Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 151 https://internationalpubls.com Scalable Fake News Detection: Implementing NLP and Embedding Models for Large-Scale Data Dr. Surjeet1, Dr.B.Vasavi2, Dr Padmesh Tripathi3, Vakaimalar Elamaran4, Ramya R5, Anandh A6 1Associate professor, Bharati Vidyapeeth's College of Engineering, New Delhi. surjeet.balhara@bharatividyapeeth.edu 2Associate Professor, Maturi Venkata subbarao(MVSR) Engineering college, Department of Information Technology, Hyderabad, vasavi.bande@gmail.com 3Professor, Department of mathematics, Delhi Technical Campus, Greater Noida, UP, India. padmesh01@rediffmail.com 4Associate Professor, Dept of IT, Kamaraj College of Engineering and Technology, Virudhunagar, Tamilnadu, India, vakaimalarit@kamarajengg.edu.in 5Associate Professor, Dept of CSE, Kamaraj College of Engineering and Technology, Virudhunagar, Tamilnadu, India, ramyacse@kamarajengg.edu.in 6Associate Professor, Dept of CSE, Kamaraj College of Engineering and Technology, Virudhunagar, Tamilnadu, India, anandhcse@kamarajengg.edu.in Corresponding author mail: vasavi.bande@gmail.com Article History: Received: 06-10-2024 Revised: 27-11-2024 Accepted: 04-12-2024 Abstract: This paper describes a scalable approach to fake news detection by employing Natural Language Processing and word embedding models for huge datasets. The collective work with different embeddings (Bag of Words, TF-IDF, Word2Vec, and Bidirectional Encoder Representations from Transformers BERT), extracting not only word frequency but also content relation in news articles. These embeddings are then combined with machine learning classifiers including logistic regression, random forests and neural networks to evaluate how different models perform. It is a scalable system using distributed processing frameworks to process large amounts of data and to enable large scale model training. Our methodology with widely adopted fake news datasets including PolitiFact and the LIAR dataset show superior classification results, in particular when employing deep learning-based embeddings such as BERT which outperforms traditional methods by accuracy and recall. The authors investigate the effect of text preprocessing methods (e.g. stop-word removal, tokenization) on classification results. Our findings call attention to the trade-offs required for launching large-scale fake news detection systems given a balance between model complexity and computational efficiency. Keywords: news, detection, system, embeddings, word, backdrop, combination, NLP, BERT, automated, tokenization, classification. INTRODUCTION The way people interact with news and media is seriously altered by the global circulation of content in the digital age. This has enabled information to be circulated faster than ever, however it has also brought a new set of hurdles in the form of fake news: misinformation or disinformation. Fake news that is often presented as if it were true and verified information intentionally spread to deceive society has become a global corrupting force, mobilizing public opinion during elections and obscuring the truth related to fundamental matters of the international community. When such content goes viral, the mailto:surjeet.balhara@bharatividyapeeth.edu mailto:vasavi.bande@gmail.com mailto:padmesh01@rediffmail.com mailto:vakaimalarit@kamarajengg.edu.in mailto:anandhcse@kamarajengg.edu.in mailto:vasavi.bande@gmail.com Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 152 https://internationalpubls.com consequences can be serious: Propagandists are able to break public trust in current journalism, tear societies even further apart or even organize violence and riots. That is why some applied areas (e.g., artificial intelligence, natural language processing and media studies) have been working hard towards the detection of fake news and extending effort in studying its propagation[1]. This is very challenging to identify fake news, as we know these days digital platforms are rapidly growing and especially social media. Social media algorithms reward engagement, sharing the content that gets people to click on it, like it, and share, whether or not those posts are accurate. Here, fake news spreads faster and reaches more people compared to accurate information. It was especially prominent for things like the 2016 U.S. presidential election and the Brexit vote, with people lapping up fake news stories, of which there were many. Given the growing sophistication of fake news, more sophisticated tools are needed for these tools to detect and flag such content in advance so that it cannot happen and cause societal damage. Given the sheer scale of content that is created daily, traditional means such as human fact-checking have definitely encountered limitations in this age. This is why automated detection systems for fake news are crucial; they can prevent the problem from continuing to spiral out of control. The automated detection of fake news faces exciting challenges which can be addressed by Natural Language Processing. Performing classification on the textual content NLP techniques can be applied, i.e., to classify the articles as real or fake based on the language of particular patterns and features. A key part of this process is word embeddings, which are numerical vectors that represent words and take into account the frequency of words and how they are related. In the course of recent years various word embedding techniques like BoW, TF-IDF, Word2Vec, GloVe and BERT are created for improving the accuracy of text classification tasks. These embeddings can then be used in combination with several machine learning algorithms like logistic regression, random forests, neural networks to build informative models to find fake news[2-5]. One of the significant hurdles faced when researching on developing automated fake news detection systems is scalability. Because of the increasing volume of online content that we are facing, fake news detection systems have to scale efficiently for large-scale datasets. It needs powerful algorithms in place to make this a possibility, and systems that can rise up to the challenge of processing huge sums of textual data on the fly. To cater for such demands, distributed computing frameworks like Apache Spark and Hadoop are widely used that can distribute the processing of large datasets across different machines. Also necessary to be scalable on a larger dataset are practices such as batch processing and using optimization algorithms so that the system is effective just with its fundamental operations[14]. Language itself is abstract and its meaning differs from one language to another. Fake news articles are designed to look similar in style and format as real or truthful content, hence a user can be forced to easily fall into their trap. These events are likely to include some accurate information but largely they may be inaccurate by the changing of scene and or a removal of vital hints. To establish whether something is fake news or not, systems for identifying fake news have to go beyond word frequency and take into account some more subtle laws of language like sentiment, coherence, relations between words. This is where advanced word embedding techniques such as BERT become important, since BERT can imbue the context with meaning using the trends of surrounding words in a sentence. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 153 https://internationalpubls.com Figure 1. Fake news on social media: from characterization to detection[3] Using Figure 1, deep learning and advances in neural networks have strengthened NLP models to detect fake news even better. They tend to be very effective for text classification when applied in combinations with types of models similar to Convolutional Neural Networks and Long Short-Term Memory networks, like both CNNs-LSTM and LSTM-CNN. Models like this can learn to recognize intricate patterns in data, even those spanning far across entire text examples, and are thus effective at contextually-based tasks. Given that it is based on a transformer architecture, BERT leverages its huge popularity within the fake news detection paradigm by being capable of dealing with sentence-level processing rather than word-by-word. Response Entity BERT can then capture bidirectional context, which essentially has access to the words before and after a target word within a sentence. Recent works employing deep learning models enhanced with sophisticated word embeddings have boosted the accuracy in identifying fake news dramatically[15-18]. Finally, even with the strides made in this field, there are still several open problems for detecting fake news. The most imminent stress point, as discussed before, is the trade off between model complexity and computational efficiency. However, one can only run models that are computationally expensive like BERT so many times before scale just becomes an issue. Less complex models like logistic regression or random forests are faster, but their performance on the fake news task is weaker. Thus, for the real-world large-scale applications, it is crucial to consider these wise trade-offs while designing the fake news detection systems. The performance of fake news detection systems is highly dependent on the quality and diversity of training data employed for their design, besides scalability and computational efficiency. Context- dependent fake news: It is a type of fake news specific to the context, which might not transfer well trained on one dataset or domain to another. E.g., a model trained on political news could fail to classify fake news in other domains, e.g., health or science. In that regard, it is necessary for researchers to train their models on different datasets covering various topics and styles. Secondly, the training data should be bias-free because biased training data led to biased predictions and consequently further increased the spread of misinformation[19]. Preprocessing is another important phase in the building of bogus news discovery systems. To perform natural language processing on a dataset, we need to first clean and preprocess the raw text of the text data as text cannot be used in its raw form by any machine learning algorithm. These are mainly used for natural language processing tasks like tokenization, stop-word removal and stemming or lemmatization. Tokenization is a step of transforming text into words or tokens where stop-word punctuation comes such as “the” or “and” that are frequently existing but do not deliver any end purpose. It is a method that reduces to the root word, so we can capture the meaning of a text data. However, even though preprocessing is necessary in the pipeline, we need to face some problems from Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 154 https://internationalpubls.com it as well. This inadvertent removal of useful phrases is particularly salient when "stop-words" are knocked out of an input example ( this happens by pruning the vocabulary, a process which only takes into account words separate from their function words; these functional words are often cherry-picked to be authoritative markers of whether or not news articles are fake) and mean disabling rows[20]. Figure 2. Estimated Trend of Publications from 2000 to 2023 in Fake news Detection and NLP One cannot deny the importance of feature engineering in fake news detection. Feature Engineering is the process of taking raw data and making useful features out of them, which provides better power to machine learning models. Using Figure 2, It shows that from the past 5 years the research on this topic is rapidly increasing so far. In the case of detecting fake news some features could be how many times a word or phrase is mentioned, if there are stylistic elements such as punctuation or capitalization and the sentiment of text. Research in recent years showed that we can really improve fake news detection systems accuracy by using stylistic features, which word embeddings learn but probably no more (this doesn't mean I am saying this is sufficient, but whatever). A good example of this is that many fake news articles attempt to elicit a strong emotion from their readers by using overly sensationalized terms or language. When these stylistic attributes are combined with the model, it can enable researchers to discern more effectively fake news from any actual one. Evaluation isn't an afterthought, though; no fake news detection system can fly without frequent testing. When a model is trained, it must be assessed so that an estimate of its performance on unseen data can be made. Main evaluation metrics are accuracy, precision, recall, F1-score. A way to quantify accuracy, precision and recall seeks to highlight how well the model performs in identifying real news versus fake news. The F1-score is a combination of precision and recall, which gives us a balanced model on the basis of its performance. Also, these measures of success are just a few of the standard ones the system has to be reliable and generalizable as well. A well-performing model in one specific dataset might not be sufficient to perform as well in many other contexts, so evaluating the model on multiple datasets is critical to warrant reliability. In this paper, we target scalable fake news detection systems by using NLP tools with advanced word embedding models for large volumes of data. The embedding techniques we will discuss are Bag of Words, TF-IDF, Word2Vec, Glove and BERT which can be attached to machine learning classifiers like logistic regression, random forests or neural networks. This system has been designed to scale thanks to the use of distributed computing, and optimization techniques in order to process massive amounts of data. The performance of the system is evaluated using benchmark fake news datasets Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 155 https://internationalpubls.com (PolitiFact and LIAR) on metrics like embedding category, factorization method used for cardinality reduction and gates classification accuracy. The results show that learnable embeddings such as BERT are more accurate compared to static embeddings in classifying fake news, and at the same time suggest the need for a balance between model complexity and computational efficiency in practical applications[21-28]. Overall, the task of detecting fake news is intricate and multifaceted and undoubtedly falls into the domain of sophisticated tools originating in NLP, machine learning, and data science. Building scalable systems for fake news classification that can work with growing data volumes will help to identify and address the problem of misinformation born in the modern age of the internet. However, the current paper aims to provide a practical and high-performance solution for this task, which fits the current environment. Thus, it can be said that the work in this document is relevant to the current research in the field since it provides a crucial tool to counteract the fake news. 1. RELATED WORK Fake news, in particular, has been a focus of much research in recent years; given the rise of misinformation fuelled by growing social media platforms which facilitate rapid information distribution. But it also comes because of increasing societal imperatives to solve the challenge of fake news detection at scale. In that context, several research studies have tried to use NLP techniques and machine learning algorithms in order to detect it. The remainder of the introduction section will include previous studies that have conducted research on fake news detection and its advantages and disadvantages regarding the current study. Foundation of Detection of Fake News: early work Project devoted to the detection of fake news initially relied on statistical methods and rule-based systems that leveraged domain-specific stylistic and syntactic characteristics of news articles. Research from Potthast et al. (2017) and Ahmed et al.[29] In (2018), a stylistic-based method for fake news detection which it identifies according to lexical features, such as part-of-speech tags, sentence structure and frequency of certain words. Early models trained on these features distinguish news based on the way it is written, rather than its content making them good at identifying some kinds of fake news, but poor at scaled data processing and generalization to diverse news topics. Moreover, we detail knowledge-based ones which seek to analyze third-party information resources that serve as the ground with which new content is verified. Shu et al. Crowd-sourcing & domain expertise-dependent approaches Holtz et al.[30] (2017) conducted a systematic review of methods using crowd-sourced verification and leveraging domain expert familiarity with the issue. Unfortunately, manual approaches are not scalable, and large datasets requiring intervention quickly become cumbersome to manage. The next step of these scalability issues was the move towards machine learning models. Word embeddings With NLP Approach A significant step towards fake news recognition nowadays is that NLP and word embedding techniques are being applied to turn the textual data into a numeric representation so that it can be processed by machine learning models. In the fake news detection tasks, some traditional embedding Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 156 https://internationalpubls.com methods like Bag of Words and Term Frequency-Inverse Document Frequency have been frequently adopted to represent textual information. For example, Hauschild and Eskridge (2024) applied the BoW model as well as the TF-IDF one to portray fake news over politically fact checked PolitiFact and LIAR datasets. BoW assumes words are independent of one another, and TF-IDF weights words based on their frequency across documents which is helpful to identify rare but meaningful terms in fake news. Although, word level models have limitations in terms of word independence and lack of awareness about the context. To fix this, more advanced embedding models including Word2Vec and Global Vectors for Word Representation (GloVe) were developed. Word2Vec employs a dense, low- dimensional neural networks that captures the context of words, by predicting the likelihood of a word appearing in a sentence given its surrounding words (skip-gram) or how likely similar words are to appear around it. It also models the local and global word co-occurrences by GloVe in a corpus, which can help to judge whether news is real or false through its context[31-37]. Truică and Apostol (2023) have used document embeddings of Word2Vec and GloVe models to recognize fake news and have shown a better output when compared with classical models [12]. Those embeddings were particularly good at learning subtle semantics: the detailed meanings of words and phrases that are often crucial for identifying fake news. Deep Learning Models In addition, the introduction of deep learning methods, especially Bidirectional Encoder Representations from Transformers (BERT), greatly helps solve the problem of fake news detection. Although BERT can analyze the context to the left and the right of a word in a sentence (bidirectional), this quality makes it well-suited for identifying nuanced differences in how fake news is written vs. real news. Using pre-trained embeddings from large corpora like Wikipedia and books, models built on BERT outperformed existing methods w.r.t accuracy and recall. Kaliyar et al. (2020) introduced FakeBERT, a system for fake news detection built on top of BERT, by adding extra layers in deep neural networks and improving their accuracy against standard fake news datasets. The system achieved state-of-the-art classification performance, especially with large datasets and intricate patterns of language. The works of other authors, e.g., Kula et al. (2021) incorporated BERT and Recurrent Neural Network (RNN) to increase robustness of fake news detection along with different text types [38]. On the other hand, works like those done by Hauschild and Eskridge [39] also demonstrated that neural network models based on deep learning (e.g., convolutional neural networks CNNs and long short term memory LSTMs) effectively captured sophisticated representations about text above simple word associations. To identify those models that are particularly useful to know, it was used on fake news using misleading narratives or fancy rhetoric. Fake News Detection for Scalability Deep Learning and NLP models have shown state-of-the-art performance in fake news detection, however scalability is a major limitation especially at scale as large datasets. The findings of Sadeghi et al. In a recent paper, Khabsa et. Al[40] (2020) demonstrates the issues of using standard machine Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 157 https://internationalpubls.com learning models for processing large datasets containing millions of articles. They highlighted the necessity for distributed processing frameworks capable of handling computation heavy embedding models like BERT and GloVe. To address this requirement, the present study illustrates how these methods can be scaled out with regard to fake news detection using distributed computing frameworks. Thanks to the integration with systems like Apache Spark, those datasets can be processed in parallel, allowing us to train intricate models like BERT on millions of news articles. This technique not only accelerates training of the model but also efficiently identifies fake news over platforms like social media, which churns out huge amounts of data every second. Source Objective Methodology Results Research Gap [5] ● Adapt fake news detectors to large language models era. ● Study interplay between human-written and machine- generated news. ● Evaluate fake news detectors trained in various scenarios ● Provide a practical strategy for robust fake news detectors ● Detectors trained on human- written articles perform well on machine-generated fake news. ● Detectors should be trained on datasets with lower machine-generated news ratio. ● Understanding interplay between human-written and machine-generated news. ● Detecting machine-generated fake news vs. human- written fake news. [6] ● Highlight dataset quality and diversity importance in fake news detection. ● Provide GitHub repository for accessible datasets in one portal. ● Dataset quality and diversity emphasized for model effectiveness. ● GitHub repository consolidates publicly accessible datasets for research. ● Dataset quality and diversity crucial for detection model effectiveness. ● GitHub repository consolidates publicly accessible datasets for research efforts. ● Dataset quality, diversity impact on model effectiveness ● Addressing biases, ethical issues, best practices in dataset creation [7] ● Investigate preprocessing techniques and model architectures for fake news detection. ● Test model performance on ● Preprocessing techniques and model architectures ● Deep learning (CNN, LSTM) and conventional ML (Random Forest, Gradient Boost) ● Investigated various preprocessing techniques and model architectures for fake news detection. ● Tested models on two ● Model implementation issues hinder effective fake news detection. ● Lack of clean, unbiased data poses a challenge in research. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 158 https://internationalpubls.com widely used datasets. widely used datasets, contributing to the field. [8] ● Develop sustainable AI solution for Fake News Detection. ● Utilize BERT technology to eradicate and control fake news. ● Backdated Neural Network Classifiers ● BERT technology combined with existing methods ● AI solution using BERT for Fake News Detection. ● Aims to detect, eliminate, and prevent threats from Fake News. ● Lack of discussion on real- world implementation challenges. ● Limited exploration of alternative AI models for fake news detection. [9] ● Evaluate LLM integration in fake news detection. ● Assess hybrid XGBoost model performance with LLM judgment. ● Conventional machine learning ● Large Language Models (LLMs) like ChatGPT-3.5 ● XGBoost model achieved 96.39% accuracy in fake news detection. ● Integration of ChatGPT-3.5 improved model performance significantly. ● Research gap in utilizing Large Language Models for fake news detection. ● Challenge of manually crafted features in conventional machine learning methods. [10] ● Utilize LLMs for news event detection framework. ● Evaluate impact of textual embeddings on clustering outcomes. ● Large Language Models (LLMs) combined with clustering analysis ● Cluster Stability Assessment Index (CSAI) for measuring clustering quality ● LLM embeddings with clustering yield best results. ● Post-event tasks provide meaningful insights for interpretation. ● Evaluate impact of textual embeddings on clustering quality. ● Introduce Cluster Stability Assessment Index (CSAI) for measuring clustering quality. [11] ● Develop model for detecting fake news. ● Assess effectiveness in recognizing false information. ● Supervised learning techniques used for model selection ● Naïve Bayes, Logistic Regression, and Random Forest algorithms applied ● Random Forest model showed best accuracy. ● Framework effectively detects fake news in various settings. ● Framework focuses on news sources and content credibility. ● Random Forest model shows highest accuracy in fake news detection. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 159 https://internationalpubls.com [12] ● Identify false news using advanced machine learning techniques. ● Improve accuracy and scalability in detecting misinformation. ● Advanced feature selection, classification algorithms, NLP approaches ● Hybrid stacking classifier: Random Forest, XGBoost, Logistic regression ● High recall rates and precision achieved. ● Improved accuracy and scalability in detecting fake news. ● Nuanced language patterns detection ● High false positive rates and scalability issues [13] ● Evaluate large language models in detecting fake news. ● Discuss implications for developers and policymaker. ● Statistical evaluation and case- by-case processing methods ● Zero-shot prompting for fair model comparison ● High- parameter LLMs effective in detecting fake news. ● Models with more parameters outperform those with fewer. ● Need for larger, diverse datasets to challenge advanced models. ● Integration of contextual and source credibility analysis for improvement. [14] ● Evaluate ChatGPT and Google Gemini models for fake news detection. ● Analyze strengths and limitations of each model for future enhancements. ● Evaluation of ChatGPT and Google Gemini models ● Comparative analysis and error examination of model strengths and limitations ● High performance metrics on LIAR dataset ● ChatGPT and Google Gemini models show substantial capabilities ● Comparative analysis highlights strengths and limitations of each model. ● Insights provided for future enhancements in fake news detection. Table 1. Literature review Comparison of the Models In experimental comparison amongst embedding from diverse models, it was noticed that deep learning based models with BERT and Word2Vec as embeddings yields better accuracy than traditional methodologies like BoW and TF-IDF in tackling detection of fake news. Hauschild & Eskridge 2024) performed the same comparison on the LIAR dataset, and found BERT had better average accuracy as well as recall than other embeddings. BERT works really well on this problem because BERT is able to capture some of the long-distance contextual dependencies between the words, which is often necessary to identify subtle manipulations of facts typical in fake news articles. Although deep learning models provide better accuracy, it is computationally expensive. The training of models like BERT is computationally intensive and requires an excessive amount of memory that may exceed resource capacity for many companies or organizations, especially in the case where real- Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 160 https://internationalpubls.com time applications are involved. As a result, many have become interested in investigating the trade- offs between model complexity and computational efficiency. The proposed study fits in this body of work meticulously, by both employing sophisticated models (BERT) and analyzing the effects of pre-processing techniques: stop-word removal, tokenization, stemming on classification outcome. The results seem to suggest that preprocessing steps can have wildly differing outcomes on the model, and that for certain models where context is key for determining a piece of fake news, they should not be done. The previous articles on Fake news detection: From simple stylistic based approaches to sophisticated deep learning models using word embeddings and contextual information In fact, with all social media and news outlets contributing to more data than we can imagine causing a considerable push for interoperability on scalability of these models. In this paper, we go beyond these approaches by constructing a scalable solution for fake news detection using word embedding models and distributed processing frameworks[41-59]. This work provides important guidance in large-scale automatic fake news detection systems by comparing model complexity, computational efficiency, and classification accuracy trade-offs. 2. PROPOSED METHODOLOGY Scaling Fake News Detection: A proposed methodology (NLP and Word Embeddings) Based on state- of-the-art text classification techniques, the approach is designed for scalability in practice and ability to process large-scale datasets in real-time environments. The framework comprises a set of components such as data preprocessing techniques, word embedding models, machine learning classifiers and distributed processing frameworks. Hereafter, an elaborate explanation shall be given for each component, then the experiments and used evaluation metrics to assess the performance of the system. The architecture of the fake news detection system embedding model is shown in Fig. 3. It is made of four main stages; ● Data Preprocessing: This is the phase where we go about cleaning the raw textual data so that it can be useful for further analysis. These steps involve such processes as removing stop-words, tokenizing the text, lemmatizing and vectorizing it. ● Word Embedding Models: The cleaned text is converted into numbers by the means of many word embedding model; Bag of Words (BoW), Term Frequency-Inverse Document Frequency (TF- IDF), Word2Vec, Bidirectional Encoder Representations from Transformers (BERT) etc. ● Machine Learning Classification: In this phase, we train different machine learning classifiers namely logistic regression, random forests and neural networks on the embedded data to predict whether the news is real or fake. ● Distributed Processing: High scalability by incorporating distributed processing frameworks like Apache Spark for handling large-scale datasets within the system. This enables the model to execute data in a very parallel fashion and hence is extremely useful for training models on high- dimensional enter facts. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 161 https://internationalpubls.com 1. Data Preprocessing We do preprocessing to make sure that the data is in a cleaner and structured shape so the Machine Learning models can understand it directly. In this study, we process the above textual data through various steps as: 1.1. Tokenization Tokenization : Split the raw text into individual words or tokens. For a document 𝐷 = {𝑑1, 𝑑2, … , 𝑑𝑛} where d_i is the i-th document in the dataset, tokenization will separate every document into its basic words as follows : 𝐷𝑖 = {𝑤1, 𝑤2, … , 𝑤𝑚} The first step of turning un-structured text into structured data is Tokenization (w_1, w_2,…, w_i: the i’th word in document D_i. ). 1.2. Stop-word Removal Stop-words are the words which usually do not have any semantic meaning and could introduce noise in the model. Net effect being: To reduce the dimensionality and thus make model training easier. The resulting document is: 𝐷𝑖 ′ = {𝑤1 ′ , 𝑤2 ′ , … , 𝑤 𝑚′ ′ }, 𝑤𝑖 ′ ∉ 𝑆𝑡𝑜𝑝𝑊𝑜𝑟𝑑𝑠 For each word w i in document D i, where D_i^' is the stop-word removed document and w_i^' is non- stop-word token. 1.3. Lemmatization and Stemming Figure 3. Flowchart of Proposed Methodology Lemmatization is the process of converting words to their dictionary or base form (i.e. lemma). For instance, "run", names like "running" and "ran" are made all to the root form. When we perform Lemmatization, the model will consider different cases of a word as single submission: 𝐿𝑒𝑚𝑚𝑎(𝑤𝑖) = {𝑤𝑟𝑜𝑜𝑡} It is a more semantically oriented process compared to stemming, which only removes common parts of the ending of the word, whereas lemmatization looks at words and considers whether they are nouns, verbs, adjectives, or adverbs. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 162 https://internationalpubls.com 1.4. Text Vectorization After the text is cleaned it needs to be converted into a machine readable numerical format for the machine learning models. Section III discusses several techniques of word embedding that are used in this study. Algorithm 1: Data Preprocessing and Embedding Generation 1. Input: Raw dataset 𝐷 = {𝑑1, 𝑑2, … , 𝑑𝑛}, where each document 𝑑𝑖 consists of a collection of words. 2. Output: Embedded vectors for each document 𝐸𝐷 = {𝑒1, 𝑒2, … , 𝑒𝑛}. Steps: 1. For each document 𝑑𝑖 ∈ 𝐷: o Tokenize document into words 𝑤 = {𝑤1 , 𝑤2, … , 𝑤𝑚}. o Remove Stop-words: Filter out common stop-words from the tokenized list. 𝑤 ′ = {𝑤𝑗 ∣ 𝑤𝑗 ∉ 𝑆𝑡𝑜𝑝𝑊𝑜𝑟𝑑𝑠, 𝑤𝑗 ∈ 𝑤} o Lemmatize/Stemming: Convert each word to its base form: 𝑤𝑙𝑒𝑚𝑚𝑎 = {𝑤𝑟𝑜𝑜𝑡 ∣ 𝑤𝑟𝑜𝑜𝑡 = 𝐿𝑒𝑚𝑚𝑎(𝑤𝑗)} o Convert to Lowercase: Ensure uniform casing for all words. 𝑤𝑙𝑜𝑤𝑒𝑟 = {𝑤𝑙𝑒𝑚𝑚𝑎 ∣ 𝑙𝑜𝑤𝑒𝑟(𝑤𝑙𝑒𝑚𝑚𝑎)} 2. For each cleaned document 𝑑𝑖 ′ ∈ 𝐷′: o Select Embedding Model: Choose between BoW, TF-IDF, Word2Vec, or BERT. o Generate Embedding: i. If BoW: 𝑣𝐵𝑜𝑊 = [𝑓(𝑤1), 𝑓(𝑤2), … , 𝑓(𝑤𝑛)] ii. If TF-IDF: 𝑣𝑇𝐹−𝐼𝐷𝐹 = 𝑇𝐹(𝑤𝑗 , 𝑑𝑖 ′) × 𝑙𝑜𝑔 𝑁 𝐷𝐹(𝑤𝑗) iii. If Word2Vec (CBoW/Skip-Gram): 𝑃(𝑤𝑗 ∣ 𝑐𝑜𝑛𝑡𝑒𝑥𝑡) = 𝑒𝑥𝑝 (𝑣𝑤𝑗 ⊤ ℎ) ∑ 𝑤𝑘∈𝑉 𝑒𝑥𝑝(𝑣𝑤𝑘 ⊤ ℎ) iv. If BERT: 𝐿𝐵𝐸𝑅𝑇 = − ∑ 𝑛 𝑖=1 𝑙𝑜𝑔𝑃(𝑤𝑖 ∣ 𝑤1, … , 𝑤𝑖−1, 𝑤𝑖+1, … , 𝑤𝑛) Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 163 https://internationalpubls.com 3. Store Embedding: o Collect the generated embeddings for all documents in 𝐸𝐷 = {𝑣1, 𝑣2, … , 𝑣𝑛}. 4. Return: Embedded document vectors 𝐸𝐷 . In Algorithm 1, the preprocessing of raw text data and word embedding using the principal models, BoW, TF-IDF, Word2Vec, and BERT need to be prepared. This covers term frequency and inverse document frequency, plus the loss functions for Word2Vec and BERT embeddings. 2. Word Embedding Models At the heart of the recommended approach is converting text data into interpretable numerical features. Word Embedding techniques convert words or text in a corpus to a vector of real numbers based on semantic meaning of the text. We use the following embedding methodology:- 2.1. Bag of Words (BoW) A Bag of Words model one of the simplest ways to turn text into vectors In BoW, every document is represented as a vector where each of the dimensions represents a particular word in the Vocabulary of the corpus. The dimensions will be counted n times of the word present in the document. 𝑣𝐵𝑜𝑊 = [𝑓(𝑤1), 𝑓(𝑤2), … , 𝑓(𝑤𝑁)] Where f(w_i ) is the frequency of word w_i in the document, and N is size of vocabulary. Still, BoW is primitive: it does not encode the contextual meaning of words and operates under a lax assumption of word independence. 2.2. TF-IDF (Term Frequency-Inverse Document Frequency) This is an improvement over BoW, where it modifies word frequencies with respect to the words in a corpus. So, they are extracted from the text file1 as follows: Term Frequency (TF): 𝑇𝐹(𝑤𝑖 , 𝐷) = 𝐹𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 𝑜𝑓 𝑤𝑖 𝑖𝑛 𝐷 𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑤𝑜𝑟𝑑𝑠 𝑖𝑛 𝐷 Inverse Document Frequency as: 𝐼𝐷𝐹(𝑤𝑖 , 𝐷) = 𝑙𝑜𝑔 𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠 𝑐𝑜𝑛𝑡𝑎𝑖𝑛𝑖𝑛𝑔 𝑤𝑖 TI-IDF score calculated as: 𝑇𝐹 − 𝐼𝐷𝐹(𝑤𝑖 , 𝐷) = 𝑇𝐹(𝑤𝑖 , 𝐷) × 𝐼𝐷𝐹(𝑤𝑖 , 𝐷) This method assigns words by their informative value when it comes to the document, thus minimizing the effects of frequent but uninformative terms. 2.3. Word2Vec This is exactly what Word2Vec, a neural network-based model does in that it gets to know the semantic relationships between words by trying to predict the context of a word given its surrounding neighbors. There are two modes of operation employed by Word2Vec: Continuous Bag of Words (CBoW), and Skip-Gram. In CBOW model, we predict word given its context, which are the surrounding words: Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 164 https://internationalpubls.com 𝑃(𝑤𝑖 ∣ 𝑤𝑖−1, 𝑤𝑖+1) = 𝑒𝑥𝑝(𝑣𝑤𝑖 ⊤ ℎ) ∑ 𝑤𝑗∈𝑉 𝑒𝑥𝑝 (𝑣𝑤𝑗 ⊤ ℎ) v_{w_i }: the vector representation of word w_ih: the hidden layer vector Skip-Gram: Given a word we will predict the surrounding words eg 𝑃(𝑤𝑖−1, 𝑤𝑖+1 ∣ 𝑤𝑖) = ∏ 𝑗=𝑖−1,𝑖+1 𝑒𝑥𝑝 (𝑣𝑤𝑗 ⊤ 𝑣𝑤𝑖 ) ∑ 𝑤𝑘∈𝑉 𝑒𝑥𝑝(𝑣𝑤𝑘 ⊤ 𝑣𝑤𝑖 ) This feature of Word2Vec gives it the ability to learn context-sensitive embeddings and to some extent semantic as well as syntactic but for more depth we can refer to GloVe. 2.4. Architecture: BERT (Bidirectional Encoder Representations from Transformers) BERT is a transformer-based model which means BERT can get the Full context of words (understand words better) because it takes both left and right context while representing the word in that sentence. Instead of treating words in isolation, the way that traditional language models do, BERT learns relationships between all words at once without requiring human labeling. BERT objective: to predict masked words (it must learn deep contextual representation!) 𝐿𝐵𝐸𝑅𝑇 = − ∑ 𝑛 𝑖=1 𝑙𝑜𝑔𝑃(𝑤𝑖 ∣ 𝑤1, … , 𝑤𝑖−1, 𝑤𝑖+1, … , 𝑤𝑛) where P(w_i∣⋅) is the probability of w given its context. Since BERT embeddings are pre-trained using massive corporate and can be fine-tuned on the tasks, they show great potential in detecting Fake News. 3. Machine Learning Classifiers After converting text into numeric vectors using the embedding models mentioned, we pass these vectors through machine learning classifiers to determine if a news article is real or fake. The research used logistic regression, random forests and a neural network classifier. 3.1. Logistic Regression Logistic regression is a linear classifier, in which the probability of a binary outcome is modelled as a function of input features. where the logistic regression encompasses from logistics importing Model: 𝑃(𝑦 = 1 ∣ 𝑥) = 1 1 + 𝑒𝑥𝑝(−(𝑤⊤𝑥 + 𝑏)) where, w is the weight vector, x is the feature vector, and b is a bias term. In this case, logistic regression works well enough for simple datasets but not when you deal with high-dimensional data as seen from word embeddings and thus couldn't capture the relations between the texts. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 165 https://internationalpubls.com 3.2. Random Forests Random Forests: Random forests are an ensemble learning method for classification, regressing and other tasks that operate by constructing a multitude of decision trees during training and outputting the mode of the classes predicted individually. The random forest model can work with high-dimensional inputs, and provide a defence against overfitting: �̂� = 𝑚𝑜𝑑𝑒({𝑇𝑖(𝑥)}𝑖=1 𝑁 ) where T_i is the i-th decision tree and y_hat the predicted label. 3.3. Neural Networks Neural networks form the most popular architecture for classification of data that can provide flexibility to capture complex patterns due to their layer by layer stacking of neurons. A Neural Network has the basic architecture of an Input layer, a few Hidden layers (1 or more), and an Output layer. When it comes to a binary classifier this is usually performed using the activation function sigmoid: �̂� = 𝜎(𝑊𝐿ℎ𝐿−1 + 𝑏𝐿) with W_L and b_L the weights and bias of the final layer, respectively, and h(L−1) is the activation of the penultimate layer. Fake news detection is complex neural networks are able to capture these non-linear relationships due to their extreme flexibility. 4. Distributed Processing for Scalability In order to make sure the fake news detection system can scale well, we have used distributed processing frameworks like Apache Spark. With this feature, the system can handle large datasets since it shares the processing to hundreds of nodes in a computing cluster. Advantages of Spark framework:- The underlying architecture of spark is optimised to parallelize the steps for both preprocessing and model training so that while dealing with large datasets the similar step can be processed on other parts concurrently. Scalability: Spark can scale to multiple nodes, which can make millions of news within minutes in real-time. If a particular node were to fail, spark is able to recover from these failures while keeping all your data and state progress through fault-tolerant properties of the architecture. The distributed processing integration ensures that our system proposal is also scalable for the large- scale real world datasets, such as the social media data. 5. Experimental Setup Experiments were performed to evaluate the performance of the proposed system using two well- known fake news datasets, PolitiFact and LIAR dataset. Both are datasets of labelled news articles tagged as either real or fake. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 166 https://internationalpubls.com In the testing phase, we evaluated model performance according to common classification metrics such as accuracy, precision, recall and F1-score. At rate 5, these metrics are defined as Accuracy: Percentage of all correctly predicted labels 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 i.e., TP = True Positive, TN = True Negative, FP = False Positive and FN = False Negative. Precision: the total number of positive instances that were actually predicted as positives out of all the predicted positive instances. 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃 𝑇𝑃 + 𝐹𝑃 Recall: The number of positive instances predicted correctly out of all actual positive instances. 𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃 𝑇𝑃 + 𝐹𝑁 F1-score: The mean of the Precision and Recall: 𝐹1 − 𝑠𝑐𝑜𝑟𝑒 = 2 × 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙 Experiments show that deep learning models, in particular the BERT model performed significantly better than the traditional logistic regression and random forests in terms of accuracy and recall. Although the most accurate model was that of BERT fine-tuned on LIAR dataset with 89% accuracy, 87% precision and 90% recall. This indicates that the context-aware embeddings work better for fake news detection. The proposed method for scalable fake news detection is to build a strong and powerful system using advanced NLP tools & machine learning classifiers, which can process large scale datasets. It offers scalability and real-time processing by using distributed processing frameworks, making it an ideal solution for social media platforms with high-throughput requirements. This acts as a very strong tool to stand against the increase in informational war with the use of different prepossessing of data along with advanced models using natural language processing and machine learning. 4. RESULTS AND EXPERIMENT This section describes the experimental setup, results and performance analysis of the proposed false information detection fake news detection model using various word embeddings to perform experiments with different machine learning classifiers We conduct experiments on two popular fake news datasets, PolitiFact and LIAR. We compared different word embedding techniques against a few machine learning models like Bag of Words, Term Frequency-Inverse-Document-Frequency, Word2Vec, and Bidirectional Encoder Representations from Transformers (BERT). We then conducted conventional evaluation based on accuracy, precision, recall and F1-score metrics for each of the models. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 167 https://internationalpubls.com The experiments were designed to evaluate the performance of conventional models and deep learning models along with different set embedding methods, so as to fit a scalable fake news detection approach. We also sized the system to see if it could scale with heterogeneously saturated grids data at larger scales by distributing processing frameworks, like Apache Spark. 1. Experimental Setup 1.1 Datasets For the experiments performed, begin by downloading two respective publicly available datasets. 1. PolitiFact Dataset- It is a dataset that consists of political-news labeled as real or fake. It is a popular dataset used in fake news detection research and suitable for binary classification tasks. 2. LIAR Dataset: The LIAR dataset is a collection of news statements from the political domain with labels in the form of “True”, “Mostly True”, ”Half True”, Full False, ”Barely True” and Half Flip. In turn, we simplified the labels as true (True, Mostly True & Half True) and fake (Barely True, False & Pants on Fire) for homogeneity and to maintain with that of the parallel dataset. Dataset # of News Articles # of True Articles # of Fake Articles Time Span PolitiFact 12,000 6,100 5,900 2007–2020 LIAR 13,000 7,500 5,500 2007–2019 Table 1: Summary statistics of the datasets used in experiments 1.2 Preprocessing Raw news articles were processed as stated above following the methodology before they were fed into machine learning models. You can perform steps like tokenization, cleansing the stop words, lemmatization and convert the text using the embedding techniques that you choose to numerical representation. During this experiment, embeddings that were using were the following: 1. BoW (Bag of Words): A simple frequency based approach. 2. TF-IDF: The importance of vocabularies in the whole corpus is weighed and measured. 3. Word2Vec (Variants: Continuous Bag of Words (CBoW) and Skip-Gram versions were employed.) 4. BERT: Contextual embedding learning based on deep learning, capturing word semantics. 1.3 Classifiers Classifiers. We used the following classifiers: Logistic Regression (LR): An algorithm for binary classification tasks. Random Forest (RF) : An ensemble method that builds multiple decision trees. Neural Networks( NN): A type of deep learning model based on fully connected layers. Support Vector Machine (SVM): A model that separate classes into two hyperplanes in a high- dimensional space. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 168 https://internationalpubls.com All the models were trained using Embedded data from various Word Embedding Techniques. In all experiments 80% of the data were used for training, and 20% for testing. All models were internally validated using 10-fold cross validation to control for overfitting. 1.4 Evaluation Metrics Using Figure 4, it represents that BERT model have highest performance among BoW, TF-IDF and Word2Vec.We used the following four metrics to evaluate the accuracy of our models: - Accuracy: The percentage of news articles that were predicted correctly • Precision: The model's ability to properly recognize positive instances (real news), Recall: How good the model is at finding all the positive instances (True news) F1 score is the harmonic mean of precision and recall, hence it gives a better measure of balance between them. Figure 4. Model Performance Comparison 2. Results 2.1 Architectures of Embedding Models Table 2 shows the results using different classifiers and Table 3 gives the results of classifier-type- specific differences as supported by each word embedding model on PolitiFact dataset and LIAR dataset respectively. The tables emphasize the accuracy, precision, recall and F1-score per combination of embedding technique/classifier. Model Embedding Accuracy Precision Recall F1-Score Logistic Regression BoW 80.5% 79.8% 81.0% 80.4% Logistic Regression TF-IDF 82.1% 81.3% 82.8% 82.0% Logistic Regression Word2Vec 85.7% 86.0% 85.4% 85.7% Logistic Regression BERT 89.3% 88.5% 90.0% 89.2% Random Forest BoW 81.2% 81.5% 80.9% 81.2% Random Forest TF-IDF 83.7% 83.2% 84.1% 83.6% Random Forest Word2Vec 87.9% 87.6% 88.1% 87.8% Random Forest BERT 91.0% 90.4% 91.3% 90.8% Neural Network BoW 82.5% 83.1% 81.9% 82.5% Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 169 https://internationalpubls.com Model Embedding Accuracy Precision Recall F1-Score Neural Network TF-IDF 85.9% 85.2% 86.4% 85.8% Neural Network Word2Vec 89.5% 89.2% 89.7% 89.4% Neural Network BERT 92.2% 91.8% 92.6% 92.2% SVM BoW 80.1% 80.2% 80.0% 80.1% SVM TF-IDF 83.0% 82.4% 83.6% 83.0% SVM Word2Vec 86.8% 86.5% 87.1% 86.8% SVM BERT 90.1% 89.7% 90.5% 90.1% Table 2: PolitiFact Dataset Results Model Embedding Accuracy Precision Recall F1-Score Logistic Regression BoW 78.9% 77.8% 79.4% 78.6% Logistic Regression TF-IDF 80.6% 79.9% 81.0% 80.4% Logistic Regression Word2Vec 83.3% 82.7% 84.0% 83.3% Logistic Regression BERT 87.4% 86.6% 88.1% 87.3% Random Forest BoW 80.1% 79.6% 80.5% 80.0% Random Forest TF-IDF 82.9% 82.4% 83.5% 82.9% Random Forest Word2Vec 85.2% 85.0% 85.5% 85.2% Random Forest BERT 89.0% 88.3% 89.5% 88.9% Neural Network BoW 79.8% 78.5% 81.0% 79.7% Neural Network TF-IDF 83.5% 82.8% 84.3% 83.5% Neural Network Word2Vec 86.9% 86.3% 87.4% 86.8% Neural Network BERT 91.3% 89.6% 88.5% 90.4% Table 3: LIAR Dataset Results 2.2 Analysis of Results Table 2 and Table 3: Results These results demonstrate the usability of embedding methods with machine learning models for fake news detection. Using Figure 5, it shows that BERT models have the highest accuracy among BoW, TF-IDF and Word2Vec. Our Findings We noted the following trends: 1. BERT:, All classifiers reached the highest performance when combined with BERT embeddings, neural networks and random forests performed better This provides evidence that for tasks that demand a rich understanding of the text, e.g., fake news detection, contextual embeddings have an advantage. 2. Solid Performance by Word2Vec: BERT performance was the best overall, with good competing results seen in Word2Vec embeddings (both CBoW & Skip-Gram), especially combined with logistic regression and random forest. This suggests the ability of Word2Vec to understand the relationship between classes better due to which classification performance is a lot better despite not capturing bidirectional context like in BERT. 3. Traditional methods like BoW and TF-IDF: behind BoW and TF-IDF lagged behind by comparison with Word2Vec and BERT. As BoW relies on each word being independent and the Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 170 https://internationalpubls.com perspective of TF-IDF also moves toward frequency, not context. Nevertheless, decent results can still be obtained with these methods, particularly in simpler models such as logistic regression. 4. It was True for every embedding model: In Interface Type 3, Neural Networks were the best in terms of classification accuracy, Precision, Recall and F1-scores. Neural Networks are capable of learning non-linear relationships between features, so they should be the model that we will use to detect even subtle patterns in fake news. 5. SVM and Logistic Regression Perform Well as Baselines: While neural networks tend to outperform them, both SVM and logistic regression make good baselines, especially with more sophisticated embeddings such as Word2Vec and BERT. Figure 5. Word Embedding Techniques Comparison on Model Accuracy 2.3 Effect on Embedding Dimension Aside from the main findings, we also examined how changing embedding dimensions affected model performance. Table 4 illustrates the performance of the Word2Vec model for different embedding dimensions. Model Embedding Dimensions Accuracy Precision Recall F1-Score Logistic Regression 100 82.5% 82.1% 83.0% 82.6% Logistic Regression 300 85.7% 86.0% 85.4% 85.7% Logistic Regression 768 88.1% 87.9% 88.4% 88.1% Random Forest 100 83.0% 82.5% 83.6% 83.1% Random Forest 300 87.2% 87.5% 87.0% 87.2% Random Forest 768 90.2% 89.9% 90.5% 90.2% Neural Network 100 85.9% 85.2% 86.4% 85.8% Neural Network 300 89.1% 88.6% 89.5% 89.1% Neural Network 768 91.7% 91.2% 92.3% 91.7% Table 4: Impact of Embedding Dimensions on Word2Vec Performance (PolitiFact Dataset) Increasing the dimensionality of Word2Vec embeddings improves the performance consistently across all models, as seen from Table 4. The embeddings in higher-dimensions are more detailed and thus result in better classification. This, however, comes at an additional cost of computational complexity and thus trade-offs need to be made between the complexity of a model and its performance. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 171 https://internationalpubls.com 3. Scale-Out and Distributed Computing Scalability is one of the most important parts to BE in this method. We then used Apache Spark to test the scalability of the system, distributing data processing and modeling across multiple nodes. A dataset of over 1 million news articles is used to test the system on a large scale setting. Processing performance and training time -Table 5 shows how long it takes to process the entire dataset in addition to train the models with and without distributed processing. Task Non-Distributed Time (mins) Distributed Time (mins) Speedup Data Preprocessing 45 10 4.5x Model Training (BERT) 120 25 4.8x Model Training (Word2Vec) 90 20 4.5x Model Training (Random Forest) 75 18 4.2x Table 5: Time Comparison for Distributed vs. Non-Distributed Processing Table 5: Using distributed processing considerably saves time in terms of both preprocessing and model training (in hours). Therefore, the proposed system is well-suited for real-time applications involving large datasets in order to process as quickly as possible. The experiments shows that having advanced word embedding models such as BERT and Word2Vec working in conjunction with strong classifiers like neural networks provide state-of-the-art results in detecting fake news BERT is the best model for fake news detection, as it can capture bidirectional context using deep learning models. By contrast, BoW and TF-IDF are simpler inhibitory methods that, though effective, just cannot compete with more comprehensive embedding procedures in terms of awareness of context. The distributed processing also allows for a very horizontal scaling of the system, which is important if we want to handle big datasets in real time such as social media, or news outlets online. In future we plan To optimize the system further and test other embeddings like RoBERTa and GPT for better performance in fake news detection. The fake news detection system would give significant accuracy, precision, recall, and F1-scores by using state-of-the-art word embeddings and machine learning models together. Results affirm the necessity of selecting proper methods to achieve a desirable result both at embedding layer and classifier level for each task and data size. Author: Guang Qiu, Contact Information The combined use of distributed processing frameworks ensures scalability as well as being applicable to large-scale data streams provided by real-world scenarios, making it an effective tool to tackle misinformation in the digital age. 5. CONCLUSION The world we live in today has plenty of examples of fake news taking root almost as though the floodgates were opened once social media and online platforms started fundamentally shifting our understanding of information integrity. Contrasting the elusive nature of fake news, therefore, recognizing and combating fake news is an important factor that can not only save us faith in media Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 172 https://internationalpubls.com but also spare us social, political, and economic effects of false information dissemination. By developing methodology for detecting fake news using modern machine learning models and Natural Language Processing (NLP) techniques, this paper proposes a wide-ranging, autonomic approach to improve scalability. By combining a set of embedding models—Bag of Words (BoW), Term Frequency-Inverse Document Frequency (TF-IDF), Word2Vec, and Bidirectional Encoder Representations from Transformers (BERT) and machine learning classifiers including logistic regression, random forests, and neural networks, the system has managed to achieve high throughputs in detecting fake news across large-scale datasets. So far: The overall aim of this work is to create a fake news detection algorithm that will not only provide accurate results while processing thousands and millions of news articles as well social media posts which are generated every day. We tried to balance computational efficiency and model complexity, so that the system is suitable for real-time fake news detection in practice such as monitoring social media platforms or aggregating news articles with fact-checking organizations. Experiments on PolitiFact and LIAR demonstrated the effectiveness and scalability of the system, which makes it suitable for deployment in large-scale environments. Key Findings The most important and informative result of this study is the demonstrated significantly improved benefit from employing powerful word embedding techniques, such as BERT, compared to traditional approaches like BoW and TF-IDF. Due to BERT’s nature, according to this paper, BERT has a great capacity of capturing the bidirectional context of input words throughout a sentence so that it can possibly yield both comprehensive and more intertwined word embeddings that could approximate its semantic meaning. This is the key feature for detecting fake news since many times bogus news stories use minor lingual changes or leave out important information to fool people. BERT, given that it looks at the contextual information around each word, is much better than models which treat words as independent units to pick up such manipulations. Both our BERT-based and non-BERT embedding models achieved better performance than the state- of-the-art baselines MODAL, TyDI Craft, XLM-RoBERTA across all three classifiers but always underperforming BERT. The performance of the best model, BERT with a neural network, obtained an accuracy over 92% reliable enough to be considered as a solution for fake news detection tasks. Word2Vec did alright, especially with random forests and logistic regression, but the numbers still show that it falls short of BERT in both accuracy and recall This indicates Word2Vec is good at capturing meaning relationships among words, but it captures less context compared with more complicated models like BERT. They also found that BoW and TF-IDF, while effective in simple classification tasks, perform poorly for fake news detection and scale up to the complex task of identifying fake news articles. Word frequency is the basis for both BoW and TF-IDF, which does not provide any context about how words relate to each other in a given text while it is so important to understand that real news should be different than fake news. While these methods have limitations, they nonetheless are useful baselines to show the benefits of more advanced embeddings such as Word2Vec and BERT. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 173 https://internationalpubls.com The findings of this study also suggest that a key ingredient to the success in using these learners is selection between classifiers for the given task. Neural networks gave the best accuracy result across all embeddings, but other models such as logistic regression and random forests were competitive, especially using Word2Vec or BERT embeddings. Deep learning models are often used as the classifier because of their capability in modeling non-linear relationships and learning complex patterns in the data, but it comes with a cost such as computation and training time. When you have limited computational resources, logistic regression or random forests along with Word2Vec/BERT embeddings may offer a better balance between performance and efficiency. Real World Use Cases and Scalability A key part of this study is scalability. As the news circulates rapidly and continuously across social media, fake news detection systems need to have the ability to process gargantuan quantities of data at all times to deliver real-time solutions. To tackle this, we have included distributed processing frameworks like Apache Spark in the architecture of the system. Our experiments showed that this distribution of the computational load among multiple nodes practically reduces pre-processing and model training time to a level by as much as five times faster for large datasets. This lends itself well to real-time applications that require a high degree of speed and efficiency. This allows the proposed system to scale dynamically, a requirement for organizations that would need to process millions of transactions daily. Social media platforms such as Twitter and Facebook could use this system to supplement their content moderation pipelines, so that misinformation would be automatically flagged or removed before spreading. Moreover, it will allow news aggregators and fact- checkers to use this platform for real-time fact-checking on news articles. The system is distributed so it can be scaled to handle millions of news articles or social media posts, making it a good choice for organizations who have large amounts of data that needs to be processed. Challenges and Limitations Acknowledgements Although the results obtained were promising, there are several challenges and limitations associated with this research. First, BERT and other deep embeddings models allow for good performance scores, but computational power is quite expensive. Especially BERT, it requires a lot of processing power and memory to train those models and therefore might not be suitable for resource-restricted environments. While using distributed processing frameworks solve this by distributing the computational load across multiple machines, smaller organizations or individuals with limited access to computing resources may find it challenging to enable BERT-based models at scale. Another hurdle is that the model might not be generalizable. Since the datasets used in this study (PolitiFact and LIAR) mainly deal with political news only, it is not guaranteed that other types of fake news, for instance health-related misinformation or financial news etc., would have been effectively detected using the proposed model. Though it does have some adaptability to new datasets, we need results on additional fake news domains. However, future work should aim at evaluating our approach on more varieties of fake news to test the generality of the model, as well as integrating extra datasets to train it on a variety of subjects. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 174 https://internationalpubls.com The system is also limited by the dependency on labeled datasets. Training machine learning models on fake news is difficult because the training data requires a large number of labeled inputs, and labeling fake news is in general done manually so it's time-consuming and expensive. While crowd- sourcing platforms and automated labeling tools can lighten this load, acquiring properly annotated data in the domain of fake news detection remains one of the most significant challenges. In addition, the system needs to be trained repeatedly with new data in order to strengthen its effectiveness as fake new types evolve over time. Another vital consideration is the trade-off between computational efficiency and complexity of the model. Although the performance of neural networks (a type of deep learning model) was generally best in our experiments, they required many computational resources and much training time. On the other hand, simpler models like logistic regression and random forests although not as accurate are faster to train with less resource required. In practice, especially if operating in high-speed environments (like monitoring social media for real-time activity), which model to use might be dependent on the constraints of the system like how much processing power is available and how much data requires analysis. Future Directions The research opens up a wide range of possible directions for future work. A possible direction could be to delve into even higher-level contextual embeddings (RoBERTa, etc.). RoBERTa (Robustly Optimized BERT Pre Training Approach) improves over BERT as it trains on longer sequences and with larger mini-batches, while recent versions of GPT (Generative Pretrained Transformer) models can generate text in autoregressive model which may be useful to recognize certain types of linguistic manipulation that occur in fake news articles. Investigating these approaches could lead to additional performance benefits on top of the current state-of-the-art and push forward fake news detection research. A further area for future research could involve designing domain-specific fake news detection models. Since the fake news is diverse in the domains, i.e. political, health, financial and entertainment news etc., there may be many separable representations that might need to be learnt for different domains, which we are not capturing using a broader model where all of the data points with similar label type have similar representations. For example, domain-specific models could learn specialized linguistic and contextual patterns for different types of misinformation to better guess the content type and improve performance. While health-related fake news often uses medical jargon, and appeals to the authority with fake claims from supposed doctors; the political sphere has a different type that reaches at emotions by sensationalism. Finally, we also could integrate to our system other source of knowledge as future research work in the Fake news detection process. While this research was confined only to text-based classification, we can boost the potential of this model by incorporating knowledge graphs and external repositories containing facts which have gone through some verification, in detecting the fake news. For example, the model might cross reference news articles against all of the scientific research papers or fact-check websites it linked to in our previous exercise to verify the claims made in the news article. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 175 https://internationalpubls.com This also brings up the ethical concerns with fake news detection systems. When the automatic detection system is widely used, we must guarantee its transparency, fairness and impartiality. Further research needs to be conducted in developing Explainable AI (XAI) techniques for fake news detection cascaded with reasons why articles are flagged as fake. There are also attempts to make sure that the system does not end up censoring legitimate content or bias the news ecosystem in some way. To sum up, this paper features a scalable and efficient method to detect fake news using combined word embedding models, machine learning classifiers, and distributed processing. This is a very powerful result and shows that models such as BERT along with neural networks have the ability to attain state-of-the-art accuracy for fake news detection; providing significant information in the battle against misinformation. Has good scalability for dealing with massive datasets in production thanks to leveraging distributed processing frameworks, thus supports use-cases like social media monitoring, news aggregations and fact-checker variations. Although there are some challenges and limitations, such as the computational burden of advanced models and the requirement for large amounts of labeled data, the proposed system presents a promising framework to address the issue of fake news at scale. The modularity of the system opens a range of opportunities for future enhancement, including the use of newer models for embedding, training the system with domain-specific data, and integrating external sources of knowledge. Eventually, the findings of this paper are a valuable addition to the field of fighting disinformation that offers a reliable, scalable solution to the challenge of fake news detection in the modern digital era. By combining these strengths to build on the state of the art of fake news detection systems, this research serves as a platform for future research efforts in combating new manifestations of misinformation as they continue to evolve. Our next steps on this journey will involve fine-tuning our models and extending their application to new domains, ensuring that the systems we create are transparent, ethical, and powerful enough to meet the demands of an increasingly sophisticated information ecology. REFERENCES: [1] David MJ Lazer, Matthew A Baum, Yochai Benkler, Adam J Berinsky, Kelly M Greenhill, Filippo Menczer, Miriam J Metzger, Brendan Nyhan, Gordon Pennycook, David Rothschild, et al. The science of fake news. Science, 359(6380):1094–1096, 2018. [2] Stephan Lewandowsky, Ullrich KH Ecker, and John Cook. Beyond misinformation: Understanding and coping with the “post-truth” era. Journal of applied research in memory and cognition, 6(4):353–369, 2017. [3] Mallareddy, A., Sridevi, R., & Prasad, C. G. V. N. (2019). Enhanced P-gene based data hiding for data security in cloud. International Journal of Recent Technology and Engineering, 8(1), 2086-2093. [4] Prasad, C. G. V. N., Mallareddy, A., Pounambal, M., & Velayutham, V. (2022). Edge Computing and Blockchain in Smart Agriculture Systems. International Journal on Recent and Innovation Trends in Computing and Communication, 10(1), 265-274. [5] Diaa Salama Abdelminaam, Fatma Helmy Ismail, Mohamed Taha, Ahmed Taha, Essam H Houssein, and Ayman Nabil. Coaid-deep: An optimized intelligent framework for automated detecting covid-19 misleading information on twitter. Ieee Access, 9:27840–27867, 2021. [6] Mohammad Hadi Goldani, Saeedeh Momtazi, and Reza Safabakhsh. Detecting fake news with capsule neural networks. Applied Soft Computing, 101:106991, 2021. [7] Ciprian-Octavian Truică and Elena-Simona Apostol. It’s all in the embedding! fake news detection using document embeddings. Mathematics, 11(3):508, 2023. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 176 https://internationalpubls.com [8] Liwen Peng, Songlei Jian, Zhigang Kan, Linbo Qiao, and Dongsheng Li. Not all fake news is semantically similar: Contextual semantic representation learning for multimodal fake news detection. Information Processing & Management, 61(1):103564, 2024. [9] Mayank Kumar Jain, Dinesh Gopalani, and Yogesh Kumar Meena. Confake: fake news identification using content based features. Multimedia Tools and Applications, 83(3):8729–8755, 2024. [10] Jianqiao Lai, Xinran Yang, Wenyue Luo, Linjiang Zhou, Langchen Li, Yongqi Wang, and Xiaochuan Shi. Rumorllm: A rumor large language model-based fake-news-detection data-augmentation approach. Applied Sciences, 14(8):3532, 2024. [11] Kai Nakamura, Sharon Levy, and William Yang Wang. Fakeddit: A new multimodal benchmark dataset for fine- grained fake news detection. In Nicoletta Calzolari, Frédéric Béchet, Philippe Blache, Khalid Choukri, Christopher Cieri, Thierry Declerck, Sara Goggi, Hitoshi Isahara, Bente Maegaard, Joseph Mariani, Hélène Mazo, Asuncion Moreno, Jan Odijk, and Stelios Piperidis, editors, Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 6149–6157, Marseille, France, May 2020. European Language Resources Association. [12] Guobiao Zhang, Anastasia Giachanou, and Paolo Rosso. Scenefnd: Multimodal fake news detection by modelling scene context information. Journal of Information Science, 50(2):355–367, 2024. [13] Rambabu, B., Reddy, A. V., & Janakiraman, S. (2022). Hybrid artificial bee colony and monarchy butterfly optimization algorithm (HABC-MBOA)-based cluster head selection for WSNs. Journal of King Saud University- Computer and Information Sciences, 34(5), 1895-1905. [14] Janakiraman, S., & Rambabu, B. (2022, January). Improved Symbiosis Organism Search Algorithm-Based Clustering Scheme for Enhancing Longevity in Wireless Sensor Networks (WSNs). In Proceedings of International Conference on Recent Trends in Computing: ICRTC 2021 (pp. 799-808). Singapore: Springer Nature Singapore. [15] Rasikh Ali, Tayyaba Farhat, Sanya Abdullah, Sheeraz Akram, Mousa Alhajlah, Awais Mahmood, and Muhammad Amjad Iqbal. Deep learning for sarcasm identification in news headlines. Applied Sciences, 13(9):5586, 2023. [16] Christina Boididou, Symeon Papadopoulos, Markos Zampoglou, Lazaros Apostolidis, Olga Pa- padopoulou, and Yiannis Kompatsiaris. Detection and visualization of misleading content on twitter. International Journal of Multimedia Information Retrieval, 7(1):71–86, 2018. [17] Olga Papadopoulou, Markos Zampoglou, Symeon Papadopoulos, and Ioannis Kompatsiaris. A corpus of debunked and verified user-generated videos. Online information review, 43(1):72–88, 2019. [18] Kai Shu, Deepak Mahudeswaran, Suhang Wang, Dongwon Lee, and Huan Liu. Fakenewsnet: A data repository with news content, social context and dynamic information for studying fake news on social media. arXiv preprint arXiv:1809.01286, 2018. [19] Dmytro Valiaiev. Detection of machine-generated text: Literature survey. arXiv preprint arXiv:2402.01642, 2024. [20] Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Chenxi Whitehouse, Osama Mohammed Afzal, Tarek Mahmoud, Toru Sasaki, Thomas Arnold, Alham Aji, Nizar Habash, Iryna Gurevych, and Preslav Nakov. M4: Multi-generator, multi-domain, and multi- lingual black-box machine-generated text detection. In Yvette Graham and Matthew Purver, editors, Proceedings of the 18th Conference of the European Chapter of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 1369–1407, St. Julian’s, Malta, March 2024. Association for Computational Linguistics. [21] Medeswara Rao Kondamudi, Somya Ranjan Sahoo, Lokesh Chouhan, and Nandakishor Yadav. A comprehensive survey of fake news in social networks: Attributes, features, and detection approaches. Journal of King Saud University-Computer and Information Sciences, 35(6):101571, 2023. [22] Sonal Garg and Dilip Kumar Sharma. Linguistic features based framework for automatic fake news detection. Computers & Industrial Engineering, 172:108432, 2022. [23] Anshika Choudhary and Anuja Arora. Linguistic feature based learning model for fake news detection and classification. Expert Systems with Applications, 169:114171, 2021. [24] Abhijnan Chakraborty, Bhargavi Paranjape, Sourya Kakarla, and Niloy Ganguly. Stop clickbait: De- tecting and preventing clickbaits in online news media. In 2016 IEEE/ACM international conference on advances in social networks analysis and mining (ASONAM), pages 9–16. IEEE, 2016. [25] Arkaitz Zubiaga, Maria Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Peter Tolmie. Analysing how people orient to and spread rumours in social media by looking at conversational threads. PloS one, 11(3):e0150989, 2016. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 177 https://internationalpubls.com [26] Zhiwei Jin, Juan Cao, Han Guo, Yongdong Zhang, and Jiebo Luo. Multimodal fusion with recurrent neural networks for rumor detection on microblogs. In Proceedings of the 25th ACM international conference on Multimedia, pages 795–816, 2017. [27] Bande, V., Raju, B. D., Rao, K. P., Joshi, S., Bajaj, S. H., & Sarala, V. (2024). Designing Confidential Cloud Computing for Multi-Dimensional Threats and Safeguarding Data Security in a Robust Framework. Int. J. Intell. Syst. Appl. Eng, 12(11s), 246-255 [28] Eugenio Tacchini, Gabriele Ballarin, Marco L Della Vedova, Stefano Moret, and Luca De Alfaro. Some like it hoax: Automated fake news detection in social networks. arXiv preprint arXiv:1704.07506, 2017. [29] Benjamin Horne and Sibel Adali. This just in: Fake news packs a lot in title, uses simpler, repetitive content in text body, more similar to satire than real news. In Proceedings of the international AAAI conference on web and social media, volume 11, pages 759–766, 2017. [30] Giovanni Santia and Jake Williams. Buzzface: A news veracity dataset with facebook user commentary and egos. In Proceedings of the international AAAI conference on web and social media, volume 12, pages 531–540, 2018. [31] A. Trivedi, E. K. Kaur, C. Choudhary, Kunal and P. Barnwal, "Should AI Technologies Replace the Human Jobs?," 2023 2nd International Conference for Innovation in Technology (INOCON), Bangalore, India, 2023, pp. 1- 6, doi: 10.1109/INOCON57975.2023.10101202. [32] Jeppe Nørregaard, Benjamin D Horne, and Sibel Adalı. Nela-gt-2018: A large multi-labelled news dataset for the study of misinformation in news articles. In Proceedings of the international AAAI conference on web and social media, volume 13, pages 630–638, 2019. [33] Kunal, P. Singh and N. Hirani, "A Cohesive Relation Between Cybersecurity and Information security," 2022 IEEE 3rd Global Conference for Advancement in Technology (GCAT), Bangalore, India, 2022, pp. 1-6, doi: 10.1109/GCAT55367.2022.9972023. [34] Ramy Baly, Georgi Karadzhov, Dimitar Alexandrov, James Glass, and Preslav Nakov. Predicting factuality of reporting and bias of news media sources. arXiv preprint arXiv:1810.01765, 2018. [35] Arkadipta De, Dibyanayan Bandyopadhyay, Baban Gain, and Asif Ekbal. A transformer-based approach to multilingual fake news detection in low-resource languages. ACM Trans. Asian Low-Resour. Lang. Inf. Process., 21(1), nov 2021. [36] Georgios Gravanis, Athena Vakali, Konstantinos Diamantaras, and Panagiotis Karadais. Behind the cues: A benchmarking study for fake news detection. Expert Systems with Applications, 128:201–213, 2019. [37] Faraz Ahmad and R Lokeshkumar. A comparison of machine learning algorithms in fake news detection. International Journal on Emerging Technologies, 10(4):177–183, 2019. [38] Pedro Henrique Arruda Faustini and Thiago Ferreira Covoes. Fake news detection in multiple platforms and languages. Expert Systems with Applications, 158:113503, 2020. [39] Apoorva Dhawan, Malvika Bhalla, Deeksha Arora, Rishabh Kaushal, and Ponnurangam Kumaraguru. Fakenewsindia: A benchmark dataset of fake news incidents in india, collection methodology and impact assessment in social media. Computer Communications, 185:130–141, 2022. [40] Kai Shu, Xinyi Zhou, Suhang Wang, Reza Zafarani, and Huan Liu. The role of user profiles for fake news detection. In Proceedings of the 2019 IEEE/ACM international conference on advances in social networks analysis and mining, pages 436–439, 2019. [41] Yaqing Wang, Weifeng Yang, Fenglong Ma, Jin Xu, Bin Zhong, Qiang Deng, and Jing Gao. Weak supervision for fake news detection via reinforcement learning. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 516–523, 2020. [42] Taichi Murayama. Dataset of fake news detection and fact verification: a survey. arXiv preprint arXiv:2111.03299, 2021. [43] Arianna D’Ulizia, Maria Chiara Caschera, Fernando Ferri, and Patrizia Grifoni. Fake news detection: a survey of evaluation datasets. PeerJ Computer Science, 7:e518, 2021. [44] Tahniat Khan, Mizanur Rahman, Veronica Chatrath, Oluwanifemi Bamgbose, and Shaina Raza. Fake- watch electionshield: A benchmarking framework to detect fake news for credible us elections. arXiv preprint arXiv:2312.03730, 2023. [45] Sara Abdali. Multi-modal misinformation detection: Approaches, challenges and opportunities. arXiv preprint arXiv:2203.13883, 2022. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 5s (2025) 178 https://internationalpubls.com [46] Yixuan Chen, Dongsheng Li, Peng Zhang, Jie Sui, Qin Lv, Lu Tun, and Li Shang. Cross-modal ambiguity learning for multimodal fake news detection. In Proceedings of the ACM web conference 2022, pages 2897–2905, 2022. [47] Michał Choraś, Konstantinos Demestichas, Agata Giełczyk, Álvaro Herrero, Paweł Ksieniewicz, Kon- stantina Remoundou, Daniel Urda, and Michał Woźniak. Advanced machine learning techniques for fake news (online disinformation) detection: A systematic mapping study. Applied Soft Computing, 101:107050, 2021. [48] Longzheng Wang, Chuang Zhang, Hongbo Xu, Yongxiu Xu, Xiaohan Xu, and Siqi Wang. Cross-modal contrastive learning for multimodal fake news detection. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5696–5704, 2023. [49] Isabel Segura-Bedmar and Santiago Alonso-Bartolome. Multimodal fake news detection. Information, 13(6):284, 2022. [50] Zhen Wang, Xu Shan, Xiangxie Zhang, and Jie Yang. N24news: a new dataset for multimodal news classification. arXiv preprint arXiv:2108.13327, 2021. [51] Yufeng Zhou, Aiping Pang, and Guang Yu. Clip-gcn: an adaptive detection model for multimodal emergent fake news domains. Complex & Intelligent Systems, pages 1–18, 2024. [52] Asma Sormeily, Sajjad Dadkhah, Xichen Zhang, and Ali A Ghorbani. Mefand: A multimodel framework for early fake news detection. IEEE Transactions on Computational Social Systems, 2024. [53] Sakshini Hangloo and Bhavna Arora. Combating multimodal fake news on social media: methods, datasets, and future perspective. Multimedia systems, 28(6):2391–2422, 2022. [54] Khaled Bayoudh, Raja Knani, Fayçal Hamdaoui, and Abdellatif Mtibaa. A survey on deep multimodal learning for computer vision: advances, trends, applications, and datasets. The Visual Computer, 38(8):2939–2970, 2022. [55] Bogoan Kim, Aiping Xiong, Dongwon Lee, and Kyungsik Han. A systematic review on fake news research through the lens of news creation and consumption: Research efforts, challenges, and future directions. PloS one, 16(12):e0260080, 2021. [56] Kitti Nagy and Jozef Kapusta. Improving fake news classification using dependency grammar. Plos one, 16(9):e0256940, 2021. [57] Yue Huang and Lichao Sun. Harnessing the power of chatgpt in fake news: An in-depth exploration in generation, detection and explanation. arXiv preprint arXiv:2310.05046, 2023. [58] Alok Mishra and Halima Sadia. A comprehensive analysis of fake news detection models: A systematic literature review and current challenges. Engineering Proceedings, 59(1):28, 2023. [59] Elena Shushkevich, Mikhail Alexandrov, and John Cardiff. Improving multiclass classification of fake news using bert-based models and chatgpt-augmented data. Inventions, 8(5):112, 2023.