Revisiting Common Assumptions about Arabic Dialects in NLP

Amr Keleg; Sharon Goldwater; Walid Magdy

REVIEW 3 major objections 4 minor 12 references

Revisiting Common Assumptions about Arabic Dialects in NLP

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Four assumptions about Arabic dialects oversimplify reality.

desk verdict A careful empirical check on four common Arabic-dialect assumptions; the qualitative finding of substantial dialect overlap is likely right, but the exact 56% figure needs uncertainty quantification before I'd trust it. read the letter →

arxiv 2505.21816 v1 pith:IWRJJ2NH submitted 2025-05-27 cs.CL

Amr Keleg , Sharon Goldwater , Walid Magdy This is my paper

classification cs.CL

keywords Arabicdialectsdialectidentificationmulti-labelclassificationALDidialectallexicalcuesannotatoragreementoverlap

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tests four assumptions common in Arabic natural language processing: that a dialectal sentence is usually valid in only one regional dialect, that only short sentences are ambiguous, that curated lexical cues identify a sentence's dialect, and that speakers of different dialects rate dialectness consistently. Using a multi-label dataset of 978 dialectal Arabic sentences judged by speakers from 11 countries, the paper finds that over 56% of sentences are valid in more than one regional dialect, and that a sentence's level of dialectness (ALDi) predicts multi-dialect validity better than its length. It also finds that published lists of distinctive dialect words have low recall and imperfect distinctiveness, and that ALDi ratings systematically differ between annotator regions. The paper concludes that Arabic dialect identification should be treated as a multi-label task at both the regional and country level, and that these assumptions may be limiting progress in Arabic NLP.

What carries the argument

The central object is the MLADI dataset, an extension of the NADI 2024 corpus in which 978 dialectal Arabic sentences carry validity labels for 11 country-level dialects, each assigned by three native speakers, plus ALDi ratings of how far each sentence is from Modern Standard Arabic on a 0-to-1 scale. The analysis machinery is comparatively simple: collapse country labels to five regional groupings, count valid labels per sentence, correlate validity counts with sentence length and ALDi, score published lexical cue lists by precision, distinctiveness, and recall, and run mean-difference and permutation tests on ALDi ratings across annotator regions. These measures are what turn the four assumptions into testable quantities.

What would settle it

Re-annotate the same 978 sentences with many more annotators per country and recompute the multi-region overlap rate on sentences where all annotators agree; if the 56% figure falls to what annotator noise alone would predict, the central overlap claim fails.

Watch

Extended reading notes

Core claim

The paper claims that four assumptions common in Arabic NLP oversimplify the dialect landscape. First, dialectal Arabic sentences are not usually confined to one regional dialect: 544 of 978 sentences (about 56%) were judged valid in at least two of five regional groupings, and about 12% were valid in all five; only 44% were single-region, and at the country level only 25% received a single label. Second, ambiguity is not mostly a short-sentence phenomenon: sentence length correlates only weakly with the number of valid dialects (Spearman $\rho = -0.28$), while the sentence's ALDi, a 0-to-1 measure of divergence from Modern Standard Arabic, correlates more strongly ($\rho = -0.52$). Third, lexical cue lists from DART and DIAL2MSA have high precision but low recall and imperfect distinctiveness, with some supposedly distinctive cues being false friends or shared across regions. Fourth, speakers from different regions rate the same sentences' dialectness differently, with statistically significant mean differences (e.g., Maghreb annotators rate shared sentences less dialectal than Levantine annotators). These four results together support the paper's conclusion that single-label framing and cue-based curation are holding back Arabic NLP.

Load-bearing premise

The validity labels, set by majority vote of only three annotators per country, reflect actual dialect overlap rather than chance agreement, but the paper reports only moderate inter-annotator agreement and cannot separate dialect effects from annotator effects at the country level.

Editorial extensions

If this is right

Dialect identification should be modeled as multi-label classification at both regional and country levels, since a single-label oracle on this data can reach only about 63% expected accuracy.
Sentence length should not be used as a proxy for dialect ambiguity; ALDi is a better predictor and could inform multi-label ADI predictions and annotation routing.
Datasets built by matching curated lexical cues need additional validation, because the tested cue lists have low recall and imperfect distinctiveness, and geolocation filtering can raise precision but lowers recall.
ALDi gold scores should not be treated as dialect-agnostic, since annotator region can shift mean ratings by up to about one third, the distance between consecutive ALDi levels.
Researchers aggregating ALDi ratings across annotators should account for regional differences, and future datasets may need dialect-aware rating normalization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the observed overlap rates hold for additional dialects and for other scripts such as Arabizi, the multi-label treatment could extend well beyond the 11 countries studied here, and the same logic likely applies to other dialect continua such as English, French, or Spanish.
The annotator-region differences reported for ALDi suggest that dialectness is partly perceptual, so averaging ratings across regions may encode a dominant-region bias rather than a property of the text alone.
The false-friend finding implies that lexical matching alone, even with high precision, can mislabel sentences whose words carry different meanings in different dialects, so context-aware or sense-aware methods would be needed for robust dialect identification.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Worth your time. This paper tests four widely held assumptions about Arabic dialects — that sentences usually belong to one regional dialect, that only short sentences are ambiguous, that curated lexical cues are distinctive, and that speakers of different dialects rate dialectness similarly. As far as I know, assumptions 2–4 have not been systematically tested before, and the regional-level overlap analysis (assumption 1) is a useful extension of earlier country-level work. The release of the extended labels and code is real evidence and makes it easy for others to build on.

The authors do honest, careful work. They run external benchmarks (DART, DIAL2MSA, TWT15DA), report inter-annotator agreement, and explicitly list limitations: annotators skew young and educated, no Arabizi, sub-dialects underrepresented, and — in section 4.4 — that country-level annotator and dialect effects cannot be separated with three annotators per country. The lexical-cue analysis is concrete and shows that supposedly distinctive words often are not, even when the lists were validated by native speakers. That alone is useful for dataset curation.

The soft spot is the headline statistic. The 56% multi-region figure is built from majority-voted validity labels with Fleiss kappa around 0.56–0.62 for the newly collected Jordan/Saudi labels, and the original NADI labels reportedly similar. With three annotators per country, majority voting reduces but does not eliminate noise, and there is no confidence interval. The stress-test note is right that annotation noise could inflate the overlap rate. I would push back only on the implied severity: even a conservative noise-corrected estimate would almost certainly leave substantial overlap — the qualitative conclusion is not an artifact. The same caveat touches the length-vs-ALDi comparison (both correlate with the same noisy dependent variable) and the precision/distinctiveness scores in Table 1. None of this undermines the main direction, but the exact numbers should be treated as provisional.

The citation pattern is fine. There is natural self-citation to the ALDi metric and NADI 2024, but the analyses are empirical and the paper is transparent about that reliance.

For whom: Arabic NLP researchers, especially anyone building or using dialect identification datasets or lexical cue lists. A broader typology audience gets the methodological lesson that single-label dialect identification may understate ambiguity.

Send it to peer review. The right revisions are modest: bootstrap confidence intervals for the overlap percentages and correlations, maybe a noise-correction or sensitivity analysis for the 56% figure. The paper is already solid and will get better with those.

Referee Report

3 major / 4 minor

Summary. The paper identifies four assumptions about Arabic dialects that are widely adopted in Arabic NLP: (Asm. 1) a dialectal sentence is usually valid in only one regional dialect, (Asm. 2) only short sentences are ambiguous, (Asm. 3) curated lists of distinctive lexical cues are reliable dialect indicators, and (Asm. 4) ALDi ratings do not depend on the annotator's native dialect. To test these, the authors extend the NADI 2024 dataset with validity and ALDi annotations from Jordan and Saudi Arabia, yielding 978 dialectal sentences labeled for validity in 11 country-level dialects. They report that 56% of these sentences are valid in multiple regional dialects, that sentence length correlates only weakly with the number of valid dialects while ALDi correlates moderately, that lexical cue lists have low recall and imperfect distinctiveness, and that ALDi ratings differ systematically across annotator regions. The paper concludes that ADI should be treated as a multi-label task and that existing assumptions oversimplify reality.

Significance. If the findings hold, this paper provides valuable quantitative evidence for rethinking how Arabic dialect identification is framed, and its released dataset and code are useful resources for the community. The analysis is transparent: the authors state limitations explicitly and make the annotation extension available. The main risk is that the headline statistics—especially the 56% multi-region overlap—are derived from majority-voted labels with only moderate inter-annotator agreement, and the paper does not quantify how sensitive those statistics are to annotation noise. This is a load-bearing concern because the central recommendation to model ADI as a multi-label task rests on those counts.

major comments (3)

[§4.1, Figure 2, Appendix B] The headline figure that 56% of sentences are valid in multiple regional dialects is computed from majority-voted country-level validity labels, but the newly collected labels have only moderate inter-annotator agreement (Fleiss kappa 0.56–0.62, Appendix B), and the authors state that the original NADI 2024 labels have similar agreement. With three annotators per country, a single false-positive annotation can become a majority label and inflate the number of regions in which a sentence is counted as valid. The paper reports no bootstrap confidence intervals, no noise-corrected estimate, and no sensitivity analysis for the 56% figure (or the 25% country-level single-label figure in Appendix C). This matters because the paper's main recommendation—ADI should be modeled as a multi-label task—rests directly on these counts. The caveat in §4.4 that dialect and annotator effects cannot be separated at the country level applies with equal force to the validity labels. I ask for an uncertainty analysis, such as bootstrapping over annotators or fitting a simple measurement-error model, to show that the qualitative conclusion is robust to plausible annotation noise.
[§4.2, Figure 5] The claim that ALDi is a better predictor of the number of valid dialects than sentence length is based on comparing Spearman correlations of -0.28 (sentence length) and -0.52 (ALDi), but the paper does not report confidence intervals for these correlations or test whether the difference is statistically significant. Moreover, both the dependent variable (number of valid dialects from majority voting) and the ALDi scores come from the same annotators, so the stronger ALDi correlation may partly reflect annotator-level response bias rather than a linguistic property. Please report bootstrap or other uncertainty intervals, test the difference between the correlations, and ideally show that the ALDi correlation remains after accounting for the measurement error in the validity labels.
[§4.1 and §3] The regional overlap statistics are computed from an incomplete set of country-level dialects: the Levant region lacks Lebanon, the Gulf region lacks the UAE, Kuwait, Qatar, Bahrain, and Oman, and 72 DA sentences that were not valid in any of the 11 annotated countries are discarded from the analysis. A sentence valid only in an unannotated country would be treated as not valid in its region and could even be dropped entirely, which biases the measured overlap counts. The paper's discussion in the Limitations is honest about this, but the headline percentages in §4.1 should be explicitly qualified as conditional on the 11 annotated countries, and a sensitivity analysis should show how the results could change if additional countries were included.

minor comments (4)

[§4.1, Figure 4] The notation 'GL/¬GL' in the Figure 4 caption is not defined in the main text; please clarify what the cells represent so the figure is self-contained.
[§2.1] The sentence 'Glottolog and Ethnologue recognize 37 and 28 Arabic dialects, respectively' lacks a citation; these counts should be attributed to the respective sources.
[Appendix A] Equation 1 assumes that when a sentence has multiple valid labels, both the gold-standard label and the oracle model's prediction are chosen uniformly at random among the valid labels; this assumption should be stated explicitly before presenting the 63.06% upper bound.
[§4.3, Table 1] The statement that precision is 'generally high (yet not perfect), except for the cues of Gulf Arabic' is not quite accurate for DART, where Egyptian Arabic also has precision 0.60; please rephrase to acknowledge this additional exception.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the empirical claims are measured from human annotations and external benchmarks, not constructed from their inputs.

full rationale

The paper's central claims are empirical measurements over a manually annotated multi-label dataset and external cue lists, not consequences of its own definitions. Asm. 1 aggregates 11 country-level majority-vote validity labels into five regional labels using the Baimukan et al. (2022) grouping and reports the histogram; the >56% overlap figure is a direct count, not an identity derived from the assumption being tested. Asm. 2 computes Spearman correlations between sentence length or ALDi and the number of valid dialects; neither variable is defined in terms of the other, and the automatic Sentence-ALDi replication (-0.45) corroborates the manual result, so there is no fitted-input prediction. Asm. 3 evaluates externally published lists (DART, DIAL2MSA, TWT15DA) by exact token matching against the validity labels; precision, distinctiveness, and recall are computed statistics, not parameters fitted to the claim. Asm. 4 is a direct comparison of mean ALDi ratings across annotator groups with a permutation test. The self-citations (ALDi, NADI 2024) are uses of prior datasets and metrics as tools, and they do not supply the conclusions. The acknowledged limitations—three annotators per country, moderate Fleiss kappa 0.56–0.62, and the §4.4 caveat that dialect and annotator effects cannot be conclusively separated—are measurement-reliability concerns, not circularity. No equation or fitted parameter reduces to the paper's own inputs.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No fitted parameters or novel postulated entities appear in the paper. The analysis depends on annotation assumptions and the adopted regional grouping, all of which are disclosed or inherited from prior work.

assumptions (5)

domain assumption Majority vote of three annotators per country yields reliable validity labels.
Used throughout §4.1-§4.4 to define whether a sentence is valid in a dialect; inter-annotator agreement is moderate (Fleiss kappa 0.56-0.62, Appendix B), so noise may affect overlap rates.
domain assumption The five-region grouping of Baimukan et al. (2022) is a valid way to aggregate the 11 country-level labels.
Used in §4.1 to compute regional overlap; the paper itself questions the Gulf of Aden grouping, indicating the grouping is not self-evident.
domain assumption ALDi ratings L0-L3 map to numeric values 0, 1/3, 2/3, 1 and can be averaged across annotators.
Used in §4.2 and §4.4; the interval-scale assumption is inherited from NADI 2024 and Keleg et al. (2023).
domain assumption Exact token matching after orthographic normalization detects the intended lexical cues.
Used in §4.3 to compute precision, distinctiveness, and recall; may miss morphological variants or match false friends.
domain assumption The 978 geolocated tweets are representative of written dialectal Arabic.
The paper acknowledges in the Limitations section that Arabizi, some sub-dialects, and certain demographics are not covered.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Revisiting Common Assumptions about Arabic Dialects in NLP." pith.science (2026). https://pith.science/paper/IWRJJ2NH

@misc{pith2026250521816,
  author       = {Pith},
  title        = {Pith review of: Revisiting Common Assumptions about Arabic Dialects in NLP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IWRJJ2NH}},
  note         = {Machine review of arXiv:2505.21816}
}

read the original abstract

Arabic has diverse dialects, where one dialect can be substantially different from the others. In the NLP literature, some assumptions about these dialects are widely adopted (e.g., ``Arabic dialects can be grouped into distinguishable regional dialects") and are manifested in different computational tasks such as Arabic Dialect Identification (ADI). However, these assumptions are not quantitatively verified. We identify four of these assumptions and examine them by extending and analyzing a multi-label dataset, where the validity of each sentence in 11 different country-level dialects is manually assessed by speakers of these dialects. Our analysis indicates that the four assumptions oversimplify reality, and some of them are not always accurate. This in turn might be hindering further progress in different Arabic NLP tasks.

Figures

Figures reproduced from arXiv: 2505.21816 by the authors.

**Figure 1.** A map of the Arab world. The black dots indicate the provinces/cities from which the annotators originate. Regional dialects (Maghreb, Nile Basin, Levant, Gulf, Gulf of Aden) are encoded as different colors according to the groupings of Baimukan et al. (2022). lines for rating the level of dialectness of sentences (Habash et al., 2008; Zaidan and Callison-Burch, 2011). After that, however, the concept was ignored … view at source ↗

**Figure 2.** The histogram of the number of valid dialects [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The total number of valid regional dialects [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: The distribution of the 2-region, 3-region, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: The distribution of the sentences (log scale) and the number of valid country-level dialects according to [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: (Left) The number of valid samples per country (with countries ordered such that same-region ones are consecutive). (Right) Mean difference (MD) of row country’s (r) and column country’s (c) ALDi scores, for the Nrc sentences valid in both (Nrc is shown as the bottom n…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

[10]

InProceedings of the Fourth Arabic Natural Language Processing Workshop, pages 285–289, Florence, Italy

Team JUST at the MADAR shared task on Ara- bic fine-grained dialect identification. InProceedings of the Fourth Arabic Natural Language Processing Workshop, pages 285–289, Florence, Italy. Associa- tion for Computational Linguistics. Jörg Tiedemann and Nikola Ljubeši´c. 2012. Efficient discrimination between closely related languages. In Proceedings of CO...

work page 2012
[2012]

words used in one or more Arabic dialects but never used in MSA, thereby distinguishing Arabic dialects from MSA

Machine translation of Arabic dialects. InPro- ceedings of the 2012 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 49–59, Montréal, Canada. Association for Computa- tional Linguistics. A Was Regional-level ADI Already Solved? When framing a multi-label task as a single-label ...

work page 2023
[2014]

In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC‘14), pages 1240–1245, Reykjavik, Iceland

A multidialectal parallel corpus of Arabic. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC‘14), pages 1240–1245, Reykjavik, Iceland. European Lan- guage Resources Association (ELRA). Houda Bouamor, Sabit Hassan, and Nizar Habash. 2019. The MADAR shared task on Arabic fine-grained di- alect identification. I...

work page 2019
[2015]

InCurrent Trends in Web Engi- neering, pages 3–14, Cham

Constructing linguistic resources for the tunisian dialect using textual user-generated contents on the social web. InCurrent Trends in Web Engi- neering, pages 3–14, Cham. Springer International Publishing. Wajdi Zaghouani and Anis Charfi. 2018. Arap-tweet: A large multi-dialect Twitter corpus for gender, age and language variety identification. InProcee...

work page 2018
[2016]

InProceedings of the Third Work- shop on NLP for Similar Languages, Varieties and Dialects (VarDial3), pages 1–14, Osaka, Japan

Discriminating between similar languages and Arabic dialect identification: A report on the third DSL shared task. InProceedings of the Third Work- shop on NLP for Similar Languages, Varieties and Dialects (VarDial3), pages 1–14, Osaka, Japan. The COLING 2016 Organizing Committee. Karen McNeil. 2018.Tunisian Arabic Corpus: Creat- ing a Written Corpus of a...

work page 2016
[2018]

InProceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan

Arabic dialect identification in the context of bivalency and code-switching. InProceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). Abdellah El Mekki, Abdelkader El Mahdaouy, Kabil Es- sefar, Nabil El Mamoun, Ismail Berrada, and Ahmed Khoums...

work page 2018
[2019]

InPro- ceedings of the Fourth Arabic Natural Language Pro- cessing Workshop, pages 214–218, Florence, Italy

Simple but not naïve: Fine-grained Arabic dialect identification using only n-grams. InPro- ceedings of the Fourth Arabic Natural Language Pro- cessing Workshop, pages 214–218, Florence, Italy. Association for Computational Linguistics. Kamel Gaanoun, Abdou Mohamed Naira, Anass Allak, and Imade Benelallam. 2024. DarijaBERT: a step forward in NLP for the w...

work page 2024
[2021]

Automatic Arabic Dialect Identification Systems for Written Texts: A Survey

Overview of the WANLP 2021 shared task on sarcasm and sentiment detection in Arabic. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 296–305, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. Kathrein Abu Kwaik and Motaz Saad. 2019. ArbDialec- tID at MADAR shared task 1: Language modelling and ensemble lear...

work page Pith review arXiv 2021

Show all 12 references

[2022]

InProceedings of the Thirteenth Language Resources and Evaluation Con- ference, pages 4586–4596, Marseille, France

Hierarchical aggregation of dialectal data for Arabic dialect identification. InProceedings of the Thirteenth Language Resources and Evaluation Con- ference, pages 4586–4596, Marseille, France. Euro- pean Language Resources Association. Timothy Baldwin and Marco Lui. 2010. Lan...

2010
[2023]

InTenth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2023), pages 142–151, Dubrovnik, Croatia

Dialect and variant identification as a multi- label classification task: A proposal based on near- duplicate analysis. InTenth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2023), pages 142–151, Dubrovnik, Croatia. Associa- tion for Computational Ling...

2023 arXiv
[2024]

Association for Computational Lin- guistics

How well do tweets represent sub-dialects of Egyptian Arabic? InProceedings of the Eleventh Workshop on NLP for Similar Languages, Varieties, and Dialects (VarDial 2024), pages 41–55, Mexico City, Mexico. Association for Computational Lin- guistics. Hamdy Mubarak. 2018. Dial2M...

2024 arXiv
[2025]

InProceedings of the 12th Workshop on NLP for Similar Languages, Varieties and Dialects, pages 168–181, Abu Dhabi, UAE

Common ground, diverse roots: The diffi- culty of classifying common examples in Spanish varieties. InProceedings of the 12th Workshop on NLP for Similar Languages, Varieties and Dialects, pages 168–181, Abu Dhabi, UAE. Association for Computational Linguistics. Leena Lulu and...

2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [10]

InProceedings of the Fourth Arabic Natural Language Processing Workshop, pages 285–289, Florence, Italy

Team JUST at the MADAR shared task on Ara- bic fine-grained dialect identification. InProceedings of the Fourth Arabic Natural Language Processing Workshop, pages 285–289, Florence, Italy. Associa- tion for Computational Linguistics. Jörg Tiedemann and Nikola Ljubeši´c. 2012. Efficient discrimination between closely related languages. In Proceedings of CO...

work page 2012

[2] [2012]

words used in one or more Arabic dialects but never used in MSA, thereby distinguishing Arabic dialects from MSA

Machine translation of Arabic dialects. InPro- ceedings of the 2012 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 49–59, Montréal, Canada. Association for Computa- tional Linguistics. A Was Regional-level ADI Already Solved? When framing a multi-label task as a single-label ...

work page 2023

[3] [2014]

In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC‘14), pages 1240–1245, Reykjavik, Iceland

A multidialectal parallel corpus of Arabic. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC‘14), pages 1240–1245, Reykjavik, Iceland. European Lan- guage Resources Association (ELRA). Houda Bouamor, Sabit Hassan, and Nizar Habash. 2019. The MADAR shared task on Arabic fine-grained di- alect identification. I...

work page 2019

[4] [2015]

InCurrent Trends in Web Engi- neering, pages 3–14, Cham

Constructing linguistic resources for the tunisian dialect using textual user-generated contents on the social web. InCurrent Trends in Web Engi- neering, pages 3–14, Cham. Springer International Publishing. Wajdi Zaghouani and Anis Charfi. 2018. Arap-tweet: A large multi-dialect Twitter corpus for gender, age and language variety identification. InProcee...

work page 2018

[5] [2016]

InProceedings of the Third Work- shop on NLP for Similar Languages, Varieties and Dialects (VarDial3), pages 1–14, Osaka, Japan

Discriminating between similar languages and Arabic dialect identification: A report on the third DSL shared task. InProceedings of the Third Work- shop on NLP for Similar Languages, Varieties and Dialects (VarDial3), pages 1–14, Osaka, Japan. The COLING 2016 Organizing Committee. Karen McNeil. 2018.Tunisian Arabic Corpus: Creat- ing a Written Corpus of a...

work page 2016

[6] [2018]

InProceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan

Arabic dialect identification in the context of bivalency and code-switching. InProceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). Abdellah El Mekki, Abdelkader El Mahdaouy, Kabil Es- sefar, Nabil El Mamoun, Ismail Berrada, and Ahmed Khoums...

work page 2018

[7] [2019]

InPro- ceedings of the Fourth Arabic Natural Language Pro- cessing Workshop, pages 214–218, Florence, Italy

Simple but not naïve: Fine-grained Arabic dialect identification using only n-grams. InPro- ceedings of the Fourth Arabic Natural Language Pro- cessing Workshop, pages 214–218, Florence, Italy. Association for Computational Linguistics. Kamel Gaanoun, Abdou Mohamed Naira, Anass Allak, and Imade Benelallam. 2024. DarijaBERT: a step forward in NLP for the w...

work page 2024

[8] [2021]

Automatic Arabic Dialect Identification Systems for Written Texts: A Survey

Overview of the WANLP 2021 shared task on sarcasm and sentiment detection in Arabic. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 296–305, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. Kathrein Abu Kwaik and Motaz Saad. 2019. ArbDialec- tID at MADAR shared task 1: Language modelling and ensemble lear...

work page Pith review arXiv 2021