Utilizing BERTopic modeling for concept discovery in the domain of gerotranscendence and solitude Authors: B. Damayanthi Jesudas, Finn Wilson, Rachel A. Mavrovich, Sean Kindya, Feng-Yu Yeh, Sam Smith, Jeremy Ravenel, Jie Zheng, Yongqun He, Hollen N. Reischer, Julie C. Bowker, John Beverley, William D. Duncan Published: 2026 ABSTRACT BackgroundOntology development is a complex, iterative process that traditionally requires extensive collaboration between ontology developers and subject matter experts. While effective, this manual approach is time-consuming, labor-intensive, and prone to cognitive bias. To streamline early-stage ontology development and uncover concepts that might be overlooked through manual review alone, we applied automated topic modeling with BERTopic to extract topics, keywords, topic labels, and summaries from The Handbook of Solitude: Psychological Perspectives on Social Isolation, Social Withdrawal, and Being Alone and Gerotranscendence: A Developmental Theory of Positive Aging. The extracted topic labels after review were used as candidate concepts for the Promoting Healthy Aging through Semantic Enrichment of Solitude Research (PHASES) Ontology.MethodsWe implemented and compared two BERTopic pipelines: (1) the default pipeline and (2) a custom preprocessing pipeline incorporating part-of-speech filtering and n-gram tuning. The pipeline is customized to flexibly extract any specified number of topics and keywords based on user-defined parameters. To compare and merge topic modeling outputs across solitude and gerotranscendence, we used semantic embeddings of topic labels, keywords, and summaries from the custom pipeline. Cosine similarity identified semantically matched topic pairs above a set threshold, enabling categorization and integration into a merged conceptual framework that bridges both domains.ResultsFrom the solitude corpus, BERTopic generated 244 initial topics, which, after SMEs review, were refined to 32 high-quality topics with the custom pipeline and 46 with the default pipeline. For the gerotranscendence corpus, the pipeline produced 172 initial topics, refined to 33 with the custom pipeline and 32 with the default pipeline. Across both corpora, BERTopic contributed 90 ontology terms, 52 from the solitude corpus and 38 from the gerotranscendence corpus. Visual evaluations, including keyword score bar charts, hierarchical clustering dendrograms, and BART-generated summaries, revealed that the custom pipeline produced more fine-grained, domain-specific topics, while the default pipeline offered broader thematic coverage and clearer labels. Certain theory-laden concepts, however, required interpretive input from SMEs.ConclusionsBERTopic provided an efficient, semi-automated approach for identifying candidate ontology terms from domain literature, supporting both breadth and specificity in concept capture. The topics produced by BERTopic modeling uncover underlying latent concepts, serving as a link between raw text data and structured ontology classes. By facilitating the formation of the hierarchical relationships, it exposes a layered semantic structure that is essential for ontology development. Further, integrating semantic similarity analysis across thematic domains revealed conceptual intersections and overlaps, enhancing the semantic foundation of the PHASES Ontology and offering a replicable method for cross-domain ontology development. Introduction Ontology development is often an iterative process that involves close collaboration between ontology developers and subject matter experts (SMEs) to identify and represent domain-relevant terms, including both concepts and relations among them. In a standard workflow, ontology developers work with SMEs to identify high-value terms for a given domain, which are then modeled using ontology techniques. The resulting models are evaluated by SMEs for domain accuracy and reviewed by knowledge engineers, who have training in the formal logics of ontology languages, to identify and correct any modeling errors [1]. Afterwards, the resulting models are evaluated by SMEs. Based on the feedback from SMEs, refinements are then made, and new terms are proposed. Many successful ontologies have been created this way, but there are drawbacks. It is labor intensive, often taking substantial time and effort to generate a core set of concepts. The process may also leave out important terms due to biases of the SMEs and developers. Recent advances in the field of natural language processing (NLP), such as the rise of large language models (LLMs), have spurred the investigation of how best to use these technologies to facilitate ontology development and address the limitations of manual approaches. DRAGON-AI, for instance, uses LLMs to assist developers in creating textual and logical components [2]. This method leverages Retrieval-Augmented Generation (RAG) to combine the latent knowledge of LLMs with sources like ontologies. While these tools can be very useful in curating ontology content, intervention from SMEs can yield more thematically coherent and domain specific terminology due to their practical experience. In this manuscript, we applied recent advances in automated text analysis to address limitations in manual approaches to domain relevant term identification for ontology development in the Promoting Healthy Aging through Semantic Enrichment of Solitude Research (PHASES) project. The project pertains to solitude, or the state of being alone and away from others, and gerotranscendence, a developmental shift in perspective toward greater personal coherence and acceptance, interpersonal and cosmic connectedness, and decreased materialism in later life [3]. Interestingly, both constructs are relevant for healthy aging, but have typically been studied separately, within different subdisciplines of psychology and/or health-relevant areas (e.g., public health, nursing). However, developing a structured, semantically rich ontology encompassing both psychological constructs could bridge these domains, facilitate integration with clinical terminologies, and support computational analyses aimed at advancing mental health care and quality‑of‑life research for aging populations. We employed BERTopic [4], an advanced topic modeling approach based on Bidirectional Encoder Representations from Transformers (BERT)-based embeddings to extract topics and associated keywords from one selected handbook on solitude and one selected book on gerotranscendence, chosen in consultation with SMEs on the PHASES project, largely for their rich and conceptually diverse domain content. The topics generated from the BERTopic modeling provide underlying latent concepts, bridging gap between raw text data and structured ontology classes. It also supports in creating hierarchical relationships revealing a layered semantic structure needed for ontology development. Two BERTopic pipelines were evaluated: one, a default pipeline parameterization and another, a custom preprocessing pipeline incorporating part-of-speech (POS) filtering and n-gram tuning to enhance conceptual specificity. The outputs from both pipelines were subsequently reviewed by SMEs to eliminate non-essential terms and refine the remaining results. We then compared the curated BERTopic derived terms with those obtained through the traditional SME-developer interaction method. To facilitate semantic integration, we implemented a cross-corpus semantic comparison approach. We proposed a data-driven method for discovering relevant themes in solitude and gerotranscendence, which can reveal important concepts in each domain that traditional ontology development workflows might overlook. This approach employed sentence embeddings and cosine similarity to identify thematically related topics between the two corpora. The resulting strongly aligned conceptual pairs were proposed as candidates for higher-level merged concepts, contributing to a more interconnected and semantically coherent ontology structure. Our results support this view, revealing 90 novel topics that had not been previously identified. The topics generated through BERTopic modeling reveal underlying latent concepts, acting as a bridge between raw text data and structured ontology classes. By enabling the establishment of hierarchical relationships, it uncovers a layered semantic structure that is crucial for ontology development. A limitation of this approach is its reliance on BERTopic modeling, which may introduce biases or omit nuanced concepts depending on parameter setting and corpus characteristics. BERTopic outputs can also include semantically light words, repetitive phrases, and citation-derived author names, particularly in smaller topic clusters. While POS filtering and preprocessing reduce much of this noise, further refinement through lemmatization, deduplication, and domain-specific stopword lists can improve topic interpretability. Background literature review The use of LLMs in ontology development addresses the need for semantically rich, context-aware knowledge representations across domains, enabling precise generation of domain-specific knowledge and facilitating complex, interoperable ontologies. For instance, the development of the PHASES Ontology within the healthy aging domain includes psychological constructs such as solitude and gerotranscendence, relevant to mental health, quality of life, and aging-related determinants. Topic and language models support ontology population and refinement to explore healthy aging facets. A 2023 BERTopic analysis of 44,343 Weibo posts identified four public concerns about active aging in China post‑COVID [5], but platform bias and exclusion of less digitally literate elders highlight the need for diverse data sources. Longitudinal modeling of 5610 “successful aging” abstracts from 1963 to 2023 revealed persistent focus on health and social domains, alongside recent growth in mental health, physical activity, and social participation [6], though limited by abstract-only data and lack of discipline-specific semantic granularity. Analysis of 63,809 tweets using BERT-based Named Entity Recognition (NER) and BERTopic captured semantically coherent topics on public perceptions of healthy ageing [7] but missed topics such as ‘frailty’ and ‘elder abuse,’ reinforcing the value of curated scholarly corpora for ontology development. A 2024 study proposed an explainable ML-based healthy aging scale built from survey data of 696 Slovenian adults aged above 50 years, with continuous input from gerontology experts [8]. Explanatory Factor Analysis identified physical, mental, and social health constructs, which experts rated via a custom web annotation tool to create ground truth. Six classifiers were tested, with XGBoost achieving the best performance (AUC = 0.92, F1 = 0.76), and SHapley Additive exPlanations (SHAP) was applied for interpretable predictions, though data were limited to a single timepoint. Kuspinar et al. [9] used NLP to identify six key domains, pain, walking, standing, stairs, sleeping, and playing with grandchildren to develop an osteoarthritis-specific Health-Related Quality of Life (HRQL) index from 102 Canadians. BERTopic captured nuanced patient concerns and enabled hierarchical topic merging, though limitations related to data completeness, recruitment approach, regional variation, and methodological validation remain. The application of LLMs in qualitative health research presents both opportunities and challenges, as highlighted by Castellanos et al. [10]. They examined how ChatGPT [11, 12] could augment thematic analysis of healthcare forum data alongside Latent Dirichlet Allocation (LDA) [13]. ChatGPT added depth through subtheme generation, uncovering hidden structures and facets that human coders might miss, but also posed risks like overfitting and misclassification. Limitations included focus on a single nurse forum, reliance on GPT3.5/4, and basic prompting strategies. LLMs acted as ontology engineers, domain experts, and evaluators, converting text into Web Ontology Language (OWL) [14] ontologies and generating outputs such as competency questions, SPARQL [15] queries, axioms, and documentation. Strengths included task mapping, model diversity, and open datasets. Limitations involved inconsistent task definitions, lack of benchmarks, limited reproducibility, and underexplored later Ontology Engineering (OE) phases. Nayyeri et al. [16] introduced Retrieval-augmented Iterative Generation of RDB Ontologies (RIGOR), a framework using LLMs to transform relational schemas into OWL ontologies with minimal human input. Context from schemas, documentation, core, and external ontologies prompts a generative LLM (Gen-LLM) to produce fragments, refined by a Judge-LLM. Applied to medical databases, RIGOR produced consistent, semantically aligned, standards compliant ontologies, outperforming baselines, though occasional issues with ambiguous properties, incomplete equivalences, and inconsistent class typing remained. Topic modeling for ontology development Topic modeling, a widely used statistical technique, uncovers latent thematic structures in large document collections by identifying clusters of related words that represent topics [17]. These themes support concept identification, thematic structuring, domain representation, and ontology evolution, facilitating semi-automated ontology development. Topic modeling enriches ontologies by capturing core concepts that may be overlooked in manual approaches. Fuellen et al. [18] highlight text mining and automated information extraction to populate aging-related databases, capturing explicit facts and inferring implicit knowledge. This supports ontology development by grouping domain-specific concepts and hypothesizing relationships for formalization. Incorporating these methods enables comprehensive, up-to-date, semantically rich ontologies for aging research. Traditional topic modeling methods include LDA, which represents each document as a mixture of latent topics with probability distributions over words; Non-negative Matrix Factorization (NMF), which extracts parts-based representations by decomposing data into additive, nonnegative components [19]; and Latent Semantic Analysis (LSA), which uncovers hidden semantic relationships by reducing term-document data [20]. Egger et al. [21] compared LDA, NMF, Top2Vec [22], and BERTopic on large-scale social media data, showing BERTopic’s superior topic coherence and contextual relevance, highlighting the value of embedding-based approaches. Advances in deep learning and embeddings have led to neural topic models like Neural LDA [23] and hybrid methods combining contextual embeddings with clustering, such as BERTopic and Top2Vec. These methods use pretrained language models to generate dense semantic embeddings capturing contextual relationships between terms and documents. High-dimensional embeddings are projected into lower-dimensional space using Uniform Manifold Approximation and Projection (UMAP) [24], enabling efficient structure discovery, followed by clustering with algorithms like Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) [25, 26] to identify coherent, contextually informed topic groupings. Applications of BERTopic in health and aging research BERTopic integrates transformer-based contextual embeddings with class-based Term Frequency - Inverse Document Frequency (c-TF-IDF) [27] to identify dense, semantically coherent document clusters. By leveraging pretrained language models for high-quality sentence embeddings, BERTopic captures contextual meaning beyond traditional bag-of-words representations. The c-TF-IDF weighting enhances interpretability by highlighting representative keywords. Its flexibility supports dynamic topic modeling (tracking evolution over time), hierarchical topic modeling (topic-subtopic relationships), multimodal modeling (multiple data types), and semi-supervised or guided modeling, which focuses discovery on user-defined thematic areas. Guided modeling is particularly useful for domain-specific applications. BERTopic has been widely applied in health and aging research, often combined with other NLP and ML techniques to improve interpretability and decision support. For example, analysis of therapist speech from YouTube transcripts used BERTopic with LLM-based summaries, Keyword extraction using BERT embeddings (KeyBERT) [28] and expert refinement to identify psychotherapy communication patterns [29]. While GPT-assisted labeling enhanced interpretability, the dataset was small and imbalanced, limiting clinical representativeness. In clinical psychotherapy, 552 transcripts from 124 patients were modeled separately for therapist and patient speech, producing 250 topics each [30]. These were applied in ML models to predict symptom severity and therapeutic alliance, with SHAP [31, 32] providing explainability. Limitations included hyperparameter tuning, topic redundancy, and predictive accuracy. In biomedical literature mining, BERTopic and LDA were applied to 1837 PubMed abstracts on opioid-related cardiovascular risks in women [33]. While BERTopic performed well for clinical text clustering, only one topic per model was manually reviewed, limiting robustness. Recently, Chung et al. [34] used dynamic BERTopic with BERT-based NLP to analyze 1332 stress-related messages from older adults via a mobile health app, identifying evolving clusters such as family stress, financial strain, and health issues, providing longitudinal insight into mental health risks and supporting early depression detection. This demonstrates BERTopic’s adaptability to continuous, time-sensitive public health data streams. Synthesis of gaps and relevance to the present study The limitations observed across prior studies, such as small or biased datasets, limited domain coverage, lack of methodological validation, insufficient interpretive depth for theory‑laden concepts, and restricted integration between topic modeling outputs and ontology engineering, are partly addressed in our approach. By applying BERTopic to two complementary, domain‑specific scholarly books rather than social media or abstract‑only datasets, we mitigate platform bias and ensure richer conceptual coverage. The dual‑pipeline strategy (custom preprocessing vs. default BERTopic) enables comparison between fine‑grained, domain‑specific topics and broader thematic patterns, directly addressing the need for methodological evaluation. SME‑guided refinement counters automated labeling bias and ensures interpretive accuracy for complex theoretical constructs. Our semantic‑embedding‑based topic matching bridges thematic domains, solitude and gerotranscendence, supporting ontology integration, an area often underexplored in earlier work. Remaining gaps to address in future work include systematic benchmarking of pipeline performance, expansion to additional aging‑related corpora for broader generalizability, and incorporation of iterative collaboration between SMEs and LLMs to further streamline ontology concept identification and validation. To operationalize these objectives, the following section details the methodological framework for applying BERTopic pipelines, semantic similarity matching, and expert-guided refinement in the development of the PHASES Ontology. Methods The PHASES Ontology development involves adding and curating concepts through several stages of iteration and refinement. The concepts are generated based on inputs, including discussions with SMEs, Competency Questions (CQs), and insights that emerge from discussions around the CQs. The CQs are natural language questions used in ontology engineering to define the scope, requirements, and intended use of a knowledge model. They capture the kinds of queries the ontology should be able to answer, ensuring relevant concepts and relationships are included, and serve as a validation tool to test whether the completed ontology meets its design objectives [35, 36]. These discussions, involving both SMEs and ontologists, highlight the complex and labor-intensive process in manual ontology development. To address this issue, we propose using NLP techniques, specifically BERTopic modeling, to automate the process of interpretable topic generation from large text corpora which can be identified as ontological concepts. These automatically extracted topics are then reviewed by the SMEs to eliminate any irrelevant terms and ensure domain completeness, resulting in a refined set of concepts for the ontology. By combining BERTopic with SME intervention, the process helps identify overlooked but relevant concepts, minimize bias in topic selection, and maintain comprehensive domain coverage while still reducing some manual effort and supporting a data-driven approach in the early stages of ontology development. Corpus selection To support the development of the PHASES Ontology, including both the Solitude Ontology and the Gerotranscendence Ontology, we selected two fundamental books that comprehensively represent the theoretical themes of their respective domains: The Handbook of Solitude: Psychological Perspectives on Social Isolation, Social Withdrawal, and Being Alone [37] and Gerotranscendence: A Developmental Theory of Positive Aging [3]. As comprehensive academic sources, these provide curated overviews and give insights of the key constructs of their respective domains. This allows conceptually rich and thematically coherent topics serving as a foundation for early-stage ontology development. These books enable the extraction of semantically meaningful topics, keywords, and summaries through the application of BERTopic modeling. Overview of BERTopic pipelines and configuration Figure 1 illustrates the BERTopic custom preprocessing pipeline with the modules used in the methodology. A yaml configuration file is used to define and customize the models and parameter values as shown in Appendix A, for each module, enabling flexible and reproducible customization that provides control over the text processing and modeling steps without modifying the underlying code. The key configuration parameters are en_core_web_sm [38], a lightweight and fast language model used in English NLP tasks within the spaCy [38] library. Term Frequency - Inverse Document Frequency (TF-IDF) [39, 40] is used for vectorization along with settings for stopword removal and n-gram range, unigrams, bigrams and trigrams. The all-MiniLM-L6-v2 [41] model, a compact variant of BERT with 6 transformer encoder layers, is used with SentenceTransformer [42] and KeyBERT. For dimensionality reduction, UMAP is used. It reduces the high-dimensional BERT embeddings while preserving data structure, which helps in enhancing clustering by revealing patterns and reducing noise. HDBSCAN then identifies dense clusters, handling varied shapes and labeling outliers as noise. After clustering, BERTopic modeling is applied to generate coherent topic representations. These topics are further refined through topic labeling using KeyBERT, and finally Bidirectional and Auto-Regressive Transformer (BART) [43] summarization is used to generate concise and readable descriptions of each topic. The resulting BERTopic results containing the topic labels, keywords, and summaries from the solitude and gerotranscendence domains are used to find the overlapping topic labels from solitude and gerotranscendence. These are embedded using all-MiniLM-L6-v2 to capture semantic meaning. Pairwise cosine similarity is computed to identify related and strongly aligned topic pairs. Fig. 1Schematic representation of the BERTopic custom preprocessing workflow Schematic representation of the BERTopic custom preprocessing workflow We implemented two versions of the topic modeling pipeline: one using the default configuration of BERTopic and another incorporating customized preprocessing. The default pipeline relies on a unigram-based CountVectorizer and standard stop word removal. In contrast, the customized pipeline improves domain relevance and topic coherence by applying part-of-speech [44, 45], filtering and n-gram tuning. The preprocessing process begins by extracting non-empty paragraphs as individual documents. In the customized pipeline, each document undergoes processing with the spaCy library to retain only specific POS categories such as nouns, verbs, and adjectives which are more semantically relevant for topic modeling. The resulting filtered token set is then used to construct a feature matrix through a TF-IDF vectorizer, configured to include n-grams ranging from unigrams to trigrams (1–3 grams). This allows the model to capture both single words and multi-word expressions enhancing the model’s contextual understanding and overall topic quality. POS filtering helps eliminate irrelevant words, enriching the vocabulary representation and increasing the interpretability of the resulting topics. In contrast, the default BERTopic pipeline uses an internal CountVectorizer without any POS filtering or n-gram customization, relying on unigrams and general-purpose stop word removal. While this approach simplifies the pipeline, it can include irrelevant tokens or miss domain-specific terms, which may lead to broader, less coherent topics. Nonetheless, the default pipeline remains effective for quick prototyping and general-purpose text analysis, where fine-grained control is less crucial. Simultaneously, each document is embedded using the all-MiniLM-L6-v2 sentence transformer model, which converts each paragraph into a dense vector representation. The model calculates semantic similarity between sentences by transforming each one into an embedding and comparing the embeddings for similarity. These embeddings are the core input for UMAP, a dimensionality reduction technique that preserves semantic proximity while reducing the vector dimensions. A min_dist value of 0.1 results in denser clusters, and cosine similarity maintains semantic closeness of documents, ensuring that semantically similar documents are grouped together. HDBSCAN then utilizes the UMAP reduced vectors to cluster semantically similar documents, with each cluster representing a distinct topic. Outliers, documents that do not fit into any cluster, are assigned a topic ID of −1, marking them as noise. HDBSCAN outputs both the topic IDs assigned to each document and the probability of each document’s membership in its assigned topic. These outputs are passed to the BERTopic, which interprets the topics based on the topic-document assignments. By default, BERTopic applies c-TF-IDF to aggregate documents within each topic and extract the most important keywords. Next, KeyBERT generates topic labels by identifying the representative document within each topic, providing a concise label that enhances understanding of the topic’s content. The final step involves summarization. Once topic labels and keywords are generated, documents assigned to each topic are aggregated and passed to a pre-trained BART model. This model produces a brief and coherent summary of each topic, offering a concise overview of the topic and the documents within it. The outcome is a comprehensive set of interpretable outputs, including topics, associated keywords, topic labels, topic assignments, automated summaries, and interactive visualizations. These outputs facilitate an effective analysis of the input documents with minimal manual intervention. To explore conceptual overlaps between topics related to solitude and gerotranscendence, we developed a semantic comparison and visualization pipeline as shown in the Fig. 2. The outputs from the BERTopic modeling analysis of both datasets, provided as structured.txt files, were parsed to extract topic identifiers, labels, keywords, and summaries. These topic labels, keywords, and summaries were then combined into full textual descriptions, which were encoded into high-dimensional semantic embeddings using the all-MiniLM-L6-v2 model from the SentenceTransformers library. Pairwise cosine similarity scores were computed between topics from the two datasets. A similarity threshold of 0.6 was set to identify conceptually related topic pairs, while a higher threshold of 0.75 indicated strong semantic alignment. Using the all-MiniLM-L6-v2 model, unrelated text pairs typically score between 0.15 and 0.35 in cosine similarity, while related pairs fall around 0.50 to 0.70, and near-paraphrases exceed 0.80. In this study, each topic combines its label, keywords, and summary, which increases similarity scores due to both lexical and contextual overlap. A threshold of 0.60 was selected to filter out chance matches while retaining conceptually related topics across corpora, ensuring enough pairs for meaningful relation extraction. Pairs scoring above 0.75 indicate strong alignment across all fields, distinguishing nearly equivalent concepts from merely related ones. This two-level approach supports both broad exploration and precise identification of high confidence matches. Matched pairs were then categorized and stored along with their similarity scores, as well as suggested merged concept labels created by combining the original topic labels. To facilitate the interpretation of these results, an interactive bipartite graph was created using NetworkX and Plotly. Topics related to solitude were positioned on the left and colored blue, while topics related to gerotranscendence were placed on the right in green. Edges between nodes reflected the strength of the semantic relationships, with red indicating strong matches and orange denoting related topics. This visualization approach enabled a systematic and interpretable comparison of thematically similar topics across the two analyses and supported the identification of higher-level conceptual linkages.Fig. 2Topic merging and relation extraction for solitude and gerotranscendence topics Topic merging and relation extraction for solitude and gerotranscendence topics Results From The Handbook of Solitude, a total of 1866 paragraphs were extracted and treated as individual documents. These were analyzed using both a customized BERTopic preprocessing and default BERTopic pipelines. Both the pipelines are customized to flexibly extract any specified number of topics and keywords based on user-defined parameters. This analysis yielded 244 topics, each represented by five keywords, resulting in a combined total of 1220 keywords. Following the removal of duplicate entries and expert review by SMEs to remove irrelevant topics, 46 high-quality topics were retained from the default pipeline and 32 from the customized pipeline, capturing both overlapping and distinct thematic structures. The same dual-pipeline approach was applied to the Gerotranscendence volume, with paragraphs similarly segmented and analyzed. A total of 1249 paragraphs were identified, and this process produced 172 topics, again represented by five keywords each, totaling 860 keywords. After duplicate removal and evaluation from SMEs, 32 high-quality topics were retained from the default pipeline and 33 from the customized pipeline. For the PHASES Ontology, after filtering the duplicates, both the BERTopic pipelines contributed a total of 90 terms, 52 from the solitude corpus and 38 from the gerotranscendence corpus. The list of terms refined through SME review is provided in Appendices B and C. Comparative analysis of default and custom BERTopic pipelines Topic bar charts depicting keyword relevance for solitude Figures 3 and 4 illustrate the topic distributions generated by the default and custom preprocessed BERTopic pipelines applied to the Solitude corpus. Each chart shows the top five keywords for each of the ten dominant topics, with horizontal bars representing each word’s relevance within its topic cluster. Each vertical segment portrays a distinct topic, capturing thematic patterns from the corpora centered on solitude and related constructs. The default pipeline captures a broad range of solitude related themes. For example, topic 0 combines terms like confinement, prisoners, solitary, and isolation, reflecting enforced solitude in carceral or institutional contexts. Topic 1 focuses on creativity, ideas, and efforts, suggesting solitude as a space for cognitive and artistic productivity. Topic 2, with keywords like singlehood, perspective and reasons, highlights individual identity and lifestyle aspects related to solitude. The presence of “paulus1”, “depaulo2” and “williams3” as keywords reflects their frequent citation in the source text, in the sentence not filtered prior to modeling and BERTopic included these author citations as keywords. Topic 3, with keywords like solitude, chapters, paradox, and psychoanalytic, suggests more abstract or interpretive discourses, possibly literary or theoretical. The absence of POS filtering allowed semantically light words, pronouns (e.g., “he”, “me”, “my”, “was”). Some emphasize conceptual or autobiographical reflections, and others center on recovery, affective states, or contain residual academic phrasing.Fig. 3Topic-word score bar chart generated by the default BERTopic pipeline for the solitude corpusFig. 4Topic-word score bar chart generated by the custom-preprocessed BERTopic pipeline for the solitude corpus Topic-word score bar chart generated by the default BERTopic pipeline for the solitude corpus Topic-word score bar chart generated by the custom-preprocessed BERTopic pipeline for the solitude corpus With custom preprocessing, these patterns become more thematically coherent and domain specific. For instance, Topic 0 is expressed with greater lexical precision emphasizing confinement, solitary confinement, and prisoners. Thematic clarity in topic 1 is enhanced through conceptually aligned keywords such as ideas, creative, creativity, efforts, and creative process. Topic 2 captures relational detachment through relationship terms like singlehood, single, single people, mating. Other topics related to psychological exclusion like ostracism, isolation and psychological traits such as shyness also emerge with more distinct separation from the broader thematic space. The appearance of the keywords “hinde,4” and “żurek5” reflects their frequent citation in the source text, as citations were not filtered prior to the custom preprocessing for topic modeling. Overall, the custom preprocessing improves topic granularity, capturing diverse affective and cognitive facets of solitude. While the default pipeline captures meaningful groupings, its output includes more generic or repetitive terms and lacks domain-sensitive precision. Nevertheless, it remains a useful baseline for exploratory thematic analysis in solitude research. The substantial overlap between the default and custom BERTopic outputs occurs because both are applied to the same corpus and the use of the same core modeling process, so distinctive terms such as confinement, prisoners, and solitary remain consistent across configurations. Differences arise from how each pipeline tokenizes, filters, and vectorizes text. The custom pipeline applies POS filtering, domain stopword removal, and n-gram extraction, producing richer multi-word expressions and removing most low-value tokens, which generally improves interpretability. However, it can introduce visible repetition such as “conclusion conclusion”, conclusion, and “conclusion conclusion conclusion” when bigram and trigram generation interact with c-TF-IDF scoring in small clusters. In these cases, limited vocabulary inflates the relative score of repeated n-grams from headings or summaries, causing duplicates to dominate the top keywords. Both pipelines also retain author citations from citation-heavy sections, especially in small topics where frequent terms have greater weight. While the custom pipeline yields more coherent and domain-relevant keywords, targeted lemmatization, n-gram de-duplication, and expanded domain-specific stopword lists would further improve topic interpretability. BART-Generated summaries providing interpretive context for solitude topics Tables 1 and 2 compile the BART-generated summaries for each topic from the default and custom preprocessed pipelines for solitude corpus respectively. These summaries integrate top keywords, topic labels, and contextual interpretations. In the default pipeline model, the summaries contextualize these topics within real-world psychological, institutional, and social frameworks. They highlight solitude’s dual nature, from its harmful expression in solitary confinement to its generative potential for creativity, while also addressing nuanced domains like singlehood, ostracism, and definitional ambiguity. These summaries enrich the interpretability of the keyword-based outputs, although the default pipeline sometimes produces broader, less domain-specific themes. The custom-preprocessed pipeline brings forward more fine-grained and conceptually precise solitude themes. These include intersections with aging, mindfulness, shyness subtypes, hikikomori, behavioral inhibition, and cognitive benefits, alongside familiar domains such as creativity and recovery. Unlike the broader narratives in the default pipeline, these summaries surface nuanced psychological and developmental contexts, revealing solitude’s multifaceted presence across life stages and experiences. In both pipelines, the BART-generated summary for the “Aloneness” topic included a direct mention of The Handbook of Solitude. This likely occurred because the paragraph chosen for summarization repeatedly contained the keyword “solitude” in the introductory section of the Handbook, prompting the model to focus on meta-textual rather than substantive content.Table 1BART-generated topic summaries for the default BERTopic pipeline applied to the solitude corpusTopicTopic LabelKeywordsSummary0Solitaryconfinement, prisoners, solitary, prison, isolationSolitary confinement is a form of solitary confinement in which prisoners are isolated from the rest of the prison population. Solitary confinement is so common in prisons that it has been dubbed a ‘socially accepted’ practice.1Creativitycreative, ideas, creativity, efforts, paulusCreative people tend to have some characteristics in common with “lone geniuses,” such as broad interests, and independence. But the key factors in many cases is simple hard work, persistence, and a resultant high quantity of creative products.2Singlehoodsinglehood, single, perspective, depaulo, reasonsSinglehood is a phenomenon that is determined by multiple factors involving social, cultural, religious, and psychological ones. This chapter will present the current knowledge on singlehood from a psychological perspective.3Alonenesssolitude, chapters, paradox, psychoanalytic, perspectivesThe Handbook of Solitude was the first academic volume to specifically focus on the diverse theoretical and empirical approaches to the psychological study of solitude. In this second edition of the Handbook, we are absolutely thrilled to present a blend of new and updated chapters that approach solitude from a myriad of theoretical perspectives and methodological approaches.4Ostracismostracism, williams, chronic, individuals, ignoredOstracism – being ignored and excluded – is a social phenomenon that occurs in a myriad cultures and contexts among humans and nonhuman social animals. Evolutionary theory offers an explanation for the robustness of ostracism in human society and its occurrence within groups of other social animals.5Solitudepresence, definition, alone, physically, solitudeSolitude can be and sometimes has been operationalized as an absence of social interaction and sometimes as physical aloneness. There is really no agreed-upon physical “distance” from others that is mandated in order for an individual to be considered alone.Table 2BART-generated topic summaries for the custom-preprocessed BERTopic pipeline applied to the solitude corpusTopicTopic LabelKeywordsSummary0Solitaryconfinement, solitary confinement, prisoners, prison, solitarySolitary confinement is a form of solitary confinement in which prisoners are isolated from the rest of the prison population. Solitary confinement is so common in prisons that it has been dubbed a ‘socially accepted’ practice.1Isolationideas, creative, creativity, efforts, creative processCreative people tend to have some characteristics in common with “lone geniuses,” such as broad interests, and independence. But the key factors in many cases is simple hard work, persistence, and a resultant high quantity of creative products.2Singlesinglehood, single, single people, mating, żurekSinglehood is a phenomenon that is determined by multiple factors involving social, cultural, religious, and psychological ones. This chapter will present the current knowledge on singlehood from a psychological perspective.3Alonenesssolitude, chapters, paradox, psychoanalytic, beneficialThe Handbook of Solitude was the first academic volume to specifically focus on the diverse theoretical and empirical approaches to the psychological study of solitude. In this second edition of the Handbook, we are absolutely thrilled to present a blend of new and updated chapters that approach solitude from a myriad of theoretical perspectives and methodological approaches.4Ostracismostracism, chronic ostracism, chronic, effects ostracism, ignoredOstracism – being ignored and excluded – is a social phenomenon that occurs in a myriad cultures and contexts among humans and nonhuman social animals. Evolutionary theory offers an explanation for the robustness of ostracism in human society and its occurrence within groups of other social animals.5Solitudedefinition, presence, definition solitude, physically, caféSolitude can be and sometimes has been operationalized as an absence of social interaction and sometimes as physical aloneness. There is really no agreed-upon physical “distance” from others that is mandated in order for an individual to be considered alone. Table 1BART-generated topic summaries for the default BERTopic pipeline applied to the solitude corpusTopicTopic LabelKeywordsSummary0Solitaryconfinement, prisoners, solitary, prison, isolationSolitary confinement is a form of solitary confinement in which prisoners are isolated from the rest of the prison population. Solitary confinement is so common in prisons that it has been dubbed a ‘socially accepted’ practice.1Creativitycreative, ideas, creativity, efforts, paulusCreative people tend to have some characteristics in common with “lone geniuses,” such as broad interests, and independence. But the key factors in many cases is simple hard work, persistence, and a resultant high quantity of creative products.2Singlehoodsinglehood, single, perspective, depaulo, reasonsSinglehood is a phenomenon that is determined by multiple factors involving social, cultural, religious, and psychological ones. This chapter will present the current knowledge on singlehood from a psychological perspective.3Alonenesssolitude, chapters, paradox, psychoanalytic, perspectivesThe Handbook of Solitude was the first academic volume to specifically focus on the diverse theoretical and empirical approaches to the psychological study of solitude. In this second edition of the Handbook, we are absolutely thrilled to present a blend of new and updated chapters that approach solitude from a myriad of theoretical perspectives and methodological approaches.4Ostracismostracism, williams, chronic, individuals, ignoredOstracism – being ignored and excluded – is a social phenomenon that occurs in a myriad cultures and contexts among humans and nonhuman social animals. Evolutionary theory offers an explanation for the robustness of ostracism in human society and its occurrence within groups of other social animals.5Solitudepresence, definition, alone, physically, solitudeSolitude can be and sometimes has been operationalized as an absence of social interaction and sometimes as physical aloneness. There is really no agreed-upon physical “distance” from others that is mandated in order for an individual to be considered alone. BART-generated topic summaries for the default BERTopic pipeline applied to the solitude corpus Table 2BART-generated topic summaries for the custom-preprocessed BERTopic pipeline applied to the solitude corpusTopicTopic LabelKeywordsSummary0Solitaryconfinement, solitary confinement, prisoners, prison, solitarySolitary confinement is a form of solitary confinement in which prisoners are isolated from the rest of the prison population. Solitary confinement is so common in prisons that it has been dubbed a ‘socially accepted’ practice.1Isolationideas, creative, creativity, efforts, creative processCreative people tend to have some characteristics in common with “lone geniuses,” such as broad interests, and independence. But the key factors in many cases is simple hard work, persistence, and a resultant high quantity of creative products.2Singlesinglehood, single, single people, mating, żurekSinglehood is a phenomenon that is determined by multiple factors involving social, cultural, religious, and psychological ones. This chapter will present the current knowledge on singlehood from a psychological perspective.3Alonenesssolitude, chapters, paradox, psychoanalytic, beneficialThe Handbook of Solitude was the first academic volume to specifically focus on the diverse theoretical and empirical approaches to the psychological study of solitude. In this second edition of the Handbook, we are absolutely thrilled to present a blend of new and updated chapters that approach solitude from a myriad of theoretical perspectives and methodological approaches.4Ostracismostracism, chronic ostracism, chronic, effects ostracism, ignoredOstracism – being ignored and excluded – is a social phenomenon that occurs in a myriad cultures and contexts among humans and nonhuman social animals. Evolutionary theory offers an explanation for the robustness of ostracism in human society and its occurrence within groups of other social animals.5Solitudedefinition, presence, definition solitude, physically, caféSolitude can be and sometimes has been operationalized as an absence of social interaction and sometimes as physical aloneness. There is really no agreed-upon physical “distance” from others that is mandated in order for an individual to be considered alone. BART-generated topic summaries for the custom-preprocessed BERTopic pipeline applied to the solitude corpus Topic bar charts depicting keyword relevance for gerotranscendence Figures 5 and 6 display topic bar charts from the default and custom-preprocessed BERTopic pipelines for gerotranscendence corpus respectively.Fig. 5Topic-word score bar chart generated by the default BERTopic pipeline for the gerotranscendence corpusFig. 6Topic-word score bar chart generated by the custom-preprocessed BERTopic pipeline for the gerotranscendence corpus Topic-word score bar chart generated by the default BERTopic pipeline for the gerotranscendence corpus Topic-word score bar chart generated by the custom-preprocessed BERTopic pipeline for the gerotranscendence corpus The later topics (topics 8-19) highlight key dimensions of gerotranscendence and aging. Topic 8 reflects identity reconstruction and social redefinition through terms like gerotranscendence, wise, and shame. Topic 10 emphasizes existential meaning-making, including reminiscence, unity, and identity. Topic 11 introduces cosmic and transcendent experiences, alongside gendered aspects. Topic 12 captures health-related challenges such as crisis and disease, indicating adaptive processes. Topic 16 strongly represents the cosmic perspective with terms like universe and cosmos. Topic 18 points to narrative and interview-based accounts, reflecting subjective life experiences. Topic 19 appears to represent structural or numerical artifacts (e.g., 196, 187, 155), likely corresponding to document indexing, coding outputs, or frequency markers rather than substantive themes. Overall, these topics collectively illustrate aging as a complex interplay of existential reflection, social engagement, health-related adaptation, and methodological inquiry, reinforcing gerotranscendence as both a lived experience and a conceptual framework. The customized topics further refine key aspects of gerotranscendence and aging. Topic 8 emphasizes social activity and meaningful relationships, linking gerotranscendence with everyday interaction. Topics 9 and 10 highlight reminiscence, unity, and existential reflection, reinforcing inner developmental processes. Topic 11 integrates cosmic transcendence with gendered experiences. Topic 12 reflects stress, crises, and health-related challenges, indicating adaptive coping. Topic 16 captures the cosmic and experiential shift through terms like universe and transcendence. Topic 18 focuses on childhood narratives and interviews, pointing to life-course and qualitative perspectives. The default pipeline produces broader, noisier topics with generic terms and numerical artifacts (e.g., Topic 19), reducing interpretability and thematic precision. In contrast, the customized pipeline removes such noise and yields more coherent, domain-focused topics (e.g., reminiscence, transcendence, life narratives), making the results more meaningful and analytically useful. BART-Generated summaries providing interpretive context for gerotranscendence topics Table 3 and 4Table 3BART-generated topic summaries for the default BERTopic pipeline applied to the gerotranscendence corpusTopicTopic LabelKeywordsSummary0Gerontologytheories, gerontology, theoretical, gerontological, departureTheory of gerotranscendence was developed from unsatisfying mismatch of common theoretical assumptions within social gerontology and some empirical findings. We will argue that the usual theoretical points of departure for gerontological research only represent a narrow corridor in a theoretical field, which is much broader.1Nursingstaff, nurses, percent, areas, impactFamiliarity with the Theory Seemed to Reduce Feelings of Insufficiency at Work. Theory Helped Many Staff Members to See Care Recipients in a New Light and to Understand Them Better.2Dissatisfactiondisengagement, theory, individual, hypothesisNewell, Dean, & McCaffrey (1960) first published their tentative disengagement theory of aging. The theory assumed an intrinsic tendency to disengage and withdraw when growing old.3Gerotranscendencegerotranscendence, practice, towards, acceler, tenthGerotranscendence can be facilitated or obstructed in different ways. The process of gerotrans transcendence can, according to the theory, be accelerated, facilitated or Obstructed.4Depressioninterviewees, sign, isolation, elderly, pleasureA lack of interest in participating in parties can be understood as a symptom of a beginning dementia (the pathology perspective) or as a way to cope with reduced mobility by means of selecting where and how often to go to parties (the SOC perspective) Robinson, 1976; Bradford, 1979; Bernard, 1982. Other indications came from subjective reports of staff members working with old people.5Cosmicwomen, men, cosmic, score, categoryWomen score higher than men on cosmic transcendence, but that this difference decreases with increasing age. The women’s higher levels of cosmic transcendence, especially in the age interval 25–44, might be partly connected with childbirth.Table 4BART-generated topic summaries for the custom-preprocessed BERTopic pipeline applied to the gerotranscendence corpusTopicTopic LabelKeywordsSummary0Definitionsgerontology, theories, points departure, gerontological, departureTheory of gerotranscendence was developed from unsatisfying mismatch of common theoretical assumptions within social gerontology and some empirical findings. We will argue that the usual theoretical points of departure for gerontological research only represent a narrow corridor in a theoretical field, which is much broader.1Nursingstaff, nurses, areas impact, percent, areasFamiliarity with the Theory Seemed to Reduce Feelings of Insufficiency at Work. Theory Helped Many Staff Members to See Care Recipients in a New Light and to Understand Them Better.2Dissatisfactiondisengagement, disengagement theory, theory, disengageNewell, Dean, & McCaffrey (1960) first published their tentative disengagement theory of aging. The theory assumed an intrinsic tendency to disengage and withdraw when growing old.3Gerotranscendencepractice gerotranscendence, gerotranscendence practice gerotranscendence, gerotranscendence practice, gerotranscendence, practiceGerotranscendence can be facilitated or obstructed in different ways. The process of gerotranscendence can, according to the theory, be accelerated, facilitated or Obstructed.4Depressioninterviewees, sign, isolation, parties, elderlyA lack of interest in participating in parties can be understood as a symptom of a beginning dementia (the pathology perspective) or as a way to cope with reduced mobility by means of selecting where and how often to go to parties (the SOC perspective) Robinson, 1976; Bradford, 1979; Bernard, 1982. Other indications came from subjective reports of staff members working with old people.5Transcendencewomen, men, cosmic transcendence, cosmic, women score higherWomen score higher than men on cosmic transcendence, but that this difference decreases with increasing age. The women’s higher levels of cosmic transcendence, especially in the age interval 25–44, might be partly connected with childbirth. display BART summarizations for the default and custom-preprocessed pipelines for the gerotranscendence corpus. These summaries present gerotranscendence as a multidimensional development process involving shifts in meaning, perspective, and selfhood in later life. These summaries distill complex themes into clear insights on how individuals transcend ego boundaries, find peace, and reinterpret aging as they grow older. Perspectives from care staff, gender-related nuances, and empirical links to mental health and purpose further enrich the interpretive landscape. The custom preprocessing pipeline, topics are highly focused and conceptually rich topics, emphasizing psychological, spiritual, and developmental aspects of gerotranscendence. This approach enhances topic coherence and separation, allowing clearer distinctions between theoretical constructs, lived experiences, and caregiving contexts. The alignment between the topics and their BART summaries, reinforcing key insights about meaning making and life integration. Overall, custom preprocessing yields a deeper, more structured understanding of gerotranscendence. Minor lexical artifacts were occasionally observed in BART-generated summaries (e.g., “gerotrans transcendence” instead of “gerotranscendence”), reflecting limitations of automated text generation. These outputs were retained to preserve the original model output and were not considered during ontology evaluation. Table 3BART-generated topic summaries for the default BERTopic pipeline applied to the gerotranscendence corpusTopicTopic LabelKeywordsSummary0Gerontologytheories, gerontology, theoretical, gerontological, departureTheory of gerotranscendence was developed from unsatisfying mismatch of common theoretical assumptions within social gerontology and some empirical findings. We will argue that the usual theoretical points of departure for gerontological research only represent a narrow corridor in a theoretical field, which is much broader.1Nursingstaff, nurses, percent, areas, impactFamiliarity with the Theory Seemed to Reduce Feelings of Insufficiency at Work. Theory Helped Many Staff Members to See Care Recipients in a New Light and to Understand Them Better.2Dissatisfactiondisengagement, theory, individual, hypothesisNewell, Dean, & McCaffrey (1960) first published their tentative disengagement theory of aging. The theory assumed an intrinsic tendency to disengage and withdraw when growing old.3Gerotranscendencegerotranscendence, practice, towards, acceler, tenthGerotranscendence can be facilitated or obstructed in different ways. The process of gerotrans transcendence can, according to the theory, be accelerated, facilitated or Obstructed.4Depressioninterviewees, sign, isolation, elderly, pleasureA lack of interest in participating in parties can be understood as a symptom of a beginning dementia (the pathology perspective) or as a way to cope with reduced mobility by means of selecting where and how often to go to parties (the SOC perspective) Robinson, 1976; Bradford, 1979; Bernard, 1982. Other indications came from subjective reports of staff members working with old people.5Cosmicwomen, men, cosmic, score, categoryWomen score higher than men on cosmic transcendence, but that this difference decreases with increasing age. The women’s higher levels of cosmic transcendence, especially in the age interval 25–44, might be partly connected with childbirth. BART-generated topic summaries for the default BERTopic pipeline applied to the gerotranscendence corpus Table 4BART-generated topic summaries for the custom-preprocessed BERTopic pipeline applied to the gerotranscendence corpusTopicTopic LabelKeywordsSummary0Definitionsgerontology, theories, points departure, gerontological, departureTheory of gerotranscendence was developed from unsatisfying mismatch of common theoretical assumptions within social gerontology and some empirical findings. We will argue that the usual theoretical points of departure for gerontological research only represent a narrow corridor in a theoretical field, which is much broader.1Nursingstaff, nurses, areas impact, percent, areasFamiliarity with the Theory Seemed to Reduce Feelings of Insufficiency at Work. Theory Helped Many Staff Members to See Care Recipients in a New Light and to Understand Them Better.2Dissatisfactiondisengagement, disengagement theory, theory, disengageNewell, Dean, & McCaffrey (1960) first published their tentative disengagement theory of aging. The theory assumed an intrinsic tendency to disengage and withdraw when growing old.3Gerotranscendencepractice gerotranscendence, gerotranscendence practice gerotranscendence, gerotranscendence practice, gerotranscendence, practiceGerotranscendence can be facilitated or obstructed in different ways. The process of gerotranscendence can, according to the theory, be accelerated, facilitated or Obstructed.4Depressioninterviewees, sign, isolation, parties, elderlyA lack of interest in participating in parties can be understood as a symptom of a beginning dementia (the pathology perspective) or as a way to cope with reduced mobility by means of selecting where and how often to go to parties (the SOC perspective) Robinson, 1976; Bradford, 1979; Bernard, 1982. Other indications came from subjective reports of staff members working with old people.5Transcendencewomen, men, cosmic transcendence, cosmic, women score higherWomen score higher than men on cosmic transcendence, but that this difference decreases with increasing age. The women’s higher levels of cosmic transcendence, especially in the age interval 25–44, might be partly connected with childbirth. BART-generated topic summaries for the custom-preprocessed BERTopic pipeline applied to the gerotranscendence corpus Both pipelines perform well but serve different purposes. The custom preprocessing pipeline yields fine-grained, domain-specific insights that are especially valuable for detailed interpretive work, though it may introduce some term redundancy. The default BERTopic pipeline, by contrast, provides a broader thematic overview with cleaner and often more distinct topic labels, making it useful for exploratory analysis. Both approaches rely on c-TF-IDF for generating topic representations, but the custom preprocessing incorporates additional preprocessing steps to improve the quality of the input features before c-TF-IDF is applied, often resulting in more domain-relevant keywords. While the default setup processes raw text with minimal preprocessing, producing broader, more general topic labels, the custom approach enhances specificity and relevance through targeted feature selection. Ultimately, the custom pipeline excels in precision and thematic depth, whereas the default offers breadth, clarity, and reduced redundancy risk. SME validation of the topics Figure 7 shows the flow diagram of the three modeling approaches, SME curated terms, topics extracted from BERTopic default pipeline, topics extracted from custom preprocessing pipeline, highlighting the steps from initial knowledge base identification to final SME validated topic inclusion. The diagram illustrates the input sizes, topic counts before and after filtration, and the number of topics retained following SME validation.Fig. 7Flow diagram showing extraction, filtering, SME validation, and final inclusion of topics from three modeling approaches Flow diagram showing extraction, filtering, SME validation, and final inclusion of topics from three modeling approaches In the initial stage of knowledge base development, once the ontologists define the scope and purpose of the ontology, relevant content was collected from books, scientific articles, clinical trial data, and the practical expertise of subject matter experts. The selected classes were then organized, refined, and thoroughly documented. The ontology was subsequently validated using real-world data and competency questions to ensure its effectiveness. During this process, manual curation across both domains by the SMEs resulted in a finalized list of 37 key terms. For the BERTopic pipelines, the knowledge for topic extraction is sourced from the handbook of solitude and gerotranscendence books. The BERTopic pipelines produced 244 initial topics from the solitude corpus, which were refined through expert review into 46 high-quality topics using the default pipeline and 32 using the custom pipeline. Similarly, for the gerotranscendence corpus, 172 initial topics were generated, later distilled into 32 and 33 high quality topics via the default and custom pipelines, respectively. The SMEs with deep expertise in their respective domains, solitude and gerotranscendence, independently reviewed the topic sets. Each expert assessed the semantic coherence and conceptual relevance of topics, filtering irrelevant and duplicate topics and refining topic labels for clarity and interpretability. For the PHASES Ontology, the BERTopic pipelines contributed a total of 90 terms, 52 from the solitude corpus and 38 from the gerotranscendence corpus. BERTopic modeling, especially when combined with expert review, can effectively support and speed up the early stages of building an ontology by automatically identifying relevant concepts from a large amount of text. While BERTopic successfully captured many broad themes, such as loneliness, shyness, solitude and aloneness that overlap with SME identified terms, the manually curated list included theory informed concepts like social withdrawal, elective solitude, imposed solitude, ambient sociability, social detachment and so on. These theoretical concepts often require deeper contextual interpretation that automated models may fail to uncover. Nonetheless, the BERTopic model outputs aligned with many foundational ideas, demonstrating its value in generating key concepts for expert refinement. The integration of both algorithmic and expert-driven insights highlighted conceptual overlaps between solitude and gerotranscendence, enriching the semantic foundation for the PHASES Ontology. Merging the results for PHASES Ontology development From the topic modeling outputs containing numbered topics with labels, keywords, and summaries representing the salient content within each domain, to explore thematic connections between solitude and gerotranscendence within the context of healthy aging, a structured text processing pipeline was developed. The analysis began with topic modeling outputs generated by BERTopic modeling for solitude and gerotranscendence, two psychological constructs which have been traditionally studied separately but are relevant for understanding variability in healthy aging. A custom parsing function was implemented to extract and organize the topic data from the BERTopic text files. Using regular expressions, the script identified topic headers and captured their components, including the topic identifier, label, keyword list, and summary. This structured representation allowed for subsequent analysis and filtering of topics based on semantic content. To operationalize the theme of healthy aging, a curated set of keywords was defined, including terms such as “healthy aging,” “well-being,” “successful aging,” “quality of life,” “longevity,” and related expressions involving mental or physical health. This list served as a thematic filter to identify topics that potentially addressed or referenced dimensions of healthy aging. Each topic’s keywords and summary were concatenated and scanned for the presence of any of the predefined healthy aging terms. Topics from both solitude and gerotranscendence models that matched the criteria were retained as relevant. These filtered topic sets were then cross-compared, producing all possible pairwise combinations between solitude-related and gerotranscendence-related topics that were associated with healthy aging. For each identified topic pair, the system recorded topic identifiers, labels, keywords, summaries, and annotated them with a common relationship tag indicating that they were “related via healthy aging theme.” The resulting relations were compiled into a structured dataset and exported as a CSV file to facilitate further analysis and interpretation. This methodology enabled systematic identification of conceptually related topics across two thematic domains, grounded in a consistent and interpretable notion of healthy aging. It provided a reproducible framework for linking topic modeling results to broader theoretical constructs. Part of the extracted topic relationships between the solitude and gerotranscendence domains as shown in the Table 5 reveal several high-confidence pairings that converge thematically around the experience and interpretation of solitude, as well as its relationship to constructs such as loneliness, aging, and existential crises. Each relationship was identified based on semantic similarity and shared textual context drawn from topic keywords and summaries, with similarity scores ranging from approximately 0.60 to 0.73. Extracted relations include verb-based relational cues such as “is,” “correlates with,” “shows,” “reduces,” and “redefines,” which serve to clarify the nature of the conceptual links between topics. A notable pattern emerges around Topic 3 (“Aloneness”) from the solitude domain, which connects consistently and with relatively high similarity to multiple gerotranscendence topics also labeled “Solitude” (e.g., Topics 6, 76, 92, 123, 162), with similarity scores reaching as high as 0.727. The extracted relations for these pairings suggest recurring themes such as the correlational significance of solitude in the context of aging and well-being (e.g., “correlated negatively with,” “is evident in,” “reduces”), and its redefinition or reinterpretation in later life (e.g., “redefines importance,” “becomes”). Topic 5 (“Solitude”) also maps onto gerotranscendence topics with the same or similar labels, particularly emphasizing its operationalization and measurable correlations. Phrases like “been operationalized as,” “correlates with,” and “need for” indicate a more theoretical or analytical treatment of solitude, possibly reflecting its evolving conceptualization in developmental aging theory. Several additional patterns like in topic 20 (“Solitude”) displays a broader set of connections not only to solitude topics but also to gerotranscendence topics labeled “Aging” (Topic 138) and “Crises” (Topic 169). These links suggest a multifaceted interpretation of solitude that extends beyond peaceful withdrawal to include associations with psychological adaptation, personal transformation, or challenge. The presence of relations like “becomes,” “typically experiences,” and “seem experienced” indicates dynamic, context-dependent meanings of solitude in late life. Topic 41 (“Loneliness”) links internally with topic 33 (“Loneliness”) in the gerotranscendence domain with a strong similarity (0.682), illustrating the nuanced overlap between solitude and loneliness. The extracted relational verbs (e.g., “lacks,” “feel,” “is with”) highlight affective and existential dimensions of this state, reinforcing the need to differentiate loneliness from solitude in ontological modeling. Finally, the appearance of topics such as “Conclusion” and “Conclusions” (Topics 8 and 15) with a high similarity (0.713) suggests a textual match likely due to structural or editorial features rather than conceptual overlap. These are useful reminders of the importance of filtering meta-content in topic modeling pipelines. Overall, this relationship matrix reinforces the theoretical overlap between solitude and gerotranscendence, particularly through the lens of healthy aging. The recurring and thematically rich links around solitude suggest that it operates as a conceptual bridge between internal psychological states and broader existential or developmental processes in later life. The extracted relations provide a valuable linguistic trace of how these connections are articulated in the source texts, which can be formally encoded into the PHASES Ontology as candidate object properties or axioms.Table 5Part of topic pairings between solitude and gerotranscendencesolitude_idsolitude_ labelgerotranscendence_idgerotranscendence_labelsimilarityrelationshipmerged_ topicextracted_ relations3Aloneness6Solitude0.6496relatedAloneness - Solitude‘are absolutely thrilled in’, ‘present’, ‘is highly evident in’, ‘is evident in’3Aloneness169Crises0.637relatedAloneness - Crises‘are absolutely thrilled in’, ‘present’, ‘seem in’, ‘experienced’5Solitude33Loneliness0.633relatedSolitude - Loneliness‘sometimes been operationalized as’, ‘lacks’, ‘may’, ‘be further’20Solitude138Aging0.607relatedSolitude - Aging‘is in’, ‘typically experiences’45Solitude169Crises0.608relatedSolitude - Crises‘presents’, ‘is in’, seem in’, ‘experienced’ A section of the interactive graph shown in Fig. 8 visually maps the semantic relationships between the solitude and gerotranscendence topics identified through similarity analysis. The complete graph is presented in Fig. 9 in Appendix D. Each topic’s label, keywords, and summary were concatenated into a single text string and encoded using the all-MiniLM-L6-v2 SentenceTransformer model, after which cosine similarity was computed between all solitude–gerotranscendence pairs. In this pipeline, a cosine similarity threshold of 0.60 was chosen to identify topic pairs that are conceptually related but not necessarily near duplicates. With the all-MiniLM-L6-v2 model, unrelated text pairs typically score around 0.15 - 0.35, while thematically related pairs often fall between 0.50 and 0.70. Because each topic representation concatenates the label, keywords, and summary, lexical overlap is naturally higher than in pure sentence-to-sentence comparisons; therefore, setting the lower cutoff at 0.60 filters out most chance overlaps while retaining cross-corpus topical neighbors that share domain vocabulary. In contrast, the higher threshold of 0.75 was used to flag “strong matches,” where the labels, keyword sets, and summaries align closely, indicating a high degree of semantic congruence. Scores above this level in “MiniLM” space generally reflect nearly equivalent concepts, even if phrasing differs, making them suitable for red high confidence edges in the visualization. This two-tier approach preserves recall for exploratory connections while distinguishing the most semantically coherent topic alignments. Table 5Part of topic pairings between solitude and gerotranscendencesolitude_idsolitude_ labelgerotranscendence_idgerotranscendence_labelsimilarityrelationshipmerged_ topicextracted_ relations3Aloneness6Solitude0.6496relatedAloneness - Solitude‘are absolutely thrilled in’, ‘present’, ‘is highly evident in’, ‘is evident in’3Aloneness169Crises0.637relatedAloneness - Crises‘are absolutely thrilled in’, ‘present’, ‘seem in’, ‘experienced’5Solitude33Loneliness0.633relatedSolitude - Loneliness‘sometimes been operationalized as’, ‘lacks’, ‘may’, ‘be further’20Solitude138Aging0.607relatedSolitude - Aging‘is in’, ‘typically experiences’45Solitude169Crises0.608relatedSolitude - Crises‘presents’, ‘is in’, seem in’, ‘experienced’ Part of topic pairings between solitude and gerotranscendence Fig. 8A section of the merged topic label graph with OpenIE relations A section of the merged topic label graph with OpenIE relations In the graph, blue nodes represent solitude topics and green nodes represent gerotranscendence topics, arranged in two vertical columns to make cross-domain links visually clear. Red edges indicate strong matches (similarity > 0.75), while orange edges denote related topics (0.60–0.75). Although some labels appear similar or even identical, the associated summaries reveal that each represents a distinctive perspective, highlighting different nuances, contexts, or implications. Descriptive labels are used to improve clarity by making topic names more specific and distinguishable across all topics. They are especially useful for differentiating topics with similar names by adding keywords, highlighting distinct perspectives within related themes. The layout facilitates immediate recognition of high confidence connections as well as looser thematic links, supporting both targeted analysis and broader exploration. Interactive features allow users to hover over nodes to view full labels and trace their connected edges, enabling direct cross-referencing with the CSV table for keywords, summaries, and extracted Open Information Extraction (OpenIE) relations. Discussion This study applied a comparative topic modeling approach using BERTopic to analyze solitude and gerotranscendence corpora, with the overarching goal of facilitating ontology development for the PHASES framework on healthy aging. Both the default BERTopic pipeline and a customized preprocessing pipeline were applied to extract, compare, and interpret meaningful topic structures across corpora. The integration of BART summarization, KeyBERT labeling, and visual analytics (e.g., bar charts, dendrograms) supported both analytical depth and interpretive clarity. Although the visualizations employed in this study are those provided natively by the BERTopic framework, several custom extensions were developed to enhance the semantic interpretability and domain applicability of the topic modeling process. In the BERTopic custom preprocessing pipeline, a POS based tokenizer was implemented using spaCy, selectively retaining nouns, verbs, adjectives, and adverbs to prioritize linguistically and semantically informative words. Second, automatic topic labeling was incorporated using KeyBERT, which extracts the most salient keyword from representative documents identified by BERTopic, thereby generating concise and interpretable topic labels in place of default numeric identifiers. Third, topic-level summarization was conducted using the BART transformer model to produce coherent natural language summaries of the aggregated content within each topic cluster, facilitating a higher-level semantic understanding of the text. Finally, a custom export module was implemented to output all results in multiple structured formats (TXT, CSV, and Excel), including topic labels, representative keywords, summaries, and document-level topic probabilities. Collectively, these methodological extensions improve the transparency, interpretability, and practical usability of topic modeling outputs. Comparative performance of default and custom BERTopic pipelines In both domains, the default BERTopic pipeline yielded broader, more generalizable topic clusters, making it particularly effective for exploratory or high-level thematic mapping. Its outputs displayed clear and interpretable clusters, though at times they suffered from superficial or repetitive keywords, and limited domain sensitivity. Nonetheless, the default approach captured key constructs such as solitude in creative spaces, social exclusion, and disengagement theory providing a useful foundation for downstream analyses. Conversely, the custom preprocessing pipeline significantly enhanced topic granularity and domain relevance. In the solitude corpus, this approach yielded topics such as positive shyness, institutional confinement, and reflective mentorship narratives, concepts that were not as distinct in the default output. For gerotranscendence, custom preprocessing uncovered intricate subthemes such as cosmic gendered experiences, behavioral manifestations of transcendence, and solitary meaning-making in later life. Despite occasional keyword redundancy, the custom method offered superior resolution for identifying and interpreting nuanced psychological and developmental constructs. Both pipelines leveraged c-TF-IDF for topic representation, but custom preprocessing enhanced this step through domain-sensitive feature selection, including POS tagging and phrase filtering. While the default pipeline benefited from minimal preprocessing that preserved general word frequency patterns, the custom pipeline foregrounded psychological, existential, and narrative terms that are central to solitude and gerotranscendence discourses. Across both corpora, solitude emerged not merely as physical aloneness, but as a multifaceted psychological and developmental phenomenon. Themes ranged from enforced isolation (e.g., solitary confinement) and social ostracism to creative solitude, philosophical inquiry, and personality traits like shyness. The custom pipeline especially surfaced solitude’s conceptual evolution, from pathological states to deliberate, enriching experiences tied to reflection, growth, and resilience. In the gerotranscendence corpus, topic modeling revealed a similarly layered structure. Topics included theoretical foundations (e.g., disengagement theory), gendered or spiritual transitions, caregiving dynamics, and the cognitive-emotional reorientation of aging individuals. The intersection of solitude and gerotranscendence became particularly salient through topics that addressed positive solitude, social withdrawal, identity shifts, and existential reinterpretations of aging. The BART-generated summaries were instrumental in deepening interpretation. They bridged the gap between keyword-based topics and human-understandable narratives, allowing for richer insights into how solitude and aging are framed across empirical, theoretical, and lived experience contexts. Ontology development and thematic integration The integration of topic modeling outputs into the PHASES Ontology framework represents a key methodological contribution. By filtering topics based on relevance to “healthy aging” operationalized via curated terms such as well-being, successful aging, and quality of life, the study identified conceptually rich topic pairs spanning both domains. These were semantically matched and annotated using relational cues (e.g., correlates with, reduces, redefines), enabling the construction of a structured, interpretable map of cross-domain relationships. High-confidence pairings, such as the repeated connections between “Aloneness” in solitude and “Solitude” in gerotranscendence, reinforced solitude’s centrality as a psychological and existential construct in later life. Topic links involving “Loneliness” and “Crises” further illustrated how affective and transitional experiences contribute to aging narratives. Additionally, the detection of meta-topics such as “Conclusion” emphasized the need for filtering structural artifacts during topic interpretation. This structured approach to relationship extraction laid the groundwork for encoding meaningful semantic relations into the PHASES Ontology. These include candidate object properties (e.g., redefines, correlates with) and potential class axioms, providing a foundation for automated reasoning and conceptual modeling in future knowledge engineering efforts. Conclusion This study employed a dual-pipeline BERTopic strategy to map overlapping and domain-specific themes in solitude and gerotranscendence literature, highlighting their conceptual interconnectedness within the broader discourse on healthy aging. The approach addressed a key limitation of traditional ontology development, its heavy reliance on SMEs, high time demands, and vulnerability to cognitive bias by introducing a semi-automated, workflow for early-stage concept extraction. BERTopic derived topics reveal latent concepts and enable hierarchical semantic structure construction connecting raw text data and ontology structured classes. Applied to The Handbook of Solitude and Gerotranscendence, the pipelines generated 417 initial topics, refined through SME review to 32 (custom) and 46 (default) high quality solitude topics, and 33 (custom) and 32 (default) high-quality gerotranscendence topics, yielding 90 ontology terms in total (52 from solitude, 38 from gerotranscendence). While both BERTopic pipelines have their strengths, the custom preprocessing version offers superior specificity and interpretive granularity, whereas the default excels in abstraction and structural coherence, beneficial for building high-level thematic maps. The custom pipeline’s richer multi-word expressions and removal of low-value tokens produce more coherent topics, enhancing interpretability for domain-specific research. The default pipeline, in contrast, provides broader exploratory coverage, though it risks more noise. Visual evaluations confirmed these differences, with the custom pipeline excelling in interpretive granularity. Future improvements to the custom approach could include post-vectorization, de-duplication, filtering of citation patterns, and merging of semantically identical n-grams to ensure each topic’s top terms are both unique and highly informative. This also helps remove unrelated content, make different forms of the same word consistent, and prepare the text for topic modeling. Certain theory-heavy concepts still required SME interpretation, underscoring the complementary role of automated methods and expert review. By systematically mapping the interrelations between solitude and gerotranscendence, this work not only contributes to empirical understanding in psychology and gerontology but also offers a reproducible framework for integrating unsupervised NLP outputs into structured ontologies. These findings may inform future interventions, policy discussions, and theoretical explorations around aging, well-being, and the transformative role of solitude across the human lifespan. To integrate outputs across the solitude and gerotranscendence domains, semantic embeddings of topic labels, keywords, and summaries from the custom pipeline were compared using cosine similarity, identifying semantically matched topic pairs above a set threshold. This enabled the construction of a merged conceptual framework that bridges both domains and strengthens the semantic foundation of the PHASES Ontology. From the findings of our study, several areas for future research can be explored. Some of them are:Automatic generation of descriptions, definitions, and relationships between the keywords for ontology development using LLMs.CQs guide what an ontology should include and give a better idea of what an ontological domain would be. Generating competency questions from information sources can supplement the competency questions generated by experts.Further incorporation of more powerful LLMs (e.g., ChatGPT). Automatic generation of descriptions, definitions, and relationships between the keywords for ontology development using LLMs. Automatic generation of descriptions, definitions, and relationships between the keywords for ontology development using LLMs. CQs guide what an ontology should include and give a better idea of what an ontological domain would be. Generating competency questions from information sources can supplement the competency questions generated by experts. CQs guide what an ontology should include and give a better idea of what an ontological domain would be. Generating competency questions from information sources can supplement the competency questions generated by experts. Further incorporation of more powerful LLMs (e.g., ChatGPT). Further incorporation of more powerful LLMs (e.g., ChatGPT). BERTopic Modeling Implementation, along with supporting documentation, can be found here: https://github.com/Buffalo-Ontology-Group/phases-nlp Appendices Appendix A – Parameter Settings for BERTopic ImplementationspaCy modelen_core_web_sm: This is a lightweight model, so POS tagging and tokenization are fast and require low memory. It’s sufficient for stopword removal, lemmatization, and part-of-speech filtering in most topic modeling tasksVectorizer parameters (vectorizer_params)scheme: tfidf: TF-IDF helps prioritize distinctive words over frequently occurring ones, improving topic separation compared to raw counts.stop_words: english: Removes common, non-informative words that otherwise dominate topics such as, ‘the’, ‘and’, ‘is’.ngram_range: [1, 3]: Includes unigrams, bigrams, and trigrams so the model can capture single words (solitude), short phrases (social isolation), and longer key expressions (quality of life in aging), which are often semantically richer.SentenceTransformer modelsall-MiniLM-L6-v2 (for embeddings and KeyBERT): This model is small and fast but still produces high-quality sentence embeddings.UMAP parameters (umap_params)n_neighbors: 5: Small neighborhood size ensures local structures are preserved, helping separate fine-grained topics.n_components: 5: Reduces high-dimensional embeddings to 5 dimensions for clustering enough to retain semantic relationships while keeping clustering efficient.min_dist: 0.1: Allows points in the same cluster to be packed fairly close while still preserving separation between clusters.metric: cosine: Best suited for embeddings, as it measures semantic similarity rather than raw Euclidean distance.random_state: 42: Ensures reproducibility across runs.HDBSCAN parameters (hdbscan_params)min_cluster_size: 2: Allows even very small but meaningful topics to be discovered.metric: euclidean: UMAP outputs Euclidean-friendly coordinates, making Euclidean distance effective for HDBSCAN clustering after dimensionality reduction.cluster_selection_method: eom: Produces cleaner, more distinct clusters.prediction_data: true: Enables soft clustering, detects outliers and assigns probabilities to clusters, which is valuable for topic quality inspection. spaCy modelen_core_web_sm: This is a lightweight model, so POS tagging and tokenization are fast and require low memory. It’s sufficient for stopword removal, lemmatization, and part-of-speech filtering in most topic modeling tasks spaCy model en_core_web_sm: This is a lightweight model, so POS tagging and tokenization are fast and require low memory. It’s sufficient for stopword removal, lemmatization, and part-of-speech filtering in most topic modeling tasks en_core_web_sm: This is a lightweight model, so POS tagging and tokenization are fast and require low memory. It’s sufficient for stopword removal, lemmatization, and part-of-speech filtering in most topic modeling tasks en_core_web_sm: This is a lightweight model, so POS tagging and tokenization are fast and require low memory. It’s sufficient for stopword removal, lemmatization, and part-of-speech filtering in most topic modeling tasks Vectorizer parameters (vectorizer_params)scheme: tfidf: TF-IDF helps prioritize distinctive words over frequently occurring ones, improving topic separation compared to raw counts.stop_words: english: Removes common, non-informative words that otherwise dominate topics such as, ‘the’, ‘and’, ‘is’.ngram_range: [1, 3]: Includes unigrams, bigrams, and trigrams so the model can capture single words (solitude), short phrases (social isolation), and longer key expressions (quality of life in aging), which are often semantically richer. Vectorizer parameters (vectorizer_params) scheme: tfidf: TF-IDF helps prioritize distinctive words over frequently occurring ones, improving topic separation compared to raw counts.stop_words: english: Removes common, non-informative words that otherwise dominate topics such as, ‘the’, ‘and’, ‘is’.ngram_range: [1, 3]: Includes unigrams, bigrams, and trigrams so the model can capture single words (solitude), short phrases (social isolation), and longer key expressions (quality of life in aging), which are often semantically richer. scheme: tfidf: TF-IDF helps prioritize distinctive words over frequently occurring ones, improving topic separation compared to raw counts. scheme: tfidf: TF-IDF helps prioritize distinctive words over frequently occurring ones, improving topic separation compared to raw counts. stop_words: english: Removes common, non-informative words that otherwise dominate topics such as, ‘the’, ‘and’, ‘is’. stop_words: english: Removes common, non-informative words that otherwise dominate topics such as, ‘the’, ‘and’, ‘is’. ngram_range: [1, 3]: Includes unigrams, bigrams, and trigrams so the model can capture single words (solitude), short phrases (social isolation), and longer key expressions (quality of life in aging), which are often semantically richer. ngram_range: [1, 3]: Includes unigrams, bigrams, and trigrams so the model can capture single words (solitude), short phrases (social isolation), and longer key expressions (quality of life in aging), which are often semantically richer. SentenceTransformer modelsall-MiniLM-L6-v2 (for embeddings and KeyBERT): This model is small and fast but still produces high-quality sentence embeddings. SentenceTransformer models all-MiniLM-L6-v2 (for embeddings and KeyBERT): This model is small and fast but still produces high-quality sentence embeddings. all-MiniLM-L6-v2 (for embeddings and KeyBERT): This model is small and fast but still produces high-quality sentence embeddings. all-MiniLM-L6-v2 (for embeddings and KeyBERT): This model is small and fast but still produces high-quality sentence embeddings. UMAP parameters (umap_params)n_neighbors: 5: Small neighborhood size ensures local structures are preserved, helping separate fine-grained topics.n_components: 5: Reduces high-dimensional embeddings to 5 dimensions for clustering enough to retain semantic relationships while keeping clustering efficient.min_dist: 0.1: Allows points in the same cluster to be packed fairly close while still preserving separation between clusters.metric: cosine: Best suited for embeddings, as it measures semantic similarity rather than raw Euclidean distance.random_state: 42: Ensures reproducibility across runs. UMAP parameters (umap_params) n_neighbors: 5: Small neighborhood size ensures local structures are preserved, helping separate fine-grained topics.n_components: 5: Reduces high-dimensional embeddings to 5 dimensions for clustering enough to retain semantic relationships while keeping clustering efficient.min_dist: 0.1: Allows points in the same cluster to be packed fairly close while still preserving separation between clusters.metric: cosine: Best suited for embeddings, as it measures semantic similarity rather than raw Euclidean distance.random_state: 42: Ensures reproducibility across runs. n_neighbors: 5: Small neighborhood size ensures local structures are preserved, helping separate fine-grained topics. n_neighbors: 5: Small neighborhood size ensures local structures are preserved, helping separate fine-grained topics. n_components: 5: Reduces high-dimensional embeddings to 5 dimensions for clustering enough to retain semantic relationships while keeping clustering efficient. n_components: 5: Reduces high-dimensional embeddings to 5 dimensions for clustering enough to retain semantic relationships while keeping clustering efficient. min_dist: 0.1: Allows points in the same cluster to be packed fairly close while still preserving separation between clusters. min_dist: 0.1: Allows points in the same cluster to be packed fairly close while still preserving separation between clusters. metric: cosine: Best suited for embeddings, as it measures semantic similarity rather than raw Euclidean distance. metric: cosine: Best suited for embeddings, as it measures semantic similarity rather than raw Euclidean distance. random_state: 42: Ensures reproducibility across runs. random_state: 42: Ensures reproducibility across runs. HDBSCAN parameters (hdbscan_params)min_cluster_size: 2: Allows even very small but meaningful topics to be discovered.metric: euclidean: UMAP outputs Euclidean-friendly coordinates, making Euclidean distance effective for HDBSCAN clustering after dimensionality reduction.cluster_selection_method: eom: Produces cleaner, more distinct clusters.prediction_data: true: Enables soft clustering, detects outliers and assigns probabilities to clusters, which is valuable for topic quality inspection. HDBSCAN parameters (hdbscan_params) min_cluster_size: 2: Allows even very small but meaningful topics to be discovered.metric: euclidean: UMAP outputs Euclidean-friendly coordinates, making Euclidean distance effective for HDBSCAN clustering after dimensionality reduction.cluster_selection_method: eom: Produces cleaner, more distinct clusters.prediction_data: true: Enables soft clustering, detects outliers and assigns probabilities to clusters, which is valuable for topic quality inspection. min_cluster_size: 2: Allows even very small but meaningful topics to be discovered. min_cluster_size: 2: Allows even very small but meaningful topics to be discovered. metric: euclidean: UMAP outputs Euclidean-friendly coordinates, making Euclidean distance effective for HDBSCAN clustering after dimensionality reduction. metric: euclidean: UMAP outputs Euclidean-friendly coordinates, making Euclidean distance effective for HDBSCAN clustering after dimensionality reduction. cluster_selection_method: eom: Produces cleaner, more distinct clusters. cluster_selection_method: eom: Produces cleaner, more distinct clusters. prediction_data: true: Enables soft clustering, detects outliers and assigns probabilities to clusters, which is valuable for topic quality inspection. prediction_data: true: Enables soft clustering, detects outliers and assigns probabilities to clusters, which is valuable for topic quality inspection. Appendix BTopic overlap between manual curation, BERTopic default and Custom preprocessing pipelines for solitude corpus Appendix BTopic overlap between manual curation, BERTopic default and Custom preprocessing pipelines for solitude corpus Topic overlap between manual curation, BERTopic default and Custom preprocessing pipelines for solitude corpus Appendix CTopic overlap between manual curation, BERTopic default and Custom preprocessing pipelines for gerotranscendence corpus Appendix CTopic overlap between manual curation, BERTopic default and Custom preprocessing pipelines for gerotranscendence corpus Topic overlap between manual curation, BERTopic default and Custom preprocessing pipelines for gerotranscendence corpus Appendix D – Complete graph of the merged topic label graph with OpenIE relationsFig. 9Complete graph of the merged topic label graph with OpenIE relations Complete graph of the merged topic label graph with OpenIE relations