Few-shot short-text classification with language representations and centroid similarity.

Aiming at the problems of insufficient labelled samples and low-generalization performance in text classification tasks, this paper studies text classification problems under the condition of few labelled samples and proposes a few-shot short-text classification method (Meta-FCS) that combines the advantages of text semantic vector representation, meta-learning, fine-tuning and vector similarity measurement. The method not only effectively transfers the common features of different fields but also highlights the individual features of this field through fine-tuning. In addition, to facilitate the downstream text classification task, a deep language representation model is proposed. On this basis, the similarity between the query set and the class centroid of the support set is compared to determine the query set category. We evaluate the proposed method on a well-studied sentiment classification dataset, an entity-relationship classification dataset and an news topic dataset. The experimental results show that on these three datasets, the proposed method significantly outperforms the existing state-of-the-art approaches. It can thus be further suggested that the combination of deep language representation, episode training mechanism, and similarity measurement can be a promising solution for few-shot learning (FSL) of natural language processing (NLP) tasks. [ABSTRACT FROM AUTHOR]