Analysis of AI patents (2002–2021) reveals four distinct classes of AI inventions and the topical pathways through which scientific research is incorporated into technology; the mapped science-to-technology links offer targeted insights for corporate R&D strategy and innovation policy.
Knowledge flows from science to AI technology: Identifying core and brokerage technological roles
Using CPC network centrality and BERTopic on AI patents (2002–2021) and their cited scientific abstracts, the paper maps four patent classes and traces the topical pathways by which scientific research feeds AI technological development.
The rapid advancement of artificial intelligence (AI) technologies has not only driven convergence with diverse technological domains but also swiftly spread across various industrial sectors. As a knowledge-intensive field, AI is particularly shaped by the flow of knowledge from scientific research to technological development, yet remains insufficiently examined in a systematic and structural way. This study addresses this gap by investigating science-to-technology knowledge flow that underpins AI’s technological evolution. We propose a semantic science-technology exploration framework specifically designed for the AI domain, consisting of the two stages: technology classification and semantic topic exploration. First, AI patents are classified into four categories using centrality measures derived from a CPC co-occurrence network. Then, we extract abstracts from both patents and their cited scientific publications to apply BERTopic modelling and generate topic labels using generative AI. Analyzing AI-related patents filed from 2002 to 2021, we trace key technological trends and elucidate the structural pathways of knowledge flow science to technology. The findings offer practical implications for corporate R&D strategies and innovation policy design in the era of AI.
Summary
Main Finding
The paper develops a two-stage semantic framework for mapping science-to-technology knowledge flows in AI. By classifying AI patents using centrality measures on a CPC co‑occurrence network and applying BERTopic to the abstracts of patents and their cited scientific publications (with generative-AI-assisted labeling), the authors trace technological trends across 2002–2021 and reveal structural pathways through which scientific knowledge feeds AI technological evolution. The approach yields actionable insight for firm R&D prioritization and for innovation policy aimed at steering AI diffusion.
Key Points
- Research gap: AI is knowledge‑intensive and shaped by flows from scientific research to technological development, but systematic, structural studies of those flows are limited.
- Framework: Two-stage pipeline
- Technology classification — AI patents are grouped into four categories based on centrality metrics computed on a CPC (Cooperative Patent Classification) co‑occurrence network.
- Semantic topic exploration — patent abstracts and abstracts of the scientific publications they cite are modeled with BERTopic; topic labels are generated/enriched using generative AI.
- Data scope: AI-related patents filed between 2002 and 2021, together with the scientific publications cited by those patents.
- Outcomes: Identification of key AI technological trends across the two-decade period and characterization of structural science→technology knowledge pathways (how particular scientific topics map into patent clusters and technological categories).
- Practical angle: The results are positioned to inform corporate R&D strategy and the design of public innovation policy in the AI era.
Data & Methods
- Data
- Patent corpus: AI-related patents (2002–2021) — patents are linked to the scientific literature via their citations to scientific publications.
- Scientific corpus: Abstracts of the scientific papers cited by those patents.
- Classification stage
- Build a CPC co‑occurrence network from patent classification codes.
- Compute centrality measures on that network to classify patents into four distinct categories (categories reflect structural/positional differences in the CPC network).
- Semantic stage
- Extract text abstracts from patents and cited papers.
- Use BERTopic (embedding + clustering topic model) to extract topics across the combined corpus.
- Use generative-AI methods to produce human‑readable topic labels and to assist interpretation.
- Analysis
- Temporal tracing of topic prevalence and patent category dynamics (2002–2021).
- Mapping of topic links between scientific publications and patents to reveal pathways of knowledge flow.
Implications for AI Economics
- Knowledge diffusion and productivity
- The mapped science→technology pathways make it possible to quantify how scientific advances diffuse into patented AI technologies, informing estimates of knowledge spillovers and their timing — crucial inputs for economic models of productivity gains from AI.
- Firm R&D strategy
- Firms can use the framework to identify which scientific topics are actively translating into technological applications, prioritize research collaborations or acquisitions, and spot emerging technological niches where scientific inputs rapidly convert into patents.
- Market structure and competition
- Structural classification of patents (via CPC centralities) highlights which areas are core/enabling versus peripheral — helping assess incumbents’ strategic positions, entry barriers, and potential for disruptive entrants.
- Innovation policy
- Policymakers can use the approach to target funding toward scientific domains with high downstream technological impact, design incentives to accelerate desirable knowledge flows, and monitor whether public investments produce translational outcomes in AI.
- Measurement and monitoring
- The method provides a scalable, reproducible way to monitor evolving AI technological ecosystems over time, which supports dynamic policy evaluation and real‑time economic analysis of AI diffusion.
If you want, I can: - Summarize specific topic clusters and example science→technology linkages found in the paper, or - Convert this into a short policy brief or slide-ready bullet list.
Assessment
Paper Typedescriptive
Evidence Strengthn/a — The study is descriptive and maps patterns of science-to-technology knowledge flow; it does not attempt causal identification of economic impacts, so causal evidence strength is not applicable. The results are informative about structure and trends but cannot support causal claims about outcomes like productivity or labor effects.
Methods Rigormedium — The paper applies reasonable and modern methods (CPC co-occurrence networks with centrality measures for patent classification, BERTopic for semantic topic extraction, and use of cited-paper abstracts), which are appropriate for mapping knowledge flows. However, rigor is limited by reliance on patent citations as a proxy for knowledge transfer, potential sensitivity to centrality metric choices and cutoff thresholds, possible instability/opacity in topic-model hyperparameters, and use of generative AI for labels which can introduce labeling noise; robustness checks, validation against alternative measures, and details on parameter selection/reproducibility would be needed to raise the rating.
SampleAI-related patents filed from 2002 to 2021 and the abstracts of scientific publications these patents cite; patents are classified via a CPC co-occurrence network and analyzed at the abstract level using BERTopic, with topic labels generated by a generative AI model. (Paper does not report full-text analysis or coverage details within the provided description.)
Themesinnovation governance adoption
GeneralizabilityPatent data bias: patents underrepresent software/algorithmic innovations that are kept as trade secrets or released as open-source rather than patented., Citation limitation: patent-to-paper citations are an imperfect proxy for knowledge flow and vary by examiner/assignee practices and jurisdiction., Geographic and sectoral coverage unclear: results may be skewed toward jurisdictions and industries with heavier patenting activity (e.g., US, Europe, large firms)., Time window stops in 2021: excludes the latest rapid advances and diffusion post-2021 (e.g., large language model commercial uptake)., Method sensitivity: findings depend on CPC network construction, centrality metric choices, BERTopic hyperparameters, and generative-AI labeling quality., Abstract-only analysis: using abstracts (not full texts) may miss substantive nuance in scientific and patent content.
Claims (9)
| Claim | Direction | Confidence | Outcome | Details |
|---|---|---|---|---|
| This study proposes a semantic science-technology exploration framework specifically designed for the AI domain, consisting of two stages: technology classification and semantic topic exploration. Other | positive | high | existence and design of a two-stage semantic science-technology exploration framework |
0.3
|
| AI patents are classified into four categories using centrality measures derived from a CPC co-occurrence network. Other | positive | high | patent classification into four categories |
0.3
|
| Abstracts from patents and their cited scientific publications were extracted and BERTopic modelling was applied; topic labels were generated using generative AI. Other | positive | high | semantic topics derived from patent and cited-publication abstracts |
0.3
|
| The analysis covers AI-related patents filed from 2002 to 2021. Other | positive | high | temporal coverage of analyzed patents |
0.3
|
| The analysis traces key technological trends in AI across the studied period. Innovation Output | positive | medium | technological trends over time |
0.11
|
| The study elucidates the structural pathways of knowledge flow from science to technology in AI. Innovation Output | positive | medium | structure/pathways of science-to-technology knowledge flow |
0.11
|
| AI is a knowledge-intensive field that is particularly shaped by the flow of knowledge from scientific research to technological development. Innovation Output | positive | high | role of scientific knowledge flow in AI development |
0.09
|
| Science-to-technology knowledge flow in AI has been insufficiently examined in a systematic and structural way. Other | negative | high | extent of systematic/structural study of science-to-technology knowledge flow in AI |
0.18
|
| The findings offer practical implications for corporate R&D strategies and innovation policy design in the era of AI. Governance And Regulation | positive | medium | practical implications for R&D strategy and policy design |
0.05
|