I am working on an NLP project about classifying offensive text data in social media. By offensive I especially mean threat words that one say to another.
"Stop doing this or else you will pay for it."
"Just wait until you see what's coming"
"I will break your legs next time I see you."
As an initial approach, I considered semantic and syntactic keyword matching. However, doing this on this very problem seemed harder because threating is an action and it is expressed in so many different ways.
My main goal is classifying text data by offensive/non-offensive text by using Machine Learning and Deep Learning algorithms. After weeks of online searching, I could not find a ready-to-use dataset. I considered manually labelling the data. However, I don't know where I should start.
What would be the best approach for making progress in this task? I also plan to do this in both English and German languages.
Also, below is a related article for fully understanding of the problem:
Deep learning for detecting inappropriate content in text