Electronic Thesis/Dissertation
 

Classifying Categories of Speech: A Machine Learning Approach to Identifying Cyberbullying, Hate Speech, Offensive and Normal Language

Open Access Deposited

Abstract of PraxisThe proliferation of online communication has led to increased occurrences of harmful digital behaviors that include hate speech, offensive language, and cyberbullying. Current machine learning models fail to distinguish between these harmful speech types accurately, resulting in high rates of missed cyberbullying incidents. This research will demonstrate that a classification model can be developed to distinguish between these types of speech having an accuracy rating of 90% or higher. This research also aims to compare three types of classification models (a deep learning model CNN, a BERT transformer, and a stacked model consisting of SVM+ BERT) to find which model is capable of accurately differentiating between four types of speech: cyberbullying, hate speech, offensive, and normal speech. All three models were pre-trained for multiclassification of text. Word embedding approaches are pivotal in multiclassification of text. The use of a BERT model simplified the word embedding approach as it is built into the BERT architecture. However, the CNN model necessitated a different approach to word embeddings, specifically utilizing GloVe to capture semantic relationships. GloVe generates static embedding, meaning each word is represented by a single fixed vector. The vector process captures the meanings of words based on their patterns of co-occurrence with other words. The dataset used in this study (Dataset 3- Four Classification) is a compilation of tweets from Twitter (now X) taken by combining two datasets. Dataset 1 Cyberbullying Comments consisting of prelabeled rows of tweets that were either identified with types of cyberbullying or no cyberbullying. Dataset 2: Hate Speech and Offensive Language also consisted of tweets and was prelabeled with rows of tweets that were identified as either hate speech, offensive language or neither hate speech or offensive language. Dataset 3- Four Classification now consists of tweets that are prelabeled with each tweet identified as one of the four categories of speech: cyberbullying, hate speech, offensive and normal language. This study leverages the advanced natural language processing techniques with the machine learning algorithms of each model to be able to: a) determine the best model that will be able to refine the detection, b) prove machine learning models are able to categorize these forms of speech properly, and c) lead to improved accurate and reliable detection of cyberbullying.

Author Language Date created Type of Work License
  • All rights reserved
Rights statement GW Unit Degree Advisor Committee Member(s) Persistent URL

Notice to Authors

If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.

Thumbnail Title Date Uploaded Visibility Actions
Preview of Mauser_gwu_0075A_16990.pdf Mauser_gwu_0075A_16990.pdf 2025-04-09 Open Access