Paper deep dive

BARRAC: Adaptation of an English Aspect-based Sentiment Analysis Approach for Classification Tasks in Arabic Dialects

Ali Almutairi, Gelareh Mohammadi, Imran Razzak, Aditya Joshi

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 87

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 10/4/2026, 3:12:47 AM

Summary

The paper introduces BARRAC, an adaptation of the English aspect-based sentiment analysis framework DS²-ABSA for Arabic dialect classification tasks. BARRAC replaces consumer-review attributes with Arabic linguistic devices and markers, and replaces noisy self-training with a two-stage intermediate training strategy. Evaluated on five Arabic dialect datasets, BARRAC achieves a mean macro-F1 of 63.93%, outperforming state-of-the-art few-label methods and GPT-4o on four out of five tasks.

Entities (11)

BARRAC · modelframework · 98%Ar-Dialects · dataset · 95%Ar-Sarcasm · dataset · 95%Ar-Sentiment · dataset · 95%DART · dataset · 95%DS²-ABSA · modelframework · 95%Sa'7r · dataset · 95%GPT-4o · model · 92%ARBERTv2 · model · 90%Cluster&Tune · modelmethod · 85%IDoFew · modelmethod · 85%

Relation Signals (10)

BARRAC → achievesmetric → 63.93% macro-F1

confidence 98% · BARRAC achieves a mean macro-F1 of 63.93%

BARRAC → adaptedfrom → DS²-ABSA

confidence 95% · We adapt an English aspect-based sentiment analysis framework... and present the adaptation as BARRAC

BARRAC → evaluatedon → Sa'7r

confidence 95% · Evaluated on five Arabic dialect datasets... Sa’7r

BARRAC → evaluatedon → DART

confidence 95% · Evaluated on five Arabic dialect datasets... DART

BARRAC → evaluatedon → Ar-Sentiment

confidence 95% · Evaluated on five Arabic dialect datasets... Ar-Sentiment

BARRAC → evaluatedon → Ar-Sarcasm

confidence 95% · Evaluated on five Arabic dialect datasets... Ar-Sarcasm

BARRAC → evaluatedon → Ar-Dialects

confidence 95% · Evaluated on five Arabic dialect datasets... Ar-Dialects

BARRAC → outperforms → GPT-4o

confidence 92% · outperforming GPT-4o on four out of five tasks

BARRAC → outperforms → Cluster&Tune

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:With the rapid growth of Arabic NLP, several models, datasets and benchmarks have been reported. This paper asks whether approaches developed for majority languages like English can be adapted to Arabic tasks. We adapt an English aspect-based sentiment analysis framework to Arabic classification tasks and present the adaptation as BARRAC: Brainstorming Alignment and Replaced Representation learning for ArabiC tasks. BARRAC replaces consumer-review attribute pools with Arabic linguistic devices and markers for dialectal sentiment, sarcasm, and dialect identification, and replaces noisy self-training with two-stage training. Evaluated on five Arabic dialect datasets, BARRAC achieves a mean macro-F1 of 63.93\%, outperforming the best few-label SOTA by 3\%, and outperforming GPT-4o on four out of five tasks. Error analysis provides insights into remaining challenges. These results demonstrate that adapting task-specific approaches is a promising direction for Arabic NLP alongside adapting models, datasets and benchmarks.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

84,404 characters extracted from source content.

Expand or collapse full text

BARRAC: Adaptation of an English Aspect-based Sentiment Analysis Approach for Classification Tasks in Arabic Dialects Ali Almutairi 1 , Gelareh Mohammadi 1 , Imran Razzak 2 , Aditya Joshi 1 , 1 University of New South Wales, Australia, 2 MBZUAI, UAE ali.almutairi@unsw.edu.au Abstract With the rapid growth of Arabic NLP, sev- eral models, datasets and benchmarks have been reported. This paper asks whether ap- proaches developed for majority languages like English can be adapted to Arabic tasks. We adapt an English aspect-based sentiment anal- ysis framework to Arabic classification tasks and present the adaptation as BARRAC: Brain- storming Alignment and Replaced Represen- tation learning for ArabiC tasks. BARRAC replaces consumer-review attribute pools with Arabic linguistic devices and markers for di- alectal sentiment, sarcasm, and dialect identi- fication, and replaces noisy self-training with two-stage training. Evaluated on five Arabic dialect datasets, BARRAC achieves a mean macro-F1 of 63.93%, outperforming the best few-label SOTA by 3%, and outperforming GPT-4o on four out of five tasks. Error analy- sis provides insights into remaining challenges. These results demonstrate that adapting task- specific approaches is a promising direction for Arabic NLP alongside adapting models, datasets and benchmarks. 1 Introduction Arabic is a Semitic language with over 1,400 years of written history (Belinkov et al.,2018) and one of the six oﬀicial languages of the United Na- tions, with more than 422 million native speak- ers ( Boudad et al.,2018). Since the early finite- state morphology ofBeesley(1996), Arabic NLP has produced dedicated pre-trained models (An- toun et al. ,2020;Abdul-Mageed et al.,2021), evaluation suites (Seelawi et al.,2021), recurring NADI dialect-identification shared tasks (Talafha et al.,2025;Abdul-Mageed et al.,2024,2023), and community datasets for sentiment and sarcasm ( Abu Farha and Magdy,2020;AlMazrua et al., 2022). Yet these resources remain limited in scale and diversity compared to English, especially for low-resource tasks such as fine-grained sarcasm detection and cross-dialect sentiment classification ( Abdhood et al.,2025). We argue this gap calls for a community-involved effort in which native speak- ers critically adapt and refine approaches first de- vised for English. We show how DS²-ABSA (Xu et al.,2025), an English aspect-based sentiment analysis frame- work, can be adapted to text classification tasks in Arabic, including low-resource dialects. Our adaptation is called BARRAC:Brainstorming Alignment andReplacedRepresentation learning forArabiCtasks. Existing responses to label scarcity, zero-shot transfer, SFT, and few-label methods, such as Cluster&Tune ( Shnarch et al., 2022), IDoFew (Alsuhaibani et al.,2024), and noisy self-training ( Xu et al.,2025), have not, to our knowledge, extended a complete English data-synthesis framework to Arabic sociolinguis- tic tasks. Benchmarking 14 baselines across dialec- tal sarcasm, sentiment, and dialect tasks with few- labels, our adaptation reaches a mean Macro-F1 of 63.93%, surpassing the best supervised baseline by 8.48%, SOTA few-label method by 3% and is com- parable to GPT-4o at a fraction of its parameter count, with consistent gains across all five datasets. Methodological contributions devised for English, we conclude, can be meaningfully adapted to serve other languages and communities. 2 Motivation Arabic NLP has grown rapidly, with dedicated venues such as the ArabicNLP conference re- flecting the field’s increasing maturity ( Abdhood et al.,2025). While acknowledging the focus on building datasets and benchmarks is valuable, we ask an alternative question: “how can an En- glish aspect-based sentiment analysis approach be adapted for classification tasks in Arabic?” DS²-ABSA is a dual-stream (key-point-driven and instance-driven) synthesis framework address- arXiv:2609.38820v1 [cs.CL] 30 Sep 2026 Brainstorming V ia GPT -4o Sentence, Label Attribute Prompting Synthetic Data Few- Labels Label Normalisation PL-FT Model FL-FT Model Seed Data Reconstruction Selective Preservation Combination & paraphrasing Sample pairs sampling Weights FlowData Flow Key-point-driven Data SynthesisStage1: Stage3: Stage2: Normalisation & intermediate training Instance-driven Data Synthesis Review Subjects Linguistic Device Linguistic Marker Prompt template: Final model output Figure 1: BARRAC framework. ing label scarcity in Aspect-Based Sentiment Analysis (ABSA). However, its attribute pools are grounded exclusively in consumer-review do- mains, and it presupposes tangible review sub- jects with discrete, reviewable aspects such as food quality, making it non-trivially applicable to tasks with a fundamentally different label space, output structure, and subject matter. Via our proposed BARRAC, we demonstrate how DS²-ABSA can be systematically adapted beyond its original scope: from tasks centred on tangible consumer entities to tasks centred on linguistic and sociolinguistic phe- nomena, specifically, sarcasm, dialects, and senti- ments. 3 Methodology 3.1 ABSA Method Given a consumer-review sentence, DS 2 -ABSA outputs one or more (aspect term, sentiment polar- ity) pairs over two domains, restaurant and laptop reviews. It addresses label scarcity by synthesising training data through two complementary streams. The key-point-driven stream builds samples from scratch: it samples an attribute tuple (review sub- ject, aspect category, aspect term, opinion term), prompts an LLM to generate a sentence satisfy- ing all four, and derives the label from the output. The instance-driven stream diversifies the seed cor- pus by combining, paraphrasing, and selectively re- constructing masked samples. A final refinement stage applies rule-based normalisation and noisy self-training, where a teacher model trained on the synthetic data re-annotates noisy instances. 3.2 Adaptation to BARRAC BARRAC (see Figure1) extends the components of DS²-ABSA as follows. 3.2.1 Attribute Linguistic Device A linguistic device is a coarse-grained rhetorical or communicative mechanism through which a speaker conveys meaning beyond the literal con- tent of an utterance. In the original framework, the aspect category serves an analogous role as a coarse-grained dimension of a product or ser- vice (e.g., screen resolution), brainstormed per do- main via GPT-4o. Such categories are inherently tied to the physical properties of reviewable en- tities and have no natural counterpart when the subject of analysis is a linguistic or sociolinguis- tic phenomenon rather than a product. We substi- tute aspect categories with linguistic devices con- ditioned on the target task: for sarcasm detection, we use coarse-grained rhetorical mechanisms such as mock_praise and verbal_irony; for dialect iden- tification, we use dialect regions (Egyptian, Gulf, Levantine, Maghrebi, and Modern Standard Ara- bic), conditioning the attribute hierarchy on the tar- get class directly; and for sentiment analysis, we use sentiment sub-categories representing domain- specific manifestations of each polarity class, such as frustration with price hikes for negative senti- ment in the economy domain. In all three cases, the substituted attribute preserves the coarse condition- ing role of the original aspect category, from which finer-grained attributes are subsequently derived, while grounding the hierarchy in the relevant lin- guistic phenomenon rather than product features. Please refer to Table 9for further details. 3.2.2 Attribute Linguistic Marker A linguistic marker is a fine-grained lexical or phrasal unit whose presence in an utterance re- liably signals the activation of a particular lin- guistic device, functioning as a surface-level cue from which deeper communicative intent can be in- ferred. In the original framework, the aspect term serves this role as a fine-grained noun phrase nam- ing the specific opinion target (e.g., decor), con- DatasetSplitTotal Classes Tok. Total Tok. Mean Tok. Med. Tok. Min Tok. Max DART Train– Real Labels10051,44714.4714.00328 Eval25553,64514.2913.00131 Test1,009514,30414.1813.00235 Synthetic Labels 429295 100278523.3624.00259 Ar-Dialects Train5,905586,63014.6715.00132 Real Labels10051,48214.8215.00328 Eval2,432536,10714.8515.00129 Test2,110531,27314.8215.00129 Synthetic Labels 429295 100278523.3624.00259 Sa’7r Train3,958257,33314.4910.00159 Real Labels 100 2 1,487 14.87 10.00 2 54 Eval3,458249,35214.2710.00159 Test1,186216,97614.3110.00259 Synthetic Labels 38542277028819.9921.00262 Ar-Sarcasm Train5,905286,81114.7015.00129 Real Labels10021,42114.2114.00225 Eval2,432235,98714.8015.00132 Test2,110231,27314.8215.00129 Synthetic Labels 38556285792022.2523.00253 Ar-Sentiment Train5,835385,98914.7415.00132 Real Labels10031,38413.8413.50229 Eval2,502336,84614.7315.00129 Test2,110331,27314.8215.00129 Synthetic Labels 485993 103914621.3822.00267 Table 1: Statistics of the train, real labels, evaluation, and test splits used in the five Arabic classification datasets. Tok. Total, Tok. Mean, Tok. Med., Tok. Min, and Tok. Max denote the total, mean, median, minimum, and maximum number of tokens per split, respectively. ditioned on the aspect category. Like aspect cat- egories, aspect terms presuppose the existence of a discrete, nameable physical entity, which is ab- sent in linguistic classification tasks. We substi- tute aspect terms with class-conditioned linguis- tic markers: for sarcasm, these are Arabic words and phrases that signal the active sarcasm device (e.g.,ل؇ ݿఈఃمfor mock_praise); for dialect iden- tification, these are phonological, morphological, and lexical features characteristic of each dialect region (e.g.,إل۬ دهfor Egyptian); and for sen- timent analysis, these are Arabic trigger phrases whose presence strongly signals a particular polar- ity (e.g.,ఈః༚ء ڣ؇ۋݷfor negative sentiment). In all three cases, markers are brainstormed con- ditioned on the class-level linguistic device, pre- serving the coarse-to-fine hierarchy of the original framework. Please refer to Table 9for further de- tails. Task InstructionThe original instruction asks the model to “label the sentence by extracting the aspect term(s) and identifying their corresponding sentiment polarity”. This is not suitable for single- label (predicting one class at a time) tasks, which have no multi-pair structure. We substitute a task- specific objective such as “identifying whether the text is sarcastic or not sarcastic”. The prompt struc- ture “Label the [sentence/text] by task instruction” is preserved verbatim, with only the objective re- placed in the attribute prompting. Additional de- tails are provided in Table 10for task instruction and Table11for attribute prompting. Domain ScopeDS²-ABSA restricts brainstorm- ing to two domains, restaurants and laptops, sourc- ing both the topic pool and category hierarchy. This suits ABSA but limits diversity and is unsuit- able for tasks spanning public discourse. We re- place it with 10 domains: economy, sports, health, environment, entertainment, transportation, gov- ernment services and infrastructure, technology and the internet, education, and food. Examples of Steps 1 and 2 in BARRAC are pro- vided in Tables 13and14, respectively. Training StrategyInstead of DS²-ABSA’s noisy self-training, which risks error accumulation across iterative re-annotation ( Zou and Caragea, 2023), we employ intermediate training (Al- suhaibani et al.,2024): The Pseudo-Label Fine- Tuning (PL-FT) model is trained using pseudo- Method/Dataset Ar-Sent. Ar-Sarc. Sa’7r DART Ar-Dial. Mean Without Fine-Tuning ArabicBERT3.2724.68 45.77 32.635.302.72 ARBERTv22.1811.41 47.56 32.92 10.77 7.88 AraBERT-Twitter4.1425.02 46.76 38.527.40 10.14 MARBERTv25.0520.17 50.40 51.74 10.77 6.03 SFT with 100 Labels BERT45.7249.34 59.91 10.12 16.02 36.22 DistilBERT43.7449.44 64.90 6.9816.02 36.22 ArabicBERT57.7758.71 64.48 25.56 26.33 46.57 ARBERTv260.2561.91 61.88 52.67 40.58 55.46 SOTA Few-label Methods with 100 Labels Cluster&Tune ∗ 57.9857.05 64.90 49.75 39.79 53.89 IDoFew ∗ 54.2260.82 63.54 53.24 32.38 52.84 IDoFew ∗ +58.7361.74 66.88 51.60 35.73 54.93 BARRAC (Ours) 65.6169.06 67.03 68.76 49.19 63.93 Table 2: Mac-F1 results on 5 Arabic datasets. “∗” indicates ARBERTv2 as SOTA backbone model and “+” indi- cates it is adapted to our training hyperparameters. labels, which will be obtained from systematic synthesis data pipelines. Once training is com- plete, we utilise the weights from the PL-FT model as initialisation weights for the subsequent Few- Labels Fine-Tuning (FT-FL) model on actual la- belled data (see Figure 1Stage 3). The synthetic corpusacts as task-adaptiveinitialisation, whilethe final decision boundary is refined solely through gold-standard supervision. 4 Experiment Setup We evaluate our adaptation to DS2-ABSA for text classification with few-labels across 5 pub- licly available Arabic datasets: Ar-Sentiment, Ar- Sarcasm, Ar-Dialects (Abu Farha and Magdy, 2020), Sa’7r (AlMazrua et al.,2022), and DART (Alsarsour et al.,2018); spanning three tasks, sar- casm detection, dialect identification, and senti- ment analysis. We fine-tune all models using 100 real labels to mimic low-resource settings. For syn- thetic data generation, we follow the hyperparame- ters reported in Xu et al.(2025) without modifica- tion, except for the LLM (GPT-4o) used for gener- ation. We release the synthetic data used in our exper- iments, We use ARBERTv2 as the backbone clas- sification model across all experiments for the few- label frameworks. We report the macro-F1 score, consistent with prior work on Arabic text classification under class imbalance. For all baselines, we use a unified con- figuration: the same number of epochs, batch size, learning rate, optimiser, and random seed across all experiments. All baselines are trained using ex- actly 100 balanced labelled examples. Table 3sum- marises the hyperparameters used for intermediate and final training across all three tasks. 4.1 Comparative Models We compare against general-purpose and Arabic- specific pre-trained language models fine-tuned on the available labelled data, namely BERT, Distil- BERT, ArabicBERT, and ARBERTv2. We further compare against Cluster&Tune, IDoFew and its Arabic-adapted variants, as well as the noisy self- training from the DS²-ABSA paper and its variants. BERT models:We have compared BARRAC with publicly available language models that include ParameterValue PL-FT model epochs1 FL-FT model epochs20 Learning rate (all other datasets)5×10 −5 Learning rate (Sa’7r)3×10 −5 Batch size32 OptimizerAdam Seed42 Table 3: Training hyperparameters used across all datasets and baselines. BERT (Devlin et al.,2019) and DistilBERT (Sanh et al.,2020). Arabic-BERT models:We have also used the Arabic BERT in our study. These models are: ArabicBERT ( Safaya et al.,2020), ARBERTv2 (Abdul-Mageed et al.,2021), AraBERT-Twitter (Antoun et al.,2020), and MARBERTv2 (Abdul- Mageed et al. ,2021). LLMs:We evaluate BARRAC against fine-tuned LLMs through Supervised Fine-Tuning (SFT), as they have shown effectiveness in managing intri- cate tasks. To achieve this, we select AllAM-7B ( Bari et al.,2025), Llama3.1-8B (Meta,2024), AceGPT (Zhu et al.,2025), Qwen2.5-72B-instruct ( Team,2024), and GPT-4o (OpenAI,2025) as rep- resentative models in our comparison. Intermediate models:In order to assess the ef- ficacy of our intermediate stage, we conducted a comparison of BARRAC’s performance with that of Cluster & Tune ( Shnarch et al.,2022) and ID- oFew ( Alsuhaibani et al.,2024), both of which utilise an inter-training task. 5 Results Table2compares BARRAC against three base- line groups. Direct inference BERT-based models (BBMs) perform near chance (means of 20.97– 27.62), confirming that task supervision is in- dispensable. Supervised fine-tuning (SFT) on the 100 labels lifts the strongest Arabic encoder, ARBERTv2, to a mean of 55.46. The BERT and DistilBERT collapse on DART (10.12 and 6.98), predicting only a subset of the five dialect classes: their English-centric vocabularies frag- ment dialect-bearing tokens. The few-label SOTA methods Cluster&Tune ( Shnarch et al.,2022) and IDoFew ( Alsuhaibani et al.,2024), both instanti- ated with ARBERTv2, do not surpass plain AR- BERTv2 SFT on average (53.89 and 54.93), sug- gesting that intermediate clustering objectives de- veloped for English text classification transfer im- perfectly to dialectal Arabic tasks. BARRAC achieves the best macro-F1 on all five datasets, with a mean of 63.93: +8.48 over the best SFT baseline and +3 over the best few-label SOTA method. The improvements are consistent across binary and multi-label spaces, dataset sizes, and both multi-dialect and Saudi-only distributions, in- dicating that the dual-stream paradigm generalises well beyond its original ABSA setting. Within BARRAC, backbone and training recipe both mat- ter: running the adapted synthesis pipeline with the original DS²-ABSA configuration, with an Ara- bic encoder under noisy self-training attains 60.92. Since this comparison varies the backbone and recipe jointly, we isolate the training-strategy ef- fect separately in (§ 5.1). Table5compares BARRAC against multiple LLMs. We outperform open-weight LLMs by a wide margin and match GPT-4o on average (63.93 vs. 63.98), surpassing it on four of the five datasets. See 5for the full breakdown. Zero-seed transfer to DART.DART provides our strictest generalisation test: unlike the other four datasets, no DART data was used to seed syn- thesis; the dialect synthetic corpus was generated from Ar-Dialects seeds, and DART contributes only its 100 gold labels, evaluation, and test sets. Despite this, our method reaches 68.76 on DART, +16.09 over ARBERTv2 SFT and +15.52 over the best SOTA method. The synthetic dialect data thus encodes transferable, class-discriminative dialec- tal evidence rather than memorising seed-set arte- facts, which is precisely the property required of synthetic data in low-resource settings. Comparison with LLMs.Table5benchmarks against LLMs, selected to span scale (7B–72B and a frontier closed model), language speciali- sation (Arabic-centric ALLaM and AceGPT vs. multilingual Llama and Qwen), and accessibility (open vs. closed weights). Our 163M-parameter encoder outperforms every open-weight LLM by a wide margin (best: ALLaM-7B at 56.77, − 7 . 16 vs. ours); notably, Arabic-centric pre-training alone does not close the gap, and scale does not help (Qwen2.5-72B trails ALLaM-7B by 17.20). Most strikingly, our model is comparable on average to GPT-4o (63.93 vs. 63.98, the same model that gen- erated our synthetic data) and surpasses it on four of the five datasets. The student thus matches its teacher at a small fraction of the parameter count ModelAr-Sent.Ar-Sarc.Sa’7rDARTAr-Dial.Mean Ours Intermediate training 65.61 69.06 67.03 68.76 49.19 63.93 Noisy self-training65.3166.43 64.20 66.43 42.21 60.92 Noisy self-training+64.0662.07 64.24 67.82 44.37 60.51 Table 4: A comparative analysis of our proposed method and the DS2-ABSA training strategies was con- ducted across five distinct evaluation datasets, maintain- ing the consistency of the ARBERTv2 backbone model. ”+” indicates SOTA baselines adapted to our training hyperparameters. Method/DatasetAr-Sent.Ar-Sarc.Sa’7rDARTAr-Dial.Mean SFT using 100 Labels (LLMs) ALLaM-7B-Instruct63.2867.07 61.72 54.99 36.79 56.77 Llama 3.1 8B56.8158.32 57.94 36.35 29.92 47.87 AceGPT-16B16.8945.55 32.54 0.35 16.02 22.27 Qwen2.5-72B-Instruct53.5356.99 34.11 27.56 25.66 39.57 GPT-4o72.4767.14 65.18 66.49 48.6363.98 BARRAC (Ours)65.6169.06 67.03 68.76 49.1963.93 Table 5: Macro-F1 (%) results on against LLMs. and inference cost, and without API dependence. GPT-4o retains a clear advantage on Ar-Sentiment (72.47 vs. 65.61). 5.1 Effect of the Synthetic–Gold Training Strategy In Table 4we contrast two ways of combining syn- thetic and gold data on the same ARBERTv2 back- bone.Noisy self-training, as in the original DS²- ABSA, mixes pseudo-labelled synthetic data with gold data throughout training; we evaluate it un- der both the original hyper-parameters (60.92) and hyper-parameters tuned to our setting (60.51). Our intermediate trainingstrategy, fine-tuning first on synthetic data, then on the 100 gold labels, reaches 63.93, a +3.01 improvement over the best self- training variant, with gains on all datasets, largest on Ar-Dialects (+4.82) and Ar-Sarcasm (+2.63) in a non-iterative learning protocol. Because syn- thetic labels are noisier than gold labels, interleav- ing them allows label noise to persist into the final model, whereas intermediate training treats synthe- sis as task-adaptive initialisation and reserves the final optimisation steps for clean supervision. The training strategy choice is consequential, account- ing for roughly a third of our total improvement over direct SFT. 5.2 Ablation Study Table6ablates BARRAC. On the synthesis side, removing the instance-driven stream costs−4.11 mean F1, more than twice the cost of removing the BARRAC w/oAr-Sent. Ar-Sarc. Sa’7r DART Ar-Dial. Mean Key-Point-Driven64.0865.16 67.55 67.62 47.18 62.32 Instance-Driven59.6964.35 62.91 66.37 45.77 59.82 Normalisation63.8365.48 65.75 67.30 47.18 61.91 Inter-model (PL-FT) 60.2561.91 61.88 52.67 40.58 55.46 Final model (FL-FT) 16.0660.98 59.95 32.90 13.47 36.67 Full Model65.6169.06 67.03 68.76 49.19 63.93 Table 6: Ablation study results reported using Macro- F1 (%). key-point-driven stream−1.61; both streams help on nearly every dataset, highlighting their com- plementarity. Removing label normalisation costs −2.02. On the training side, removing the syn- thetic stage entirely costs−8.47. Conversely, re- moving the gold fine-tuning stage is catastrophic −27.26: Ar-Sentiment and Ar-Dialects collapse to 16.06 and 13.47, while binary sarcasm degrades more gracefully (60.98 and 59.95). This asym- metry suggests that synthetic data alone suﬀices to learn coarse binary distinctions but cannot cal- ibrate fine-grained decision boundaries under the distribution shift between LLM-generated and real tweets; even 100 gold labels are decisive for an- choring the label space. Together, the two training- side ablations show that synthetic and gold data are not interchangeable but strictly complemen- tary: each stage alone underperforms their sequen- tial combination by 8.5 and 27.3 points, respec- tively. 6 Error Analysis To better understand the remaining failure modes, we manually examined misclassified instances from each dataset. Specifically, we randomly sam- pled 20 misclassified examples per dataset and grouped them according to the underlying source of error. As shown in Table 7, most errors fall into four broad categories: Insuﬀicient Context, where very short texts provide limited evidence; Linguis- tic Overlap, where lexical cues are shared across classes; Label Ambiguity, where multiple interpre- tations are plausible; and Annotation Ambiguity, where the gold label appears debatable. See Ap- pendix A(Figure2, Table7) for details. 7 Conclusion We asked whether a task-specific data-synthesis framework devised for English aspect-based sen- timent analysis could be adapted to Arabic soci- olinguistic classification. By regrounding the com- ponents of DS²-ABSA in task-specific Arabic lin- guistic markers across a ten-domain space, and re- placing noisy self-training with intermediate train- ing followed by few-label fine-tuning, our adapta- tion attains a mean macro-F1 of 63.93 across five datasets and three tasks. In our experiments, it improves on the strongest supervised baseline by 8.48% and the best few-label method by +3%, and performs comparably to GPT-4o 63.98 on average while exceeding it on four of five datasets, at a sub- stantially smaller parameter count. These gains hold across binary, three-way, and five-way label spaces and transfer to a zero-seed dataset. 8 Limitations Several factors limit the scope of our conclusions. The generation pipeline itself, however, relies on a single proprietary model (GPT-4o), so regenerat- ing the data from scratch or extending the approach to new tasks incurs API cost, offers limited trans- parency into the generation process, and may ef- fectively bound the student’s performance by the teacher’s competence; the gap that remains on Ar- Sentiment would be consistent with such a ceiling, though we cannot establish this from our experi- ments alone. Our evaluation covers a single lan- guage, five datasets, three tasks, and a 100-label budget, so we make no claims about behaviour be- yond this particular low-resource setting, and re- sults may not transfer to other languages, dialects, or tasks. The class-balanced synthetic corpus that appears to strengthen minority-class evidence may also shift decision thresholds, which we tentatively associate with the precision cost observed in sar- casm detection; with only 100 gold labels, the fine- tuning stage may be insuﬀicient to fully recalibrate it. Performance on low-support dialects (e.g., Gulf, Levantine, and Maghrebi) is likewise limited, plau- sibly in part because of an imperfect match be- tween synthetic and real usage. Finally, some of the residual error may not be purely learnable: prior work suggests the neutral–polar boundary in Ar-Sentiment and the irony-hashtag construction of Sa’7r introduce label ambiguity in the source corpora, which could upper-bound achievable per- formance independently of the model. 9 Ethical Considerations All datasets used are publiclyavailableand were, to our knowledge, employed in accordance with their licenses and intended research use. 10 Generative AI usage disclosure The authors disclose the use of generative AI tools to assist with translation, LaTeX code cleanup, for- matting, and the addition of Arabic script. References Sura F. Abdhood, Nazlia Omar, and Sabrina Tiun. 2025.Data augmentation for arabic text classifi- cation: A review of current methods, challenges and prospective directions.PeerJ Computer Sci- ence, 11:e2685. Muhammad Abdul-Mageed, AbdelRahim El- madany, and El Moatez Billah Nagoudi. 2021. ARBERT & MARBERT: Deep bidirectional transformers for Arabic. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Lan- guage Processing (Volume 1: Long Papers), pages 7088–7105, Online. Association for Computational Linguistics. Muhammad Abdul-Mageed, AbdelRahim El- madany, Chiyu Zhang, El Moatez Billah Nagoudi, Houda Bouamor, and Nizar Habash. 2023. NADI 2023: The fourth nuanced Arabic dialect identification shared task. In Proceedings of ArabicNLP 2023, pages 600– 613, Singapore (Hybrid). Association for Computational Linguistics. Muhammad Abdul-Mageed, Amr Keleg, Abdel- Rahim Elmadany, Chiyu Zhang, Injy Hamed, Walid Magdy, Houda Bouamor, and Nizar Habash. 2024. NADI 2024: The fifth nuanced Arabic dialect identification shared task. In Proceedings of the Second Arabic Natural Lan- guage Processing Conference, pages 709–728, Bangkok, Thailand. Association for Computa- tional Linguistics. Ibrahim Abu Farha and Walid Magdy. 2020. From Arabic sentiment analysis to sarcasm detection: The ArSarcasm dataset. InProceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Of- fensive Language Detection, pages 32–39, Mar- seille, France. Halah AlMazrua, Najla AlHazzani, Amaal AlDa- wod, Lama AlAwlaqi, Noura AlReshoudi, Hend Al-Khalifa, and Luluh AlDhubayi. 2022.Sa‘7r: A saudi dialect irony dataset. InProceedinsg of the 5th Workshop on Open-Source Arabic Cor- pora and Processing Tools with Shared Tasks on Qur’an QA and Fine-Grained Hate Speech De- tection, pages 60–70, Marseille, France. Euro- pean Language Resources Association. Israa Alsarsour, Esraa Mohamed, Reem Suwaileh, and Tamer Elsayed. 2018. DART: A large dataset of dialectal Arabic tweets. InLREC 2018. Abdullah Alsuhaibani, Hamad Zogan, Imran Razzak, Shoaib Jameel, and Guandong Xu. 2024. Idofew: Intermediate training using dual- clustering in language models for few labels text classification. InWSDM, page 18–27. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. AraBERT: Transformer-based model for Arabic language understanding. InProceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection, pages 9– 15. M Saiful Bari, Yazeed Alnumay, Norah A. Alzahrani, Nouf M. Alotaibi, Hisham Abdullah Alyahya, Sultan AlRashed, Faisal Abdulrah- man Mirza, Shaykhah Z. Alsubaie, Hassan A. Alahmed, Ghadah Alabduljabbar, Raghad Alkhathran, Yousef Almushayqih, Raneem Al- najim, Salman Alsubaihi, Maryam Al Mansour, Saad Amin Hassan, Dr. Majed Alrubaian, Ali Alammari, Zaki Alawami, and 7 others. 2025. ALLam: Large language models for arabic and english. InThe Thirteenth International Conference on Learning Representations. Kenneth R. Beesley. 1996.Arabic finite-state mor- phological analysis and generation. InCOLING 1996 Volume 1: The 16th International Confer- ence on Computational Linguistics. Yonatan Belinkov, Alexander Magidow, Alberto Barrón-Cedeño, Avi Shmidman, and Maxim Romanov. 2018.Studying the history of the arabic language: Language technology and a large-scale historical corpus.Preprint, arXiv:1809.03891. Naaima Boudad, Rdouan Faizi, Rachid Oulad Haj Thami, and Raddouane Chiheb. 2018. Sen- timent analysis in arabic: A review of the literature.Ain Shams Engineering Journal, 9(4):2479–2490. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. InNAACL, pages 4171–4186. Meta. 2024.Introducing meta llama 3: The most capable openly available llm to date. OpenAI. 2025.GPT-Family (April 14 version). Large language model. Ali Safaya, Moutasem Abdullatif, and Deniz Yuret. 2020. KUISAIL at SemEval-2020 task 12: BERT-CNN for offensive speech identification in social media. InProceedings of the Four- teenth Workshop on Semantic Evaluation, pages 2054–2059, Barcelona (online). Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2020.Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter .Preprint, arXiv:1910.01108. Haitham Seelawi, Ibraheem Tuffaha, Mahmoud Gzawi, Wael Farhan, Bashar Talafha, Riham Badawi, Zyad Sober, Oday Al-Dweik, Abed Al- hakim Freihat, and Hussein Al-Natsheh. 2021. ALUE: Arabic language understanding evalua- tion. InProceedings of the Sixth Arabic Natu- ral Language Processing Workshop, pages 173– 184, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. Eyal Shnarch, Ariel Gera, Alon Halfon, Lena Dankin, Leshem Choshen, Ranit Aharonov, and Noam Slonim. 2022. Cluster & tune: Boost cold start performance in text classification. InACL, pages 7639–7653. Bashar Talafha, Hawau Olamide Toyin, Peter Sul- livan, AbdelRahim A. Elmadany, Abdurrah- man Juma, Amirbek Djanibekov, Chiyu Zhang, Hamad Alshehhi, Hanan Aldarmaki, Mustafa Jarrar, Nizar Habash, and Muhammad Abdul- Mageed. 2025. NADI 2025: The first multidi- alectal Arabic speech processing shared task. In Proceedings of The Third Arabic Natural Lan- guage Processing Conference: Shared Tasks, pages 720–733, Suzhou, China. Association for Computational Linguistics. Qwen Team. 2024.Qwen2.5: A party of founda- tion models. Hongling Xu, Yice Zhang, Qianlong Wang, and Ruifeng Xu. 2025.DS 2 -ABSA: Dual-stream data synthesis with label refinement for few-shot aspect-based sentiment analysis. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15460–15478, Vienna, Austria. Association for Computational Linguistics. Jianqing Zhu, Huang Huang, Zhihang Lin, Juhao Liang, Zhengyang Tang, Khalid Almubarak, Ab- dulmohsen Alharthik, Bang An, Juncai He, Xi- angbo Wu, Fei Yu, Junying Chen, Zhuoheng Ma, Yuhao Du, He Zhang, Emad A. Alghamdi, Lian Zhang, Ruoyu Sun, Haizhou Li, and 2 oth- ers. 2025.Second language (arabic) acquisi- tion of llms via progressive vocabulary expan- sion.Preprint, arXiv:2412.12310. Henry Peng Zou and Cornelia Caragea. 2023. JointMatch: A unified approach for diverse and collaborative pseudo-labeling to semi- supervised text classification. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7290– 7301. Association for Computational Linguis- tics. A Extended Error Analysis Figure2reports our model’s error rate per gold class (misclassified instances over class support) on each test set; Table 8presents representative in- stances that our model classifies correctly while all competing few-label baselines fail. Examples are drawn from the intersection of baseline errors and our model’s correct predictions to characterise the source of the gains in § 5; our model’s own dom- inant failure modes, visible in Figure2, are dis- cussed throughout this section. On Ar-Sentiment, errors are far from uniform across polarities: the neutral class is misclassified at 43.4%, against 27.2% for positive and 18.6% for negative. This mirrors the documented sub- jectivity of neutral annotations in the source cor- pus, where re-annotation shifted a large fraction of neutral and positive labels ( Abu Farha and Magdy, 2020): the neutral–polar boundary is partly an an- notation artefact rather than a purely learnable dis- tinction. The qualitative examples reflect the same phenomenon; all baselines mislabel a congratula- tory news headline as positive on surface lexis (ฟฃዛኗ, “congratulates”), whereas our model, exposed to synthetic neutral instances containing polar trigger words, correctly returns neutral. Residual neutral errors of this kind plausibly account for much of the remaining gap to GPT-4o on this task (§5). On Ar-Sarcasm, error rates are nearly balanced across classes (24.2% on not-sarcasm vs. 22.3% on sarcasm), i.e., sarcastic recall reaches 77.7%, a sharp contrast with the recall bottleneck charac- teristic of this dataset, whose original baseline de- tected only 38% of sarcastic tweets (Abu Farha and Magdy,2020). The cost is precision: because the sarcastic class comprises only 16% of the test set, the 24.2% false-alarm rate on the majority class ac- counts for 427 of the 504 total errors. We attribute this trade-off to our class-balanced synthetic cor- pus, which strengthens minority-class evidence but shifts the decision threshold toward sarcasm; with only 100 gold labels, the subsequent fine-tuning stage cannot fully re-calibrate it. On Sa’7r the false-alarm tendency is amplified (42.1% on not- sarcasm vs. 28.5% on sarcasm), which we trace to the corpus construction: Sa’7r was collected via irony-indicative hashtags, keywords, and phrases, so its non-ironic tweets contain the very surface markers that signal irony elsewhere (AlMazrua et al. ,2022), weakening exactly the marker-level evidence our synthesis injects. The Zamalek ex- ample in Table 8illustrates the converse skill: all baselines misread enthusiastic, hyperbolic fan dis- course as sarcastic, while our model correctly with- holds the sarcasm label. On Ar-Dialects, error rates are inversely related to gold-class support: MSA, with 1,410 test in- stances, is misclassified at only 8.7%, while Gulf (n=105) and Levantine (n=112) reach 55.2% and 62.5%, and all four Maghrebi instances are missed, though with n=4 this cell measures support scarcity, not model behaviour, and it depresses macro-F1, partly explaining why Ar-Dialects re- mains our lowest-scoring task (49.19). With 100 gold labels spread over five classes (≈20 per class), the gold stage cannot calibrate rare dialects whose synthetic representations imperfectly match real usage. Zero-seed DART, whose test set is dialect-balanced, shows a strikingly different pro- file: error rates are uniform and moderate (14.7%– 32.7%; the 2/9 MSA cell is again too small to interpret), with no class collapsing, further ev- idence that the synthetic dialect evidence trans- fers as generalisable markers rather than corpus- specific artefacts (§ 5). Maghrebi remains the hard- est dialect (32.7%), consistent with DART’s own annotation study, which found Maghrebi the least reliably identifiable group even for human annota- tors (Alsarsour et al.,2018). The dialect examples in Table8show the characteristic baseline failure our model avoids: collapsing dialectal tweets onto MSA or Levnt, whereas our model recovers the gold dialect from sparse function-word evidence (e.g., Levantine݁ݱ؇ري,مఈ႙ၽܳ؇۱). 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa 18.6% (133/716) 43.4% (468/1078) 27.2% (86/316) (a) Ar-Sentiment NegativeNeutralPositive 24.2% (427/1765) 22.3% (77/345) (b) Ar-Sarcasm Not SarcasmSarcasm 42.1% (241/572) 28.5% (175/614) (c) Sa'7r Not SarcasmSarcasm 44.1% (211/479) 55.2% (58/105) 62.5% (70/112) 100.0% (4/4) 8.7% (123/1410) (d) Ar-Dialects Egypt Gulf Levant Magreb Msa 17.7% (50/282) 14.7% (40/272) 22.7% (56/247) 32.7% (65/199) 22.2% (2/9) (e) DART Egypt Gulf Levant Magreb Msa Figure 2: Misclassified instances distribution across five Arabic datasets. Percentages are computed over the test set. DatasetError TypeExampleGoldPred. DART / Ar-Dialects Insuﬀicient Context༡ߺࠊ ༡ߺࠊ Nice, nice. MGHLEV Linguistic Overlapڢܹب ݿ؇ًگً؇ وأாணر I said it before and I repeat it. MGHMSA Linguistic Overlapاذا ࣖࢻ ఋః۱ّ؜ب If you excelled now. LEVGLF Questionable Annota- tion د༠ܹ٭ጥ጑ اذا ۱ࣖࢻܾ ෠ຬިঌॻ༟ ނ؇ᄩႍ၍ ۏ٭ܹଫଊت زوદઊ وَ؇لᄩᄥ ݁أިض ۱ఈఃً؇ ۱؇ܳگݱ۬ Please, if they are going to come in the style of Gilbert Zwain and Nayla Mouawad, forget about this whole story. MSALEV Label Ambiguityر਍ಸ؇ لިڣگި ان ނ؇ء اՄ៰Ղ May God grant him success, God willing. LEVMGH Ar-Sarcasm / Sa’7r Questionable Annota- tion ۏٴب اܳـ୍ଲاݿ۬ وۏٴب ا৖৑ܳިان ؜ލ؇ن ۋݱ۬ ا྾ངීෂ ނ؇ޗݠ ྵื اܳٺڎرྵะ واݬިୖ୒؇ ف ݁٭ب ؜گٴ۬ وྵื ي ܋ި݁ٴ؇رس I brought the notebook and colours for the art class. You’re good at that, but teaching and its principles are a hundred obstacles away, you extra. Not-Sarc. Sarc. Label Ambiguityڢܭ ڢ྾ཏ ۱۬ Say “I swear”, haha. Not-Sarc. Sarc. Linguistic Overlap๤ཛྷ ৖৑ا༡۬ ۱ި ༥؇ب اܳأ٭ڎ ݁ؕ ً؇رྵะ اۊٺ٭؇راّ۬ ݿ྘٪۬ ೑಻Ⴄ၍ Honestly, he completely messed it up with Paris; his choices were bad. Not-Sarc. Sarc. Questionable Annota- tion ا৵৥ৠڎᄩፁዧ ا৵৥ৠڎᄩፁዧ ا۳ይዧܾ ዻዧ ا৵৥ৠڎ واܳލ୍ଲ Praise be to God, praise be to God; all thanks and gratitude belong to You, O God. Sarc. Not-Sarc. Label Ambiguity৖৑ ل؇ً؇ً؇ ᄔცا༚ ܹޔ ॷॖळ اڣݱܭ اܳۇٴب ؜ۇٴ۹ No, baby, that’s wrong. Do you want me to disconnect the internet from you? Sarc. Not-Sarc. Insuﬀicient Context๴ཇ؇ৎ৊ع ا اᄕც ۰༥؇༡ Just something quick. Not-Sarc. Sarc. Sentiment Questionable Annota- tion اܳٴۇٴ؇ت اঌፇዧ م ݬ؇݁ި ًگިܳـુળ ߙߵا ر݁ݯ؇ن ڢݠب و اܳިڢب ڢܹ٭ܭ ف اঌፇዧ َ؇ݿ٭ඔ൹ ۋٴ྘ب اذ۱ாணܾ Girls who have not started fasting, Ramadan is near and time is short, so I wanted to remind those who forgot. Neu.Pos. Label Ambiguityا৵৥ৠڎ اՄ៰Ղ وًأڎ ݁ލިار ޗި لܭ ؜݆ ොໍາت #ًި܋٭݄ިن ༥؇رى اܳٺ݄ۜ٭ܭ Thank God, after a long search, Pokémon is finally downloading. Neu.Pos. Questionable Annota- tion ޗگݴ Ⴄ၍ذب لگިܳި ٔߺࠊج و لޚܹؕ රඞ #ޗگݴ_اܳأݠب False weather forecast; they said there would be snow but it turned out hot. Neu.Neg. Insuﬀicient Contextاى واՄ៰Ղ :( #اݿߺࠊب_ۋ٭؇ة Yes indeed :) #Lifestyle Pos.Neu. Questionable Annota- tion ا঺঒ ሒᇀ؇ب ۱؇ري ًިߙߵ :) ৎ৊݆ َگܹۇٴ؇ ྘ྲྀٺۇٴ؇ ذا ݪ؇ع اܳݱۇٴڎوق اሒᇿ ڣ٭۬ ܋ٺม฀ I want my Harry Potter book :( When we moved house, the box containing my books got lost. Pos.Neg. Table 7: Representative errors made by our model. Most errors arise from insuﬀicient context, linguistic overlap between classes, label ambiguity, or questionable annotations. Text / English TranslationGoldC&TIDoFewNSTOurs Dialect Identification ّأޚ٭۹ ݁ݱ؇ري আॻ༟ ۱؇ܳఈ႙ၽم They give you money for saying this. LEVGLFGLFGLFLEV ݁ڎام أن ۏۇٴ؇۹ً ނ؇م ࠍ੊ۇٴ؇ሒᇀ وأᆙᆍ ೑಻ب ܋ଫଃ༠ ଫ଒ك وأَ؇ ݁؇༥৖৑ ೞ಻؇لؕ و৖৑ ݪ؇ሒᇧ Since you and your side is leaving mine, I’l leave you. May you be blessed with much good; I am neither in need nor asking. GLFMGHMGHMGHGLF ݁گ؇ޗأ۰ ڣݠ૭૙؇ اৎ৊؞ݠب ਍ಱޙܾ إሌᇿ اොູ৖৑؇د ا৕৑ڣݠ لࠔࠫ ڢݠਊಱ؇ ً... France boycott, Morocco is joining the African Union soon ... MGHMSAMSAMSAMGH وزඔ൹َఈః༟ اذا رڣأۇٴ؇ ݿأݠ اܳٺڍاாண .. ؗݱص ؜ۇٴ۹ ࣖࣁڣؕ وا೑಻ ݿ؇܋ب ؜ލ؇ن ૰૜ިف اොູ৖৑؇د و۳ᆇᅹިره And they are upset if we raise ticket prices. You will pay whether you like it or not and stay quiet so you can watch Al-Ittihad and its fans. GLFEGYEGYEGYGLF Sarcasm Detection اࠍ༠ࣖࣁ ඔ൹᛻੆ߺࠊن رو݁؇ت أَࡤࡲ ا৵৥ৠڎᄩፁዧ واܳލ୍ଲ Now you enter online chat rooms; thank God. sarcasm not-sarcasm not-sarcasm not-sarcasmsarcasm ؜ڰިا ڣ۳ڎ ᆇ୚୘ દઑڎ اނިف ఈ႙၍݁۬ ؜݆ ଫଃ༚ ا۳ᆙᆊ؇ Excuse me, Fahad bin Mohammed’s words are about something else. not-sarcasm sarcasmsarcasmsarcasmnot-sarcasm انت مهما كبرت ݁ݷ ๴ཏྡྷ۱ اي ৖৑؜ص ܳྟݴ྘ོ ލଫଃت اෑෂ݁ዻዧ؇ وܳأص ༟ܹލ؇ن اෑෂ݁ዻዧ؇ ݁ݷ ༟ܹލ؇ن اܳڰߺࠊس وۋص اෑෂ݁ዻዧ؇ و۳ᆇᅹިر اෑෂ݁ዻዧ؇ ݆݁ ڢܹٴ۬ What matters is that any player wears Zamalek’s shirt and plays for Zamalek ... not-sarcasm sarcasmsarcasmsarcasmnot-sarcasm Sentiment Classification اොູ৖৑؇د ا৙৑وروฟฃዛኗ ሒᇀ ܳٴۇٴ؇ن ൑ശ૰૜ আॻ༟ܭ༡ ܝި݁ٺ۬ اࠍ੊ڎࣖࢴة ߓߵف؇ݿ۰ ݿأڎ اࠍ੆ݠߌߵي The European Union congratulates Lebanon on forming its new government headed by Saad Hariri. neutralpositivepositivepositiveneutral ޗگݴ اܳأݠب أۊٴ؇رቕመ داஓ୴؇ ً݁ٺ؊රඝة و৖৑ ؜ۇٴڎቕመ ݁ݱڎر ܳٺۜݠي اܳޚگݴ ෛຬݧુળً Arab Weather, your news is always late, and you do not have your own source for checking the weather. negativeneutralneutralpositivenegative Table 8: Examples of correctly classified instances by the proposed model where competing baselines make incor- rect predictions. Arabic texts are shown with English translations. C&T denotes Cluster&Tune, and NST denotes noisy self-training. AttributePrompt Tweet TopicBrainstorm a list of Arabic Twitter tweet topics (at least 200). Please adhere to the following guidelines: - Summarise each topic as a short, neutral description of a real-world event, situation, or subject that Arabic Twitter users commonly discuss. - Cover diverse domains: economy, sports, health, environment, entertainment, transportation, government services and infrastructure, technology and the internet, education, and food culture. - Topics must be neutral, without sentiment, dialect, or rhetorical bias. Your output should be a Python list of strings, with each element being a topic description. Output ONLY the Python list, nothing else. Linguistic De- vice (Sarcasm) Brainstorm a list of commonly used sarcasm devices in Arabic tweets about the domain: topic domain. Please adhere to the following guidelines: - Sarcasm devices are coarse-grained rhetorical mechanisms, not specific words or phrases. - They describe the way sarcasm is expressed, e.g. verbal irony, hyperbole, mock praise. - Always include ‘direct_statement’ as the non-sarcastic baseline device. Your output should be a Python list of strings, with each element being a device name. Out- put ONLY the Python list, nothing else. Linguistic De- vice (Sentiment) Brainstorm a list of commonly used sentiment categories in Arabic tweets about the domain: topic domain. Please adhere to the following guidelines: - Sentiment categories are coarse-grained expressive modes, not specific words or phrases. - They describe the way a polarity is expressed, e.g. frustration, celebration, factual reporting. - Provide categories for each of the three polarities: positive, neutral, and negative. Your output should be a Python dict with keys “positive”, “neutral”, and “negative”, each mapping to a list of category names. Output ONLY the Python dict, nothing else. Linguistic Marker Brainstorm a list of Arabic words or phrases that signal linguistic device in tweets about topic domain. Please adhere to the following guidelines: - Markers are fine-grained and concrete surface cues, not broad descriptions. - A marker is text a native speaker would immediately associate with the given device, dialect, or polarity. - Include lexical items, multiword phrases, and, for dialects, characteristic morphological and phonological forms. - Provide diverse markers spanning formal, colloquial, and mixed register. - Write each marker in Arabic script. Your output should be a Python list of Arabic strings. Output ONLY the Python list, nothing else. Table 9: Brainstorming prompts used in BARRAC. The linguistic device prompt is task-specific; the tweet topic and linguistic marker prompts are shared across tasks, with the device slot instantiated per task. For dialect, the device set is fixed a priori by the label space (EGY, GLF, LEV, MGH, MSA) TaskTask Instruction / Label Space Sarcasmidentifying whether the tweet is sarcastic or not sarcastic. Label space:sarcastic, not sarcastic Dialectidentifying which Arabic dialect the tweet is written in. Label space:EGY, GLF, LEV, MGH, MSA Sentimentidentifying the sentiment of the tweet. Label space:positive, neutral, negative Table 10: Task instructions and label spaces substituted into the task instruction and label space slots of all BARRAC prompts. MethodPrompt Attribute Prompting Write a tweet for the Arabic Twitter: tweet topic Label the tweet by task instruction. Requirements: - Keep a consistent style and annotation standard with the examples. - Convey linguistic device in the tweet, using the marker type ’linguistic marker’ as a signal. - Keep a consistent tweet style with examples. - Express expression pattern across the tweet. Here are some examples: Tweet: tweet Label: label ... ... Tweet: <Arabic tweet text> Label: <label space> Table 11: Prompt template for attribute prompting in BARRAC. task instruction and label space are instantiated per task (Table 10). expression pattern takes the values consistent, mixed, or implicit for sarcasm and sentiment, and formal, colloquial, or mixed for dialect. MethodPrompt Sample Combination Given 2 Arabic Twitter example tweets with the labels, please combine them to generate K diverse tweets. Label each tweet by task instruction. Requirements: - Keep a consistent style and annotation standard with the examples. - Maintain the same format as the examples. - Combine the aspects and meanings of both examples in every generated tweet. - If both examples share the same label, all K generated tweets must carry that label; otherwise, generate an equal number of tweets for each label. Examples: Tweet: tweet Label: label Tweet: tweet’ Label: label’ K Diverse Combined Tweets with Labels: 1. Tweet: Given an Arabic Twitter example tweet with the label, please paraphrase it to generate K diverse tweets. Label each tweet by task instruction. Requirements: - Keep a consistent style and annotation standard with the example. - Maintain the same format as the example. - The meaning of the example tweet should be unchanged. - All generated tweets must carry the same label as the example. Example: Tweet: tweet Label: label K Diverse Paraphrased Tweets with Labels: 1. Tweet: Selective Reconstruction Given a partially masked Arabic Twitter tweet, please fill and reconstruct it to generate K diverse tweets following the example. Label each tweet by task instruction. Masked tweet: masked tweet Requirements: - Keep a consistent style and annotation standard with the example. - Maintain the same format as the example. - The unmasked part of the tweet should be unchanged. Example: Tweet: tweet Label: label K Diverse Reconstructed Tweets with Labels: 1. Tweet: Table 12: Prompts for instance-driven data synthesis in BARRAC. The reconstruction prompt is applied to both masking strategies, differing only in the masked tweet supplied. MethodInput Prompt / Output Attribute Prompting (Sarcasm) Write a tweet for the Arabic Twitter:a government announcement about electricity price increases Label the tweet byidentifying whether the tweet is sarcastic or not sarcastic. Requirements: - Keep a consistent style and annotation standard with the examples. - Conveymock_praisein the tweet, using thesarcasm marker‘ل؇ ݿఈఃم’ (“how wonderful”) as a signal. - Keep a consistent tweet style with examples. - Expressconsistentacross the tweet. Tweet: <Arabic tweet text> Label: <sarcastic or not sarcastic> ... ... Tweet:ل؇ ݿఈఃمআॻ༟ اܳگݠار اࠍ੆ܝࡗࡲ، رڣأިا اܳـ۳᠌ݠً؇ء وᎂۋۇٴ؇ ྡྷྲྀލ۱୍ଲܾ How wonderful, a wise decision, they raised electricity prices and we thank them Label: sarcastic Table 13: Example of key-point-driven synthetic data from LLMs in BARRAC. Slots instantiated from the sampled attribute tuple and the task instruction are shown inblue; English glosses are given in italics and do not form part of the prompt. The output below the dashed line is the model response. MethodInput Prompt / Output Sample Combination (Sentiment) Given 2 Arabic Twitter example tweets with the labels, please combine them to generate4diverse tweets. Label each tweet by identifying the sentiment of the tweet. Requirements: - Keep a consistent style and annotation standard with the examples. - Maintain the same format as the examples. - Combine the aspects and meanings of both examples in every generated tweet. - If both examples share the same label, all 4 generated tweets must carry that label; otherwise, generate an equal number of tweets for each label. Examples: Tweet:أݿأ؇ر اࠍ੅ݯ؇ر ارّڰأب ႟ၽ૰૖ ۏۇٴިሒᇃ ۱ڍا ا৙৑ݿٴިع Vegetable prices rose insanely this week Label: negative Tweet:اሒᇭ ۰ᆇᅵෑෂ اܳޚݠ لݑ ༠؇َگ۰ ݆݁ اܳݱٴں The traﬀic on the road has been suffocating since morning Label: negative 4 Diverse Combined Tweets with Labels: 1. Tweet: ... ... وڢڰب ݿ؇ሒᇭ ۰༟ ا۰ᆇᅵෑෂ ؜ލ؇ن أނଫଐي ۊݯ؇ر ً؊ݿأ؇ر ۏۇٴި۰ਃ಻ I stood an hour in traﬀic just to buy vegetables at insane prices Label: negative 2. Tweet:ا༠ ۰ᆇᅵෑෂ؇َگ۰ وا৙৑ݿأ؇ر ሒᇭ اܳފިق ݁؇ ّޚ؇ق، لިم ݁ٺأص The traﬀic is suffocating and market prices are unbearable, an exhausting day Label: negative 3. Tweet:݆݁ ز۰ᆇᅵ اܳݱٴں ܳ؞ఈఃء اࠍ੅ݯ؇ر، ๴ཇ ႟၍ء ݬ؇ر أݬأص From the morning traﬀic to the cost of vegetables, everything has become harder Label: negative 4. Tweet:أݿأ؇ر اࠍ੅ݯ؇ر ۏۇٴި۰ਃ಻ واܳޚݠ لݑ ༠؇َݑ، اܳިݪؕ ݬأص ۱؇৙৑ݿٴިع Vegetable prices are insane and the road is choked, the situation is hard this week Label: negative Selective Reconstruction (Dialect) Given a partially masked Arabic Twitter tweet, please fill and reconstruct it to generate4diverse tweets following the example. Label each tweet byidentifying which Arabic dialect the tweet is written in. Masked tweet:إل۬ ده<mask> <mask> <mask>݆݁ اܳݱٴں What is this <mask> <mask> <mask> since morning Requirements: - Keep a consistent style and annotation standard with the example. - Maintain the same format as the example. - The unmasked part of the tweet should be unchanged. Example: Tweet:إل۬ ده ᄩႍ၍ زᄕც ۰ᆇᅵه ݆݁ اܳݱٴں What is all this traﬀic since morning Label: EGY 4 Diverse Reconstructed Tweets with Labels: 1. Tweet: ... ... إل۬ ده ᄩႍ၍ زᄕც ۰ᆇᅵه ݆݁ اܳݱٴں What is all this traﬀic since morning Label: EGY 2. Tweet:إل۬ ده اࠍ੊ިරඞ أوي ݆݁ اܳݱٴں What is this, the weather has been so hot since morning Label: EGY 3. Tweet:إل۬ ده اܳޚ؇ًިر ݁ݷ ౪౏భܹݧ ݆݁ اܳݱٴں What is this, the queue hasn’t ended since morning Label: EGY 4. Tweet:إل۬ ده اܳލ؞ܭ ܋ᄕც ଫଐه ݆݁ اܳݱٴں What is this, the work has piled up like this since morning Label: EGY Table 14: Examples of instance-driven synthetic data from LLMs in BARRAC. Slots instantiated per task are shown in blue; English glosses are given in italics and do not form part of the prompt. The output below each dashed line is the model response forK= 4. The output below the dashed line is the model response.

Instant research discovery