The Linguistic Ideologies of Deep Abusive Language Classification

Abstract

This paper brings together theories from sociolinguistics and linguistic anthropology to critically evaluate the so-called "language ideologies"-the set of beliefs and ways of speaking about language-in the practices of abusive language classification in modern machine learning-based NLP. This argument is made at both a conceptual and empirical level, as we review approaches to abusive language from different fields, and use two neural network methods to analyze three datasets developed for abusive language classification tasks (drawn from Wikipedia, Facebook, and Stack-Overflow). By evaluating and comparing these results, we argue for the importance of incorporating theories of pragmatics and metapragmatics into both the design of classification tasks as well as in ML architectures.

Developing an online hate classifier for multiple social media platforms

“…Furthermore, standards (i.e., what is interpreted as hate) differ by the online community, so that an expression that is hateful in one community may be considered neutral, humor, or typical discourse in another community [55]. For example, in online Q&A platforms like StackOverflow, a concise form of expression can appear rude to outsiders but be perfectly acceptable given the community standards [1].…”

“…The output feature vector was then fed into a Gated Recurrent Unit (GRU) layer followed by global max pooling and a softmax layer. The previous studies indicate that CNNs' potential to capture the local patterns of features benefits online hate detection [1]. Finally, while most previous work relies on text features, there are also studies using other features.…”

“…In addition to the selected datasets, we considered the following datasets from prior literature: Waseem [9], Fortuna [74], Ross et al [70], Salminen et al [48], and the StackOverflow dataset [1]. The Waseem dataset, containing 6910 tweets labeled for racism, sexism, or neither, was omitted because our manual inspection revealed a high rate of false positives.…”

“…Finally, the StackOverflow dataset, released in conjunction with the 2nd Workshop on Abusive Language Online, is made accessible upon accepted application from the company. This dataset was omitted because previous research has shown it to be inconsistent [1]. This is because the dataset is based on a type of self-selection rather than an objective coding procedure and is thus subjected to social dynamics taking place in the social media platform in question-for example, some users may flag a reply as offensive to retaliate, while others do not bother to flag an offensive response at all.…”

“…This is because the dataset is based on a type of self-selection rather than an objective coding procedure and is thus subjected to social dynamics taking place in the social media platform in question-for example, some users may flag a reply as offensive to retaliate, while others do not bother to flag an offensive response at all. When annotated by independent coders, Castelle [1] found a high disagreement between the original flagging and the ratings of the independent coders. Finally, the dataset by Salminen et al [48] was omitted because the authors used a keyword-based detection of hateful comments.…”