Neural Character-based Composition Models for Abuse Detection

By Pushkar Mishra, Helen Yannakoudakis, Ekaterina Shutova. 2018

63 citations recorded by the source corpus.

The advent of social media in recent years has fed into some highly undesirable phenomena such as proliferation of offensive language, hate speech, sexist remarks, etc. on the Internet. In light of this, there have been several efforts to automate the detection and moderation of such abusive content. However, deliberate obfuscation of words by users to evade detection poses a serious challenge to the effectiveness of these efforts. The current state of the art approaches to abusive language detection, based on recurrent neural networks, do not explicitly address this problem and resort to a generic OOV (out of vocabulary) embedding for unseen words. However, in using a single embedding for all unseen words we lose the ability to distinguish between obfuscated and non-obfuscated or rare words. In this paper, we address this problem by designing a model that can compose embeddings for unseen words. We experimentally demonstrate that our approach significantly advances the current state of the art in abuse detection on datasets from two different domains, namely Twitter and Wikipedia talk page.

Subfields: Artificial Intelligence, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

Citations in the corpus

Towards Safer Generative Language Models: A Survey on Safety Risks, Evaluations, and Improvements (2023, background): “…RNNs, and LSTMs (Georgakopoulos et al., 2018; van Aken et al., 2018; Gunasekara and Nejadgholi, 2018; Joulin et al., 2017; Kshirsagar et al., 2018; Mishra et al., 2018; Mitrovi´c et al., 2019; Sigurbergsson and Derczynski, 2020), contemporary approaches increasingly favor fine-tuning pre-trained…”
Detection of Abusive Language: the Problem of Biased Datasets (2019, background): “On this particular dataset, such a strategy leads to good results: Both Qian et al. (2018) and Mishra et al. (2018a) proposed classiﬁcation approaches that add author information to common text-level features.”
Joint Modelling of Emotion and Abusive Language Detection (2020, background): “Character-level features have also been shown to be beneficial in tackling the issue of Out-of-Vocabulary (OOV) words (Mishra et al., 2018b), since abusive comments tend to contain obfuscated words.”
HateBERT: Retraining BERT for Abusive Language Detection in English (2020, background): “The development of systems for the automatic identiﬁcation of abusive and offensive language have followed a common trend in NLP: feature-based linear classiﬁers (Waseem and Hovy, 2016; Ribeiro et al., 2018), neural network architectures (e.g., CNN or Bi-LSTM) (Kshirsagar et al., 2018; Mishra et al.”
Developing an online hate classifier for multiple social media platforms (2020, methodology): “[15] used Twitter and Wikipedia datasets for the detection of non-obvious hateful comments—as such, their approach did not consider general detection of all hateful comments.”
Tackling Online Abuse: A Survey of Automated Abuse Detection Methods (2019, methodology): “However, working with a similar setup, Mishra et al. (2018a) recently reported that GL o V e initialization provided superior performance; a mismatch is attributed to the fact that Badjatiya et al. tuned the embeddings on the entire dataset (including the test set), hence allowing for the…”
Exploring Deep Multimodal Fusion of Text and Photo for Hate Speech Classification (2019, methodology): “The automatic identiﬁcation of hate speech has been mostly formulated as a natural language processing problem (e.g. Mishra et al., 2018; Gu-nasekara and Nejadgholi, 2018; Kshirsagar et al., 2018; Magu and Luo, 2018; Sahlgren et al., 2018).”
A Gentle Introduction to Deep Learning for Graphs (2019, background): “thorship as arcs. User representations are of great interest in a variety of tasks, for example to detect whether an actor in the graph is the potential source of misinformation or unhealthy behavior [91, 94]. For these reasons, social networks are arguably the richest source of information for grap”