The rapid expansion of artificial intelligence has created a fragmented landscape of information. Academic researchers, software engineers, and industry practitioners each document their progress through distinct channels—such as scientific preprints, open-source repositories, and newsletters—that vary significantly in technical vocabulary and depth. Traditional scientometric tools have struggled to synthesize these disparate archives, often relying on flat classification systems that fail to capture the multi-level evolution of the field.

A new methodology, published in EPJ Data Science, addresses these limitations by mapping documents from diverse sources into shared hierarchical taxonomies. This framework allows researchers to conduct rigorous, multi-resolution comparative analyses without the constraints of single-source evaluation or ad-hoc labeling.

Operational Architecture and Hierarchical Mapping

The system functions by organizing information into directed acyclic graphs. In these structures, higher-level nodes represent broad disciplinary categories, while child nodes capture increasingly specialized technical subfields. When a document is assigned to a specific node, the framework automatically propagates that assignment to all ancestor nodes along the branch. This upward aggregation ensures statistical stability at broad conceptual levels while maintaining the granularity required to inspect specialized developments.

To avoid compressing complex technological properties into a single structure, the framework supports multiple parallel taxonomies. This allows analysts to classify the same document along orthogonal dimensions, such as model architecture, learning paradigm, problem formulation, and application domain.

Quantifying Thematic Patterns

The researchers developed five core metrics to measure dynamics within these taxonomy graphs, three of which provide critical insights into community behavior:

  • Topic Popularity: Measures the proportion of a stream’s total documents assigned to a specific node or its descendants, reflecting the relative prominence of a subject within a community.
  • Average Detail Depth: Quantifies the technical granularity of a source. By calculating the mean relative depth of document labels, the framework reveals whether a source provides high-level overviews or deep, specialized technical documentation.
  • Topical Diversity: Assesses how evenly document coverage is distributed across subtopics. By using a uniform baseline to cap the influence of dominant branches, this metric allows for fair comparisons of thematic balance across different sources, regardless of their total document volume.

Cross-Source Insights and Model Validation

When applied to real-world data from Hugging Face, arXiv, and Deep Learning Weekly, the framework revealed distinct community characteristics. Hugging Face showed a high concentration in transformer architectures and fine-tuning, while arXiv displayed a broader distribution with a stronger emphasis on computer vision and supervised learning. Deep Learning Weekly demonstrated the highest topical diversity, reflecting its role as a curated source for industry applications.

To support further investigation, the team developed the AI Taxonomy Trend Explorer, an interactive tool that utilizes hierarchical sunburst diagrams and temporal curves. By providing a standardized way to monitor technical domains, this framework offers a scalable solution for tracking the shifting priorities of the global artificial intelligence community.