Memory-enhanced text style transfer with dynamic style learning and calibration
Fuqiang Lin,Yiping Song,Zhiliang Tian,Wangqun Chen,Diwen Dong,Bo Liu
DOI: https://doi.org/10.1007/s11432-022-3726-0
2024-04-02
Science China Information Sciences
Abstract:Text style transfer aims to rephrase a sentence to match the desired style while retaining the original content. As a controllable text generation task, mainstream approaches use content-independent style embedding as control variables to guide stylistic generation. Nonetheless, stylistic properties are context-sensitive even under the same style. For example, "delicious" and "helpful" convey positive sentiments, although they are more likely to describe food and people, respectively. Therefore, desired style signals must vary with the content. To this end, we propose a memory-enhanced transfer method, which learns fine-grained style representation concerning content to assist transfer. Rather than employing static style embedding or latent variables, our method abstracts linguistic characteristics from training corpora and memorizes subdivided content with the corresponding style representations. The style signal is dynamically retrieved from memory using the content as a query, providing a more expressive and flexible latent style space. To address the imbalance between quantity and quality in different content, we further introduce a calibration method to augment memory construction by modeling the relationship between candidate styles. Experimental results obtained using three benchmark datasets confirm the superior performance of our model compared to competitive approaches. The evaluation metrics and case study also indicate that our model can generate diverse stylistic phrases matching context.
computer science, information systems,engineering, electrical & electronic