Detecting Tweets with Offensive Content Using Document Embeddings in Supervised Learning Approaches

Abstract

            As more and more people gain access to the internet, there is an increase in user-created content (e.g. comments, blogs, posts & microblogs) on online forums and social media platforms. A challenge that comes with this increased volume in content is preventing the use of these platforms for abusive behavior. Due to the high rate at which content is being generated, manual moderation and intervention to detect offensive language and prevent abusive behavior are not scalable. Previous studies have shown that techniques from supervised learning can help in detecting offensive language at scale, nevertheless, several of these analyses have concentrated on the usage of lexical and/or syntactic characteristics, whereas the semantic interaction between words is largely overlooked. This study aims to understand the utility of Doc2Vec document embedding technique in offensive language identification tasks by building models to classify given tweets into offensive or non-offensive classes. In this study, we have used SemEval-2019 Task 6’s official dataset called OLID (Offensive Language Identification Dataset). Document embedding vectors of various lengths are created for every tweet in OLID using Word2vec and Doc2Vec. These vectors are then utilized as feature vectors in common classification algorithms like Logistic regression, SVM, Random Forest, etc. We have compared several metrics for each classification technique. The experimental results show significance of use of Doc2Vec embeddings in offensive language identification.