<p>The rapid proliferation of digital content piracy poses significant challenges to the protection of intellectual property. Digital watermarking provides an effective countermeasure by embedding imperceptible identifiers into textual content to enable ownership verification. With the advancement of large language models such as BERT, semantic-based watermarking has emerged as a promising research direction. However, semantic-based watermarking often introduces unnatural or erroneous sentence structures, compromising text quality. To address this limitation, we propose a three-stage textual watermarking framework consisting of keyword extraction, candidate synonym generation, and semantic filtering. Specifically, BERT-based masked language modeling is used to generate context-aware substitution candidates for selected keywords. A multi-level semantic filtering strategy, combining word-level similarity, context-aware similarity, and sentence-level similarity, is then applied to retain only semantically reliable candidates and ensure stable watermark embedding and extraction. Experimental results demonstrate that the proposed method not only maintains semantic accuracy and readability but also achieves superior performance in six evaluation metrics: semantic relevance, embedding capacity, text quality, extraction accuracy, recoverability, and imperceptibility. Quantitatively, the proposed method achieves BPW values of 0.124, 0.106, and 0.112 on WikiText-2, AgNews, and IMDB, respectively, and improves sentence-level embedding capacity by more than 15% compared with the existing method, while maintaining competitive text quality.</p>