Lower Couplet Generation Model Based on Double-Layers Attention

Abstract:
Couplets are the artistic treasures of Chinese traditional culture. The generation of lower couplets based on deep learning has important research value. The existing methods generally increased the aesthetic feeling of couplets by fusing auxiliary information, they focused on the overall meaning of the upper and lower couplets and splitted words into characters in the word embedding step, which splits the overall semantics of words. To solve these problems, we propose a new method of generating lower couplet based on double-layers attention mechanism. Firstly, we use "word" embedding to preserve the word semantics, and adopt the three-layer text structure of "corpus-phrase-word" to ensure the semantic correspondence and same character count of each phrase in the upper and lower couplets. Then, the double-layers attention mechanism is embedded into the three-layer text structure, which focus on the key information by capturing the attention information of both phrase layer and word layer respectively. Finally, the joint learning method is used to train the model. We conducted a large number of experiments on the Chinese couplet dataset, and used the BLEU score and perplexity degree as evaluation indicators for quantitative measurement. The experimental results show that, compared with the existing methods, the proposed method has the best performance in each indicator, and the generated couplets are very similar to human couplets in semantics and form.