Modelling Language Acquisition through Syntactico-Semantic Pattern Finding

Findings of the Association for Computational Linguistics: EACL 2023, pages 125–132 May 2-6, 2023 ©2023 Association for Computational Linguistics Data-Efficient Methods For Improving Hate Speech Detection Sumegh Roychowdhury ∗ IIT Kharagpur, India sumeghtech@gmail.com Vikram Gupta ShareChat, India vikramgupta@sharechat.co Abstract Scarcity of large-scale datasets, especially for resource-impoverished languages encouraged exploration of data-efficient methods for hate speech detection. In this work, we progress implicit and explicit hate speech detection us- ing an input-level data augmentation technique, task reformulation using entailment and cross- learning across five languages. Our proposed data augmentation technique EasyMixup, im- proves the F1 performance across languages by 0.5-9%. We also observe substantial F1 gains of 1-8% by reformulating hate speech detec- tion as Entailment-style problem. We further probe the contextual models and ob- serve that higher layers encode implicit hate while lower layers focus on explicit hate, high- lighting the importance of token-level under- standing for explicit and context-level for im- plicit hate speech detection. 1 1 Introduction Deep learning based methods (Badjatiya et al., 2017; Zhang et al., 2018; Kshirsagar et al., 2018) have shown impressive results in detecting hate speech. Transformer based models (Caselli et al., 2021; Tekiro˘glu et al., 2020; Aluru et al., 2020; Mozafari et al., 2019; Dutta et al., 2022) have fur- ther pushed the state-of-the-art by leveraging large amount of unlabeled data in a self-supervised man- ner. Various hate speech detection datasets have been contributed in textual (Gibert et al., 2018; Davidson et al., 2017; Founta et al., 2018), au- dio (Gupta et al., 2022) and visual (Gomez et al., 2020) domains. However, these algorithms are data-hungry and motivate development of algo- rithms which are data-efficient. To tackle this, we introduce an input-level data augmentation technique EasyMixup and im- prove hate speech detection in monolingual and ∗Work done during internship at ShareChat 1Code and Dataset splits - https://github.com/ Sumegh-git/data_efficient_hatedetect multilingual settings. EasyMixup is inspired by mixup based augmentation techniques which are broadly categorized into input-level mixup (Yun et al., 2019; Kim et al., 2020; Uddin et al., 2021; Walawalkar et al., 2020) and hidden-level mixup (Verma et al., 2019). EasyMixup follows the input-level paradigm and leverages a simple ob- servation that the label of a hateful instance is preserved on concatenation with a hateful or non- hateful instance. Similarly, label of a non-hateful instance does not change on concatenation with another non-hateful instance. We also study the efficacy of reformulating hate speech detection as Entailment-style prob- lem. We extend the work by (Wang et al., 2021) and perform detailed experiments under implicit, explicit and multilingual settings. We observe that monolingual entailment performs better than En- glish based entailment. This observation is intu- itive because the models are pretrained using pair of sentences from same language and monolingual entailment reflects the same settings. Majority of the existing textual datasets focus on explicit hate speech where swear, cuss, abusive words are used to express the hateful intent. In contrast, implicit hate speech employs subtle, indi- rect and contextual ways for expressing hate speech making it …

Content truncated to 500 words