Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Efficient Text Encoders for Labor Market Analysis

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read With only 109 million parameters, ConTeXT-match closes the performance gap between lightweight encoders and LLM-based skill extraction, reporting state-of-the-art results on most SkillSpan-ESCO metrics at roughly 1/7660 of the cost.

desk verdict A well-ablated token-level contrastive model and a useful new benchmark, but the SOTA claim against LLMs rests on published numbers and loses on HOUSE. read the letter →

arxiv 2505.24640 v1 pith:PGQ7ZV6A submitted 2025-05-30 cs.CL cs.AI

Jens-Joris Decorte , Jeroen Van Hautte , Chris Develder , Thomas Demeester This is my paper
classification cs.CLcs.AI
keywords skillextractioncontrastivelearningtoken-levelattentionextrememulti-labelclassificationESCOtaxonomyjobtitlenormalizationadvertisementslabormarketanalysis
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 argues that skill extraction from job advertisements—matching each sentence to the right skills in a taxonomy of nearly 14,000 labels—can be done at LLM-level quality with a 109-million-parameter bi-encoder, at a tiny fraction of the compute and latency. The vehicle is ConTeXT-match, a contrastive training objective that scores a sentence against a skill by weighting each token's similarity to the skill through attention instead of averaging the whole sentence into one vector. A sympathetic reader would care because labor market analysis needs to process millions of ads cheaply and in real time, and current state-of-the-art systems rely on expensive LLM pipelines. To make such comparisons trustworthy, the paper also builds Skill-XL, a sentence-level benchmark whose annotations explicitly cluster redundant skill labels, and it updates the JobBERT job-title normalizer to use the extracted skills.

What carries the argument

The central object is the ConTeXT-match score, a label-dependent token-level attention mechanism. For a sentence with token embeddings $z_{x_j}$ and a skill averaged over its tokens as $z_s$, the match score is $\text{match}(x,s)=\sum_j \alpha_j \cos(z_{x_j}, z_s)$, where the weights $\alpha_j$ are softmax-normalized dot products of each token with the skill embedding. This replaces the usual fixed sentence vector, so the model can attend to the few tokens that justify a skill, and those same attention weights double as an explanation and as the basis for a redundancy filter that keeps only skills with a high dot product against some token. Training pairs this score with a symmetric InfoNCE loss in-batch, uses gradient caching for batch size 4,096, and adds a multi-task objective matching skills to their taxonomy descriptions via plain cosine similarity.

What would settle it

Take a fresh sample of real job ads that were not part of the synthetic training corpus or the Skill-XL development set, annotate them exhaustively with the same clustered-label protocol, and run the released ConTeXT-match model; if the redundancy-filtered F1 falls clearly below the reported 0.40-0.41 on Skill-XL, or if the token-level explanations correlate with human rationales no better than chance, the transfer claim is disproved.

Watch

Extended reading notes

Core claim

On the SkillSpan-ESCO test sets (HOUSE, TECH, TECHWOLF), ConTeXT-match reports higher MRR and RP@K than every encoder baseline and, on most metrics, higher than the LLM-based IReRa pipeline and the GPT-3.5/GPT-4 baselines. The method achieves this with a 109M-parameter MPNet-based bi-encoder, using a symmetric InfoNCE contrastive loss trained on 138,260 synthetic sentence-skill pairs plus skill-description matching, with token-level attention replacing sentence averaging at inference. A calibrated threshold plus a token-attention-based redundancy filter yields an F1 of 0.4389 on the Skill-XL development set with redundancy reduced from 27.32% to 13.46%, and F1 of 0.407/0.401 on the Skill-XL RANDOM and UNIQUE test sets. The paper presents these as the first baseline results for Skill-XL, and reports that the same skill signal improves job title normalization, with JobBERT V2 reaching MRR 0.390 on the JobBERT benchmark.

Load-bearing premise

The model is trained almost entirely on 138,260 LLM-generated synthetic job-ad sentences, and the paper assumes those sentences are representative enough of real job ads that the token-level attention learned on them transfers to live data.

Editorial extensions

If this is right

  • Skill extraction at LLM-comparable accuracy becomes feasible with a 109M-parameter model running on a $0.10/hour VM, roughly 7 sentences per second, which the paper demonstrates against the IReRa pipeline (19 seconds and $0.00053 vs. 14m44s and $4.06 on 136 sentences).
  • The token-level attention weights provide per-prediction explanations whose alignment with human rationales (mean Spearman $\rho \approx 0.53$) approaches human-human agreement, supporting interpretable skill tagging in production.
  • ConTeXT-match is a general mechanism for extreme multi-label classification, not a skill-specific pipeline, so the same loss and attention can be applied to other large label spaces.
  • The Skill-XL benchmark defines redundancy-aware metrics (cluster-level precision/recall/F1) that quantify how many near-duplicate labels a model emits, a dimension missing from prior span-labeling benchmarks.
  • JobBERT V2 shows that high-quality extracted skills can replace noisy skill co-occurrence signals, reaching state-of-the-art job title normalization on the original JobBERT benchmark.

Reading between the lines

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

  • Because the training data is synthetic, a stress test the paper does not run is a fully real, non-synthetic training mix; if real job-ad sentences contain vocabulary and sentence structures the LLM generator never produced, the reported Skill-XL gains and the redundancy filter's 13.46% figure may not transfer.
  • If ConTeXT-match generalizes as claimed, the same token-level redundancy filter could serve other extreme-label domains such as medical coding or product categorization, where label overlap similarly inflates evaluation.
  • The token-level attention is a byproduct rather than a trained span labeler; a natural extension would be to convert the attention weights into hard span predictions and measure them against span-level ground truth, which would test whether the explanatory signal is as reliable as the ranking signal.
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

2 major / 6 minor

Summary. The paper proposes ConTeXT-match, a bi-encoder model with token-level attention for extreme multi-label skill classification, applied to ESCO skill extraction from job-ad sentences. It also introduces Skill-XL, a sentence-level benchmark with exhaustive, redundancy-aware annotations, and JobBERT V2, a job title normalization model trained on skills extracted by ConTeXT-match. The skill extraction model is trained on 138,260 synthetic sentence-skill pairs from prior work and evaluated on SkillSpan-ESCO and Skill-XL, reporting state-of-the-art results on most metrics with a 109M-parameter model, including claims of outperforming the LLM-based IReRa pipeline on the TECH and TECHWOLF test sets at roughly 1/7660 of the cost.

Significance. The central technical idea, replacing averaged sentence embeddings with token-level attention over sentence tokens in a contrastive objective, is well motivated and supported by the ablation in Table 5, which shows a large performance drop when the ConTeXT mechanism is removed. The public release of the models and the Skill-XL benchmark is a concrete contribution to the labor-market NLP community, and the efficiency comparison is suggestive for practical deployment. If the claimed state-of-the-art results over LLM-based pipelines are substantiated by a controlled comparison under the same evaluation protocol, this would be a valuable paper. However, the current comparison with IReRa is not sufficiently rigorous to support the headline claim as stated.

major comments (2)
  1. [Section VI-A, Table 2, Appendix C] The claim that ConTeXT-match 'outperforms LLM-based systems on most metrics' and 'closes the performance gap' rests on comparing with IReRa results taken from [5] rather than recomputed under the authors' evaluation script. Appendix C compares only on 136 validation sentences and reports no quantitative RP@K or MRR for IReRa on that common set, only qualitative observations (subsections a-d). Since IReRa's published numbers may depend on sentence splitting, candidate-set size, gold alignment, and the evaluation protocol, the Table 2 comparison is not controlled. Moreover, on the HOUSE test set IReRa's published RP@5 (56.50) and RP@10 (66.51) both exceed ConTeXT-match's (51.09 and 65.84), so the 'closing the gap' statement is not uniform across datasets. The authors should either run IReRa under their own evaluation harness and report the resulting numbers, or weaken the conclusion to 'competitive on TECH and TECHWOLF, inferior on HOUSE.'
  2. [Table 2 caption, Section VI-A] The caption states that asterisks flag results that are significantly (p<0.05) stronger than the second-best method, but the paper never describes the statistical test, the number of random seeds, or any multiple-comparison correction. No error bars or confidence intervals are reported for any model in Table 2, so the reader cannot assess whether the observed differences are reliable beyond the point estimates. This is load-bearing because the asterisks are used to support the state-of-the-art claim. Please provide the test procedure, the number of runs, and variance estimates for the reported metrics.
minor comments (6)
  1. [Section II-B] Typo: 'explit skill data requirement' should be 'explicit skill data requirement.'
  2. [Section III-A, Eqs. (2) and (4)] Equation (2) uses cosine similarity while Equation (4) uses the raw dot product in the softmax for the attention weights; this inconsistency is not explained. Please state whether this is intentional and discuss the effect, or align the two definitions.
  3. [Section IV] The inter-annotator agreement F1 of 0.4395 is reported but not compared with a chance baseline. Given the large label space (close to 14,000 ESCO skills), a discussion of how this value compares with expected agreement for multi-label annotation would help readers interpret the benchmark quality.
  4. [Appendix D] The sentence 'approaching human-human consistency' is somewhat strong: the mean model-human Spearman correlation is about 0.53, while human-human correlation is 0.60. A more cautious wording would better reflect the small sample size and the gap.
  5. [Section VI-A] The model is trained entirely on synthetic job-ad sentences from [16]; a brief discussion of potential domain shift or artifacts in the synthetic corpus, and how the SkillSpan-ESCO results mitigate that risk, would be useful for readers planning to deploy the model.
  6. [Appendix C] The phrase 'cost reduction with a factor of 1/7660' is confusing; it should be 'cost is reduced by a factor of about 7660' or 'cost is 1/7660 of IReRa's.'

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the model is trained on synthetic data and evaluated on independent benchmarks; the IReRa comparison is an evaluation-support gap, not a circular reduction.

full rationale

The paper's central skill-extraction claim is not circular by construction. ConTeXT-match is trained on synthetic sentence-skill pairs from the authors' prior work [16] and on ESCO skill descriptions, while the headline evaluation numbers are computed on separately annotated benchmarks (SkillSpan-ESCO, introduced in the authors' earlier work [14], and the newly released Skill-XL) whose test labels do not overlap with the training pairs. The threshold and redundancy filter are selected on the Skill-XL development split and then applied to the test split, which is standard model selection rather than fitting the reported test metrics. JobBERT V2 is trained on skills extracted by the authors' own skill-extraction model, but it is evaluated on an independent job-title-normalization benchmark whose labels come from the original JobBERT task, not from the skill model; this is a training-data dependency, not a definitional equivalence. The main weakness is evaluative rather than circular: the claim of outperforming LLM-based systems imports IReRa's published RP@K scores from [5], a paper overlapping in authorship with the present paper, without re-running IReRa under the same evaluation protocol, and Appendix C reports only qualitative observations on 136 validation sentences. That is a reproducibility and support gap, not a case where the prediction reduces to its own inputs. The paper's many self-citations supply training data, benchmark provenance, and baselines, but they do not define the reported metrics or force the outcome. Therefore the derivation is self-contained in the relevant sense, and the circularity score is low.

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

The central claims rest on a small number of fitted hyperparameters (threshold, redundancy rule, batch size) and on domain assumptions about ESCO, synthetic training data, and proprietary preprocessing. None of these are independently verified in the paper, but they are standard for applied ML benchmarks.

free parameters (5)
  • threshold tau = 0.53 (without redundancy filtering), 0.48 (with filtering)
    Selected on the Skill-XL development set to maximize F1 (Section VI-A).
  • redundancy filtering decision rule = keep only predicted skills with the highest dot product against at least one sentence token
    Designed and validated on the Skill-XL development set; it changes the decision boundary and reduces redundancy from 27.32% to 13.46% (Section VI-A).
  • batch size (skill extraction) = 4096
    Chosen by comparing RP@5 on the SkillSpan-ESCO development set across batch sizes (Fig. 2).
  • scale hyperparameter = 20
    Default temperature for the InfoNCE softmax used in both skill extraction and JobBERT V2 (Sections III-A and VI-B).
  • JobBERT V2 projection and batch size = 1024-dim projection, batch size 2048
    Set without systematic tuning (Section V and VI-B); the final MRR gain over the previous SOTA is 0.003.
assumptions (6)
  • standard math Softmax attention weights (Eq. 4) provide a valid, differentiable aggregation for matching sentence tokens to skills.
    Standard softmax and cosine similarity properties; needed for the InfoNCE loss in Eqs. 5-7.
  • domain assumption ESCO v1.1.0 with 13,981 skills and descriptions is the correct ground-truth taxonomy for skill normalization.
    The paper adopts ESCO without validating the taxonomy itself (Section VI-A).
  • domain assumption The synthetic job-ad sentence corpus from [16] is representative enough of real job ads to train a transferable skill extractor.
    All training data is synthetic (138,260 sentence-skill pairs, Section VI-A); the paper does not assess distribution shift.
  • domain assumption In-batch negatives sampled from the synthetic corpus are valid hard negatives for contrastive learning.
    Used in Eq. 6 and 7; false negatives could distort the ranking loss (Section III-A).
  • domain assumption The proprietary segmentation model and proprietary feature representation correctly identify relevant sentences and diverse job ads in the TechWolf data lake.
    Used to construct Skill-XL (Section IV) and JobBERT V2 training data (Section V-B); these components are not described or released.
  • domain assumption An inter-annotator F1 of 0.4395 is high enough to support the Skill-XL benchmark as a reliable evaluation set.
    The paper asserts this without a statistical justification (Section IV).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Text Encoders for Labor Market Analysis." pith.science (2026). https://pith.science/paper/PGQ7ZV6A

@misc{pith2026250524640,
  author       = {Pith},
  title        = {Pith review of: Efficient Text Encoders for Labor Market Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PGQ7ZV6A}},
  note         = {Machine review of arXiv:2505.24640}
}
read the original abstract

Labor market analysis relies on extracting insights from job advertisements, which provide valuable yet unstructured information on job titles and corresponding skill requirements. While state-of-the-art methods for skill extraction achieve strong performance, they depend on large language models (LLMs), which are computationally expensive and slow. In this paper, we propose \textbf{ConTeXT-match}, a novel contrastive learning approach with token-level attention that is well-suited for the extreme multi-label classification task of skill classification. \textbf{ConTeXT-match} significantly improves skill extraction efficiency and performance, achieving state-of-the-art results with a lightweight bi-encoder model. To support robust evaluation, we introduce \textbf{Skill-XL}, a new benchmark with exhaustive, sentence-level skill annotations that explicitly address the redundancy in the large label space. Finally, we present \textbf{JobBERT V2}, an improved job title normalization model that leverages extracted skills to produce high-quality job title representations. Experiments demonstrate that our models are efficient, accurate, and scalable, making them ideal for large-scale, real-time labor market analysis.

Figures

Figures reproduced from arXiv: 2505.24640 by the authors.

Figure 1
Figure 1. FIGURE 1 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FIGURE 2 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Multilingual JobBERT for Cross-Lingual Job Title Matching

    cs.CL 2025-07 conditional novelty 4.0 of 10

    A multilingual contrastive model matches job titles across four languages using synthetic translations and shared skill labels.

Reference graph

Works this paper leans on

40 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [5]

    In-context learning for extreme multi-label classification, 2024

    Karel D’Oosterlinck, Omar Khattab, François Remy, Thomas Demeester, Chris Develder, and Christopher Potts. In-context learning for extreme multi-label classification, 2024

    work page 2024

  2. [1]

    Data science for job market analysis: A survey on applications and techniques.Expert Systems with Applications, 251:124101, 2024

    Ibrahim Rahhal, Ismail Kassou, and Mounir Ghogho. Data science for job market analysis: A survey on applications and techniques.Expert Systems with Applications, 251:124101, 2024

    work page 2024

  3. [2]

    Wendy V Cunningham and Paula Villaseñor. Employer voices, employer demands, and implications for public skills development policy connecting the labor and education sectors.The World Bank Research Observer, 31(1):102–134, 2016

    work page 2016

  4. [3]

    A survey on skill identification from online job ads.IEEE Access, 9:118134–118153, 2021

    Imane Khaouja, Ismail Kassou, and Mounir Ghogho. A survey on skill identification from online job ads.IEEE Access, 9:118134–118153, 2021

    work page 2021

  5. [4]

    Large language models as batteries-included zero-shot ESCO skills matchers

    Benjamin Clavié and Guillaume Soulié. Large language models as batteries-included zero-shot ESCO skills matchers. In Mesut Kaya, Toine Bogers, David Graus, Chris Johnson, and Jens-Joris Decorte, editors, Proceedings of the 3rd Workshop on Recommender Systems for Human Resources (RecSys in HR 2023) co-located with the 17th ACM Conference on Recommender Sys...

    work page 2023

  6. [6]

    Automated occupation coding with hierarchical features: a data-centric approach to classification with pre-trained language models

    Parisa Safikhani, Hayastan Avetisyan, Dennis Föste-Eggers, and David Broneske. Automated occupation coding with hierarchical features: a data-centric approach to classification with pre-trained language models. Discover Artificial Intelligence, 3:6, 02 2023

    work page 2023

  7. [7]

    JobBERT : understanding job titles through skills

    Decorte, Jens-Joris and V an Hautte, Jeroen and Demeester, Thomas and Develder, Chris. JobBERT : understanding job titles through skills. In FEAST, ECML-PKDD 2021 Workshop, Proceedings, page 9, 2021

    work page 2021

  8. [8]

    Skill: A system for skill identification and normalization.Proceedings of the AAAI Conference on Artificial Intelligence, 29(2):4012–4017, Jan

    Meng Zhao, Faizan Javed, Ferosh Jacob, and Matt McNair. Skill: A system for skill identification and normalization.Proceedings of the AAAI Conference on Artificial Intelligence, 29(2):4012–4017, Jan. 2015

    work page 2015

Show all 40 references
  1. [9]

    Representation of job-skill in artificial intelligence with knowledge graph analysis

    Shanshan Jia, Xiaoan Liu, Ping Zhao, Chang Liu, Lianying Sun, and Tao Peng. Representation of job-skill in artificial intelligence with knowledge graph analysis. In2018 IEEE Symposium on Product Compliance Engi- neering - Asia (ISPCE-CN), pages 1–6, 2018

    2018

  2. [10]

    Learning representations for soft skill matching

    Luiza Sayfullina, Eric Malmi, and Juho Kannala. Learning representations for soft skill matching. In Wil M. P . van der Aalst, Vladimir Batagelj, Goran Glavaš, Dmitry I. Ignatov, Michael Khachay, Sergei O. Kuznetsov, Olessia Koltsova, Irina A. Lomazova, Natalia Loukachevitch, ...

    2018

  3. [11]

    SkillSpan: Hard and soft skill extraction from English job postings

    Mike Zhang, Kristian Jensen, Sif Sonniks, and Barbara Plank. SkillSpan: Hard and soft skill extraction from English job postings. In Marine Carpuat, Marie-Catherine de Marneffe, and Ivan Vladimir Meza Ruiz, editors, Proceedings of the 2022 Conference of the North American Chap...

    2022

  4. [12]

    NNOSE: Nearest neighbor occupational skill extraction

    Mike Zhang, Rob van der Goot, Min-Y en Kan, and Barbara Plank. NNOSE: Nearest neighbor occupational skill extraction. In Yvette Gra- ham and Matthew Purver, editors,Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (V o...

    2024

  5. [13]

    Skill-llm: Repurposing general-purpose llms for skill extraction, 2024

    Amirhossein Herandi, Yitao Li, Zhanlin Liu, Ximin Hu, and Xiao Cai. Skill-llm: Repurposing general-purpose llms for skill extraction, 2024

    2024

  6. [14]

    Design of negative sampling strategies for dis- tantly supervised skill extraction, 2022

    Jens-Joris Decorte, Jeroen V an Hautte, Johannes Deleu, Chris Develder, and Thomas Demeester. Design of negative sampling strategies for dis- tantly supervised skill extraction, 2022

    2022

  7. [15]

    Fine-grained extraction and classification of skill requirements in German- speaking job ads

    Ann-sophie Gnehm, Eva Bühlmann, Helen Buchs, and Simon Clematide. Fine-grained extraction and classification of skill requirements in German- speaking job ads. In David Bamman, Dirk Hovy, David Jurgens, Katherine Keith, Brendan O’Connor, and Svitlana V olkova, editors,Proceedi...

  8. [16]

    Extreme multi-label skill extraction training using large language models

    Decorte, Jens-Joris and V erlinden, Severine and V an Hautte, Jeroen and Deleu, Johannes and Develder, Chris and Demeester, Thomas. Extreme multi-label skill extraction training using large language models. InAI4HR & PES 2023 : International Workshop on AI for Human Resources ...

    2023

  9. [17]

    Lightxml: Transformer with dynamic negative sampling for high-performance extreme multi-label text classification

    Ting Jiang, Deqing Wang, Leilei Sun, Huayi Y ang, Zhengyang Zhao, and Fuzhen Zhuang. Lightxml: Transformer with dynamic negative sampling for high-performance extreme multi-label text classification. InProceed- ings of the AAAI conference on artificial intelligence, volume 35,...

    2021

  10. [18]

    Deep- xml: A deep extreme multi-label learning framework applied to short text documents

    Kunal Dahiya, Deepak Saini, Anshul Mittal, Ankush Shaw, Kushal Dave, Akshay Soni, Himanshu Jain, Sumeet Agarwal, and Manik V arma. Deep- xml: A deep extreme multi-label learning framework applied to short text documents. InProceedings of the 14th ACM international conference o...

    2021

  11. [19]

    Fast multi-resolution transformer fine-tuning for extreme multi-label text clas- sification.Advances in Neural Information Processing Systems, 34:7267– 7280, 2021

    Jiong Zhang, Wei-Cheng Chang, Hsiang-Fu Y u, and Inderjit Dhillon. Fast multi-resolution transformer fine-tuning for extreme multi-label text clas- sification.Advances in Neural Information Processing Systems, 34:7267– 7280, 2021

    2021

  12. [20]

    Dimensionality reduction by learning an invariant mapping

    Raia Hadsell, Sumit Chopra, and Y ann LeCun. Dimensionality reduction by learning an invariant mapping. In2006 IEEE computer society con- ference on computer vision and pattern recognition (CVPR’06), volume 2, pages 1735–1742. IEEE, 2006

    2006

  13. [21]

    Representation learning with contrastive predictive coding, 2018

    Aaron van den Oord, Y azhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2018

    2018

  14. [22]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PmLR, 2020

    2020

  15. [23]

    Learning transferable visual models from natural lan- guage supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural lan- guage supervision. InInternational conference on machine learning, p...

    2021

  16. [24]

    High-precision phrase-based docu- ment classification on a modern scale

    Ron Bekkerman and Matan Gavish. High-precision phrase-based docu- ment classification on a modern scale. InProceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’11, page 231–239, New Y ork, NY , USA, 2011. Association for Comp...

    2011

  17. [25]

    Carotene: A job title classification system for the online recruitment domain

    Faizan Javed, Qinlong Luo, Matt McNair, Ferosh Jacob, Meng Zhao, and Tae Seung Kang. Carotene: A job title classification system for the online recruitment domain. In2015 IEEE First International Conference on Big Data Computing Service and Applications, pages 286–293, 2015

    2015

  18. [26]

    Deepcarotene -job title classification with multi-stream convolutional neural network

    Jingya Wang, Kareem Abdelfatah, Mohammed Korayem, and Janani Bal- aji. Deepcarotene -job title classification with multi-stream convolutional neural network. In2019 IEEE International Conference on Big Data (Big Data), pages 1953–1961, 2019

    1953

  19. [27]

    Learning Job Titles Similarity from Noisy Skill Labels

    Rabih Zbib, Lucas Alvarez Lacasa, Federico Retyk, Rus Poves, Juan Aizpuru, Hermenegildo Fabregat, V aidotas Šimkus, and Emilia García- Casademont. Learning Job Titles Similarity from Noisy Skill Labels. In FEAST, ECML-PKDD 2022 Workshop, Proceedings, 2022

    2022

  20. [28]

    Bocharova, Eugene V

    Maiia Y . Bocharova, Eugene V . Malakhov, and Vitaliy I. Mezhuyev. V a- cancysbert: the approach for representation of titles and skillsfor semantic similarity search in the recruitment domain.Applied Aspects of Information Technology, 6(1):52–59, April 2023

    2023

  21. [29]

    Learning job title rep- resentation from job description aggregation network

    Napat Laosaengpha, Thanit Tativannarat, Chawan Piansaddhayanon, At- tapol Rutherford, and Ekapol Chuangsuwanich. Learning job title rep- resentation from job description aggregation network. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Findings of the Association ...

    2024

  22. [30]

    Scaling deep contrastive learning batch size under memory limited setup

    Luyu Gao, Y unyi Zhang, Jiawei Han, and Jamie Callan. Scaling deep contrastive learning batch size under memory limited setup. In Anna Rogers, Iacer Calixto, Ivan Vulić, Naomi Saphra, Nora Kassner, Oana- Maria Camburu, Trapit Bansal, and V ered Shwartz, editors,Proceedings of ...

    2021

  23. [31]

    Deep learning-based computational job market analysis: A survey on skill ex- traction and classification from job postings

    Elena Senger, Mike Zhang, Rob van der Goot, and Barbara Plank. Deep learning-based computational job market analysis: A survey on skill ex- traction and classification from job postings. In Estevam Hruschka, Thom Lake, Naoki Otani, and Tom Mitchell, editors,Proceedings of the ...

    2024

  24. [32]

    Retrieving skills from job descriptions: A language model based extreme multi-label classification framework

    Akshay Bhola, Kishaloy Halder, Animesh Prasad, and Min-Y en Kan. Retrieving skills from job descriptions: A language model based extreme multi-label classification framework. In Donia Scott, Nuria Bel, and Chengqing Zong, editors,Proceedings of the 28th International Confer- e...

    2023

  25. [33]

    Apley, and Sanjay Mehrotra

    Boyang Shang, Daniel W. Apley, and Sanjay Mehrotra. Diversity subsam- pling: Custom subsamples from large data sets.INFORMS Journal on Data Science, 2(2):161–182, October 2023

    2023

  26. [34]

    Establishing annotation quality in multi-label annotations

    Marian Marchal, Merel Scholman, Frances Y ung, and V era Demberg. Establishing annotation quality in multi-label annotations. In Nicoletta Calzolari, Chu-Ren Huang, Hansaem Kim, James Pustejovsky, Leo Wan- ner, Key-Sun Choi, Pum-Mo Ryu, Hsin-Hsi Chen, Lucia Donatelli, Heng Ji,...

    2022

  27. [35]

    De- identification of privacy-related entities in job postings

    Kristian Nørgaard Jensen, Mike Zhang, and Barbara Plank. De- identification of privacy-related entities in job postings. In Simon Dobnik and Lilja Øvrelid, editors,Proceedings of the 23rd Nordic Conference on Computational Linguistics (NoDaLiDa), pages 210–221, Reykjavik, Icel...

    2021

  28. [36]

    Extreme multi-label legal text clas- sification: A case study in EU legislation

    Ilias Chalkidis, Emmanouil Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, and Ion Androutsopoulos. Extreme multi-label legal text clas- sification: A case study in EU legislation. In Nikolaos Aletras, Elliott Ash, Leslie Barrett, Daniel Chen, Adam Meyers, Daniel Preoti...

    2019

  29. [37]

    Mpnet: masked and permuted pre-training for language understanding

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Y an Liu. Mpnet: masked and permuted pre-training for language understanding. InPro- ceedings of the 34th International Conference on Neural Information Pro- cessing Systems, NIPS ’20, Red Hook, NY , USA, 2020. Curran Associates Inc

    2020

  30. [38]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, ICML’20. JMLR.org, 2020. JENS-JORIS DECORTEreceived the M.S. degre...

    2020

  31. [2020]

    degree with the Internet Technology and Data Science Labo- ratory (IDLab), Ghent University–imec, in collab- oration with TechWolf

    He is currently pursuing a Ph.D. degree with the Internet Technology and Data Science Labo- ratory (IDLab), Ghent University–imec, in collab- oration with TechWolf. His supervisors are Prof. Chris Develder and Prof. Thomas Demeester. His research focuses on AI for HR applicati...

    2005

  32. [2022]

    Association for Computational Linguistics

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.