Team Rouges at SemEval-2020 Task 12: Cross-lingual Inductive Transfer to Detect Offensive Language

Abstract

With the growing use of social media and its availability, many instances of the use of offensive language have been observed across multiple languages and domains. This phenomenon has given rise to the growing need to detect the offensive language used in social media crosslingually. In OffensEval 2020, the organizers have released the multilingual Offensive Language Identification Dataset (mOLID), which contains tweets in five different languages, to detect offensive language. In this work, we introduce a cross-lingual inductive approach to identify the offensive language in tweets using the contextual word embedding XLM-RoBERTa (XLM-R). We show that our model performs competitively on all five languages, obtaining the fourth position in the English task with an F1-score of 0.919 and eighth position in the Turkish task with an F1-score of 0.781. Further experimentation proves that our model works competitively in a zero-shot learning environment, and is extensible to other languages.

A survey on multi-lingual offensive language detection

“…On the other hand, cross-lingual language models like XLM were also widely employed, where we found implementation of XLM-RoBERTa (XLM-R) ( Roy et al, 2021a ; Bhatia et al, 2021 ; Wang et al, 2020 ; De la Peña Sarracén & Rosso, 2022 ; Zia et al, 2022 ; Tita & Zubiaga, 2021 ; Ranasinghe & Zampieri, 2021b ; Dadu & Pant, 2020 ; Eronen et al, 2022 ; Ranasinghe & Zampieri, 2021a , 2020 ; Mozafari, Farahbakhsh & Crespi, 2022 ; Barbieri, Espinosa Anke & Camacho-Collados, 2022 ; Awal et al, 2024 ; Stappen, Brunn & Schuller, 2020 ), and both XLM-R and XLM-T ( Montariol, Riabi & Seddah, 2022 , Riabi, Montariol & Seddah, 2022 ). These approaches have all been shown to improve performance on tasks involving multilingual/cross-lingual hate speech detection because they are more likely able to capture semantic and syntactic features across languages thanks to their pre-training on multilingual large volumes of texts.…”

OffensEval 2023: Offensive language identification in the age of Large Language Models

“…The second best team (Wang et al 2020) also used an ensemble approach, where they first finetuned two multilingual XLM-RoBERTa models, XLM-RoBERTa-base, and XLM-RoBERTa-large (Conneau et al 2019). In comparison, the third placed team fine-tuned only one multilingual XLM-RoBERTa model (Dadu and Pant 2020).…”

A systematic review of Hate Speech automatic detection using Natural Language Processing

“…The second team Wang et al [143] achieved an F1 score of 0.9204, and it used RoBERTa-large that was fine-tuned with the dataset in an unsupervised way. The third team Dadu and Pant [33], achieved an F1 score of 0.9198, using an ensemble that combined XLM-RoBERTa-base and XLM-RoBERTa-large trained on Subtask A data for all languages. The top-10 teams were close to each other and employed BERT, RoBERTa or XLM-RoBERTa models; sometimes CNNs and LSTMs were also been mentioned either for comparison or hybridization purpose.…”