Abstract:
In the last two decades, the penetration of Social Networking Sites and Social
Media (SNS/SM) platforms has risen to include more than one-third of the global
population. However, the use of SNS/SM has produced both positive and negative
results so much so that there have been calls to researchers pay immediate and far
greater attention to these contradictory effects of SNS/SM capabilities. A key negative
impact is the fast and significant rise of cyberbullying.
Cyberbullying has emerged as a serious act on social networking sites/social
media (SNS/SM) platforms in today’s digital society. Statistics underscore this whereby
42% of individuals indicate they have experienced cyberbullying, more specifically
38% of females and 54% of males have also experienced some form of bullying. This
form of dysfunctional social act is expressed via aggression, harassment, and toxic
behavior poses severe consequences to increasing penetration of SNS/SM.
Simultaneously, there has been a great awareness of the need for moderating
contents on (SNS/SM) to detect and reduce cyberbullying contents. It is also recognized
that human content moderation on SNS/SM is impractical and too costly. Therefore,
there is an increasing need for accurate methods of content moderation that are less
reliant on human judgement and instead employ sophisticated machine learning
methods. A key shortcoming of the current machine learning approaches to
cyberbullying detection is that their accuracy needs to be improved significantly to be
relied upon for practical deployment of non-human content moderation on SNS/SM.
Problem of using advanced data analytics and AI-based techniques in
cyberbullying detection in the service of reducing the latter has been extensively
studied. However, existing research method have faced challenges with the issue of
false positives whereby many identified instances may not be cyberbullying. For
instance, “I hate you” and “I hate thinking about the future” both contain the word
“hate”, yet the second sentence is a non-cyberbully phrase/sentence that contains the
word hate as a metaphor for personal discomfort with the uncertainty associated with
one’s future.
Indeed, more accurate detection can come from a deeper contextually sensitive
detection methods going beyond identifying simple hate words, whereby distinguishing
3
between harmful and innocuous expressions becomes the focal research problem. To
address this, we will conduct a complex phrasal analysis to identify cyberbullying
through the detection of the underlying emotion behind the seemingly toxic phrases.
This approach aims to mitigate false positive predictions, which can occur when relying
on hate keywords or an online hate corpus.
To achieve our goal, we employ an emotion recognition transfer learning model
to comprehend the underlying emotion trigger of cyberbullying and enhance its
detection. To accomplish this, we use a Bi-LSTM pre-trained emotion detection model
with a 92% accuracy on the training set. Then, we adapt the knowledge gained from the
first model to improve learning on a different but related task which is improving
cyberbullying detection by integrating elements of context based on emotion
identification within the expressed phrase. Based on extensive testing and training of
the model on the data, we propose that LSTM and CNN multi-label-based classification
model which is exhibiting an 80% accuracy on the training as superior approach to
cyberbullying detection.
To evaluate and validate our results, we randomly split our dataset into training
and validation sets, testing against an unseen dataset. A comparative analysis with
(Gencoglu, 2021) model reveals a significant improvement in performance using
McNemar’s test and the Paired T-test. Notably, our model shows an 18% improvement
on (Gencoglu, 2021) model.
In summary, this research has addressed the limitations inherent in keyword-based cyberbullying detection methodologies, thereby making a meaningful
contribution to the field of cyberbullying detection in SNS/SM research.