This study focuses on the implementation of a web-based application prototype designed to extend the visualization of modern and contemporary Korean poetry from quantitative statistics to sensory and interpretive dimensions, based on the Korean Poetry Emotion Mapping (KPoEM) dataset. The KPoEM dataset (Lim et al. 2025) consists of a total of 7,622 emotion annotation instances : 7,007 are line-level ; 615 are work-level annotations.  Five annotators performed emotion labeling at both the line and work levels according to the 44 emotion categories defined in the Korean Online That-gul Emotions (KOTE) dataset taxonomy (Jeon et al. 2024), allowing up to ten co-assigned emotions in order to reflect the multilayered and contextual nature of poetic emotion. This study integrates the KPoEM dataset and the outputs of its corresponding emotion classification model into a web-based prototype, within which Co-Reading is implemented as a functional module that enables human–AI collaborative interpretation of poetic emotion (Lim 2025b).

The Drama Corpora Project (Fischer et al. 2019)

The Drama Corpora Project (DraCor) is an open research infrastructure that provides richly annotated dramatic corpora and tools for computational literary analysis. See https://dracor.org

demonstrates how TEI-encoded literary texts can be transformed into an open visualization platform that enhances scholarly reproducibility and accessibility. Building on this approach, the present study extends this framework toward the visualization of emotion data in poetry. In particular, the finely segmented line-based structure of poetry provides a more suitable analytical environment for visualizing the hierarchical organization of emotions and their micro-level variations. Accordingly, the KPoEM visualization prototype is designed to explore and support digital humanities openness and reproducibility by experimentally providing an interactive visualization web interface in which poetry and emotion function as the primary dimensions for user-driven, sensory engagement and exploration.

Drucker (2014) emphasizes that data visualization interfaces function not merely as efficient tools but as mediating spaces that actively shape users’ modes of thinking and directions of discourse. In this regard, the present study proposes a web-based approach to visualizing emotions in modern Korean poetry grounded in the KPoEM dataset. The technical realization of this study consists of three reading modules (see Figure 1).

Figure 1. Conceptual Architecture of the KPoEM Prototype for Close Reading, Distant Reading, and Co-Reading

An illustrated interface prototype based on the architecture shown in Figure 1 is available at:[https://drive.google.com/file/d/1onRJACtuTAE6nwNyiJy68fAlvvcYlDB5/view?usp=drive_link ]

First, the Close Reading module is designed to enable detailed exploration of line-level emotion annotations. Users can browse emotion labels for individual lines through tag and filter-based interfaces and compare differences among annotators’ interpretations for selected sample works by five poets—Kim Sowol, Yun Dong-ju, Yi Sang, Im Hwa, and Han Yong-un. In this module, line-level emotional tagging results are presented as interactive textual and visual elements, allowing users to intuitively grasp the flow of emotion within each poem. Through this process, static poetic texts are reconfigured into interactive structures for scholarly archiving and visual emotion exploration.

Second, the Distant Reading module provides a statistical and visual overview of emotional structures across poets and works. For the poets included in KPoEM, emotion trends are aggregated at both the poet and work levels and visualized particularly through heatmaps and SHAP (SHapley Additive exPlanations)-based word clouds, enabling users to grasp distributions, intensities, contrasts, and key lexical elements. SHAP

SHAP (SHapley Additive exPlanations) is an explainability framework for interpreting the outputs of machine learning models based on Shapley values from cooperative game theory. It provides feature-level contribution scores that help explain individual model predictions. See https://github.com/shap/shap

is employed as a basic explainability tool to derive lexical elements that contribute to the emotion predictions generated by the KPoEM-based classification model (Lim et al. 2025). These visualizations serve as a foundation for exploring emotional tendencies across poets and relational networks among the 44 emotion categories, thereby enabling a macro-level quantitative interpretation of large-scale poetic emotion patterns.

Third, the Co-Reading module presents an interactive environment in which users interpret the emotions of input poetic texts collaboratively with artificial intelligence and experience them through color-based visualization. In this module, colors are  determined through a combination of a human-curated emotion–color dataset and computational processes performed by AI. The input poem is first analyzed by the KPoEM emotion classification model with the resulting emotion categories mapped to corresponding hues based on the Korean Color Emotion Mapping (KCoEM) dataset (Lim 2025a). Subsequently, a large language model (LLM) applies context engineering (Mei et al. 2025) to select a single image adjective from the 94-adjective list of the I.R.I. Color Image Scale. This adjective probabilistically modulates brightness and saturation within the I.R.I. color scheme, producing a final two-color palette. For the purposes of this study, these Co-Reading outputs are precomputed and mapped onto the interface as part of an experimental prototype implementation. In this way, poetic emotion is reconstructed through hue, saturation, and brightness, allowing users to experience a multimodal mapping among language, emotion, and color through human–AI co-interpretation.

While Close Reading and Distant Reading (Moretti 2013) have long served as representative methodologies for micro-level textual analysis and macro-level statistical exploration in literary and computational literary studies, this study draws on these traditions by proposing Co-Reading, which positions AI as a co-reader that translates poetic emotion into color. This approach complements efficiency-oriented data visualization practices (Drucker 2014) by reconfiguring literary interpretation as a sensory and affective experience. The color transformation process realized through Co-Reading does not merely function as a visualization procedure but experimentally explores its potential as a mediating interface through which humans and AI jointly interpret poetry and share emotional resonance.

Accordingly, the KPoEM web prototype is designed as an integrated web interface environment in which Close Reading, Distant Reading, and Co-Reading are experienced in a unified manner, centered on the qualitative value of poetic emotion. Built upon a multilayered, modular web architecture, the prototype allows for intuitive exploration of emotion data through visualizations such as emotion distributions, keyword-based representations, and color-based visualizations. Through these interactions, users move beyond merely viewing data to experiencing how multiple layers of interpretation emerge through processes of data generation and analysis.

Based on this architectural design, the KPoEM prototype brings together three modes of reading—Close Reading, Distant Reading, and Co-Reading—within a unified web environment, functioning as an experimental web interface in which literary texts, emotion data, and AI models interact. This prototype allows researchers to iteratively examine and further extend processes of data production, analysis, and interpretation in a reproducible manner, while also offering non-specialist users new sensory and interpretive ways to explore poetry and emotion.