Paper deep dive
SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension
Daniele Raimondi, Feichi Lu, Oliver Grun, Mariia Eremina, Andrea Perlato
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/11/2026, 3:40:20 AM
Summary
The paper introduces SCALE, a framework that extends the OpenAlex taxonomy by adding a new level of 'Scientific Concepts' below existing Topics. Using LLMs, text embeddings (SPECTER2), and graph-based community detection (Leiden algorithm) on MDPI author keywords, SCALE creates approximately 114,000 stable, fine-grained conceptual units to bridge the gap between broad disciplinary topics and fragmented individual keywords.
Entities (8)
Relation Signals (7)
SCALE → extends → OpenAlex
confidence 95% · SCALE extends the OpenAlex taxonomy with a new level of scientific Concepts below Topics.
SCALE → generates → Scientific Concepts
confidence 95% · The resulting taxonomy provides a basis for future applications... approximately 114,000 Scientific Concepts.
Scientific Concepts → isbelow → OpenAlex Topics
confidence 95% · They form a new level below existing OpenAlex Topics.
SCALE → uses → LLM
confidence 95% · The framework combines scientific text embeddings, large language models, and graph-based community detection.
MDPI → provides → Author Keywords
confidence 90% · MDPI author keywords include approximately three million distinct keywords extracted from nearly two million scientific articles from MDPI.
SCALE → uses → Leiden algorithm
confidence 90% · we apply the Leiden algorithm to identify semantic communities.
SCALE → uses → SPECTER2
confidence 90% · The keyword and its definition are then jointly encoded with SPECTER2.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The increasing specialization of scientific research challenges existing classification systems, which provide effective representations of broad disciplines and research topics but often fail to capture the fine-grained conceptual structure of contemporary science. Author keywords offer greater specificity, but their fragmentation, redundancy, and terminological variability limit their use as stable units of knowledge organization. We introduce SCALE (Scientific Concept Aggregation via LLMs and Embeddings), a framework that extends the OpenAlex taxonomy with a new level of scientific Concepts below Topics. Rather than treating keywords as isolated descriptors, SCALE organizes semantically related terms into coherent and interpretable conceptual units and integrates them within the existing disciplinary hierarchy. The framework combines scientific text embeddings, large language models, and graph-based community detection to construct this additional layer at scale. The resulting taxonomy enables scientific literature to be read through an intermediate conceptual level between broad research topics and individual documents. This perspective provides a more detailed representation of how scientific knowledge is structured, specialized, and connected across disciplines. By transforming heterogeneous author terminology into reusable hierarchical units, SCALE offers a foundation for fine-grained scholarly classification, scientometric analysis, research monitoring, and future ontology development.
Tags
Links
- Source: https://arxiv.org/abs/2608.07254v2
- Canonical: https://arxiv.org/abs/2608.07254v2
Trouble viewing inline? Open PDF directly →
Full Text
48,994 characters extracted from source content.
Expand or collapse full text
SCALE: SCIENTIFIC CONCEPT AGGREGATION VIA LLMS AND EMBEDDINGS FOR FINE-GRAINED TAXONOMY EXTENSION Daniele Raimondi daniele.raimondi@mdpi.com Feichi Lu feichi.lu@mdpi.com Oliver Grun oliver.grun@mdpi.com Mariia Eremina mariia.eremina@mdpi.com Andrea Perlato ∗ andrea.perlato@mdpi.com MDPI Data Intelligence Team MDPI AG Grosspeteranlage 5 CH-4052 Basel, Switzerland ABSTRACT The increasing specialization of scientific research challenges existing classification systems, which provide effective representations of broad disciplines and research topics but often fail to capture the fine-grained conceptual structure of contemporary science. Author keywords offer greater specificity, but their fragmentation, redundancy, and terminological variability limit their use as stable units of knowledge organization. We introduce SCALE (Scientific Concept Aggregation via LLMs and Embeddings), a framework that extends the OpenAlex taxonomy with a new level of scientific Concepts below Topics. Rather than treating keywords as isolated descriptors, SCALE organizes semantically related terms into coherent and interpretable conceptual units and integrates them within the existing disciplinary hierarchy. The framework combines scientific text embeddings, large language models, and graph-based community detection to construct this additional layer at scale. The resulting taxonomy enables scientific literature to be read through an intermediate conceptual level between broad research topics and individual documents. This perspective provides a more detailed representation of how scientific knowledge is structured, specialized, and connected across disciplines. By transforming heterogeneous author terminology into reusable hierarchical units, SCALE offers a foundation for fine-grained scholarly classification, scientometric analysis, research monitoring, and future ontology development. Keywords Scientific taxonomy·Large language models·Text embeddings·Graph clustering·Knowledge organization 1 Introduction The rapid expansion of scholarly literature has made organizing research increasingly difficult. Scientific output continues to grow in volume, while research fields become more specialized and new interdisciplinary areas emerge at the boundaries of established disciplines [Bornmann and Mutz, 2015]. As a result, simply assigning publications to broad disciplinary categories is no longer sufficient. Organizing research has become essential for information retrieval, scientometric analysis, the detection of emerging trends, and evidence-informed decision-making. Research communities have proposed several approaches to make the structure of science more visible and easier to analyze. Classification systems, citation-based methods, term co-occurrence analysis, knowledge graphs, ontologies, and semantic representations have each contributed to this goal from different perspectives [Fortunato et al., 2018]. Large-scale infrastructures such as OpenAlex show how hierarchical structures can connect disciplines, subdisciplines, and research topics across the global scholarly record [Priem et al., 2022]. These systems provide an important foundation for navigating and analyzing scientific literature at scale. ∗ Corresponding author. arXiv:2608.07254v2 [cs.DL] 10 Aug 2026 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension However, a key limitation remains. Existing hierarchical classifications represent broad domains and research topics well, yet they rarely offer enough granularity to describe highly specialized areas of contemporary research [Clarivate, 2025, Elsevier, 2024, Australian Bureau of Statistics, 2020]. Author keywords offer a more detailed view of individual publications, but they are fragmented, redundant, and strongly affected by terminological variation. The same concept may appear under different names, while similar terms may carry different meanings across fields. This makes author keywords unreliable as stable units at scale. These limitations create a gap between high-level classifications and the highly specific descriptions associated with individual documents. What is needed is a new level, positioned below existing topics, specific enough to capture specialized research themes but still stable enough to fit into a coherent hierarchy. Such a level should also be scalable and compatible with existing classification infrastructures, rather than requiring a completely new system. In this work, we introduce Concepts as a new hierarchical level designed to bridge this gap. They form a new level below existing OpenAlex Topics: coherent groups of related terms, built from the raw vocabulary of author keywords but organized into a far more stable and reusable structure. To construct them, we propose a scalable framework that combines semantic representations, graph-based clustering, and automated procedures for annotating and integrating them into the hierarchy. Applied to the OpenAlex taxonomy, the framework extends the existing structure with approximately 114,000 Scientific Concepts connected to the Topics already present in the OpenAlex hierarchy [OpenAlex, a]. The resulting taxonomy provides a basis for future applications in scientometric analysis, semantic exploration, and large-scale research intelligence. The next section reviews the relevant literature and clarifies the methodological context that led to the development of Scientific Concepts. 2 Related Work This section reviews prior work relevant to the construction of Concepts, organized around four lines of research: science mapping, semantic representation of scientific text, topic modeling, and knowledge organization systems. It then examines OpenAlex specifically, before identifying the gap that Concepts are designed to address. 2.1 Mapping and Visualization of Science Understanding the structure of science has long been a central concern in scientometrics. Science mapping methods aim to identify, visualize, and analyze relationships between disciplines, research areas, and topics, transforming large bibliographic collections into interpretable representations of research activity [Petrovich, 2021]. A major line of work relies on citation-based methods, bibliographic coupling, and term co-occurrence analysis to reveal emerging structures and thematic relationships within the literature. Tools such as VOSviewer have contributed significantly to the adoption of these approaches by enabling the construction of science maps based on bibliometric and semantic similarities [van Eck and Waltman, 2010]. More recent comparative approaches have extended this perspective by supporting systematic comparisons of different representations of scientific structure and their evolution over time. These methods are valuable for describing and visualizing how research areas are connected. Still, they are usually designed to reveal patterns, clusters, or relationships in bibliographic data rather than to create persistent conceptual units that a hierarchical classification system can incorporate. Science mapping therefore provides an important foundation for understanding the structure of research, without addressing the need for stable, fine-grained, hierarchically organized units. 2.2 Semantic Representation of Scientific Literature Advances in semantic representation have changed the way scientific literature is analyzed and organized. Traditional approaches relied mainly on keywords, term frequencies, and co-occurrence patterns. More recent developments in Natural Language Processing have introduced models that represent scientific documents as dense vectors in high-dimensional semantic spaces, enabling the capture of relationships among publications beyond lexical similarity. Transformer-based models such as SPECTER have played an important role in this development. SPECTER introduced document representations that combine textual information with citation-based signals. By embedding related articles close to one another in a shared vector space, these representations have supported tasks such as information retrieval, document classification, and large-scale analysis of scientific literature [Cohan et al., 2020]. Semantic representation models provide a powerful technological basis for analyzing scientific literature at scale. Still, these models primarily represent documents or concepts in a semantic space rather than construct explicit structures for 2 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension knowledge organization. They can capture meaningful relationships among publications without directly producing an interpretable, maintainable conceptual hierarchy. 2.3 Topic Modeling Approaches Researchers have long used topic modeling to discover latent thematic structure in large text collections. Classical approaches such as Latent Dirichlet Allocation (LDA) represent documents as mixtures of topics, where a distribution over words defines each topic, without requiring predefined categories. More recent neural approaches, such as BERTopic, combine transformer-based embeddings with clustering to produce topics that better capture semantic similarity beyond word co-occurrence [Blei et al., 2003, Grootendorst, 2022]. These methods are effective at uncovering patterns in unlabeled corpora. Still, the resulting topics typically remain corpus-specific, can vary across different runs, and do not align with any pre-existing classification hierarchy. As a result, they do not directly produce reusable, hierarchically organized units for large-scale scholarly classification. 2.4 Knowledge Organization Systems and Scientific Ontologies Alongside advances in semantic representation, researchers have developed explicit frameworks to organize concepts and their relationships. Knowledge Organization Systems (KOS) include taxonomies, ontologies, and classification schemes that represent knowledge in a structured and interpretable form. Recent computational approaches have addressed the automatic construction and completion of topic taxonomies. TaxoGen constructs corpus-specific topic hierarchies by combining adaptive term embeddings with recursive clustering of semantically coherent terms, whereas TaxoCom uses a partial taxonomy to guide the discovery of both existing and novel subtopic clusters. More recently, TaxoAdapt combines LLM-based multidimensional taxonomy generation with corpus-grounded hierarchical classification and iterative expansion to represent evolving scientific domains [Zhang et al., 2018, Lee et al., 2022, Kargupta et al., 2025]. Unlike these approaches, our framework does not construct or recursively complete an entire corpus-specific hierarchy. Instead, it preserves the four-level OpenAlex taxonomy and introduces a controlled fifth level of fine-grained Concepts below Topics through LLM-based granularity and field classification, field-specific semantic graphs, and Leiden clustering. In the scientific domain, the Computer Science Ontology (CSO) provides an important example of how large-scale representations of research areas can combine automated methods with expert knowledge. Related work, such as Klink-2, has demonstrated how computational methods can identify, link, and update relationships among scientific topics over time, supporting the construction of dynamic semantic research networks [Salatino et al., 2020, Osborne and Motta, 2015]. Even so, combining broad disciplinary coverage, fine granularity, and continuous updating within a single system remains difficult in practice. 2.5 OpenAlex and the Challenge of Fine-Grained Classification OpenAlex is one of the largest infrastructures for organizing and classifying the global scholarly record. Its dataset is released under the C0 license, allowing unrestricted reuse [OpenAlex, b]. The OpenAlex taxonomy is structured into Domains, Fields, Subfields, and Topics, providing a broad and coherent representation of contemporary research areas [OpenAlex, a]. Despite its scale and utility, the OpenAlex hierarchy still leaves the challenge of representing more granular research themes open. Earlier concept-based approaches inherited from the Microsoft Academic Graph (MAG) showed how difficult it is to maintain conceptual units that remain stable, interpretable, and consistent over time [Wang et al., 2020]. For this reason, OpenAlex deprecated Concepts in favor of the current Topics hierarchy, complemented by an auxiliary entity called Keywords, derived from each topic [OpenAlex, c]. These OpenAlex keywords, however, are a fixed, closed set (currently over 26,000, ten per topic) generated to describe existing topics and assigned to individual works for retrieval purposes; they do not constitute a hierarchical level of stable, reusable concept units with an explicit path to subfields, fields, and domains [OpenAlex, d]. This choice reflects a broader limitation of large-scale scholarly classification. Existing hierarchical levels often remain too general for highly specialized research areas. In contrast, keyword- and similarity-based approaches provide greater detail but do not necessarily produce a stable and interpretable structure. As a result, existing systems do not produce a coherent, maintainable layer beneath the Topics hierarchy from this material. 3 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension 2.6 Toward Concepts The literature reviewed in the previous sections shows a clear progression in the methods used to represent and organize scientific literature. Science mapping has provided tools for identifying and visualizing the structure of research. Semantic representation models have improved the ability to capture conceptual relationships among documents, while Knowledge Organization Systems have introduced explicit and interpretable structures for organizing concepts. However, each of these approaches addresses only part of the problem: detecting structure, representing meaning, or organizing concepts, but rarely all three at once. Concepts address this gap without replacing existing taxonomies, ontologies, or classification infrastructures. Rather, they extend the hierarchy downward, increasing the granularity of scientific classification while remaining coherent, interpretable, and easy to maintain. 3 Methods This section describes our framework for constructing a fifth Concept level and integrating it below the existing OpenAlex taxonomy. The core challenge is to transform noisy, heterogeneous author keywords into stable and reusable scientific Concepts that are fine-grained but not too specific for specialized research themes. Starting from MDPI author keywords introduced in Section 3.1, our framework addresses the challenging problem through four main steps. In Section 3.2, we identify terms that match the target level of granularity for a fifth-level taxonomy. In Section 3.3, we construct field-specific semantic graphs for keyword community detection. Section 3.4 focuses on graph clustering techniques and concept renaming. In Section 3.5, we attach resulting Concepts to OpenAlex Topics. Additionally, Section 3.6 defines an incremental maintenance procedure that allows new Concepts to be added over time. Fig. 1 summarizes the complete methodological workflow. Figure 1: Overview of the Concepts framework. Author keywords are progressively transformed into Concepts through semantic processing, clustering, and hierarchical integration with the OpenAlex taxonomy. 4 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension 3.1 Data Sources: OpenAlex Topics and MDPI Author Keywords The OpenAlex Taxonomy organizes scientific knowledge into 4 levels: Domains, Fields, Subfields, and Topics. In this work, Topics [OpenAlex, a] serve as the parent level for the new concepts to be attached. MDPI author keywords include approximately three million distinct keywords extracted from nearly two million scientific articles from MDPI [MDPI, 2026]. The data provide a broad, detailed, and multidisciplinary representation of specialized research concepts and methods, making it a suitable starting point for constructing a fine-grained scientific taxonomy level. 3.2 Identifying the Target Level of Granularity A central design decision is to determine which author keywords are suitable to be included as fifth-level Concepts. The target level should be narrower than existing OpenAlex Topics, but broader than too specific entities, such as products, biological species, genes, chemical compounds, and long contextual phrases. The goal is therefore to identify reusable scientific classification units rather than to preserve every keyword that appears in the source data. The author keywords are then classified according to their level of granularity. Each keyword is assigned to one of three categories: high-level, Concept-level, or low-level. High-level terms include existing OpenAlex 4 levels, examples such as Computer Science, Physics, Artificial Intelligence, and Epidemiology. Low-level terms are those too specific or contextual to serve as stable taxonomy units. Concept-level terms represent the target granularity of the fifth level, including methods, techniques, research themes, and specialized application areas. We use OpenAI models to classify the keywords into different conceptual granularity [OpenAI, 2024, 2025]. Only keywords classified as Concept-level are retained for subsequent field-specific clustering. This filtering step ensures that the resulting taxonomy is not a simple keyword inventory, but a controlled layer of scientific Concepts positioned between OpenAlex Topics and raw author keywords. 3.3 Construction of Field-Specific Semantic Graph After identifying candidate Concept-level keywords, the next step is to detect keyword communities with similar meanings that can form the basis of the Concept layer. To do so, we first construct a semantic graph over the candidate keywords, where nodes represent keywords and edges represent semantic relationships between them. This graph-based strategy is well suited to scientific terminology because conceptual communities can vary substantially in density, size, and internal structure. Unlike direct clustering approaches, which often impose stronger assumptions about cluster shape or scale, graph clustering allows Concepts to emerge from local semantic relationships among keywords [Ester et al., 1996, Traag et al., 2019]. However, a global semantic graph may ignore disciplinary context. Author keywords do not always carry a single stable meaning across all fields: the same or similar keyword may refer to different concepts in different disciplines, and keywords from different Fields may appear close in embedding space even when they belong to distinct research contexts. To preserve semantic coherence, we therefore construct field-specific semantic graphs rather than a single global graph. Each candidate keyword is assigned to one or more of the 26 OpenAlex Fields using an LLM [OpenAI, 2024]. When a keyword is relevant to more than one discipline, it is processed independently within each assigned Field, rather than being forced into a single disciplinary path. To build the semantic graph for each field, we first enrich the contexts of the author keywords with a short field-specific definition generated by an LLM. The keyword and its definition are then jointly encoded with SPECTER2 [Singh et al., 2023], producing a contextualized semantic representation that captures both the keyword itself and its disciplinary meaning. To create semantic edges, we perform nearest-neighbor search using the efficient HNSW algorithm [Malkov and Yashunin, 2020], which identifies related terms in high-dimensional embedding space. To reduce weak or accidental connections, two keyword nodes are linked only when they appear in each other’s nearest-neighbor lists, and their semantic similarity exceeds a predefined threshold. 3.4 Graph Clustering Towards Concept Layer Based on the field-specific semantic graph, we apply the Leiden algorithm [Traag et al., 2019] to identify semantic communities. These communities group together closely related keywords and provide the empirical basis for constructing the final Concepts. A key challenge in this step is controlling the granularity of the clusters. If clusters are too large, the resulting Concepts may become overly general; if clusters are too small, the resulting Concepts may become too specific and fragmented. Therefore, clustering parameters, especially the Leiden resolution parameter, must be tuned to obtain an appropriate level of conceptual granularity. 5 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension To address this trade-off, we manually annotate two sample sets of Concept pairs: must-link pairs, which should be placed in the same cluster, and should-not-link pairs, which should be separated into different clusters. We then perform a grid search over the graph construction and clustering parameters, and select the configuration that achieves the best balance between must-link precision, should-not-link precision, and a reasonable total number of clusters. The selected configuration and its evaluation are reported in Section 4.1. The semantic communities detected by the Leiden algorithm provide the structural basis for constructing the final Concept layer. However, these communities are initially only clusters of related candidate keywords and must be converted into interpretable taxonomy units. For each cluster, we prompt an LLM with the constituent candidate keywords and their field-specific definitions to generate a representative Concept name and a short description. The prompt is designed to avoid overly broad labels and to ensure that the generated name closely reflects the shared meaning of the constituent terms. Through this step, each Leiden community is formalized as a named and described Concept, producing the fine-grained Concept layer that is subsequently attached below OpenAlex Topics. 3.5 Attaching Concepts to OpenAlex Topics After name and description assignment, each Concept is attached to OpenAlex Topics. First, the name and description are used to generate a vector representation of the Concept [Singh et al., 2023]. This representation is then compared with the vector representations of OpenAlex Topics using cosine similarity, and only the most similar Topics are retained as candidates for hierarchical attachment. This retrieval step narrows the search space before LLM evaluation, avoiding the need to ask the LLM to consider all OpenAlex Topics. GPT-4o-mini then evaluates this limited candidate set and selects the most appropriate Topic based on the Concept and disciplinary context. When a Concept matches more than one Topic above a similarity threshold, all matching Topics are retained rather than only the single best one. The selected links are finally checked to reduce redundant or incoherent associations. At the end of this stage, each Concept is connected to one or more OpenAlex Topics and receives a complete path within the taxonomy. In this way, the OpenAlex hierarchy is extended to include a fifth level comprising approximately 114,000 Concepts. Fig. 2 illustrates the final five-level taxonomy and, using Industry 5.0 as an example, shows how a Concept is integrated into the OpenAlex hierarchy. OpenAlex hierarchy extended by SCALEExample: Industry 5.0 Domains 4 Physical Sciences Fields 26 Engineering Subfields 252 Electrical and Electronic Engineering Topics 4,516 Advanced Data and IoT Technologies Concepts 113,892 Industry 5.0 one of three possible paths new level Figure 2: Five-level taxonomy extending the OpenAlex hierarchy with Concepts. 3.6 Taxonomy Maintenance Scientific literature evolves rapidly, so the taxonomy must be updated to reflect emerging scientific trends. Our maintenance procedure starts by identifying emerging keywords from the MDPI keywords in Section 3.1. A keyword is 6 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension defined as emerging when the gain of its recent frequency percentile relative to its historical percentile exceeds a fixed threshold. We then conduct similarity comparisons both among emerging keywords and between emerging keywords and existing Concepts, in order to filter out redundant terms and identify keywords that may serve as new Concepts in the taxonomy. The remaining candidates are processed through a granularity filtering step similar to Section 3.2. Candidate Concepts with proper granularity are then attached to OpenAlex Topics following Section 3.5. Each new Concept is tagged with a version label identifying the update in which it was introduced. This incremental approach keeps the taxonomy up to date while preserving the stability of the existing structure and providing a foundation for future applications such as article classification. 4 Results This section reports the selected graph-clustering configuration, its performance on curated must-link and should-not- link pairs, the resulting Concept layer, and its integration into the OpenAlex hierarchy. The Taxonomy Explorer and Atlantis interfaces are then presented, together with the public release of the taxonomy dataset. Downstream evidence of operational use and expert assessment of Concept applicability is also reported. 4.1 Clustering Optimization and Hierarchical Integration The quality of the generated Concepts depends on the parameters that control graph construction and clustering. To identify the final configuration, the optimization procedure tested three parameters through grid search: the number of nearest neighbors, the minimum similarity threshold, and the Leiden resolution parameter. The evaluation used manually curated pairs of candidate terms. Must-link pairs represented terms expected to belong to the same semantic community, whereas should-not-link pairs represented terms expected to belong to different communities. The pairs focused on difficult boundary cases observed during clustering development, rather than on obvious synonym or unrelated pairs. Must-link precision was defined as the proportion of must-link pairs assigned to the same community. In contrast, should-not-link precision was defined as the proportion of should-not-link pairs assigned to different communities. Overall accuracy was defined as the proportion of correctly classified pairs among all evaluated pairs. Fig. 3 reports the results of the grid search. Each point represents a specific parameter configuration, with must-link precision on the x-axis and should-not-link precision on the y-axis; bubble size indicates the number of detected communities, corresponding to candidate Concepts. The dashed line marks configurations in which the two precision values are equal. The final configuration was selected by jointly considering the two precision scores and the resulting number of communities. The highlighted point represents the configuration adopted for constructing the taxonomy. 7 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension Figure 3: Performance of the clustering configurations evaluated during grid search. The highlighted point represents the final parameter configuration adopted by the framework.K: the number of nearest neighbors,sim_min: the minimum similarity threshold, leiden_resolution: the Leiden resolution parameter. Based on this evaluation of 68 must-link and 68 should-not-link pairs curated by the authors, the final configuration used k=50, a minimum similarity threshold of 0.85, and a Leiden resolution parameter of 0.6, achieving 69.5% must-link precision and 87.5% should-not-link precision, serving as a consistency check on challenging aggregation and separation decisions rather than a random-sample benchmark of clustering quality. The subsequent clustering and naming stages produced 113,892 Concepts, each representing a specific research theme formed by grouping semantically related candidate concepts within the corresponding Field. After hierarchical attachment, at least one Concept was associated with 98.8% of the 4,516 OpenAlex Topics (4,461 Topics). Because each OpenAlex Topic belongs to a Subfield, Field, and Domain, this attachment gives each Concept a complete position within the OpenAlex hierarchy. The resulting structure extends OpenAlex with a fifth level while preserving its original organization. When semantically appropriate, the taxonomy can place a Concept under more than one Topic, allowing interdisciplinary research themes to appear across multiple disciplinary paths. At the Domain level, the 113,892 Concepts have the following distribution across the OpenAlex hierarchy: Physical Sciences accounted for 65,615 Concepts (57.6%), followed by Social Sciences with 18,814 (16.5%), Health Sciences with 15,437 (13.6%), and Life Sciences with 14,026 (12.3%). 4.2 Public Release and Exploration Tools The generated taxonomy is made available through three access channels: the Taxonomy Explorer for hierarchical navigation, Atlantis for map-based exploration of Concepts, and an open dataset containing the full hierarchy and Concept descriptions. 8 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension 4.2.1 Taxonomy Explorer To demonstrate the practical use of the generated taxonomy, a prototype web application, named Taxonomy Explorer, was developed. The application allows users to navigate the five-level hierarchy, from Domains to Concepts, and explore the organization of research areas in greater detail than the original OpenAlex taxonomy. Fig. 4 shows the Taxonomy Explorer interface. The left panel provides a radial visualization of the hierarchy, allowing users to move across Domains, Fields, Subfields, Topics, and Concepts. When users select an item in the hierarchy, the right panel displays the corresponding taxonomic path, together with the description and related Concepts. Figure 4: Interface of the Taxonomy Explorer prototype for navigating the five-level taxonomy from Domains to Concepts. In addition to hierarchical navigation, the Taxonomy Explorer provides search, filtering, and export functions. A research prototype is available online, with access credentials available upon request from the corresponding author. 2 4.2.2 Atlantis Atlantis provides a semantic map of the Concept layer. In Atlantis, each point corresponds to one Concept in a low-dimensional layout obtained from Concept embeddings. The interface supports 2D and 3D views, search, and coloring by Domain or Field. Selecting a Concept displays its label, explanation, and hierarchical path, and retrieves papers tagged with that Concept. The map is intended for navigation and inspection. Distances in the projected space should not be interpreted as a formal measure of taxonomic relatedness. Taxonomy Explorer and Atlantis therefore address different tasks: the former exposes the explicit hierarchy, while the latter shows Concepts in embedding space. Atlantis is publicly available without registration or password. 3 Fig. 5 shows the interface of Atlantis. 2 https://taxonomy.di.mdpi.ai/ 3 https://atlantis.mdpi.com 9 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension Figure 5: Interface of Atlantis. Each point is a Concept; selecting a Concept shows its taxonomic path and linked papers. 4.2.3 Open taxonomy dataset A versioned snapshot of the taxonomy is released as an open dataset on Hugging Face. 4 The package includes tables for Domains, Fields, Subfields, Topics, and Concepts, a flat hierarchy view, Concept explanations, and stable identifiers. Levels 1–4 follow the OpenAlex hierarchy; level 5 contains the Concepts produced by SCALE. Explanations are short AI-generated descriptions intended for browsing and labeling, not expert-curated definitions. The release contains taxonomy metadata only. Paper full texts, abstracts, author lists, and human-evaluation microdata are not included. The dataset is distributed under a C0 1.0 license. Table 1 summarizes the released resource. Table 1: Size of the released taxonomy snapshot. LevelCount Domains4 Fields26 Subfields252 Topics4,516 with≥ 1 Concept4,461 (98.8%) Concepts113,892 4.3 Downstream Use and Expert Applicability Two additional forms of evidence are reported for the released taxonomy: operational use in paper classification, and an expert evaluation of the Concepts returned for sampled papers. The taxonomy has been used in a production paper classification pipeline for the MDPI corpus of nearly two million papers. 94% of Concepts were assigned to at least one paper. Papers received a median of 5 Concepts (IQR 5–6). A human evaluation was conducted with 22 domain experts across 11 fields (110 papers; 220 paper×evaluator units; 3,958 Concept judgments). For each paper, experts inspected Concepts returned by two tagging configurations based 4 https://huggingface.co/datasets/mdpi-di/taxonomy 10 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension on different language models, GPT-5.4 mini and Qwen3 [OpenAI, 2026, Yang et al., 2025], and judged each returned Concept as relevant or not relevant to the paper. Because production tagging returns a median of five Concepts per paper, applicability is reported as precision@5. Across all units, mean precision@5 was 88.9% for GPT-5.4 mini and 88.7% for Qwen3. Precision@1 was 94.5% and 95.9%, respectively. Precision@5 varied by field (Table 2), ranging from 75.0% to 97.0% for GPT-5.4 mini and from 79.0% to 98.0% for Qwen3. Agreement between the two experts assigned to the same field was examined on Concepts rated by both. Observed agreement wasp o = 0.76. Under high positive prevalence (TRUE rate= 0.82), prevalence-adjusted coefficients were moderate (PABAK= 0.52, Gwet AC1= 0.66), while Cohen’sκwas low (0.19) [Byrt et al., 1993, Gwet, 2008, Cohen, 1960]. Table 2: Precision@5 by evaluation field (expert-study labels; not OpenAlex Fields). Values are means over paper×evaluator units. n is the number of papers per field. Overall is the mean over all paper×evaluator units. FieldnP@5(GPT)P@5(Qwen) Physics1075.085.0 Materials Science 11078.079.0 Mathematics1080.881.0 Atmospheric/Environmental Science1090.085.0 Engineering1090.088.0 Botany/Plant science1091.088.0 Drug Development and Delivery1091.095.0 Molecular and Cellular Biology1093.087.0 Materials Science 21095.093.0 Biochemistry1097.098.0 History1097.097.0 Overall11088.988.7 5 Discussion The results demonstrate the feasibility of constructing a fine-grained Concept layer below OpenAlex Topics at scale. The selected configuration achieved 69.5% must-link precision and 87.5% should-not-link precision, while the final hierarchy associated at least one Concept with 98.8% of OpenAlex Topics. These findings support the broad hierarchical integration achieved by the framework, although the lower must-link score indicates that some semantically related candidate terms remain separated across communities. 5.1 Scientific Contribution The main contribution of this work is the introduction of Concepts as a new unit for organizing scholarly knowledge. Unlike individual author keywords, Concepts consolidate semantically related candidate terms into named units located within explicit disciplinary paths. They are designed to provide a finer-grained and more reusable representation of specialized research areas than OpenAlex Topics. Existing taxonomy construction approaches primarily organize terms or documents from text corpora, which makes them difficult to integrate with established scholarly knowledge organization systems [Zhang et al., 2018, Lee et al., 2022, Kargupta et al., 2025]. In contrast, SCALE extends OpenAlex through an automated taxonomy construction pipeline that discovers fine-grained Concepts and links them to the existing topic hierarchy. SCALE also differs from the OpenAlex Keywords system [OpenAlex, d]. Whereas OpenAlex Keywords are derived from Topics and assigned to individual works, SCALE Concepts introduce a separate, finer-grained structural level that is constructed bottom-up from author keywords and integrated below the existing Topic hierarchy. This design enables SCALE Concepts to capture more specific research areas and to be updated as new scientific trends emerge. SCALE helps reduce the gap between broad disciplinary classifications and highly specific document-level descriptions. The result is not only a quantitative extension of the taxonomy, but also the introduction of a conceptual layer designed to support more detailed analyses of the scientific literature. Concepts may therefore contribute to new applications in scientometrics and the monitoring of research evolution. 11 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension 5.2 From Taxonomic Tagging to Ontological Extension A first operational application of SCALE is its use for systematically tagging articles published by MDPI and indexed in Scilit, an MDPI-developed scholarly discovery and indexing platform that aggregates metadata for scientific publications [Scilit, 2026]. In operational use, papers are typically associated with around five Concepts, consistent with the production default. The tagging pipeline is deployed across the MDPI corpus of nearly two million papers. 94% of Concepts appear in at least one assignment, indicating that the taxonomy is already used broadly rather than only in a narrow subset of branches. Large-scale tagging also supports examination of how Concepts are distributed across disciplines, journals, countries, institutions, and time periods, and thus monitoring of research themes, emerging areas, and interdisciplinary connections in greater detail than classifications based only on Topics. Existing knowledge graph infrastructures may provide a natural environment for this extension. MarmotGraph, developed within EBRAINS, provides infrastructure for managing metadata, controlled vocabularies, ontologies, and semantic relations across scientific data [EBRAINS]. From this perspective, Concepts tagged across Scilit publications could provide a broad scholarly layer to extend such graph-based representations beyond neuroscience and toward a more general organization of scientific knowledge. At a later stage, this direction could support the evolution of the taxonomy toward a richer ontological representation. In the current implementation, Concepts are hierarchical taxonomic units rather than formally defined ontological entities. Future work could decompose Concepts into finer ontological entities and use citation patterns, co-occurrence signals, publication metadata, and other semantic evidence to infer explicit, interpretable relations among them. Some limitations should also be addressed in the future. The must-link and should-not-link Concept pairs used for parameter selection were curated by the authors and may therefore introduce subjective bias, while the quality and conceptual coherence of sampled keyword clustering results would benefit from independent human evaluation. In addition, the current distribution of Concepts across Fields and Domains is uneven; although this may partly reflect differences in disciplinary breadth and publication activity, future work should investigate methods to mitigate this imbalance and improve coverage of underrepresented areas. Overall, SCALE provides a structured and interpretable layer between OpenAlex Topics and individual scientific documents. It supports a more fine-grained organization of scholarly knowledge and offers a concrete basis for future applications in article tagging, scientometric analysis, and ontology development. Acknowledgements The authors thank MDPI for providing the AI infrastructure used in this work. The authors also thank Dietrich Rordorf and Luke Sui for their continuous financial support of this project. Finally, we thank the volunteer editors at MDPI who participated in the user evaluation and generously contributed their time and expertise. References Lutz Bornmann and Rüdiger Mutz. Growth rates of modern science: A bibliometric analysis based on the number of publications and cited references. Journal of the Association for Information Science and Technology, 66(11): 2215–2222, 2015. doi:10.1002/asi.23329. Santo Fortunato, Carl T. Bergstrom, Katy Börner, James A. Evans, Dirk Helbing, Staša Milojevi ́ c, Alexander M. Petersen, Filippo Radicchi, Roberta Sinatra, Brian Uzzi, Alessandro Vespignani, Ludo Waltman, Dashun Wang, and Albert-László Barabási. Science of science. Science, 359(6379):eaao0185, 2018. doi:10.1126/science.aao0185. Jason Priem, Heather Piwowar, and Richard Orr. OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. arXiv preprint arXiv:2205.01833, 2022. URLhttps://arxiv.org/abs/2205.01833. Clarivate. Web of science subject categories, 2025. URLhttps://webofscience.zendesk.com/hc/en-us/ar ticles/27505726032017-Web-of-Science-Subject-Categories. Web of Science Help Center. Updated August 6, 2025. Accessed July 8, 2026. Elsevier. What are scopus subject area categories and asjc codes?, 2024. URLhttps://service.elsevier.com/a p/answers/detail/a_id/12007/supporthub/scopus/ . Scopus Support Center. Updated August 15, 2024. Accessed July 8, 2026. Australian Bureau of Statistics. Australian and new zealand standard research classification (anzsrc), 2020, 2020. URL https://w.abs.gov.au/statistics/classifications/australian-and-new-zealand-standard-r esearch-classification-anzsrc/latest-release . Latest release. Released June 30, 2020. Accessed July 8, 2026. 12 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension OpenAlex. Topics. OpenAlex Help Center, a. URLhttps://help.openalex.org/hc/en-us/articles/24736 129405719-Topics. Accessed: 2026-08-06. Eugenio Petrovich.Science mapping and science maps.Knowledge Organization, 48(7–8):535–562, 2021. doi:10.5771/0943-7444-2021-7-8-535. Nees Jan van Eck and Ludo Waltman. Software survey: VOSviewer, a computer program for bibliometric mapping. Scientometrics, 84(2):523–538, 2010. doi:10.1007/s11192-009-0146-3. Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, and Daniel S. Weld. SPECTER: Document-level repre- sentation learning using citation-informed transformers. In Proceedings of the 58th Annual Meeting of the Associ- ation for Computational Linguistics, pages 2270–2282, Online, 2020. Association for Computational Linguistics. doi:10.18653/v1/2020.acl-main.207. URL https://aclanthology.org/2020.acl-main.207/. David M. Blei, Andrew Y. Ng, and Michael I. Jordan. Latent dirichlet allocation. Journal of Machine Learning Research, 3:993–1022, 2003. URL https://w.jmlr.org/papers/v3/blei03a.html. Maarten Grootendorst. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv preprint arXiv:2203.05794, 2022. URL https://arxiv.org/abs/2203.05794. Chao Zhang, Fangbo Tao, Xiusi Chen, Jiaming Shen, Meng Jiang, Brian Sadler, Michelle Vanni, and Jiawei Han. TaxoGen: Unsupervised topic taxonomy construction by adaptive term embedding and clustering. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2701–2709, New York, NY, USA, 2018. Association for Computing Machinery. doi:10.1145/3219819.3220064. Dongha Lee, Jiaming Shen, SeongKu Kang, Susik Yoon, Jiawei Han, and Hwanjo Yu. TaxoCom: Topic taxonomy completion with hierarchical discovery of novel topic clusters. In Proceedings of the ACM Web Conference 2022, pages 2819–2829, New York, NY, USA, 2022. Association for Computing Machinery. doi:10.1145/3485447.3512002. Priyanka Kargupta, Nan Zhang, Yunyi Zhang, Rui Zhang, Prasenjit Mitra, and Jiawei Han. TaxoAdapt: Aligning LLM-based multidimensional taxonomy construction to evolving research corpora. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29834–29850, Vienna, Austria, 2025. Association for Computational Linguistics. doi:10.18653/v1/2025.acl-long.1442. URL https://aclanthology.org/2025.acl-long.1442/. Angelo Antonio Salatino, Thiviyan Thanapalasingam, Andrea Mannocci, Aliaksandr Birukou, Francesco Osborne, and Enrico Motta. The computer science ontology: A comprehensive automatically-generated taxonomy of research areas. Data Intelligence, 2(3):379–416, 2020. doi:10.1162/dint_a_00055. Francesco Osborne and Enrico Motta. Klink-2: Integrating multiple web sources to generate semantic topic networks. In The Semantic Web – ISWC 2015, volume 9366 of Lecture Notes in Computer Science, pages 408–424, Cham, 2015. Springer International Publishing. doi:10.1007/978-3-319-25007-6_24. OpenAlex. Pricing. OpenAlex Help Center, b. URL https://help.openalex.org/hc/en-us/articles/24397 762024087-Pricing. Accessed: 2026-07-14. Kuansan Wang, Zhihong Shen, Chi-Yuan Huang, Chieh-Han Wu, Yuxiao Dong, and Anshul Kanakia.Mi- crosoft academic graph: When experts are not enough. Quantitative Science Studies, 1(1):396–413, 2020. doi:10.1162/qss_a_00021. OpenAlex. Concepts (deprecated) overview. OpenAlex Developers, c. URLhttps://developers.openalex.org/ api-reference/concepts. Accessed: 2026-07-14. OpenAlex. Keywords. OpenAlex Help Center, d. URLhttps://help.openalex.org/hc/en-us/articles/247 36201130391-Keywords. Accessed: 2026-07-14. MDPI. MDPI: Publisher of open access journals. https://w.mdpi.com/, 2026. Accessed: 2026-07-09. OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence.https://openai.com/index/gpt-4o-mini-advan cing-cost-efficient-intelligence/, July 2024. Accessed: 2026-07-09. OpenAI. Introducing gpt-5 for developers.https://openai.com/index/introducing-gpt-5-for-developer s/, August 2025. Accessed: 2026-07-09. Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, KDD’96, page 226–231. AAAI Press, 1996. Vincent A. Traag, Ludo Waltman, and Nees Jan van Eck. From louvain to leiden: Guaranteeing well-connected communities. Scientific Reports, 9:5233, 2019. doi:10.1038/s41598-019-41695-z. 13 SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension Amanpreet Singh, Mike D’Arcy, Arman Cohan, Doug Downey, and Sergey Feldman. SciRepEval: A multi-format benchmark for scientific document representations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5548–5566, Singapore, 2023. Association for Computational Linguistics. doi:10.18653/v1/2023.emnlp-main.338. URL https://aclanthology.org/2023.emnlp-main.338/. Yu. A. Malkov and D. A. Yashunin. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):824–836, 2020. doi:10.1109/TPAMI.2018.2889473. OpenAI. Introducing gpt-5.4 mini and nano.https://openai.com/index/introducing-gpt-5-4-mini-and-n ano/, 2026. Accessed: 2026-08-07. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao, Shixuan Liu, Shuang Luo, Tianhao Li, Tianyi Tang, Wenbiao Yin, Xingzhang Ren, Xinyu Wang, Xinyu Zhang, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388. Ted Byrt, Janet Bishop, and John B Carlin. Bias, prevalence and kappa. Journal of clinical epidemiology, 46(5): 423–429, 1993. Kilem Li Gwet. Computing inter-rater reliability and its variance in the presence of high agreement. British Journal of Mathematical and Statistical Psychology, 61(1):29–48, 2008. Jacob Cohen. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1): 37–46, 1960. Scilit. Scilit. https://w.scilit.com/, 2026. An MDPI scholarly literature database. Accessed: 2026-08-06. EBRAINS. Marmotgraph. EBRAINS Data, Tools and Services. URLhttps://ebrains.eu/data-tools-servi ces/tools/marmotgraph. Accessed 14 July 2026. 14