Linguistic Methods for the Detection of Implicit Abuse

Wider research context: Recent years have seen a massive rise in abusive content on the web. Automatic classification methods are sought to assist moderators in finding such content. Since much abusive content is expressed in the form of written comments, natural language processing is a key technology in tackling this issue. The effectiveness of state-of-the-art methods for abusive language detection is limited. While explicit abuse, that is, abuse conveyed by unambiguously abusive words (e.g. the f-word or n-word), can now be fairly reliably detected, we currently have no indication that classifiers can also detect implicit forms of abuse. A recent survey on implicit abuse has suggested that either existing datasets do not contain sufficient instances of implicit abuse, or the datasets have been sampled in such a biased way that using them to effectively learn to classify a particular subtype of implicit abuse is not possible. In this project, we want to address the classification of a set of subtypes of implicit abuse to fill this important gap in current research.
Research questions: In this project, we want to address the three following research questions: How can we create datasets that reflect particular subtypes of implicitly abusive language adequately? What types of linguistic phenomena are involved in particular subtypes of implicit abuse and how can they be effectively used to build a classifier? How can we build classifiers that do not overfit to artefacts of a particular dataset?
Approach: Much of the previous work in abusive language detection follows a one-size-fits-all approach. However, significant shortcomings have been previously identified in the resulting datasets and evaluations involving the detection of implicitly abusive language. In this project, we will pursue a divide-and-conquer approach. We will address individual subtypes of implicit abuse independently. By carefully extracting data for each particular subtype and taking into consideration issues like acquiring expressive negative data, we hope to come up with less biased datasets and thus to be able to build more effective classifiers. Given the nature of implicit abuse, significant linguistic feature engineering is expected to be necessary.
Level of originality: There has been hardly any research dedicated to the task of detecting implicitly abusive language. Unlike much of prior work, we may depart from generic classification approaches, such as deep learning, and also consider more linguistically informed classification approaches. The output of this project will be new datasets dedicated to implicitly abusive language, which currently are still missing, new classification approaches and a better understanding of how individual subtypes of implicit abuse can be characterized from a linguistic point of view.
Primary researchers involved: Univ.-Prof. Dr. Michael Wiegand