Introduction
The trajectory of Artificial Intelligence (AI) from theoretical breakthroughs in academic laboratories to practical applications in the global market is often obscured by fragmented data. While academic research is well-documented in scholarly databases, industrial innovation is largely confined to patent registries. Bridging these two domains is essential for understanding the lifecycle of innovation. This paper introduces DeepInnovationAI, a comprehensive global dataset that maps the journey of AI by linking over 3.5 million academic papers to 2.3 million industrial patents using large language models (LLMs) and hypergraph analysis.
The Innovation Gap
A persistent challenge for policymakers, economists, and technologists is the "silo" problem: academic research and industrial patents exist in separate ecosystems. Traditional methods of tracking knowledge flow—such as keyword matching—often fail to capture modern AI terminology or the nuanced semantic relationships between a paper's findings and a patent's claims. Furthermore, existing datasets are frequently limited by geographic scope (often focusing solely on the US) or rely on lagging indicators like citation counts, which do not reflect the immediate "combinatorial novelty" of an invention.
DeepInnovationAI addresses these limitations by providing a unified, global, and semantically rich infrastructure designed to track the AI innovation lifecycle across sixty years of data.
Methodology
The construction of DeepInnovationAI involved a sophisticated pipeline to process nearly 100 million raw records into a curated, AI-specific dataset. The methodology is built upon three technical pillars:
1. Hierarchical Classification
To identify AI-related patents efficiently, the researchers employed a "coarse-to-fine" approach. Initial filtering was conducted via International Patent Classification (IPC) codes and keywords. A high-quality subset was then annotated using GPT-4 to serve as training data for a dual-layer BERT architecture. This system utilized BERT_outer for broad filtering and BERT_inner for precise classification, ensuring high accuracy without the prohibitive cost of running LLMs on the entire 100-million-document corpus.
2. Hypergraph-Driven Novelty
Rather than relying on citations—which can take years to accumulate—the team utilized hypergraphs to measure "combinatorial novelty." By representing technical keywords as nodes and documents as hyperedges, they calculated the statistical rarity of specific knowledge combinations. A patent that successfully integrates previously unrelated concepts receives a higher novelty score, allowing for the identification of transformative research in near real-time.
3. Semantic Similarity Mapping
Using KeyBERT and Doc2Vec, the researchers calculated the cosine similarity between papers and patents. To ensure logical consistency, the team enforced "temporal validity": a patent is only matched to a paper if the patent filing occurred after the paper’s publication, establishing a clear direction for the potential flow of knowledge.
Dataset Structure
The DeepInnovationAI dataset is organized into three core components:
- DeepPatentAI: Contains 2,356,204 patent records, including metadata and calculated novelty scores.
- DeepDiveAI: Comprises 3,511,929 academic publications with 13 distinct metadata fields.
- DeepCosineAI: Provides 3,511,929 pairs representing the most semantically relevant patent match for each academic paper.
The patent data was sourced from the Intelligent Innovation Dataset (IIDS), drawing from the European Patent Office’s Espacenet to ensure a truly global perspective.
Key Results and Validation
The framework demonstrates high technical rigor and reliability:
- Classification Accuracy: The
BERT_innerclassifier achieved an overall accuracy of 97.5% on the test set. - Expert Validation: The model was tested against a "gold standard" set of 6,000 samples manually annotated by domain experts, maintaining a robust 90.8% accuracy and a 99.2% recall for AI patents.
- Consensus Reliability: In instances where keyword matching, LLMs, and BERT models all reached a consensus, classification accuracy reached 98%.
- Longitudinal Depth: The dataset spans over 60 years (up to 2020), allowing researchers to track AI from its symbolic roots to the modern era of deep learning.
Discussion and Limitations
While DeepInnovationAI provides a powerful lens into the AI ecosystem, the authors acknowledge several constraints. The current dataset has a temporal cutoff of 2020, meaning the most recent surge in Generative AI is not yet fully captured. Additionally, while high semantic similarity suggests knowledge flow, it does not definitively prove causality. Future work will focus on integrating inventor-author matching to confirm direct technology transfer.
Conclusion
DeepInnovationAI represents a significant step forward in the quantitative study of innovation. By moving beyond simple citations and embracing semantic and hypergraph-based metrics, this dataset provides a high-resolution map of how theoretical "aha moments" in the lab transform into the industrial technologies that shape our world.
Summary updated