A new concordance maps patents to market-facing trademarks, giving researchers a practical way to trace how patented technologies diffuse into products and markets; the validated mapping can be used to study technology adoption at firm, regional and national levels.

Main Finding

The authors produce and validate a content-based, probabilistic concordance (PAT2TM) that links patent technology classes (4‑digit CPC; ~662 classes) to fine‑grained trademark market subclasses (616 Nice subclasses derived from EUIPO HDB terms). The concordance enables direct many‑to‑many mapping from patented technologies to the markets where they are commercialized, allowing researchers and policymakers to trace technology diffusion from invention to market deployment.

Key Points

• Two public data sources: PATSTAT Global (DOCDB patent families, 1996–2020; ~21.6M documents) and EUIPO trademark filings (1996–2020; ~1.65M filings).
• The standard 45 Nice classes were expanded into 616 Nice subclasses by clustering EUIPO HDB terms (85k terms → 62,944 unique HDB codes) using Doc2Vec (semantic vectors) within each Nice class and K‑means (elbow method).
• Patent texts (title + abstract) and HDB terms were preprocessed (lowercasing, stopword removal, stemming, removal of sentences with negatives, etc.).
• Matching strategy: extract bigrams from HDB subclass terms and search patent titles/abstracts to create weighted co‑occurrence links (result: ~400k CPC–NiceSubclass relations before probabilistic weighting).
• Convert co‑occurrence counts into probabilistic links using a null expectation model, producing many‑to‑many CPC → NiceSub probabilities (PAT2TM).
• Validation: content‑based links are compared against firm‑level patent–trademark co‑occurrences and prior indirect concordances (patent ↔ industry ↔ trademark) to assess plausibility (authors report validation steps and consistency with firm portfolios).
• Outputs shared: the PAT2TM concordance and the Nice subclass taxonomy (HBS2Nicesub).

Data & Methods

• Data
  • Patents: EPO PATSTAT Global (DOCDB families), patents filed 1996–2020; CPC classification at 4‑digit level (662 classes).
  • Trademarks: EUIPO filings 1996–2020; Nice classification plus harmonized database (HDB) of descriptors (85,233 terms; 62,944 HDB codes).
• Preprocessing
  • Text normalization: lowercase, remove stopwords/punctuation/numbers, remove sentences with explicit negation, stemming.
• Creating Nice subclasses
  • For each Nice class, embed HDB terms using Doc2Vec to capture semantic similarity in context.
  • Cluster embeddings with K‑means; choose cluster count by elbow method; produce 616 Nice subclasses, each labeled with Nice class, subclass index, and share of terms.
• Linking patents to Nice subclasses
  • Extract bigrams from HDB terms for each subclass.
  • Search patent titles and abstracts for these bigrams (title+abstract chosen for standard practice and to reduce noise).
  • Build a weighted bipartite graph of CPC 4‑digit classes ↔ Nice subclasses based on matched counts (roughly 400k initial relations).
  • Compute probabilistic link weights by comparing observed co‑occurrence to a null expectation (accounts for baseline frequencies), yielding conditional probabilities/intensities rather than binary matches.
• Validation
  • Compare content‑based CPC ↔ NiceSubclass links with firm‑level patent–trademark portfolios (companies that file both) and with existing indirect patent↔industry↔trademark concordances.
  • Reported diagnostics to confirm meaningful overlap and to surface where content links differ from firm‑based links (useful for ecosystem analyses).

Implications for AI Economics

• Mapping AI inventions to markets
  • Researchers can identify which market subclasses are most likely downstream destinations of AI‑related patent classes (e.g., G06F, G06N, subfields in ML/AI), enabling quantitative tracking of AI commercialization.
• Measuring diffusion and commercialization rates
  • Combine patent counts in AI CPC classes with PAT2TM probabilities to estimate expected trademark activity in particular markets — useful for measuring how patenting translates into market introductions over time and across geographies.
• Firm strategy and commercialization gaps
  • At the firm level, compare a firm’s patent portfolio (AI technologies) to its trademark footprint (markets). Probabilistic links highlight markets where a firm’s technologies are under‑exploited (commercialization opportunities) or where other firms are active market converters.
• Regional and industrial policy targeting
  • Use concordance to detect regional complementarities (tech specialization vs market specialization). For AI policy, this can identify local markets likely to absorb AI inventions (e.g., healthcare services, automotive systems, fintech), and inform place‑based innovation policy or ecosystem building.
• Labor, skills, and demand forecasting
  • Link AI technology classes to downstream markets to forecast sectoral demand for skills, services, and complementary inputs as AI diffuses into consumer and business markets.
• Competition, market structure, and welfare
  • Trace how patented AI capabilities map to consumer services and products (trademark markets) to study market concentration, barriers to entry, and the role of IP complementarities across ecosystems (specialized technology firms vs commercializer firms).
• Advantages over indirect concordances
  • Direct, content‑based, many‑to‑many probabilistic links avoid relying on coarse industry codes; better suited for rapidly evolving domains like AI where industry boundaries blur.
• Caveats and limitations for AI economics users
  • Geographic/office coverage: PAT2TM is based on PATSTAT global patents and EUIPO trademark filings — trademark side is EU‑centric; results may underrepresent markets outside EU trademark filings.
  • Time window: data through 2020 (paper accessed in 2021); recent post‑2020 AI commercialization trends (e.g., LLM deployment, new platforms) may not be captured.
  • Text matching limits: matching relies on bigrams within titles/abstracts and HDB terms; some AI inventions with opaque abstracts or different phrasing could be missed, and false positives may occur.
  • Granularity choices: 4‑digit CPC and clustering parameters (Doc2Vec/K‑means) influence results — users should inspect subclass labels and probabilistic weights for their focal AI topics.
  • Validation: concordance is probabilistic and should be used as a signal (not deterministic proof) of tech→market linkages; where possible, triangulate with firm‑level portfolios, product datasets, or web/app data.
• Practical suggestions
  • For empirical analyses, weight trademark counts by PAT2TM probabilities to estimate expected market exposure of specific AI technology classes.
  • Use the Nice subclass taxonomy to refine market categories when constructing dependent variables (e.g., market entry events, product launches).
  • Combine PAT2TM with firm identifiers (patent assignees, trademark owners) to study firm roles across technology development vs market commercialization in AI ecosystems.

If helpful, I can: - Extract likely AI‑related CPC classes and list their top linked Nice subclasses from PAT2TM (example mapping). - Provide a short code recipe (pseudocode) for applying PAT2TM to compute expected trademark markets for a set of CPC classes or for a firm portfolio.