Hate speech becomes dangerous for social media users. Detection of hate speech becomes more complex with the existing models. This paper focused on detecting hate speech from social media content using the Deep Ensemble Algorithm (DEA). The proposed DEA is integrated with updated Hierarchical Attention Networks (HANs) that detect hate speech effectively based on the external hate keywords list. The hate keywords list helps the algorithm find the appropriate words in the given input and improves the detection of hate speech rate. The pre-trained model DistilBERT (Distilled version of BERT) is used for training on hate speech data. Preprocessing techniques, such as removing punctuation, Stopping Words, stemming, and minimizing, help to improve data quality by eliminating noise from the input dataset. Feature extraction techniques such as FastText and N-grams extract significant data from the input hate speech data—finally, the two hate speech datasets are used for experimental analysis. The proposed updated HAN analyzed hate speech based on grammar, emotion, and scenario. The experimental study shows the comparative performance in terms of hate detection rate.