Digital work on classical Chinese poetry has tried to move beyond literal explanation by adding background, scene, and multimodal support. Cao et al. built a interoperable data model for classical Chinese poetry [8]. Shih et al. used context-aware mobile guidance for the appreciation of scenic poetry [25]. Chao et al. used augmented reality to support motivation in poetry learning [9]. Li et al. created a virtual learning space for ancient Chinese poetry learning [23]. Wang et al. showed that emotional design can influence multimedia learning and appreciation of Chinese poetry [26]. These studies show the value of contextual and interactive support. Still, they leave open a question: how should poems, background notes, place, imagery, and comparison be organised so that learners can return to them during interpretation?

Adaptive hypermedia describes systems that adapt presentation or navigation in linked material [5]. In educational settings, these techniques have been used to tailor explanation, recommend next steps, and organise learner-specific routes through complex content [6]. A review of educational hypermedia shows that non-linear access can support learning when tasks and structure remain clear [11]. Hypertext reading studies make a similar point. Link selection affects comprehension [24]. Weak structure can make networked material harder to follow [2]. For our study, the lesson is clear: poetry learning systems need not only rich material, but also clear organisation, guided movement, and opportunities to return to earlier material.

AI-enabled adaptive learning systems adjust content, feedback, or sequence with AI methods [19]. Generative AI can also produce short explanations and learning paths on demand [12]. Adaptive interface research suggests that adaptive features need to be discoverable, intuitive, and integrated into the core learning loop if users are to perceive their value [18]. Explainable AI research in education makes a similar point about visible sources and understandable explanations [20]. This matters for classical Chinese poetry because contextual knowledge is central to interpretation [7]. There is still little design knowledge on how GenAI-supported adaptive hypermedia should combine linked context, adaptive explanation, and learner control in this domain.

The education interviews focused on how classical Chinese poetry is currently taught, what adult learners are able to gain from it, where common learning barriers arise, what kinds of linked background materials are useful, what hypermedia has been used for current teaching and learning, and what opportunities or limits experts saw in generative AI. These interviews were intended to clarify the learning priorities that GenAI-supported adaptive hypermedia should support and the design tensions that such a system would need to address.

The second interview focused on how linked materials might be structured, how explanation depth and navigation could vary across learners, how sources should be shown, what roles AI agents or interface elements should play, and what technical constraints would shape implementation. These interviews were intended to ground the design considerations in current interaction, system, and implementation concerns rather than pedagogical aims alone.

All interviews were conducted online in a one-to-one format, lasted between 30 and 60 minutes, and were audio-recorded with consent. Participants first received an information sheet and were free to stop at any time. Interviews were conducted in Mandarin, transcribed, anonymised, and analysed using reflexive thematic analysis [3, 4]. Two bilingual members of the research team read the transcripts closely, generated initial codes, compared interpretations, and iteratively refined candidate themes. NVivo was used to support organisation and retrieval. In line with reflexive thematic analysis, we did not calculate inter-rater reliability.

The education interviews yielded six themes covering learning barriers, learning pathways, multimedia and linked materials, AI roles, learning value, and adaptive hypermedia design considerations. The interviews with GenAI-supported adaptive hypermedia design and development experts yielded five themes covering knowledge reliability, agent and interface design, adaptive scaffolding, contextual framing, and implementation constraints. Full code lists and definition entries appear in Table 2 and Table 3.

Table 2: Codebook A of classical Chinese poetry education experts

Table 3: Codebook B of GenAI-supported adaptive hypermedia design and development experts

Connected knowledge rather than isolated facts. Educational experts emphasised the importance of making explicit links to rhetorical devices, imagery, historical context, and aesthetic appreciation. Because poetry is the expression of poets situated in specific personal and historical contexts, learners are expected to develop cultural and emotional experiences: “feel the rhythm and mood.” (P9) and “a way to grasp history.” (P11). P12 called for a background and biographical “knowledge network” before formal study (identified under the theme Learning barriers). They also emphasised the importance of constructing interconnected knowledge networks during the learning process itself: “relationship networks” among poets of the same period (P18), and “different poets may have visited the same place but produced different works, which provides a basis for comparison and helps learners to establish a contextual framework.” (P19). Taken together, these accounts indicate that learners are suggested to be supported in building structures that connect history, geography, literature, and other domains, rather than focusing on isolated poems in a vacuum. Such cross-cutting connections, in turn, can deepen their understanding of individual poems and situate them within broader cultural and spatial contexts.

Cultural and emotional connection. In this study, cultural -emotional experience denotes the affective dimension of cultural engagement, encompassing feelings of resonance, belonging, nostalgia, pride, and aesthetic pleasure that arise when encountering and interpreting cultural forms. Such experiences often transcend purely cognitive understanding, shaping identity(P18) and fostering deeper cultural appreciation (P11 and 17). In our analysis, they are closely linked to codes such as Cultural identity and confidence, Cultural and historical heritage, and Cultural‑industry literacy within the theme “Learning potentials: aims and value”. These experiences frequently emerge from the creation of scenes and atmospheres that serve as windows and lenses for cultivating interest in, and motivation to engage with, poetry and classical Chinese culture more broadly. Moreover, the cultural and emotional impact of learning classical Chinese poetry often manifests with a temporal delay. The theory of narrative transportation [15] describes this delayed resonance, that when learners are “carried” by an evocative scene, imagery and narrative help them form coherent interpretations that can later be reactivated in lived experience. As P19 explained, “Later on, I realised that some of the poems I had learned earlier had indeed influenced me. In certain situations, like being in York and far away from home, particular verses would resurface in my mind, and I felt they reflected and described my inner state of mind.”

Sustained interest. Experts argued that adult learners need short, achievable, and varied forms of interaction that fit fragmented study time. Interviewees report that multimedia is “more direct and interactive… [and] boosts interest” (P9, P10), and that reframing familiar content can “shake up old views” (P18). They also stressed learner fit: “Match the audience; start with concrete hooks, even pop bits, to spark interest” (code: Audience fit and cognitive activation, P18). Interactivity was treated as a practical lever: “It helps teach unfamiliar ideas and builds memory through interaction.” (code: Advantages of interactive pathways, P15). This shows that interest develops when interactions trigger attention, and sustained attention often points to the linked materials for poetry understanding, fit learning needs, and satisfy interest.

AI answers may be fast, but may still miss key background. Both groups saw value in AI as a fast and structured source of support, but they also worried that model outputs may be unreliable, incomplete, or weakly grounded in the poem's historical and cultural setting. AI experts also described cases in which learners move quickly across periods, poets, or topics and the system responds smoothly but without enough grounding. One participant put it plainly: “Some outputs aren't reliable” (P5). At the same time, experts also valued AI's ability to provide “structured knowledge” (P8). This means in the adaptive hypermedia, staying tied to inspectable and traceable source units rather than free-form answers alone.

Context and experience gap. Education experts described multiple obstacles, including limited life experience, the linguistic distance of Classical Chinese, weak affective engagement, and difficulty transferring what has been learned into a new situation. As P12 noted, “Classical word order and meanings differ a lot,” while P13 observed that many learners “can't really feel the emotion.” These accounts suggest that poetry-learning systems need to do more than answer factual questions. They need to reconstruct enough cultural and historical context for learners to understand why a scene or image matters.

Pathway from perception to resonance. Interviewees also revealed a common and recommended pathway in the learning of classical Chinese poetry, summarised under the theme Learning pathway from perception to resonance: initial engagement through reading and recitation; translation and comprehension of Classical Chinese into modern Chinese; situating or evoking personal experiences; extracting and comparing imagery; and ultimately achieving affective–aesthetic resonance. Within this sequence, contextualisation plays a pivotal role: it was described not only as a way to “enhance retention” (P19), but also as a “scaffold” (P4, P5) that links linguistic, experiential, and emotional aspects of learning. These accounts suggest that hypermedia is a promising medium for supporting this pathway because it enables learners to move between a poem and linked contextual materials—such as individual lines, imagery notes, place notes, biography notes, and archive entries—while keeping those materials anchored to the learner's current interpretive focus.

Building on our findings that learners benefit from explicit, linked knowledge, cultural-emotional resonance, and sustained interest, yet face fragmented access and difficulty achieving situated understanding, we organise the considerations below around two complementary parts of an adaptive hypermedia system: GenAI as adaptive presentation and guidance, and hypermedia structure as linked contextual traversal.

4.3.1 Role A. Generative AI as adaptive presentation and guidance. Provide layered explanations through progressive disclosure. Experts wanted support that begins with a short gloss, then expands to background, and only then moves to deeper interpretation. For adult learners, this layered structure can keep information manageable while still allowing deeper study when needed. This form of adaptive presentation helps adjust depth without forcing every learner through the same full explanation [5, 21].

Keep background in small linked units and make sources visible. Rather than presenting background as one long passage, the system should store small units such as biography notes, place notes, imagery notes, and historical notes. AI output should show where a background fact comes from at the moment of use, so that learners can inspect and revisit it. This directly answers interviewees’ requests for a visible knowledge network rather than a single opaque explanation.

Use learner state to vary hint depth and suggested next steps. Learners differ in prior knowledge and confidence. Support should therefore be sensitive to current performance, giving fuller worked hints to novices and shorter cues to more advanced learners. When needed, the system can also suggest the next most useful note, scene, or task rather than leave the learner with one undifferentiated set of options. Knowledge tracing offers one way to model changing learner state [10]. Worked steps can then fade as competence grows [1]. Timely task-specific feedback also supports learning [16].

4.3.2 Role B. Hypermedia as linked contextual traversal. Build meaning through scene, place, and background. To convert background knowledge into felt understanding, the system should reconstruct a place, time, and situation that makes the poem interpretable. This does not require one long linear story. Instead, it can be achieved through a short path across linked cues, such as place, weather, figure, object, and background note. This supports interviewees’ emphasis on scene-based evocation and supports interpretation through structured traversal rather than through explanation alone.

Use interaction steps to make links explicit. Interaction should encourage learners to connect wording, imagery, place, and emotion rather than answer isolated quiz items. This can be done through exploration, collection, matching, reconstruction, and short comparison steps. The point is not to add interaction for its own sake, but to make relations among linked materials visible and usable during learning.

Support short comparison trails across poems. Experts often mentioned comparison across poems, particularly when poems share a place or image but differ in feeling. Design should therefore support brief comparison and reflection trails that help learners articulate differences and build a broader knowledge structure. These short trails can connect one poem to another through shared imagery, location, or mood, while still keeping the learner in control of interpretation.

Keep paths short, resumable, and under learner control. For adult learners, sessions should be short enough to pause and resume easily. Progress indicators, small steps, and immediate feedback matter because they reduce friction and make return visits more likely. Adaptive support should guide, but it should not lock the learner into one fixed route. Learners need room to reopen earlier notes, compare readings, and decide when to go deeper.

Two interface screenshots from Poetictok. The left screenshot shows map-based exploration with poets and locations. The right screenshot shows a poem reconstruction task with phrase selection and ordering.

Exploration in Poetictok begins from a stylised map of China that functions as an overview of the linked poetic knowledge space. The map marks historical place names and the poems associated with them, allowing learners either to follow a poet across multiple locations and stages of life or to explore a single place through poems written there by different poets. At the level of an individual poem, learners move through a short place-linked interface related to the target poem. Along the way, they encounter historical figures, trigger brief dialogues to collect small background units related to place, event, imagery, and poet biography if needed (Figure 2, a, map-based exploration with poets and locations). The design goal is to build contextual understanding through linked notes rather than through one long explanation. To check how many parts of the poem the learners have understood, we have a poem reconstruction phase(Figure 2, b) after the exploration phase. Learners complete a poem through wording or phrase selection and ordering tasks. The goal is to assess the imagery and interpretation in a structured way while linking them back to the background gathered earlier. Each poem session follows this two-part repeatable flow to fit learners’ needs, while still connecting context and text.

Four panels showing the four AI agent types in Poetictok. The Narrative agent explains the current situation, the Instructional agent provides hints during poem tasks, the Exploratory agent answers short object- or context-related questions, and the Social agent supports short in-character conversation.

Poetictok design centres on two core features: linked contextual traversal to help learners build poetic scene and background knowledge, and generative AI-powered adaptive presentation and guidance. This design directly responds to two core tensions surfaced in our interviews. First, learners need connected poetic knowledge, yet access to relevant information is often fragmented and untimely. Second, in self-paced adult learning, learners frequently struggle to engage with poems without emotional support and lived context. Figure 4 presents Poetictok’s end-to-end learner workflow, showing how the underlying LLM and knowledge base support four AI agents across the learning journey.

Poetictok's technical architecture and end-to-end learner workflow, showing how the underlying LLM and knowledge base support four AI agents across the six stages of the learning journey

Linked background units instead of one-time exposition. Rather than presenting background as a long reading passage, the exploration phase breaks context into small, inspectable encounters. At each stop, the learner completes a lightweight action, such as opening a landmark, talking to a historical figure, or collecting an object cue. Each action yields an archive card that can be reopened later, such as a short biography note, a place note, or an imagery note. These cards function as persistent linked units that remain available during later interpretation.

Guided interpretation through reconstruction tasks. To translate interviewees’ emphasis on imagery identification and expression into interaction, the reconstruction phase uses two task types: selecting a missing character or phrase from close alternatives, and ordering phrase blocks under a time limit. The selection task asks learners to distinguish between near-synonyms and link wording to imagery and mood. The ordering task tests understanding of the poem and rehearses structure in a way that is visible and checkable. In both cases, brief feedback points back to the relevant line feature or background unit instead of giving a detached final answer.

Role-bounded GenAI agents as adaptive presentation and guidance. AI functions are separated into four agents (Figure 3), each invoked at a predictable moment. The Narrative agent links the current location to the poem's situation. The Exploratory agent answers short “what is this?” questions and points learners to relevant archive cards. The Instructional agent provides hints during poem tasks and adjusts hint depth based on recent performance. The Social agent supports short in-character conversations for engagement and reflection, while remaining separate from factual guidance. This role separation helps learners predict what kind of support they will receive.

Visible sources and bounded context. To address concerns about unreliable AI output, the interface provides lightweight source cues. When the system presents a background fact, it can attach a small source label that points back to an in-system archive entry. When it offers interpretation, it signals that more than one reading may be possible rather than presenting a single fixed answer. In this way, AI support remains tied to linked source units and stays open to learner inspection.