Investigating patterns of change, stability, and interaction among scientific disciplines using embeddings

The paper addresses how to describe and quantify relationships among scientific disciplines and how these relationships evolve over time, given that disciplines are fluid, context-dependent, and often interact in multi- or interdisciplinary ways. The authors propose treating disciplines as empirical phenomena emerging from article-level behaviors and co-classifications, drawing on distributional semantics from NLP. They set three research questions: (1) how to generate accurate computational representations of disciplines; (2) how similarity between disciplines changes over time overall and within categories; and (3) how individual disciplines’ profiles change over time. The study frames multiple disciplinarity as a gradient, system-level phenomenon and aims to provide scalable, ontology-informed yet data-driven measures that align micro-level article co-occurrences with macro-level disciplinary structures.

The paper reviews approaches to defining and measuring disciplines, including co-authorship, co-citation, and textual similarity methods, often relying on ontologies such as Web of Science/JCR categories, noting coverage limitations. Microsoft Academic Graph is cited as a broad ontology that nonetheless imposes categorical constraints. Prior work on interdisciplinarity employs reference integration/diffusion metrics (e.g., Solomon et al., 2016) and co-authorship networks (Xie et al., 2018). NLP-based approaches have modeled similarity among topics and papers using embeddings and network methods (e.g., Word2Vec on titles for physics; topic embeddings tracking changes; paper embeddings such as P2V), but typically either domain-specific or not aligned with broad ontologies. Research on modeling disciplinary interaction has used top-down ontology or bottom-up article/journal modeling; qualitative studies are rich but do not scale; network models can be challenging to scale. The authors draw on NLP literature on diachronic semantic change (e.g., alignment of temporal embeddings, changepoint detection) to inspire tracking profile changes of disciplines. They argue type embeddings (e.g., Word2Vec) are preferable to contextual embeddings (e.g., BERT) for their task due to lack of token order, subword issues, and empirical performance in semantic change detection.

Data: Over 21,201,258 English-language research articles (document type: article) from the Dimensions database (1990–2019), grouped into ten 3-year intervals (1990–1992 through 2017–2019). Each article is assigned up to four ANZSRC Fields of Research (FoR) codes at level 2 (154 disciplines mapped to 22 divisions and 8 clusters). Co-occurrence is defined when multiple FoR codes are assigned to the same article. Embedding model: Word2Vec Skip-gram implemented in Gensim. Inputs are per-article lists of up to four level-2 FoR codes treated as co-occurring items. Hyperparameters: window size 2, min count 1, embedding dimension 12, negative sampling size selected via intrinsic evaluation (best: 1). Training is performed separately for each 3-year interval, producing temporal embedding spaces. Embedding evaluation: (1) Similarity: For 2017–2019 embeddings, nearest neighbors were retrieved with Annoy; top-3 neighbors showed mean cosine similarity 0.68 (SD 0.10) versus random pairs mean 0.07 (SD 0.31); difference highly significant (t=39.01, p<0.0001). (2) Analogy: Exploratory vector arithmetic produced meaningful analogies (e.g., Cognitive Sciences−Psychology≈Language Studies/Data Format; Communication and Media Studies−Performing Arts & Creative Writing + Visual Arts and Crafts≈Sociology/Journalism and Professional Writing). (3) Higher-level structure: PCA of 2017–2019 embeddings yielded a horseshoe-shaped continuum aligning with STEM↔HSS, with PC1 41% and PC2 27% variance explained. FoR code validity check: For 12 linguistics-related journals (2019 articles), the mean overlap of FoR codes between articles and their references correlated with journal scope: multidisciplinary (33%), general linguistics (47.7%), sub-field journals (60.7%), supporting that FoR co-occurrence reflects disciplinary breadth. Similarity analysis: For each interval, compute pairwise cosine similarity among all discipline embeddings; analyze the average over pairs over time with Kendall rank correlation. Neighborhood(e,d) defined as all embeddings with cosine similarity ≥ 1−d; neighborhood size computed per discipline with threshold d=0.2 (cosine ≥0.8), averaged over disciplines and tracked over time. Temporal alignment and profile change: Build embeddings per interval; align spaces for comparison using Orthogonal Procrustes (for first vs last) and Generalized Procrustes Alignment (for full time series). Define self-similarity as cosine between the same discipline across intervals relative to a reference space (last interval). Use PELT changepoint detection (penalty 0.5, jump 1) on self-similarity time series to identify significant shifts. Additionally, define a discipline’s profile as its neighborhood set; measure changes by tracking neighbor identity changes between first and last intervals. Baseline: Mann–Kendall test on mean number of FoR codes per document over time to assess simple multi-label trend (r=0.64, p<0.001).

• Representation quality: Discipline embeddings capture known relationships and support analogy operations, enabling compositional analysis of multi-disciplinary fields. PCA reveals a STEM–HSS continuum (PC1 41%, PC2 27%). Nearest-neighbor similarity far exceeds random (mean 0.68 vs 0.07; t=39.01, p<0.0001).
• Similarity over time: The average cosine similarity between all discipline pairs increases significantly over time (Kendall’s r=0.56, p=0.044), indicating global convergence. STEM disciplines show a clearer upward trend than HSS; at the cluster/division level, many groups display flat or decreasing within-group similarity, implying increases are largely across-group.
• Neighborhood size: Average neighborhood size (cosine ≥0.8) decreases significantly over time (Kendall’s r=−0.94, p<0.05), consistent in both HSS and STEM and across clusters; exceptions at division level include Technology, Physical Sciences, and Chemical Sciences, which show increasing neighborhood sizes. This pattern suggests increasing specialization (smaller close-neighbor sets) alongside global convergence.
• Profile change: Disciplines with the greatest profile shifts include Communication and Media Studies, Computer Hardware, Building, and Food Sciences. Example: Computer Hardware neighbors shifted from biology-related fields (e.g., Plant Biology, Physiology, Genetics) in 1990–1992 to computational fields (e.g., Distributed Computing, Computer Software, Information Systems) in 2017–2019, indicating a move from application-focused to more inward computational proximity. Communication and Media Studies shifted from early neighbors linked to health/risk and applied areas to later neighbors in journalism, religion, anthropology, film/visual arts, reflecting maturation and identity formation; analogies reveal facets aligning with sociology, journalism, law, and marketing.
• Baseline trend: Mean number of FoR codes per article increased over time (Mann–Kendall r=0.64, p<0.001), indicating rising multidisciplinarity in labeling, though raw counts alone do not capture relational structure.
• Hyperparameters: Intrinsic evaluation using level-1/level-2 mappings favored window=2 and negative sampling=1, with an average of 19/20 level-1 categories showing higher within-category similarity than overall baselines.

The findings address the research questions by demonstrating that distributional, embedding-based representations can model disciplines in a way that aligns with established taxonomies while revealing nuanced, compositional relationships not captured by static ontologies. The observed increase in average pairwise similarity indicates a global convergence across disciplines, more pronounced in STEM than HSS. Concurrently, the significant decline in neighborhood size implies local specialization: disciplines become more similar overall while maintaining or deepening distinct close-knit neighborhoods. This reconciles debates on whether research is becoming more or less specialized by decomposing multiple disciplinarity into similarity/analogy structure, neighborhood size (breadth/depth of close ties), and neighborhood identity (which fields interact). Temporal profile analyses reveal substantive shifts in certain disciplines’ positions and neighbors, consistent with historical accounts (e.g., communication studies’ maturation; computer hardware’s evolving proximity). These results are relevant for researchers, institutions, and policymakers to understand evolving disciplinary landscapes, inform collaboration strategies, curriculum design, and hiring, and potentially guide funding and journal scope decisions.

The study contributes a scalable, data-driven framework for representing disciplines via embeddings computed from FoR co-occurrences, enabling analysis of similarity, specialization, and temporal profile changes across more than 21 million articles (1990–2019). Key contributions include: (1) discipline embeddings that capture known relations and support compositional analogies; (2) evidence for global convergence (r=0.56, p=0.044) alongside local specialization (neighborhood size decline r=−0.94, p<0.05), with STEM showing stronger convergence; and (3) identification and interpretation of major profile shifts in specific disciplines. Future research directions include: expanding evaluation and analogy benchmarks for discipline embeddings; deeper case studies of additional disciplines; applying embeddings to predictive tasks (e.g., journal scope overlap, co-authorship); exploring compositionality and its relation to interdisciplinarity metrics; translating between ontologies (e.g., FoR to MAG) via embedding alignment; and integrating metadata embeddings with text-based embeddings for richer scientometric analyses.

The approach depends on the quality and consistency of FoR code assignment in Dimensions; the underlying ML assignment process and its temporal stability are not fully transparent, and accuracy may vary across fields (especially under-represented ones). The analysis excludes pre-1990 publications, potentially missing longer-term trends and introducing period-specific biases. While embeddings offer powerful distributional insights, they are black-box models sensitive to hyperparameters and data distributions. Co-occurrence-based proximity may conflate methodological similarity with topical co-labeling, and neighborhood thresholds (e.g., cosine ≥0.8) are heuristic choices.