Online Social media is exploiting with large number of abusive comment on proliferation of the user-generated contents. Social media platform has been empowered user to create, share, exchange content with each for socializing but it leads to various challenges to network user on propagation of abusive content from cyber bullies, and haters to their post as it spread to various audience. However many conventional approaches using machine learning model has been proposed to handle these challenges but it provides wrong misinterpretation. In order to mitigate those challenges, a new deep learning architecture entitled as Autoencoder is designed to detect and encrypt the unwanted content in social media. In this architecture, dataset is collected from twitter. It contains comments for various event and news and it is transformed as CSV file. Transformed file is preprocessed using data normalization, stop word removal and stemming process for removing number and emojis. Preprocessed data is projected to feature extraction model, which process the data on computation of the feature from part of speech tagging. Computed feature vector is employed to deep learning model along BERT based sentiment analysis architecture to identify the polarity of the term. Positive polarity of the vector is classified as normal comment and negative polarity of the feature vector is classified as unwanted comment using deep learning model. Classified content is encodedusing encoderin deep learning model represent the abusive content in the encoded form to other audience. Experimental analysis of the proposed approach is compared with conventional approaches to evaluate the performance. Further performance of the approaches is computed using performance metric such as precision, recall and f measure. Proposed model produces 98% efficiency is identifying the abusive comment compared state of art approaches.

Unwanted Review Classification, Autoencoder, Deep learning, BERT, Principle Component Analysis, Sentiment analysis