Fake news detection: state-of-the-art review and advances with attention to Arabic language aspects

Search and survey methodology summary

The following search phrases were used to find literature from 2020 to 2024: “Fake news”, “Fake news approaches”, “Fake news life cycle”, “Fake news detection”, “Arabic fake news detection”, “Fake news Datasets”, “Optimal fake news dataset” and “Machine learning methods for Fake news detection”. The search was conducted on six scientific databases: IEEE, Springer Link, Taylor & Francis Online, ACM digital library, MDPI, and Elsevier. These databases were chosen considering their extensive coverage and relevance in the domain of computer science. Google Scholar was also searched to increase publication coverage. Search results were filtered based on the following criteria: articles had to be peer-reviewed, published from 2020 onward, and relevant to fake news detection. Relevancy was assessed by reviewing the abstracts, methodologies, and findings to ensure the inclusion of studies introducing novel datasets, advanced detection techniques, or addressing issues specific to Arabic content. Duplicate results were removed, and articles with limited applicability to fake news detection were excluded. In order to maintain focus on the most recent advancements, peer-reviewed publications published in 2020 and later were prioritized over earlier or irrelevant research.

Research questions

• RQ1: What domains and traits define fake news, and how do they influence its spread on social media?

• RQ2: How can understanding the fake news life cycle enhance detection methods?

• RQ3: How do linguistic, topic-agnostic, visual, and social approaches compare in detecting fake news, and what are their strengths and limitations?

• RQ4: How do social context and user behavior impact Arabic fake news detection?

• RQ5: Which machine learning models and datasets are most effective for detecting fake news in English and Arabic?

• RQ6: Do pre-processing techniques improve Arabic fake news detection accuracy?

• RQ7: What challenges exist in building reliable datasets for Arabic fake news detection?

• RQ8: How can early fake news prediction be improved, and which features are most influential?

• RQ9: What unique challenges does Arabic pose for fake news detection, and how can they be overcome?

Domains of fake news

The propagation of misleading content spans various domains, including politics, health, art, economy, and entertainment. Fake news may be more prominent in certain fields, such as politics, but its impact remains significant across all domains. Globally, events like COVID-19 and the 2016 U.S. presidential election have heightened the focus on detecting fake news. For instance, misleading content about global conflicts, such as the wars between Russia and Ukraine (Shin et al., 2023), the recent Hamas-Israel conflict (Shahi, Jaiswal & Mandl, 2024), the Facebook revolution (https://rb.gy/307rns) in the Arab region following the Arab Spring, and various other global events, underscores the widespread impact of misleading information. Table 1 shows examples of fake news and their respective domains. One recent example, illustrated in Fig. 3, involves fake news shared on a Facebook profile (http://tiny.cc/4toa001) about a genuine earthquake in the Arab region in 2023. The post falsely claimed that the “British Meteorological Center” predicted a 6.2-magnitude earthquake in the Arabian Gulf region and Yemen in the coming hours. Fatabyyano (https://rb.gy/fdf98v), a popular Arabic fact-checking organization, debunked this fabricated post.

Characteristics of fake news

Understanding the characteristics of misleading content propagation is crucial for individuals and governments to control, target, and reduce its impact. Especially with the widespread and intense use of social media, key characteristics of untruth news include: intention to deceive (Thota et al., 2018), sensationalist headlines with exaggerated or unbelievable claims (Wei & Wan, 2017), echo chambers reinforcing existing beliefs, lack of credible sources (Shao et al., 2018), rapid spread (Vosoughi, Roy & Aral, 2018), manipulated content, poor writing quality, biased writing supporting a particular point of view or agenda (Pennycook & Rand, 2021b), and distinct linguistic features, such as the use of more proper nouns, comparatives, conjunctions, and adverbs. For instance, word pairs, rather than single words, can better indicate fake news due to the significant role of semantic properties (Shrestha & Spezzano, 2021).

Fake news life cycle

It is important first to understand the fake news life cycle: creation, dissemination, early detection, and propagation. The way users interact, respond, share, support, or ignore news tainted with falsehoods is crucial. Figure 4 illustrates the fake news life cycle, spanning four stages: Creation involves fabricated, inaccurate, or misleading content to manipulate or entertain people through sensationalist or exaggerated headlines. Dissemination refers to the spread of fake content through various traditional and, more significantly, social media platforms, which facilitate the sharing of misleading content. Early Detection aims to prevent widespread propogation and minimize its impact based on real-time analysis (Zubiaga et al., 2016). A multimodal approach for early fake news detection was proposed, addressing the limitations of previous methods that relied on specific models. This approach focused on the characteristics of fake news content and propagation data, utilizing graph neural networks (GNNs) and BERT (Bidirectional Encoder Representations from Transformers). By extracting propagation patterns and analyzing news content for social media posts, the method achieved impressive results on public datasets (Sormeily et al., 2024). In contrast, research in the field of early fake news detection in Arabic is considered rare, with only a few projects, such as those recently presented by Qatar University, supporting this area (https://n9.cl/6oifi).

Knowledge-based approach

The knowledge-based approach is widely used for verifying the truthfulness of news by consulting experts, professionals, crowdsourced (https://www.fiskkit.com/), and fact-checking websites or organizations. Many fact-checking sites employ manual detection, such as FactCheck (https://www.factcheck.org/) and Politifact (https://www.politifact.com/). In the Arab world, for instance, several organizations have been founded for fact-checking, particularly in the fields of politics and health, such as Fatabyyano. Other organizations, like DaBegad (https://dabegad.com/), Matsda2sh (https://matsda2sh.com/), and Misbar, have also found developing automated techniques for fake news detection necessary, although these sites still operate primarily based on manual detection (Alkhair et al., 2019). Table 2 provides examples of popular online fact-checking organizations in English and Arabic. However, manual fact-checking is characterized by several key aspects. It is time-consuming, requires daily updates on ongoing news, and is subject to bias and subjectivity, such as in crowdsourced fact-checking, which is less reliable.

Linguistic approach

This approach focuses on grammar, syntax, and semantics using various computational techniques to analyze and understand human language. It analyzes linguistic patterns such as specific writing patterns in news content, sentiment patterns and word frequencies, stylometric features, lexicon textual characteristics (Zhang & Ghorbani, 2020), syntax, semantics, discourse levels and grammatical structure to detect anomalies in the data. Features that can be extracted based on the language approach include lexical features such as character and word counts, large and unique words, sentence features, phrases, punctuation, and parts of speech (POS) tagging. For example, extracting lexical features from Arabic textual content has yielded promising results, with 78% accuracy in identifying fake news in a study by Himdi et al. (2022). However, the linguistic approach has limitations in Arabic due to its rich morphology, diverse dialects, and lack of standardized spelling, particularly in dialectal texts. These challenges hinder feature extraction, complicate generalization, and impede the detection of nuanced semantics or authentic mimicry, reducing its scalability and robustness for fake news detection (Shaalan et al., 2019; Saadany, Mohamed & Orasan, 2020).

Topic-agnostic approach

In the topic-agnostic approach, the focus is more on identifying and flagging fake news regardless of the specific topic or content, such as trustworthiness of the news source, the writing style and tone, and, moreover, the presence of certain patterns or elements meant to elicit emotional responses (Horner et al., 2021). Psychological features, such as those related to social interactions or family and friends, and biological process features, such as health or sexual references, can be analyzed to address fake news. Presence of an author’s name, long words, and URLs are also considered as core elements for assessing source credibility, writing style which involve techniques used to create highly biased and one side news include: emotional language, selective facts, exaggeration, attacks against opposing people views (Hoy & Koulouri, 2021). However, when applied to the Arabic language, this approach faces challenges due to dialectal variations and ambiguous writing styles. The meaning is often concealed behind words, with shifts in tone and lack of clarity. The use of emotional cues and diverse writing styles across dialects further complicates the identification of patterns indicative of fake news (Shaalan et al., 2019).

Visual-based approach

Visual material can significantly enhance the credibility of a news article; textual content is not sufficient when used alone. Visual statistical modeling techniques and statistical features effectively assess news credibility. An overview of fake image detection is discussed, along with benchmarking tools by Sharma et al. (2023). Deep fake images are one of the key challenges of this approach, as they require specialized algorithms to analyze metadata, content, lighting, shadows, and facial expressions (Yu et al., 2021).

Social context approach

The social context approach relies on the social context of misleading content along with network-based features. Profile characteristics such as verified status, number of followers, account age, user behavior (including frequency of posts, retweet patterns, user interactions), temporal patterns (such as time of posting, temporal spread patterns), and integration types (including comments, likes, and shares) are considered. Additionally, it explores the propagation path of misleading content, such as content and contextual information, as investigated by Passaro et al. (2022). Various social media elements were introduced and investigated by Nielsen & McConville (2022). However, changes in user behavior, the nature of noisy data, account characteristics and their credibility, the propagation mechanisms of fake news, its ambiguity, the volume of real-time news, and its multilingual characteristics are all challenges faced by this approach (Shu, Wang & Liu, 2019; Shu et al., 2020a).

For Arabic fake news detection, a multi-approach strategy can be used. Knowledge-based systems could integrate Arabic fact-checking organizations for quicker verification (Murayama, 2021), while linguistic models analyze unique Arabic linguistic markers, such as emotional tone (Hamed, Ab Aziz & Yaakub, 2023). Topic-agnostic methods assess credibility across topics by tracking authorship and style patterns (Liu et al., 2021), visual-based tools can verify Arabic visuals and detect manipulated images (Giachanou, Zhang & Rosso, 2020), and social context analysis would examines Arabic social media behaviors, focusing on user profiles and interaction patterns to identify and track fake news spread (Shu, Wang & Liu, 2019), Fig. 5 illustrates these approaches and their primary applications in fake news detection.

Arabic fake news detection

The Arabic language presents challenges for natural language processing due to its features, including intricate grammar, dialectal differences, limited data availability, resource constraints, and complex morphology. For instance, the same morphological Arabic words can convey different meanings depending on their position and the diacritic marks used in the sentence (Nassif et al., 2022). Arabic is written from right to left and is considered a semantic language, rich in morphological structures. It consists of 28 letters, of which three are considered vowels: (أ, و, ي). Diacritical marks are placed on words to indicate pronunciation, providing a specific meaning to each word. For example, the word (عُقد) with a $\underset{.}{\mathrm{d}}$ammah on the first letter means “necklace,” while (عَقد) with a fat $\underset{.}{\mathrm{h}}$ah on the same letter means “contract.” Despite having the same letters, the meanings differ due to the type of diacritical mark used. Thus, Arabic often presents variations in meaning depending on the diacritical marks. The complexity of Arabic grammar and its rich semantics adds further complications, and with approximately 10,000 independent roots (Elkateb et al., 2006), semantic analysis becomes significantly harder. Arabic sentences can be nominal (subject–verb), like الولد أكل التفاحة (Al-walad akala at-tuffaha/“The boy ate the apple”), or verbal (verb–subject), like أكل الولد التفاحة (Akala al-walad at-tuffaha/“Ate the boy the apple”) with a free word order. This variability contrasts with the fixed subject–verb order in English, increasing the load on NLP techniques and ML models for better classification of fake news (Shaalan et al., 2019). Most words in the Arabic language are derived from what is called a “root”; some words cannot be derived and remain as they are, such as question words and pronouns. For example, words like (ملعب)/stadium, (لاعب)/player, and (يلعب)/plays share the root (لعب), meaning “play.” Other words, such as (تكنولوجيا)/technology, are borrowed from English and lack a root. Arabic also allows the addition of suffixes, prefixes, and infixes to words, often creating new meanings for the same word. For instance, the word (سعيدة), meaning “happy,” contains a suffix (the last letter). By removing it, the word changes from an adjective to a noun (Farghaly & Shaalan, 2009). In addition, the Arabic language still presents unique challenges, such as the use of dialectal Arabic (DA) by social media users instead of formal Arabic or Modern Standard Arabic (MSA). This usage, especially on social media, does not follow specific Arabic language rules and is instead used in a disorganized manner, often containing spelling errors and high noise. Additionally, it is highly diverse, with dialects varying widely between Arab countries, and even within the same country, there are multiple dialects between the north, center, and south. This results in a wide variety of potential words used in spreading fake content. Dialectal variation further complicates semantic analysis, introducing differences between MSA and regional dialects (Azzeh, Qusef & Alabboushi, 2024). For instance, “car” is سيارة ,مركبة (Markaba or Syarah) in MSA, and عربية (Arabeya) in Egyptian dialect or كرهبه (Karhba) in Tunisian. This causes semantic ambiguities, which require further investigation into various NLP solutions such as context-aware embedding models (Habib et al., 2021), dialect-specific lexicons (Bouamor et al., 2018), and advanced disambiguation techniques to accurately interpret and process the meaning in different contexts such as MADAMIRA, a morphological analysis and disambiguation tool (Nagoudi et al., 2020).

Prepossessing

Pre-processing Arabic text involves converting raw data into a suitable format for advanced classification. The quality of pre-processing is often an indicator of excellent results and performance for classification algorithms (El Kah & Zeroual, 2021). The process begins with an initial data review, followed by data cleaning, which includes removing missing values, punctuation, duplicate letters, numbers, spaces, non-Arabic digits, symbols, and non-Arabic words. Additionally, common Arabic stop words (https://github.com/mohataher/arabic-stop-words) are removed, along with diacritics (https://en.wikipedia.org/wiki/Arabic_diacritics). Diacritics include: (Tanween Fatha) ( ), (Tanween Damma) ( ), (Tanween Kasra) ( ), (Fatha) ( ), (Damma) ( ), (Kasra) ( ), (Sukun) ( ) and (Shadda) ( ). These marks are placed above or below letters to indicate short vowels, pronunciation, or emphasis. Repeated letters such as (ااااا) are reduced to a single occurrence. Normalizing Arabic letters such as (إ), (أ), and (ًأً) are changed to (ا). The letter (ة) is changed to (ه), and the letter (ى) is replaced with (ي) (Soliman, Eissa & El-Beltagy, 2017), as it reduces orthographic variations and ensures consistent text representation. This normalization minimizes noise and ambiguity, thereby enhancing model effectiveness in tasks like classification and search by treating similar forms uniformly. Tokenization is then performed on the data by splitting text into tokens. For example, the Arabic sentence فاز المنتخب الأقوى في المباراة would be tokenized into separate words: "فاز" ,"المنتخب" ,"الأقوى" ,"في" ,"المباراة". Some pre-processing operations involve stemming, which returns words to their original root using various techniques, such as Tashaphyne (https://pypi.org/project/Tashaphyne/): Arabic Light Stemmer (Matti & Yousif, 2023). Lemmatization refers to returning words to their root forms when they are identical both morphologically and semantically, unlike stemming, which reduces words to their root independently and without considering the meaning. Techniques like stemming, lemmatization, and stop-word removal yield better results in Arabic text classification, especially when used together (El Kah & Zeroual, 2021). Figure 6 illustrates the general steps for preprocessing Arabic text, which are crucial for effective fake news detection. The process begins with data cleaning through an initial data review, followed by the removal of missing values, punctuation, numbers, non-Arabic characters, diacritic marks, and extra spaces to ensure consistency and reduce noise. Normalization addresses variations in writing styles, while stop word removal eliminates non-informative words. The final step, tokenization, segments the text into smaller units for analysis. These steps are particularly important for Arabic compared to English, as its complex morphology, extensive use of diacritics, and dialectal variations can obscure linguistic patterns critical for classification models. Effective preprocessing ensures the input data is clean, uniform, and representative of Arabic’s unique linguistic features, enhancing the model’s capacity to accurately detect fake news.

Datasets

The availability of optimal datasets in the Arabic language remains limited, and the existing ones are not as accessible to researchers compared to non-Arabic datasets. Datasets are often limited to short tweets or a single domain, such as politics or health. The annotation process is sometimes unclear, with unbalanced classes, and the cleaning and pre-processing phase is not always sufficient. Moreover, there is a need for datasets that leverage Modern Standard Arabic (MSA) alongside Arabic dialects, with cross-dataset analysis, to enhance model accuracy and achieve generalization (Tommasi & Tuytelaars, 2015). To address this concern, researchers have developed Arabic datasets (Nagoudi et al., 2020). Investigators such as Righi et al. (2022) and Shaker & Alqudsi (2024) have adeptly translated high-quality English datasets into Arabic, allowing access to a broader range of labeled datasets. Researchers use careful translation and adaptation techniques to preserve linguistic and contextual accuracy, ensuring that these datasets maintain their integrity in Arabic while supporting cross-lingual analysis. Data augmentation (DA) techniques, such as synonym substitution, back-translation, paraphrasing, and style transfer, improve identification performance by exposing models to diverse language expressions. Generative adversarial networks (GANs) further enhance robustness by generating realistic fake news samples. However, challenges remain in ensuring the authenticity of generated samples, avoiding biases, and maintaining a balance between synthetic and real data to preserve dataset integrity and reliability (Kuntur et al., 2024). Data augmentation (DA) techniques have proven valuable in providing more generalized datasets and improving model robustness for Arabic fake news detection (Mohamed et al., 2024). Additionally, establishing partnerships with news and social media outlets provides access to authentic, well-balanced Arabic content for annotation and analysis (Nagoudi et al., 2020). Other strategies include using crowdsourcing to annotate data with diverse dialects, as presented by Alsarsour et al. (2018), merging datasets across domains (e.g., politics, health) to build richer, multi-domain resources that improve model generalization (Khalil et al., 2021). Future efforts should prioritize incorporating dialectal diversity into dataset development by including MSA alongside various regional dialects. Crowdsourcing can be employed for annotation, ensuring linguistic accuracy and cultural relevance through diverse annotator participation. Additionally, collaboration with news and social media outlets can facilitate access to authentic, well-balanced content, while leveraging advanced data augmentation techniques and cross-domain merging will further enhance dataset richness and model generalization.

Fake news early prediction

In Arabic-speaking countries, social media behavior and platform differences significantly impact the spread of fake news, making it essential to tailor detection models to these unique dynamics. Platforms such as WhatsApp, Facebook, Twitter, and Instagram play crucial roles in content dissemination (Agrawal & Sharma, 2021), with WhatsApp being a primary channel for news sharing in private, group-based contexts. Users often share unverified content, influenced by personal beliefs and social networks, amplifying misinformation. The linguistic diversity of Arabic dialects and regional cultural nuances further complicate detection, as content can take on different meanings across various dialects. Fake news spreads rapidly, often with little fact-checking, making early detection crucial to prevent harm. Early signals of fake news, such as questionable sources, unprofessional headlines, or unreliable content, should be flagged early on to prevent widespread dissemination (Cavalcante et al., 2024). Researchers indicate that personal information, for example, is the most significant indicator for early fake news detection (Liu & Wu, 2020). Given the fast-paced and heterogeneous nature of misinformation on social media, it is critical to develop automated systems capable of detecting fake news at an early stage. Analyzing the propagation of fake news over time and considering the reputation of publishers is a promising approach for early identification (Algabri et al., 2024). By incorporating measures like removing malicious accounts and providing fact-checking information, we can better control the spread and mitigate the impact of fake news. Thus, early prediction, propagation analysis, and half-truth detection are essential areas of research for tackling misinformation before it reaches a broader audience (Zannettou et al., 2019).

Feature extraction

Dynamic content that includes text, audio, images, links, and symbols poses significant challenges for feature extraction and classification models, various feature extraction methods are employed for numerical representation, such as TF-IDF, bag of words (BOW), and word embeddings like Word2Vec. Advanced contextual embeddings from transformers like BERT further enhance model performance (Shishah, 2022; Algabri et al., 2024; Mujahid et al., 2023). In Arabic text processing, models like AraVec, trained on data from Twitter, the Arabic Web, and Wikipedia articles, have shown strong results. Additional Arabic-specific embeddings, such as Mazajak (Farha & Magdy, 2019), trained on large datasets of tweets, also perform well in representing the language. Meanwhile, deep learning methods, including CNN, RNN, LSTM, and CNN-LSTM, bypass manual feature engineering by automatically learning and extracting relevant features from raw data, achieving effective representation of Arabic text (Mohamed et al., 2024). Furthermore, fake news detection models typically rely on feature extraction using independent approaches, such as linguistic or contextual methods. However, combining multiple feature extraction techniques has proven to yield promising results (Sahoo & Gupta, 2021; Almarashy, Feizi-Derakhshi & Salehpour, 2023), as these approaches complement each other and enhance classification accuracy.