The Role of Data Papers in Scientific Data Dissemination and Reuse

Abstract

Data papers have emerged as a distinctive publication format in the open science era, yet their actual role in fostering scientific data reuse remains uncertain. Drawing on extensive citation-context datasets and large language models, this study investigates the specific contributions and citation purposes of data papers, complemented by a comparative analysis across diverse disciplinary domains. The findings show that while data papers significantly enhance dataset visibility and scholarly recognition, most citations emphasize descriptive or contextual information rather than facilitating direct computational, comparative, or synthetic reuse. Moreover, despite persistent advocacy for a “data-driven” research paradigm, the contribution of data papers to catalyzing novel scientific insights appears to be modest and often indirect. Notable disciplinary differences reveal that data-intensive fields, such as Earth and Life Sciences, embed datasets directly into research workflows, whereas more conceptual domains primarily treat data papers as methodological guides or contextual references. Additionally, publishing attributes—spanning journal practices to national data infrastructures—emerge as critical determinants of reuse patterns, with dissemination effectiveness shaped more by publication management and the thematic priorities of editors than by output volume alone.

Introduction

The rapid transformation of scientific paradigms underscores the significance of scientific data. Promoting the open sharing of scientific data can foster knowledge innovation and enhance research efficiency. Traditional approaches such as the development of data repositories and the publication of data statements and attachments have, to some extent, facilitated data accumulation and reuse. However, the ongoing dissemination of scientific data has increasingly revealed a number of pressing issues¹: researchers’ reluctance to share data^2,3, unclear data ownership⁴, the fragmentation of data resources⁵, and challenges in controlling data quality⁶. To address these concerns and promote the reproducibility of research findings, the data paper has emerged as an important academic publication format.

As a response to these challenges, data papers have gained prominence as a scholarly publication format dedicated to documenting datasets in a structured and standardized manner^7,8. Typically, they include detailed descriptions of data collection and processing methods, data structure and storage format, usage instructions, and access pathways⁹, effectively serving as a comprehensive manual for the underlying data. By integrating both narrative and structured information, data papers significantly improve the interpretability, discoverability, and reusability of research data. They also provide transparent accounts of dataset provenance, the experimental procedures used in their generation, and direct links to the resulting data files. It is also important to note that the adoption and prominence of data papers have varied historically across disciplines. As Candela et al.¹⁰ pointed out, although the term data paper has only recently become standardized, similar publication formats—such as data descriptors, data notes, or database articles—have long been used across disciplines, reflecting an established practice of dataset description. Moreover, the growing number of data journals suggests that the publication of data papers is particularly emphasized in fields such as medicine, biochemistry, and agricultural sciences.

From a theoretical standpoint grounded in research behavior motivation, data papers are considered effective tools for promoting researchers’ active engagement in data sharing. Reluctance to share and reuse scientific data generally stems from two key factors. First, data producers often lack incentives, as they are uncertain whether data sharing will lead to appropriate recognition or rewards^2,11,12. Second, inadequate documentation and insufficient context often hinder effective data reuse^13,14. Compared to other data publication formats, data papers appear to address these challenges at their root by enhancing researchers’ intrinsic and extrinsic motivations. By enabling formal citation and securing the producers’ rights to discovery priority and academic recognition, data papers provide a tangible incentive for sharing. Moreover, their peer-review process helps ensure data quality, thereby facilitating more reliable reuse^7,15.

Some studies have explored the actual dissemination impact of data papers on scientific data reuse, particularly through analyses of citation practices. For example, Jiao and Darch¹⁶ manually examined the citation contexts of 103 data papers in earth sciences and physics, and found that data papers have not yet fully realized their potential in promoting data reuse. Similarly, Li et al.¹⁷ observed that most citations to data papers were unrelated to the datasets described in the papers, instead focusing on the research findings or methodologies. Nevertheless, other studies have shown that data papers still enhance the academic visibility of scientific datasets by providing access points to associated data and offering detailed metadata descriptions¹⁸. Research on citation behaviors in the biomedical field indicates a steady increase in the formal citation of data papers specifically for data usage purposes¹⁹, while in the humanities and social sciences, data papers have been found to positively influence the scholarly impact of datasets referenced in research articles²⁰. Overall, existing literature presents several limitations: there is a lack of in-depth investigation into the specific roles and characteristics of data papers, and many studies are confined to single disciplinary contexts and small sample sizes, limiting the generalizability of their findings.

A more direct and systematic approach to assessing whether data papers substantively promote the dissemination and reuse of scientific data is to trace and measure the extent to which scholarly communities engage with the underlying datasets through data paper citations. Citation behavior serves as a key indicator of such engagement, especially when analyzed through citation contexts—the textual environments in which citations occur—which can reveal the contribution and citation purpose of the cited data paper^21,22. This perspective enables a deeper understanding of the scientific functions of data papers, viewed through the lenses of academic rhetoric and semantic cognition. With recent advances in NLP and large language models, large-scale citation context analysis has become increasingly feasible²², offering a robust methodological foundation for investigating how data papers are cited and what roles they fulfill within scholarly narratives.

Building on the above, the purpose of this study is to investigate the role of data papers in facilitating the open sharing and informed reuse of scientific data within scholarly communication. Specifically, the study addresses the following three research questions: (1) Do data papers contribute data-related value to the studies that cite them, that is, data-oriented contributions such as providing data context, facilitating data processing or integration, and assisting in data validation? (2) If so, for what specific purposes do the citing works utilize the associated scientific data? (3) Does the functional role of data papers vary across different disciplinary domains? To answer these questions, this research conducts a large-scale citation context analysis of data papers across multiple disciplines. The analysis focuses on identifying the actual contributions made by data papers and examining the purposes behind their citation, with the aim of gaining a deeper understanding of how data papers support the dissemination and reuse of scientific data in diverse academic contexts.

Data and Methods

Data

In our study, we restricted the document type to “data papers” in the Web of Science Core Collection, covering the period from 1980 to 2024. In February 2025, we retrieved a total of 17,318 data papers, of which 82.24% (14,243) were cited, with an average of 15.99 citations per paper. To categorize these data papers by discipline, we employed the InCites citation topics (macro) schema, thereby mapping them to 10 disciplines. After importing the data papers into InCites, 16,898 papers were found to possess the corresponding field labels (with a loss of 420 papers, accounting for approximately 2.4% of the total). We use citation context analysis because it offers richer explanatory power for understanding data papers compared with traditional citation-frequency indicators and altmetrics. Citation contexts reveal not only quantitative patterns but also how and why data papers are cited, thereby providing insight into the underlying research processes. In addition, this approach captures multiple dimensions of the use and value of data papers, leading to a more comprehensive understanding of their roles in scholarly communication. We obtained citation context data from their citing documents through the Scite platform (https://scite.ai). Scite is an intelligent citation database that uses machine learning to classify and contextualize citation statements in academic publications. In collaboration with major publishers such as Wiley and SAGE, Scite has indexed over 280 million publications (including both open-access and subscription content) and 1.6 billion citation contexts. Citation context data for the data papers were collected using a paid account on the Scite platform. Using the DOIs of the target data papers, the citation contexts were retrieved through the Scite API (https://api.scite.ai/docs#section/Terms-of-Use). In this study, a citation context refers to the sentence in which a target paper is mentioned within the citing document, together with its surrounding sentences—typically the preceding and the following ones. Ultimately, 12,422 data papers were matched with 219,707 citation context entries. The experimental dataset generated in this study is available in the Science Data Bank repository²³.

To ensure the reliability and representativeness of the citation context data used in this study, we conducted an assessment of the Scite database as the primary data source, in comparison with the citation records available in the Web of Science (WoS) Core Collection. Among the 16,898 data papers included in the analysis, 14,051 had citation records in WoS, corresponding to 188,909 citing documents. Scite provided citation context information for 12,422 of these data papers, associated with 113,331 citing documents. Matching by DOI revealed that 75.3% (10,580 out of 14,051) of the data papers had at least one overlapping citing document between WoS and Scite, while the overlapping citing documents represented about 36% of all citing records in WoS. On average, the overlap of citing documents per paper was approximately 30%.

Although Scite does not cover all citing documents indexed in WoS, its citation context data encompass a diverse set of publication types, including journal articles, conference papers, theses, and books. This broader coverage enables a more comprehensive investigation of how scientific data papers are cited and discussed across different scholarly communication formats. Achieving complete citation-context coverage for all citing documents indexed in WoS would require full-text access to approximately 188,000 publications, which is currently constrained by copyright and data availability restrictions.

To test the representativeness of the Scite-indexed citation data, we compared metadata attributes of the citing documents found in Scite (about 36%) with those not covered (about 63%) in WoS. The two groups showed similar distributions in publication year, language, discipline, and journal impact factor (see Fig. 1). Specifically, both were dominated (>99%) by English-language publications; the temporal distributions between 2016 and 2023 were largely consistent, with slightly lower representation in Scite for 2024–2025; in disciplinary terms, Scite-covered citing documents showed marginally higher representation in Clinical & Life Sciences; and in terms of journal impact factor, Scite data were somewhat richer in journals with IF between 2 and 6, while slightly underrepresented in IF between 6 and 10. Overall, these observations indicate that the subset of citing documents indexed by Scite is structurally consistent with those recorded in the WoS Core Collection, supporting its adequacy for subsequent citation-context analysis.

Fig. 1.

Distributions of publication year, discipline, and journal impact factor for Scite-covered and WoS-only citing documents.

Accordingly, this study focuses on citation contexts available in the Scite index as the empirical basis for examining how data papers are cited and discussed in scholarly literature. While partial coverage introduces certain limitations, the observed structural similarity and transparent methodological framing support the validity of employing Scite for large-scale citation context analysis.

Methods

This study developed a multidimensional textual analysis framework based on the citation context data of data papers (see Fig. 2), with the aim of investigating the scholarly roles and contributions of data papers in the process of academic communication and knowledge dissemination. This framework consists of several components, including rhetorical analysis of in-text citations, classification of actual contribution types, identification of citation purposes, and, ultimately, citation behavior analysis and role attribution of data papers. In the following parts of this section, we provide a detailed account of the measurement of actual contribution types and citation purposes.

Fig. 2.

A multidimensional analytical framework for the citation roles of data papers.

Classification of actual contribution types for data papers

The identification of the actual contributions in data papers was conducted using an automatic recognition method developed in our previous study²¹. This method builds upon a comprehensive synthesis of prior research on the categorization of scientific contributions^24–26, and classifies actual scientific contributions into five distinct types: Theoretical, Experimental, Methodological, Data-based, and Other (Table 1). The classification approach employs a self-supervised fine-tuning strategy based on the open-source LLaMA2-13B model. The fine-tuning process utilized approximately 12,000 high-quality training instances, curated through a combination of manual annotation and GPT-4-assisted labeling. To enable parameter-efficient adaptation, the model was fine-tuned using the Low-Rank Adaptation (LoRA) technique. Following fine-tuning, the model achieved a macro-average F1 score of 0.94 on the scientific contribution type classification task, indicating robust and reliable performance.

Table 1.

Classification of actual contribution types for data papers.

Scientific contribution types	Description	Example
Theoretical	The citing paper draws upon theories, conceptual models, ideas, or viewpoints proposed by the target cited reference	An economics study builds upon rational choice theory < reference > to interpret patterns of socio-economic behavior.
Methodological	The citing paper employs or mentions experimental methods, methodological parameters, or frameworks designed by the target cited reference	A biological study uses the CRISPR-Cas9 protocol described in a prior publication < reference > .
Experimental	The citing paper utilizes, assesses or discusses the experimental findings or results derived from the target cited reference	A chemistry paper discusses the catalytic performance results previously reported in a previous study < reference > .
Data-based	The citing paper uses datasets, sample collections, or other experimental resources constructed by the target cited reference	A social science study uses the SciSciNet dataset < reference > for international collaboration research.
Other	Other contributions of the target cited reference to the citing paper, or contributions that cannot be categorized into the previous four types	A citing paper refers to a prior study < reference > for contextual information, historical background, or general review purposes.

Identification of citation purposes for data papers

A substantial body of research has explored the definition and identification of citation purposes or citation functions in scholarly communication²⁷. Existing classification frameworks range from broad, general-purpose annotation schemes^28–30 to fine-grained, domain-specific taxonomies developed for targeted semantic analysis^31–33. Corresponding identification methods have evolved from small-scale manual annotations to large-scale, automated approaches powered by AI models^34–36. While these citation function classification systems are well established, they are not fully suited to the unique characteristics of data papers, which focus on the systematic description, curation, and sharing of scientific datasets.

The primary aim of data papers is to facilitate data discovery, reuse, and deep analysis. However, existing taxonomies often fail to capture how citing papers actually engage with the underlying data, overlooking the specific roles that scientific datasets play in subsequent research. To address this gap, we propose a tailored citation purpose taxonomy specifically designed for data papers. This scheme draws on prior studies examining how researchers use scientific data^16,37 and on scholarship that delineates the functional roles of dataset citations^38,39, while also incorporating concepts from classical citation classification frameworks. It retains commonly observed citation purposes (e.g., providing background), but extends them to highlight purposes unique to data papers, including supporting computational experiments and enabling data integration. This approach aligns with the analytical logic adopted in studies of data citation functions. For example, Lafia et al.³⁸ categorized citation functions into critique, describe, illustrate, interact, and legitimize. Table 2 presents the complete set of citation-purpose types proposed in this study, along with their definitions and illustrative examples. It should be emphasized that such a rich categorization of citation purposes becomes feasible precisely because data papers provide detailed information about the corresponding datasets—their composition, usage, validation, and access links. These elements offer strong evidence of the quality, verifiability, and reproducibility of the research outcomes that cite them.

Table 2.

Classification of citation purpose types for data papers.

Purpose Type	Definition	Example
Background	Providing research context or reference background	Historical climate data < reference > as model background in geoscience study.
Inspiration	Proposing new questions or reinterpretations through existing data	Rediscovery of dark matter evidence from archival astronomy data < reference > .
Calculation	Data as input for modeling, algorithm training, or computational formulas	Genomic data < reference > driving disease prediction models in biomedicine study.
Comparison	Comparing datasets or results	Cross-national economic data < reference > comparisons for policy evaluation in social sciences research.
Integration	Aggregating multi-source data to generate new datasets	Multi-telescope data < reference > fusion in astronomy study.
Verification	Testing hypotheses, calibrating tools, or assessing consistency	Climate model parameter tuning with geological data < reference > .
Other	Applications beyond research contexts (e.g., education, policy, outreach)	Medical imaging datasets < reference > for educational training.

Given the large scale of our dataset—nearly 220,000 citation contexts—and the relatively high number of categories in our annotation scheme, we opted to use large language models (LLMs) to automate citation purpose classification, balancing computational complexity and classification efficiency. Our approach combines pre-trained LLMs with prompt engineering, adopting a few-shot prompting strategy. After preliminary testing of several mainstream LLMs, we focused on two models: DeepSeek-R1-67.1B and GPT-4o. To evaluate model performance, a stratified sampling approach was applied to select 1,227 citation context records, ensuring that sample sizes across major purpose categories were appropriately balanced. Two annotators independently classified the citation purpose type for each citation context, and disagreements were resolved by an additional expert adjudicator. The final labeled dataset constitutes the gold-standard test set for this study.

Following extensive testing of different prompt designs (Figs. 3 & 4) across both models (see Table 3), we found that GPT-4o, when paired with optimized prompts, delivered the best performance, achieving an F1-score of 0.81 on the test set. Accordingly, we used this model and prompt configuration for large-scale citation purpose classification across the full dataset.

Fig. 3.

The prompt scheme 1 for identifying citation purposes of data papers.

Fig. 4.

The prompt scheme 2 for identifying citation purposes of data papers.

Table 3.

Large language models for citation purpose identification and their performance.

Prompt scheme	Models	Accuracy	Precision	Recall	F1-score
Prompt1	DeepSeek-R1:671b	0.63	0.60	0.58	0.56
	GPT-4o	0.79	0.72	0.73	0.72
Prompt2	DeepSeek-R1:671b	0.75	0.68	0.69	0.67
	GPT-4o	0.85	0.82	0.80	0.80

^*Bold text highlights the best-performing prompt–model pairs and their F1-scores.

Identification for self-citations of data papers

To determine the institutional names and corresponding country information for each data paper, we first extracted the affiliations and addresses fields from the Web of Science (WoS) metadata. The affiliations field enables direct matching of institutional names, while the addresses field allows for straightforward extraction of country information. However, given that this study groups publications by country and subsequently calculates institution-level self-citations within each group, it was also necessary to retrieve the institution–country relationship for every data paper.

To efficiently process large-scale extraction of institution–country data, we adopted an automated approach combining a large language model (LLM) with prompt engineering. Specifically, we used LLaMA-13B as the LLM, and the prompt scheme is illustrated in Fig. 5. After obtaining the country information for each data paper, we applied a fractional counting method to compute the publication volume of data papers for each country.

Fig. 5.

The prompt scheme 2 for national and institutional self-citations identification.

To determine disciplinary self-citations of data papers, all citing records retrieved from the WoS were imported into InCites, and the disciplinary categories of the citing documents were assigned according to the citation topics (macro) schema.

Results

In-text rhetorical characteristics of data papers

To understand how data papers are rhetorically positioned within scholarly communication, we first examined their in-text linguistic features and structural patterns, conducting a comparative analysis across disciplines. An analysis of in-text mention frequency reveals that, on average, data papers are mentioned 1.67 times per citing article—comparable to that frequency of traditional academic paper⁴⁰. However, notable disciplinary differences emerge. Data papers in Earth Sciences exhibit a higher average number of in-text mentions, indicating their more integral role in supporting empirical research, while those in Arts and Humanities and Social Sciences are mentioned less frequently, reflecting the relatively lower centrality of structured scientific data in these domains.

We further analyzed the positional distribution of data paper mentions within the structure of citing documents. As shown in Fig. 6, data papers are most frequently mentioned in the Materials and Methods (24.28%) and Introduction (23.33%) sections. This pattern contrasts with that of traditional research articles, which are predominantly mentioned in the Introduction and Discussion sections^41,42. The higher proportion of data paper citations in methodological sections highlights their functional role in providing data-related support, reinforcing their position as resources rather than conceptual references.

Fig. 6.

The distribution of in-text locations for data papers within the citing studies.

Disciplinary breakdowns further illustrate these rhetorical tendencies (see Fig. 7). In fields such as Mathematics, Earth Sciences, and Clinical & Life Sciences, data papers are most frequently mentioned in the Materials and Methods section, while those in the Art & Humanities and Social Sciences as well as Engineering and Materials Science are more often mentioned in the Introduction, suggesting their use as contextual or background resources rather than as direct data inputs. Interestingly, in disciplines like Chemistry and Clinical and Life Sciences, data papers are relatively more frequently cited in the Results section—possibly because researchers in these areas often employ external datasets for comparative analysis or to validate their experimental outcomes, thereby integrating data papers more deeply into the core findings of the study. This pattern may also be partly attributable to disciplinary conventions in manuscript organization, as the structure of sections (e.g., strict or flexible adoption of the IMRaD format) differ across fields, shaping where citations to data papers tend to appear.

Fig. 7.

The distribution of in-text locations for data papers across 10 disciplines.

The actual contribution roles of data papers

Building upon the rhetorical analysis, we next explored the actual contribution roles that data papers play in citing documents, with the aim of revealing how these roles function within the broader process of scholarly knowledge production. Our finding indicates that in 42.7% of cases (Table 4), data papers definitely make a data-based contribution to the citing work. However, in over half of the citation context instances, data papers fulfill a variety of additional roles beyond simply supporting data. These include facilitating experimental comparisons and result validation, offering methodological guidance, and providing contextual or background information relevant to the research topic. This diversity of functions highlights the multifaceted value of data papers and suggests that their utility extends well beyond data application alone.

Table 4.

The proportion of various actual contributions of data papers.

Contribution types	Proportion of citation context instances (%)
4.Data-based	42.70%
3.Experimental	22.50%
2.Methodological	20.65%
1.Theoretical	11.70%
5.Other	2.46%

When analyzed by discipline (Fig. 8), the proportion of data-based contributions is notably higher in fields such as Electrical Engineering, Electronics and Computer Science, where empirical data often form the foundation of research. In contrast, disciplines such as Engineering and Materials Science, Mathematics, and Chemistry exhibit a lower proportion of data-based contributions in citation contexts. Data papers in these fields are more frequently cited for their methodological or experimental components—for instance, mathematical models, computational scripts, or detailed laboratory protocols. This pattern suggests that data papers in these disciplines may not contribute directly through the reuse of datasets themselves, but rather through supporting other aspects of the citing research, such as analytical workflows and data preparation processes. In some cases, the contribution may be indirect, reflecting the utility of the data paper in enabling or informing experimental setups rather than serving as a primary data source.

Fig. 8.

Distribution of actual contribution types made by data papers in 10 research areas.

The purpose of citing data papers

To gain deeper insights into the motivations behind citing data papers, we conducted a citation purpose analysis focusing specifically on contexts where data papers clearly make a data-based contribution. As illustrated in Fig. 9, nearly half of these citations—despite involving a data-based contribution—are used merely to provide background information about the datasets. A notable portion of citation contexts reflects the use of data papers in experimental calculations, suggesting that the datasets described in these papers may have been directly involved in the computational or experimental procedures of the citing research. Additionally, data papers are cited for purposes such as data integration (11.99%), comparative data analysis (11.79%), and data benchmarking or validation (11.54%), all of which exhibit relatively similar mention frequencies. These patterns indicate that the detailed descriptions and documentation in data papers primarily enhance the transparency and usability of the datasets they describe, which in turn facilitates their adoption in citing studies and enables the execution of experimental procedures directly linked to those datasets, ultimately advancing the rigor and scope of subsequent research.

Fig. 9.

Citation purposes distribution of data papers with data-based contributions.

Interestingly, only a very small number of citations indicate that data papers serve as sources of research inspiration—that is, sparking new research questions or conceptual frameworks. This finding highlights a potential limitation in the current role of data within the scholarly ecosystem: although scientific research is increasingly described as ‘data-driven,’ citation behavior suggests that data rarely acts as a primary catalyst for innovation. Instead, data papers tend to support or reinforce existing research agendas rather than initiate them.

When examining disciplinary differences, we found that the overall distribution of citation purposes for data papers remains relatively consistent across disciplines, with some notable variations in specific fields (Fig. 10). In particular, data papers in the Clinical and Life Sciences and Earth Sciences exhibit a higher proportion of citations for usage-related purposes—including experimental calculations, data integration, comparative analysis, and validation—compared to the cross-disciplinary average. In contrast, such usage-oriented citations are considerably less common in fields like the Arts and Humanities and Mathematics. This disparity may be attributed to differences in research practices and epistemological frameworks. For instance, in the Arts and Humanities, data may play a more contextual or interpretive role rather than serving as input for formal computation. Similarly, in Mathematics, research is often theoretical or model-based, relying less on external datasets and more on logical derivation or abstract reasoning. These findings highlight how disciplinary cultures and methodological norms shape the ways in which data papers are cited, emphasizing that the value and function of research data are not uniform but context-dependent across academic fields.

Fig. 10.

Citation purposes distribution of data papers with data contributions (10 disciplines).

Characteristics of cited data papers

To complement the preceding analyses, we investigated potential factors influencing the citation patterns of data papers by examining the characteristics of cited data papers themselves, including self-citation behaviors and relevant external attributes. As a first step, we assessed the dissemination scope of the cited data papers. Specifically, we identified the top ten countries in terms of data paper output and analyzed the national-level, institutional-level and discipline-level self-citation rates for these papers.

At the national level (Fig. 11a), the analysis of data papers across the ten countries shows a wide range of national self-citation rates, with several countries—such as China, the United States, and India—exceeding median values exceeding 50%, and Italy approaching that level with a median of around 45%. In contrast, the remaining countries generally exhibit rates near or below 40%, indicating a comparatively stronger dependence on citations from the broader international research community. At the institutional level (Fig. 11b), the self-citation rates appear more consistent across countries, with most averages falling within the 30%–40% range. This relative uniformity may reflect similar institutional publishing and citation practices, regardless of the country’s overall output, and implies that institutional self-citation behavior is less variable than national self-citation behavior.

Fig. 11.

The distribution of the proportion of self-citations (country-level and institution-level) in data papers of various countries.

We further examined the disciplinary self-citation patterns of data papers across ten disciplines (Fig. 12). Overall, the median self-citation rates in most disciplines are high, generally exceeding 80%, indicating that scientific data—particularly in the natural sciences and engineering—tend to circulate predominantly within the disciplinary system, reflecting a high degree of specialization. Based on the observed characteristics, the disciplines can be broadly classified into three types:

1. High self-citation with narrow variability: Clinical and Life Sciences, Electrical Engineering and Computer Science, Earth Sciences, Mathematics, and Physics exhibit high median self-citation rates. Notably, Clinical and Life Sciences, Electrical Engineering and Computer Science, and Mathematics display relatively small interquartile ranges (IQRs), suggesting a pronounced disciplinary “inwardness” in data use, where data papers are predominantly cited within the same field.

2. High self-citation with wide variability: Agriculture, Environmental and Ecology, Chemistry, and Social Sciences also show high self-citation rates, but their IQRs are considerably larger. This pattern suggests more diverse data usage: while certain subfields rely heavily on data generated within their own discipline, others produce research data with greater interdisciplinarity and general applicability, supporting scientific experiments across multiple fields.

3. Low citation and low self-citation: Arts and Humanities data papers receive the fewest citations overall, with the lowest median self-citation rate. Cross-disciplinary citation analysis indicates that, beyond citations from within the field, these data papers are cited relatively frequently by Clinical and Life Sciences and Social Sciences. This may be due to the nature of the data—such as historical archives, cultural heritage records, linguistic corpora, and visual materials—which hold substantial value for interdisciplinary research in areas including medical humanities, social behavior studies, public health, and psychology.

Fig. 12.

The distribution of the proportion of self-citations (discipline-level) in data papers of various disciplines.

Next, we examined the relationship between the data dissemination performance of data papers and their key attributes. To differentiate the dissemination performance, we classified data papers into two categories based on the proportion of data-based contributions, using a threshold of 50%. Papers categorized as data-centric exhibit a higher level of data-based contribution, indicating that they play a more direct and effective role in disseminating the associated datasets. In contrast, data-complementary papers place greater emphasis on contextual information, methodologies, or supporting analyses, and thus exert a relatively weaker influence on the dissemination of the datasets. We then explored how varying levels of data reuse relate to specific paper attributes.

The results reveal that, in terms of funding support, data-centric papers tend to secure slightly higher levels of funding—82.27% of such papers acknowledge financial support—compared to 78.34% for data-complementary papers. Although modest, this difference may suggest that projects with more direct dataset dissemination are perceived as having greater potential impact by funding agencies, possibly due to their higher visibility and utility for the broader research community. Regarding journal distribution (Fig. 13), while Data in Brief and Scientific Data are the leading outlets in terms of overall data paper volume, the proportion of data-centric papers is notably higher in Scientific Data (37.29%). Moreover, in Ecology and Geoscience Data Journal, data-centric papers account for more than 50% of publications. These results indicate that the data dissemination effectiveness of these journals is not necessarily tied to their overall publication volume; even journals with relatively small numbers of data papers can play a “small but strong” role in promoting dataset reuse when the content is highly focused and impactful. From a disciplinary perspective (Fig. 14), data papers in fields such as Agriculture, Earth Sciences, and Arts and Humanities display notably higher levels of data reuse. These patterns may reflect the inherently data-intensive nature of research in these domains, as well as the strong demand for accessible datasets to facilitate comparative analyses and longitudinal studies.

Fig. 13.

Journal distributions of the two types of data papers. The journals on the vertical axis are ordered from top to bottom according to decreasing publication volume.

Fig. 14.

The normalized proportion of data-centric data papers in each discipline.

Discussion and Conclusion

This study offers a comprehensive examination of the contribution roles, citation purposes, and disciplinary variations of data papers within scholarly communication. By integrating rhetorical analysis, citation context analysis, and attribute-based dissemination evaluation, we present a multi-dimensional perspective on how data papers function as vehicles for research data reuse and knowledge exchange. Overall, our findings indicate that data papers play a meaningful role in disseminating and documenting the scientific datasets they describe, thereby enhancing their visibility and accessibility within the research community. However, their current function falls short of fully embodying the role of active data reusers or integrators. In most cases, citations to data papers emphasize descriptive or contextual information rather than direct reuse of data in computational, integrative, or comparative research activities. This gap reveals that data papers function as important catalysts in facilitating dataset use, while their capability to stimulate deeper and more innovative applications has not been fully harnessed.

The roles of data papers in scholarly communication

Our findings confirm that data papers provide clear and tangible data-related contributions to the works that cite them; however, their utility extends beyond the direct application of data. In most cases, the citations primarily involve references to the research results and methodological aspects of the data papers, which is consistent with the findings reported by Li et al.¹⁷. Moreover, in more than half of the citation contexts, data papers fulfil broader scholarly communication functions—serving as sources of methodological guidance, experimental comparison, and result validation—thus exhibiting characteristics more typical of research articles. Nevertheless, when narrowing the focus to citations involving explicit data-based contributions, the predominant role of data papers is to provide background information about the datasets, with relatively fewer cases of deeper integration into computational workflows, data synthesis, comparative analysis, or benchmarking. This result aligns with the observations of Lafia et al.³⁸, who similarly found that data papers more often act as “additional context” that help citing researchers to gain a clearer understanding of the underlying data.

From this perspective, data papers often operate as “dataset appendices” to research articles—offering reusable protocols, comprehensive documentation, and benchmark datasets that enhance transparency and reproducibility. The substantial proportion of background-oriented citations further indicates that this publication format significantly increases the visibility and scholarly recognition of data producers, enabling their work to be more readily discovered and cited. However, the relatively low share of citations reflecting an inspirational or generative role suggests that data papers primarily enhance research efficiency rather than stimulate novelty—accelerating the progress of “normal science” but only rarely acting as triggers for paradigm shifts.

For scientometric evaluation, these observations underscore the importance of assessing not only the practical utility of data papers and the datasets they describe, but also their intellectual and conceptual impact. By fostering richer dataset integration, cross-domain reuse, and more creative applications, data papers could evolve from primarily serving as background resources into active drivers of conceptual innovation.

Disciplinary differences and epistemic cultures of data papers

Our findings show that the role of data papers varies substantially across disciplines, which is consistent with the observations of Jiao and Darch¹⁶. This disciplinary specificity may, on the one hand, be related to differences in data reusability, while on the other hand reflecting distinct epistemic cultures⁴³. In Earth Sciences and Life Sciences, high rates of data-based contributions and frequent use in computational or experimental procedures point to data-intensive research paradigms, where datasets are treated as central research objects. In Mathematics and Chemistry, data papers function more often as sources of methodological reference—providing technical guidance such as algorithms, models, or protocols rather than datasets for direct reuse. In the Arts and Humanities, citations to data papers are primarily contextual—they tend to provide background information or conceptual framing rather than operational inputs. Such citation motivations may represent a way for researchers to acknowledge intellectual debt³⁹.

These disciplinary variations demonstrate the influence of methodological norms on citation behavior. Instrumental fields (e.g., engineering and applied sciences) tend to treat datasets as reusable, manipulable research objects, whereas more conceptual fields (e.g., mathematics and theoretical sciences) engage with data papers primarily as process blueprints or contextual frameworks. This disciplinary divergence not only helps explain current developmental imbalances, but also suggests that each field can leverage its epistemic culture to implement targeted strategies for scientific data management and open sharing—thereby unlocking greater value from data resources.

For example, in the life sciences, adopting standardized metadata schemas and jointly publishing data papers with methodological articles can enhance cross-project dataset reuse and foster multi-platform collaboration. In the social sciences, where data interpretability is tied closely to specific research contexts and temporal–spatial boundaries, data papers could place greater emphasis on documenting the generalizable attributes of datasets and providing reproducible analytical procedures, thus lowering the barriers to secondary analysis.

Properties, dissemination, and reuse patterns of data papers

Our analysis of the attributes of cited data papers provides insights into how these characteristics relate to the dissemination and reuse of scientific data. Across the dataset, self-citation rates are generally high, indicating that a substantial proportion of reuse occurs within the same communities that produced the data. When examined from a disciplinary perspective, pronounced self-citation patterns are observed in fields such as the life sciences and the earth sciences, suggesting that these domains may have more cohesive research networks, stronger thematic continuity, and well-established norms for citing prior work within the same field. Such discipline-specific tendencies may be influenced by several factors, including the degree of specialization of research topics, the prevalence of domain-specific data repositories, and the maturity of data-sharing practices within each field. While national-level differences in self-citation patterns exist, institutional-level rates remain relatively consistent—typically within the 30–40% range—implying that local collaborative networks and organizational practices play a stabilizing role. Overall, these findings suggest that the reuse of data papers is shaped more by disciplinary cultures and collaboration structures than by national policy environments alone.

Furthermore, our classification of data papers into data-centric and data-complementary types indicates that those with higher data-based contributions tend to receive slightly greater funding support and are more frequently published in certain journals and disciplines. Notably, journals such as Scientific Data, Ecology, and Geoscience Data Journal—despite varying in their overall publication output—exhibit a superior dissemination impact, indicating that content quality and management of publishing can outweigh sheer quantity in driving dataset reuse.

Despite the diverse roles and disciplinary patterns identified above, several systemic challenges continue to constrain the full scholarly potential of data papers and the datasets they describe. First, the overall visibility and recognition of data papers remain limited, with many researchers still perceiving them as peripheral rather than integral components of scholarly communication¹⁰. Second, the accessibility of datasets across various storage platforms is uneven, with interoperability and long-term availability yet to be fully ensured⁴⁴. Third, citation and attribution practices for both data papers and underlying datasets lack sufficient standardization⁴⁵, hindering accurate tracking of reuse and impact. Finally, effective data reuse is fundamentally grounded in trust among research communities; yet, when the authenticity or quality of shared datasets cannot be readily verified⁴⁶, scholars often default to relying on self-generated data.

Looking forward, the potential of data papers could be further unlocked through coordinated efforts in publication management, the establishment of clearer policy standards, and the cultivation of shared epistemic awareness within scholarly communities. Publishers can play a pivotal role by adopting rigorous metadata requirements, encouraging transparent peer review of datasets, and fostering cross-journal interoperability for data linking. Policymakers may strengthen open data mandates and incentivize international repository integration, thereby reducing fragmentation in data accessibility. Equally important, academic communities should recognize data papers not merely as ancillary outputs, but as primary scholarly contributions—integral to research reproducibility, innovation, and cross-disciplinary collaboration. Together, these measures could enhance both the visibility and reusability of scientific datasets, amplifying their impact across national and disciplinary boundaries.

Our study has certain limitations. First, regarding the experimental data, due to copyright restrictions, we were unable to obtain all citation contexts from studies citing data papers. The use of proprietary sources like Web of Science (WoS) and Scite limits full reproducibility. We recognize that open infrastructures such as OpenAlex and the OpenAIRE Graph offer valuable and transparent alternatives for identifying data papers. In future work, we plan to explore these resources and collaborate with domain publishers and experts to build larger and more open citation-context datasets. Second, concerning the coverage of scientific disciplines, although the study includes ten major fields, this selection does not represent the entire scholarly landscape. Our goal was not exhaustive coverage, but rather to focus on a sufficiently diverse disciplinary sample to reveal and illustrate patterns in how data papers are cited across fields. Future research could expand this work by incorporating additional or currently underrepresented areas—such as Social Sciences and Arts and Humanities—to examine whether these patterns hold in a broader disciplinary context. Third, the identification of contribution types and citation purposes also has limitations. The classification framework still requires further refinement and contextual adaptation. Moreover, with the rapid advancement of large language models (LLMs), there remains significant potential to improve the automated identification of these categories.

Building on the current findings, our future research will further examine the evolving academic role of data papers, assessing whether their functions within scholarly communication have shifted over time in response to changes in data-sharing practices and research workflows. We will also conduct a more detailed investigation into the publishing attributes of cited data papers and their citing literature, with the aim of identifying patterns in editorial policies, disciplinary contexts, and citation behaviors that may influence dataset reuse. This will include correlation analyses to reveal statistical associations, as well as causal analyses to examine whether specific publishing factors directly drive dataset reuse. In addition, we plan to develop and apply robust indicators to measure the intellectual impact and novelty contribution of data papers and the scientific datasets they describe. Such efforts will not only enrich our understanding of how data papers shape knowledge dissemination, but also provide empirical evidence to guide publication strategies, policy-making, and community recognition of data-centered scholarship.

Author contributions

L.C. and X.L. designed research; L.C. performed research; L.C. and Z.Q. analyzed data; L.C. and X.L. wrote the paper; X.L. and N.M. reviewed the manuscript.

Funding

This work was supported by the China Association for Science and Technology Foundation (Grant No. 2023QNRC001).

Data availability

The dataset used in this study has been deposited in the Science Data Bank, accessible via the following link: https://doi.org/10.57760/sciencedb.31287. This dataset includes the metadata of data papers retrieved from the Web of Science (WoS), citation contexts obtained from Scite together with the computed actual contribution type and citation purpose type, as well as the standardized national and institutional self-citation data of data papers. Due to copyright restrictions, only partial sample data from WoS and Scite are provided. In addition, the dataset comprises the primary statistical data used in this research.

Code availability

The codes used for data processing and analysis in this study have been deposited in the Science Data Bank repository²³. The coeds can be accessed via the following link: https://doi.org/10.57760/sciencedb.31287.

Competing interests

The authors declare no competing interests.