CAMsterdam at SemEval-2019 Task 6: Neural and graph-based feature extraction for the identification of offensive tweets

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 556–563 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 556 CAMsterdam at SemEval-2019 Task 6: Neural and graph-based feature extraction for the identification of offensive tweets Guy Aglionby†, Christopher Davis†, Pushkar Mishra‡, Andrew Caines†, Helen Yannakoudakis†, Marek Rei†, Ekaterina Shutova* & Paula Buttery† † Department of Computer Science & Technology, University of Cambridge, U.K. {ga384,ccd38,apc38,hy260,mr472,pjb48}@cam.ac.uk ‡ Facebook AI, London, U.K. pushkarmishra@fb.com * Institute for Logic, Language and Computation, University of Amsterdam, Netherlands e.shutova@uva.nl Abstract We describe the CAMsterdam team entry to the SemEval-2019 Shared Task 6 on offen- sive language identification in Twitter data. Our proposed model learns to extract tex- tual features using a multi-layer recurrent net- work, and then performs text classification us- ing gradient-boosted decision trees (GBDT). A self-attention architecture enables the model to focus on the most relevant areas in the text. We additionally learn globally optimised em- beddings for hashtags using node2vec, which are given as additional tweet features to the GBDT classifier. Our best model obtains 78.79% macro F1-score on detecting offensive language (subtask A), 66.32% on categorising offence types (targeted/untargeted; subtask B), and 55.36% on identifying the target of of- fence (subtask C). 1 Introduction The SemEval-2019 shared task 6 (‘OffensEval’) involved three sub-parts: the classification of tweets as offensive or not (subtask A), classify- ing whether they are targeted insults or not (sub- task B), and finally whether the targeted insults are aimed at an individual, group or otherwise (sub- task C). Further details may be found in the shared task report (Zampieri et al., 2019b). Here we de- scribe CAMsterdam’s competition entry. In recent years, there has been a growing inter- est in the automatic detection of offensive opinions expressed in online texts, including those posted in discussion forums, news article comment sec- tions, and social networks. Such detection is not straightforwardly a matter of identifying texts con- taining obscene words (Malmasi and Zampieri, 2018); offensiveness often arises from the con- text, current affairs, world knowledge, the use of acronyms and slang, and the identity of the authors and audience. Therefore the task is a challenging one, but one with real world impact: if measures can be taken to identify and curtail trolling, the toxicity of the internet can to some extent be re- duced. There is evidence that online harassment is connected with oppression, violence and sui- cide (Dinakar et al., 2011; Sood et al., 2012; Wul- czyn et al., 2017), and there may moreover be rea- sons for concern about the perpetrator’s wellbeing along with that of the victims (Cheng et al., 2017). Our approach to the task extends the work of Mishra et al. (2018b), who extract features from tweets using an RNN for subsequent use in a gradient-boosted decision tree (GBDT) (Ke et al., 2017). Firstly, we experiment with changes to the RNN, including the use of self-attention (Rei and Søgaard, 2019) and ELMo embeddings (Pe- ters et al., 2018). Secondly, we add …

Content truncated to 500 words