<![CDATA[The growing popularity of social media has made it simpler for videos with strong or offensive language to spread quickly, which has led to a need for automatic tools to find and group this type of content. This study uses deep learning to identify videos with harsh language by looking at their visual, audio, and text parts using a detailed method that covers all these aspects. The model uses EfficientNetB0 to get visual features, a CNN to handle audio features from spectrogram images, and an Embedding layer along with a BiGRU to deal with text information. These three types of features are combined using an Adaptive Multimodal Fusion technique and then sorted into categories with the help of a Deep Neural Network (DNN). The final dataset includes 1,473 videos, with 738 videos marked as Harsh and 735 videos marked as Safe. The model is trained using the Adam optimizer and Binary Crossentropy loss function, and it also uses label smoothing. For evaluation purposes, 295 videos were used, which included 148 Harsh videos and 147 Safe videos. The model's accuracy was 83%, and it had average precision, recall, and F1-score values of 0.83 each. The model also achieved an ROC-AUC score of 0.8686, showing it can effectively tell the difference between the two types of videos. These results show that the new deep learning method that uses multiple types of data can work well in identifying harsh language in videos by combining visual, audio, and text information, and it can help with automatically checking and managing video content. Keywords: Deep Learning, Video Classification, EfficientNetB0, Adaptive multimodal fusion, Harsh Language]]>
