Comparative Analysis of Methods for Sentiment Labeling with Limited Labeled Data

This paper presents a comparative analysis of various methods for automatic sentiment labeling. We investigate the performance of three semi-supervised learning methods alongside a state-of-the-art zero-shot learning model on four datasets. Our experiments reveal the strengths and weaknesses of each approach, particularly in scenarios with limited labeled data. While traditional semi-supervised methods demonstrate exemplary performance in leveraging unlabeled data, zero-shot classification offers a compelling alternative without requiring labeled instances for the target task. However, our findings also highlight significant limitations of zero-shot learning, mainly when applied to datasets with complex textual structures. These insights underscore the need for careful method selection based on the specific characteristics and constraints of the sentiment analysis task.