Pith. sign in

REVIEW 4 major objections 5 minor 64 references

FewTopNER: Integrating Few-Shot Learning with Topic Modeling and Named Entity Recognition in a Multilingual Framework

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read FewTopNER claims that adding document-level topic context to few-shot NER raises F1 by 2.5 to 4.0 points across five languages and improves topic coherence.

desk verdict A sensible architecture that combines topic modeling with few-shot NER, but the empirical evidence is too unstable to support the headline gains until code and baseline details are released. read the letter →

arxiv 2502.02391 v1 pith:3IDZCYEH submitted 2025-02-04 cs.CL cs.AI

Ibrahim Bouabdallaoui , Fatima Guerouate , Samya Bouhaddour , Chaimae Saadi , Mohammed Sbihi This is my paper · ORCID
classification cs.CLcs.AI
keywords few-shotlearningnamedentityrecognitiontopicmodelingcross-lingualtransfermultilingualNLPprototypenetworksconditionalrandomfieldsXLM-RoBERTa
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 is trying to establish that a few-shot named-entity-recognition system improves when it is made topic-aware. It presents FewTopNER, in which a shared XLM-RoBERTa encoder feeds both a prototype-based entity branch (BiLSTM plus CRF) and a topic branch that combines LDA with neural features, all connected by a cross-task bridge. On English, French, Spanish, German, and Italian, the authors report F1 gains of 2.5 to 4.0 percentage points over ProtoNER and MetaNER, along with better topic coherence as measured by normalized pointwise mutual information (NPMI). If this claim holds, low-resource and cross-lingual NER can be improved by exploiting cheap document-level semantic signal instead of relying only on additional labeled entity examples.

What carries the argument

The load-bearing component is the cross-task bridge: a bidirectional attention and gated feature-fusion module that projects entity and topic representations into a shared space, lets each branch attend to the other, and gates how much cross-task information enters. It sits between the prototype/BiLSTM/CRF entity branch and the LDA-plus-neural topic branch, and it is what turns topic modeling into a disambiguation signal for few-shot NER. Language-specific calibration on top of the shared XLM-RoBERTa encoder is the second carrier of the argument, because it is what makes the topic-aware representations transferable across English, French, Spanish, German, and Italian.

What would settle it

Run the reported 5-way 5-shot episodes on the same WikiNeural-style splits using the baselines' own published configurations, and check whether FewTopNER still beats ProtoNER and MetaNER by 2.5 to 4.0 F1 points with NPMI -0.28 versus -0.32 and -0.33; if the gaps disappear or reverse, the central claim is falsified.

Watch

Extended reading notes

Core claim

FewTopNER's central claim is that topic context is not a byproduct but a working disambiguation signal for few-shot NER: the system couples entity recognition with topic modeling so that document-level themes help resolve entity mentions, while entity cues in turn sharpen topic representations. Concretely, the paper reports that FewTopNER reaches an average F1 of 67.5% in the 5-way 5-shot setting, against 65.0% for ProtoNER and 63.5% for MetaNER; it also reports NPMI of -0.28 versus -0.32 and -0.33 for those baselines. In cross-lingual transfer from English, FewTopNER is said to beat MetaNER by 3.0 to 4.0 F1 points depending on the target language. The authors attribute these gains mainly to the cross-task bridge and the language-calibrated shared encoder, with ablations showing that removing the bridge costs about 3.5 F1 points and removing the shared encoder costs about 13.5 points.

Load-bearing premise

The load-bearing premise is that ProtoNER and MetaNER were evaluated in exactly the same setting as FewTopNER—same data splits, same N-way K-shot episodes, same label sets, and same hyperparameter budget—so the reported 2.5 to 4.0 point gaps reflect the architecture rather than the evaluation setup.

Editorial extensions

If this is right

  • Document-level topic context can serve as a usable disambiguation signal for few-shot NER, so annotation budgets could be spent on entity types rather than on additional contexts.
  • Joint training of NER and topic modeling becomes an attractive option for low-resource pipelines, since the paper's ablation attributes about 3.5 F1 points to the cross-task bridge alone.
  • The shared multilingual encoder is the single most important component, so cross-lingual representation quality, not just the topic branch, drives the reported gains.
  • The stated cross-lingual improvements are smallest for English-to-German transfer, suggesting that morphologically rich target languages benefit less from English-only transfer.
  • The reported gains come with higher computational cost, so practical deployment will involve trading resources against the 2.5 to 4.0 point F1 advantage.

Reading between the lines

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

  • If the reported gains replicate, the same cross-task bridge could be applied to other sequence-labeling tasks such as slot filling, event extraction, or clinical concept extraction, where document theme often helps disambiguate mentions.
  • A cheap extension would be to test whether the topic branch works when it uses only unlabeled in-domain text, since LDA needs no entity annotations; that would separate the contribution of topic signal from the cost of extra labeled data.
  • The paper reports aggregate F1 numbers only; running multiple random support-set seeds and reporting variance would show whether the 2.5 to 4.0 point gaps are stable or driven by a few favorable episodes.
  • The mutual-reinforcement claim implies that feeding recognized entities back into the topic branch should raise NPMI beyond static LDA; this feedback direction could be measured directly by ablating it.
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

4 major / 5 minor

Summary. The paper proposes FewTopNER, a multilingual few-shot named entity recognition framework that couples a prototype/CRF entity branch with a topic-modeling branch (language-specific LDA plus neural topic prototypes) through cross-task attention and gating. The authors claim that adding topic-aware context improves few-shot NER F1 by 2.5–4.0 points over ProtoNER and MetaNER on WikiNeural data for English, French, Spanish, German, and Italian, and also improves topic coherence as measured by NPMI. The presentation includes a full architecture description, training objectives, ablations, and a comparative table.

Significance. If the claimed gains were reproducible, the paper would provide a useful architectural suggestion: document-level topic context can help disambiguate entities in few-shot, cross-lingual settings. The general idea is plausible and the ablation structure is sensible in outline. However, the current evidence is not sufficient to support the central claim: no code, configuration, or data splits are released; the reported full-model F1 values are internally inconsistent; and the topic-coherence comparisons are partly circular because the training loss directly maximizes the metric being reported. The paper is a reasonable draft for a workshop-style contribution, but as a journal submission it needs substantial additional experimental rigor before the headline claims can be accepted.

major comments (4)
  1. [Section 4.4, Table 7] The comparison of FewTopNER with ProtoNER and MetaNER is not a controlled evaluation. The paper reports no WikiNeural split version, no N-way/K-shot episode sampling scheme, no support/query ratio, no label set, no random seeds, and no hyperparameter budgets for the baselines, and no code is provided. Few-shot NER F1 is highly sensitive to these choices, so the advertised 2.5–4.0 point gains could arise from evaluation mismatch alone. The same concern applies to the cross-lingual transfer results in Section 4.2.1, Table 4.
  2. [Tables 3, 5, 6, 7 and Table 1] The full-model F1 score is not stable across the paper: Table 3 reports 68.5 for the with-bridge model, Table 5 and Table 7 report 67.5, Table 6 reports 70.0 with data augmentation, and the mean of the per-language 5-way 5-shot scores in Table 1 is 66.84 rather than 67.5. Since the central quantitative claim is a gain of 2.5–4.0 points over baselines, this internal inconsistency means the gain estimate is not pinned down. No standard deviations or significance tests are reported for any of these differences.
  3. [Section 3.4.2 / Section 3.6.2 / Tables 2 and 7] The topic-coherence improvement is partly by construction. Equations in Sections 3.4.2 and 3.6.2 define Lcoherence and Ltopic as direct objectives that maximize PMI and topic diversity, while the LDA baseline in Table 2 and the ProtoNER/MetaNER rows in Table 7 have no such objective or topic branch. Comparing NPMI under these conditions does not demonstrate that cross-task integration produces more coherent topics; it largely reflects the chosen training objective. A controlled comparison would require a topic branch trained without the cross-task bridge or with the same coherence objective but no entity branch.
  4. [Section 3.6 and Section 4] The introduction states that FewTopNER integrates MAML and an Active Learning Interface, but neither component is defined with equations or evaluated in the experiments. If these mechanisms are part of the framework, their contribution to the reported results is unmeasured; if they are not used, they should be removed from the framework description.
minor comments (5)
  1. [Section 3.7.1] The citation placeholder "[ ?]" appears in the Dataset Management subsection and should be replaced with a proper reference or removed.
  2. [Section 4.4] The sentence "The results in Table 7 clearly that FewTopNER outperforms..." is missing a verb and should read "clearly show that".
  3. [Section 3.8] Sections 3.8.2 and 3.8.3 repeat the same headings 'Cross-Lingual Adaptation' and 'Multi-Task Prediction' with overlapping but slightly different content; this duplication should be resolved.
  4. [Figures 1 and 2] The text and captions describe Figure 1 and Figure 2 in detail, but the corresponding figure images are not present in the submitted manuscript; if the figures were lost in compilation, they should be restored.
  5. [Section 3.2.2 and Section 3.6.2] The symbol Lalign is used for two different objectives (cross-lingual projection alignment and cross-task alignment), and Ldiversity appears with different definitions in Sections 3.2.3 and 3.4.2; distinct names or subscripts would improve clarity.

Circularity Check

1 steps flagged · score 4.0 of 10

Topic-coherence gain reduces to the PMI training objective; the NER F1 comparison is external but internally unstable.

  1. fitted input called prediction [Section 3.4.2 (Eq. Lcoherence) and Section 3.6.2 (Eq. Ltopic); reported in Table 2 and Table 7]
    "Lcoherence = -\sum_c \sum_{w,w'} PMI(w,w')P(w|c)P(w'|c) (Sec. 3.4.2); Ltopic = -\sum_k \sum_{i,j} PMI(w_i,w_j)P(w_i|z_k)P(w_j|z_k) + \lambda_{div}||PP^T - I||_F^2, 'The first term maximizes the pointwise mutual information (PMI)' (Sec. 3.6.2); 'FewTopNER attains a superior topic coherence, with an NPMI of -0.28, compared to -0.32 for ProtoNER and -0.33 for MetaNER' (Table 7)."

    The paper's headline 'enhanced topic coherence' is measured by NPMI, a normalized version of PMI. Yet the model's own training objectives directly maximize PMI over topic-word distributions: Lcoherence in Sec. 3.4.2 and Ltopic in Sec. 3.6.2 both contain negative PMI terms. The reported NPMI gain is therefore the value of the objective being optimized, not an independent measurement of an emergent property. The LDA baseline is not trained with a PMI-maximizing term, and ProtoNER/MetaNER have no topic branch, so the comparison is not controlled. The coherence improvement is forced by construction rather than empirically discovered.

full rationale

The only concrete circular step is the topic-coherence claim: the NPMI scores in Tables 2 and 7 are produced by a model whose Lcoherence/Ltopic losses directly maximize PMI, so the evaluation metric coincides with the training signal. The central few-shot NER F1 claim (2.5–4.0 points over ProtoNER/MetaNER) is based on external baselines and is not circular; concerns about unstated baseline configurations, absent code, and the full-model F1 varying across Tables 3, 5, 6, and 7 are reproducibility and correctness risks, not circularity. No load-bearing self-citation is present. A score of 4 rather than 6 reflects that the main NER result retains independent empirical content, while a secondary but headline metric reduces to its own optimization objective.

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

The empirical claims rest on pretrained representations, a specific dataset setup, and a set of unstated hyperparameters. None of the central numbers are accompanied by code, data splits, or significance tests; the free parameters and assumptions above are the minimum needed to reproduce the claimed results.

free parameters (6)
  • Number of LDA topics K = not reported
    Sets topic granularity and directly changes NPMI and UMass values in Table 2; not reported in Section 3.4.1.
  • LDA Dirichlet prior alpha = not reported
    Used in the document-topic distribution formula in Section 3.4.1; not reported, affecting the topic branch behavior.
  • Coherence and diversity weights alpha1, alpha2 in topic prototype loss = not reported
    Control Lcoherence and Ldiversity in Section 3.4.2, so they directly shape the NPMI gains in Table 2; values not reported.
  • Cross-task alignment weight lambda_mutual and alignment lambda = not reported
    Balance alignment and task loss in Sections 3.2.2 and 3.6.2; not reported, making the loss balance unreproducible.
  • Contrastive temperature tau and curriculum sampling beta = not reported
    Control contrastive alignment and curriculum sampling in Section 3.2.3; not reported.
  • Dynamic loss weights wNER, wtopic, walign, wcontrast, wreg = not reported
    Multi-objective loss in Section 3.6.2 uses learned dynamic weights; final values are not reported, so the training balance is unreproducible.
assumptions (5)
  • domain assumption XLM-RoBERTa pretrained representations transfer useful entity information across English, French, Spanish, German, and Italian.
    The shared-encoder design in Section 3.2 assumes cross-lingual transfer from the pretrained model.
  • domain assumption Document-level LDA topics carry signal about entity types that is not already present in the token embeddings.
    The cross-task bridge in Section 3.5 only helps if topics add information; the paper motivates this but does not measure redundancy.
  • domain assumption ProtoNER and MetaNER baselines in Tables 4 and 7 were run under the same episodes, label sets, and splits as FewTopNER.
    Comparability of the central F1 gains depends on this; no baseline configs are provided in Section 4.4.
  • domain assumption NPMI and UMass coherence are valid proxies for topic quality in the comparison of LDA and FewTopNER.
    Table 2 uses these metrics without validating them against downstream tasks; the topic branch directly optimizes a PMI-like objective.
  • domain assumption The WikiNeural and WikiMedia corpora have compatible document and entity distributions across languages.
    Cross-lingual transfer and topic enrichment in Sections 3.7 and 4.2 assume comparable multilingual data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FewTopNER: Integrating Few-Shot Learning with Topic Modeling and Named Entity Recognition in a Multilingual Framework." pith.science (2026). https://pith.science/paper/3IDZCYEH

@misc{pith2026250202391,
  author       = {Pith},
  title        = {Pith review of: FewTopNER: Integrating Few-Shot Learning with Topic Modeling and Named Entity Recognition in a Multilingual Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3IDZCYEH}},
  note         = {Machine review of arXiv:2502.02391}
}
read the original abstract

We introduce FewTopNER, a novel framework that integrates few-shot named entity recognition (NER) with topic-aware contextual modeling to address the challenges of cross-lingual and low-resource scenarios. FewTopNER leverages a shared multilingual encoder based on XLM-RoBERTa, augmented with language-specific calibration mechanisms, to generate robust contextual embeddings. The architecture comprises a prototype-based entity recognition branch, employing BiLSTM and Conditional Random Fields for sequence labeling, and a topic modeling branch that extracts document-level semantic features through hybrid probabilistic and neural methods. A cross-task bridge facilitates dynamic bidirectional attention and feature fusion between entity and topic representations, thereby enhancing entity disambiguation by incorporating global semantic context. Empirical evaluations on multilingual benchmarks across English, French, Spanish, German, and Italian demonstrate that FewTopNER significantly outperforms existing state-of-the-art few-shot NER models. In particular, the framework achieves improvements of 2.5-4.0 percentage points in F1 score and exhibits enhanced topic coherence, as measured by normalized pointwise mutual information. Ablation studies further confirm the critical contributions of the shared encoder and cross-task integration mechanisms to the overall performance. These results underscore the efficacy of incorporating topic-aware context into few-shot NER and highlight the potential of FewTopNER for robust cross-lingual applications in low-resource settings.

Figures

Figures reproduced from arXiv: 2502.02391 by the authors.

Figure 1
Figure 1. Entity-Type Specific Performance Breakdown for Mu [PITH_FULL_IMAGE:figures/full_fig_p025_1.png] view at source ↗
Figure 2
Figure 2. Entity Recognition Accuracy Across Configuration [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 60 canonical work pages

  1. [1]

    Comprehensive overview of named entit y recognition: Models, domain-specific applications and challenges

    Kalyani Pakhale. Comprehensive overview of named entit y recognition: Models, domain-specific applications and challenges. ArXiv, abs/2309.14084, 2023

    arXiv 2023

  2. [2]

    Integrating structural priors into transformer for named entity recognition

    Li Y an, Linlin Tang, Wenbin Zhang, Gaozhou Wang, Fading P an, Hailin Hu, and Jing Liu. Integrating structural priors into transformer for named entity recognition. 2023 IEEE 11th Joint International Information T echnology and Artificial Intelligence Conference (ITAIC) , 11:220–224, 2023

    work page 2023

  3. [3]

    Uri Katz, Matan V etzler, Amir D. N. Cohen, and Y oav Goldbe rg. Neretrieve: Dataset for next generation named entity recognition and retrieval. ArXiv, abs/2310.14282, 2023

    work page Pith review arXiv 2023

  4. [4]

    Gliner: Generalist model for named entity recognition using bidirectional transformer

    Urchade Zaratiana, Nadi Tomeh, Pierre Holat, and Thierr y Charnois. Gliner: Generalist model for named entity recognition using bidirectional transformer. ArXiv, abs/2311.08526, 2023. FewTopNER

    arXiv 2023

  5. [5]

    Large language model-driven sem i-automated construction of named entity datasets in specific knowledge domain

    Y ahan Liang and Hua Cao. Large language model-driven sem i-automated construction of named entity datasets in specific knowledge domain. 2024 9th International Symposium on Computer and Informati on Processing T echnology (ISCIPT), pages 312–318, 2024

    work page 2024

  6. [6]

    Named- entity recognition in bengali

    Apurbalal Senapati, Arjun Das, and Utpal Garain. Named- entity recognition in bengali. Proceedings of the 4th and 5th Annual Meetings of the F orum for Information Retriev al Evaluation, 2013

    work page 2013

  7. [7]

    Danhui and Y ude Bi

    Y . Danhui and Y ude Bi. Rule-based recognition of vietnam ese named entities. 2014

    work page 2014

  8. [8]

    Sanglikar and D

    M. Sanglikar and D. Kothari. Named entity recognition sy stem for hindi language: A hybrid approach. 2011

    work page 2011

Show all 64 references
  1. [9]

    Chinese named entity r ecognition using support vector machines

    Xu-Dong Lin, Hong Peng, and Bo Liu. Chinese named entity r ecognition using support vector machines. 2006 International Conference on Machine Learning and Cybernet ics, pages 4216–4220, 2006

    2006

  2. [10]

    R. Paper. A hybrid approach of similarity based svm and c rf for named entity recognition. 2015

    2015

  3. [11]

    Z. Guodong. Geospatial named entities recognition usi ng combination of crf and rules. Computer Engineering, 2011

    2011

  4. [12]

    Named entity recognition of electronic me dical records based on multi-feature fusion

    Xiaoqin Tan. Named entity recognition of electronic me dical records based on multi-feature fusion. Frontiers in Computing and Intelligent Systems , 2023

    2023

  5. [13]

    Clin ical named entity recognition using deep learning models

    Y onghui Wu, Min Jiang, Jun Xu, Degui Zhi, and Hua Xu. Clin ical named entity recognition using deep learning models. AMIA ... Annual Symposium proceedings. AMIA Symposium , 2017:1812–1819, 2017

    2017

  6. [14]

    Conesa, and Cecile Pereira

    Qile Zhu, Xiaolin Li, A. Conesa, and Cecile Pereira. Gra m-cnn: a deep learning approach with local context for named entity recognition in biomedical text. Bioinformatics, 34:1547 – 1554, 2017

    2017

  7. [15]

    Bissyandé, and Jacques Klein

    Cedric Lothritz, Kevin Allix, Lisa V eiber, Tégawendé F . Bissyandé, and Jacques Klein. Evaluating pretrained transformer-based models on the task of fine-grained named e ntity recognition. pages 3750–3760, 2020

    2020

  8. [16]

    Enriching urdu ner with bert embedding, data augmentation, and hybrid encoder-cnn architecture

    Anil Ahmed, Degen Huang, Syed Y asser Arafat, and Imran H ameed. Enriching urdu ner with bert embedding, data augmentation, and hybrid encoder-cnn architecture. ACM Transactions on Asian and Low-Resource Lan- guage Information Processing, 23:1 – 38, 2024

    2024

  9. [17]

    Multi-task learning of q uery intent and named entities using transfer learning

    Shalin S Shah and Ryan Siskind. Multi-task learning of q uery intent and named entities using transfer learning. ArXiv, abs/2105.03316, 2021

    2021 arXiv

  10. [18]

    A novel fusion of machine learning methods for enhancing named entity recognition in indonesi an language text

    Widyawan Widyawan, Bayu Prasetiyo Utomo, and Muhammad Nur Rizala. A novel fusion of machine learning methods for enhancing named entity recognition in indonesi an language text. Jurnal Sistem Informasi Bisnis , 2024

    2024

  11. [19]

    Chinese named entity recognition in the ship news field based on adversarial transfer learning

    Zhihong Zhu, Weiwen Zhang, Hongbin Zhang, and Lianglun Cheng. Chinese named entity recognition in the ship news field based on adversarial transfer learning. Proceedings of the 2024 16th International Conference on Machine Learning and Computing , 2024

    2024

  12. [20]

    Campbell

    Cyprien de Lichy, Hadrien Glaude, and W . Campbell. Meta -learning for few-shot named entity recognition. Proceedings of the 1st W orkshop on Meta Learning and Its Appl ications to Natural Language Processing , 2021

    2021

  13. [21]

    Li, Billy Chiu, Shanshan Feng, and Hongya Wang

    J. Li, Billy Chiu, Shanshan Feng, and Hongya Wang. Few-s hot named entity recognition via meta-learning (extended abstract). 2023 IEEE 39th International Conference on Data Engineerin g (ICDE), pages 3805–3806, 2023

    2023

  14. [22]

    Zhao, and Ch in-Y ew Lin

    Tingting Ma, Huiqiang Jiang, Qianhui Wu, T. Zhao, and Ch in-Y ew Lin. Decomposed meta-learning for few-shot named entity recognition. pages 1584–1596, 2022

    2022

  15. [23]

    Kcl: Few-shot named en tity recognition with knowledge graph and con- trastive learning

    Shan Zhang, Bin Cao, and Jing Fan. Kcl: Few-shot named en tity recognition with knowledge graph and con- trastive learning. pages 9681–9692, 2024

    2024

  16. [24]

    Copner: Contrastive learning with prompt guiding for few-shot named entity reco gnition

    Y ucheng Huang, Kai He, Yige Wang, Xianli Zhang, Tielian g Gong, Rui Mao, and Chen Li. Copner: Contrastive learning with prompt guiding for few-shot named entity reco gnition. pages 2515–2527, 2022

    2022

  17. [25]

    Metaner: Named entit y recognition with meta-learning

    Jing Li, Shuo Shang, and Ling Shao. Metaner: Named entit y recognition with meta-learning. In Proceedings of the web conference 2020 , pages 429–440, 2020

    2020

  18. [26]

    Exploiting named enti ty recognition via pre-trained language model and adver- sarial training

    Jia Fu, Jie Liu, and Wen-Jian Shi. Exploiting named enti ty recognition via pre-trained language model and adver- sarial training. 2021 IEEE International Conference on Computer Science, El ectronic Information Engineering and Intelligent Control T echnology (CEI), pages 665–669, 2021

    2021

  19. [27]

    Logacheva, and M

    Alexander Fritzler, V . Logacheva, and M. Kretov. Few-s hot classification in named entity recognition task. Proceedings of the 34th ACM/SIGAPP Symposium on Applied Com puting, 2018

    2018

  20. [28]

    Wang, Chengcheng Han, Chengyu Wang, Chuanqi Tan, Min ghui Qiu, Songfang Huang, Jun Huang, and Ming Gao

    J. Wang, Chengcheng Han, Chengyu Wang, Chuanqi Tan, Min ghui Qiu, Songfang Huang, Jun Huang, and Ming Gao. Spanproto: A two-stage span-based prototypical netwo rk for few-shot named entity recognition. ArXiv, abs/2210.09049, 2022. FewTopNER

    arXiv

  21. [29]

    Few-shot named entity recognition with entity-level prototypical network enhanced by dispersedl y distributed prototypes

    Bin Ji, Shasha Li, Shaoduo Gan, Jie Y u, Jun Ma, and Huijun Liu. Few-shot named entity recognition with entity-level prototypical network enhanced by dispersedl y distributed prototypes. ArXiv, abs/2208.08023, 2022

    2022 arXiv

  22. [30]

    Few-shot learning for named entity recognition based on bert and two-level model fusion

    Y uan Gong, Lu Mao, and Changliang Li. Few-shot learning for named entity recognition based on bert and two-level model fusion. Data Intelligence, 3(4):568–577, 2021

    2021

  23. [31]

    Heproto: A hierarchical enhancing protonet based on multi-task learning for few-shot named en tity recognition

    Wei Chen, Lili Zhao, Pengfei Luo, Tong Xu, Yi Zheng, and E nhong Chen. Heproto: A hierarchical enhancing protonet based on multi-task learning for few-shot named en tity recognition. Proceedings of the 32nd ACM International Conference on Information and Knowledge Man agement, 2023

    2023

  24. [32]

    Llm-rm at semeval-2023task 2: Multilingual complex ner using xlm-roberta

    Rahul Mehta and V asudeva V arma. Llm-rm at semeval-2023task 2: Multilingual complex ner using xlm-roberta. ArXiv, abs/2305.03300, 2023

    2023 arXiv

  25. [33]

    U nsupervised cross-lingual model transfer for named entity recognition with contextualized word representati ons

    Huijiong Y an, Tao Qian, Liang Xie, and Shanguang Chen. U nsupervised cross-lingual model transfer for named entity recognition with contextualized word representati ons. PLoS ONE, 16, 2021

    2021

  26. [34]

    Malmasi, Anjie Fang, B

    S. Malmasi, Anjie Fang, B. Fetahu, Sudipta Kar, and Oleg Rokhlenko. Multiconer: A large-scale multilingual dataset for complex named entity recognition. pages 3798–3 809, 2022

    2022

  27. [35]

    Rivera Zavala, P aloma Martínez, Claudia Moro, and E

    Elisa Terumi Rubel Schneider, Renzo M. Rivera Zavala, P aloma Martínez, Claudia Moro, and E. Paraiso. Uc3m- pucpr at semeval-2022 task 11: An ensemble method of transfo rmer-based models for complex named entity recognition. pages 1448–1456, 2022

    2022

  28. [36]

    Rinalds Viksna and I. Skadina. Multilingual transform ers for named entity recognition. Balt. J. Mod. Comput. , 10, 2022

    2022

  29. [37]

    Hedderich, David Ifeoluwa Adelani, D

    Michael A. Hedderich, David Ifeoluwa Adelani, D. Zhu, J esujoba Oluwadara Alabi, Udia Markus, and D. Klakow. Transfer learning and distant supervision for mu ltilingual transformer models: A study on african languages. pages 2580–2591, 2020

    2020

  30. [38]

    Jayanth, Bharathi Mohan, and R

    K. Jayanth, Bharathi Mohan, and R. P . Kumar. Indian lang uage analysis with xlm-roberta: Enhancing parts of speech tagging for effective natural language preprocessi ng. 2023 Seventh International Conference on Image Information Processing (ICIIP), pages 850–854, 2023

    2023

  31. [39]

    Uppsala nlp a t semeval-2021 task 2: Multilingual language models for fine-tuning and feature extraction in word-in-context d isambiguation

    Huiling Y ou, Xingran Zhu, and Sara Stymne. Uppsala nlp a t semeval-2021 task 2: Multilingual language models for fine-tuning and feature extraction in word-in-context d isambiguation. ArXiv, abs/2104.03767, 2021

    2021 arXiv

  32. [40]

    Pali at semeval-2021 task 2: Fine-tune xlm-roberta for word in context disambiguation

    Shu-Yi Xie, Jian Ma, Haiqing Y ang, Lian-Xin Jiang, Y ang Mo, and Jianping Shen. Pali at semeval-2021 task 2: Fine-tune xlm-roberta for word in context disambiguation. pages 713–718, 2021

    2021

  33. [41]

    Cross-ling ual ability of multilingual bert: An empirical study

    Zihan Wang, Stephen Mayhew, Dan Roth, et al. Cross-ling ual ability of multilingual bert: An empirical study. arXiv preprint arXiv:1912.07840 , 2019

    1912 arXiv

  34. [42]

    Understanding contra stive representation learning through alignment and uniformity on the hypersphere

    Tongzhou Wang and Phillip Isola. Understanding contra stive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning , pages 9929–9939. PMLR, 2020

    2020

  35. [43]

    Contributions of transformer attention heads in multi-and cross-lingual tasks

    Weicheng Ma, Kai Zhang, Renze Lou, Lili Wang, and Sorous h V osoughi. Contributions of transformer attention heads in multi-and cross-lingual tasks. arXiv preprint arXiv:2108.08375 , 2021

    2021 arXiv

  36. [44]

    Multi-level contrastive learning for cross-lingual alignment

    Beiduo Chen, Wu Guo, Bin Gu, Quan Liu, and Y ongchao Wang. Multi-level contrastive learning for cross-lingual alignment. ICASSP 2022 - 2022 IEEE International Conference on Acousti cs, Speech and Signal Processing (ICASSP), pages 7947–7951, 2022

    2022

  37. [45]

    Lacn er: Enhancing few-shot named entity recognition with label words and contrastive learning

    Y uhui Xiao, Qun Y ang, Jianjian Zou, and Sichi Zhou. Lacn er: Enhancing few-shot named entity recognition with label words and contrastive learning. 2024 International Joint Conference on Neural Networks (IJ CNN), pages 1–8, 2024

    2024

  38. [46]

    Moscato, Marco Postiglione, and Giancarlo Sperlí

    V . Moscato, Marco Postiglione, and Giancarlo Sperlí. F ew-shot named entity recognition: Definition, taxonomy and research directions. ACM Transactions on Intelligent Systems and T echnology, 14:1 – 46, 2023

    2023

  39. [47]

    Benchmarking neura l topic models: An empirical study

    Thanh-Nam Doan and Tuan-Anh Hoang. Benchmarking neura l topic models: An empirical study. pages 4363– 4368, 2021

    2021

  40. [48]

    Few-shot learning for topic modeling

    Tomoharu Iwata. Few-shot learning for topic modeling. ArXiv, abs/2104.09011, 2021

    2021 arXiv

  41. [49]

    Dual cross-attention learning for fine-grained visual categorization and object re-identification

    Haowei Zhu, Wenjing Ke, Dong Li, Ji Liu, Lu Tian, and Yi Sh an. Dual cross-attention learning for fine-grained visual categorization and object re-identification. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4682–4692, 2022

    2022

  42. [50]

    He, Jia Chen, and Guoqiang Xiao

    Xian-Feng Han, Z. He, Jia Chen, and Guoqiang Xiao. Cross -level cross-scale cross-attention network for point cloud representation. ArXiv, abs/2104.13053, 2021. FewTopNER

    2021 arXiv

  43. [51]

    Chen, Wenyi Qin , and Richong Zhang

    Zhuoran Li, Chunming Hu, Xiaohui Guo, J. Chen, Wenyi Qin , and Richong Zhang. An unsupervised multiple- task and multiple-teacher model for cross-lingual named en tity recognition. pages 170–179, 2022

    2022

  44. [52]

    Joint cros s-document information for named entity recognition with multi-task learning

    Dongsheng Wang, Hongjie Fan, and Junfei Liu. Joint cros s-document information for named entity recognition with multi-task learning. 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 1146–1150, 2020

    2020

  45. [53]

    Passonne au, and Rui Zhang

    Sarkar Snigdha Sarathi Das, Arzoo Katiyar, R. Passonne au, and Rui Zhang. Container: Few-shot named entity recognition via contrastive learning. ArXiv, abs/2109.07589, 2021

    2021 arXiv

  46. [54]

    Mitra, Ruiyi Zhang, and Ricardo Henao

    Rui Wang, Tong Y u, Handong Zhao, Sungchul Kim, S. Mitra, Ruiyi Zhang, and Ricardo Henao. Few-shot class-incremental learning for named entity recognition. pages 571–582, 2022

    2022

  47. [55]

    Large-scal e label interpretation learning for few-shot named entity recognition

    Jonas Golde, Felix Hamborg, and Alan Akbik. Large-scal e label interpretation learning for few-shot named entity recognition. pages 2915–2930, 2024

    2024

  48. [56]

    Example-based named entity recognition

    Morteza Ziyadi, Y uting Sun, Abhishek Goswami, Jade Hua ng, and Weizhu Chen. Example-based named entity recognition. ArXiv, abs/2008.10570, 2020

    2008 arXiv

  49. [57]

    Li, Hui Li, Jingguo Ge, Lei Zhang, Liangxiong Li, and B ingzhen Wu

    W . Li, Hui Li, Jingguo Ge, Lei Zhang, Liangxiong Li, and B ingzhen Wu. Cdaner: Contrastive learning with cross-domain attention for few-shot named entity recognit ion. 2023 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2023

    2023

  50. [58]

    Prompt-based metric learning for few-shot ner

    Y anru Chen, Y anan Zheng, and Zhilin Y ang. Prompt-based metric learning for few-shot ner. ArXiv, abs/2211.04337, 2022

    2022 arXiv

  51. [59]

    Protoner: Few shot incremental learning for named entity recognition using prototypical networks

    Ritesh Kumar, Saurabh Goyal, Ashish V erma, and V atche I sahagian. Protoner: Few shot incremental learning for named entity recognition using prototypical networks. In International Conference on Business Process Management, pages 70–82. Springer, 2023

    2023

  52. [60]

    Manner: A variational memory-augmented model for cross domain few-shot named ent ity recognition

    Jinyuan Fang, Xiaobin Wang, Zaiqiao Meng, Pengjun Xie, Fei Huang, and Y ong Jiang. Manner: A variational memory-augmented model for cross domain few-shot named ent ity recognition. pages 4261–4276, 2023

    2023

  53. [61]

    A multi-task semantic decomposition frame- work with task-specific pre-training for few-shot ner

    Guanting Dong, Zechen Wang, Jinxu Zhao, Gang Zhao, Daic hi Guo, Dayuan Fu, Tingfeng Hui, Chen Zeng, Keqing He, Xuefeng Li, Liwen Wang, Xinyue Cui, and Weiran Xu. A multi-task semantic decomposition frame- work with task-specific pre-training for few-shot ner. Proceedings of the...

    2023

  54. [62]

    Dfs-ner: Description enhanced few-shot ner via prompt learning and meta-learning

    Huinan Huang, Y uming Feng, Xiaolong Jin, Saiping Guan, and Jiafeng Guo. Dfs-ner: Description enhanced few-shot ner via prompt learning and meta-learning. 2022 IEEE/WIC/ACM International Joint Conference on W eb Intelligence and Intelligent Agent T echnology (WI-IAT), pages 79...

    2022

  55. [63]

    Cllmfs: A contrastive learning enhanced large language model framework for few-shot named entity recognition

    Y afeng Zhang, Zilan Y u, Y uang Huang, and Jing Tang. Cllmfs: A contrastive learning enhanced large language model framework for few-shot named entity recognition. ArXiv, abs/2408.12834, 2024

    2024 arXiv

  56. [64]

    Redko, Romaric Audigier, Angéliqu e Loesch, and Amaury Habrard

    Quentin Bouniot, I. Redko, Romaric Audigier, Angéliqu e Loesch, and Amaury Habrard. Improving few-shot learning through multi-task representation learning theo ry. pages 435–452, 2020

    2020

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.