PaperlayerView exact revision on arXiv

Source-linked AI summary

A Large Language Model LLM Approach to COVID-19 Fake News Detection

Muhammad Arif, Samah Fodeh-Jarad, Elyas Irankhah, Uroosa Fatima, Lama Ayash, Sreeraj Ramachandran

arXiv:2610.23800v1cs.CY

TL;DR

Fake news spreads through social media, while detecting the authenticity of news remains challenging and existing approaches have limitations. The study applies transfer-learning and large language models to binary classification of COVID-19 English tweets and explores optimization techniques for model performance. GPT-2 achieved an accuracy of 0.9631, a precision of 0.9612, a recall of 0.9610, and an F1-score of 0.9600.

  • Problem

    Fake news spreads through social media, while detecting the authenticity of news remains challenging and existing approaches have limitations.

    Source: §I · p. 2

  • Method

    The study applies transfer-learning and large language models to binary classification of COVID-19 English tweets and explores optimization techniques for model performance.

    Source: §VI · p. 7

  • Results

    GPT-2 achieved an accuracy of 0.9631, a precision of 0.9612, a recall of 0.9610, and an F1-score of 0.9600.

    Source: §VI · p. 7

  • Takeaways & Limitations

    GPT-2 was the best-performing model and demonstrated effectiveness in classifying real and fake news.

    Source: §VI · p. 7

    The study focuses on binary classification of English tweets; future work plans larger datasets, multiclass classification, other languages, and real-time tweets.

    Source: §VI · p. 7

Abstract

from arXiv · show

Text classification is a widely studied problem in information retrieval and data mining, with applications ranging from healthcare diagnosis to the detection of COVID-19-related fake news. Social media platforms generate vast volumes of user-generated text, and the rapid spread of misinformation during the pandemic has made automated fake news detection an important task. In this study, we apply pretrained transformer-based transfer learning (TL) models and large language models (LLMs), namely BERT-base, ALBERT-base, XLM-RoBERTa-base, DistilBERT, GPT-2, XLNet, and T5, to the COVID-19 English Tweets dataset, which contains 6,420 tweets labeled as real or fake. To further improve performance, we apply six late-fusion optimization techniques: weighted fusion, particle swarm optimization (PSO) fusion, the Nelder-Mead algorithm, Powell's method, the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm, and trust-region constrained optimization. Models are evaluated using accuracy, precision, recall, and F1-score, and confusion-matrix heatmaps are used to visualize the predictions of each model. GPT-2 achieved the best performance, with an accuracy of 0.9631, precision of 0.9612, recall of 0.9610, and F1-score of 0.9600, while Nelder-Mead fusion achieved the best result among the fusion methods with an accuracy of 0.9610. These results highlight the value of LLMs for COVID-19 fake news detection. We conclude by outlining future directions, including multiclass classification, multilingual datasets, and real-time social media streams.

I. INTRODUCTION

The introduction frames fake-news detection as a challenge intensified by social media and motivates transformer-based transfer learning for text classification. It also presents optimization techniques as an approach to improving model performance.

I. INTRODUCTION

  • Transfer learning uses information from related domains to improve performance on tasks with few labeled samples, a setting where transformers are described as outperforming traditional ML and DL techniques.The passage says prior information can support high accuracy with few labeled data points.Source: §I · p. 2
  • Transformers have been used to distinguish real from fabricated information, including COVID-19 news, and are described as effective for this task.Source: §I · p. 2
  • Existing fake-news detection approaches have limitations, motivating researchers to develop hybrid methods to improve accuracy and consistency.The introduction also notes that social media has exacerbated the spread of fake news.Source: §I · p. 2
  • The study applies optimization techniques to refine model parameters, reduce losses, and improve overall accuracy and prediction robustness.Source: §I · p. 2

II. RELEVANT LITERATURE

The literature review covers transfer learning and transformer methods across text-classification tasks, alongside studies of COVID-19 misinformation and other social-media content. It highlights varied model strategies, datasets, and evaluation settings.

II. RELEVANT LITERATURE

  • Adapters added 3.6% trainable parameters per task while attaining near-state-of-the-art performance across 26 BERT classification tasks [14].Source: §II · p. 2
  • COVID-19 credibility studies reported F1-scores of 98.69 for CT-BERT and 98.55 for a transfer-learning approach that ranked 5th among 160 teams [16].Source: §II · p. 2
  • An unsupervised fake-news method was developed because reliable labeled datasets were difficult to obtain; it achieved accuracy of 0.759 on LIAR and 0.679 on BuzzFeed data.Source: §II · pp. 2–3
  • Further pre-training BERT and DistilBERT on COVID-19 tweets outperformed fine-tuning only on the dataset, and the resulting BERT model performed better than COVID-Twitter-BERT [18].That study evaluated 10,700 social-media posts and articles, using news content without user-level or propagation signals.Source: §II · p. 3
  • On an official test set, RoBERTa and ELECTRA achieved F1-scores of 0.9864 and 0.9827, respectively, in a COVID-19 misinformation study [19].Source: §II · p. 3
  • Other reviewed work includes transfer-learning studies in hate-speech detection and BERT4TC, which incorporated domain-specific and task-oriented knowledge and outperformed conventional feature-based and several fine-tuned approaches.A finance sentiment study reported a BART F1-score of 0.85, while an expert-annotated CNN–BERT model was strongest among four models for political text classification.Source: §II · p. 3

III. MATERIAL AND METHODS

The methodology classifies COVID-19 news as real or fake through a workflow spanning dataset collection, preprocessing, data splitting, and model evaluation.

III. MATERIAL AND METHODS

  • The workflow begins with dataset collection and data preprocessing.Source: §III · p. 3
  • Data splitting is a named stage in the classification methodology.Source: §III · p. 3
  • The workflow includes model evaluation for the real-versus-fake COVID-19 news classification task.Source: §III · p. 3

A. Dataset

The study uses the COVID-19 English Tweets dataset, containing 6,420 tweets labeled as real or fake. Its two columns hold tweet text and authenticity labels, which are encoded numerically for processing.

A. Dataset

  • The dataset contains 6,420 tweets: 3,360 labeled real and 3,060 labeled fake.Source: A. Dataset · p. 3
  • The public dataset has a “Tweets” column containing text and a “Labels” column distinguishing authentic from fake tweets.Source: A. Dataset · p. 3
  • The dataset documentation identifies pandemic updates, statistics, precautions, and opinions among the tweet content, and the paper visualizes frequent words and a word cloud from raw data.Source: A. Dataset · p. 3
  • Real labels are encoded as 1 and fake labels as 0 using a Python script.Source: A. Dataset · p. 3

B. Data Preprocessing

The study cleans the COVID-19 tweet dataset, splits it for training, validation, and testing, and evaluates model predictions with standard classification metrics and confusion matrices.

B. Data Preprocessing

  • URLs, special tags, and numbers are removed, multiple spaces are collapsed, and characters are lowercased; no class balancing is applied.The authors state the two classes contain enough samples for the experiments.Source: B. Data Preprocessing · p. 3

C. Training, Validation, and Test Splits

D. Model Evaluation

  • Precision, recall, accuracy, and F1-score evaluate performance, while confusion-matrix heatmaps compare predictions with actual values.The matrix represents true positives, true negatives, false positives, and false negatives.Source: D. Model Evaluation · p. 4

E. Classification Algorithms

The study evaluates four transfer-learning models and three LLMs, then applies optimization techniques to seek improved results.

E. Classification Algorithms

  • The seven evaluated models are BERT-base, XLM-RoBERTa-base, ALBERT-base, DistilBERT, GPT-2, T5, and XLNet.Source: E. Classification Algorithms · pp. 4–5
  • The selected transfer-learning models and LLMs are fine-tuned, mostly with default settings, with some changes to epochs, learning rate, and batch size.GPT-2’s learning rate was set to 2 × 10−5 rather than its stated default of 5 × 10−5.Source: E. Classification Algorithms · p. 5
  • Optimization techniques are applied after initial model evaluation to compare performance before and after optimization.Source: E. Classification Algorithms · p. 5

1) BERT-Base:

This section describes BERT and related transformer models, including their architecture or training characteristics and reported uses in language tasks.

1) BERT-Base:

  • BERT uses masked-language modeling and next-sentence prediction, learning bidirectional context from unlabeled text; task adaptation adds a task-specific output layer.BERT-Base is described as requiring comparatively less processing and computation time.Source: 1) BERT-Base · p. 5

2) ALBERT-Base:

  • ALBERT reduces parameters relative to BERT, accelerating training and reducing memory usage while performing well on classification and sentiment-analysis tasks.Source: 2) ALBERT-Base · p. 5

3) DistilBERT:

  • DistilBERT is described as 60% faster and 40% smaller than BERT while retaining 97% of its language-understanding capability.Source: 3) DistilBERT · p. 5

4) XLM-RoBERTa-Base:

  • XLM-RoBERTa is trained across 100 languages using more than two terabytes of filtered data and has been applied to multilingual and other language tasks.Source: 4) XLM-RoBERTa-Base · pp. 5–6

6) XLNet:

The section outlines XLNet and T5 characteristics and describes late fusion, including weighted averaging and optimization-based weight selection.

6) XLNet:

  • XLNet learns bidirectional contexts by maximizing expected likelihood over permutations of factorization order, addressing BERT’s masked-position dependency limitation.Source: 6) XLNet · p. 6
  • T5 treats every task as text-to-text and uses an encoder–decoder transformer pretrained with an unsupervised span-corruption objective.Methods for extracting high-quality text representations from T5 are still being studied.Source: 6) XLNet · p. 6

F. Optimization Techniques and Fusion

1) Weighted Fusion:

  • Weighted fusion averages model predictions with weights that favor better-performing models, giving greater influence to models with higher precision or reliability.Source: 1) Weighted Fusion · p. 6

2) PSO Fusion:

This section describes optimization methods for refining fusion solutions, including PSO, Nelder–Mead, Powell’s method, and BFGS. Their approaches differ in how they search for minima and the conditions they suit.

2) PSO Fusion:

  • PSO iteratively improves candidate solutions using a fitness function, updating particle positions and velocities from personal-best and global-best positions.It begins with a diverse population of candidate solutions.Source: 2) PSO Fusion · p. 6

3) Nelder–Mead Algorithm:

  • Nelder–Mead is a derivative-free method for minimizing an objective, including validation classification error to obtain fusion weights.It is suited to n-dimensional problems.Source: 3) Nelder–Mead Algorithm · p. 6

4) Powell’s Method:

  • Powell’s method uses successive line searches along conjugate directions to approach a minimum, but its effectiveness can depend on step-size assumptions.It iteratively refines an initial guess.Source: 4) Powell’s Method · p. 6

5) BFGS:

  • BFGS is a quasi-Newton method for minimizing multidimensional functions, suited to smooth, well-behaved problems when gradients are available.Source: 5) BFGS · p. 6

6) Trust-Region Constrained (trust-constr) Algorithm:

The paper reports model results and comparisons for COVID-19 tweet classification, and describes trust-region constrained optimization. It notes that similar model results may reflect dataset or training similarities.

6) Trust-Region Constrained (trust-constr) Algorithm:

  • Trust-region constrained optimization minimizes or maximizes an objective while respecting equality and inequality constraints, using trust regions for smooth constrained problems.Source: 6) Trust-Region Constrained… · p. 6

IV. EXPERIMENTS AND RESULTS

  • Experiments report accuracy, precision, recall, and F1-score with confusion-matrix heatmaps; Tables II–V cover transfer-learning, optimization, LLM, and state-of-the-art comparisons.Source: §IV · p. 6

V. DISCUSSION

  • GPT-2 led with 96.31% accuracy, 96.12% precision, 96.10% recall, and 96.00% F1-score; transformers and LLMs outperformed the compared state-of-the-art ML approaches.Source: §V · pp. 6–7, Table IV · p. 7
  • Similar model performance may be attributed to the small, balanced dataset or to similar hyperparameters or initialization methods causing convergence toward similar solutions.Source: §V · p. 7

VI. CONCLUSION

The study evaluates transfer-learning and large language models, along with several fusion optimizers, for binary real-or-fake classification on COVID-19 English tweets. GPT-2 performs best, while future work targets broader datasets, languages, and streams.

  • The study applied four transfer-learning models and three LLMs to binary classification on COVID-19 English tweets, exploring six fusion optimization techniques.The models were evaluated using accuracy, recall, precision, and F1-score.Source: §VI · p. 7
  • GPT-2 was best, with accuracy of 0.9631, precision of 0.9612, recall of 0.9610, and F1-score of 0.9600.Source: §VI · p. 7
  • Future work will test larger datasets for multiclass classification, datasets in different languages, real-time tweets, and other social media platforms.Source: §VI · p. 7

Related papers

Loading 2610.23800v1…