BrainChat: Decoding Semantic Information from fMRI using Vision-language Pretrained Models