In this paper we aim to develop a state-of-the-art method
for detecting abusive language in user comments, while also
addressing the above deficiencies in the field. Specifically,
this paper has the following contributions:
• We develop a supervised classification methodology
with NLP features to outperform a deep learning approach.
We use and adapt several of the features used
in prior art in an effort to see how they perform on the
same data set. We also extend this feature set with features
derived from distributional semantics techniques.
• We make public a new data set of several thousand user
comments collected from different domains. This set
includes three judgments per comment and for comments
which are labeled as abusive, a more fine-grained
classification on how each is abusive.
• Prior work has evaluated on a fixed, static data set.
However, given the issues with language changing over
time and also with users trying to cleverly evade keywordbased
approaches, we perform several analyses of how
models trained on different types and sizes of data perform
over the span of one year, across two different
domains. To our knowledge, this is the first longitudinal
study of a computational approach to abusive
language detection.