RetrieveAll: A Multilingual Named Entity Recognition Framework with Large Language Models

Fan Gao; Gadeng Luosang; Jin Zhang; Linyu Li; Nyima Tashi; Xiangxiang Wang; Yongbin Yu

REVIEW 3 major objections 5 minor 33 references

RetrieveAll: A Multilingual Named Entity Recognition Framework with Large Language Models

T0 review · 3 major / 5 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 RetrieveAll claims per-language LoRA adapters chosen by retrieved examples lift multilingual NER F1 by 12.1% on PAN-X.

desk verdict The headline PAN-X gain is likely confounded by retrieval-augmented prompting that baselines don't get, though the framework idea is worth a second look. read the letter →

arxiv 2505.19128 v1 pith:4C6S4PRV submitted 2025-05-25 cs.CL cs.AI

Jin Zhang , Fan Gao , Linyu Li , Yongbin Yu , Xiangxiang Wang , Nyima Tashi , Gadeng Luosang This is my paper

classification cs.CLcs.AI

keywords namedentityrecognitionmultilingualNLPlow-resourcelanguagesLoRAretrieval-augmentedlearningcross-lingualtransferprompt-drivenparameter-efficientfine-tuning

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 proposes RetrieveAll, a multilingual named-entity-recognition framework that keeps one compact low-rank adapter per language and, at inference, picks the adapter whose language matches the input by retrieving similar training examples. It adds a cross-granularity knowledge augmentation stage in which hierarchical prompts insert retrieved entity-level and context-level examples into both training and inference inputs, shifting the role of prompts from guiding inference to driving learning. The paper reports that on the PAN-X dataset RetrieveAll improves average F1 by 12.1% over existing baselines, with the largest per-language gains in German and Korean, and that the gains hold across base models of different sizes. This matters because it offers a way to scale multilingual NER to new languages by adding a small adapter rather than retraining a shared model, and it claims to do so without any external knowledge base.

What carries the argument

The mechanism has three pieces: a pool of language-specific low-rank adaptation (LoRA) modules—small trainable matrices that add a cheap update to frozen base-model weights; an input-aware retrieval step that encodes input and candidate examples with a large-language-model encoder, keeps the top-5 entity and context examples above similarity thresholds, and chooses the LoRA whose language is the mode of the retrieved contexts' languages (Equation 11); and cross-granularity knowledge augmented learning (CKAL), where hierarchical prompts concatenate those retrieved entity and context examples into both the fine-tuning and inference inputs. The retrieval step is the load-bearing identity: it simultaneously supplies the knowledge-injection exemplars and the language signal that selects the adapter, eliminating the need for an explicit language tag.

What would settle it

Take a language whose script is poorly covered by the LLM encoder, or add noise to the retrieved embeddings for one of the tested languages, and rerun the CKAL pipeline. If LoRA-selection accuracy (the mode of retrieved language labels) falls toward chance and the F1 gap over a no-retrieval baseline collapses, the load-bearing assumption is that similarity-retrieved examples are same-language and useful.

Watch

Extended reading notes

Core claim

RetrieveAll's central claim is that language interference—the suppression of low-resource features by high-resource languages in a shared multilingual model—can be removed by decoupling NER knowledge into per-language low-rank adapters and then routing each input to its adapter through an input-aware retrieval mechanism. The same retrieval mechanism also supplies the training signal: during fine-tuning, each input is prefixed with the top-5 most similar entity examples and context examples, selected by cosine similarity in the embedding space of an LLM encoder, and the language of the retrieved context examples (their mode) selects the LoRA without an explicit language label. On PAN-X, per-language F1 gains over the strongest existing baseline range from 5.3 to 19.5 percentage points, and the framework's language-selection accuracy exceeds 99% on several languages; on MultiCoNER, training-stage example injection produces F1 gains of 13.4% to 44% while inference-stage example expansion adds almost nothing.

Load-bearing premise

The framework assumes that the top-k examples retrieved by embedding cosine similarity are in the same language as the input and are genuinely useful exemplars, because that retrieval both picks the LoRA adapter and supplies the knowledge-injection contexts.

Editorial extensions

If this is right

Training-stage knowledge injection drives the gains: increasing training examples from 0+5 to 5+5 raises F1 by 13.4% to 44% on MultiCoNER, while tripling inference examples changes F1 by roughly 2% at most.
Monolingual CKAL training beats joint multi-language training on most tested languages, supporting the claim that per-language decoupling, not shared training, is what suppresses interference.
The framework's sensitivity to base model size is small: moving from a 7B to a 14B base model changes PAN-X F1 by less than 1%, so the gains come from the framework rather than from model scale.
Input-aware LoRA retrieval reaches over 99% selection accuracy on some languages, which enables batched inference with several adapters active at once, reducing the latency cost of per-language models.

Reading between the lines

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

If retrieval can route to the correct adapter without an explicit label, the same mechanism could select adapters by domain, genre, or register rather than only language, with no architectural change.
The comparison may understate or overstate the gain from example injection: baselines do not receive retrieved exemplars at inference, so a head-to-head against a retrieval-augmented baseline would isolate whether the 12.1% improvement comes from decoupling or from simply giving the model more examples.
Because CKAL draws only on the training data itself, its ceiling is set by that data; for extremely low-resource languages, a small external seed lexicon could plausibly extend the gains, though the paper deliberately avoids external resources.
The mode-based language identification implicitly assumes each input is single-language; code-switched inputs would be routed to one adapter, so an interpolation across top-LoRAs is a natural testable extension.

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.

Desk Editor's Note

read the letter

Quick take: the headline 12.1% average F1 gain on PAN-X is not actually attributable to the proposed LoRA/CKAL framework. Section 5.1 says inference uses the same input format as the CKAL stage, which prefixes every test sentence with entity and context examples retrieved from the training set (carrying gold labels). None of the five baselines get that augmentation, and no retrieval-augmented control is reported. So the comparison measures retrieval-augmented prompting plus LoRA/CKAL against unaugmented baselines, and the claimed gain cannot be assigned to the framework. That is a load-bearing flaw.

What is genuinely new: combining per-language LoRA retrieval with cross-granularity example injection during fine-tuning is a reasonable idea that I haven't seen exactly this way. The framework is coherent and the ablations (entity-only vs context-only, training-vs-inference augmentation, zero-shot joint vs monolingual) are thoughtful. They also report confidence intervals and explicitly avoid test-set leakage in retrieval.

The other problems are smaller but real. No code is released, and they don't report LoRA rank, learning rate, or training steps, so the work is not independently reproducible. The choice of 8 of 282 PAN-X languages and 6 of 11 MultiCoNER languages is never justified, which raises selection concerns. And the 12.1% average doesn't exactly match the per-language numbers in Section 5.2 (those average about 11.8%). These are fixable with transparency.

If the authors clarified the protocol and added a baseline where they prefix the same retrieved examples to, say, the strongest baseline, the central claim could be tested. As written, the evidence doesn't support the title claim. That said, the paper is not a waste of time: the idea has legs, and the framework description is clear enough to reimplement (modulo missing hyperparameters).

I'd send it to peer review because the idea deserves scrutiny, but I would not accept it as is. For a reading group, it's a good case study in evaluation confounds in retrieval-augmented LLM methods. I would not cite it until the evaluation is redone.

Referee Report

3 major / 5 minor

Summary. The paper proposes RetrieveAll, a multilingual NER framework built on per-language LoRA adapters with input-aware LoRA retrieval, and a cross-granularity knowledge augmented learning (CKAL) stage that builds hierarchical prompts from retrieved entity-level and context-level examples drawn from the training set. The framework is evaluated on eight PAN-X languages and six MultiCoNER languages using LLaMA3-8B and Qwen2.5-7B/14B backbones; the authors report that RetrieveAll beats five baselines and, on PAN-X, achieves an average F1 improvement of 12.1%.

Significance. If the reported gains were properly attributed, the paper would make a useful contribution: dynamically selecting per-language LoRA adapters while injecting retrieved training knowledge at both training and inference is a plausible way to mitigate language interference in multilingual NER. The paper includes useful empirical ingredients: ten-run averaged results with confidence intervals, ablations on entity-level versus context-level examples, a table of LoRA-retrieval accuracy, and experiments across three base models. However, the central comparison is confounded because, at inference, RetrieveAll receives retrieved training examples with gold labels, while the baselines receive no such augmentation, and no retrieval-augmented control is reported. The claimed 12.1% improvement is therefore not currently attributable to the proposed mechanisms.

major comments (3)

[§5.1 and Eq. (5)] The inference setup confounds the headline result. Section 5.1 states: "During inference, we adopt the same input format as used in the cross-granularity knowledge augmented learning stage." In Eq. (5), that format prefixes every input with entity examples (Pe, se1..sen) and context examples (Pc, sc1..scm). At inference these examples are retrieved from the training set and therefore carry gold NER labels. None of the five baselines (XLM-RoBERTa, LS-unLLaMA, PromptNER, GPT-NER, CascadeNER) receives this augmentation, and the paper reports no retrieval-augmented baseline control. Consequently, the 12.1% average F1 gain on PAN-X cannot be assigned to dynamic LoRA decoupling or CKAL; it may be driven largely by giving RetrieveAll gold-annotated few-shot examples at test time. The authors should add at least two controls: (a) the base LLM with the same retrieval-augmented prompt but without LoRA modules, and (b) RetrieveAll without the inference-time retrieval prefix, to isolate the contribution of each component.
[§5.2 and Table 1] The claimed average improvement of 12.1% is not reproducible from the reported per-language numbers. Section 5.2 lists improvements of 5.3, 13.4, 12.7, 12.9, 16.0, 7.2, 7.2, and 19.5 percent, which average to about 11.8%, and these values do not match the differences between RetrieveAll (LLaMA3-8B, CKAL) and the strongest per-language baselines in Table 1. For example, English is 96.6 versus 91.7 for LS-unLLaMA or 91.0 for CascadeNER, a difference of 4.9 or 5.6 points, not 5.3; Spanish is 98.7 versus 87.0, a difference of 11.7, not 13.4. The authors should specify which baseline and averaging formula produce the 12.1% figure and report the per-language deltas consistently.
[§5.1 and Appendix A] The implementation description omits essential hyperparameters: the LoRA rank r, learning rate, number of training steps or epochs, batch size, optimizer, and the number of training examples used per language to fit each LoRA module. The selection of 8 of 282 PAN-X languages and 6 of 11 MultiCoNER languages is also unexplained. These omissions make the dynamic-LoRA results difficult to reproduce and leave open the possibility of selection bias. Please provide a complete hyperparameter table and a clear specification of the language-selection criterion.

minor comments (5)

[§5.3, 'Decoupling Capability'] The text says that the decoupling ablation was done "by removing the entity-level and context-level examples in CKAL," but Table 2 actually compares joint training with monolingual training; the description should be aligned with the table.
[References] The baseline called CascadeNER in the text is cited to Luo et al. 2024, but the bibliography entry for that reference is titled "Geic: Universal and multilingual named entity recognition with large language models"; please reconcile the method name and citation.
[Figure 3] The figure lacks axis labels and a definition of the notation "0+5-shot", "5+5-shot", and "5+15-shot", which is necessary for interpreting the training-stage versus inference-stage comparison.
[Table 2] The term "zero-shot" is potentially misleading because the models in Table 2 are trained with CKAL; clarify what is zero-shot in this setting.
[Appendix A] The appendix says it presents "error bars and confidence intervals," but Tables 5 and 6 contain numeric standard deviations and confidence bounds; either add error-bar plots or rephrase the description.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework is an empirical supervised method and its benchmark comparisons do not reduce to a fitted target.

full rationale

RetrieveAll is an empirical supervised framework evaluated on external benchmarks (PAN-X, MultiCoNER), and its central claims are performance comparisons rather than derivations from its own inputs. The CKAL mechanism retrieves entity and context examples from the training set and injects them as prompts; because those examples carry gold labels and are also used at inference, this is a standard supervised resource and a possible evaluation-fairness concern relative to baselines, but it is not a circular reduction of the kind where an output is defined by its input. The input-aware LoRA selection in Equation 11 is a nearest-neighbor mode over retrieved contexts, not a quantity fitted to the reported F1 scores, and the thresholds tau_e=0.65, tau_c=0.7 and k=5 are hyperparameters rather than fitted predictions. No load-bearing self-citation chain appears in the paper: LoRA, LLM2Vec, and batched LoRA machinery are cited as independent external prior work. No equation reduces to its own target, so no specific circular step can be exhibited; the skeptic's retrieval-augmentation critique is better treated as an evaluation-control risk than as circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on a set of hand-set retrieval hyperparameters (tau_e, tau_c, k) and an unreported LoRA rank, plus assumptions about embedding similarity, language identification, and fair baseline comparison. No new physical or conceptual entities are introduced; the LoRA candidate pool is an engineering artifact, not an invented entity.

free parameters (4)

Entity example similarity threshold tau_e = 0.65
Hand-set in Section 5.1; controls which entity examples are selected during CKAL, with no sensitivity analysis reported.
Context example similarity threshold tau_c = 0.7
Hand-set in Section 5.1; controls context example selection, with no sensitivity analysis reported.
Number of retrieved examples k = 5
Set in Section 5.1; top-k entity and context examples per sample, with no full ablation over k.
LoRA rank r
Never reported in the paper; the rank of per-language adapters is a capacity hyperparameter that directly affects the decoupling claim and reproducibility.

assumptions (5)

domain assumption Cosine similarity in LLM2Vec embedding space selects semantically useful entity and context examples (Section 4.2, Equations 9-10).
The whole CKAL mechanism assumes embedding similarity corresponds to task-relevant similarity, especially for low-resource languages.
domain assumption The mode of the language labels of the top-k retrieved context examples identifies the language of the input (Section 4.3, Equation 11).
LoRA selection depends on this; code-mixed or short inputs may violate it, and retrieval errors directly cause wrong adapter selection.
domain assumption Including retrieved training examples in prompts during both training and inference is a fair basis for comparing against baselines that do not use this augmentation (Sections 4.2 and 5.1).
If the gains come mostly from retrieval augmentation, the dynamic LoRA contribution is not isolated; ablations partially address this, but the baselines are not given the same exemplars.
domain assumption Per-language LoRA modules trained independently on limited data can capture language-specific NER features without overfitting (Section 4.1).
For medium- and low-resource languages, training a separate adapter requires sufficient data; the paper's own Limitations section notes extremely scarce target languages may fail.
standard math LoRA weight update Delta-W equals B times A with low rank r is sufficient for per-language adaptation (Equation 1, Section 3).
Standard LoRA assumption, not unique to this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RetrieveAll: A Multilingual Named Entity Recognition Framework with Large Language Models." pith.science (2026). https://pith.science/paper/4C6S4PRV

@misc{pith2026250519128,
  author       = {Pith},
  title        = {Pith review of: RetrieveAll: A Multilingual Named Entity Recognition Framework with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4C6S4PRV}},
  note         = {Machine review of arXiv:2505.19128}
}

read the original abstract

The rise of large language models has led to significant performance breakthroughs in named entity recognition (NER) for high-resource languages, yet there remains substantial room for improvement in low- and medium-resource languages. Existing multilingual NER methods face severe language interference during the multi-language adaptation process, manifested in feature conflicts between different languages and the competitive suppression of low-resource language features by high-resource languages. Although training a dedicated model for each language can mitigate such interference, it lacks scalability and incurs excessive computational costs in real-world applications. To address this issue, we propose RetrieveAll, a universal multilingual NER framework based on dynamic LoRA. The framework decouples task-specific features across languages and demonstrates efficient dynamic adaptability. Furthermore, we introduce a cross-granularity knowledge augmented method that fully exploits the intrinsic potential of the data without relying on external resources. By leveraging a hierarchical prompting mechanism to guide knowledge injection, this approach advances the paradigm from "prompt-guided inference" to "prompt-driven learning." Experimental results show that RetrieveAll outperforms existing baselines; on the PAN-X dataset, it achieves an average F1 improvement of 12.1 percent.

Figures

Figures reproduced from arXiv: 2505.19128 by the authors.

**Figure 2.** RetrieveAll injects cross-granularity knowledge by retrieving entity-level and context-level examples via [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Comparative analysis of the impact of exam [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: The standardized input and output format [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 10 canonical work pages

[1]

Ahmet Yusuf Alan, Enis Karaarslan, and \"O mer Aydin. 2024. A rag-based question answering system proposal for understanding islam: Mufassirqas llm. arXiv preprint arXiv:2401.15378

arXiv 2024
[2]

Dhananjay Ashok and Zachary C Lipton. 2023. Promptner: Prompting for named entity recognition. arXiv preprint arXiv:2305.15444

arXiv 2023
[3]

Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961

arXiv 2024
[4]

Ruirui Chen, Chengwei Qin, Weifeng Jiang, and Dongkyu Choi. 2024. Is a large language model a good annotator for event extraction? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17772--17780

work page 2024
[5]

Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116

arXiv 2019
[6]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

arXiv 2024
[7]

Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen

J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. 2021. https://api.semanticscholar.org/CorpusID:235458009 Lora: Low-rank adaptation of large language models . ArXiv, abs/2106.09685

arXiv 2021
[8]

Wenming Huang, Dengrui Hu, Zhenrong Deng, and Jianyun Nie. 2020. Named entity recognition for chinese judgment documents based on bilstm and crf. EURASIP Journal on Image and Video Processing, 2020:1--14

work page 2020

Show all 33 references

[9]

Guochao Jiang, Zepeng Ding, Yuchen Shi, and Deqing Yang. 2024. https://api.semanticscholar.org/CorpusID:269626679 P-icl: Point in-context learning for named entity recognition with large language models . ArXiv, abs/2405.04960

2024 arXiv
[10]

Imed Keraghel, Stanislas Morbieu, and Mohamed Nadif. 2024. Recent advances in named entity recognition: A comprehensive survey and comparative study

2024
[11]

Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. Unifiedqa: Crossing format boundaries with a single qa system. arXiv preprint arXiv:2005.00700

2020 arXiv
[12]

Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691

2021 arXiv
[13]

u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Informati...

2020
[14]

Peng Li, Tianxiang Sun, Qiong Tang, Hang Yan, Yuanbin Wu, Xuanjing Huang, and Xipeng Qiu. 2023 a . Codeie: Large code generation models are better few-shot information extractors. arXiv preprint arXiv:2305.05711

2023 arXiv
[15]

Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. 2019. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855

2019 arXiv
[16]

Zongxi Li, Xianming Li, Yuzhang Liu, Haoran Xie, Jing Li, Fu-lee Wang, Qing Li, and Xiaoqin Zhong. 2023 b . Label supervised llama finetuning. arXiv preprint arXiv:2310.01208

2023 arXiv
[17]

Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM computing surveys, 55(9):1--35

2023
[18]

Jinghui Lu, Rui Zhao, Brian Mac Namee, and Fei Tan. 2023. Punifiedner: A prompting-based unified ner system for diverse datasets. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 13327--13335

2023
[19]

Hanjun Luo, Yingbin Jin, Xuecheng Liu, Tong Shang, Ruizhe Chen, and Zuozhu Liu. 2024. Geic: Universal and multilingual named entity recognition with large language models. arXiv preprint arXiv:2409.11022

2024 arXiv
[20]

Shervin Malmasi, Anjie Fang, Besnik Fetahu, Sudipta Kar, and Oleg Rokhlenko. 2022. Multiconer: A large-scale multilingual dataset for complex named entity recognition. arXiv preprint arXiv:2208.14536

2022 arXiv
[21]

Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th annual meeting of the association for computational linguistics (volume 1: long papers), pages 1946--1958

2017
[22]

Zeqi Tan, Shen Huang, Zixia Jia, Jiong Cai, Yinghui Li, Weiming Lu, Yueting Zhuang, Kewei Tu, Pengjun Xie, Fei Huang, and 1 others. 2023. Damo-nlp at semeval-2023 task 2: A unified retrieval-augmented system for multilingual named entity recognition. arXiv preprint arXiv:2305.03688

2023 arXiv
[23]

Shuhe Wang, Xiaofei Sun, Xiaoya Li, Rongbin Ouyang, Fei Wu, Tianwei Zhang, Jiwei Li, and Guoyin Wang. 2023. Gpt-ner: Named entity recognition via large language models. arXiv preprint arXiv:2304.10428

2023 arXiv
[24]

Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, and 1 others. 2022. Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tas...

2022 arXiv
[25]

Yeming Wen and Swarat Chaudhuri. 2023. Batched low-rank adaptation of foundation models. arXiv preprint arXiv:2312.05677

2023 arXiv
[26]

Shuang Wu, Xiaoning Song, Zhenhua Feng, and Xiao-Jun Wu. 2022. Nflat: Non-flat-lattice transformer for chinese named entity recognition. arXiv preprint arXiv:2205.05832

2022 arXiv
[27]

An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

2024 arXiv
[28]

Urchade Zaratiana, Nadi Tomeh, Pierre Holat, and Thierry Charnois. 2023. Gliner: Generalist model for named entity recognition using bidirectional transformer. arXiv preprint arXiv:2311.08526

2023 arXiv
[29]

Ziyu Zhao, Leilei Gan, Guoyin Wang, Wangchunshu Zhou, Hongxia Yang, Kun Kuang, and Fei Wu. 2024. Loraretriever: Input-aware lora retrieval and composition for mixed tasks in the wild. arXiv preprint arXiv:2402.09997

2024 arXiv
[30]

Xingyu Zhu, Feifei Dai, Xiaoyan Gu, Bo Li, Meiou Zhang, and Weiping Wang. 2024 a . Gl-ner: Generation-aware large language models for few-shot named entity recognition. In International Conference on Artificial Neural Networks, pages 433--448. Springer

2024
[31]

Yuqi Zhu, Xiaohan Wang, Jing Chen, Shuofei Qiao, Yixin Ou, Yunzhi Yao, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2024 b . Llms for knowledge graph construction and reasoning: Recent capabilities and future opportunities. World Wide Web, 27(5):58

2024
[32]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[33]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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

[1] [1]

Ahmet Yusuf Alan, Enis Karaarslan, and \"O mer Aydin. 2024. A rag-based question answering system proposal for understanding islam: Mufassirqas llm. arXiv preprint arXiv:2401.15378

arXiv 2024

[2] [2]

Dhananjay Ashok and Zachary C Lipton. 2023. Promptner: Prompting for named entity recognition. arXiv preprint arXiv:2305.15444

arXiv 2023

[3] [3]

Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961

arXiv 2024

[4] [4]

Ruirui Chen, Chengwei Qin, Weifeng Jiang, and Dongkyu Choi. 2024. Is a large language model a good annotator for event extraction? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17772--17780

work page 2024

[5] [5]

Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116

arXiv 2019

[6] [6]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

arXiv 2024

[7] [7]

Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen

J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. 2021. https://api.semanticscholar.org/CorpusID:235458009 Lora: Low-rank adaptation of large language models . ArXiv, abs/2106.09685

arXiv 2021

[8] [8]

Wenming Huang, Dengrui Hu, Zhenrong Deng, and Jianyun Nie. 2020. Named entity recognition for chinese judgment documents based on bilstm and crf. EURASIP Journal on Image and Video Processing, 2020:1--14

work page 2020