Hate Speech Detection on Social Networks Using Machine Learning

U jeku razvoja Interneta podstrek dobivaju i nepoželjeni oblici komunikacije, izmedu ostaloga govor mržnje. Automatizaciji njegova otkrivanja pristupili su brojni znanstvenici s područja obrade prirodnog jezika i strojnog učenja. U obzir uzimaju mnogobrojne značajke i primjenjuju različite modele s uglavnom solidnom, no ne i idealnom ucinkovitošću. Ovaj rad eksperimentalno pokazuje da model treniran na jednom skupu podataka teško daje dobre rezultate na ulazu iz drugog, nesrodnog skupa podataka. Bolji rezultati ostvarivi su uz prilagodbu okružju treniranjem na prikladnim podatcima.The development of Internet stimulates, among other things, undesirable forms of communication, including hate speech. The automation of its detection has been attempted by numerous scientists from the field of natural language processing and machine learning. They have tried extracting many features and applying different models with mostly , but not ideal efficiency. This thesis experimentally shows that models trained on one dataset often fail when utilized on other, unrelated datasets. Better results can be achieved by adaptation through training on appropriate data