codertimo / paper-log

읽어야 하는 논문들을 관리하고, 읽은 논문들의 기록을 남기는 공간
31 stars 5 forks source link

Domain Adaptive Text Style Transfer #11

Open codertimo opened 4 years ago

codertimo commented 4 years ago

어떤 내용의 논문인가요? 👋

Abstract (요약) 🕵🏻‍♂️

Text style transfer without parallel data has achieved some practical success. However, in the scenario where less data is available, these methods may yield poor performance. In this paper, we examine domain adaptation for text style transfer to leverage massively available data from other domains. These data may demonstrate domain shift, which impedes the benefits of utilizing such data for training. To address this challenge, we propose simple yet effective domain adaptive text style transfer models, enabling domain-adaptive information exchange. The proposed models presumably learn from the source domain to: (i) distinguish stylized information and generic content information; (ii) maximally preserve content information; and (iii) adaptively transfer the styles in a domain-aware manner. We evaluate the proposed models on two style transfer tasks (sentiment and formality) over multiple target domains where only limited non-parallel data is available. Extensive experiments demonstrate the effectiveness of the proposed model compared to the baselines.

이 논문을 읽어서 무엇을 배울 수 있는지 알려주세요! 🤔

레퍼런스의 URL을 알려주세요! 🔗

https://arxiv.org/abs/1908.09395