Hierarchical Attention-enhanced Contextual CapsuleNet for Multilingual Hope Speech Detection

Abstract

Social media was initially intended for creative purposes, but a notable dissemination of offensive material adversely affects users of these platforms. It is imperative to spotlight and endorse positive and uplifting instances, often overshadowed by the massive influx of user-generated content. To address the challenge of detecting hopeful messages in Dravidian languages such as Tamil, Malayalam, and Kannada that have limited resources, we propose HopeCap, a novel multilingual hope speech detection framework. HopeCap employs a hierarchical attention-enhanced novel capsule network. The proposed CapsuleNet leverages the integrated representation of the prediction vector from the child capsule and the final vector obtained through dynamic routing. This allows CapsuleNet to capture spatial information more effectively. The hierarchical attention module captures the word-level and sentence-level attention features that are integrated with capsule features and auxiliary features. The proposed method computes three classification probabilities that are computed for translated, transliterated original script, and transliterated Roman script versions of the comment. Original script refers to the Tamil, Malayalam, and Kannada scripts for the respective language scripts. The mean of three probabilities provides enhanced classification performance. Through rigorous analysis of HopeCap on three low-resource Dravidian languages, the study sheds light on the effectiveness of the proposed approach. HopeCap outperforms the existing state-of-the-art methods with an average increase of 6.13%, 6.58%, and 4.26% in terms of weighted-F1 for Tamil, Malayalam, and Kannada languages, respectively. © 2024 Elsevier Ltd