Current methods for multimodal sentiment analysis task usually overlook the early interaction between acoustic and vision modalities before fusing with text modality. To overcome this problem, this paper proposed a model with a BAG-LSTM module, which consists of a bimodal adaption gate (BAG) and two modality-specific LSTM units, and other two attention modules focusing on improving the input and output sequences of BAG-LSTM. Through early interaction between acoustic and vision modalities, BAG calculates two shift vectors for these two modalities, respectively, to stimulate interactive sentimental information and facilitate implicit alignment of acoustic and vision modalities. The comprehensive experiments demonstrated that the proposed method has achieved the new state-of-the-art (SOTA) results. Our code is publicly available at https://github.com/mlmmwym/BAG-LSTM .