During my previous publications I presented to you language modeling and how machine learning breakthrough enables text generation even confusing to humans (and somehow dangerous). Today I wanted to share with you an interesting article released by Harvard NLP and MIT-IBM Lab. The authors promoted a tool to help people to detect if a text is fake (ie generated by a machine) or not (written by a human being). visit here
By highlighting the probability distribution of words used within the context of the sequence we can see how machine stays in a “comfort zone” when compared to human texts. Therefore preventing the fake text to be easily detected by a reader. Nice work and a step forward in preparing defenses against malicious use of machine learning. Linguistic is awesome and doing research in NLP at Disaitek gives us the chance to learn much about that every day. If you have any interest in understanding how their demonstration works or how language generation is done, we would be glad to talk to you about it. 😉