The selection of optimal feature representations is a critical step in the use of machine learning in text classification. Traditional features (e.g. bag of words and n-grams) have dominated for decades, but in the past five years, the use of learned distributed representations has become increasingly common. In this paper, we summarise and present a categorisation of the stateof-the-art distributed representation techniques, including word and sentence embedding models. We carry out an empirical analysis of the performance of the various feature representations using the scenario of detecting abusive comments. We compare classification accuracies across a range of off-the-shelf embedding models using 10 labelled datasets gathered from different social media platforms. Our results show that multi-task sentence embedding models perform best with consistently highest classification results in comparison to other embedding models. We hope our work can be a guideline for practitioners in selecting appropriate features in text classification task, particularly in the domain of abuse detection.