Abstract
The massive amount of data residing in social networks, becomes a fertile source of much relevant knowledge. Violent and criminal language is a very important knowledge that can be extracted from tweets. It is very interesting to predict suspicious threat language that can threaten the privacy or the integrity of a person or a community. Threats posted via social networks has possible to cause suffering and harm on an individual and society. Systematic studies of threat from a computer science perspective, is still recent. This paper will present a related works on automatic threat detection, including algorithms, methods, and text analysis features used. Additionally, we introduce a threat dataset consisting of 2440 tweet messages in English. Each tweet is manually annotated as either being a Threat or Non-Threat. The threat dataset is useful for training algorithms such as machine learning and deep learning, and for studying the nature of using threat words in society. This paper also discusses challenges of automatic threat detection. The development of shared resources, such as annotated datasets, algorithms and open-source code and platforms is a very important step to advance the automatic threat detection.