HateCheckHIn: Evaluating Hindi Hate Speech Detection Models

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5378–5387 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5378 HateCheckHIn: Evaluating Hindi Hate Speech Detection Models Mithun Das, Punyajoy Saha, Binny Mathew, Animesh Mukherjee Department of Computer Science & Engineering Indian Institute of Technology, Kharagpur West Bengal, India – 721302 {mithundas,punyajoys,binnymathew}@iitkgp.ac.in, animeshm@cse.iitkgp.ac.in Abstract Due to the sheer volume of online hate, the AI and NLP communities have started building models to detect such hateful content. Recently, multilingual hate is a major emerging challenge for automated detection where code-mixing or more than one language have been used for conversation in social media. Typically, hate speech detection models are evaluated by measuring their performance on the held-out test data using metrics such as accuracy and F1-score. While these metrics are useful, it becomes difficult to identify using them where the model is failing, and how to resolve it. To enable more targeted diagnostic insights of such multilingual hate speech models, we introduce a set of functionalities for the purpose of evaluation. We have been inspired to design this kind of functionalities based on real-world conversation on social media. Considering Hindi as a base language, we craft test cases for each functionality. We name our evaluation dataset HateCheckHIn. To illustrate the utility of these functionalities , we test state-of-the-art transformer based m-BERT model and the Perspective API. Keywords: Hate speech, evaluation, multilingual, code-mixed, test cases, functionalities 1. Introduction Hate speech is a serious concern that is plaguing on- line social media. With the increasing amount of hate speech, automatic detection of such content is receiving significant attention from the AI and NLP communities, and models are being developed to detect hate speech online. While earlier efforts in hate speech detection focused mostly on English, recently researchers have begun to develop multilingual models of hate speech de- tection. However, even state-of-the-art models demon- strate substantial weaknesses (Mishra et al., 2019; Vid- gen et al., 2019). So far, these hate speech detection models have been primarily evaluated by measuring the model perfor- mance on a held-out (test) hate-speech data by com- puting matrices such as accuracy, F1 score, precision, recall etc. at the aggregate level (Waseem and Hovy, 2016; Davidson et al., 2017; Kumar et al., 2020). Higher values of these metrics indicate more desirable performance. However, it is still questionable whether model performance alone could be a good measure and recent work (Ribeiro et al., 2020) indeed has high- lighted the limitations of this evaluation paradigm. Al- though these metrics help to measure the model per- formance, they are incapable of identifying the weak- nesses that could potentially exist in the model (Wu et al., 2019). Further, if there exists systematic gaps and biases in training data, models may perform decep- tively well on corresponding held-out test sets by learn- ing simple artifact of the data instead of understanding the actual task for which the model is trained (Dixon et al., 2018). Existing research has …

Content truncated to 500 words