Arabic Dialect Identification

The Arabic language is a loose term that refers to the many existing varieties of Arabic. Those varieties include one “written” form, Modern Standard Arabic (MSA), and many “spoken” forms, each of which is a regional dialect. MSA is the only variety that is standardized, regulated, and taught in schools, necessitated by its use in written communication and formal venues. The regional dialects, used primarily for day-to-day dealings and spoken communication, remain somewhat absent from written communication compared with MSA. That said, it is certainly possible to produce dialectal Arabic text, by using the same letters used in MSA and the same (mostly phonetic) spelling rules of MSA.

One domain of written communication in which both MSA and dialectal Arabic are commonly used is the on-line domain: Dialectal Arabic has a strong presence in blogs, forums, chatrooms, and user/reader commentary. Harvesting data from such sources is a viable option for computational linguists to create large data sets to be used in statistical learning setups. However, because all Arabic varieties use the same character set, and furthermore much of the vocabulary is shared among different varieties, it is not a trivial matter to distinguish and separate the dialects from each other.

In this article, we focus on the problem of Arabic dialect identification. We describe a large data set that we created by harvesting a large amount of reader commentary on on-line newspaper content, and describe our annotation effort on a subset of the harvested data. We crowdsourced an annotation task to obtain sentence-level labels indicating what proportion of the sentence is dialectal, and which dialect the sentence is written in. Analysis of the collected labels reveals interesting annotator behavior patterns and biases, and the data are used to train and evaluate automatic classifiers for dialect detection and identification. Our approach, which relies on training language models for the different Arabic varieties, greatly outperforms baselines that use (much more) MSA-only data: On one of the classification tasks we considered, where human annotators achieve 88.0% classification accuracy, our approach achieves 85.7% accuracy, compared with only 66.6% accuracy by a system using MSA-only data.

The article is structured as follows. In Section 2, we provide an introduction to the various Arabic varieties and corresponding data resources. In Section 3, we introduce the dialect identification problem for Arabic, discussing what makes it a difficult problem, and what applications would benefit from it. Section 4 provides details about our annotation set-up, which relied on crowdsourcing the annotation to workers on Amazon's Mechanical Turk. By examining the collected labels and their distribution, we characterize annotator behavior and observe several types of human annotator biases. We introduce our technique for automatic dialect identification in Section 5. The technique relies on training separate language models for the different Arabic varieties, and scoring sentences using these models. In Section 6, we report on a large-scale Web crawl that we performed to gather a large amount of Arabic text from on-line newspapers, and apply our classifier on the gathered data. Before concluding, we give an overview of related work in Section 7.

Although the Arabic language has an official status in over 20 countries and is spoken by more than 250 million people, the term itself is used rather loosely and refers to different varieties of the language. Arabic is characterized by an interesting linguistic dichotomy: the written form of the language, MSA, differs in a non-trivial fashion from the various spoken varieties of Arabic, each of which is a regional dialect (or a lahjah, lit. “accent”; also darjah, lit. “modern”). MSA is the only variety that is standardized, regulated, and taught in schools. This is necessitated because of its use in written communication in formal venues.¹ The regional dialects, used primarily for day-to-day dealings and spoken communication, are not taught formally in schools, and remain somewhat absent from traditional, and certainly official, written communication.

Unlike MSA, a regional dialect does not have an explicit written set of grammar rules regulated by an authoritative organization, but there is certainly a concept of grammatical and ungrammatical.² Furthermore, even though they are “spoken” varieties, it is certainly possible to produce dialectal Arabic text, by spelling out words using the same spelling rules used in MSA, which are mostly phonetic.³

There is a reasonable level of mutual intelligibility across the dialects, but the extent to which a particular individual is able to understand other dialects depends heavily on that person's own dialect and their exposure to Arab culture and literature from outside of their own country. For example, the typical Arabic speaker has little trouble understanding the Egyptian dialect, thanks in no small part to Egypt's history in movie-making and television show production, and their popularity across the Arab world. On the other hand, the Moroccan dialect, especially in its spoken form, is quite difficult to understand by a Levantine speaker. Therefore, from a scientific point of view, the dialects can be considered separate languages in their own right, much like North Germanic languages (Norwegian/Swedish/Danish) and West Slavic languages (Czech/Slovak/Polish).⁴

One possible breakdown of regional dialects into main groups is as follows (see Figure 1):

• Egyptian: The most widely understood dialect, due to a thriving Egyptian television and movie industry, and Egypt's highly influential role in the region for much of the 20th century (Haeri 2003).

• Levantine: A set of dialects that differ somewhat in pronunciation and intonation, but are largely equivalent in written form; closely related to Aramaic (Bassiouney 2009).

• Gulf: Folk wisdom holds that Gulf is the closest of the regional dialect to MSA, perhaps because the current form of MSA evolved from an Arabic variety originating in the Gulf region. Although there are major differences between Gulf and MSA, Gulf has notably preserved more of MSA's verb conjugation than other varieties have (Versteegh 2001).

• Iraqi: Sometimes considered to be one of the Gulf dialects, though it has distinctive features of its own in terms of prepositions, verb conjugation, and pronunciation (Mitchell 1990).

• Maghrebi: Heavily influenced by the French and Berber languages. The Western-most varieties could be unintelligible by speakers from other regions in the Middle East, especially in spoken form. The Maghreb is a large region with more variation than is seen in other regions such as the Levant and the Gulf, and could be subdivided further (Mohand 1999).

There are a large number of linguistic differences between MSA and the regional dialects. Some of those differences do not appear in written form if they are on the level of short vowels, which are omitted in Arabic text anyway. That said, many differences manifest themselves textually as well:

• • 

  MSA's morphology is richer than dialects' along some dimensions such as case and mood. For instance, MSA has a dual form in addition to the singular and plural forms, whereas the dialects mostly lack the dual form. Also, MSA has two plural forms, one masculine and one feminine, whereas many (though not all) dialects often make no such gendered distinction.⁵ On the other hand, dialects have a more complex cliticization system than MSA, allowing for circumfix negation, and for attached pronouns to act as indirect objects.

• • 

  Dialects lack grammatical case, whereas MSA has a complex case system. In MSA, most cases are expressed with diacritics that are rarely explicitly written, with the accusative case being a notable exception, as it is expressed using a suffix (+A) in addition to a diacritic (e.g., on objects and adverbs).

• • 

  There are lexical choice differences in the vocabulary itself. Table 1 gives several examples. Note that these differences go beyond a lack of orthography standardization.

• • 

  Differences in verb conjugation, even when the triliteral root is preserved. See the lower part of Table 1 for some conjugations of the root š-r-b (to drink).

This list, and Table 1, deal with differences that are expressed at the inidividual-word level. It is important to note that Arabic varieties differ markedly in style and sentence composition as well. For instance, all varieties of Arabic, MSA, and otherwise, allow both SVO and VSO word orders, but MSA has a higher incidence of VSO sentences than dialects do (Aoun, Benmamoun, and Sportiche 1994; Shlonsky 1997).

Despite the fact that speakers are usually less comfortable communicating in MSA than in their own dialect, MSA content significantly dominates dialectal content, as MSA is the variant of choice for formal and official communication. Relatively little printed material exists in local dialects, such as folkloric literature and some modern poetry, but the vast majority of published Arabic is in MSA. As a result, MSA's dominance is also apparent in data sets available for linguistic research. The problem is somewhat mitigated in the speech domain, since dialectal data exists in the form of phone conversations and television program recordings, but, in general, dialectal Arabic data sets are hard to come by.

The abundance of MSA data has greatly aided research on computational methods applied to Arabic, but only the MSA variant of it. For example, a state-of-the-art Arabic-to-English machine translation system performs quite well when translating MSA source sentences, but often produces incomprehensible output when the input is dialectal. For example, most words of the dialectal sentence shown in Figure 2 are transliterated, whereas an equivalent MSA sentence is handled quite well. The high transliteration rate is somewhat alarming, as the first two words of the dialectal sentence are relatively frequent function words: Aymtý means ‘when’ and rH corresponds to the modal ‘will’.

Figure 3 shows another dialectal sentence, this time in Egyptian, which again causes the system to produce a poor translation even for frequent words. Case in point, the system is unable to consistently handle any of Ayh (‘what’), Ally (the conjunction ‘that’), or dh (‘this’). Granted, it is conceivable that processing dialectal content is more difficult than MSA, but the main problem is the lack of dialectal training data.⁶

This is an important point to take into consideration, because the dialects differ to a large enough extent to warrant treating them as more or less different languages. The behavior of machine translation systems translating dialectal Arabic when the system has been trained exclusively on MSA data is similar to the behavior of a Spanish-to-English MT system when a user inputs a Portuguese sentence. Figure 4 illustrates how MT systems behave (the analogy is not intended to draw a parallel between the linguistic differences MSA-dialect and Spanish-Portuguese). The MT system's behavior is similar to the Arabic example, in that words that are shared in common between Spanish and Portuguese are translated, while the Portuguese words that were never observed in the Spanish training data are left untranslated.

This example illustrates the need for dialectal data to train MT systems to handle dialectal content properly. A similar scenario would arise with many other NLP tasks, such as parsing or speech recognition, where dialectal content would be needed in large quantities for adequate training. A robust dialect identifier could sift through immense volumes of Arabic text, and separate out dialectal content from MSA content.

One domain of written communication in which MSA and dialectal Arabic are both commonly used is the on-line domain, because it is more individual-driven and less institutionalized than other venues. This makes a dialect much more likely to be the user's language of choice, and dialectal Arabic has a strong presence in blogs, forums, chatrooms, and user/reader commentary. Therefore, on-line data is a valuable resource of dialectal Arabic text, and harvesting this data is a viable option for computational linguists for purposes of creating large data sets to be used in statistical learning.

We created the Arabic On-line Commentary Data Set (AOC) (Zaidan and Callison-Burch 2011) a 52M-word monolingual data set by harvesting reader commentary from the on-line versions of three Arabic newspapers (Table 2). The data is characterized by the prevalence of dialectal Arabic, alongside MSA, mainly in Levantine, Gulf, and Egyptian. These correspond to the countries in which the three newspapers are published: Al-Ghad is from Jordan, Al-Riyadh is from Saudi Arabia, and Al-Youm Al-Sabe' is from Egypt.⁷

Although a significant portion of the AOC's content is dialectal, there is still a very large portion of it that is in MSA. (Later analysis in Section 4.2.1 shows dialectal content is roughly 40%.) In order to take full advantage of the AOC (and other Arabic data sets with at least some dialectal content), it is desirable to separate dialectal content from non-dialectal content automatically. The task of dialect identification (and its automation) is the focus for the remainder of this article. We next present the task of Arabic dialect identification, and discuss our effort to create a data set of Arabic sentences with their dialectal labels. Our annotation effort relied on crowdsourcing the annotation task to Arabic-speakers on Amazon's Mechanical Turk service (Section 3).

The discussion of the varieties of Arabic and the differences between them gives rise to the task of automatic dialect identification (DID). In its simplest form, the task is to build a learner that can, given an Arabic sentence S, determine whether or not S contains dialectal content. Another form of the task would be to determine in which dialect S was written, which requires identification at a more fine-grained level.

In many ways, DID is equivalent to language identification. Although language identification is often considered to be a “solved problem,” DID is most similar to a particularly difficult case of language ID, where it is applied to a group of closely related languages that share a common character set. Given the parallels between DID and language identification, we investigate standard statistical methods to establish how difficult the task is. We discuss prior efforts for Arabic DID in Section 7.

Despite the differences illustrated in the previous section, in which we justify treating the different dialects as separate languages, it is not a trivial matter to automatically distinguish and separate the dialects from each other. Because all Arabic varieties use the same character set, and because much of the vocabulary is shared among different varieties, identifying dialect in a sentence is not simply a matter of, say, compiling a dialectal dictionary and detecting whether or not a given sentence contains dialectal words.

This word-level source ambiguity is caused by several factors:

• A dialectal sentence might consist entirely of words that are used across all Arabic varieties, including MSA. Each of the sentences in Figure 5 consists of words that are used both in MSA and dialectally, and an MSA-based dictionary would not (and should not) recognize those words as out of vocabulary (OOV). Nevertheless, the sentences are heavily dialectal.

• Some words are used across the varieties with different functions. For example, Tyb is used dialectally as an interjection, but is an adjective in MSA. (This is similar to the English usage of okay.)

• Primarily due to the omission of short vowels, a dialectal word might have the same spelling as an MSA word with an entirely different meaning, forming pairs of heteronyms. This includes strongly dialectal words such as dwl and nby: dwl is either Egyptian for these (pronounced dowl) or the MSA for countries (pronounced duwal); nby is either the Gulf for we want (pronounced nibi) or the MSA for prophet (pronounced nabi).

It might not be clear for a non-Arabic speaker what makes certain sentences, such as those of Figure 5, dialectal, even when none of the individual words are. The answer lies in the structure of such sentences and the particular word order within them, rather than the individual words themselves taken in isolation. Figure 6 shows MSA sentences that express the same meaning as the dialectal sentences from Figure 5. As one could see, the two versions of any given sentence could share much of the vocabulary, but in ways that are noticeably different to an Arabic speaker. Furthermore, the differences would be starker still if the MSA sentences were composed from scratch, rather than by modifying the dialectal sentences, since the tone might differ substantially when composing sentences in MSA.

Being able to perform automatic DID is interesting from a purely linguistic and experimental point of view. In addition, automatic DID has several useful applications:

• Distinguishing dialectal data from non-dialectal data would aid in creating a large monolingual dialectal data set, exactly as we would hope to do with the AOC data set. Such a data set would aid many NLP systems that deal with dialectal content, for instance, to train a language model for an Arabic dialect speech recognition system (Novotney, Schwartz, and Khudanpur 2011). Identifying dialectal content can also aid in creating parallel data sets for machine translation, with a dialectal source side.

• A user might be interested in content of a specific dialect, or, conversely, in strictly non-dialectal content. This would be particularly relevant in fine-tuning and personalizing search engine results, and could allow for better user-targeted advertising. In the same vein, being able to recognize dialectal content in user-generated text could aid in characterizing communicants and their biographic attributes (Garera and Yarowsky 2009).

• In the context of an application such as machine translation (MT), identifying dialectal content could be quite helpful. Most MT systems, when faced with OOV words, either discard the words or make an effort to transliterate them. If a segment is identified as being dialectal first, the MT system might instead attempt to find equivalent MSA words, which are presumably easier to process correctly (e.g., as in Salloum and Habash [2011] and, to some degree, Habash [2008]). Even for non-OOV words, identifying dialectal content before translating could be critical to resolve the heteronym ambiguity of the kind mentioned in Section 3.1.

In this section, we discuss crowdsourcing Arabic dialect annotation. We discuss how we built a data set of Arabic sentences, each of which is labeled with whether or not it contains dialectal content. The labels include additional details about the level of dialectal content (i.e., how much dialect there is), and of which type of dialect it is. The sentences themselves are sampled from the AOC data set, and we observe that about 40% of sentences contain dialectal content, with that percentage varying between 37% and 48%, depending on the news source.

Collecting annotated data for speech and language applications requires careful quality control (Callison-Burch and Dredze 2010). We present the annotation interface and discuss an effective way for quality control that can detect spamming behavior. We then examine the collected data itself, analyzing annotator behavior, measuring agreement among annotators, and identifying interesting biases exhibited by the annotators. In Section 5, we use the collected data to train and evaluate statistical models for several dialect identification tasks.

The annotation interface displayed a group of Arabic sentences, randomly selected from the AOC. For each sentence, the annotator was instructed to examine the sentence and make two judgments about its dialectal content: the level of dialectal content, and its type, if any. The instructions were kept short and simple:

The instructions were accompanied by the map of Figure 1, to visually illustrate the dialect breakdown. Figure 7 shows the annotator interface populated with some actual examples, with labeling in progress. We also collected self-reported information such as native Arabic dialect and age (or number of years speaking Arabic for non-native speakers). The interface also had built-in functionality to detect each annotator's geographic location based on their IP address.

Of the 3.1M sentences in the AOC, we randomly⁹ selected a “small” subset of about 110,000 sentences to be annotated for dialect.

For each sentence shown in the interface, we asked annotators to label which dialect the segment is written in and the level of dialect in the segment. The dialect labels were Egyptian, Gulf, Iraqi, Levantine, Maghrebi, other dialect, general dialect (for segments that could be classified as multiple dialects), dialect but unfamiliar (for sentences that are clearly dialect, but are written in a dialect that the annotator is not familiar with), no dialect (for MSA), or not Arabic (for segments written in English or other languages). Options for the level of dialect included no dialect (for MSA), a small amount of dialect, an even mix of dialect and MSA, mostly dialect, and not Arabic. For this article we use only the dialect labels, and not the level of dialect. Zaidan (2012) incorporates finer-grained labels into an “annotator rationales” model (Zaidan, Eisner, and Piatko 2007).

The sentences were randomly grouped into sets of 10 sentences each, and when Workers performed our task, they were shown the 10 sentences of a randomly selected set on a single HTML page. As a result, each screen contained a mix of sentences across the three newspapers presented in random order. As control items, each screen had two additional sentences that were randomly sampled from the article bodies. Such sentences are almost always in MSA Arabic, and so their expected label is MSA. Any worker who frequently mislabeled the control sentences with a non-MSA label was considered a spammer, and their work was rejected. Hence, each screen had twelve sentences in total.

We offered a reward of $0.05 per screen (later raised to $0.10), and had each set redundantly completed by three distinct Workers. The data collection lasted about 4.5 months, during which 33,093 Human Intelligence Task (HIT) Assignments were completed, corresponding to 330,930 collected labels (excluding control items). The total cost of annotation was $3,050.52 ($2,773.20 for rewards, and $277.32 for Amazon's commission).

With the aid of the embedded control segments (taken from article bodies) and expected dialect label distribution, it was possible to spot spamming behavior and reject it. Table 3 shows three examples of workers whose work was rejected on this basis, having clearly demonstrated they are unable or unwilling to perform the task faithfully. In total, 11.4% of the assignments were rejected on this basis. In the approved assignments, the embedded MSA control sentence was annotated with the MSA label 94.4% of the time. In the remainder of this article, we analyze only data from the approved assignments.

We note here that we only rejected assignments where the annotator's behavior was clearly problematic, opting to approve assignments from workers mentioned later in Section 4.2.3, who exhibit systematic biases in their labels. Although these annotators' behavior is non-ideal, we cannot assume that they are not working faithfully, and therefore rejecting their work might not be fully justified. Furthermore, such behavior might be quite common, and it is worth investigating these biases to benefit future research.

Overall, 454 annotators participated in the task, 138 of whom completed at least 10 HITs. Upon examination of the provided labels for the commentary sentences, 40.7% of them indicate some level of dialect, and 57.1% indicate no dialectal content (Figure 8a). Note that 2.14% of the labels identify a sentence as being non-Arabic, non-textual, or as being left unanswered. The label breakdown is a strong confirmation of our initial motivation, which is that a large portion of reader commentary contains dialectal content.¹⁰

Figure 8 also illustrates the following:

• The most common dialectal label within a given news source matches the dialect of the country of publication. This is not surprising, since the readership for any newspaper is likely to mostly consist of the local population of that country. Also, given the newspapers' countries of publication, there is almost no content that is in a dialect other than Levantine, Gulf, or Egyptian. For this reason, other dialects such as Iraqi and Maghrebi, all combined, correspond to less than 0.01% of our data, and we mostly drop them from further discussion.

• The three news sources vary in the prevalence of dialectal content. The Egyptian newspaper has a markedly larger percentage of dialectal content (46.6% of labels) compared with the Saudi newspaper (40.1%) and the Jordanian newspaper (36.8%).

• A nontrivial amount of labels (5–8%) indicate General dialectal content. The General label was meant to indicate a sentence that is dialectal but lacks a strong indication of a particular dialect. Although many of the provided General labels seem to reflect an intent to express this fact, there is evidence that some annotators used this category in cases where choosing the label Not sure would have been more appropriate but was ignored (see Section 4.2.3).

• Non-Arabic content, although infrequent, is not a rare occurrence in the Jordanian and Egyptian newspapers, at around 3%. The percentage is much lower in the Saudi newspaper, at 0.8%. This might reflect the deeper penetration of the English language (and English-only keyboards) in Jordan and Egypt compared with Saudi Arabia.

We can associate a label with each segment based on the majority vote over the three provided labels for that segment. If a sentence has at least two annotators choosing a dialectal label, we label it as dialect. If it has at least two annotators choosing the MSA label, we label it as MSA.¹¹ In the remainder of the article, we will report classification accuracy rates that assume the presence of gold-standard class labels. Unless otherwise noted, this majority-vote label set is used as the gold-standard in such experiments.

In experiments where the dialectal label set is more fine-grained (i.e., LEV, GLF, and EGY instead of simply dialect), we assign to the dialectal sentence the label corresponding to the news source's country of publication. That is, dialectal sentences in the Jordanian (respectively, Saudi, Egyptian) are given the label LEV (respectively, GLF, EGY). We could have used dialect labels provided by the annotators, but chose to override those using the likely dialect of the newspaper instead. It turns out that sentences with an EGY majority, for instance, are extremely unlikely to appear in either the Jordanian or Saudi newspaper—only around 1% of those sentences have an EGY majority. In the case of the Saudi newspaper, 9% of all dialectal sentences were originally annotated as LEV but were transformed to GLF. Our rationales for performing the transformation is that no context was given for the sentences when they were annotated, and annotators had a bias towards their own dialect. We provide the original annotations for other researchers to re-analyze if they wish.

Even when a sentence would receive a majority-vote label that differs from the news source's primary dialect, inspection of such sentences reveals that the classification was usually unjustified, and reflected a bias towards the annotator's native dialect. Case in point: Gulf-speaking annotators were in relatively short supply, whereas a plurality of annotators spoke Levantine (see Table 4). Later in Section 4.2.3, we point out that annotators have a native-dialect bias, whereby they are likely to label a sentence with their native dialect even when the sentence has no evidence of being written in that particular dialect. This explains why a non-trivial number of LEV labels were given by annotators to sentences from the Saudi newspaper (Figure 8). In reality, most of these labels were given by Levantine speakers over-identifying their own dialect. Even if we were to assign dialect labels based on the (Levantine-biased) majority votes, Levantine would only cover 3.6% of the sentences from the Saudi newspaper.¹²

Therefore, for simplicity, we assume that a dialectal sentence is written in the dialect corresponding to the sentence's news source, without having to inspect the specific dialect labels provided by the annotators. This not only serves to simplify our experimental set-up, but also contributes to partially reversing the native dialect bias that we observed.

The annotators exhibit a decent level of agreement with regard to whether a segment is dialectal or not, with full agreement (i.e., across all three annotators) on 72.2% of the segments regarding this binary dialect/MSA decision. This corresponds to a kappa value of 0.619 (using the definition of Fleiss [1971] for multi-rater scenarios), indicating very high agreement.¹³ The full-agreement percentage decreases to 56.2% when expanding the classification from a binary decision to a fine-grained scale that includes individual dialect labels as well. This is still quite a reasonable result, since the criterion is somewhat strict: It does not include a segment labeled, say, {Levantine, Levantine, General}, though there is good reason to consider that annotators are in “agreement” in such a case.

So how good are humans at the classification task? We examine their classification accuracy, dialect recall, and MSA recall. The classification accuracy is measured over all sentences, both MSA and dialectal. We define dialect (MSA) recall to be the number of sentences labeled as being dialectal (MSA), over the total number of sentences that have dialectal (MSA) labels based on the majority vote. Overall, human annotators have a classification accuracy of 90.3%, with dialect recall at 89.0%, and MSA recall at 91.5%. Those recall rates do vary across annotators, as shown in Figure 9, causing some accuracy rates to drop as low as 80% or 75%. Of the annotators performing at least five HITs, 89.4% have accuracy rates greater than or equal to 80%.

Most annotators have both high MSA recall and high dialect recall, with about 70% of them achieving at least 80% in both MSA and dialect recall. Combined with the general agreement rate measure, this is indicative that the task is well-defined—it is unlikely that many people would agree on something that is incorrect.

We note here that the accuracy rate (90.3%) is a slight overestimate of the human annotators' accuracy rate, by virtue of the construction of the gold labels. Because the correct labels are based on a majority vote of the annotators' labels themselves, the two sets are not independent, and an annotator is inherently likely to be correct. A more informative accuracy rate disregards the case where only two of the three annotators agreed and the annotator whose accuracy was being evaluated contributed one of those two votes. In other words, an annotator's label would be judged against a majority vote that is independent of that annotator's label. Under this evaluation set-up, the human accuracy rate slightly decreases, to 88.0%.

Examining the submitted labels of individual workers reveals interesting annotation patterns, and indicates that annotators are quite diverse in their behavior. An annotator can be observed to have one or more of the following bias types:¹⁴

• MSA bias/dialect bias: Figure 9 shows that annotators vary in how willing they are to label a sentence as being dialectal. Whereas most workers (top right) exhibit both high MSA and high dialect recall, other annotators have either a MSA bias (top left) or a dialect bias (bottom right).

• Dialect-specific bias: Many annotators over-identify a particular dialect, usually their native one. If we group the annotators by their native dialect and examine their label breakdown (Table 4), we find that Levantine speakers over-identify sentences as being Levantine, Gulf speakers over-identify Gulf, and Egyptian speakers over-identify Egyptian. This holds for speakers of other dialects as well, as they over-identify other dialects more often than most speakers. Another telling observation is that Iraqi speakers have a bias for the Gulf dialect, which is quite similar to Iraqi. Maghrebi speakers have a bias for Egyptian, reflecting their unfamiliarity with the geographically distant Levantine and Gulf dialects.

• TheGeneralbias: The General label is meant to signify sentences that cannot be definitively classified as one dialect over another. This is the case when enough evidence exists that the sentence is not in MSA, but contains no evidence for a specific dialect. In practice, some annotators make very little use of this label, even though many sentences warrant its use, whereas other annotators make extensive use of this label (see, for example, Table 5). One interesting case is that of annotators whose General label seems to mean they are unable to identify the dialect, and a label like Not sure might have been more appropriate. Take the case of the Maghrebi worker in Table 5, whose General bias is much more pronounced in the Jordanian and Saudi newspapers. This is an indication she might have been having difficulty distinguishing Levantine and Gulf from each other, but that she is familiar with the Egyptian dialect.

From a computational point of view, we can think of dialect identification as language identification, though with finer-grained distinctions that make it more difficult than typical language ID. Even languages that share a common character set can be distinguished from each other at high accuracy rates using methods as simple as examining character histograms (Cavnar and Trenkle 1994; Dunning 1994; Souter et al. 1994), and, as a largely solved problem, the one challenge becomes whether languages can be identified for very short segments.

Due to the nature and characteristics and high overlap across Arabic dialects, relying on character histograms alone is ineffective (see Section 5.3.1), and more context is needed. We will explore higher-order letter models as well as word models, and determine what factors determine which model is best.

Note that the decision process takes into account the prior distribution of the classes, which is estimated from the training set. The training set is also used to train probabilistic models to estimate the probability of S given a particular class. We rely on training n-gram language models to compute such probabilities, and apply Kneser-Ney smoothing to these probabilities and also use that technique to assign probability mass to unseen or OOV items (Chen and Goodman 1998). In language model scoring, a sentence is typically split into words. We will also consider letter-based models, where the sentence is split into sequences of characters. Note that letter-based models would be able to take advantage of clues in the sentence that are not complete words, such as prefixes or suffixes. This would be useful if the amount of training data is very small, or if we expect a large domain shift between training and testing, in which case content words indicative of MSA or dialect might not still be valuable in the new domain.

Although our classification method is based only on language model scoring, and is thus relatively simple, it is nevertheless very effective. Experimental results in Section 5.3 (e.g., Figure 10) indicate that this method yields accuracy rates above 85%, only slightly behind the human accuracy rate of 88.0% reported in Section 4.2.2.

To properly evaluate classification performance trained on dialectal data, we compare the language-model classifiers to two baselines that do not use the newly collected data. Rather, they use available MSA-only data and attempt to determine how MSA-like a sentence is.

The first baseline is based on the assumption that a dialectal sentence would contain a higher percentage of “non-MSA” words that cannot be found in a large MSA corpus. To this end, we extracted a vocabulary list from the Arabic Gigaword Corpus, producing a list of 2.9M word types. Each sentence is given a score that equals the OOV percentage, and if this percentage exceeds a certain threshold, the sentence is classified as being dialectal. For each of the cross validation runs in Section 5.3.1, we use the threshold that yields the optimal accuracy rate on the test set (hence giving this baseline as much a boost as possible). In our experiments, we found this threshold to be usually around 10%.

The second approach uses a more fine-grained approach. We train a language model using MSA-only data, and use it to score a test sentence. Again, if the perplexity exceeds a certain threshold, the sentence is classified as being dialectal. To take advantage of domain knowledge, we train this MSA model on the sentences extracted from the article bodies of the AOC, which corresponds to 43M words of highly relevant content.

In this section, we explore using the collected labels to train word- and letter-based DID systems, and show that they outperform other baselines that do not utilize the annotated data.

We measure classification accuracy at various training set sizes, using 10-fold cross validation, for several classification tasks. We examine the task both as a general MSA vs. dialect task, as well as when restricted within a particular news source. We train unigram, bigram, and trigram (word-based) models, as well as unigraph, trigraph, and 5-graph (letter-based) models. Table 6 summarizes the accuracy rates for these models, and includes rates for the baselines that do not utilize the dialect-annotated data.

Generally, we find that a unigram word model performs best, with a 5-graph model slightly behind (Figure 11). Bigram and trigram word models seem to suffer from the sparseness of the data and lag behind, given the large number of parameters they would need to estimate (and instead resort to smoothing heavily). The letter-based models, with a significantly smaller vocabulary size, do not suffer from this problem, and perform well. This is a double-edged sword though, especially for the trigraph model, as it means the model is less expressive and converges faster.

Overall though, the experiments show a clear superiority of a supervised method, be it word- or letter-based, over baselines that use existing MSA-only data. Whichever model we choose (with the exception of the unigraph model), the obtained accuracy rates show a significant dominance over the baselines.

It is worth noting that a classification error becomes less likely to occur as the length of the sentence increases (Figure 12). This is not surprising given prior work on the language identification problem (Řehůřek and Kolkus 2009; Verma, Lee, and Zakos 2009), which points out that the only “interesting” aspect of the problem is performance on short segments. The same is true in the case of dialect identification: a short sentence that contains even a single misleading feature is prone to misclassification, whereas a long sentence is likely to have other features that help identify the correct class label.¹⁵

One could also observe that distinguishing MSA from dialect is a more difficult task in the Saudi newspaper than in the Jordanian paper, which in turn is harder than in the Egyptian newspaper. This might be considered evidence that the Gulf dialect is the closest of the dialects to MSA, and Egyptian is the farthest, in agreement with the conventional wisdom. Note also that this is not due to the fact that the Saudi sentences tend to be significantly shorter—the ease of distinguishing Egyptian holds even at higher sentence lengths, as shown by Figure 12.

The experiments reported earlier focused on distinguishing MSA from dialect when the news source is known, making it straightforward to determine which of the Arabic dialects a sentence is written in (once the sentence is determined to be dialectal). If the news source is not known, we do not have the luxury of such a strong prior on the specific Arabic dialect. It is therefore important to evaluate our approach in a multi-way classifiation scenario, where the class set is expanded from {MSA, dialect} to {MSA, LEV, GLF, EGY}.

Under this classification set-up, the classification accuracy decreases from 85.7% to 81.0%.¹⁶ The drop in performance is not at all surprising, since four-way classification is inherently more difficult than two-way classification. (Note that the classifier is trained on exactly the same training data in both scenarios, but with more fine-grained dialectal labels in the four-way set-up.)

Table 7 is the classifier's confusion matrix for this four-way set-up, illustrating when the classifier tends to make mistakes. We note here that most classification errors on dialectal sentences occur when these sentences are mislabeled as being MSA, not when they are misidentified as being in some other incorrect dialect. In other words, dialect→dialect confusion constitutes a smaller proportion of errors than dialect→MSA confusion. Indeed, if we consider a three-way classification setup on dialectal sentences alone (LEV vs. GLF vs. EGY), the classifier's accuracy rate shoots up to 88.4%. This is a higher accuracy rate than for the general two-way MSA vs. dialect classification (85.7%), despite involving more classes (three instead of two), and being trained on less data (0.77M words instead of 1.78M words). This indicates that the dialects deviate from MSA in various ways, and therefore distinguishing dialects from each other can be done even more effectively than distinguishing dialect from MSA.

Figure 13 lists, for each news source, the word types with the highest and lowest dialectness factor. The most dialectal words tend to be function words, and they also tend to be strong indicators of dialect, judging by their very high DF. On the other hand, the MSA word group contains several content words, relating mainly to politics and religion.

One must also take into account the actual frequency of a word, as DF only captures relative frequencies of dialect/MSA, but does not capture how often the word occurs in the first place. Figure 14 plots both measures for the words of Al-Ghad newspaper. The plot illustrates which words are most important to the classifier: the words that are farthest away from the point of origin, along both dimensions.

As for letter-based features, many of the longer ones (e.g., 5-graph features) are essentially the same words important to the unigram word model. The letter-based models are, however, able to capture some linguistic phenomenon that the word model is unable to: the suffixes +š (not in Levantine) and +wn (plural conjugation in Gulf), and the prefixes H+ (will in Egyptian), bt+ (present tense conjugation in Levantine and Egyptian), and y+ (present tense conjugation in Gulf).

Figure 15 sheds some light on why even the unigraph model outperforms the baselines. It picks up on subtle properties of the MSA writing style that are lacking when using dialect. Namely, there is closer attention to following hamza rules (distinguishing A, Â, and ѡ from each other, rather than mapping them all to A), and better adherence to (properly) using + instead of +h at the end of many words. There is also a higher tendency to use words containing the letters that are most susceptible to being transformed when pronounced dialectally: (usually pronounced as z), Ď (pronounced as D), and θ (pronounced as t).

On the topic of spelling variation, one might wonder if nomalizing the Arabic text before training language models might enhance coverage and therefore improve performance. For instance, would it help to map all forms of the alef hamza to a single letter, and all instances of to h, and so on? Our pilot experiments indicated that such normalization tends to slightly but consistently hurt performance, so we opted to leave the Arabic text as is. The only type of preprocessing we performed was more on the “cleanup” side of things rather than computationally motivated normalization, such as proper conversion of HTML entities (e.g., " to ”) and mapping Eastern Arabic numerals to their European equivalents.

We conducted a large-scale Web crawl to gather Arabic text from the on-line versions of newspapers from various Arabic-speaking countries. The first batch contained 319 on-line Arabic-language newspapers published in 24 countries. This list was compiled from http://newspapermap.com/ and http://www.onlinenewspapers.com/, which are Web sites that show the location and language of newspapers published around the world. The list contained 55 newspapers from Lebanon, 42 from Egypt, 40 from Saudi Arabia, 26 from Yemen, 26 from Iraq, 18 from Kuwait, 17 from Morocco, 15 from Algeria, 12 from Jordan, and 10 from Syria. The data were gathered from July–Sept 2011.

We mirrored the 319 Web sites using wget, resulting in 20 million individual files and directories. We identified 3,485,241 files that were likely to contain text by selecting the extensions htm, html, cmff, asp, pdf, rtf, doc, and docx. We converted these files to text using xpdf's pdftotext for PDFs and Apple's textutil for HTML and Doc files. When concatenated together, the text files contained 438,940,861 lines (3,452,404,197 words). We performed de-duplication to remove identical lines, after which 18,219,348 lines (1,393,010,506 words) remained.

We used the dialect-annotated data to train a language model for each of the four Arabic varieties (MSA, LEV, GLF, EGY), as described in the previous section. We used these models to classify the crawled data, assigning a given sentence the label corresponding to the language model under which that sentence received the highest score. Table 8 gives the resulting label breakdown. We see that the overwhelming majority of the sentences are classified as MSA, which comes as no surprise, given the prevalence of MSA in the newspaper genre. Figure 16 shows some sentences that were given non-MSA labels by our classifier.

Habash et al. (2008) presented annotation guidelines for the identification of dialectal content in Arabic content, paying particular attention to cases of code switching. They present pilot annotation results on a small set of around 1,600 Arabic sentences (19k words), with both sentence- and word-level dialectness annotations.

The Cross Lingual Arabic Blog Alerts (COLABA) project (Diab et al. 2010) is another large-scale effort to create dialectal Arabic resources (and tools). They too focus on on-line sources such as blogs and forums, and use information retrieval tasks to measure their ability to properly process dialectal Arabic content. The COLABA project demonstrates the importance of using dialectal content when training and designing tools that deal with dialectal Arabic, and deal quite extensively with resource creation and data harvesting for dialectal Arabic.

Chiang et al. (2006) investigate building a parser for Levantine Arabic, without using any significant amount of dialectal data. They utilize an available Levantine–MSA lexicon, but no parses of Levantine sentences. Their work illustrates the difficulty of adapting MSA resources for use in a dialectal domain.

Zbib et al. (2012) show that incorporating dialect training data into a statistical machine translation system vastly improves the quality of the translation of dialect sentences when compared to a system trained solely on an MSA-English parallel corpus. When translating Egyptian and Levantine test sets, a dialect Arabic MT system outperforms a Modern Standard Arabic MT system trained on a 150 million word Arabic–English parallel corpus—over 100 times the amount of data as their dialect parallel corpus.

As far as we can tell, no prior dialect identification work exists that is applied to Arabic text. However, Lei and Hansen (2011) and Biadsy, Hirschberg, and Habash (2009) investigate Arabic dialect identification in the speech domain. Lei and Hansen (2011) build Gaussian mixture models to identify the same three dialects we consider, and are able to achieve an accuracy rate of 71.7% using about 10 hours of speech data for training.

Biadsy, Hirschberg, and Habash (2009) utilize a much larger data set (170 hours of speech data) and take a phone recognition and language modeling approach (Zissman 1996). In a four-way classification task (with Iraqi as a fourth dialect), they achieve a 78.5% accuracy rate. It must be noted that both works use speech data, and that dialect identification is done on the speaker level, not the sentence level as we do.

Social media, like reader commentary on on-line newspapers, is a rich source of dialectal Arabic that has previously not been studied in detail. We have harvested this type of resource to create a large data set of informal Arabic that is rich in dialectal content. We selected a large subset of this data set, and had the sentences in it manually annotated for dialect. We used the collected labels to train and evaluate automatic classifiers for dialect identification, and observed interesting linguistic aspects about the task and annotators' behavior. Using an approach based on language model scoring, we develop classifiers that significantly outperform baselines that use large amounts of MSA data, and we approach the accuracy rates exhibited by human annotators.

In addition to n-gram features, one could imagine benefiting from morphological features of the Arabic text, by incorporating analyses given by automatic analyzers such as BAMA (Buckwalter 2004), MAGEAD (Habash and Rambow 2006), ADAM (Salloum and Habash 2011), or CALIMA (Habash, Eskander, and Hawwari 2012). Although the difference between our presented approach and human annotators was found to be relatively small, incorporating additional linguistically motivated features might be pivotal in bridging that final gap.

In future annotation efforts, we hope to solicit more detailed labels about dialectal content, such as specific annotation for why a certain sentence is dialectal and not MSA: Is it due to structural differences, dialectal terms, and so forth? We also hope to expand beyond the three dialects discussed in this article, by including sources from a larger number of countries.

Given the recent political unrest in the Middle East (2011), another rich source of dialectal Arabic are Twitter posts (e.g., with the #Egypt tag) and discussions on various political Facebook groups. Here again, given the topic at hand and the individualistic nature of the posts, they are very likely to contain a high degree of dialectal data.

The Arabic transliteration scheme used in the article is the Habash-Soudi-Buckwalter transliteration (HSBT) mapping (Habash, Soudi, and Buckwalter 2007), which extends the scheme designed by Buckwalter in the 1990s (Buckwalter 2002). Buckwalter's original scheme represents Arabic orthography by designating a single, distinct ASCII character for each Arabic letter. HSBT uses some non-ASCII characters for better readibility, but maintains the distinct 1-to-1 mapping.

Figure 17 lists the character mapping used in HSBT. We divide the list into four sections: vowels, forms of the hamzah (glottal stop), consonants, and pharyngealized consonants. Pharyngealized consonants are “thickened” versions of other, more familiar consonants, voiced such that the pharynx or epiglottis is constricted during the articulation of the sound. Those consonants are present in very few languages and are therefore likely to be unfamiliar to most readers, which is why we place them in a separate section—there is no real distinction in Arabic between them and other consonants.

HSBT also allows for the expression of short vowels and other Arabic diacritics, but because those diacritics are only rarely expressed in written (and typed) form, we omit them for brevity.

This research was supported in part by the DARPA GALE program under contract no. HR0011-06-2-0001, the DARPA BOLT program contract no. HR0011-12-C-0014, the EuroMatrixPlus project funded by the European Commission (7th Framework Programme), the Human Language Technology Center of Excellence, and by gifts from Google and Microsoft. The views and findings are the authors' alone. They do not reflect the official policies or positions of the Department of Defense or the U.S. Government. The authors would like to thank the anonymous reviewers for their extremely valuable comments on earlier drafts of this article, and for suggesting future work ideas.