Cross-lingual Offensive Language Identification for Low Resource Languages: The Case of Marathi

Proceedings of Recent Advances in Natural Language Processing, pages 437–443 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_050 437 Cross-lingual Offensive Language Identification for Low Resource Languages: The Case of Marathi Saurabh Gaikwad1, Tharindu Ranasinghe2, Marcos Zampieri1, Christopher M. Homan1 1Rochester Institute of Technology, USA 2University of Wolverhampton, UK T.D.RanasingheHettiarachchige@wlv.ac.uk Abstract The widespread presence of offensive lan- guage on social media motivated the devel- opment of systems capable of recognizing such content automatically. Apart from a few notable exceptions, most research on au- tomatic offensive language identification has dealt with English. To address this shortcom- ing, we introduce MOLD1, the Marathi Of- fensive Language Dataset. MOLD is the first dataset of its kind compiled for Marathi, thus opening a new domain for research in low- resource Indo-Aryan languages. We present results from several machine learning experi- ments on this dataset, including zero-short and other transfer learning experiments on state-of- the-art cross-lingual transformers from exist- ing data in Bengali, English, and Hindi. 1 Introduction The presence of hate speech, cyber-bullying, and other forms of offensive language in online com- munities is a global phenomenon. Even though thousands of languages and dialects are widely used in social media, most studies on the auto- matic identification of such content consider En- glish only, a language for which datasets and other resources such as pre-trained models exist (Rosen- thal et al., 2021). In the past few years researchers have studied this problem on languages such as Arabic (Mubarak et al., 2021), French (Chiril et al., 2019), and Turkish (C¸ ¨oltekin, 2020) to name a few. In doing so, they have created new datasets for each of these languages. Competitions such as Offen- sEval (Zampieri et al., 2020) and TRAC (Kumar et al., 2020a) provided multilingual datasets, which enabled the use of data augmentation methods (Ghadery and Moens, 2020), multilingual word em- beddings (Pamungkas and Patti, 2019), and cross- 1MOLD is available at: https://github.com/ tharindudr/MOLD lingual contextual word embeddings (Ranasinghe and Zampieri, 2020) to tackle this problem. In this paper, we revisit the task of offensive language identification for low resource languages, focusing on Marathi, an Indo-Aryan language spoken by over 80 million people, most of whom live in India. Even though Marathi is spoken by a large population, it is relatively low-resourced compared to other languages spoken in the region. We collect and annotate the first Marathi offensive language identification dataset to date and we train a number of monolingual models on this dataset. Finally, we explore state-of-the-art cross-lingual learning methods to project predictions to Marathi from Bengali, Hindi, and English. We address two research questions in this paper: RQ1: What is the impact of the dataset size in monolingual and cross-lingual models for offen- sive language identification? While the Marathi dataset is relatively small, cross-lingual transfer learning methods allow us to take advantage of larger available datasets in other languages. RQ2: What is the influence of language similarity in cross-lingual predictions for offensive language identification? Previous work used English as the base language to make predictions in …

Content truncated to 500 words