LaSTUS/TALN at SemEval-2019 Task 6: Identification and Categorization of Offensive Language in Social Media with Attention-based Bi-LSTM model

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 672–677 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 672 LaSTUS/TALN at SemEval-2019 Task 6: Identification and Categorization of Offensive Language in Social Media with Attention-based Bi-LSTM model Lutfiye Seda Mut Altin, `Alex Bravo, Horacio Saggion Large Scale Text Understanding Systems Lab / TALN Research Group Department of Information and Communication Technologies (DTIC) Universitat Pompeu Fabra Tanger 122, Barcelona (08018), Spain lutfiyeseda.mut01@estudiant.upf.edu, {alex.bravo, horacio.saggion}@upf.edu Abstract This paper describes a bidirectional Long- Short Term Memory network for identifying offensive language in Twitter. Our system has been developed in the context of the Se- mEval 2019 Task 6 which comprises three dif- ferent sub-tasks, namely A: Offensive Lan- guage Detection, B: Categorization of Offen- sive Language, C: Offensive Language Target Identification. We used a pre-trained Word Embeddings in tweet data, including informa- tion about emojis and hashtags. Our approach achieves good performance in the three sub- tasks. 1 Introduction As the amount of user generated content in so- cial media is increasing at an exponential pace, detecting offensive language and harmful content automatically in an efficient way is a very impor- tant issue for the society. Recent work has shown that offensive language in various forms such as hate speech, cyberbullying, profanity and harass- ment has negative effects especially in adolescents (Hamm et al., 2015). The shared task, Categorizing Offensive Lan- guage in Social Media (SemEval 2019 - Task 6), focuses on improving identification of offensive language by considering type and target of the of- fense into account (Zampieri et al., 2019b). The task is composed of three sub-tasks. Sub-task A aims to identify if a given tweet is offensive or not (annotated as OFF or NOT). Sub-task B aims to categorize the offense type in offensive tweets into two categories: targeted (TIN) or untar- geted (UNT) meaning that if a tweet contains an insult or threat to an individual, a group or some- thing else or if a tweet contains non-targeted of- fense such as general profanity or non-acceptable language. Lastly, Sub-task C aims to identify the target type of targeted offensive posts. The tar- get type is supposed to be classified as individ- ual, group or other for the rest (annotated as IND, GRP or OTH). We submitted three different runs for each sub-task. The training dataset released by the shared task organizers, consists of 14,100 English tweets with one annotation layer per task with a hierarchical annotation scheme where each annotation level is related to an independent sub-task. The meth- ods used to collect this dataset is described in (Zampieri et al. (2019a)). Examples from the dataset with annotations at the end are given be- low: ”@USER That shit weird! Lol OFF (of- fensive) - -” ”@USER @USER You are an embar- rassing citizen!! OFF TIN -” ”@USER @USER Liberals ruin every- thing! OFF TIN GRP” This paper describes a bidirectional Long Short Term Memory network (biLSTM) model with an Attention layer to identify offensive …

Content truncated to 500 words