Introduction

Poetry is a defining feature of Chinese vernacular fiction in the late imperial and early Republican periods (1368–1949). Canonical novels such as Honglou meng 紅樓夢 incorporate verses into their narratives to shape characters, storylines, and guide interpretation. Recent scholarship describes vernacular fiction as freely traversing classical and vernacular registers, with poetry serving as a nexus of aesthetic and narrative effects (Yue, 2024). A central but understudied dimension of this topic is poetic perspective, i.e., whether a poem is voiced by the narrator or by a character, which illustrates how fiction negotiates voice, emotion, and authority.

This attention to poetic perspective builds on long-standing work at the intersection of oral performance and written fiction. Studies on “spoken-and-sung literature” 講唱文學 show how prosimetric traditions integrated spoken narrative with chanted verse (Hana 1973, 1981; Børdahl 1996; Chan 1997; Xu 2025). Verse functioned as a structural and emotive device, and its narrator-voiced use in novels echoes the storyteller persona central to these traditions. At the same time, novels used character-voiced poems as part of storytelling, contributing to the rise of psychological realism (Shang 2003; Epstein 2001).

Additionally, poetic perspectives shaped moral and interpretive framing of vernacular fiction. Narrator-spoken poems often reinforce didactic commentary, while character-spoken poems introduce emotions, thought, or self-perception (Yu 2001; Rolston 1997). Examining “who speaks the poem” thus offers a productive lens for understanding narrative dynamics and stylistic practices across vernacular fiction.

To examine these perspectives at scale, we propose a systematic empirical approach for analyzing how narrator- and character-voiced poems are distributed across periods using large datasets. This approach is grounded in digital humanities and computational literary studies, where large-scale analysis complements close reading by identifying patterns, distributions, and historical tendencies that are difficult to observe through canonical examples alone (Jockers, 2013; Underwood, 2019). More recently, scholars have begun to explore the use of large language models (LLMs) for literary research, including corpus annotation, classification, thematic analysis, narrative analysis, and the study of characters, genres, and stylistic patterns (Jannidis et al., 2025; Pichler et al., 2026). By modeling narrator–character perspective using LLMs, this study treats poetic voice as a computable yet interpretive feature, linking large-scale analysis to questions of voice, agency, and narrative mediation in Chinese vernacular fiction.

In previous work, we annotated 339 embedded poems by position, perspective, and content and classified them using LLMs. Of the dimensions annotated, narrator–character perspective yielded the clearest signal: inter-annotator agreement reached 0.89, and model performance achieved 0.75 accuracy and an F1 score of 0.73. This suggests the feasibility of a large-scale analysis of poetic perspectives in this study.

Data and Methods

From 144 works of Chinese vernacular fiction (most of which date to the Qing dynasty, see Figure 1), we were able to extract 6,233 poems and these poems are mainly found in titles from Ming and Qing dynasties (Figure 2). By examining the relationship between the length of each novel and the number of poems found within it, we can observe a weak positive correlation between them. (Figure 3).

Figure 1. Number of novels from different periods in our corpus.

Figure 2. Number of extracted poems across time.

Figure 3. Relationship between text length and number of found poems in each novel.

Using the 339 annotated poems from the previous study, we evaluated six open-source LLMs on classifying poetry perspective to find the best model. We then used that model to classify the perspectives of the extracted 6233 poems. For all classification tasks, we used the open-source tool Ollama and the models published on it.

See: https://ollama.com/.

“The following Chinese text contains a poem. The beginning and end of the poem are marked with 'p_s' and 'p_e' respectively. Determine whether the poem is composed or recounted from the first-person perspective of a character in the story or a third-person perspective of the author or a storyteller. If the former, please answer with ‘character,’ if the latter, please answer with ’narrator.’ You must choose only one option as your answer. You do not have to explain your answer. Here is the text with the poem:”

As presented in Table 1, although llama3.3:70b, qwen3:4b and qwen3:8b achieved similar performance, qwen3:4b is much smaller than the other two models. Therefore, we used qwen3:4b in the next step.

Table 1. Comparison of classification results using six LLMs.

Findings

Figure 4 illustrates how character- and narrator-voiced poems are distributed in each book. Ming works exhibit the highest and most variable counts per title, indicating that inserted poems, especially narrator-voiced poems, were used more extensively in this period. Several Ming titles contain unusually large numbers of narrator-voiced poems, such as Complete Biography of Eunuch Wei Zhongxian 魏阉全傳 and Romance of Emperor Yang of Sui 隋炀帝艳史. By contrast, Qing fiction works, although more numerous in the corpus, generally contain fewer inserted poems per title. Character and narrator counts remain generally similar, though narrator perspectives appear slightly more often. One exception is Book of Beauties 美人书, a Qing “scholar and beauty” 才子佳人 fiction, which contains the largest number of character-voiced poems in the corpus (n = 162). In ROC, works show a still narrower distribution, suggesting a reduced and more standardized use of inserted poems. Pairwise comparisons indicate significant declines from Ming to Qing for both character-voiced poems (0.01 < p ≤ 0.05) and narrator-voiced poems (0.001 < p ≤ 0.01), as well as a significant decline from Ming to Republican-era works for narrator-voiced poems (0.001 < p ≤ 0.01). No significant differences are observed between character- and narrator-voiced poems within the same period.

Figure 4. Distribution of character- and narrator-voiced inserted poems in each book across Ming, Qing, Republican-era, and unknown fiction.

P-value annotation legend: ns: 0.05 < p <= 1.00; *: 0.01 < p <= 0.05; **: 0.001 < p <= 0.01.

Overall, the results reveal a consistent pattern across time: Ming vernacular fiction shows the most frequent insertion of verses and narrator perspectives predominate across all periods. These results offer important insights into the shifting yet continuous functions of poetic insertions in Chinese vernacular fiction. The greater density of poetic insertion in Ming works supports longstanding claims that early vernacular fiction remained deeply anchored in oral‐performative traditions, where verse operated as a structural and rhetorical device integral to storytelling practice. The subsequent decline in verse frequency across Qing and ROC titles corresponds with the genre’s progressive literarization, as fiction became increasingly shaped by written conventions and a more text-oriented reading public. Yet the persistent dominance of narrator-voiced poems across all periods points to a durable narrative logic: regardless of stylistic evolution, verse continued to serve primarily as an instrument of authorial or storyteller mediation, rather than a vehicle for character description. This pattern underscores the extent to which prosimetric fiction retained a storyteller-centered mode of narration, even as it adapted to new literary, commercial, and cultural conditions (Johnson et al., 1986).

Future work can extend this quantitative framework to finer-grained analyses, such as identifying authorial or regional stylistic fingerprints to assess whether certain writers use narrator-voiced poems for moral commentary–patterns that may clarify authorship or networks of influence. Likewise, comparing genre-specific traditions, from shuhua-style storytelling fiction to chivalric novels and romances, may clarify how different narrative forms shaped patterns of verse usage and voice.