Abstract
In the modern era of technological development, the emergence of Web 2.0 applications, related to social media, the dissemination of opinions, feelings, and participation in discussions on various issues have become very easy, which have led to a boom in text mining and natural language processing research. YouTube is one of the most popular social sites for video sharing. This may contain different types of unwanted content such as violence, which is the cause of many social problems, especially among children like aggression and bullying at home, in school and in public places. The research work reports performance of two different sentiment lexicons when they were applied on video transcripts to detect violence in YouTube videos. The automation of process to detect violence in videos can be helpful for censor boards that can use the technology to restrict violent video for a certain age group or can fully block entire video regardless of age. The models were built using the existing sentiment lexicons. The dataset consists of 100 English video transcripts collected from the web and was annotated manually as violent and non-violent. Various experiments were performed on the dataset using English SentiWordNet (ESWN) and Vader Package with different text preprocessing settings. The Vader package outperformed the ESWN by providing 75% accuracy. ESWN results for all POS tagging with 66% accuracy were better than its result for adjectives POS tagging with 58% accuracy.