Working with Few Samples: Methods that Help Analyze Social Attitude and Personal Emotion

Mengwen Liang, Jie Zhou, Yuling Sun, Liang He

Abstract: In the past decade, sentiment analysis on social media has attracted great attention and has been used in many studies in CSCW and related fields. Recently, with the rapid development of machine learning, using machine learning methods to analyze sentiment has become an efficient experiment framework. Now, the existing sentiment analysis methods in machine learning are mainly based on supervised learning, and they need enough training data to ensure high accuracy. They encounter a common problem that they cannot recognize and calculate the emotion of samples with unseen labels, which don't belong to the training set. However, most data collected from social media is unstructured and unlabeled, which challenges the effectiveness and usability of existing methods. In our study, we first refer to existing sentiment analysis methods and zero-shot learning for addressing the problem. After that, we propose two zero-shot sentiment analysis methods and design an experiment to compare our methods and strong baselines. In conclusion, our methods obtain better results and we try to apply these methods to future social media research.