How to Find Social Robots exactly?

With the rapid development of artificial intelligence and natural language processing, there are more and more social robots applied in the social networks such as Twitter, intended to lead public opinion or crawling private information illegally. The problem of detection social robots, which is automated social accounts governed by artificial intelligence software, pretend to be a human user. There are some technologies proposed to detect the social robots automatically applied to the real social network for verification. Hence, conventional social robot detecting technologies proposed before are applied to detect by the account's metadata or account posted tweet content respectively. With the help of pre-trained language model such as BERT, this paper propose a deep neural network model based on contextual long short-term memory (LSTM) architecture named DeepBot, which exploits tweet content and account's metadata features. The architecture of DeepBot contains three phases: (1) it uses the pretrained model such as BERT to extract the embedding vector from the tweet content of the specific account, and (2) it choose more discriminative account metadata to extract a metadata vector, and then (3) it combines the auxiliary embedding vector and metadata vector into decoder layer to train a detecting model. What's more, in this paper, we review the labelling social robots datasets proposed in public, and get a mixture datasets of labelling social datasets to verify and compare the experimental results of our proposed DeepBot and other conventional methods. We also present empirical results of DeepBot and our ongoing experimentation with it, as we have gained experience applying it to the mixture labeling social robot dataset, including over 10000 accounts. The experimental results show that DeepBot outperforms previous state-of-the-art methods, with leveraging a small and interpretable set of features.