Pith. sign in

REVIEW 5 major objections 6 minor 52 references

FullRecall: A Semantic Search-Based Ranking Approach for Maximizing Recall in Patent Retrieval

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that FullRecall, a three-phase semantic ranking pipeline, retrieves every known relevant prior-art patent in all five test cases.

desk verdict The 100% recall claim is an artifact of cutting the ranked list at the worst known relevant rank; the pipeline is novel and worth testing, but this evaluation does not support the headline result. read the letter →

arxiv 2507.14946 v1 pith:DMPSDEV4 submitted 2025-07-20 cs.IR cs.LG

Amna Ali , Liyanage C. De Silva , Pg Emeroylariffion Abas This is my paper
classification cs.IRcs.LG
keywords patentretrievalrecallsemanticsearchrankingkeyphraseextractionIPC-basedfilteringpriorartnounphrase
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces FullRecall, a three-phase patent prior-art retrieval approach, and claims it achieves 100% recall in all five test cases, meaning every patent that examiners cited against the query patent appears in the final ranked list. The motivation is that patent search is recall-critical: missing even one relevant document can invalidate a claim or trigger costly litigation. FullRecall's reported recall exceeds that of the two baselines, HRR2 and ReQ-ReC, on the same test patents. If the result holds, the approach suggests that complete coverage can be obtained from IPC descriptions, ranked noun phrases, and semantic re-ranking, without query-specific training data.

What carries the argument

The load-bearing object is the ranked noun-phrase list $NP_{query}$: a set of 12 to 16 technical phrases extracted from the query patent and ranked by a composite score $R(np_i)=\alpha P(np_i)+\beta D(np_i)+\delta B(np_i)+(1-\alpha-\beta-\delta)S(np_i)$ that mixes PageRank (a graph-based importance measure), degree centrality, and betweenness centrality with a semantic score based on cluster and graph centroids. Those phrases come from sentences of the query patent that cosine-match against key phrases derived from IPC descriptions, and they are manually assembled into a query that is run over a same-IPC patent dataset. Retrieved documents are then compared with the phrase list using cosine similarity of patent-domain transformer embeddings, and ranked by a final score $final\_score(dp)=weighted\_score(dp)+\lambda\cdot NP\_matches(dp)$ that rewards both the number of matching phrases and the similarity of the strongest matches. This ranking converts a large retrieved set into a smaller final list while, according to the paper, keeping every known relevant patent.

What would settle it

Recompute recall at fixed cutoffs rather than at the depth of the final ranked set. Using the rank positions reported by the paper, a top-100 cutoff would show none of the ten examiner-cited patents for the first query, since its relevant ranks start at 174; if the central claim is that full recall holds at a practical review depth, that observation would settle the question.

Watch

Extended reading notes

Core claim

The central claim is that recall-oriented patent retrieval can be made complete by splitting the task into three stages: deriving a small set of informative noun phrases from the IPC descriptions of the query patent, using those phrases to form a query that narrows the candidate set, and then re-ranking the retrieved patents by semantic similarity to the query's key phrases. On five patents used as test queries, the final ranked set contains every examiner-cited prior art patent; the paper reports 100% recall in each case, with the lowest-ranked relevant patent appearing at position 895, 1030, 1473, 1568, or 1660 depending on the query. The authors present this as evidence that the approach ensures no relevant prior art is overlooked, while keeping the ranked list small enough that an examiner can review fewer documents than the full retrieved set.

Load-bearing premise

The 100% recall result depends on defining the final ranked set to be exactly long enough to include the lowest-ranked relevant patent, so the perfect recall score is partly a consequence of the list length chosen for the evaluation.

Editorial extensions

If this is right

  • If FullRecall's 100% recall holds, patent examiners can restrict the search to a same-IPC candidate set plus a semantic re-ranking step without losing any known relevant citation.
  • The reported reduction from large retrieved sets to smaller ranked sets, for example from 3,635 to 895 documents for one query, means complete coverage can still produce a reviewable list.
  • The finding that 15 to 16 key phrases capture all target patents in all five test cases suggests query length can itself be tuned as a recall lever.
  • Because the pipeline works with IPC codes either assigned by an examiner or inferred by an inventor, the method is aimed at both pre-filing novelty searches and post-filing examination.
  • On the five test patents, FullRecall's reported recall of 100% is higher than HRR2's recorded values of 0% to 33.3% and ReQ-ReC's recorded values of 0% to 50%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Re-running the same five queries with fixed cutoffs chosen in advance (e.g., top 100 or top 500) would show how the method performs under the standard evaluation convention of a bounded ranked list.
  • Because the ground truth is examiner citations, a next test is to judge relevance with independent labels; that would reveal whether the approach finds all relevant prior art or only the cited subset.
  • Extending the candidate dataset beyond the query patent's own IPC codes would test the framework's behaviour when relevant prior art lives in neighbouring classifications.
  • Automating the manual query-formulation step would show whether the reported recall depends on the chosen phrasing or on the ranked phrase list itself.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. This paper proposes FullRecall, a three-phase pipeline for recall-oriented patent prior-art search. Phase 1 starts from the IPC codes of a patent under observation, extracts bi- and tri-gram keyphrases from IPC descriptions with YAKE, selects sentences from the patent text by cosine similarity, extracts noun phrases with spaCy, embeds them with BERT for Patents, clusters them with HDBSCAN, and ranks the phrases with a composite score combining connectivity, uniqueness, PageRank, degree centrality, and betweenness centrality. Phase 2 selects the top K noun phrases (12 ≤ K ≤ 16) and forms a structured query with human oversight. Phase 3 constructs a dataset from documents sharing the query patent's IPC classes, executes the query, and ranks the retrieved documents by cosine similarity against the ranked noun phrases. The evaluation uses five patents with their examiner citations as ground truth; the paper reports 100% recall for all five cases and compares the method with HRR2 and ReQ-ReC.

Significance. Recall-oriented retrieval is a legitimate and practically important goal for patent search, and the authors have assembled a nontrivial five-patent case study with examiner-citation ground truth and two baseline comparisons. The pipeline description is detailed, and the use of IPC descriptions as domain priors is a reasonable design choice. However, the headline claim of 100% recall does not follow from the experiments as reported: the ranked set is truncated at the last relevant document, the candidate dataset contains the ground-truth documents by construction, and the query length is selected per patent after seeing the results. No code, parameter values, or fixed-cutoff results are provided, so the central claim is not reproducible as stated.

major comments (5)
  1. [3 (Tables 7 and 9)] The 100% recall result is an artifact of the evaluation cutoff. In Table 7, the 'Ranked set with FullRecall' sizes are 1660, 895, 1030, 1473, and 1568 for the five patents, and in Table 9 the lowest-ranked target patent in each case sits exactly at that size (ranks 1660, 895, 1030, 1473, and 1568). Recall is therefore measured at a cutoff chosen to be the maximum target rank, so any system that places every target in its ranked output automatically obtains 100% recall. Under a fixed cutoff such as top 1000, Table 9 shows the claim would fail: P1_UO has targets at ranks 1331, 1439, and 1402, and P5_UO has targets at ranks 1273, 1463, and 1568. The Abstract's '100% recall' is a consequence of this data-dependent truncation, not a demonstrated property of the ranking method.
  2. [4] The baseline comparison uses the same data-dependent boundary. Section 4 states that for ReQ-ReC the value of n was set to 1660, 895, 1030, 1473, and 1568, exactly matching FullRecall's final ranked-set sizes. This does not create a meaningful comparison at a fixed operating point: FullRecall's output size is defined by the position of its last relevant document, while ReQ-ReC is forced to return lists of the same externally imposed length. No fixed-cutoff comparison is reported, so Table 8 cannot support the conclusion that FullRecall 'outperforms' the baselines in any practically useful evaluation regime.
  3. [3 (dataset construction)] The candidate dataset is constructed so that the ground truth is guaranteed to be present. The paper states: 'These datasets also include the examiner-cited patents identified earlier in Table 2, ensuring that relevant prior art is captured.' Because S_d is built from the query patent's own IPC codes and is explicitly seeded with the relevant documents, the experiment measures retrieval from a small, domain-filtered set that is known to contain the target items. This does not simulate a realistic full-corpus prior-art search and removes the main difficulty the paper claims to address.
  4. [2.2 and Table 6] The query length k is tuned per patent to produce the reported result. Table 6 shows that full citation coverage is reached only at specific query lengths: P1_UO achieves 10/10 only at k=16, P3_UO reaches 6/6 only at k=15-16, P4_UO reaches 5/5 only at k=16, and P5_UO reaches 7/7 only at k=15-16. The paper reports the successful k values after the fact, while the manual query construction f(·) in Eq. (30) is not described operationally. In addition, no values are given for alpha_c, alpha_u, alpha, beta, delta, tau, lambda, or the HDBSCAN hyperparameters, so the result cannot be reproduced or checked for sensitivity to these choices.
  5. [2.1 (Eqs. 18-28)] The Phase-1 ranking formula is internally inconsistent in its polarity. Equations (19)-(20) define connectivity in terms of Euclidean distances for which lower values indicate greater centrality, and Eqs. (22)-(23) define uniqueness in terms of cosine similarities for which lower values indicate greater distinctiveness. Yet Eq. (28) adds the semantic score S(npi) with a positive coefficient to PageRank, degree centrality, and betweenness centrality, all of which are higher-is-better. Under the stated equations, phrases that are less central and less unique receive higher R(npi), which is the opposite of the intended ranking. The manuscript does not state that S is negated or normalized before combination, so the top-K phrase selection that drives the query is not well defined as written.
minor comments (6)
  1. [2.2] Equation (29) states '12 ≤ K ≥ 16'; this should be '12 ≤ K ≤ 16'.
  2. [2.1 (Algorithm Phase 1c)] Algorithm Phase 1c line 1 contains a typo: 'HSBSCAN' should be 'HDBSCAN'.
  3. [4] The baseline is referred to inconsistently as HRR2, HHR2, and, in the text after Table 8, 'HRR2 [3]' instead of [1]; the references and terminology should be harmonized.
  4. [Table 1] The notation table lists 'Set on m clusters' under the entry for C; this should be 'Set of m clusters'.
  5. [3] The five patents are described as 'randomly selected across diverse IPC domains,' but no sampling procedure, inclusion criteria, or confidence statement is given; this limits the generalizability claim for all five test cases.
  6. [3 (Table 9)] No ranking-quality metric such as NDCG, precision@k, or reciprocal rank is reported, so the claim of 'meaningful document ranking' is not quantitatively supported; Table 9 lists target ranks but does not measure ranking quality against the full relevant set.

Circularity Check

3 steps flagged · score 8.0 of 10

FullRecall's headline 100% recall is forced by the evaluation design: the candidate corpus is seeded with the examiner-cited targets, the comparison cutoff is set to the lowest-ranked target, and the query length k is selected after seeing which values achieve full recall.

  1. fitted input called prediction [Section 3, final phase of FullRecall (candidate dataset S_d construction)]
    "These datasets also include the examiner-cited patents identified earlier in Table 2, ensuring that relevant prior art is captured."

    The claimed outcome, 100% recall, is measured over S_d, a corpus that is explicitly seeded with every examiner-cited target patent before retrieval begins. The ground-truth relevant documents are therefore part of the search space by construction, making the 'complete recall' reported in the abstract partly an input to the experiment rather than a discovered property of the ranking algorithm.

  2. self definitional [Section 4, Performance Comparison; compare Table 7 with Table 9]
    "To ensure a meaningful comparison, the retrieval boundary for ReQ-ReC was expanded to match the point at which FullRecall achieved 100% recall for each query patent [1]. For instance, in the case of query patent P1_UO in FullRecall setting, all relevant documents were retrieved within the top 1660 documents of the ranked list. Therefore, the value of n for ReQ-ReC was set to 1660 to evaluate how many relevant documents ReQ-ReC could identify within that range."

    The evaluation cutoff n is not predefined; it is set to the maximum rank occupied by a target patent in FullRecall's own output. Table 9 lists US20190363843 A1 at rank 1660 for P1_UO, and Table 7 gives FullRecall's final ranked set size as 1660; the analogous equality holds for 895, 1030, 1473 and 1568. At cutoff n = max rank of the targets, any output containing all targets has recall 1 by definition, because every relevant document is above the cutoff. Under a fixed cutoff of, say, top-1000, P1_UO would retain only 5 of 10 targets, so the perfect-recall result is an artifact of the data-dependent boundary.

1 more flagged steps
  1. fitted input called prediction [Section 3, Table 6 and the surrounding discussion of query length k]
    "All five patents under observation achieve full citation recall at this range, validating the ranked noun phrase approach employed during query formulation."

    Query length k is not fixed in advance. Table 6 varies k from 12 to 16 and shows full recall only at selected values: P1_UO finds 3/10 targets at k=12-14 and 10/10 only at k=16; P4_UO finds 4/5 at k=15 and 5/5 at k=16; P5_UO reaches 7/7 only at k>=15. The paper then reports 100% recall at 'this range', where the range is chosen because it contains the successful k values. The headline result is thus a post-hoc best-case selection over a tuned parameter, not a measurement at an a priori design setting.

full rationale

The mathematical derivation in Section 2 is not itself circular: equations (1)-(37) define the pipeline and no equation reduces to its own output. The circularity is entirely in the evaluation protocol. The candidate dataset is constructed so that it includes every examiner-cited target; the cutoff at which recall is reported is set equal to the worst rank of those targets in FullRecall's output; and the query-length parameter k is chosen after observing which values recover all targets. Each of these choices makes the '100% recall' result true by construction rather than by an independent retrieval benchmark. The comparison with ReQ-ReC is also biased in FullRecall's favor because the same data-dependent cutoff is imposed on the baseline. There are self-citations in the reference list, but they are not load-bearing for the central claim, so the score reflects evaluation circularity rather than citation circularity. A fixed-cutoff evaluation against an independent corpus would be needed before the perfect-recall claim can be treated as a property of the method.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The ledger shows the paper's contribution is largely a new assembly of existing tools rather than a derivation. The free parameters are mostly unreported, and the most consequential choice, the ranked-set cutoff n, is determined by the ground truth ranks. The axioms are domain assumptions that make the evaluation tractable but not general. No invented entities are introduced.

free parameters (8)
  • alpha_c (connectivity weight) = not reported
    Eq. 18 interpolates between cluster centroid distance and global centroid distance; no value is given.
  • alpha_u (uniqueness weight) = not reported
    Eq. 21 combines intra-cluster and cross-cluster uniqueness; no value is given.
  • alpha, beta, delta (composite ranking weights) = not reported
    Eq. 28 combines PageRank, degree centrality, betweenness centrality, and the semantic score; no values or constraints are given.
  • cosine threshold tau = not reported
    The same symbol tau is used in sentence selection (Eq. 9), graph edge construction (Eq. 15), and valid matching (Eq. 33); the threshold values are never specified.
  • K (number of query phrases) = 15 or 16 per query
    Table 6 varies k from 12 to 16, and 100% recall is only reached at k=15 or k=16; the query length is chosen after seeing which values succeed.
  • lambda (match count weight) = not reported
    Eq. 36 weights the number of valid noun phrase matches; no value is given.
  • ranked-set cutoff n per query = P1 1660, P2 895, P3 1030, P4 1473, P5 1568
    Tables 7 and 9 show the final ranked set ends at the worst-positioned target patent; Section 4 confirms this cutoff is used as the recall boundary, making 100% recall true by construction.
  • HDBSCAN hyperparameters = not reported
    Algorithm Phase 1c invokes HDBSCAN without specifying min_cluster_size or other parameters.
assumptions (7)
  • domain assumption Examiner citations are the complete ground truth for relevant prior art.
    Used in Section 3 and Table 2; the recall denominator for each query is only these 4 to 10 citations, ignoring all other potentially relevant documents.
  • domain assumption The candidate dataset S_d is the correct search universe.
    Built from IPC codes and explicitly includes the examiner-cited patents, so all known relevant documents are inside the pool by construction (Section 3).
  • domain assumption Embedding cosine similarity measures patent relevance.
    Used throughout Eqs. 8, 15, and 32 without calibration against human relevance judgments.
  • domain assumption Noun phrases represent the core technical content of patents.
    Justified in Section 2.1 by the claim that over 90% of technical phrases are noun-based, cited from reference [44].
  • domain assumption Independent claims alone are sufficient to judge document relevance.
    Phase 3 scores each document only from its independent claims (Section 2.3), assuming no relevant information is missed from other sections.
  • domain assumption Human query formulation is an acceptable part of the method.
    Section 2.2 introduces manual structuring of the query, making the system dependent on expert judgment and reducing reproducibility.
  • ad hoc to paper Recall can be measured on a ranked list truncated at the last relevant document.
    Section 4 sets the comparison cutoff n to the rank of FullRecall's last target patent; without this assumption, 100% recall is not defined for a fixed system output.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FullRecall: A Semantic Search-Based Ranking Approach for Maximizing Recall in Patent Retrieval." pith.science (2026). https://pith.science/paper/DMPSDEV4

@misc{pith2026250714946,
  author       = {Pith},
  title        = {Pith review of: FullRecall: A Semantic Search-Based Ranking Approach for Maximizing Recall in Patent Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DMPSDEV4}},
  note         = {Machine review of arXiv:2507.14946}
}
read the original abstract

Patent examiners and inventors face significant pressure to verify the originality and non-obviousness of inventions, and the intricate nature of patent data intensifies the challenges of patent retrieval. Therefore, there is a pressing need to devise cutting-edge retrieval strategies that can reliably achieve the desired recall. This study introduces FullRecall, a novel patent retrieval approach that effectively manages the complexity of patent data while maintaining the reliability of relevance matching and maximising recall. It leverages IPC-guided knowledge to generate informative phrases, which are processed to extract key information in the form of noun phrases characterising the query patent under observation. From these, the top k keyphrases are selected to construct a query for retrieving a focused subset of the dataset. This initial retrieval step achieves complete recall, successfully capturing all relevant documents. To further refine the results, a ranking scheme is applied to the retrieved subset, reducing its size while maintaining 100% recall. This multi-phase process demonstrates an effective strategy for balancing precision and recall in patent retrieval tasks. Comprehensive experiments were conducted, and the results were compared with baseline studies, namely HRR2 [1] and ReQ-ReC [2]. The proposed approach yielded superior results, achieving 100% recall in all five test cases. However, HRR2[1] recall values across the five test cases were 10%, 25%, 33.3%, 0%, and 14.29%, while ReQ-ReC [2] showed 50% for the first test case, 25% for the second test case, and 0% for the third, fourth, and fifth test cases. The 100% recall ensures that no relevant prior art is overlooked, thereby strengthening the patent pre-filing and examination processes, hence reducing potential legal risks.

Figures

Figures reproduced from arXiv: 2507.14946 by the authors.

Figure 2
Figure 2. Significance of Independent Claims of a Patent Document For each retrieved patent document 𝑑𝑝 ∈ 𝑅𝑠𝑑 and for each noun phrase 𝑛𝑝𝑖 ∈ 𝑁𝑃𝑞𝑢𝑒𝑟𝑦 from the ranked set of extracted noun phrases, the cosine similarity is computed between their respective embeddings: cos (𝑒𝑑𝑝 , 𝑒𝑛𝑝𝑖 ) = 𝑒𝑑𝑝 ⋅ 𝑒𝑛𝑝𝑖 ‖𝑒𝑑𝑝 ‖ . ‖𝑒𝑛𝑝𝑖 ‖ (32) where 𝑒𝑑𝑝 represents the embedding of the retrieved patent document 𝑑𝑝 ∈ 𝑅𝑠𝑑, while 𝑒𝑛𝑝𝑖 represents the embed… view at source ↗
Figure 4
Figure 4. Second phase of the algorithm - Retrieval of ranked relevant documents with Full Recall 3. Results and Discussions Empirical experiments were conducted to evaluate the effectiveness of the proposed FullRecall framework in simulating real-world patent retrieval scenarios for both patent examiners and inventors. Patent examiners perform prior art searches for patent applications Algorithm Phase 1c: Ranking of noun phr… view at source ↗
Figure 3
Figure 3. Graphical representation of noun phrases for 𝑃1_𝑈𝑂 [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Graphical representation of noun phrases for 𝑃2_𝑈𝑂 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Graphical representation of noun phrases for 𝑃3_𝑈𝑂 [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Graphical representation of noun phrases for 𝑃4_𝑈𝑂 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 40 canonical work pages

  1. [1]

    J., Lee, W., & Afshar, J

    Song, J. J., Lee, W., & Afshar, J. (2019). An effective High Recall Retrieval method. Data & Knowledge Engineering, 123, 101603. https://doi.org/https://doi.org/10.1016/j.datak.2017.07.006

    work page doi:10.1016/j.datak.2017.07.006 2019

  2. [2]

    Li, C., Wang, Y., Resnick, P., & Mei, Q. (2014). ReQ-ReC: high recall retrieval with query pooling and interactive classification. https://doi.org/10.1145/2600428.2609618

    arXiv 2014

  3. [3]

    Losee, R. M. (1991). An analytic measure predicting information retrieval system performance. Information Processing & Management, 27(1), 1-13. https://doi.org/https://doi.org/10.1016/0306-4573(91)90027-J

    work page doi:10.1016/0306-4573(91)90027-j 1991

  4. [4]

    King, D., Downey, D., & Weld, D. S. (2020). High-Precision Extraction of Emerging Concepts from Scientific Literature Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, China. https://doi.org/10.1145/3397271.3401235

    arXiv 2020

  5. [5]

    C., & Abas, P

    Ali, A., Tufail, A., De Silva, L. C., & Abas, P. E. (2024). Innovating Patent Retrieval: A Comprehensive Review of Techni ques, Trends, and Challenges in Prior Art Searches. Applied System Innovation, 7(5), 91. https://www.mdpi.com/2571-5577/7/5/91

    work page 2024

  6. [6]

    Varma, B. K. (2014). Chapter 3.4 - Intellectual Property Rights and the Technology Transfer Process. In S. Seetharaman (Ed.), Treatise on Process Metallurgy (pp. 1249-1289). Elsevier. https://doi.org/https://doi.org/10.1016/B978-0-08-096988-6.00039- 0

    work page doi:10.1016/b978-0-08-096988-6.00039- 2014

  7. [7]

    J., & Lee, W

    Song, J. J., & Lee, W. (2020). Relevance maximization for high-recall retrieval problem: finding all needles in a haystack. J. Supercomput., 76(10), 7734–7757. https://doi.org/10.1007/s11227-016-1956-8

    work page doi:10.1007/s11227-016-1956-8 2020

  8. [8]

    Deng, Y., Qi, Y., & Guo, Q. (2024). The impact of patent infringement litigation decisions on firms’ innovation performance in China. Scientific Reports, 14(1), 23564. https://doi.org/10.1038/s41598-024-74924-1

    work page doi:10.1038/s41598-024-74924-1 2024

Show all 52 references
  1. [9]

    (2010, March 28, 2010)

    Magdy, W., & Jones, G. (2010, March 28, 2010). A new metric for patent retrieval evaluation AsPIRe’10, 1st International Workshop on Advances in Patent Information Retrieval, Milton Keynes. https://doras.dcu.ie/16181/1/A_New_Metric_for_Patent_Retrieval_Evaluation.pdf

    2010

  2. [10]

    Magdy, W., & Jones, G. J. F. (2010). PRES: a score metric for evaluating recall-oriented information retrieval applications Proceedings of the 33rd international ACM SIGIR conference on Research and development in information retrieval, Geneva, Switzerland. https://doi.org/10....

    2010

  3. [11]

    C., J.; Park, S.; Jang, D

    Kim, J. C., J.; Park, S.; Jang, D. (2018). Patent Keyword Extraction for Sustainable Technology Management. Sustainability 8(10),

    2018

  4. [12]

    Paranjpe, P. (2012). Patent Information and Search. DESIDOC Journal of Library & Information Technology, 32, 197-203. https://doi.org/10.14429/djlit.32.3.2375

    2012 doi

  5. [13]

    I., and Noriko Kando (2007, May 15-18, 2007)

    Atsushi Fujii, M. I., and Noriko Kando (2007, May 15-18, 2007). Overview of the patent retrieval task at the NTCIR-6 workshop Proceedings of NTCIR-6 Workshop Meeting, Tokyo, Japan http://research.nii.ac.jp/ntcir/workshop/OnlineProceedings6/NTCIR/78.pdf

    2007

  6. [15]

    Shalaby, W., & Zadrozny, W. (2019). Patent retrieval: a literature review. Knowledge and Information Systems, 61(2), 631-660. https://doi.org/10.1007/s10115-018-1322-7

    2019 doi

  7. [16]

    Xue, X., & Croft, W. B. (2009). Transforming patents into prior-art queries Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, Boston, MA, USA. https://doi.org/10.1145/1571941.1572139

    2009

  8. [17]

    Mahdabi, P., & Crestani, F. (2014). The effect of citation analysis on query expansion for patent retrieval. Information Retrieval, 17(5), 412-429. https://doi.org/10.1007/s10791-013-9232-5

    2014 doi

  9. [18]

    Jones, W. M. a. G. J. F. (October 24, 2011, October 24, 2011). A Study on Query Expansion Methods for Patent Retrieval PaIR’11, Proceedings of the 4th workshop on Patent information retrieval, Glasgow, Scotland, UK. https://doras.dcu.ie/16517/1/A_Study_on_Query_Expansion_Metho...

    2011

  10. [19]

    Mahdabi, P., & Crestani, F. (2014). Patent Query Formulation by Synthesizing Multiple Sources of Relevance Evidence. ACM Transactions on Information Systems, 32(4), 1-30. https://doi.org/http://dx.doi.org/10.1145/2651363

    2014 doi

  11. [20]

    (2024, July 14–18)

    Björkqvist, S. (2024, July 14–18). Relevance Feedback Method For Patent Searching Using Vector Subspaces SIGIR ’24, Washington, DC, USA https://dl.acm.org/doi/pdf/10.1145/3626772.3661365

    2024

  12. [21]

    Bashir, S., & Rauber, A. (2009). Improving retrievability of patents with cluster-based pseudo-relevance feedback documents selection Proceedings of the 18th ACM conference on Information and knowledge management, Hong Kong, China. https://doi.org/10.1145/1645953.1646250

    2009

  13. [22]

    Narges Neshat, A. K. (2024). Using the Citation-Content-Based Approach to Patent Clustering. International Journal of Information Science and Management, 22(2), 139-149. https://doi.org/https://dorl.net/dor/20.1001.1.20088302.2024.22.2.9.8

    2024

  14. [23]

    M., Lee, C

    Kang, D. M., Lee, C. C., Lee, S., & Lee, W. (2020, 12 August 2020). Patent prior art search using deep learning language model IDEAS '20: Proceedings of the 24th Symposium on International Database Engineering & Application, https://api.semanticscholar.org/CorpusID:221299629

    2020

  15. [24]

    J., & Lee, W

    Song, J. J., & Lee, W. (2015). High Recall-Low Cost Model for Patent Retrieval Proceedings of the 2015 International Conference on Big Data Applications and Services, Jeju Island, Republic of Korea. https://doi.org/10.1145/2837060.2837098

    2015

  16. [25]

    Hu, Z., Zhou, X., & Lin, A. (2023). Evaluation and identification of potential high-value patents in the field of integrated circuits using a multidimensional patent indicators pre-screening strategy and machine learning approaches. Journal of Informetrics, 17(2), 101406. http...

    2023

  17. [26]

    Ali, A., Humayun, M., De Silva, L., & Abas, P. E. (2025). Optimizing Patent Prior Art Search: An Approach Using Patent Abstract and Key Terms. Information, 16, 145. https://doi.org/10.3390/info16020145

    2025 doi

  18. [27]

    Khode, A., & Jambhorkar, S. (2017). A Literature Review on Patent Information Retrieval Techniques. Indian Journal of Science and Technology, 10, 1-13. https://doi.org/10.17485/ijst/2017/v10i37/116435

    2017 doi

  19. [28]

    and Zhang, L

    Dessi, R., Aras, H. and Zhang, L. (2023, September 4, 2023). DeepKEA: Employing Deep Learning Models for Keyword Extraction from Patent Documents LIRAI’23: 1st Legal Information Retrieval meets Artificial Intelligence Workshop, Rome, Italy. https://ceur-ws.org/Vol-3594/paper3.pdf

    2023

  20. [29]

    Stamatis, V. (2022). End to End Neural Retrieval for Patent Prior Art Search. In (pp. 537-544). https://doi.org/10.1007/978-3-030- 99739-7_66

    2022 doi

  21. [30]

    Lee, J.-S., & Hsiang, J. (2020). Patent classification by fine-tuning BERT language model. World Patent Information, 61, 101965. https://doi.org/10.1016/j.wpi.2020.101965

    2020

  22. [31]

    Reimers, N., & Gurevych, I. (2020). Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation. https://doi.org/10.48550/arXiv.2004.09813

    2020 doi arXiv

  23. [32]

    (2012, 19-21 September, 2012)

    Tannebaum, W., & Rauber, A. (2012, 19-21 September, 2012). Acquiring Lexical Knowledge from Query Logs for Query Expansion in Patent Searching 2012 IEEE Sixth International Conference on Semantic Computing, Palermo, Italy. https://sci- hub.se/10.1109/icsc.2012.15

    2012 doi

  24. [33]

    Jin, Z., Yang, Z., Tang, G., Liu, T., & Xun, E. (2024). Research on Keyword-Based Element Extraction for Chinese Patent Retrieval. https://doi.org/10.1109/IALP63756.2024.10661116

    2024

  25. [34]

    Kanoulas, J. R. a. M. W. a. E. (2018). Query Generation for Patent Retrieval with Keyword Extraction based on Syntactic Features. In Legal Knowledge and Information Systems, 210-214. https://arxiv.org/abs/1906.07591

    2018 arXiv

  26. [35]

    and Jurowetzki, R

    Bekamiri, H., Hain, D.S. and Jurowetzki, R. (2024). Patentsberta: A deep nlp based hybrid model for patent distance and classification using augmented sbert. Technological Forecasting and Social Change, 206, 123536. https://doi.org/https://doi.org/10.1016/j.techfore.2024.123536

    2024

  27. [36]

    Nogueira, R., Yang, W., Lin, J., & Cho, K. (2019). Document Expansion by Query Prediction. https://doi.org/10.48550/arXiv.1904.08375

    2019 doi arXiv

  28. [37]

    M., Perego, R., Tonellotto, N., Goharian, N., & Frieder, O

    MacAvaney, S., Nardini, F. M., Perego, R., Tonellotto, N., Goharian, N., & Frieder, O. (2020, July, 25-30, 2020). Efficient Document Re-Ranking for Transformers by Precomputing Term Representations SIGIR ’20: Conference on Research and Development in Information Retrieval, Vir...

    2020 arXiv

  29. [38]

    Stamatis, V., Salampasis, M., & Diamantaras, K. (2024). A novel re-ranking architecture for patent search. World Patent Information, 78, 102282. https://doi.org/https://doi.org/10.1016/j.wpi.2024.102282

    2024

  30. [39]

    Wang, S., Hu, L., Wang, Y., He, X., Sheng, Q., Orgun, M., Cao, L., Ricci, F., & Yu, P. (2021). Graph Learning based Recommender Systems: A Review. https://doi.org/10.24963/ijcai.2021/630

    2021 doi

  31. [40]

    Kuwa, T., Schamoni, S., & Riezler, S. (2020). Embedding Meta-Textual Information for Improved Learning to Rank. https://doi.org/10.48550/arXiv.2010.16313

    2020 doi arXiv

  32. [41]

    Huang, Z., & Xie, Z. (2022). A patent keywords extraction method using TextRank model with prior public knowledge. Complex & Intelligent Systems, 8(1), 1-12. https://doi.org/10.1007/s40747-021-00343-8

    2022 doi

  33. [42]

    Vowinckel, K., & Hähnke, V. D. (2023). SEARCHFORMER: Semantic patent embeddings by siamese transformers for prior art search. World Patent Information, 73, 102192. https://doi.org/https://doi.org/10.1016/j.wpi.2023.102192

    2023

  34. [43]

    Lu, Y., Tong, X., Xiong, X., & Zhu, H. (2024). Knowledge graph enhanced citation recommendation model for patent examiners. Scientometrics, 129(4), 2181-2203. https://doi.org/10.1007/s11192-024-04966-9

    2024 doi

  35. [44]

    Liu, Y., Wu, H., Huang, Z., Wang, H., Ma, J., Liu, Q., Chen, E., Tao, H., & Rui, K. (2020). Technical Phrase Extraction for Patent Mining: A Multi-level Approach. https://doi.org/10.1109/ICDM50108.2020.00139

    2020

  36. [45]

    Team, W. D. (2020). Patent Claim. Legal Information Institute. https://www.law.cornell.edu/wex/patent_claim#:~:text=Independent%20claims%20can%20%E2%80%9Cstand%20alone,nec essary%20to%20define%20the%20invention

    2020

  37. [46]

    (2024, Jan 2025)

    Minesoft origin, Advanced AI Patent Search. (2024, Jan 2025). ©minesoft. https://minesoft.com/

    2024

  38. [47]

    Tran, B. (2025). The Role of International Patent Classifications (IPC) in Global Filings. https://patentpc.com/blog/the-role-of- international-patent-classifications-ipc-in-global-filings

    2025

  39. [48]

    Nadim, M., Akopian, D., & Matamoros, A. (2023). A Comparative Assessment of Unsupervised Keyword Extraction Tools. IEEE Access, 11, 144778-144798. https://doi.org/10.1109/ACCESS.2023.3344032

    2023

  40. [49]

    Honnibal, M. (2017). spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. https://spacy.io/

    2017

  41. [50]

    H., Steve Astels

    Leland McInnes, J. H., Steve Astels. (2017). hdbscan: Hierarchical density based clustering. Journal of Open Source Software, 11(2). https://doi.org/doi:10.21105/joss.00205

    2017 doi

  42. [51]

    (2020, 25-27 September, 2020)

    Deng, D. (2020, 25-27 September, 2020). DBSCAN Clustering Algorithm Based on Density 2020 7th International Forum on Electrical Engineering and Automation (IFEEA), Hefei, China. https://sci-hub.se/10.1109/ifeea51475.2020.00199

    2020

  43. [52]

    (2019, Aug 14, 2019)

    Bhasin, J. (2019, Aug 14, 2019). Graph Analytics — Introduction and Concepts of Centrality. https://medium.com/towards-data- science/graph-analytics-introduction-and-concepts-of-centrality-8f5543b55de3

    2019

  44. [1287]

    https://doi.org/https://doi.org/10.3390/su10041287

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.