GL at SemEval-2019 Task 5: Identifying hateful tweets with a deep learning approach.

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 416–419 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 416 GL at SemEval-2019 Task 5: Identifying hateful tweets with a deep learning approach. Gretel Liz De la Pe˜na Sarrac´en Universitat Polit`ecnica de Val`encia, Spain Abstract This paper describes the system we developed for SemEval 2019 on Multilingual detection of hate speech against immigrants and women in Twitter (HatEval - Task 5). We use an approach based on an Attention-based Long Short-Term Memory Recurrent Neural Net- work. In particular, we build a Bidirectional LSTM to extract information from the word embeddings over the sentence, then apply at- tention over the hidden states and finally feed this vector to another LSTM model to get a representation from de data. Then, the output obtained with this model is used to get the pre- diction of each of the sub-tasks with models based on neural networks and linguistic char- acteristics. 1 Introduction Nowadays, the number of content generated by users on social networks is growing rapidly. In this context, the problem of detecting and limit- ing the dissemination of the Hate Speech is be- coming a matter of great importance. Therefore, many efforts are dedicated to studying and treating this phenomenon. A large number of workshops on this topic have been developed in recent years, which reflects the interest of many researchers. Some examples are the Workshop on Trolling, Aggression and Cyberbullying (Kumar et al., 2018), that included a shared task on aggression identification; the tracks on Automatic Misog- yny Identification (AMI) (Fersini et al., 2018a) and on Autohorship and Aggressiveness Analy- sis (MEX-A3T) ( ´Alvarez-Carmona et al., 2018) proposed at IberEval 2018; the Automatic Misog- yny Identificationtask at EVALITA 2018 (Fersini et al., 2018b), the Workshop on Abusive Language (Waseem et al., 2017) and the GermEval Shared Task on the Identification of Offensive Language (Wiegand et al., 2018). The proposed works have used different fea- tures and models. Among them, models based on deep learning, such as Recurrent Neural Net- works (RNN) and Convolutional Neural Networks (CNN) have been widely used. This paper presents a strategy based on RNN, which is an extension of previous models pro- posed for the tasks MEX-A3T and EVALITA 2018 (Cuza et al., 2018; la Pe˜na Sarrac´en et al., 2018). Those models use an Attention-based LSTM in- spired by the work (Yang et al., 2016). In this work the authors use a hierarchical attention net- work for document classification and their experi- ments show that the architecture outperforms pre- vious methods. This last model has two levels of attention mechanisms applied at the word and sen- tence level, enabling it to attend differentially to more and less important content when construct- ing the document representation. The aim of the proposed extension in this work is not only the hate language identification, but also the study of other features of hateful mes- sages. In this case, the model based on the Atten- tion mechanism and …

Content truncated to 500 words