Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Examining and Adapting Time for Multilingual Classification via Mixture of Temporal Experts

T0 review · 4 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 The paper claims that a mixture of temporal experts, MoTE, improves multilingual classification over time by treating time as domains and adapting to future data without target labels.

desk verdict Genuine attempt at multilingual temporal adaptation with real gains, but the temporal-expert interpretation rests on cluster-to-time alignment the paper never verifies. read the letter →

arxiv 2502.08825 v1 pith:XS6QLZM2 submitted 2025-02-12 cs.CL

Weisi Liu , Guangzeng Han , Xiaolei Huang This is my paper · ORCID
classification cs.CL
keywords temporaldomainadaptationmultilingualclassificationmixtureofexpertstimeshiftEURLEXXLM-RoBERTatextunsupervised
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

Time drift degrades multilingual text classification, and the paper proposes a mixture-of-experts model, MoTE, that adapts a classifier to future time periods without any labels from those periods. MoTE treats each source time interval as a cluster, measures how far a target document sits from each cluster centroid, and routes the document to specialized temporal experts. On the 23-language EURLEX legal corpus the paper reports macro-F1 gains of 6.79 to 17.64 absolute points over the ChronosLex baseline, and on four review corpora it reports gains of 5.13 to 8.17 F1 points over the best baseline. The authors conclude that time should be treated as a domain and that temporal shifts can be learned from data distribution alone.

What carries the argument

The central machinery is the Clustering-based Shift Evaluator paired with a Temporal Router Network. The shift evaluator encodes source texts with XLM-RoBERTa, applies K-Means with T clusters where T matches the number of source time domains, labels each source point with its nearest cluster to create a warmup set, and computes shift vectors $v_{ij} = z_i - c_j$ between a target embedding and each cluster centroid $c_j$. The router is a top-K gating network $G(z) = \mathrm{TopK}(\mathrm{softmax}(W_g z), K)$ trained with an auxiliary load-balancing loss, and each expert is a transformer block; the shift vector is concatenated with the expert's $[cls]$ output before classification, and the final prediction is $\frac{1}{|K|}\sum_{k \in T} G_k(z^*) p_k(x)$. The warmup labels teach the router which expert corresponds to which cluster, so at test time routing can be driven by the shift vectors alone.

What would settle it

Run MoTE on EURLEX after randomly permuting the cluster labels used in router warmup, or set T to the true number of source time periods and measure the agreement between cluster assignments and actual source years; if macro-F1 gains over ChronosLex persist with shuffled labels, or if the cluster-time agreement is near chance, the shift vectors are not carrying temporal information.

Watch

Extended reading notes

Core claim

The paper's central claim is that temporal shifts in multilingual text can be made explicit and exploited: classifiers trained on earlier time periods lose accuracy on later ones, with larger drops over longer intervals, and MoTE recovers most of that loss without target labels. The method runs K-Means on XLM-RoBERTa embeddings of the source documents to form T clusters, treats each cluster as a pseudo-temporal domain, and represents each target document by its difference from each cluster centroid. A top-K gating router dispatches each document to the relevant experts, each a transformer block with a classification layer, and the experts receive the shift vectors as extra input before the final routing-weighted prediction. The paper reports that this outperforms ChronosLex, Self-Labeling, and Anti-CF across 23 EURLEX languages and four review languages, and that ablations show both the router and the shift evaluator contribute to the gain.

Load-bearing premise

The load-bearing premise is that grouping the source texts into T clusters by embedding similarity yields clusters that correspond to the T time periods, so the distance from a new text to a cluster center measures temporal drift; if the clusters track topic, style, or language instead, the experts are not truly temporal.

Editorial extensions

If this is right

  • A single multilingual architecture can adapt to temporal drift across short informal reviews and long formal legal documents without any labels from the target period.
  • On the 23-language EURLEX benchmark, macro-F1 gains over the ChronosLex baseline range from 6.79 to 17.64 absolute points, with the largest relative gain on Maltese.
  • On four review corpora, MoTE beats the best baseline by 5.13 to 8.17 F1 points and reduces fairness disparities relative to the source model.
  • Ablations show that removing the router or the shift evaluator lowers macro-F1 by up to 10.14 points, confirming both modules carry weight.
  • The gains are largest on class-imbalanced data such as the German review corpus, suggesting temporal expert routing particularly helps minority classes.

Reading between the lines

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

  • The temporal interpretation depends on K-Means clusters actually aligning with time; a useful check is to measure the agreement between cluster assignments and true source time periods, for instance with adjusted mutual information, since low agreement would mean the shift vectors encode topic or register rather than drift.
  • The paper never states how the 1958-2010 EURLEX training split is divided into T source domains, so the exact number of temporal experts for the legal benchmark is not reproducible from the text alone.
  • Some of the reported gain may come from the added capacity of a mixture of experts rather than from temporal structure; a control with randomly initialized experts or shuffled time labels would isolate the temporal component.
  • The shift-vector mechanism could transfer to other timestamped domains such as clinical notes or social media streams, wherever future labels are unavailable.
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 studies temporal distribution shift in multilingual text classification. It first reports cross-time-domain performance drops for XLM-RoBERTa on four review corpora (English, French, German, Danish). It then proposes MoTE, which clusters source XLM-R embeddings with K-Means into T clusters, uses cluster assignments to warm up a router, computes shift vectors as differences between target representations and cluster centroids, and routes inputs to a mixture of transformer-block experts. MoTE is evaluated on the review corpora and on 23 languages of EURLEX against ChronosLex, Anti-CF, Self-Labeling, and a source model, with additional ablations. The paper reports large macro-F1 gains in both settings.

Significance. If the reported results are reproducible, the paper would make a useful empirical contribution: temporal adaptation for multilingual classification is underexplored, the evaluation covers 23 languages, the code is released, and the ablation structure is sensible. However, the evidence as presented is not yet reliable: the motivating temporal-effect observation contradicts the paper's own tables, the main quantitative claims contain arithmetical inconsistencies, no statistical significance or variance is reported, and the method's temporal interpretation rests on an unvalidated assumption that K-Means clusters correspond to time periods. These issues are load-bearing for the central claim and require correction before the contribution can be assessed.

major comments (4)
  1. [§3.1, Algorithm 1; §4.3] The central claim that MoTE adapts classifiers to temporal drift requires that the T K-Means clusters index time periods, because the cluster labels L are used to warm up the router and the shift vectors v_ij are the only time signal given to the experts. No evidence is provided that the clusters correspond to time rather than to topic, register, or language: the paper does not report cluster-to-time alignment metrics (e.g., NMI or purity against time intervals), visualizations of cluster assignments over time, or a control experiment with shuffled or non-temporal cluster labels. Moreover, T is never specified for EURLEX: Section 4.3 says only that the training split (1958-2010) is used as the source time domain, but Section 3 requires dividing the source into T time-ordered domains. This ambiguity is load-bearing for the temporal interpretation.
  2. [§5, Tables 2 and 3] The headline quantitative claims are not supported by the reported tables. Section 5 states macro-F1 improvements 'ranging from 6.79% to 17.64%' over ChronosLex on EURLEX, but Table 3 lists Finnish at 20.76 and Croatian at 8.17, so the upper bound should be 20.76. The same paragraph states MoTE beats the best baseline by '5.13% to 8.17%' F1 on reviews, while Table 2 gives French MoTE 44.97 vs ChronosLex 41.48, a 3.49-point difference. These inconsistencies must be corrected before the reported gains can be assessed.
  3. [§5; §4.3; Appendix E] No significance testing, confidence intervals, or standard deviations are reported for any table, despite the paper claiming that MoTE 'significantly improves' model generalizability and despite Appendix E stating that all experiments are run with three seeds. Given that the central contribution is empirical and the gains are large, the paper should report per-seed variation or paired tests across seeds and languages; otherwise the improvements over baselines cannot be distinguished from optimization noise.
  4. [§2, Figure 1] The text's key motivating evidence for temporal effects is inconsistent with Figure 1. The manuscript states that 'English, French, and Danish classifiers drop more than 3% when training on D1 and testing on D4'; the corresponding F1-change entries in Figure 1 are +0.14, +2.15, and -1.72, and the AUC-change entries are -0.32, +3.90, and -1.64. Either the figure or the text is wrong, and this observation motivates the entire paper.
minor comments (5)
  1. [Table 2 caption] The sentence 'For Fair, negative values indicate worse performance as lower is better' is self-contradictory; if lower Fair scores are better, then negative deltas indicate better performance, and the caption should say so.
  2. [§2, Table 1] The text says the label imbalance ratio is 'the number of minority labels divided by the number of majority labels,' but Table 1 and Table 5 define it as majority over minority, and the reported values (e.g., 33.7, 50.1) are only possible under the latter definition. Please define the ratio consistently.
  3. [§4.3 and Appendix E] The maximum token length for legal data is given as 512 in Section 4.3 but as 521 in Appendix E; also Section 4.3 mentions early stopping while Appendix E says checkpoints of the last epoch are saved. Please align these experimental details.
  4. [Throughout] There are several typos and minor wording errors: 'reciew' (Appendix C), 'lagal' (Conclusion), 'cross-enropy' (Appendix E.2), 'absoluate' (Table 3 caption), and 'The The label imbalance ratio' (Section 2).
  5. [Table 5 caption] The caption mentions 'train, development, and test splits,' but the table lists only train/test document counts; clarify whether a development split exists and how it was used.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MoTE's improvements are evaluated on held-out target labels against external baselines; the cluster-to-time alignment question is a construct-validity risk, not a derivation loop.

full rationale

MoTE is evaluated on held-out target-domain labels using public benchmarks (EURLEX and Trustpilot review corpora) against external baselines (ChronosLex, Anti-CF, Self-Labeling). The adaptation signal is an unsupervised K-Means clustering of source embeddings plus shift vectors v_ij = z_i - c_j; no target label enters the adaptation objective, and no reported performance number is a refit of the target metric. The router warmup uses cluster labels derived from source data, but those labels are auxiliary routing targets, not the classification labels being predicted, so the macro-F1 gains are not forced by construction. The main vulnerability is interpretive—clusters are assumed to track time—and the paper explicitly acknowledges in Limitations that alternative domain-splitting strategies were not explored; that is a validity risk, not circular reasoning. Self-citations (Liu et al. 2024; Jones et al. 2024; Huang and Paul 2019) appear only as background and do not carry the derivation. No equation in the paper reduces to its own input, and no fitted parameter is renamed as a prediction.

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

The central method rests on user-chosen structural hyperparameters (T, K), a post hoc clustering choice, and the assumption that embedding-space clusters represent time. No new physical or conceptual entities are introduced.

free parameters (4)
  • T: number of temporal clusters and experts = 4 for review data; unspecified for EURLEX
    The method sets T equal to the number of source time domains; EURLEX uses one train split, so the actual T used there is not reported in the paper (Sections 3.1, 4.3).
  • K: top-K experts selected by router = 2
    Set to 2 for all experiments (Appendix E.2), chosen by hand and not tuned per language.
  • lambda: auxiliary load-balancing loss weight = 0.01
    Follows Shazeer et al. (2017) as stated in Section 3.2; not fitted here.
  • Clustering algorithm = K-Means
    Selected after comparing with Spectral Clustering because 'K-Means achieved the fastest parameter convergence and the best performance' (Section 3.1), a post hoc choice on the same datasets.
assumptions (4)
  • domain assumption K-Means with T clusters on XLM-R embeddings yields clusters that align with temporal periods.
    The warmup and routing logic assume cluster identity is a proxy for time; no validation of cluster-time alignment is provided (Section 3.1, Algorithm 1).
  • domain assumption Unlabeled target time-domain data is available as a batch at adaptation and test time.
    MoTE computes shift vectors from target representations and routes them before prediction, which presumes transductive access to the target distribution (Section 3, Figure 2).
  • domain assumption XLM-RoBERTa-base embeddings carry language-general semantic information sufficient for shift measurement across the 23 EURLEX languages.
    The method uses a single frozen encoder for all languages; the paper acknowledges other encoders may behave differently (Limitations, Section 8).
  • ad hoc to paper The EURLEX temporal split (train 1958-2010, test 2012-2016) can be used with an unspecified internal partitioning of the source into T time domains.
    EURLEX is described as a single source domain, but MoTE requires T>1 source domains; the paper does not state how the source is divided (Section 4.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Examining and Adapting Time for Multilingual Classification via Mixture of Temporal Experts." pith.science (2026). https://pith.science/paper/XS6QLZM2

@misc{pith2026250208825,
  author       = {Pith},
  title        = {Pith review of: Examining and Adapting Time for Multilingual Classification via Mixture of Temporal Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XS6QLZM2}},
  note         = {Machine review of arXiv:2502.08825}
}
read the original abstract

Time is implicitly embedded in classification process: classifiers are usually built on existing data while to be applied on future data whose distributions (e.g., label and token) may change. However, existing state-of-the-art classification models merely consider the temporal variations and primarily focus on English corpora, which leaves temporal studies less explored, let alone under multilingual settings. In this study, we fill the gap by treating time as domains (e.g., 2024 vs. 2025), examining temporal effects, and developing a domain adaptation framework to generalize classifiers over time on multiple languages. Our framework proposes Mixture of Temporal Experts (MoTE) to leverage both semantic and data distributional shifts to learn and adapt temporal trends into classification models. Our analysis shows classification performance varies over time across different languages, and we experimentally demonstrate that MoTE can enhance classifier generalizability over temporal data shifts. Our study provides analytic insights and addresses the need for time-aware models that perform robustly in multilingual scenarios.

Figures

Figures reproduced from arXiv: 2502.08825 by the authors.

Figure 1
Figure 1. Visualizations of temporal effects (performance variations by macro-F1 and AUC) of the cross-domain [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The MoTE method overview. Dt is in the source time domain data that has true labels, D is temporal ordered data excluding target domain, and Dtarget is the target time domain data without labels. Blue and grey lines indicate the training process, and pink line represent predicting data flow in the target time domain. centroids of the source data clusters as a data shift representation, and this representation is pro… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. FUTURE: Flexible Unlearning for Tree Ensemble

    cs.LG 2025-08 conditional novelty 6.0 of 10

    FUTURE forgets training samples from tree ensembles by optimizing sigmoid-smoothed split thresholds and copying them back to the original discrete trees.

Reference graph

Works this paper leans on

55 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Oshin Agarwal and Ani Nenkova. 2022. Temporal effects on pre-trained models for language processing tasks. Transactions of the Association for Computational Linguistics, 10:904--921

    work page 2022

  4. [4]

    Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. 2010. A theory of learning from different domains. Machine learning, 79:151--175

    2010

  5. [5]

    John Blitzer, Ryan McDonald, and Fernando Pereira. 2006. Domain adaptation with structural correspondence learning. In Proceedings of the 2006 conference on empirical methods in natural language processing, pages 120--128

    work page 2006

  6. [6]

    Ilias Chalkidis. 2023. https://arxiv.org/abs/2304.12202 Chatgpt may pass the bar exam soon, but has a long way to go for the lexglue benchmark . Preprint, arXiv:2304.12202

    work page Pith review arXiv 2023

  7. [7]

    Ilias Chalkidis, Manos Fergadiotis, and Ion Androutsopoulos. 2021. https://arxiv.org/abs/2109.00904 Multieurlex -- a multi-lingual and multi-label legal document classification dataset for zero-shot cross-lingual transfer . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    arXiv 2021

  8. [8]

    Ilias Chalkidis, Abhik Jana, Dirk Hartung, Michael Bommarito, Ion Androutsopoulos, Daniel Katz, and Nikolaos Aletras. 2022. https://doi.org/10.18653/v1/2022.acl-long.297 L ex GLUE : A benchmark dataset for legal language understanding in E nglish . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long P...

    doi:10.18653/v1/2022.acl-long.297 2022

Show all 55 references
  1. [9]

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

    2019 arXiv

  2. [10]

    Emilio Cueva, Adrian Lopez Monroy, Fernando S \'a nchez-Vega, and Thamar Solorio. 2024. https://doi.org/10.18653/v1/2024.naacl-long.460 Adaptive cross-lingual text classification through in-context one-shot demonstrations . In Proceedings of the 2024 Conference of the North Am...

    2024 doi

  3. [11]

    Hal Daume III and Daniel Marcu. 2006. Domain adaptation for statistical classifiers. Journal of artificial Intelligence research, 26:101--126

    2006

  4. [12]

    Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W

    Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. 2022. https://doi.org/10.1162/tacl_a_00459 Time-aware language models as temporal knowledge bases . Transactions of the Association for Computational Linguistics,...

    2022 doi

  5. [13]

    Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2018. Measuring and mitigating unintended bias in text classification. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 67--73

    2018

  6. [14]

    Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. 2021. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877--894

    2021

  7. [15]

    Yaroslav Ganin and Victor Lempitsky. 2015. https://proceedings.mlr.press/v37/ganin15.html Unsupervised domain adaptation by backpropagation . In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 1...

    2015

  8. [16]

    Huan He, Owen Queen, Teddy Koker, Consuelo Cuevas, Theodoros Tsiligkaridis, and Marinka Zitnik. 2023. Domain adaptation for time series under feature and label shifts. In International Conference on Machine Learning, pages 12746--12774. PMLR

    2023

  9. [17]

    Dirk Hovy, Anders Johannsen, and Anders S gaard. 2015. User review sites as a resource for large-scale sociolinguistic studies. In Proceedings of the 24th international conference on World Wide Web, pages 452--461

    2015

  10. [18]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  11. [19]

    Xiaolei Huang and Michael Paul. 2019. Neural temporality adaptation for document classification: Diachronic word embeddings and domain adaptation models. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4113--4123

    2019

  12. [20]

    Jing Jiang and ChengXiang Zhai. 2007. https://aclanthology.org/P07-1034 Instance weighting for domain adaptation in NLP . In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 264--271, Prague, Czech Republic. Association for Computat...

    2007

  13. [21]

    Yiqiao Jin, Minje Choi, Gaurav Verma, Jindong Wang, and Srijan Kumar. 2024. Mm-soc: Benchmarking multimodal large language models in social media platforms. In ACL

    2024

  14. [22]

    Yiqiao Jin, Xiting Wang, Yaru Hao, Yizhou Sun, and Xing Xie. 2023. Prototypical fine-tuning: Towards robust performance under varying data sizes. In Proceedings of the AAAI Conference on Artificial Intelligence

    2023

  15. [23]

    Precious Jones, Weisi Liu, I-Chan Huang, and Xiaolei Huang. 2024. https://arxiv.org/abs/2412.17803 Examining imbalance effects on performance and demographic fairness of clinical language models . Preprint, arXiv:2412.17803

    2024 arXiv

  16. [24]

    Adam King. 2024. https://arxiv.org/abs/2405.05478 Using machine translation to augment multilingual classification . Preprint, arXiv:2405.05478

    2024 arXiv

  17. [25]

    Fanshuang Kong, Richong Zhang, Ziqiao Wang, and Yongyi Mao. 2024. https://doi.org/10.1609/aaai.v38i16.29800 On unsupervised domain adaptation: Pseudo label guided mixup for adversarial prompt tuning . Proceedings of the AAAI Conference on Artificial Intelligence, 38(16):18399--18407

    2024 doi

  18. [26]

    Viet Dac Lai, Nghia Trung Ngo, Amir Pouran Ben Veyseh, Hieu Man, Franck Dernoncourt, Trung Bui, and Thien Huu Nguyen. 2023. https://arxiv.org/abs/2304.05613 Chatgpt beyond english: Towards a comprehensive evaluation of large language models in multilingual learning . Preprint,...

    2023 arXiv

  19. [27]

    Shichen Li, Zhongqing Wang, Xiaotong Jiang, and Guodong Zhou. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.22 Cross-domain sentiment classification using semantic representation . In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 289--299...

    2022 doi

  20. [28]

    Weisi Liu, Zhe He, and Xiaolei Huang. 2024. https://arxiv.org/abs/2407.17638 Time matters: Examine temporal effects on biomedical language models . Preprint, arXiv:2407.17638

    2024 arXiv

  21. [29]

    David Lopez-Paz and Marc Aurelio Ranzato. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/f87522788a2be2d171666752f97ddebb-Paper.pdf Gradient episodic memory for continual learning . In Advances in Neural Information Processing Systems, volume 30. Curran Assoc...

    2017

  22. [30]

    Xiuwei Lv, Zhiqiang Wang, and Lei Ju. 2023. Review generation combined with feature and instance-based domain adaptation for cross-domain aspect-based sentiment analysis. In Natural Language Processing and Chinese Computing, pages 813--825, Cham. Springer Nature Switzerland

    2023

  23. [31]

    Weicheng Ma, Kai Zhang, Renze Lou, Lili Wang, and Soroush Vosoughi. 2021. https://doi.org/10.18653/v1/2021.acl-long.152 Contributions of transformer attention heads in multi- and cross-lingual tasks . In Proceedings of the 59th Annual Meeting of the Association for Computation...

    2021 doi

  24. [32]

    OpenAI et al. 2024. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . Preprint, arXiv:2303.08774

    2024 arXiv

  25. [33]

    Felix Ott, David R \"u gamer, Lucas Heublein, Bernd Bischl, and Christopher Mutschler. 2022. Domain adaptation for time-series classification to mitigate covariate shift. In Proceedings of the 30th ACM international conference on multimedia, pages 5934--5943

    2022

  26. [34]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu ...

    2019

  27. [35]

    Hossein Rajaby Faghihi and Parisa Kordjamshidi. 2021. Time-stamped language model: Teaching language models to understand the flow of events. In The 2021 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-2021)

    2021

  28. [36]

    Olesya Razuvayevskaya, Ben Wu, Jo \ a o A Leite, Freddy Heppell, Ivan Srba, Carolina Scarton, Kalina Bontcheva, and Xingyi Song. 2024. Comparison between parameter-efficient techniques and full fine-tuning: A case study on multilingual news article classification. Plos one, 19...

    2024

  29. [37]

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. 2019. https://proceedings.neurips.cc/paper_files/paper/2019/file/fa7cdfad1a5aaf8370ebeda47a1ff1c3-Paper.pdf Experience replay for continual learning . In Advances in Neural Information Processin...

    2019

  30. [38]

    Paul R \"o ttger and Janet Pierrehumbert. 2021. Temporal adaptation of bert and performance on downstream document classification: Insights from social media. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2400--2412

    2021

  31. [39]

    TYS Santosh, Tuan-Quang Vuong, and Matthias Grabmair. 2024. https://arxiv.org/abs/2405.14211 Chronoslex: Time-aware incremental training for temporal generalization of legal classification tasks . Preprint, arXiv:2405.14211

    2024 arXiv

  32. [40]

    Chao Shang, Guangtao Wang, Peng Qi, and Jing Huang. 2022. Improving time sensitivity for question answering over temporal knowledge graphs. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8017--8026

    2022

  33. [41]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538

    2017 arXiv

  34. [42]

    Oleh Shliazhko, Alena Fenogenova, Maria Tikhonova, Vladislav Mikhailov, Anastasia Kozlova, and Tatiana Shavrina. 2023. https://arxiv.org/abs/2204.07580 mgpt: Few-shot learners go multilingual . Preprint, arXiv:2204.07580

    2023 arXiv

  35. [43]

    Ian Stewart, Dustin Arendt, Eric Bell, and Svitlana Volkova. 2017. https://doi.org/10.1609/icwsm.v11i1.14938 Measuring, predicting and visualizing short-term change in word representation and usage in vkontakte social network . Proceedings of the International AAAI Conference ...

    2017 doi

  36. [44]

    Yi Su, Yixin Ji, Juntao Li, Hai Ye, and Min Zhang. 2023. Beware of model collapse! fast and stable test-time adaptation for robust question answering. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12998--13011

    2023

  37. [45]

    Xiaofei Sun, Xiaoya Li, Jiwei Li, Fei Wu, Shangwei Guo, Tianwei Zhang, and Guoyin Wang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.603 Text classification via large language models . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 899...

    2023 doi

  38. [46]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in Ne...

    2017

  39. [47]

    Kwok, and Lionel M

    Yaqing Wang, Quanming Yao, James T. Kwok, and Lionel M. Ni. 2020. https://doi.org/10.1145/3386252 Generalizing from a few examples: A survey on few-shot learning . ACM Comput. Surv., 53(3)

    2020 doi

  40. [48]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020

  41. [49]

    Siheng Xiong, Ali Payani, Ramana Kompella, and Faramarz Fekri. 2024 a . https://doi.org/10.18653/v1/2024.acl-long.563 Large language models can learn temporal reasoning . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long...

    2024 doi

  42. [50]

    Siheng Xiong, Yuan Yang, Ali Payani, James C Kerce, and Faramarz Fekri. 2024 b . https://doi.org/10.1609/aaai.v38i14.29544 Teilp: Time prediction over knowledge graphs via logical reasoning . Proceedings of the AAAI Conference on Artificial Intelligence, 38(14):16112--16119

    2024 doi

  43. [51]

    Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. https://doi.org/10.18653/v1/2021.naacl-main.41 m T 5: A massively multilingual pre-trained text-to-text transformer . In Proceedings of the 2021 Conferenc...

    2021 doi

  44. [52]

    Kerce, and Faramarz Fekri

    Yuan Yang, Siheng Xiong, Ali Payani, James C. Kerce, and Faramarz Fekri. 2024. https://doi.org/10.24963/ijcai.2024/400 Temporal inductive logic reasoning over hypergraphs . In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24 ,...

    2024 doi

  45. [53]

    Wenpeng Yin, Jamaal Hay, and Dan Roth. 2019. https://doi.org/10.18653/v1/D19-1404 Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th In...

    2019 doi

  46. [54]

    Huimin Zeng, Zhenrui Yue, Lanyu Shang, Yang Zhang, and Dong Wang. 2024. https://doi.org/10.1109/TETC.2024.3354419 Unsupervised domain adaptation via contrastive adversarial domain mixup: A case study on covid-19 . IEEE Transactions on Emerging Topics in Computing, pages 1--12

    2024

  47. [55]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong W...

    2023 arXiv

Pith tools

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