Style Rewriter: A Three-Stage Approach for Stylistic Generation of Open-domain Conversational Responses
Nana Zhu, Caihai Zhu, Qingfu Zhu, Yuanxing Liu · Data Intelligence · 2025
Generating conversational responses in an appropriate style is a crucial but challenging task. Due to the lack of parallel post and stylized response data, existing studies on stylized response gen- eration are either less style-specific or less content-relevant. To address this issue, in this paper, we propose a Style Rewriter model to bridge the semantic gap between conversational data and non-parallel style sentences. The proposed model attends to the latent space of conversation and style and generates a stylized response via a fusion decoder with a rewriting scheme. Experimen- tal results show that the proposed Style Rewriter model outperforms the competitive approaches on automatic and human evaluation metrics in a benchmark dataset.