PaperlayerView exact revision on arXiv

Source-linked AI summary

BLANC: Discovering Patent White Space via Changes in Normalized Pointwise Mutual Information Between Multi-View Clusters

Shuichi Miyazawa, Kensuke Fujii

arXiv:2608.26685v1cs.IRcs.CLcs.DL

TL;DR

Patent white-space discovery is hindered by laborious review, manual mapping, and methods that lack combined multi-view and quantitative gap detection. BLANC addresses this with multi-view clustering, NPMI association measurement, and keyword-conditional ΔNPMI ranking. In removal-driven tests, it recovered depleted combinations while size-matched controls essentially did not, and it surfaced an expert-confirmed industrial candidate.

  • Problem

    Existing patent-landscape analysis relies on laborious review or mapping and lacks a method combining multi-view clustering with normalized association and keyword-conditional gap detection.

  • Method

    BLANC independently clusters application/use, novelty, and inventive-step views, computes cross-dimensional NPMI, and ranks pairs whose conditional NPMI drops after keyword filtering.

  • Results

    34.1% (G06N) and 27.3% (C03C) of three-quarters-depleted target pairs were recovered, while different-combination depletion reached 0 of 191 top-20 recoveries and controls essentially did not recover targets.

  • Takeaways & Limitations

    BLANC presents combinations established in the general patent landscape but underexplored for a specified keyword as candidates for new patent applications and supports expert discussion.

  • Takeaways & Limitations

    NPMI estimates become unstable for small filtered corpora, and broader industrial validation beyond one expert-confirmed candidate is needed.

Abstract

from arXiv · show

Identifying white space --- the unexplored but potentially valuable regions of a patent landscape --- is essential for strategic R&D planning, yet existing methods rely on manual patent mapping or apply single-view clustering without quantitative gap detection. We propose BLANC (Blank Landscape Analysis through NPMI Conditioning), a three-phase pipeline combining (1) multi-view neural topic modeling along three semantic dimensions (application/use, novelty, inventive step); (2) Normalized Pointwise Mutual Information (NPMI) to quantify cross-dimensional cluster association; and (3) conditional detection that flags combinations whose NPMI drops when the corpus is filtered by a user-specified keyword. The drop is captured by a new metric, $Δ$NPMI, which identifies combinations "established globally, unexplored locally." Because white space has no ground truth, we evaluate BLANC on two public USPTO corpora --- machine learning/AI (5,417 patents, CPC G06N) and glass compositions (1,982 patents, CPC C03C) --- by artificially depleting known technology combinations and testing recovery. When three-quarters of a target pair's documents are removed, BLANC recovers 34.1% (ML/AI) and 27.3% (glass) of the depleted combinations, whereas size-matched removals not aimed at them (random documents, or those of a different established combination) essentially never do: the target is never recovered in 191 decoy trials. Collapsing the three semantic views into one recovers nothing, while prior co-occurrence measures also flag the target under random removal, offering no specificity. In a proprietary case (302 float glass / glass-ceramics patents), the keyword "fluorine" reveals a fluorine surface treatment $\times$ warpage suppression candidate ($Δ$NPMI up to 0.48) that experts had independently identified.

1. Introduction

BLANC addresses the challenge of finding patent-landscape combinations that are established broadly but absent or underrepresented for a user’s keyword. It combines multi-view clustering, normalized cross-dimensional co-occurrence, and keyword-conditional gap detection to identify candidate white space.

  • Motivation: Patent review and conventional mapping are laborious, expertise-dependent, and often produce few actionable insights.Standard AI classification can also omit an organization’s own technical knowledge, yielding generic perspectives.
  • Prior work: Existing white-space methods use morphological maps, inter-cluster or SAO structures, neural topic models, or co-occurrence measures, but not their combined capabilities.The paper identifies a gap involving multi-view clustering, normalized association measurement, and keyword-conditional detection.
  • BLANC pipeline: BLANC independently clusters patents by application/use, novelty, and inventive step using BERTopic-based neural topic modeling.The clustering pipeline embeds documents with Sentence-BERT, reduces dimensionality with UMAP, and groups them with HDBSCAN.
  • BLANC pipeline: NPMI quantifies associations between clusters from different dimensions, while conditional recomputation after keyword filtering uses ΔNPMI to flag weakened combinations as white space.The target concept is “established globally, unexplored locally.”
  • Evaluation: BLANC recovers injected white space under removal-driven evaluation, whereas size-matched random or different-combination removals essentially do not.The evaluation spans public ML/AI and glass-composition USPTO corpora and compares BLANC with component and external-method baselines.

2. Related Work

Prior patent white-space methods span map-based vacancy detection, cluster- and function-based gap inference, and neural topic modeling, but lack a unified framework combining multi-view clustering, normalized association, and keyword-conditional detection.

  • Research gap: BLANC addresses a gap left by methods that use traditional clustering without normalized association measures or single-view clustering without statistical gap detection.The paper identifies these as limitations of its closest precursors.
  • Prior approaches: Prior methods flag empty or sparse keyword/classification regions, infer gaps from cluster connectivity or SAO structures, and more recently apply transformer-based topic models.These approaches represent three successive methodological streams in patent white-space analysis.
  • BLANC framework: BLANC treats application/use, novelty, and inventive step text as independent semantic views corresponding to established patent-document distinctions.The views are analyzed independently and can be supplied from structured fields or generated from full text when fields are absent.
  • BLANC framework: The pipeline independently embeds and clusters each view, constructs cross-view co-occurrence tensors and matrices, and computes NPMI for cluster-pair association.Its clustering sequence uses Sentence-BERT, UMAP, HDBSCAN, and topic keywords before co-occurrence quantification.
  • Conditional detection: ΔNPMI identifies cluster pairs whose association is strong globally but drops in documents containing a user-specified keyword.Candidates require positive ΔNPMI, at least one conditional joint assignment, and global NPMI at least the default threshold of 0.3.
  • Computational profile: For typical corpora of thousands of documents and tens of clusters per dimension, BLANC’s post-embedding steps complete in seconds, while embedding is the dominant cost.The post-embedding operations scale as O(N1N2N3|D|), O(MN) per 2D NPMI matrix, and O(|D|) for keyword filtering.

4. Experiments

BLANC is evaluated across two public USPTO technology domains using clustering, white-space case studies, removal-based recovery, and baseline comparisons. Results show that multi-view clustering and NPMI are important for specific detection, while sensitivity estimates vary with the depletion threshold.

  • Dataset and clustering: The evaluation covers G06N machine learning/AI and C03C glass-composition corpora, with 5,417 and 1,982 applications respectively.The corpora span distinct technology domains and use application, novelty, and inventive-step dimensions.
  • Dataset and clustering: G06N produced 62, 47, and 51 clusters across application, novelty, and inventive step, while C03C produced 63, 63, and 48.Noise rates were 36.8%, 33.2%, and 28.3% for G06N and 21.3%, 21.3%, and 17.3% for C03C.
  • Global associations: 0.978 was the strongest G06N application×novelty association, linking quantum-computing clusters and indicating an established technology combination.Other highly associated pairs included decision trees at 0.958, finite automata at 0.948, time-series forecasting at 0.939, and social networking at 0.930.
  • White-space case studies: 0.338 was the ΔNPMI for the top “neural” candidate, where neuromorphic hardware paired with spiking-neural-network claims.Global NPMI fell from 0.417 to conditional NPMI 0.079; four of five candidates shared the spiking-NN novelty cluster.
  • Removal-based validation: 34.1% and 27.3% of targeted pairs were recovered at 75% removal in G06N and C03C, while the target was recovered in 0 of 191 decoy trials.Random-removal controls were 0% in every cell but one, whereas the depleted decoy entered the top-20 in 25.6–40.5% of G06N and 30.0–54.5% of C03C trials.
  • Baseline comparison: 0% recovery resulted from collapsing the three semantic views into one BERTopic clustering run on concatenated text.The single view could not express cross-dimensional combinations, leaving no pairs to deplete.
  • Baseline comparison: NPMI was the only tested association measure with 0% random-removal recovery on both corpora, unlike Jaccard, cosine, and raw connectivity.The reported spurious-recovery rates were 15.9%/27.3% for Jaccard, 22.7%/13.6% for cosine, and 29.5%/54.5% for connectivity.

5. Discussion

BLANC structures patent brainstorming as a reproducible, keyword-conditioned search for specific technology gaps, while retaining expert judgment for interpreting candidates. Its cross-domain evaluation supports consistent behavior, but statistical stability, preprocessing, dimension choice, and industrial validation remain bounded.

  • Practical implications: BLANC turns brainstorming preparation into a focused, data-driven workflow that can articulate gaps as specific cross-dimensional combinations rather than blank map regions.Once the corpus is prepared, the workflow runs in minutes compared with weeks for traditional patent-map construction.
  • Practical implications: Keyword selection remains a critical user skill because domain-specific terms substantially outperform generic terms.The paper recommends starting from Phase-1 cluster keywords, preferring multiword technical nomenclature, and consulting CPC subclass definitions.
  • Practical implications: BLANC supports rather than replaces experts, who must determine whether an underexplored combination reflects opportunity, infeasibility, weak demand, avoidance, or terminology differences.The proprietary float-glass case matched a candidate independently identified by experts, illustrating judgment focused by the tool.
  • Broader interpretation: 34.1% versus 27.3% targeted recovery at δ=0.75 across ML/AI and glass supports similar qualitative behavior across technologically distinct domains.The comparison used no tuning beyond an HDBSCAN size adjustment and a 0% random-removal control in both domains.
  • Future work: Systematic evaluation of LLM-generated explanations for white-space candidates remains future work.Preliminary commercial-LLM experiments produced plausible explanations from cluster-keyword and co-occurrence context.
  • Limitations: NPMI estimates become unstable in small filtered corpora, especially for pairs with co-occurrence counts ≤12, while the primary evaluation does not probe the smallest-filtered-corpus regime.The three-view decomposition also lacks formal sensitivity analysis, and a proprietary novelty dimension required manually specified k-means with k=10.

6. Conclusion

BLANC combines multi-view clustering, NPMI co-occurrence analysis, and conditional NPMI drops to identify technology combinations established globally but underexplored for a user keyword. Removal-driven tests across two USPTO domains show targeted recovery and specificity, while field-specific preprocessing and further extensions remain important.

  • Conclusion: BLANC integrates multi-view BERTopic clustering, NPMI-based co-occurrence analysis, and conditional gap detection through the ΔNPMI metric.It presents combinations established in the general patent landscape but underexplored for a specified keyword as candidates for new applications.
  • Conclusion: 34.1% and 27.3% of target pairs were recovered at three-quarter removal in G06N and C03C, while different-pair depletion never recovered the target in 0 of 191 trials.The evaluation used 5,417 ML/AI patents and 1,982 glass-composition patents under a keyword-independent removal protocol.
  • Conclusion: Field-specific text preprocessing is essential for effective multi-view clustering, and future work includes temporal, multi-keyword, bootstrap-confidence, and LLM-explanation analyses.Without domain-aware preprocessing, the dimensions collapse into uninformative clusters.

Data availability

Quantitative experiments use the public Harvard USPTO Patent Dataset, while proprietary float-glass data come from a licensed commercial patent database and cannot be redistributed.

  • Data availability: The quantitative experiments use the publicly available Harvard USPTO Patent Dataset (HUPD).The dataset is accessible through its Hugging Face repository.
  • Data availability: The proprietary float-glass patent records and underlying search query cannot be redistributed because they were retrieved under institutional license.Corpus composition is described in Section 4.5.

Funding

The research reports no specific grant funding from public, commercial, or not-for-profit funding agencies.

  • Funding: This research did not receive any specific grant from public, commercial, or not-for-profit funding agencies.

Declaration of competing interest

The authors declare no known competing financial interests or personal relationships that could have influenced the reported work; part of the methodology is the subject of a Japanese patent application.

  • No competing interests or potentially influential personal relationships were declared, while part of the methodology is covered by Japanese patent application JP2025148721A.The application was filed by Miyazawa, Fujii, Fujii, and Tomiyori and assigned to AGC Inc.

CRediT authorship contribution statement

Miyazawa led conceptualization, methodology, software, validation, visualization, and the original draft, while Fujii contributed to conceptualization, validation, and review. The authors also disclosed using Claude Code for language editing and manuscript consistency checks, followed by human review.

  • Miyazawa contributed to conceptualization, methodology, software, validation, visualization, and original-draft writing; Fujii contributed to conceptualization, validation, and review.
  • Claude Code assisted with language editing, paragraph restructuring, and consistency checks, after which the authors reviewed and edited the manuscript.

A. Public data experiment parameters

The public experiments use HUPD patent corpora and three text views clustered with BERTopic-based tooling, followed by diversified keyword scoring and NPMI-filtered theme selection. The evaluation reports strong recovery on two domains, while validating the third dimension after field-specific preprocessing.

  • Implementation: BERTopic, Sentence-Transformers, UMAP, and HDBSCAN implement clustering with fixed random seeds, making reported values deterministic for a given corpus and configuration.Exact NPMI values remain sensitive to library versions, whereas rank-based recovery results are robust.
  • Data and views: The experiments use HUPD corpora containing 5,417 G06N patent applications filed during 2013–2016 and 1,982 C03C applications filed during 2012–2017.Application/use uses abstracts, novelty uses claims, and inventive step uses summaries.
  • Preprocessing: Preprocessing removes claims boilerplate and summary headings, figure sentences, and formulaic language, increasing novelty clusters from 2 to 47 and inventive-step clusters from 8 to 51.
  • Keyword scoring: Cluster keywords combine within-cluster frequency, term–cluster NPMI, and cluster-IDF, then apply greedy stemming-based diversification to suppress redundant or boilerplate terms.A term unique to one cluster receives cIDF = log N_c, while terms appearing in every cluster receive cIDF = 0.
  • Clustering and selection: The G06N clustering produces 62 application, 47 novelty, and 51 inventive-step clusters; 58 application×novelty pairs meet NPMI ≥0.3 and co-occurrence ≥8, with 46 meeting the stricter ≥12 floor.
  • Evaluation: 93.8% Hit Rate and 0.469 MRR are achieved on G06N, while C03C reaches 81.2% Hit Rate and 0.594 MRR; third-dimension validation reaches 87.5% Hit Rate on C03C.Theme selection ranks application×novelty pairs by global NPMI, filters co-occurrence, and stops at 16 nonredundant keyword-linked pairs.

B. Theme-keyword sensitivity diagnostic

The diagnostic removes documents from established cluster pairs, recomputes NPMI, and tests whether a related keyword places the depleted target among the top 20. Results show that keyword specificity and partial depletion are central to detection.

  • Protocol: A target is successfully detected when a related keyword places the depleted cluster pair in the top-20 after recomputing NPMI on the reduced corpus.The protocol repeats across 16 diverse themes in each HUPD corpus and reports Hit Rate and MRR.
  • Keyword specificity: 56.2% Hit Rate and 0.287 MRR result from automatically generated glass keywords, versus 81.2% and 0.594 with domain-specific keywords.Generic terms such as “glass” and “layer” match many pairs and dilute the conditional signal.
  • Removal fraction: ΔNPMI rises monotonically from 0.726 to 0.786 as removal increases from 0.10 to 0.90 for the quantum-computing pair, which remains rank 1 until complete removal.At δ = 1.00, the target is excluded because conditional co-occurrence becomes zero.
  • Results tables: Table 7 reports synthetic injection outcomes for G06N at δ = 0.50 using application×novelty targets.
  • Controls: Under no removal, hand-selected targets appear in the top-20 for 100% of G06N and 87.5% of C03C themes, and random removal reproduces those rates exactly.Removing target documents lowers the rates only slightly, to 93.8% for G06N and 81.2% for C03C.
  • Results tables: Table 8 reports C03C injection outcomes at δ = 0.50 for application×novelty pairs with domain-specific keywords.
  • Results tables: Table 10 records assigned keywords, keyword-selected document counts |D_q|, and target-pair shares within the filtered documents.

C. Keyword assignment in the removal-driven protocol

The keyword-assignment rule is deterministic but not injective, so a few generic terms can cover all targets. Target pairs occupy uniformly small shares of selected corpora, with “computer” concentration motivating an independence caveat.

  • The Step 2 keyword-assignment rule is deterministic but not injective, allowing a few generic terms to cover all targets.
  • Target pairs occupy at most 7.0% of their assigned keyword corpora, making them undetectable before removal.
  • Concentration on “computer” forms the basis of the independence caveat in §5.4.
Loading 2608.26685v1…