Hope Speech detection in Under-resourced Kannada language

Numerous methods have been developed to monitor the spread of negativity in modern years by eliminating scurrilous, offensive, and trenchant comments from social media platforms. However, there is little to no study on embracing positivity, reinforcing supportive and reassuring content in online forums. Consequently, we aim to bridge the gap amidst research on hope speech and knowledge retrieved from social media by proposing KanHopeEDI, a code mixed hope speech dataset for equality, diversity, and inclusion in Kannada, an under-resourced Dravidian language. The dataset consists of 6,176 user-generated comments in code mixed Kannada crawled from YouTube and manually labelled as bearing hope speech or not-hope speech. We introduce DC-BERT4HOPE, a dual-channel model that uses the English translation of KanHopeEDI for supplementary training to promote hope speech detection. The approach achieves the best weighted F1-score 0.756, bettering other models. Henceforth, KanHopeEDI aims to instigate research in Kannada while broadly promoting researchers to take a pragmatic approach towards online content that encourages, positive, and supportive.