Convolutional neural networks have been used for low-light image enhancement (LLIE), but cannot effectively capture global context relationships. Transformers are good at capturing long-range dependencies, but involve the costly computation of the attention map. In addition, existing transformers-based  LLIE methods pay less attention to frequency-domain components, and neglect the role of multi-modal information in image enhancement. To overcome these issues, we propose a new LLIE network based on Two-stage Frequency-domain Decomposition with Text Guidance (TFDTG-Net). Specifically, we decompose the LLIE task into two stages and separately handle the amplitude and phase information  in the frequency domain. In both stages, we embed a frequency-domain attention module to efficiently estimate the attention map of transformers for global context modeling. Meanwhile, we introduce a text guidance module to leverage the vision-language model to extract semantic information for image enhancement. As such, we comprehensively consider spatial-frequency features, local-global context relationships and semantic information in our TFDTG-Net. The experimental results quantitatively and qualitatively demonstrate that TFDTG-Net is effective for LLIE compared with other methods.