Nora Al-Twairesh

Machine reading comprehension aims at understanding a passage and answer a given question by selecting a span from the passage. Recently, pre-trained language models achieved state-of-the-art results on Arabic machine reading comprehension, yet a broad body of works suggests that BERT-based variant models fail to encode and associate common sense facts and world knowledge. To alleviate this weakness, we propose an Arabic knowledge enriched reader model, which fuses external knowledge into contextual representation using an attention and gating mechanism. We learn and generate Arabic knowledge graph embeddings that represent information from Arabic Wikidata and utilize this representation when fusing knowledge. We adopted a knowledge graph embedding scoring function to select the most relevant concepts to the context from the knowledge graph. We evaluated our approach on multiple Arabic machine reading comprehension datasets. Despite leveraging a comparatively smaller pre-trained language model, our approach significantly outperforms large language models in Arabic machine reading comprehension across multiple benchmark datasets, achieving substantial gains in both EM and F1 scores.

One of the main tasks in the field of natural language processing (NLP) is the analysis of affective states (sentiment and emotional) based on written text, and attempts have improved dramatically in recent years. However, in studies on the Arabic language, machine learning or deep learning algorithms were utilised to analyse sentiment and emotion more often than current pre-trained language models. Additionally, further pre-training the language model on specific tasks (i.e., within-task and cross-task adaptation) has not yet been investigated for Arabic in general, and for the sentiment and emotion task in particular. In this paper, we adapt a BERT-based Arabic pretrained language model for the sentiment and emotion tasks by further pre-training it on a sentiment and emotion corpus. Hence, we developed five new Arabic models: QST, QSR, QSRT, QE3, and QE6. Five sentiment and two emotion datasets spanning both small- and large-resource settings were used to evaluate the developed models. The adaptation approaches significantly enhanced the performance of seven Arabic sentiment and emotion datasets. The developed models showed excellent improvements over the sentiment and emotion datasets, which ranged from 0.15–4.71%.

Over the last decade, many methods have been developed to address the domain dependency problem of sentiment classification under domain shift. This problem is exacerbated in Arabic by its feature sparsity induced by morphological complexity and dialect variability. However, only a few studies have proposed sentiment domain adaptation methods for Arabic, with inconsistent comparisons resulting from different datasets and settings, making it difficult to identify the most effective approaches. This is the first comparative study of the most effective domain adaptation methods for Arabic sentiment classification. We replicate the existing methods proposed for Arabic and compare their effectiveness on the standard dataset settings. To further examine the extent to which adaptation performance differs between Modern Standard Arabic (MSA) and Dialectal Arabic (DA), we employ two public multi-domain sentiment datasets. We also test two well-established methods that have been thoroughly utilized in English-related studies and examine if they maintain the same levels of performance when applied to Arabic. Our findings indicate that adaptation performanace on MSA is better than on DA for all traditional approaches. However, implementing adaptation on top of transformer-based language models shows superior performance on DA. Finally, methods that have proven to excel in English suffer from low performance when applied to Arabic and exhibit negative transfer in most cases.

A series of mitigation efforts were implemented in response to the COVID-19 pandemic in Saudi Arabia, including the development of mobile health applications (mHealth apps) for the public. Assessing the acceptability of mHealth apps among the public is crucial. This study aimed to use Twitter to understand public perceptions around the use of six Saudi mHealth apps used during COVID-19: “Sehha”, “Mawid”, “Sehhaty”, “Tetamman”, “Tawakkalna”, and “Tabaud”. We used two methodological approaches: network and sentiment analysis. We retrieved Twitter data using specific mHealth apps-related keywords. After including relevant tweets, our final mHealth app networks consisted of a total of 4995 Twitter users and 8666 conversational relationships. The largest networks in size (i.e., the number of users) and volume (i.e., the conversational relationships) among all were “Tawakkalna” followed by “Tabaud”, and their conversations were led by diverse governmental accounts. In contrast, the four remaining mHealth networks were mainly led by the health sector and media. Our sentiment analysis approach included five classes and showed that most conversations were neutral, which included facts or information pieces and general inquires. For the automated sentiment classifier, we used Support Vector Machine with AraVec embeddings as it outperformed the other tested classifiers. The sentiment classifier showed an accuracy, precision, recall, and F1-score of 85%. Future studies can use social media and real-time analytics to improve mHealth apps’ services and user experience, especially during health crises.

The Arabic language has many spoken dialects. However, until recently, it was primarily written in Modern Standard Arabic (MSA), which is the formal variant of Arabic. Social media platforms have changed the face of written Arabic where users converse freely in various dialects, thus offering a massive number of resources for the study of dialectal text. The Arabic dialects differ from MSA in morphology, syntax, and phonetics. Consequently, since the effectiveness of NLP tasks—like sentiment analysis—is dependent on the availability of representative resources, there is currently a great need for such resources in these dialects. In this paper, we present MARSA—the largest sentiment annotated corpus for Dialectal Arabic (DA) in the Gulf region, which consists of 61,353 manually labeled tweets that contain a total of 840 K tokens. The tweets were collected from trending hashtags in four domains: political, social, sports, and technology to create a multi-domain corpus. The importance of such a corpus is to facilitate the study of domain-dependent sentiment analysis in Arabic. In addition to this corpus, the annotators extracted indicator words to form affect lexicons for each domain. We draw insights from these lexicons regarding contextual polarity of certain words. Furthermore, we present benchmark experiments on the MARSA corpus in order to establish a baseline for further studies.

BACKGROUND In Saudi Arabia, the first novel coronavirus disease (COVID-19) confirmed case was reported on March 2, 2020, which followed a series of mitigation efforts imposed by the government. The development of specific mobile health applications (mHealth apps) for public use was one of the response strategies employed by the Saudi government. Assessing the impact of these mHealth apps through the opinions of the public posted on social media is crucial to improve mHealth services offered by governments. OBJECTIVE Our aim was to utilize Twitter, as a source of data, to understand conversations and perceptions of users around the use of six mHealth apps developed by the Saudi Ministry of Health, by conducting a network and sentimental analysis of Tweets. The mHealth apps included in our study were “Sehha”, “Mawid”, “Sehhaty”, “Tetamman”, “Tawakkalna”, and “Tabaud”. METHODS We collected mHealth-related Twitter data on December 12, 2020. After including relevant tweets, our final mHealth app networks consisted of a total of 4,995 Twitter users and 8,666 relationships. We used NodeXL to perform the network analysis and visualization. We performed a sentiment analysis using a human-in-the-loop and machine learning approaches. Our manual annotation initially included five classes (positive, neutral, negative, indeterminate, and sarcasm). We excluded indeterminate and sarcasm classes as they usually cause ambiguity for the sentiment classifier. We applied data augmentation techniques to ensure sentiment polarity (positive, negative) in the tweets. The sentiment classifier dataset consisted a total of 4,719 tweets with 26.6% positive, 52.2% neutral, and 21.2% negative. Data preprocessing and normalization were also performed. For building the sentiment classifier, we used the Support Vector Machine with the word2vec embeddings of AraVec. RESULTS Our network analysis showed that “Sehhaty”, “Tawakkalna”, and “Tabaud” had similar patterns and more interactions in conversations than other networks. “Tawakkalna” and “Tabaud” were the largest networks among all, and their conversations were led by various governmental accounts. In comparison, “Sehha”, “Mawid”, “Sehhaty”, and “Tetamman” networks were mainly led by a health sector and media. Our sentiment analysis showed that the majority of Twitter conversations around the six mHealth apps were neutral, which encompassed facts or information pieces, neutral suggestions, and general inquires. Positive tweets focused on appreciation, positive opinions, and expressions around government trust. In contrast, negative tweets included suggestions to overcome weaknesses, issues faced with apps, negative opinions, and negative psychological impact. Our sentiment classifier showed an accuracy, precision, recall, and an F1-score of 85%. CONCLUSIONS Social media can be used as a data source to understand public perceptions on the use of mHealth apps during pandemics. Real-time analytics of social media can help health authorities to address issues and concerns about mHealth apps during public health crises.

The field of natural language processing (NLP) has witnessed a boom in language representation models with the introduction of pretrained language models that are trained on massive textual data then used to fine-tune downstream NLP tasks. In this paper, we aim to study the evolution of language representation models by analyzing their effect on an under-researched NLP task: emotion analysis; for a low-resource language: Arabic. Most of the studies in the field of affect analysis focused on sentiment analysis, i.e., classifying text into valence (positive, negative, neutral) while few studies go further to analyze the finer grained emotional states (happiness, sadness, anger, etc.). Emotion analysis is a text classification problem that is tackled using machine learning techniques. Different language representation models have been used as features for these machine learning models to learn from. In this paper, we perform an empirical study on the evolution of language models, from the traditional term frequency–inverse document frequency (TF–IDF) to the more sophisticated word embedding word2vec, and finally the recent state-of-the-art pretrained language model, bidirectional encoder representations from transformers (BERT). We observe and analyze how the performance increases as we change the language model. We also investigate different BERT models for Arabic. We find that the best performance is achieved with the ArabicBERT large model, which is a BERT model trained on a large dataset of Arabic text. The increase in F1-score was significant +7–21%.

Approaches for developing Dialogue Systems (DSs) are typically categorized into rule-based and data-driven. Data-driven DSs require a massive quantity of training data, while rule-based DSs rely on a predefined set of rules and keywords that are to be detected in the user's utterances. The data-driven approaches show more promising results, but owing to a lack of available training data for Arabic task-oriented DSs, development of an Arabic task-oriented DS has typically been conducted using the rule-based approach. In this article, we propose a hybrid rule-based and data-driven approach in a text-based flight booking DS capable of handling customer's utterances. The proposed DS was built through utilizing the Wit.ai natural language interface. The conversation flow was configured using the Wizard of Oz technique, and the DS intents and entities were developed through the use of crowdsourcing training examples. The evaluation results show that the system developed was able to understand user utterances and to self-feed efficiently.

Sentiment Analysis of Arabic tweets is a complex task due to the rich morphology of the Arabic language and the informal nature of language on Twitter. Previous research on sentiment analysis of tweets mainly focused on manually extracting features from the text. Recently, neural word embeddings have been utilized as less labor-intensive representations than manual feature engineering. Most of these word-embeddings model the syntactic information of words while ignoring the sentiment context. In this paper, we propose to learn sentiment specific word embeddings from Arabic tweets and use them in Arabic Twitter sentiment classification. Moreover, we propose a feature ensemble model of surface and deep features. The surface features are manually extracted features, and the deep features are generic word embeddings and sentiment-specific word embeddings. Extensive experiments are performed to test the effectiveness of the surface and deep features ensemble, pooling functions, embeddings size, and cross-dataset models. The recent language representation model BERT is also evaluated on the task of Sentiment Analysis of Arabic tweets. The models are evaluated on three different datasets of Arabic tweets, and they outperform previous results on all these datasets with a significant increase in the F-score. The experimental results demonstrate that (1) the highest performing model is the ensemble of surface and deep features, and (2) the approach achieves state-of the-art results on several benchmarking datasets.