TU Berlin

Service-centric NetworkingSunder, S. (2020). Sentiment Polarization in Online Social Networks: The Flow of Hatespeech. Master Thesis, Technische Universität Berlin


zur Navigation

Es gibt keine deutsche Übersetzung dieser Webseite.

Master Thesis: Sentiment Polarization in Online Social Networks: The Flow of Hatespeech


Sentiment Polarization in Online Social Networks: The Flow of Hatespeech


The influence of sentiment polarization and exchange in online social networks has been grow-ing  and  studied  by  many  researchers  and  organizations  alike.   Detecting  communities  andtracking their evolution over time is the modern method of sentiment analysis combining mul-tiple domains of data mining, deep learning, graph theory and computational algorithms.In this thesis, we show sentiment polarization in online networks with an emphasis on hatespeech and discuss the methodologies adopted for text classification followed by communitydetection algorithms. The concept of transfer learning continues to be a novel approach and isembraced in this study to perform multiclass classification on text derived from social mediaplatform, Twitter.  The sentiments expressed in a text with respect to a topic in discussion isunderstood to influence a community when a Twitter user retweets the original text causing achain of reactions in a network.  The sentiments picked are closely related to concerns raisedfrom global organizations, legislative bodies and law enforcement, particularly the presence ofhate speech.

Traditional machine learning has encountered challenges for being unable to scale up to sliceand dice the complexities of language. Communication on social media contain obfuscationsand non-vernacular terms that require a robust algorithm and system to study and learn thepatterns.   We use a model built on long short term memory (LSTM), a variant of RNN thatcontains the ability to stretch back several iterations making learning process much more suitedfor languages semantics.   We use a pre-trained language model ULMFiT (AWD-LSTM) thatuses DropConnect for regularization and optimization and familiarize it with the construct ofsocial media text, followed by fine-tuning the classifiers to apply classifications.

Subsequently, we use the classified dataset to detect networks and hidden communities withinwhich users inadvertently become members of, through their tweet-retweet actions.  We jus-tify the use of Louvain algorithm since it follows computation of maximum modularity whichproves useful in detecting communities based on social media interactions. Lastly, we create atemporal visualization combining several static graphs for individual time instances to presentthe evolution of communities and flow of sentiments within. Future analysis might focus on the adoption of Graph Neural Network to track communitychanges and predict the next stage of community evolution using a Time Series model.

Supervisor: Katerina Katsarou

Type:  Master Thesis

Duration: 6 months



Schnellnavigation zur Seite über Nummerneingabe