Evaluating Deep Learning Approaches for Covid19 Fake News Detection

1 Introduction

• Fake news is defined as intentionally shared, verifiably false information, and the introduction notes serious harms and particular concern during the Covid-19 crisis.The text cites political, social, and economic bias, mob lynchings and riots, and reported suicides after Covid diagnoses,,,.Source: §1 · p. 2
• Manual verification is described as reliable but too slow for the large volumes of online content, motivating automated detection.The introduction notes that existing approaches may use content as well as user, network, polarity, and behavior signals, [27],.Source: §1 · p. 2
• The paper evaluates Covid-19 fake-news detection systems for the Contraint@AAAI 2021 shared task, using data collected from social media.The task classifies Covid-19-related news as fake or real.Source: §1 · p. 2
• The paper treats fake-news detection as text classification, evaluating CNN, LSTM, and BERT-based models and Covid-related pretraining or word embeddings.The models aim to capture discriminative linguistic, style, and polarity features from news text.Source: §1 · p. 2

2 Related Works

• Social-context approaches use credibility and behavior cues; other work combines URLs, propagation features, and content, or represents topic spread as a diffusion network,.The cited diffusion-network work uses social-network features with classifiers such as SVM, random forest, or decision tree.Source: §2 · p. 3
• Comment-language analysis clusters inquiry phrases such as “Really?” and “Is it true?” to help classify rumors.Source: §2 · p. 3
• TriFN combines news, user, user-news, and publisher-news embeddings with a semi-supervised classifier for fake-news detection.Source: §2 · p. 3
• Propagation-path work represents fake-news spread as vectors and uses RNNs for global and CNNs for local path variation.Source: §2 · p. 3
• Content-focused work uses explicit and latent text features or CNN-derived contextual information to classify fake news,.Source: §2 · p. 3

3 Architecture Details

• The architecture overview groups two simple models on the left and two transformer-based models on the right.The section introduces the techniques and hyperparameters used for text classification.Source: §3 · pp. 3–4, Figure 1 · p. 4

3.1 CNN

• The CNN extracts features from 300-dimensional fastText embeddings using parallel 1D convolutions, followed by convolution, pooling, and dense layers [14].The parallel kernels have sizes 2–6, with 128 filters each; dropout is used after later convolutional and early dense layers.Source: §3.1 · p. 4

3.2 LSTM

• The LSTM model embeds sequences sized to the longest training tweets, then uses an LSTM layer, dropout, and two dense layers.The LSTM has 128 units; the dense layers have 128 and 2 units.Source: §3.2 · pp. 4–5

3.3 Bi-LSTM + Attention

• The Bi-LSTM + Attention model reads sequences in both directions, then applies attention and dense layers.Each direction has 256 units; the dense layers have 128 and 2 units.Source: §3.3 · p. 5

3.4 HAN

• HAN encodes words and sentences with bidirectional LSTMs, using attention at both levels to form a document vector [29].Samples are limited to 40 sentences, with up to 50 words per sentence.Source: §3.4 · p. 5

3.5 Transformers

• Transformer models use self-attention for contextual word representations; classification uses the encoder output, with BERT and DistilBERT adaptations,.BERT classifies the final encoder’s [CLS] representation [10]; DistilBERT halves the layers and is described as 40% smaller than BERT [21].Source: §3.5 · pp. 5–6, BERT · p. 6, DistilBERT · p. 6

4.1 Dataset details

• The shared-task dataset contains 10,700 labeled social-media posts, split into 6,420 training, 2,140 validation, and 2,140 test samples.The test and validation sets each contain 1,020 fake and 1,120 real samples; 2,998 test tokens and 2,888 validation tokens are absent from training.Source: §4.1 · p. 6
• An unlabelled corpus of COVID-19 tweets gathered using the #covid19 hashtag was used for further BERT pretraining and FastText experiments.Source: §4.1 · pp. 6–7

4.2 Preprocessing of the dataset

• Preprocessing removed HTML tags, converted accented characters, expanded contractions, removed special characters, and filtered noise.Source: §4.2 · p. 7
• Text was lowercased, stop words were removed, and Porter’s Stemmer was used to reduce words to their roots.Source: §4.2 · p. 7
• Sequential models used trainable embeddings: 100-dimensional pretrained GloVe or 300-dimensional FastText trained on task and COVID-tweet corpora.Source: §4.2 · pp. 7–8

4.3 Training Details

• Transformer experiments included fine-tuned BERT and DistilBERT, publicly shared COVID-pretrained BERT models, and further BERT or DistilBERT language-model pretraining on COVID tweets.Models were trained for up to 10 epochs, with validation loss used to select the best epoch.Source: Transformer-based architectures · p. 8, §4.3 · p. 8

• BERT and DistilBERT pretrained on COVID-19 tweets performed better than versions only fine-tuned on the dataset.Source: §5 · p. 8
• COVID-tweet-pretrained bert-cased gave the best results, followed by Covid-Twitter-Bert; transformer models outperformed non-transformer word-based models overall.Source: §5 · pp. 8–9, Table 2 · p. 9
• Among non-transformer models, HAN gave the best results, while target-corpus FastText vectors performed slightly better than pretrained GloVe embeddings.Source: §5 · p. 9

• 98.41% accuracy from language-model-pretrained BERT exceeded the 93.32% baseline, while transformers outperformed other basic models by an absolute 3–4% in accuracy.Source: §6 · p. 9
• Manually pretraining on a subject-related corpus and then adapting the model to the task gave the best accuracy.Source: §6 · p. 9