Controlled Vocabulary Implementation

Controlled Vocabulary Implementation in AI Discoverability Architecture refers to the systematic application of standardized, predefined terminology sets to enhance the findability, accessibility, and semantic understanding of AI systems, models, and datasets ¹². Its primary purpose is to establish consistent metadata frameworks that enable efficient indexing, retrieval, and interoperability across diverse AI resources and platforms ³. This implementation matters critically because it addresses the growing challenge of AI resource fragmentation, where thousands of models, datasets, and tools exist across disparate repositories without unified discovery mechanisms ⁴. By establishing controlled vocabularies, organizations can create semantic bridges that allow both human researchers and automated systems to locate, evaluate, and integrate AI resources more effectively ⁵.

Overview

The emergence of controlled vocabulary implementation in AI discoverability architecture stems from the exponential growth of machine learning resources and the resulting discovery crisis ¹². As AI research accelerated in the 2010s, repositories like Hugging Face, TensorFlow Hub, and Papers with Code proliferated, each developing independent classification schemes that hindered cross-platform resource discovery ³. The fundamental challenge addressed by controlled vocabularies is the vocabulary mismatch problem—where semantically equivalent concepts are described using different terminology, preventing users from finding relevant resources ⁴⁵.

Historically, early AI repositories relied on free-text tagging and ad-hoc categorization, which created inconsistent metadata that degraded search effectiveness as collections scaled ²⁶. The practice evolved significantly with the adoption of semantic web technologies and linked data principles in the late 2010s, enabling vocabularies to be published as machine-readable resources with unique identifiers ³⁷. Contemporary implementations now integrate ontology engineering methodologies, faceted classification approaches, and automated metadata generation techniques to manage vocabularies that must accommodate rapidly evolving AI terminology while maintaining semantic precision ¹⁴.

Key Concepts

Term Standardization

Term standardization involves mapping synonymous concepts to single preferred terms with explicit scope definitions ²³. This process establishes canonical forms that eliminate ambiguity and ensure consistent indexing across resources. For example, a controlled vocabulary might designate "transformer architecture" as the preferred term, mapping variants like "transformer model," "transformer network," and "attention-based architecture" to this single authorized descriptor. When researchers at a pharmaceutical company search for transformer-based drug discovery models, standardization ensures they retrieve all relevant resources regardless of which terminology variant the original creators used in their documentation.

Hierarchical Relationships

Hierarchical relationships establish parent-child taxonomic structures that organize terms from general to specific ¹⁴. These broader/narrower term relationships enable users to navigate from high-level concepts to specialized subcategories. Consider a vocabulary where "neural network" serves as a broader term with narrower terms including "convolutional neural network," "recurrent neural network," and "transformer network," each of which has further subdivisions. A computer vision researcher searching for image segmentation models can start with the broad "convolutional neural network" category and drill down to "U-Net architecture" and "Mask R-CNN," discovering specialized resources they might not have known to search for initially.

Semantic Equivalence Mapping

Semantic equivalence mapping creates explicit connections between preferred and non-preferred terms, ensuring comprehensive retrieval regardless of terminology variations ³⁵. This concept recognizes that different research communities may use distinct terminology for identical concepts. For instance, a vocabulary might map "ANN," "artificial neural network," and "connectionist model" as non-preferred synonyms of the preferred term "neural network." When a researcher from a neuroscience background searches using "connectionist model"—terminology common in cognitive science—the system automatically retrieves resources tagged with "neural network," bridging disciplinary vocabulary differences.

Faceted Classification

Faceted classification organizes AI resources along multiple independent dimensions that users can combine to refine searches ²⁶. Unlike hierarchical taxonomies with single classification paths, faceted approaches allow resources to be described simultaneously by architecture type, task category, domain application, training methodology, and computational requirements. A medical imaging AI platform might implement facets for "Model Architecture" (CNN, transformer, hybrid), "Clinical Task" (diagnosis, segmentation, prediction), "Imaging Modality" (CT, MRI, X-ray), and "Validation Status" (FDA-approved, research-only). Radiologists can combine "transformer + diagnosis + CT + FDA-approved" to find precisely the regulatory-compliant diagnostic models relevant to their clinical workflow.

Metadata Schema Integration

Metadata schema integration specifies which controlled vocabulary terms apply to particular resource attributes, creating structured descriptions ⁴⁷. This concept ensures that vocabularies are applied consistently across different metadata fields. For example, an AI model repository might define a schema where the "architecture" field accepts only terms from the model architecture vocabulary, the "task" field uses the task taxonomy, and the "domain" field draws from a domain classification. When a researcher uploads a BERT-based sentiment analysis model for financial text, the schema guides them to select "transformer" for architecture, "text classification" for task, and "finance" for domain, ensuring complete and consistent metadata that supports precise discovery.

Vocabulary Versioning and Provenance

Vocabulary versioning and provenance tracking documents the evolution of controlled vocabularies over time, maintaining backward compatibility while accommodating new concepts ¹³. This concept recognizes that AI terminology evolves rapidly, requiring systematic management of vocabulary changes. A research institution's AI catalog might implement version 1.0 of their vocabulary in 2020, add "few-shot learning" and "prompt engineering" terms in version 2.0 (2022), and introduce "constitutional AI" in version 3.0 (2024). Each model's metadata records which vocabulary version was used for tagging, allowing the system to interpret historical metadata correctly while new submissions benefit from current terminology. When a 2020 model tagged with deprecated terms is retrieved, the system can display both the original metadata and equivalent current terminology.

Cross-Vocabulary Mapping

Cross-vocabulary mapping enables translation between different controlled vocabularies, facilitating interoperability across systems using distinct standards ⁵⁶. This concept is essential when integrating resources from multiple repositories with independent classification schemes. For instance, a federated AI search platform might map between Hugging Face's task taxonomy, Papers with Code's method categories, and an institutional repository's custom vocabulary. When a user searches for "object detection" in the federated system, cross-vocabulary mappings translate this to equivalent terms in each repository—"object-detection" in Hugging Face, "Object Detection" in Papers with Code, and "visual_object_recognition" in the institutional system—enabling comprehensive retrieval across all platforms despite their different terminology conventions.

Applications in AI Resource Discovery

Model Repository Discovery Systems

Controlled vocabularies are extensively applied in model repositories to enable precise discovery among thousands of pre-trained models ²³. Hugging Face implements controlled tags for model architectures (BERT, GPT, T5), tasks (text-classification, translation, question-answering), and languages, allowing users to filter over 100,000 models efficiently. The vocabulary includes hierarchical relationships where "transformer" serves as a parent category with specific architectures as children, and semantic mappings that connect "seq2seq" with "sequence-to-sequence" and "encoder-decoder." This implementation enables both novice users browsing high-level categories and expert researchers seeking specific architectural variants to discover relevant models effectively.

Dataset Cataloging and Search

Dataset discovery platforms leverage controlled vocabularies to describe data characteristics, collection methodologies, and domain applications ⁴⁷. Google Dataset Search employs Schema.org vocabularies extended with AI-specific terms to index datasets across the web, using controlled terms for data modalities (image, text, tabular), collection methods (crowdsourced, sensor-generated, synthetic), and domain applications (healthcare, autonomous-vehicles, climate). When a robotics researcher searches for "sensor-generated navigation datasets," the controlled vocabulary ensures retrieval of relevant resources regardless of whether creators described them as "sensor data," "telemetry data," or "perception datasets," while filtering out irrelevant crowdsourced or synthetic alternatives.

Experiment Tracking and Reproducibility

Machine learning experiment platforms implement controlled vocabularies to standardize descriptions of training configurations, hyperparameters, and evaluation metrics ¹⁵. OpenML uses comprehensive vocabularies covering algorithm types (decision-tree, neural-network, ensemble), hyperparameter categories (learning-rate, regularization, architecture-depth), and performance metrics (accuracy, F1-score, AUC-ROC). This standardization enables meta-learning research that analyzes patterns across thousands of experiments—for instance, identifying which hyperparameter configurations consistently perform well for specific dataset characteristics. The controlled vocabulary ensures that experiments from different research groups can be meaningfully compared despite variations in documentation practices.

AI Ethics and Governance Documentation

Controlled vocabularies increasingly support AI ethics and governance by standardizing terminology for bias assessment, fairness metrics, and regulatory compliance ³⁶. A healthcare AI platform might implement vocabularies for bias categories (selection-bias, measurement-bias, algorithmic-bias), fairness metrics (demographic-parity, equalized-odds, individual-fairness), and regulatory frameworks (HIPAA-compliant, FDA-approved, GDPR-aligned). When hospital administrators search for "HIPAA-compliant diagnostic models with documented demographic-parity assessment," the controlled vocabulary enables precise filtering to models meeting specific ethical and regulatory requirements, supporting responsible AI deployment in clinical settings.

Best Practices

Implement Agile Vocabulary Maintenance Cycles

Establish frequent review processes—monthly or quarterly rather than annual—to incorporate emerging AI terminology and respond to user feedback ¹². The rationale is that AI terminology evolves exceptionally rapidly, with new architectures, techniques, and paradigms emerging continuously. Annual review cycles result in vocabularies that lag significantly behind current usage, reducing their utility and encouraging users to bypass controlled terms in favor of free-text descriptions. A practical implementation involves creating a vocabulary governance committee that meets monthly to review term proposals submitted through a web form, analyze usage statistics from search logs, and monitor AI research literature for emerging terminology. For example, when "diffusion models" gained prominence in 2022, agile processes enabled repositories to add this term within weeks rather than waiting for annual reviews, ensuring timely discoverability of resources using this approach.

Balance Specificity with Usability Through Progressive Disclosure

Design vocabulary interfaces that present high-level categories initially while allowing drill-down to specialized terms, preventing overwhelming complexity ³⁴. The rationale is that overly granular vocabularies displayed simultaneously create cognitive overload for users and metadata creators, while overly coarse vocabularies fail to distinguish meaningfully different resources. Implementation involves creating hierarchical vocabularies with 5-10 top-level categories, each containing 10-20 second-level terms, with additional depth available on demand. For instance, a model repository might display top-level categories like "Computer Vision," "Natural Language Processing," and "Reinforcement Learning" initially. When users select "Computer Vision," second-level terms like "Object Detection," "Image Segmentation," and "Image Generation" appear. Only when "Object Detection" is selected do specialized terms like "YOLO," "Faster R-CNN," and "DETR" become visible, guiding users progressively toward precise terminology without initial overwhelm.

Integrate Auto-Suggestion and Intelligent Defaults

Implement interfaces that recommend appropriate controlled terms as users type and suggest defaults based on resource characteristics ⁵⁶. The rationale is that adoption barriers—including learning curves and resistance to constrained terminology—significantly impede controlled vocabulary usage. Auto-suggestion reduces friction by helping users discover appropriate terms without memorizing the entire vocabulary. A practical implementation involves developing an auto-complete interface that searches both preferred and non-preferred terms, displaying matches with scope notes. For example, when a researcher types "conv," the system suggests "convolutional neural network (CNN architecture for processing grid-structured data)" and "conversational AI (systems for dialogue interaction)," helping disambiguate similar terms. Additionally, the system analyzes uploaded model code or documentation using natural language processing to suggest relevant terms—detecting TensorFlow Conv2D layers and recommending "convolutional neural network" as a likely architecture descriptor.

Establish Cross-Vocabulary Mappings for Interoperability

Create explicit mappings between your controlled vocabulary and widely-used external vocabularies to enable federated discovery ²⁷. The rationale is that AI resources exist across multiple repositories with independent classification schemes, and users benefit from searching across platforms simultaneously. Implementation involves identifying major external vocabularies relevant to your domain—such as Schema.org for general datasets, Papers with Code taxonomies for research models, or domain-specific ontologies—and creating equivalence mappings. For instance, an institutional AI repository might map its "supervised_classification" term to Schema.org's "classification" property, Papers with Code's "Classification" task, and Hugging Face's "text-classification" tag. These mappings are encoded in machine-readable formats like SKOS, enabling federated search systems to translate queries across vocabularies. When a user searches the institutional repository for "supervised_classification," the system can simultaneously query external repositories using their equivalent terms, providing comprehensive discovery across the AI ecosystem.

Implementation Considerations

Tool and Format Selection

Selecting appropriate tools and formats for vocabulary management significantly impacts implementation success and long-term sustainability ¹³. Organizations must choose between lightweight formats like simple CSV files or JSON structures versus sophisticated semantic web technologies like SKOS (Simple Knowledge Organization System) and OWL (Web Ontology Language). For small-scale implementations with fewer than 500 terms and limited hierarchical complexity, CSV-based vocabularies managed in spreadsheet software may suffice, offering low technical barriers and easy editing. However, as vocabularies grow beyond 1,000 terms or require complex semantic relationships, dedicated vocabulary management platforms like PoolParty, Synaptica, or open-source tools like Skosmos become essential. These platforms support SKOS serialization, enabling vocabularies to be published as linked data with unique URIs for each concept. For example, a research consortium implementing a shared AI model vocabulary might use Skosmos to publish their taxonomy at https://vocab.example.org/ai-models/, where each term like "transformer-architecture" receives a persistent URI (https://vocab.example.org/ai-models/transformer-architecture) that can be referenced in metadata across multiple repositories, facilitating true semantic interoperability.

Audience-Specific Customization

Tailoring vocabulary granularity and terminology to specific user communities enhances adoption and effectiveness ⁴⁵. Different audiences—from novice practitioners to domain experts—require different levels of specificity and may use distinct terminology conventions. Implementation involves creating vocabulary views or subsets appropriate for each audience while maintaining a comprehensive master vocabulary. For instance, a healthcare AI platform might maintain a detailed master vocabulary with highly specific terms like "U-Net-with-attention-gates" and "ResNet50-pretrained-ImageNet," but present clinical users with a simplified view using terms like "segmentation-model" and "diagnostic-classifier" that align with clinical workflows. Conversely, the platform might offer AI researchers an expanded view including architectural details and training methodology terms. This customization can be implemented through role-based interfaces that filter vocabulary displays based on user profiles, or through multiple vocabulary "lenses" that users can select. Documentation and training materials should be similarly customized—clinical users receive guides focused on task-oriented discovery, while AI developers receive technical documentation explaining architectural distinctions.

Organizational Maturity Assessment

Successful implementation requires assessing organizational readiness across technical infrastructure, staff expertise, and cultural factors ²⁶. Organizations with limited information management maturity may struggle with sophisticated ontology engineering approaches, while those with established metadata practices can leverage advanced semantic technologies. Implementation should begin with a maturity assessment evaluating factors including existing metadata practices, staff knowledge organization expertise, technical infrastructure capabilities, and stakeholder buy-in. For example, a startup AI company with minimal metadata infrastructure might begin with a simple flat list of 50-100 essential terms managed in a shared document, focusing on establishing consistent tagging habits before investing in complex hierarchical structures. In contrast, a large research institution with established library science expertise and semantic web infrastructure might immediately implement a comprehensive SKOS-based ontology with thousands of terms, hierarchical relationships, and integration with institutional knowledge graphs. The assessment should inform a phased implementation roadmap—perhaps starting with a minimal viable vocabulary for critical use cases, demonstrating value through improved discovery metrics, then expanding scope and sophistication as organizational capabilities mature.

Multilingual and Cross-Cultural Considerations

As AI research globalizes, implementing multilingual vocabularies with culturally appropriate terminology becomes increasingly important ³⁷. This consideration extends beyond simple translation to addressing conceptual differences across languages and research traditions. Implementation involves creating parallel vocabularies in target languages with explicit equivalence mappings, ideally developed by native-speaker domain experts who understand both linguistic nuances and AI concepts. For instance, a European AI research network might implement vocabularies in English, German, French, and Spanish, where the English term "deep learning" maps to German "tiefes Lernen," French "apprentissage profond," and Spanish "aprendizaje profundo." However, implementation must address cases where direct equivalents don't exist—some languages may lack established translations for emerging concepts like "prompt engineering" or "constitutional AI," requiring either transliteration, descriptive phrases, or adoption of English terms. Technical implementation typically uses SKOS's multilingual capabilities, where each concept has a unique identifier with labels in multiple languages. User interfaces detect browser language settings or allow explicit language selection, displaying vocabulary terms in the user's preferred language while maintaining consistent underlying concept identifiers that enable cross-language discovery.

Common Challenges and Solutions

Challenge: Terminology Volatility and Rapid Evolution

AI terminology evolves exceptionally rapidly, with new architectures, techniques, and paradigms emerging continuously ¹². Terms that are cutting-edge today may become obsolete within months, while entirely new concepts require immediate vocabulary incorporation to maintain relevance. This volatility creates significant maintenance burden and risks vocabularies becoming outdated quickly. For example, between 2020 and 2023, terms like "vision transformer," "diffusion model," "prompt engineering," and "retrieval-augmented generation" emerged and became central to AI discourse, while earlier terms like "capsule network" declined in prominence. Organizations struggle to keep vocabularies current without destabilizing existing metadata or overwhelming users with constant changes.

Solution:

Implement a tiered term status system with categories including "established" (stable, widely-used terms), "emerging" (new concepts gaining traction), and "deprecated" (outdated terms maintained for backward compatibility) ³⁴. Establish monthly vocabulary review meetings where a governance committee evaluates term proposals, monitors AI research literature and conference proceedings, and analyzes search log data to identify terminology gaps. Create a public term proposal system where users can suggest new terms with justification and usage examples. For emerging terms, implement a provisional acceptance process—add the term with an "emerging" flag and monitor its usage for 3-6 months before promoting to "established" status. For example, when "diffusion model" emerged in 2022, it could be added as "emerging" immediately, then promoted to "established" after demonstrating sustained usage across multiple resources. Maintain deprecated terms as non-preferred synonyms mapped to current equivalents, ensuring historical metadata remains interpretable. Communicate vocabulary changes through release notes and email notifications, but implement changes gradually—perhaps quarterly vocabulary version releases—to avoid constant disruption while maintaining currency.

Challenge: Granularity Calibration and Scope Creep

Determining appropriate vocabulary granularity presents persistent challenges—overly coarse vocabularies fail to distinguish meaningfully different resources, while excessive granularity overwhelms users and creates maintenance burden ²⁵. Organizations often experience scope creep, where vocabularies expand uncontrollably as stakeholders request increasingly specific terms for niche use cases. For instance, a model repository might debate whether to include separate terms for "BERT," "RoBERTa," "ALBERT," "DistilBERT," and dozens of other BERT variants, or simply use "BERT-based model" as a broader category. Too much specificity creates vocabularies with hundreds of rarely-used terms, while too little prevents users from finding precisely relevant resources.

Solution:

Establish explicit granularity criteria based on usage thresholds and semantic distinctiveness ³⁶. Implement a rule that terms should be added only when they would apply to at least 10-15 resources in the collection, preventing proliferation of single-use terms. Use usage analytics to guide subdivision decisions—if a term accounts for more than 30% of resources, consider whether meaningful subcategories exist. For example, if "transformer" applies to 40% of models in a repository, subdividing into "encoder-only transformer," "decoder-only transformer," and "encoder-decoder transformer" may be warranted. Conversely, consolidate terms with fewer than five associated resources into broader categories. Implement faceted classification to manage specificity without vocabulary explosion—rather than creating separate terms for every combination of characteristics, use independent facets that users can combine. For instance, instead of terms like "supervised-CNN-image-classification" and "supervised-transformer-image-classification," use separate facets for learning paradigm (supervised/unsupervised), architecture (CNN/transformer), and task (image-classification), allowing users to combine them as needed. Regularly audit vocabulary usage, identifying orphaned terms (no associated resources) and consolidating underused terms into broader categories.

Challenge: User Adoption and Metadata Quality

Even well-designed controlled vocabularies fail if users don't apply them consistently or accurately ⁴⁷. Metadata creators may resist constrained terminology, preferring free-text descriptions that feel more natural. They may lack domain knowledge to select appropriate terms, or find vocabulary interfaces cumbersome. This results in incomplete metadata, incorrect term application, or users bypassing controlled vocabularies entirely. For example, researchers uploading models might skip controlled vocabulary fields, provide only minimal categorization, or misapply terms—tagging a transformer model as "RNN" because both process sequential data, despite fundamental architectural differences.

Solution:

Reduce friction through intelligent interfaces and workflow integration while providing education and incentives ¹⁵. Implement auto-suggestion interfaces that recommend appropriate terms as users type, searching both preferred and non-preferred terms with scope notes to aid selection. Develop automated metadata generation tools that analyze model code, documentation, or performance characteristics to suggest relevant controlled terms—for instance, detecting PyTorch transformer layers and recommending "transformer architecture." Make controlled vocabulary fields required for resource submission while providing "unsure" options that flag resources for expert review rather than blocking submission. Integrate vocabulary application into existing workflows rather than creating separate metadata entry systems—for example, if researchers already document models in README files, provide tools that parse these files and suggest controlled terms based on content. Offer training through short video tutorials demonstrating vocabulary usage for common scenarios, and provide contextual help within metadata entry interfaces. Create incentives for quality metadata—highlight well-documented resources in search results, feature exemplary metadata in newsletters, or implement gamification with badges for comprehensive metadata. Establish a metadata quality review process where information professionals periodically audit resource descriptions, correct errors, and provide feedback to creators, gradually improving community metadata practices.

Challenge: Interoperability Across Heterogeneous Systems

AI resources exist across diverse repositories, each potentially using different controlled vocabularies, metadata schemas, and technical implementations ²⁶. Achieving interoperability—enabling discovery across these heterogeneous systems—requires technical and semantic alignment that individual repositories struggle to accomplish independently. For example, a researcher seeking sentiment analysis models might need to search Hugging Face (using its custom tag taxonomy), Papers with Code (using task categories), institutional repositories (using local vocabularies), and commercial platforms (using proprietary classification schemes), each requiring different query formulations and returning results in incompatible formats.

Solution:

Implement cross-vocabulary mappings using semantic web standards and participate in community vocabulary harmonization efforts ³⁷. Encode vocabularies using SKOS, which provides standard properties for expressing equivalence (skos:exactMatch, skos:closeMatch) and hierarchical relationships (skos:broader, skos:narrower) between terms in different vocabularies. Create explicit mappings between your vocabulary and widely-adopted external vocabularies—for instance, mapping institutional terms to Schema.org properties, Hugging Face tags, and Papers with Code categories. Publish these mappings as machine-readable linked data that federated search systems can consume. For example, an institutional repository might publish SKOS mappings indicating that its "sentiment_analysis" term is an exact match to Hugging Face's "sentiment-analysis" tag and a close match to Schema.org's "SentimentAnalysis" type. Participate in community standardization initiatives like the ML Schema working group or domain-specific ontology development efforts, contributing to shared vocabularies that multiple repositories can adopt. Implement content negotiation in vocabulary services, serving vocabularies in multiple formats (HTML for humans, JSON-LD for machines, RDF/XML for semantic web applications) to maximize interoperability. For organizations building federated search systems, develop vocabulary reconciliation services that translate queries across mapped vocabularies—when a user searches for "sentiment analysis," the system automatically queries each repository using its equivalent term, aggregating results into a unified interface.

Challenge: Balancing Stability with Adaptability

Controlled vocabularies require stability to maintain consistent metadata over time, yet must adapt to accommodate rapidly evolving AI concepts ¹⁴. Excessive change disrupts users and creates metadata inconsistency, while insufficient evolution renders vocabularies obsolete. Organizations struggle to balance these competing requirements—for instance, when a new concept like "retrieval-augmented generation" emerges, should it be added immediately (risking premature standardization of unstable terminology) or delayed until usage stabilizes (risking vocabulary irrelevance)?

Solution:

Implement versioned vocabulary releases with clear deprecation policies and backward compatibility mechanisms ²⁵. Adopt semantic versioning for vocabularies—major versions for breaking changes (term deletions, significant restructuring), minor versions for additions (new terms, new relationships), and patch versions for corrections (typo fixes, scope note clarifications). Release new minor versions quarterly, incorporating accumulated term additions and refinements, while reserving major versions for annual comprehensive reviews. Maintain all vocabulary versions as accessible resources with persistent identifiers, ensuring metadata created with earlier versions remains interpretable. For example, a model tagged with "sequence-to-sequence" using vocabulary version 2.1 (2022) remains valid even after version 3.0 (2024) introduces "encoder-decoder-transformer" as a more specific term. Implement deprecation policies that never delete terms but instead mark them as deprecated with mappings to current equivalents—when users encounter deprecated terms in historical metadata, systems display both the original term and its current equivalent. Create a "provisional terms" section for emerging concepts, clearly labeled as experimental and subject to revision, allowing rapid incorporation of new terminology while signaling instability. Provide vocabulary change logs and migration guides with each release, helping users understand what changed and how to update metadata. Implement automated metadata migration tools that can upgrade historical metadata to current vocabulary versions when appropriate, while preserving original metadata for provenance.

References

1. Publio, G. C., Esteves, D., Lawrynowicz, A., Panov, P., Soldatova, L., Soru, T., Vanschoren, J., & Zafar, H. (2020). ML-Schema: Exposing the Semantics of Machine Learning with Schemas and Ontologies. arXiv:1807.05685. https://arxiv.org/abs/1807.05685
2. Brickley, D., Burgess, M., & Noy, N. (2019). Google Dataset Search: Building a search engine for datasets in an open Web ecosystem. The World Wide Web Conference. https://research.google/pubs/pub46826/
3. Vanschoren, J., van Rijn, J. N., Bischl, B., & Torgo, L. (2020). OpenML: Networked Science in Machine Learning. ACM SIGKDD Explorations Newsletter. arXiv:1909.03011. https://arxiv.org/abs/1909.03011
4. Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumé III, H., & Crawford, K. (2020). Datasheets for Datasets. Communications of the ACM. arXiv:1803.09010. https://arxiv.org/abs/2010.14337
5. Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2020). Model Cards for Model Reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency. https://aclanthology.org/2020.acl-main.447/
6. Chen, J., Lécué, F., Pan, J. Z., & Chen, H. (2020). Learning from Ontology Streams with Semantic Concept Drift. IEEE International Conference on Data Engineering. https://ieeexplore.ieee.org/document/9006350
7. Martínez-Romero, M., O'Connor, M. J., Shankar, R. D., Graybeal, J., Pazos, A., Musen, M. A., & Cheung, K. H. (2021). Using association rule mining and ontologies to generate metadata recommendations from multiple biomedical databases. Database. https://www.sciencedirect.com/science/article/pii/S0306457321001527