ARHINET

A System for Generating and Processing Semantically-Enhanced

Archival eContent

Ioan Salomie, Mihaela Dinsoreanu, Cristina Pop

Sorin Suciu, Tudor Vlad and Ioana Iacob

Department of Computer Science, Technical University of Cluj-Napoca, 15 C. Daicoviciu Street, Cluj-Napoca, Romania

Keywords: Archival domain model, Knowledge acquisition, Historical domain ontology, Semantic annotation,

Semantic query, Reasoning.

Abstract: This paper addresses the problem of generating and processing of eContent from archives and digital

libraries. We present a system that adds semantic mark-up to the content of historical documents, thus

enabling document and knowledge retrieval as response to semantic queries. The system functionality

follows two main workflows: eContent generation and knowledge acquisition on one hand and knowledge

processing and retrieval on the other hand. Within the first workflow, the relevant domain information is

extracted from documents written in natural languages, followed by semantic annotation and domain

ontology population. In the second workflow, ontologically guided queries trigger reasoning processes that

provide relevant search results.

1 INTRODUCTION

Most of the existing archives-related information is

written in natural language and is available in large

amounts, distributed in archives and digital libraries.

Although content management systems capable of

dealing with distributed data exist for years, making

them able to process natural language documents is

still a challenge. Documents in natural language

human readable format contain unstructured and

heterogeneous information, which makes it hard for

machines to automatically process their contents.

Our system addresses these challenges by

adopting Semantic Web techniques in the context of

archive management. We add a layer of machine-

processing semantics over the content of raw

documents contained in the archives by using (semi)

automated knowledge acquisition. The machine-

processable semantics, captured in a domain

knowledge base, is further used as support for

intelligent searches that provide the most relevant

results to agent queries. These relevant results

represent documents, information and knowledge

related to the semantics of the keywords specified in

the queries.

In this context, our system offers a solution for

managing domain-specific archival data. Our system

was specialized to generate archival eContent, to

semantically enhance and process it, from the

available medieval documents regarding the history

of Transylvania. For the semantic enhancement of

the documents we have built an ontology core of

concepts and relations, which is continuously

expanded as new documents are processed.

Historians and archivists can use the system to find

relevant documents, information and knowledge.

The rest of the paper is organized as follows. In

Section 2, we introduce related work. Section 3

presents our generic model of the archival domain.

The proposed system architecture is presented in

Section 4, while the associated workflows are

described in Section 5 and Section 6. Section 7

contains a case study that illustrates the system's

functionality in the context of historical archives.

The paper ends with our conclusions and future

work proposals.

2 RELATED WORK

The OntoPop methodology provides a single-step

151

Salomie I., Dinsoreanu M., Pop C., Suciu S., Vlad T. and Iacob I.

ARHINET - A System for Generating and Processing Semantically-Enhanced Archival eContent.

DOI: 10.5220/0001833401510158

In Proceedings of the Fifth International Conference on Web Information Systems and Technologies (WEBIST 2009), page

ISBN: 978-989-8111-81-4

solution for (i) semantically annotating the content

of documents and (ii) populating the ontology with

the new instances found in the documents

(Amardeilh, 2007). The solution uses domain-

specific knowledge acquisition rules which link the

results obtained from the information extraction

tools to the ontology elements, thus creating a more

formal representation (RDF or OWL) of the

document content (Amardeilh, 2006). The OntoPop

methodology has certain limitations regarding the

resolving of synonyms on one hand and the

resolving of multiple instances with the same lexical

representation on the other hand. In this paper, we

address the identified limitations by extending the

OntoPop methodology with new processing steps

before populating the ontology.

SOBA is a system designed to create a soccer

specific knowledge base from heterogeneous sources

(Buitelaar et al., 2006). The system performs (i)

automatic document retrieval from the Web, (ii)

linguistic annotation and information extraction

using the Heart-of-Gold approach (Schäfer, 2007)

and (iii) mapping of the annotated document parts

on ontology elements (Buitelaar et al., 2006). Our

approach performs information extraction from

unstructured text and document annotation for a

specific domain and uses reasoning on the ontology

to infer properties for the newly added instances.

Ontea performs semi-automatic annotation using

regular expressions combined with lemmatization

and indexing mechanisms (Laclavik et al., 2007).

The methodology was implemented and tested on

English and Slovak content. Our system was

designed to process multilingual documents,

including Latin languages, and so far it has provided

good results for a corpus of Romanian documents,

by using resources specific to the Romanian

language.

3 ARCHIVAL DOMAIN MODEL

This paper proposes a generic representation of the

archival domain as illustrated in Figure 1. The

archival domain is modelled starting from the raw

medieval documents provided by the Cluj County

National Archives (CCNA, 2008). These documents

are hand written and contain many embellishments,

making them hard to be automatically processed.

Due to this difficulty, in our case studies we have

used document summaries generated by the

archivists (see Figure 2).

Within our model, the central element is the

document. Documents belong to a specific domain

such as the historical domain or the medical domain.

In our research we have used the historical archival

domain, formally represented as domain knowledge

by means of domain ontology (concepts and

relations) and rules. Documents can be obtained

from several data sources like external databases,

Web sites or digitized manuscripts.

Figure 1: The archival domain model.

The document content (see Figure 2) is expressed

in natural language in an unstructured manner. In

our case study, the document content actually

represents a summary of the associated original

document. Several documents may be related to one

another by referring information about the same

topics even if they are not containing the same

lexical representations (e.g. names, events, etc.). The

document also features a set of technical data, such

as the date of issue, archival fund or catalogue

number. In the case of the document shown in

Figure 2, the technical data specifies the document

number (“235”), the language in which the raw

document was written (“Latin”) and the edition in

which the original document has appeared

(“Zimmermaan-Werner 1892 –I, nr.169”).

When searching in the archival documents it is

important to identify all documents that are related

to a specified topic. To enable information retrieval

from all relevant documents, the domain knowledge

is used to add a semantic mark-up level to the

documents content.

Figure 2: Example of a document which contains technical

data and the summary of the original archival document.

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

152

The domain knowledge core (domain ontology

and rules) is captured by processing and analyzing a

large repository of archival documents, focusing on

identifying their common concepts and

relationships. Next, based on information extraction

techniques applied on the raw documents, the

domain knowledge is enriched through instance

population.

4 SYSTEM ARCHITECTURE

AND FUNCTIONALITY

The system is structured on three interacting

processing layers: the raw data acquisition and

representation layer, the knowledge acquisition

layer and the knowledge processing and retrieval

layer. The layers and their associated resources and

processes are shown in Figure 3. The Primary

DataBase (PDB) is used for raw document

persistence, while the Knowledge Server (KS) is

used for learning and reasoning tasks.

The Raw Data Acquisition and Representation

layer provides support for collecting and storing data

in the Primary DataBase from multiple sources by

means of Optical Character Recognition (OCR)

techniques on raw documents, data import from

external databases or by means of the system’s

integrated user interface.

Figure 3: An overview of resources and processes.

The Knowledge Acquisition layer uses pattern-

matching to extract relevant data from the raw

documents. Based on the domain ontology and on a

set of semantic rules, the documents are then

semantically annotated. New concepts and instances

are identified and added to the domain ontology as a

result of this process.

The Knowledge Processing and Retrieval layer

enables ontologically-guided intelligent searches

over the annotated documents.

The system’s main workflows capture the

processing steps of the Knowledge Acquisition layer

and of the Knowledge Processing and Retrieval

layer.

5 KNOWLEDGE ACQUISITION

The objective of the Knowledge Acquisition layer is

to extend the domain knowledge by identifying,

extracting and annotating the relevant domain-

specific information from the summaries of archival

documents (see Figure 4).

Knowledge Acquisition uses text mining

techniques (tokenization, pattern matching and data

structuring processes) applied in a pipeline fashion

over the documents’ content. Actually the

Knowledge Acquisition layer extends the OntoPop

(Amardeilh, 2006) (Amardeilh, 2007) methodology

with two additional processing steps. The first step,

synonyms population, is required for identifying and

processing ontology instances having several lexical

forms with the same meaning (i.e. they are

synonyms) in different documents. For example, the

names “Palostelek” present in one document and

“Paulusteleky” in another document have been

identified and further processed as synonyms. The

second step, homonym identification and

representation, deals with common lexical

representations for different instances. As an

example, the name “Mihai” may refer either to the

same person or to different persons in different

documents.

In the following, we describe the main activities

of the knowledge acquisition workflow.

Technical Data Extraction. This activity is

responsible for separating the document technical

data from its content (for technical data examples

see Section 3 and Figure 2).

Lexical Annotation. The objective of this activity is

to identify and annotate the relevant lexical elements

in the content based on pattern-matching rules. A

pattern matching rule defines relationship between

the lexical elements and their annotation elements.

The output of the lexical annotation activity consists

of annotated lexical data represented in a hierarchic

format of extracted words along with their

annotation elements according to the pattern

matching rule (see Figure 8 in the Case Study

ARHINET - A System for Generating and Processing Semantically-Enhanced Archival eContent

153

Section for an example).

Knowledge Extraction. The objective of

knowledge extraction activity is to use the domain

ontology in order to semantically annotate the

hierarchical structure of annotated lexical elements

obtained in the previous activity. This activity is

supported by a set of mapping rules. Each mapping

rule defines (i) ways of associating the annotated

lexical elements to ontology concepts and (ii) a set

of actions for populating the ontology with instances

and relations. The result of the knowledge extraction

activity is an RDF structure stored in a file

associated to the original document content (see

Figures 10 and 11 in the Case Study Section for

examples).

Figure 4: Knowledge acquisition.

Ontology Population and Management. Ontology

population activity integrates the new instances,

synonyms, homonyms and properties identified in

the knowledge extraction activity into the domain

ontology. By using a dictionary of synonyms, the

domain ontology is populated with all the synonyms

of an instance. The OWL-Lite ontology

representation allows synonym definition through

the “sameAs” property, which specifies that an

instance X is equivalent with another instance Y. To

address the problem of homonyms we defined a

distance function that takes as arguments the

document technical data, attributes and relations of

the ontology-stored potential instances and the

current instance being verified for the homonymous

relationship. If the computed function value exceeds

a certain threshold we consider that the two

instances are identical, otherwise they are

homonyms.

Figure 5: Example of logical inference rules.

Ontology management activities aim (i) to infer

new relations and properties as a result of ontology

modification due to previous population processes

and (ii) to preserve ontology consistency. For the

inferring of new properties we have used the Jess

rule engine (Sandia National Laboratories, 2008).

For example, in a newly processed document we

have identified the new instance “Mihail” and its

associated property “hasFather” having the range

“Albert de Juc”. After populating the ontology with

this information, the Jess rule engine infers the

inverse property “hasSon” with the domain “Albert

de Juc” and the range “Mihail”. To enable these

logical inferences, the Jess rule engine requires

SWRL (Horrocks et al., 2004) rules to be defined on

the domain ontology. An example of ontology

associated SWRL rules is illustrated in Figure 5. The

first rule defines the “hasSon” relation between two

persons as the inverse of the “hasFather” relation

between the same persons. The second rule shows

that the “hasBrother” relation between two persons

is symmetrical.

6 KNOWLEDGE PROCESSING

AND RETRIEVAL

The aim of the knowledge processing and retrieval

layer is to provide support for intelligent queries that

enable searching for the most relevant information

available in archival documents. Document

searching is performed at two levels: one level relies

on the technical data, which narrows the set of

documents, while the other level relies on the

semantic meaning of the user input query. The user

input query triggers a complex reasoning process

that includes synonym search, logical inferences and

subclass / super class searches. As a result, the set of

query relevant documents is identified and new

Person(?s) ^ Person(?f) ^ hasSon(?f, ?s) -> hasFather(?s, ?f)

Person(?p) ^ Person(?b) ^ hasBrother(?p, ?b)

-> hasBrother(?b, ?p)

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

154

query relevant knowledge may be generated. We

used the Jess rule engine and a set of SQWRL

(SQWRL, 2008) rules as the main tools for

knowledge processing and retrieval.

Usually, in historical documents, several terms,

such as person or location names, have different

representations around a common root and it is

essential to identify all documents containing these

synonyms. To address this problem, the initial

search query is improved to search for instances

connected by the “sameAs” property, thus enhancing

the search process. Figure 6 presents an example of

a SQWRL enhanced query that enables synonym

search. The query searches in the ontology a person

that is Magister (“magistru”), taking in consideration

all the possible representations (synonyms) of the

person’s name.

Figure 6: Search query example.

The initial search query is further improved by

considering a sub-tree of the ontology that contains

the searched concept (class). For example, if the

instance “Transylvania” belonging to the ontology

concept “Principality” is the search key, the query is

refined to enable searching for the instances of its

super class “TerritorialDivision”.

7 CASE STUDY

The system proposed in this paper was used for

developing and processing semantically enhanced

archival eContent from documents capturing the

history of Transylvania, starting from the medieval

period. The historical documents have been obtained

from the Cluj County National Archives (CCNA,

2008). The original documents are found in Latin,

Hungarian, German and Romanian. Each document

is associated with a document summary in

Romanian which highlights the events and

participants. These summaries were used in our

system as the raw documents, the main source of

information.

An example of such a document summary which

will be used for further exemplification throughout

this section is presented in Figure 7. In English, this

summary reads: “Carol Robert, the king of Hungary,

Figure 7: Medieval document summary.

donates to Mihail and Nicolae, the sons of Albert of

Juk, the Palostelek domain and the Imbuz (Omboz)

forest, in the Dabaca County, for their faithful

military services carried out together with the

magister Stefan, against Moise, a rebel against the

crown”.

Figure 8: The core of the domain ontology.

In order to specialize the generic knowledge

acquisition workflow presented in Figure 4 for the

historical domain, we have used a corpus of several

documents for creating (i) a core for the historical

domain ontology, (ii) a set of specific pattern-

matching rules for the annotation of the lexical

elements and (iii) a set of mapping rules between the

annotated lexical elements and ontological concepts.

Figure 8 presents the core of the domain

ontology that was manually created by using the

Protégé Ontology Editor (Horridge et al., 2007),

based on the analysis of several raw documents.

We created the set of specific pattern-matching rules

for the annotation of lexical elements as JAPE

grammars (Tablan et al., 2004). A JAPE grammar

groups in phases the rules that specify actions to be

performed when certain patterns are matched. Such

a JAPE rule can be seen in Figure 9. The rule

searches for instances of the child–parent

relationship by looking for specific linguistic

construction patterns. The presented rule,

Person(?p0) ^ rdfs:label(?p0, ?p0_name)

^ sameAs(?p0,?p0s) ^ rdfs:label(?p0s, ?p0s_name)

^ hasTitle(?p0,?ha0) ^ rdfs:label(?ha0,"magistru")

-> sqwrl:selectDistinct(?p0s_name, "Person")

ARHINET - A System for Generating and Processing Semantically-Enhanced Archival eContent

155

Figure 9: Example of a JAPE rule

CandidateKinship_XSonOfY, finds phrasal patterns

of the form “X son of Y”, in order to, annotate the

lexical elements X and Y as Person, Person

Collection (several persons connected by commas

and conjunctions) or Complex Person (a lexical

construct consisting of a name and a title).

For the identification of proper names in the raw

documents, we used an existing Romanian gazetteer

(Tablan et al., 2004), that provides lists of Romanian

words. We have enriched the gazetteer with

additional lists that contain information specific to

the addressed historical periods, such as events,

kinship relations, titles, estates, etc.

Within the process of annotating the lexical

elements, the raw document is passed along with the

gazetteer lists through the pipeline of JAPE

grammars for extracting and structuring the relevant

information. Inside this process, the JAPE grammars

are used with the ANNIE (Tablan et al., 2004)

information extraction system integrated in our

system using the GATE API (Tablan et al., 2004).

Figure 10: Information extraction results.

For the raw document shown in Figure 7, the

identified lexical elements are presented in Figure

10a. The result of the annotation of the lexical

elements is the XML file shown in Figure 10b. It

contains a hierarchic structure of the identified

lexical elements that are further semantically

annotated by using a set of mapping rules (see

Figure 11). Each mapping rule associates to a

specific pattern defined in the XML file (i) a set of

ontology concepts that semantically annotate the

lexical elements and (ii) a set of operations that need

to be performed on the ontology in order to store the

identified information (instance population,

definition of properties and relations).

In the mapping rule shown in Figure 11, for the

lexical tag Kinship_XSonOfY, the child elements X

and Y are semantically annotated with the ontology

concept Person. The mapping rule also specifies the

actions of (i) adding X and Y as instances of Person

into the ontology and (ii) defining the hasFather

relation between the two instances.

For the currently processed raw document, the

Phase: Kinship_XSonOfY

Input: Lookup Token SpaceToken TempPerson Title

TitleComplex

Options: control = appelt

Macro: PERSON_COLLECTION

(

({TempPerson}

({Token.kind == punctuation, Token.string == ","} |

(SPACE {Token.string == "si"}))?

SPACE )+

{TempPerson}

)

Macro: PERSON_COMPLEX

(

{TempPerson}

({Token.kind == punctuation, Token.string == ","} |

{Token.string == "si"})?

SPACE

({TitleComplex} | {Title})

)

Rule: CandidateKinship_XSonOfY

(

((PERSON_COLLECTION) |

PERSON_COMPLEX |

{TempPerson})

({Token.kind == punctuation, Token.string == ","})?

SPACE

{Lookup.majorType == kinship_relations}

SPACE

({Token.string == "lui"}

SPACE)?

{TempPerson}

):kinship -->

:kinship.Kinship_XSonOfY =

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

156

Figure 11: Example of a mapping rule.

knowledge extraction process also generates an RDF

file (see Figure 12) that contains RDF statements

capturing the semantic annotations of the document.

Figure 12: Example of RDF file.

After processing several documents within the

knowledge acquisition workflow, the domain

ontology is populated with new instances and

properties (see Figures 8 and 13).

Figure 14 illustrates an example of a semantic

query that aims to search all documents that provide

information about the territorial division

“Palostelek”. Most of the archival documents related

to Transylvania may contain names and terms

written in different forms because of linguistic and

and phonetic influences (Romanian, Hungarian,

German or Slavic). Consequently, the location name

“Palostelek” may also appear as “Paulusteleky”.

Similarly, the location name “Juc” may appear as

“Szuc”, “Suzuluk”, “Zuk” or “Suk”.

Figure 13: Ontology population results.

Figure 14: Query example.

When performing queries about a term, like

“Palostelek”, it is essential to also identify all

documents about “Paulusteleky” and other possible

formats. This problem is solved within the synonym

search functionality implemented as part of the

knowledge processing and retrieval workflow (see

Section 6).

The whole set of relevant document contents

obtained after executing the query illustrated in

Figure 14 is presented in Figure 15.

Figure 15: Query results.

TerritorialDivision(?t0) ^ rdfs:label(?t0, ?t0_name)

^ swrlb:contains(?t0_name, "Palostelek")

^ sameAs(?t0,?t0s) ^ rdfs:label(?t0s, ?t0s_name)

-> sqwrl:selectDistinct(?t0s_name, TerritorialDivision")

<sem_tag>Kinship_XSonOfY</sem_tag>

<child_tag>Person</child_tag>

</context>

<atype>addInstance</atype>

<aclass>Person</aclass>

<aobject child="yes" index="0">Person</aobject>

</action>

<atype>addInstance</atype>

<aclass>Person</aclass>

<aobject child="yes" index="1">Person</aobject>

</action>

<atype>hasFather</atype>

<aclass>Person</aclass>

<aobject index="1">Person</aobject>

</action>

</actions>

</rule>

ARHINET - A System for Generating and Processing Semantically-Enhanced Archival eContent

157

8 CONCLUSIONS AND FUTURE

WORK

The present paper proposes a generic model of the

archival domain and offers a technical solution for

generating, semantically enhancing and processing

archival eContent. The solution follows two main

workflows: knowledge acquisition and knowledge

processing and retrieval. Within the knowledge

acquisition workflow we extend the OntoPop

methodology by adding two processing steps that

regard synonym and homonym population.

Reasoning techniques applied in knowledge

processing and retrieval enable ontology-guided

intelligent queries aiming at the identification of

relevant documents and knowledge.

For future work, we intend to extend our system

to enable ontologically-guided natural language

query processing and multilingual transparency.

Moreover, the domain ontology will be

automatically extended with new knowledge

extracted from documents.

ACKNOWLEDGEMENTS

This work was supported by the ArhiNet project

within the framework of the “Research of

Excellence” program initiated by the Romanian

Ministry of Education and Research.

REFERENCES

Amardeilh, F., 2007. Web Sémantique et Informatique

Linguistique: propositions méthodologiques et

réalisation d’une plateforme logicielle. These de

Doctorat, Universite Paris X-Nanterrere.

Amardeilh, F., 2006. OntoPop or how to annotate

documents and populate ontologies from texts. in

Proceedings of the ESWC 2006 Workshop on

Mastering the Gap: From Information Extraction to

Semantic Representation, Budva, Montenegro, June

12, 2006. CEUR Workshop Proceedings,

ISSN 1613-0073.

Buitelaar, P., Cimiano, P., Racioppa S., Siegel, M., 2006.

Ontology-based Information Extraction with SOBA.

In Proceedings of the International Conference on

Language Resources and Evaluation, pp. 2321-2324.

Laclavik M., Ciglan M, Seleng M, Krajei S., 2007. Ontea:

Semi-automatic Pattern based Text Annotation

empowered with Information Retrieval Methods. In

Tools for acquisition, organisation and presenting of

information and knowledge: proceedings in

Informatics and Information Technologies, Kosice:

Vydavatelstvo STU, Bratislava. ISBN 978-80-227-

2716-7, part 2, pp. 119-129.

Schäfer, U., 2007. Integrating Deep and Shallow Natural

Language Processing Components – Representations

and Hybrid Architectures. Saarbrücken Dissertations

in Computational Linguistics and Language Te, DFKI

GmbH and Computational Linguistics Department,

Saarland University, Saarbrücken, Germany.

Tablan V., Maynard D., Bontcheva K., Cunningham H.,

2004. Gate – An Application Developer’s Guide.

Available online: http://gate.ac.uk/.

Sandia National Laboratories, 2008. Jess the Rule Engine

for the Java Platform, Version 7.1. Available online:

http://www.jessrules.com/jess/docs/Jess71.pdf.

Horrocks I., et al., 2004. SWRL: A Semantic Web Rule

Language Combining OWL and RuleML. Available

online: http://www.w3.org/Submission/SWRL/.

SQWRL: Semantic Query-Enhanced Web Rule Language.

Available online: http://protege.cim3.net/cgi-

bin/wiki.pl?SQWRL, Date accessed 01/06/2008.

Horridge, M., et al., 2007. A Practical Guide to Building

OWL Ontologies Using Protégé 4 and CO-ODE

Tools. Available online: http://www.co-

ode.org/resources/tutorials/ProtegeOWLTutorial-

p4.0.pdf.

CCNA, Cluj County National Archives, 2008,

Online: http://www.clujnapoca.ro/arhivelenationale/.

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

158