VietQuill: Quality-Controlled Paraphrase Generation for Vietnamese Language

Abstract

Paraphrase generation is a core task in natural language processing, yet most existing methods lack explicit control over linguistic properties, limiting their applicability in user-driven settings. This challenge is particularly pronounced for Vietnamese, a low-resource analytic language with distinctive lexical and syntactic characteristics. We propose VietQuill, a quality-controlled paraphrase generation framework that enables fine-grained control over lexical overlap, syntactic similarity, and semantic preservation through explicit control signals. The framework incorporates Vietnamese-aware quality estimation and is supported by ViPQC, a corpus of approximately 300K Vietnamese sentence pairs automatically annotated with continuous lexical, syntactic, and semantic quality scores. Experiments on three Vietnamese paraphrase benchmarks (ViQP, ViSP, and vnPara) demonstrate consistent responsiveness to the proposed control signals, with semantic control improving semantic similarity by approximately 15–25% and syntactic control increasing structural variation by 40–60%. Compared with its ViT5-base backbone, VietQuill improves BERTScore by 11.2% on ViQP and 4.3% on ViSP, while increasing structural diversity by 82.6% on ViQP. LLM-based evaluation further shows improvements of 11.9% in meaning preservation and 15.1% in diversity over the backbone. These findings demonstrate the effectiveness of language-aware quality modeling for controllable paraphrase generation in Vietnamese. The source code is publicly available at https://github.com/ngwgsang/vietquill.