Abstract

Large text corpora rarely arrive with the structure a reader needs to navigate them. We present an end-to-end system that induces navigable trails through a flat collection of 323,552 quizbowl questions — short academic quiz texts, each with a canonical answer but no authored links — and serves them to learners through a deployed, multimodal study platform. An offline pipeline segments each question into sentences, embeds them with an instruction-tuned encoder, and clusters each answer's sentences into concept nodes. Rather than linking concepts by co-occurrence, we order them with optimal leaf ordering (OLO) over their embedding hierarchy, producing, for each subject and difficulty level, a single guided tour that minimizes the total dissimilarity between consecutive concepts. Each tour is at once the navigable structure and the first-exposure sequence; an FSRS-7 spaced-repetition scheduler then governs revisitation at concept granularity. The platform scaffolds each encounter along a five-level ladder of fading support, from highlighted reading with synchronized audio to audio-only recall. The pipeline and platform are the artifacts. A sampled transition audit finds 96% of trail steps land on a related concept; a formative usage observation shows recall accuracy rising with repeated spaced encounters. Establishing pedagogical efficacy remains future work.

1 Introduction

A recurring problem in hypertext research is how to derive navigable link structure from a corpus that arrives without one — and what makes automatically discovered links, as opposed to hand-authored ones, useful for navigation. Early work discovered links semi-automatically [4] and generated them from raw text with information-retrieval techniques [1, 2, 12, 30], and the question has remained live as corpora and representation methods have changed [17]. A complementary tradition organizes hyperspace not as a web of associative links but as a trail — an authored path that walks a reader through related material, from Bush's memex trails [10] to Trigg's guided tours [28]. This paper joins the two traditions: we induce the trail rather than author it, at the granularity of concepts rather than documents, using contemporary text embeddings to order a corpus's concepts into coherent guided tours in which each step lands on a closely related idea.

Our setting is competitive quizbowl, a buzzer game played across a recognizable canon of academic subjects (Section 3). Quizbowl is well-suited to hypertext induction for two reasons. First, it is large and flat: hundreds of thousands of short texts with no authored links between them. Second, its questions are pyramidal — ordered from obscure to well-known clues — and graded by difficulty, so the corpus carries a latent notion of conceptual depth that a navigable structure can expose. No single question contains the shape of a subject (which ideas sit near which); the corpus as a whole encodes it, and reading it at that scale rather than question by question is a form of Moretti's distant reading [19].

Three research questions organize the paper.

RQ1 (ordering): which traversal of the induced concept nodes produces the best trail — an order whose consecutive concepts stay related from end to end? We compare a random baseline, a nearest-neighbor greedy walk, and optimal leaf ordering (OLO).\

RQ2 (unification): can the winning ordering serve a deployed learning platform as both its navigable structure and its first-exposure sequence?\

RQ3 (revisitation): how should the platform organize revisitation so that it follows learning-science principles — spaced retrieval practice at concept granularity, a spiral climb through difficulty levels, and multimodal scaffolds that fade from highlighted, audio-synchronized reading to audio-only recall?\

The contribution is the end-to-end artifact that embodies our answers. First, an offline pipeline induces navigable concept trails — one per subject category and difficulty level — from a flat 323,552-question corpus. Second, a deployed web platform serves each trail as a multimodal, adaptive study environment. Section 5 addresses RQ1 with an ordering comparison and a sampled transition audit; RQ2 and RQ3 are answered in design by the platform (Section 4) and checked with formative usage signals. Demonstrating pedagogical efficacy would require a controlled study beyond this paper's scope.

2 Related Work

Deriving links from unlinked text is a long-standing thread in hypertext research. Bernstein's apprentice [4] proposed link discovery as a semi-automatic process; Salton et al. [26] showed that IR similarity measures could structure and decompose texts [2]; and Green [12] demonstrated that lexical-semantic similarity yields better links than raw term repetition. Wilkinson and Smeaton [30] survey this line of work, which continues at web scale in systems that learn to link text to encyclopedia articles [17]. Allan [1] also studied the typing of automatically generated links — an emphasis on link semantics we return to as future work (Section 6). Our work differs from this tradition in granularity and in form. Classical link generators connect whole documents whose terms overlap into a document-to-document web. We instead induce sub-document nodes — concepts, each a cluster of sentences that share a canonical answer — and, rather than a web of associative links, we produce a single linear trail through them (Section 3).

Producing a trail places us in the guided-tour line of hypertext [10, 28]. Trails have historically been authored by hand, though Guinan and Smeaton [13] planned guided tours dynamically over an existing hypertext's nodes and links. What has not been attempted, to our knowledge, is inducing the tour with no hypertext to plan over — discovering the nodes themselves from flat text, at concept granularity, and committing to a single global order. That induction is the central hypertextual novelty of our approach. Mechanically, ordering items so that neighbors are similar is the seriation problem; we order with optimal leaf ordering, the standard fast algorithm for linearizing a clustering dendrogram [3], carrying a technique established in clustering visualization into trail construction.

Adaptive hypermedia is the hypertext community's own tradition of guided learning: it maintains a model of the learner and adapts navigation to that model through direct guidance, link sorting, hiding, annotation, or generation [7, 8], as in the adaptively annotated electronic textbook ELM-ART [9]. The tradition presupposes a closed hyperspace whose pages, links, and domain model are authored by hand; extending adaptation to open corpora remains a recognized challenge [8]. Our pipeline approaches the challenge from the opposite direction, inducing concepts, links, and a default order from the flat corpus, while the platform's only adaptation is the per-learner review schedule.

Spaced-repetition tools (Anki, SuperMemo, Duolingo) schedule review over decks a human has already written, and adaptive tutors typically presume a prerequisite graph. Where that graph is not authored, a line of work infers pairwise prerequisite relations between concepts from text and course data [16, 20], yielding relation graphs rather than a navigable, deployed trail. Our system instead induces the study material's structure end to end: it discovers the concepts, clusters their clue sentences, orders them into the trail, and only then schedules them. The novelty is not the scheduler but that the navigable structure and the sequence it schedules are both derived automatically from a flat corpus that arrived with neither.

The platform layer draws on well-established ideas. Retrieval practice — testing as a learning event rather than a measurement — reliably improves retention [15, 25]. Modern spaced repetition schedules that practice over expanding intervals; we use FSRS-7, the current generation of the Free Spaced Repetition Scheduler [31], in the lineage of trainable review models [27] (mechanics in Section 4.2). We combine this with Bruner's spiral curriculum [6], in which core concepts are revisited at increasing depth, and Vygotsky's zone of proximal development [29], which motivates graduated scaffolding. Our contribution here is not these mechanisms but their composition over an automatically induced hypertext structure.

3 Inducing the Trail

Quizbowl is a buzzer-based academic quiz game [5] descended from the 1950s radio and television program College Bowl (licensed in Britain as University Challenge); today school, collegiate, and open tournaments run on packets written fresh each year by players and alumni, so hundreds of thousands of questions accumulate in public archives. Play alternates between two question types. A tossup is read aloud word by word, and any player on either team may interrupt — buzz — the moment they recognize the answer, with five seconds to respond; an early buzz wins the points, while a wrong one eliminates the buzzing player's team from that tossup, leaving it for the opposing team. To make early buzzing reward knowledge rather than reflexes, tossups are written pyramidally: clues run from obscure to famous, so the deepest knowledge buzzes first. A correct tossup earns the team a bonus — three linked parts of graded difficulty on one theme; after each part is read, the team confers and has seven seconds to answer.

3.1 Corpus

We scraped 323,552 such questions (174,715 tossups, 148,837 bonus parts) from 653 publicly available tournaments via QBReader [22]. Each question carries one of 13 top-level categories (e.g. History, Literature, Science) and one of eleven difficulty levels, from 0 (ungraded) to 10 (open, the hardest); History, Literature, and Science anchor 61% of the corpus. The corpus is flat — no authored links between questions — but not structureless in the way that matters: because tossups are pyramidal and tournaments are difficulty-graded, the same answer recurs across thousands of packets at many depths.

Pyramidality is easiest to see in an example:

This process occurs in the zone of proximal development in a model by Lev Vygotsky. Edward Thorndike's law of effect influenced a theory of this process tested with () rats in the Skinner Box. A ringing bell causing dogs to salivate exemplified a type of this process for Ivan Pavlov. For 10 points, name this process in which experience produces new behavior. Answer: learning

2023 SHOW-ME More (difficulty 3)

The early clue (Vygotsky's ZPD) rewards deeper knowledge; the late clues (Pavlov, then the bare definition) are giveaways. A bonus instead steps through one tight conceptual neighborhood. The pipeline recovers exactly such neighborhoods as clusters, and the trail visits them in sequence:

Its introduction proposes “analysis into units,” like analyzing the individual molecules of water. For 10 points each: [10h] Name this book criticizing Piaget's theory of the two title concepts in children.Answer: Thought and Language [10m] This Soviet psychologist wrote that book and demarcated the Zone of Proximal Development.Answer: Lev Vygotsky [10e] Vygotsky held that this Bruner-named support helps children expand their ZPD.Answer: scaffolding

2011 ACF Nationals (difficulty 9)

3.2 From Sentences to Nodes

Quizbowl text is hostile to rule-based tokenizers: pronunciation guides (“RILL-kuh” for Rilke), bracketed power markers (“()”, before which a buzz earns extra points), stage directions (“[emphasize]”), and long parentheticals routinely break punctuation rules. We therefore split each question into sentences with SaT-12l-sm, a neural punctuation-agnostic segmenter [18], after a normalization pass that strips brackets and boilerplate. Each sentence is concatenated with its answer before embedding, so the vector encodes the complete piece of information — the clue and what it points to — and the same fact can cluster with restatements of it in questions that ask for something else. Sentences are encoded with Qwen3-Embedding-8B, an instruction-tuned encoder producing 4096-dimensional vectors [23], yielding a 1.89M × 4096 matrix stored in float16.

Answerlines are written inconsistently across authors — “Tony Hoare” may appear as “Sir Charles Antony Richard Hoare” or “C. A. R. Hoare” — so an answerline cannot serve as a key. For each answerline we instead build a tolerant regular expression and collect every sentence whose answerline or text matches it, then cluster the collected sentence embeddings with average-linkage agglomerative clustering over cosine distance. The distance threshold of 0.20 was chosen by sweeping values and inspecting the resulting cluster sizes. Each resulting sentence cluster is a concept node: a set of clues that refer to the same answer in semantically coherent ways. Because collection ranges over the whole corpus, a cluster is kept only if it contains a sentence from the target category and spans at least two questions; singletons are dropped as noise. Finally, answer nodes whose sentence sets overlap heavily are merged by single-link absorption at Jaccard ≥ 0.8, collapsing residual answer variants into one node. Each concept node is filed at the difficulty level given by the 25th percentile of its source questions’ difficulties.

3.3 Ordering Concepts into a Trail

Given the concept nodes of one category or subcategory at one difficulty level, we induce a single navigable trail through them. We represent each concept by the centroid of its sentence embeddings, build an agglomerative hierarchy over those centroids with the same average-linkage, cosine-distance recipe, and apply optimal leaf ordering (OLO) to the resulting dendrogram [3]. OLO flips the dendrogram's internal nodes to minimize the total dissimilarity between adjacent leaves — the best linear order consistent with the hierarchy — yielding a path that stays locally smooth from end to end. The trail is thus a graph traversal in the literal sense: a path over the concept nodes, with successive nodes joined by similarity, that visits each concept exactly once. The tour is treated as a cycle and rotated so the trail opens on the category's most frequently asked concept. For scale, the three per-level trails audited in Section 5 comprise 1,692 concept nodes.

We deliberately order by embeddings rather than by cross-concept co-occurrence — how often one concept's answer appears as a clue in another's question. Co-occurrence yields a dense, untyped web (on the order of forty neighbors per concept) whose links are real but indiscriminate, mixing tight relations with tangential ones. The OLO trail instead commits to a single, locally coherent successor at each step; Section 5 compares OLO against random and nearest-neighbor-greedy orderings of the same concepts and audits the transitions it produces. The order the learner navigates is thus the same order in which concepts are first encountered (Section 4); the co-occurrence graph is kept for visualizing a category's structure, but plays no role in traversal.

3.4 Salient-Phrase Highlighting

Within a concept, recurring stock clues often pinpoint the moment a player should recognize the answer. We surface them by extracting frequent n-grams (bigrams through quadgrams) per concept and ranking them by pointwise mutual information [11] weighted by corpus rarity, which favors phrases distinctive to a concept over generic ones; common and quizbowl-specific stopwords (“for 10 points”, “this author”) are stripped first. The top-ranked phrases are stored with the concept and highlighted in the reading interface (Section 4.4). In the null pointers concept, for instance, independently written clues converge on Tony Hoare's 2009 apology, and the shared phrase is highlighted wherever it recurs:

[ ... ] In 2009, Hoare apologized for inventing pointers that have this property, calling it a “billion dollar mistake.” [ ... ] Answer: null pointer

2021 ACF Nationals Qualifier (difficulty 8)

[ ... ] In 2009, Sir Anthony Hoare claimed that his invention of this concept was a “billion dollar mistake” that has “led to innumerable errors, vulnerabilities, and system crashes.” [ ... ] Answer: null pointers

2018 PACE NSC (difficulty 5)

The reader learns the exact wording on which a prepared player buzzes; highlighting thus turns each concept into a quickly legible unit. Figure 1 shows the reading interface in use.

Figure 1: The deployed platform's Sentences traversal page, reading the Photography trail at difficulty 1. The current clue is revealed word by word at the top; completed concepts stack beneath it with salient phrases highlighted (Section 3.4 ) and their source tournament, difficulty, and type. The sidebar exposes difficulty, spaced-repetition, and scaffolding-mode controls.

4 The Platform

4.1 Implementation and scale

All computationally heavy stages run offline. The pipeline embeds the corpus into the ≈ 15 GB matrix of Section 3, clusters the concepts, and constructs the OLO trails, with clustering and ordering completing in minutes per category. The serving side is an ordinary web application: a thin client and an API backed by a document store (Figure 2). The store holds the questions, the concept clusters, the rendered audio of Section 4.3, user accounts, and per-interaction activity logs. The FSRS optimizer is the only custom component, implemented as a standalone PyTorch HTTP service. It runs outside the serving process, so its memory footprint and update cadence never block the request path.

4.2 Traversing the trail

A learner walks the OLO trail of their chosen category and difficulty level (Section 3.3). The client presents one concept at a time (Figure 1), its clues highlighted as in Section 3.4, and exposes the trail directly: the learner advances and backtracks at sentence or question granularity, skips the rest of a concept, or undoes back to the previous one. Controls work by button, keyboard, or hardware media keys, so the audio-only rung (Section 4.4) remains operable from a phone's lock screen. At any point the learner may buzz: the reading pauses and a typed answer is checked against the answerline by a fuzzy matcher, mirroring competition play. The learner's position in every category–difficulty trail is persisted server-side, so trails are resumable across sessions, and progress can be reset per level or wholesale.

The sequence a learner encounters over time is governed by FSRS-7 [31]. FSRS models, for each item, the probability that the learner still recalls it as a function of elapsed time, and schedules the next review when that probability decays to a target retention; successful reviews flatten the subsequent decay, so intervals lengthen as a concept is mastered. Critically, a spaced-repetition card is keyed at the concept (cluster) level, not the individual question: all clues that agglomerated into one node share a single card. The scheduler therefore revisits concepts at a learner-specific cadence, rather than their surface realizations. Outcomes map to FSRS grades — a wrong buzz rates again, a correct buzz good, and advancing without needing to buzz easy — and every interaction (buzz, correctness, response latency, difficulty) is logged and fed to the optimizer of Section 4.1, which refits the learner's parameters every 100 reviews. Within a session, the client interleaves new trail material with due reviews from earlier difficulty levels, so a session at difficulty 3 resurfaces level-1 and level-2 concepts as they come due; over weeks the learner climbs the ladder while the scheduler keeps mastered concepts from fading — the spiral [6], enacted without manual tracking.

Figure 2: End-to-end architecture. The offline pipeline induces the trail and audio; the platform serves it and closes a feedback loop: learner interactions are logged, an out-of-process FSRS-7 optimizer fits per-user parameters, and these drive the next scheduling pass.

4.3 Audio and alignment

For the audio modalities, an offline stage renders every question and answer to speech with a self-hosted neural text-to-speech server (Chatterbox [24]), using fixed decoding parameters so the voice is consistent across the corpus. A forced-alignment pass (stable-whisper [14]) then assigns word-level timestamps, which let the client reveal highlighted text in lock-step with the spoken audio at the most-scaffolded level (Section 4.4). The audio served by the platform comprises 1.1 million low-bitrate Opus clips, roughly 20 GB and 2,000 hours of speech.

4.4 Graduated modalities

The platform operationalizes graduated scaffolding, in the spirit of Vygotsky's zone of proximal development [29], as a five-level ladder composed from three flags: whether audio is played, whether the question text is hidden, and whether high-information n-grams are visually highlighted. At maximum scaffolding (level 0) the learner sees highlighted text revealed in lock-step with synthesized audio (Section 4.3). At minimum scaffolding (level 4) the text is hidden and the learner must recall from audio alone. The same concept can be presented at any rung, and the learner descends the ladder as mastery develops.

5 Evaluation

We evaluate the system against the research questions of Section 1. RQ1 receives a quantitative ordering comparison and a human transition audit. RQ2 and RQ3 are answered chiefly by the deployed platform itself (Section 4); the evaluation adds a formative usage signal for each. For RQ2 we ask whether learners can sustain the trail as a study order, and for RQ3 whether within-system retention behaves as the scheduler intends.

5.1 RQ1: Trail ordering and coherence

We compare three traversals of the same concept nodes by the cosine similarity of consecutive concepts at each step of the walk: a random baseline, a nearest-neighbor greedy walk (each step moves to the closest unvisited concept), and OLO. Figure 3 plots the three on the Sociology category, averaged over 94 groups of 13 concepts. The greedy walk starts tightest (0.73) but strands dissimilar outliers and decays into a long-jump tail (0.57 by the final step). The random order never rises above the corpus's background similarity (≈ 0.62). OLO orders globally, never collapses, and ends as coherent as it begins (≈ 0.68). Consecutive similarity is the quantity OLO optimizes, so the comparison shows that the optimization succeeds, not that the resulting transitions read as related.

Figure 3: Consecutive-concept cosine similarity by traversal step for three orderings of the same concept nodes (Sociology; mean over 94 groups of 13 concepts). The nearest-neighbor greedy walk starts tightest but decays into a long-jump tail; the random order sits at background similarity; OLO stays coherent from end to end.

We sampled 150 transitions — adjacent concept pairs — uniformly at random from the 1,689 transitions of three induced trails (Computer Science at difficulty 2; Architecture at difficulties 2 and 3). We classified each by inspecting the two concepts and their clues: specific (a direct, nameable relation, e.g. designer–of, located–in, or invented–by), associative (genuinely related and a smooth step, but not a tight pedagogical dependency, e.g. two works of the same era), or jarring (no defensible relation). One author labelled all 150 transitions; the sample and script allow the labels to be re-derived and contested. Adjacencies between answer-variants of one concept (e.g. I. M. Pei / Ieoh Ming Pei) count as specific. This three-way split lets us report navigational smoothness and pedagogical specificity separately, rather than scoring any non-jarring step as a success.

Of the 150 sampled transitions, 73% were specific, 23% associative, and only 4% jarring. Read for navigation, the trail is sound: 96% of steps land on a genuinely related concept, and associative steps are an asset rather than a defect: they are the lateral moves that give a guided tour its continuity. Two caveats temper the pedagogical reading. First, roughly half of all adjacencies (49% of the sample) merely join answer-variants of the same concept, which cohere trivially; among the distinct-concept steps the split is 48% specific, 44% associative, and 8% jarring. These variants persist by design: the answer merge of Section 3 is deliberately conservative, because transitive merging produced runaway clusters that fused unrelated concepts, and we prefer residual duplicates to false fusions. Second, only the specific fraction corresponds to the tight conceptual dependency a learning sequence would ideally honor, and ordering by embedding similarity cannot, by construction, name which relation links two adjacent concepts. We therefore claim the induced trail is a good navigation structure and an incomplete pedagogical one — a gap we attribute to its untyped, similarity-only links and address in Section 6.

5.2 RQ2 and RQ3: Formative usage signals

Two learners used the deployed platform to study quizbowl categories of their own choosing, in self-directed sessions motivated by upcoming competition. Over three weeks they logged 12,373 interactions across 45 categories: one, a generalist, built foundations at difficulty levels 1–5 with the scheduler enabled, while the other, a Literature specialist, probed gaps at difficulty 8 without scheduling. In practice both settled on the ladder's extreme rungs (Section 4.4): full support at level 0 or audio-only recall at level 4.

Because every review is logged (Section 4), we measure recall accuracy by encounter number: the fraction of times a learner correctly recalled a concept on its k-th spaced encounter, pooled across concepts. A recall counts as correct when the answer the learner types within five seconds of buzzing matches the concept's answerline under the platform's fuzzy checker (Section 4.2). The result is a within-system retention curve.

Toward RQ3, Figure 4 pools 332 concepts: recall accuracy rises from roughly 70% on the first few encounters to 94% by the fourteenth, consistent with spaced revisitation strengthening recall. Toward RQ2, and even more tentatively, the learners found following the OLO trail easier to sustain than a randomized concept order. The smoother transitions appeared to lower the effort of each step. Both observations are promising signals only: with two non-independent participants and no control condition, they motivate a controlled study (Section 6) rather than establishing efficacy.

Figure 4: Within-system retention: recall accuracy by spaced-encounter number, pooled across 332 concepts in the deployed platform. The upward trend is the behavior FSRS-scheduled revisitation is designed to produce; it is a formative signal, not an efficacy result. Later encounter numbers pool fewer concepts, since only the longest-studied concepts reach them.

6 Limitations and Future Work

This work has four principal limitations. First, the usage observation involves only two non-independent learners, one of them an author, and the transition audit was labelled by a single author. Demonstrating learning benefit requires a controlled study with independent participants and raters, a randomized trail-vs.-shuffled condition, and a held-out retention test. We regard such a study as the essential next step. Second, the trail's links are untyped: adjacency means only “similar in embedding space.” Hypertext research has long stressed that link semantics matter [1]; recovering typed relations (designer–of or influenced–by) would let the trail express pedagogical dependencies, not just similarity, and is the most direct route to closing the navigation–pedagogy gap of Section 5. Third, a trail is strictly linear: it commits to one global path and offers no branching, so a learner cannot follow an alternative thread out of a concept the way an associative hypertext would allow. A typed, branchable graph that still admits a coherent default traversal would restore that affordance; the co-occurrence graph of Section 3.3 is one candidate substrate. Fourth, quizbowl is a favorable corpus: it is deliberately authored so that concepts carry clear, reusable clues, which likely flatters the induction pipeline. Establishing any domain-general claim requires running the same pipeline on a corpus not authored around canonical answers and re-auditing trail coherence there.

Beyond these, concept nodes carry only text and synthesized speech, which limits categories such as Fine Arts and Music that depend on images or original audio; multimodal nodes are a natural extension. The per-concept activity logs are also a substrate for downstream knowledge tracing [21].

7 Discussion and Conclusion

We presented an end-to-end hypertext system whose contributions are, to our knowledge, novel in combination. First, we induce navigable structure at the granularity of concepts rather than documents, turning a flat 323,552-question corpus into concept nodes of clue sentences grouped by answer. Second, instead of emitting an associative link web, we linearize those concepts into navigable trails via optimal leaf ordering over their embeddings — inducing from raw text the guided tours that earlier systems authored by hand or planned over an existing hypertext's links, and carrying a seriation technique from clustering visualization into automatic link generation. Third, the trails are delivered as a deployed adaptive learning environment in which a single object plays three roles at once — the navigable structure, the first-exposure sequence, and (under FSRS) the spiral-curriculum review schedule — beneath a five-level ladder of fading scaffolds. This union of automatic hypertext induction, spaced repetition, and graduated multimodal scaffolding is absent from prior link-generation, guided-tour, and flashcard systems, each of which addresses only one of these elements.

Our evaluation supports RQ1 directly and RQ2 and RQ3 formatively: OLO yields the most coherent trails, learners sustained them as a study order, and recall rose across spaced encounters. More broadly, the pipeline assumes little beyond a corpus in which concepts recur across many short texts at varied depths. Encyclopedic entries, lecture transcripts, and flashcard collections share that shape, and each could in principle be served as its own spiral curriculum. Eighty years after Bush imagined readers blazing trails through the memex [10], a corpus can begin to blaze its own.

Acknowledgments

We thank the Northwestern University Quizbowl team for the competitive context that motivated this work; the volunteer question writers whose packets constitute the corpus; the maintainers of QBReader; and the platform's other learner — you know who you are. Computation was performed on Northwestern's Quest cluster.

Source

ACM full-text HTML

ACM Digital Library

Imported from ACM’s structured HTML source. ACM Reference Format: Jacob Puthipiroj. 2026. Learning as Graph Traversal: Inducing Navigable Hypertext Structure from a Flat Corpus. In 37th ACM Conference on Hypertext (HT '26), September 14--18, 2026, London, United Kingdom. ACM, New York, NY, USA 7 Pages. https://doi.org/10.1145/3800935.3830839

References

[1] James Allan. 1996. Automatic Hypertext Link Typing. In Proceedings of the Seventh ACM Conference on Hypertext (HYPERTEXT ’96). ACM, 42–52. https://doi.org/10.1145/234828.234833

[2] James Allan. 1997. Building Hypertext Using Information Retrieval. Information Processing & Management 33, 2 (1997), 145–159. https://doi.org/10.1016/S0306-4573(96)00059-3

[3] Ziv Bar-Joseph, David K. Gifford, and Tommi S. Jaakkola. 2001. Fast Optimal Leaf Ordering for Hierarchical Clustering. Bioinformatics 17, suppl1 (2001), S22–S29. https://doi.org/10.1093/bioinformatics/17.suppl1.S22

[4] Mark Bernstein. 1990. An Apprentice That Discovers Hypertext Links. In Hypertext: Concepts, Systems and Applications (Proceedings of the First European Conference on Hypertext, ECHT ’90). Cambridge University Press, 212–223.

[5] Jordan Boyd-Graber, Shi Feng, and Pedro Rodriguez. 2018. Human-Computer Question Answering: The Case for Quizbowl. In The NIPS ’17 Competition: Building Intelligent Systems. Springer, 169–180. https://doi.org/10.1007/978-3-319-94042-79

[6] Jerome S. Bruner. 1960. The Process of Education. Harvard University Press, Cambridge, MA.

[7] Peter Brusilovsky. 1996. Methods and Techniques of Adaptive Hypermedia. User Modeling and User-Adapted Interaction 6, 2–3 (1996), 87–129. https://doi.org/10.1007/BF00143964

[8] Peter Brusilovsky. 2001. Adaptive Hypermedia. User Modeling and User-Adapted Interaction 11, 1–2 (2001), 87–110. https://doi.org/10.1023/A:1011143116306

[9] Peter Brusilovsky, Elmar Schwarz, and Gerhard Weber. 1996. ELM-ART: An Intelligent Tutoring System on World Wide Web. In Intelligent Tutoring Systems (ITS ’96)(Lecture Notes in Computer Science, Vol. 1086). Springer, 261–269. https://doi.org/10.1007/3-540-61327-7123

[10] Vannevar Bush. 1945. As We May Think. The Atlantic Monthly 176, 1 (1945), 101–108. https://www.theatlantic.com/magazine/archive/1945/07/as-we-may-think/303881/.

[11] Kenneth Ward Church and Patrick Hanks. 1989. Word Association Norms, Mutual Information, and Lexicography. In Proceedings of the 27th Annual Meeting of the Association for Computational Linguistics (ACL). 76–83. https://doi.org/10.3115/981623.981633

[12] Stephen J. Green. 1999. Building Hypertext Links by Computing Semantic Similarity. IEEE Transactions on Knowledge and Data Engineering 11, 5 (1999), 713–730. https://doi.org/10.1109/69.806932

[13] Catherine Guinan and Alan F. Smeaton. 1992. Information Retrieval from Hypertext Using Dynamically Planned Guided Tours. In Proceedings of the ACM Conference on Hypertext (ECHT ’92). ACM, Milan, Italy, 122–130. https://doi.org/10.1145/168466.168506

[14] Jian (jianfch). 2024. stable-ts: Stabilizing Timestamps for Whisper. https://github.com/jianfch/stable-ts. Forced-alignment wrapper around Whisper for word-level timestamps. Accessed 2026-01-15.

[15] Jeffrey D. Karpicke and Janell R. Blunt. 2011. Retrieval Practice Produces More Learning Than Elaborative Studying with Concept Mapping. Science 331, 6018 (2011), 772–775. https://doi.org/10.1126/science.1199327

[16] Chen Liang, Zhaohui Wu, Wenyi Huang, and C. Lee Giles. 2015. Measuring Prerequisite Relations Among Concepts. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). 1668–1674. https://doi.org/10.18653/v1/D15-1193

[17] David Milne and Ian H. Witten. 2008. Learning to Link with Wikipedia. In Proceedings of the 17th ACM Conference on Information and Knowledge Management (CIKM ’08). ACM, 509–518. https://doi.org/10.1145/1458082.1458150

[18] Benjamin Minixhofer, Jonas Pfeiffer, and Ivan Vulić. 2023. Where's the Point? Self-Supervised Multilingual Punctuation-Agnostic Sentence Segmentation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Toronto, Canada, 7215–7235. https://doi.org/10.18653/v1/2023.acl-long.398 wtpsplit library, SaT model family.

[19] Franco Moretti. 2013. Distant Reading. Verso, London.

[20] Liangming Pan, Chengjiang Li, Juanzi Li, and Jie Tang. 2017. Prerequisite Relation Learning for Concepts in MOOCs. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL). 1447–1456. https://doi.org/10.18653/v1/P17-1133

[21] Chris Piech, Jonathan Bassen, Jonathan Huang, Surya Ganguli, Mehran Sahami, Leonidas J. Guibas, and Jascha Sohl-Dickstein. 2015. Deep Knowledge Tracing. In Advances in Neural Information Processing Systems, Vol. 28. 505–513. https://arxiv.org/abs/1506.05908.

[22] QBReader. 2026. QBReader: Online Quizbowl Question Database. https://www.qbreader.org. Last accessed 2026-07-13.

[23] Qwen Team. 2025. Qwen3-Embedding-8B. https://huggingface.co/Qwen/Qwen3-Embedding-8B. Instruction-tuned 8B text encoder, 4096-d embeddings. Accessed 2026-01-15.

[24] Resemble AI. 2025. Chatterbox: Open-Source Neural Text-to-Speech. https://github.com/resemble-ai/chatterbox. Local neural TTS server. Accessed 2026-01-15.

[25] Henry L. Roediger and Jeffrey D. Karpicke. 2006. Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science 17, 3 (2006), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x

[26] Gerard Salton, James Allan, Chris Buckley, and Amit Singhal. 1994. Automatic Analysis, Theme Generation, and Summarization of Machine-Readable Texts. Science 264, 5164 (1994), 1421–1426. https://doi.org/10.1126/science.264.5164.1421

[27] Burr Settles and Brendan Meeder. 2016. A Trainable Spaced Repetition Model for Language Learning. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 1848–1858. https://doi.org/10.18653/v1/P16-1174

[28] Randall H. Trigg. 1988. Guided Tours and Tabletops: Tools for Communicating in a Hypertext Environment. In Proceedings of the 1988 ACM Conference on Computer-Supported Cooperative Work (CSCW ’88). ACM, 216–226. https://doi.org/10.1145/62266.62283

[29] Lev S. Vygotsky. 1978. Mind in Society: The Development of Higher Psychological Processes. Harvard University Press, Cambridge, MA.

[30] Ross Wilkinson and Alan F. Smeaton. 1999. Automatic Link Generation. Comput. Surveys 31, 4es, Article 27 (1999). https://doi.org/10.1145/345966.346024

[31] Jarrett Ye. 2024. Free Spaced Repetition Scheduler (FSRS). https://github.com/open-spaced-repetition/fsrs4anki. FSRS-7 spaced-repetition scheduling algorithm. Accessed 2026-01-15.