Poetry Generation Combining Poetry Theme Labels Representations

Proceedings of Recent Advances in Natural Language Processing, pages 1246–1255 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_132 1246 Abstract Ancient Chinese poetry is the earliest literary genre that took shape in Chinese literature and has a dissemination effect, showing China's profound cultural heritage. The current work in the field of poetry generation is mainly aimed at improving the fluency and structural accuracy of words and sentences, ignoring the theme unity of poetry generation results. In order to solve this problem, this paper proposes a graph neural network poetry theme representation model based on label embedding. Based on the network representation of poetry, the topic feature representation of poetry is constructed and learned from the granularity of words. Then, the features of the poetry theme representation model are combined with the autoregressive language model to construct a theme-oriented ancient Chinese poetry generation model TLPG (Poetry Generation with Theme Label). Through experiment and evaluation by experts in related fields, the model proposed in this paper has significantly improved the topic consistency on the premise of ensuring the fluency and format accuracy of poetry. 1 Introduction Ancient Chinese poetry is a short, vivid form of literary genre, and the depictions in ancient poetry are a living reflection of the daily life of the Chinese ancestors as well as their inner prayers and expectations. In world literature, ancient Chinese poetry is also an important means of demonstrating the power of the Chinese language. It is interesting to note that although automated machine poetry compositions fall short of human beings in terms of rhyme, mood, and feeling (Zhang et al., 2023). However automated poetry generation is still worth researching. On the one hand, machine-generated ancient poems can assist students in generating a deeper understanding of poetry (Ma et al., 2023). On the other hand, the emergence of multimodal technologies has made poetry learning not only limited to texts (Wu et al., 2021) a variety of information can be utilized in the study of poetry teaching and learning. In terms of poetic format, poems of different genres and eras often have different formats (e.g., poetic style, rhyme, etc.), which involve grammatical, semantic, and phonological aspects. Current research tends to establish grammatical analysis as well as symbolic representation of phonology for ancient poems, which makes poetry often contain explicit thematic information (Yang et al., 2023). Themes in poetry are implicit textual information are important references for poetry appreciation and creation, so this paper focuses on the representation and integration of theme features in the process of poetry generation. At present, the mainstream poetry generation models are mainly based on the Transformer (Zhao et al.,2022) and generate poems in the form of autoregressive language models. However, unlike general text generation tasks, poetry is a highly structured literary genre (Li,2020). In the poetry generation task, the input part of the model is designed with a special identifier to learn the specific format of the poem (Yang et al., 2022). During the learning process this special identifier learns potential formatting information in …

Content truncated to 500 words