Pith. sign in

REVIEW 3 major objections 5 minor 68 references

Towards Explainable Job Title Matching: Leveraging Semantic Textual Relatedness and Knowledge Graphs

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that adding a skill-based knowledge graph to fine-tuned sentence embeddings produces the most accurate similarity scores for highly related job title pairs, and that this combination also makes matches explainable by surfac

desk verdict Useful stratified evaluation of job-title STR, but the headline gain is measured against SBERT-generated labels, so treat the practical claim as conditional. read the letter →

arxiv 2509.09522 v1 pith:IDGMJQP5 submitted 2025-09-11 cs.CL cs.AI

Vadim Zadykian , Bruno Andrade , Haithem Afli This is my paper
classification cs.CLcs.AI
keywords semantictextualrelatednessjobtitlematchingknowledgegraphself-supervisedlearningsentenceembeddingsneuralnetworksstratifiedevaluationexplainability
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 claims that a self-supervised hybrid of sentence embeddings and a job-skill knowledge graph yields more precise job-title relatedness scores, especially for highly similar titles, and that the improvement is visible only when results are examined separately by relatedness region. On their validation set, the hybrid configuration reaches a high-relatedness RMSE of 0.11, compared with 0.18 for the strongest text-only fine-tuned baselines, a gain the authors attribute to the graph's explicit job-skill structure. The work also argues that training and evaluating with stratified STR regions—low, medium, and high—reveals strengths and weaknesses that global metrics hide. This matters because HR recommender systems need to distinguish among near-duplicate job titles, not just rank broadly, and need explanations to justify recommendations.

What carries the argument

The central machinery is a self-supervised pipeline (Algorithm 1) that: (1) summarizes job descriptions with BART; (2) encodes them with SBERT to compute cosine-similarity relatedness labels; (3) builds a bipartite job-skill knowledge graph from top semantically matched skills; (4) learns node embeddings on that graph with a relational graph convolutional network (R-GCN); and (5) trains a lightweight MLP to project SBERT text embeddings into the graph embedding space. At inference, job-title relatedness is the cosine similarity of the projected graph embeddings. The analytic lens that exposes the KG's advantage is stratified evaluation, partitioning scores into low (0.0–0.5), medium (0.5–0.7

What would settle it

Collect a set of job-title pairs with human-rated relatedness scores and re-run the comparison: if MPNET+RGCN's high-STR RMSE advantage over MPNET-F does not persist against human labels, the central claim is falsified. A more surgical test would randomize the job-skill edges in the knowledge graph and retrain; if the high-STR RMSE stays at 0.11, the graph structure itself is not the cause.

Watch

Extended reading notes

Core claim

The paper's central claim is that a system which aligns fine-tuned Sentence-BERT (SBERT) embeddings with a domain-specific job-skill knowledge graph is the most precise way to score highly related job title pairs. On a validation set of job-title pairs, the MPNET+RGCN configuration achieves a high-STR RMSE of 0.11, versus 0.18 for the best fine-tuned baselines (MPNET-F and JOBBERT-F). The authors attribute this gain to the knowledge graph's explicit job-skill edges, which supply relational structure that text-only encoders miss, and to the learned mapping that projects text embeddings into the graph embedding space. They further claim that the knowledge graph makes matches explainable by rev

Load-bearing premise

The load-bearing premise is that SBERT-computed cosine similarities between job description summaries are a valid proxy for human-perceived job-title relatedness; if that proxy fails, the reported RMSE gains may not transfer to real HR systems.

Editorial extensions

If this is right

  • In re-ranking stages where irrelevant jobs have already been filtered, the high-STR improvement translates to better ordering among near-equivalent job titles.
  • Systems can choose different models per STR region: a general-purpose pretrained encoder for coarse filtering in the low range and a graph-augmented fine-tuned model for fine-grained ranking in the high range.
  • Skill specificity scores attached to graph explanations let recruiters see whether a match rests on meaningful shared skills or only generic ones, supporting auditability of high-risk AI systems.
  • The same self-supervised recipe can be transplanted to other STR tasks, such as product matching or legal case retrieval, wherever a skill or type ontology exists.

Reading between the lines

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

  • Because the training labels and the knowledge-graph edges are both generated from SBERT cosine similarities, the reported error reductions could partly reflect self-confirmation within one embedding family; a human-annotation benchmark is the natural check.
  • The graph's benefit may be largest for rare or idiosyncratic job titles, where text embeddings have sparse neighbors; this could be tested by slicing high-STR performance by title frequency.
  • A deployment concern: the MLP projection into graph space might compress score distributions toward the high range, so calibration curves on held-out pairs should accompany adoption.
  • If graph structure is what matters, replacing R-GCN with lighter graph encoders or even hand-built skill-overlap features should preserve most of the high-STR gain—a cheap ablation to run.
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 / 5 minor

Summary. The manuscript proposes a self-supervised pipeline (Algorithm 1) for job-title Semantic Textual Relatedness (STR) matching. BART summarizes job descriptions; SBERT encodes them; cosine similarities generate training and evaluation STR labels; a job-skill knowledge graph is built from the same SBERT similarities; an R-GCN learns graph embeddings; and an MLP maps fine-tuned SBERT embeddings into the graph space. Evaluation stratifies the STR score range into low, medium, and high regions and reports RMSE (Table 2). The central empirical claim is that KG-augmented fine-tuned MPNET (MPNET+RGCN) reduces high-STR RMSE by about 25% over strong baselines.

Significance. The stratified evaluation framework is a useful corrective to global RMSE, and the availability of code/data plus the region-wise t-tests are strengths. If the high-STR gain survives evaluation against human-judged relatedness, KG augmentation would be a practical contribution for HR recommender systems. However, because the labels, KG edges, and evaluation set all derive from the same SBERT model, the current evidence does not establish the central practical claim. The paper itself acknowledges weak-supervision label noise only as future work, but this issue is load-bearing and needs to be addressed directly.

major comments (3)
  1. [Section 3.2, Algorithm 1 (Steps 2, 3, 5, 6); Section 3.6] The evaluation is self-referential. Every STR label is generated as cosine similarity of SBERT embeddings on BART summaries (Step 2), and the KG edges for job-skill relations are selected from the same SBERT cosine similarities (Step 5). The validation set (Step 3) is drawn from this same generated label space, with only job-title disjointness. MPNET+RGCN therefore has direct access to the geometry of the teacher that produced the labels, and its high-STR RMSE advantage in Table 2 may reflect reconstructing that geometry rather than modeling human-perceived relatedness. The Future Work concession about weak supervision does not address this core validity threat. Please add human-rated or externally validated job-title STR labels (e.g., a small expert-annotated set, or O*NET/ESCO-derived gold pairs) and report the same stratified RMSE.
  2. [Section 4, Table 2] The abstract's '25% reduction' is not precisely reproduced by Table 2. High-STR RMSE for MPNET+RGCN is 0.11, versus 0.15 for JOBBERT (a 27% reduction) and 0.18 for MPNET-F and JOBBERT-F (39%). The paper should specify which baseline yields 25%, or revise the claim. Additionally, no confidence intervals or model-to-model significance tests are reported. The paired t-tests in Table 3 compare regions within a model, not MPNET+RGCN against MPNET-F/JOBBERT-F on the same evaluation pairs. Without such comparisons, the 0.11 versus 0.15 difference cannot be distinguished from noise.
  3. [Section 3.2/3.3, MPNET-F vs MPNET+RGCN] The comparison between MPNET-F and MPNET+RGCN changes several components at once: fine-tuned SBERT plus R-GCN plus MLP alignment. The paper attributes the improvement to KG integration, but no ablation controls for the extra MLP capacity, the graph-embedding training objective, or the choice of R-GCN. Please add ablations (e.g., MPNET-F + MLP without KG, MPNET+RGCN with randomly rewired KG edges, and a second graph model such as ComplEx) to isolate the contribution of structured skill knowledge. With only one graph model tested, the causal claim that KG augmentation produces the improvement is not yet supported.
minor comments (5)
  1. [Abstract and Section 1] Typographical issues: 'explainabilityis' in Section 1 is missing a space; the author-footer contains a corrupted string '/envel⌢pe-⌢penvadim.zadykian@mymtu.ie'. Please fix.
  2. [Table 3] The model name is written as 'MPNET-RGCN' in Table 3, while everywhere else it is 'MPNET+RGCN'. Please unify the notation.
  3. [Section 4.2] The high-STR example 'Senior Performance and Project Analyst' vs. 'Director, eCommerce & Retail' is not obviously highly related. If this pair comes from the auto-generated labels, it further illustrates the label-validity concern and should be justified or replaced with a more representative example.
  4. [Algorithm 1, Step 7; Section 3.4.1] Clarify the exact inference-time composition: is the MLP applied to the fine-tuned SBERT embedding before cosine similarity is computed? Table 1 lists vectorizer sizes 500 versus 768 without specifying which space the final cosine is computed in.
  5. [Figure 1] Report the actual distribution of the training STR scores and add axis labels/units. Currently the figure is referenced but its content is not described in the text.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the KG-augmented model's high-STR gain is an empirical result on a self-generated benchmark, not a constructional identity.

full rationale

The paper's pipeline is self-supervised: Algorithm 1 Step 2 generates STR labels as cosine similarities of SBERT embeddings on BART-summarized job descriptions; Step 5 selects job–skill edges using the same kind of cosine similarities; Step 6 learns RGCN node embeddings; Step 7 trains a neural network to map fine-tuned SBERT embeddings into that graph space; inference computes cosine in graph space. Although the target labels and the KG edges both derive from SBERT cosine geometry, the predicted STR is a learned composition (fine-tuned SBERT → MLP → graph embedding → cosine) that is not equal by construction to the label-generating cosine. The train/eval split uses disjoint job titles (Section 3.3), so the reported RMSE values are genuine held-out numbers, and the improvement of MPNET+RGCN in the high-STR region is an empirical result, not a forced reduction. The main caveat is that the ground-truth labels are machine-generated from SBERT, so the results measure fidelity to SBERT's similarity geometry rather than human-perceived relatedness; the paper openly acknowledges this weak-supervision limitation in Section 4.4. The only self-citation is the GitHub repository [56] for code/data, which is not load-bearing. Hence no meaningful circularity, though the external-validity threat should be noted as a correctness risk rather than a circularity.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The paper relies on a self-generated label space (SBERT cosine similarities) as ground truth, and on pretrained models and taxonomies as inputs. The free parameters listed above are hand-set thresholds and hyperparameters that directly shape the graph and the evaluation regions. No new physical or conceptual entities are invented.

free parameters (8)
  • STR region boundaries = 0.5 and 0.75
    Hand-chosen thresholds to define low, medium, and high relatedness zones (Table 4); the paper says they are guided by domain expertise, but no sensitivity analysis is performed.
  • Job-skill STR threshold = 0.5
    Minimum cosine similarity for a skill to be considered related to a job (Table 4); controls graph connectivity.
  • Skill-skill STR threshold = 0.25
    Minimum cosine similarity for a child skill to be connected to a parent skill (Table 4).
  • Number of skills per job = 10
    Limits the maximum number of skills assigned to a job (Table 4).
  • Generic-skill pruning threshold = 20% job share
    Skills appearing in more than 20% of jobs are removed from the graph (Section 3.4.2).
  • Text epochs = 5
    SBERT fine-tuning epochs (Table 4).
  • Graph epochs = 15
    R-GCN training epochs (Table 4).
  • KG embedding dimension = 500
    Dimensionality of the MPNET+RGCN output shown in Table 1.
assumptions (5)
  • domain assumption Cosine similarity between SBERT embeddings is a valid proxy for semantic textual relatedness between job descriptions and job titles.
    Used to generate all training and evaluation labels (Algorithm 1, Steps 2 and 5) and to compute the final STR score.
  • domain assumption BART summarization preserves the functional content of job descriptions.
    Step 0 of Algorithm 1 relies on BART summaries as input to the embedding step.
  • domain assumption ESCO and Indeed skill taxonomies provide a complete enough set of job-skill relationships for the graph.
    Section 3.6 states the graph is built from these sources; the paper acknowledges that domains without formalized ontologies may not work (Section 4.4).
  • domain assumption R-GCN node embeddings, combined with a learned MLP projection, capture semantic relatedness in a way that supports cosine comparison.
    Section 3.4.1 assumes the learned mapping from SBERT space to KG space preserves relatedness.
  • domain assumption The disjoint-title train/eval split prevents leakage of job titles between training and evaluation.
    Section 3.3 states the validation set contains job titles not seen in training; this is an assumption about the data generation process.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Explainable Job Title Matching: Leveraging Semantic Textual Relatedness and Knowledge Graphs." pith.science (2026). https://pith.science/paper/IDGMJQP5

@misc{pith2026250909522,
  author       = {Pith},
  title        = {Pith review of: Towards Explainable Job Title Matching: Leveraging Semantic Textual Relatedness and Knowledge Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IDGMJQP5}},
  note         = {Machine review of arXiv:2509.09522}
}
read the original abstract

Semantic Textual Relatedness (STR) captures nuanced relationships between texts that extend beyond superficial lexical similarity. In this study, we investigate STR in the context of job title matching - a key challenge in resume recommendation systems, where overlapping terms are often limited or misleading. We introduce a self-supervised hybrid architecture that combines dense sentence embeddings with domain-specific Knowledge Graphs (KGs) to improve both semantic alignment and explainability. Unlike previous work that evaluated models on aggregate performance, our approach emphasizes data stratification by partitioning the STR score continuum into distinct regions: low, medium, and high semantic relatedness. This stratified evaluation enables a fine-grained analysis of model performance across semantically meaningful subspaces. We evaluate several embedding models, both with and without KG integration via graph neural networks. The results show that fine-tuned SBERT models augmented with KGs produce consistent improvements in the high-STR region, where the RMSE is reduced by 25% over strong baselines. Our findings highlight not only the benefits of combining KGs with text embeddings, but also the importance of regional performance analysis in understanding model behavior. This granular approach reveals strengths and weaknesses hidden by global metrics, and supports more targeted model selection for use in Human Resources (HR) systems and applications where fairness, explainability, and contextual matching are essential.

Figures

Figures reproduced from arXiv: 2509.09522 by the authors.

Figure 1
Figure 1. Distribution of STR in Training dataset 3.6. Data The raw data are obtained from several open-source col￾lections: a Kaggle dataset [57], a list of granular skills and competencies by ESCO [58], and a list of broad job functions Indeed job site [59] [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 4
Figure 4. Distribution of Prediction Errors for JOBBERT-F [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Distribution of Prediction Errors for MPNET 4.1.4. MPNET-F MPNET-F shows strong, significant differences across all region pairs, with particularly large negative t-values for low–medium and low–high comparisons. However, it shows no significant difference between medium and high STR, suggesting more consistent performance at the higher end of the similarity spectrum ( [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Distribution of Prediction Errors for JOBBERT 4.1.2. JOBBERT-F JOBBERT-F exhibits fewer significant differences: low–medium and low–high contrasts are significant, but medium–high differences are not. This stability in medium and high STR regions may reflect the benefi…
Figure 6
Figure 6. Figure 6: Distribution of Prediction Errors for MPNET-F [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 8
Figure 8. Figure 8: Explanation Graph - Good Job Title match. The skill specificity (shown in brackets) is the inverse of the centrality degree [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: Explanation Graph - Poor Job Title match. The skill specificity (shown in brackets) is the inverse of the centrality degree job pairs (e.g., in deduplication or anomaly detection tasks), performance in the low STR range is of greater interest ( [PITH_FULL_IMAGE:figure…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 8 canonical work pages

  1. [1]

    Chandrasekaran, V

    D. Chandrasekaran, V. Mago, Evolution of seman- tic similarity—a survey, ACM Comput. Surv. 54 (2021). URL: https://doi-org.mtu.idm.oclc.org/10.1145/ 3440755. doi:10.1145/3440755

    work page doi:10.1145/3440755 2021

  2. [2]

    J. Chen, Q. Huang, C. Wang, C. Li, Sensemap: Urban performance visualization and analytics via semantic textual similarity, IEEE Transactions on Visualization and Computer Graphics 30 (2024) 6275–6290. doi:10. 1109/TVCG.2023.3333356

    arXiv 2024

  3. [3]

    R. Gaur, A. Dwivedi, Knowledge graph-based eval- uation metric for conversational ai systems: A step towards quantifying semantic textual similarity, in: S. Dhar, S. Goswami, U. Dinesh Kumar, I. Bose, R. Dubey, C. Mazumdar (Eds.), AGC 2023, Springer Nature Switzerland, Cham, 2024, pp. 112–124

    2023

  4. [4]

    D. A. Cruse, Lexical semantics, Cambridge university press, 1986

    1986

  5. [5]

    Webb, Word families and lemmas, not a real dilemma: Investigating lexical units, Studies in Sec- ond Language Acquisition 43 (2021) 973–984

    S. Webb, Word families and lemmas, not a real dilemma: Investigating lexical units, Studies in Sec- ond Language Acquisition 43 (2021) 973–984. doi:10. 1017/S0272263121000760

    2021

  6. [6]

    Zikánová, E

    Š. Zikánová, E. Hajičová, B. Vidová-Hladká, P. Jínová, J. Mírovsk `y, A. Nědolužko, L. Poláková, K. Rysová, M. Rysová, J. Václ, Discourse and coherence: from the sentence structure to relations in text, Ústav formální a aplikované lingvistiky, 2015

    2015

  7. [7]

    Mikolov, K

    T. Mikolov, K. Chen, G. Corrado, J. Dean, Efficient estimation of word representations in vector space, arXiv preprint arXiv:1301.3781 (2013). URL: https:// arxiv.org/pdf/1301.3781

    arXiv 2013

  8. [8]

    Devlin, M

    J. Devlin, M. Chang, K. Lee, K. Toutanova, BERT: pre-training of deep bidirectional transformers for language understanding, in: J. Burstein, C. Doran, T. Solorio (Eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, NAACL-HLT 2019, Minneapolis, MN, USA, Ju...

    doi:10.18653/v 2019

Show all 68 references
  1. [9]

    Reimers, I

    N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks, arXiv preprint arXiv:1908.10084 (2019). doi:10.18653/v1/ d19-1410

    1908 arXiv

  2. [10]

    M. A. Abdalla, K. Vishnubhotla, S. M. Mohammad, What makes sentences semantically related: a textual relatedness dataset and empirical study (2021). doi:10. 48550/arxiv.2110.04845

    2021 doi arXiv

  3. [11]

    Sloan, Relational ambiguity between sentences, College Composition and Communication 39 (1988) 154–165

    G. Sloan, Relational ambiguity between sentences, College Composition and Communication 39 (1988) 154–165. URL: http://www.jstor.org/stable/358025

    1988

  4. [12]

    Miyabe, H

    Y. Miyabe, H. Takamura, M. Okumura, Identifying cross-document relations between sentences, in: Pro- ceedings of the Third International Joint Conference on Natural Language Processing: Volume-I, 2008

    2008

  5. [13]

    Lombardo, A

    P. Lombardo, A. Boiardi, L. Colombo, A. Schiavone, N. Tamagnone, Top-rank-focused adaptive vote col- lection for the evaluation of domain-specific semantic models, in: B. Webber, T. Cohn, Y. He, Y. Liu (Eds.), Pro- ceedings of the 2020 Conference on Empirical Methods in Natura...

    2020

  6. [14]

    M. Ye, Z. Tang, J. Xu, L. Jin, Recommender system for e-learning based on semantic relatedness of con- cepts, Information 6 (2015) 443–453. doi: 10.3390/ info6030443

    2015

  7. [15]

    Likavec, F

    S. Likavec, F. Osborne, F. Cena, Property-based seman- tic similarity and relatedness for improving recommen- dation accuracy and diversity, International Journal on Semantic Web and Information Systems (IJSWIS) 11 (2015) 1–40. doi:10.4018/ijswis.2015100101

    2015 doi

  8. [16]

    Natarajan, S

    S. Natarajan, S. Vairavasundaram, K. Kotecha, V. In- dragandhi, S. Palani, J. R. Saini, L. Ravi, Cd-semmf: Cross-domain semantic relatedness based matrix fac- torization model enabled with linked open data for user cold start issue, IEEE Access 10 (2022) 52955– 52970. doi:10.1...

    2022

  9. [17]

    Rahhal, K

    I. Rahhal, K. M. Carley, I. Kassou, M. Ghogho, Two stage job title identification system for online job ad- vertisements, IEEE Access 11 (2023) 19073–19092. doi:10.1109/ACCESS.2023.3247866

    2023

  10. [18]

    Tanberk, S

    S. Tanberk, S. S. Helli, E. Kesim, S. N. Cavsak, Resume matching framework via ranking and sorting using nlp and deep learning, in: 2023 8th International Confer- ence on Computer Science and Engineering (UBMK), 2023, pp. 453–458. doi: 10.1109/UBMK59864.2023. 10286605

    2023

  11. [19]

    S. A. Pias, M. Hossain, H. Rahman, M. M. Hossain, Enhancing job matching through natural language processing: A bert-based approach, in: 2024 Inter- national Conference on Innovations in Science, En- gineering and Technology (ICISET), 2024, pp. 1–6. doi:10.1109/ICISET62123.202...

    2024

  12. [20]

    M. Kaya, T. Bogers, An exploration of sentence- pair classification for algorithmic recruiting, in: Proceedings of the 17th ACM Confer- ence on Recommender Systems, RecSys ’23, Association for Computing Machinery, New York, NY, USA, 2023, p. 1175–1179. URL: https: //doi-org.mt...

    2023

  13. [21]

    Rezaeipourfarsangi, E

    S. Rezaeipourfarsangi, E. E. Milios, Ai-powered resume-job matching: A document ranking approach using deep neural networks, in: Proceedings of the ACM Symposium on Document Engineering 2023, DocEng ’23, Association for Computing Machinery, New York, NY, USA, 2023. URL: https:...

    2023

  14. [22]

    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, Expert Sys- tems with Applications (2025) 127043. URL: https://www.proquest.com/scholarly-journals/ careerbert...

    2025

  15. [23]

    G. L, R. M L, G. H B, K. Mathada, M. B, Intelligent resume scrutiny using named entity recognition with bert, in: 2023 International Conference on Data Sci- ence and Network Security (ICDSNS), 2023, pp. 01–08. doi:10.1109/ICDSNS58469.2023.10245304

    2023

  16. [24]

    Dhobale, V

    P. Dhobale, V. Bhoir, A. Vyavhare, P. Yelkar, P. A. Dharmadhikari, Resumatcher: An intelligent resume ranking system, in: 2025 3rd International Confer- ence on Intelligent Data Communication Technologies and Internet of Things (IDCIoT), 2025, pp. 1778–1783. doi:10.1109/IDCIOT...

    2025

  17. [25]

    J. Tang, H. Chen, Z. Chen, J. Pan, Y. He, J. Zhao, Z. Li, A person-job matching method based on bm25 and pre-trained language model, in: Proceedings of the 2023 6th International Conference on Machine Learning and Natural Language Processing, MLNLP ’23, Association for Computi...

    2023

  18. [26]

    Ramyar, G

    R. Ramyar, G. Nagarani, S. Natarajan, Deep learn- ing based approach to streamline resume categoriza- tion and ranking, in: 2024 International Confer- ence on IoT Based Control Networks and Intelligent Systems (ICICNIS), 2024, pp. 840–845. doi:10.1109/ ICICNIS64247.2024.10823187

    2024

  19. [27]

    M. A. Aleisa, N. Beloff, M. White, Implement- ing airm: a new ai recruiting model for the saudi arabia labour market, Journal of Innova- tion and Entrepreneurship 12 (2023) 59. URL: https://www.proquest.com/scholarly-journals/ implementing-airm-new-ai-recruiting-model-saudi/ d...

    2023

  20. [28]

    M. Zhao, P. Dufter, Y. Yaghoobzadeh, H. Schütze, Quan- tifying the contextualization of word representations with semantic class probing, Findings of the Asso- ciation for Computational Linguistics: EMNLP 2020 (2020). doi:10.18653/v1/2020.findings-emnlp. 109

    2020 doi

  21. [29]

    Misra, A

    K. Misra, A. Ettinger, J. T. Rayz, Exploring bert’s sensitivity to lexical cues using tests from semantic priming, Findings of the Association for Compu- tational Linguistics: EMNLP 2020 (2020) 4625–4635. doi:10.18653/v1/2020.findings-emnlp.415

    2020 doi

  22. [30]

    A. Zou, Z. Zhang, H. Zhao, Eventbert: incorporat- ing event-based semantics for natural language under- standing, Lecture Notes in Computer Science (2022) 66–80. doi:10.1007/978-3-031-18315-7_5

    2022 doi

  23. [31]

    Hammami, S

    A. Hammami, S. B. Abdallah Ben Lamine, H. Baazaoui, Bert-based semantic relations extraction from large- scale medical datasets, in: 2024 IEEE International Conference on Big Data (BigData), 2024, pp. 6460–6468. doi:10.1109/BigData62323.2024.10825162

    2024

  24. [32]

    1–157, 2024

    Regulation (eu) 2024/1689 of the european parliament and of the council of 12 july 2024 laying down har- monised rules on artificial intelligence (artificial intel- ligence act), Official Journal of the European Union, L 168, 12 July 2024, p. 1–157, 2024. URL: https:// artific...

    2024

  25. [33]

    H. Min, B. Yang, D. G. Allen, A. A. Grandey, M. Liu, Wisdom from the crowd: can recommender systems predict employee turnover and its destinations?, Per- sonnel Psychology 77 (2022) 475–496. doi:10.1111/ peps.12551

    2022

  26. [34]

    Zhang, X

    Y. Zhang, X. Chen, Explainable recommendation: A survey and new perspectives, Foundations and Trends®in Information Retrieval (2020). doi: 10. 1561/1500000066

    2020

  27. [35]

    G. Bied, S. Nathan, E. Perennes, M. Hoffmann, P. Cail- lou, B. Crépon, C. Gaillac, M. Sebag, Toward Job Recommendation for All, in: IJCAI 2023 - The 32nd International Joint Conference on Artificial In- telligence, International Joint Conferences on Artifi- cial Intelligence O...

    2023 doi

  28. [36]

    L. D. Chao Yang, Intelligent talent recommendation al- gorithm for college students for the future job market, Jes (2024). doi:10.52783/jes.1721

    2024 doi

  29. [37]

    J. Yao, Y. Xu, J. Gao, A study of reciprocal job recom- mendation for college graduates integrating semantic keyword matching and social networking, Applied Sciences (2023). doi:10.3390/app132212305

    2023 doi

  30. [38]

    Schellingerhout, Explainable multi-stakeholder job recommender systems (2024)

    R. Schellingerhout, Explainable multi-stakeholder job recommender systems (2024). doi: 10.1145/ 3640457.3688014

    2024

  31. [39]

    Schellingerhout, F

    R. Schellingerhout, F. Barile, N. Tintarev, A co-design study for multi-stakeholder job recom- mender system explanations (2023). doi: 10.1007/ 978-3-031-44067-0\_30

    2023

  32. [40]

    L. Yao, C. Mao, Y. Luo, Kg-bert: Bert for knowledge graph completion, 2019. URL: https://arxiv.org/abs/ 1909.03193.arXiv:1909.03193

    2019 arXiv

  33. [41]

    X. Liu, H. Hussain, H. Razouk, R. Kern, Effective use of bert in graph embeddings for sparse knowledge graph completion, in: Proceedings of the 37th ACM/SIGAPP Symposium on Applied Computing, SAC ’22, Associa- tion for Computing Machinery, New York, NY, USA, 2022, p. 799–802. ...

    2022

  34. [42]

    Y. Xu, Q. Su, Boosting bert-based knowledge graph completion with contrastive learning and hard sample training, Procedia Computer Science 222 (2023) 71–80. URL: https://www.sciencedirect.com/science/article/ pii/S1877050923009109. doi: https://doi.org/10. 1016/j.procs.2023.08...

    2023

  35. [43]

    B. Kim, T. Hong, Y. Ko, J. Seo, Multi-task learning for knowledge graph completion with pre-trained lan- guage models, in: D. Scott, N. Bel, C. Zong (Eds.), Proceedings of the 28th International Conference on Computational Linguistics, International Committee on Computational ...

    2020 doi

  36. [44]

    H. Gul, A. G. Naim, A. A. Bhat, A contextu- alized bert model for knowledge graph comple- tion, 2024. URL: https://arxiv.org/abs/2412.11016. arXiv:2412.11016

    2024 arXiv

  37. [45]

    Zhang, J

    D. Zhang, J. Liu, H. Zhu, Y. Liu, L. Wang, P. Wang, H. Xiong, Job2vec: Job title benchmarking with col- lective multi-view representation learning, in: Pro- ceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM ’19, Association for Com...

    2019

  38. [46]

    D. Lavi, V. Medentsiy, D. Graus, consultantbert: Fine- tuned siamese sentence-bert for matching jobs and job seekers, 2021. URL: https://arxiv.org/abs/2109.06501. arXiv:2109.06501

    2021 arXiv

  39. [47]

    M. Kaya, T. Bogers, Effectiveness of job title based embeddings on résumé to job ad recommendation, in: CEUR Workshop Proceedings, volume 2967, CEUR Workshop Proceedings, 2021. URL: https://vbn.aau.dk/ ws/portalfiles/portal/464971521/Kaya_Bogers.pdf

    2021

  40. [48]

    Decorte, J

    J.-J. Decorte, J. Van Hautte, T. Demeester, C. Develder, Jobbert: Understanding job titles through skills, 2021. URL: https://www.proquest.com/working-papers/ jobbert-understanding-job-titles-through-skills/ docview/2574960980/se-2

    2021

  41. [49]

    J. Liu, Y. C. Ng, Z. Gui, T. Singhal, L. 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). doi:10.1186/s40537-022-00649-5

    2022 doi

  42. [50]

    R. Zbib, A. L. Lucas, F. Retyk, R. Poves, J. Aizpuru, H. Fabregat, V. Simkus, E. García-Casademont, Learn- ing job titles similarity from noisy skill labels, 2023. doi:10.48550/arxiv.2207.00494

    2023 doi arXiv

  43. [51]

    Lewis, Y

    M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Mo- hamed, O. Levy, V. Stoyanov, L. Zettlemoyer, BART: Denoising sequence-to-sequence pre-training for nat- ural language generation, translation, and comprehen- sion, in: D. Jurafsky, J. Chai, N. Schluter, J. Tetreault (Eds.), ...

    2020 doi

  44. [52]

    Schlichtkrull, T

    M. Schlichtkrull, T. N. Kipf, P. Bloem, R. van den Berg, I. Titov, M. Welling, Modeling relational data with graph convolutional networks, 2017. URL: https: //arxiv.org/abs/1703.06103.arXiv:1703.06103

    2017 arXiv

  45. [53]

    google.com/, 2024

    Google, Google colaboratory, https://colab.research. google.com/, 2024. Accessed: 2025-08-02

    2024

  46. [54]

    Python Core Team, Python: A dynamic, open source programming language, Python Software Foundation,

  47. [55]

    Integrated development environment (IDE) by Microsoft

    Microsoft Corporation, Visual studio, https: //visualstudio.microsoft.com/, 2025. Integrated development environment (IDE) by Microsoft. Accessed: 2025-08-07

    2025

  48. [56]

    Zadykian, Job title relatedness, https://github.com/ vadimzadykian/JobTitleRelatedness, 2025

    V. Zadykian, Job title relatedness, https://github.com/ vadimzadykian/JobTitleRelatedness, 2025. Accessed: 2025-08-07

    2025

  49. [57]

    Koneru, Linkedin job postings (2023 - 2024), 2024

    A. Koneru, Linkedin job postings (2023 - 2024), 2024. doi:10.34740/KAGGLE/DSV/9200871

    2023

  50. [58]

    Commission, S

    E. Commission, S. A. Directorate-General for Employ- ment, Inclusion, ESCO handbook – European skills, competences, qualifications and occupations, Publica- tions Office of the European Union, 2019. doi:doi/10. 2767/451182

    2019

  51. [59]

    I. E. Team, 230 job titles in 17 industries to include on your resume, 2025. URL: https://www.indeed.com/ career-advice/resumes-cover-letters/job-title

    2025

  52. [60]

    Department of Labor, O*net online, 2025

    U.S. Department of Labor, O*net online, 2025. URL: https://www.onetonline.org/, accessed: 2025-08-06

    2025

  53. [61]

    URL: https:// scite.ai/assistant

    Scite, Inc., Ai research assistant, 2024. URL: https:// scite.ai/assistant

    2024

  54. [62]

    URL: https:// openai.com, large language model used for code gen- eration and assistance

    OpenAI, Openai code assistant, 2025. URL: https:// openai.com, large language model used for code gen- eration and assistance

    2025

  55. [63]

    URL: https://www.grammarly.com/grammar-check, accessed: 2025-04-21

    Grammarly, Grammarly grammar checker, n.d. URL: https://www.grammarly.com/grammar-check, accessed: 2025-04-21

    2025

  56. [64]

    Accessed: 2025- 08-07

    Microsoft, Copilot (gpt-4) [large language model], https://copilot.microsoft.com/, 2025. Accessed: 2025- 08-07

    2025

  57. [67]

    Appendices A. Tables Parameter Description Value High STR Range Data region in which STR is considered ’high’ [0.75, 1.0] Medium STR Range Data region in which STR is considered ’medium’ [0.5, 0.75] Low STR Range Data region in which STR is considered ’low’ [0, 0.5] Number of ...

  58. [68]

    Source Skill Hier- archy Input

    and a list of 50 skill categories [59]. Source Skill Hier- archy Input. File ’source_skill_hierarchy.csv’ de- fines relationships between skill categories. For example, ‘Marketing’ and ‘Sales’ are grouped under ‘Sales & Marketing’ Training Job Title Pairs Output. File ’train_j...

  59. [2019]

    URL: https://www.python.org/

  60. [3093]

    URL: https://aclanthology.org/2020.emnlp-main. 249/. doi:10.18653/v1/2020.emnlp-main.249

    2020 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.