Challenge
The internet has connected people through platforms that enable information exchange. To prevent hate and offensive speech, most platforms have community guidelines, moderation teams, and monitoring techniques. TikTok, for example, classifies content as restrictible (inappropriate for youth) or removable (violating policies). Our goal is to create an AI-powered solution that effectively classifies content as normal, offensive, or hate speech. A key challenge is accurately identifying and categorizing text per TikTok's hate speech policy. The complexity of natural language, including sarcasm, slang, and cultural references, complicates this task.
Our Solution:
Data Processing
Our team began by taking multiple initial datasets of comments with three varying classification criteria for comments of similar nature. The datasets were combined where data preparation and validation were performed to standardize all data entries into one form, removing any junk symbols or unsuitable entries to leave a homogenous list of text comments.
The sizes of light and heavy datasets
Each entry was labelled in a systematic manner, using numerical values to represent each category, where the number 0, 1 and 2 represented normal, offensive and hate speech respectively.
Here you can see the label distribution for out datasets:
The research component consisted of using classical machine learning algorithms, for which we employed a light dataset model consisting of 20 thousand entries. The production component centred around a deep learning algorithm using a heavy dataset model of over 2 million entries. Datasets are available here.
Classical Machine Learning
To ascertain the best type of machine algorithm to utilise in production, a light dataset was used to train various learning models such as multilayer perceptron neural network, gradient booster classifier and support vector machine. The research was focused mainly on assessing the accuracy and f1 score of the models, however all models yielded unsatisfactory results with scores averaging around 50% across all models. It hence became evident that a more sophisticated algorithm, such as deep learning, was necessitated to attain the desired quality of results.
picture taken from here
picture taken from here
picture taken from here
Deep Learning
To implement the deep learning algorithm, a BERT model was selected because of it being a bidirectional model, which tends to capture more of the text’s context and meaning.
Initial tests on light datasets showed results of 65-70%, so we concluded that a heavy dataset would be capable of producing results up to 90%, however our personal machines lacked sufficient computing power to support the training of such a model on a large number of dataset entries such as ours. Despite this, we are confident that training with a larger dataset, would allow the algorithm to more accurately differentiate between various classifications of comments, and recognise linguistic elements such as sarcasm, slang, and cultural references.
Here you can see the differences for accuracies between Classiacal and Deep Learning approaches:
Conclusion
In the process of our research, we have successfully created a BERT model algorithm, with demonstrated capabilities to achieve an accuracy up to 70% in language and text classification. We have also developed a prototype of a website, that ideally, upon full functionality would be able to load videos from Tik-Tok and classify them in a similar manner without loss of accuracy.