Meta-Learning for Efficient Cross-Domain Sentiment Categorization: Leveraging Small Training Datasets for Enhanced Classifier Performance

Researchers and technologists working on sentiment categorization algorithms need to collect and organize data for every new domain. Annotating these corpora is expensive work, particularly in areas where features change over time. Therefore, this study explores the concept of training classifiers on a small range of domains before applying them to more domains. One problem is that when the data of training and testing distribution differ, classifiers usually perform poorly. A further problem is that it's not obvious how to select domains that will be appropriate for training and testing other domains based on similarities. To overcome this challenge, we develop classifiers that are adaptable and still perform well on test sets obtained from various distributions by applying several, smaller training datasets, we propose to improve cross-domain embedding learning for the Few-Shot Learning (FSL) environment by using the most recent advancements in Meta-Learning (MetaL).