Stratigraphic Semantics: Considering Context in Paleolithic Archaeology

Formation processes theory (Schiffer Reference Schiffer1972, Reference Schiffer1987) has long emphasized that the archaeological record results from the interaction of cultural and natural processes, and that the distinction between systemic context and archaeological context is crucial when interpreting stratigraphy. From this perspective, lithostratigraphic units represent only one part of a more complex formation history that also includes postdepositional and cultural transformations.

In this light, researchers are making efforts to integrate methods of establishing context through the application of different clustering techniques (Clark Reference Clark2017), lithic or faunal refitting (Deschamps and Zilhão Reference Deschamps and Zilhão2018; Sossa-Ríos et al. Reference Sossa-Ríos, Mayor, Sánchez-Romero, Mallol, Vaquero and Hernández2024), machine learning (Arteaga-Brieba et al. Reference Arteaga-Brieba, Courtenay, Cobo-Sánchez, Rodríguez-Hidalgo, Saladié, Ollé and Mosquera2023), and the integration of taxonomy/typology/taphonomy along vertical sequence of a site (Discamps et al. Reference Discamps, Thomas, Dancette, Gravina, Plutniak, Royer and Angelin2023; McPherron et al. Reference McPherron, Dibble and Goldberg2005; Mielgo et al. Reference Mielgo, Huguet, Laplana, Martín-Perea, Moclán, Márquez, Arsuaga, Pérez-González and Baquedano2024). Although these techniques retain the spatial logic of horizontal continuity and superposition, they are also marked by two key differences: they are defined from the spatial relationships among archaeological remains rather than sedimentary properties, and they are typically established or finalized after excavation is complete. However, such advances expose a persistent problem: the lack of clear terminological boundaries. Without explicit definitions that specify both the analytical purpose of a unit and the methodological criteria by which it is established, concepts risk being conflated or applied inconsistently. This undermines reproducibility and can lead to conflicting interpretations (Bicho and Cascalheira Reference Bicho, Cascalheira, Cascalheira and Picin2020).

An illustrative case is the effort to associate Nubian technology and Neanderthals at Shukbah cave (Blinkhorn et al. Reference Blinkhorn, Zanolli, Compton, Groucutt, Scerri, Crété, Stringer, Petraglia and Blockley2021). Blinkhorn and colleagues use varying terminology related to Shukbah’s sequence, which includes “layer,” “occupation,” and “archaeological horizon” interchangeably. In contrast, the original excavator of the site, Dorothy Garrod, exclusively uses “layers” as the main contextual units, describing them as units of different geological attributes (Garrod Reference Garrod1942). In their reply, Hallinan and colleagues (Reference Hallinan, Barzilai, Bicho, Cascalheira, Demidenko, Goder-Goldberger, Hovers, Marks, Oron and Rose2022) argue that Shukbah’s stratigraphic units represent time-averaged deposits and as such cannot support a secure association between the Neanderthal remains and the Nubian artifacts. The exchange shows that questioning the chronological resolution of a deposit does not necessarily challenge the broader assumptions used to define archaeological context. Addressing this issue requires recognizing that methodological innovation alone cannot resolve ambiguity unless it is grounded in a coherent and explicit theoretical framework.

What’s in a Name?

Although definitions of stratigraphic and contextual units exist (Darvill Reference Darvill2003; Gasche and Tunca Reference Gasche and Tunca1983; Harris Reference Harris1979), their application often depends on the author’s background (Gilbert and Goldberg Reference Gilbert and Goldberg2017). This points to a deeper issue in archaeological interpretation, because terminology should convey both the analytical purpose of a unit and the methodological criteria used to establish it. Instead, definitions are often inferred indirectly from archaeological dictionaries (Darvill Reference Darvill2003; Kipfer Reference Kipfer2021) or excavation methodology manuals (Harris Reference Harris1979; Renfrew and Bahn Reference Renfrew and Bahn2004).

As a result, researchers frequently use different terms when describing the same stratigraphic unit (SU), or the same term for units established through very different methods. Terms such as “level,” “layer,” or “horizon” may appear superficially interchangeable, yet the methods used to define them can vary substantially (Table 1), particularly with the rise of PES approaches. More complexity arises when these are conflated with interpretative terms such as “occupation,” “episode,” or “event.” A “layer” identifies a bounded deposit or spatial distribution, whereas an “occupation” implies human presence and activity within that unit. “Episode” suggests a more discrete phase of activity, whereas “event” implies an even narrower and more closely contemporaneous occurrence; related concepts such as living floors carry similar fine-scale behavioral implications (Malinsky-Buller et al. Reference Malinsky-Buller, Hovers and Marder2011; Vaquero Reference Vaquero2008; Vaquero et al. Reference Vaquero, Gema Chacón, Dolores García-Antón, de Soler, Martínez and Cuartero2012). For example, there is often no explicit step in determining when a layer bearing Magdalenian tools becomes a Magdalenian occupation, despite the very different inferential weight these terms carry. The former may only designate the stratigraphic unit in which Magdalenian material was recovered, whereas the latter implies that the material represents a meaningful phase or episode of human activity. In practice, they are often used interchangeably to denote human presence (Bailey Reference Bailey2007). When this inferential transition is left unstated, sedimentary, excavational, analytical, and behavioral units may be treated as equivalent despite being defined through different forms of evidence. This matters because archaeological analysis routinely operates across multiple unit types (e.g., layers, levels, assemblages, phases) that organize time and structure subsequent interpretation (Bailey Reference Bailey2007; Discamps et al. Reference Discamps, Thomas, Dancette, Gravina, Plutniak, Royer and Angelin2023). Formation-process theory has long demonstrated that archaeological deposits are not direct reflections of past behavior (Schiffer Reference Schiffer1972, Reference Schiffer1987); consequently, different kinds of units support different levels of behavioral or temporal inference.

Table 1.

Definitions, Analytical Bases, and Common Inferential Issues for Key Stratigraphic and Contextual Terms in Paleolithic Archaeology.

Table 1 Long description

Columns are Term, Defined as, Based on, Issues, and References. Layer is defined as a geological unit based on physical traits of sediments, based on observations in the field and geoarchaeological analysis; issues note it is often treated as an occupation proxy, boundaries can be coarse or time averaged, and layer is sometimes used as a generic synonym for level or horizon; references are Harris 1979, Goldberg and Macphail 2005, and Gasche and Tunca 1983. Level is defined as an excavation unit or slice of arbitrary thickness, based on methodological choices; issues note it is defined as a nongeological excavation unit but often treated as a stratigraphic entity in writing, and its mechanical nature can cut across lithostratigraphic boundaries; references are Harris 1979 and Gasche and Tunca 1983. Horizon is defined as a thin horizontally traceable feature or zone, often a correlation unit rather than a deposit unit, based on observations in the field and interpretation of materials; issues note it can be defined by both geological and archaeological entities in the sequence; reference is Gasche and Tunca 1983. Occupation is defined as an interpretive unit implying human presence or activity and not a sedimentary unit by default, based on interpretation of materials; issues note a stratigraphic unit is often equated to occupation, it implies an event or episode but is usually time averaged, and it is often inferred with weak operational criteria; references are Bailey 2007, Bicho and Cascalheira 2019, Discamps et al. 2023, and Machado et al. 2013.

Efforts to impose greater terminological precision are not new. In the 1970s and 1980s, Fedele (Reference Fedele, Davidson and Shackley1976, Reference Fedele1984) proposed the concept of the “elemental sediment unit” (ESU), defined as the smallest geologically recognizable entity. Around the same time, Gasche and Tunca (Reference Gasche and Tunca1983), following a series of workshops, developed the Guide to Archaeostratigraphic Classification and Terminology. Their framework defined the lithologic unit as the base stratigraphic unit: a three-dimensional homogeneous entity composed of a dominant lithologic type or a consistent combination of several types. Both initiatives highlighted a central problem: semantic inconsistency, in which identical units were given different names, or distinct units were referred to with the same term.

Addressing such inconsistencies is especially important today, given that nonlithological approaches to defining stratigraphy have become increasingly prominent. Incorporating new types of SUs into existing interpretive frameworks without terminological standardization risks conflating methodological tools with interpretive outcomes. Establishing consistent baseline vocabulary would not only align Paleolithic archaeology with calls for greater standardization in geoarchaeology (Stein and Holliday Reference Stein, Holliday, Gilbert, Goldberg, Mandel and Aldeias2020) but would also mirror successful efforts in other disciplines. Controlled vocabularies in ecology, linguistics, and the biomedical sciences, for example, have been crucial for ensuring interoperability and cumulative research (e.g., American Psychological Association 2001; Ashburner et al. Reference Ashburner, Ball, Blake, Botstein, Heather Butler and Davis2000; Farrar and Langendoen Reference Farrar and Langendoen2003). Despite recognition of this need, a standard practice for defining and naming SUs in Paleolithic archaeology has yet to be implemented.

Terminological Instability in Archaeological Contexts: A Case for NLP Integration

The persistent problem of vague and inconsistent terminology in archaeology requires systematic analysis across a wide range of subfields and research traditions. Conducting such a review manually would be prohibitively time consuming, given the breadth of the archaeological corpus. Natural language processing (NLP) is a field concerned with the computational analysis of human language. Drawing on linguistics, computer science, and artificial intelligence, it enables large-scale analysis of terminology across diverse publications (Devlin et al. Reference Devlin, Chang, Lee and Toutanova2019; Martín-Rodilla Reference Martín-Rodilla, Gonzalez-Perez, Martin-Rodilla and Pereira-Fariña2023).

Large language models (LLMs) such as BERT (Bidirectional Encoder Representations from Transformers) or GPT (Generative Pre-Trained Transformer) are well-established tools that encode language and represent it as measurable vectors (Brown et al. Reference Brown, Mann, Ryder, Subbiah, Kaplan, Dhariwal and Neelakantan2020; Devlin et al. Reference Devlin, Chang, Lee and Toutanova2019; Mikolov et al. Reference Mikolov, Chen, Corrado and Dean2013). Although generative models have recently become integral to producing scientific text, other capabilities of NLP (e.g., semantic similarity detection, clustering, terminology mapping) remain underutilized in archaeology (see Brandsen Reference Brandsen, Gonzalez-Perez, Martin-Rodilla and Pereira-Fariña2023; Martín-Rodilla Reference Martín-Rodilla, Gonzalez-Perez, Martin-Rodilla and Pereira-Fariña2023).

Because data mining and NLP can rapidly process vast corpora, detecting language patterns, sentiment, or named entities, they are increasingly applied in the social sciences (Franzosi et al. Reference Franzosi, Dong and Dong2022; Macanovic Reference Macanovic2022). Recent studies have used LLMs to investigate gender and ethnic stereotypes (Garg et al. Reference Garg, Schiebinger, Jurafsky and Zou2018), cultural markers of class (Kozlowski et al. Reference Kozlowski, Taddy and Evans2019), and representations of intangible cultural heritage (Ning and Chen Reference Ning and Chen2023). In archaeology, ArchaeoBERT has demonstrated promise, demonstrating effectiveness in tasks such as named entity recognition (NER) and fill-mask completion. This model was fine-tuned on archaeological texts in English, Dutch, and German (Brandsen Reference Brandsen2024; Brandsen et al. Reference Brandsen, Verberne, Lambers and Wansleeben2022).

Our analysis focuses on the classification of predefined terms related to archaeological context. Each term was observed within a defined window of neighboring words, converted into a high-dimensional vector, and then visualized using dimensionality reduction techniques such as t-SNE (Van der Maaten and Hinton Reference Van der Maaten and Hinton2008), Uniform Manifold Approximation and Projection (UMAP; McInnes et al. Reference McInnes, Healy and Melville2020), and principal component analysis (PCA). By systematically mining archaeological literature in this way, NLP can help standardize terminology, highlight inconsistencies, and generate tangible representations of implied or assumed meanings.

Materials and Methods

Data Mining and Preprocessing

To compile a dataset suitable for semantic analysis of archaeological texts, we implemented a Python workflow that automated data collection, cleaning, and formatting of publications for NLP. We used the Scholarly Python library to interact programmatically with Google Scholar, creating an Excel database that allowed for systematic screening of the queried articles. Because bulk downloading of articles can lead to IP address bans from Google Scholar, each article was manually downloaded using the URLs provided in the metadata. Although time consuming, this approach preserved the integrity of the data collection process and allowed for further control and screening of the sampled articles. The data collection was performed in July 2024, which defines the temporal boundary of the dataset.

Using the search string “paleolithic+occupation,” we retrieved 960 records from Google Scholar. Because query-based searches introduce topical bias, this focus was deliberate: the aim was to examine terminology in occupation-related Paleolithic contexts. For each record, we documented author, year, title, and URL. Study selection followed PRISMA-style reporting principles (Page et al. Reference Page, McKenzie, Bossuyt, Boutron, Hoffmann, Mulrow and Shamseer2021; Figure 1). Records were screened by title and abstract, and subsequently, by full text. Articles were included if they were peer-reviewed and addressed Paleolithic archaeology. No temporal restrictions were applied so as to capture long-term patterns of terminological usage. Purely geological or nonarchaeological publications were excluded. Although the search query centered on “occupation,” eligibility remained broad within Paleolithic research (including site reports, syntheses, and methodological studies) to capture varied uses of stratigraphic terminology.

Figure 1.

PRISMA-style flowchart outlining the corpus selection and screening process. Adapted from the PRISMA 2020 flow diagram (Page et al. Reference Page, McKenzie, Bossuyt, Boutron, Hoffmann, Mulrow and Shamseer2021).

The collected articles were processed according to a workflow adapted from Hickman et alia (Reference Hickman, Thapa, Tay, Cao and Srinivasan2022) and Brandsen (Reference Brandsen, Gonzalez-Perez, Martin-Rodilla and Pereira-Fariña2023) for extracting text from scientific articles (Table 2). The library PyMuPDF was used to extract the text from articles and prepare it for further processing. As an initial step, the first page and references were removed. Stopwords and words shorter than three characters were removed, and the remaining words were tokenized—that is, transformed into individual units using the Natural Language Toolkit (NLTK) library (Bird et al. Reference Bird, Klein and Loper2009). Each token was then lemmatized (e.g., “running” to “run”). We established the LBFC dataset (lithic, bone, flake, charcoal) under the assumption that these four terms are semantically distinct material-related terms. Although “flake” and “lithic” are related terms, including them was done intentionally to test whether the NLP pipeline can detect more nuanced distinctions within closely related terminology. For the main analysis, we generated the OLLH dataset (occupation, layer, level, horizon), based on the terms identified above as central to defining archaeological context.

Table 2.

Workflow for Text Extraction, Semantic Analysis, and Visualization of the Results.

Table 2 Long description

Three columns list Action, Example, and Resources Used for five numbered steps. Step 1, Access an article, includes creating a database, downloading articles, removing the first page, and removing references; the example is a sample text excerpt, and resources are Google Scholar, Scholarly, PyMuPDF, and Regex. Step 2, Process, lists lowercase, stopword removal, citation removal, and lemmatization; the example is an article represented as a list of lemmatized terms, and resources are Natural Language Toolkit, labeled NLTK, and Regex. Step 3, Querry, lists locating terms and tokenizing terms within a 20 word window; the example is Layer represented as a list of terms, and the resource is Transformers. Step 4, SciBERT, lists vectorizing the window and representing each term by a 768 dimension vector; the example is Layer represented as a list of decimal values, and resources are Transformers and PyTorch. Step 5, Data analysis and visualization, lists Venn diagram, UMAP, Silhouette score, and K NN; the example cell is blank, and resources are Matplotlib and scikit learn.

NLP and UMAP Analysis

The results of our analysis display two distinct patterns between the LBFC and OLLH datasets. The two datasets are shown to be distinct in both UMAP (Figure 4, Supplementary Figure 1), as well as k-NN, which displays high separability between datasets, with an accuracy of 0.945 ± 0.006 and a macro-F1 of 0.945 ± 0.006 (Figure 4).

LBFC Dataset

Testing the NLP pipeline on the LBFC dataset demonstrated the expected outcome: clear semantic separation among terms. Across all n_neighbors settings, the clusters remained stable in structure, with only their positions shifting (Figure 4a, c, e, g). Contextualized embedding clusters for “lithic,” “flake,” and “charcoal” remain proximate (Figure 4a), but their separation remains stable with increasing n_neighbors. Contextualized embeddings were consistently distinct across all settings, forming a well-defined cluster throughout. The silhouette scores remain between 0.48 and 0.49, showing moderate but meaningful separation of terms. This is confirmed by k-NN, displaying a very high accuracy (Figure 3; Accuracy = 0.952 ± 0.009; M-F1 = 0.937 ± 0.012)

Figure 3.

Cross-validated k-NN results.

Lithic contextualized embeddings produced two separate clusters, which become more distant with the increase of n_neighbors (Figure 4a, c, e, g). Clusters of “lithic,” “flake,” and “charcoal” remain proximate while maintaining separation. They start to diffuse as the n_neighbors value increases, implying that their separation is stronger at a more global scale. The rest of the term embeddings remain separated and stable throughout different n_neighbors values.

Figure 4.

UMAP structure across different neighborhood sizes.

Figure 4 Long description

Eight scatter plots labeled a) through h), comparing LBFC and OLLH datasets across neighborhood sizes of 15, 30, 50 and 75. Each plot has Component 1 on the horizontal axis and Component 2 on the vertical axis, both ranging from negative 15 to positive 15. The plots illustrate clustering patterns with silhouette scores indicating clustering quality. Plots a), c), e) and g) show the LBFC dataset with silhouette scores of 0.46, suggesting moderate clustering. The data points form distinct clusters, with some overlap, particularly as the number of neighbors increases. The densest areas are visible in the lower right and upper left regions. Plots b), d), f) and h) display the OLLH dataset with silhouette scores of 0.08, indicating less distinct clustering. The points are more dispersed, with less clear separation between categories. The densest regions are less pronounced compared to LBFC. Each plot uses color-coding to represent categories: incorporation, no incorporation, no incorporation 2 and no incorporation 3. In LBFC plots, colors form more distinct regions, while in OLLH plots, colors are more mixed. The figure aims to show how neighborhood size affects clustering compactness and separation in both datasets.

Although the clusters remained stable overall, limited intermixing was observed, such as small numbers of “bone” and “charcoal” contextualized embeddings appearing within the larger lithic cluster. This behavior is consistent with the design of the embeddings, which capture semantic context rather than strict lexical similarity, and it reflects the heterogeneous character of the source literature.