Development of Multi-lingual Models for Detecting Hope Speech Texts from Social Media Comments

Comments on social media can be written in any number of languages, and many of them may also be written in languages with few resources. Hope Speech comments are kind expressions that support or critique a viewpoint without offending the individual or the community. On the other hand, non-hope speech is made up of harsh, mocking, or demotivating words. Since the Covid-19 pandemic, the need for positive reinforcement on the internet has made the field of natural language processing pay more attention to hope speech detection. Hope speech detection looks for words and phrases in social media comments that make people feel good. In this paper, an attempt to share content on these platforms that is positive and helpful is made. The models that are based on transformers to figure out whether a social media comment is “hope speech” or “non-hope speech” has been used. The objective of this work is to find the “hope speech” comments in YouTube datasets that were made as part of the “LT-EDI-ACL 2022: Hope Speech Detection for Equality, Diversity, and Inclusion” shared task. The shared task dataset was suggested in five different languages: Malayalam, Tamil, English, Spanish, and Kannada. The model based on a transformer was used as both a fine-tuner and an adapter transformer. In the end, adapters and fine-tuners do the same thing, but adapters add layers to the main model that has already been trained and freeze the weights of those layers. This study shows that models that are based on adapters do better than models that are fine-tuned. The proposed model classifies the Tamil dataset with an accuracy of 51.7% and the English dataset with an accuracy of 92.1%, which is the highest among all the datasets.