![Institutional Repository [SANDBOX]](images/dias.logo.png "Institutional Repository [SANDBOX]"){kind=logo}
Institutional Repository [SANDBOX]
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/CD32B0A5-46BF-4DE9-8855-CBD48F8B53CC | ||
| Year | 2024 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Panagiota Koutra, "Social media analysis targeting Troll detection", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024 https://doi.org/10.26233/heallink.tuc.101219 | ||
| Appears in Collections |
Social networks have transformed the way people interact, share information, andexpress opinions; however, they have introduced significant challenges, notablythe rise of trolling behavior. Trolling involves posting inflammatory or disruptivemessages that degrade online discourse, spread misinformation, and foster a toxicenvironment. This thesis addresses the challenge of troll detection using artificialintelligence (AI) and natural language processing (NLP) techniques.Troll detection is approached through text classification, which distinguishes between troll and non-troll content using various machine learning algorithms. Thisstudy compares Logistic Regression, Naive Bayes, SVM, and the pre-trained DeBERTa model. The results show that DeBERTa outperforms the other methods interms of accuracy, precision, F1-score, and recall.Furthermore, the thesis examines the generalization capabilities of the DeBERTamodel by testing it on unseen data from different sources through zero-shot andone-shot predictions. The model shows satisfactory performance in zero-shot predictions and good results in one-shot predictions.The final phase focuses on developing a general model that accurately predictstrolling behavior across different datasets. Scale-invariant fine-tuning (SiFT) wasapplied to improve the model’s performance and showed significant improvement.
Copyright © DIAS 2013
