Offensive Language Detection Using Brown Clustering

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5079–5087 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5079 Offensive Language Detection Using Brown Clustering Zuoyu Tian, Sandra K¨ubler Indiana University zuoytian@indiana.edu, skuebler@indiana.edu Abstract In this study, we investigate the use of Brown clustering for offensive language detection. Brown clustering has been shown to be of little use when the task involves distinguishing word polarity in sentiment analysis tasks. In contrast to previous work, we train Brown clusters separately on positive and negative sentiment data, but then combine the information into a single complex feature per word. This way of representing words results in stable improvements in offensive language detection, when used as the only features or in combination with words or character n-grams. Brown clusters add important information, even when combined with words or character n-grams or with standard word embeddings in a convolutional neural network. However, we also found different trends between the two offensive language data sets we used. Keywords: Brown clustering, offensive language detection, machine learning 1. Introduction Over the last two decades, the increase in the use of social media has led to an increase in offensive language. Social media often allow anonymous access, and users increas- ingly use this anonymity to publicize aggressive or offen- sive attitudes. Because of the vast volume of data produced on social media platforms every day, we need automated methods that can detect hate speech without restricting peo- ple’s right to freedom of expression. Research on offensive language detection has mostly con- centrated on using machine learning algorithms with a set of shallow lexical features such as word or character n- grams (e.g., (Warner and Hirschberg, 2012; Malmasi and Zampieri, 2018)). One of the challenges that these ap- proaches face is data sparsity since such shallow features do not generalize well: We may not have seen the exact word that characterizes a tweet as hate speech, but we may have seen similar or related words in training. This problem has been approached by using word embeddings. However, most of these approaches use pre-trained models, trained on large sets of English language data (Peters et al., 2018; Devlin et al., 2018; Mikolov et al., 2018). Such models cur- rently provide state of the art models (Badjatiya et al., 2017; Liu et al., 2019; Zhu et al., 2019; Mishra et al., 2019). Since these embeddings were not created from data sets specifi- cally targeting the detection of hate speech, it is not clear if they actually generalize over hate speech related words, or if they generalize over specific domains in which hate speech is common. The Waseem data set (Waseem and Hovy, 2016), for example, was sampled based on specific topics that had created large amounts of hate speech, such as football or an Australian cooking show. Since the ap- proaches using word embeddings often require more data and compute power than most people have access to, it is impossible or at …

Content truncated to 500 words