HATERecognizer at SemEval-2019 Task 5: Using Features and Neural Networks to Face Hate Recognition

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 409–415 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 409 HATERecognizer at SemEval-2019 Task 5: Using Features and Neural Networks to Face Hate Recognition Victor Nina-Alcocer Department of Computer Systems and Computation Universitat Polit`ecnica de Val`encia, Spain vicnial@inf.upv.es Abstract This paper presents a detailed description of our participation in task 5 on SemEval-20191. This task consists of classifying English and Spanish tweets that contain hate towards wo- men or immigrants. We carried out several ex- periments; for a finer-grained study of the task, we analyzed different features and designing architectures of neural networks. Additionally, to face the lack of hate content in tweets, we include data augmentation as a technique to in- crease hate content in our datasets. 1. Introduction HatEval (Basile et al., 2019) aims to identify ca- ses of aggressiveness and hate speech towards wo- men and immigrants in social media considering tweets messages written in English and Spanish. This task defines two main sub-tasks: Task1. Hate Speech Detection against Immi- grants and Women: predict whether a tweet in English or in Spanish is hateful or not ha- teful. Task2. Aggressive behavior and Target Clas- sification: to identify if a tweet is aggressive or not aggressive, and to identify the target harassed as individual or generic. To tackle the subtasks mentioned above we will use Natural Language Processing (NLP) and Ma- chine Learning (ML) fields to propose two approa- ches. The first approach pretends to know more about valuable features that allow us to get a good understanding of its embedded knowledge. To get this knowledge we study four features that we con- sider important: its structure, its embedded emo- tions, patterns on its pos tagging and skip-grams. Each of these features will allow us to know if 1http://alt.qcri.org/semeval2019/index.php?id=tasks tweets have some patterns that can be used to dis- criminate tweets that contain hate or not. The second approach will try to recognize pat- terns using weights among neurons. To implement this focus we designed several architectures of neural networks (NN), which were fed with dif- ferent kinds of corpora that were processed consi- dering many aspects such as, lemmatization, stem- ming, normalized hashtags, etc. This work is organized into three sections. The first section provides an introduction to this paper. The second one describes the proposed approa- ches, the experiments conducted, and the results that we achieved. The third section presents some conclusions. 2. Systems Description In this section, we present the main features of the two approaches considered in this paper and its respective experiments. The organizers provided a training dataset of 9000 and 4500 tweets written in English and Spa- nish labeled with hate speech (HS), Aggressive behavior (AG) and Target (TR). For what concerns HS the distribution is almost balanced among hate(42 %) and no-hate (58 %) tweets in both lan- guages. Regarding AG and TR, the distribution is skewed towards tweets that do not contain TR (75 %) or …

Content truncated to 500 words