Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Small Language Model Makes an Effective Long Text Extractor

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

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

Pith's one-line read A small span-based model, SeNER, claims state-of-the-art accuracy for extracting entities from long documents.

desk verdict A genuinely useful memory-saving mechanism for long-text NER, but the SOTA claim is contradicted by the paper's own table and the baseline comparison is apples-to-oranges. read the letter →

arxiv 2502.07286 v1 pith:AOQ52BP3 submitted 2025-02-11 cs.CL cs.AI

Yelin Chen , Fanjin Zhang , Jie Tang This is my paper · ORCID
classification cs.CLcs.AI
keywords long-textnamedentityrecognitionspan-basedNERarrowattentionLogN-ScalingBiSPAtoken-pairspantensorGPUmemoryefficiency
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

SeNER is a span-based named-entity-recognition model designed for inputs of thousands of tokens, a regime where ordinary span enumeration explodes GPU memory and where LLM prompting generates long entity spans inaccurately. The paper argues that two changes remove the bottleneck: an 'arrow attention' encoder—global attention only on the [CLS] token, local sliding-window attention on every other token, with LogN-Scaling to stabilise attention entropy—and a bidirectional sliding-window plus-shaped attention (BiSPA) that operates on a span tensor compressed to keep only spans of length at most 128. On three long-text NER datasets (Scholar-XL, SciREX, Profiling-07), the method reports state-of-the-art F1 while using a DeBERTa-V3-large backbone of roughly 0.3B parameters, and it supports input lengths about six times longer than the UTC-IE span-based baseline under the same hardware. The result matters because it suggests that long-text entity extraction, including multi-hundred-word 'entity blocks' such as work experience and awards, does not require a large generative model.

What carries the argument

The load-bearing object is the compressed token-pair span tensor together with the BiSPA mechanism. Standard span-based NER forms an $L\times L$ tensor of all start–end pairs, costing $O(L^2)$ memory and $O(L^3)$ interaction; SeNER instead keeps only spans whose length does not exceed the unilateral window $w'=128$, producing $S_h \in \mathbb{R}^{L\times w'\times c}$. BiSPA runs horizontal self-attention on $S_h$ and vertical self-attention on a transformed $S_v$, then fuses the two with an MLP and a $3\times 3$ CNN, reducing interaction complexity to $O(L(w')^2)$. The encoder side is arrow attention: the [CLS] token attends globally to all tokens, every other token attends within a sliding window, and LogN-Scaling stabilises [CLS] attention entropy as input length varies. This pair of mechanisms is what lets the model claim long-input capacity without full attention or full span enumeration.

What would settle it

On Scholar-XL and Profiling-07, compute exact-match recall restricted to ground-truth spans longer than 128 words; because SeNER never represents such spans, recall on that subset should be zero, and the SOTA claim would then hold only because very long entities are rare in the test sets. A second check: retrain the same model on a version of Profiling-07 where the 'Education info' and 'Contact info' long-type entities are oversampled, and see whether F1 collapses relative to a full-span baseline.

Watch

Extended reading notes

Core claim

The paper's central claim is that SeNER—a lightweight span-based extractor that replaces full bidirectional attention with arrow attention and replaces full plus-shaped interaction on the complete $L\times L$ span tensor with BiSPA on an $L\times 128$ compressed tensor—achieves state-of-the-art extraction accuracy on three long NER datasets while remaining GPU-memory-friendly. With this design, the model can encode texts of several thousand tokens, extract entity spans of more than one hundred tokens (awards, education and work-experience blocks), and do so with a small language model backbone rather than a large generative one. The authors further claim that SeNER supports inputs three times longer than CNN-NER and six times longer than UTC-IE on a single A100, and that removing either arrow attention or BiSPA leads to out-of-memory failures on the longer datasets.

Load-bearing premise

The load-bearing premise is that no target entity longer than 128 tokens matters enough to hurt the score, because all longer candidate spans are discarded before classification—yet the datasets contain ground-truth entities of up to 480 words.

Editorial extensions

If this is right

  • Span-based NER can move from sentence-level to document-level inputs of several thousand tokens on a single 80GB GPU without sacrificing accuracy.
  • Entity blocks of more than 100 tokens—awards, education and work-experience entries—are extractable by a roughly 0.3B-parameter encoder, a regime where generation-based LLM methods are reported to fail.
  • The memory savings come from both modules: removing arrow attention or BiSPA triggers OOM on SciREX and Profiling-07, so long-input capacity depends on the joint design.
  • SeNER runs inference about as fast as CNN-NER and about 20% faster than UTC-IE, while supporting three times and six times longer inputs, respectively.
  • Longer input context helps short-entity NER too, since extending SeNER's input length on SciREX improves F1 by about 1.2 points.

Reading between the lines

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

  • Editorial inference: the hard 128-token span cutoff means SeNER's advantage is benchmark-specific; a corpus with frequent 200+ token entities would expose a recall ceiling that the current datasets only hint at in the Education Experience and Work Experience categories.
  • Editorial inference: the arrow-attention design, with [CLS] as a global attention sink and LogN-Scaling for entropy stability, is a general recipe that could carry over to other long-document encoders, not just NER span tensors.
  • Editorial inference: a natural extension is a hierarchical or two-pass span proposal that first finds coarse regions with BiSPA and then re-ranks long spans, which would remove the length cap while keeping memory subquadratic.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes SeNER, a lightweight span-based NER method for extracting entities from long texts. SeNER combines arrow attention with LogN-Scaling on the [CLS] token for efficient long-input encoding, and a bidirectional sliding-window plus-shaped attention (BiSPA) mechanism to reduce the computational cost of token-pair span interaction from O(L^3) to O(L(w')^2). The authors evaluate SeNER on three long NER datasets (Scholar-XL, SciREX, Profiling-07), reporting F1 scores competitive with or better than existing span-based and generation-based baselines, and claim state-of-the-art extraction accuracy on all three. They also report ablation studies supporting each component and an analysis of maximum supported input length and inference time.

Significance. If the results are validated, SeNER is a useful contribution to long-text NER: it offers a practical way to apply span-based methods to inputs of several thousand tokens while keeping memory and compute manageable, and it demonstrates that a relatively small model can outperform much larger generation-based models on this task. The complexity reduction is clearly derived, the ablations are informative, and the code is provided for reproducibility. However, the headline claim of state-of-the-art on three datasets is contradicted by the paper's own Table 2 on Profiling-07, where SeNER's F1 (67.34) is below UTC-IE (67.43). In addition, the baseline evaluation is asymmetric: baselines receive at most 512 input tokens while SeNER receives up to 5120, which confounds the reported accuracy gains. These issues undermine the central claim as stated, although the underlying method appears sound and the problems are fixable with a revised comparison and more careful wording.

major comments (3)
  1. [Abstract; Conclusion; Table 2] The paper claims 'state-of-the-art extraction accuracy on three long NER datasets' in both the Abstract and the Conclusion, but Table 2 shows that on Profiling-07 SeNER achieves an F1 of 67.34, while UTC-IE achieves 67.43. SeNER is therefore second-best, not state-of-the-art, on that dataset. This is an internal inconsistency between the reported numbers and the headline claim. The claim should be revised to 'state-of-the-art on two datasets and competitive on Profiling-07' or equivalent.
  2. [Experimental Setup; Detailed Experimental Setup] The comparison is asymmetric: the paper states that for baseline methods 'the maximum length of the input text is set to 512' with a sliding window, while SeNER uses full text on Scholar-XL and inputs up to 5120 tokens on the other datasets. Because SeNER sees the full document context and baselines see only 512-token chunks, the reported F1 differences could reflect input length rather than architectural superiority. The authors should include a controlled experiment where baselines receive the same input length as SeNER, or where SeNER is truncated to 512 tokens, to isolate the effect of the model design.
  3. [Token-Pair Span Interaction Module; Eq. (9); Eq. (10)] The method description is ambiguous about how spans longer than the unilateral window w' (set to 128) are handled. The text says 'we propose preserving only the hidden features of spans whose lengths do not exceed w′' and compresses S to Sh in R^(L x w' x c), but then Eq. (9) 'recovers' S'' to the full L x L size and Eq. (10) adds the original S as a residual. If long spans (beyond w') are still scored from the Biaffine output S, then the claim that such candidates are 'discarded' is inaccurate, and the memory savings are not as clear as suggested. If long spans are instead zeroed in the final prediction, then SeNER cannot extract entities longer than 128 words, which conflicts with the motivation of extracting long entities (Table 1 reports entity maximum lengths of 480 and 307 on Scholar-XL and Profiling-07). The authors should clarify the forward pass for spans longer than w' and provide a recall breakdown by entity length.
minor comments (5)
  1. [Long Input Encoding] Typo: 'incuring' should be 'incurring'.
  2. [Entity Types] Typo: 'The Scholarst-XL dataset' should be 'The Scholar-XL dataset'.
  3. [Eq. (3)] The LogN-Scaling factor is written as 'log512L sqrt(d)', which is ambiguous; it should be expressed as (log_512 L) / sqrt(d) or with clear parentheses and a division sign.
  4. [Throughout] Several instances of 'O(L2)' and 'O(L3)' lack superscripts; these should be typeset as O(L^2) and O(L^3) for clarity.
  5. [Table 3] The ablation rows 'w/o Arrow' and 'w/o BiSPA' report OOM on two datasets, so the claimed benefit of these components on those datasets is not directly measured; the authors should state this explicitly in the text, as they do implicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: SeNER is an empirical system combining cited components and evaluated on public benchmarks; the SOTA wording is contradicted by Table 2 on Profiling-07, but that is a correctness/reporting issue, not a circular derivation.

full rationale

SeNER's design is not derived from its target results. The encoder (arrow attention with LogN-Scaling) and the token-pair interaction module (BiSPA) are defined by explicit equations (Eq. 1-13) and trained with binary cross-entropy on annotated spans; no component is defined in terms of the F1 scores it is claimed to achieve. Components are taken from prior work (UTC-IE plus-shaped attention, LogN-Scaling from Su 2021, LoRA, whole word masking) and tested against public baselines. The two benchmarks Scholar-XL and Profiling-07 originate from the authors' group, and hyperparameters are tuned on their validation sets; this is a data-familiarity/overfitting concern, not circularity, because the labels and baseline numbers are external to the model's derivation. The paper explicitly preserves only spans with length at most w'=128 while reporting entity max lengths of 480 and 307 words, and Figure 6 admits SeNER falls behind on the longest entity types; this is a real limitation of the approximation, not a circular step. The abstract's "state-of-the-art on three datasets" wording is contradicted by Table 2 on Profiling-07 (UTC-IE 67.43 vs SeNER 67.34), and baseline truncation at 512 tokens versus SeNER's 5120 creates an asymmetric comparison; these are correctness/reporting issues outside the circularity definition. No equation in the paper reduces to its input, and no load-bearing claim rests solely on an unverified self-citation.

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

The central claim rests on a small number of hyperparameters and domain assumptions; no new physical or formal entities are introduced. The most consequential choice is the 128-token span window, which trades recall on very long entities for memory.

free parameters (4)
  • Arrow attention window size w = 128
    Unilateral window for local tokens in the PLM encoder; chosen from {32, 64, 128, 256, 512} by validation F1 (Hyper-parameter Search; Figure 8).
  • BiSPA window size w' = 128
    Maximum stored span length in the token-pair span tensor; chosen from the same search range; directly determines which entities can be extracted.
  • LogN-Scaling base (512) = 512
    Reference length in the scaling factor log_{512} L / sqrt(d) applied to the [CLS] attention; chosen so that at L=512 the scaling matches standard attention.
  • LoRA rank = 8
    LoRA rank for Q and V projections; chosen in Experimental Setup without a reported search.
assumptions (4)
  • domain assumption A single global [CLS] token plus local sliding-window attention conveys sufficient global context for long-text NER.
    Invoked in Long Input Encoding; the model replaces full attention with arrow attention, and the only global path is the [CLS] token.
  • domain assumption Candidate spans longer than w'=128 can be discarded without materially harming the benchmark F1.
    Invoked in Token-Pair Span Interaction Module and Experimental Setup; contradicts the observed max entity lengths, so it is a risky but load-bearing assumption.
  • domain assumption LogN-Scaling on the [CLS] token stabilizes entropy across input lengths.
    Borrowed from Su (2021); asserted in Long Input Encoding and verified only through the ablation w/o LogN, not through an entropy measurement.
  • domain assumption The three datasets (Scholar-XL, SciREX, Profiling-07) are a representative sample of long-text NER.
    Used in Datasets and Main Results; two of the three originate from the authors' research group, which could inflate results through implicit dataset knowledge.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Small Language Model Makes an Effective Long Text Extractor." pith.science (2026). https://pith.science/paper/AOQ52BP3

@misc{pith2026250207286,
  author       = {Pith},
  title        = {Pith review of: Small Language Model Makes an Effective Long Text Extractor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AOQ52BP3}},
  note         = {Machine review of arXiv:2502.07286}
}
read the original abstract

Named Entity Recognition (NER) is a fundamental problem in natural language processing (NLP). However, the task of extracting longer entity spans (e.g., awards) from extended texts (e.g., homepages) is barely explored. Current NER methods predominantly fall into two categories: span-based methods and generation-based methods. Span-based methods require the enumeration of all possible token-pair spans, followed by classification on each span, resulting in substantial redundant computations and excessive GPU memory usage. In contrast, generation-based methods involve prompting or fine-tuning large language models (LLMs) to adapt to downstream NER tasks. However, these methods struggle with the accurate generation of longer spans and often incur significant time costs for effective fine-tuning. To address these challenges, this paper introduces a lightweight span-based NER method called SeNER, which incorporates a bidirectional arrow attention mechanism coupled with LogN-Scaling on the [CLS] token to embed long texts effectively, and comprises a novel bidirectional sliding-window plus-shaped attention (BiSPA) mechanism to reduce redundant candidate token-pair spans significantly and model interactions between token-pair spans simultaneously. Extensive experiments demonstrate that our method achieves state-of-the-art extraction accuracy on three long NER datasets and is capable of extracting entities from long texts in a GPU-memory-friendly manner. Code: https://github.com/THUDM/scholar-profiling/tree/main/sener

Figures

Figures reproduced from arXiv: 2502.07286 by the authors.

Figure 1
Figure 1. An example of entity/attribute extraction from an [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Performance of entity recognition with respect to input length and the number of model parameters of NER methods. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. An overview of the SeNER model. a. Full Attention b. Sliding Window Attention c. Arrow Attention [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustration of arrow attention, full attention, and [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: Diagram of the transformation for the token-pair [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Performance of different entity types on the [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Blue bar: Maximum input length comparison of [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Impact of window sizes of arrow attention and [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 30 canonical work pages

  1. [1]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Anthropic. 2024. The Claude 3 Model Family: Opus, Sonnet, Haiku

    work page 2024

  3. [3]

    Ashok, D.; and Lipton, Z. C. 2023. Promptner: Prompting for named entity recognition. arXiv preprint arXiv:2305.15444

    arXiv 2023

  4. [4]

    E.; and Cohan, A

    Beltagy, I.; Peters, M. E.; and Cohan, A. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150

    arXiv 2020

  5. [5]

    Cui, Y.; Che, W.; Liu, T.; Qin, B.; and Yang, Z. 2021. Pre-training with whole word masking for chinese bert. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29: 3504--3514

    work page 2021

  6. [6]

    S.; Ceder, G.; Persson, K

    Dagdelen, J.; Dunn, A.; Lee, S.; Walker, N.; Rosen, A. S.; Ceder, G.; Persson, K. A.; and Jain, A. 2024. Structured information extraction from scientific text with large language models. Nature Communications, 15(1): 1418

    work page 2024

  7. [7]

    Dao, T.; Fu, D.; Ermon, S.; Rudra, A.; and R \'e , C. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35: 16344--16359

    work page 2022

  8. [8]

    Dozat, T.; and Manning, C. D. 2017. Deep Biaffine Attention for Neural Dependency Parsing. In International Conference on Learning Representations

    work page 2017

Show all 52 references
  1. [9]

    Gu, X.; Yang, H.; Tang, J.; Zhang, J.; Zhang, F.; Liu, D.; Hall, W.; and Fu, X. 2018. Profiling Web users using big data. Social Network Analysis and Mining, 8: 1--17

    2018

  2. [10]

    He, P.; Gao, J.; and Chen, W. 2023. DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing

    2023

  3. [11]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  4. [12]

    Jain, S.; van Zuylen, M.; Hajishirzi, H.; and Beltagy, I. 2020. S ci REX : A Challenge Dataset for Document-Level Information Extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 7506--7516

    2020

  5. [13]

    Jiang, G.; Luo, Z.; Shi, Y.; Wang, D.; Liang, J.; and Yang, D. 2024. ToNER: Type-oriented Named Entity Recognition with Generative Language Model. arXiv preprint arXiv:2404.09145

    2024 arXiv

  6. [14]

    Li, J.; Fei, H.; Liu, J.; Wu, S.; Zhang, M.; Teng, C.; Ji, D.; and Li, F. 2022. Unified named entity recognition as word-word relation classification. In proceedings of the AAAI conference on artificial intelligence, 10965--10973

    2022

  7. [15]

    Li, X.; Feng, J.; Meng, Y.; Han, Q.; Wu, F.; and Li, J. 2019. A unified MRC framework for named entity recognition. arXiv preprint arXiv:1910.11476

    2019 arXiv

  8. [16]

    Loshchilov, I.; Hutter, F.; et al. 2017. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101

    2017 arXiv

  9. [17]

    Lou, C.; Yang, S.; and Tu, K. 2022. Nested Named Entity Recognition as Latent Lexicalized Constituency Parsing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 6183--6198

    2022

  10. [18]

    Lu, Y.; Liu, Q.; Dai, D.; Xiao, X.; Lin, H.; Han, X.; Sun, L.; and Wu, H. 2022. Unified structure generation for universal information extraction. arXiv preprint arXiv:2203.12277

    2022 arXiv

  11. [19]

    Ma, X.; and Hovy, E. 2016. End-to-end sequence labeling via bi-directional lstm-cnns-crf. arXiv preprint arXiv:1603.01354

    2016 arXiv

  12. [20]

    Miwa, M.; and Bansal, M. 2016. End-to-end relation extraction using lstms on sequences and tree structures. arXiv preprint arXiv:1601.00770

    2016 arXiv

  13. [21]

    Moll \'a , D.; Van Zaanen, M.; and Smith, D. 2006. Named entity recognition for question answering. In Australasian Language Technology Association Workshop, 51--58

    2006

  14. [22]

    Qi, Y.; Peng, H.; Wang, X.; Xu, B.; Hou, L.; and Li, J. 2024. ADELIE: Aligning Large Language Models on Information Extraction. arXiv preprint arXiv:2405.05008

    2024 arXiv

  15. [23]

    Rajbhandari, S.; Rasley, J.; Ruwase, O.; and He, Y. 2020. Zero: Memory optimizations toward training trillion parameter models. In International Conference for High Performance Computing, Networking, Storage and Analysis, 1--16

    2020

  16. [24]

    E.; Adi, Y.; Liu, J.; Remez, T.; Rapin, J.; et al

    Roziere, B.; Gehring, J.; Gloeckle, F.; Sootla, S.; Gat, I.; Tan, X. E.; Adi, Y.; Liu, J.; Remez, T.; Rapin, J.; et al. 2023. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950

    2023 arXiv

  17. [25]

    L.; Rigau, G.; and Agirre, E

    Sainz, O.; Garc \' a-Ferrero, I.; Agerri, R.; de Lacalle, O. L.; Rigau, G.; and Agirre, E. 2023. Gollie: Annotation guidelines improve zero-shot information-extraction. arXiv preprint arXiv:2310.03668

    2023 arXiv

  18. [26]

    Schiaffino, S.; and Amandi, A. 2009. Intelligent user profiling. In Artificial Intelligence An International Perspective: An International Perspective, 193--216

    2009

  19. [27]

    Shen, Y.; Song, K.; Tan, X.; Li, D.; Lu, W.; and Zhuang, Y. 2023. Diffusionner: Boundary diffusion for named entity recognition. arXiv preprint arXiv:2305.13298

    2023 arXiv

  20. [28]

    Shen, Y.; Wang, X.; Tan, Z.; Xu, G.; Xie, P.; Huang, F.; Lu, W.; and Zhuang, Y. 2022. Parallel instance query network for named entity recognition. arXiv preprint arXiv:2203.10545

    2022 arXiv

  21. [29]

    Strakov \'a , J.; Straka, M.; and Haji c , J. 2019. Neural architectures for nested NER through linearization. arXiv preprint arXiv:1908.06926

    2019 arXiv

  22. [30]

    Su, J. 2021. Analyzing the Scale Operation of Attention from the Perspective of Entropy Invariance

    2021

  23. [31]

    Su, J.; Ahmed, M.; Lu, Y.; Pan, S.; Bo, W.; and Liu, Y. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568: 127063

    2024

  24. [32]

    Su, J.; Murtadha, A.; Pan, S.; Hou, J.; Sun, J.; Huang, W.; Wen, B.; and Liu, Y. 2022. Global pointer: Novel efficient span-based approach for named entity recognition. arXiv preprint arXiv:2208.03054

    2022 arXiv

  25. [33]

    Tan, Z.; Shen, Y.; Zhang, S.; Lu, W.; and Zhuang, Y. 2021. A sequence-to-set network for nested named entity recognition. arXiv preprint arXiv:2105.08901

    2021 arXiv

  26. [34]

    Tang, J.; Zhang, D.; and Yao, L. 2007. Social network extraction of academic researchers. In Seventh IEEE International Conference on Data Mining, 292--301

    2007

  27. [35]

    Tang, J.; Zhang, J.; Yao, L.; Li, J.; Zhang, L.; and Su, Z. 2008. Arnetminer: extraction and mining of academic social networks. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining, 990--998

    2008

  28. [36]

    Wang, S.; Sun, X.; Li, X.; Ouyang, R.; Wu, F.; Zhang, T.; Li, J.; and Wang, G. 2023 a . Gpt-ner: Named entity recognition via large language models. arXiv preprint arXiv:2304.10428

    2023 arXiv

  29. [37]

    Wang, X.; Zhou, W.; Zu, C.; Xia, H.; Chen, T.; Zhang, Y.; Zheng, R.; Ye, J.; Zhang, Q.; Gui, T.; et al. 2023 b . Instructuie: Multi-task instruction tuning for unified information extraction. arXiv preprint arXiv:2304.08085

    2023 arXiv

  30. [38]

    Xiao, G.; Tian, Y.; Chen, B.; Han, S.; and Lewis, M. 2023. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453

    2023 arXiv

  31. [39]

    Xie, T.; Li, Q.; Zhang, J.; Zhang, Y.; Liu, Z.; and Wang, H. 2023. Empirical Study of Zero-Shot NER with C hat GPT . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 7935--7956

    2023

  32. [40]

    Xu, B.; Xu, Y.; Liang, J.; Xie, C.; Liang, B.; Cui, W.; and Xiao, Y. 2017. CN-DBpedia: A never-ending Chinese knowledge extraction system. In International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems, 428--438

    2017

  33. [41]

    Yan, H.; Deng, B.; Li, X.; and Qiu, X. 2019. TENER: adapting transformer encoder for named entity recognition. arXiv preprint arXiv:1911.04474

    2019 arXiv

  34. [42]

    Yan, H.; Gui, T.; Dai, J.; Guo, Q.; Zhang, Z.; and Qiu, X. 2021. A unified generative framework for various NER subtasks. arXiv preprint arXiv:2106.01223

    2021 arXiv

  35. [43]

    Yan, H.; Sun, Y.; Li, X.; and Qiu, X. 2023 a . An Embarrassingly Easy but Strong Baseline for Nested Named Entity Recognition. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, 1442--1452

    2023

  36. [44]

    Yan, H.; Sun, Y.; Li, X.; Zhou, Y.; Huang, X.; and Qiu, X. 2023 b . UTC - IE : A Unified Token-pair Classification Architecture for Information Extraction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, 4096--4122

    2023

  37. [45]

    Yang, S.; and Tu, K. 2022. Bottom-Up Constituency Parsing and Nested Named Entity Recognition with Pointer Networks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2403--2416

    2022

  38. [46]

    Yuan, Z.; Tan, C.; Huang, S.; and Huang, F. 2022. Fusing Heterogeneous Factors with Triaffine Mechanism for Nested Named Entity Recognition. In Findings of the Association for Computational Linguistics, 3174--3186

    2022

  39. [47]

    A.; Ainslie, J.; Alberti, C.; Ontanon, S.; Pham, P.; Ravula, A.; Wang, Q.; Yang, L.; et al

    Zaheer, M.; Guruganesh, G.; Dubey, K. A.; Ainslie, J.; Alberti, C.; Ontanon, S.; Pham, P.; Ravula, A.; Wang, Q.; Yang, L.; et al. 2020. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33: 17283--17297

    2020

  40. [48]

    Zhang, F.; Shi, S.; Zhu, Y.; Chen, B.; Cen, Y.; Yu, J.; Chen, Y.; Wang, L.; Zhao, Q.; Cheng, Y.; et al. 2024. OAG-Bench: A Human-Curated Benchmark for Academic Graph Mining. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2024

  41. [49]

    X.; Zhou, K.; Li, J.; Tang, T.; Wang, X.; Hou, Y.; Min, Y.; Zhang, B.; Zhang, J.; Dong, Z.; et al

    Zhao, W. X.; Zhou, K.; Li, J.; Tang, T.; Wang, X.; Hou, Y.; Min, Y.; Zhang, B.; Zhang, J.; Dong, Z.; et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223

    2023 arXiv

  42. [50]

    Zhu, E.; and Li, J. 2022. Boundary smoothing for named entity recognition. arXiv preprint arXiv:2204.12031

    2022 arXiv

  43. [51]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  44. [52]

    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 8, 2026 · model on record in the stance chip above.