Abstract

Data annotation is a time-consuming and labor-intensive process in classification tasks. Recently, numerous studies have explored the few-shot learning approach using meta-learning, particularly the MAML algorithm. Most research aimed at improving MAML has focused on image classification rather than text data, and the proposed enhancements often involve complex models that require significant processing resources. Furthermore, there is a notable scarcity of research attempting to apply few-shot learning methodologies to the Arabic language. This research paper aims to enhance the performance of the Model-Agnostic Meta-Learning (MAML) algorithm in the domain of few-shot sentiment analysis, with a specific focus on the Arabic language, which suffers from resource scarcity and a lack of multi- domain labeled datasets. This paper addresses two primary challenges: the instability of the MAML algorithm during training, and the importance of measuring divergence between training domains. To improve training stability without requiring substantial processing resources, we propose using Cosine Annealing to schedule the learning rate in the outer loop of MAML. Additionally, we present a significant empirical finding demonstrating that the homogeneity of training domains has a substantial impact on MAML performance. The validity of these contributions is verified through extensive experiments on sentiment analysis datasets in both English and Arabic, including the Amazon Reviews dataset and a multi-domain Arabic dataset compiled from several other research studies, processed, and formatted to be suitable for the MAML algorithm. The results demonstrate the effectiveness of the proposed method in improving the stability and performance of MAML and underscore the importance of training domain homogeneity in few-shot learning scenarios with low processing resources.