Identifying green, digital, and twin-transition patents is essential for tracking innovation and assessing policy impact, yet existing code-based and machine-learning approaches often yield non-overlapping results, undermining comparability and reproducibility. This study introduces a scalable framework that combines configurable keyword and technology rules for candidate identification, a rule-guided seed and antiseed definition, and bidirectional citation expansion. Patent texts are encoded with a domain-specific transformer, and final selection is achieved through topic-guided pruning based on a contrastive cosine rule applied to topic-level representations. Validation against proxy labels on a held-out split indicates high precision under a conservative threshold and balanced performance under a data-driven threshold. The workflow is automated, largely unsupervised, and tractable at the scale of millions of patent families, with results robust to sensible hyperparameter choices and threshold selection, thereby improving transparency and comparability for patent landscaping in the green and digital domains.
Automated Patent Landscaping
Identifying patents at the intersection of environmental and digital technologies (“twin patents”) is
pivotal for tracking innovation dynamics and evaluating policies that foster sustainability and
digitalization. Yet current approaches—ranging from examiner- or expert-selected technological codes and
curated keywords, to citation-based heuristics and machine-learning pipelines—often select markedly
diferent sets of documents [
To address fragmentation while preserving transparency and scalability, this work introduces an automated workflow that minimizes manual intervention and combines rule-guided seed construction, two-level bidirectional citation expansion of the candidates, and topic-guided semantic pruning with patent-specific transformer embeddings.
First, candidate “twin” patents are identified by systematically crossing green selection strategies from the literature—namely the Cooperative Patent Classification (CPC) code Y02 and targeted sustainability Classification (IPC), and digital/AI keywords). This Candidates set is then used to automatically derive three working sets: (i) the seed set, comprising high-precision exemplars of twin patents; a family enters the Seed if it is flagged by more than two independent modules (i.e., at least three; strict voting rule); (ii) the expansion set, obtained by expanding the Candidates via forward and backward family-level citations in two waves to collect plausibly related families; and (iii) the antiseed set, a size-matched control sampled outside both the Seed and the Expansion, designed to include mostly random non-twin patents and a 10% share of hard negatives. These hard-to-classify negatives are sampled from patents tagged as green or digital through Y02 and Y04 CPC codes, but not classified as jointly green-and-digital by https://sites.google.com/view/francescaghinami/ (F. Ghinami)
CEUR Workshop
ISSN1613-0073
any method (see Section 3.2). A patent-tailored encoder (PaECTER, [
This article (i) proposes a reproducible, weakly-unsupervised seed definition procedure that integrates diferent sets of conditions derived from the literature, with a strict voting rule to balance breadth and precision; (ii) develops a citation-network aware expansion and matched random antiseed construction to enable unsupervised, contrastive pruning at scale; (iii) introduces a topic-guided semantic pruning strategy that couples domain-specific patent embeddings (PaECTER) with BERTopic topic distributions and a seed-vs-antiseed cosine decision rule, yielding cleaner, more coherent landscapes; and (iv) delivers a scalable, transparent pipeline with practical diagnostics and an accompanying Python implementation for researchers and policy analysts.
Section 2 reviews related methods and the overlap problem; Section 3 details seed rules, expansion, and topic-guided pruning; Section 4 describes data; Section 5 reports diagnostics; Section 5.5 presents the pseudo-label evaluation and robustness; Section 6 discusses limitations and future directions; Section 7 concludes.
The classification and identification of green and digital technologies have become central to
understanding innovation dynamics in the context of the “twin transition”, which couples sustainability
objectives with digital transformation [
One group of approaches relies primarily on classification codes such as the International Patent
Classification (IPC) or the Cooperative Patent Classification (CPC). These codes are assigned by examiners
and enable systematic, replicable searches [
Keyword-based searches ofer a more flexible alternative, capable of capturing emerging or
crosscutting technologies [
Citation-based techniques leverage references among patents to identify related inventions and trace
knowledge flows [
Recent work has increasingly combined these strategies. Integrated approaches, such as those
underpinning the IPC Green Inventory or ENV-TECH classification systems, attempt to blend the
strengths of codes, keywords, and expert rules to improve recall and precision [
Finally, advances in machine learning have introduced new possibilities for automating patent
classification. Supervised and semi-supervised models trained on expert-labeled seed sets can extend
coverage and reduce manual efort [
No single approach fully resolves coverage, accuracy, and scalability. This fragmentation, together with low overlap across methods, motivates integrated, open, and reproducible pipelines such as the one proposed here.
This study builds on prior semi-supervised patent landscaping methods [
Earlier approaches typically relied on human-curated seeds and antiseeds [
Finally, given the heterogeneity of twin patents combining digital and sustainability-related
technologies, novel unsupervised methods are introduced to test the adherence of the expanded set to the seed.
Specifically, BERTopic modeling [
The seed set forms the foundation of the patent landscape, making its composition critical to ensure
both accuracy and representativeness. To reduce reliance on manual curation and enhance replicability,
a rule-based approach inspired by the strategy proposed by [
In their method, patents are identified as “twin” if they combine digital and green characteristics, based on CPC codes and keyword presence. Specifically, they apply six selection rules that combine Y02 (green technologies) and Y04 CPC codes (digital technologies), with keywords. Additionally, IPC codes are used to capture relevant groups and subclasses. Given the comprehensive nature of this framework1, their 1Details of the keyword and code lists are reported in Table 1. method is adopted and extended to allow broader generalization to other technological domains. Unlike the original formulation, which applied a fixed set of combinations, here an adaptation is suggested to systematically combine any green identification strategy with any digital strategy in all possible pairings. This means that every sustainability-related rule (e.g., Y02 codes or green keywords) was crossed with every digital or AI-related rule (e.g., Y04 codes, digital IPC codes, or digital keywords), generating an expanded and more granular set of inclusion criteria. This ensures a balanced and comprehensive coverage of patents that may reflect diverse configurations of digital and green technologies. A patent was thus included in the initial candidate pool if it matched any of these combined rules. To improve precision and avoid reliance on any single identification method, a stringent inclusion criterion is applied: only patents identified as twin by more than two of the resulting combinations were retained in the seed set. This procedure yielded a final set of 9,847 unique patent families. These patents were considered suficiently diverse and representative of the target technological intersection for subsequent expansion and pruning phases.
The expansion methodology implemented in this study largely follows the approach outlined by Abood and Feltenberger (2018), which involves a two-tiered expansion process. However, the first level of expansion based on the most relevant CPC codes is here excluded, to mitigate the risk of over-relying on this information, as CPC codes are already incorporated as a rule in the definition of seed patents.
The first level of expansion (Level 1) involves identifying patents related to seed patents through backward and forward citations. This level includes all patents that cite the seed patents or are cited by them. Moreover, it is augmented by all the patents identified by any of the candidate selection method, but not included in the seed. The second level of expansion (Level 2) further extends this network by including patents that are related to the patents identified in Level 1 through their own backward and forward citations.
The antiseed set, serving as negative examples for later semantic comparison, was generated by randomly sampling an equal number of patent families not included in either the seed or expansion sets.
In patent classification and pruning, models can be trained from scratch or adapted from domain-specific
encoders. Training from scratch ofers flexibility but typically requires substantial compute. A practical
alternative is a pre-trained model tailored to patents. This study uses the PaECTER encoder [
Documents in the Seed, Antiseed, and Expansion sets—formed by concatenating title and abstract2—are
2All numerals are removed and text is lowercased.
encoded jointly with PaECTER to obtain a shared embedding space that enables comparable topic
modeling and pruning. They are then organized with BERTopic [
The BERTopic mapping serves two roles: (i) it enables unsupervised diagnostics via topic-level metrics (Section 3.5); and (ii) it provides the representation on which the topic-guided pruning operates (Section 3.4 below).
Prior work [
To select relevant documents from a large, noisy expansion set in an unsupervised manner, we
apply a topic-guided pruning strategy inspired by weak supervision and semi-supervised learning
[
() }. Each document is represented by its topic-probability vector p ; let p̃ = p /‖p ‖2 The pruning score is a best-seed vs. best-antiseed cosine margin: Δ() = max ⟨p̃ , ⏟≤⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟ () ⟩ − max ⟨p̃ , ⏟ℓ⏟≤⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟ () ⟩ ℓ
.
closest seed prototype closest antiseed prototype A document is retained when Δ() ≥ . The primary operating point on the unlabeled corpus is the zero-contrast rule (=0 ), which keeps a document when it is at least as similar to seed prototypes as to antiseed prototypes. This choice is parameter-light, interpretable, and does not require labels.
We evaluate pruning within the BERTopic framework using established topic-model diagnostics. First,
an intertopic distance map projects topic representations into two dimensions to visualize distinctiveness
and dispersion, assessing whether pruning improves semantic coherence and separation [
This approach combines the high recall of unsupervised expansion with the precision gains of
topic-guided selection, and builds on guided topic discovery [
In the absence of costly and hard-to-replicate human annotation, we validate our topic-guided pruning against a proxy gold standard derived from CPC Y02/Y04 codes. These codes are curated jointly by two patent authorities, the European Patent Ofice (EPO) and the United States Patent and Trademark Ofice (USPTO), and assigned by trained examiners, providing a practical proxy for human labels.
The evaluation pool comprises 29,032 patent families tagged with both Y02 and Y04 (treated as positives) and 9,847 antiseed families (treated as negatives, by construction a mix of random non-twins and hard negatives). We report standard retrieval metrics: precision (the fraction of selected patents that are true twins), recall (the fraction of true twins that are selected), and F1 (the harmonic mean of precision and recall, summarizing the precision–recall trade-of).
The evaluation pool also allows us to fine-tune and test diferent thresholds for pruning.
For quantitative assessment, a pseudolabeled evaluation pool is built as Y02∩Y04 CPC families (positives) together with antiseeds (negatives). A single 75/25 stratified train–test split is drawn from the pseudo-labeled evaluation pool. Thresholds are selected by maximizing MCC on the train split; all metrics are reported on the held-out test split.
On the training split, an MCC-optimal operating point is selected by
MΔCC ∈ arg max MCC( , {Δ ≥ } ), and all supervised-style metrics (precision, recall, F1, MCC, accuracy, Jaccard and the confusion matrix) are reported for both =0 and MΔCC.
As a robustness check, a sparse lexical baseline
is included that operates at the individual-document level in TF–IDF space [
⏟⏟∈⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟ closest seed doc closest antiseed doc retain if () ≥ . serving as robustness checks.
On the same training split, MCC
TF–IDF is obtained by maximizing MCC, and test metrics are reported at =0 and MTFC–CIDF. Because TF–IDF compares documents directly at the lexical level, it commonly achieves slightly higher supervised metrics on the evaluation split; however, Δ remains the preferred primary scorer for the full unlabeled corpus, as it enables unsupervised diagnostics, topic-level interpretability, and a principled =0 operating point independent of labels, with both MΔCC and the TF–IDF baseline
The empirical analysis is conducted using the Patstat Autumn 2024 database [
An example of the data collected is shown in Table 2.
As an initial analysis, the identification strategies proposed by [
Applying the most stringent criterion —requiring identification by more than two methods to be included in the seed— results in a final set of 9,847 unique seed patent families, which can be considered as the most representative and thus suitable for a robust seed definition. By expanding the candidates set twice through bidirectional citations, an expansion set comprising 1,918,509 unique patent families is derived.
After preprocessing (lowercasing and removing numerals), the seed, antiseed, and expansion documents are encoded and modeled with BERTopic. The model yields 86 topics for the overall set (Figure 2). The intertopic distance visualization (Figure 2, panel (a)) provides an overview of the topics generated by the BERTopic model for the overall set of seed, expansion and antiseed documents. Each point on the visualization represents a topic, with its position determined by the model’s learned embeddings. Topics that are close to one another on the map suggest similar themes, while those that are farther apart indicate distinct conceptual areas. This visualization revealed a well-structured topic space, with several clusters of topics grouped around core themes relevant to the twin patent corpus.
Overall, the BERTopic model’s initial output provides a coherent representation of the twin patent data’s thematic landscape. The visualizations demonstrates a structured topic space with clear thematic areas and revealed areas for further refinement in the subsequent pruning phase, where the expansion set would be filtered based on cosine similarity to seed topics.
By applying the pruning strategy described in Section 3.3 to the expansion set, 575,441 patent families are retained.
To assess the impact of the pruning strategy in an unsupervised setting, id est in absence of labeling
that hinders the use of recall and precision analysis, a number of tests can be performed. First, the
topic analysis on the whole and pruned expansion set can be repeated, in order to compare the
Intertopic Distance Map [
Finally, comparing the panels (b) and (c) in Figure 2, a greater average distance is found in the expansion set compared to the pruned one, and that broader and disjointed topics were efectively dropped.
Subsequently, to evaluate the semantic alignment of the expansion set with seed and antiseed topics,
the Kernel Density Estimation (KDE) plots of the maximum cosine similarity between each expansion
document and (i) seed topics and (ii) antiseed topics are produced. These plots provide a smooth
estimation of similarity density across the corpus [
These diagnostics confirm that pruning was both selective and semantically discriminative, reducing noise and enhancing alignment with the seed topics.
In the first two rows of Table 4 we report validation of the baseline method—the multi-prototype topic contrast Δ —at the conservative threshold =0 and at a data-driven threshold MΔCC selected by maximizing the Matthews correlation coeficient on a separate development split of the labeled pool. At the conservative operating point (=0 ), Δ attains Precision=0.973, F1=0.928, and Jaccard=0.866. Using the data-driven threshold, Δ improves to F1=0.956 with higher recall (Recall=0.981) and slightly lower precision (Precision=0.932). This operating point typically increases recall (and overall F1) while allowing a controlled rise in false positives. Although the tuned threshold is favored in terms of precision, accuracy, and overlap, the conservative rule performs well and remains attractive when labels are unavailable. Practitioners can choose the operating point that matches their objective: the label-free =0 rule when high precision and simplicity are paramount, or the MCC-selected threshold when broader coverage (higher recall) is preferred, and labels are available.
a) Precision: Δ vs. TF-IDF b) Recall: Δ vs. TF-IDF c) F1: Δ vs. TF-IDF
The same table also reports test performance for a TF–IDF max-margin variant evaluated at =0 and at its own TF–IDF–specific MCC-selected threshold. Δ operates in topic space to provide interpretable, prototype-aligned pruning, whereas TF–IDF ofers a document-space robustness baseline. At =0 , Δ achieves high precision with slightly higher recall than TF–IDF (F1 = 0.928 vs. 0.935). With a data-driven threshold, Δ improves to F1 = 0.956 with a strong recall gain, while TF–IDF reaches a higher balanced score overall (F1 = 0.978).
Despite TF–IDF’s gain on this proxy-labeled test, we retain Δ as the primary pruning rule because it (i) operates in the same topic space that underpins our diagnostics, yielding interpretable prototype-level decisions; (ii) reduces lexical leakage from seed phrasing and is less sensitive to vectorizer settings and vocabulary drift; and (iii) provides a transparent, label-free =0 policy that performs well on the unlabeled corpus. We therefore report both scorers—using Δ for the main selection and TF–IDF as a robustness check—and include threshold-sensitivity curves in Figure 4 to make the trade-ofs explicit. Both scorers display the expected trade–of: as increases, precision rises while recall falls, and F1 peaks on a broad plateau. The TF–IDF margin reaches the highest peak F1 and attains near–perfect precision at more conservative , dominating the upper–right region of the curves. The multi–prototype Δ increases precision more gradually and preserves relatively higher recall near =0 , providing a smooth, interpretable operating range aligned with the topic–space diagnostics. In practice, TF–IDF at its data–driven MTFC-CIDF is preferable for balanced performance when labels (or proxy labels) are available, whereas Δ near =0 is a well-behaved default for label–free, topic–aligned pruning. Both methods are stable over a wide plateau of values, so small threshold shifts do not materially afect results.
To assess how the pruning threshold in the topic–space contrast score Δ = cos(p, s̄) − cos(p, ā) afects retrieval quality—and how this behavior varies with modeling choices— the topic model is refit over a grid spanning the three stages of the pipeline: (i) Text vectorization (bag–of–words with grams), controlling vocabulary granularity via min_df ∈ {2, 5}; (ii) Low–dimensional projection (UMAP), controlling local neighborhood size and layout via neighbors ∈ {15, 30, 50} and min_dist ∈ {0.0, 0.1}; (iii) Density–based clustering (HDBSCAN), controlling cluster granularity and treatment of noise via min_cluster_size ∈ {10, 30} and min_samples ∈ {None, 5}. For each configuration, topic probabilities are obtained, the contrast score Δ is recomputed, and is swept to trace precision/recall/F1 curves. Across configurations, curves cluster tightly around = 0 and around a fixed MCC (chosen once on a default configuration by maximizing Matthews correlation on a held-out split), with only minor dispersion. Peak F1 consistently occurs near ≈ 0 –0.1; performance degrades only for large positive where recall collapses (a regime not used operationally). Overall, selection performance is insensitive to reasonable variation in vectorizer, projection, and clustering hyperparameters, supporting = 0 as a default operating point and MCC as a robustness check.
a) Precision vs. threshold .
b) Recall vs. threshold .
c) 1 vs. threshold .
While the proposed pipeline demonstrates strong performance in identifying thematically coherent and contextually relevant patents, several limitations open avenues for further enhancement. In the expansion and pruning phases, citation networks were leveraged to build a semantically rich candidate set. To avoid redundancy and overfitting, these same features were excluded from the classification phase, opting instead for a single-input BERT architecture that processes concatenated abstracts and titles. This design ensured that distinct sets of information were exploited in diferent phases, reducing the risk of data leakage or circular logic.
However, treating all textual fields as a single input may dilute domain-specific signals, such as
those embedded in CPC codes or reference patterns. A multi-input BERT architecture, where each
modality (e.g., abstract, CPC, references) is processed independently before feature fusion, could better
preserve the structural and semantic nuances of each information type. For instance, recent work
by [
In the absence of human annotations, a pseudo-labeled set based on overlap with CPC codes is used. While defensible as a proxy for human curation, these labels are not independent of the construction rules; reported metrics should therefore be interpreted as upper-bound estimates. Future work will validate beyond CPC proxies and explore multi-input models that fuse text, classification codes, and citation structure.
This work introduces an automated, scalable framework for identifying twin (green ∩ digital) patents that integrate green and digital technologies. By combining rule-based seed selection, bidirectional citation expansion, transformer-based embeddings, and topic-guided pruning, the methodology addresses persistent limitations of earlier approaches, including limited coverage and overlap and lack of reproducibility.
Empirical results indicate that the proposed framework yields patent sets with greater topical relevance and improved semantic coherence, as shown by clearer intertopic separation and a KDE shift toward seed topics after pruning. Specifically, the combination of PaECTER embeddings, BERTopic modeling and topic-guided pruning, is shown to enable the efective filtering of heterogeneous patent corpora while minimizing human intervention. Sensitivity analyses indicate small variation across reasonable UMAP/HDBSCAN settings and cosine-similarity thresholds, suggesting a robust pruning score.
Overall, this approach provides a scalable, transparent basis for monitoring innovation dynamics and informing policy in the twin transition.
This study was funded by the European Union - NextGenerationEU, Mission 4, Component 2, in the framework of the GRINS -Growing Resilient, INclusive and Sustainable project (GRINS PE00000018 – CUP F53C22000760007). The views and opinions expressed are solely those of the authors and do not necessarily reflect those of the European Union, nor can the European Union be held responsible for them.
During the preparation of this work, the author used ChatGPT4.5 in order to perform grammar and spelling check, paraphrase and reword. After using this tool/service, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content.
epo.org/en/searching-for-patents/business/patstat, 2024. Accessed 2025-08-25. [32] European Patent Ofice, PATSTAT 2024 Autumn Edition Data Catalog, v5.24, Technical Report, European Patent Ofice, 2024. URL: https://link.epo.org/web/searching-for-patents/business/patstat/ data-catalog-patsat-global-en.pdf, accessed 2025-01-10.