Conventional Orthography for Dialectal Arabic
Nizar Habash, Mona Diab, Owen Rambow
Center for Computational Learning Systems
New York, NY, USA
Dialectal Arabic (DA) refers to the day-to-day vernaculars spoken in the Arab world. DA lives side-by-side with the official
language, Modern Standard Arabic (MSA). DA differs from MSA on all levels of linguistic representation, from phonology and
morphology to lexicon and syntax. Unlike MSA, DA has no standard orthography since there are no Arabic dialect academies,
nor is there a large edited body of dialectal literature that follows the same spelling standard. In this paper, we present CODA, a
conventional orthography for dialectal Arabic; it is designed primarily for the purpose of developing computational models of Arabic dialects. We explain the design principles of CODA and provide a detailed description of its guidelines as applied to Egyptian Arabic.
Keywords: Arabic, Dialects, Orthography
Dialectal Arabic (DA) refers to the day to day vernaculars
spoken in the Arab world. DA lives side by side with Modern Standard Arabic (MSA). As spoken varieties of Arabic,
DAs differ from MSA on all levels of linguistic representation, from phonology and morphology to lexicon and syntax. Most differences are at the phonological, morphological and lexical levels. MSA is the language of education in
the Arab world, while DA is perceived as a lower form of
expression; this has implications on the way DA is used in
daily written venues. On the other hand, being the natively
spoken language, DAs have been the object of many efforts
to study their patterns and regularities (Erwin, 1963; Cowell, 1964; Abdel-Massih et al., 1979; Holes, 2004). Most
of such studies have been field work or theoretical in nature
with limited transcribed data.
In current statistical Natural Language Processing (NLP)
there is an inherent need for large-scale annotated resources. For DA, the absence of such resources creates a
pronounced bottleneck for processing and building robust
tools and applications. Applying NLP tools designed for
MSA directly to DA yields significantly low performance,
making it imperative to build resources and dedicated tools
for DA processing.
In recent years, DA has emerged as the language of informal communication online, in emails, blogs, discussion forums, SMS, etc. These genres pose significant challenges
to NLP in general for any language including English. The
challenge arises from the fact that the language is less controlled and more speech-like while many of the textually
oriented NLP techniques are designed for processing edited
text. The problem is compounded for Arabic precisely because of the use of DA in these genres. Unlike MSA, DAs
have no standard published orthographies since there are no
Arabic dialect academies nor is there a large body of edited
dialectal literature that follows the same spelling standard.
There is a wide range of conventions used by native speakers in naturally occurring text and by creators of various
DA computational resources (tools, transcript collections).
These conventions are often inconsistent, a problem for efforts in DA computational processing.
In this paper, we present CODA, a conventional orthography for dialectal Arabic that aims at filling this gap; it is
designed primarily for the purpose of developing computational models of Arabic dialects. The paper is organized as
follows. Section 2. discusses previous efforts. Section 3.
presents a sketch of MSA orthography. Section 4. outlines
relevant differences between MSA and DA. Section 5. highlights the goals and principles of CODA. Section 6. details
CODA decisions for one dialect, Egyptian Arabic (EGY).
The issue of standardization of DA orthography is politically loaded, since it is seen by many as an attack on
MSA hegemony and Arab nationalism. One extreme example is that of the Lebanese poet Said Akl, who proposed a Latin-based orthography for Lebanese (Arabic) in
the 1960s (Arkadiusz, 2006). On the other end of the spectrum, the Asaakir system, which is the only approach to
Arabic dialect orthography approved by the Arabic Language Academy of Egypt, utilizes additional diacritics to
add on top of standard Arabic words to produce their dialectal forms (‘Asaakir, 1950). This standard is not used outside
of very limited circles (Al-Tonsi and Al-Sawi, 1990). Various DA dictionaries utilize Arabic, Latin or mixed script orthographies (Badawi and Hinds, 1986). These resources often focus on lemmatized (uninflected) forms. Resources developed for DA automatic speech recognition are typically
phonological transcriptions that are not readily usable for
modeling written text (Kilany et al., 2002; Maamouri et al.,
2004). Our CODA guidelines are inspired by the Linguistic
Data Consortium (LDC) guidelines for transcribing Levantine (LEV) and Iraqi (IRQ) Arabic (Maamouri et al., 2004).
They differ from them in that, whereas the LDC guidelines
are for transcription, and thus focus more on phonological
variations in sub-dialects, CODA is intended for general
purpose writing in a way that abstracts from these variations when possible. CODA is intended and designed as a
common convention for all DAs, making choices that minimize differences among them. We extend the LDC guidelines to cover EGY in detail – for which we profited from
the work on CallHome Egyptian (Kilany et al., 2002).
In a previous publication (Diab et al., 2010), we presented a
different conventional orthography (CCO: COLABA Conventional Orthography). CCO differs from CODA in many
respects, the most important of which is that CCO is intended to capture specifics of dialectal phonology and morphology. This goal, however, is very hard to achieve as the
annotator/transcriber training process was long and tedious
and annotators had a very hard time learning what some
described as a “foreign” system of writing. Also, interannotator agreement was rather low, especially over short
vowels that are often ignored in Arabic orthography.
A Sketch of MSA Orthography
We present a general sketch of Arabic orthography starting
with a brief description of MSA phonology followed by a
presentation of Arabic script and MSA orthographic rules.
For more details, see Habash (2010).
Consonants and Vowels MSA’s phonological profile includes 28 consonants, three short vowels, three long vowels
and two diphthongs (/ay/ and /aw/). Some of the consonants
are emphatic versions of other consonantal phonemes. Em
phasis ( Õæ
j®JË@ Altafxiym)1 is a bass effect giving an acoustic impression of hollow resonance to the basic sounds
(Holes, 2004). MSA vowel phonemes are limited in number compared to English or French; however, there are
many allophones to each of them depending on the consonantal context, such as becoming emphatic near emphatic
consonants. Another interesting phenomenon, called Waqf,
allows for optionally dropping the word-final short vowels
marking syntactic case in utterance-final words.
Morphotactics There are numerous additional phonological variations that are limited to specific morphological
contexts, i.e., they are constrained morpho-phonemically as
opposed to phonologically. The most common example of
such phenomena is the assimilation of the Arabic definite
article proclitic + È@ Al+ to the first consonant in the noun
or adjective it modifies if this consonant is an alveolar, dental or inter-dental phoneme (except for /j/). This set of 14
consonants is called the Sun Letters. It includes among oth t, H θ, P z, and š. For example, the word ÒË@
Al+šams ‘the sun’ is pronounced /aššams/ not */alšams/.
The rest of the consonants are called the Moon Letters. A
less common example is the phoneme /t/ in verbal pattern
VIII (Ai1ta2a3)2 which becomes voiced (/d/) when adjacent to specific root consonants such as /z/: Aiztahar becomes Aizdahar ‘it flourished’.
Syllabic Structure and Stress Syllabically, MSA is
rather simple having mostly CV and CVC syllables and a
few CVCC syllables in some word final positions. Stress is
not phonemic in Arabic.
Arabic transliteration is presented in the Habash-SoudiBuckwalter scheme (Habash et al., 2007): (in alphabetical order)
P ¨ ¨ ¬ ¼ È Ð à è ð ø
@ H. H H h. h p X XP
Â b t θ j H x d ð r z s š S D T Ď ς γ f q k l m n h w y
and the additional symbols: ’ Z, Â @, Ǎ @, Ā @, ŵ ð'
, ŷ Zø', ~ è, ý ø.
The digits 1/2/3 refer to root radicals.
The Arabic script is a right-to-left alphabet. There are two
types of symbols in the Arabic script for writing words:
letters and diacritics. Arabic letters are written in cursive
style in both print and script (handwriting). Diacritics are
additional zero-width symbols that appear above or below
the letters. MSA uses 36 letters and nine diacritics. We
discuss the different types of letters and diacritics in more
detail below as part of the orthography of MSA. There are
a few additional letters that are not officially part of Arabic
script for MSA. Most commonly seen are H
p, h c ¬ v and
À g. These are borrowings from other languages typically
used to represent sounds not in MSA.
An orthography is a specification of how the sounds of a
language are mapped to/from a particular script. We present
an account of standard MSA orthography using the Arabic script. The correspondence between writing and pronunciation in MSA falls somewhere between that of languages such as Spanish and Finnish, which have an almost
one-to-one mapping between letters and sounds, and languages such as English and French, which exhibit a more
complex letter-to-sound mapping (El-Imam, 2004). Most
Arabic letters and diacritics have a one-to-one mapping to
MSA phonemes. However, there is a number of common
important exceptions (El-Imam, 2004; Habash et al., 2007;
Biadsy et al., 2009).
3.3.1. Basic Phonemic Map
Consonants All of the consonants except for the glottal
stop (aka, Hamza) have a unique mapping into an Arabic
Short Vowels The three short vowels /a/, /u/ and /i/ are
written using the three short-vowel diacritics, a, u, and
Long Vowels Long vowels are written as a combination
of a short vowel and a glide consonant. The long vowels /ū/,
/ı̄/ and /ā/ are written as ñ uw, ù iy and A aA, respectively.
The diphthongs /ay/ and /aw/ are written as ñ aw and ù ay.
No Vowels The Sukun . diacritic marks vowel absence.
It is typically used to mark syllable boundaries. In the
case of two identical consecutive consonants with no vowel
between them, the second repeated consonant is replaced
with the Shadda, the consonant doubling diacritic, e.g.,
H. b∼ (/bb/).
Vowels at the Beginning of Words Arabic diacritics can
only appear after a letter. As such, word-initial vowels are
preceded with an extra silent Alif ( @ A) called Hamzat-Wasl.
The following are some examples: the word /kattaba/ ‘he
. J» kat∼aba, the word /maktūb/ ‘letdictated’ is written as I
H. ñJºÓ mak.tuwb, and the word
. Jº [email protected] Ain.kataba.
/inkataba/ ‘it was written’ is written as I
ter/written’ is written as
3.3.2. Hamza Spelling
The consonant Hamza (glottal stop /’/) has multiple forms
in Arabic script: Z ’, @ Ā, @ Â, ð ŵ, @ Ǎ and ø ŷ. The different
forms are governed by a set of complex spelling rules that
reflect word position, vocalic context and neighboring letter
forms (Habash and Rambow, 2007). For example, consider
the different Hamza forms in the following word meaning
‘his glory’ when its case marker changes: èZAîE. bahA’ahu
. bahAŵuhu /bahā’uhu/ (nomi/bahā’ahu/ (accusative), èðAîE
native), and éKAîE. bahAŷihi /bahā’ihi/ (genitive).
Arabic orthography distinguishes between two types of
) is always proHamzas. The Real Hamza ( ©¢¯ èQÒë
nounced as a glottal stop regardless of whether it is at
the beginning or in the middle of a word. The Tempo ) – see above, is a
rary Hamza, or Hamzat-Wasl ( Éð èQÒë
word-initial glottal-stop vowel allophone that only appears
if the word is at the beginning of a sentence/utterance.
3.3.3. Clitic Spelling
A clitic is a morpheme that has the syntactic characteristics of a word, but shows evidence of being phonologically bound to another word (Loos et al., 2004). In this
respect, a clitic is distinctly different from an affix, which
is phonologically and syntactically part of the word. MSA
has a small number of such clitics which are written attached to the word. Proclitics (prefixing clitics) are typically single-letter particles, such as the conjunction + ð wa+
‘and’, the preposition + H
. bi+ ‘in/with’, the future particle + sa+ ‘will’ and the definite article + È@ Al+ ‘the’.
Enclitics (suffixing clitics) are generally object/possessive
pronouns, e.g. Ñë+ +hum ‘them/their’. Multiple clitics
can appear in a word. For example, the word AîEñJ
wa+sa+yaktubuwna+hA ‘and they will write it’ has two
proclitics and one enclitic. Clitics generally do not modify
the spelling of the word base they attach to, although there
are a few exceptions, which are presented below.
3.3.4. Morpho-phonemic Spelling
The Arabic script contains a small number of common morphophonemic spellings. These are cases that spell a morpheme with multiple allomorphs using a form that reflects
the phonology of the most common allomorph or that of
some combination of allomorphs.
Definite Article The Arabic definite article is always
spelled as + È@ Al+ even though it phonologically assimilates to the first consonant in the noun or adjective it attaches to (as discussed above). The Alif of the definite article remains written when additional proclitics are added to
the word except with the prepositional proclitic + È li+, e.g.,
AJºËA¿ ka+Al+kitAb /kalkitāb/ ‘like the book’
. AJºÊË li+l+kitAb /kilkitāb/ ‘for
Ta-Marbuta The Ta-Marbuta ( è ~) is typically a feminine ending. It can only appear at the end of a word. In
MSA, it is pronounced as /t/ unless it is not followed by a
vowel (as in Waqf), in which case it is silent. For example,
éJ . JºÖÏ @ Almaktaba~u ‘the library’ is pronounced /’almaktabatu/ (normal) or /’almaktaba/ (Waqf). When the morpheme it represents is in word-medial position, such as be ). For
fore an enclitic, it is written using the letter Ta ( H
example, Ñë+ éJ.JºÓ mktb~+hm ‘library+their’ is written as
ÑîDJ. ºJÓ mtkbthm ‘their-library’.
Alif-Maqsura The Alif-Maqsura ( ø ý) is a silent deriva-
tional marker marking a range of morphological information from feminine endings to underlying word roots. AlifMaqsura always follows a short vowel /a/ at the end of a
word. In word-medial positions, it may be written using
the letters Alif ( @) or a Ya ( ø). For example, Ñë+ ù® Ó
mstšfý+hm ‘hospital+their’ is written ÑëA® Ó mstšfAhm
‘their-hospital’; however, Ñë+ úÍ@ Ǎlý+hm ‘to+them’ is
Ë@ Ǎlyhm ‘to-them’.
Waw of Plurality A silent Alif appears in the morpheme
Ì '@ ð@ð wAw AljamAςa~) which indicates
@ð+ +uwA /ū/ ( é«AÒm
a masculine plural conjugation in verbs. For example, @ñJ.J»
katabuwA ‘they wrote’ is pronounced /katabū/. This Alif
is deleted if followed by an enclitic, e.g., AëñJ.J» katabuwhA
‘they wrote it’.
Nunation Nunation is a nominal indefiniteness morpheme in MSA. It has the form of a word-final /n/, which is
written using the nunation diacritics ã, ũ and ı̃. These
diacritics combine the short vowel (case marker) preceding
the nominal indefiniteness morpheme: they are pronounced
. AJ» kitAbũ
/an/, /un/ and /in/, respectively. For example, H
is pronounced /kitābun/. A silent Alif appears word finally
with some nunated nouns (before or after the diacritic), e.g.,
AK. AJ» kitAbAã or kitAbãA /kitāban/.
3.3.5. Exceptional Spelling
There are few cases of exceptional spelling that are outside
the rules presented above. Archaic spellings of some com
mon words, e.g., é<Ë@ All∼áh ‘Allah’ and @ Yë háðA ‘this’,
use a diacritic called the Dagger Alif ( éK
Qj.J mÌ '@ Ë [email protected] á),
which represents a long /a/ vowel (/ā/). Another common
odd spelling is that of the proper name ðQÔ« ςamrw /ςamr/
‘Amr’ where the final ð w is silent.
3.3.6. Notes on Consistency and Standardization
Diacritic Optionality Whereas letters are always written, diacritics are optional: written Arabic can be fully
diacritized, partially diacritized, or entirely undiacritized.
Over 98% of written Arabic words are diacritic free
(Habash, 2010). This is not so much a problem when mapping from phonology to script but it poses a challenge in
the other direction.
Suboptimal Spelling A few letters are not spelled consistently. Arabic writers often replace hamzated letters with
the un-hamzated form, e.g., @ Â ⇔ @ A, or through two-letter
spelling, e.g., ø ŷ ⇔ Zø ý’. And the word-final letters ø y
and ø ý are often used interchangeably (Buckwalter, 2007).
Regional Standards MSA orthography is largely standardized. However, a few variations remain across and
within different Arab countries. For example, there are two
common spellings for names of geographic entities ending with an /a/ vowel: /sūrya/ ‘Syria’ appears as AK
swryA and éK
Pñ swry~. Hamza spelling rules may have
some exceptions also. For example, the word for ‘offi masŵuwl (common in
cial/responsible’ appears as Èð ñÓ
masŷuwl (common in Egypt).
the Levant) and ÈñJÓ
Dialectal Arabic vs. MSA
We present below a listing of important differences between
DAs and MSA. For more information on Arabic dialects,
see (Harrell, 1962; Erwin, 1963; Cowell, 1964; AbdelMassih et al., 1979).
4.1. Phonological Variations
Arabic dialects vary phonologically from MSA and from
each other. Some of the common variations include the following (Holes, 2004; Habash, 2006; Biadsy et al., 2009;
• The MSA alveolar affricate h. /j/ is realized as /g/
in EGY, as /ž/ in LEV and as /y/ in Gulf Arabic
(GLF). For example, ÉJ
Ôg. ‘handsome’ is pronounced
/jamı̄l/ (MSA, IRQ), /gamı̄l/ (EGY), /žamı̄l/ (LEV)
and /yamı̄l/ (GLF). The EGY and LEV pronunciations
are used for MSA in those regions.
• The MSA consonant /q/ is realized as a glottal stop
/’/ in EGY and LEV and as /g/ in GLF and IRQ.
For example, K
Q£ ‘road’ appears as /Tarı̄q/ (MSA),
/Tarı̄’/ (EGY and LEV) and /Tarı̄g/ (GLF and IRQ).
These changes do not apply to modern and religious
borrowings from MSA. For instance, à @Q¯ ‘Qur’an’ is
never pronounced anything but /qur’ān/.
/θ/ is pronounced as /t/ in
• The MSA consonant H
LEV and EGY (or /s/ in more recent borrowings from
MSA), e.g., éKCK ‘three’ is pronounced /θalāθa/ in
MSA versus /talāta/ in EGY.
• The MSA consonant X /ð/ is pronounced as /d/ in
LEV and EGY (or /z/ in more recent borrowings from
MSA), e.g., MSA I
. K X ðanb ‘fault’ and H. Y» kiðb
‘lies’ are pronounced /zanb/ and /kidb/, respectively.
• The MSA consonants /D/ (emphatic d) and /Ď/
(emphatic /ð/) are both normalized to /D/ in EGY and
LEV and to /Ď/ in GLF and IRQ. In modern borrowings from MSA, /Ď/ is pronounced /Z/ (emphatic z) in
EGY and LEV. For instance, ¡. A£ ‘police officer’ is
/ĎābiT/ in MSA but /ZābiT/ in EGY and LEV.
• Change in or complete drop of short vowels, e.g.,
‘he writes’ is pronounced /yaktubu/ MSA versus /yiktib/ (EGY and IRQ) or /yuktub/ (LEV). MSA
diphthongs /aw/ and /ay/ have mostly become /ō/ and
• Predictable shortening of long vowels under certain
conditions such as word-final position, loss of stress
or syllabic constraints. For example, compare the following forms of the same verb (stress vowel is bolded:
šAf /šāf/ ‘he saw’, šAf+hA /šafha/ ‘he saw her’, and
mA+šAf+hA+š /mašafhāš/ ‘he did not see her’.
4.2. Morphological Variations
There are a lot of differences between MSA and DAs morphologically. Some of these differences are a result of a
simplification of complex MSA paradigms. Others are the
opposite: more complex structures arising in the dialects
with no correlates in MSA. Some examples of the simplifying direction are the disappearance of the nominal case
marking system altogether in DAs. This is an important
change that has syntactic consequences. Similarly, verbal
mood and voice have disappeared. It is interesting to note
that the form of the indicative mood still survives as the default form in some dialects, whereas the subjunctive/jussive
mood form is used in others. Other simplification phenomena include the loss of the dual form in verb conjugation in
the dialects and the consolidation of feminine and masculine in the plural form.
In the rest of this section, we present some of important
specific examples of morphological differences. A verbal progressive particle, which has no correspondence in
MSA, appears as + H
. bi+ in EGY and LEV, as + X da+ in
IRQ and + ¼ ka+ in Moroccan Arabic (MOR). The MSA
future proclitic + sa+ is replaced by + h Ha+ in EGY
and LEV (appearing also as + ë ha+ occasionally in EGY)
and + ¨ γa in MOR. LEV, IRQ and GLF have a demonstrative proclitic + ë ha+ which strictly precedes with the
definite article + È@ Al+. Several dialects include the proclitic + ¨ ςa+, a reduced form of the preposition úÎ« ςalaý
‘on/upon/about/to’. Also, several dialect include the nonMSA negation circum-clitic + + AÓ mA+ +š.
There are also specific patterns that appear in some dialects
but not in MSA, e.g., it1a2a3 as in I
. Jº[email protected] Aitkatab ‘it was
written’. The form of some pronominal clitics and affix
has also changed. For example MSA Õç'+/ Õ»+ +tum/+kum
‘you [nominative]/[accusative]’ becomes EGY ñ»+/ @ñK+
+tuwA/+kuw. Some sub-paradigm changes also occur, e.g.,
mad∼a/madadtu ‘he/I extended’ becomes
MSA YÓ/ HXYÓ
YÓ mad∼/mad∼ayt in EGY and LEV.
Lexically, the number of differences is quite significant.
The following are a few examples: EGY . bas ‘only’,
èQ K Q£ tarabayza~ ‘table’, [email protected]
QÓ mirAt ‘wife [of]’ and
ÈðX dawl ‘these’, correspond to MSA ¡® ¯ faqaT, éËðA£
TAwila~, ék. ð P zawja~ and ZBñë
For comparison, the LEV forms of the above words are
bas (like EGY), TAwli~ (closer to MSA), mart and hadawl.
Given the lack of an orthographic standard, there is a lot of
orthographic variation in DA. DA writers are often inconsistent even with themselves. The differences in phonology
between MSA and EGY are often responsible: words can
be spelled phonologically or etymologically (using their re
lated MSA form), e.g., H
. Y» kidb or H. Y» kiðb. Furthermore, some cases of regular phonological assimilation are
written to reflect their phonology or underlying morphol
ogy, EGY I
is also written as I
. Jk. janb ‘side’
. Ôg. jamb;
while the plural form H
. AJk. jinAb does not have a similar
alternate form. Some clitics have multiple common forms,
e.g., the future particle h Ha appears as a separate word
or as a proclitic k/ë Ha+/ha+, reflecting different pronunciations. The different spellings may add some confusion,
e.g., ñJ.J» ktbw may be @ñJ.J» katabuwA ‘they wrote’ or éJ.J»
katabuh ‘he wrote it’. Finally, shortened long vowels can
be spelled long or short, e.g., Aê¯A/ Aê® šAf+hA/šf+hA ‘he
CODA Goals and Principles
In this section, we outline CODA goals and principles and
discuss some relevant practical considerations for the creation of a CODA annotated corpus.
PñK. A£ TAbuwr ‘line/queue’ (pronounced /Tabūr/) is
written using the MSA pattern /1A2uw3/ and not as
® KQK. burtuqAn ‘oranges’ (pronounced /burtu’ān/) is
the EGY word for ÈA®KQK. burtuqAl in MSA. CODA
CODA Design Principles
An Ad Hoc Convention CODA is an ad hoc convention. There are numerous decisions that could have been
made differently especially when it comes to the phonology/orthography interface. These principles make CODA
comparable to English spelling (a bit phonological, a bit
historical, with some exceptions). In some cases, we followed decisions that have been made by previously published efforts.
Arabic Script CODA uses only the inventory of Arabic
script characters including the diacritics used for writing
MSA. CODA does not use extended Arabic characters, e.g.,
from Persian or Urdu. Just like MSA, CODA can be written
undiacritized or diacritized.
Consistent Each DA word has a unique orthographic
form in CODA that represents its phonology and morphology.
MSA-like As a general rule, CODA uses MSA-like orthographic decisions (rules, exceptions and ad hoc choices),
e.g., cliticizing single letter particles, using Shadda for
phonological gemination, using Ta-Marbuta and AlifMaqsura, and spelling the definite article morphemically.
Generally Phonemic CODA generally preserves the
phonological form of dialectal words given the unique
phonological rules of each dialect (e.g., vowel shortening),
and the limitations of Arabic script (e.g., using a diacritic
and a glide consonant to write a long vowel). Two examples of important ad hoc exceptions pertain to specific root
radical letters that happen to be highly variant across di
alects, e.g., q, and to long pattern vowels that can be
shortened deterministically in the dialects, e.g., the pattern
«@ñ¯ 1awA2iy3. For these cases, the word is written using
the MSA cognate root radicals or pattern. The following are
idiosyncratic examples from EGY:
H. AJ» kitAb ‘book’ is written the same way it is in MSA
since the word does not vary in phonology or morphology.
Ég. @P rAjil ‘man’ is not written using the MSA variant
Ég. P rajul.
I.Jº[email protected] Aitkatab ‘was written’ is not written using the
MSA form I
. J» kutib.
qaSr ‘palace’ is written using MSA root radicals even through it is pronounced /’aSr/ (and can be
spelled more phonologically as Qå @).
We identify five goals for CODA: (i) CODA is an internally
consistent and coherent convention for writing DA; (ii)
CODA is created for computational purposes; (iii) CODA
uses the Arabic script; (iv) CODA is intended as a unified
framework for writing all DAs; and finally, (v) CODA aims
to strike an optimal balance between maintaining a level of
dialectal uniqueness and establishing conventions based on
spells this word using the MSA root radical for the
q/’ but not for n/l.
miš ‘not’ is uniquely dialectal and is not replaced
by one of its MSA equivalents: AÓ/ ÕË/ áË
Morphologically Faithful CODA preserves dialectal
morphology (e.g., dialectal clitics Èñ®Jk Ha+tiquwl ‘she
will say’ instead of the MSA variant Èñ®J sa+taquwlu).
The only exception here is separating the negation and indirect object pronouns although they are part of the word’s
¯ AÓ mA qult liphonological utterance, e.g., EGY AêË
hAš /ma’ultilhAš/ ‘I did not tell her’.
Syntactically Faithful CODA preserves dialectal syntax,
i.e., there is no change in word order.
Easily Learnable CODA is easy to learn and write. The
more CODA looks like what a dialect speaker may write,
Pan-Arabic but Specific Although most of the principles
of CODA are the same for all DAs, each dialect will have
its unique CODA Map (a list of rules and exceptions) where
the relevant phonology and morphology of the dialect are
outlined with the full diacritized inventory together with a
list of idiosyncratic exception cases.
Easily Readable CODA is not a purely phonological representation; however, text in CODA can be read perfectly
in DA given the specific dialect and its CODA Map.
5.3. Practical Considerations
The following are some consideration that arise when
working on creating annotated text with CODA, where the
raw text is converted manually to a CODA form to create
data that can be used to train automatic CODAfication.
Code Switching Since MSA and DA coexist, we often
find a lot of code switching between the two. CODA for
MSA text is the accepted MSA Arabic spelling. In the example in Figure 1, a joke is set up in MSA but the punchline
is in DA. DA words that are not CODA-compliant but happen to mimic MSA spelling of a cognate word in context
should be changed to a CODA-compliant form. For ex
ample, in the following EGY sentence Ð Qm× èX Ég. QË@ Alrjl dh mHtrm ‘this man is respectable’, the MSA-spelled
word Ég. QË@ Alrjl should be changed to Ég. @QË@ AlrAjl. It
may not always be easy to distinguish between the two
cases. For a discussion of the issues of dialect identification, see (Habash et al., 2008; Zaidan and Callison-Burch,
2011; Diab and Elfardy, 2012).
CODA Diacritization We expect CODA to be rendered
diacritized for morphological representations and be rendered undiacritized for large-scale creation of orthographically normalized training data (annotation).
Consistency CODA idiosyncratic decisions must be followed strictly. There is no room for improvisation by annotators and tool creators. New cases that are not handled can
be identified and added to the CODA exception lists and
CODA Map as needed.
root. Only mapping into MSA q, ð, θ and the emphatic
consonants S, T, D and Ď are allowed. These cases are
chosen because they are often variable across DAs. The
following are some illustrative examples:
Typographical Errors Typos such as split or merged
words (e.g., H
yArb instead of H. P AK
yA rb ‘oh Lord’),
misspelled words where some letters are missing, added or
transposed (e.g., Q.J
» kybr vs. Q
J.» kbyr ‘big’), should be
corrected as part of the annotation process. The directive is
to render them in a CODA-compliant orthography.
Other Issues The data to annotate may have other types
of issues due to the nature of the noisy input stream such as
URLs, html markup, speech effects (such as Q
‘very’), internet language, emoticons. These phenomena,
though they do touch on CODA, are considered outside
the scope of CODA definition. That said, these phenomena need to be handled as part of an initial preprocessing
round following guidelines specific to the general task.
CODA Guidelines for Egyptian Arabic
In this section, we present a summary of specific CODA
guidelines for EGY as an example of CODA guidelines.
For the full guidelines, see (Habash et al., 2012). An example of EGY in CODA is presented in Figure 1. In the rest
of this section, we consider Cairene the default EGY. Generally, EGY follows the same orthographic rules as MSA
(Section 3.3.) with the following exceptions and extensions.
6.1. Phonological Exceptions
The Egyptian "Geem" The phoneme symbol /j/ and the
corresponding letter h. j are used to represent the voiced
velar stop [g] in both MSA and dialect in Egypt.
Long Vowels The long vowels /ē/ and /ō/ which do not
exist in MSA are spelled as ù ay and ñ aw. There are
a few cases that will be ambiguous, but are all a result of
Ëð X dawlat is pronounced /dawlat/
MSA influences, e.g., I
‘Dawlat’ (a proper name from standard Arabic through
Turkish) or /dōlat/ ‘these’.
Vowel Shortening EGY long vowels shorten under certain conditions such as being word final, losing stress or
being followed by two consonants. Only one long vowel
is maximally allowed per word. Adding affixes and clitics changes stress patterns and interacts with vowel length.
Long vowel phonemes have short allophones, but short
vowel phonemes do not have long allophones. Vowel allophones involving shortening or emphasis are written phonemically, i.e., phonetically shortened long vowels are still
¯A AÓ mA šAf+hA+š /ma šafhāš/ ‘he
written long, e.g., Aê
did not see her’.
6.2. Phono-Lexical Exceptions
Etymologically Spelled Consonants A limited number
of consonants may be spelled differently from their
phonology if the following two conditions are met: (1) the
consonant must be an EGY root radical and (2) the EGY
root must have a cognate MSA root. If the conditions are
met, then we spell the consonant using the corresponding
radical from the cognate MSA root of the dialectal word’s
/t/ or /s/
/d/ or /z/
/Z/ or /d/ or /z/
/D/ or /z/
I.Ê¯ qalb /’alb/ ‘heart’
J» kiθiyr /kitı̄r/ ‘lots’
È X ðul∼ /zull/ ‘oppression’
¡. A DAbiT /ZābiT/ ‘officer’
£ Ďalma~ /Dalma/ ‘darkness’
©K A SAyig /sāyig/ ‘jeweler’
All other phonological differences from MSA are written
phonologically even though there are cases where there are
shared cognates, e.g., ½m» kaHk ‘cookies’ (not using the
MSA form ½ª» kaςk). In some cases, some consonants
are spelled etymologically but others are not, e.g., àA
burtuqAn ‘oranges’ (not using the MSA ÈA®KQK. burtuqAl).
MSA Pattern Vowels and Consonants A number of patterns in MSA have multiple long vowels which are not allowed in EGY. However since EGY phonology shortens
some of these vowels regularly, we write the word with
KA ¯ qAnuwn ‘law’ (pronounced
the MSA pattern, e.g., àñ
/qanūn/ not like MSA /qānūn/). The same principle applies
to pattern consonants (except for pattern Ai1ta2a3 as in
MSA), e.g., Q®K naftariD ‘we suppose’ has the inflected
pattern na+1ta2i3 and is pronounced /nafTariD/ (with the t
Hamza Spelling The Hamza spelling rules for EGY are
the same as MSA (Section 3.3.2.), when the Hamza is pronounced in EGY. However, EGY words that have hamzated
MSA cognates but no Hamza in EGY are written as pro
nounced in EGY, e.g., @P rAs ‘head’ (not like MSA @P
K. biyr ‘well’ (not like MSA QK. biŷr), @Q¯ qarA ‘he
read’ (not like MSA @Q¯ qaraÂa), ÉK
AÓ mAyil ‘leaning’ (not
mAŷil), and XBð wilAd ‘children’ (not like
like MSA ÉKAÓ
XBð @ ÂawlAd).
Alif-Maqsura The letter ø ý is often used in Egypt to
write word-final ø/ ø ý/y (even when writing MSA). This
is not allowed in CODA. All rules for using Alif-Maqsura
are the same as MSA.
Attached Clitics EGY uses almost all the attached clitics in MSA, e.g. the definite article + È@ Al+. EGY also
has a few additional attached clitics not in MSA, e.g., the
progressive particle proclitic + H
. bi+, the future particle
proclitic + h Ha+ and the negation particle enclitic + +š.
Some of the EGY pronominal enclitics have multiple contextual forms (allomorphs). Clitics are generally written in
their allomorphic phonemic form, with a few exceptions,
e.g., compare ½¯A šAf+ik ‘he saw you’, and ú»ñ¯A šA-
fuw+kiy ‘they saw you’; and éJ® šuft+uh ‘I saw him’,
¯A AÓ mA šAèñ¯A šAfuw+h ‘they saw him’; and ñëñ
fuw+huw+š ‘they didn’t see him’. The Shadda rule is disabled across stem-clitic boundaries (except for ø+ +ya),
ð wAHšiynnA ‘we miss you’ and úÎ« ςalay∼a.
AJ J [email protected]
Separated Clitics The indirect object enclitics and the
negation proclitic AÓ mA are written separately, e.g.,
Ë ÈA¯ AÓ mA qAl liyš /ma+’al+lı̄+š/ ‘he did not tell me’.
Ta-Marbuta The Ta-Marbuta has four forms in EGY.
Three are similar to MSA: word-final non-construct é
a~ /a/, word-final construct é i~ /it/, and word-medial
it /it/, e.g., éÊj.« ςajala~ ‘bicycle’, éÊj
. « ςajalithA ‘her bicycle’. The
ςajali~ ‘bicycle of’, and AîDÊj
fourth case is dialectal: word-medial non-construct A A /ā/,
. AJºË@ é[email protected] dArsa~ AlkitAb ‘she studied the book’ /
èA[email protected] dArsA+h ‘she studied it’.
Waw of Plurality The silent Alif added at the end of the
3rd person plural affix @ð+ +uwA in MSA is also used in
EGY. It is also added to the 2nd person plural affix @ñK+
+tuwA which is not in MSA. This silent Alif is not added
to the pronominal enclitic ñ» kuw ‘your/you [acc]’ or the
pronoun ñ[email protected] Aintuw ‘you [nom]’.
Nunation Nunation has disappeared from EGY as a productive inflection. It remains as an adverbial derivational
morpheme, e.g., AJ
ÊÔ« ςamaliy∼Aã ‘practically’.
6.4. Lexical Exceptions
EGY CODA guidelines include a word list specifying ad
hoc spellings of EGY words that may be inconsistent with
the default mapping outlined above or that have multiple commonly used spellings. Examples include pronouns
such as ñ[email protected] Aintuw ‘you’ (not @ñ[email protected] AintuwA), demonstratives such as èX dah ‘this’ (not @X dA), limited cliticiza d+AHnA ‘so+we’ (not AJkX
daHnA), adtions such as [email protected]
. barDuh ‘also’ (not èXQK. barduh, or ñQK
verbs such as éQK
barDaw), and special partly ambiguous cases such as the
¯ fiyh ‘there is’ and its negative
‘there is not’ contrasted with the closely related preposi
¯ AÓ mA
¯ fiyh ‘in it’ and its negative ñîD
fiyhuwš ‘not in it’.
More Details, More Dialects We plan on continuously
improving the CODA guidelines. There will naturally be
additional issues to address in EGY. We are also working
on developing CODA guidelines for other dialects.
Resources In terms of developing resources that are annotated for CODA, we will use a graphical user interface
tool developed under the COLABA project for annotation
(Benajiba and Diab, 2010) to annotate a large collection of
We conducted a preliminary annotation experiment using
four annotators trained with an earlier version of the guidelines.
We do not expect major differences in the statistics we report
here. The annotation covered over 110K words. 15.1% of all
words were changed. 12.1% of the changes involved a merge action (removing incorrect space between two words) and 13.2% involved a split action (adding a space to separate two incorrectly
Tools We plan to use the annotations we will create to
develop automatic CODAfication tools that can be used as
part of general preprocessing of DA data for a variety of
NLP applications. For a published early attempt at this effort, see (Dasigi and Diab, 2011).
We would like to thank Abdelati Hawwari, Sondos Krouna,
Mohamed Maamouri, Reem Faraj, Ahmed El Kholy and
Ramy Eskander for helpful feedback. This work was partially funded under DARPA project number HR0011-12-C0014.
Ernest T. Abdel-Massih, Zaki N. Abdel-Malek, and ElSaid M. Badawi. 1979. A Reference Grammar of Egyptian Arabic. Georgetown University Press.
Abbas Al-Tonsi and Laila Al-Sawi. 1990. An Intensive
Course in Egyptial Colloquial Arabic. American University in Cairo.
Plonka Arkadiusz. 2006. Le nationalisme linguistique au
liban autour de sa‘id ‘aql et l’idée de langue libanaise
dans la revue “lebnaan” en nouvel alphabet. Arabica,
Khalil ‘Asaakir. 1950. A Method for Writing Modern Arabic Dialects with Arabic Letters. (in Arabic). The Arab
Academy Magazine, 8(181–192), January.
El-Said Badawi and Martin Hinds. 1986. A Dictionary of
Egyptian Arabic. Librairie du Liban.
Yassine Benajiba and Mona Diab. 2010. A web application for dialectal arabic text annotation. In Proceedings
of the LREC Workshop for Language Resources (LRs)
and Human Language Technologies (HLT) for Semitic
Languages: Status, Updates, and Prospects.
Fadi Biadsy, Nizar Habash, and Julia Hirschberg. 2009.
Improving the Arabic Pronunciation Dictionary for
Phone and Word Recognition with Linguistically-Based
Pronunciation Rules. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the
North American Chapter of the Association for Computational Linguistics, pages 397–405, Boulder, Colorado.
Tim Buckwalter. 2007. Issues in Arabic Morphological
Analysis. In A. van den Bosch and A. Soudi, editors,
Arabic Computational Morphology: Knowledge-based
and Empirical Methods. Springer.
Mark W. Cowell. 1964. A Reference Grammar of Syrian
Arabic. Georgetown University Press.
Pradeep Dasigi and Mona Diab. 2011. Codact: Towards
identifying orthographic variants in dialectal arabic. In
Proceedings of the International Joint Conference on
Natural Language Processing, Chiang Mai, Thailand.
attached words). Among character substitutions, changing the
Alif form into one of its variants is the most common change
(22.1%) followed by cases involving the Ta-Marbuta (14.4%) and
Alif-Maqsura/Ya (8.5%); these are expected results given Arabic
orthography (Habash, 2010). Among the less common but interesting cases linguistically, we find that 1.7% of the words have a
H t → H θ change and 0.8% of all words have a change involving
the letter q. Inter-annotator agreement is about 98%.
ð éë éë
J« ù®K . èY» AJkX
øX èYÓAm.Ì '@ ©J
@ ñÖÏ @ Q
[email protected] ÐñÒªË@ úÎ« ÉÓA¾JÓ ÉÓA¿ ½JK . AKY
JË@ ©Ó @ñºJ.
. á Ëñ®JÓ AªJ.£ YJ
. k. ð HYî
I.ËAªK é@ Q¯ úÎ« Ég. QË@ Y¯P AJë . Iº
AÖÏð . AK
AªÓ úæ®¯ð HY
¯[email protected] AÖÏ : CKA¯ AîE
K. ½ÓA¯ àA Jm '. éJ
Ë@ Q¢ J Kð éJ.KAm .'. éJk. ð P Yg. ð A¯@ AÒÊ¿ð éK. ñJ.J
m ½[email protected] Ag [email protected] éJ
Ë ¬PA« Ó . úæ.Ô . èXQK. úæ[email protected] úæ¯ñËXð . úæ.Ô . úæJ» Qm @ AJJ
K. AÖÏð . úæ.Ôg.
hh hh wAll~ Alς Ďym fTst mn AlDHk Ay~ yAςbyr AlmwADyς AljAmd~ dý dHnA kd~ bqý ςndnA bnk kAml mtkAml ςlý Alςmwm
AnA Ajthdt wjbt šwy~ nkt wAkyd TbςA mnqwlyn bs yArb yςjbwkwA AsybkwA mς Alnkt . hnA rqd Alrjl ςlý frAš~ yγAlb
Alγybwb~ wklmA AfAq wjd zwjt~ bjAnb~ wtnĎr Aly~ bHnAn fAmsk bydyhA qAŷlA : lmA Atrfdt wqftý mςAyA . wlmA šrktý
flst kntý jmbý . wlmA bytnA AtHrq kntý jmbý . wdlwqtý Antý brd~ jmbý . mš ςArf ly~ AnA HAšš Ank nHs
J« ù®K . èY» [email protected]
ø X èYÓAm.Ì '@ ©J
@ ñÖÏ @ Q
é<Ë@ð éë éë
ÐñÒªË@ úÎ« ÉÓA¾JÓ ÉÓA¿ ½JK . AKY
. @ [email protected]
. k. ð HYî
JË@ ©Ó ñºJ.
H. P AK
. á Ëñ® JÓ AªJ.£ YJ
» @ð Iº
é@ Q¯ úÎ« Ég. QË@ Y¯P AJë . Iº
AÖÏð . ø AªÓ úæ® ¯ð
I.ËAªK ¯Q [email protected] AÖÏ : CKA ¯ AîE
K. ½Ó A¯ àA Jm '. éJ
Ë@ Q¢ J Kð éJ.KAm .'. éJk. ð P Yg. ð A ¯ @ AÒÊ¿ð éK . ñJ.J
' . úæJk
m' ½[email protected] Ag [email protected] éJ
. . éQK. ú
æ[email protected] ú
æ¯ñËXð . ú
æJ» Qm @ AJJ
K. AÖÏð . ú
hh hh wAllh Alς Ďym fTst mn AlDHk Ǎyh yA ςbyr AlmwADyς AljAmd~ dy dAHnA kdh bqý ςndnA bnk kAml mtkAml ςlý
Alςmwm AnA Ajthdt wjbt šwy~ nkt wÂkyd TbςA mnqwlyn bs yA rb yςjbwkw Asybkw mς Alnkt . hnA rqd Alrjl ςlý frAšh
yγAlb Alγybwb~ wklmA ÂfAq wjd zwjth bjAnbh wtnĎr Ǎlyh bHnAn fÂmsk bydyhA qAŷlA : lmA Atrfdt wqfty mςAy .
wlmA šrkty flst knty jnby . wlmA bytnA AtHrq knty jnby . wdlwqty Anty brDh jnby . mš ςArf lyh AnA HAss Ank nHs
English ha ha [,] I swear to God [,] I died from laughter [,] Abeer [,] what cool topics [!] we now have a complete comprehensive
bank [.] any way [,] I put some effort and got some jokes that are of course copied [,] but hopefully you will like them [.] I
leave you with the jokes . [MSA] There lied a man on his bed coming in and out of a coma [;] and every time he woke up he
found his wife by his side looking at him lovingly [.] so he held her hands and said [/MSA]: when I got fired [,] you stood by
me . And when my company went bankrupt you were by my side . And when our house burnt down you were by my side .
And now also you are by my side . I don’t know why I feel you’re bad luck [.]
Figure 1: An Egyptian Arabic snippet in raw and CODA orthography. Bracketed punctuation and comments are added in
the English translation to help the reader. The region between [MSA] and [/MSA] is in MSA. Bolding in the CODA row
marks modified words.
Mona Diab and Heba Elfardy. 2012. Simplified guidelines
for the creation of large scale dialectal arabic annotations. In Language Resources and Evaluation Conference (LREC), Istanbul.
Mona Diab, Nizar Habash, Owen Rambow, Mohamed AlTantawy, and Yassine Benajiba. 2010. Colaba: Arabic dialect annotation and processing. In Proceedings of
the LREC Workshop for Language Resources (LRs) and
Human Language Technologies (HLT) for Semitic Languages: Status, Updates, and Prospects.
Y. A. El-Imam. 2004. Phonetization of Arabic: Rules
and Algorithms. In Computer Speech and Language 18,
Wallace Erwin. 1963. A Short Reference Grammar of Iraqi
Arabic. Georgetown University Press.
Nizar Habash and Owen Rambow. 2007. Morphophonemic and Orthographic Rules in a Multi- Dialectal Morphological Analyzer and Generator for Arabic Verbs. In
International Symposium on Computer and Arabic Language (ISCAL), Riyadh, Saudi Arabia.
Nizar Habash, Abdelhadi Soudi, and Tim Buckwalter.
2007. On Arabic Transliteration. In A. van den Bosch
and A. Soudi, editors, Arabic Computational Morphology: Knowledge-based and Empirical Methods.
Nizar Habash, Owen Rambow, Mona Diab, and Reem
Kanjawi-Faraj. 2008. Guidelines for Annotation of Arabic Dialectness. In Proceedings of the LREC Workshop
on HLT & NLP within the Arabic world.
Nizar Habash, Mona Diab, and Owen Rambow. 2012.
Conventional Orthography for Dialectal Arabic: Principles and Guidelines – Egyptian Arabic. Technical Report
CCLS-12-02, Columbia University Center for Computational Learning Systems.
Nizar Habash. 2006. On Arabic and its Dialects. Multilingual Magazine, 17(81).
Nizar Habash. 2010. Introduction to Arabic Natural Language Processing. Morgan & Claypool Publishers.
Richard Harrell. 1962. A Short Reference Grammar of Moroccan Arabic. Georgetown University Press.
Clive Holes. 2004. Modern Arabic: Structures, Functions,
and Varieties. Georgetown Classics in Arabic Language
and Linguistics. Georgetown University Press.
H. Kilany, H. Gadalla, H. Arram, A. Yacoub, A. ElHabashi, and C. McLemore. 2002. Egyptian Colloquial
Arabic Lexicon. LDC catalog number LDC99L22.
Eugene E. Loos, Susan Anderson, Jr. Dwight H., Day,
Paul C. Jordan, and J. Douglas Wingate. 2004. Glossary
of Linguistic Terms.
Mohamed Maamouri, Tim Buckwalter, and Christopher
Cieri. 2004. Dialectal Arabic Telephone Speech Corpus: Principles, Tool design, and Transcription Conventions. In NEMLAR International Conference on Arabic
Language Resources and Tools.
Omar F. Zaidan and Chris Callison-Burch. 2011. The arabic online commentary dataset: an annotated dataset of
informal arabic with high dialectal content. In Proceedings of the Association for Computational Linguistics,
Portland, Oregon, USA.