On annotating multi-dialect Arabic datasets, it is common to randomly assign the samples across a pool of native Arabic speakers. Recent analyses recommended routing dialectal samples to native speakers of their respective dialects to build higher-quality datasets. However, automatically identifying the dialect of samples is hard. Moreover, the pool of annotators who are native speakers of specific Arabic dialects might be scarce. Arabic Level of Dialectness (ALDi) was recently introduced as a quantitative variable that measures how sentences diverge from Standard Arabic. On randomly assigning samples to annotators, we hypothesize that samples of higher ALDi scores are harder to label especially if they are written in dialects that the annotators do not speak. We test this by analyzing the relation between ALDi scores and the annotators’ agreement, on 15 public datasets having raw individual sample annotations for various sentence-classification tasks. We find strong evidence supporting our hypothesis for 11 of them. Consequently, we recommend prioritizing routing samples of high ALDi scores to native speakers of each sample’s dialect, for which the dialect could be automatically identified at higher accuracies.

On annotating multi-dialect Arabic datasets, it is common to randomly assign the samples across a pool of native Arabic speakers. Recent analyses recommended routing dialectal samples to native speakers of their respective dialects to build higher-quality datasets. However, automatically identifying the dialect of samples is hard. Moreover, the pool of annotators who are native speakers of specific Arabic dialects might be scarce. Arabic Level of Dialectness (ALDi) was recently introduced as a quantitative variable that measures how sentences diverge from Standard Arabic. On randomly assigning samples to annotators, we hypothesize that samples of higher ALDi scores are harder to label especially if they are written in dialects that the annotators do not speak. We test this by analyzing the relation between ALDi scores and the annotators’ agreement, on 15 public datasets having raw individual sample annotations for various sentence-classification tasks. We find strong evidence supporting our hypothesis for 11 of them. Consequently, we recommend prioritizing routing samples of high ALDi scores to native speakers of each sample’s dialect, for which the dialect could be automatically identified at higher accuracies.

We describe the findings of the fifth Nuanced Arabic Dialect Identification Shared Task (NADI 2024). NADI’s objective is to help advance SoTA Arabic NLP by providing guidance, datasets, modeling opportunities, and standardized evaluation conditions that allow researchers to collaboratively compete on prespecified tasks. NADI 2024 targeted both dialect identification cast as a multi-label task (Subtask 1), identification of the Arabic level of dialectness (Subtask 2), and dialect-to-MSA machine translation (Subtask 3). A total of 51 unique teams registered for the shared task, of whom 12 teams have participated (with 76 valid submissions during the test phase). Among these, three teams participated in Subtask 1, three in Subtask 2, and eight in Subtask 3. The winning teams achieved 50.57 F1 on Subtask 1, 0.1403 RMSE for Subtask 2, and 20.44 BLEU in Subtask 3, respectively. Results show that Arabic dialect processing tasks such as dialect identification and machine translation remain challenging. We describe the methods employed by the participating teams and briefly offer an outlook for NADI.

A few benchmarking datasets have been released to evaluate the factual knowledge of pretrained language models. These benchmarks (e.g., LAMA, and ParaRel) are mainly developed in English and later are translated to form new multilingual versions (e.g., mLAMA, and mParaRel). Results on these multilingual benchmarks suggest that using English prompts to recall the facts from multilingual models usually yields significantly better and more consistent performance than using non-English prompts. Our analysis shows that mLAMA is biased toward facts from Western countries, which might affect the fairness of probing models. We propose a new framework for curating factual triples from Wikidata that are culturally diverse. A new benchmark DLAMA-v1 is built of factual triples from three pairs of contrasting cultures having a total of 78,259 triples from 20 relation predicates. The three pairs comprise facts representing the (Arab and Western), (Asian and Western), and (South American and Western) countries respectively. Having a more balanced benchmark (DLAMA-v1) supports that mBERT performs better on Western facts than non-Western ones, while monolingual Arabic, English, and Korean models tend to perform better on their culturally proximate facts. Moreover, both monolingual and multilingual models tend to make a prediction that is culturally or geographically relevant to the correct label, even if the prediction is wrong.

Transcribed speech and user-generated text in Arabic typically contain a mixture of Modern Standard Arabic (MSA), the standardized language taught in schools, and Dialectal Arabic (DA), used in daily communications. To handle this variation, previous work in Arabic NLP has focused on Dialect Identification (DI) on the sentence or the token level. However, DI treats the task as binary, whereas we argue that Arabic speakers perceive a spectrum of dialectness, which we operationalize at the sentence level as the Arabic Level of Dialectness (ALDi), a continuous linguistic variable. We introduce the AOC-ALDi dataset (derived from the AOC dataset), containing 127,835 sentences (17% from news articles and 83% from user comments on those articles) which are manually labeled with their level of dialectness. We provide a detailed analysis of AOC-ALDi and show that a model trained on it can effectively identify levels of dialectness on a range of other corpora (including dialects and genres not included in AOC-ALDi), providing a more nuanced picture than traditional DI systems. Through case studies, we illustrate how ALDi can reveal Arabic speakers’ stylistic choices in different situations, a useful property for sociolinguistic analyses.

Counterspeech has been proposed as a solution to the proliferation of online hate. Research has shown that natural language processing (NLP) approaches could generate such counterspeech automatically, but there are competing ideas for how NLP models might be used for this task and a variety of evaluation metrics whose relationship to one another is unclear. We test three different approaches and collect ratings of the generated counterspeech for 1,740 tweet-participant pairs to systematically compare the counterspeech on three aspects: quality, effectiveness and user preferences. We examine which model performs best at which metric and which aspects of counterspeech predict user preferences. A free-form text generation approach using ChatGPT performs the most consistently well, though its generations are occasionally unspecific and repetitive. In our experiment, participants’ preferences for counterspeech are predicted by the quality of the counterspeech, not its perceived effectiveness. The results can help future research approach counterspeech evaluation more systematically.

Automatic Arabic Dialect Identification (ADI) of text has gained great popularity since it was introduced in the early 2010s. Multiple datasets were developed, and yearly shared tasks have been running since 2018. However, ADI systems are reported to fail in distinguishing between the micro-dialects of Arabic. We argue that the currently adopted framing of the ADI task as a single-label classification problem is one of the main reasons for that. We highlight the limitation of the incompleteness of the Dialect labels and demonstrate how it impacts the evaluation of ADI systems. A manual error analysis for the predictions of an ADI, performed by 7 native speakers of different Arabic dialects, revealed that ≈ 67% of the validated errors are not true errors. Consequently, we propose framing ADI as a multi-label classification task and give recommendations for designing new ADI datasets.

Previous sarcasm generation research has focused on how to generate text that people perceive as sarcastic to create more human-like interactions. In this paper, we argue that we should first turn our attention to the question of when sarcasm should be generated, finding that humans consider sarcastic responses inappropriate to many input utterances. Next, we use a theory-driven framework for generating sarcastic responses, which allows us to control the linguistic devices included during generation. For each device, we investigate how much humans associate it with sarcasm, finding that pragmatic insincerity and emotional markers are devices crucial for making sarcasm recognisable.

Aspect Based Sentiment Analysis is a dominant research area with potential applications in social media analytics, business, finance, and health. Prior works in this area are primarily based on supervised methods, with a few techniques using weak supervision limited to predicting a single aspect category per review sentence. In this paper, we present an extremely weakly supervised multi-label Aspect Category Sentiment Analysis framework which does not use any labelled data. We only rely on a single word per class as an initial indicative information. We further propose an automatic word selection technique to choose these seed categories and sentiment words. We explore unsupervised language model post-training to improve the overall performance, and propose a multi-label generator model to generate multiple aspect category-sentiment pairs per review sentence. Experiments conducted on four benchmark datasets showcase our method to outperform other weakly supervised baselines by a significant margin.

Recently, author-annotated sarcasm datasets, which focus on intended, rather than perceived sarcasm, have been introduced. Although datasets collected using first-party annotation have important benefits, there is no comparison of human and machine performance on these new datasets. In this paper, we collect new annotations to provide human-level benchmarks for these first-party annotated sarcasm tasks in both English and Arabic, and compare the performance of human annotators to that of state-of-the-art sarcasm detection systems. Our analysis confirms that sarcasm detection is extremely challenging, with individual humans performing close to or slightly worse than the best trained models. With majority voting, however, humans are able to achieve the best results on all tasks. We also perform error analysis, finding that some of the most challenging examples are those that require additional context. We also highlight common features and patterns used to express sarcasm in English and Arabic such as idioms and proverbs. We suggest that to better capture sarcasm, future sarcasm detection datasets and models should focus on representing conversational and cultural context while leveraging world knowledge and common sense.

Data annotation is the foundation of most natural language processing (NLP) tasks. However, data annotation is complex and there is often no specific correct label, especially in subjective tasks. Data annotation is affected by the annotators’ ability to understand the provided data. In the case of Arabic, this is important due to the large dialectal variety. In this paper, we analyse how Arabic speakers understand other dialects in written text. Also, we analyse the effect of dialect familiarity on the quality of data annotation, focusing on Arabic sarcasm detection. This is done by collecting third-party labels and comparing them to high-quality first-party labels. Our analysis shows that annotators tend to better identify their own dialect and they are prone to confuse dialects they are unfamiliar with. For task labels, annotators tend to perform better on their dialect or dialects they are familiar with. Finally, females tend to perform better than males on the sarcasm detection task. We suggest that to guarantee high-quality labels, researchers should recruit native dialect speakers for annotation.

The Qur’an QA 2022 shared task aims at assessing the possibility of building systems that can extract answers to religious questions given relevant passages from the Holy Qur’an. This paper describes SMASH’s system that was used to participate in this shared task. Our experiments reveal a data leakage issue among the different splits of the dataset. This leakage problem hinders the reliability of using the models’ performance on the development dataset as a proxy for the ability of the models to generalize to new unseen samples. After creating better faithful splits from the original dataset, the basic strategy of fine-tuning a language model pretrained on classical Arabic text yielded the best performance on the new evaluation split. The results achieved by the model suggests that the small scale dataset is not enough to fine-tune large transformer-based language models in a way that generalizes well. Conversely, we believe that further attention could be paid to the type of questions that are being used to train the models given the sensitivity of the data.

iSarcasmEval is the first shared task to target intended sarcasm detection: the data for this task was provided and labelled by the authors of the texts themselves. Such an approach minimises the downfalls of other methods to collect sarcasm data, which rely on distant supervision or third-party annotations. The shared task contains two languages, English and Arabic, and three subtasks: sarcasm detection, sarcasm category classification, and pairwise sarcasm identification given a sarcastic sentence and its non-sarcastic rephrase. The task received submissions from 60 different teams, with the sarcasm detection task being the most popular. Most of the participating teams utilised pre-trained language models. In this paper, we provide an overview of the task, data, and participating teams.

SemEval 2021 Task 7, HaHackathon, was the first shared task to combine the previously separate domains of humor detection and offense detection. We collected 10,000 texts from Twitter and the Kaggle Short Jokes dataset, and had each annotated for humor and offense by 20 annotators aged 18-70. Our subtasks were binary humor detection, prediction of humor and offense ratings, and a novel controversy task: to predict if the variance in the humor ratings was higher than a specific threshold. The subtasks attracted 36-58 submissions, with most of the participants choosing to use pre-trained language models. Many of the highest performing teams also implemented additional optimization techniques, including task-adaptive training and adversarial training. The results suggest that the participating systems are well suited to humor detection, but that humor controversy is a more challenging task. We discuss which models excel in this task, which auxiliary techniques boost their performance, and analyze the errors which were not captured by the best systems.

The introduction of transformer-based language models has been a revolutionary step for natural language processing (NLP) research. These models, such as BERT, GPT and ELECTRA, led to state-of-the-art performance in many NLP tasks. Most of these models were initially developed for English and other languages followed later. Recently, several Arabic-specific models started emerging. However, there are limited direct comparisons between these models. In this paper, we evaluate the performance of 24 of these models on Arabic sentiment and sarcasm detection. Our results show that the models achieving the best performance are those that are trained on only Arabic data, including dialectal Arabic, and use a larger number of parameters, such as the recently released MARBERT. However, we noticed that AraELECTRA is one of the top performing models while being much more efficient in its computational cost. Finally, the experiments on AraGPT2 variants showed low performance compared to BERT models, which indicates that it might not be suitable for classification tasks.

This paper provides an overview of the WANLP 2021 shared task on sarcasm and sentiment detection in Arabic. The shared task has two subtasks: sarcasm detection (subtask 1) and sentiment analysis (subtask 2). This shared task aims to promote and bring attention to Arabic sarcasm detection, which is crucial to improve the performance in other tasks such as sentiment analysis. The dataset used in this shared task, namely ArSarcasm-v2, consists of 15,548 tweets labelled for sarcasm, sentiment and dialect. We received 27 and 22 submissions for subtasks 1 and 2 respectively. Most of the approaches relied on using and fine-tuning pre-trained language models such as AraBERT and MARBERT. The top achieved results for the sarcasm detection and sentiment analysis tasks were 0.6225 F1-score and 0.748 F1-PN respectively.

We introduce Chandler, a system that generates sarcastic responses to a given utterance. Previous sarcasm generators assume the intended meaning that sarcasm conceals is the opposite of the literal meaning. We argue that this traditional theory of sarcasm provides a grounding that is neither necessary, nor sufficient, for sarcasm to occur. Instead, we ground our generation process on a formal theory that specifies conditions that unambiguously differentiate sarcasm from non-sarcasm. Furthermore, Chandler not only generates sarcastic responses, but also explanations for why each response is sarcastic. This provides accountability, crucial for avoiding miscommunication between humans and conversational agents, particularly considering that sarcastic communication can be offensive. In human evaluation, Chandler achieves comparable or higher sarcasm scores, compared to state-of-the-art generators, while generating more diverse responses, that are more specific and more coherent to the input.

Sarcasm is one of the main challenges for sentiment analysis systems. Its complexity comes from the expression of opinion using implicit indirect phrasing. In this paper, we present ArSarcasm, an Arabic sarcasm detection dataset, which was created through the reannotation of available Arabic sentiment analysis datasets. The dataset contains 10,547 tweets, 16% of which are sarcastic. In addition to sarcasm the data was annotated for sentiment and dialects. Our analysis shows the highly subjective nature of these tasks, which is demonstrated by the shift in sentiment labels based on annotators’ biases. Experiments show the degradation of state-of-the-art sentiment analysers when faced with sarcastic content. Finally, we train a deep learning model for sarcasm detection using BiLSTM. The model achieves an F1 score of 0.46, which shows the challenging nature of the task, and should act as a basic baseline for future research on our dataset.

This paper provides an overview of the offensive language detection shared task at the 4th workshop on Open-Source Arabic Corpora and Processing Tools (OSACT4). There were two subtasks, namely: Subtask A, involving the detection of offensive language, which contains unacceptable or vulgar content in addition to any kind of explicit or implicit insults or attacks against individuals or groups; and Subtask B, involving the detection of hate speech, which contains insults or threats targeting a group based on their nationality, ethnicity, race, gender, political or sport affiliation, religious belief, or other common characteristics. In total, 40 teams signed up to participate in Subtask A, and 14 of them submitted test runs. For Subtask B, 33 teams signed up to participate and 13 of them submitted runs. We present and analyze all submissions in this paper.

Offensive language and hate-speech are phenomena that spread with the rising popularity of social media. Detecting such content is crucial for understanding and predicting conflicts, understanding polarisation among communities and providing means and tools to filter or block inappropriate content. This paper describes the SMASH team submission to OSACT4’s shared task on hate-speech and offensive language detection, where we explore different approaches to perform these tasks. The experiments cover a variety of approaches that include deep learning, transfer learning and multitask learning. We also explore the utilisation of sentiment information to perform the previous task. Our best model is a multitask learning architecture, based on CNN-BiLSTM, that was trained to detect hate-speech and offensive language and predict sentiment.

The use of pre-trained language models such as BERT and ULMFiT has become increasingly popular in shared tasks, due to their powerful language modelling capabilities. Our entry to SemEval uses ERNIE 2.0, a language model which is pre-trained on a large number of tasks to enrich the semantic and syntactic information learned. ERNIE’s knowledge masking pre-training task is a unique method for learning about named entities, and we hypothesise that it may be of use in a dataset which is built on news headlines and which contains many named entities. We optimize the hyperparameters in a regression and classification model and find that the hyperparameters we selected helped to make bigger gains in the classification model than the regression model.

Emoji are widely used to express emotions and concepts on social media, and prior work has shown that users’ choice of emoji reflects the way that they wish to present themselves to the world. Emoji usage is typically studied in the context of posts made by users, and this view has provided important insights into phenomena such as emotional expression and self-representation. In addition to making posts, however, social media platforms like Twitter allow for users to provide a short bio, which is an opportunity to briefly describe their account as a whole. In this work, we focus on the use of emoji in these bio statements. We explore the ways in which users include emoji in these self-descriptions, finding different patterns than those observed around emoji usage in tweets. We examine the relationships between emoji used in bios and the content of users’ tweets, showing that the topics and even the average sentiment of tweets varies for users with different emoji in their bios. Lastly, we confirm that homophily effects exist with respect to the types of emoji that are included in bios of users and their followers.

The choice of the corpus on which word embeddings are trained can have a sizable effect on the learned representations, the types of analyses that can be performed with them, and their utility as features for machine learning models. To contribute to the existing sets of pre-trained word embeddings, we introduce and release the first set of word embeddings trained on the content of Urban Dictionary, a crowd-sourced dictionary for slang words and phrases. We show that although these embeddings are trained on fewer total tokens (by at least an order of magnitude compared to most popular pre-trained embeddings), they have high performance across a range of common word embedding evaluations, ranging from semantic similarity to word clustering tasks. Further, for some extrinsic tasks such as sentiment analysis and sarcasm detection where we expect to require some knowledge of colloquial language on social media data, initializing classifiers with the Urban Dictionary Embeddings resulted in improved performance compared to initializing with a range of other well-known, pre-trained embeddings that are order of magnitude larger in size.

We consider the distinction between intended and perceived sarcasm in the context of textual sarcasm detection. The former occurs when an utterance is sarcastic from the perspective of its author, while the latter occurs when the utterance is interpreted as sarcastic by the audience. We show the limitations of previous labelling methods in capturing intended sarcasm and introduce the iSarcasm dataset of tweets labeled for sarcasm directly by their authors. Examining the state-of-the-art sarcasm detection models on our dataset showed low performance compared to previously studied datasets, which indicates that these datasets might be biased or obvious and sarcasm could be a phenomenon under-studied computationally thus far. By providing the iSarcasm dataset, we aim to encourage future NLP research to develop methods for detecting sarcasm in text as intended by the authors of the text, not as labeled under assumptions that we demonstrate to be sub-optimal.

Previous work has shown how to effectively use external resources such as dictionaries to improve English-language word embeddings, either by manipulating the training process or by applying post-hoc adjustments to the embedding space. We experiment with a multi-task learning approach for explicitly incorporating the structured elements of dictionary entries, such as user-assigned tags and usage examples, when learning embeddings for dictionary headwords. Our work generalizes several existing models for learning word embeddings from dictionaries. However, we find that the most effective representations overall are learned by simply training with a skip-gram objective over the concatenated text of all entries in the dictionary, giving no particular focus to the structure of the entries.

We investigate the impact of using author context on textual sarcasm detection. We define author context as the embedded representation of their historical posts on Twitter and suggest neural models that extract these representations. We experiment with two tweet datasets, one labelled manually for sarcasm, and the other via tag-based distant supervision. We achieve state-of-the-art performance on the second dataset, but not on the one labelled manually, indicating a difference between intended sarcasm, captured by distant supervision, and perceived sarcasm, captured by manual labelling.

Speech acts are the actions that a speaker intends when performing an utterance within conversations. In this paper, we proposed speech act classification for asynchronous conversations on Twitter using multiple machine learning methods including SVM and deep neural networks. We applied the proposed methods on the ArSAS tweets dataset. The obtained results show that superiority of deep learning methods compared to SVMs, where Bi-LSTM managed to achieve an accuracy of 87.5% and a macro-averaged F1 score 61.5%. We believe that our results are the first to be reported on the task of speech-act recognition for asynchronous conversations on Arabic Twitter.

Sentiment analysis (SA) is one of the most useful natural language processing applications. Literature is flooding with many papers and systems addressing this task, but most of the work is focused on English. In this paper, we present “Mazajak”, an online system for Arabic SA. The system is based on a deep learning model, which achieves state-of-the-art results on many Arabic dialect datasets including SemEval 2017 and ASTD. The availability of such system should assist various applications and research that rely on sentiment analysis as a tool.

In this paper, we present our work on detecting abusive language on Arabic social media. We extract a list of obscene words and hashtags using common patterns used in offensive and rude communications. We also classify Twitter users according to whether they use any of these words or not in their tweets. We expand the list of obscene words using this classification, and we report results on a newly created dataset of classified Arabic tweets (obscene, offensive, and clean). We make this dataset freely available for research, in addition to the list of obscene words and hashtags. We are also publicly releasing a large corpus of classified user comments that were deleted from a popular Arabic news site due to violations the site’s rules and guidelines.

This paper describes the development of a structured document collection containing user-generated text and numerical metadata for exploring the exploitation of metadata in information retrieval (IR). The collection consists of more than 61,000 documents extracted from YouTube video pages on basketball in general and NBA (National Basketball Association) in particular, together with a set of 40 topics and their relevance judgements. In addition, a collection of nearly 250,000 user profiles related to the NBA collection is available. Several baseline IR experiments report the effect of using video-associated metadata on retrieval effectiveness. The results surprisingly show that searching the videos titles only performs significantly better than searching additional metadata text fields of the videos such as the tags or the description.

This paper describes the development of an Arabic document image collection containing 34,651 documents from 1,378 different books and 25 topics with their relevance judgments. The books from which the collection is obtained are a part of a larger collection 75,000 books being scanned for archival and retrieval at the bibliotheca Alexandrina (BA). The documents in the collection vary widely in topics, fonts, and degradation levels. Initial baseline experiments were performed to examine the effectiveness of different index terms, with and without blind relevance feedback, on Arabic OCR degraded text.