Analyzing Cross-Lingual Approaches: a Case Study for Detecting Multilingual Hope Expressions in YouTube Comments

Designing natural language processing (NLP) based systems for under-resource languages opens up new opportunities for marginalized communities. However, these NLP systems usually face several issues concerning under-resource languages such as lack of annotated data and available processing resources that result in poor NLP performance. These concerns can be addressed with cross-lingual learning (CLL) approach and multilingual learning is one of the special cases of CLL. Hope expression identification is a new emerging research direction that supports and promotes positive material on social media to encourage positivity and harmony in society. However, it is rarely handled as a multilingual paradigm and the Urdu language is ignored by the literature. This paper deals with the task of detecting multilingual hope expression using a transfer learning technique with fine-tuning. For this purpose, the joint multilingual methodology is considered which suggests designing a universal classifier for different languages. The potential of multilingual bidirectional encoder representations from transformers (mBERT) is exploited with fine-tuning and a grid search approach is used to find optimal values of hyperparameters. An annotated Urdu corpus is also designed. The experiments conducted on the bilingual hope expression corpus revealed that the proposed methodology presented benchmark performance and outperformed the nine baselines. Furthermore, it improved hope class identification by 13.44%, not-hope by 0.07%, and macro by 6.66% considering the F1-score.