Arabic NLP

We developed an original approach to Arabic traditional morphology, involving new concepts in Semitic lexicology, morphology, and grammar for standard written Arabic. This new methodology for handling the rich and complex Semitic languages is based on good practices in Finite-State technologies (FSA/FST) by using Unitex, a lexicon-based corpus processing suite. For verbs (Neme, 2011), I proposed an inflectional taxonomy that increases the lexicon readability and makes it easier for Arabic speakers and linguists to encode, correct, and update it. Traditional grammar defines inflectional verbal classes by using verbal pattern-classes and root-classes. In our taxonomy, traditional pattern-classes are reused, and root-classes are redefined into a simpler system. The lexicon of verbs covered more than 99% of an evaluation corpus. For nouns and adjectives (Neme, 2013), we went one step further in the adaptation of traditional morphology. First, while this tradition is based on derivationa...

Vowels in Arabic are optional orthographic symbols written as diacritics above or below letters. In Arabic texts, typically more than 97 percent of written words do not explicitly show any of the vowels they contain; that is to say, depending on the author, genre and field, less than 3 percent of words include any explicit vowel. Although numerous studies have been published on the issue of restoring the omitted vowels in speech technologies, little attention has been given to this problem in papers dedicated to written Arabic technologies. In this research, we present Arabic-Unitex, an Arabic Language Resource, with emphasis on vowel representation and encoding. Specifically, we present two dozens of rules formalizing a detailed description of vowel omission in written text. They are typographical rules integrated into large-coverage resources for morphological annotation. For restoring vowels, our resources are capable of identifying words in which the vowels are not shown, as well as words in which the vowels are partially or fully included. By taking into account these rules, our resources are able to compute and restore for each word form a list of compatible fully vowelized candidates through omission-tolerant dictionary lookup. In our previous studies, we have proposed a straightforward encoding of taxonomy for verbs (Neme, 2011) and broken plurals . While traditional morphology is based on derivational rules, our description is based on inflectional ones. The breakthrough lies in the reversal of the traditional root-and-pattern Semitic model into pattern-and-root, giving precedence to patterns over roots. The lexicon is built and updated manually and contains 76,000 fully vowelized lemmas. It is then inflected by means of finite-state transducers (FSTs), generating 6 million forms. The coverage of these inflected forms is extended by formalized grammars, which accurately describe agglutinations around a core verb, noun, adjective or preposition. A laptop needs one minute to generate the 6 million inflected forms in a 340-Megabyte flat file, which is compressed in two minutes into 11 Megabytes for fast retrieval. Our program performs the analysis of 5,000 words/second for running text (20 pages/second).

We describe a lexicon of Arabic verbs constructed on the basis of Semitic patterns and used in a resource-based method of morphological annotation of written Arabic text. The annotated output is a graph of morphemes with accurate linguistic information. An enhanced FST implementation for Semitic languages was created. This system is adapted also for generating inflected forms. The language resources can be easily updated. The lexicon is constituted of 15 400 verbal entries. We propose an inflectional taxonomy that increases the lexicon readability and maintainability for Arabic speakers and lingui sts. Traditional grammar defines inflectional verbal classes by using verbal pattern-classes and root-classes, related to the nature of each of the triliteral root-consonants. Verbal pattern-classes are clearly defined but root-classes are complex. In our taxonomy, traditional pattern-classes are reused and root-classes are simply redefined. Our taxonomy provides a straightforward encoding scheme for inflectional variations and orthographic adjustments due to assimilation and agglutination. We have tested and evaluated our resource against 10 000 diacriticized verb occurrences in the Nemlar corpus and compared it to Buckwalter resources. The lexical coverage is 99.9 % and a laptop needs two minutes in order to generate and compress the inflected lexicon of 2.5 million forms into 4 Megabytes.

A natural path for Arabic morphology consists in adopting or adapting both the traditional Semitic model and finite-state technologies. On the one hand, we have to facilitate the linguist’s task of lexical encoding by proposing a familiar formalism: the Semitic model for morphology. On the other hand, computer scientists, in general, point to FSTs as standard devices for inflection; and FSTs have shown their simplicity and efficiency in inflectional morphology for European languages. Nevertheless, there are countless complexities in the implementation of this model with such a technique. This is due to the richness of Arabic morphology and to the actual details of the traditional root-and-pattern model. In fact, there is an opposition between the requirement to be faithful to the essence of the Semitic model, for the sake of lexicon encoders, and the necessity to curb the complexity of its traditional version. Yet, no trade-off has been found.Indeed, we have achieved and created fro...

Data annotation is an important and necessary task for all NLP applications. Designing and implementing a web-based application that enables many annotators to annotate and enter their input into one central database is not a trivial task. These kinds of web-based applications require a consistent and robust backup for the underlying database and support to enhance the efficiency and speed of the annotation. Also, they need to ensure that the annotations are stored with a minimal amount of redundancy in order to take advantage of the available resources(e.g, storage space). In this paper, we introduce WASA, a web-based annotation system for managing large-scale multilingual Code Switching (CS) data annotation. Although WASA has the ability to perform the annotation for any token sequence with arbitrary tag sets, we will focus on how WASA is used for CS annotation. The system supports concurrent annotation, handles multiple encodings, allows for several levels of management control, and enables quality control measures while seamlessly reporting annotation statistics from various perspectives and at different levels of granularity. Moreover, the system is integrated with a robust language specific date prepossessing tool to enhance the speed and efficiency of the annotation. We describe the annotation and the administration interfaces as well as the backend engine.

This paper presents a language identification system designed to detect the language of each word, in its context, in a multilingual documents as generated in social media by bilingual/multilingual communities, in our case speakers of Algerian Arabic. We frame the task as a sequence tagging problem and use supervised machine learning with standard methods like HMM and Ngram classification tagging. We also experiment with a lexicon-based method. Combining all the methods in a fall-back mechanism and introducing some linguistic rules, to deal with unseen tokens and ambiguous words, gives an overall accuracy of 93.14%. Finally, we introduced rules for language identification from sequences of recognised words.

DIRA is a query expansion tool that generates search terms in Standard Arabic and/or its dialects when provided with queries in English or Standard Arabic. The retrieval of dialectal Arabic text has recently become necessary due to the increase of dialectal content on social media. DIRA addresses the challenges of retrieving information in Arabic dialects, which have significant linguistic differences from Standard Arabic. To our knowledge, DIRA is the only tool in existence that automatically generates dialect search terms with relevant morphological variations from English or Standard Arabic query terms.

Most developing economies continue to be overdependent on agriculture, yet farmers fail to access professional knowledge and information in a timely fashion that may help them generate higher yields, reduce risks, and make decisions. In the past, farmer-advising systems have relied on face-to-face meetings, call centers, or fixed platforms in relaying information. Another revolutionary solution is the appearance of artificial intelligence that can be used in the form of multilingual chatbots with the help of Natural Language Processing (NLP). These chatbots provide farmers with highly personalized, context-sensitive, and readily available consulting services in their own languages. This study explores the concept, design, and effectiveness of multilingual chatbots in terms of advising farmers using natural language processing. It reviews the existing studies, discusses methods of chatbot development, analyzes challenges, including a variety of dialects and low-resource languages, and presents potential conclusions in the context of usability, functionality, and socioeconomic consequences. To ensure trust, scalability, and inclusivity, the discussion ends with a glimpse of the future and the need to focus on explainable AI, advanced NLP models, and interaction with Internet of Things (IoT) devices..

The rapid growth of social media in recent years has fed into some highly undesirable phenomena such as proliferation of abusive and offensive language on the Internet. Previous research suggests that such hateful content tends to come from users who share a set of common stereotypes and form communities around them. The current state-of-the-art approaches to hate speech detection are oblivious to user and community information and rely entirely on textual (i.e., lexical and semantic) cues. In this paper, we propose a novel approach to this problem that incorporates community-based profiling features of Twitter users. Experimenting with a dataset of 16k tweets, we show that our methods significantly outperform the current state of the art in hate speech detection. Further, we conduct a qualitative analysis of model characteristics. We release our code, pre-trained models and all the resources used in the public domain.

The rapid growth of social media in recent years has fed into some highly undesirable phenomena such as proliferation of hateful and offensive language on the Internet. Previous research suggests that such abusive content tends to come from users who share a set of common stereotypes and form communities around them. The current state-of-the-art approaches to abuse detection are oblivious to user and community information and rely entirely on textual (i.e., lexical and semantic) cues. In this paper, we propose a novel approach to this problem that incorporates community-based profiling features of Twitter users. Experimenting with a dataset of 16k tweets, we show that our methods significantly outperform the current state of the art in abuse detection. Further, we conduct a qualitative analysis of model characteristics. We release our code, pre-trained models and all the resources used in the public domain.

Extracting contextual information from low-resource languages such as Igbo remains a significant challenge due to limited linguistic data. This paper proposes a novel hybrid approach that leverages both global and subword-level information to address this limitation. A hybrid embedding framework, combining GloVe and FastText embeddings, is employed to capture rich semantic and syntactic information. These embeddings are then integrated into a Compact Convolutional Transformer (CCT) architecture, which replaces the computationally intensive self-attention mechanism with efficient convolutional layers. This design enables effective capture of local and global dependencies while reducing computational costs. Experimental results on small, domain-specific Igbo datasets, including customer support and medical dialogues, demonstrate the superior performance of the proposed model over baseline approaches. The hybrid model achieves higher accuracy and F1 scores, highlighting its potential to improve NLP performance in low-resource settings. This work contributes to the advancement of natural language processing for underrepresented languages.
Keywords: Low-resource languages, Natural Language Processing (NLP), Contextual information extraction, Compact Convolutional Transformer (CCT).

This research investigates the use of Machine Learning (ML) and Deep Learning, including BiLSTM approaches, for Sentiment Analysis (SA) of consumer reviews on social media sites. Businesses are increasingly depending on online reviews to determine customer satisfaction due to social media's explosive growth. We used three classification models, the assessment of these attitudes uses Naive Bayes (NB), Support Vector Machine (SVM), and a BiLSTM model. Customer reviews categorized as neutral, negative, or positive feelings made up the dataset used for this study, implying that positive reviews are related to satisfied customers. Text cleaning, tokenization, Bert, and TF-IDF feature extraction were among the preprocessing procedures. Our results show that MNB and CNB had accuracy rates of 80.21 and 81.44%, respectively, whereas linear and RBF SVM had slightly higher accuracy rates of 84.28 and 88.64%, respectively. Also, BiLSTM achieves 87.72%. Besides, this research sheds light on unbalanced dataset issues. Therefore, we apply random oversampling to the research dataset. The evaluated outcomes show that sentiment classification and knowledge extraction through machine learning models and deep learning methods yield beneficial insights that enable businesses to understand their customers better.

The relevance of textual analysis appears in numerous case studies across fields of social, business and academic communication. A central question in multilingual research is to develop a universal concept representation using a variety of models. Typologically different languages may have differing numbers of values for the same concept. There is an increased interest how typologically different languages encode morphology and syntax features across different layers. Language specific features as subject-verb agreement, flexibility of word order, morphological type require more parameters to represent. Crosslingual abstractions of morphosyntactic concepts serve as a basis for disentangling latent grammatical concepts across typologically diverse languages.

This overview presents the Author Profiling and Deception Detection in Arabic (APDA) shared task at PAN@FIRE 2019. Two have been the main aims of this years task: i) to profile the age, gender and native language of a Twitter user; ii) to determine whether an Arabic text is deceptive or not in two different genres: Twitter and news headlines. For this purpose we have created three corpora in Arabic. Altogether, the approaches of 13 participants are evaluated.

In recent times, machine learning and deep learning have quickly risen to prominence as highly effective instruments across a multitude of domains, encompassing areas such as image and speech interpretation, the processing of natural language, and even applications within the medical field. This paper offers an examination of the methodologies and practical uses of both machine learning and deep learning, detailing their respective advantages and disadvantages, alongside their prospective future trajectories. Furthermore, we delve into the inherent challenges linked with these technologies, such as concerns about data confidentiality, ethical questions, and the imperative for transparency within their operational decision-making frameworks. As two of the most transformative technologies within the artificial intelligence sphere, machine learning and deep learning have garnered significant popularity lately, largely due to their capacity for generating predictions, scrutinizing extensive datasets, and yielding insights previously unattainable. This document will investigate the fundamental principles of machine learning and deep learning, their distinguishing features, their varied applications, and their influence across different industrial sectors. The way we engage with technology is being fundamentally reshaped by machine learning and deep learning, which are also unlocking novel avenues for innovation. These technologies have already exerted considerable effects in numerous industries and possess the capability to continue revolutionizing our world. This review delivers a thorough overview of the core concepts of machine learning and deep learning, their differences, their applications, and their broader societal impact. By concentrating on contemporary literature and research, this article endeavors to foster a deeper comprehension of the potential held by machine learning and deep learning and their ramifications for the times to come.

This paper introduces SERTUS (Speech Emotion Recognition TUnisian Spontaneous), an extensive dataset collection intended to propel research in Speech Emotion Recognition (SER), particularly within the realm of Tunisian Dialect (TD). SERTUS encompasses both registers of the Tunisian Dialect: the Popular (familiar) register and the intellectual register, capturing a diverse range of emotions in spontaneous environments and natural interactions across different regions of Tunisia. This work delineates the methodology utilized in crafting SER-TUS, highlighting the challenges and strategies involved in capturing spontaneous interactions. Moreover, we underscore the importance of including TD and the multidomain nature of the dataset, illustrating its potential applications in various domains such as sports, politics, and culture. It's imperative to note that this corpus collection adhered to a rigorous protocol to ensure corpus quality, as it will undergo annotation in future research endeavors.

Emotional expressions are a fundamental aspect of human communication, with speech being one of the most natural modes of interaction. Speech Emotion Recognition (SER) is a significant research topic in Natural Language Processing (NLP), aimed at identifying emotions such as satisfaction, frustration, and anger from speech audio using multiple classifiers. This paper presents a method to emotion recognition from spontaneous Tunisian Dialect (TD) speech, marking the first work in the SER field to utilize spontaneous speech for emotion recognition in this dialect. The dataset was created from freely available YouTube videos across multiple domains and labeled with four perceived emotions: anger, satisfaction, frustration, and neutral. To address the data scarcity issue, we implemented data augmentation techniques, specifically Vocal Tract Length Perturbation (VTLP). The preprocessing of the speech signals involved cleaning the data from ambient and unwanted noises. We extracted and selected various spectral features, including mel-frequency cepstral coefficients (MFCC) and Linear Prediction Cepstral Coefficients (LPCC). Subsequently, we applied several classification methods: Support Vector Machine (SVM), Bidirectional Long Short-Term Memory (BiLSTM), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Random Forest. Our experiments demonstrated that the Random Forest classifier achieved the highest F-score of 58.75%. The results were thoroughly discussed, analyzed, and compared across the five models using different feature extractions. This study provides valuable insights and advancements in the SER field, particularly for the TD, future research directions for improving emotion recognition systems.

Speech Emotion Recognition (SER) using Natural Language Processing (NLP) for underrepresented dialects faces significant challenges due to the lack of annotated corpora. This research addresses this issue by constructing and annotating SERTUS (Speech Emotion Recognition in TUnisian Spontaneous speech), a novel corpus of spontaneous speech in the Tunisian Dialect (TD), collected from various domains such as sports, politics, and culture. SERTUS includes both registers of TD: the popular (familiar) register and the intellectual register, capturing a diverse range of emotions in spontaneous settings and natural interactions across different regions of Tunisia. Our methodology uses a categorical approach to emotion annotation and employs inter-annotator agreement measures to ensure the reliability and consistency of the annotations. The results demonstrate a high level of agreement among annotators, indicating the robustness of the annotation process. The study's core contribution lies in its comprehensive and rigorous approach to the development of a dataset of spontaneous emotional speech in this dialect. The constructed corpus has significant potential applications in various fields, such as human-computer interaction, mental health monitoring, call center analytics, and social robotics. It also facilitates the development of more accurate and culturally nuanced SER systems. This work contributes to existing research by providing a high-quality annotated corpus while emphasizing the importance of including underrepresented dialects in NLP research.

artificial intelligence (ai) tools such as deepseek r1 and chatGPT 4.5 have emerged as promising aids in arabic-english literary translation. This study aims to compare the translation performance of these two systems using a mixed-methods approach. Quantitative analysis was conducted through five established evaluation metrics, Bleu, coMeT, MeTeor, BerTscore, and sacreBleu, to assess accuracy, fluency, and semantic coherence. complementing these measures, a qualitative evaluation was carried out with 80 undergraduate students from cadi ayyad university, who critically assessed anonymized ai-generated translations against their versions and published human translations using a structured rubric. results indicate that deepseek r1 achieved consistently higher automated metric scores across literary genres (novels, plays, poems). however, qualitative analysis highlighted persistent challenges in pragmatic coherence, cohesion, and emotional depth, especially in poetry and dramatic texts. While deepseek r1 demonstrated potential in lexical accuracy and fluency, significant human intervention remains essential for achieving high-quality literary translations. Future research should integrate larger-scale human evaluations to comprehensively assess the capabilities and limitations of ai translation tools in diverse literary contexts.

Named Entity Recognition (NER) is among the main tasks of Natural Language Processing (NLP). NER is a critical and fundamental component for several NLP applications including Information Retrieval (IR), Question-Answering (QA) and Machine Translation (MT). While several NER models for formal languages such as English and Modern Standard Arabic (MSA) have emerged, Arabic dialects remain in their infancy. We noted that the most recent researches focused on the study of a single Arabic dialect, hence the absence of a perfect multidialectal NER model. In this paper, we present the ARDIAL-BERT, the first multidialectal NER model which was built upon a continuous pretraining and then a finetuning on Arabic dialect publicly available datasets grouped by region (Levantine, Maghrebi, Egyptian and Gulf). The model was built upon the last updated version of BERT transfer transformers after several experiments on various BERT NER models. We approached our contribution on two different tasks: first we built ARDIAL-NER, an Arabic multidialect dataset extracted from existing NER datasets. ARDIAL-NER was manually annotated and contains a total of 53,539 entities of 369,372 tokens composing 21683 sentences. Second, we conducted a continual pretraining process using additional unannotated data, and then we guided a finetuning on new annotated NER datasets. The continuous learning system can be applied at different levels: updating model parameters, incorporating new language data, and training on new labels. Our results demonstrate its effectiveness to varying degrees. Our approach showed that it exhibited a greater ability by achieving superior results compared with both baselines and previous models. This demonstrates the capabilities of grouping Arabic dialects by region and the good selection of data that matches well with the baseline transformers. Impact Statement-ARDIAL-BERT represents a significant breakthrough in Arabic multidialectal Named Entity Recognition (NER), providing the first comprehensive model capable of accurately recognizing entities across major Arabic dialects, including Levantine, Maghrebi, Egyptian, and Gulf dialects. By leveraging a novel approach that combines continual pretraining with finetuning on regionally grouped datasets, this work addresses a critical gap in Natural Language Processing (NLP), enabling advancements in applications such as information retrieval, machine translation, and question-answering for Arabic dialects.

Irony and satire in Orwell's Animal Farm are lexically investigated in the current paper, in order to find out the correlation between both concepts. The researcher adopts a qualitative method of analysis, focusing on chapter 10. The findings show that there is an integrative relationship between irony and satire in the novel, where they both portray the image of corrupt leaders. While irony shows the contrast between their subjects' life before and after them, and how it gets worse, satire displays the common characteristics between in corrupt animal leaders and those of humans.

Paraphrase plagiarism is a significant and widespread problem and research shows that it is hard to detect. Several methods and automatic systems have been proposed to deal with it. However, evaluation and comparison of such solutions is not possible because of the unavailability of benchmark corpora with manual examples of paraphrase plagiarism. To deal with this issue, we present the novel development of a paraphrase plagiarism corpus containing simulated (manually created) examples in the Urdu language - a language widely spoken around the world. This resource is the first of its kind developed for the Urdu language and we believe that it will be a valuable contribution to the evaluation of paraphrase plagiarism detection systems.

In this paper, we describe the process of building
a corpus for Tunisian Speech Emotion Recognition (SER). To
the best of our knowledge, it is the first work in the SER field
that uses spontaneous speech emotion in Tunisian dialect. SER
represents an active research problem in the field of Natural
Language processing (NLP). It aims to detect different emotions
such as satisfaction, frustration and anger from audio speeches
using various classifiers. Speech signal preprocessing is the first
and the most important step in the SER process. Moreover,
Pre-Processing of Speech is very crucial in the applications
where silence or ambient noise is completely undesirable. Voice
activity detection is a common procedure that plays a key role in
preprocessing speech signals and noise cancellation. Pre-emphasis
of speech helps the system be computationally more effective
[1].This work proposed a preprocessing method to extract fea
tures from natural Tunisian speech. Speech preprocessing consists
of cleaning the speech signal from ambient and unwanted noises,
detecting speech activity and normalizing the length of the vocal
tract.

The field of artificial intelligence (AI) is evolving at an unprecedented pace, with transfer learning and transformer-based models now forming the backbone of many state-of-the-art systems.
This book, 200 Questions About Transfer Learning and Transformers, is written to help learners, practitioners, and enthusiasts navigate this exciting landscape through a clear, question-driven format. Rather than presenting dense theory or overwhelming technical detail, the book offers focused expla- nations—each grounded in a specific, practical question—making the material accessible and easy to absorb.
Let me briefly share my own journey. I began in applied mathematics and statistical modeling, later transitioning into data science, machine learning, and natural language processing. As the field matured, I saw how transfer learning and transformers revolutionized both research and industry. I also noticed a recurring challenge: while many people wanted to understand these tools, they struggled to find clear, structured answers to common questions.

Question Answering (QA) is a specialized area in the field of Information Retrieval (IR). The QA systems are concerned with providing relevant answers in response to questions proposed in natural language. QA is therefore composed of three distinct modules, each of which has a core component beside other supplementary components. These three core components are: question classification, information retrieval, and answer extraction. Question classification plays an essential role in QA systems by classifying the submitted question according to its type. Information retrieval is very important for question answering, because if no correct answers are present in a document, no further processing could be carried out to find an answer. Finally, answer extraction aims to retrieve the answer for a question asked by the user. This survey paper provides an overview of Question-Answering and its system architecture, as well as the previous related work comparing each research against the oth...

In the Philippines, parents refused their children having an anti-measles and anti-dengue vaccines, which created a medical outbreak. This may not happen if product warnings have been given and explained to the parents. Indeed, product warnings are found to be in their optimal position in safeguarding the life of consumer-patients. This paper anatomizes the lexical features of medicine product warnings in the Philippines which are crucial in the response discourses. A range of linguistic frameworks were applied and significant findings were drawn. Gaps were identified on the use of noun abstractness, synthetic personalization, field continuum, adjectives, and adverbs. Such an investigation brought up the transparency of communicative features of medicine safety texts. In the end, linguistic components create a vital impact on the legal content adequacy of medicine product warnings, unfolding the vitalities of these messages in facilitating informed decision-making among consumer-pat...

Websites are regarded as domains of limitless information which anyone and everyone can access. The new trend of technology has shaped the way we do and manage our businesses. Today, advancements in Internet technology has given rise to the proliferation of e-commerce websites. This, in turn made the activities and lifestyles of marketers/vendors, retailers and consumers (collectively regarded as users in this paper) easier as it provides convenient platforms to sale/order items through the internet. Unfortunately, these desirable benefits are not without drawbacks as these platforms require that the users spend a lot of time and efforts searching for best product deals, products updates and offers on ecommerce websites. Furthermore, they need to filter and compare search results by themselves which takes a lot of time and there are chances of ambiguous results. In this paper, we applied web crawling and scraping methods on an e-commerce website to obtain HTML data for identifying p...

The widespread dissemination of fake news across digital platforms has emerged as a critical issue, undermining public trust and influencing societal discourse. This paper presents a lightweight yet effective fake news detection system using natural language processing (NLP) techniques and logistic regression. We constructed a labeled dataset of news articles and employed preprocessing steps such as tokenization, stopword removal, and TF-IDF vectorization to extract informative text features. Based on these characteristics, the logistic regression model, the main classifier, was trained to distinguish between authentic and fraudulent news. On held-out test data, our model performed well across important evaluation measures like precision, recall, F1score, and AUC, achieving an accuracy of over 95%. The precision-recall curve, ROC curve, and confusion matrix are examples of visual assessments that show how reliable the classifier is. Furthermore, we used Flask and Python to create an intuitive online interface that allows for real-time predictions on news inputs.Comparative analysis with state-of-the-art models such as BERT and ensemble classifiers is included, confirming that even a simple linear model, when combined with appropriate text representations, can yield competitive results. This work highlights the potential of interpretable, resource-efficient models in combating misinformation and lays a foundation for future improvements using deep contextual embeddings and hybrid approaches.

In this work, we examine the limitations of digital tools in facilitating cross-linguistic and crosscultural research from a humanistic perspective. Our primary objective is to draw comparisons between the TenTen corpora, assessing their degree of similarity. In order to achieve this goal, we will conduct cluster analysis on the 43 corpora within the TenTen Corpus Family using a set of parameters that characterize this family membership. This analysis pinpoints the TenTen corpora that exhibit the most similar characteristics within the family, bringing to the surface an implicit hierarchy within the Sketch Engine platform, a multilingual digital tool environment. This hierarchy is structured into four distinct clusters, definable by size, number of functional tools, versions, and Part-of-Speech (PoS) tagging. The findings of the current study call for prudence when comparing the TenTen corpora, but also suggest a way of improving a multilingual environment; the examination and establishment of connections among the TenTen corpora are imperative for a comprehensive understanding of multilingualism in Digital Humanities. Digital Studies/Le champ numérique is a peer-reviewed open access journal published by the Open Library of Humanities.

This paper explores the interplay between artificial intelligence (AI) in natural language processing (NLP) and linguistics, offering NLP engineers actionable methodologies (e.g., syntactic probes, evaluation metrics) and linguists theoretical insights (e.g., cognitive plausibility, universal grammar). By examining historical foundations, modern synergies, and critical challenges, we propose interdisciplinary strategies to enhance model robustness and test linguistic hypotheses at scale, fostering collaboration across fields.

The increasing diffusion of misinformation in online media has raised alarm as a significant threat to information credibility and societal trust. The ease of disseminating false information across social media platforms, news websites, and digital forums has led to severe consequences, including political manipulation, financial fraud, and public misinformation. This research outlines a robust strategy for detecting misinformation using Natural Language Processing. To take the detection process a step further, the study analyses the implementation of ensemble models. Ensemble learning combines multiple classifiers to improve generalization and robustness, reducing the likelihood of misclassification and helps the model focus on critical words and phrases that contribute to determining the authenticity of news articles. The operational efficacy of the model is measured exploiting standard evaluation assessment metric fidelity, precision, recall, and F1-score to confirm consistent performance. Inclusion of ensemble learning further improves classification accuracy by reducing biases inherent in individual models. Future work in this domain can focus on Live Monitoring for Misinformation, multilingual analysis, and incorporation of context-aware models to further refine detection capabilities. By continuously evolving NLP-based approaches, researchers and technology developers can serve an important function in mitigating the effect of misinformation on social dynamics.

Social media is used as a dominant source of news distribution among users. The world's preeminent decisions such as politics are acclaimed by social media to influence users for enclosing users' decisions in their favor. However, the adoption of social media is much needed for awareness but the authenticity of content is an unknown factor in the current scenario. Therefore, this research work finds it imperative to propose a solution to fake news detection and classification. In the case of fake news, content is the prime entity that captures the human mind towards trust for specific news. Therefore, a linguistic model is proposed to find out the properties of content that will generate language-driven features. This linguistic model extracts syntactic, grammatical, sentimental, and readability features of particular news. Language driven model requires an approach to handle time-consuming and handcrafted features problems in order to deal with the curse of dimensionality problem. Therefore, the neural-based sequential learning model is used to achieve superior results for fake news detection. The results are drawn to validate the importance of the linguistic model extracted features and finally combined linguistic feature-driven model is able to achieve the average accuracy of 86% for fake news detection and classification. The sequential neural model results are compared with machine learning based models and LSTM based word embedding based fake news detection model as well. Comparative results show that features based sequential model is able to achieve comparable evaluation performance in discernable less time.

In this paper we implement a document retrieval system using the Lucene tool and we conduct some experiments in order to compare the efficiency of two different weighting schema: the well-known TF-IDF and the BM25. Then, we expand queries using a comparable corpus (wikipedia) and word embeddings. Obtained results show that the latter method (word embeddings) is a good way to achieve higher precision rates and retrieve more accurate documents.

The emergence and subsequent development of deep learning, specifically transformer-based architectures, Generative Adversarial Networks (GANs), and attention mechanisms, have had revolutionary implications on Natural Language Processing (NLP) and multimodal learning. Transformer models are neural network architectures that change an input sequence into an output sequence. Transformer architectures like the Generative Pre-Training Transformer (GPT) and Bidirectional Encoder Representations from Transformers (BERT) leverage self-attention mechanisms to enable highlevel contextual learning as well as long-range dependencies. GANs are a kind of AI algorithm that is designed to solve generative modeling problems. Different GANs, such as StyleGAN and BigCAN, study a collection of training data and learn the distribution probabilities used to generate such datasets. Attention mechanisms, acting as the unifying thread between Transformers and GANs in multimodal learning, optimize deep learning models to attend to the most relevant parts of the input data. This paper explores the synergy between these technologies, emphasizing their combined potential in multimodal learning frameworks. In addition, the paper analyzes recent advancements, key innovations, and practical implementations that leverage Transformers, GANs, and attention mechanisms to enhance natural language understanding and generation.

The paper presents the creation of an end-toend voice assistant system designed for a lesser-resourced dialect of Arabic, Libyan Tripolitanian, which does not receive local support in commercial ASR and NLP applications. To remediate this lack, we built a demographically balanced and phonemically rich corpus of speech data containing over 13,000 audio samples. It contains both natural and semi-structured utterances and is annotated using the CODA* orthography for dialectal Arabic. Using this dataset, we trained the OpenAI Whisper model with the Hugging Face Transformers, achieving a WER (Word Error Rate) reduction of 2.045 → 0.040. To assist in managing smartphone commands and having simple conversations in Tripolitanian Arabic, the ASR output is passed to a Rasa-based chatbot that is trained on intent-annotated queries. The chatbot was able to perform with 100% intent accuracy and a 0.998 entity F1-score. This modular pipeline is confirmed by evaluation results on standard ASR and NLU metrics. These findings show that it is possible to create high-performance, specific voice interfaces based on training for specific dialect inquiry through domain-adapted training, data augmentation, and system integration. Future expansions include extending the dataset to suit use in speech synthesis in the Libyan dialect and broader Libyan dialect support.

This study introduces a novel approach to sentiment classification of Arabic tweets regarding educational reforms in Saudi Arabia. The complexity of the Arabic language, with its numerous dialects, poses challenges for natural language processing tasks, particularly when large volumes of data require manual annotation. To overcome the limitations of traditional labeling methods, we developed a weakly supervised learning framework that combines LLMs (GPT-3.5) and pre-trained language models (MarBERT and XLM-RoBERTa) to generate high-quality weakly labeled training data using the Snorkel framework. We fine-tuned the AraBERT model with this weakly labeled data for sentiment classification. Our experimental results demonstrated the effectiveness of the proposed approach, achieving 83% precision, 76% recall, and an 85% F1 score in classifying tweets as positive, negative, or neutral. Comparative analysis showed that GPT-3.5 outperformed Llama 2 in prompting tasks, and our weakly supervised model surpassed baseline machine learning methods. These findings highlight the potential of weakly supervised learning in analyzing public opinion on Arabic social media platforms without relying on large, labeled datasets. INDEX TERMS AraBERT, large language models, natural language processing, social media data, weak supervision, weakly training data.

This paper describes a comprehensive set of experiments conducted in order to classify Arabic Wikipedia articles into predefined sets of Named Entity classes. We tackle using four different classifiers, namely: Naïve Bayes, Multinomial Naïve Bayes, Support Vector Machines, and Stochastic Gradient Descent. We report on several aspects related to classification models in the sense of feature representation, feature set and statistical modelling. The results reported show that, we are able to correctly classify the articles with scores of 90% on Precision, Recall and balanced F-measure.

This paper presents Small Synthetic Embedding Dataset, a fully synthetic dataset in Ukrainian designed for training, fine-tuning, and evaluating text embedding models. The use of large language models (LLMs) allows for controlling the diversity of generated data in aspects such as NLP tasks, asymmetry between queries and documents, the presence of instructions, support for various languages, and avoidance of social biases. A zero-shot generation approach was used to create a set of Ukrainian query-documents pairs with corresponding similarity scores. The dataset can be used to evaluate the quality of multilingual embedding models, as well as to train or fine-tune models to improve their effectiveness when working with Ukrainian texts. The paper covers a comprehensive description of the dataset construction process, including the parameters influencing the diversity of generated texts, the large language models used for actual generation of the data, and an example of using the dataset to evaluate and compare selected multilingual embedding models on the task of semantic text similarity. Unlike existing Ukrainian datasets, which are mainly based on real texts, SED-UA-small is fully synthetic, providing greater flexibility in controlling the diversity and specificity of data for the needs of training and evaluating embedding models, and allowing for feast and cost-effective expansion of the dataset with high-quality entries if needed. We used a combination of open and proprietary large language models of different sizes to generate the first version of the dataset, consisting of 112 thousand text pairs, divided into training (~50 %), testing (25 %), and validation (25 %) sets. The data is publicly available at https://huggingface.co/datasets/suntez13/sed-ua-small-sts-v1.

Automatic Text Classification is a machine learning task that automatically assigns a given text document to a set of pre-defined categories based on the features extracted from its textual content. Most online communication forums, including social media, enable users to express themselves freely, and most times, anonymously. The ability to freely express oneself is a human right that should be cherished, but people always induce and spread hate or illegal words towards another group as an abuse of this liberty. For instance many online forums such as Facebook, YouTube, and Twitter consider hate speech harmful, and have policies to remove hate speech content. This paper attempts to automatically classify the textual entries made by bloggers on various topics into hate speech and non-hate speech. This was achieved by following steps like pre-processing, feature extraction and support vector machine classification. Empirical evaluation of this binary classification has resulted in anaccuracy of approximately 83% over the test set. In addition to classifying the textual entries of the blogs, it is proposed that the extracted features themselves be further classified under more meaningful heads which results in generation of a semantic resource that lends greater understanding to the classification task. This semantic resource can be used for data mining requirements that arise in the future.

Arabic nouns can be marked for definiteness or indefiniteness. The definite article is the prefix ''Al-,'' which confines the determiner class to a single element ''Al-.'' This topic is generally discussed under noun inflections, such as Gender, Number, Definiteness, and Case (GNDC), in grammar textbooks. The primary aim of this paper is to expand the Arabic determiner class (DET) by incorporating additional lexical items and providing a detailed description of their syntactic context within noun phrases (NPs). In traditional grammar, these lexical items are typically classified as noun adjectives and, at best, are referred to as noun specifiers since they modify the head noun. However, these ''nouns'' exhibit limited inflection compared to regular adjectives and occupy a specific position within the NP sequence. Additionally, the modified head nouns are constrained in their inflectional attributes, Number, and Definiteness. Our approach is qualitative and guided by morpho-syntactic considerations. We conduct an in-depth analysis of the grammatical features of ten lexical items, focusing particularly on the dependencies between the determiner and the following noun. This analysis also addresses some semantic properties. By focusing on context-sensitive grammatical rules, the study shows how these methods can enhance precision in parsing and reduce ambiguity in NLP tasks, highlighting the potential for developing more refined grammar for Arabic. This work is a prototype for comprehensive studies in linguistics and NLP.

Arabic nouns can be marked for definiteness or indefiniteness. The definite article is the prefix ''Al-,'' which confines the determiner class to a single element ''Al-.'' This topic is generally discussed under noun inflections, such as Gender, Number, Definiteness, and Case (GNDC), in grammar textbooks. The primary aim of this paper is to expand the Arabic determiner class (DET) by incorporating additional lexical items and providing a detailed description of their syntactic context within noun phrases (NPs). In traditional grammar, these lexical items are typically classified as noun adjectives and, at best, are referred to as noun specifiers since they modify the head noun. However, these ''nouns'' exhibit limited inflection compared to regular adjectives and occupy a specific position within the NP sequence. Additionally, the modified head nouns are constrained in their inflectional attributes, Number, and Definiteness. Our approach is qualitative and guided by morpho-syntactic considerations. We conduct an in-depth analysis of the grammatical features of ten lexical items, focusing particularly on the dependencies between the determiner and the following noun. This analysis also addresses some semantic properties. By focusing on context-sensitive grammatical rules, the study shows how these methods can enhance precision in parsing and reduce ambiguity in NLP tasks, highlighting the potential for developing more refined grammar for Arabic. This work is a prototype for comprehensive studies in linguistics and NLP.

Fake News (FN) dissemination on websites and online platforms influences human behaviours, sociopolitical domains, and the sovereignty of a country. The outpour of biased news and propaganda on online portals can be addressed by restricting online propaganda using an automated mechanism. Proving the authenticity of news and information on online platforms in regional languages, such as Urdu, with limited resources and datasets, is challenging. Furthermore, limited research in resourceconstrained languages has created language bias in Artificial Intelligence (AI) research, which is concentrated in this study. Natural Language Processing (NLP) techniques have been used for Fake News Detection (FND) for English news and various language-related tasks. Previous studies used Machine Learning (ML), Deep Learning (DL), and individual Pre-trained Language Models (PLMs) for Urdu FND. ML-based ensemble model showed better performance than pre-trained models for Urdu FND. We propose a methodology for Urdu FND by applying stacked ensemble learning of PLMs, ELECTRA, mBERT and XLM-RoBERTa after apposite fine-tuning and hyperparameter optimization. To overcome the limitations of each pre-trained transformer model, these are fine-tuned individually using a publicly available Urdu dataset. The prediction performance results of the proposed stacking approach surpass the performance of each pre-trained model. An Accuracy of 0.914, a Matthews Correlation Coefficient (MCC) value of 0.898, and an F1-score of 0.904 validate the efficacy of the proposed ensemble model.

This article gives mathematical pseudocodes for large language model training based on a dataset from a corpus, inference, and chat with possibly lengthy human prompts and generated replies. It introduces the concepts of "microlect" and "resonant community." Most generally, a microlect is a specialized behavioral "language" consisting of expressions called "cores" built upon units called "keys." It is a minimal and adequate representation of a mental model. An expression may resonate more or less with a mental model. A resonant-community is a set of human beings whose expressions resonate with one another. These notions are helpful in understanding of how the "frozen matrices" of a large language model are a conduit for resonance between human-beings and a zeitgeist represented in a corpus. The primary objectives are to encapsulate the functionality of large language models in a succinct technical representation, and to explain how this functionality connects human beings to their ambient cultural zeitgeist as represented in a textual corpus. The research draws on relevant literature in psychology, linguistics, mathematics, physics, and software engineering. The significance of the research is to embed the explanation in a perspective on the entirety of human expressiveness. The study can provide valuable insights based on resonances between human beings among themselves and with computing machines.

The goal of natural language processing (NLP), which has recently gained popularity, is to improve the capacity of computers to comprehend and interact with human language. Consequently, to converse using natural language, it is crucial that spoken language be grammatically correct, especially for Geéz language. Geéz language sentences must follow certain norms of agreement in terms of number, person, gender, tense, and other factors to be considered grammatically correct. If the input sentence in Geéz language is improper, then it can have problems with subject-verb agreement, object-verb agreement, adjective-noun agreement, and adverb-verb agreement. The goal of the proposed work is to provide a neural machine translation approach for detecting and correcting grammar errors in Geéz sentences. We have prepared manually 11,490 Geéz parallel corpuses (Geéz language grammatically incorrect and grammatically correct sentences). After we have prepared a parallel Geéz sentence, we have used normalization, tokenization, padding, and one hot encoding as preprocesses. We have used two deep learning algorithms, including a bidirectional long shortterm memory encoder-decoder and a long short-term memory encoder-decoder, for training the proposed model. Keras and TensorFlow were used for importing the required libraries, and we used the Python 3.7 environment for implementation. Two test cases are used for the evaluation technique. The first one is for the long short-term memory encoder-decoder model, and the second one is for the bidirectional long short-term memory encoder-decoder model. Finally, the bidirectional long short-term memory encoder-decoder model achieved best results with an accuracy of 82%, recall of 82%, precision of 85%, and F1-measure of 83% with balanced error type classes. CCS Concepts: • Computing methodologies → Machine translation ;

Financial Named Entity Recognition (NER) presents a pivotal task in extracting structured information from unstructured financial data, especially when extending its application to languages beyond English. In this paper, we present AMWAL, a named entity recognition system for Arabic financial news. Our approach centered on building a specialized corpus compiled from three major Arabic financial newspapers spanning from 2000 to 2023. Entities were extracted from this corpus using a semi-automatic process that included manual annotation and review to ensure accuracy. The total number of entities identified amounts to 17.1k tokens, distributed across 20 categories, providing a comprehensive coverage of financial entities. To standardize the identified entities, we adopt financial concepts from the Financial Industry Business Ontology (FIBO, 2020), aligning our framework with industry standards. The significance of our work lies not only in the creation of the first customized NER system for Arabic financial data but also in its potential to streamline information extraction processes in the financial domain. Our NER system achieves a Precision score of 96.08, a Recall score of 95.87, and an F1 score of 95.97, which outperforms state-of-the-art general Arabic NER systems as well as other systems for financial NER in other languages.

Question answering system (QAS) is essential to satisfy the need to query information available in various formats, including structured data (ontology, databases) or unstructured data (document, web). The QAS provides a correct response to the question asked by a user in natural language. QAS uses natural language processing (NLP) techniques to interface with the system user. In this paper, we survey various QAS such as Natural Language Interfacing to DataBases (NLIDB), ontology-based question answering and question answering systems for unstructured data. We give also statistics and analysis. This can help researchers to choose an appropriate solution to their issues. In case of insufficiency, they can propose new systems for complex queries and adapt or reuse QAS techniques for specific research issues. We give also our point of view on how can QAS deal with Arabic linked data.