Many people are using Twitter for thought expression and information sharing in real-time. Twitter is one of the trendiest social media applications that cybercriminals also widely use to harass the victim in the form of cyberstalking. Cyberstalkers target the victim through sexism, racism, offensive language, hate language, trolling, and fake accounts on Twitter. This paper proposed a framework for automatic cyberstalking detection on Twitter in real-time using the hybrid approach. Initially, experimental works were performed on recent unlabeled tweets collected through Twitter API using three different methods: lexicon-based, machine learning, and hybrid approach. The TF-IDF feature extraction method was used with all the applied methods to obtain the feature vectors from the tweets. The lexicon-based process produced maximum accuracy of 91.1%, and the machine learning approach achieved maximum accuracy of 92.4%. In comparison, the hybrid approach achieved the highest accuracy of 95.8% for classifying unlabeled tweets fetched through Twitter API. The machine learning approach performed better than the lexicon-based, while the performance of the proposed hybrid approach was outstanding. The hybrid method with a different approach was again applied to classify and label the live tweets collected by Twitter Streaming in real-time. Once again, the hybrid approach provided the outstanding result as expected, with an accuracy of 94.2%, recall of 94.1%, the precision of 94.6%, f-score of 94.1%, and the best AUC of 98%. The performance of machine learning classifiers was measured in each dataset labeled by all three methods. Experimental results in this study show that the proposed hybrid approach performed better than other implemented approaches in both recent and live tweets classification. The performance of SVM was better than other machine learning algorithms with all applied approaches.