NLP@UIOWA at SemEval-2019 Task 6: Classifying the Crass using Multi-windowed CNNs

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 704–711 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 704 NLP@UIOWA at SemEval-2019 Task 6: Classifying the Crass using Multi-windowed CNNs Jonathan Rusert and Padmini Srinivasan Department of Computer Science University of Iowa Iowa City, IA, USA {jonathan-rusert, padmini-srinivasan}@uiowa.edu Abstract This paper proposes a system for OffensEval (SemEval 2019 Task 6), which calls for a sys- tem to classify offensive language into several categories. Our system is a text based CNN, which learns only from the provided training data. Our system achieves 80 - 90% accu- racy for the binary classification problems (of- fensive vs not offensive and targeted vs untar- geted) and 63% accuracy for trinary classifica- tion (group vs individual vs other). 1 Introduction Background. Social media (e.g. Twitter, Face- book) is widely used today. For example, 68% of all Americans report owning a Facebook account in 2018 (Smith and Anderson, 2018), while 71% of Americans (within the ages of 18-24) report us- ing Twitter. Online gaming is a second popular use of the internet, reaching upwards of 80 - 100 million monthly users depending on game (Goslin, 2018). These uses demonstrate the internet as a way for humans to connect with others. How- ever, connecting with others can carry a downside. More than 1 in 3 young people have been cyber- bullied online (cyb, 2018), this extends to around half of teens. Besides cyberbullying, offensive language, on a public forum, can cause users to stay away from certain platforms. Because of this, companies have increased their efforts to remove offensive language from their platforms (Terdi- man, 2018). With the large amount of traffic these platforms see, a purely manual approach to detect- ing/removing offensive language is impossible, which means an automated approach is needed to help. OffensEval (Zampieri et al., 2019b) provides a community driven opportunity to build such sys- tems. We approach this problem of classifying of- fensive tweets with a CNN architecture trained on the provided training dataset. 2 Proposed Approach Our system is a variation of a Convolutional Neu- ral Network (CNN), which was chosen since it has seen success previously with classification tasks CNN Infrastructure. We experimented with two different Convolutional Neural Networks. The first whose architecture is based on the CNN originally proposed in (Kim, 2014) (CNN 1), and the second a combination of multiple CNNs (CNN 2). We further discuss each of these CNNs and their comparison on the training data. The vi- sual structure for CNN 1 and CNN 2 can be found in figure 1 and figure 2 respectively. Preprocessing. Both CNN 1 and CNN 2 begin by preprocessing the text of the tweet. As noted in section 3, URLs and user mentions are already denoted as URL and USER. Basic cleaning of the text is applied, includes removal of punctuation, converting text to lowercase, and filtering of stop- words via NLTK’s stopword list1. Finally, all sep- arated words are tokenized via nltk’s tokenize() function2. …

Content truncated to 500 words