FASST: Few-Shot Abstractive Summarization for Style Transfer

Abstract

Unsupervised text style transfer methods aim to transfer the style of the text without affecting its fundamental meaning using non-parallel data. Although previous work has explored few-shot learning for this task, incorporating few-shot abstractive summarization and its benefits have not yet been explored. Hence, we propose a novel unsupervised text style transfer approach using few-shot abstractive summarization. In our method, we infer a vector space embedding for the corpora and align the source-target embeddings using their vector space centroids. A set of nearest neighbors is retrieved for every source text unit from the target style based on their semantic similarity in the aligned vector space. Multiple subsets of nearest neighbors are extracted and summarized using a language model with a reranking procedure to optimize the style transfer quality, which achieves state-of-the-art results on automatic evaluation metrics.