Pith. sign in

REVIEW 3 major objections 7 minor 40 references

Ontology-Aligned Embeddings for Data-Driven Labour Market Analytics

T0 review · 3 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fine-tuning a sentence-embedding model on official German occupation codes produces an alignment space in which nearest-neighbour search classifies free-form job titles to KldB 2010 and ISCED 2011 with high macro-F1.

desk verdict A plausible, incremental method for aligning German job titles with KldB and ISCED, but the evaluation is in-distribution on official BAfA terms and the 'any free-form title' claim outruns the evidence. read the letter →

arxiv 2509.04942 v1 pith:A63KSUVQ submitted 2025-09-05 cs.LG

Heinke Hihn , Dennis A. V. Dittrich , Carl Jeske , Cayo Costa Sobral , Helio Pais , Timm Lochmann This is my paper
classification cs.LG
keywords embeddingmodelSentenceTransformerssemanticsearchcontrastivelearningapproximatek-nearest-neighbourKldB2010ISCED2011labourmarketanalytics
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

This paper tries to show that a lightweight embedding-based system can replace hand-maintained occupational ontologies for German labour-market analytics. It fine-tunes a sentence-embedding model on roughly 1.5 million triplet samples derived from 525,207 official occupational terms, so that embeddings respect the structure of the KldB 2010 classification. A nearest-neighbour search then classifies free-form job titles into the 5-digit KldB 2010 codes and, through a rule-based mapping, into ISCED 2011 education levels. The reported macro-F1 scores are strong: 0.926 at the 4-digit subgroup level, 0.886 at the 5-digit type level, and 0.961 for the requirement-level digit. If the transfer to genuinely free-form titles holds, this matters because it offers a fast, explainable, and extensible way to reason across occupational and educational data at scale.

What carries the argument

The load-bearing object is an ontology-aligned embedding space: a Sentence-BERT variant fine-tuned so that job titles sharing a KldB 2010 code are pulled together and titles with different codes are pushed apart. Training uses triplet contrastive learning with Multiple Negatives Ranking Loss plus Matryoshka Loss, and the input is structured with separator tokens for job title, qualification, and skills. At inference, an approximate nearest-neighbour index (HNSW) treats classification as a semantic search problem: the majority vote of the top-k neighbours' KldB and ISCED codes answers the query. This makes adding new classes as simple as inserting new labelled embedding points.

What would settle it

Build or obtain a gold-standard set of free-form German job titles with expert-assigned KldB 2010 and ISCED 2011 codes, for example from social security contribution records, and measure the system's macro-F1 at the subgroup level. If it lands near the reported 0.926, the transfer claim is supported; if it falls well below, the random-holdout evaluation on official titles overstated real-world performance.

Watch

Extended reading notes

Core claim

A single fine-tuned sentence-embedding model can align free-form German job titles with two official classifications: the five-digit German Klassifikation der Berufe (KldB 2010) and the International Standard Classification of Education (ISCED 2011). Training data comes from about 525,000 official occupational terms from the German public employment service's occupational database, organised into anchor-positive-negative triplets where positive titles share the same KldB code and negatives do not. The fine-tuned model is evaluated by k-nearest-neighbour search with HNSW and majority vote; it reaches macro-F1 0.926 at the four-digit subgroup level, 0.886 at the five-digit type level, and 0.96

Load-bearing premise

The evaluation assumes that a random hold-out of official job titles from the public occupational database represents real free-form job titles; the paper itself notes that there is no gold-standard set of free-form titles to verify this.

Editorial extensions

If this is right

  • Real-time labour-market applications—job recommendation, pay-band calibration, skill-gap detection—can classify free-form German titles to KldB and ISCED in sub-second time without human curation.
  • The method beats the evaluated baselines at every KldB hierarchy level when compared on the official-terms test split, with particularly strong performance at the requirement-level digit.
  • The requirement-level digit (job complexity) is the most reliably predicted dimension, and the ablations suggest the model generalises beyond surface gender and word-order variants.
  • Because classification is k-NN search in a graph, adding new classes, new ontology codes, or additional languages requires adding labelled embedding points rather than redesigning the classifier.
  • The ISCED mapping makes occupational and educational taxonomies interoperable, enabling education-aware and cross-nationally comparable labour-market analyses.

Reading between the lines

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

  • The same learned embedding space could serve as a similarity measure for career-transition or skill-gap analyses, not just classification; the paper does not test this use.
  • If the transfer assumption fails in practice, the failure would likely show up first at the 5-digit type level, where lexical variability is highest and class sizes are smallest.
  • The rule-based ISCED mapper is the least learned part of the pipeline; replacing it with a model that predicts ISCED directly, or running ISCED prediction first and using it to constrain KldB, could improve ambiguous titles—an inversion the paper mentions as future work.
  • The graph formulation suggests a cheap multilingual extension: embed foreign-language titles and connect them to the existing German nodes by nearest-neighbour links, then propagate codes; this aligns with the planned work on additional international taxonomies but is not demonstrated here.
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

3 major / 7 minor

Summary. The paper proposes an ontology-alignment method for German job titles. It fine-tunes Sentence-BERT on contrastive triplets built from 525,207 official BAfA terms, using a query structure with job title, qualification, and skill fields, and then performs classification via HNSW k-nearest-neighbour search over the resulting embeddings. KldB 2010 codes are predicted directly and ISCED 2011 levels are inferred through a hand-crafted rule-based mapping. Evaluation on an 80/20 random split of the BAfA terms reports macro-F1 of 0.926 at subgroup (4-digit), 0.886 at type (5-digit), and 0.961 on the requirement digit, plus ablations on gender variants, management nouns, and word-order reversal. The authors frame the contribution as a lightweight, queryable, and extendable knowledge graph for real-time labour market analytics.

Significance. The problem is practically important, and the design is reasonable: the use of publicly available BAfA data, explicit separator-token query structure, Matryoshka losses for flexible embedding dimensions, and HNSW for sub-second inference are sensible engineering choices. If the central transfer claim were validated on genuinely free-form job titles, the approach would offer a useful lightweight alternative to LLM-based classifiers. The authors also honestly discuss the absence of a gold-standard free-form evaluation set. However, the reported evidence is entirely in-distribution on the official BAfA vocabulary, and the ISCED mapping is validated only by anecdote. These gaps directly affect the headline claims of the abstract.

major comments (3)
  1. [§4.1, §3.3, Table 5, §5.2] The central claim to link 'any free-form German job title' is not supported by the evaluation. The test set is a random 20% split of the same 525,207 BAfA official terms used to build the training triplets in §3.3. Since training and test queries are drawn from the same canonical vocabulary, lexical near-duplicates of an occupation can straddle the split; with k=1 retrieval, the model can match surface forms rather than learned semantics. The authors themselves state in §5.2 that no gold-standard free-form title set exists. Thus the macro-F1 values (subgroup 0.926, type 0.886) are upper bounds for in-vocabulary retrieval, not evidence of transfer to employer-style titles. An external held-out set of truly free-form titles, or at least a split that removes lexical variants across train/test, is needed before the abstract's claim can stand.
  2. [§3.3.2, §3.4.1, Table 4] The ISCED 2011 mapping is rule-based and is never quantitatively evaluated. Table 4 gives only hand-picked examples, and the inference examples in §3.4.1 are anecdotal. There is no labelled ISCED test set, no inter-annotator agreement, and no comparison against a baseline. Since inferring ISCED ranges is one of the two headline contributions, the absence of any quantitative validation is load-bearing. At minimum, the authors should evaluate on an expert-annotated sample or an established crosswalk and report per-level precision/recall.
  3. [§4.2, Table 7] The ablation experiments begin from official BAfA titles and perturb only gender inflection, management nouns, and word order. They do not simulate the forms that dominate employer-generated free text: abbreviations, compounds, misspellings, dialect, or missing whitespace. Moreover, the affected subsets are small (765, 3,324, and 1,005 titles) and no confidence intervals or multiple-seed results are reported, so the observed 0.6–1.9 pp drops may not be statistically distinguishable. These ablations therefore do not establish robustness to the target distribution and should not be used to support the transfer claim.
minor comments (7)
  1. [Abstract / §1] The phrase 'any free-form German job title' overstates the demonstrated scope; consider softening to 'job titles from the BAfA vocabulary' or adding an external validation.
  2. [§3.3] Typo: 'by the first four digits of the KldB 2019nd' should read 'KldB 2010'.
  3. [§2] The related-work comparison mentions 'Baskaran and Müller' and 'Safikhani et al.' without publication years in the text; please add years for readability.
  4. [§3.3.2 / Table 3] The mapping table lists 'Bamtenausbildung' (typo for 'Beamtenausbildung') and 'helfer-/anlerntätigkeiten' with inconsistent capitalization; these should be corrected.
  5. [§3.4.1] Typo: 'Combing them' should be 'Combining them'. Also, 'Construction Supersivor' and 'Contruction Manager' are misspelled in the example; if these are deliberately simulating noisy input, state so.
  6. [Table 6] The comparison with OJRD is informative, but §4.1 notes OJRD was trained on full job ads. Please make explicit in the table caption that OJRD's input modality differs, to avoid an apples-to-oranges impression.
  7. [§4.1 / Tables 5-7] No standard deviations or confidence intervals are reported for any metric. Given the deterministic k-NN setting, at least a few random seeds or bootstrap intervals would help assess stability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: KldB labels are supervised targets with a held-out test split; limitations concern external validity, not circularity.

full rationale

The paper's pipeline is a standard supervised fine-tuning and retrieval setup. Training triplets are constructed from KldB 2010 codes ('positives share a KldB 2010 code, and negatives have different codes'), but the model never sees test labels during training, and evaluation is performed on a random 20% hold-out ('The data set is divided into 80% for training and 20% for validation and testing'). The ISCED mapping is explicit and rule-based rather than learned from labels, so it is not a fitted-parameter-then-prediction loop. The paper itself concedes the main validity gap: 'Reliable evaluation remains difficult because there is no publicly available, gold-standard set of free-form titles mapped to all 1300 KldB 2010 codes,' and that truly open-ended titles still require expert annotation. That is an external-validity limitation, not a circularity. The one self-citation ([27], Hihn & Braun) supports a peripheral point about batch sizes in contrastive learning and is not load-bearing for any central claim. No self-definitional, fitted-input-called-prediction, or self-citation-chain circularity is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central results rest on 525k official terms, ground-truth KldB labels, and a hand-built ISCED rule set. No new physical or theoretical entities are introduced. The free parameters are mostly standard ML hyperparameters plus the hand-crafted qualification grouping and ISCED rules.

free parameters (6)
  • k_nearest_neighbors = 1
    k in the HNSW majority-vote classifier; authors state k=1 is optimal on the validation split and use it for reported numbers.
  • embedding_dimension = 64
    Matryoshka losses train multiple dimensions; the paper reports the 64-dim embedding as the evaluated configuration.
  • num_positives_and_negatives_per_anchor = 3
    Each anchor gets three randomly selected positives and three negatives, giving about 1.5M triplets.
  • qualification_consolidation_groups = See Table 3
    Hand-mapped grouping of original BERUFENET qualification values into nine simplified categories, used to build the query string.
  • ISCED_rule_mapping = Hand-crafted rules (Table 4)
    Set of if-then rules from KldB requirement level, qualification, job-title keywords, and DQR level to ISCED 2011 codes; not fully specified and not evaluated.
  • baseline_pretrained_model = deutsche-telekom/gbert-large-paraphrase-cosine
    Selected by evaluating several German SentenceTransformer models on a fraction of the data (Section 3.3).
assumptions (5)
  • domain assumption KldB 2010 codes are a valid ground-truth grouping for semantic similarity of job titles.
    Used to define anchor/positive/negative triplets (Section 3.3, Figure 1) and to evaluate classification.
  • domain assumption The BAfA BERUFENET term list is complete and accurate for German occupations.
    The entire training set and evaluation set come from this public dataset (Section 3.1).
  • domain assumption A random split of the official term list yields a test set representative of real-world free-form job titles.
    The reported metrics are computed on this split (Section 4.1); if free-form titles differ, the accuracy does not transfer.
  • domain assumption DQR levels can be mapped to ISCED levels with simple keyword/requirement-level rules.
    Section 3.3.2: ISCED mapping is derived from DQR-based rules, but no independent validation is provided.
  • standard math Fine-tuning a sentence encoder with contrastive triplets makes the embedding space reflect the ontology hierarchy.
    Relies on standard contrastive learning methodology (Multiple Negatives Ranking Loss, Matryoshka Loss), Section 3.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ontology-Aligned Embeddings for Data-Driven Labour Market Analytics." pith.science (2026). https://pith.science/paper/A63KSUVQ

@misc{pith2026250904942,
  author       = {Pith},
  title        = {Pith review of: Ontology-Aligned Embeddings for Data-Driven Labour Market Analytics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A63KSUVQ}},
  note         = {Machine review of arXiv:2509.04942}
}
read the original abstract

The limited ability to reason across occupational data from different sources is a long-standing bottleneck for data-driven labour market analytics. Previous research has relied on hand-crafted ontologies that allow such reasoning but are computationally expensive and require careful maintenance by human experts. The rise of language processing machine learning models offers a scalable alternative by learning shared semantic spaces that bridge diverse occupational vocabularies without extensive human curation. We present an embedding-based alignment process that links any free-form German job title to two established ontologies - the German Klassifikation der Berufe and the International Standard Classification of Education. Using publicly available data from the German Federal Employment Agency, we construct a dataset to fine-tune a Sentence-BERT model to learn the structure imposed by the ontologies. The enriched pairs (job title, embedding) define a similarity graph structure that we can use for efficient approximate nearest-neighbour search, allowing us to frame the classification process as a semantic search problem. This allows for greater flexibility, e.g., adding more classes. We discuss design decisions, open challenges, and outline ongoing work on extending the graph with other ontologies and multilingual titles.

Figures

Figures reproduced from arXiv: 2509.04942 by the authors.

Figure 1
Figure 1. We collect publicly available data from the Bundesagentur für Arbeit, enrich it, build a training data set, and fine-tune an embedding model. Anchor-Positive-Negative triplets are determined by matching KldB 2010 codes. In this example, we link to the anchor with KldB 62422 another job title from the KldB subgroup 6262 (“sales occupations (retail) selling medical supplies and healthcare goods”) as the positive and w… view at source ↗
Figure 2
Figure 2. 2D -projection using t-SNE [36] of the embedding of the test dataset learned by our model, showing the structure it has learned for KldB 2010 areas (left) and requirement levels (right). We can see clusters evolving - especially in the requirement level groupings. to a 𝑘-NN classifier fit on a variety of pre-trained models on occupational data. We acknowledge that the comparison may not be fair, as only our model ha… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages

  1. [1]

    Christoph, B

    B. Christoph, B. Matthes, C. Ebner, Occupation-based measures—an overview and discus- sion, Kölner Zeitschrift für Soziologie und Sozialpsychologie 72 (2020) 41–78. doi: 10.1007/ s11577-020-00673-4

    work page 2020

  2. [2]

    Bundesagentur für Arbeit, Klassifikation der Berufe 2010 - überarbeitete Fassung 2020 Band 1: Systematischer und alphabetischer Teil mit Erläuterungen, Bundesagentur für Arbeit, Nürnberg, 2021

    work page 2010

  3. [3]

    Dörpinghaus, J

    J. Dörpinghaus, J. Binnewitt, S. Winnige, K. Hein, K. Krüger, Towards a German labor market ontology: Challenges and applications, Applied ontology 18 (2023) 343–365. doi: 10.3233/ ao-230027

    work page 2023

  4. [4]

    URL: https://www.dqr.de/dqr/shareddocs/downloads/media/content/dqr_handbuch_01_08_2013.pdf? __blob=publicationFile&v=2

    Handbuch zum Deutschen Qualifikationsrahmen, Bund-Länder-Koordinierungsstelle für den Deutschen Qualifikationsrahmen für lebenslanges Lernen, Berlin, 2013. URL: https://www.dqr.de/dqr/shareddocs/downloads/media/content/dqr_handbuch_01_08_2013.pdf? __blob=publicationFile&v=2

    work page 2013

  5. [5]

    URL: https://www.dqr.de/dqr/shareddocs/downloads/media/content/2024_dqr_liste_ zugeordnete_qualifik_01082024.pdf?__blob=publicationFile&v=2

    Deutscher Qualifikationsrahmen für lebenslanges Lernen: Liste der zugeordneten Qualifikationen, Bund-Länder-Koordinierungsstelle für den Deutschen Qualifikationsrahmen für lebenslanges Lernen, 2024. URL: https://www.dqr.de/dqr/shareddocs/downloads/media/content/2024_dqr_liste_ zugeordnete_qualifik_01082024.pdf?__blob=publicationFile&v=2

    work page 2024

  6. [6]

    Méhaut, C

    P. Méhaut, C. Winch, The european qualification framework: skills, competences or knowledge?, European educational research journal 11 (2012) 369–381

    work page 2012

  7. [7]

    doi:10.15220/978-92-9189-123-8-en

    UNESCO Institute for Statistics, International Standard Classification of Education (ISCED) 2011, UNESCO Institute for Statistics, 2012. doi:10.15220/978-92-9189-123-8-en

    work page doi:10.15220/978-92-9189-123-8-en 2011

  8. [8]

    Zhang, J

    D. Zhang, J. Liu, H. Zhu, Y. Liu, L. Wang, P. Wang, H. Xiong, Job2Vec: Job title benchmarking with collective multi-view representation learning, in: Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM ’19), ACM, Beijing, China, 2019, pp. 2763–2772. URL: https://doi.org/10.1145/3357384.3357825. doi:10.1145/33...

    arXiv 2019

Show all 40 references
  1. [9]

    J. Liu, Y. C. Ng, Z. Gui, T. Singhal, L. T. M. Blessing, K. L. Wood, K. H. Lim, Title2Vec: A contextual job title embedding for occupational named entity recognition and other applications, Journal of Big Data 9 (2022). URL: https://journalofbigdata.springeropen.com/articles/1...

    2022 doi

  2. [10]

    Reimers, I

    N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks, in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2019, p...

    2019

  3. [11]

    A. Khan, T. Wu, X. Chen, LLM+KG@VLDB’24 Workshop Summary: Data Management Opportu- nities in Unifying Large Language Models and Knowledge Graphs, in: Proceedings of the VLDB 2024 Workshops, 2024. URL: https://arxiv.org/abs/2410.01978, arXiv:2410.01978

    2024 arXiv

  4. [12]

    R. S. Dehal, M. Sharma, E. Rajabi, Knowledge Graphs and Their Reciprocal Relationship with Large Language Models, Machine Learning and Knowledge Extraction 7 (2025) 38. URL: https: //doi.org/10.3390/make7020038. doi:10.3390/make7020038

    2025 doi

  5. [13]

    URL: https://www.thestepstonegroup.com/english/, accessed 2025-06-11

    The Stepstone Group, The stepstone group — we help companies to hire the right talent, 2025. URL: https://www.thestepstonegroup.com/english/, accessed 2025-06-11

    2025

  6. [14]

    Dawson, M.-A

    N. Dawson, M.-A. Williams, M.-A. Rizoiu, Skill-driven recommendations for job transition path- ways, PloS one 16 (2021) e0254722. doi:10.1371/journal.pone.0254722

    2021 doi

  7. [15]

    Clemens, C

    T. Clemens, C. Dibben, A method for estimating wage, using standardised occupational classifi- cations, for use in medical research in the place of self-reported income, BMC medical research methodology 14 (2014) 59. doi:10.1186/1471-2288-14-59

    2014 doi

  8. [16]

    R. Zbib, L. A. Lacasa, F. Retyk, R. Poves, J. Aizpuru, H. Fabregat, V. Simkus, E. García-Casademont, Learning job titles similarity from noisy skill labels, arXiv [cs.IR] (2022).arXiv:2207.00494

    2022 arXiv

  9. [17]

    Brunello, P

    G. Brunello, P. Wruuck, Skill shortages and skill mismatch: A review of the literature, Journal of Economic Surveys 35 (2021) 1145–1167. doi:10.1111/joes.12424

    2021 doi

  10. [18]

    Baskaran, J

    R. Baskaran, J. Müller, Classification of German Job Titles in Online Job Postings Using the KldB 2010 Taxonomy, Technical Report, &effect data solutions GmbH, 2023. URL: https://www. and-effect.com/publications/2022-11-21_technical_report_kldb.pdf, version 1, last updated 07 Nov 2023

    2010

  11. [19]

    Safikhani, H

    P. Safikhani, H. Avetisyan, D. Föste-Eggers, D. Broneske, Automated occupation coding with hierarchical features: A data-centric approach to classification with pre-trained language models, Discover Artificial Intelligence 3 (2023). doi:10.1007/s44163-023-00050-y

    2023 doi

  12. [21]

    de Groot, J

    M. de Groot, J. Schutte, D. Graus, Job posting-enriched knowledge graph for skills-based matching, arXiv preprint arXiv:2109.02554 (2021). URL: https://arxiv.org/abs/2109.02554

    2021 arXiv

  13. [22]

    A. Seif, S. Toh, H. K. Lee, A dynamic jobs–skills knowledge graph, in: Proceedings of the 4th Workshop on Recommender Systems for Human Resources (RecSys in HR 2024), CEUR-WS, 2024. URL: https://recsyshr.aau.dk/wp-content/uploads/2024/10/RecSysHR2024-paper_1.pdf

    2024

  14. [23]

    D. T. A. May, Philipp, German bert large paraphrase cosine, 2025. URL: https://huggingface.co/ deutsche-telekom/gbert-large-paraphrase-cosine, huggingFace repository, accessed 2025-12-06

    2025

  15. [24]

    Jaiswal, A

    A. Jaiswal, A. R. Babu, M. Z. Zadeh, D. Banerjee, F. Makedon, A survey on contrastive self- supervised learning, Technologies 9 (2020) 2

    2020

  16. [25]

    Rojas, A

    E. Rojas, A. N. Kahira, E. Meneses, L. B. Gomez, R. M. Badia, A study of checkpointing in large scale training of deep neural networks, arXiv preprint arXiv:2012.00825 (2020)

    2012 arXiv

  17. [26]

    Soydaner, A comparison of optimization algorithms for deep learning, International Journal of Pattern Recognition and Artificial Intelligence 34 (2020) 2052013

    D. Soydaner, A comparison of optimization algorithms for deep learning, International Journal of Pattern Recognition and Artificial Intelligence 34 (2020) 2052013

    2020

  18. [27]

    H. Hihn, D. A. Braun, Online continual learning through unsupervised mutual information maximization, Neurocomputing 578 (2024) 127422

    2024

  19. [28]

    Henderson, R

    M. Henderson, R. Al-Rfou, B. Strope, Y.-H. Sung, L. Lukács, R. Guo, S. Kumar, B. Miklos, R. Kurzweil, Efficient natural language response suggestion for smart reply, arXiv preprint arXiv:1705.00652 (2017)

    2017 arXiv

  20. [29]

    Kusupati, G

    A. Kusupati, G. Bhatt, A. Rege, M. Wallingford, A. Sinha, V. Ramanujan, W. Howard-Snyder, K. Chen, S. Kakade, P. Jain, et al., Matryoshka representation learning, Advances in Neural Information Processing Systems 35 (2022) 30233–30249

    2022

  21. [30]

    S. L. Schneider, The international standard classification of education 2011, in: Class and stratification analysis, volume 30, Emerald Group Publishing Limited, 2013, pp. 365–379

    2011

  22. [31]

    Fischer, BERUFENET.API: Dokumentation zur berufenet-api der bundesagentur für arbeit,

    A. Fischer, BERUFENET.API: Dokumentation zur berufenet-api der bundesagentur für arbeit,

  23. [32]

    Y. A. Malkov, D. A. Yashunin, Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs, IEEE transactions on pattern analysis and machine intelligence 42 (2018) 824–836

    2018

  24. [33]

    Y. A. Malkov, D. A. Yashunin, hnswlib: Header-only C++/python library for fast approximate nearest neighbors, 2023. URL: https://github.com/nmslib/hnswlib, gitHub repository, accessed 2025-06-11

    2023

  25. [34]

    Sparck Jones, A statistical interpretation of term specificity and its application in retrieval, Journal of Documentation 28 (1972) 11–21

    K. Sparck Jones, A statistical interpretation of term specificity and its application in retrieval, Journal of Documentation 28 (1972) 11–21. URL: https://doi.org/10.1108/eb026526. doi:10.1108/ eb026526

    1972 doi

  26. [35]

    Krüger, Using explainable ai for robustness checks in requirement level classification for german online job advertisements, Springer Special Issue (forthcoming)

    K. Krüger, Using explainable ai for robustness checks in requirement level classification for german online job advertisements, Springer Special Issue (forthcoming). In press; publication details pending

  27. [36]

    van der Maaten, G

    L. van der Maaten, G. Hinton, Visualizing data using t-SNE, Journal of Machine Learning Research 9 (2008) 2579–2605. URL: http://www.jmlr.org/papers/v9/vandermaaten08a.html

    2008

  28. [37]

    Rosenberger, L

    J. Rosenberger, L. Wolfrum, S. Weinzierl, M. Kraus, P. Zschech, CareerBERT: Matching resumes to ESCO jobs in a shared embedding space for generic job recommendations (forthcoming), Expert Systems With Applications (2025)

    2025

  29. [38]

    Decorte, J

    J. Decorte, J. V. Hautte, T. Demeester, C. Develder, Jobbert: Understanding job titles through skills, CoRR abs/2109.09605 (2021). URL: https://arxiv.org/abs/2109.09605.arXiv:2109.09605

    2021 arXiv

  30. [39]

    Varga, K

    R. Varga, K. Reuter, H.-J. Bürger, S. Broeckx, ESCO: Towards a Semantic Web for the European Labor Market, in: Proceedings of the 1st Workshop on Open Data and Knowledge Graphs in Practice (LDK 2014), volume 1268 ofCEUR Workshop Proceedings, CEUR-WS.org, 2014. URL: http://ceur...

    2014

  31. [40]

    Krüger, Kkrueger/oja_reqlevel_de, 2025

    K. Krüger, Kkrueger/oja_reqlevel_de, 2025. URL: https://huggingface.co/KKrueger/oja_reqlevel_de, huggingFace repository, accessed 2025-22-06

    2025

  32. [2025]

    URL: https://github.com/AndreasFischer1985/berufenet-api, gitHub repository, accessed 2025-12-06

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.