BERTopic-Based Topic Modeling and Knowledge Graph Framework for Semantic Mapping of the Scoliosis Literature

Summary
This manuscript presents a large-scale semantic analysis of scoliosis research (2000–2025) by integrating BERTopic-based topic modeling with automated knowledge graph construction. The study analyzes 31,836 Scopus abstracts, derives latent research themes, constructs a topic-aware knowledge graph via automated triplet extraction, performs similarity and temporal analyses, and reports a stratified manual validation of relation extraction quality.
The work aims to bridge neural topic modeling and knowledge representation for semantic bibliometric analysis in a biomedical domain.
(1) Originality
The integration of contextual topic modeling (BERTopic) with knowledge graph construction for domain-level literature mapping is a relevant and timely idea. The explicit attempt to combine latent semantic structure (topics) with explicit relational knowledge is conceptually interesting.
However:
-The topic modeling component relies on established BERTopic methodology without methodological novelty.
-The relation extraction pipeline is described at a high level and does not introduce a novel extraction framework.
-The knowledge graph is implemented as a property graph “compatible with RDF principles,” but no formal semantic modeling (RDF/OWL/SHACL) or ontology alignment is provided.
Thus, the originality lies primarily in the application-level integration rather than in a novel Semantic Web method and it is considered as "moderate".
(2) Significance of the Results
The empirical analysis (topic structure, similarity network, temporal evolution, and relation validation) provides a useful macro-level view of scoliosis research. The validation protocol (relation-wise sampling with explicit TRUE/FALSE/AMBIGUOUS categories) is a positive aspect and shows methodological awareness.
However, the following factors limit the broader significance:
-The relation extraction methodology is underspecified, limiting confidence in reproducibility.
-Evaluation is limited to precision (excluding ambiguous cases); no recall estimation, inter-annotator agreement, or confidence intervals are provided.
-The knowledge graph is not formally grounded in existing biomedical ontologies and is not demonstrated as interoperable or "queryable" in a Semantic Web sense.
-Threshold selection for similarity network construction is heuristic and not supported by sensitivity analysis.
The results are informative as a semantic case study, but the Semantic Web contribution seems to be underdeveloped and without formal semantic grounding.
(3) Quality of Writing
The manuscript is clearly structured and generally well written. The methodological flow (data → topics → similarity → temporal analysis → knowledge graph → validation) is coherent and easy to follow.
Few minor issues include:
-A typographical error at the beginning of Section 3.4 (“o examine how…”).
-A likely mismatch between Figure 8 in-figure title (“Property Graph: Author-Institution-Topic-journal…”) and its caption describing a topic-aware semantic knowledge graph.
Overall, the writing quality is good and the paper is readable.
(4) Assessment of Data and Resource Availability
The manuscript was submitted as a full paper and, therefore, requires assessment of the “Long-term stable URL for resources.” The manuscript and submission materials were carefully reviewed but any repository link, DOI, data file, or supplementary resource corresponding to this requirement were not located.
Given that the manuscript presents a computational pipeline involving topic modeling, relation extraction, and knowledge graph construction, the absence of reproducibility artifacts significantly limits transparency and replicability. Even if raw Scopus data cannot be redistributed due to licensing restrictions, derived artifacts (e.g., extraction scripts, configuration parameters, relation schema, annotated validation samples, or graph snapshots) should be made available.

Overall Recommendation: Major Revision
The paper is promising as a large-scale semantic analysis and has a clear structure and reasonable validation effort. However, to meet the standards of a full research contribution in Semantic Web Journal, the authors should:
-Provide reproducibility resources in a stable repository,
-Clarify and fully specify the relation extraction methodology,
-Strengthen evaluation reporting (including reliability measures),
-Improve semantic formalization and interoperability of the knowledge graph.
Addressing these issues would substantially improve the technical robustness and Semantic Web relevance of the work.

(1) Originality

The paper studies how topic modeling and knowledge graph techniques can be used to analyse scoliosis research literature. The authors collect 31,836 abstracts from Scopus published between 2000 and 2025 and analyse them using BERTopic. They then extract entities and relations from the text and construct a graph that represents semantic relationships in the literature. The study also examines topic similarity and topic evolution over time.

The idea of combining topic modeling with knowledge graph construction is interesting and relevant to the topic of the special issue, which focuses on bridging machine learning and knowledge representation. The paper tries to connect latent topics discovered through machine learning with explicit relations extracted from text.

However, the individual methods used in the study are already well known. BERTopic, embedding models, and relation extraction pipelines are widely used in text mining research. The novelty mainly comes from combining these methods and applying them to scoliosis literature rather than proposing a new method or algorithm.

The paper also describes the constructed structure as a knowledge graph, but the manuscript does not clearly describe a semantic schema, ontology alignment, or RDF representation that is usually expected in Semantic Web research. The graph appears to function mainly as a relation network extracted from text rather than a formally defined knowledge graph supporting semantic reasoning or linked data interoperability.

(2) Significance of the Results

The study uses a large dataset of more than 30,000 biomedical abstracts and focuses on an important clinical research domain. The analysis shows topic clusters in the literature, semantic relations between biomedical concepts, and how topics evolve over time.

The results provide a descriptive overview of scoliosis research themes. However, the paper does not clearly demonstrate how the proposed framework produces insights beyond what traditional bibliometric or topic modeling approaches could produce. The manuscript also does not compare its results with other literature analysis approaches.

The validation of extracted relations is a positive aspect of the work. The authors manually review a sample of extracted triplets and report precision values for several relation types. However, the validation is limited to a small sample size and reports only precision. The study does not report recall or inter-annotator agreement. Because of this, the evaluation measures relation extraction accuracy but does not fully assess the quality of the knowledge graph itself.

The integration between machine learning and knowledge representation also appears mostly sequential. Topic modeling identifies thematic clusters, and relation extraction is then used to populate the graph. The manuscript does not demonstrate deeper interaction between topic representations and the knowledge graph structure.

(3) Quality of Writing

The manuscript is generally well structured and follows a clear organization with introduction, methods, results, and discussion sections. The motivation of the study is understandable and the limitations are partly acknowledged.

However, some parts of the methodology lack important technical details. The paper does not clearly report the configuration of the BERTopic model, the exact embedding model used, or the clustering parameters. The relation extraction process is described only at a high level and the specific rules or dependency patterns are not fully explained.

Several figures are also difficult to interpret.

For example:

Figures 2 and 3 show topic modeling results, but the topic labels are small and partially truncated. Because of this, it is difficult for the reader to clearly identify the themes represented by each topic cluster.

Figures 4 and 5 present the topic similarity network and the topic dendrogram. The paper states that these figures reveal a “core–periphery structure” of the research field, but the visualization does not clearly show which topics belong to the core and which belong to the periphery. The threshold used for constructing the similarity network is also not explained in detail.

Figure 6 shows topic prevalence over time, but the dataset is described as covering the period 2000–2025 while the Sankey diagram in Figure 7 presents topic flows from 2000–2004 up to 2020–2024. The time ranges used in these figures are not fully consistent and should be clarified.

Figure 8 is described as a “topic-aware knowledge graph” where nodes represent topics and edges represent semantic relations derived from extracted triplets. However, the methods section explains that the knowledge graph is constructed from extracted entities and relations in the form of head–relation–tail triplets. This creates confusion about whether the nodes in the graph represent topics or biomedical entities.

Clearer figure labeling and more precise explanations would help readers interpret the results.

Assessment of Data and Resources
(A) Organization of the data file and README

The reviewer guidelines request evaluation of the data provided under the “Long-term stable URL for resources”. In the material provided for review, a repository link or downloadable dataset is not clearly visible. Because of this, it is not possible to evaluate whether the data files are well organized or whether a README file is provided to explain the dataset and preprocessing steps.

(B) Completeness of resources for replication

Since the repository and resources are not accessible in the review material, it is not possible to determine whether the provided resources are sufficient to reproduce the experiments. Replication would require access to the dataset, preprocessing scripts, topic modeling configuration, and the relation extraction pipeline.

(C) Repository and long-term discoverability

Because the repository location is not clearly available in the review material, it is not possible to assess whether the chosen repository ensures long-term discoverability. If the authors provide a repository, it should ideally be hosted on a stable platform such as GitHub, Zenodo, or Figshare and include documentation describing the dataset and the experimental workflow.

(4) Completeness of data artifacts

The completeness of the data artifacts cannot be evaluated because the dataset and associated resources are not accessible in the review materials. Clear access to the dataset and code would allow proper assessment of the reproducibility of the study.

Summary
------

The paper proposes an integrated semantic framework for analyzing the scoliosis research literature. Starting from 31,836 English-language abstracts indexed in Scopus (2000–2025), the authors apply BERTopic to identify latent research themes, construct a topic-aware property graph through automated head-relation-tail extraction, quantify inter-topic semantic proximity via cosine similarity on centroid embeddings, and trace temporal trends in knowledge intensity across year-binned periods. A stratified validation of 250 extracted triplets (50 per predicate across five relation types) yields precision estimates ranging from 0.78 (Treats) to 0.92 (Associated_with). The integration of implicit topic structure with explicit relational representation is presented as the paper's primary methodological contribution, with scoliosis serving as the application domain. The motivation is sound and the corpus scale is a genuine asset. Overall, I found the paper interesting and well organized. However, my principal reservations concern the systematic omission of quantitative details that make reproducibility impossible, a validation design that does not establish the credibility of its precision claims, the absence of any comparative baseline, a purely qualitative temporal analysis, and the near-total absence of the resource metadata required by the journal's long-term stable resource criteria. None of these are insurmountable, but they must be addressed before the paper can be accepted.

Strengths
------

The corpus scale - 31 836 abstracts spanning 25 years - is a genuine asset and positions the work well as a large-scale semantic analysis rather than a proof of concept. The decision to combine topic modeling with explicit relational extraction is conceptually well motivated: the paper correctly identifies that distributional representations alone do not encode typed semantic links, and that knowledge graphs alone do not capture latent thematic organization. The validation design distinguishes explicitly between incorrect and ambiguous relations rather than conflating uncertainty with error, which is methodologically honest and aligned with best practices in information extraction evaluation. The temporal perspective - examining how thematic intensity and diversity evolved across decades - adds a longitudinal dimension that is absent from most prior semantic bibliometric work. The Sankey diagram (Figure 7) is an effective visualization device that integrates time, topic, and journal dimensions in a single view.

Major limits
------

1. Several methodological parameters that are essential for reproducibility are either missing or insufficiently emphasized in the manuscript. In particular, the final number of retained topics produced by BERTopic is not explicitly stated in the main text, the sentence embedding model is described only generically ("a sentence-level transformer embedding model"), the total size of the resulting knowledge graph is not reported, and the cosine similarity threshold used in Section 2.4 is discussed qualitatively without providing its numerical value. Likewise, the temporal aggregation used in Section 2.5 can only be inferred from Figure 7 rather than being explicitly described. Reporting these parameters explicitly would greatly improve the reproducibility of the proposed framework.

2. Section 2.6 mentions "annotators" in the plural but never states how many annotators evaluated the triplets, what their domain expertise was, or whether inter-annotator agreement was measured. For a paper presenting precision values as evidence of moderate-to-high reliability, this is a fundamental gap. A single judge with high familiarity with the source material would inflate precision on precisely the relation types that are most ambiguous. Authors should report the annotation design (number of judges, qualification, instructions), and provide at least one agreement metric (Cohen’s κ or Krippendorff’s α) on a common subset. If the validation was effectively single-annotator, this must be acknowledged as a limitation bearing directly on the precision figures.

3. The reported accuracy values are difficult to interpret without further context. Although relation-specific accuracy values are provided (0.78-0.92), the article does not explain how these results compare with those reported in previous biomedical relation extraction studies or in comparable semantic exploration frameworks from the literature. Such a comparison would allow readers to understand whether the reported performance reflects the expected level of accuracy for this type of task or represents a significant improvement. Even if a direct experimental comparison is not possible, a discussion situating these results within the existing literature would greatly strengthen the evaluation.

4. Section 3.4 reports observations such as "a gradual but consistent increase" and "accelerated growth trajectories" without any statistical test (e.g., Mann-Kendall, Spearman rank correlation, or change-point detection). The "intensity proxy" defined as the count of extracted triplets per bin is novel but not validated against any external measure of knowledge richness. Without statistical backing, the temporal claims reduce to visual inspection of Figure 6. Authors should apply appropriate trend-detection methods to distinguish genuine directional trends from noise, and should validate or at least discuss the assumptions underlying the triplet-count proxy.

5. Section 2.3 describes the extraction technique as "rule-based and dependency-aware natural language processing technique" without naming the NLP framework (spaCy, CoreNLP, etc.), the dependency patterns or rules used to identify each relation type, or any error analysis at the extraction stage prior to validation. The five relation types (Treats, Used_for, Associated_with, Measured_by, Predicts) are introduced without operational definitions. Readers cannot assess whether the patterns are specific enough to avoid spurious triplets, whether they are appropriate for abstract-level text, or whether the schema is grounded in any biomedical ontology. A dedicated subsection describing the extraction rules and any ontological grounding (UMLS, SNOMED-CT, MeSH) is needed.

6. The integration between BERTopic and the knowledge graph is well illustrated throughout the manuscript and constitutes the main conceptual contribution of the work. However, its added value over applying topic modelling and knowledge graph construction independently is never quantitatively demonstrated. For example, the manuscript does not investigate whether the topic-aware graph improves the interpretability of the extracted relations, facilitates exploration, or provides better analytical capabilities than a conventional knowledge graph. A dedicated experiment or, at least, a more explicit discussion of the practical benefits brought by the topic-aware representation would considerably strengthen the central contribution of the paper.

7. Figures 2 and 4 presumably display topic labels and IDs, but the text of the paper never states how many topics BERTopic produced, what labels were assigned to them, or what the five largest topics are. Figure 7 references labels T0-T10, implying at most 11 topics, which is inconsistent with the "top-20 topics" used for the similarity analysis in Section 3.3. Authors should include a table listing all topics (ID, label, document count, coherence score) and resolve the apparent inconsistency between the top-20 mention and the T0-T10 labeling in Figure 7.

Minor limits
------

The sentence opening Section 3.4 reads "o examine how scoliosis research has evolved" - the initial "T" is missing. This must be corrected.

The abstract states that validation "demonstrated moderate-to-high precision across relation categories" without giving the numerical range. Given that the abstract is the only part of the paper many readers will see, the range (0.78-0.92) should be reported there.

Section 2.4 says the threshold was chosen based on "empirical inspection of network density, connectedness, and the emergence of visually and analytically coherent communities." This is a reasonable heuristic, but the specific threshold value and the resulting network density (edges retained/total possible edges) should be stated so that other researchers can reproduce the network.

The outlier category (topic −1) is mentioned briefly: "outlier documents were excluded from subsequent analyses." The number of outlier documents should be reported alongside the 31,836 total so readers know what fraction of the corpus was excluded.

The co-authorship network analysis (Section 3.5 and Figure 9) is introduced and briefly described but then plays no role in the discussion or conclusion. If it is not analytically central, it should either be developed into a substantive analysis with network statistics (degree distribution, clustering coefficient, identified communities) or removed to sharpen the paper's focus.

Assessment of Long-term Stable Resources
------

A. Availability. I could not identify any persistent URL or repository associated with the proposed resource in either the manuscript or the supplementary material. Section 2.6 notes that preprocessing and validation scripts were implemented in Python, but no code repository is mentioned. Without a publicly accessible location, the primary resource claimed by the paper is inaccessible to readers. A citable, stable repository (e.g., Zenodo, GitHub with a tagged release, or an institutional SPARQL endpoint) must be provided before the resource can be assessed.

B. Licensing. No license is specified for any component — the knowledge graph, the code, or any derived dataset. Given that the underlying abstracts are Scopus-indexed and may be subject to commercial copyright, authors must clarify what they are actually permitted to distribute and under which license.

C. Standards compliance and format. Section 2.3 states that the property graph was designed to be "compatible with Resource Description Framework (RDF) principles," but no actual RDF serialization is provided, no ontology or vocabulary alignment is described (no SKOS for topics, no UMLS for relation types), and no SPARQL endpoint or VOID descriptor is mentioned. For a paper submitted to the Semantic Web Journal, the gap between "designed to be compatible" and "serialized and queryable as RDF" is significant. Authors should either provide a concrete RDF dump with documented prefixes and schema, or revise the resource claims accordingly.

Typos and Editorial Suggestions
------

"o examine how scoliosis research has evolved"  "To examine how scoliosis research has evolved" (Section 3.4, first sentence)
The reference list mixes DOI formats: some entries use "doi: 10.XXX" and others "Available: https://arxiv.org/..." — standardize to one format per entry type.
Reference [9] (Angelov, Top2Vec) is cited in Section 2.2 to support BERTopic's use of contextual embeddings but Top2Vec is a competing method, not supporting evidence; replace with a more appropriate citation for transformer-based topic modeling advantages.

Overall Impression
------

The paper addresses a real need - scalable semantic analysis of domain-specific biomedical literature - and the combination of BERTopic with knowledge graph construction is conceptually appealing. The corpus scale and the explicit handling of ambiguity in validation are genuine positives. However, the paper currently presents the framework at a level of description that cannot support the reproducibility, comparability, or resource-contribution claims that a Semantic Web Journal submission requires. The omission of basic quantitative parameters (topic count, embedding model, KG size, threshold) across every component is systematic rather than incidental, and it must be addressed comprehensively. The validation, as described, is insufficient to establish the reliability of the extraction pipeline to the standard expected for a resource paper. The paper would benefit substantially from a revision that adds one comparative experiment, tightens the resource description to meet FAIR standards, and completes the missing quantitative reporting.