Team Oulu at SemEval-2020 Task 12: Multilingual Identification of Offensive Language, Type and Target of Twitter Post Using Translated Datasets

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1628–1637 Barcelona, Spain (Online), December 12, 2020. 1628 Team Oulu at SemEval-2020 Task 12: Multilingual Identification of Offensive Language, Type and Target of Twitter Post Using Translated Datasets Md. Saroar Jahan University of Oulu, Faculty of Information Tech., CMVS PO Box 4500, Oulu 90014 FINLAND mjahan18@student.oulu.fi Mourad Oussalah University of Oulu, Faculty of Information Tech., CMVS PO Box 4500, Oulu 90014 FINLAND mourad.oussalah@oulu.fi Abstract With the proliferation of social media platforms, anonymous discussions together with easy online access, reports on offensive content have caused serious concern to both authorities and research communities. Although there is extensive research in identifying textual offensive language from online content, the dynamic discourse of social media content, as well as the emergence of new forms of offensive language, especially in a multilingual setting, calls for future research in the issue. In this work, we tackled Task A, B, and C of Offensive Language Challenge at SemEval2020. We handled offensive language in five languages: English, Greek, Danish, Arabic, and Turkish. Specifically, we pre-processed all provided datasets and developed an appropriate strategy to handle Tasks (A, B, & C) for identifying the presence/absence, type and the target of offensive language in social media. For this purpose, we used OLID2019, OLID2020 datasets, and generated new datasets, which we made publicly available. We used the provided unsupervised machine learning implementation for automated annotated datasets and the online Google translation tools to create new datasets as well. We discussed the limitations and the success of our machine learning-based approach for all the five different languages. Our results for identifying offensive posts (Task A) yielded satisfactory accuracy of 0.92 for English, 0.81 for Danish, 0.84 for Turkish, 0.85 for Greek, and 0.89 for Arabic. For the type detection (Task B), the results are significantly higher (.87 accuracy) compared to target detection (Task C), which yields .81 accuracy. Moreover, after using automated Google translation, the overall efficiency improved by 2% for Greek, Turkish, and Danish. 1 Introduction The emergence of Web 2.0 platform that enabled user-generated content and participatory culture has witnessed the proliferation of online hate speech at an unprecedented level, increasing the likelihood of a random people of any age group to be subject to online harassment and abuse through some internet forum, message board or social network platform. On the other hand, offensive language fosters discrimination against particular categories on the basis of gender, ethnicity, or race (Nockleby, 2000), which violates the Universal Declaration of Human Rights. This often leads to a profound negative impact on the society as a whole, especially for youth and vulnerable groups, as well as on the policy-makers and business entities. To overcome this issue, many internet companies put forward standards and generic guidelines that users must adhere to when publishing content on such platforms; at the same time, they employ manual annotators to identify offensive language and remove the post (s) accordingly. Nevertheless, such manual effort is inevitably costly, non-scalable, and non-sustainable, which …

Content truncated to 500 words