Classification of offensive social media video texts using turkish natural language processing

In the Turkish social media ecosystem, offensive expressions such as insults, profanity, sexism, and racism are becoming increasingly visible, creating a security challenge that is difficult to monitor in real time, especially on video-based platforms. This thesis addresses this gap by constructing a medium-scale text dataset of 5,725 samples, categorized into five classes (PROFANITY, INSULT, SEXIST, RACIST, OTHER), compiled from YouTube content. The class distribution of the dataset is as follows: 4,810 samples (84.0%) non-toxic content (OTHER), 631 samples (11.0%) insults/inciting expressions (INSULT), 104 samples (1.8%) profanity (PROFANITY), 100 samples (1.7%) racist expressions (RACIST), and 80 samples (1.4%) sexist expressions (SEXIST). This distribution shows that toxic content mainly appears as minority classes in social media, highlighting the critical importance of the class imbalance problem. In the modeling phase, ≈110M parameter BERTurk (uncased-base, 128k) and the 17-billion-parameter LLaMA-4 Maverick were compared in terms of accuracy, macro-F1, and sub-class recall metrics. The results demonstrated that LLaMA-4 achieved an overall accuracy of 88.7% and a weighted F1 score of 0.896, providing a ≈11-point improvement compared to BERTurk. Moreover, the +0.267 increase in macro-F1 reflects a balancing effect, especially in the minority classes PROFANITY, RACIST, and SEXIST. Furthermore, latency and cost analyses revealed that BERTurk, with a p50 latency of ≈100 ms and low operational cost, is efficient for high-volume pre-screening tasks. On the other hand, LLaMA-4, despite its ≈350 ms p50 latency, can be used as a secondary verification layer in risky content detection scenarios, achieving a balanced optimization of cost and performance. The study thus proposes a practically applicable two-layer moderation architecture, turning academic findings directly into sectoral solutions. Supported by a comprehensive bibliometric analysis of online hate speech literature between 2018 and 2025, this technical research provides a reliable dataset and operational modeling framework for Turkish offensive language detection, while also offering a guiding atlas that summarizes the structural dynamics of the field for researchers and policymakers.