Jarrar.lecture notes.arabicontology

Lecture Notes
Birzeit University, Palestine
2012

Advanced Topics in Ontology Engineering

Arabic Ontology

Dr. Mustafa Jarrar
Sina Institute, University of Birzeit
mjarrar@birzeit.edu
www.jarrar.info

Jarrar © 2012 1

Reading

Mustafa Jarrar: Building A Formal Arabic Ontology (Invited Paper) . In proceedings of
the Experts Meeting On Arabic Ontologies And Semantic Networks. Alecso, Arab League.
Tunis, July 26-28, 2011.Article http://www.jarrar.info/publications/J11.pdf.htm
Slides: http://mjarrar.blogspot.com/2011/08/building-formal-arabic-ontology-invited.html

Mustafa Jarrar: Towards The Notion Of Gloss, And The Adoption Of Linguistic
Resources In Formal Ontology Engineering. In proceedings of the 15th International World
Wide Web Conference (WWW2006). Edinburgh, Scotland. Pages 497-503. ACM Press. ISBN:
1595933239. May 2006.http://www.jarrar.info/publications/J06.pdf.htm

Jarrar © 2012 2

Arabic Ontology Team (Current)

Fateh Badran Collaborators
Maather Alkqam Jamal Daher
Rabaa Yusuf
Prof. Mahi Arrar
Razan Marwan
Haneen Somoom
Naser Musleh
Jarrar © 2012 3

The Arabic Ontology Project
http://sites.birzeit.edu/comp/ArabicOntology

• A project started in 2010, at Birzeit University, Palestine.
• The ArabicOntology is more than an Arabic WordNet
• Unlike WordNet, the ArabicOntology is logically and philosophically well-
founded, as it follows strict ontological principles.  but can be used an
Arabic WordNet.

 The project is partially funded
(Seed funding) by Birzeit
University (VP academic
Office, Research Committee).

Jarrar © 2012 4

Arabic Ontology: Data Model (Simplified)
• ConceptID (as a synsetID in WordNet) to identify a concept.
• Polysemy and synonymy: like in WordNet, several words (i.e., lexical
units) can be used to lexicalize one concept (synonymy); and one word
might be used to lexicalize several concepts.

Lexical Unit

Gloss: describes a concept

Semantic Relations
Concept ID: concept unique reference

Jarrar © 2012 5

Lexical vs. Semantic Relationships
• Semantic relations are relationships between concepts (not
words), e.g., subtype, part-of, etc.
• Lexical relations are relationships between words (not
concepts), e.g., synonym-of, root-of, abbreviation-of, etc.
• Ontologies are mainly concerned with semantic relations.

Lexical Unit

Gloss: describes a concept
Semantic Relations

Concept ID: concept unique reference
Jarrar © 2012 6

Arabic Ontology

• Arabic Ontology: the set of concepts (of all Arabic terms), and the
semantic (not lexical) relationships between these concepts.
• To build an Arabic Ontology: Identify the set of concepts for every
Arabic word (Polysemy), and define semantic relations between these
concepts.
• Most important relation is the subtype relation,
which leads to a (tree of concepts) .

Jarrar © 2012 7

Arabic Ontology: Subtype Relationships
• Subtype relation: is a mathematical relations (subset: A B ), such
that every instance in A must also be an instance of B.
• Inheritance: subtypes inherit all properties of their super types.
• “Hyponymy” in WordNet is close to (but not the same as) the subtype relation.
• “General-Specific” relations, as in thesauri, are not subtype relations.

world
. . . .. .
. . .
. . .. . .. . . .
. . . 10 . . 3 . . .
. 6 . . .
. . . . .. . 4 . .. .
. . . . .. . . . . .
. .
. . . .. .
Jarrar © 2012 8

Arabic Ontology: Subtype Relationships

• It is recommended to use proper subtypes, as it is more strict.
• That is, A and B are never equal, B is always a super set of A.
• It is recommended to classify concepts based on “rigidity”.
• For example it is wrong to say that a „WorkTable‟ is type of „Table‟.
as being a work table is a non-rigid property.
• As such, subtypes form a tree.

Jarrar © 2012 9

Arabic Ontology: Core (Top Levels).

Arabic Core Ontology: the top levels of the Arabic Ontology, - built
manually based on DOLCE and SUMO upper level ontologies, and
taking into account, carefully, the philosophical and historical aspects of
the Arabic conceptsterms.
Top 3 levels shown here, for simplicity

10 levels, 550 concepts
• The 10th level of this core ontology should top all Arabic concepts and levels.
• This allow us to detect any problems in the tree/relations!
• The core Ontology governs the correctness and the evolution of the whole
Arabic Ontology.
Jarrar © 2012 10

Arabic Ontology: Glosses
according to strict ontological guidelines[J06]

A gloss: is an auxiliary informal (but controlled) account of the intended
meaning of a linguistic term, for the commonsense perception of humans.

A gloss is supposed to render factual knowledge that is critical to understand a concept, but that
e.g. is implausible, unreasonable, or very difficult to formalize and/or articulate explicitly. (NOT) to
catalogue general information and comments, as e.g. conventional dictionaries and encyclopedias
usually do, or as <rdfs:comment>. Jarrar © 2012 11

Arabic Ontology: Gloss Guidelines

What should and what should not be provided in a gloss:
1. Start with the principal/super type of the concept being defined.
E.g. „Search engine‟: “A computer program that …”, „Invoice‟: “A business document
that…”, „University‟: “An institution of …”.

3. Focus on distinguishing characteristics and intrinsic prosperities that
differentiate the concept out of other concepts.
E.g. Compare, „Laptop computer‟:
“A computer that is designed to do pretty much anything a desktop computer can do, it runs for a
short time (usually two to five hours) on batteries”.
“A portable computer small enough to use in your lap…”.

2. Written in a form of propositions, offering the reader inferential
knowledge that help him to construct the image of the concept.
E.g. Compare „Search engine‟:
“A computer program for searching the internet, it can be defined as one of the most useful aspects
of the World Wide Web. Some of the major ones are Google, ….”;
A computer program that enables users to search and retrieves documents or data from a database
or from a computer network…”.
Jarrar © 2012 12


4. Use supportive examples :
- To clarify cases that are commonly known to be false but they are true, or
that are known to be true but they are false;

- To strengthen and illustrate distinguishing characteristics (e.g. define by
examples, counter-examples).

Examples can be types and/or instances of the concept being defined.

5. Be consistent with formal definitions/axioms.

6. Be sufficient, clear, and easy to understand.

 WordNet glosses do not follow such ontological guidelines

Jarrar © 2012 13


As a gloss starts with a supertype of concept being defined, try to read
the gloss as the following, to verify what you do is correct:

Jarrar © 2012 14

ArabicOntology Vs WordNet

Unlike WordNet, the Arabic Ontology is:

1. Philosophically well founded:
• Focuses on intrinsic properties;
• All types are rigid;
• The top level is derived from known Top Level Ontologies.

2. Strictly formal:
• Semantic relations are well-defined mathematical relations.

3. Strictly-controlled glosses
• The content and structure of the glosses is strictly based on
ontological principles.

Jarrar © 2012 15

Our Approach to Building the
ArabicOntology
Roughly:

Step1:
Mine Arabic concepts/glosses from dictionaries.

Step 2:
Automatically map between these Arabic concepts and WordNet
concepts, thus inherit semantic relations from WordNet.

Step 3:
Link all concepts with the Arabic Core Ontology.

Step 4:
Re-formulate these glosses, according to strict ontological guidelines.

Jarrar © 2012 16

Step1-Mining Arabic Concepts from
Dictionaries
• Collect as much glosses/concepts as possible from specialized and
general dictionaries.
• Manual extraction from dictionaries, then basic cleaning done
automatically.

Mining
concepts

• 75k glosses ready.
• We have ~100 students typing dictionaries now!
• +150K more glosses (expected this year)

Jarrar © 2012 17

Dictionaries
• Collect as much glosses/concepts as possible from specialized and
general dictionaries.
• Manual extraction from dictionaries, then basic cleaning done
automatically.

Mining
concepts

Jarrar © 2012 18

Dictionaries
• Most Arabic dictionaries are not useful, but some are a good start.
The dictionaries we need should:
 Focus on the semantic aspects.
 Multiple meanings are not mixed up.
 Structure of quality of the meaning.

Mining
concepts

Jarrar © 2012 19

Examples (Good & Bad Resources)

Wiktionary 











Jarrar © 2012 20

The Matching Function is used for:

1- Based on the previous mapping, we can inherit Semantic Relations
from WordNet.
Arabic Concepts WordNet Concepts

A
L
H
J B
R
C D Q

 Remark: This is only a good start, as these inherited relations need to be
cleaned using the Arabic Top Levels, and using the OnToClean Methodology.

2- Same function is used to detect redundant concepts, within the
Arabic Ontology itself.
Jarrar © 2012 21

Step2: Map Arabic concepts to WordNet
(Matching Function)

We developed a smart algorithm, such that:
Input: (Arabic gloss, 117k English glosses in WordNet).
Output: (best match, rank)
Accuracy: +90% (being improved)
WordNet (English)
The territory occupied by one of the constituent
administrative districts of a nation
The way something is with respect to its main
attributes
The group of people comprising the government
of a sovereign state
A politically organized body of people under a
single government
A compilation of the known facts regarding
something or someone
….

A politically organized
body of people under a
single government
Jarrar © 2012 22

Step 3: Link concepts with the Arabic Core
Ontology
Each Arabic concept (from previous steps) is mapped to a concept in the
10th level.
That is, the 10th level of this core ontology should top all Arabic concepts
and levels, so to enable automatic detection of problems in the hierarchy.


A C
J
J
Jarrar © 2012 23

Automatic Detection of Inconsistencies
Subtype links from Arabic concepts to the core ontology (done manually)
Subtypes links between Arabic concepts (derived via the mappings to WordNet)
Now we can automatically detect whether the links are correct?
 If (J A) and (A ) then it‟s most likely true that (J
) , thus no need to have (J ).
 However, as H and C don‟t share a supertype, (H C) is likely incorrect.

X!
A C X!
J
Jarrar © 2012
H 24

Step 4- Re-Formulate Glosses,
according to strict ontological guidelines[J06]

Glosses are re-formulated semi-manually, to meet our strict rules.

Gloss-cleaning can be done automatically to a certain point.

 While the manual-cleaning (=re-formulating) glosses, mistakes in
subtype relation can be detected.

Jarrar © 2012 25

Outline

• Arabic Ontology Project

• Design principles

• Methodology and progress

• Ongoing Research

• Matching Function

• The Top levels

• Finding Arabic Synonyms

Jarrar © 2012 26

Why the Matching Function

Arabic concepts WordNet:
without subtype Arabic Concepts WordNet Concepts English concepts
links between with subtype
them relations between
A them
L
H
J B
R
C D Q

Assumption: If we could know that the (English concept B = Arabic concept A)
and (English concept D = Arabic concept C), and (D SubtypeOf B) then we can
automatically deduce that (C SubtypeOf A).

Research Problem: given an Arabic concept, automatically find its equivalent
English concept (if exists).

Jarrar © 2012 27

The Matching Function
(Map Arabic concepts to WordNet)

A smart algorithm that maps a gloss of a concept written in Arabic into
its English equivalent in WordNet, such that:
Input: (Arabic gloss, 117k English glosses in WordNet).
Output: (best match, rank)

WordNet (English)
The way something is with respect to its main
attributes
The group of people comprising the government
of a sovereign state
A politically organized body of people under a
single government
A compilation of the known facts regarding
something or someone
….

Country
A politically organized
body of people under a
single government
Jarrar © 2012 28

Matching Function: Main Steps

Translate the Arabic gloss into English (using Google or Bing)

Find the minimal set of English Concepts(Synsets) that contains
the best match for an Arabic concept (i.e. Search Domain).

Rank the concepts in the search domain (according to how
much they are close to the Arabic Gloss)

 The best match of an Arabic concept is the English concept
with the highest score.

Jarrar © 2012 29

Translating Arabic Glosses into English

A country with defined
borders, the people and
Government and institutions

 This step is done easily, as Google and Bing provide APIs for
automatic translation.

 The translation is not always good, but satisfactory. It depends on
the quality of the gloss being translated.

Jarrar © 2012 30

Determine the Search Domain

Matching the translated gloss into an English gloss among (117k gloss)
is:
 Time consuming.
 The algorithm might be misled as the many English glosses.

Instead, Find the minimal set of English Concepts(Synsets) that
contains the best match for an Arabic concept (i.e. Search Domain).

 Translate the word “ ” into English using the Google and Bing
{state, country, nation, polity, land } =

Now, Search Domain = the set of all glosses of this set words + two
levels up and to levels down.

Jarrar © 2012 31

Search Domain

(~2000 glosses)
The way something is with respect to its main attributes
English Terms The group of people comprising the government of a sovereign
state
country A politically organized body of people under a single
government
state A compilation of the known facts regarding something or
someone
land
Put before
nation
A state of depression or agitation
polity
The territory occupied by a nation
earth
Express in words

…

 Our translated target gloss should be in in this search
domain
Jarrar © 2012 32

Search Domain

(~2000 glosses)
state
government
someone
land
Put before
nation
polity
earth
Express in words

• How many levels to go for?
…
– Too many levels increases the number of irrelevant synsets.
– Too few levels lowers the possibility of finding the best match.
• We need to find the minimal set of synsets that contains the best match
for the Arabic concept. Jarrar © 2012 33

The Ranking Step

Search Domain

(~2000 glosses)
state
government
someone
land
Put before
nation
polity
earth
Express in words

…

Jarrar © 2012 34

The Ranking Step

The ranking step is divided into several steps
Convert the translated gloss into an Array of words: original
words, their Synonyms, super/subtypes.

Give a weight for each word depending on its type. (Synonym,
direct Subtype, direct Supertype, second level Supertype)

Match the words in each concept in the search domain with the
word in the array.

For each hit, add the weight of the word to the synset score.
The synset with the highest score is the best match.
Jarrar © 2012 35

Convert the translated gloss into an Array of words
Array
A country with defined country
borders, the people defined
1-m
and Government and Original words
…
institutions Body politic
m-n
Commonwheath
Synonyms of
Array … Original words
Stop words
country 1 People
are removed n-o
defined Political entity
subtypes of
… Original words
borders
Asian country
people o-p
City state
Government subtypes of
… Original words
institutions m countries
defines p-q
The array 1-q might be 1500 words or defining stems of all
more. the idea is to have the relevant … words (1-p)
words that one may to describe the
concept being matched. Jarrar © 2012 36

Array
country
1-m
defined
Original words
… weight=a
Body politic
Give a weight for each word depending on Commonwheath
m-n
its type. (Synonym, direct Subtype, direct Synonyms of
… Original words
Supertype, second level Supertype) People weight=b
n-o
Political entity
subtypes of
… Original words
Asian country weight=c
o-p
City state
subtypes of
… Original words
countries weight=d
defines p-q
defining stems of all
… words (1-p)
weight=e

Jarrar © 2012 37

Array
Match the words in each concept in the country
search domain with the word in the array. defined
1-m
Original words
… weight=a
Search Domain Body politic
The territory occupied by one of the constituent administrative districts of a nation
m-n
Commonwheath
Synonyms of
… Original words
The group of people comprising the government of a sovereign state
People weight=b
A politically organized body of people under a single government n-o
Political entity
A compilation of the known facts regarding something or someone subtypes of
Put before
… Original words
Asian country weight=c
o-p
The territory occupied by a nation City state
subtypes of
Express in words … Original words
… countries weight=d
defines p-q
defining stems of all
… words (1-p)
For each hit, add the weight of the word to the array. weight=e

Jarrar © 2012 38

Ordered Search Domain

 The synset with the highest score is the best match.
Search Domain rank
The territory occupied by one of the constituent administrative districts of a
90
nation
The way something is with respect to its main attributes 89
The group of people comprising the government of a sovereign state 89
A politically organized body of people under a single government 78
A compilation of the known facts regarding something or someone 70
Put before 65
A state of depression or agitation 50
The territory occupied by a nation 20
Express in words 0

…

Jarrar © 2012 39

Ordered Search Domain

 The synset with the highest score is the best match.
Search Domain rank
The territory occupied by one of the constituent administrative districts of a
90
nation
The way something is with respect to its main attributes 89
The group of people comprising the government of a sovereign state 89
A politically organized body of people under a single government 78
A compilation of the known facts regarding something or someone 70
Put before 65
A state of depression or agitation 50
Express in words 0

…

 The results (until this step) were not too bad, but we added an extra
step, called “Rank x Centrality” to improve the accuracy.

Jarrar © 2012 40

The Rank X Centrality Step (Rank+)

 Rebuild the subtype links between concepts
Search Domain rank
50 65
The territory occupied by one of the P S
constituent administrative districts of a 90 Q A
nation 90
The way something is with respect to its
89 70
main attributes T B
The group of people comprising the W Q
government of a sovereign state 89
A politically organized body of people
78 20 89
under a single government F U
A compilation of the known facts 0 R E
regarding something or someone 70
O 78 89
Put before 65
A state of depression or agitation 50 G J
Express in words 0
…

R: the rank computed in the previous step
C: the centrality C of each concept in its tree, how many nodes above/under
Rank+ = R x C (the idea is to use centrality to compromise ranks that are
close to each other) Jarrar © 2012 41

The Experiment

Exp N° N° of glosses Match percentage
160 gloss More than 90%
1st experiment
1000 gloss In progress…
2 experiment
nd

Jarrar © 2012 42

Ongoing Research (Use Machine learning and Neural networks)
By Mohammed Mellhem

Static weights Dynamic weights

For each word found, a
weight is given as follows
Type Weight Learning to find the best
Value
weights and levels to expand
Keyword 1st 3 words 3
(using neural networks).
Keyword 4th and above 1

Super-type 0.7  The algorithm tries to find the
Sub-Type 0.4 maximum level of expansion that
Synonym 0.3 can help to find the best match
Sub-word 0.2

Bigram (not-imp.) 1.2

Jarrar © 2012 43

Progress so far

• Considering three parts:
1. Find best depth for Search Space expansion
2. Find best depth for Query expansion
3. Find best weights for best matching

• For point 1 and 2; the algorithm try to find the maximum level of
expansion that can help to find best match.

• For Point 3; the algorithm is still in the learning process, initial
results shows that in more than 85% of 162 processed till now,
found by concept keyword or keyword stem.

• Improve performance by loading matching data on demand.

Jarrar © 2012 44


What are the top 10 levels of the ArabicOntology? And how to build it?


• The 10th level of this core ontology should top all Arabic concepts and levels.
Arabic Ontology.
Jarrar © 2012 46


What are the top 10 levels of the ArabicOntology? And how to build it?

Methodology
(by Rana Rashmawi)

Arabic Core Ontology: the top levels of the Arabic Ontology, - built manually
based on DOLCE and SUMO upper level ontologies, and taking into
account, carefully, the philosophical and historical aspects of the Arabic
conceptsterms.

Phase 1. Determining the top (Core) Arabic concepts.

Phase 2. Constructing the Top Levels.

• Phase 3. Verifyingthis core ontology should top all Arabic concepts and levels.
The 10th level of the correctness and completeness of the top levels.
Arabic Ontology.
Jarrar © 2012 47

Phase 1. Determining the top (Core) Arabic
concepts
1. Both SUMO & DOLCE were translated to Arabic separately.
Where each English upper term was mapped to (3-5) Arabic
Concepts, based on the relevance of the meaning.
 The result of this step was a pool of 1200 Arabic terms, that
necessarily demonstrate the upper terms of the Arabic Language.

≈ SUMO Terms
Entity Arabic Terms ≈
Abstract
Physical
Attribute
…

= 80 DOLCE Terms
Entity
Abstract
Object
Event
… Jarrar © 2012 48

concepts
2. We investigated all the meanings for each Arabic term in the resulted
pool of terms, using Arabic Lexicons. Allowing some editing and
reforming of the glosses following strict Ontological guidelines.

 The result of this step was a larger pool of glosses (about 6000 glosses)
of the most general and comprehensive concepts in the Arabic Language.

Arabic Meaning ≈
Arabic Terms ≈
…

.

.

… Jarrar © 2012 49

concepts
3. The last step in this phase was to choose one Arabic concept for
each English concept for both SUMO & DOLCE.
 This step could be considered as a process of semantic mapping
between English and Arabic Concepts.

Arabic Terms Arabic Meaning
…

.

.

…

Jarrar © 2012 50

Phase 2. Constructing the Top Levels

• After determining the Arabic upper (core) Concepts that correspond
to the English Upper Concepts from both SUMO & DOLCE. We
derived the relations between these Arabic concepts depending on
the existing relations in both foreign Ontologies. As both of them
depend on the (sub/super-type relation) in their structure.

Entity

Abstract Object Info. Entity Event Quality

• The same methodology was used to extract the semantic relations
from both DOLCE & SUMO to complete all the lower levels, up until
the 10th.

Jarrar © 2012 51

Phase 3. Verifying the correctness and
completeness of the top levels

• A manual mapping of 6000 Arabic meaning was
done to verify the top levels.

Jarrar © 2012 52

Further/Ongoing Research

 Given many Arabic-English, Arabic-French, Arabic-Italian dictionaries

 Can we derive an Arabic-Arabic thesaurus? For example

 Then Categorize very-related words (maybe using WordNet) as the
following:

This will help finding possible Arabic synsets, which help detecting
possible subtype relations and/or validate the existing relations.

Jarrar © 2012 54

Table
sort ‫رتت‬ tidy ‫أنيق‬
arrange ‫رتت‬ tidy ‫ضخم‬
order ‫رتت‬ tidy ‫نظيف‬
set ‫رتت‬ tidy ‫منهجي‬
tidy ‫رتت‬ tidy ‫منظم‬
pack ‫رتت‬ tidy ‫مهندم‬
shape ‫رتت‬ pack ‫حشمخ‬
form ‫رتت‬ pack ‫رسمخ‬
sort ‫ نىع‬pack ‫وضت‬
sort ‫ شكم‬pack ‫تكىم‬
sort ‫ ضزة‬pack ‫حشب‬
sort ‫ هذا اننىع‬shape ‫شكم‬
sort ‫ صنف‬shape ‫حبنخ‬
sort ‫ طزيقخ‬shape ‫هيئخ‬
arrange ‫ نظم‬shape ‫مظهز‬
arrange ‫ اتخذ‬shape ‫تجسد‬
arrange ‫ سىي انخالف‬form ‫شكم‬
arrange ‫ ػدل‬form ‫استمبرح‬
order ‫ اننظبم‬form ‫صىرح‬
order ‫ تزتيت‬form ‫نىع‬
order ‫ أمز‬form ‫هيئخ‬
set ‫ مجمىػخ‬form ‫صيغخ‬
set ‫ وضغ‬tune ‫رتت‬
set ‫ ضجط‬form ‫ضزة‬
set ‫2102 © ثدأ‬
Jarrarorganize ‫رتت‬ 55

Jarrar.lecture notes.arabicontology

Recommended

Recommended

More Related Content

Similar to Jarrar.lecture notes.arabicontology

Similar to Jarrar.lecture notes.arabicontology (20)

Recently uploaded

Recently uploaded (20)

Jarrar.lecture notes.arabicontology