Remake/Remodel: Topic Modelling to examine shifting Archaeological Heritage Discourse

1.1 Text Mining Archaeological Heritage Multipasts

Research previously conducted within the Digital Humanities provide text mining techniques for historical corpus bodies (Murrieta-Flores & Gregory 2015). These examinations are helpful indications to the depth of unanalysed data within these texts (geospatial, linguistic, etc.; Murrieta-Flores et al. 2015). The advancement of quantitative methodologies has been concurrent with accessible software, evidenced through continued research within the Digital Humanities. These advancements provide insight into the level of digital literacy and accessibility for researchers within various subfields within the Humanities (Gad 2014). The results are advanced analyses of both the past itself and of social engagement between the past and the present in a multivocal way (Rivolta et al. 2014) while supporting the use of quantitative methodologies. Exploratory methodologies applied to humanities and social science texts (Murrieta-Flores et al. 2015) are invaluable guides for developing methods to provide a more neutral context, seen in other fields such as “critical” digital archaeological heritage management (Bonacchi 2021). Successful applications include the development of Geographical Text Analysis (GTA; Gregory et. al. 2015), improving GTA with historical texts (Porter, Atkinson & Gregory 2015), and evaluating spatial data linked to descriptive keywords (Murrieta-Flores & Gregory 2015).

Additionally, there has been research on semantic values of terms in historical texts (McGillivray 2020) and the evaluation of benefits of different techniques applied to archaeological heritage discourse (Plets, Huijnen & van Oeveren 2021). One method, Topic Modelling, is an approach in which themes and topics in a text body (i.e., corpora) are uncovered (Onan, Korukoglu & Bulut 2016). This approach is popular in many fields (Nikolenko, Kolycov & Koltsova 2017; Brookes & McEnery 2019), further gaining traction within archaeological research (Schwartz, Jardón Giner & Rubio Campillo 2024; Bellat, Tagizadeh-Mehrjardi & Scholten 2024). This paper aims to showcase a further examination of Topic Modelling as a technique applied in archaeological heritage contexts from a critical perspective. This is demonstrated by evaluating how the specific algorithm of Latent-Dirichlet Allocation is beneficial in the investigation of historiographical texts by focusing on hidden biases that academics have unknowingly incorporated within specific discourse. By analysing a corpus of academic texts pertaining to the archaeological heritage site and landscape of Stonehenge, the progression of shifting narratives associated with this site is determined within what can be considered the ‘academic gaze’ (the view from the construction of academic narrative) from a multivocality perspective (Hanson et al. 2022). A further aim is to present a methodological workflow and dataset, made openly accessible for replication.

1.2 Case Study: The Stonehenge Landscape

Today’s Stonehenge is not the Stonehenge created by past people, nor is it the Stonehenge that existed in the 1990s. Stonehenge and its ever-evolving narratives stem from new excavations, discoveries, and research, resulting in a plethora of publications providing a ‘textual timeline’ of new Stonehenge iterations. Stonehenge is ideal for exploring the evolution of constructed narratives, due to the amount of accessible textual sources. Previous investigations into the evolution of Stonehenge’s discourse were situated in exploring the evolution of archaeological thought within academic discourse (Schwartz 2023). These results highlighted the potential to use Stonehenge to explore the impact that the evolution the academic gaze has had on shifting narratives, contexts, and interpretations of the preserved past, especially as these shifts can reformulate perceptions, value, and meaning assigned to archaeological heritage landscapes (Schwartz 2025). Additionally, Stonehenge is a unique marker for standard practices in management, preservation, and usage in the commodification and authentication of archaeological heritage. Authenticity is linked with interpretations and perceptions (McIntosh & Prentice 2017), and as these shift within texts, modern relationships with the past change.

Topic Modelling is an approach to determine different themes and topics as they appear in different corpus bodies and differing iterations of the same corpus. This method was applied to academic abstracts about Stonehenge published between 1990 and 2022. The interest is in revealing potential thematic overlaps and evolutions within subsections of the same corpus. The detection of overlaps defines discussions of policies and management practices co-occurring with any excavations or projects present in the corpus. A decree from the House of Commons in 1993 declared the presentation of the Stonehenge landscape ‘as a disgrace’. This decree was used as an analytic marker to trace the potential impact on academic research. How can Topic Modelling aid in depicting shifts in Stonehenge narratives and academic multivocality, and its forthcoming impact in shaping knowledge following this decree?

3.1 Latent Dirichlet Allocation

The Latent-Dirichlet Allocation (LDA) algorithm uses a probabilistic approach to determine topics and themes within a corpus through term occurrence and frequency (Daud et al. 2010). Specifically, LDA is structured to generate iterative models representing texts in a corpus as random collections of topics that emerge as a result of probability distributions of terms (Blei, Ng & Jordan 2003). The output is that every term is assigned a weight per topic with every topic assigned a weight per record (an abstract, in our case). Every topic is the result of the model searching for patterns in term occurrences within the observed data (i.e., every word in the corpus). This algorithmic behaviour renders outputs in which multiple topics can appear within a single abstract, rather than only detecting one topic per abstracts, which allows for thematic overlap in the analysis of the corpus.

While other algorithms share behaviours, the interest was in exploring LDA’s flexibility in uncovering said topics. This functionality is the predominant reason for our interest in both utilizing this specific algorithm and exploring the benefits for this study’s context. LDA supports analysis of interdisciplinary papers where nuance is necessary, such as within archaeology. Topics are further identified by the unsupervised co-location of words generating unnamed topics and their terms for the researcher to interpret. As highlighted previously, there are a variety of Topic Modelling algorithms with overlapping features where each algorithm is selected by individual researchers based on their perspectives and topics of study. Data was stored in SQLite with the database queried into R. Abstracts were then processed to remove noise. One example is removing stop words (i.e., ‘the’, ‘and’) allowing the algorithm to run smoothly (Jelodar et al. 2019) and used with other preprocessing techniques, like stemming and lemmatization (Khyani et al. 2021). For this study, a list of predetermined English stop words were used from a list accessible in the tm R package (Feinerer 2013). A secondary, personalized stop words list was generated to remove generic words missed by the pre-existing list and non-generic words that did not add value to the results (i.e., ‘Stonehenge’, ‘Archaeology’, etc.).

The decision regarding the removal of stop words resulted from consulting previous Topic Modelling studies’ recommendations. Stemming is a process in which shared suffixes are removed, leaving just the stem, whereas lemmatization takes context into account, so there is a meaningful representation of the term. For this research, the successful – or unsuccessful – application of stemming shows that similar terms with or without suffixes impact interpretations. Additional parameters and metrics testing were utilised in this processing stage before running each model. First, metrics tests were run using the Gibbs sampling method (Gao, Chen & Zhu 2016) and the results from these tests were interpreted and used to guide parameters settings. These include assigning a minimum frequency for term occurrence and setting the number of topics for the algorithm to search for (Onan, Korukoglu & Bulut 2016). The final processed model was split into three corpus bodies to determine themes depending on paper publication. If there is a small number of papers about a specific theme, these could be masked in a larger corpus. Lastly, final parameters were set in the ldatuning package (Nikita & Chaney 2020) – including k, iteration, alpha, seed, etc. – before running the three LDAs. This information, along with scripts and data used within this study is accessible in an online repository (see Data Availability Statement).

LDA was combined with a similarity analysis to identify the extent of thematic overlaps between the three periods. The method to explore the rate of topics that are potentially shared between periods was based on a dissimilarity measure between pairwise topics. A final list of the twenty-five terms that have the highest weights per topic was selected and then used to create an absence/presence matrix of the listed terms that were found within each topic. Horn distance (Wolda 1981; Giovas 2021) was computed with this binary matrix and visualized with a heatmap, where the colour of each pairwise topic mapped its similarity ranging from 0 (same top 25 terms) to 1 (no overlap between top 25 terms). A variety of R packages were used to carry out the entire process.

5.1 Period 1: 1990–2001

Terms in Period 1 present Stonehenge before and in the years directly following the decree. The most notable repeated terms are: sites, monuments, society, and management – appearing with thematic demonstrative terms. These include visitor, political, government, symbolic, conservation and heritage. Texts published during this period discuss Stonehenge in broader, more generalized terms depicting the monument’s role within the public and academic spheres (Wainwright 2000; Pollard & Ruggles 2001). These terms reflect narratives describing public perception and current interventions linked to the presentation and urgent preservation of Stonehenge (see also Blandford 2006).

In Topic 6, management appears with Avebury, landscape, English, visitor, and national. Terms appearing with management indicate discourse surrounding the management of Stonehenge for public consumption and as a domestic heritage symbol. Stonehenge is also discussed comparatively with Avebury. The use of visitor is interesting, as they could be local or international – and the context of the site-visit could be for any number of reasons. In Topic 7, management appears with: societies, space, sites, and prehistoric. In this context, management relates to past people’s management. The reference to prehistoric and societies could pertain to how Stonehenge was managed by prehistoric people. Additionally, Topic 4 contains monuments appearing with political, conservation, and government. These terms can be interpreted as discourse highlighting the political impact on conservation and preservation policies for archaeological heritage monuments (Holtorf & Schadla-Hall 1999). This is important to highlight, due to the adoption of the ‘practice into policy’ framework, based on the development of values-based archaeological heritage management and conservation practices within the UK – gaining support and traction within both academia and heritage institutions – impacting adopted policies between local and national governments (Clark 2019).

Topic 9 has Wainwright appearing with heritage, site, England and current. It is fair to interpret Wainwright as archaeologist and prehistorian, Geoffrey Wainwright. During Period 1, Wainwright was the Chief Archaeologist for Historic England and is remembered for establishing planning standards within archaeological practices. How heritage sites within England are managed, conserved, and valued are attributed to the standards and practices developed during this time. The appearance of the term Wainwright could be further indicative of academic responses to the establishment and acceptance of new management and conservations practices, including potential commentary in how certain practices were translated into government policy within England. Lastly, there is the appearance of space and place. Within Topic 7, space appears with societies, setting and recognition in comparison to place, in Topic 8, appearing with ritual, symbolic, interest and region. Space, within the cultural heritage-geographic sphere, pertains more to the physical environment itself, whereas place is used to highlight human assignation of cultural value (Mallarach & Varschuuren 2019). Period 1 topics correspond to the lens which Stonehenge maintained visibility within the physical landscape and modern socio-cultural landscape. The terms further highlight how academic authors assigned value, authenticity, and agency in line with current archaeological thought.

5.2 Period 2: 2002–2012

During Period 2, the terms reflect new projects carried out and resulting research. Terms in multiple topics include: management, prehistoric, sites, landscape, and data – with data appearing for the first time. The first project was the ‘Stonehenge Riverside Project’, from 2003 to 2009 (Pearson et al. 2006). This project re-examined Stonehenge during the Bronze age by analysing the potential relationship of Stonehenge with nearby monuments and landscape features. Key results include the discovery of a previously unknown stone circle and new construction timeline. Overall, texts published during this period depict knowledge production reshaping Stonehenge as a monument linked with its surrounding features (Bewley, Crutchley & Shell 2005; Gaffney et al. 2012). Here, Stonehenge, the monument and landscape, is in motion as both a heritage site and researchable archaeological landscape (Fairclough 2006). Key topics include Topic 4, Topic 5, Topic 6, and Topic 8. Both Topic 4 and Topic 5 pertain to activities, based on the terms: data, landscape, source, found, material, and excavations.

Topic 6 focuses on new knowledge production. Terms include: project, new, interpretations and academic. In Topic 8 management appears again with: cultural, tourism, visitors, global, and experience. Previously, management pertained to Stonehenge as a site, its role within landscape, and concept of a visitor, whereas, in Period 2, management appears with terms in relation to a tourist-gaze and experience. It is important to note tourism appears with visitors, linking the two concepts. A tourist and a visitor are not intrinsically the same, each having different connotations (Dragouni & Fouseki 2018). While both visitors and tourists are participating in travel, the reasoning differs. Tourists are typically interested in more recreational and leisure activities in comparison to visitors that can be travelling for any reason (Poria, Butler & Airey 2004). It is furthermore important to note that shifting terminology pertaining to how individuals or groups are categorised when experiencing Stonehenge not only highlights shifts within academic theory and debates, but acts as a reflection on public perception and political discourse regarding Stonehenge’s framing and accepted social use. Discourse discussing what type of individual may experience Stonehenge indicates a shift in Stonehenge’s public presentation (Mason & Kuo 2008) which could further be reflected by the substantial interest by visitors and tourism activities at Stonehenge, with an estimated 1.253.397 number of people making their way to the Salisbury Plain in 2025 (ALVA 2025). Additionally, there are increasingly appealing tourism activities beyond solstice events, highlighted by the upcoming “Playhenge at Stonehenge” event for children (English Heritage 2026) and offering specialised British Sign Language (BSL) tours (English Heritage 2026). Other terms appearing for the first time are: research, significance, context, understanding, results, and knowledge.

The appearance of significance is positive when considering the changing valuation of an archaeological heritage site and landscape (ICOMOS Australia 2013) and with the importance of reflecting a multiplicity of voices in the discourses associated with archaeological sites (Colwell 2016, Rivolta et al. 2014). As time, socio-cultural contexts, and academic thought evolves – so does the assignations of value (Pastor Pérez et al. 2021; Díaz-Andreu & Pastor Pérez 2024). As theories and discussions within academic discourse develop, this reshapes the narratives of archaeological heritage sites and landscapes (Avrami 2020). Additionally, the ‘Stonehenge Hidden Landscape Project’ started in 2010 and continuing until 2016. This project utilized surveys beyond the monument to better understand the surrounding physical environment and features. Significant publications appear in Period 3. Topics and terms appearing in Period 2 highlight shifts of the Stonehenge site experience and physically reshaping the landscape. The interpretation of terms indicates authors’ changing perspectives and presentations of archaeological heritage knowledge. This period delineates a reinvention of the Stonehenge monument and landscape under an academic gaze that begins to want to include the lessons learned from public archaeology (Moshenska 2013).

5.3 Period 3: 2013–2022

The final period contains the conclusion of the ‘Stonehenge Hidden Landscape Project’, additional findings, and new publications. There are terms appearing in all three subsects. Terms correspond to previously mentioned projects about Stonehenge as a scientific object of study (Golding 1989; Pearson 2013). These include features, geochemical, analyses, examines, approach, identified, evidence, and theory. Specifically, this shift is most prevalent in Topic 3, Topic 10 and Topic 11. However, Topic 2 and Topic 5 support a transition of Stonehenge discourse from a tourist attraction to active archaeological intervention. Additionally, the following terms appear: present, presence, and presented in three different topics. For example, present appears with the terms: prehistoric, recent, development, suggest, and past. Whereas presence appears with construction, Neolithic, provide, and approach. Finally, presented appears with evidence, interpretation, site, and identified. What is most interesting is the difference between the associated terms for both present and presented. With present, terms primarily pertain to current work analysing Stonehenge in relation to its place within the past. With presented, the theme of the topic is different and more in line with discussions on the interpretations of new discoveries at the site.

Besides excavations, a new visitor centre was built at Stonehenge. This occurred during a period when the terms visitor, tourism and management disappeared. It is also worth noting that conservation disappeared after Period 1, with context disappearing after Period 2. The terms experience and interpretation remain. The House of Commons’ decree in 1993 highlighted a disappointment in the presentation and management of Stonehenge to the public. The outcome was a significant 30-year period of new research and new discoveries modifying Stonehenge – physically and conceptually – further within the academic gaze (Ellis 2015). The disappearance of certain terms – management, tourism, visitor, context, conservation – is a fascinating departure from previous discussions and ‘versions’ of Stonehenge. this is further interesting when considering the continued, evolving relationship of public and political perceptions and interactions attributed to monuments and landscapes such as Stonehenge. Additionally, with shifting theoretical debates and new perspectives applied to the study of tangible archaeological heritage, the above missing terms were unanticipated. The Stonehenge in Period 3 is a different iteration of the Stonehenge that the House of Commons was unimpressed with. The presence, role and understanding of Stonehenge is both relative to continued research and continued accepted experiences and management of the site and its landscape, which has evolved immensely in the 30-year period that was the focal point. The final period pertains to the Stonehenge site and landscape together with surrounding landscape features, nearby sites, and origin of stones.

5.4 Patterns between Periods

While there are clear differences between each period, there are still similarities, as is visible in Figure 6. The top three similar pairs determined exploring the Horn distance are: Pair A) Bluestone research (Period 2) and Bluestone Origins (Period 3); Pair B) Heritage Management (Period 2) and World Heritage (Period 3); and Pair C) Bluestone landscape (Period 1) and Bluestone Origins (Period 3) [second appearance]. The topic themes for Pair A and Pair C discuss the bluestones, with minor thematic deviations. The topics focus on the bluestones’ location of origin, movement to the Stonehenge landscape, and the research lens. Pair B instead presents topics referring to Stonehenge as a world heritage site and associated values, with emphasis on both management and the built environment. It is notable that both Pair A and Pair C contain topics, themes, and terms that appear in all three periods. The key takeaway regarding the themes that transcend the temporal relationship between research interests pertain to fascination with the bluestones used for construction and Stonehenge’s role as an integral facet to world heritage. Regardless of the human activities reshaping and revaluing Stonehenge, new research perspectives and engagement practices do not impact the appearance of the highlighted themes within ongoing Stonehenge discourse.

The evolution of topics reveals the diverse histories that have coexisted at Stonehenge, giving this space a unique multivocality where academics have created a space for the discourses of different social groups to find their place in current times. This movement into the present, into a new version – whether through deeper theoretical analysis (Hamilakis 2011), a new perceptive lens (Smith 2012), or physically changing interactions and experiences with the past (McIntosh & Prentice 1999) – creates the multitemporality of past objects and landscapes, a multipast landscape. This corpus presents a clear timeline of Stonehenge as an archaeological heritage landscape, including the transformation from a Neolithic monument to a global icon through contestation of its potential as a national disgrace. Stonehenge is a symbol no longer solely representing the monument. It has become a symbol of the preserved past itself.

The progression of subjects in academic texts studied in this article addressed two main questions, on the one hand indicating a transition towards incorporating a technological gaze in the examination of this case study, as well as a focus on material culture (Olsen 2012) and on the other hand a mapping of the evolution of archaeological trends that permeate into the core evolution of archaeological heritage management (de la Torre et al. 2005; Vienni 2014; Egloff 2019). These two-fold dimensions may support a better understanding in how best to utilize machine learning techniques to explore the construction of archaeological heritage management narratives over time (Högberg et al. 2017), particularly in reference to artefacts that may be retrieved from unaltered environments, such as the Neolithic period (Pearson 2021). The gradual increase in terms related to people and development is linked both to the site’s inclusion on the World Heritage List (Cleere 2010), and to the growth of public archaeology in the UK (Faulkner 2000). Additionally, these terms strengthened Stonehenge’s cultural heritage protection agencies (Clark 2006), which are increasingly aware of the social value of cultural heritage (Parga-Dans, Alonso González & Otero Enríquez 2020). However, we must also consider what outside factors can impact the trajectory of academic research within the fields of archaeology and heritage. In attempting to address these dimensions of dispute, utilizing quantitative text analysis measures, this manuscript explored the historical and contemporary ways in which academics perceive how people have used and engaged with landscapes, while also preserving the significance and value attributed to these landscapes by previous generations (Avrami & Mason 2019).

In addition, within the Digital Humanities some of the observed concepts correspond with different descriptions that various experts have detected in the evolution of approaches to modern considerations of relevance in describing the relationship between archaeological sites and present-day societies (Schwartz 2025). Management trends derived from critical studies place archaeological heritage on a continuum or process mutating in line with social changes, when following a values-based approach. However, the academic perception of these changes shows us a maintenance in terms of the symbolic value of a space that is appreciated historically at an early stage. This opens the door to future comparisons with other text compendia.

5.5 Benefits and Limitations of Methodological Approach

Quantitative textual analysis and Topic Modelling are not novel computational approaches within the humanities, and are not exempt from our own opinions, bias and criteria. More research conducted following FAIR and CARE principles leads to a higher rate of accessible workflows and guides for archaeological heritage researchers. Collaborative practices further aid researchers, in light of the continued development and availability of Artificial Intelligence (AI) technologies and tools, such as Generative AI. These technologies could affect usage of traditional NLP and text analysis methods. As far as limitations are concerned, Topic Modelling is a probabilistic model and for this reason it presents all usual challenges of modelling methods, the most important being that the modeller needs to set a list of parameter values for the model to work. The values of these parameters allows the user to tune and control the output, but at the same time they may generate different results. To address this issue, two strategies were developed. First, several runs of the algorithm were performed to explore to what extent decisions such as the minimum frequency of a term or the number of topics were modifying the result. Our findings suggested that these results were robust enough, provided that the parameter values were reasonable and aligned with the dataset sample size and the qualitative exploration of the model. Second, the source code and dataset are provided to allow for the reproducibility of the analysis. This final step is required to promote transparency in modelling methods so other modellers may be able to learn about our approach, test its results, and apply it to other relevant case studies. While reproducibility strategies are becoming common in archaeological modelling it is important to note that they are far less widespread within text-based Digital Humanities research (Marwick, 2017; Da, 2019; Ries, Van Dalen-Oskam & Offert 2024).

Beyond parameter values, the choice of a different algorithm than LDA structures the model in different ways resulting in similar yet varying results for the same corpora. Other algorithms include: Latent Semantic Analysis (LSA), BERTopic, and Negative Matrix Factorization (NMF). Researchers need to carefully consider and justify the selected algorithm, parameters and interpretations that contribute to useful results. Each method comes with its own assumptions, advantages, and limitations regarding the type and structure of the corpus. It is up to the individual researcher to evaluate which is the best method to use depending on the corpus and research questions. The use of LDA for this specific case study highlighted the advantages of this widespread technique, including the robustness of results for corpora with relatively large documents and the identification of topics when there is significant overlap between them (Hankar, Kasri & Beni-Hssane 2025). These traits allow researchers to generate scalable interpretations of archaeological heritage corpora combining data visualization and close reading of relevant documents. The ability to interpret results through close readings of pertinent texts or using critical heritage perspectives is integral for the value of results, yet this is linked with limitations. The interpretation and scalability of both the methods and results are challenging when considering the evolution of topics overtime, presence of overlapping topics and evaluation of parameters by researchers (Hankar, Kasri & Beni-Hssane 2025). Topic Modelling can also be used with other quantitative methods for a deeper dive into the sentiments, semantics and context in a corpus, and is an ideal complement to subsequent qualitative analyses.