Person Name Entity Recognition for Arabic

Khaled Shaalan Hafsa Raza

Institute of Informatics Institute of Informatics

The British University in Dubai The British University in Dubai

P O Box 502216, Dubai, UAE P O Box 502216, Dubai, UAE

Khaled.shaalan@buid.ac.ae hafsa.raza@gmail.com

tactic Parsing/Chunking, Question-Answering,

Abstract among others. The valuable information in text is
usually located around proper names, to collect this
Named entity recognition (NER) is nowa-
information it should be found first (Abuleil, 2004;
days an important task, which is responsi-
Chinchor, 1998). In our presentation, we will con-
ble for the identification of proper names in
centrate on the role of NER in Information Extrac-
text and their classification as different
tion (IE). IE is the NLP task that retrieves relevant
types of named entity such as people, loca-
information from unstructured texts and produces
tions, and organizations. In this paper, we
as a result a structured set of data.
present our attempt at the recognition and
This paper describes work on recognizing and
extraction of the most important proper
extracting the most important entities, that is, per-
name entity, that is, the person name, for
son names for the Arabic language. We have
the Arabic language. We developed the
adopted the rule-based approach using linguistic
system, Person Name Entity Recognition
grammar-based techniques to develop PERA. This
for Arabic (PERA), using a rule-based ap-
approach provides flexibility and adaptability fea-
proach. The system consists of a lexicon, in
tures in our system and it can be easily configured
the form of gazetteer name lists, and a
to work with different languages, NLP applications,
grammar, in the form of regular expres-
and domains. In order to determine the best rules
sions, which are responsible for recogniz-
for recognition of person names, various Arabic
ing person name entities. The PERA sys-
text corpora were analyzed. Phrases containing
tem is evaluated using a corpus that is
person names were retrieved, the underlying pat-
tagged in a semi-automated way. The sys-
tern was learned and person indicators such as ti-
tem performance results achieved were sat-
tles were identified. Apart from this, person names
isfactory and confirm to the targets set
were extracted from the available corpora and
forth for the precision, recall, and f-
other resources to build up a lexicon, in the form of
measure.
gazetteer name lists, or gazetteer for short. The
1 Introduction various Arabic naming conventions and the person
indicators identified helped in deriving fine rules
The recognition and classification of proper names that gave high-quality recognition of person names
in text (e.g. persons, locations, and organizations) in Arabic text. The recognition was done in two
has recently become considered of major impor- cycles using first the gazetteer and then the gram-
tance in Natural Language Processing (NLP) as it mar rules. The PERA system is evaluated using a
plays a significant role in various types of NLP reference corpus that is tagged with person names
applications, especially in Information Extraction, in a semi-automated way. The achieved system
Information Retrieval, Machine Translation, Syn- performance results were satisfactory and confirm
to the targets set forth for the precision, recall, and used in question-answering systems. This work
f-measure. was done in three main stages: 1) marking the
The paper is structured as follows. Section 2 phrases that might include names, 2) building up
presents the related work. Section 3 describes the graphs to represent the words in these phrases and
naming conventions of person names used in Ara- the relationships between them, and 3) applying
bic language. Section 4 presents methods of data rules to generate the names, classify each of them,
collection used. Section 5 explains the system ar- and saves them in a database.
chitecture and implementation. Section 6 presents Larkey et al. (2003) have conducted a study that
the experiment performed to evaluate the system showed the importance of the proper names com-
and finally Section 7 concludes the paper, ponent in cross language tasks involving searching,
summarizes our achievements, and highlights our tracking, retrieving, or extracting information. In
plans for future work.. particular, they have concluded that a combination
of static proper name (English-Arabic) translation
2 Related Work plus transliteration provides a successful solution.
Pouliquen et al. (2005) developed a tool for mul-
As in other NLP techniques, there are two main
tilingual person name recognition that focuses on
approaches to NER (Toral, 2005). One is based on
the "Who" part of the analysis of large news text.
linguistic knowledge, in particular grammar rules
As multilingual NER is concerned, the translitera-
and hence called rule-based, while the other is
tion of the NE has included alternative spelling
based on machine learning techniques. The re-
variants where the origin language of the name is
quired resources for the knowledge approach are
usually not known. Several variants could also be
usually gazetteers and rules whereas the learning
found in the same language.
approach needs an annotated (tagged) corpus. The
Samy et al. (2005) has used parallel corpora in
linguistic knowledge-based model achieve better
Spanish, and Arabic and an NE tagger in Spanish
results in specific domains, as the gazetteers can be
to tag the names in the Arabic corpus. For each
adapted very precisely, and it is able to detect
sentence pair aligned together, they use a simple
complex entities, as the rules can be tailored to
mapping scheme to transliterate all the words in
meet nearly any requirement. However, if we deal
the Arabic sentence and return those matching with
with an unrestricted domain, it is better to choose
NEs in the Spanish sentence as the NEs in Arabic.
the machine learning approach, as it would be inef-
While they report high precision and recall, it
ficient to acquire and/or derive rules and gazetteers
should be noted that their approach is applicable
in this case.
only when a parallel corpus is available.
Name identification has been worked on quite
Zitouni et al. (2005) has adopted a statistical ap-
intensively for the past few years, and has been
proach for the entity detection and recognition
incorporated into several products. Many research-
(EDR). In this work, a mention can be either
ers have attacked this problem in a variety of lan-
named (e.g. John Mayor), nominal (the president)
guages but only a few limited researches have fo-
or pronominal (she, it). An entity is the aggregate
cused on NER for Arabic text. This is due to the
of all the mentions (of any level) which refer to
lack of resources for Arabic NE and the limited
one conceptual entity. This extended definition of
amount of progress made in Arabic NLP in gen-
the entity has proved the suitability of the ap-
eral.
proach.
Maloney and Niv (1998) developed TAGARAB
an Arabic name recognizer that uses a pattern- 3 Components of an Arabic Full Name
recognition engine integrated with morphological
analysis. The role of the morphological analyzer is Arabic has well-defined naming practices. The
to decide where a name ends and the non-name Arabic name elements may be divided into five
context begins. The decision depends on the part- main categories, Ibn Auda (2003):
of-speech of the Arabic word and/or its inflections. 1. An ism (pronounced IZM, as the final syllable
Abuleil (2004) presented a technique to extract in the word dogmatism), a personal, proper
proper names from text to build a database of name given shortly after birth, i.e. the given
names along with their classification that can be name. Examples of such names are Muham-
 mad [Mohammed], Musa [Moses], Ibrahim 4 Data Collection
[Abraham].
2. A kunya (pronounced COON-yah), an honor- The development of the system PERA depends on
ific name or surname, as the father or mother collecting dictionaries of proper nouns and their
of someone; e.g., abu Da'ud [the father of related indicators. Techniques used for acquiring
David], umm Salim [the mother of Salim]. It is such data to build the dictionaries include:
meant as a prefix of respect or reverence. Mar- 1. Automatic collection of person names from
ried persons (especially married ladies) are, as annotated corpus. The person entities in the
a general rule, simply called by their kunya ACE1 and Treebank corpus2 were recognized
(abu or umm + the name of their first-born and extracted using regular expression patterns
child). When using a person's full name, the coded within Python scripts. Python is a strong
kunya precedes the personal (given) name: Abu string processing language and widely used in
Yusuf Hasan [the father of Joseph, Hasan], developing NLP applications and tools.
Umm Ja’far Aminah [the mother of Ja’far, 2. Identification of person indicators. Apart from
Aminah]. extracting the person names, these corpora
3. By a nasab (pronounced NAH-sahb), a pedi- were used also to extract noun phrases contain-
gree, as the son or daughter of someone; e.g., ing the person names. The surrounding se-
ibn 'Umar [the son of Omar], bint 'Abbas [the quence of words around person names was
daughter of Abbas]. The nasab follows the ism analyzed to identify indicators of person
in usage: Hasan ibn Faraj [Hasan the son of names. A dictionary of these indicators was
Faraj], Sumayya bint Khubbat [Sumayya the formed which represented contextual cues of
daughter of Khubbat]. Many historical person- person names.
ages are more familiar to us by their nasab 3. Name Database provided by government or-
than by their ism: e.g., the historian ibn ganization. The person name dictionary was
Khaldun, the traveler ibn Battuta, and the phi- also build from names collected from some or-
losopher ibn Sina [Avicenna]. ganizations including Immigration Depart-
Nasabs may be extended for several genera- ments, Educational bodies, and Brokerage
tions, as may be noted in the example below companies.
containing two generations nasab: 4. Internet Resources. Names were retrieved fur-
Abu al-Qasim Mansur ibn al-Zabriqan ibn ther from various websites3 containing lists of
Salamah al-Namari Arabic names. Some of these names are Ro-
4. A laqab (pronounced LAH-kahb), a combina- manized (written using the Latin alphabet) and
tion of words into a byname or epithet, usually had to be transliterated from English to Arabic.
religious, relating to nature, a descriptive, or of This was done using the online translation
some admirable quality the person had (or software ‘Tarjim’ provided by Sakhr Software
would like to have); e.g., al-Rashid [the Company. Notice that the variations in Roman-
Rightly-guided], al-Fadl [the Prominent]. ized Arabic due to the lack of one to one corre-
Laqabs follow the ism: Harun al-Rashid spondence between Arabic letters and Roman
[Aaron the Rightly-guided]. letters have also been reflected in the translit-
5. A nisba (pronounced NISS-bah), a name de- eration, in reverse, from Romanized Arabic to
rived from a person's: trade or profession, Arabic Script.
place of residence or birth, religious affiliation, The raw data received had to be further proc-
among others; e.g. al-Hallaj [the dresser of essed to make it suitable for building gazetteers to
cotton], Al Msri [The Egyptian], Islami [Is-
lamic]. Nisbas follow the ism or, if the name 1
ACE reference: http://projects.ldc.upenn.edu/ace/
contains a nasab (of however many genera- 2
Treebank Corpus reference:
tions), generally follow the nasab. http://www.ircs.upenn.edu/arabic/
Both software are available to BUiD under license agreement.
3
Web sites include:
http://en.wikipedia.org/wiki/List_of_Arabic_names ,
http://www.islam4you.info/contents/names/fa.php, and
http://www.mybabynamessite.com/list.php?letter=a
be incorporated within the system. Some of the • A dictionary of job titles (19245 entries)
automated preprocessing performed on these data • A dictionary of honorifics used before
includes: names (173 entries)
ƒ Removing extra whitespaces between first • A dictionary of country names including
and last names, or beginning and end of variations in spellings (923 entries)
names for the efficient processing of the • A dictionary of nick names and laqabs
main gazetteer (dictionary) of complete (8169 entries)
person names. • A dictionary of person titles (20 entries)
ƒ Creating separate dictionaries (i.e. first, • a dictionary of words and phrases that act
last and middle names) without redun- as person indicators such as ‘ ‫اﻟﻤ ﺸﺮف‬
dancy because the full names had to be ‫( ’اﻟﺮﻳﺎﺿ ﻲ‬The sports supervisor) (421 en-
parsed. The extraction of each of these in- tries)
dividual components from full person
names was based on Python code and 5 System Architecture and Implementa-
common sense. tion
4.1 Typographic Variants Figure 1 shows the architecture of the PERA sys-
In order to be able to recognize variant Arabic tem. Our system has two major components: the
name entities, we added extra expressions in rules gazetteers and the Grammar. A filtration mecha-
and lexicon entries which lead to recognizing nism is employed that enables revision capabilities
named entities and their typographic variants. Ex- in the system.
amples of typographic variants include:
• The drop of hamza initially, medially, and Text

finally (e.g. ‫ اﺣﺴﺎن‬vs ‫ إﺣﺴﺎن‬- [Ehessan])

• Two dots inserted on aleph maqsura, and
two dots removed from yaa (e.g. ‫ ﻣﻮﺳ ﻲ‬vs Data Collection
Rule-based System
‫ﻣﻮﺳﻰ‬-[Mousa]) Whitelist
• Dropping the madda from the aleph (e.g. Dictionary
Acquisition from ACE
& Treebank corpus
‫ ال ﺧﻠﻴﻔﺔ‬vs ‫ آل ﺧﻠﻴﻔﺔ‬- [Al Khalifa]) (1) Gazetteer
• Hamza insertion below vs. above aleph Internet Resources

(e.g. ‫ أﺳﺮاء‬vs ‫إﺳﺮاء‬-[Essraa]) Names Databases

• Two dots inserted on final haa, and two Dictionaries
dots removed from taa marbouta (e.g. ‫ﻓﺎﻃﻤ ﻪ‬
(2) Grammar
vs ‫ﻓﺎﻃﻤﺔ‬-[Fatma]) Configuration
Arabic
script
• Diacritics: partial, full, or none. In the cur-
rent version we remove diacritics.
• Typing hamza followed by aleph maqsura
(3) Filter
separately vs. together (e.g. ‫ ه ﺎﻧﻰء‬vs ‫ه ﺎﻧﺊ‬-
[Hani]). Blacklist
Dictionary
4.2 Dictionaries
The following dictionaries (gazetteers) are derived
using the aforementioned data collection tech- Annotated
Text
niques. A total of 472617 entries were collected.
• A dictionary of full person names (263598 Figure 1: Architecture of the System
entries)
5.1 Gazetteers
• A dictionary of first names (78956 entries)
• A dictionary of middle names (67595 en- The main gazetteer (dictionary) of complete person
tries) names plays the role of a fixed static dictionary of
• A dictionary of last names (33517 entries) full person names. It recognizes person name enti-
ties by being applied as a Whitelist mechanism that Description:
accepts matches which are reported as a result of • The names should be verified against their
an intersection between the dictionary and the in- respective dictionaries (i.e. first, middle,
put text. A Whitelist is a list of strings that must be and last names).
recognized independent of the rules. It contains • The indicator pattern is composed of an
entries in the following format: honorific such as "‫[ "اﻟﻤﻠﻚ‬The king] fol-
‫|ﻋﺒﺪاﻟﺮﺣﻤﻦ ﻗﺎﺳﻢ اﻟﺸﻴﺮاوى‬Abdulrahman Qasim lowed by an optional Nisba derived from a
Mohammed Alshirawi location name such as "‫[ "اﻷردﻧﻲ‬Jordanian].
These act as trigger words to recognize the
Since the system being developed can be incor- person name and should be verified against
porated in various applications independent of lan- their respective dictionaries of honorific
guage constraints, the English transliterations of and locations.
the Arabic names are included in the dictionary as • The rule also matches an optional ordinal
meta data. number appearing at the end of some
names such as "‫[ "اﻟﺜﺎﻧﻲ‬II].
5.2 Grammar
• The Arabic suffix letters "‫ "ﻳﺔ‬and "‫ "ي‬used
The grammar performs recognition and extraction in the above pattern parses the inflections
of person entities from the input text based on attached to Nisba derived from locations
combinations of regular expression patterns. This that are commonly found in Arabic text.
rule definition is particularly challenging for the
Arabic language due to reasons such as: Implementation:
• Arabic writing systems do not exhibit dif-
ferences in orthographic case, such as ini- (($honorific$ws*($location
tial capitalized letters to indicate the pres- (\x{064A}|\x{0629})*$ws*)?)+
ence of a proper name. This lack of spe- $firsts_v(($ws*$middle_vv)|
($ws*$lasts_v))?$ws*($number)?)
cific internal structure in the Arabic lan-
guage poses great challenge for recogniz- Writing conventions:
ing person entities. • $: reference to a slave schema.
• Arabic is a highly inflected language • Firsts_v: dictionary of first names.
which entails a requirement of understand- • Middle_vv: dictionary of middle names.
ing of its morphological nature. The in-
• Lasts_v: dictionary of last names.
flected Arabic word maybe composed of
prefixes such as prepositions and suffixes • Ws: whitespace.
such as pronouns. These affixes need to be • Honorific: dictionary of honorifics ap-
addressed to ensure recognition of person pearing before names.
names alone. • Location: dictionary of locations.
Due to the above complexities in the Arabic • Number: Arabic ordinal numbers.
language a deep contextual analysis of various
Arabic scripts was performed using Python scripts Example:
to build grammar rules based on keywords or trig- The following name would be recognized by the
ger words forming a window around a person above rule:
name. ‫اﻟﻤﻠﻚ اﻷردﻧﻲ ﻋﺒﺪ اﷲ اﻟﺜﺎﻧﻲ‬
[The Jordanian king Abdullah II]
An Example Rule:
The following rule recognizes a person name com- Apart from contextual cues, the typical Arabic
posed of a first name followed by optional middle naming elements were used to formulate rules such
and last names based on a preceding person indica- as nasab, kunya, etc. Thereby the rules resulted in a
tor pattern. good control over critical instances by recognizing
((honorfic+ws(location(‫)ي|ﻳﺔ‬+ws)?) complex entities.
+firsts_v((ws+middle_vv)|
(ws+lasts_v))?ws+(number)?)
5.3 Filter Politics of Ukraine
In July 1994, Leonid Kuchma was elected as Ukraine's second president in
free and fair elections. Kuchma was reelected in November 1999 to
A filtration mechanism is used in the form of a another five-year term, with 56 percent of the vote. International observers
criticized aspects of the election, especially slanted media coverage;
Blacklist (rejecter) within the grammar configura- however, the outcome of the vote was not called into question. In March
2002, Ukraine held its most recent parliamentary elections, which were
tion to filter matches that appear after person titles characterized by the Organization for Security and Cooperation in Europe
(OSCE) as flawed, but an improvement over the 1998 elections. The pro-
but are invalid person names. In the following ex- presidential For a United Ukraine bloc won the largest number of seats,
followed by the reformist Our Ukraine bloc of former Prime Minister Viktor
ample: Yushchenko, and the Communist Party. There are 450 seats in parliament,
with half chosen from party lists by proportional vote and half from individ-
‘‫[ ’وزﻳﺮ اﻟﺨﺎرﺟﻴﺔ اﻟﻌﺮاﻗﻲ اﻻﻣﻴﻦ اﻟﻌﺎم‬The Iraqi Foreign ual constituencies

Minister the Secretary-General]

The sequence of words ‘‫[ ’وزﻳﺮ اﻟﺨﺎرﺟﻴﺔ اﻟﻌﺮاﻗﻲ‬The

Iraqi Foreign Minister] acts as a person indicator PIPELINE
and the word immediately following it is usually a
valid person name. However, in this example, the Person Person
Person
words following the person indicator that is, ‘ ‫اﻻﻣﻴﻦ‬
‫( ’اﻟﻌﺎم‬the Secretary-General) is not a valid person
name. Hence the role of the blacklist comes into
play by rejecting the incorrect matches recognized
by certain rules.
5.4 The Implementation Platform
The PERA system was implemented through in-
corporation into the FAST ESP framework, Figure 2: PERA incorporated into FAST ESP pipe-
(FAST,_). FAST ESP is an integrated software line to produce Tagged text
application that provides searching and filtering
services. It is a distributed system that enables in-
6.1 Reference Corpus
formation retrieval from any type of information,
combining real-time searching, advanced linguis- The text within the ACE and Treebank corpus was
tics, and a variety of content access options into a used for creating the entity tagged reference corpus
modular, scalable product suite. for evaluating PERA. The text was chosen ran-
The document processing stage within FAST domly from files with ‘sgm’ extension (containing
ESP system provides support for Entity Extraction. the Arabic script) within ACE & Treebank corpus.
PERA is implemented through the customizable The tagging was automatically performed with a
document processing pipelines within FAST ESP, Python script and further a post manual check was
which consists of multiple document processing performed to correct any invalid tags or identify
stages. A new search pipeline was created and the missing ones. The end product was an anno-
stages containing the grammar configuration and tated text corpus in the xml format with the UTF-8
gazetteers were added to this pipeline. Figure 2 encoding. This was divided into a 46 test sets and
indicates the functionality of the PERA system each evaluated individually with hurricane. The
incorporated in the pipeline within FAST ESP for total size of the reference corpus build is around
recognizing and tagging person entity in text. 4MB. The size and content of the corpus is such
that it contains a representative amount of occur-
6 The Experiment rences of the person entity.
In evaluating the PERA system we follow the 6.2 Evaluation Method
standard practice in the IE field of comparing sys-
We have adopted the evaluation measures that are
tem output against a reference corpus and measur-
standard in the IE community (De Sitter et al.,
ing the performance of the Arabic person named
2004), to evaluate and compare the results (preci-
entity.
sion, recall and F-measures):
correct entities recognized
Precision= total entities recognized
 Recall=
correct entities recognized gazetteers but being recognized by the grammar
total correct entities rules.
2 x recall x precision Table1 indicates the performance figures pro-
F-measure= recall + precision duced by 6 out of the 46 sets used for Hurricane
Precision indicates how many of the extracted evaluation.
entities are correct. Recall indicates how many of The average Precision and Recall for the total 46
the entities that should have been found, are effec- sets in recognizing person names is 85.5% and
tively extracted. Usually there is a trade off of re- 89%, respectively. And the average f-measure is
call against precision. Therefore, often an average 87.5%.
accuracy is reported in the form of the F-measure,
a harmonic mean which weights recall and preci- Test Set Precision Recall F-measure
sion equally. It was introduced to provide a single Treebank set 1 91.2 90.3 90.7
figure to compare different systems’ performances. Treebank set 2 94 96.3 95.1
Treebank set 3 84.2 84.7 84.4
The PERA system implemented within the FAST
ACE set 1 89.6 96.8 93.1
ESP pipeline was evaluated using an Information
ACE set 2 88.4 94.2 91.2
Extraction testing tool called ‘hurricane’ that ap-
ACE set 3 86.7 89 87.8
plies these standard measures.
Table 1: Evaluation result for 6 test sets.
6.3 Results
The missing accuracy can be overcome in the
Figure 3 is a snapshot of the evaluation performed following ways:
by hurricane in terms of the above mentioned • Expanding the dictionary of person
measure. names further.
• More Arabic text/corpus can be analyzed
to identify strings that act as person indi-
cators.
• Reducing negative effects on evaluation
results (true positive being treated as
false positives) caused due to incomplete
annotation of the test corpus. The refer-
ence corpus can be further fine tuned to
tag all person entities completely.
• Enhancing quality of transliterated
names used.
• Using Arabic text with error free spell-
ing.
• Including all possible spelling variations
Figure 3: An Extraction from Hurricane Evaluation used for names in Arabic written text.

The extraction quality of the pipeline created for 7 Conclusion and Future Work
the person name extractor confirms to the initial The work done in this project is an attempt to
target set. The required degree of precision (80%) broaden the coverage for entity extraction by in-
and recall (70%), for the Person name extractor, corporating the Arabic language, thereby paving
has been achieved with the hurricane evaluation. the path towards enabling search solutions to the
Some of the entries within the gazetteers were ex- Arabian market.
tracted from the same corpus used also for creating Various data collection techniques were used for
the reference corpus for evaluation. However, the acquiring gazetteer name lists. The rule-based ap-
results achieved are accurate since they indicated proach employed with great linguistic expertise
recognition of person entities not included in the provided a successful implementation of the PERA
system. Rules are capable of recognizing inflected
forms by breaking them down into stems and af- Technical Report IR-278. Available at
fixes. A filtration mechanism is employed in the http://ciir.cs.umass.edu/pubfiles/ir-278.pdf
form of a rejecter within the grammar configura- John Maloney and Michael Niv. 1998. TAGARAB: A
tion that helps in deciding where a name ends and Fast, Accurate Arabic Name Recogniser Using High
Precision Morphological Analysis. In Proceedings of
the non-name context begins. We have evaluated
the Workshop on Computational Approaches to Se-
our system performance using a reference corpus mitic Languages. Montreal, Canada. August, pp. 8-
that is tagged in a semi-automated way. The aver- 15.
age Precision and Recall achieved for recognizing
person names was 85.5% and 89%, respectively. Bruno Pouliquen, Ralf Steinberger, Camelia Ignat, Irina
Temnikova, Anna Widiger, Wajdi Zaghouani, and
Suggestions for improving the system performance
Jan Zizka. 2005. Multilingual person name recogni-
were provided. tion and transliteration. Journal CORELA-Cognition,
This work is part of a new system for Arabic Représentation, Langage, Vol. 2, ISSN 1638-5748.
NER. It has several ongoing activities, all con- Available at http://edel.univ-poitiers.fr/corela/
cerned with extending our research to recognize
Doaa Samy, Antonio Moreno and Jose M. Guirao. 2005.
and categorize other entity Arabic named entities
A Proposal for an Arabic Named Entity Tagger Lev-
such as locations, organization. eraging a Parallel Corpus, International Conference
RANLP, Borovets, Bulgaria, pp. 459-465.
Acknowledgement
An De Sitter, Toon Calders, and Walter Daelemans.
This work is funded by the "Named Entity Rec- 2004. A Formal Framework for Evaluation of Infor-
ognition for Arabic" joint project between The mation Extraction, University of Antwerp, Dept. of
British Univ. in Duabi, Dubai, UAE and FAST Mathematics and Computer Science, Technical Re-
search & Transfer Inc., Oslo, Norway. We thank port, TR 2004-0. Available at
the FAST team. In particular, we would like to http://www.cnts.ua.ac.be/Publications/2004/DCD04
thank Dr. Petra Maier and Dr. Jürgen Oesterle for Antonio Toral. 2005. DRAMNERI: a free knowledge
their technical support. based tool to Named Entity Recognition. In Proceed-
Any opinions, findings and conclusions or rec- ings of the 1st Free Software Technologies Confer-
ommendations expressed in this material are the ence. A Coruña, Spain. pp. 27-32.
authors, and do not necessarily reflect those of the Imed Zitouni, Jeffrey Sorensen, Xiaoqiang Luo and
sponsor. Radu Florian, 2005 The Impact of Morphological
Stemming on Arabic Mention Detection and
References Coreference Resolution, In the Proceedings of the
ACL workshop on Computational Approaches to Se-
Saleem Abuleil 2004. Extracting Names from Arabic mitic Languages, 43rd Annual Meeting of the Asso-
Text for Question-Answering Systems, In Proceed- ciation of Computational Linguistics (ACL05). June,
ings of Coupling approaches, coupling media and Ann Arbor, Michigan, USA, pp. 63-70.
coupling languages for information retrieval (RIAO
2004), Avignon, France. pp. 638- 647.
Da'ud Ibn Auda. 2003. Period Arabic Names and Nam-
ing Practices, In Proceedings of the Known World
Heraldic Symposium (SCA: KWHS Proceedings,
2003), pp. 42-56, St. Louis, USA.
FAST ESP
http://www.fastsearch.com/thesolution.aspx?m=376
Nancy Chinchor 1998. Overview of MUC-7. In Pro-
ceedings of the Seventh Message Understanding
Conference (MUC-7). Available at:
http://www.itl.nist.gov/iaui/894.02/related_projects/
muc/
Leah S. Larkey, Nasreen Abdul Jaleel, Margaret Con-
nell. 2003. What's in a Name?: Proper Names in
Arabic Cross Language Information Retrieval CIIR