NLPDove at SemEval-2020 Task 12: Improving Offensive Language Detection with Cross-lingual Transfer

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1576–1586 Barcelona, Spain (Online), December 12, 2020. 1576 NLPDove at SemEval-2020 Task 12: Improving Offensive Language Detection with Cross-lingual Transfer Hwijeen Ahn1*, Jimin Sun2*, Chan Young Park3*, Jungyun Seo 1 1Sogang University, Republic of Korea 2Seoul National University, Republic of Korea 3Language Technologies Institute, Carnegie Mellon University, USA {hwijeen, seojy}@sogang.ac.kr, jiminsun@dm.snu.ac.kr, chanyoun@cs.cmu.edu Abstract This paper describes our approach to the task of identifying offensive languages in a multilingual setting. We investigate two data augmentation strategies: using additional semi-supervised labels with different thresholds and cross-lingual transfer with data selection. Leveraging the semi- supervised dataset resulted in performance improvements compared to the baseline trained solely with the manually-annotated dataset. We propose a new metric, Translation Embedding Distance, to measure the transferability of instances for cross-lingual data selection. We also introduce various preprocessing steps tailored for social media text along with methods to fine-tune the pre-trained multilingual BERT (mBERT) for offensive language identification. Our multilingual systems achieved competitive results in Greek, Danish, and Turkish at OffensEval 2020. 1 Introduction Online social media has become one of the most important means of communication. Unfortunately, the discourse is often laden with abusive language that can have damaging effects on social media users (Zampieri et al., 2019a). As online abuse grows as a serious social problem, many social media platforms are employing human moderators to track inappropriate contents and comments. However, such human moderation comes at the cost of additional operation and the mental suffering of the workers (Roberts, 2014; Dosono and Semaan, 2019). In response, the automatic identification of offensive language has been a rising area of interest to both academia and industry. Computational methods presented in several prior studies (Park and Fung, 2017; Pitsilis et al., 2018; Wulczyn et al., 2017; Founta et al., 2018) have shown promising results on how we can automatically detect a large number of offensive contents with minimal human intervention. As an effort to promote this line of research, OffensEval was organized at SemEval-2019 (Zampieri et al., 2019b). The organizers released a dataset, Offensive Language Identification dataset (OLID), that contains English tweets with human annotations over three sub-tasks: 1) whether a tweet is offensive or not, 2) whether an offensive tweet is targeted or not, and if so 3) whether the target is an individual or a group. The second edition of OffensEval, OffensEval 2020 : Multilingual Offensive Language Identification in Social Media, holds the same task but extends the OLID to multiple languages (English, Arabic, Danish, Greek, and Turkish) (Zampieri et al., 2020). This paper describes our system, NLPDove, submitted to OffensEval 2020 . We present two different data augmentation strategies. First, for English, we augment the training set with reliable samples filtered out from the provided semi-supervised dataset. Second, we use samples from languages other than the target language to enhance performance in the target language, namely cross-lingual transfer. Specifically, we propose a new metric, Translation Embedding Distance(TED), to measure the data transferability of each example in …

Content truncated to 500 words