Pith. sign in

REVIEW 4 major objections 4 minor 31 references

Large Language Models for Zero-Shot Multicultural Name Recognition

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A prompt-engineered fine-tuning framework for LLMs reaches 93.1% overall and 89.5% zero-shot accuracy on multicultural name recognition, beating Bi-LSTM baselines with cultural tags.

desk verdict A plausible recipe for LLM-based name-origin classification, but the headline zero-shot numbers are uninterpretable as reported. read the letter →

arxiv 2507.04149 v1 pith:DILR5ZYT submitted 2025-07-05 cs.CL

Thanakorn Phonchai , Surasakdi Siripong , Nicholas Patterson , Owen Campbell This is my paper
classification cs.CL
keywords multiculturalnamerecognitionzero-shotlearningprompt-engineeredfine-tuningadversarialdataaugmentationculturalknowledgegraphLoRAnamedentitylargelanguagemodels
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 tries to show that a large language model can be turned into a reliable classifier of the cultural origin of names, including names it was never fine-tuned on, by combining prompt engineering, low-rank adaptation (LoRA) as a parameter-efficient fine-tuning method, adversarial character-level augmentation, and explicit cultural knowledge injected from a knowledge graph. The headline results are 93.1% overall accuracy and 89.5% accuracy on a zero-shot set of unseen names, surpassing the best Bi-LSTM baseline with cultural tags by 2.8 and 4.3 percentage points, respectively. If true, this makes LLM-based name recognition a practical alternative to specialized sequence models: fine-tuning touches only 1.5 million parameters and takes about two hours. An ablation study attributes the zero-shot gain to both adversarial augmentation and cultural knowledge integration, which work better together than either does alone.

What carries the argument

The load-bearing object is the composed prompt $P(N) = [\text{Task Prefix}] \oplus K(N_C) \oplus N \oplus [\text{Instruction Suffix}]$, where $N$ is the input name and $K(N_C)$ is a natural-language snippet drawn from a cultural knowledge graph, for example "French names often feature hyphens and silent final consonants." The training objective is negative log-likelihood of the target label string, plus a $\lambda$-weighted term for adversarially augmented names. LoRA keeps the base model frozen and updates only low-rank matrices $B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$, so each adapted weight matrix adds $d r + r k$ trainable parameters. This prompt-plus-augmentation-plus-LoRA mechanism is what carries the claimed 89.5% zero-shot accuracy.

What would settle it

A concrete check is to probe the unfine-tuned base model on the zero-shot test names: if the base model alone classifies a large fraction of them correctly, or shows strong membership signals such as low perplexity and confident completions, then the zero-shot result mostly reflects pretraining memorization. A second check is to retrain the framework without the cultural-knowledge prompt at inference and verify whether the drop matches the ablation's 2.0-point prediction.

Watch

Extended reading notes

Core claim

The central claim is that the paper's PEFT framework (their acronym for Prompt-Engineered Fine-Tuning, which itself uses parameter-efficient LoRA adaptation) improves zero-shot multicultural name recognition by reframing classification as constrained generation: a composed prompt carries the name plus dynamically retrieved cultural context, and the model outputs a cultural label or "Not a Name". Adversarial augmentation generates typos, transpositions, and cross-cultural character fusions, filters them by a plausibility model, and adds them to the training loss. The method reports 93.1% accuracy on the full test set and 89.5% on names absent from the fine-tuning data, compared with 85.2% for the strongest Bi-LSTM baseline. In a human evaluation on 500 hard unseen names, the method scores 88.0% versus 91.0% for human majority vote, with 87.1% agreement with that vote. The paper's conclusion is that the combination of the LLM's implicit linguistic knowledge, explicit cultural guidance, and robustness training is what produces the gain.

Load-bearing premise

The load-bearing premise is that the zero-shot test names are genuinely new to the model, yet the paper only ensures they are absent from the fine-tuning data, not from the language model's original training corpus, so if the model memorized them during pretraining the 89.5% figure reflects memory rather than generalization.

Editorial extensions

If this is right

  • If the reported numbers hold, the framework improves zero-shot name recognition by 4.3 percentage points over the strongest cultural-tag Bi-LSTM baseline (89.5% versus 85.2%).
  • Ablations imply that both main components matter: removing adversarial augmentation costs 2.7 points, removing cultural knowledge integration costs 2.0 points, and removing both drops zero-shot accuracy from 89.5% to 84.1%.
  • The LoRA configuration keeps trainable parameters at 1.5 million versus 25 million for the baseline, with fine-tuning time of about 2.0 hours versus 1.5 hours, so the approach is feasible without large training budgets.
  • Per-category results show the largest F1 gains on African names (4.2 points) and Middle Eastern names (3.3 points), the categories where training data are scarcer.
  • Human evaluation indicates the model's judgments align with expert intuition: 87.1% agreement with the human majority vote on hard unseen names, versus 81.2% for the strongest baseline.

Reading between the lines

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

  • A stronger protocol would evaluate on synthetic names generated so their exact strings cannot appear in any pretraining corpus, which would separate genuine generalization from recall of memorized strings.
  • The same recipe of prompt fine-tuning, adversarial surface perturbations, and explicit knowledge injection should transfer to sibling tasks where surface form encodes a hidden category, such as surname-origin disambiguation in genealogical records or dialect attribution in short text.
  • The ablation pattern (2.7-point loss without augmentation, 2.0-point loss without knowledge, 5.4-point loss without both) hints that the two components are partially redundant, so a cost-benefit curve could tell practitioners whether one component alone is enough for most deployments.
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 / 4 minor

Summary. The paper proposes a Prompt-Engineered Fine-Tuning (PEFT) framework for multicultural name recognition, combining LoRA-based fine-tuning of an unnamed LLM with adversarial data augmentation and dynamic cultural knowledge graph prompts. The central claim is that this framework achieves 93.1% overall accuracy and 89.5% accuracy on a zero-shot test set of unseen names, outperforming Bi-LSTM baselines with cultural tags (Abstract, Tables I and II). The authors also present ablation results, a human evaluation, per-culture F1-scores, and efficiency comparisons.

Significance. If the experimental claims were fully supported, the proposed framework would be a useful contribution to multicultural name recognition, especially in leveraging LLM pretrained knowledge with prompt engineering. The paper contains a detailed method description with formal equations, several evaluation tables, and attempts at ablation and human evaluation. However, the current significance is limited by the absence of essential experimental details: the base LLM is not named, dataset size and composition are not given, hyperparameters are not reported, no variance or significance tests are provided, and neither code nor data are released. More critically, the zero-shot claim is not established because the test names may have been seen during LLM pretraining, so the reported 89.5% figure cannot be interpreted as evidence of generalization to unseen names.

major comments (4)
  1. [§IV.A, Tables I–II] The zero-shot claim is not established. The paper states that zero-shot names are 'chosen to be absent from the training data,' but it does not rule out their presence in the base LLM's pretraining corpus. Because LLMs are known to memorize named entities from web-scale text, the 89.5% zero-shot accuracy could reflect retrieval of memorized knowledge rather than generalization to genuinely unseen names. The decisive control experiment—querying the frozen base model on the exact zero-shot test set—is absent. Without this, the paper's central claim is unsupported.
  2. [§IV.A, Tables I–III and VIII] The experimental section omits information necessary to verify or reproduce the reported numbers: the specific base LLM is never identified, the dataset size (number of names per culture, train/validation/test splits) is not reported, hyperparameters (λ, τ, LoRA rank r, perturbation probabilities) are not given, and no standard deviations, confidence intervals, or significance tests accompany the accuracy differences claimed to be 'statistically significant.' These omissions are load-bearing because the main contribution is empirical, and the missing details make the comparison to baselines impossible to audit.
  3. [§I and §IV.B] The Introduction describes the results with prospective language: 'The expected results demonstrate' and 'our method is projected to achieve an overall accuracy of 93.1%.' This phrasing suggests the reported numbers may be anticipated rather than measured. The paper must clarify whether all results in Tables I–VIII are actual experimental outcomes; if so, the wording should be corrected, and if not, the empirical claims are not supported.
  4. [§III.B.1, Eq. (3)] The combined loss function has a sign error that reverses the intended optimization. Equation (2) defines L_NLL as a negative log-likelihood, but Eq. (3) adds a positive term λ Σ log P(...) for augmented data. Minimizing this combined loss would push the model to assign low probability to correct augmented labels, which is the opposite of the stated goal. The augmented term should be −λ Σ log P(...). This formal inconsistency affects the described training procedure.
minor comments (4)
  1. [References] References [1] and [3] are identical; the list should be de-duplicated and renumbered.
  2. [§IV.A] The text refers to 'as detailed in Section 2,' but the method is presented in Section III; the cross-reference is incorrect.
  3. [Table IV] The 'Agreement with Human Majority Vote' cell for the Human Annotators row is empty; either report the value (which by construction should be 100% for the majority vote itself) or replace it with a dash and explain.
  4. [§III.A] The notation K(NC) is used without defining NC; it should be explicit that NC is the cultural label associated with name N, and the cultural knowledge graph G and its construction are never described beyond a single French example.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical and the zero-shot validity concern is an external benchmark issue, not a derivation-level circular reduction.

full rationale

I walked the paper's claimed derivation chain and found no step in which a prediction or first-principles result is equivalent, by the paper's own equations or by construction, to its inputs. The method is a standard PEFT/LoRA fine-tuning pipeline with prompt engineering, adversarial augmentation, and knowledge-context insertion; the reported accuracies (93.1% overall, 89.5% zero-shot) are empirical measurements on a held-out test set, not quantities derived from fitted parameters. The zero-shot set is defined only as absent from the fine-tuning training data (Section IV.A), which raises a legitimate validity concern about pretraining memorization in the base LLM, but that is a benchmark-construction/external-validity issue, not a circularity internal to the derivation; no equation in the paper reduces the zero-shot result to the training objective. The paper contains no load-bearing self-citations: the reference list includes no authors overlapping with the present authors, and no uniqueness theorem or ansatz is imported from prior work by the same team. The cultural knowledge term K(NC) is injected during training using ground-truth labels, and the paper explicitly states that at inference this component can be omitted or replaced with a generic prompt, so there is no label leakage by construction in the reported test-time evaluation. The ablation study is an empirical removal test and the human evaluation is an independent comparison. The main weaknesses are reproducibility and the uncontrolled pretraining corpus, but these do not constitute circular reasoning, so the honest finding is no significant circularity (score 0).

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The method rests on three unfree assumptions: the LLM has useful latent cultural knowledge, the cultural knowledge graph is accurate and complete, and augmented names inherit labels correctly. No graph or data source is provided. The free parameters (lambda, tau, LoRA rank, perturbation probabilities) are all unreported, so the results cannot be reproduced.

free parameters (4)
  • lambda (augmentation loss weight) = not reported
    Controls the contribution of augmented data in Eq. 3; no value or schedule given.
  • tau (plausibility threshold) = not reported
    Filters augmented names by a character-level model; threshold never specified.
  • LoRA rank r = not reported
    Determines trainable parameter count in Eq. 5; no rank or target modules specified.
  • perturbation probabilities = not reported
    Probabilities for insertion, deletion, substitution, and transposition are mentioned but never set.
assumptions (3)
  • domain assumption The pre-trained LLM contains useful implicit cultural knowledge about names.
    The whole method relies on leveraging this latent knowledge, but the paper provides no evidence for its extent.
  • domain assumption The cultural knowledge graph G contains accurate and sufficiently complete facts for each culture.
    Only two example prompts are shown; no graph schema, source, or coverage is described.
  • domain assumption Augmented names that pass the plausibility filter inherit the original cultural label.
    This label inheritance is stated in Section III.B.2 but could be wrong for cross-cultural fusions.
invented entities (1)
  • Cultural Knowledge Graph G
    purpose: Provides natural language cultural context prompts during fine-tuning.
    No actual graph is defined, released, or evaluated; the paper only shows a couple of hand-written example rules, so this is an unspecified placeholder rather than a grounded resource.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models for Zero-Shot Multicultural Name Recognition." pith.science (2026). https://pith.science/paper/DILR5ZYT

@misc{pith2026250704149,
  author       = {Pith},
  title        = {Pith review of: Large Language Models for Zero-Shot Multicultural Name Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DILR5ZYT}},
  note         = {Machine review of arXiv:2507.04149}
}
read the original abstract

The robust and accurate recognition of multicultural names, particularly those not previously encountered, is a critical challenge in an increasingly globalized digital landscape. Traditional methods often falter when confronted with the vast diversity and novel permutations of names across different linguistic and cultural backgrounds. This paper introduces a novel framework, Prompt-Engineered Fine-Tuning (PEFT) for Large Language Models (LLMs) with Adversarial Data Augmentation and Cultural Knowledge Graph Integration, designed to significantly enhance zero-shot multicultural name recognition. Our approach leverages the powerful linguistic understanding of pre-trained LLMs, transforming the recognition task into a guided generation problem. Through meticulous prompt engineering, dynamic integration of explicit cultural knowledge derived from knowledge graphs, and the strategic application of adversarial data augmentation, we equip the LLM with an unprecedented ability to infer the cultural origin of unseen names. Extensive experiments demonstrate that our PEFT method consistently outperforms established deep learning baselines, including advanced Bi-LSTM models with cultural tags, achieving an impressive 93.1\% overall accuracy and a remarkable 89.5\% accuracy on challenging zero-shot name identification. An in-depth ablation study confirms the synergistic contribution of each component, while a human evaluation highlights our method's performance approaching human expert judgment. This work signifies a substantial leap in multicultural name recognition, offering a highly effective and scalable solution for real-world applications.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 17 canonical work pages

  1. [3]

    Deep neural architectures for joint named entity recognition and disambiguation,

    Q. Wang and M. Iwaihara, “Deep neural architectures for joint named entity recognition and disambiguation,” in IEEE International Conference on Big Data and Smart Computing, BigComp 2019, Kyoto, Japan, February 27 - March 2, 2019 . IEEE, 2019, pp. 1–4. [Online]. Available: https://doi.org/10.1109/BIGCOMP.2019.8679233

    work page arXiv 2019

  2. [2]

    Character level full convolution neural network for text classifi- cation,

    S. Li, “Character level full convolution neural network for text classifi- cation,” 2020

    work page 2020

  3. [4]

    Crowdsourcing the character of a place: Character-level convolutional networks for multilingual geographic text classification,

    B. Adams and G. McKenzie, “Crowdsourcing the character of a place: Character-level convolutional networks for multilingual geographic text classification,” Trans. GIS, vol. 22, no. 2, pp. 394–408, 2018. [Online]. Available: https://doi.org/10.1111/tgis.12317 9

    work page doi:10.1111/tgis.12317 2018

  4. [5]

    Cross- lingual name tagging and linking for 282 languages,

    X. Pan, B. Zhang, J. May, J. Nothman, K. Knight, and H. Ji, “Cross- lingual name tagging and linking for 282 languages,” in Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers , R. Barzilay and M. Kan, Eds. Association for Computational Linguistics,...

    doi:10.18653/v1/p17-1178 2017

  5. [6]

    Liberating seen classes: Boosting few-shot and zero-shot text classification via anchor generation and classification reframing,

    H. Liu, S. Zhao, X. Zhang, F. Zhang, W. Wang, F. Ma, H. Chen, H. Yu, and X. Zhang, “Liberating seen classes: Boosting few-shot and zero-shot text classification via anchor generation and classification reframing,” in Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intell...

    work page 2024

  6. [7]

    Named entity recog- nition for code mixed social media sentences,

    Y . Sharma, R. Bhargava, and B. V . Tadikonda, “Named entity recog- nition for code mixed social media sentences,” International Journal of Software Science and Computational Intelligence (IJSSCI) , vol. 13, no. 2, pp. 23–36, 2021

    work page 2021

  7. [8]

    Deep learning for multi- label learning: A comprehensive survey,

    A. N. Tarekegn, M. Ullah, and F. A. Cheikh, “Deep learning for multi- label learning: A comprehensive survey,” CoRR, vol. abs/2401.16549,

    arXiv

  8. [9]

    Score: Story coherence and retrieval enhancement for ai narratives,

    Q. Yi, Y . He, J. Wang, X. Song, S. Qian, X. Yuan, M. Zhang, L. Sun, K. Li, K. Lu et al., “Score: Story coherence and retrieval enhancement for ai narratives,” arXiv preprint arXiv:2503.23512 , 2025

    arXiv 2025

Show all 31 references
  1. [10]

    Divide-then-aggregate: An efficient tool learning method via parallel tool invocation,

    D. Zhu, W. Shi, Z. Shi, Z. Ren, S. Wang, L. Yan, and D. Yin, “Divide-then-aggregate: An efficient tool learning method via parallel tool invocation,” arXiv preprint arXiv:2501.12432 , 2025

    2025 arXiv

  2. [11]

    Weak to strong generalization for large language models with multi-capabilities,

    Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generalization for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2025. [Online]. Available: https://openreview.net/forum?id=N1vYivuSKq

    2025

  3. [12]

    A survey of textual cyber abuse detection using cutting-edge language models and large language models,

    J. A. Diaz-Garcia and J. P. Carvalho, “A survey of textual cyber abuse detection using cutting-edge language models and large language models,” CoRR, vol. abs/2501.05443, 2025. [Online]. Available: https://doi.org/10.48550/arXiv.2501.05443

    2025 doi arXiv

  4. [13]

    Large language models: a comprehensive survey of its applications, challenges, limitations, and future prospects,

    M. U. Hadi, R. Qureshi, A. Shah, M. Irfan, A. Zafar, M. B. Shaikh, N. Akhtar, J. Wu, S. Mirjalili et al. , “Large language models: a comprehensive survey of its applications, challenges, limitations, and future prospects,” Authorea Preprints, vol. 1, pp. 1–26, 2023

    2023

  5. [14]

    Mam: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration,

    Y . Zhou, L. Song, and J. Shen, “Mam: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration,” arXiv preprint arXiv:2506.19835 , 2025

    2025 arXiv

  6. [15]

    Large language models in medicine: Applications, challenges, and future directions,

    E. Yu, X. Chu, W. Zhang, X. Meng, Y . Yang, X. Ji, and C. Wu, “Large language models in medicine: Applications, challenges, and future directions,” International Journal of Medical Sciences , vol. 22, no. 11, p. 2792, 2025

    2025

  7. [16]

    Collaborative large language models for automated data extraction in living systematic reviews,

    M. A. Khan, U. Ayub, S. A. A. Naqvi, K. Z. R. Khakwani, Z. bin Riaz Sipra, A. Raina, S. Zhou, H. He, A. Saeidi, B. Hasan, R. B. Rumble, D. S. Bitterman, J. L. Warner, J. Zou, A. J. Tevaarwerk, K. Leventakos, K. L. Kehl, J. M. Palmer, M. H. Murad, C. Baral, and I. B. Riaz, “Col...

    2025 doi

  8. [17]

    Training medical large vision- language models with abnormal-aware feedback,

    Y . Zhou, L. Song, and J. Shen, “Training medical large vision- language models with abnormal-aware feedback,” arXiv preprint arXiv:2501.01377, 2025

    2025 arXiv

  9. [18]

    Enhancing code llms with reinforcement learning in code generation,

    J. Wang, Z. Zhang, Y . He, Y . Song, T. Shi, Y . Li, H. Xu, K. Wu, G. Qian, Q. Chen et al. , “Enhancing code llms with reinforcement learning in code generation,” arXiv preprint arXiv:2412.20367 , 2024

    2024 arXiv

  10. [19]

    Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,

    Y . Zhou, J. Yuan, and Q. Wang, “Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,” arXiv preprint arXiv:2505.24787 , 2025

    2025 arXiv

  11. [20]

    Less is more: Vision representation compression for efficient video generation with large language models,

    Y . Zhou, J. Zhang, G. Chen, J. Shen, and Y . Cheng, “Less is more: Vision representation compression for efficient video generation with large language models,” 2024

    2024

  12. [21]

    Vislinginstruct: Elevating zero-shot learning in multi-modal lan- guage models with autonomous instruction optimization,

    D. Zhu, X. Tang, W. Han, J. Lu, Y . Zhao, G. Xing, J. Wang, and D. Yin, “Vislinginstruct: Elevating zero-shot learning in multi-modal lan- guage models with autonomous instruction optimization,” arXiv preprint arXiv:2402.07398, 2024

    2024 arXiv

  13. [22]

    The debate over understanding in ai’s large language models,

    M. Mitchell and D. C. Krakauer, “The debate over understanding in ai’s large language models,” CoRR, vol. abs/2210.13966, 2022. [Online]. Available: https://doi.org/10.48550/arXiv.2210.13966

    2022 doi arXiv

  14. [23]

    Sda: simple discrete augmentation for contrastive sentence representation learning,

    D. Zhu, Z. Mao, J. Lu, R. Zhao, and F. Tan, “Sda: simple discrete augmentation for contrastive sentence representation learning,” arXiv preprint arXiv:2210.03963, 2022

    2022 arXiv

  15. [24]

    A recent survey of arabic named entity recognition on social media,

    B. A. B. Ali, S. Mihi, I. E. Bazi, and N. Laachfoubi, “A recent survey of arabic named entity recognition on social media,” Rev. d’Intelligence Artif. , vol. 34, no. 2, pp. 125–135, 2020. [Online]. Available: https://doi.org/10.18280/ria.340202

    2020 doi

  16. [25]

    An overview of named entity recognition,

    P. Sun, X. Yang, X. Zhao, and Z. Wang, “An overview of named entity recognition,” in 2018 International Conference on Asian Language Processing (IALP). IEEE, 2018, pp. 273–278

    2018

  17. [26]

    Named entity recognition with bidirectional lstm-cnns,

    J. P. C. Chiu and E. Nichols, “Named entity recognition with bidirectional lstm-cnns,” Trans. Assoc. Comput. Linguistics , vol. 4, pp. 357–370, 2016. [Online]. Available: https://doi.org/10.1162/tacl a 00104

    2016 doi

  18. [27]

    Character level and word level embedding with bidirectional lstm–dynamic recurrent neural net- work for biomedical named entity recognition from literature,

    S. Gajendran, D. Manjula, and V . Sugumaran, “Character level and word level embedding with bidirectional lstm–dynamic recurrent neural net- work for biomedical named entity recognition from literature,” Journal of Biomedical Informatics , vol. 112, p. 103609, 2020

    2020

  19. [28]

    Combining named entity recognition and emotion analysis of tweets for early warning of violent actions,

    M. El Barachi, S. S. Mathew, and M. AlKhatib, “Combining named entity recognition and emotion analysis of tweets for early warning of violent actions,” in 2022 7th International Conference on Smart and Sustainable Technologies (SpliTech). IEEE, 2022, pp. 1–6

    2022

  20. [29]

    Style-aware contrastive learning for multi-style image captioning,

    Y . Zhou and G. Long, “Style-aware contrastive learning for multi-style image captioning,” in Findings of the Association for Computational Linguistics: EACL 2023 , 2023, pp. 2257–2267

    2023

  21. [30]

    Stock market prediction with new generation deep contextualized word representations and deep learning models using user sentiments,

    D. Othan and Z. H. Kilimci, “Stock market prediction with new generation deep contextualized word representations and deep learning models using user sentiments,” in International Conference on INnovations in Intelligent SysTems and Applications, INISTA 2021, Kocaeli, Turkey, ...

    2021

  22. [31]

    Improving cross-modal alignment for text- guided image inpainting,

    Y . Zhou and G. Long, “Improving cross-modal alignment for text- guided image inpainting,” in Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , 2023, pp. 3445–3456. 10

    2023

  23. [2024]

    Available: https://doi.org/10.48550/arXiv.2401.16549

    [Online]. Available: https://doi.org/10.48550/arXiv.2401.16549

    doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.