An abusive text detection system based on enhanced abusive and non-abusive word lists

摘要

Abusive text (indiscriminate slang, abusive language, and profanity) on the Internet is not just a message but rather a tool for very serious and brutal cyber violence. It has become an important problem to devise a method for detecting and preventing abusive text online. However, the intentional obfuscation of words and phrases makes this task very difficult and challenging. We design a decision system that successfully detects (obfuscated) abusive text using an unsupervised learning of abusive words based on word2vec's skip-gram and the cosine similarity. The system also deploys several efficient gadgets for filtering abusive text such as blacklists, n-grams, edit-distance metrics, mixed languages, abbreviations, punctuation, and words with special characters to detect the intentional obfuscation of abusive words. We integrate both an unsupervised learning method and efficient gadgets into a single system that enhances abusive and non-abusive word lists. The integrated decision system based on the enhanced word lists shows a precision of 94.08%, a recall of 80.79%, and an f-score of 86.93% in malicious word detection for news article comments, a precision of 89.97%, a recall of 80.55%, and an f-score 85.00% for online community comments, and a precision of 90.65%, a recall of 93.57%, and an f-score 92.09% for Twitter tweets. We expect that our approach can help to improve the current abusive word detection system, which is crucial for several web-based services including social networking services and online games.