Abstract:

As user-generated Web content increases, the amount of inappropriate and/or objectionable content also grows. Several scholarly communities are addressing how to detect and manage such content: research in computer vision focuses on detection of inappropriate images, natural language processing technology has advanced to recognize insults. However, profanity detection systems remain flawed. Current list-based profanity detection systems have two limitations. First, they are easy to circumvent and easily become stale–that is, they cannot adapt to misspellings, abbreviations, and the fast pace of profane slang evolution. Secondly, they offer a one-size fits all solution; they typically do not accommodate domain, community and context specific needs. However, social settings have their own normative behaviors–what is deemed acceptable in one community may not be in another. In this paper, through analysis of comments from a social news site, we provide evidence that current systems are performing poorly and evaluate the cases on which they fail. We then address community differences regarding creation/tolerance of profanity and suggest a shift to more contextually nuanced profanity detection systems.