ArticlePDF Available

Abstract and Figures

Name identification has been worked on quite intensively for the past few years, and has been incorporated into several products revolving around natural language processing tasks. Many researchers have attacked the name identification problem in a variety of languages, but only a few limited research efforts have focused on named entity recognition for Arabic script. This is due to the lack of resources for Arabic named entities and the limited amount of progress made in Arabic natural language processing in general. In this article, we present the results of our attempt at the recognition and extraction of the 10 most important categories of named entities in Arabic script: the person name, location, company, date, time, price, measurement, phone number, ISBN, and file name. We developed the system Named Entity Recognition for Arabic (NERA) using a rule-based approach. The resources created are: a Whitelist representing a dictionary of names, and a grammar, in the form of regular expressions, which are responsible for recognizing the named entities. A filtration mechanism is used that serves two different purposes: (a) revision of the results from a named entity extractor by using metadata, in terms of a Blacklist or rejecter, about ill-formed named entities and (b) disambiguation of identical or overlapping textual matches returned by different name entity extractors to get the correct choice. In NERA, we addressed major challenges posed by NER in the Arabic language arising due to the complexity of the language, peculiarities in the Arabic orthographic system, nonstandardization of the written text, ambiguity, and lack of resources. NERA has been effectively evaluated using our own tagged corpus; it achieved satisfactory results in terms of precision, recall, and F-measure.
Content may be subject to copyright.
NERA: Named Entity Recognition for Arabic
Khaled Shaalan and Hafsa Raza
Faculty of Informatics, The British University in Dubai, P.O. Box 502216, Dubai, United Arab Emirates.
Name identification has been worked on quite inten-
sively for the past few years, and has been incorporated
into several products revolving around natural language
processing tasks. Many researchers have attacked the
name identification problem in a variety of languages,
but only a few limited research efforts have focused on
named entity recognition for Arabic script. This is due
to the lack of resources for Arabic named entities and
the limited amount of progress made in Arabic natu-
ral language processing in general. In this article, we
present the results of our attempt at the recognition
and extraction of the 10 most important categories of
named entities in Arabic script: the person name, loca-
tion, company, date, time, price, measurement, phone
number, ISBN, and file name. We developed the system
Named Entity Recognition for Arabic (NERA) using a rule-
based approach. The resources created are: a Whitelist
representing a dictionary of names, and a grammar, in the
form of regular expressions, which are responsible for
recognizing the named entities. A filtration mechanism
is used that serves two different purposes: (a) revision of
the results from a named entity extractor by using meta-
data, in terms of a Blacklist or rejecter, about ill-formed
named entities and (b) disambiguation of identical or
overlapping textual matches returned by different name
entity extractors to get the correct choice. In NERA, we
addressed major challenges posed by NER in the Arabic
language arising due to the complexity of the language,
peculiarities in the Arabic orthographic system, nonstan-
dardization of the written text, ambiguity, and lack of
resources. NERA has been effectively evaluated using
our own tagged corpus; it achieved satisfactory results
in terms of precision, recall, and F-measure.
ANamed Entity Recognition (NER) system is a signif-
icant tool in natural language processing (NLP) research
since it allows identification of proper nouns in open-domain
(i.e., unstructured) text. For the most part, such a system
is simply recognizing instances of linguistic patterns and
collating them. Larkey, Abdul Jaleel, and Connell (2003)
Received October 18, 2008; revised March 12, 2009; accepted March 12,
© 2009 ASIS&T Published online 22 April 2009 in Wiley InterScience
( DOI: 10.1002/asi.21090
conducted a study that showed the importance of the proper
names component in language tasks involving searching,
tracking, retrieving, or extracting information.Another study
by Crestan and de Loupy (2004) showed that named entity
extraction helps users to more quickly and efficiently browse
large document collections. This seems plausible because
according to Gey (2000), 30% of the content-bearing words in
news are proper names. Abuleil (2004) and Chinchor (1998)
stated that the valuable information in text is usually located
around proper names, so identifying proper names is an
important first step.
In the 1990s, the NER concept was introduced at the
Message Understanding Conferences (MUCs), which were
financed by the Defense Advanced Research Projects Agency
to encourage the development of new and better methods
of information extraction. At the sixth conference (MUC-6; the task of
named entity recognition was defined as three subtasks:
ENAMEX (for the person, location, and organization names),
TIMEX (for date and time expressions), and NUMEX (for
monetary amounts and percentages). Until now, NER sys-
tems developed in various languages have evolved around
these three subtasks; however, we have broadened the cov-
erage of the named entities with our system NERA, which
identifies 10 types of phenomena, including Person names,
locations, companies, dates, time, prices, measurements,
phone numbers, ISBNs, and file names.
The work presented in this article concentrates on the
role of NER in an information-extraction task that retrieves
relevant information from a large amount of diverse data.
We have adopted the rule-based approach using linguistic
grammar-based techniques to develop NERA. The approach
is motivated by the characteristics and peculiarities of the
Arabic language. The recognition process requires two cycles
(Shaalan & Raza 2007, 2008): (a) using the Whitelist compo-
nent for matching relatively simple linguistic items such as
person names and (b) applying the grammar rules involving
relatively complex linguistic structures such as NE indicators.
The set of grammar rules was derived by analyzing the local
lexical context of a large amount of diverse data. A comple-
mentary process, which uses metadata (Blacklist or rejecter)
about ill-formed NEs, is applied to filter recognition results
TABLE 1. Examples of inflections in Arabic text.
Arabic example English translation Entity type Affix (clitics)
and the United Arab Emirates Location (Waw)
to Pakistan Location (laam)
for the United States Location (baa, alif-laam)
by ABC Network for the American television Company (baa)
for the British Broadcasting Corporation “BBC” Company (laam)
the 2925 meter Measurement (alif-al)
for 3 years Measurement (laam)
for $20,266 Price (baa)
to discard incorrect matches. Sometimes identical or overlap-
ping textual matches are inevitable, resulting in ambiguous
NEs. In this case, a heuristic disambiguation technique is
applied to get the correct choice with respect to the context in
which an ambiguous situation arises. This open-architecture
approach provides flexibility and adaptability features in our
system so that it can be easily configured to work with differ-
ent languages, NLP applications, and domains. The NERA
system has been evaluated using a reference corpus that is
tagged with names in a semi-automated way. The system
performance results achieved were satisfactory when eval-
uated against the standard measures: precision, recall, and
The rest of this article is structured as follows. We first
highlight how NERA provides solutions to challenges posed
by the Arabic language and then present previous related
work in Arabic NER. Next, the data-collection methods
are described. The following section explains in detail our
approach to NER in terms of system architecture. Then, we
briefly present an idea about the implementation platform.
The subsequent section is dedicated to describing the refer-
ence corpora we built to carry out our experimental work. We
present the results of our experiments, and then draw some
conclusions and discuss future work.
Challenges Tackled by NERA
In NERA, we addressed major challenges posed by NER in
the Arabic language arising due to the complexity of the mor-
phological system, peculiarities in the Arabic orthographic
system, nonstandardization of the written text, ambiguity, and
lack of resources. The following subsections discuss these
issues and how we deal with them in NERA.
Complex Morphological System
Arabic has as a rich and complex morphological system
due to its highly inflected nature (Shaalan, 2005). Any given
Arabic lemma has usually more than one word form to repre-
sent it, which includes a root, its internal structure, prefixes,
suffixes, and clitics. Since we deal with real, publishedArabic
text that has not been preprocessed in any way, NEs appear
in their real context; one important issue, in this respect,
is that NEs as other nouns in Arabic may appear preceded
by clitics. These clitics may be a conjunction (Waw,
and), a preposition (Laam, for), (baa, with), or both
(Waw-Laam, and-for), and so on. The internal struc-
ture itself includes short vowels and vocalic length, which
together carry the bulk of the morphological and morphosyn-
tactic structures, and a consonantal skeleton, which bears the
weight of the lexical (semantic) structure. This concatena-
tive strategy to form words inArabic causes data sparseness;
hence, this peculiarity of the Arabic language poses a great
challenge to NER systems.
These inflected forms should not be recognized as a part
of the extracted NE. To handle this issue in NERA, we use a
heuristic method within the pattern-matching engine, which
takes into consideration affixes of words within the pattern
being processed. Consequently, within the handcrafted rules,
we had to expand the possibilities of matching by indicat-
ing that the string might be preceded by one or more preclitics
that should be stripped from the recognized NEs in the final
output. This method performs morphological analysis before
recognizing the NE. Morphological analysis is necessary to
look into the affixes and see whether a word is an NE. The
rules recognize the inflected NE forms by breaking them
down into stems and affixes. Since rules were written using
a real-data context, the accuracy achieved is quite authentic.
Table 1 shows some inflected NE examples that have been
dealt with in NERA’s grammar for the respective entity type.
Peculiarities in the Arabic Orthographic System
Arabic does not have capital letters; this characteristic rep-
resents a considerable obstacle for the NER task because in
other languages, capital letters represent a very important
feature in identifying proper nouns. Thus, the problem of
identifying proper names is particularly difficult for Arabic
because we cannot recognize them in the text by looking at
the first letter of the word.
DOI: 10.1002/asi
TABLE 2. Examples of variations in Arabic text.
Arabic example English translation Entity type
Indonesia Location
Los Angeles Location
Johannesburg Location
Guilder Price (currency)
Mobile no. 3546575 Phone no.
TABLE 3. Examples of typographic variations in Arabic text.
Arabic example English translation Entity type Typographic variation
Australia Location Drop of hamza initially, medially, or finally
Saudi Arabia Location Two dots removed from taa marbouta
Asia Location Drop of the letter madda from the aleph
American dollar Price (currency) Drop of hamza initially, medially, or finally
lira Price (currency) Two dots inserted on final haa
Swiss franc Price (currency) Two dots removed from yaa
4th Date (day) Hamza insertion below vs. above aleph
Hence, to tag proper names in Arabic text, we used key-
words or indicator words to guide us to the place where one
could find them in the text. By using keywords, we marked
name phrases that might contain a certain name, then we pro-
cessed these phrases to extract names. The method adopted
in NERA to analyze these phrases and extract the names was
the derivation of a set of heuristic rules and their applica-
tion to parse the phrases to extract the name entities. Some
examples of keywords used for identifying the names are:
Personal names (title): Mr. John Adams
Personal names (job title): President John Adams
Nonstandardization of the Arabic Written Text
Arabic text includes many translated and transliterated
NEs. Spelling of translated and transliterated proper names in
general tends to be inconsistent in Arabic text.Table 2 shows
some examples of the inconsistency, although some can be
considered as typographical errors.
The extractor can handle, to some extent, the aforemen-
tioned spelling variants. Such issues were dealt with within
the context-sensitive rules and dictionary-building rules for
the NERA system.
Additionally, the extractor is capable of recognizing vari-
ations in written Arabic text for the various named entities
being recognized. Table 3 contains some example NEs
indicating typographic variations.
The loss of the internal diacritics (e.g., short vowels or
shadda) leads to different types of ambiguity in Arabic texts
(both structural and lexical) because different diacritics rep-
resent different meanings. These ambiguities can be resolved
only by contextual information and an adequate knowledge
of the language. Apart from ambiguity due to missing diacrit-
ics, Arabic—like many other languages—faces the problem
of ambiguity between two or more named entities. The fol-
lowing example indicates an ambiguous situation in Arabic
(Ahmed Abad has a keen
interest in philosophy.)
In the previous example, the boldface text fragment,
(Ahmed Abad), represents both a person name
and a location, thereby giving rise to an ambiguous situa-
tion. These situations can be handled in NERA by specifying
a filter rule that gives preference on one extractor over the
other. Table 4 shows some of the ambiguous situations that
the system can handle.
Lack of Resources
We carried out research on the Arabic language NLP tools
and resources in general (e.g., corpora, gazetteers, POS tag-
gers, etc.). This led us to conclude that in comparison with
other languages, Arabic lacks mature linguistic resources,
DOI: 10.1002/asi
TABLE 4. Ambiguous examples.
Ambiguous example English translation Incorrect Correct
1.6985 Swiss francs Person Price
15th of Ramadan Al karim 2005 Person Date
Jassim united for real estate and general maintenance Person Company
1.5 billion Singapore dollars Location Price
Saudi Aramco Location Company
Racheal Victoria Queen Location Person
In the evening Elizabeth II Time Person
...a turning point in September 1954 Martin presented ... Measurement Date
especially free resources available for research purposes.
These resources are often limited in both capability and cov-
erage. Thus, efforts were required in building up resources:
evaluation corpora and Whitelist dictionaries of NEs and, as
a preprocessing task, to build NERA.
As mentioned earlier, the nonstandardization of written
Arabic text causes further bottlenecks; the lack of control
over written forms of Arabic script leads to the unstructured
nature of Arabic text, thereby making Arabic NLP research
far more challenging as compared to other languages.
Related Work
Name identification has been worked on quite intensively
for the past few years and has been incorporated into several
products. Many researchers have attacked this problem in a
variety of languages, but only a few limited research efforts
have focused on NER forArabic text. This is due to the lack of
resources for Arabic NE and the limited amount of progress
made in Arabic NLP in general. Next, we present some of the
successful systems that have been produced in this endeavor.
Maloney and Niv (1998) developed TAGARAB, an Ara-
bic name recognizer that uses a pattern-recognition engine
integrated with morphological analysis. The role of the mor-
phological analyzer is to decide where a name ends and the
nonname context begins. The decision depends on the part of
speech of the Arabic word and/or its inflections. For this test
set, 14 texts from the AI-Hayat CD-ROM were selected ran-
domly. In addition to manually tagging them, the authors also
ran TAGARAB over these 14 texts and used a standard MUC-
style scoring program to compare the morphological output
of TAGARAB with the “answers” in the hand-tagged version.
The evaluation corpus contains 3,214 tokens, of which 2,324
are Arabic words; 1,879 of the latter received morphological
features when hand-tagged. The performance achieved for
precision, recall, and F-measure for Person NE recognition
was 86.2, 76.2, and 80.9%, respectively; for Location NE:
94.5, 85.3, and 89.7%, respectively; for Number NE: 97.7,
97, and 97.3%, respectively; and for Time NE: 91, 80.7, and
85.4%, respectively.
Abuleil (2004) presented a technique to extract proper
names from text to build a database of names along with
their classification that can be used in question-answering
systems. This work was done in three main stages: (a) mark-
ing the phrases that might include names; (b) building up
graphs to represent the words in these phrases and the rela-
tionships between them; and (c) applying rules to generate
the names, classify each of them, and save them in a database.
The module has been tested on 500 articles from the Al-Raya
newspaper, published in Qatar. In total, it has identified 335
names, missed 92 names, and extracted 8 names mistakenly.
The NER accuracy was calculated in terms of precision by the
author: People (90.4%), Location (93%), and Organization
Samy, Moreno, and Guirao (2005) used parallel corpora in
Spanish and in Arabic, and an NE tagger in Spanish to tag the
names in the Arabic corpus. For each sentence pair aligned
together, they used a simple mapping scheme to transliterate
all the words in the Arabic sentence and return those match-
ing with NEs in the Spanish sentence as the NEs in Arabic.
The size of the subcorpus used for the experiment is not large
(1,200 sentence pairs), but due to its nature and its source, it
contains a considerable number of NEs. From the 1,200 pairs
of sentences, 300 sentences from the Spanish corpus were
selected randomly with their equivalentArabic sentences. For
each pair, the output of the NE tagger was compared to the
manually annotated gold-standard set. They have improved
the precision by applying a filter to the Arabic words, which
omitted the stop words from the possible transliterated candi-
dates. While they reported high precision (i.e., 84% improved
to 90%) and recall (97.5%), note that their approach is
applicable only when a parallel corpus is available.
Zitouni, Sorensen, Luo, and Florian (2005) adopted a
statistical approach for the entity detection and recognition
(EDR). In this work, a mention can be either named (e.g.,
John Mayor), nominal (e.g., the president), or pronominal
(e.g., she, it). An entity is the aggregate of all the mentions (of
any level) that refer to one conceptual entity. This extended
definition of the entity has convinced us of the suitability of
the approach. The system was trained and evaluated on the
DOI: 10.1002/asi
Arabic Automatic Content Extraction (ACE) 2003 and part
of the 2004 data. The test dataset consists of 178 documents
from three sources: 38 Arabic Treebank (ATB) documents,
76 broadcast (bnews) documents, and 64 newswire (nwire)
documents. The objective of the evaluation was to investigate
the usefulness of stem n-gram features in the mention detec-
tion system. The stemming n-gram features gave interesting
improvement in terms of precision (64.2 vs. 64.4%), recall
(55.3 vs. 55.7%), and F-measure (59.4 vs. 59.7%).
A very recent work by Benajiba and Rosso (2008) also
experimented with the statistical approach towards NER
(person, location, and organization) using probabilistic mod-
els; maximum entropy and then further conditional random
(CRF) fields. The authors used their own corpus, called
ANERcorp, to train and test the CRF model. ANERcorp
is composed of a training corpus and a test corpus anno-
tated especially for the NER task. The overall performance
combining all features in terms of precision, recall, and F-
measure was 86.9, 72.77, and 79.21%, respectively. The
results obtained an accuracy improvement by more than 10
points as compared to the entropy model. In a later work
(Benajiba, Diab, & Rosso, 2008), the authors reported that
ANERsys is subject to further comparative study between
many probabilistic models (e.g., SVM, HMM, Maximum
Entropy, CRF, etc.) and also experiments using a combination
of different models.
Data Collection
Various methods and techniques were used for acquiring
data for building up the Whitelist component. This includes:
Automatic collection of named entity instances and indicators
from annotated corpora. The ACE (http://projects.ldc.upenn.
edu/ace/) and the ATB (
are some great resources that facilitate corpus-based studies
of many interesting linguistic phenomena in Modern Standard
Arabic. These corpora were exploited for the data-collection
task. These corpora, which are tagged with many linguis-
tic details, were first analyzed and the commonly occurring
patterns studied. These identified patterns were then used to
extract useful data.
Acquisition of named entities from a database provided
by a government organization. The person and company-
name dictionaries also were built from names collected
from some organizations including immigration departments,
educational bodies, and brokerage companies.
Automatic acquisition of named entities from Internet
resources. Names were retrieved further from various Web
sites2containing lists of Arabic names, company names, and
locations. Some of these names are Romanized (written using
the Latin alphabet) and had to be transliterated from English
to Arabic.
Once NEs were compiled from the corpora processing,
Internet resources, and various organizations, they had to be
1Both software systems are available to BUiD under license agreement.
2Web sites included:
names,, and http://www.
FIG. 1. Architecture of the System.
further processed to ensure that the compiled data were clean.
The raw data received had to be further processed to make it
suitable for incorporation into the system.
Architecture of the NERA System
The NERA system requires two main processing
resources: a Whitelist (gazetteer) and a finite state transduc-
tion grammar.Afiltration mechanism also is employed that
enables revision capabilities in the system. Figure 1 shows the
abstract architecture of the NERA system. The system con-
verts the unstructured input Arabic text into structured form
by producing the annotations of the Arabic NE as a result of
the recognition task.
The recognition techniques employed include the follow-
ing two major steps: (a) a lookup procedure, called Whitelist,
that performs the recognition based on a gazetteer containing
lists of known named entities; and (b) a finite state transducer,
called Grammar Configuration, based on a set of grammar
rules derived by analyzing the local lexical context.
The Whitelist plays the role of fixed static dictionaries of
various NEs. It is a mechanism that accepts matches that are
reported as a result of an intersection between the dictionary
and the input text. A Whitelist is a list of strings that must be
recognized independent of the rules. It contains entries in the
|Abdulrahman Qasim Mohammed
DOI: 10.1002/asi
Since the system being developed can be incorporated in
various applications independent of language constraints, the
English transliterations of the Arabic names are included in
the dictionary as metadata.
The grammar performs recognition and extraction of Ara-
bic named entities from the input text based on derived rules.
It describes patterns to match NEs, annotations being cre-
ated as a result. Due to the peculiarities and complexity of
the Arabic language, grammar rules are a vital processing
resource for the recognition system. For instance, the lack
of capitalization for proper nouns can be largely dealt with
by using NE indicators to formulate recognition rules. These
NE indicators were obtained as a result of the deep contextual
analysis of various Arabic scripts that were performed dur-
ing the data-collection phase. The indicators are referred to as
trigger words within our system, forming a window around
a named entity, which helps in identifying an NE within text,
but does not get recognized itself. The following are examples
of indicators used within rules.
Person title: (Mrs.), (Mrs.)
Job title: (the doctor), (the sciences
Company indicator: (LLC)
Country postindicators: (the federal),
(the democracy)
City postindicators: (the finance capital)
Measurement: (mg), (km)
Price: (Egyptian pound),
(Emirate dirham)
Being an agglutinative language, Arabic has highly
inflected forms. Hence, the grammar rules build encoded
morphological information that describes the inflected struc-
ture of the candidate-word forms. This enables stripping off
of the prefixes and suffixes from the word stem, thereby
ensuring the recognition of the actual NE instance alone.
For example, the following pattern describes an optionally
defined (masculine or feminine) adjective derived from a
location (DAL): ()? + country name + ( ),
where country name is an entry in the country name dictio-
nary that lists all the country name base forms along with their
variants. This pattern can, for example, be used to indicate
a company (Al-Ahram Egyptian
newspaper), a person (The
Egyptian President Hosni Mubarak), a price
(Egyptian pound), and so on.
For each type of NE, several rules were built, and each one
was applied in a particular order to ensure that the most com-
prehensive recognition result was achieved. Some examples
of NERA’s grammar rules are provided next.
Example rule for Person name recognition
((honorfic +ws(location( )+
This rule recognizes a person name composed of a first
name followed by optional last name based on a preceding
person-indicator pattern, or the trigger words.
The following name entities would be recognized by the
previous rule:
[The King Abdullah]
[The Jordanian King Abdullah]
[The Jordanian King Abdul-
lah II]
[The Jordanian Queen Rania]
Apart from contextual cues, the typical Arabic naming ele-
ments were used to formulate rules such as nasab,kunya, and
so on (Shaalan & Raza, 2007). Thus, the rules resulted in a
good control over critical instances by recognizing complex
Example rule for Company name recognition
preceding_known_part) + ((unknown + (DAL|
company_following_indicator| ((prefix_
busines)?businessType + (DAL)?))
This rule recognizes a company name based on prefix
words such as (Company) or (newspa-
per) from the dictionary of company preceding indicators,
which form a part of the recognized company name. Addi-
tionally, business-type indicators such as
(for Internet services) or suffix words/phrases such as
(LLC) appearing at the end of the company names aid in the
recognition process.
The following named entities would be recognized by the
previous rule:
(for BB company for Inter-
net services)
(and Al-Ahram Egyptian news-
(Jofind trading company LLC)
Example rule for Date recognition
(Weekday)? + ws + (DayFigArabic-Indic|
DayFigArabic) + ws + (( + (DayFigEng|
DayFigArabic))| )? + ws + (MonthName(
[-/](MonthName))?) + ws + ([-/] ( )?
()? + ws + (yearFigArabic|
yearWord) (year_range)?)
In its primitive format, the previous rule matches dates
in the format “day month year” with an optional weekday
name at the beginning. The day of month in Arabic dates can
be represented in either Arabic-Indic or Arabic digits; both
variations are matched by this rule.
Moreover, Arabic dates can optionally be comprised of
two different month names representing different calendar
DOI: 10.1002/asi
schemes; one of them is usually from the Gregorian calendar.
The rule is capable of recognizing this peculiarity in Arabic
dates. Further, the year inArabic dates can be stated in either
words or figures (Arabic-Indic or Arabic numerals), or be
represented by a relative word such as (previous).
All these variations are dealt with well by the aforementioned
The following name entities would be recognized by the
previous rule:
(Saturday, 19th of
Kanoun, the 2nd of January 1999)
2002 6 (Saturday, 6th of January
of Year 2002)
24 (24th of last May)
2001 1999 (May from Year
1999 till the end of Year 2001)
19 18 (18th and 19th of January)
28 [28th of September (Aylol)]
Example rule for Location recognition
((( | Administrative division) + ws)?
+ city name +ws + direction)
This rule recognizes a city name (existing in the dictio-
nary of city names). The following name entity would be
recognized by the this rule:
(Agadir City south of ...)
Afiltration mechanism is used that serves two different
purposes: revision of the NE extractor results and disam-
biguation of matches returned by different NE extractors.
The Revision capability is based on a Blacklist (rejecter) dic-
tionary within the grammar configuration to filter matches,
returned by rules that appear before or after NE indicators or
trigger words but are invalid entities. These invalid entities
are derived by analyzing the local lexical context of named
entities during grammar rule formulation. This process is
illustrated by the following example:
(The Iraqi Foreign
Minister the Secretary-General)
The sequence of words (The
Iraqi Foreign Minister) acts as a person indicator, and the
word immediately following it is usually a valid person name.
In this example, however, the sequence of words following the
person indicator, [i.e., (the Secretary- General)],
is not a valid person name; it acts as an appositive. Hence, the
role of the Blacklist, another set of rules, comes into play by
rejecting the incorrect matches returned by certain grammar
Apart from the Blacklist component, certain heuristic fil-
ter rules are used for postprocessing the system’s extraction
results to disambiguate extracted named entities. These rules
make it possible to disambiguate matches returned by dif-
ferent NE extractors by heuristic prioritization rules. When
applying a set of single-slot extraction rules to the input text
(i.e., sets of rules that extract particular types of named enti-
ties one after the other), one cannot exclude the possibility
of identical or overlapping textual matches within the doc-
ument, among different rules for different named entities.
For instance, different sets of rules for extracting instances
of both the named entities Person and Location names may
overlap or exactly match in certain text fragments, resulting in
ambiguous named entities. Among these named entities, the
correct choice must be made. The filter rule is an intelligent
way of making the correct choice, with respect to the con-
text in which the ambiguous situation arises. The following
example indicates an ambiguous situation in Arabic script:
(Ahmed Abad has a keen
interest in philosophy)
In this example, the boldface text fragment
(Ahmed Abad) represents both a person name and a location
name. Hence, when NERA is applied here, both the Per-
son and the Location Extractors within NERA will return
matches as (Ahmed Abad), thereby giving rise
to an ambiguous situation. Sometimes, the required behav-
ior is to have exactly one result. In this case, the following
filter rule can be used to disambiguate the aforementioned
If a possible match M1 for a location entity reported by the
location extractor intersects with a match M2 of a person
entity that is also reported by the person extractor, then the
match as a location name will be discarded.
So, in case of an intersection, the match for person names is
preferred over location names. Thus, the filter rules defined
within the system play a significant role in handling such
situations and resolving ambiguity.
FAST ESP—NERA Implementation Platform
The NERA system was implemented and incorporated
into the FAST ESP framework (FAST, 2008). FAST ESP
is an integrated software environment for development and
deployment of searching and filtering services. It is a
distributed system that enables information retrieval from
any type of information, combining real-time searching,
advanced linguistics, and a variety of content-access options
into a modular, scalable product suite. FAST ESP supports a
set of rule-based tools that we used to deploy our system. It
also includes the “hurricane” evaluation tool, which we used
to perform our NERA evaluation using a reference corpus.
NERA is implemented within the entity-extraction com-
ponent of the Content Pipeline, in the Document Processing
Engine of FAST ESP. Figure 2 indicates the functionality of
DOI: 10.1002/asi
FIG. 2. NERA incorporated into the FAST ESP (2008) pipeline to
recognize named entities in text.
the NERA system incorporated in the pipeline within FAST
ESP for recognizing and tagging named entities in text.
Resources Built for Arabic NER Within NERA
To develop the Arabic NER, we had to build our own
corpora due to the unavailability of free Arabic corpora
for research purposes. Moreover, the commercially available
Arabic corpora are oriented towards the newswire domain,
which we found lacks equal coverage of the 10 named enti-
ties involved in our research. Further, we also have built the
Whitelist (gazetteer) component, which is a vital processing
resource for many NLP tasks. In this section, we present the
main characteristics of the resources developed for Arabic.
Corpora for Person, Location, Date, Time, Price, and
Measurement NEs
ACE (Version 5.3.3 2005.05.31) and ATB (Version 2.0,
LDC Catalog No. LDC2003T06) corpora are standard Arabic
resources built by LDC for Arabic NLP tasks. These cor-
pora mainly contain text taken from newswire documents
and broadcast news which was used to create the entity tagged
reference corpora for evaluating the followingextractors: Per-
son, Location, Date, Time, Price, and Measurement within
The tagset used by LDC within these corpora provides very
detailed and sophisticated annotation, with markup based
on Arabic linguistics associated with the Arabic language.
Using Python scripts and a pattern-matching algorithm, we
first acquired NEs from LDC’s original tagset. For instance,
for extracting Person NEs, the script was programmed to
match the “PER” tag within the ACE corpus and the “Prop-
Noun” tag within the ATB corpus. The acquired NEs were
then used to create our NE tagged reference corpora, with 10
different tagsets (e.g., <person>...</person> tags for
Person NE). The tagging was done in a semi-automated way
as follows:
The Person names and Locations contained within the Source
Arabic Text from ACE and ATB was automatically tagged
using Python scripts and the acquired NEs.
The same reference corpus was further hand-tagged again
to mark the Date,Time,Price, and Measurement NEs. The
manual tagging was done for two reasons:
The tagset in ACE and ATB uses a generic POS tag “Numeric”
for entities such as price, measurement, and percentages.
A common tag “TIMEX” is used by ACE and ATB to tag both
date and time entities in a combined way.
For efficiency, the reference corpus that we built was
divided into sets of test corpora, each being approximately
100 KB in size. The total number of test sets for these named
entities is 34, with 24 created from the ACE corpus and 10
created from the ATB corpus. The total size of the reference
corpus is around 4 MB, composed of 300,000 distinct words.
The size and content of the corpus are such that it contains
a representative number of occurrences of the following NE
types: The person name category includes 500+entities, the
location category includes 500+entities, the date category
includes 394 entities, the time category includes 110 entities,
the price category includes 400 entities, and the measurement
category includes 386 entities.
Corpus for Company-Named Entities
The ACE and ATB corpora do not include a representa-
tive number of entities for company names. Thus, we sought
another corpus, the Corpus of Contemporary Arabic (CCA3),
to create the reference corpus for evaluating the company
extractor. This choice was based on the fact that the text
within CCA gave a good, varied coverage of company names,
thereby ensuring a more reliable evaluation of the company
extractor. For building up the company test corpus, we cre-
ated two reference corpus sets (each 100 KB in size) from
randomly selected text from the CCA corpus. Both sets were
hand tagged to mark company names within them. A total of
226 company-name instances have been tagged.
Named Entity Corpus for Phone Numbers, ISBNs, and
File Names
Available corpus resources in Arabic are quite limited and
restricted to coverage of the most important NEs such as
Person, Location, and so on. Hence, various Arabic Web
sites (e.g., Real Estate, Newspaper, etc.) were analyzed to
collect text containing phone number, ISBN, and file-name
entities. The corpus built was hand-tagged with 191 phone
number entities, 100 entities for ISBNs, and 139 entities for
file names.
3CCA can be freely downloaded online from Latifa Al-Sulaiti’s web site, As indicated by the
developers, the Arabic text within this corpus was mainly acquired from
magazine and newspaper web sites.
DOI: 10.1002/asi
In summary, the reference corpora for evaluating the 10
types of named entities (person, location, company, date,
time, price, measurement, phone number, file name, and
ISBN) within NERA are divided in the following way:
34 corpus sets for person, location, date, time, price, and mea-
surement extractor evaluation (created from ACE and ATB
corpus text)
2 sets of corpora for company extractor evaluation (created
from CCA corpus text)
3 individual reference corpora each for phone number, ISBN,
and file name extractor evaluation(created from text at various
Arabic Web sites)
The corpora created are in the XML format with UTF-8
encoding, in accordance with the guidelines set forth at the
beginning of the project. Additionally, the size and content of
the corpora are such that they contain a representative number
of occurrences of all 10 entity types.
Whitelist/Dictionaries Built
NERA gathers three different manually built gazetteers:
Person gazetteer: This contains a list of 263,598 complete
names of people collected from various government organiza-
tions, existing Arabic corpora, and Internet resources. Further,
the names were split into dictionaries of first and last names,
omitting the repeated names; the final list contains 175,502
first names and 33,517 last names.
Location gazetteer: This consists of 4,900 names of conti-
nents, countries, cities, states, political regions, towns, and
villages found in the Arabic version of Wikipedia and other
Web sites.
Organization gazetteer: This consists of a list of 273,491
names of companies, including those in areas of media and
newspapers, construction, banks and insurance, airlines,
and telecommunications, among others.
The evaluation of the NERA extractors was performed
using our own reference corpora, which highlight the Arabic
resources built during this project.
As mentioned in the previous section, the Whitelist built
for Person, Location, and Company NE extractors contains
certain entries extracted from the ACE and ATB corpora.
The evaluation corpus, to some extent, was built using the
same Arabic corpora resources; however, since the corpora
were huge in size, the overlap between texts used forWhitelist
and evaluation corpora building was kept minimal. Addition-
ally, the positive recognition results achieved can be attributed
mainly to the grammar rules, as compared to the gazetteer,
since the pattern matching developed was able to deal with
issues peculiar to the Arabic language, including inflections,
typographic variation, and so on.
The Evaluation Method
The performance was measured by Precision, Recall,
and F-measures, which are the standard measures for NER
(De Sitter, Calders, & Daelemans, 2004):
Precision =correct entities recognized
total entities recognized
Recall =correct entities recognized
total correct entities
F-measure =2×Recall ×Precision
Recall +Precision
Another way to look at Precision and Recall is:
Precision =true positives
true positives +false positives
Recall =true positives
true positives +false negatives
Precision indicates how many of the extracted entities are
correct. Recall indicates how many of the entities that should
have been found are effectively extracted. Usually, there is
a trade-off of recall against precision. Therefore, an average
accuracy is often reported in the form of the F-measure,a
harmonic mean that equally weights recall and precision. It
was introduced to provide a single figure to compare different
systems’ performances.
Since the corpora were tagged in a semi-automated way,
certain named entities were left untagged. In the recognition
results, these NEs were recognized correctly by the system,
but since they were not tagged in the test corpora, the eval-
uation tool marked these entities as false positives when in
reality they were true positives. To overcome this issue, the
entities marked as false positives by this tool were identi-
fied and retagged (i.e., manually corrected) in the reference
corpora. This iterative tagging of the corpus ensured quality.
The NERA system implemented within the FAST ESP
pipeline was evaluated using an information-extraction-
testing tool called Hurricane that applies the aforementioned
standard measures. This tool can perform evaluation on a
corpus with a size limit to 100 KB. Hence, the 5MB of eval-
uation corpora built were divided into 46 sets of corpus files.
Each test set was then individually given as input to Hurri-
cane, and separate accuracy results were produced by each.
The average of the results was estimated to reach conclusions
about each NE’s recognition accuracy.
At the beginning of the project, we set minimum accep-
tance criteria based on previous experience of FAST gained
from various NER systems for languages other than Arabic.
Table 5 shows the comparison between the achieved accuracy
and these minimum acceptance criteria (Excellent =90–
100%, Good =80–89%, Fair =70–79%, Poor =<70%).
From this table, note that the precision achieved is almost
the same as planned whereas the recall achieved was higher
than that of the initial plan.
DOI: 10.1002/asi
TABLE 5. Comparison of accuracy achieved and acceptance criteria set.
Acceptance criteria Achieved accuracy
Entity type Precision Recall Precision Recall
Person name Good Fair Good Good
Organization Fair Fair Good Good
Locations Fair Fair Fair Good
Date Excellent Good Excellent Excellent
Time Excellent Good Excellent Excellent
Price Excellent Good Excellent Excellent
Measurements Excellent Good Excellent Excellent
Phone no. Excellent Good Excellent Good
ISBN Excellent Good Excellent Excellent
File name Excellent Good Excellent Excellent
TABLE 6. Accumulated accuracy of the 10 NEs.
No. Entity type Precision (%) Recall (%) F-measure (%)
1 Person 86.3 89.2 87.7
2 Location 77.4 96.8 85.9
3 Company 81.45 84.95 83.15
4 Date 91.2 92.3 91.6
5 Time 97.25 94.5 95.4
6 Price 100 99.45 98.6
7 Measurement 97.8 97.3 97.2
8 Phone no. 94.9 87.9 91.3
9 ISBN 94.8 95.8 95.3
10 File name 95.7 97.1 96.4
Table 6 summarizes the accumulative recognition accu-
racy, in terms of precision and recall, achieved by each of
the 10 extractors built within NERA against the reference
With respect to the results of the extractors handling the
person, location, and company types, some of the entries
within the Whitelist component built were extracted from
the same corpus also used for creating the reference corpora
for evaluation. However, the evaluation results achieved are
accurate since they indicated recognition of named entities
not included in the Whitelist but being recognized by the
grammar rules within the pattern-matching component. After
careful analysis of the evaluation results, we found that the
accuracy can be further improved in the following ways:
Expanding the Whitelist dictionary of Person, Location, and
Company Names further.
More Arabic text/corpora can be analyzed to identify strings
that act as named entity indicators.
Reducing negative effects on evaluation results (e.g., true pos-
itive being treated as false positives) because of incomplete
annotation of the test corpora. The reference corpora can be
further fine-tuned to tag each and every named entity instance.
Enhancing the quality of transliterated names used.
Using Arabic text with error-free spelling.
Including all possible spelling variations used for names in
Arabic written text in an automated way.
One important factor that has greatly influenced the results
achieved is the nonstandardization of written Arabic text.
The majority are unstructured and are loaded with incon-
sistencies due to the lack of control over written forms of
Arabic script. Standard practices in publishing written Arabic
resources can help achieve far better accuracy results.
Arabic is a relatively complex and difficult language to
analyze, not so much because of its difficult morphological
structure but mostly because of how that structure is impacted
and made more complex by the orthographic issues of its writ-
ten form coupled with the drawbacks of limited research done
for the Arabic language. This work is an attempt to broaden
the coverage for entity extraction incorporating the Arabic
language by overcoming the language-specific challenges
to a great extent, thereby paving the path towards enabling
search solutions for the Arabic market.
Various data-collection techniques were used for acquir-
ing dictionary name lists. The rule-based approach employed
with great linguistic expertise led to a successful implemen-
tation of the NERA system by overcoming the challenges
posed by Arabic language.A set of grammar rules was derived
by analyzing the local lexical context of a large amount of
diverse data. Rules are capable of recognizing inflected forms
by breaking them down into stems and affixes. A filtration
mechanism is employed in the form of a rejecter within
the grammar configuration that helps in deciding where a
name ends and the nonname context begins. Further, the
DOI: 10.1002/asi
intelligent use of filter rules helps in dealing with recogni-
tion ambiguity between named entities. We have evaluated
our system performance using our own corpora tagged in a
semi-automated way. Moreover, these corpora could be used
as a standard evaluation dataset for Arabic NER approaches.
The evaluation results thus far look very promising; NERA
achieved high average precision and recall for each named
entity type against the reference corpora. Suggestions for
improving the system performance based on analyzing the
results were provided.
This work was funded by the “Named Entity Recognition
for Arabic” joint project between The British University in
Duabi, Dubai, United Arab Emirates and FAST Search &
Transfer Inc., Oslo, Norway. FAST was recently acquired by
Microsoft. We thank the FAST team; in particular, Dr. Petra
Maier and Dr. Jürgen Oesterle for their technical support.
Any opinions, findings, and conclusions or recommenda-
tions expressed in this material are the authors, and do not
necessarily reflect those of the sponsor.
Abuleil, S. (2004). Extracting names fromArabic text for question-answering
systems. In Proceedings of the 7th International Conference on Coupling
Approaches, Coupling Media, and Coupling Languages for Information
Retrieval (pp. 638–647), University ofAvignon (Vaucluse), France.
Benajiba, Y., Diab, M., & Rosso, P. (2008). Arabic named entity recogni-
tion: An SVM-based approach. In Proceedings of 2008Arab International
Conference on Information Technology (ACIT) (pp. 16–18). Amman,
Jordan: Association of Arab Universities.
Benajiba, Y., & Rosso, P. (2008). Arabic named entity recognition using
conditional random fields. Proceedings of the Workshop on HLT & NLP
Within the Arabic World. Arabic Language and Local Languages Pro-
cessing: Status Updates and Prospects, 6th International Conference on
Language Resources and Evaluation (pp. 26–31). Marrakech, Morocco.
Chinchor, N. (1998). Overview of MUC-7. In Proceedings of the 7th
Message Understanding Conference (pp. 2–5). Retrieved January 26,
2009, from
Crestan, E., & de Loupy, C. (2004). Browsing help for a faster retrieval.
Proceedings of the 20th International Conference on Computational
Linguistics (pp. 576–582). New York: ACM Press.
De Sitter, A., Calders, T., & Daelemans,W. (2004). A formal framework for
evaluation of information extraction. University of Antwerp, Department
of Mathematics and Computer Science, Technical Report TR 2004–0.
Retrieved April 15, 2009, from
FAST. (2008). FAST ESP: The world’s most intelligent, secure, high-
performance search platform. FAST, A Microsoft Subsidiary in Oslo,
Norway. Retrieved January 26, 2009, from
Gey, F. (2000). Research to improve cross-language retrieval. Position paper
for CLEF. In C. Peters (Ed.), Proceedings of the cross-language informa-
tion retrieval and evaluation. Workshop of Cross-Language Evaluation
Forum, Lisbon, Portugal. Lecture Notes in Computer Science, 2069
(pp. 83–88). Berlin: Springer.
Larkey, L., Abdul Jaleel, N., & Connell, M. (2003). What’s in a name?
Proper names in Arabic cross language information retrieval. CIIR Tech-
nical Report No. IR-278. Retrieved January 26, 2009, from http://ciir.cs.
Maloney, J., & Niv, M. (1998). TAGARAB: A fast, accurate arabic name
recogniser using high precision morphological analysis. In Proceedings
of the Workshop on Computational Approaches to Semitic Languages
(pp. 8–15). Montreal.
Samy, D., Moreno,A., & Guirao, J. (2005). A proposal for anArabic named
entity tagger leveraging a parallel corpus. Proceedings of the Interna-
tional Conference on Recent Advances in Natural Language Processing
(pp. 459–465). Borovets, Bulgaria: Benjamins.
Shaalan, K. (2005). Arabic GramCheck: A grammar checker for Arabic.
Software Practice and Experience, 35(7), 643–665. Chichester, England:
Shaalan, K., & Raza, H. (2007). Person name entity recognition for Arabic.
Proceedings of the ACL 2007 Workshop on ComputationalApproaches to
Semitic Languages: Common Issues and Resources (pp. 17–24). Prague,
Czech Republic: Association for Computational Linguistics.
Shaalan, K., & Raza, H. (2008). Arabic named entity recognition from
diverse text types. In B. Nordström & A. Ranta (Eds.), Proceedings
of the 6th International Conference on Natural Language Processing,
Gothenburg, Sweden. Lecture Notes in Computer Science/Lecture Notes
in Artificial Intelligence (LNCS/LNAI): Advances in Natural Language
Proceedings (Vol. 5221, pp. 440–451). Berlin, Germany: Springer-Verlag.
Zitouni, I., Sorensen, J., Luo, X., & Florian, R. (2005). The impact
of morphological stemming on Arabic mention detection and corefer-
ence resolution. Proceedings of the ACL Workshop on Computational
Approaches to Semitic Languages, 43rd annual meeting of the Asso-
ciation of Computational Linguistics (pp. 63–70). Ann Arbor, MI:
Association for Computational Linguistics.
DOI: 10.1002/asi
Dictionaries of NE Indicators
The following dictionaries were derived using the aforementioned data-collection techniques. The following specifies
various indicator dictionaries along with their respective number of entries for eight types of named entities covered by
NERA. The other two types of named entities that do not have indicator dictionaries and rely only on grammar rules are ISBN
and file name.
TableA1. Dictionaries of NE indicators.
Person Company Location Date Time Price Measurement Phone no.
Job titles Business type Administrative Month names Time zones Currency Unit 1 (481) Phone indicators
(19,245) (1,410) division (23) (156) (13) name (37) (160)
Person Company following City Weekday (23) Time word (37) Power of 10 Unit 2 (14) Phone-related
titles (20) known part (114) preindicators (12) in words (13) words (32)
Honorifics Company following City Related Time units (39) Locations (39) Rejecter
(173) indicator (37) postindicators (10) words (11) units (1,579)
Country Company preceding Country pre- Days in word Tens in
names (923) known part (163) indicators (77) 1–31 (118) word (20)
LaqabsaCompany preceding Country Hundreds in fractions (13)
(8,169) indicator (4) postindicators (22) words (18)
Person Location names (4,909) Location Tens in
indicators Blacklist (167) words (20)
Business prefix (4) Direction1 (17) Units in
words (43)
Company rejecters Direction2 (8) Date rejecter
(4,980) words (546)
Company part Direction3 (4)
rejecters (4,997)
Normalization Direction4 (4)
base form (22)
Direction5 (4)
Location base
form (534)
inhabitant (44)
aAlaqab (pronounced LAH-kahb), a combination of words into a byname or epithet, usually religious, relating to nature, a descriptive, or of some
admirable quality the person had (or would like to have) [e.g., al-Rashid (the Rightly-guided), al-Fadl (the Prominent). Laqabs follow the ism:Harun
al-Rashid (Aaron the Rightly-guided)].
DOI: 10.1002/asi
... e whitelists were dictionaries of NEs that matched the target texts and that were not dependent on the rules. Whitelists contain complete names that are not found anywhere else, and dictionaries contain single names that can be found in different places [29,36]. Examples of gazetteers are shown in Table 6. ...
... Blacklist. (Reject Word) A filtration procedure was completed during the last stage of the NER in the NERA system to create a list of rejected words [36]. Incorrect words used to identify NE were found and filtered out. ...
Full-text available
Named entity recognition (NER) is fundamental in several natural language processing applications. It involves finding and categorizing text into predefined categories such as a person's name, location, and so on. One of the most famous approaches to identify named entity is the rule-based approach. is paper introduces a rule-based NER method that can be used to examine Classical Arabic documents. e proposed method relied on triggers words, patterns, gazetteers, rules, and blacklists generated by the linguistic information about entities named in Arabic. e method operates in three stages, operational stage, preprocessing stage, and processing the rule application stage. e proposed approach was evaluated, and the results indicate that this approach achieved a 90.2% rate of precision, an 89.3% level of recall, and an F-measure of 89.5%. is new approach was introduced to overcome the challenges related to coverage in rule-based NER systems, especially when dealing with Classical Arabic texts. It improved their performance and allowed for automated rule updates. e grammar rules, gazetteers, blacklist, patterns, and trigger words were all integrated into the rule-based system in this way.
... If the total number of NEs identified by a technique is N(T, A), then T and N(T, C) represent the number of NEs that were correctly identified by the technique, while N(T, F) represents the number of NEs that were incorrectly identified by the technique. Precision is formalized in Equation (12). ...
Full-text available
Named entity recognition (NER) is an important task in natural language processing, as it is widely featured as a key information extraction sub-task with numerous application areas. A plethora of attempts was made for NER detection in Western and Asian languages. However, little effort has been made to develop techniques for the Urdu language, which is a prominent South Asian language with hundreds of millions of speakers across the globe. NER in Urdu is considered a hard problem owing to several reasons, including the paucity of large, annotated datasets; an inaccurate tokenizer; and the absence of capitalization in the Urdu language. To this end, this study proposed a conditional-random-field-based technique with both language-dependent and language independent features, such as part-of-speech tags and context windows of words, respectively. As a second contribution, we developed an Urdu NER dataset (UNER-I) in which a large number of NE types were manually annotated. To evaluate the effectiveness of the proposed approach , as well as the usefulness of the dataset, experiments were performed using the dataset we developed and an existing dataset. The results of the experiments showed that our proposed technique outperformed the baseline technique for both datasets by improving the F1 scores by 1.5% to 3%. Furthermore, the results demonstrated that the enhanced dataset was useful for learning and prediction in a supervised learning approach.
... These rules make use of gazetteers and contextual lexical triggers to identify and classify the NE. The main advantage of the rule-based NER systems is that they are based on a combination of typographical features and linguistic knowledge [24]. Although various rule-based approaches were implemented in the early stages of NER, the necessity for handcrafted rules that are very time-consuming to create, the difficulty of handling ambiguity, along with the fact that the rules do not scale well to new contexts or previously unseen instances, has meant that it has not been a widely used approach in recent years. ...
Named entity recognition has been one of the most widely researched natural language processing technologies over the last two decades. For the South African languages, however, relatively little research and development work has been done. This changed with the release of the NCHLT named entity annotated resources, a collection of named entity annotated data and Conditional Random Field-based named entity recognisers for ten of the official languages. In this work we provide a detailed description and linguistic analysis of the named entity (NE) annotated data for the agglutinative isiXhosa language, by analysing the morphosyntactic features relevant to the three main types of NE, viz. person, location, and organisation. From the data we identify suffix and capitalisation features that may be good predictors of the different NE types. Based on these features, we describe the named entity recogniser and feature set developed as part of the NCHLT release. The recogniser has high precision, 0.9713 overall, but relatively low recall, 0.7409, especially for person names, 0.5963, resulting in an overall F-score of 0.8406. Although there are various avenues to improve the named entity recogniser, this is a significant release for a historically under-resourced language.
... Researchers have used various methods and techniques in an attempt to improve Arabic NER. Traditional approaches to the problem have relied on rule-based models (Zaghouani, 2012;Shaalan and Raza, 2009;Oudah and Shaalan, 2016;Abdallah et al., 2012). These approaches were followed by models that were based on machine learning (Darwish, 2013;Benajiba et al., 2010;Abdelali et al., 2016), as well as by hybrid approaches (Oudah and Shaalan, 2012;Shaalan and Oudah, 2014). ...
The main objective of this paper is to compare and evaluate the performances of three open Arabic NER tools: CAMeL, Hatmi, and Stanza. We collected a corpus consisting of 30 articles written in MSA and manually annotated all the entities of the person, organization, and location types at the article (document) level. Our results suggest a similarity between Stanza and Hatmi with the latter receiving the highest F1 score for the three entity types. However, CAMeL achieved the highest precision values for names of people and organizations. Following this, we implemented a "merge" method that combined the results from the three tools and a "vote" method that tagged named entities only when two of the three identified them as entities. Our results showed that merging achieved the highest overall F1 scores. Moreover, merging had the highest recall values while voting had the highest precision values for the three entity types. This indicates that merging is more suitable when recall is desired, while voting is optimal when precision is required. Finally, we collected a corpus of 21,635 articles related to COVID-19 and applied the merge and vote methods. Our analysis demonstrates the tradeoff between precision and recall for the two methods.
... Universal [8] 2020 A survey of named-entity recognition methods for food information extraction Arabic [9] 2009 NERA: Named entity recognition for Arabic Arabic [10] datasets is from 3 to 18. Enlarging the quantity and the diversity of the CNER dataset would greatly facilitate relevant research. In addition, the CNER datasets are also suffer from the labeling ambiguity problem, i.e. the definition of entity types may vary between datasets. ...
Full-text available
Named Entity Recognition(NER), one of the most fundamental problems in natural language processing, seeks to identify the boundaries and types of entities with specific meanings in natural language text. As an important international language, Chinese has uniqueness in many aspects, and Chinese NER (CNER) is receiving increasing attention. In this paper, we give a comprehensive survey of recent advances in CNER. We first introduce some preliminary knowledge, including the common datasets, tag schemes, evaluation metrics and difficulties of CNER. Then, we separately describe recent advances in traditional research and deep learning research of CNER, in which the CNER with deep learning is our focus. We summarize related works in a basic three-layer architecture, including character representation, context encoder, and context encoder and tag decoder. Meanwhile, the attention mechanism and adversarial-transfer learning methods based on this architecture are introduced. Finally, we present the future research trends and challenges of CNER.
... En témoignent les nombreuses campagnes d'évaluation (MUC-6 et MUC-7, CoNLL, ACE, ESTER-2 ) et les abondants travaux sur différentes langues (Nadeau et Sekine 2007, Sharnagat 2014, Shaalan 2014, Etaiwi, Awajan et Suleiman 2017. 35 Outils existants pour la REN en arabe 36 Comme pour les autres langues, les systèmes de REN de l'arabe ont été développés en utilisant principalement trois approches : approches linguistiques à base de règles manuellement créées (Mesfar 2008, Shaalan et Raza 2009, Ben Mesmia et al. 2017 ; approches statistiques ou à base d'apprentissage automatique avec des données pré-étiquetées (Benajiba et al. 2007, Pasha et al. 2014, Abdelali et al. 2016, Helwe et Elbassuoni 2017 ; approches hybrides combinant les deux approches précédentes (Oudah et Shaalan 2012, Alotaibi et Lee 2014. ...
Full-text available
We present in this paper an automated method to map out positive or negative semantic modalities associated with place names in Arabic travelogue literature. This research sits at the crossroads of Natural Language Processing, Literary Studies, and Digital Humanities. Our pipeline identifies place named entities, analyzes their semantic context (with regard to opinions, sentiments and emotions), and locates the place names on geographic maps. Our corpus includes six travel writings on Paris from some of the most influential Arab writers of the 19th and 20th centuries. We evaluate rule-based and machine-learning approaches for their efficacy in named entity recognition and semantic analysis. The results of our automated analysis confirm, to a great extent, the judgements and interpretations of traditional critical scholarship on these Arabic literary texts.
... Early studies on NER often used conditional random fields (CRF) [15], support vector machines (SVM) [16], hidden Markov models (HMM) [17] and rules-based approaches [18]. Recently, in NLP, some deep neural networks methods have attracted considerable attention because of their better performances and less feature engineering [19]. ...
Identifying adverse drug reaction (ADR) entities from texts is a crucial task for pharmacology, and it is the basis for the ADR relation extraction task. The publicly available resources on this task include PubMed abstracts, social media, and other resources. Among these resources, social media data can reflect the reactions of drug users after taking medicine in real-time and update quickly. However, a very small quantity of annotated social media data leads to less research on these data. Moreover, social media data have colloquialism and informal vocabulary expression problems, which pose a major challenge for ADR named entity recognition (NER). In this work, we present an adversarial transfer learning architecture for the ADR NER task. Our model improves the performance on Twitter data (target resource) by incorporating biomedical domain information from PubMed (source resource). Additionally, we set the scale parameter in the final loss function to address the problem of bias in model training caused by imbalanced amounts of data. Without adding any additional manually designed features, our approach achieves state-of-the-art performance with an F1 on Twitter ADR data of 68.58%.
... As a continuation of PERA work, NERA was introduced by (Shaalan and Raza, 2008) and Shaalan and Raza (2009). NERA is a rule-based system to recognize NEs of 10 types: person, location, organization, date, time, ISBN, price, measure, phone numbers and filenames. ...
The Named Entity Recognition (NER) is an integrated task in many NLP applications such as machine translation, Information extraction and question answering. Arabic is one of the authorised spoken languages in the united nation. Currently, there is much Arabic information on the internet, so, nowadays the need for tools which process this information becomes significant. In this study, we have examined the impact of the conditional random field and the structured support vector machine in the task of Arabic NER. The structured support vector machine is the first time to be applied in the Arabic name entity recognition. Our proposed system has three stages: Preprocessing, extracting features and building model. We have used simple features like the bag of words in the [-1,1] window, the bag of part of speech tag in the [-1,1] window to enable our system to detect the multi-words entities. Also, we have tried to enhance the Stanford part of speech tagger to enhance the tagger output tags, which enabled our system to differentiate between the name entities from the nonentities. In addition, we have employed the binary features of: Is a person, is a prename, is a pre-location, is a location and is an organization. Our system has been trained and tested on part of ANER Crop. The results have proved that the conditional random field-based Arabic NER system outperforms the structured support vector machine-based Arabic NER using the same features set.
Extraction of meaningful information from a huge amount of data available on the web is a quite challenging task. The challenges faced in information extraction can be overcome with the help of an efficient named entity recognition (NER) system. Named entities are the proper names that play an important role in searching important information of interest. In this study, an efficient deep learning-based NER technique has been proposed which recognizes the named entities belonging to the general domain from Hindi, Punjabi, and bilingual Hindi and Punjabi text. An important variant of recurrent neural network, namely bidirectional long short-term memory-based model using improved word embeddings has been developed. Improved word embeddings are the combination of character convolutional neural network embeddings and part of speech embeddings. The main findings of the study include the development of a NER system that can extract named entities not only from Hindi and Punjabi datasets individually but also from mixed Hindi and Punjabi text. Besides, improved word embeddings are the combination of character-level features and word-level features which we find as the novel work as per our knowledge. Improved word embeddings are found to be effective in achieving better results than the results obtained by earlier NER models with deep feature extraction tasks.
Full-text available
Arabic presents an interesting challenge to natural language processing, being a highly inflected and agglutinative language. In particular, this paper presents an in-depth investigation of the entity detection and recognition (EDR) task for Arabic. We start by highlighting why segmentation is a necessary prerequisite for EDR, continue by presenting a finite-state statistical segmenter, and then examine how the resulting segments can be better included into a mention detection system and an entity recognition system; both systems are statistical, build around the maximum entropy principle. Experiments on a clearly stated partition of the ACE 2004 data show that stem-based features can significantly improve the performance of the EDT system by 2 absolute F-measure points. The system presented here had a competitive performance in the ACE 2004 evaluation.
Full-text available
Name identification has been worked on quite intensively for the past few years, and has been incorporated into several products. Many researchers have attacked this problem in a variety of languages but only a few limited re-searches have focused on Named Entity Recognition (NER) for Arabic text due to the lack of resources for Arabic named entities and the limited amount of progress made in Arabic natural language processing in general. In this paper, we present the results of our attempt at the recognition and extraction of 10 most important named entities in Arabic script; the person name, location, com-pany, date, time, price, measurement, phone number, ISBN and file name. We developed the system, Name Entity Recognition for Arabic (NERA), using a rule-based approach. The system consists of a whitelist representing a diction-ary of names, and a grammar, in the form of regular expressions, which are re-sponsible for recognizing the named entities. NERA is evaluated using our own corpora that are tagged in a semi-automated way, and the performance results achieved were satisfactory in terms of precision, recall, and f-measure.
Full-text available
The Named Entity Recognition (NER) task consists in determining and classifying proper names within an open-domain text. This Natural Language Processing task proved to be harder for languages with a complex morphology such as the Arabic language. NER was also proved to help Natural Language Processing tasks such as Machine Translation, Information Retrieval and Question Answering to obtain a higher performance. In our previous works we have presented the first and the second version of ANERsys: an Arabic Named Entity Recognition system, whose performance we have succeeded to improve by more than 10 points, from the first to the second version, by adopting a different architecture and using additional information such as Part-Of-Speech tags and Base Phrase Chunks. In this paper, we present a further attempt to enhance the accuracy of ANERsys by changing the probabilistic model from Maximum Entropy to Conditional Random Fields which helped to improve the results significantly.
Full-text available
The term Named Entity (NE), first intro-duced in 1995 by the Message Under-standing Conference (MUC-6), is widely used in the field of Natural Language Processing and Information Retrieval. Since 1995, a lot of studies have ad-dressed NE recognition, tagging and clas-sification. These studies reflected its efficient role in IE systems (Sekine, 2004; Grishman and Sundheim, 1996; Hase-gawa et al., 2004) as well as its effective-ness when used as anchor points in alignment techniques (Melamed, 2001; Samy et al., 2004). In this paper, we cover three main aspects concerning Arabic NE recognition and tagging. First, we present an overview of the linguistic nature and the studies concerning NE in Arabic texts. Second, we highlight the methodology of developing tools leveraging parallel cor-pora and previously developed tools for other languages. Third, we present our proposal for an Arabic NE tagger; its dif-ferent modules, its coverage scope and the methodology used for its implementation. However, it could also be considered a method for aligning NE in parallel cor-pora. Finally, we evaluate the results against a gold standard. At the end, we discuss the final conclusions and future work.
Full-text available
The Named Entity Recognition (NER) task has been garnering significant attention as it has been shown to help improve the performance of many Natural Language Processing (NLP) applications. More recently, we are starting to see a surge in developing NER systems for languages other than English. With the relative abundance of resources for the Arabic language and a certain degree of maturation in the state of the art for processing Arabic, it is natural to see interest in developing NER systems for the language. In this paper, we investigate the impact of using different sets of features that are both language independent and language specific in a discriminative machine learning framework, namely, Support Vector Machines. We explore lexical, contextual and morphological features and nine data-sets of different genres and annotations. We systematically measure the impact of the different features in isolation and combined. We achieve the highest performance using a combination of all features. Combining all the features, our system yields an F1=82.71. Essentially combining language independent features with language specific ones yields the best performance on all the genres of text we investigate.
Full-text available
Arabic is a Semitic language that is rich in its morphology and syntax. The very numerous and complex grammar rules of the language may be confusing for the average user of a word processor. In this paper, we report our attempt at developing a grammar checker program for Modern Standard Arabic, called Arabic GramCheck. Arabic GramCheck can help the average user by checking his/her writing for certain common grammatical errors; it describes the problem for him/her and offers suggestions for improvement. The use of the Arabic grammatical checker can increase productivity and improve the quality of the text for anyone who writes Arabic. Arabic GramCheck has been successfully implemented using SICStus Prolog on an IBM PC. The current implementation covers a well-formed subset of Arabic and focuses on people trying to write in a formal style. Successful tests have been performed using a set of Arabic sentences. It is concluded that the approach is promising by observing the results as compared to the output of a commercially available Arabic grammar checker. Copyright © 2005 John Wiley & Sons, Ltd.
Conference Paper
Full-text available
Tagging and extracting proper names is an important key for improving the effectiveness of question- answering systems. The valuable information in the text usually is located around proper names, to collect this information it should be found first. By extracting proper names from the text we provide question- answering systems with both the proper name found in the text, some information about it and where it was found. The proper names in Arabic do not start with capital letter as in many other languages so special treatment is needed to find them in a text. Little research has been conducted in this area; most efforts have been based on a number of heuristic rules used to find names in the text. In this paper we present a new technique to extract names from text by building a database and graphs to represent the words that might form a name and the relationships between them. First we mark the phrases that might include names, second we build graphs to represent the words in these phrases and the relationships between them, third we apply rules to find the names.
Conference Paper
Proper names are problematic for cross language information retrieval. Standard bilingual dictionaries typically have poor coverage of proper names. On the other hand, IR tasks involving news corpora, like TDT and TREC cross language IR, have proper names at their core. In this study, we demonstrate the importance of proper names in one such task, the TREC 2002 (Arabic-English) cross language track, by showing that performance degrades a tremendous amount when the bilingual lexicons do not have proper names. We then examine several different sources of proper name translations from English to Arabic, both static and generative (transliteration) and explore their effectiveness in the context of the TREC 2002 cross language IR task. We support a conclusion that a combination of static translation resources plus transliteration provides a successful solution.