Μεταπτυχιακές Εργασίες

Μόνιμο URI για αυτήν τη συλλογήhttps://pyxida.aueb.gr/handle/123456789/7

Περιήγηση

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω

Τώρα δείχνει 1 - 1 από 1

Toxicity detection on Greek tweets
Anagnostopoulos, Theodoros; Athens University of Economics and Business, Department of Informatics; Pavlopoulos, Ioannis; Malakasiotis, Prodromos; Vassalos, Vasilios
In this thesis, we experimented with detecting verbally offensive (toxic) greek tweets withgreek & multilingual transformer models. We firstly compared said BERT models on apublicly available greek dataset on toxic tweet detection. Having been trained on that publicdataset, we tested their ability to generalize their knowledge on our newly annotated datasetconstructed, in collaboration with PALOservices. Furthermore, we applied Semi-Supervisedlearning techniques on 10,000 unlabeled data to augment the existing training data, andfurther increase performance on our dataset. We also further pretrained a BERT model on300,000 unlabeled tweets to examine its capacity to learn new tokens and detect toxicity in amore reliable way. Lastly, we experimented with sentiment and emotion in tweets, not only toinvestigate possible correlations with toxicity but also, once again, to maximize our model’spredictive ability on offensive language detection.