Few-Shot Semi-Supervised Robust Text Classification with MAML

Abstract

The need for few-shot semi-supervised text classification arises in a variety of applications, including, e.g., recommendation systems classifying textual content such as product descriptions or news articles based on limited amounts of user feedback. In such settings, existing supervised methods lack a way to leverage unlabeled data, which may be available in larger amounts.

We develop a method for improving the accuracy and robustness of a supervised meta-learning algorithm (Model-Agnostic Meta-Learning) applied to few-shot natural language text classification tasks. We also detail a way to incorporate semi-supervised learning into MAML by designing a procedure to create self-supervised tasks from unlabeled text examples. We present the test accuracies in experimental results for sentiment classification and topic classification. As a representative example, we achieved gains in accuracy ranging from 1% to 3% on Amazon review and news headline datasets