Harnessing the Power of Text Mining for the Detection of Abusive Content in Social Media

Abstract

The issues of cyberbullying and online harassment have gained considerable coverage in the last number of years. Social media providers need to be able to detect abusive content both accurately and efficiently in order to protect their users. Our aim is to investigate the application of core text mining techniques for the automatic detection of abusive content across a range of social media sources include blogs, forums, media-sharing, Q&A and chat -using datasets from Twitter, YouTube, MySpace, Kongregate, Formspring and Slashdot. Using supervised machine learning, we compare alternative text representations and dimension reduction approaches, including feature selection and feature enhancement, demonstrating the impact of these techniques on detection accuracies. In addition, we investigate the need for sampling on imbalanced datasets. Our conclusions are: (1) Dataset balancing boosts accuracies significantly for social media abusive content detection; (2) Feature reduction, important for large feature sets that are typical of social media datasets, improves efficiency whilst maintaining detection accuracies; (3) The use of generic structural features common across all our datasets proved to be of limited use in the automatic detection of abusive content. Our findings can support practitioners in selecting appropriate text mining strategies in this area.

Presenting a labelled dataset for real-time detection of abusive user posts

“…In addition, our dataset is unbalanced, with the 'abusive' comments and 'undecided' comments occurring far less than 'nonabusive' comments (see Table 2). Therefore, we applied resampling [3] to randomly oversample the minority instances before training the model. To minimise the impact of conducting random oversampling, we resampled the training data 3 times, and used the average performance across the 3 result sets.…”

“…Given that comments in the dataset are typically conversational in style and short, we did not apply stemming or stop word removal. Following results from our previous work [3], Document Frequency (DF) feature reduction was used to cut down on high or low frequency occurrence terms without jeopardising model performance. Term Frequency (TF) was used to normalise the feature values.…”

“…We applied normalisation technique (MinMax [16]) to transform these values to the same range level. Once a nal feature representation for each comment has been obtained, we used Singular Value Decomposition (SVD) [3] to reduce the feature space before feeding them into the supervised machine learning algorithm.…”

“…We required labellers to categorise our data into three classes (abusive, non-abusive, and undecided) and then to identify the severity of abusive posts on a scale. Next, a set of text classi cation techniques based on our previous work [3] was used to generate a benchmark for this dataset on detection accuracy. The techniques include feature representation, feature reduction and dataset balancing.…”

Supervised machine learning for classification mining and ranking of illegal web contents = Aprendizaje automático supervisado para la clasificación, extracción de conocimiento y ordenación de contenidos web ilegales

“…Nowadays, a considerable number of research lines and applications require accurate text classification. For example, spam emails detection (Mohamad and Selamat, 2015;Wang et al, 2011), abusive text detection (Chen et al, 2017), sentiment analysis (Pang et al, 2008), detecting suspicious activities in the Darknet network , and classifying illegal content in the Online Notepad Services (ONS) . However, several factors collaborate to make this task challenging, such as the possibility of access to a sufficient number of labeled examples (Ohana et al, 2012;, and the quality and the length of the addressed text (Li et al, 2016b).…”