Transformer couplet generation model based on joint word segmentation

Chinese couplet is an art form with concise and strict antithesis. Given a sentence (namely an antecedent clause), people reply with another sentence (namely a subsequent clause) equal in length. It has made the machine-generated couplets readable and consistent for the result of existing deep learning methods. However, the generation process is character based which is different from the word-based use of Chinese. It is still a gap in semantic consistency between the results and the artificial creation. In this paper, a joint word segmentation of couplet is designed for the symmetry of couplet word segmentation results, and a word-based transformer couplet generation model is built to improve the semantic coherence of subsequent clauses generated. Moreover, word count information and part of speech information are added into the word vectors to provide the features of generating. Finally, the effectiveness of our model was confirmed in BLEU, Perplexity and human evaluation.