeks, and
the classification of named entities (NERC production) was done also in 4 weeks. Regarding
NERC production only for English collections done, since for Spanish NERC is performed
when producing KAF. These two subtasks where parallelized in order to speed the whole
process up. Finally, in the last week the extracted information was filtered out to SAF files.
This last step was run in a single CPU.

46


Table 16 shows the estimation of the throughput for different linguistic processing and SAF
converting. Note that the ratio of items per minute is estimated on a single CPU

#collections #CPU #weeks Item/minutes
KAF 4 7 4 7.5
English NERC 2 7 4 3.1

SAF 4 1 1 295
TOTAL - - - 7.4
Table 17: Estimation of the throughput for different linguistic processing and SAF converting. Note that the
ratio of items per minute is estimated on a single CPU.

Regarding SAF size, the 4 collections are 9GB big, being on average about 4.9KB the size of
each item. The relative to the size of ESE each SAF item is almost 4 times bigger that an ESE
item.

Taking into account the actual estimations, the pre-processing of datasets 10 times larger
would be tedious, while the size would still manageable. More specifically, in that case, we
would need about 50 weeks to process the whole dataset using 7 CPUs.

The figures show that the process is as expected for the size and nature of the collection. We
have identified some possibilities to make the process faster, which we plan to deploy for the
next version.

8.2 Ontology extension
The quantitative study of the amount of matches is shown in Section 5, where the number of
matches per vocabulary and collection are given. Moreover, a detailed description of
coverage and the different examples of matched words is also depicted.

Regarding the size, vocabularies SAF files are about 0.2KB each, which in total would take
around 0.5GB. On average, the relative size to ESE input item is about 6 times smaller than
an ESE item.

The performance in time of matching algorithm is not as expensive as the linguistic
processing. Nevertheless, the algorithm could be expensive enough for bigger datasets, as for
example the whole set of Europeana, in the case we are using a large vocabulary for
matching. Roughly, the throughput of the system is about 3,500 items per minute.

47


8.3 Intracollection links
The various methods used are evaluated against a gold-standard data set of 30 pairs of items
selected randomly from Europeana collections. This data set includes human judgements on
the similarity between the items.

We obtained human judgements by conducting an online survey. Annotators were asked to
rate similarity of a pair of items from 0 to 4 where 0 denotes completely unrelated items and
4 indicates completely related ones. A total of 74 responses received from which 38 were
partially completed and apparently discarded. We considered the rest of 36 responses for
creating the gold-standard. A number of obvious similar pairs (same title, image) were used to
ensure that the answers have not been given randomly. Therefore, for each pair of item the
gold-standard score is the average of all the human ratings.

We applied similarity measures using item titles and descriptions. We measure performance
by computing Pearson's correlation coefficient between results of each method and the gold-
standard. Table 7 shows the results of each method.

Measure Pearson’s Correlation Score
0.722

0.705

0.67
0.807

0.651
Table 18: Evaluation of similarity.

Results show that LDA-vector outperforms the rest of the measures and is chosen as the
similarity measure.

The output size of Alinari data is 27.8MB and it took approximately 1.5 hours to complete.
Therefore, the throughput ratio is about 125 items/min and 309 KB/min.

48


8.4 Background links
It takes approximately 24 hours to annotate all the Europeana items with Wikipedia links.
This means a single item takes on average 0.1 seconds to process. Using a probability
threshold of 0.5, the number of links found per item in Europeana follows this distribution:

Num. of links Freq. (%)

0 33.56

1 27.27

2 19.56

3 10.1

4 5.03

5 2.16

6 1.07

7 0.54

8 0.26

9 0.15

>=10 0.29

Table 19: Number of links per item in Europeana.

More than 90% of the items have 3 or fewer links. Over a third has no links at all.

Evaluation
For evaluation purposes, a gold standard set of Europeana items was created. In total 100
items were randomly selected from Europeana. 25 of the items were filtered out since there
was insufficient text to be processed, leaving 75 remaining. For each the title, subject and
description text was aggregated together. The text was then manually annotated with a set of
links to relevant Wikipedia articles (if any). It was found that 65 of these 75 had one or more
relevant Wikipedia links (in the judgement of the annotator). The output of the Wikipedia
Miner system was then judged against this gold standard. The evaluation methodology used
here follows that of (Milne and Witten 2008), where only the set of links is judged, and not
the anchor positions of text. As mentioned earlier, finding the correct anchors for the links in
the text is a separate problem in itself. The system was evaluated using different probability
cut-off points. Precision and recall and F-measure were calculated.

49


Probability cut-off Precision Recall F

p > 0 (all links) 0.29 0.87 0.44

p > 0.2 0.46 0.31 0.37

p > 0.5 0.44 0.25 0.32

Table 20: Evaluation of links according to cut-off.

Using all links gives a high recall value of 87%. Eliminating low probability links results in a
large drop in recall performance, but precision is improved.

50


9 Summary
This report accompanies and describes the contents of Deliverable 2.1 “Processing and
Representation of Content for First Prototype”. The deliverable comprises the data produced
by WP2 “Content Processing and Enrichment” which is to be used in the first prototype. The
data has been released in DVDs and is also available from the subversion server of the
project:

http://paths.avinet.no:81/svn/paths/tags/D2.1-2011-11-30

experience. The additional information includes links between items in the collection and to
external sources like Wikipedia. The resulting data forms the basis for the paths used to
navigate the collection, providing a collection of content for the first PATHS prototype and
defining the standards for content format and descriptive metadata.

In this first release of the data, we include a private collection from Alinari, and four
collections from Europeana: the Culture Grid and SCRAN collections from the UK and the
Hispana and Cervantes collections from Spain.

The content of the target collection, both metadata and textual information, was analysed
with Natural Language Processing tools to identify relevant pieces of information that can be
used for generating links. The information identified includes specific attributes of the items
as typically found in the metadata and also additional information that gives context to the
item as found in the textual information that accompanies the items (people, organizations,
dates and locations).

The subject field was also processed and automatically compared with some relevant
vocabularies (two for English and one for Spanish). The items in the target collections were
linked to each other based on the information provided by the data analysis task (links
available for the Alinari data alone). In addition, the items have been linked to background
information as made available in Wikipedia. The background links to other sources and the
intra-collection links for Europeana items will be incorporated later in the project. The
respective tasks were planned to start in November and the first results will be ready soon.

The contents of the deliverable will be updated with the outcome of WP2, which will
continue to work concurrently to the development of the first prototype. The final outcome of
WP2 will be released on the beginning of April, so it can be used in the first prototype.

51


Glossary
• ESE: Europeana Semantic Elements (ESE) specification format.
• KAF: Kyoto Annotation Format.
• Latent Dirichlet Allocation: A generative model that allows sets of observations to
be explained by unobserved groups that explain why some parts of the data are
similar.
• LDA: Latent Dirichlet Allocation
• Lemmatization: The process of grouping together the different inflected forms of a
word so they can be analysed as a single item.
• Linguistic Processor: A program which make use of natural language processing
techniques in order to extract information from text.
• Named Entity: Usually predefined categories of atomic elements such as the names
of persons, organization, locations and temporal expresions.
• Named Entity Recognition and Classification: Also known as entity identification
and entity extraction, is a subtask of information extraction that seeks to locate and
classify atomic elements in text into predefined categories such as the names of
persons, organizations, locations, expressions of times, quantities, monetary values,
percentages, etc.
• NERC: Named Entity Recognition and Classification.
• Multiword expression: Made up of a sequence of two or more lexemes that has
properties that are not predictable from the properties of the individual lexemes or
their normal mode of combination.
• Part of Speech tagging: Part-of-speech tagging (POS tagging or POST), also called
grammatical tagging or word-category disambiguation, is the process of marking up a
word in a text (corpus) as corresponding to a particular part of speech, based on both
its definition, as well as its context.
• SAF: Simple Annotation Format.

52


References
Agirre, E., Alfonseca, E., Hall, K., Kravalova, J., Pasca, M. and Soroa, A. (2009), “A study on
similarity and relatedness using distributional and WordNet-based approaches”, Proceedings of
Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the
Association for Computational Linguistics on - NAACL ’09 , Association for Computational
Linguistics, Morristown, NJ, USA, p. 19.

Blei, D. M., Ng, A. Y. and Jordan, M. I. (2003), “Latent Dirichlet allocation”, Journal of Machine
Learning Research , Vol. 3, JMLR.org, pp. 993–1022.

Bosma, W. E. , P. Vossen, A. Soroa, G. Rigau, M. Tesconi, and A. Marchetti, and M. Monachini, and
C. Aliprandi. KAF: A Generic Semantic Annotation Format. In Proceedings of the GL2009 Workshop
on Semantic Annotation, 2009.

Budanitsky, A. and Hirst, G. (2006), “Evaluating wordnet-based measures of lexical semantic
relatedness”, Computational Linguistics , Vol. 32, pp. 13–47.

Cunningham, H., D. Maynard, K. Bontcheva and V. Tablan. GATE: A Framework and Graphical
Development Environment for Robust NLP Tools and Applications. In Proceedings of the 40th
Anniversary Meeting of the Association for Computational Linguistics, 2002.

Grieser, K., Baldwin, T., Bohnert, F. and Sonenberg, L. (2011), “Using ontological and document
similarity to estimate museum exhibit relatedness”, Journal on Computing and Cultural Heritage
(JOCCH) , Vol. 3, ACM, p. 10.

Harrington, B. (2010), A semantic network approach to measuring relatedness, in ‘International
Conference on Computational Linguistics’, pp. 356–364.

Mihalcea R. and A. Csomai. Wikify!: Linking Documents to Encyclopedic Knowledge. In
CIKM, volume 7, pages 233–242, 2007.

Mihalcea, R., Corley, C. and Strapparava, C. (2006), Corpus-based and knowledgebased measures of
text semantic similarity, in ‘National Conference on Artificial Intelligence’.

Medelyan, O. , Witten, I.H. , and D. Milne. Topic Indexing with Wikipedia. In Proceedings of
the AAAI WikiAI workshop, 2008.

Milne, D. (2007), Computing semantic relatedness using wikipedia link structure, in ‘Proceedings of
the New Zealand Computer Science Research Student Conference’, Citeseer.

Milne, D. and I.H. Witten. Learning to Link with Wikipedia. In Proceeding of the 17th ACM
conference on Information and knowledge management, pages 509–518. ACM, 2008.

53


Padró , L., M. Collado, S.Reese, M. Lloberes, and I. Castellón. FreeLing 2.1: Five Years of Open-
Source Language Processing Tools. In Proceedings of 7th Language Resources and Evaluation
Conference (LREC), 2010.

Patwardhan, S., Banerjee, S. and Pedersen, T. (2003), Using measures of semantic relatedness for
word sense disambiguation, in ‘Proc. of the 4th Int’l. Conf. on Intelligent Text Processing and
Computational Linguistics’, pp. 241–257.

J.R. Quinlan. C4. 5: Programs for Machine Learning. Morgan Kaufmann, 1993.

54


Annex 1: Summary of Europeana data
Summary of English collections
Total records = 985346

dc:contributor
Values for 421786 records (42.8%)
Top values:
Value Freq.

North East Midland Photographic Record 80765

The Scotsman Publications Ltd 66101

Royal Commission on the Ancient and Historical Monuments of Scotland 28112

Kirklees 26132

The Scotsman newspaper publisher 21060

The Scotsman' newspaper publisher 17774

National Museums of Scotland 16983

Scottish Motor Museum 10722

National Library of Scotland 9958

Newsquest (Herald& Times) 9909
Tokens: Mean=3.6, Mode = 0, Median = 0

55


dc:creator
Top values:
Value Freq.

Root 62703

Taunt, Henry 12914

Unknown 7594

Donna Nicoll 7394

unknown 7011

Lambeth Libraries 5700

Minter, Faye - Portable Antiquities Scheme 5695

Abigail Evans 5601

Andrews-Wilson, Liz - Portable Antiquities Scheme 5347

Daubney, Adam - Portable Antiquities Scheme 4714

Tokens: Mean = 2.0 Mode = 0, Median = 0

56


dc:date
Top values:
Value Freq.

[2008] 26833

[2007] 25199

[2006] 16121

[2005] 15503

[2003] 12020

[2004] 9560

[2009] 8974

01/01/90 00:00 6257

31/12/10 00:00 4444

31/12/36 00:00 4241
Tokens: mean = 2.4, Mode = 0, Median = 1

57


dc:description
Top values:
Value Freq.

Scotland, Edinburgh headquarters of The Scotsman newspaper 14907

Scotland, Glasgow, River Clyde 3609

Single-decker 3511

Scotland, Edinburgh location 2700

World War I taken during 1903

Scotland, City of Edinburgh, Edinburgh locality 1704

Double-decker 1473

Scotland, Dundee location of publication 1469

Scotland locality 1375

Scotland, Angus, Dundee depicted 1328
Tokens: Mean = 26.7, Mode = 0, Median = 10

58


dc:format
Top values:
Value Freq.

image/jpeg 271865

text/html 162356

JPEG/IMAGE 112082

application/imgzoom 31396

Image 15788

video/mpeg 5832

audio/mpeg 2226

application/pdf 2212

video/quicktime 400

Christmas card 128


59


dc:publisher
Top values:

Value Freq.

The Fitzwilliam Museum, Cambridge, UK 162356

North East Midland Photographic Record 80765

Kirklees 26132

Wolverhampton Archives 7301

Gateshead Council 4144

Wolverhampton Arts and Museums Service 4048

Tyne & Wear Archives & Museums 3916

Dudley Archives 3686

Dudley Museums Service 2371

Walsall Archives / Local History Centre 1627

60


dc:rights
Top values:

Value Freq.

Copyright 2005 Portable Antiquities Scheme 125487

The Scotsman Publications Ltd 66028

Reproduced by permission of English 23644
Heritage.NMR

National Museums Scotland 23064

Royal Commission on the Ancient and 16427
Historical Monuments of Scotland

Corpus Vitrearum Medii Aevi / Courtauld 15043
Institute of Art

English Heritage.NMR 14468

Scottish Motor Museum Trust 10722

National Library of Scotland 10315

Newsquest (Herald& Times) 9911

dc:source
Top values:

Value Freq.

Fitzwilliam Museum 162356

Portable Antiquities 125562

Vads 93105

61


dc:subject
Top values:

Value Freq.

coin 35530

print 33081

Kirklees 26127

drawing 17666

letter 14971

Unknown 14832

Photograph 14708

letter, handwriting 13367

Transport 10702

Photograph from The 9918
Herald newspaper archive
Tokens: mean = 4.7, Mode = 0, Median = 3

62


dc:title
Top values:

Value Freq.

Coin 26818

The Corpus of Romanesque Sculpture in Britain and 18964
Ireland

coin 15417

letter 14252

COIN 7874

Corpus Vitrearum Medii Aevi (CVMA) 6992

Design Council Slide Collection 4479

Buckle 4345

Brooch 4298

Vessel 4110

63


dc:type
Top values:
Value Freq.

Image 413538

Physical Object 93105

Unknown 2369

PhysicalObject 1158

Sound 1038

MovingImage 515

Text 465

64


dcterms:alternative
Top values:
Value Freq.

from James Wotherspoon's 'In The Track Of The Comet' 3608

Jean Jenkins collection 1499

As described in Ordnance Gazetteer of Scotland, 1882-1885, by Francis H. Groome 1339

From the Roy Map, 1747-1755, 1:36,000 1034

a photograph by Thomas Annan of Glasgow (1829-1887) 671

Scanned from the 1853 edition of the 'Scots Musical Museum', James Johnson and 600
Robert Burns (Edinburgh and London: W. Blackwood& Sons, 1853)

Printed plate 550

Single-decker 538

from the Hutton Collection 534

minted in Rome 457

65


dcterms:isPartOf
Top values:
Value Freq.

Fitzwilliam Museum 162356

Portable Antiquities 125562

Vads 93105

Picture the Past 80774

Leodis 55976

English Heritage 50208

Kirklees Image Archive 31326

Black Country History 21152

Durham County Council 15788

Bowes Museum 14780

66


dcterms:medium
Top values:
Value Freq.

Paper 20784

Photograph 18682

Print, monochrome 8261

Video 4082

Engraving 3039

Broadside 1874

Glass negative 1741

Paper, ink 1634

Ink on paper 1620

Book 1577

67


dcterms:provenance
Top values:
Value Freq.

bequeathed by Reitlinger, Henry Scipio, 1950, Bequeathed by H.S. Reitlinger, 1950 252

bequeathed by Reitlinger, Henry Scipio, 1950, H. S. Reitlinger Bequest 1950 113

bequeathed by Reitlinger, Henry Scipio, 1950, Bequeathed by Henry Scipio Reitlinger, 102
1950

bequeathed by Reitlinger, Henry Scipio, 1950, H.S. Reitlinger Bequest, 1950 10

given by Ricketts and Shannon 6

bequeathed by Reitlinger, Henry Scipio, 1950, H.S. Reitlinger Bequest 1950 6

bequeathed by Reitlinger, Henry Scipio, 1905-05-03, H. S. Reitlinger Bequest 1950 4

bequeathed by Reitlinger, Henry Scipio, 1991-04-29, H.S. Reitlinger Bequest, 1950 2

bequeathed by Harris, Peter, 1976, with a life interest to his widow 2

bequeathed by Leverton Harris Fund, 1926 [3777] 1
Tokens: Mean = 1.2 Mode = 0, Median = 0

68


dcterms:spatial
Top values:
Value Freq.

Europe 62410

United Kingdom 62409

World 62408

England 6817

London 5713

Greater London 5700

Lambeth 5700

Darlington 3128

Durham City 2190

World, Europe, United Kingdom, England, 2174
Tyne and Wear, Gateshead, Gateshead

69


dcterms:temporal
Top values:
Value Freq.

2001 to date 6114

1901 - 1925 2503

1951 - 1975 2002

1976 - 2000 1607

name=Edwardian; start=1901; end=1910; 1391

1926 - 1950 1384

1951-01-01T00:00:00/1970-12-31T23:59:00 1359

1926-01-01T00:00:00/1950-12-31T23:59:00 1141

1901-01-01T00:00:00/1925-12-31T23:59:00 1077

1876 - 1900 730

europeana:provider
Values for 947712 records (100%)
Top values:
Value Freq.

CultureGrid 674544

Scran 310802

europeana:type
Top values:
Value Freq.

IMAGE 851210

TEXT 91739

70


VIDEO 39133

SOUND 3264

europeana:year
Top values:

Value Freq.

2008 26843

2007 25199

2006 16607

2005 16283

2003 12074

2004 9567

2009 8974

1900 6560

1920 3958

1951 3851

71


Summary of Spanish records
Total records = 1254411

dc:contributor
Values for 231494 records (18.5%).
Top values:
Value Freq.

Santa Ana, Manuel María de (1820-1894) 50692

Cámara Agrícola de Carrión de los Condes (Palencia) 18605

Partido Liberal (Gerona) 11573

Federación de Sindicatos Católicos-Agrarios (Palencia) 10053

Confederación Regional del Trabajo de Baleares 9599

Sindicato Único de Trabajadores de Mahón 9597

Comité Comarcal de Sindicatos de Menorca 9591

Partido Liberal Dinástico (Tarragona) 8815

Tipografía de Antonio Arqueros (Badajoz) 8199

Tipografía La Minerva Extremeña (Badajoz) 8197

72


dc:coverage
Top values:
Value Freq.

-19º-20º siglos 10330

-19º-20º séculos 10330

-1887-1938 9970

-1900-1919 4362

-1897-1947 1927

-Galicia 1927

-S.XIX 1371

-1906-1938 1076

-S.XIX-XX 981

-S. XIX 898

73


dc:creator
Top values:
Value Freq.

Anonymous 3378

Anónimo 2059

Lorente Páramo, Francisco Javier 1772

Sorolla Bastida, Joaquín ( Lugar de Nacimiento: Valencia, 27/08/1863 - TL04 1286
Cercedilla, 10/08/1923)

Gadir 757

Pardo Bazán, Emilia Condesa de 1851-1921 655

Atribuido a 430

PLOSSU, Bernard ( Lugar de Nacimiento: Dalat, 26/02/1945) 392

SERRANO AGUILAR, Pablo ( Lugar de Nacimiento: Crivillén, 1908 - TL04 Madrid 385
(m), 1985)

Real Fábrica de Cristales de la Granja 362

74


dc:date
Top values:
Value Freq.

1913 24136

1914 23036

1917 22766

1912 22701

1916 22128

1909 21939

1915 21787

1910 21778

1908 21472

1911 21426

75


dc:description
Top values:
Value Freq.

Copia digital. Madrid: Ministerio de Cultura. Subdirección General de 331786
Coordinación Bibliotecaria, 2004


Copia digital: realizada por la Biblioteca de Andalucía 85408

Copia digital. Valladolid: Junta de Castilla y León. Consejería de Cultura y 6883
Turismo. Dirección General de Promociones e Instituciones Culturales, 2009-2010

Copia digital: realizada por la Fundación Provincial de Cultura. Diputación 2229
Provincial de Cádiz

Digitalización. Vitoria-Gasteiz: Fundación Sancho el Sabio, 2008 925

Texto a dos col. 917

Port. con orla tip. 714


Rústica 485
Tokens: Mean=9.7, Mode=0, Median=8

76


dc:format
Top values:
Value Freq.

image 939507

application/pdf 807069

image/tiff 129659

image/jpeg 60620

video/flv 121

audio/mp3 96

text/plain 53

application/msword 4

imagen 3

application/epub+zip 1

77


dc:language
Top values:
Value Freq.

spa 1226712

cat 19927

lat 2453

fre 1192

eng 764

baq 755

val 349

glg 346

por 220

ita 196

78


dc:publisher
Top values:
Value Freq.

Ministerio de Cultura 101884

Zuloaga, Hilarión de 50721

Hilarión de Zuloaga 50295

Diputación Provincial de Guadalajara] 16123

Alicante Biblioteca Virtual Miguel de Cervantes 14450

: [s. n. 13333

[s. n. 6172

Asociación de la Prensa 5457

[FET.JONS 5342

Prensa y Radio del Movimiento 5340

79


dc:rights
Top values:
Value Freq.

Ministerio de Cultura 101884

VEGAP 1426

VEGAP, Visual Entidad de Gestión de Artistas Plásticos 51

Cabré Aguiló, Juan 16

Todos los públicos. 8

Cifuentes, Ramón 2

Pérez Rioja, Aurelio 1

Alviach, Manuel 1

Huerta, M. 1

Cánovas del Castillo Vallejo, Antonio 1

80


dc:source
Top values:
Value Freq.

Biblioteca Virtual de Prensa Histórica 848932

Galiciana: Biblioteca Digital de Galicia 129674

Hispana; Spain 101884

Repositorio Biblioteca virtual de Andalucía 98020

Biblioteca Digital de Madrid 22856

CER.ES: Red Digital de Colecciones de museos de España (Museo de América) 17341

CER.ES: Red Digital de Colecciones de museos de España (Museo Nacional de Artes 16823
Decorativas)

Biblioteca Virtual del Patrimonio Bibliográfico 9201

CER.ES: Red Digital de Colecciones de museos de España (Museo del Traje. Centro de 8387
Investigación del Patrimonio Etnológico)

Biblioteca Digital de Castilla y León 6927

81


dc:subject
Top values:
Value Freq.

Secciones - Hemeroteca - Prensa 91537

Catálogos - Publicaciones periódicas - Prensa 91504

Prensa 33658

Prensa gallega 29837

Prensa galega 29287

Cerámica 17808

Edad Contemporánea 14349

-Administración-Publicaciones periódicas 14204

Bronce 9357

Numismática 8722

Tokens: Mean = 1.96, Mode =0, Median = 0

82


dc:title
Top values:
Value Freq.

Moneda 7270

Vasija 3206

Plato 2777

Aguja 2164

Fotografía 2074

Cuenco 1722

Vaso 1325

Fusayola 1226

Figura antropomorfa 888

Moeda 888

Tokens: Mean=14.2, Mode = 13, Median
=14

83


dc:type
Top values:
Value Freq.

text 17107

Moneda 7276

Cuadro 6096

Fotografía 3998

Book 3590

Vasija 3206

Estampa 2961

Plato 2814

Escenografía 2474

Aguja 2164

84


dcterms:created
Top values:
Value Freq.

1801=1900 2076

1900-1965 1622

1974-1978 1552

1701=1800 1473

1601=1700 1433

100 AC-700 DC 1303

1901=2000 1247

200[ac]=1[ac] 665

1100-1450 DC 659

1970[ca] 594

85


dcterms:isPartOf
Top values:
Value Freq.

En: La Correspondencia de España : diario universal de noticias. - [Madrid] : 50072
Hilarión de Zuloaga , 1859-[1925] = ISSN 1137-1188

En: El Defensor de Granada : diario político independiente. - [S.l. : s.n. , 1880- 23541
1984]

En: Boletín Oficial de la Provincia de Oviedo . - Oviedo : [s.n.] , 1836-1982 = ISSN 23216
1579-7279

En: Diario de Córdoba de comercio, industria, administración, noticias y avisos . - 22708
[S.l. : s.n. , 1849-1938]

En: Crónica Meridional : diario liberal independiente y de intereses generales. - 19646
[S.l. : s.n. , 1860- ]

Museo de América 17382

Museo Nacional de Artes Decorativas 16873

En: Boletín Oficial de la Provincia de Guadalajara . - [S.l. : Diputación Provincial 16075
de Guadalajara] ,

En: El regional : diario de Lugo . - Lugo : [s.n.] , 1884- 15085

En: Boletín de Segovia . - [S.l. : s.n.] , 1833- 14076
Tokens, Mean = 15, Mode = 15, Median = 15

86


dcterms:issued
Top values:
Value Freq.

[s.a.] 733

s.a.] 720

[19--?] 397

[18--?] 297

[1906] 178

[1903] 171

1884 153

1885 148

[S. XVIII] 146

[1917] 142

87


dcterms:medium
Top values:
Value Freq.

Bronce 9355

Cerámica 8562

Arcilla 8224

Papel 8056

Madera 6420

Lienzo 5037

Plata 4698

Hueso 3994

Pasta cerámica 3529

Loza 3371

88


dcterms:provenance
Top values:
Value Freq.

Donación 15430

Depósito 5654

Dación 3317

2004 3104

Fábrica Pickman S.A. 3046

Legado fundacional 1416

Legado fundacional de la Vda. de Sorolla, 1931 1070

Asignación: Donación 866

Egner, Günther Adolf 832

03/05/2007 825

89


dcterms:spatial
Top values:
Value Freq.

España-Comunidad de Madrid-Madrid-Madrid 70855

España-Andalucía-Granada-Granada 63628

-Galicia 62834

España-Comunidad Valenciana-Alicante-Alicante 49968

España-Andalucía-Córdoba-Córdoba 48398

España-Illes Balears-Illes Balears-Palma de Mallorca 47519

España-Andalucía-Almería-Almería 44929

España-Canarias-Santa Cruz de Tenerife-Santa Cruz de Tenerife 34383

España-Cataluña-Tarragona-Tarragona 32956

España-Cataluña-Girona-Girona 31691

90


dcterms:temporal
Top values:
Value Freq.

Edad Contemporánea 14349

Edad Moderna 5480

Cultura romana 4620

Andes centrales 4384

1801=1900 2076

Intermedio temprano 2000

Alto Imperio Romano 1738

Andes septentrionales 1678

1900-1965 1622

Alto Imperio Romano. Romano 1591

europeana:hasObject
Top values:
Value Freq.

true 946515

false 307896

europeana:provider
Value Freq.

Hispana 1235133

Biblioteca Virtual Miguel de Cervantes 19278

91


europeana:type
Value Freq.

TEXT 1148918

IMAGE 105493

europeana:year
Top values:
Value Freq.

1913 24151

1914 23048

1917 22774

1912 22710

1916 22139

1909 21957

1915 21810

1910 21805

1908 21497

1911 21450

92


Annex 2: Alinari records
Currently we have a set of 11628 records provided by Alinari. This contains the following
fields:

Image ID
Caption
Date of photography
Place of photography
Detailed place of photography
Photographer
Object
Technique
Date of artwork
Artist
Period and style
Artwork support
Location
Events
People
Credit
Keywords
Permission and Restrictions
Labels
Format (b/w - col)
Orientation (portrait or landscape)

Below is a list of some of the fields and summary statistics of occurring values for each field.

93


Caption
Top values:
Value Freq.

The Pine trees of Rome 23

Construction of the Transiberian railway around Lake Baikal 14

View of Genoa 9

Electric poles of the S.A.E., Società Anonima Elettrificazione of Milan, on 8
the Milano-Domodossola line

A construction phase of the Cignana Dam, work of the Umberto Girola 8
building firm

View of the city of Perugia 8

A part of the historic parade on the occasion of the ""Regata delle antiche 8
Repubbliche Marinare"" held in Amalfi in 1957

Water consortium of Agro Monfalconese: work on the water pump on the 7
Isonzo river in Sagrado

Genoa's port 6

Hunting on horseback with dogs 5
Tokens: Mean=11.45, Median = 10.0, Mode=7

94


Date of photography
Top values:
Value Freq.

1890 ca. 993

1900 ca. 832

1920-1930 ca. 511

1915-1920 ca. 341

1920 - 1930 ca. 339

1880 ca. 331

1890 - 1910 ca. 242

1910 ca. 207

1930 ca. 197

1870 ca. 160

95


Place of photography
Top values:
Value Freq.

France 567

Rome 466

Naples 347

Great Britain 303

Florence 279

Greece 264

Switzerland 259

United States 216

Italy 206

Africa 191

96


Detailed place of photography
Values for 6612 tokens (58.7%)
Top values:
Value Freq.

Environs 184

environs 142

Paris 98

Somalia 53

New York 52

Lake Como 47

Athens 43

Via Appia Antica 35

Vallombrosa 34

Corfù 34

97


Photographer
Top values:
Value Freq.

Unidentified Author 3803

Alinari, Fratelli 2510

Brogi 1478

P.z 358

Ojetti, Ugo 217

Graham, James 87

Mauri, Achille 73

Fürst Lwoff, Eugen 71

Sommer, Giorgio 64

Baldinetti, Modestino 63

Object
Top values:
Value Freq.

Print on double-weight paper 6256

Glass plate 4052

Photomechanic print 713

Carte de visite 151

Stereoscopic Photography 43

Panoramic 24

Cabinet mount 7

98


Photomechanic matrix 6

Engraving from a daguerreotype 5

Postcard 4

Technique
Top values:
Value Freq.

N Silver salt gelatin 3541

P Albumen 2883

P Silver salt gelatin 1762

P Collodio-chloride paper 431

P Chromolithography 380

P Collotype 211

P Phototype 161

P Photomechanics 156

P Aristotype 143

P Albumenised salt paper 83

99


Date of artwork
Values for 2087 rcords (18.5%)
Top values:
Value Freq.

XIX sec. 42

XVIII sec. 28

XIV sec. 28

XVI sec. 27

XV sec. 26

XII sec. 23

XIII sec. 22

1837-1846 18

I sec. 17

II sec. 16

100


Artist
Top values:
Value Freq.

Ammannati Bartolomeo (1511-1592) 19

Canzio, Michele 18

Junker, Carl 15

Gizdulich, Riccardo 11

Fanzago, Cosimo 11

Kallikrates 11

Brizzi, Emilio 11

Fidia 10

Vignola 10

Roebling, John Augustus 10

101


Period and style
Top values:
Value Freq.

Europe 1910

First and Second Millennium A.D. 1458

Renaissance-Baroque styles and periods 583

Ancient Civilization 544

Middle Ages 405

Modern European Styles and Movements 362

Roman Art 358

Renaissance 299

Gothic 154

Imperial Period 153

102


Artwork support
Top values:
Value Freq.

Stationary Modern Work 1297

Stationary Archaeological Work 517

Castle 255

Movable Modern Work 229

Building 179

Church 176

Temple 130

Bridge 125

Villa 109

Painting 66

103


Location
Top values:
Value Freq.

Rome, 68

Florence, 60

Venice, 36

Pompeii, 36

Greece, Athens 31

Naples, 31

Turin, 30

Great Britain, 27

France, Paris 23

Bellagio, 23

104


Events
Top values:
Value Freq.

World War I, 1914-1918 302

The Paris Commune, March-May 1871 32

Eruption of Vesuvius, April 1906 24

World War II, 1939 - 1945 21

Universal Exhibition of Paris, 1900 18

Spanish Civil War, 1936-1939 18

Earthquake of Diano Marina e Diano Castello, 23 February 1887 16

Official visit of Adolf Hitler to Italy, 3-9 May 1938 14

Earthquake of Messina (Sicilian and Calabrian Earthquake), 1908 13

Flood in Verona, 22 september 1882 12

105


People
Top values:
Value Freq.

Lorena Karl Stephan of 10

Ojetti Ugo 5

Antinori Piero 3

Victor Emanuel III of Savoy 3

Barrès Maurice 3

Hitler Adolf 3

Vittorio Emanuele Prince of Naples 2

D'Annunzio Gabriele 2

Gallone Soava 2

Maria of Savoy 2

106


Credit
Value Freq.

Fratelli Alinari Museum Collections, Florence 7162

Alinari Archives-Alinari Archive, Florence 2537

Alinari Archives-Brogi Archive, Florence 1517

Gabba Raccolta Acquisto Fratelli Alinari Museum Collections, Florence 20

Aragozzini Vincenzo Archive Fratelli Alinari Museum Collections, Florence 12

Verchi Marialieta Collection Fratelli Alinari Museum Collections, Florence 6

Abetti Fondo Acquisto Fratelli Alinari Museum Collections, Florence 5

Reteuna Collection Fratelli Alinari Museum Collections, Florence 5

Ojetti Archive Fratelli Alinari Museum Collections, Florence 1

Pasta Fernando Archivio Fratelli Alinari Museum Collections, Florence 1

107


Keywords
Top values:
Value Freq.

Landscape 11249

Architecture 7660

Structure 7323

Vegetation 3687

View 3677

Housing 3517

Tree 3093

Building Elements,Decoration and Sections 2878

Transportation 2283

Landscape (Seascape) 2191

Note: The meta data contains a list of distinct keywords for each item. Therefore it makes
more sense to consider number of keywords rather than number of tokens:

Keywords: Mode = 12, Median = 13, Mean = 14.1

108


Labels
Value Freq.

WARNING: Permission must be required for non editorial use. Please 11233
contact Alinari Archives;

Copyright: ARTIST'S COPYRIGHT MUST ALSO BE CLEARED; 30
WARNING: Permission must be required for non editorial use. Please
contact Alinari Archives;

109


Format (b/w - col)
Top values:
Value Freq.

B/W 10781

Col 487

Orientation (portrait or landscape)
Top values:

Value Freq.

L 9796

P 1472

110

PATHS Content processing 1st prototype

More Related Content

Viewers also liked (9)

Similar to PATHS Content processing 1st prototype (20)

More from pathsproject (20)

Recently uploaded (20)

PATHS Content processing 1st prototype