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

Shuichi Miyazawa; Kensuke Fujii

cs.IRAug 27, 2026

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

Authors: Shuichi Miyazawa, Kensuke Fujii

Organizations: Digital & Innovation Management Division, AGC Inc., Tokyo, Japan · Intellectual Property Division, AGC Inc., Tokyo, Japan

Abstract

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.

Figures & tables

Figure 1: System architecture and data flow of the BLANC pipeline. Phase 1 independently clusters patents along three semantic dimensions using BERTopic. Phase 2 constructs a 3D co-occurrence tensor and computes pairwise NPMI matrices. Phase 3 filters by a user keyword, recomputes conditional NPMI, and ranks cluster pairs by $\Delta\text{NPMI}$ .

Study	Clustering	Co-occ. measure	Multi-view	Cond. detection
Kim et al. (2014)	Traditional	Connectivity $\dagger$	—	—
Choi & Jun (2014)	Bayesian	—	—	—
Jeon et al. (2023)	BERTopic	—	—	—
Yang & Chen (2025)	BERTopic	—	—	—
This work	BERTopic	NPMI	✓	✓
† Unnormalized inter-cluster connectivity, not a normalized co-occurrence association measure.

Table 1: Comparison with closest prior work. ✓ = present; — = absent.

Figure 2: Cluster size distribution for the G06N corpus after field-specific preprocessing. Each dimension resolves into many clusters (62/47/51); the application dimension is finely graded, while novelty and inventive step each retain one larger leading cluster but still spread across dozens of themes rather than collapsing into a single dominant cluster as they do without preprocessing, enabling meaningful cross-dimensional NPMI analysis. Note: Y-axis scales differ across panels to accommodate the different cluster size ranges.

	G06N	C03C
	(ML/AI)	(Glass)
Documents	5,417	1,982
Filing years	2013–16	2012–17
Avg. words (abstract)	114	100
Avg. words (claims)	1,216	764
Avg. words (summary)	684	729

Table 2: Dataset statistics for the two HUPD corpora.

Figure 3: NPMI heatmap for application $\times$ novelty cluster pairs (G06N corpus, top-20 application clusters by co-occurrence; vertical-axis numbers are the original cluster IDs and are therefore non-consecutive, since only these top-20 clusters are shown). Dark-red cells mark cluster pairs whose application and novelty themes co-occur far more than chance (coherent technology combinations), whereas most cells are weak (blue), indicating that the majority of cross-dimensional combinations are only loosely associated.

Application theme	Novelty theme	NPMI	Co.
Quantum comp.	Quantum proc.	0.978	106
Decision tree	Decision tree	0.958	33
Finite automaton	DFA/NFA	0.948	30
Time series	Forecast model	0.939	44
Social network	Social graph	0.930	61
Energy cons.	Energy anomaly	0.926	60

Table 3: Top-10 app

\times

novelty pairs by global NPMI. Co.: co-occurrence count.

Figure 4: Top-5 white space candidates for keyword “neural” (G06N, application $\times$ novelty; same candidates as Table 4 ). Blue bars show global NPMI; salmon bars show conditional NPMI after keyword filtering. The gap ( $\Delta\text{NPMI}$ , annotated) represents the degree of underrepresentation among “neural”-keyword patents.

Application	Novelty	$\mathrm{NPMI}$	$\mathrm{NPMI}_{q}$	$\Delta$	$n_{q}$
Neuromorphic	Spiking NN	0.417	0.079	0.338	19
Ensemble learn.	Classif.	0.480	0.168	0.312	1
Spiking NN	Spiking NN	0.395	0.089	0.306	37
NN augment.	Spiking NN	0.412	0.115	0.298	68
Neurosynaptic	Spiking NN	0.375	0.090	0.285	28

Table 4: Top-5 white space candidates for keyword “neural” (app

\times

novelty).

\mathrm{NPMI}

: global association;

\mathrm{NPMI}_{q}

: conditional association on the keyword-filtered subcorpus;

\Delta=\mathrm{NPMI}-\mathrm{NPMI}_{q}

;

n_{q}

: conditional co-occurrence count.

			Random		Decoy
Corpus	$\delta$	Targeted	corpus	$D_{q}$	target	decoy
G06N ( $n=44$ )	0.50	13.6%	0%	0%	0%	25.6%
	0.75	34.1%	0%	0%	0%	40.5%
	1.00	0%	0%	2.3%	0%	35.7%
C03C ( $n=22$ )	0.50	9.1%	0%	0%	0%	36.4%
	0.75	27.3%	0%	0%	0%	54.5%

Table 5: Removal-driven injection: percentage of established, initially undetected pairs newly entering the top-20 after targeted removal versus three size-matched controls. “Decoy” depletes a different established pair in the same

D_{q}

; its two columns report whether the target is spuriously detected (specificity) and whether the depleted decoy is (sensitivity). Decoy denominators are smaller where no size-matched decoy exists. Because the decoy arm removes as many documents as the target arm, a decoy larger than the target is only partly depleted and can still be recovered, which is why the decoy column stays non-zero at

\delta=1.00

while the target column is 0%. Depletion is complete for 22 of 42 G06N decoys and 6 of 20 on C03C at that setting, and none of those fully depleted decoys is recovered, reproducing the total-absence regime on the decoy side.

	G06N (ML/AI)		C03C (Glass)
Method	Targeted	Random	Targeted	Random
Multi-view + NPMI (BLANC)	34.1%	0.0%	27.3%	0.0%
Multi-view + Jaccard	29.5%	15.9%	31.8%	27.3%
Multi-view + Cosine	52.3%	22.7%	45.5%	13.6%
Multi-view + connectivity [ 4 ]	29.5%	29.5%	22.7%	54.5%
Single-view + NPMI [ 8 ]	0.0%	0.0%	0.0%	0.0%

Table 6: Baseline comparison on the removal-driven protocol (

\delta=0.75

): percentage of initially undetected target pairs newly entering the top-20 under targeted versus size-matched random removal, the latter drawn from the whole corpus (the corpus-random control of § 4.6.2 ). Only NPMI exhibits zero spurious recovery under random removal; a useful detector needs Targeted

\gg

Random.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Table 11

$\delta$	Removed	Rank	$\Delta$ NPMI
0.10	10	1	0.726
0.25	27	1	0.728
0.50	54	1	0.733
0.75	81	1	0.747
0.90	98	1	0.786
1.00	109	MISS	—

Appendix

Table 9:

\delta

sweep for the quantum-computing pair (app

\times

nov). Removal counts are taken over all 109 documents assigned to the pair in the application and novelty dimensions. The co-occurrence count of 106 in Table 3 is smaller because three of these documents are unassigned (noise) in the inventive step dimension and are thus excluded from the tensor (§ 3.2 ).

				Share of $D_{q}$
Corpus	Keyword	Pairs	$\|D_{q}\|$	median	max
G06N	computer	37	4,118	0.56%	2.21%
	data	4	3,789	0.50%	0.69%
	information	1	2,417	0.54%	0.54%
	memory	1	2,307	1.34%	1.34%
	state	1	984	7.01%	7.01%

Appendix

Table 10: Keywords assigned by Step 2 of § 4.6.2 . Pairs: targets receiving the keyword.

|D_{q}|

: documents selected by it. Share:

|\text{pair}\cap D_{q}|/|D_{q}|

over those targets.

Legal IR

Legal information retrieval finds and ranks relevant statutes, cases, patents, and other legal materials for queries and evidence-grounded systems.

See all 53 papers

Benchmarking Patent Embeddings: A Multi-Task Evaluation of 22 Models Across Retrieval, Classification, and Clustering
May 22, 2026Amirhossein Yousefiramandi, Ciaran CooneyLegal IR Clustering
Clarivate, Intellectual Property Barcelona, Spain 08025
Benchmarking Embedding Geometries and Clustering Algorithms for Unsupervised Patent Claim Classification
Oct 9, 2026Sujith Subramanian, P N Bhargav Teja, Susen Sarkar +1Information Retrieval Hyperbolic Representation Learning
Samsung R&D Institute Bangalore, India
Citation-Driven Multi-View Training for Patent Embeddings: QaECTER and Sophia-Bench
Apr 24, 2026Younes Djemmal, You Zuo, Kim Gerdes +1Legal IR Text Embedding Models
ALMAnaCH · 1AI Lab, Questel, Paris, France · 3Inria, Paris, France +3

May 22, 2026J/K move · Enter open · S save

Benchmarking Patent Embeddings: A Multi-Task Evaluation of 22 Models Across Retrieval, Classification, and Clustering

Amirhossein Yousefiramandi, Ciaran Cooney

Clarivate, Intellectual Property Barcelona, Spain 08025

Legal IR Clustering Text Embedding Evaluation Text Classification Information Retrieval IR Evaluation

Open paper AskPDF arXiv

Two questions regarding practitioners' use of patent embeddings arise: (i) Does one fine-tuning recipe suffice for all downstream applications? (ii) Is fine-tuning on one patent landscape sufficient for downstream application on other landscapes? By evaluating 22 pre-trained embedding models (ranging from 22M to 12B parameters) on three tasks -- information retrieval, classification, and clustering -- on 113,148 WIPO patents for assistive technology (46,069 citation queries) and on an external DAPFAM dataset, we find that two results cast doubt on the prevailing wisdom. (i) The optimal fine-tuning recipe depends on the downstream task: cross-sectional alignment (recipe R3) provides the largest improvements to retrieval performance (+7.1% nDCG@10), whereas a combined signal recipe (recipe R4) is better suited to classification (+7.1 F1) and clustering (+10.9 V-measure); a matched data control confirms that differences in training dataset size are not a contributing factor. (ii) Single-landscape fine-tuning hampers cross-landscape information retrieval: fine-tuning on one landscape significantly degrades cross-domain retrieval for 5 of 8 model-recipe combinations on the DAPFAM corpus, with the stronger zero-shot models suffering most. While within-family scaling is consistent (Qwen3 0.6B->4B->8B; Llama-Nemotron 1B->8B), cross-family scaling is erratic; the 12B KaLM-Gemma3 is ranked 8th on TAC retrieval performance, following prefix modification. Title+Abstract+Claims is the ubiquitous best text view, and all models suffer from a 55-65% gap between IN and OUT-of-domain performance which cannot be mitigated by hybrid BM25-dense fusion. Code and evaluation framework are publicly available.

Abstract

Legal IR

Explore similar work

Benchmarking Patent Embeddings: A Multi-Task Evaluation of 22 Models Across Retrieval, Classification, and Clustering

Benchmarking Embedding Geometries and Clustering Algorithms for Unsupervised Patent Claim Classification

Citation-Driven Multi-View Training for Patent Embeddings: QaECTER and Sophia-Bench