IRLab_DAIICT at SemEval-2020 Task 12: Machine Learning and Deep Learning Methods for Offensive Language Identification

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 2006–2011 Barcelona, Spain (Online), December 12, 2020. 2006 IRLab DAIICT at SemEval-2020 Task 12: Machine Learning and Deep Learning Methods for Offensive Language Identification Apurva Parikh DA-IICT Gandhinagar, India apurvakparikh @gmail.com Abhimanyu Singh Bisht DA-IICT Gandhinagar, India bisht2492 @gmail.com Prasenjit Majumder DA-IICT Gandhinagar, India p majumder @daiict.ac.in Abstract The paper describes systems that our team IRLab DAIICT employed for shared task OffensEval 2020: Multilingual Offensive Language Identification in Social Media shared task. We conducted experiments on the English language dataset which contained weakly labelled data. There were three sub-tasks but we only participated in sub-tasks A and B. We employed Machine learning techniques like Logistic Regression, Support Vector Machine, Random Forest and Deep learning techniques like Convolutional Neural Network and BERT. Our best approach achieved a MacroF1 score of 0.91 for sub-task A and 0.64 for sub-task B. 1 Introduction Since the past few years, rapid improvement in communication technologies has led to an exponential rise in the number of people who have unrestricted access to the internet. This newfound abundance has seen the emergence of the ”Netizen”, and Social Media Platforms like Twitter, Facebook, Instragram etc, the cyber-agoras where Netizens assemble to discuss and debate ideas. But in several cases the exchange on these platforms is not civil and at times the behaviour exhibited by individuals on social media platforms can be hostile, targeting individuals or communities by posting insults or threats. Twitter receives more than 500 million tweets daily1, which can be classified as offensive or hate speech. This makes the detection of offensive and hateful content on social media platforms an important research question in the field of Natural Language Processing. To promote research in this field several competitions have been organised such as OffensEval 2019 (Zampieri et al., 2019b), HASOC 2019 (Modha et al., 2019), HatEval 2019 (Basile et al., 2019). OffensEval 2020 (Zampieri et al., 2020) task was organised for five languages English, Arabic, Danish, Greek and Turkish. The organizers proposed a hierarchical task which consists of three sub-tasks as below: • Sub-task A: Identification of Offensive Language, posts are classified as follows: – Offensive (OFF): if post contains non-acceptable language or targeted offense. – Not Offensive (NOT): post that do not contain profanity or offensive language. • Sub-task B: Categorization of Offensive, posts identified as OFF in sub-task A are further catego- rized as: – Targeted Insult (TIN): posts containing an insult/threat to an individual, group, or others. – Untargeted Insult (UNT): posts containing non-targeted profanity and swearing. • Sub-task C: Identification of target,in which posts identified as TIN in sub-task B, are further classified on basis of type of targeted insult. – Individual(IND): posts targeting an individual. This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/. 1https://sproutsocial.com/insights/social-media-statistics/ 2007 – Group (UNT): posts targeting group of people. – Other (OTH): target that does not fall into any of above two category. All three sub-tasks were organised for …

Content truncated to 500 words