A.1 Basic Information
Ontea tool analyze document or text using a regular expression patterns and detects equivalent semantics elements according to defined domain ontology. Several cross application patterns are defined but to achieve good results new patterns need to be defined for each application. Ontea also creates new ontology individual of defined class and assignees detected ontology elements/individuals as properties of defined ontology class.
A.1.1 Basic Terms
Semantic Annotation of Text
Identification of formalized objects in texts
Formalized model of problem environment understandable by computer system
Regular expression patterns
regular expression is a string that is used to describe or match a set of strings, according to certain syntax rules
A.1.2 Method Description
Ontea works on text, in particular domain described by domain ontology and uses regular expression patterns for semi-automatic semantic annotation. In Ontea we try to detect ontology elements within the existing application/domain ontology model. It means that with the Ontea annotation engine we want to achieve the following objectives:
§ Detect meta data from the text
§ Prepare improved structured data for later computer processing
§ Structured data are based on the existing application ontology model
The tool results can be made more precise by connecting Ontea with other tools for lemmatization (e.g. The Tvaroslovnik tool for lemmatization of Slovak) and also by estimating relevance of new created individuals by information retrieval tools such a s RFTS or Lucene.
Ontea can be executed in 3 different scenarios:
§ Ontea: searching relevant concepts in knowledge base (KB) according to generic patterns
§ Ontea creation: creating new individuals of concrete application specific objects found in text
§ Ontea IR: Similar as previous with the feedback of RTFS or other information retrieval tool (e.g. Lucene) to get relevance computed above word occurrence.
Use of Ontea, 1st scenario is described on example from Job Offer Application.
Figure 1 Job Offer Ontology
On Figure 1 main components of Job Offer ontology are displayed. Important fragments on ontology are Location or Skills individuals which can be then detected by annotation.
Figure 2 Web Document of Job Offer
Figure 3 Offer Individual Created by Ontea
On Figure 3 the individual of the Job Offer is created based on the semantic annotation of a Job Offer document (Figure 1), using simple regular expression patterns (see regular expression patterns chapter) where main individuals can be detected by the title property such as skillXML or skillPHP individuals. Other specialized patterns such as pattFullTime are used to detect concrete job offer properties – jtPermanent individual, which represents a permanent job position. In this example the job offer location – San Francisco id identified by a regular expression ([-A-Za-z0-9]+ [ ]+[-A-Za-z0-9]+), because individual locSF has the property title „San Francisco“. Similarly, other ontology elements are detected. Some regular expressions search for ontology individuals, other ontology classes and others such as pattFullTime annotate a job offer by a concrete individual jtPermanent if expression (Full[ -]Time) is found. Systems detect ontology elements based on domain ontology. In this example it is ontology of job offers.
Detected ontology individuals are then assigned as properties of job offer, thus ontology instance of job offer is created out of its text representation in the NAZOU pilot application.
§ Ontology based Text Annotation – OnTeA; Michal Laclavik, Martin Seleng, Emil Gatial, Zoltan Balogh, Ladislav Hluchy; Information Modelling and Knowledge Bases XVIII. IOS Press, Amsterdam, Marie Duzi at al., Frontiers in AI, Vol. 154, February 2007, pp.311-315. ISBN 978-1-58603-710-9, ISSN 0922-6389.
§ Hatcher E., Gospodnetić O., Lucene in Action, Manning (12 Jan 2005), ISBN: 1932394281
§ Log4J, Java-based logging utility, Apache Software Foundation. (http://logging.apache.org/log4j)
§ Jena, http://jena.sf.net/
§ Sesame, http://openrdf.org/
A.2 Integration Manual
Ontea is developed in Java (Standard Edition 5) and distributed as a jar archive. Access to the functionality of the tool is provided through Java Interface. Ontea is not a stand-alone application; the tool is proposed to be included in other application/tool, which will call the Ontea methods, however several classes are implemented with main() method to run some test and example functionality.
Ontea uses following libraries:
§ Log4J logging utility
§ Jena or Sesame semantic web library depending on used version
§ Lucene information retrieval library in case of use in IR scenario
§ RFTS tool in case of IR scenario
§ Tvaroslovnik tool in case of connection with lemmatizator
Deploying Ontea into other application requires the following steps (Java Integrated Development Environment and Apache Ant should be used):
1. download all Ontea files
2. ant start to start the demo or ant dist to create jar file and use library
Ontea use property file which need to be defined while creating onto.sesame.Memory instance. See Java Doc for more details.
Values such as Sesame repository type, location or ontology namespaces need to be defined. See onto.sesame.Config Java Doc for more details.
ontea.config package contain of pattern.property files for different languages or applications where regular expression patterns can be modified
A.2.4 User Guide
If you want to use Ontea without deep knowledge of its structure and functionalities, you should build your code based on examples in ontea.example package. Class FilesAnnotation shows example how to annotate text files in the defined directory by applying regular expression patterns from user defined property file. Classes DistributedFilesAnnotationStepSearch and DistributedFilesAnnotationStepCreate are equivalent to Scenarios of use “Ontea create” and “Ontea search” described in chapter A.1.3.
Annotation examples are also provided within JUnit test classes e.g. ontea.core.test. TestPatternRegExp class, showing annotation of String using defined regular expression.
Main Ontea classes and interfaces are located in ontea.core package. Main objects are:
Ontea annotation is build on pattern based approach, thus Pattern interface is one of main Ontea objects. Currently only one implementation of Pattern is provided: PatternRegExp, which annotates using regular expression patterns. Each Pattern implementation has to implement annotates method, which annotates given text and returns Set of Results. Result represents result of annotation, which is in fact individual of certain type or class. So far there are two Result super classes: ResultRegExp and ResultOnto. While Result is general individual not depending of used pattern or ontology/model implementation, Result extensions are created or found by applying result transformers on Result instances or Result sets.
These transformers are located in ontea.transform package. Package contain ResultTransformer interface with methods for transforming Result and Result set and also its implementations. Some of implementations are annotation result lemmatization, ontology knowledge base mapping or relevance identification using information retrieval methods.
Other packages and classes are related to transformers implementation, e.g manipulating sesame repository or integrating lucene indexing functionality to identify relevance.
Use of transformers can be seen in nazou.integration.Ontea class, which shows use of sesame transformers or ontea.example.FilesAnnotationEvaluation class, which show use of several transformers in a row.
A.3 Developer Manual
A.3.1 Tool Structure
Architecture of the system contains similar elements as the main annotation algorithm described in next chapter. Inputs are text resources (HTML, email, plain text) which need to be annotated as well as corresponding patterns from property files. An output is a new ontology individual, which corresponds to the annotated text. Properties of this individual are filled with detected ontology individuals according to defined patterns.
When extending the code of Ontea some of following classes and interfaces need to be extended and implemented:
§ ontea.core.Pattern: interface to addpot different pattern annotation tehniques
§ onetea.core.Result: class representing results of pattern annotation – instances of defined type.
§ ontea.transform.ResultTransformer: interface transforming results of annotation to different type or quality of results e.g. concrete ontology mapping, knowledge base implementation or result quality checking.
Ontea works with RDF/OWL Ontologies. It is implemented in Java using Jena Semantic Web Library or Sesame library. In both implementation inference is used to achieve better results.
Figure 4 Ontea architecture
The Ontea tool can be connected with other tools as it is also done in NAZOU project. Scheme of such connection can be seen on Figure 4 where following steps are performed:
1. retrieval of document for indexing
2. conversion of documents to plain text
3. Language identification
4. execution of lemmatization tool based on detected language
5. writing of information about document to the databaze (e.g. language, location or indexes)
6. reading data for conversion from relational database to ontology (e.g. language or file location)
7. Writing of empty document instances to ontology with info from previous step
8. Reading of empty document instances with file location
9. Reading text from file for annotation
10. lemmatization of words from document
11. Relevance of new created individuals using IR methods
12. writing of detected And created individual to document instance and to ontology knowledge base
Figure 5 Ontea integration with other tools.
A.3.2 Method Implementation
The underlying principle of the Ontea algorithm can be described by the following steps:
The algorithm also uses inference in order enable assignment of a found individual to the corresponding property also if the inferred type of a found individual is the same as the property type. The weak point of the algorithm is that if the ontology definition corresponding with the detected text contains several properties of the same type, in this case detected individuals cannot be properly assigned. This problem can be overcome if algorithm is used only on creation of individuals of different property types. Crucial steps of the algorithms as well as inputs and outputs can be seen also on Figure 4.
Regular expression patterns are the key element of the Ontea algorithm. Usually for each problem domain we need to define new, problem specific patterns to match the ontology elements in the text. However several cross application patterns exist:
· Matching one word starting with capital letter: ([A-Z][-A-Za-z0-9]+)
· Two words pattern: ([A-Z][A-Za-z0-9]+[\s]+[A-Z][A-Za-z0-9]+)
· Similar for 3 and 4 words pattern.
When individuals in the reference domain ontology contain plain text labels describing or naming the individuals, they can be detected by Ontea using the patterns mentioned above. Even when using such simple patterns, we can achieve satisfactory results.. If the reference ontology contains critical amount of individuals with assigned text labels, results of annotation are satisfactory with above mentioned cross application patterns. A good example is the location ontology, which is used in both examples provided in this paper. The location ontology contains concepts such as regions, countries, settlements, mountains, rivers or lakes as well as individuals of such classes. It is possible to create such ontology with concrete individuals of towns, settlements, mountains or rivers. Such data are available on the internet. We have also created such ontology containing all geographical data for Slovakia.
When keywords such as “Danube” or “Bratislava” appear in the text, correct individuals are detected by Ontea, where the Danube is an individual of a river and Bratislava is an individual of the Capital subclass of the settlement and the town. Similar detection can be achieved also with concepts such as a skill, a company or a category when ontology consists of a critical mass of individuals.
As already described, Ontea not only detects but also creates individuals if patterns are set up that way. A good example is again the location ontology. In many web pages for instance job offers, location is referenced by text “Location: City or Region Name”. When we convert web pages to a plain text, a regular expression pattern can be easily set up as follows: Location:[\s]*([A-Z][-a-zA-Z]+[\s]*[A-Za-z0-9]*). This will match one or two words after the “Location:” string. If document contains e.g. “Location: New York” and “New York” text is not found in any individual in the reference ontology, we can create a new simple individual of the region type (this is set up in Class property of pattern) with rdfs:label “New York”. In the future if New York is found in another document, the same individual is detected. Note that if we would create e.g. “New York City” individual, one can be sure that New York City is not a region but rather it’s subclass - the town in our location ontology. If we would change manually such individual to be an individual of the town class, it would be updated automatically in all detected data.
We think that with Ontea it is possible to detect or create ontology elements within the reference ontology with satisfactory results. This can be achieved automatically or semi automatically, when an expert can review and update the results.
Each pattern in Ontea contains of several properties which are defined in java property files:
§ PATTERN: contain regular expression pattern which is search in text
§ CLASS: represents URI of ontology class. Individuals are searched within this class.
§ CREATE: If set to True, and PATTERN was found in text but not found in knowledge base, Ontea will create simple individual of CLASS type in knowledge base with only property of found text as rdfs:label.
§ INDIVIDUAL: if not empty and contains of some URI and PATTERN is detected in text, individual is added to set of detected individuals
§ KEYWORD: if not empty and new individual is being created due CREATE, it will use keyword to find relevance using RFTS tool. For example if term „Google“ is often near term „Inc.“. Base on this relevance and set up threshold individual is created or not.
A.4 Manual for Adaptation to Other Domains
The Ontea tool can be used in different application domains. Used annotation method is generic and it can work with any text and OWL based ontology model.
A.4.1 Configuring to Other Domain
When using Ontea in other domain it is necessary to provide following modifications:
§ To change application or domain ontology
§ To change or modify used regular expressions
The tool search or create ontology individuals in scope of domain ontology model, thus change of ontology model is related to Ontea core functionality.
When changing regular expression patterns, we can still use some generic patterns or reuse patterns for specific individuals/objects. See section on regular expression patterns.
If Ontea is used with Information Retrieval tools such as Lucene or RFTS, depending on language of processed texts, appropriate lemmatization tool need to be used.
Log4J is involved domain independently into the Ontea.
Lemmatization tool need to be used based on language of texts in the domain. E.g. Tvaroslovnik for Slovak or Porter’s algorithm for English.
 GEOnet Names Server, Names, 2006, http://earth-info.nga.mil/gns/html/cntry_files.html