This presentation designs a system capable of generating a long-form science fiction novel by training a language model on the poetic prose of Han Kang, a Nobel Laureate in Literature from Korea. This revised abstract outlines the original objectives of our first attempt, discusses the challenges encountered along the way, and proposes potential solutions to address them.

Han Kang (1970—) is prominent Korean author who initially debuted as a poet. Through her delicate and sensuous prose, she has consistently addressed the oppression of minorities by the majority, the domination of women by men, and state violence against historical victims. The Swedish Academy cited her selection with the words, “her intense poetic prose that confronts historical traumas and exposes the fragility of human life,” emphasizing how she elevates these specific Korean issues into a universal sense of loss experienced by the human body and soul. Our rationale for using Han Kang’s work as input is quite simple: she has never written an SF novel. What new work might emerge when her dense, interior-focused description is projected onto an SF imagination through automatic AI generation?

Initially, we planned to generate a short story, scaling it into a novella featuring a plot twist. To produce the short stories, the framework utilizes a collaborative process between a “Story Planner” that outlines synopses and a “Scene Describer” fine-tuned on a specialized dataset of Han Kang’s works to flesh out the details. This approach yielded relatively high-quality stories, such as “Ghosts of the White City,” which successfully replicated dense emotional resonance and sensory contrast.

However, we faced two primary challenges. First, the output generated via fine-tuning was comparable to results obtained solely through prompt engineering, rendering the fine-tuning largely ineffective. Second, attempting style transfer without a clear, working definition of Han Kang’s defining poetic prose made it difficult to determine whether the stylistic characteristics of the generated text genuinely reflected her literary style.

To address these challenges, we propose an inductive methodology that derives an empirical taxonomy of Han Kang’s poetic prose directly from existing critical scholarship, supplying both the working definition our prior pipeline lacked and an evaluation rubric for AI-generated outputs.

Figure 1. Overview of our methodology for deriving an empirical taxonomy of Han Kang's poetic prose from Korean literary criticism.

(A) Data Collection. We assemble a corpus of Korean literary criticism through Model Context Protocol (MCP) servers for DBPIA and KCI, the two principal Korean academic databases. An AI agent (Claude Code) issues queries combining Han Kang’s authorship markers with poetic-prose terminology variants yielding 58 candidate papers from KCI. After automated subject-area filtering to retain literary-studies categories only, candidates are screened paper-by-paper: a reviewer reads each title, author keywords, and abstract to determine whether the paper directly and centrally analyzes the poetic-prose features of Han Kang’s prose works, discarding translation/reception studies, thematic analyses without stylistic focus, and works on Han Kang’s poetry (rather than her prose). This strict screening yields 11 papers, each of which subsequently undergoes PyPDF text extraction and DeepSeek-OCR processing for structurally faithful representation. The agent then constructs (passage, interpretation) pairs—each consisting of a verbatim Han Kang quotation and the surrounding critical analysis identifying its poetic-prose features. For a comparative robustness analysis, we additionally collected 13 papers analyzing the poetic-prose features of other Korean literary works, identified through a parallel KCI query of 1,272 candidates filtered by the same screening criterion.

(B) Analysis. Building on RELiC (Thai et al. 2022), the principal English-language precedent for paired literary-claim/quotation modeling, we encode passages and interpretations through separate Transformer encoders. For Korean text we primarily use KoSimCSE, with multilingual-E5 serving as a comparative baseline to probe how embedding-model choice affects downstream clustering and to verify that the pipeline generalizes to cross-lingual extensions in future work. Because individual critical paragraphs typically bundle several distinct claims, we first decompose interpretations into atomic style attributes via LLM prompting, following the LISA framework (Patel et al. 2023). Analysis is planned at three levels: (i) within-paper recovery, validating that a single critic’s section labels are detectable in embedding space; (ii) cross-paper alignment, mapping how critics working from divergent theoretical frameworks converge on or diverge from shared formal features; and (iii) unsupervised discovery of latent dimensions absent from any individual critic’s vocabulary. Cluster validity will be assessed via silhouette analysis, dual-space coherence between passage and interpretation embeddings, and expert review by Korean literature scholars. The corpus construction described in (A) is currently in progress; the analytical pipeline in (B) is currently being implemented, with empirical results to be reported in a subsequent paper.

In future research, we plan to apply this methodology to English literature, thereby enabling a quantitative classification of what constitutes poetic prose. Furthermore, as the application of such classifiers allows language models to generate novels in desired styles at scale, we intend to discuss the ethical implications of this automated generation and the paradigm shifts it necessitates in literary criticism.