Pith. sign in

REVIEW 3 major objections 5 minor 32 references

QU-NLP at CheckThat! 2025: Multilingual Subjectivity in News Articles Detection using Feature-Augmented Transformer Models with Sequential Cross-Lingual Fine-Tuning

T0 review · 3 major / 5 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 feature-augmented transformer with a learned gate over TF-IDF features and sequential cross-lingual fine-tuning wins subjectivity detection for English and Romanian.

desk verdict A competent shared-task system paper whose main analytical claim about language order is undermined by a recency confound. read the letter →

arxiv 2507.21095 v1 pith:VQNJ2XS2 submitted 2025-07-01 cs.CL

Mohammad AL-Smadi This is my paper
classification cs.CL
keywords subjectivitydetectioncross-lingualtransferfeature-augmentedtransformerTF-IDFgatingsequentialfine-tuningmultilingualNLPnewsarticles
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

This paper claims that subjectivity detection in news sentences is improved by a feature-augmented transformer that combines contextual embeddings with TF-IDF lexical features through a learned gate. The central evidence is a set of shared-task results: macro-F1 of 0.8052 on English, 0.8126 on Romanian in the zero-shot setting, and 0.8013 on German, with weaker results on Italian and on Polish, Ukrainian, and Greek. The paper also argues that the order of sequential cross-lingual fine-tuning and the gating mechanism both materially affect performance. A sympathetic reader would care because the recipe is concrete, competitive for multilingual subjectivity detection, and points to language order as a design choice in cross-lingual pipelines.

What carries the argument

The load-bearing mechanism is a gating fusion of two feature streams. For non-Arabic languages, term frequency-inverse document frequency (TF-IDF) features over character n-grams (3-7) are projected to 128 dimensions, and a scalar gate computed from the DeBERTa V3 encoder's CLS embedding (the model's sentence-level summary vector) multiplies the projected lexical vector before concatenation with the contextual embedding and classification. The gate is what lets the model downweight lexical cues when the contextual signal is strong and upweight them when sparse lexical cues carry the subjectivity signal. The second mechanism is sequential cross-lingual fine-tuning: the model is fine-tuned on one language, then the next, then the next, and the paper evaluates the claim that this order transfers subjectivity cues across languages. For Arabic, the architecture instead concatenates part-of-speech and TF-IDF projections with the AraELECTRA CLS vector.

What would settle it

Re-run the exact same training procedure for each configuration in Tables 5 and 6 with several random seeds and compute the spread of macro-F1 scores on the official test set; if the spread across seeds is comparable to or larger than the reported gaps, such as Italian 0.7139 versus 0.8033 across language orders, then the claimed dependence on the gating mechanism and on language order is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a DeBERTa V3 large encoder (a pretrained transformer language model) augmented with a TF-IDF branch modulated by a scalar gate, and fine-tuned sequentially on German, then Italian, then English, is a competitive multilingual subjectivity detector. The gating scalar is computed from the CLS representation and decides per instance how much the lexical signal should contribute; the ablation shows that removing the gate or the TF-IDF branch lowers macro-F1 on English and German, while the cross-lingual sequence improves English and Italian scores relative to no sequence training. The architecture is language-dependent: Arabic uses AraELECTRA, an Arabic ELECTRA variant, with part-of-speech tags and TF-IDF projections concatenated directly. The strongest zero-shot transfer appears for Romanian, which the paper attributes to linguistic proximity to the training languages, while Polish, Ukrainian, and Greek fall at or below the provided baseline.

Load-bearing premise

The load-bearing premise is that the single-run macro-F1 scores on the official test set are stable enough that the differences between configurations, especially the language-order differences, reflect real effects rather than training noise.

Editorial extensions

If this is right

  • Teams building multilingual subjectivity systems should treat the gated TF-IDF branch and the order of language fine-tuning as design choices worth tuning, not incidental details.
  • Zero-shot transfer is strongest to languages close to the training sequence, such as Romanian, and can be below the baseline for more distant languages such as Polish, Ukrainian, and Greek, so cross-lingual gains should be expected to be language-pair dependent.
  • The multilingual model trained on the same language sequence underperforms the monolingual English and German models, suggesting that a single shared model trades away per-language accuracy.
  • The Arabic model is architecturally separate and weaker, so the recipe for non-Arabic languages does not automatically extend to Arabic; the feature-augmentation approach needs a language-specific design there.

Reading between the lines

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

  • An inference beyond the paper: the language-order results are based on a single training run per order, so the headline claim about order sensitivity should be read as a hypothesis until repeated runs with different seeds confirm it.
  • If order sensitivity is real, a natural testable extension is to select the fine-tuning order per target language, since the best order differs for German, Italian, and English in the reported table.
  • The same gated-fusion recipe could be tried on other classification tasks with sparse lexical cues, such as stance detection or media bias detection, where a per-instance gate might also balance lexical and contextual signals.
  • The below-baseline zero-shot scores for Polish, Ukrainian, and Greek suggest that the linguistic-proximity explanation would be stronger if tested across more Romance and Slavic target languages with confidence intervals.
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. This paper describes the QU-NLP submission to the CheckThat! 2025 Task 1 on sentence-level subjectivity detection in news. The authors propose two feature-augmented transformer architectures: an AraELECTRA model with POS and TF-IDF features for Arabic, and a DeBERTa-V3-large model with a learned gating mechanism for TF-IDF features for the other languages. Non-Arabic models are trained with sequential cross-lingual fine-tuning in a fixed order (German -> Italian -> English) and evaluated in monolingual, multilingual, and zero-shot settings. The paper reports first-place macro-F1 scores on English (0.8052) and Romanian (0.8126), third place on German (0.8013), and fourth place on Arabic (0.5771), with weaker results on Polish, Ukrainian, and Greek. An ablation study (Table 5) and a language-order study (Table 6) are presented to support claims about the importance of gating and of cross-lingual fine-tuning order.

Significance. If the descriptive claims are taken at face value, the paper demonstrates a competitive and reproducible-in-spirit recipe for a multilingual shared task: feature-augmented transformers with gated TF-IDF and sequential fine-tuning achieve state-of-the-art results on English and Romanian and competitive results on German, while the authors honestly report below-baseline scores for Polish, Ukrainian, and Greek and explicitly note that Arabic was absent from the multilingual training sequence. The paper also provides a useful, if limited, ablation decomposition. However, the analytical claims about language order and linguistic proximity, and to a lesser extent about the gating mechanism, go beyond what the reported single-run experiments can support; these are the load-bearing points that need revision.

major comments (3)
  1. [§5.3, Table 6] The claim that language order and linguistic proximity drive transfer is confounded by a recency effect. In Table 6, for every target language the highest macro-F1 is obtained when that language is the last element of the fine-tuning sequence: German scores 0.8195 under (English -> Italian -> German), Italian scores 0.8033 under (German -> English -> Italian), and English scores 0.8052 under (German -> Italian -> English). This pattern is fully predicted by the target language being the most recently fine-tuned language, i.e., by minimal catastrophic forgetting, rather than by any pairwise typological similarity. The prose in §5.3 attributes the Italian drop under (German -> Italian -> English) to "language interference," but in that order Italian is also the middle language and would suffer a recency disadvantage. The experimental design needs control sequences that hold the target language's position fixed while varying the preceding languages; without such controls, or at least an explicit acknowledgment of the confound, the linguistic-proximity conclusion in the abstract and §5.3 is not supported.
  2. [§5.2, Table 5] The ablation results partially contradict the paper's summary claims. The "Full Model (without cross-lingual training)" row reports a German macro-F1 of 0.8013, exactly equal to the full model with cross-lingual sequence training, so there is no evidence in this table that cross-lingual training benefits German. Moreover, for Italian, "DeBERTa + TF-IDF (No Gating)" achieves 0.7234, which is higher than the full model with gating (0.7139), directly contradicting the abstract's statement that "combining TF-IDF features with the gating mechanism" is important for performance. The text in §5.2 acknowledges that gating improved only English and German, but the abstract and conclusion overgeneralize. These inconsistencies must be addressed either by revising the claims or by providing additional evidence that the observed pattern is not noise.
  3. [§3.5.2, Tables 5 and 6] All ablation and language-order results come from a single training run per configuration, with checkpoints selected on development loss, and no standard deviations, seed averages, or significance tests are reported. Given test sets of roughly 300 sentences per language (Table 1), differences such as the Italian scores of 0.7139 versus 0.8033 across language orders, or the gating effect in Table 5, could be within run-to-run variance. The paper should either report variance estimates (e.g., multiple seeds with mean and standard deviation) or explicitly weaken the causal and comparative claims to "preliminary observations." As written, the statements that language order "significantly" affects performance and that each component "contributes" are stronger than the experimental design supports.
minor comments (5)
  1. [§3.1] The bullet list for the monolingual setting states "five languages: English, Arabic, Italian, and German," which lists only four languages; the number should be corrected to four.
  2. [§5.2] The sentence "Trainings of the monolingual languages forllowed the same sequence" contains a typo ("forllowed" should be "followed"), and the term "consequently" is used where "respectively" is meant in several places in the discussion of Table 5.
  3. [§3.3.2] The text says the model uses "deberta-v3-large" but the tokenizer is described as "DeBERTaV2Tokenizer"; please confirm that this is intentional, since DeBERTa V3 and V2 tokenizers may differ in vocabulary.
  4. [§4.1, Table 2] There is a typographical issue in the first sentence of §4.1: "our system achieved the best performance for English" is fine, but earlier in the abstract the English score is reported as "0.8052" while in the same paragraph it appears as "0.8052 (1st)"; no inconsistency is present, but the repeated formatting of scores should be unified.
  5. [References] Several references lack DOIs or URLs (e.g., [8], [9], [26]); adding persistent identifiers would improve reproducibility for readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: an empirical shared-task paper whose claimed results are evaluated on held-out official test data; ablations and order comparisons are empirical, not derived from fitted parameters or self-citations.

full rationale

This is a shared-task system paper, not a derivation chain. The headline numbers (English macro-F1 0.8052, Romanian 0.8126, etc.) are scores on the official held-out CheckThat! 2025 test set, and no predicted value is defined in terms of a fitted parameter or constructed from the test labels. The gating mechanism and TF-IDF branch are trained on labeled training data and evaluated by the task organizers, so the main results are externally grounded. The ablation study in Table 5 compares explicit model configurations (DeBERTa-only, DeBERTa+TF-IDF without gating, full model, full model without cross-lingual training) and is an empirical comparison, not a renaming of inputs. The language-order analysis in Table 6 is likewise an experimental covariate; ordering is varied and performance is measured, so the claim of sensitivity to order is not circular. The only self-citation, reference [18] on Arabic aspect-based sentiment analysis, appears in related work and is not load-bearing for any central claim. There is no imported uniqueness theorem, no ansatz smuggled in via citation, and no fitted-input-called-prediction step. Two non-circular evidentiary weaknesses are worth noting: in Table 5, the German score with and without cross-lingual training is identical (0.8013), so the prose claim that cross-lingual sequence training positively enhanced monolingual results is not supported for German; and in Table 6, each language scores best exactly when it is the last language in the fine-tuning sequence, which is a recency confound for the Section 5.3 linguistic-proximity interpretation. These are correctness and statistical-interpretation concerns, not circularity, and do not change the verdict.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims are empirical and rest on the assumption that the official annotations and macro-F1 scores are correct, that pretrained transformer representations transfer to this task and to unseen languages, and on hand-set hyperparameters such as TF-IDF settings, learning rate, early stopping, and the language order German then Italian then English. No new theoretical entities are introduced; the gating mechanism is a learned scalar weight, not an invented physical or conceptual entity.

free parameters (3)
  • Language order for sequential fine-tuning = German then Italian then English
    Chosen by hand as the primary training sequence. Section 5.3 shows macro-F1 changes by up to about 0.09 across orders, so this choice partly determines the headline results.
  • TF-IDF vectorizer settings = max_features=3000, min_df=2, character n-grams (3,7)
    Hand-set in Section 3.5.1 for Arabic and used for the lexical branch in other languages; the importance of lexical features is a core claim, so these choices matter.
  • Training hyperparameters = learning rate 1e-5, batch size 8 or 16, warmup 100, early stopping patience 2 or 3
    Standard values chosen without a reported search. Final scores and ablations depend on them, and no seed averaging is reported.
assumptions (4)
  • domain assumption Gold labels in the CheckThat 2025 Task 1 dataset correctly implement the language-agnostic subjectivity definition of Ruggeri et al.
    Every reported F1 score is computed against these labels. If annotations are noisy or inconsistent across languages, all rankings and transfer conclusions inherit that noise. Invoked in Section 3.2.
  • domain assumption AraELECTRA and DeBERTa-v3-large pretrained representations transfer to subjectivity detection in the covered languages.
    The method replaces heavy feature engineering with fine-tuning these backbones. If their cross-lingual alignments do not hold for subjectivity markers, zero-shot results would collapse to random or baseline behavior. Invoked in Sections 3.3.1 and 3.3.2.
  • domain assumption Macro-F1 is the agreed evaluation measure and the official leaderboard scores are computed correctly.
    All claims of rank and effectiveness depend on the shared task score computation. Invoked in Section 3.1 and Section 4.
  • domain assumption Sequential fine-tuning on German, Italian, and English creates a usable intermediate representation for Romanian and other unseen languages.
    Zero-shot success is interpreted as transfer from this particular sequence. The paper explicitly attributes Romanian success to linguistic similarity with the training languages. Invoked in Section 3.5.2 and Section 5.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QU-NLP at CheckThat! 2025: Multilingual Subjectivity in News Articles Detection using Feature-Augmented Transformer Models with Sequential Cross-Lingual Fine-Tuning." pith.science (2026). https://pith.science/paper/VQNJ2XS2

@misc{pith2026250721095,
  author       = {Pith},
  title        = {Pith review of: QU-NLP at CheckThat! 2025: Multilingual Subjectivity in News Articles Detection using Feature-Augmented Transformer Models with Sequential Cross-Lingual Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VQNJ2XS2}},
  note         = {Machine review of arXiv:2507.21095}
}
read the original abstract

This paper presents our approach to the CheckThat! 2025 Task 1 on subjectivity detection, where systems are challenged to distinguish whether a sentence from a news article expresses the subjective view of the author or presents an objective view on the covered topic. We propose a feature-augmented transformer architecture that combines contextual embeddings from pre-trained language models with statistical and linguistic features. Our system leveraged pre-trained transformers with additional lexical features: for Arabic we used AraELECTRA augmented with part-of-speech (POS) tags and TF-IDF features, while for the other languages we fine-tuned a cross-lingual DeBERTa~V3 model combined with TF-IDF features through a gating mechanism. We evaluated our system in monolingual, multilingual, and zero-shot settings across multiple languages including English, Arabic, German, Italian, and several unseen languages. The results demonstrate the effectiveness of our approach, achieving competitive performance across different languages with notable success in the monolingual setting for English (rank 1st with macro-F1=0.8052), German (rank 3rd with macro-F1=0.8013), Arabic (rank 4th with macro-F1=0.5771), and Romanian (rank 1st with macro-F1=0.8126) in the zero-shot setting. We also conducted an ablation analysis that demonstrated the importance of combining TF-IDF features with the gating mechanism and the cross-lingual transfer for subjectivity detection. Furthermore, our analysis reveals the model's sensitivity to both the order of cross-lingual fine-tuning and the linguistic proximity of the training languages.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 22 canonical work pages

  1. [1]

    Ruggeri, F

    F. Ruggeri, F. Antici, A. Galassi, A. Korre, A. Muti, A. Barron, On the definition of prescriptive annotation guidelines for language-agnostic subjectivity detection, Proceedings of Text2Story — Sixth Workshop on Narrative Extraction From Texts 3370 (2023) 103–111

    work page 2023

  2. [2]

    F. Alam, J. M. Struß, T. Chakraborty, S. Dietze, S. Hafid, K. Korre, A. Muti, P. Nakov, F. Ruggeri, S. Schellhammer, V. Setty, M. Sundriyal, K. Todorov, V. V., The clef-2025 checkthat! lab: Subjectivity, fact-checking, claim normalization, and retrieval, in: C. Hauff, C. Macdonald, D. Jannach, G. Kazai, F. M. Nardini, F. Pinelli, F. Silvestri, N. Tonellot...

    2025

  3. [3]

    Wiebe, T

    J. Wiebe, T. Wilson, R. Bruce, M. Bell, M. Martin, Learning subjective language, Computational Linguistics 30 (2004) 277–308

    work page 2004

  4. [4]

    Artetxe, H

    M. Artetxe, H. Schwenk, Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond, Transactions of the Association for Computational Linguistics 7 (2019) 597–610

    work page 2019

  5. [5]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies 1 (2019) 4171–4186

    work page 2019

  6. [6]

    Conneau, K

    A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, V. Stoyanov, Unsupervised cross-lingual representation learning at scale, in: D. Jurafsky, J. Chai, N. Schluter, J. Tetreault (Eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Association for Computation...

    doi:10.18653/v1/2020 2020

  7. [7]

    L. Xue, N. Constant, A. Roberts, M. Kale, R. Al-Rfou, A. Siddhant, A. Barua, C. Raffel, mt5: A massively multilingual pre-trained text-to-text transformer, Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (2021) 483–498

    work page 2021

  8. [8]

    S. Wu, M. Dredze, Beto, bentz, becas: The surprising cross-lingual effectiveness of bert, Proceed- ings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (2019) 833–844

    work page 2019

Show all 32 references
  1. [9]

    Pires, E

    T. Pires, E. Schlinger, D. Garrette, How multilingual is multilingual BERT?, in: A. Korhonen, D. Traum, L. Màrquez (Eds.), Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, Florence, Italy, 2019,...

    2019 doi

  2. [10]

    M. Yan, H. Zhang, D. Jin, J. T. Zhou, Multi-source meta transfer for low resource multiple- choice question answering, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 7331–7341

    2020

  3. [11]

    Pfeiffer, I

    J. Pfeiffer, I. Vulić, I. Gurevych, S. Ruder, Mad-x: An adapter-based framework for multi-task cross-lingual transfer, Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (2020) 7654–7673

    2020

  4. [12]

    Q. Do, J. Gaspers, Cross-lingual transfer learning with data selection for large-scale spoken language understanding, in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Proce...

    2019

  5. [13]

    Nooralahzadeh, G

    F. Nooralahzadeh, G. Bekoulis, J. Bjerva, I. Augenstein, Zero-shot cross-lingual transfer with meta learning, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020, pp. 4547–4562

    2020

  6. [14]

    Wiebe, R

    J. Wiebe, R. Bruce, T. P. O’Hara, Development and use of a gold-standard data set for subjectivity classifications, in: Proceedings of the 37th annual meeting of the Association for Computational Linguistics, 1999, pp. 246–253

    1999

  7. [15]

    Recasens, C

    M. Recasens, C. Danescu-Niculescu-Mizil, D. Jurafsky, Linguistic models for analyzing and detect- ing biased language, Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics 1 (2013) 1650–1659

    2013

  8. [16]

    A. M. El-Halees, Arabic text classification using maximum entropy, in: The International Arab Conference on Information Technology (ACIT), 2011

    2011

  9. [17]

    Abdul-Mageed, M

    M. Abdul-Mageed, M. Diab, Subjectivity and sentiment analysis of modern standard arabic, in: Proceedings of the ACL Workshop on Computational Approaches to Subjectivity and Sentiment Analysis, ACL, 2011, pp. 35–44. URL: https://aclanthology.org/W11-1703/

    2011

  10. [18]

    Al-Smadi, M

    M. Al-Smadi, M. Al-Ayyoub, H. Al-Sarhan, Y. Jararweh, An aspect-based sentiment analysis approach to evaluating arabic news affect on readers, Journal of Universal Computer Science 22 (2016) 630–649

    2016

  11. [19]

    Spinde, L

    T. Spinde, L. Rudnitckaia, J. Mitrović, F. Hamborg, M. Granitzer, B. Gipp, K. Donnay, Automated identification of bias inducing words in news articles using linguistic and context-oriented features, Information Processing & Management 58 (2021) 102505

    2021

  12. [20]

    Antici, F

    F. Antici, F. Ruggeri, A. Galassi, A. Korre, A. Muti, A. Bardi, A. Fedotova, A. Barrón-Cedeño, et al., A corpus for sentence-level subjectivity detection on english news articles, in: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language...

    2024

  13. [21]

    Suwaileh, M

    R. Suwaileh, M. Hasanain, F. Hubail, W. Zaghouani, F. Alam, Thatiar: Subjectivity detection in arabic news sentences, arXiv preprint arXiv:2406.05559 (2024)

    2024 arXiv

  14. [22]

    Antoun, F

    W. Antoun, F. Baly, H. Hajj, Araelectra: Pre-training text discriminators for arabic language understanding, arXiv preprint arXiv:2012.15516 (2020)

    2020 arXiv

  15. [23]

    Clark, Electra: Pre-training text encoders as discriminators rather than generators, arXiv preprint arXiv:2003.10555 (2020)

    K. Clark, Electra: Pre-training text encoders as discriminators rather than generators, arXiv preprint arXiv:2003.10555 (2020)

    2020 arXiv

  16. [24]

    J. D. M.-W. C. Kenton, L. K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of naacL-HLT, volume 1, Minneapolis, Minnesota, 2019, p. 2

    2019

  17. [25]

    Inoue, B

    G. Inoue, B. Alhafni, N. Baimukan, H. Bouamor, N. Habash, The interplay of variant, size, and task type in Arabic pre-trained language models, in: Proceedings of the Sixth Arabic Natural Language Processing Workshop, Association for Computational Linguistics, Kyiv, Ukraine (On...

    2021

  18. [26]

    Glorot, A

    X. Glorot, A. Bordes, Y. Bengio, Deep sparse rectifier neural networks, in: Proceedings of the fourteenth international conference on artificial intelligence and statistics, JMLR Workshop and Conference Proceedings, 2011, pp. 315–323

    2011

  19. [27]

    P. He, J. Gao, W. Chen, Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing, 2023. URL: https://arxiv.org/abs/2111.09543. arXiv:2111.09543

    2023 arXiv

  20. [28]

    R. K. Srivastava, K. Greff, J. Schmidhuber, Highway networks, 2015. URL: https://arxiv.org/abs/ 1505.00387. arXiv:1505.00387

    2015 arXiv

  21. [29]

    E. P. Stabler, E. L. Keenan, Structural similarity within and among languages, Theoretical Computer Science 293 (2003) 345–363

    2003

  22. [30]

    Z. Li, D. Hoiem, Learning without forgetting, IEEE transactions on pattern analysis and machine intelligence 40 (2017) 2935–2947

    2017

  23. [31]

    Kirkpatrick, R

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al., Overcoming catastrophic forgetting in neural networks, Proceedings of the national academy of sciences 114 (2017) 3521–3526

    2017

  24. [32]

    Muennighoff, T

    N. Muennighoff, T. Wang, L. Sutawika, A. Roberts, S. Biderman, T. Le Scao, M. S. Bari, S. Shen, Z. X. Yong, H. Schoelkopf, et al., Crosslingual generalization through multitask finetuning, in: The 61st Annual Meeting Of The Association For Computational Linguistics, 2023

    2023

Pith tools

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