2.1 Classical Machine Learning Approach for Detection of Fake News about Covid-19

The most commonly used algorithms for the fake news detection, and fake news related with Covid-19 in particular, are classical machine learning algorithms. They are really helpful in case of binary classification, when we need to identify if a message is fake news or real news. In a range of classical machine learning algorithms it is reasonable to highlight Logistic Regression with the basic idea of linear classifier, which is able to divide a feature’s space into two spaces by a hyperplane, and each half-space is reflecting each class of a binary classification. Support Vector Machines are very popular in the framework of classification problems, and this classifier is creating a hyperplane or set of hyperplanes in a multidimensional or infinite-dimensional space. For fake new detection, methods, which are based on classification trees, are also used. The main representatives of such methods are Gradient Boosting Classifier and Random Forest Classifier. The Random Forest Classifier in contrast to Gradient Boosting Classifier builds each tree independently and combines results at the end of the process.

In the context of using classical machine learning algorithms to identify fake news about Covid-19 the authors [⁵] analyzed fake news connected with COVID pandemic. They collected a dataset from 150 users, extracting data from their social media accounts including Twitter, email, mobile, Whatsapp and Facebook for 4 months from March 2020 to June 2020. At the preprocessing stage they removed information not related to the Covid-19 data and deleted incomplete news. K-Nearest Neighbour was used to perform the classification. As a result, they achieved the best prediction results for June with 0.91 F1-score and the worst ones for March with 0.79 F1-score.

In [⁶,⁷] the researchers aimed to detect fake news connected with Covid-19 on small datasets. The researchers compared Logistic Regression, Support Vector Machine, Gradient Boosting and Random Forest on a limited dataset of 1,000 fake and real messages. Support Vector Machine and Random Forest classifiers performed the best results with 69% micro-F1 score. The authors noted that despite the fact that the results are not as strong as the results obtained on the full dataset of fake news about Covid-19, such an approach could be helpful for the researchers who do not have a big enough dataset or sufficient time to collect a big dataset, but where the classification decisions have to be taken in the shortest span of time possible.

In [⁸] the authors compare four machine learning baselines (Decision Tree, Logistic Regression, Gradient Boost and Support Vector Machine (SVM)) on the dataset devoted to fake news about Covid-19 detection. The authors collected themselves from Facebook, Twitter and other social media platforms. The best result - 0.93 F1-score - was achieved with the SVM model.

From the above, we can conclude that there are a lot of algorithms of machine learning, which showed quite good results on the binary classification for fake and real news.

2.2 Neural Networks for Covid-19 Fake News Detection

There is a large variety of algorithms based on neural networks which have been successfully applied to fake news detection, including that related to Covid-19. Researchers use a lot of interesting linguistic models, the most important of which is BERT [⁹], which stands for Bidirectional Encoder Representations from Transformers, and which can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks.

DistilBERT [¹⁰] model allows to reduce the size of a BERT model by 40%, while retaining 97% of its language understanding capabilities and being 60% faster.

COVID-Twitter-BERT (CT-BERT) model [¹¹] was pretrained on a large corpus of Twitter messages on the topic of COVID-19 and has proved very useful in case of detecting such messages.

In addition, an important model which is often used is RoBERTa [¹²], which is a robust BERT trained on a larger dataset for a larger number of iterations and with a larger batch size.

Some interesting and useful models are ELECTRA [¹³] – Pre-training Text Encoders as discriminators rather than generators, and AlBert [¹⁴] – a Lite BERT for Self-supervised Learning of language representations.

There is the XLNet model [¹⁵], which is similar to BERT but learns bidirectional context along with autoregressive formulation.

Hierarchical Attention Networks (HAN) is based on LSTM and comprises four sequential levels – word encoder, word-level attention, sentence encoder and sentence-level attention [¹⁶].

ELMO [¹⁷] is a deep contextualized word representation that models both complex characteristics of word use (e.g. syntax and semantics), and how these uses vary across linguistic contexts (i.e. to model polysemy).

A successful use of neural networks for Covid-19 fake news detection is shown in the paper [¹⁸]. Its authors collected 4.8K expert-annotated social media posts related to COVID-19 and 86 common misconceptions about the disease to evaluate the performance of misinformation detection related to the Coronavirus pandemic. Each message was labeled in a context of chosen misconceptions as "misinformative" (tweet is a positive expression of the misconception), "informative" (tweet disagrees with the misconception), or "irrelevant" (tweet is not relevant to the misconception).

For the modelling, the authors used TF-IDF and GloVe parameterizations to obtain vectorized representations. Researchers used the RoBERTa-base implementation with two models of textual similarity (cosine similarity computed by averaging over the token vectors, and BERTScore which involves adding cosine similarities between RoBERTa token embeddings) to obtain contextual word embeddings. In addition, the authors tuned the RoBERTa-base model using the dataset of tweets connected with COVID-19.

As a result, the domain-adapted BERTScore achieved the best results among the similarity models.

In [¹⁹] the authors created an ensemble of linguistic models XLNet, RoBERTa, XLM-RoBERTa, DeBERTa, ERNIE 2.0, and ELECTRA for the task of Covid-19 fake news detection. They also used a tweet-preprocessor library from Python [²⁰] to filter out noisy data such as usernames, URLs, emojis, etc.

Furthermore, the authors implemented Heuristic Post-Processing, which takes Soft-voting prediction vectors into account. Thanks to the preprocessing step, the construction of such an ensemble, and by choosing the soft-voting through prediction vectors instead of the hard-voting approach, the researchers achieved an F1-score of 0.9831.

To deal with a problem of Covid-19 fake news detection, the authors of [²¹] implemented a combination of topical distributions from Latent Dirichlet Allocation (LDA) with contextualized representations from XLNet.

For the implementation, the authors used the Transformers library^fn maintained by the researchers and engineers at Hugging Face [²²], which provides the PyTorch interface for XLNet. The created model allowed researchers to achieve 0.967 F1-score on the test dataset. In addition, to compare the results the authors implemented SVM model with contextualized representations of the input using Universal Sentence Encoder (USE)^fn, BERT with document-topic distributions from LDA, fine-tuned pretrained XLNet model, and the combination of BERT and BERT+topic models. During the comparison, the model of XLNet with topic distributions showed the best results.

2.3 Other features for Covid-19 Fake News Detection

Both classical machine learning models and neural networks play an important role in the problem of fake news detection connected with Covid-19, but there are a lot of other features and approaches, which are popular in natural language processing and which could be very helpful for fake news detection. For example, we can mention algorithms using n-grams of words or characters [²³], GloVe [²⁴], an unsupervised learning algorithm for obtaining vector representations for words, Fast-text [²⁵], the classifier, which is often on par with deep learning classifiers in terms of accuracy, and many orders of magnitude faster for training and evaluation, Label Smoothing – a regularization technique that introduces noise for the labels [²⁶]; adversarial training [²⁷], a novel regularization method for classifiers to improve model robustness for small, approximately worst case perturbations; and tax2vec [²⁸], a Semantic space vectorization algorithm.

There are some interesting researches describing the impact of fake news about Covid-19, for example [²⁹], which is devoted to researching the social impact of fake news in a context of health information in social media. Researchers also try to define false information and information, which is the evidence of the social impact shared in social media.

The authors analyzed data from three social media platforms: Twitter, Facebook and Reddit. The authors' research questions were about ways that social media messages focused on fake health information, and how interactions based on health evidence with social impacts help overpower fake health information.

The authors used a methodology of social impact in social media, which combined quantitative and qualitative content analysis of the collected messages. The messages' selection corresponded to three criteria: relevance of the number of active users in millions according to Statista 2019 data, availability of public messages, and suitability for online discussion. The selection data contained the hashtags “health”, “vaccines”, “nutrition” and “Ebola”, which matched well with the interests of researchers.

The authors divided the data in four groups: ESISM (the message is an example of evidence of the social impact shared in social media), MISFA (the message is a fake news), OPINION (the message is a user’s opinion) and INFO (the message is a fact or news).

As a result, the authors found that misinformation and fake news is higher in Twitter (19%) than in Facebook (4%) or Reddit (7%). The researchers also found that messages focused on false health information are mostly aggressive in the analyzed data, while messages based on evidence of social impact are more peaceful and respectful.

In addition, messages based on evidence of social impact successfully challenged false information in cases when the authors of the misinformation messages were polite and open to discussion. However, this is not the case when the authors of misinformation messages showed disrespect and an aggressive position against science.

In [³⁰], the authors used natural language processing to analyze messages connected with COVID-19 and to indicate the main topics, which people discuss in social media in the frames of the pandemic. To build the dataset, the researchers extracted messages and comments from Twitter, Youtube, Facebook and three online discussion forums.

At the preprocessing stage the authors removed URLs, HTML tags, number words, special characters that are not required for sentence boundary detection, hash tags, and mentions.

The authors expanded contractions, compressed words with repeated characters, and converted slangs to English words.

Afterwards, the researchers implemented the Keyphrase Extractor algorithm using Python, which contains 7 steps: (1) grammar determination (where all words are marked as corresponding parts of speech); (2) sentence breaking and tokenization (where documents are separated into sentences using Python NLTK’s tokenize library and sentences into words or tokens); (3) POS tagging (where each token is assigned a POS tag); (4) lemmatization (where each token is converted to its root word); (5) chunking (where words are combined in chunks using special rules); (6) transformation and filtering (where stopword keyphrases are deleted); and (7) sentiment scoring and filtering (the last step where the authors assigned a sentiment score to each keyphrase using VADER lexicon-based algorithm).

As a result, the authors using keyphrases and experts assessments revealed 34 negative and 20 positive categories. The top-5 negative themes included concerns about social distancing and isolation policies, misinformation, political influence, financial issues, and poor governance. The top-5 positive themes connected with the pandemic were public awareness, spiritual support, encouragement, charity, and entertainment.

3.1 Challenge and Dataset

The main goal of the mentioned challenge was to create a system for binary classification of fake news and real news. The challenge contained the same task for two languages: English and Hindi, but in this survey, we concentrated our attention only on the English task and dataset.

The dataset for Covid-19 fake news detection contains 10,700 messages, with 5,100 of them real and 5,600 of them fake.

Real news was collected from reliable sources such as the World Health Organization (WHO), Centers for Disease Control and Prevention (CDC), and others, whereas fake news was collected from social media such as Facebook posts, Twitter tweets, Instagram posts, etc. The dataset contains 37,503 unique words, and the principal statistics are presented in Table 1.

The interesting observation from the dataset is that on average real news tend to be longer than the fake ones by approximately 10 words and the number of characters in a fake message is lower than in the real one by nearly 75 characters.

Some examples of real and fake news from the dataset (which were collected from Twitter) are presented in Table 2.

The shared task’s participants’ submissions were ranked according to their weighted average F1-score. F1-score was calculated for fake news’ class and real news’ class, while the average is weighted by the number of true instances for that class.

3.2 Best Approaches from the Challenge

There were 166 teams who took part in the challenge with the English dataset, and 114 of them beat the baseline of 93% F1-score. The best results are very close to each other and are higher than 98% F1-score. The top three results are presented in Table 3.

The winner of the challenge g2tmn team [³³] achieved 98.69% F1-score using the ensemble of three pretrained CT-BERT models with random seed values and with different data splitting into training and validation samples. In addition, at the preprocessing step, the researchers used the Python emoji library to replace the emoji with short textual descriptions [³⁴], URLs tokenizations (a replacement with $URL$ token), and converting texts of messages to lowercase.

The second result 98.65% F1-score was achieved by the saradhix team [³⁵]. During the research, the authors used several classical machine learning methods such as Naive Bayes, Logistic Regression, Random Forest, XGBoost, Support Vector Machine, and also several Transformer models, including BERT, DB-BERT, RoBERTa, Electra and XLNet. The best results were obtained using a RoBERTa-based model with 12-layer, 768-hidden, 12-heads, 125M parameters. Researchers noted that RoBERTa is a Robust BERT model, which has been trained on a much larger dataset and for a much larger number of iterations with a larger batch size.

The third place with 98.6% F1-score was obtained by the xiangyangli team [³⁶]. In this case, the authors also used Text Transformers for their research. Additionally, the authors used a Pseudo Label Algorithm to do data augmentation, because the dataset was quite small. As a test data was predicted with a probability greater than 0.95, the authors assumed that the data was predicted correctly with a relatively high confidence and added it into the training set. The authors used Text Transformers to create the ensemble, which is made of BERT, Ernie, XL-Net, RoBERTa and Electra models, and cross validation with Pseudo Label Algorithm to achieve their best result.

To summarize the best approaches from the Constraint@2021 Fake News Detection open shared task, we see that: