使用隱空間校準實現非平行文本風格轉移

Abstract

風格轉換是人工智慧研究領域內一個熱門的議題。不過，在自然語言處理研究上的非平行文本風格轉換受限於一些既有的問題使得研究仍非常具有挑戰性。其中一個問題在於在文字中區別風格(style) 和內文(content) 是非常困難的，而且現有的模型並沒有適當的機制去在文本中保持與風格不相關的內文。另一個缺點在於主要的方式僅專注於兩種類別的風格轉換, 因此在處理多類別風格轉移時，需要建立成對的模型在任意兩種方面。本論文提出一個自編碼模型以及一個統一的生成對抗網路用以處理多類別間的風格轉換，同時設計在隱空間(latent space) 的正規化損失函數(regularization loss) 用於保持內文的特徵。實驗結果顯示我們的模型能夠達到更加多樣且一般性的風格轉移。

Style transfer is a popular topic in artificial intelligence research. However, several issues of non-parallel style transfer in natural language processing remain challenging. One is that separating styles with content of texts is difficult, and current models have no proper mechanism to remain style-unrelated content of texts. Another drawback is that main approaches focus on transfer styles between two classes, since pairwise models among any two aspects should be built when dealing with transfer of multiple classes. In this work, we propose an auto-encoder model including a unified generative adversarial network for multi-class style transfer and a designed regularization losses on latent space to preserve content representation. Empirical results show that our models achieve more diverse and general style transfer.