nlpUP at SemEval-2019 Task 6: A Deep Neural Language Model for Offensive Language Detection

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 722–726 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 722 nlpUP at SemEval-2019 Task 6: A Deep Neural Language Model for Offensive Language Detection Jelena Mitrovi´c, Bastian Birkeneder, Michael Granitzer Faculty of Computer Science and Mathematics University of Passau, Germany jelena.mitrovic@uni-passau.de | birkeneder@fim.uni-passau.de michael.granitzer@uni-passau.de Abstract This paper presents our submission for the SemEval shared task 6, sub-task A on the identification of offensive language. Our pro- posed model, C-BiGRU, combines a Convolu- tional Neural Network (CNN) with a bidirec- tional Recurrent Neural Network (RNN). We utilize word2vec to capture the semantic sim- ilarities between words. This composition al- lows us to extract long term dependencies in tweets and distinguish between offensive and non-offensive tweets. In addition, we evaluate our approach on a different dataset and show that our model is capable of detecting online aggressiveness in both English and German tweets. Our model achieved a macro F1-score of 79.40% on the SemEval dataset. 1 Introduction The ever-increasing amount of user-generated data introduces new challenges in terms of automatic content moderation, especially regarding hate speech and offensive language detection. User content mostly consists of microposts, where the context of a post can be missing or inferred only from current events. The challenge of automatic identification and detection of online aggressive- ness has therefore gained increasing popularity in the scientific community over the last years. Several recent workshops and conferences such as TRAC (Kumar et al., 2018), ALW2 (Fiˇser et al., 2018), and GermEval (Wiegand et al., 2018) show the growing importance of this subject. The SemEval 2019 shared task 6 (Zampieri et al., 2019b) further addresses this topic by introduc- ing the Offensive Language Identification Dataset (OLID), which consists of tweets, labeled with a three-level annotation model (Zampieri et al., 2019a). Sub-task A is composed of a binary clas- sification problem of whether a tweet in the dataset is offensive or not. Sub-task B focuses on differ- ent categories of offensive language and the goal of sub-task C is to identify the targeted individual of an offensive tweet. In the following paper, we present our contri- bution to sub-task A. After the related work sec- tion, we outline our conducted experiments in section 3 and further describe the used baseline model, as well as the submitted model. In sec- tion 4 we report the results of our experiments on the OLID dataset and the additionally used GermEval dataset. Section 5 discusses our results and section 6 concludes our work and describes possible future work. 2 Related Work Several methods and models have been presented in literature over the last decade to address the predicament of identifying hate speech, offensive language, and online aggressiveness. In the fol- lowing section, we present the most notable con- tributions related to our work. The tweets collected by Davidson et al. (2017) were divided into Hate, Offensive, and Neither. Their proposed algorithm uses unigram, bigram, and trigram tokens as features, weighted by …

Content truncated to 500 words