An Emotion-Aware Multitask Approach to Fake

News and Rumor Detection Using Transfer Learning

ABSTRACT

Social networking sites, blogs, and online articles are instant sources of news for
internet users globally. However, in the absence of strict regulations mandating the
genuineness of every text on social media, it is probable that some of these texts are
fake news or rumours. Their deceptive nature and ability to propagate instantly can
have an adverse effect on society. This necessitates the need for more effective
detection of fake news and rumours on the web. In this work, we annotate four fake
news detection and rumour detection datasets with their emotion class labels using
transfer learning. We show the correlation between the legitimacy of a text with its
intrinsic emotion for fake news and rumour detection, and prove that even within the
same emotion class, fake and real news are often represented differently, which can
be used for improved feature extraction. Based on this, we propose a multi-task
framework for fake news and rumour detection, predicting both the emotion and
legitimacy of the text. We train a variety of deep learning models in single-task and
multi-task settings for a more comprehensive comparison. We further analyze the
performance of our multi-task approach for fake news detection in cross-domain
settings to verify its efficacy for better generalization across datasets, and to verify
that emotions act as a domain-independent feature. Experimental results verify
that our multi-task models consistently outperform their single-task counterparts in
terms of accuracy, precision, recall, and F1 score, both for in-domain and cross-
domain settings. We also qualitatively analyze the difference in performance in
single-task and multi-task learning models.
 EXISTING SYSTEM

In recent years, significant work has been done to tackle the problem of fake news
and rumour detection. Bhutani, Rastogi, Sehgal and Purwar (2019) used a
combination of cosine similarity, Tf-Idf and sentiment scores as features for the
detection of fake news. Pérez-Rosas, Kleinberg, Lefevre and Mihalcea (2018)
presented two novel datasets, FakeNews AMT and Celeb, for fake news detection
across multiple domains, and used hand-crafted linguistic features and an SVM
model for fake news detection. Saikh, De, Ekbal and Bhattacharyya (2019) treated
fake news detection as a text classification task and presented two deep learning
models for fake news detection on FakeNews AMT and Celeb datasets.

Bhattacharya, Patel, Gupta, Tanwar and Rodrigues (2021) proposed a Blockchain

and LSTM-based classification technique incorporating Q-GloVe word embeddings
for the detection of fake news. They efficiently used Blockchain to protect real news
from adversaries to prevent its alteration. Verma, Agrawal, Amorim and Prodan
(2021) curated a new generalized dataset by combining four previously available
benchmark datasets, to reduce bias associated with domains specific to each dataset.
They generated embeddings from the linguistic features of the combined dataset and
were used for classification using an ensemble of multiple machine learning
classifiers.

Shahid, Jamshidi, Hakak, Isah, Khan, Khan and Choo (2022) further presented a
detailed survey of topics, challenges, and future areas of research associated with
fake news detection, providing a detailed overview of dissemination, detection, and
mitigation of fake news. They further focused on a specific potential future research
agenda, i.e., using artificial intelligence explainable fake news credibility system.
Some contemporary works have studied rumour detection from the direction of
network sciences, focusing on its evolution and impact on users. Cheng, Yin,
Nazarian and Bogdan (2021b) modelled a misinformation network for the spread of
rumours during the COVID-19 pandemic. Each node in the network represented a
different rumour, while the edges indicated the similarity between the rumours. This
work showed that information propagation could be controlled by deleting some
central nodes in the misinformation graph, and further proposed a deep learning
framework to predict Twitter posts that could act as central nodes.

Verma, Bhardwaj, Aledavood, Choudhury and Kumar (2022), on the other hand,
studied the impact of misinformation about the COVID-19 pandemic on the mental
health of people and found that individuals who consumed a higher amount of
misinformation suffered from high stress and anxiety, and exhibited an increased
tendency for suicidal thoughts. Recent research works have incorporated advanced
mainstream deep learning approaches like Generative Adversarial Networks (GANs)
and Transformers for fake news detection and rumour detection to train detectors in
an unsupervised manner.

Cheng, Li, Nazarian and Bogdan (2021a) proposed a GAN-based layered model for
rumour detection, which was trained in a label-free manner without using any
verified database. The Generator and the discriminator were trained adversarially in a
sequential manner, where the generator produced rumours by fabricating the non-
rumours, and the discriminator learnt to detect rumours and non-rumours with the
help of the produced rumours. Roy, Bhanu, Saxena, Dandapat and Chandra (2022)
proposed a transformer-based architecture with a modified multi-head attention
mechanism used to learn multiple contextual correlations for enhancing feature
representation learning.

Kochkina et al. (2018a) proposed a branch LSTM-based multi-task learning model

trained using veracity and stance detection as auxiliary tasks to rumour detection.
The model trained with both veracity and stance as auxiliary tasks outperformed
single-task models, as well as multi-task models with either veracity or stance as an
auxiliary task. Building upon this work, Cheng, Nazarian and Bogdan (2020) also
used multi-task learning for the rumour classification, but unlike Kochkina et al.
(2018a), they tackled all four component tasks related to rumour detection using the
same model: rumour detection, rumour tracking, stance classification, and veracity
classification.

Disadvantages
The spread of misinformation has become a significant problem for society with
adverse consequences. Our existing research further improves upon previous
machine learning-based fake news and rumour detection approaches by incorporating
the correlated features between the legitimacy of a text and its intrinsic emotion, to
more accurately detect fake news or rumours. Instead of taking the sentiment or
emotion features as input to the models like in previous works, we apply a multi-task
learning approach to detect fake news and rumours.

Proposed System
• We verify the correlation between the legitimacy of a text and its portrayed
emotions by dimensionality reduction. We experimentally use this correlation for
improved fake news and rumour detection by proposing a multi-task framework with
fake news or rumour detection as the primary task, and multi-class emotion detection
as the auxiliary task.
• We show that for the same emotion, rumours and non-rumours are often
represented slightly differently in the feature space upon dimensionality reduction.
These differences will vary from dataset to dataset due to differences in the domain
and other biases. We suggest that these differences can aid the model to better
incorporate this correlation for improved rumours and fake news detection.
• We further discover that emotion-guided implicit features are more domain
independent than features extracted by models without the use of emotions, due to
improved performance observed in cross-domain fake news detection.
• We evaluate both Ekman’s and Plutchik’s emotion classes for fake news and
rumour detection, showing that Ekman’s emotion classes lead to similar or better
performance than Plutchik’s emotion classes. To the best of our knowledge, no
previous work on fake news detection considers Ekman’s emotion classes.
• We evaluate the performance of the single-task and multi-task models in in-domain
and cross-domain settings for different deep learning architectures and datasets.
Based on the results obtained, we verify that emotionguided implicit features
extracted by our multi-task models aid in cross-domain fake news and rumour
detection.
Advantages
The proposed system incorporates emotion-guided features in our fake news
detection model by implementing a multi-task framework incorporating a multi-class
emotion classification task in addition to the original fake news or rumour detection
task. We further hypothesize that there is a prominent correlation between the
legitimacy of a text and its intrinsic Ekman’s and Plutchik’s emotion, and within the
same emotion, real and fake texts are sometimes represented differently. We further
suggest that this correlation can positively affect a classifier’s capability in the
detection of fake news or rumours.

SYSTEM REQUIREMENTS

➢ H/W System Configuration:-

➢ Processor - Pentium –IV

➢ RAM - 4 GB (min)
➢ Hard Disk - 20 GB
➢ Key Board - Standard Windows Keyboard
➢ Mouse - Two or Three Button Mouse
➢ Monitor - SVGA

Software Requirements:
 Operating System - Windows XP
 Coding Language - Java/J2EE(JSP,Servlet)
 Front End - J2EE
 Back End - MySQL