Hope Speech Detection for Dravidian Languages Using Cross-Lingual Embeddings with Stacked Encoder Architecture

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Media Analysis

Background:

• The COVID-19 pandemic highlighted the need for online positive reinforcement.
• Work-induced stress necessitates external inspiration, often found online.
• Hope speech detection aims to identify content invoking positive emotions.

Purpose of the Study:

• To propose a multilingual model for automatic hope speech detection.
• To focus on Dravidian languages, including Tamil and Malayalam.
• To address the challenge of code-mixed social media data.

Main Methods:

• Employed a stacked encoder architecture.
• Utilized language-agnostic cross-lingual word embeddings for code-mixed data.
• Conducted empirical analysis comparing against traditional, transformer, and transfer learning methods.
• Performed a k-fold paired t-test to validate model superiority.

Main Results:

• Achieved an F1-score of 0.61 for Tamil and 0.85 for Malayalam.
• The proposed model demonstrated superior performance compared to other tested approaches.
• The methodology is competitive with state-of-the-art methods.

Conclusions:

• The developed multilingual model effectively detects hope speech in Dravidian languages.
• The stacked encoder architecture with cross-lingual embeddings is suitable for code-mixed data.
• This work contributes to increasing positive online content and mental well-being support.