Multi-Dimensional Classification

Multi-dimensional classification in AI discoverability architecture represents a sophisticated approach to organizing, categorizing, and retrieving artificial intelligence systems, models, and resources across multiple simultaneous taxonomic dimensions ¹. Unlike traditional single-axis classification systems, this methodology enables AI artifacts to be tagged and discovered through various independent yet complementary facets such as task type, domain application, architectural approach, performance characteristics, and deployment requirements ². The primary purpose is to enhance the findability, accessibility, and appropriate matching of AI solutions to specific use cases in increasingly complex AI ecosystems ³. This approach matters critically as the proliferation of AI models, frameworks, and applications has created an urgent need for robust discovery mechanisms that can navigate the multifaceted nature of modern AI systems while supporting diverse stakeholder needs from researchers to practitioners.

Overview

Multi-dimensional classification in AI discoverability architecture emerged from the convergence of information science principles and the explosive growth of AI systems in the late 2010s ¹². As AI model repositories expanded from hundreds to millions of artifacts, traditional keyword-based search and simple hierarchical taxonomies proved inadequate for helping users locate appropriate solutions. The Model Cards framework, introduced by Google researchers in 2018, represented a pivotal moment in recognizing that AI systems require documentation across multiple independent dimensions including intended use, performance characteristics, ethical considerations, and limitations ¹³.

The fundamental challenge this approach addresses is the inherent complexity and multifaceted nature of AI systems ². A single AI model might simultaneously be characterized by its architecture (transformer-based), task (text generation), domain (healthcare), computational requirements (GPU-intensive), ethical properties (fairness-audited), and licensing terms (open-source). Users approach discovery from different perspectives depending on their context—a researcher might prioritize architectural novelty, while a practitioner focuses on deployment constraints and compliance requirements ⁴. Single-dimensional classification forces artificial choices about which aspect to prioritize, while multi-dimensional approaches accommodate these varied access points simultaneously.

The practice has evolved significantly from early model cards to comprehensive metadata schemas and standardized ontologies ²⁴. Initial implementations focused primarily on technical dimensions like architecture and performance metrics. However, growing awareness of AI ethics and governance has driven expansion into dimensions covering fairness, transparency, privacy preservation, and regulatory compliance ¹. Modern systems increasingly employ hybrid approaches combining automated metadata extraction from code repositories and documentation with expert curation for nuanced dimensions like ethical considerations ⁵⁶.

Key Concepts

Faceted Classification

Faceted classification refers to organizing information using multiple independent dimensions or "facets" that represent distinct aspects of the classified items ². Each facet contains a controlled vocabulary of mutually exclusive values, and items can be assigned one or more values from each facet independently. In AI discoverability, common facets include task type, domain application, architectural paradigm, data requirements, and computational resources.

Example: The Hugging Face Model Hub classifies a BERT model for sentiment analysis across multiple facets simultaneously: Task facet (text classification), Architecture facet (transformer-encoder), Domain facet (natural language processing), Training approach facet (supervised learning), Language facet (multilingual), and License facet (Apache 2.0). A user searching for "transformer models for text classification in healthcare that are privacy-preserving" can filter across these independent facets to narrow results precisely to their requirements.

Metadata Enrichment

Metadata enrichment is the process of systematically adding classificatory information to AI artifacts across defined dimensions ¹³. This involves capturing structured data about models, datasets, and systems that enables multi-dimensional discovery. Enrichment can occur through automated extraction, semi-automated approaches using natural language processing, or manual expert curation.

Example: When a researcher publishes a new computer vision model on GitHub, an automated pipeline extracts metadata from the repository's README file, configuration files, and code structure to populate initial classifications: detecting the framework (PyTorch), architecture type (convolutional neural network), and task (object detection). The system then prompts the researcher to complete additional dimensions through a guided form, including performance metrics on standard benchmarks, computational requirements (GPU memory needed), and ethical considerations (whether the model was tested for demographic bias). This enriched metadata enables the model to be discovered through multiple search pathways.

Orthogonal Dimensions

Orthogonal dimensions are classification facets that are independent of one another, meaning the value assigned in one dimension does not constrain or determine values in other dimensions ². This independence is crucial for enabling flexible, multi-criteria search without creating logical contradictions or unnecessary complexity.

Example: In an enterprise AI catalog, the "Deployment Environment" dimension (values: cloud, edge, hybrid) is orthogonal to the "Compliance Status" dimension (values: GDPR-compliant, HIPAA-compliant, SOC2-certified, pending-review). A model can be classified as "edge-deployed" and "HIPAA-compliant" independently—the deployment location doesn't determine compliance status, and vice versa. This allows users to search for "edge-compatible models that are GDPR-compliant" without the classification system imposing artificial constraints between these dimensions.

Controlled Vocabularies

Controlled vocabularies are standardized, predefined sets of terms used for each classification dimension to ensure consistency and enable effective retrieval ⁴. Unlike free-text tagging, controlled vocabularies limit classification values to approved terms with clear definitions, reducing ambiguity and improving interoperability across systems.

Example: For the "Model Architecture" dimension, a controlled vocabulary might include precisely defined terms: "transformer-encoder" (BERT-style models), "transformer-decoder" (GPT-style models), "transformer-encoder-decoder" (T5-style models), "convolutional-neural-network," "recurrent-neural-network," and "hybrid-architecture." When a user classifies a new model, they must select from these predefined terms rather than entering free text. This prevents inconsistencies like having some models tagged "CNN," others "ConvNet," and others "convolutional neural network" for the same architectural concept, ensuring all convolutional models can be reliably discovered regardless of terminology variations.

Semantic Interoperability

Semantic interoperability refers to the ability of different classification systems to understand and exchange classification information meaningfully, even when using different vocabularies or schemas ⁴. This is achieved through ontologies that define relationships between terms and mappings between different classification schemes.

Example: The Papers with Code platform uses one vocabulary for classifying machine learning tasks ("image-classification," "object-detection," "semantic-segmentation"), while TensorFlow Hub uses slightly different terms ("image-classifier," "object-detector," "image-segmentation"). A semantic interoperability layer maintains mappings between these vocabularies, recognizing that "image-classification" and "image-classifier" refer to the same concept. When a federated search queries both repositories for "image classification models," the system automatically translates the query to match each repository's vocabulary, returning relevant results from both sources despite their different classification schemes.

Cross-Dimensional Navigation

Cross-dimensional navigation enables users to explore AI artifacts by moving between different classification perspectives, refining searches by progressively adding constraints across multiple dimensions ⁶. This supports exploratory discovery where users may not initially know all their requirements but refine them iteratively.

Example: A healthcare data scientist begins by browsing models in the "Healthcare" domain facet, viewing 500 results. They then filter by the "Task" dimension to "medical-image-analysis," reducing results to 150 models. Noticing many require extensive computational resources, they add a constraint in the "Computational Requirements" dimension for "edge-compatible" models, narrowing to 30 results. Finally, they filter by "Explainability" dimension for "interpretable" models, yielding 8 highly relevant candidates. The interface shows how many models match each dimension value, guiding the user's navigation through the multi-dimensional space.

Metadata Schema

A metadata schema defines the structure, format, and requirements for capturing classification information about AI artifacts ¹³. It specifies which dimensions are mandatory versus optional, data types for each field, validation rules, and relationships between metadata elements.

Example: An organization's AI model registry implements a metadata schema based on the Model Cards framework with extensions for internal requirements. The schema mandates completion of: model_name (string), task_type (controlled vocabulary), architecture (controlled vocabulary), intended_use (free text, minimum 100 characters), performance_metrics (structured object with metric names and values), computational_requirements (structured object with memory, GPU, and latency specifications), training_data (reference to dataset registry entry), ethical_review_status (controlled vocabulary: completed, in-progress, not-required), and business_owner (reference to employee directory). Optional fields include fairness_metrics, privacy_preservation_techniques, and regulatory_compliance_certifications. This schema is enforced through validation at submission time, ensuring consistent, comprehensive metadata across all registered models.

Applications in AI Discovery Contexts

Enterprise AI Model Governance

Organizations with multiple AI development teams implement multi-dimensional classification to maintain centralized catalogs of internal AI assets ⁴. Classification dimensions typically include business unit ownership, data sensitivity levels, compliance status, deployment environment, and approval stage. This enables governance teams to audit which models handle sensitive data, track compliance with regulatory requirements, and prevent redundant development by surfacing existing solutions.

A financial services company maintains an AI model registry where each model is classified across dimensions including: regulatory framework (GDPR, CCPA, financial regulations), data classification (public, internal, confidential, restricted), model risk tier (high, medium, low based on impact), deployment status (development, staging, production, retired), and business function (fraud detection, credit scoring, customer service, trading). When a new project requires a credit scoring model, developers can search for "production-deployed, medium-risk, credit-scoring models that are CCPA-compliant and handle restricted data," immediately identifying three existing models that meet requirements rather than building from scratch.

Research Literature Discovery

Academic platforms employ multi-dimensional classification to help researchers navigate the rapidly expanding AI literature ⁵⁶. Dimensions include research methodology, datasets used, benchmark performance, reproducibility status, and code availability. This supports systematic literature reviews and meta-analyses by enabling researchers to identify papers sharing specific methodological characteristics.

The Papers with Code platform classifies research papers across dimensions including: task taxonomy (over 3,000 specific tasks organized hierarchically), datasets (linking papers to the datasets they use), methods (specific techniques employed), results (benchmark performance metrics), and code availability (official implementation, community implementation, no code). A researcher investigating few-shot learning approaches for medical image classification can filter to papers that: use few-shot learning methods, apply to medical imaging tasks, report results on specific medical imaging datasets, and have available code implementations. This multi-dimensional filtering reduces 50,000 computer vision papers to 23 highly relevant papers with reproducible implementations.

Cloud AI Service Marketplaces

Cloud providers use multi-dimensional classification to help customers discover appropriate pre-trained models and AI services from extensive catalogs ³. Dimensions include pricing model, supported regions, integration options, performance tiers, and certification status. This enables customers to find solutions matching both technical requirements and business constraints.

Google Cloud's AI Hub classifies pre-trained models across dimensions including: ML framework (TensorFlow, PyTorch, scikit-learn), task category (vision, language, structured data), deployment target (Cloud AI Platform, on-premises, edge devices), pricing tier (free, pay-per-use, enterprise), support level (community, standard, premium), and industry vertical (retail, healthcare, manufacturing, finance). A retail company seeking a product recommendation model can filter for: task="recommendation-system," industry="retail," deployment="Cloud AI Platform," pricing="pay-per-use," and support="standard," discovering five models with detailed comparisons of their performance characteristics, pricing structures, and integration requirements.

AI Ethics and Fairness Tools

Specialized platforms focus on ethical dimensions of AI systems, classifying models across fairness metrics, bias testing results, transparency features, and privacy preservation techniques ¹. This supports organizations in identifying AI systems that meet ethical guidelines and regulatory requirements for responsible AI deployment.

An AI ethics platform maintains a registry where models are classified across dimensions including: fairness_testing_status (tested, not-tested), bias_mitigation_techniques (reweighting, adversarial-debiasing, fairness-constraints, none), explainability_methods (SHAP, LIME, attention-visualization, inherently-interpretable, black-box), privacy_preservation (differential-privacy, federated-learning, encrypted-computation, none), demographic_groups_evaluated (age, gender, race, disability-status, socioeconomic-status), and audit_status (third-party-audited, internal-review, not-audited). A healthcare provider required to deploy only audited, explainable models can search for "third-party-audited models with SHAP explainability that use differential-privacy and have been tested for age and disability-status bias," ensuring compliance with their responsible AI framework.

Best Practices

Design Orthogonal, User-Centered Dimensions

Classification dimensions should be independent of one another and aligned with how users actually think about and search for AI systems ². Conduct user research to understand discovery patterns and mental models before defining dimensions. Avoid creating dimensions that are derivative of others or that users don't find meaningful for their decision-making.

Rationale: Orthogonal dimensions prevent logical contradictions and enable flexible multi-criteria search. User-centered design ensures the classification system supports actual workflows rather than imposing artificial organizational structures that don't match user needs.

Implementation Example: Before implementing an AI model catalog, a research institution conducts interviews with 30 researchers, data scientists, and engineers about how they currently discover and evaluate models. Analysis reveals users think in terms of: "What task does it perform?" (task dimension), "What data does it need?" (data requirements dimension), "Can I run it with my resources?" (computational requirements dimension), "Is it trustworthy?" (validation status dimension), and "Can I legally use it?" (licensing dimension). The team designs the classification schema around these five orthogonal dimensions rather than imposing a hierarchical taxonomy based on technical architecture, which users found less intuitive for discovery.

Implement Hybrid Automated and Manual Classification

Combine automated metadata extraction with expert human curation to balance scalability with accuracy ⁵⁶. Use automation for objective, easily extractable dimensions while reserving human judgment for nuanced aspects requiring domain expertise or ethical assessment.

Rationale: Purely manual classification doesn't scale to large repositories, while purely automated approaches lack the nuance and accuracy needed for complex dimensions. Hybrid approaches leverage the strengths of both methods.

Implementation Example: An AI model repository implements a three-tier classification pipeline. Tier 1 (fully automated) extracts technical metadata from code repositories: framework detection from import statements, architecture type from model class names, input/output specifications from function signatures, and computational requirements from profiling runs. Tier 2 (semi-automated with human validation) uses NLP to extract performance metrics, intended use cases, and limitations from README files and papers, presenting suggestions to model authors for confirmation. Tier 3 (expert manual curation) requires domain experts to assess and classify ethical considerations, fairness testing results, and appropriate use contexts—dimensions where automated extraction would be unreliable and errors could have serious consequences.

Version and Evolve Classification Schemes

Treat classification schemas as living artifacts that evolve with the AI landscape ⁴. Implement versioning for controlled vocabularies, provide migration paths when terms change, and establish governance processes for adding new dimensions or values based on community feedback and emerging needs.

Rationale: AI technology evolves rapidly, with new architectures, tasks, and considerations emerging continuously. Static classification schemes quickly become outdated and fail to accommodate innovations. Versioning enables evolution while maintaining backward compatibility.

Implementation Example: A model hub implements semantic versioning for its classification schema (currently v2.3.1). When transformer architectures emerged, the team added new values to the "Architecture" dimension vocabulary in v2.0.0, a major version indicating new capabilities. When users requested finer-grained task classifications, the team subdivided "text-generation" into "text-generation-conditional" and "text-generation-unconditional" in v2.1.0, a minor version. The system maintains mappings so models classified under the old "text-generation" term automatically appear in searches for either new term, ensuring backward compatibility. A governance committee meets quarterly to review community proposals for schema changes, evaluating each against criteria of user value, orthogonality with existing dimensions, and maintainability.

Provide Progressive Disclosure in User Interfaces

Design discovery interfaces that present essential dimensions prominently while making advanced dimensions available without overwhelming users ⁶. Support both simple search for novice users and complex multi-dimensional queries for power users.

Rationale: Exposing all classification dimensions simultaneously creates cognitive overload and poor user experience. Progressive disclosure accommodates users with varying expertise levels and specificity of requirements.

Implementation Example: An AI marketplace implements a three-level interface. Level 1 (simple search) presents a single search box with autocomplete and three prominent filters: task type, domain, and license. This serves 70% of users who have straightforward requirements. Level 2 (guided discovery) offers a wizard-style interface asking users questions about their use case, then progressively revealing relevant dimensions: "Will this run in the cloud or on edge devices?" reveals computational requirement filters; "Do you need to explain model decisions?" reveals explainability filters. Level 3 (advanced search) provides access to all 15 classification dimensions with boolean operators and saved search capabilities for power users conducting systematic evaluations. Analytics show appropriate distribution across levels, with users self-selecting the complexity they need.

Implementation Considerations

Tool and Format Selection

Choose metadata formats and storage technologies that balance machine-readability, human-readability, and interoperability ¹³. Common formats include JSON-LD for web-based systems, RDF for semantic web integration, and specialized formats like Model Cards. Storage technologies should support efficient multi-dimensional querying, with options including document databases (MongoDB), graph databases (Neo4j), and search engines (Elasticsearch).

Example: A research consortium building a federated AI discovery system selects JSON-LD as the metadata format because it provides both human-readable JSON structure and semantic web compatibility through linked data principles. They implement storage using Elasticsearch for fast faceted search across millions of models, with a graph database layer (Neo4j) for capturing complex relationships between models, datasets, and papers. This hybrid approach enables sub-second faceted search queries while supporting sophisticated relationship traversal like "find all models trained on datasets derived from ImageNet that cite papers by authors from institution X."

Audience-Specific Customization

Tailor classification dimensions and vocabularies to specific user communities while maintaining core interoperability ²⁴. Different stakeholder groups may require different dimension emphasis—researchers prioritize methodological details, practitioners focus on deployment characteristics, and governance teams need compliance dimensions.

Example: A healthcare AI platform maintains a single underlying classification schema but presents customized views for different audiences. Clinical researchers see dimensions emphasizing medical specialty, clinical task, validation study design, and regulatory approval status. IT operations teams see dimensions highlighting deployment architecture, integration requirements, computational resources, and security certifications. Compliance officers see dimensions covering regulatory frameworks, audit status, data governance, and ethical review completion. All views draw from the same underlying metadata, but the interface prioritizes dimensions relevant to each audience's decision-making needs.

Organizational Maturity and Context

Align classification complexity with organizational maturity in AI adoption ⁴. Organizations early in AI adoption benefit from simpler classification schemes focused on basic discovery, while mature AI organizations require sophisticated multi-dimensional systems supporting governance, compliance, and optimization use cases.

Example: A company beginning AI adoption implements a minimal classification schema with five dimensions: business function, development status, data sensitivity, owner, and last updated. This provides basic discovery and governance without overwhelming teams still learning AI development practices. After two years and 50 deployed models, they expand to 12 dimensions adding: model risk tier, regulatory requirements, fairness testing status, computational costs, and integration dependencies. The phased approach prevents initial complexity from hindering adoption while scaling classification sophistication as organizational needs mature.

Integration with Existing Systems

Design classification systems to integrate with existing development workflows, model registries, and governance tools rather than creating isolated catalogs ³. Leverage APIs, webhooks, and automation to capture classification metadata as a natural byproduct of development processes.

Example: An organization integrates their AI classification system with existing tools: GitHub repositories automatically trigger metadata extraction when models are committed; CI/CD pipelines populate performance metrics and computational requirements from automated testing; the model serving platform updates deployment status and usage statistics; and the data governance system provides data lineage and sensitivity classifications. This integration means developers don't manually maintain a separate model catalog—classification metadata is automatically aggregated from systems they already use, reducing maintenance burden while ensuring accuracy and currency.

Common Challenges and Solutions

Challenge: Classification Consistency Across Contributors

When multiple people classify AI artifacts, inconsistencies emerge from different interpretations of dimension definitions, varying levels of expertise, and subjective judgment on ambiguous cases ². A model might be classified as "high-performance" by one contributor and "medium-performance" by another based on different reference points. These inconsistencies degrade discovery effectiveness and user trust.

Solution:

Implement comprehensive classification guidelines with concrete examples and decision trees for ambiguous cases ⁴. Provide training for contributors and establish validation workflows where experienced curators review classifications before publication. Use automated consistency checking to flag potential errors, such as logical contradictions (a model classified as "unsupervised" but also requiring "labeled training data"). For subjective dimensions, consider using multiple independent classifiers and flagging items with low inter-rater agreement for expert review.

Example: A model repository creates a 50-page classification guide with specific examples for each dimension value. For the "Computational Requirements" dimension, they define "edge-compatible" as "runs inference in under 100ms on a Raspberry Pi 4 with less than 1GB memory," "mobile-compatible" as "runs on iOS/Android devices with under 500ms latency," and "cloud-required" as "requires GPU or more than 8GB RAM." New contributors complete a training module where they practice classifying 10 example models and receive feedback on discrepancies. The system automatically flags models classified as both "edge-compatible" and "requires-GPU" for curator review, catching logical inconsistencies.

Challenge: Vocabulary Evolution and Terminology Drift

AI terminology evolves rapidly, with new architectures, techniques, and concepts emerging continuously ⁵. Classification vocabularies quickly become outdated, missing important new categories while retaining obsolete ones. Additionally, terminology drift occurs when the community's understanding of existing terms shifts over time.

Solution:

Establish a governance process for vocabulary evolution with clear criteria for adding, modifying, or deprecating terms ⁴. Implement versioning for controlled vocabularies with migration paths that map old terms to new ones. Monitor research literature and community discussions to identify emerging terminology requiring incorporation. When terms are deprecated, maintain mappings so historical classifications remain discoverable and provide automated suggestions to update classifications to current vocabulary.

Example: When vision transformers emerged as a significant new architecture, the classification governance committee evaluated adding a new term to the "Architecture" dimension. They established that the term had appeared in over 100 papers, represented a meaningfully distinct approach from existing "transformer" and "convolutional-neural-network" categories, and was being actively searched for by users. They added "vision-transformer" as a new vocabulary term in schema v3.0.0, created documentation distinguishing it from related architectures, and implemented a suggestion system that prompted owners of models previously classified as "transformer" in computer vision tasks to consider reclassifying as "vision-transformer." Over six months, 200 models were reclassified, improving discovery precision for this architecture type.

Challenge: Cold Start and Incomplete Metadata

New AI artifacts often lack comprehensive classification metadata, creating a cold start problem where they're poorly discoverable until fully classified ⁶. Contributors may skip optional dimensions, provide minimal information, or lack expertise to accurately classify certain aspects. This results in incomplete metadata that limits discovery effectiveness.

Solution:

Design submission workflows that make metadata creation as frictionless as possible while encouraging completeness ¹³. Implement smart defaults based on automated extraction and similarity to existing artifacts. Use progressive metadata enrichment where basic classification enables initial discovery, with prompts and incentives for contributors to add more detail over time. Consider gamification or reputation systems that reward comprehensive classification.

Example: When a researcher uploads a new model, the system automatically extracts basic metadata from the code repository (framework, architecture type, task) and suggests initial classifications based on similarity to existing models with similar code patterns. The submission workflow presents a two-stage form: Stage 1 (required) captures essential dimensions for basic discovery—task, domain, license, and intended use. This minimal metadata makes the model immediately discoverable. Stage 2 (encouraged but optional) requests additional dimensions—performance metrics, computational requirements, ethical considerations, and limitations. The system emails contributors one week after submission with a personalized prompt: "Your model has been viewed 47 times. Adding performance metrics and computational requirements would help 23 users who filtered for these dimensions find your work." Contributors who complete comprehensive metadata receive "Detailed Documentation" badges on their profiles, creating social incentives for completeness.

Challenge: Scalability of Multi-Dimensional Indexing

As the number of AI artifacts and classification dimensions grows, query performance can degrade significantly ⁶. A repository with millions of models across 20 dimensions creates a massive search space where complex multi-dimensional queries become computationally expensive, leading to slow response times that frustrate users.

Solution:

Implement optimized indexing strategies specifically designed for multi-dimensional queries ⁶. Use inverted indices for each dimension, enabling efficient filtering. Consider approximate methods like locality-sensitive hashing for very large repositories where exact search becomes prohibitive. Implement query result caching for common dimension combinations. Provide query cost estimation to help users understand performance implications of complex queries and suggest optimizations.

Example: A large AI model hub with 5 million models implements a tiered indexing strategy. Tier 1 uses Elasticsearch with carefully tuned inverted indices for the 10 most commonly queried dimensions, enabling sub-second response for 90% of queries. Tier 2 employs a columnar database (Apache Druid) for analytical queries across all 20 dimensions, optimized for aggregations and facet counting. Tier 3 uses vector embeddings to represent models in a high-dimensional space, enabling approximate similarity search for exploratory discovery when users don't have precise requirements. The system monitors query patterns and automatically promotes frequently accessed dimension combinations to cached materialized views, reducing repeated computation. When users construct very complex queries, the interface shows estimated result counts and response times, suggesting simplifications if performance would be poor.

Challenge: Cross-Repository Interoperability

Different AI repositories use varying classification schemes, vocabularies, and metadata formats, making federated discovery across repositories difficult ⁴. A user searching for "sentiment analysis models" might need to query multiple repositories that use different terms ("sentiment-classification," "opinion-mining," "sentiment-detection") and different metadata structures.

Solution:

Develop and adopt shared ontologies and metadata standards that enable semantic interoperability across repositories ¹⁴. Implement mapping layers that translate between different vocabularies, recognizing synonymous terms and equivalent concepts. Participate in community standardization efforts and contribute to open metadata schemas. For federated search, use schema.org markup or JSON-LD with shared vocabularies to enable cross-repository queries.

Example: A consortium of AI research institutions develops a shared ontology for NLP tasks, defining canonical terms and relationships. Each institution's repository maintains its own internal vocabulary but implements a translation layer to the shared ontology. When a user searches the federated discovery portal for "sentiment-analysis" models, the query is translated to each repository's vocabulary: Repository A's "sentiment-classification," Repository B's "opinion-mining," and Repository C's "sentiment-detection." Results are aggregated and presented with normalized metadata using the shared ontology terms. The system maintains bidirectional mappings, so users can browse using any repository's terminology and discover relevant models across all repositories. This approach preserves each repository's autonomy while enabling effective cross-repository discovery.

References

1. Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2018). Model Cards for Model Reporting. https://arxiv.org/abs/1810.03993
2. Paleyes, A., Urma, R. G., & Lawrence, N. D. (2021). Towards a Comprehensive Understanding of AI Model Documentation. https://arxiv.org/abs/2108.07258
3. Google Research. (2019). Model Cards for Model Reporting. https://research.google/pubs/pub48120/
4. IEEE. (2020). IEEE Standards for AI Model Documentation. https://ieeexplore.ieee.org/document/9321694
5. Elsken, T., Metzen, J. H., & Hutter, F. (2020). A Survey on Neural Architecture Search. https://arxiv.org/abs/2011.03395
6. Balog, K., Radlinski, F., & Arakelyan, S. (2021). Information Retrieval and AI Systems. https://www.sciencedirect.com/science/article/pii/S0306457321001035
7. Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., & Amodei, D. (2020). Language Models are Few-Shot Learners. https://arxiv.org/abs/2005.14165
8. NeurIPS. (2021). ML Reproducibility and Discoverability. https://neurips.cc/Conferences/2021/Schedule?showEvent=21895
9. Dodge, J., Gururangan, S., Card, D., Schwartz, R., & Smith, N. A. (2020). Metadata and Classification in NLP Systems. https://aclanthology.org/2020.acl-main.442/