Hate speech detection has provoked a vast amount of attention and research in recent decades with the growing prevalence of hateful contents. When the texts contain distinct semantic features, such as hateful words or expressions, it is easy to achieve satisfying results, but how to accurately detect implicit hate speech in nuanced or contextual forms remains a challenge. Despite the fine-tuning pre-trained language models (PLMs), such as BERT and HateBERT, can achieve fairy good results on various downstream tasks including implicit hate speech detection, some recent studies find one of its critical challenges is the significant gap of objective forms in pretraining and fine-tuning, which restricts taking full advantage of knowledge in PLMs. In this work, we propose a novel Implicit Hate Speech detection method via Soft Prompt-tuning (IHS-SP), which can further stimulate the rich knowledge distributed in PLMs to serve implicit detection tasks better. We employ three different strategies to expand the label words space for exploring the true intention behind original texts, and the integration of these strategies is used in the final prompt-tuning. In contrast to existing methods that either introduce crafted template but are time-consuming and labor-intensive, or automatic prompt generation methods cannot achieve satisfactory performance, our method considers both the template generation and detection performance to construct prompts for addressing the linguistic diversity in implicit hate speech. Despite being automatic, experimental results show that our method achieved more desirable performance even than the crafted template methods on implicit hate speech detection. In particular, our approach significantly improves the model’s sensitivity to subtle and context-dependent expressions, reducing false negatives and enhancing interpretability over baseline methods.