Abstract

Travel literature is a unique form of hypertext that extends beyond its medium into real-world physical exploration. While conventional computational methods can easily extract surface-level ontological entities (e.g., locations, dates), the deeper epistemic and judgmental subtext that guides the discovery and evaluation of places has traditionally required close human reading. To address this gap, this paper explores the capacity of Large Language Models (LLMs) to complement classical computing by performing interpretive analysis of deeper textual characteristics. We present a proof-of-concept case study using an LLM and Retrieval-Augmented Generation (RAG) on a corpus of over 600 digitised travel literature. By employing LLMs as hypertext engines to structure and apply analytical frameworks in collaboration with scholars, we demonstrate the feasibility of using generative AI for the interpretative study of unstructured corpora. This contribution provides actionable insights into system architecture, AI roles, and interaction design, advancing the broader integration of LLM technologies within the hypertext paradigm and Digital Humanities research infrastructures.

1 Introduction

Travel literature is a special type of hypertext that combines non-linear reading with physical exploration and first-hand experience of places of both the author and readers. Travel diaries report a specific traversal as experienced (or as narrated) by an author who encounters places, communities and events. Travel guides offer a structured interface for places: what, where, how, when and why. Despite the more narrative or structured form, both types of travel literature are realised through a unique experience, a long tail after reading that stretches to the reader's planning, travelling, voting, and experiencing places firsthand. While neglected as a research object by the community, the “travel” metaphor plays a central role in conceptualising hypertext, e.g., as a bridge between geographical and intellectual spaces [9].

To address this gap, we provide a first formal framing of travel literature as hypertext that escapes its medium to offer a first-hand experience of physical space, people, and activities. By this definition, travel literature is particularly hard to conceptualise, catalogue, explore and study as a research object. Even in classic hypertext, the personal experience is hard to predict and replicate, but the confines of human-medium interaction can facilitate, e.g., logs and other workarounds to create evidence of traversals. Experiences elicited by travel literature, however, are by nature outside the medium, spread across the physical world or other digital media (travel booking, reviews, documentaries) used in preparation for or as a proxy for a first-hand visit (elicited knowledge or physical exploration).

As argued, the most interesting aspect is the first-hand experience of how the literature comes to life through the reader's interaction with places. Hence, the focus can only be on elements that emerge from the text itself. On the surface, travel literature is most commonly characterised by referred ontological entities: locations, language, events, people, dates, timing, and costs. Below the surface, they offer epistemic guidance about how to discover (when and where) and engage (how) with places and communities. On a deeper level, travel guides provide judgments and an approach to evaluation that is often the key to a positive engagement with foreign cultures, foods, and lifestyles. The surface level is the most analysed and easiest to extract using conventional computational means, e.g., ontologies, Name Entity Recognition techniques, and Gazeteers. The epistemic and judgmental layers are left to close reading by human researchers, as they are harder to define and, hence, to extract in objective terms that could be used to characterise and enable the exploration of travel literature.

To this end, this contribution explores the use of LLMs to determine whether they can complement classical computing to perform an interpretative analysis of travel literature. Specifically, we would like to know whether an LLM is sensitive to the epistemological and judgmental characteristics of travel literature, as defined by a user.

This question is addressed through a case study examining a corpus of 600+ travel literature examples, digitised and made available to AI via a Retrieval-Augmented Generation (RAG) approach using a vector-based database. LLMs are used as hypertext engines, supporting (1) the structuring of the epistemic and judgmental frameworks of analysis through a conversation with scholars, and then (2) running such frameworks on a corpus to curate a list of resources through a justification and extracts that explain their relevance. The case study was developed as a proof-of-concept to inform and guide the development of a research infrastructure.

The proof-of-concept allowed the collection of evidence on the feasibility of this approach in a real-life setting: the study of unstructured, novel corpora. While limited in its scope and generalisability, the presented study is self-contained and sufficient to inform decisions regarding the system architecture, the types of AI roles and tasks, and the interaction design for researcher-readers with AI and, through the AI, with the corpora. Those considerations and decisions are provided to contribute to the ongoing community effort to integrate LLM technologies within the hypertext paradigm.

The state of the art focuses on travel and place-based hypertext applications, clarifying the above-mentioned research gaps. The case study is presented thereafter, including details about the methodology and experiments. Then, we list the results of the experiments and discuss the viability of using LLMs for interpretative research and for specifically analysing travel literature. Finally, we provide a brief conclusion on the implications for research infrastructures in the Digital Humanities and on potential AI-powered hypertext applications for travel literature.

2 Travel Metaphors, Travel Literature & Hypertext

Travel metaphors are foundational to the hypertext community. Early researchers relied on concepts like traversal, paths, and guided tours to counter the disorientation of feeling “lost in hyperspace” [5, 6]. By borrowing constructs from physical travel, the community grounded abstract digital architectures in familiar acts of wayfinding and spatial navigation [4, 10, 14].

Yet, a striking paradox emerges. While the field has spent decades using spatial traversal to explain digital interactions, it has largely ignored actual travel literature. Texts explicitly designed to dictate real-world physical and cultural traversal remain overlooked as a primary hypertextual research object. Unpacking how hypertext currently treats spatial and locative data exposes this specific conceptual gap.

2.1 Spatial Hypertext

Spatial hypertext emerged as a solution to the “tyranny of the link”. Foundational systems allowed relationships to surface organically through spatial parsing [10]. The 2D plane acted as a workspace for information triage, enabling users to organise unstructured data without defining formal links [11]. Over time, researchers integrated spatio-temporal parsing to capture the dynamic sequence of human reasoning [13]. Most recently, the field has abstracted this concept into a method of inquiry [2], framing spatial structures as cognitive scaffolding for AI-driven discovery [1].

Despite this evolution, a critical conceptual gap remains. Spatial hypertext continues to treat space as a container for intellectual arrangement rather than a medium for physical traversal. While these frameworks successfully model how scholars map information clusters, they lack a foundation for the cultural and physical journeys inherent to travel literature. In a journey, the relationship between locations depends on narrative arcs and physical transitions. The current workspace metaphor fundamentally fails to capture these elements.

2.2 Location-focused & Ubiquoitous Hypertext Systems

As mobile computing advanced, the hypertext community expanded its focus from screen-bound spatiality to physical geography, giving rise to locative hypermedia. Researchers recognised that a user's physical location could act as a dynamic context for information delivery. Systems demonstrated how narrative and data could be anchored to GPS coordinates, effectively turning the physical world into a hypertextual canvas [12, 15]. In these location-aware systems, the reader's physical traversal triggers the delivery of digital nodes, merging the act of reading with the act of walking.

Locative hypermedia represent the closest intersection of hypertext with physical travel, while still retaining key differences from how traditional travel literature approaches place exploration. Locative systems generally require simultaneous physical presence and technological mediation: digital resources overlay the physical space in real time. A clear example of augmented reality technologies, from smart glasses to digital maps, literally overlaps digital assets with digital representations of the physical space (e.g., pictures, videos or cartography). In contrast, travel literature operates asynchronously, acting as an epistemic guide and a proxy for the journey, read in preparation for, or in reflection upon, a physical visit. While the community has engineered systems that react to location, it has spent less time analysing the inherent hypertextuality of static texts that describe and evaluate those locations.

2.3 Authoring Styles and E-literature Inspired by Travel

Electronic literature frequently co-opts the aesthetics of travel. e.g., positioning the reader as a tourist, framing interaction as a journey of discovery [8] or deliberately “inducing disorientation” to make readers receptive to difficult ideas [3]. Similarly, locative narratives actively exploit the tension between reading and physical movement [7].

Crucially, these creative works deploy travel as a metaphor for fictional or experimental ends. Travel literature, conversely, is pragmatic and epistemological. It provides actionable frameworks for engaging with real foreign cultures and physical environments. While e-literature proves hypertext excels at simulating the feeling of a journey, the community has yet to apply these analytical frameworks to the non-fictional texts that actually direct human travel.

2.4 Travel Literature as Hypertext

As argued so far, while sharing common elements, travel literature is quite unique and specific. In the introduction, we mentioned how travel literature escapes the medium as its intent is realised outside, in the reader's exploration of the real world, engaging first-hand with communities, places and cultures. While first-hand exploration may never occur, it is the nature of travel literature, hence a definition of travel literature as hypertext must account for it. Travel literature is a structured representation of a traveller's experience (the author) designed to facilitate a reader's own discovery through an evaluative and epistemic framework, manifested as a journey-oriented relational network of places. Worth highlighting from this formulation are the following core characteristics:

Projection, not a model that articulates a phenomenological (rather than an ontological) representation of places, as traversals that connect affinities, travellers’ needs and goals over objective, impersonal relationships\

Evaluation of places that, through omissions (more than inclusions), exemplify an approach to relevance and judgment of what is worth, and, e.g., how to spend travel resources\

Guide to discovering places using the structured presentation to clarify how to engage and interact with them outside the medium.\

In conclusion, travel literature re-orients structures, connections and journey cues outside as an overarching structure to classic hypertextual media features.

3 Case Study – Interactive Explorations for Scholarly Research

In the tradition of Hypertext, the study combines theoretical investigation of Travel Literature (see Section 2.4) with experimental, exploratory work. This last was carried out through a case study aimed at assessing the feasibility of a novel research infrastructure to support the study of a corpus of travel literature, newly acquired and yet to be curated.

To test this framework, we conducted a case study using a secure, locally hosted Gemma 3 (27B) integrated with Qwen 3 (4B) for vector embeddings. This setup allowed researchers to process internal institutional corpora safely while maintaining strict control over intellectual property and computational resources.

The opportunity for the study was provided by the recent digitisation of a significant corpus of more than 600 historic travel diaries and guides, which tell the story of centuries of visits to Rome and visitors from all around Europe. The text corpus in Markdown format, with Wikidata hyperlinks on entity mentions, is made available by the Bibliotheca Hertziana – Max Planck Institute for Art History. It is worth clarifying that, in a typical Digital Humanities research, this corpus would have been made available through a research case study, hence curated in light of a specific research agenda and analytical framework. It is worth stressing that the exploration of a novel, non-curated corpus is quite uncommon and difficult to approach: indexing systems rely on metadata often produced as a by-product of a first curatorial work, annotating, e.g., places, events, periods, and other significant components of the work. Indeed, it is the research curatorial work that, in essence, builds the implicit and explicit hypertextual structure that enables the exploration of research archives.

This peculiar scenario, a novel AI infrastructure, and a virgin dataset provided the background for the case study, which aimed to evaluate whether a human-AI interactive exploration would provide an alternative approach to a corpus lacking structure. To this end, we carried out the following activities:

baseline analysis of AI's understanding of travel literature focused on clarifying what notions and tendencies the LLM manifests toward this specific genre\

development of a meta-theory of travel literature for AI in the form of structured prompts that identify and clarify their core elements, to be used as a blueprint for explorations\

interpretative challenges exploring AI's ability to answer non-trivial questions about the scope, perspective and narrative qualities of the works\

interactive instantiation of a research analytical framework structuring the process into two stages, to extract from researchers their unique point of view, to be run on the corpora to produce research briefs, curating a selection of relevant resources.\

Steps 1 and 2 were preparatory to Step 3, which shed light on the ability of LLM to interpret the user-researcher analytical framework. Step 4 was instead aimed at collecting evidence about how to structure, integrate and scale up an LLM-driven exploration as part of the infrastructure.

4 Results

The experimental evidence gathered from the case study concerned four primary areas.

Interpretative Depth: Uncovering the Unspoken. The LLM demonstrated a significant ability for interpretative analysis, moving beyond surface-level data to identify the underlying reasons and structures within travel guides. It successfully established links between the ‘what’, ‘how’, and ‘why’ of a text, advancing hypotheses on implicit elements.

Bias Detection, the model identified implicit prejudices and ideological influences\

Rhetorical Sensitivity, it captured nuanced emotional states—such as disillusionment or disdain—conveyed through subtle adjectives and allusions rather than explicit statements.\

When asked to detect sources with implicit disdain concerning Rome, the AI proved able to justify its answer. For instance, in the “Un viaggio a Roma senza vedere il papa” of Giovanni Faldella, the author's disillusionment is detected through the use of specific expressions:

the panteon being “angustiato” [squeezed] in between two very common roads\

“mancomale” [luckily, better than expected] about Rome still being there and standing\

Rome appears a “fastello” [disorganised mess] of architectonic elements\

Methodological Flexibility and the AI Guide. The LLM proved highly adaptable to the researcher's interpretative lens to structure corpus-wide analyses.

User lens and framework, the AI structured an analysis framework based on the researcher's interests\

Proactive Inquiry, When prompted via the system instructions, the AI acted as a research guide by suggesting follow-up questions to refine the study's focus.\

AI was asked about foreign travellers and their prejudices towards 19th-century Rome and, in repose, it produced a set of questions to clarify the investigation line, such as travellers from a specific origin (Germany, England, US), what we meant by prejudice in this context or if we were interested in negative or all sorts of prejudices.

Comparative Criticism and Resource Selection. AI was able to perform a comparative analysis, identifying unique, recurring themes across the corpus as a justification for the selection of sources in response to the researcher's prompt.

Comparative criticism, AI identifies the peculiarity of the sources and compares them, highlighting their main differences\

Source Justification, the model coherently justified why certain resources were more suitable than others for specific queries\

Reveal patterns, the AI detects recurring themes throughout the corpus\

Operational Limitations and Complacency. Despite its strengths, two significant limitations were observed:

Complacency Trap, the LLM occasionally exhibited "complacency", attempting to satisfy user requests even at the cost of incoherence or arbitrary interpretation. This led to quantitative errors, such as repeatedly suggesting the same source when a specific number of examples is requested\

Instruction Adherence: the model did not consistently follow system prompts for proactive questioning; without explicit, repeated requests, it tended to limit its analysis to the user's immediate points rather than expanding the inquiry.\

AI was asked to find a travel route along the Via Appia Antica from outside Rome to its centre: it found a correct source but narrated leaving rather than arriving in Rome (the opposite of centre out).

5 Lessons Learned & Recommendations

From the experiment described above it emerged that generative AI may not necessarily offer a direct grounding for every sensible appraisal theory of travel literature, rather, a theory in and of itself which results from engineered contexts. Expert input is still required in the feedback loop to inject actual art historical research questions, thereby controlling the construction of said appraisal theory.

Implicit Bias Detection. The AI successfully identifies underlying reasons and ideological influences within the sources. It can articulate exactly how these implicit factors shape the author's adopted perspective.\

Explicit Prompting Requirements. To extract these hidden purposes, the researcher must explicitly instruct the AI to look for them. Prompts must direct the model to analyse specific rhetorical devices and demand clear justifications for its interpretative choices.\

The aforementioned issues determine a significantly greater problem in scope than one that can be addressed by prompt engineering alone. We argue, in fact, that the factors at play transcend the configurability that "traditional" combinations of system and user prompts can provide. It is therefore proposed that the appraisal methodology of travel literature closely follow an agentic paradigm, yet allowing an element of context engineering, in that:

System prompts are used to define the behaviour of an agentic pipeline that, first, engages with the user in defining the context and scope of the research and, then, carries out the analysis. The system prompts define the structure of the output as well, serving as a form of curation of the selected sources (the research briefs).\

Context definition is arguably the cornerstone of the activity, lying in the ability to extract and make explicit the researcher's vision, needs, and assumptions for the research. This is the most difficult part, as often researchers do not come with a clear idea of what they are looking for and a limited understanding of AI's lack of understanding of human phenomena, dynamics, and experience, which travel is a very hard case of.\

6 Conclusions

Our contribution focuses on a clear gap in the study and understanding of travel literature. Despite the centrality of travel-related metaphors in hypertext, no previous work attempted to frame travel literature as a hypertext. Beyond the hypertext community, the study of travel literature is generally underserved, with effort concentrated on the most common, objective elements that, at best, serve as a basic search index. The contributions presented – the framing of travel literature (see Section 2.4) and the analysis of the exploratory work (see Section 5) – addressed both issues, providing a path forward.

The work was very limited: one corpus with strict geographical limitations, but, from a Humanities perspective, significantly broad due to the wide range of years, authors, and their provenance. The number of experiments was also very limited in number, but drastically novel in terms of the challenge posed by the questions. The proof of concept used an internal, bespoke AI infrastructure using a local model, developed to serve a large national network of research labs and, hence, realistic in terms of capabilities, resources and constraints/limitations. Overall, for the purpose of our aim, this work was sufficient to draw the conclusions presented and justify future developments.

Lastly, travel literature is a particularly hard genre for AI due to the intrinsic relationship with physical, embodied experience. In this view, an argument can be made about how to revise or complement their structure and content to facilitate AI's understanding to support their effective use in supporting human activities. This structure and content should clarify, e.g., (a) what travel literature is and where it comes from, (b) what the limits of a place-based description are, and (c) what the scope and purpose of the different types of travel literature are and how those change throughout cultures, places, and media.

Source

ACM full-text HTML

ACM Digital Library

Imported from ACM’s structured HTML source. ACM Reference Format: Virginia Orlando, Alessandro Adamou, and Alessio Antonini. 2026. Travel Literature as Hypertext: Definition and Case Study of AI-enabled Interactive Explorations. In 37th ACM Conference on Hypertext (HT '26), September 14--18, 2026, London, United Kingdom. ACM, New York, NY, USA 5 Pages. https://doi.org/10.1145/3800935.3830882

References

[1] Claus Atzenbeck, Eelco Herder, and Daniel Roßner. 2023. Breaking the routine: spatial hypertext concepts for active decision making in recommender systems. New Review of Hypermedia and Multimedia 29, 1 (2023), 1–35. https://doi.org/10.1080/13614568.2023.2170474

[2] Claus Atzenbeck and Peter J. Nürnberg. 2019. Hypertext as Method. In Proceedings of the 30th ACM Conference on Hypertext and Social Media: Tear Down The Wall (Hof, Germany) (HT ’19). Association for Computing Machinery, New York, NY, USA, 29–38. https://doi.org/10.1145/3342220.3343669

[3] Mark Bernstein. 1998. Hypertext gardens: Delightful vistas. Eastgate Systems Web site, accessed at www. eastgate. com/garden, February (1998).

[4] Mark Bernstein. 1998. Patterns of Hypertext. In Proceedings of the Ninth ACM Conference on Hypertext and Hypermedia: Links, Objects, Time and Space—Structure in Hypermedia Systems (Pittsburgh, Pennsylvania, USA) (HYPERTEXT ’98). Association for Computing Machinery, New York, NY, USA, 21–29. https://doi.org/10.1145/276627.276630

[5] Jeff Conklin. 1987. Hypertext: An Introduction and Survey. Computer 20, 9 (1987), 17–41. https://doi.org/10.1109/MC.1987.1663693

[6] Deborah M. Edwards and Lynda Hardman. 1989. ’Lost in Hyperspace’: Cognitive Mapping and Navigation in a Hypertext Environment. In Hypertext: Theory into Practice, Ray McAleese (Ed.). Intellect Books, 105–125.

[7] Jeremy Hight. 2003. Narrative archaeology. StreetnotesSummer (2003).

[8] Michael Joyce. 1988. Siren shapes: Exploratory and constructive hypertexts. Academic computing 3, 4 (1988), 10–14.

[9] George P Landow. 1991. HyperText: the convergence of contemporary critical theory and technology (parallax: re-visions of culture and society series). Johns Hopkins University Press.

[10] Catherine C. Marshall and Frank M. Shipman. 1995. Spatial Hypertext: Designing for Change. Commun. ACM 38, 8 (1995), 88–97. https://doi.org/10.1145/208344.208350

[11] Catherine C. Marshall and Frank M. Shipman. 1997. Spatial Hypertext and the Practice of Information Triage. In Proceedings of the Eighth ACM Conference on Hypertext (HT ’97). ACM, 124–133. https://doi.org/10.1145/267437.267451

[12] David E Millard and Hugh C Davis. 2000. Navigating spaces: the semantics of cross domain interoperability. In International Workshop on Structural Computing. Springer, 129–139.

[13] Thomas Schedel and Claus Atzenbeck. 2016. Spatio-Temporal Parsing in Spatial Hypermedia. In Proceedings of the 27th ACM Conference on Hypertext and Hypermedia (Halifax, Nova Scotia, Canada) (HT ’16). Association for Computing Machinery, New York, NY, USA, 149–157. https://doi.org/10.1145/2914586.2914596

[14] Randall H. Trigg. 1988. Guided Tours and Tabletops: Tools for Communicating in a Hypertext Environment. In Proceedings of the Second ACM Conference on Hypertext (HT ’89). ACM, 398–414. https://doi.org/10.1145/58566.59299 Originally presented at HT ’87.

[15] Mark J Weal, Danius T Michaelides, Kevin Page, David C De Roure, Eloise Monger, and Mary Gobbi. 2011. Semantic annotation of ubiquitous learning environments. IEEE Transactions on Learning Technologies 5, 2 (2011), 143–156.