Offensive Video Detection: Dataset and Baseline Results

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4309–4319 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4309 Offensive Video Detection: Dataset and Baseline Results Cleber de S. Alcˆantara, Diego Feij´o, Viviane P. Moreira Instituto de Inform´atica Universidade Federal do Rio Grande do Sul (UFRGS) – Porto Alegre – RS – Brazil {cleber.alcantara,dvfeijo,viviane}@inf.ufrgs.br Abstract Web-users produce and publish high volumes of data of various types, such as text, images, and videos. The platforms try to restrain their users from publishing offensive content to keep a friendly and respectful environment and rely on moderators to filter the posts. However, this method is insufficient due to the high volume of publications. The identification of offensive material can be performed automatically using machine learning, which needs annotated datasets. Among the published datasets in this matter, the Portuguese language is underrepresented, and videos are little explored. We investigated the problem of offensive video detection by assembling and publishing a dataset of videos in Portuguese containing mostly textual features. We ran experiments using popular machine learning classifiers used in this domain and reported our findings, alongside multiple evaluation metrics. We found that using word embedding with Deep Learning classifiers achieved the best results on average. CNN architectures, Naive Bayes, and Random Forest ranked top among different experiments. Transfer Learning models outperformed Classic algorithms when processing video transcriptions, but scored lower using other feature sets. These findings can be used as a baseline for future works on this subject. Keywords: offensive content, hate speech, dataset, video, Portuguese 1. Introduction The wide adoption of social media platforms popularized the creation of user-generated content. Together with the democratization of content creation enabling users to ex- press their ideas, came the dissemination of hate speech and other types of offensive material and behavior such as profanity, cyberbullying, and harassment. When users publish and disseminate offensive content, they are con- tributing to a hostile environment. This type of content might be harmful for users and discourage them from us- ing the platforms. An unpleasant environment could also cause loss of revenue to the owners. Additionally, compa- nies do not wish to be associated with this type of content, which could happen if their advertisements get displayed in an offensive video, for example. To tackle these problems, researchers from both companies and academia have pro- posed approaches aiming at detecting offensive content on different platforms (Schmidt and Wiegand, 2017). On the Web, there is a variety of platforms that enable user content production in various formats, such as text, photo, audio, and video. So far, the text has been the most popular format used by people to do so, thanks to its input and stor- age simplicity, and the diffusion of comment sections sup- ported by social networks. As a result, the vast majority of the existing works focused on identifying offensive content in text (social network posts, news comments, tweets, etc.). However, videos also play an essential role in the …

Content truncated to 500 words