The datasets used to train hate speech detection models affect the models’ performance. The majority of existing datasets are created using a small number of hate domains or instances to define hate subjects. Large-scale analysis and transfer learning for hate domains are hampered as a result. In this work, we create massive twitter datasets with 100k human-labeled tweets each for the purpose of detecting hate speech in English and Turkish, a language with limited resources. We have scattered an equal amount of tweets over five domains in our datasets. In terms of large-scale hate speech identification, Transformer-based language models perform at least 5% better in English and 10% better in Turkish when compared to conventional bag-of-words and neural models, according to experimental data corroborated by statistical testing. Additionally, the performance is adaptable to various training sizes; with 20% of training instances, 98% of the English performance and 97% of the Turkish performance are recovered. We also investigate the capacity of cross-domain transfer between hate domains to generalize. We demonstrate that, on average, other domains recover 96% of a target domain’s performance for English and 92% for Turkish. Sports struggle the hardest to generalize to other domains, but gender and religion fare better.

https://paperswithcode.com/paper/large-scale-hate-speech-detection-with-cross-1/review

By author

Leave a Reply

Your email address will not be published. Required fields are marked *