Apprehension of Spurious Information - A Hybrid Architecture to Detect COVID Fake News

doi:10.21203/rs.3.rs-1873766/v2

Propagation of false information has an enduring effect on the economy and can lead to political tensions. In order to counterattack this unexpected repercussion, a system has to be modelled to distinguish the hoax from the real. This paper proposes a deep learning based hybrid model that can determine if the news is veracious. The performance of the model is tested on four datasets including COVID fake news to ensure that it is able to generalise across different domains and achieve best results. Bidirectional Gated Recurrent Units emerges to be the ideal decoder to work along with the Convolutional Neural Network in this hybrid architecture while the words being represented using the glove.6B.50D embeddings provided to be the best representation when juxtaposed with Word2Vec, BERT Embeddings, glove.6B.100D and keras embedding layer. Experiments on the datasets show promising results with an overall accuracy of 80.17% on Disaster Tweets, 99% on ISOT dataset, 59.86% on LIAR and 90.05% on COVID Fake News dataset.

COVID

Hybrid Architecture

Fake News

Word Embedding

With the onset of the technological boom and the affordability of smart gadgets, the dispersion of information has become easier than ever. Indeed the objective of fast delivery of the intended message has been accomplished. While this can be deemed a successful venture, there have been spin-offs which have led to a quandary. The question of whether the information received in a fraction of seconds is reliable or not. This has been a reassurance of the fact that one innovation paves way for the need of many more. Seizing the spurious news amongst the sea of details being passed on every moment has captured the limelight of the technologists. It is imperative that these news are stopped from further propagation using state-of-the-art mechanisms.

In the first place, this task of demarcating the fake from the real has become more challenging considering the succinct nature of news which has been more preferable for the fast moving generation at present. Several initiatives have been undertaken such as the ‘Chain Breaking Brigade’ to fight against viral fake news. This movement gained more prominence with the surge in the fake news of the details of vaccines against the CoronaVirus when the vaccination campaign rolled out, engendering more chaos over the safety and reliability of the vaccines. Moreover, for any robust system capable of seizing these details, understanding the context and extracting details out of it is a key component.

The advent of deep learning has ameliorated the technicalities involved by eradicating handcrafted rules. This is accomplished using the domain of Natural Language Processing (NLP), a subset of Deep Learning Techniques. NLP has wide applications in numerous scenarios like information extraction and retrieval, textual summarization, intent detection and slot filling in voice assistants etc. Its ability to work on natural human language after transmuting into machine decipherable format makes it an ideal candidate to work on the task at hand.

Although in the literature, there exists a significant number of works related to fake news detection, there still is room for progress, especially with regard to making these models more domain-invariant as possible. The problem is more challenging considering the dynamically evolving changes on the forms of fake news. In most of the existing works, this task has been carried out using a CNN or RNN in isolation.

In order to detect spurious news, a hybrid Natural Language Processing model has been put forth in this paper. The hybrid nature of the architecture ensures that the local features are encoded using a convolutional neural network. The recurrent part of this network decodes the feature map whose representations are used to determine if the news is reliable or hoax. Hence this conundrum is modelled as a binary classification problem.

The major contributions of this work are:

We propose a deep hybrid architecture which classifies the input information as Fake or Real by employing different RNN architectures such as LSTM, GRU, Bi-LSTM and Bi-GRU upon the local feature encoding with the Convolutional Neural Network.
Understand the working of different word embeddings and identify the one with the optimal performance among the Keras Embedding layer, Word2Vec, glove.6B.50D, glove.6B.100D and the BERT Embeddings.

The remainder of this paper is organized as follows: In the “Related Work” section, a survey on the approaches taken to tackle this task is briefed upon. We describe our proposed approach in the “Methodology” section. Experimental setup and the datasets are reported in the “Dataset and Experiments” section. Further under the "Results and Analysis" section, the performance of the system is noted under various scenarios and analysis are derived out of the same. Finally, the concluding remarks and directions for future research are presented in the “Conclusion” section.

Though the pioneer works in this domain of Fake News Detection revolved around machine learning techniques, the advent of Deep Learning could be attributed to the surge in the improvements recorded for the task. Inspired by the memory power of Vanilla Recurrent Neural Networks (RNN) state of the art results were recorded in standard datasets coupled with a refined word representation technique. Liu et al. [10] modelled the classification of propagation paths for an early curb of Fake News using RNN. Also a deeper understanding of the propagation of spurious social content was put forth by Wu et al. [18] using these vanilla networks.

Furthermore an advancement of the existing memory networks are the Long Short Term Memory Cells (LSTM) which are capable of retaining a superior amount of details due to the gating mechanism involved. Bangyal et al. [6] adopted various Machine Learning and Deep Learning approaches with TF-IDF Vectorization strategy but achieved the optimal performance with LSTMs. The works which emphasized on LSTMs varied in terms of the embedding strategy used. Ahmad et al. [1] experimented LSTMs with traditional methods like term frequency–inverse document frequency, count vector, character level vector and N-Gram level vector upon the Fake News Detection corpus. Rodríguez et al. [14] used BERT word embeddings and exploited various models like LSTM and CNN. Other works used one hot representation to convert text into numerical form before feeding into the LSTM network.

As a counterpart to these LSTM networks, the Gated Recurrent Units (GRU) gained prominence down the lane as in Kaliyar et al. [7] work on FNDNet with Word2Vec encoding. Moreover the comprehensive study of neural network and transformer based language models carried out by Al-Yahya et al. [3] used GRUs over the Arabic Corpora on Covid Fake News. A plethora of embedding strategies were tested with GRU for this task. Glove was resorted to as the embedding mechanism by Bajaj et al. [5] on the signal media news dataset, while on the other hand fasttext was used by Verma et al. [17] in their work to detect fake news using Natural Language Processing.

While memory retention had been attended to with the above networks, the challenge was to instill further semantic understanding of the sentence in hand. One way to bring about this was the inclusion of the bidirectional feature, where a better understanding is achieved upon scanning from both directions of the query. Bahad et al. [4] experimented Bidirectional LSTMs on structured news articles for this task using Glove.6B embeddings, while Kashyap et al. [11] put forth an end-to-end model for evidence aware credibility using the same network. Reddy et al. [13] attempted a Bi-GRU based mis-information detector for the Urdu Corpus. Appending attention to these networks Kumar et al. [8] proposed the Fake News Net with an ensemble approach with CNN, Bi-LSTM and attention.

Having exploited the various advantages of the available recurrent networks, further work was carried out by incorporating an amalgamation of convolutional networks with the recurrent ones. Convolutional Neural Networks are primarily used for dealing with images. However, their ability to extract local features had proven to be cardinal as in the experiments conducted by Lin et al. [9] where semantic information is captured with a multi-level CNN architecture and Yang et al. [20] novel approach on TI-CNN which uses the information from both Text and Image to detect fake news.

Despite experiments incorporating the hybrid structure of Convolutional and Recurrent Networks as in [1, 2, 8, 12, 15, 16, 19, 21], the existing implementations fall short in generalizing across news from disparate domains. The correlation between the encoding strategies on the results obtained have been difficult to assert and hence an unanimous decision can’t be arrived at. In this work, we propose a hybrid architecture which classifies the input information as Fake or Real and also to explore the suitable architectures for the decoder of the model, having a CNN encoder. We aim to understand the working of different word embeddings and identify the one with the optimal performance for the task at hand.

The overall block diagram of our proposed architecture is depicted in Fig. 1 Our model is a deep hybrid architecture which performs the task of fake news detection using the underlying representations.

Figure 1. Deep Hybrid Fake News Detection Workflow

4.1. Preprocessing

This binary classification task entails the labels to be encoded as integers as neural networks can’t directly processing the string labels. Hence this is achieved using the label encoder. Real is encoded with the label 1, while fake is encoded with the label 0. As a part of the pre-processing, the text is converted into lower case after which the white spaces and decimals in the text are removed. The stopwords in English are removed with the help of NLTK upon which stemming is carried out. For the LIAR dataset, the statement and label attributes are alone considered to make the prediction. Though this dataset has 6 output classes, it has been modelled as a binary classification problem with true, mostly-true and half-true being represented as real and false, barely-true and pants-fire denoted as fake. In the ISOT dataset, texts are first cleared from publisher details, using the re python package for regular expressions. Then further processing is carried out.

4.2. Formulating Word Embeddings

The label matrices are encoded and therefore the text is vectorized by the tok-enizer available in keras. It helps in converting the input sentence into a sequence of integers. The length of sequences is set to 300 and a post-padding is applied so as to use it for model training. This is often done because the length of every sequence varies. By fixing the length of every text sequence to 300, it’s necessary to append zeros in each sequence that’s shorter than the fixed length. The ones longer than the fixed length are truncated. To facilitate the feature extraction by the CNN, word embeddings are used which are either pretrained on a standard corpus or trained on the dataset at hand. In this work pretrained BERT, Glove and Word2Vec embeddings are used. The keras embeddings are obtained based on the representations from the training data.

4.3. Feature Extraction Using CNN

The word representations from the previous step are passed through a Conv1D layer. Rectified Linear Units Activation is used for this layer having 32 filters with a kernel size of 8. Despite being used predominantly for computer vision related tasks, their ability to convolve over a fixed data window makes them suitable for tasks related to Natural Language Processing. Each of the filter aids in detecting various details related to the text. In order to bring about a representation of higher order, max pooling layer is used with a pooling size of 2 to reduce the parameters involved.

4.4. Feature Map Decoding using RNN

In order to get a semantic understanding of the sentence, decoders like Long Short Term Memory (LSTM), Bidirectional LSTM, Gated Recurrent Units (GRU) and Bidirectional GRU were appended with the previous CNN to determine the veracity of the news. Though the LSTM architectures overcome the vanishing gradients problem faced by vannila RNN, their Bidirectional Versions help in providing more context as the sentence is traversed in both directions contributing to a better semantic understanding.

4.1. Datasets

We evaluate our proposed deep hybrid model on four datasets. The first dataset is the well-known COVID Fake News corpus, which is followed by the Disaster Tweets dataset and the ISOT Dataset. Finally the performance is also tested on the LIAR dataset modelled as a binary classification task. The distribution for all the datasets are given in Table 1 - Table 4.

Covid Fake News Dataset

This dataset has social media posts and articles of real and fake news on COVID-19. Its attributes include id, tweet and label and is composed of 12836 records with two possible output classes - Fake and Real.

Table 1

Covid Fake News Dataset Composition
Dataset	#Train	#Test	#Val
Covid Fake News	6420	2429	2380

Disaster Tweets Dataset

This dataset has manually labelled short sentences collected to determine if the information is about a real disaster or not. Its attributes include id, keyword, location, text and target and is composed of 12836 records with two possible output classes - Fake and Real.

Table 2

Disaster Tweets Dataset Composition
Dataset	#Train	#Test
Disaster Tweets	6981	3109

ISOT Dataset

This dataset has news from politifact.com. Its attributes include title, text, subject and date and is composed of 44835 records with two possible output classes - Fake and Real.

Table 3

ISOT Dataset Composition
Dataset	#real	#fake
ISOT	21417	23481

LIAR Dataset

This dataset has manually labelled short sentences collected from politi-fact.com for fact checking research. Its attributes include id, label, statement, subject, speaker and is composed of 12836 records with six possible output classes - true, mostly-true, half-true, barely-true, false, pants-fire.

Table 4

LIAR Dataset Composition
Dataset	#Train	#Test	#Val
LIAR	10269	1283	1284

4.2. Training Details

We have used the python-based neural network package, Keras for the implementation. In our work, a pair of CNN and max pooling layer was followed by one layer of RNN and a MLP. We fix the number of neurons on the Bi-LSTM/Bi-GRU layer to be 32. The dimensions of the experimented word embedding ranges from 50 for glove to 300 for Word2Vec. We use ReLU activations for the intermediate layers of our model and sigmoid activation for the output layer.

Two regularisation mechanisms have been put into use to ensure that the model avoids overfitting and the bestfit is chosen. One is Early Stopping where we stop training the model if the performance constantly degrades after a fixed number of epochs. Another is Dropout which randomly sets the outgoing edges of the hidden layers to zero at each update during the training phase. The performance of the model has been observed on both conditions - Including Dropout and Not including dropout and has been recorded. We used 40% dropout and ‘Adam’ optimizer for regularisation and optimization. Model parameters are updated using the binary cross-entropy. Table 5 lists the different parameters that we experimented with and the final chosen parameters of the proposed model.

The Sequential interface has been utilised to incorporate the fake news detection model. Its structure has the following layers - Embedding, Conv1D, Max Pooling1D, Bi-GRU, Flatten, Dense, Sigmoid. The pre-trained embeddings are put into use with the help of the embedding layer provided by Keras. ReLU activation is used in the Convolutional layer which does the process of extracting local features from the text. To sample down the feature vectors, a max-pooling layer is also used. These down-sampled feature maps are fed into the RNN (LSTM / GRU / Bi-GRU / Bi-LSTM) layer that follows. In the end, a binary classification is performed using sigmoid activation. The various hyperparameters tuned are presented in Table. 5 which covers the various experimentations carried out and the best performing hyperparameter.

Table 5

Hyperparameter Tuning
Hyperparameters	Range	Final
Word Embedding	Keras Embedding , Word2Vec, glove.6B.50D, glove.6B.100D, BERT	glove.6B.50D
Embedding Dimension	50 / 100 / 300	50
Padding	Post	Post
Truncating	Post	Post
Regularization Mechanisms	Dropout, Early Stopping	Dropout, Early Stopping
Dropout	0-0.6	0.4
CNN #Filters	32	32
CNN Filter size	8	8
RNN	GRU / LSTM	GRU
Bi-Directional	True / False	True
Hidden Size	32	32
Optimizer	Adam	Adam
Loss Function	Binary Cross Entropy	Binary Cross Entropy
Epochs	15–25	25
Batch Size	128 / 256	128

5.1. Effect of Word Embeddings

Word2Vec takes texts as training data for a neural network. The resulting embedding captures whether words appear in similar contexts. GloVe focuses on word co-occurrences over the whole corpus. Its embeddings relate to the probabilities that two words appear together. This has enabled it to outperform Word2Vec embeddings. GloVe works with the traditional word-like tokens, whereas BERT segments its input into subword units called word-pieces. While it ensures there are no out-of-vocabulary tokens, on the other hand, totally unknown words get split into characters and BERT was not able to make much sense of them either.From the presented experimental results, it can be observed that Glove.6B.50D word embeddings record the best performance for three datasets - LIAR, ISOT and Disaster Tweets.

5.2. Effect of Dropout

Surprisingly dropout hasn’t impacted the performance of the model to a major extent. This can be attributed to the hybrid nature of the CNN - RNN architecture adopted. Since convolutional layers have few parameters, they need less regularization to begin with. Furthermore, because of the spatial relationships encoded in feature maps, activations had become highly correlated. This has rendered dropout ineffective.

5.3. Error Analysis

To get an idea where our system fails, we performed a detailed error analysis of the models. Upon analysing it has been found that the model fails to achieve the desired results when there isn’t substantial information in the input news. It falls short in understanding the nuances in smaller phrases like raining flooding Florida TampaBay Tampa 18 or 19 days; Says Paul Ryan is still endorsing Trump. ; Building the perfect tracklist to life leave the streets ablaze ; Whereas optimal performance is achieved when supplementing data is provided like in 9 Mile backup on I − 77 South. accident blocking the Right 2 Lanes at Exit 31 Langtree Rd. consider NC 115 or NC 150 to NC 16 as alternate. ; The nuclear test conducted in our nation this time is the Earth’s 2,054th nuclear test. The five permanent members of the U.N. Security Council have conducted

99.99 percent of the total nuclear tests.

This explains its better performance in datasets like ISOT, Covid Fake News where more than 90% accuracy is reported due to its inherent comprehensive nature. On the other hand, datasets like Disaster Tweets and LIAR display subordinate performances due to its succinct nature of the input news. For the ISOT dataset, a certain amount of data cleaning has proven effective in the model’s predictions. This is due to the fact that the majority of the labelled real news has the source of the publication attached, which is missing in Fake news. Hence to bring about a level playing field, the publication’s source details have been removed and this has impacted the performance of the model. Table 6 to Table 13 summarises the performance of the baseline methods for various datasets.

Table 6

Disaster Tweets – Performance Metrics
Model Description		With DropOut			Without DropOut
Decoder	Embedding	Precision	Recall	F1	Precision	Recall	F1
Bi-GRU	BERT	0.66	0.79	0.79	0.6	0.8	0.8
LSTM		0.67	0.78	0.78	0.65	0.79	0.78
GRU		0.59	0.78	0.78	0.69	0.76	0.76
Bi-LSTM		0.65	0.77	0.77	0.53	0.8	0.78
BiGRU	Glove.6B.100d	0.47	0.81	0.8	0.46	0.8	0.8
BiLSTM		0.48	0.79	0.79	0.47	0.8	0.8
LSTM		0.5	0.79	0.79	0.47	0.79	0.79
GRU		0.49	0.78	0.78	0.45	0.8	0.8
BiGRU	Glove.6B.50d	0.45	0.8	0.8	0.45	0.8	0.8
LSTM		0.46	0.79	0.8	0.46	0.79	0.8
GRU		0.45	0.8	0.8	0.45	0.8	0.8
BiLSTM		0.47	0.79	0.79	0.46	0.81	0.8
GRU	Keras Embedding	0.45	0.8	0.8	0.46	0.79	0.79
LSTM		0.48	0.79	0.79	0.48	0.79	0.79
BiGRU		0.46	0.79	0.79	0.46	0.8	0.79
BiLSTM		0.63	0.78	0.78	0.48	0.8	0.79
GRU	Word2Vec	0.51	0.8	0.79	0.64	0.76	0.75
BiGRU		0.85	0.79	0.79	0.79	0.75	0.74
LSTM		0.72	0.77	0.77	0.74	0.77	0.77
BiLSTM		0.82	0.76	0.76	0.79	0.78	0.78

Table 7

Disaster Tweets (Accuracy in %)
Embedding	LSTM		Bi-LSTM		GRU		Bi-GRU
Embedding	With DropOut	No DropOut	With DropOut	No DropOut	With DropOut	No DropOut	With DropOut	No DropOut
Keras Embedding	78.6	78.9	77.7	79.1	79.8	78.9	78.8	79.3
Word2Vec	77.1	76.9	76.2	77.9	79.3	74.9	79.0	73.7
Glove50D	79.9	79.5	79.0	80.5	79.5	80.2	80.2	80.1
Glove100D	78.9	79.1	79.5	79.7	78.5	80.0	79.8	80.0
BERT	78.2	78.4	77.2	77.9	77.9	76.2	78.6	79.7

Table 8

ISOT – Performance Metrics
Model Description		With DropOut			Without DropOut
Decoder	Embedding	Precision	Recall	F1	Precision	Recall	F1
Bi-GRU	BERT	0.04	0.99	0.99	0.04	0.99	0.99
Bi-LSTM		0.04	0.99	0.99	0.08	0.98	0.98
LSTM		0.04	0.99	0.99	0.04	0.99	0.99
GRU		0.07	0.98	0.98	0.07	0.98	0.98
BiGRU	Glove.6B.100d	0.11	0.96	0.96	0.11	0.96	0.96
LSTM		0.11	0.96	0.96	0.11	0.96	0.96
GRU		0.11	0.96	0.96	0.12	0.96	0.96
BiLSTM		0.13	0.96	0.96	0.11	0.96	0.96
BiGRU	Glove.6B.50d	0.09	0.99	0.99	0.05	0.99	0.99
GRU		0.25	0.95	0.95	0.19	0.95	0.95
BiLSTM		0.16	0.96	0.96	0.12	0.96	0.96
LSTM		0.20	0.95	0.95	0.21	0.96	0.96
BiGRU	Keras Embedding	0.11	0.96	0.96	0.13	0.96	0.96
GRU		0.11	0.96	0.96	0.19	0.95	0.95
BiLSTM		0.16	0.96	0.96	0.13	0.96	0.96
LSTM		0.20	0.95	0.95	0.11	0.96	0.96
LSTM	Word2Vec	0.05	0.99	0.99	0.08	0.98	0.98
GRU		0.06	0.99	0.99	0.07	0.98	0.98
BiLSTM		0.06	0.99	0.99	0.05	0.99	0.99
BiGRU		0.05	0.99	0.99	0.06	0.99	0.99

Table 9

ISOT (Accuracy in %)
Embedding	LSTM		Bi-LSTM		GRU		Bi-GRU
Embedding	With DropOut	No DropOut	With DropOut	No DropOut	With DropOut	No DropOut	With DropOut	No DropOut
Keras Embedding	95.4	96.4	96.0	95.7	96.3	95.3	96.2	95.8
Word2Vec	98.9	98.5	98.8	98.7	98.9	98.3	98.5	98.6
Glove50D	98.1	98.9	99.0	98.9	99.0	98.9	99.0	99.0
Glove100D	96.2	95.6	95.7	96.3	96.0	95.9	96.2	96.5
BERT	98.8	98.8	98.9	98.5	98.3	98.2	99.0	98.8

Table 10

Covid Fake News (Accuracy in %)
Embedding	LSTM		Bi-LSTM		GRU		Bi-GRU
Embedding	With DropOut	No DropOut	With DropOut	No DropOut	With DropOut	No DropOut	With DropOut	No DropOut
Keras Embedding	89.4	90.1	89.9	89.6	89.2	89.8	89.8	90.1
Word2Vec	89.5	89.2	88.6	89.9	90.6	88.7	90.1	89.8
Glove50D	86.0	85.2	84.7	85.1	86.0	86.2	85.7	85.6
Glove100D	84.9	83.5	85.2	84.1	84.6	84.1	84.6	85.8
BERT	88.7	90.2	88.6	90.5	90.3	89.7	87.5	89.4

Table 11

Covid Fake News – Performance Metrics
Model Description		With DropOut			Without DropOut
Decoder	Embedding	Precision	Recall	F1	Precision	Recall	F1
Bi-GRU	BERT	0.89	0.89	0.89	0.90	0.90	0.90
Bi-LSTM		0.88	0.88	0.88	0.89	0.88	0.88
LSTM		0.87	0.87	0.87	0.88	0.87	0.87
GRU		0.88	0.87	0.87	0.88	0.88	0.88
BiGRU	Glove.6B.100d	0.85	0.85	0.85	0.84	0.84	0.84
LSTM		0.86	0.85	0.86	0.83	0.83	0.83
GRU		0.85	0.85	0.85	0.84	0.84	0.84
BiLSTM		0.85	0.85	0.85	0.85	0.85	0.85
BiGRU	Glove.6B.50d	0.89	0.85	0.87	0.85	0.88	0.86
GRU		0.89	0.85	0.87	0.87	0.87	0.87
BiLSTM		0.87	0.86	0.87	0.87	0.87	0.87
LSTM		0.86	0.85	0.86	0.87	0.85	0.86
BiGRU	Keras Embedding	0.90	0.90	0.90	0.90	0.90	0.90
GRU		0.92	0.92	0.92	0.90	0.90	0.90
BiLSTM		0.90	0.89	0.89	0.90	0.90	0.90
LSTM		0.89	0.89	0.89	0.90	0.90	0.90
LSTM	Word2Vec	0.91	0.91	0.91	0.90	0.90	0.90
GRU		0.90	0.90	0.90	0.90	0.90	0.90
BiLSTM		0.89	0.89	0.89	0.89	0.89	0.89
BiGRU		0.89	0.89	0.89	0.90	0.90	0.90

Table 12

LIAR (Accuracy in %)
Embedding	LSTM		Bi-LSTM		GRU		Bi-GRU
Embedding	With DropOut	No DropOut	With DropOut	No DropOut	With DropOut	No DropOut	With DropOut	No DropOut
Keras Embedding	57.9	56.0	56.4	56.6	58.2	56.8	55.6	56.8
Word2Vec	58.0	54.3	53.8	58.6	58.2	56.5	55.4	57.5
Glove50D	54.2	58.7	56.0	56.8	57.5	58.2	59.9	59.4
Glove100D	56.3	58.3	57.3	57.7	58.0	58.1	58.2	59.3
BERT	58.6	57.3	58.1	58.3	58.1	55.1	58.2	58.4

Table 13

**LIAR -** Performance Metrics
Model Description		With DropOut			Without DropOut
Decoder	Embedding	Precision	Recall	F1	Precision	Recall	F1
Bi-GRU	BERT	0.54	0.59	0.44	0.57	0.57	0.57
Bi-LSTM		0.58	0.58	0.58	0.58	0.58	0.58
LSTM		0.58	0.58	0.58	0.57	0.58	0.58
GRU		0.58	0.58	0.58	0.55	0.55	0.55
BiGRU	Glove.6B.100d	0.58	0.58	0.58	0.58	0.59	0.58
LSTM		0.57	0.58	0.58	0.58	0.58	0.58
GRU		0.58	0.57	0.57	0.58	0.58	0.58
BiLSTM		0.57	0.56	0.57	0.58	0.58	0.58
BiGRU	Glove.6B.50d	0.59	0.60	0.59	0.58	0.59	0.58
GRU		0.57	0.58	0.57	0.58	0.58	0.58
BiLSTM		0.57	0.56	0.56	0.58	0.57	0.57
LSTM		0.56	0.54	0.54	0.57	0.59	0.57
BiGRU	Keras Embedding	0.58	0.58	0.58	0.57	0.57	0.57
GRU		0.58	0.58	0.58	0.57	0.56	0.56
BiLSTM		0.57	0.56	0.57	0.58	0.57	0.57
LSTM		0.57	0.56	0.56	0.57	0.57	0.57
LSTM	Word2Vec	0.58	0.58	0.58	0.57	0.56	0.57
GRU		0.57	0.58	0.58	0.57	0.54	0.55
BiLSTM		0.57	0.55	0.56	0.57	0.58	0.57
BiGRU		0.57	0.54	0.54	0.57	0.59	0.57

This paper proposes a deep hybrid approach for the task of fake news detection. Baseline models like Bi-LSTM, Bi-GRU, GRU and LSTM are used to learn the representations of the sequence. Bidirectional Gated Recurrent Units exhibit advantages over the other architectures when its performance was checked on the four chosen datasets. By using pre-trained word embedding, it further helps in better representations of the input and we have found glove.6B.50D to perform the best for this task. The generalization ability of the model has been exhibited as the model performs well across datasets from different domains. In future, we plan to work on state of the art NLP architectures like transformers and also try out a multi-modal approach for fake news detection.

Ethical Approval NA

Consent to Participate NA

Consent to Publish NA

Competing Interests NIL

Author Contributions Statements

S. Srihari and V. Shakthi Prarthana were responsible for idea formation and Implementation

G.S. Mahalakshmi supervised the work

S. Srihari edited the manuscript

G.S. Mahalakshmi and S. Sendhilkumar reviewed the manuscript

Funding NA

Availability of Data and Materials

Covid fake news - https://paperswithcode.com/dataset/covid-19-fake-news-dataset

Disaster tweets - https://www.kaggle.com/datasets/vbmokin/nlp-with-disaster-tweets-cleaning-data

ISOT dataset - https://paperswithcode.com/dataset/isot-fake-news-dataset

LIAR Dataset- https://paperswithcode.com/dataset/liar

Ahmad, I., Yousaf, M., Yousaf, S., & Ahmad, M. O. (2020). Fake news detection using machine learning ensemble methods. Complexity, 2020.
Ajao, O., Bhowmik, D., & Zargari, S. (2018, July). Fake news identification on twitter with hybrid cnn and rnn models. In Proceedings of the 9th international conference on social media and society (pp. 226-230).
Al-Yahya, M., Al-Khalifa, H., Al-Baity, H., AlSaeed, D., & Essam, A. (2021). Research Article Arabic Fake News Detection: Comparative Study of Neural Networks and Transformer-Based Approaches.
Bahad, P., Saxena, P., & Kamal, R. (2019). Fake news detection using bi-directional LSTM-recurrent neural network. Procedia Computer Science, 165, 74-82.
Bajaj, S. (2017). The pope has a new baby!. Fake news detection using deep learning, 1-8.
Bangyal, W. H., Qasim, R., Ahmad, Z., Dar, H., Rukhsar, L., Aman, Z., & Ahmad, J. (2021). Detection of fake news text classification on COVID-19 using deep learning approaches. Computational and mathematical methods in medicine, 2021.
Kaliyar, R. K. (2018, December). Fake news detection using a deep neural network. In 2018 4th International Conference on Computing Communication and Automation (ICCCA) (pp. 1-7). IEEE.
Kumar, S., Asthana, R., Upadhyay, S., Upreti, N., & Akbar, M. (2020). Fake news detection using deep learning models: A novel approach. Transactions on Emerging Telecommunications Technologies, 31(2), e3767.
Li, Q., Hu, Q., Lu, Y., Yang, Y., & Cheng, J. (2020). Multi-level word features based on CNN for fake news detection in cultural communication. Personal and Ubiquitous Computing, 24(2), 259-272.
Liu, Y., & Wu, Y. F. (2018, April). Early detection of fake news on social media through propagation path classification with recurrent and convolutional networks. In Proceedings of the AAAI conference on artificial intelligence (Vol. 32, No. 1).
Popat, K., Mukherjee, S., Yates, A., & Weikum, G. (2018). Declare: Debunking fake news and false claims using evidence-aware deep learning. arXiv preprint arXiv:1809.06416.
Qawasmeh, E., Tawalbeh, M., & Abdullah, M. (2019, October). Automatic identification of fake news using deep learning. In 2019 Sixth international conference on social networks analysis, Management and Security (SNAMS) (pp. 383-388). IEEE.
Reddy, S. M., Suman, C., Saha, S., & Bhattacharyya, P. (2020). A GRU-based Fake News Prediction System: Working Notes for UrduFake-FIRE 2020. In FIRE (Working Notes) (pp. 464-468).
Rodríguez, Á. I., & Iglesias, L. L. (2019). Fake news detection using deep learning. arXiv preprint arXiv:1910.03496.
Shishah, W. (2021). Fake News Detection Using BERT Model with Joint Learning. Arabian Journal for Science and Engineering, 46(9), 9115-9127.
Umer, M., Imtiaz, Z., Ullah, S., Mehmood, A., Choi, G. S., & On, B. W. (2020). Fake news stance detection using deep learning architecture (CNN-LSTM). IEEE Access, 8, 156695-156706.
Verma, A., Mittal, V., & Dawn, S. (2019, August). FIND: Fake information and news detections using deep learning. In 2019 twelfth international conference on contemporary computing (IC3) (pp. 1-7). IEEE.
Wu, L., & Liu, H. (2018, February). Tracing fake-news footprints: Characterizing social media messages by how they propagate. In Proceedings of the eleventh ACM international conference on Web Search and Data Mining (pp. 637-645).
Wu, L., Rao, Y., Yu, H., Wang, Y., & Nazir, A. (2018, September). False information detection on social media via a hybrid deep model. In International Conference on Social Informatics (pp. 323-333). Springer, Cham.
Yang, Y., Zheng, L., Zhang, J., Cui, Q., Li, Z., & Yu, P. S. (2018). TI-CNN: Convolutional neural networks for fake news detection. arXiv preprint arXiv:1806.00749.
Oyebode, O., Ndulue, C., Mulchandani, D. et al. COVID-19 Pandemic: Identifying Key Issues Using Social Media and Natural Language Processing. J Healthc Inform Res 6, 174–207 (2022). https://doi.org/10.1007/s41666-021-00111-w

No competing interests reported.

Apprehension of Spurious Information - A Hybrid Architecture to Detect COVID Fake News

Status:

Version 2

Abstract

Figures

1. Introduction

2. Related Works

3. Methodology

4.1. Preprocessing

4.2. Formulating Word Embeddings

4.3. Feature Extraction Using CNN

4.4. Feature Map Decoding using RNN

4. Datasets And Experiments

4.1. Datasets

4.2. Training Details

5. Results & Discussion

5.1. Effect of Word Embeddings

5.2. Effect of Dropout

5.3. Error Analysis

6. Conclusion

Declarations

References

Additional Declarations

Status:

Version 2