Although the existing stego-poems that meet basic thematic and structural requirements have been generated, they have not demonstrated consistent semantic emotions, which reduces the perceptual-imperceptibility of the stego-poems and their transmission security. To balance the tradeoff between hiding capacity and concealment in stego-text, a steganography model for the automatic generation of poems with controllable themes and emotions is constructed in this paper, and a visual testing system based on PyQt5 is developed. The poem generation model adopts bidirectional encoder representations from transformers (BERT) word embedding and an attention mechanism to achieve a full understanding of the dynamitic and context information in the input sentences. To our knowledge, emotion analysis is introduced for the first time in generative model. As a result, stego-poems are generated under the guidance of keywords and the constraints of metrical/emotional/mutual information. The experimental results show that the algorithm proposed in this paper not only ensures that the theme of the generated poem is consistent and the emotion is controllable but also effectively hides and correctly extracts secret information. The proposed algorithm improves the embedding rates of the state-of-the-art steganography algorithms for generating quatrains by 9% and 3%, respectively, while also achieving higher security.