Pith. sign in

REVIEW 4 major objections 5 minor 15 references

Small Models, Big Impact: Efficient Corpus and Graph-Based Adaptation of Small Multilingual Language Models for Low-Resource Languages

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 Lightweight adapter training on small multilingual models outperforms both full fine-tuning and much larger prompted models on low-resource language tasks, with as little as 1 GB of text or a few MB of knowledge-graph data.

desk verdict A solid empirical comparison of adapters for low-resource languages, but the abstract overstates the results; the body is honest enough to deserve serious review. read the letter →

arxiv 2502.10140 v1 pith:CCBKDXAO submitted 2025-02-14 cs.CL

Daniil Gurgurov , Ivan Vykopal , Josef van Genabith , Simon Ostermann This is my paper · ORCID
classification cs.CL
keywords low-resourcelanguagesparameter-efficientadaptationadaptersmultilinguallanguagemodelsknowledgegraphsmaskedmodelingXLM-RmBERT
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 argues that the most effective way to build NLP systems for low-resource languages is not to scale up models but to adapt small multilingual models with lightweight parameter-efficient adapters. The authors try to establish that adapter-based adaptation delivers the central practical results: a few hundred MB to 1 GB of free text or even a few MB of knowledge-graph data yields reliable gains on masked language modeling, topic classification, sentiment analysis, and named entity recognition across 30 low-resource languages. They further claim that these adapted small models match or beat full fine-tuning while training far fewer parameters, and that they outperform massive models such as LLaMA-3, GPT-4, and DeepSeek-R1 distillates, which are typically only prompted rather than adapted. If this is right, the practical implication is that effective low-resource NLP is within reach of modest data and compute budgets, and that pre-training coverage, not model size, is the main driver of performance.

What carries the argument

The central mechanism is the language adapter: a small bottleneck module inserted into each layer of a frozen transformer, trained with a masked-language-modeling objective on adaptation text and then stacked with a task adapter for the downstream task. Three variants carry the comparison: Sequential Bottleneck (a two-layer down-and-up projection), Sequential Bottleneck with Invertible Layers (adding an invertible layer that reorders and mixes features, improving input-output embedding alignment), and Low-Rank Adaptation (low-rank updates to weight matrices). The structured knowledge enters through the same adapter mechanism: ConceptNet triples are verbalized into sentences with fixed predicates and used as adaptation text. Adapter fusion, which combines text-trained and graph-trained adapters, is used for named entity recognition. This machinery isolates what adaptation data and architecture contribute while leaving the base model frozen, which is what makes the parameter and data efficiency claims measurable.

What would settle it

Run the same task-adapter training recipe (the paper's Seq_bn_inv setup with identical target-language training splits) on a large model such as LLaMA-3-8B across all 30 languages and all three downstream tasks, and compare per-language F1 with the adapted XLM-R; if the large model matches or exceeds the small one when both receive the same supervised adaptation data, the central small-beats-large claim would be contradicted.

Watch

Extended reading notes

Core claim

The paper reports a systematic comparison of three adapter architectures — Sequential Bottleneck, Sequential Bottleneck with Invertible Layers, and Low-Rank Adaptation — inserted into frozen mBERT and XLM-R models and pre-trained on two data types: unstructured text from GlotCC and structured knowledge from ConceptNet converted into natural-language sentences. The central discovery is that Sequential Bottleneck adapters give the largest improvements in masked-language-modeling pseudo-perplexity, while Invertible Bottleneck adapters give the best downstream task scores due to better embedding alignment and more trainable parameters; both match or approach full fine-tuning at a fraction of the parameters. On the comparison with large models, Glot-adapted XLM-R exceeds the zero- and few-shot performance of much larger models on topic classification and named entity recognition for overlapping low-resource languages, and adapter-tuned LLaMA-3-8B still trails the small adapted models. The paper also establishes that pre-training data size dominates performance, with adaptation data yielding diminishing returns for languages already well covered during pre-training, and that pseudo-perplexity is a moderately reliable proxy of downstream task quality after adaptation.

Load-bearing premise

The claim that small models beat massive ones rests on comparing task-adapted small models with large models that were only prompted, which may not give the large models their best chance.

Editorial extensions

If this is right

  • Up to 1 GB of free text or a few MB of knowledge-graph data is enough to produce measurable gains on all four tasks, so low-resource NLP pipelines no longer need large monolingual corpora.
  • Adapter-based adaptation matches or exceeds full fine-tuning with orders of magnitude fewer trainable parameters, making continued pre-training of the full model hard to justify for low-resource languages.
  • For languages with more than roughly 20 GB of pre-training data, adaptation gives diminishing returns, so adaptation effort should concentrate on under-represented languages.
  • Pseudo-perplexity after adaptation correlates moderately with downstream performance, so it can serve as a cheap early filter when deciding whether an adapter is worth deploying.
  • Fusing text-based and knowledge-graph-based adapters lifts named entity recognition for XLM-R, suggesting complementary information in the two data sources.

Reading between the lines

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

  • The strongest comparison in the paper — adapter-tuned LLaMA-3 versus adapter-tuned XLM-R — still favors the small model, but it covers only five languages and three tasks; a fair test would adapt a large model on the full 30-language setup to see where the small-model advantage saturates.
  • Because ConceptNet adapters help most on named entity recognition, injecting structured knowledge for other relation-heavy tasks, such as relation extraction or question answering, is a natural next test.
  • The diminishing-returns threshold tied to pre-training coverage suggests a practical rule for resource allocation: adapt where pre-training data is scarce, and spend compute on vocabulary extension or larger adaptation corpora where it is not.
  • The paper leaves open whether the small-model advantage comes from capacity fit to scarce data or from the fact that large models were only prompted; a matched training-data experiment would separate these explanations.
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 parameter-efficient adapter-based adaptation of small multilingual language models (mBERT, XLM-R-base) to 30 low-resource languages, using unstructured text from GlotCC and structured knowledge from ConceptNet. It evaluates three adapter architectures (Sequential Bottleneck, Invertible Bottleneck, LoRA) on masked language modeling pseudo-perplexity and on topic classification, sentiment analysis, and named entity recognition, and compares the adapted small models against full fine-tuning and against several large language models under prompting or adapter-based adaptation. The central claims are that adapter-based methods match or outperform full fine-tuning with far fewer trainable parameters, and that smaller mLMs are more effective than massive LLMs for these languages.

Significance. If the headline claims hold, the paper offers practical guidance for low-resource NLP: adaptation with a small fraction of trainable parameters can be sufficient, and model scale is not the main lever for these languages. The work is valuable for its breadth: 30 languages, two base models, three adapter types, two adaptation data sources, and four evaluation tasks. The authors also release code, which supports reproducibility, and the limitations section is candid about hyperparameter sensitivity, data-size caps, and omitted experiments. The main significance is therefore conditional on correcting the overstatement in the abstract and on providing uncertainty estimates for the small observed differences.

major comments (4)
  1. [Abstract; §5.2.1; Table 1] The abstract claim that "Adapter-based methods match or outperform full fine-tuning" is not supported by the paper's own results. In Table 1, full fine-tuning beats the best adapter on topic classification for both seen and unseen languages (mBERT: 81.73 vs. 79.35 seen, 43.65 vs. 42.40 unseen; XLM-R: 85.61 vs. 84.06 seen, 57.30 vs. 51.43 unseen) and on sentiment analysis for seen languages for both models (mBERT: 84.07 vs. 83.64; XLM-R: 88.56 vs. 88.20) and for unseen XLM-R languages (68.19 vs. 65.88). Section 5.2.1 explicitly concedes that "full fine-tuning provided better average results" for topic classification. The body is more careful than the abstract, but the abstract's wording overstates the evidence and should be revised to describe adapters as competitive while using far fewer parameters, with the specific task-by-task gaps reported.
  2. [§6; Table 2; Table 3] The "small mLMs beat massive LLMs" conclusion rests on an uneven comparison. The small models receive target-language supervised task adapters (e.g., 701 training examples for topic classification), while GPT-4, GPT-3.5, and DeepSeek models are only evaluated with zero-shot or few-shot prompting, in part using numbers taken from prior work rather than from a common evaluation protocol. Table 3 provides fairer support by showing that adapter-tuned LLaMA-3-8B still trails XLM-R on topic classification for five languages, but the abstract's general claim "smaller mLMs prove more effective ... than massive LLMs" should be qualified to specify that the comparison is primarily against prompting, not against task-adapted LLMs.
  3. [Tables 15, 16, 18, 19, 21, 22; §4.3] No error bars or significance tests are reported even though downstream results are averaged over three random seeds. Several head-to-head differences that support the narrative are small (e.g., XLM-R sentiment analysis seen languages: 88.20 for Seq_bn and Seq_bn_inv vs. 88.56 for full fine-tuning; mBERT sentiment analysis seen: 83.64 vs. 84.07). Without variance or significance tests, the reader cannot tell whether the reported ordering among adapters and full fine-tuning is reliable. The authors should report standard deviations or confidence intervals, and ideally pairwise significance tests, at least for the key comparisons in Tables 1 and 3.
  4. [§7.1; Tables 10-11; Appendix G-H] The MLM pseudo-perplexity results for languages unseen by mBERT are dominated by the UNK-token artifact that the authors themselves describe: Sinhala and Amharic show artificial pre-adaptation pseudo-perplexity that increases after adaptation, and per-language values such as 56052.75 for Amharic and 15640.68 for Sinhala make average pseudo-perplexity for the unseen group not comparable across configurations. The paper does acknowledge this in Section 7.1, but the aggregate numbers in Table 1 still mix these unreliable values into the averages used in the main comparison. The unseen-language MLM averages should either be reported with the artifact documented in the main table or excluded from the headline comparisons.
minor comments (5)
  1. [§7.4] There is a typo, "surprisinlgy", in the sentence about adaptation data volumes; it should read "surprisingly".
  2. [Table 9] The table header says "Langauge adapter" for LLaMA-3; this should be "Language adapter".
  3. [Appendix A] The appendix title "ConceptNet Tripple Conversion Mapping" contains a typo; it should be "Triple".
  4. [Table 2] The column header "#Params (B)" is ambiguous: for mBERT and XLM-R it lists total model size in billions (0.177 and 0.279), while for LLMs it lists parameter counts; clarifying that these are total model parameters, not trainable adapter parameters, would help readers interpret the efficiency claim.
  5. [§4.2; Appendix F] The training step caps of 100,000 for GlotCC and 25,000 for ConceptNet are stated, but there is no early stopping criterion or discussion of whether all languages reached the cap; reporting actual training steps or convergence behavior would strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical comparisons against external benchmarks, and self-citations are methodological rather than load-bearing.

full rationale

The paper makes no formal derivation that reduces to its inputs. All central results are measured against external datasets: MLM pseudo-perplexity on FLORES-200, topic classification on SIB-200, sentiment datasets from multiple sources, and NER on WikiANN (Tables 1-3, Sections 4.3 and 6). The adapter-versus-full-fine-tuning and small-versus-large-LLM comparisons are empirical evaluations, not quantities fitted from the data being 'predicted.' The only self-citations (Gurgurov et al. 2024a,b) are used to justify the ConceptNet-to-text conversion and to contrast prior trend differences (Section 3.2, Section 7.4); they do not supply a theorem or constrain the outcome, and the current experiments independently re-run the adaptation pipeline. The abstract's claim that adapters 'match or outperform full fine-tuning' is contradicted by the paper's own Table 1 on topic classification and sentiment analysis, and Section 5.2.1 concedes full fine-tuning is better on TC; this is an internal-consistency or correctness concern, not circularity. Similarly, comparing adapter-tuned mLMs with prompted LLMs is an uneven baseline choice, but it is not a definitional equivalence. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors, and no ansatz is smuggled in via citation. Consequently the derivation chain is self-contained against external benchmarks and the circularity score is 0.

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

The central results rest on standard adapter architectures and public datasets; the main uncontrolled choices are fixed hyperparameters, training step caps, and the 1 GB data cap. No new entities are introduced.

free parameters (3)
  • Adapter hyperparameters = reduction=16, LoRA r=8 alpha=8, LR=1e-4, batch=16
    Chosen from AdapterHub defaults and held fixed; the relative performance of adapter architectures could shift with different capacities.
  • Training step caps = 100,000 steps for GlotCC; 25,000 for ConceptNet
    Arbitrary limits on adaptation; more training could change results, as the authors acknowledge in Limitations.
  • GlotCC per-language cap = 1 GB
    Defines the low-resource setting; results may differ without truncation, as the Limitations section notes.
assumptions (4)
  • domain assumption Pseudo-perplexity on FLORES-200 devtest is a valid proxy for language modeling quality.
    Used as intrinsic metric; Section 7.1 notes it is unreliable for languages with scripts unseen by mBERT, so some MLM gains may reflect tokenization artifacts.
  • domain assumption GlotCC training text does not overlap with the evaluation sets used for MLM and downstream tasks.
    No deduplication against FLORES, SIB-200, or WikiANN is reported; CommonCrawl-derived corpora can contain benchmark content, which would inflate scores.
  • domain assumption Fixed default adapter hyperparameters provide a fair comparison across architectures and against full fine-tuning.
    Appendix F fixes reduction factor 16, LoRA r=8 alpha=8, learning rates, and batch sizes; different hyperparameters could reorder the adapter ranking.
  • domain assumption The 30 selected languages represent the low-resource spectrum described by Joshi et al. (2020).
    The set includes relatively well-resourced languages such as Thai, Romanian, and Bulgarian; the paper partially controls for this by reporting seen/unseen splits.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Small Models, Big Impact: Efficient Corpus and Graph-Based Adaptation of Small Multilingual Language Models for Low-Resource Languages." pith.science (2026). https://pith.science/paper/CCBKDXAO

@misc{pith2026250210140,
  author       = {Pith},
  title        = {Pith review of: Small Models, Big Impact: Efficient Corpus and Graph-Based Adaptation of Small Multilingual Language Models for Low-Resource Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CCBKDXAO}},
  note         = {Machine review of arXiv:2502.10140}
}
read the original abstract

Low-resource languages (LRLs) face significant challenges in natural language processing (NLP) due to limited data. While current state-of-the-art large language models (LLMs) still struggle with LRLs, smaller multilingual models (mLMs) such as mBERT and XLM-R offer greater promise due to a better fit of their capacity to low training data sizes. This study systematically investigates parameter-efficient adapter-based methods for adapting mLMs to LRLs, evaluating three architectures: Sequential Bottleneck, Invertible Bottleneck, and Low-Rank Adaptation. Using unstructured text from GlotCC and structured knowledge from ConceptNet, we show that small adaptation datasets (e.g., up to 1 GB of free-text or a few MB of knowledge graph data) yield gains in intrinsic (masked language modeling) and extrinsic tasks (topic classification, sentiment analysis, and named entity recognition). We find that Sequential Bottleneck adapters excel in language modeling, while Invertible Bottleneck adapters slightly outperform other methods on downstream tasks due to better embedding alignment and larger parameter counts. Adapter-based methods match or outperform full fine-tuning while using far fewer parameters, and smaller mLMs prove more effective for LRLs than massive LLMs like LLaMA-3, GPT-4, and DeepSeek-R1-based distilled models. While adaptation improves performance, pre-training data size remains the dominant factor, especially for languages with extensive pre-training coverage.

Figures

Figures reproduced from arXiv: 2502.10140 by the authors.

Figure 1
Figure 1. Correlation between the pre-training data sizes for mBERT and XLM-R and downstream task results [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Correlation between the pre-training data sizes for mBERT and XLM-R and the pseudo-perplexities with [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗
Figure 3
Figure 3. Correlation between the downstream performance for mBERT and XLM-R pre- and post-adaptation and [PITH_FULL_IMAGE:figures/full_fig_p030_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Correlation between the downstream performance for mBERT and XLM-R and the pre-training data and [PITH_FULL_IMAGE:figures/full_fig_p033_4.png]
Figure 5
Figure 5. Figure 5: Correlation between the downstream performance for mBERT and XLM-R and the pre-training data and [PITH_FULL_IMAGE:figures/full_fig_p036_5.png]
Figure 6
Figure 6. Figure 6: Correlation between the downstream performance for mBERT and XLM-R and the pre-training data [PITH_FULL_IMAGE:figures/full_fig_p039_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

  1. [4]

    Retrofitting Word Vectors to Semantic Lexicons

    InProceedings of the 5th Workshop on Noisy User-generatedText(W-NUT2019) ,pages364–369, Hong Kong, China. Association for Computational Linguistics. DeepSeek-AI,DayaGuo,DejianYang,HaoweiZhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li,...

    work page Pith review arXiv 2025

  2. [5]

    In International Conference on Machine Learning, pages 2790–2799

    Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790–2799. PMLR. Junjie Hu, Sebastian Ruder, Aditya Siddhant, Gra- ham Neubig, Orhan Firat, and Melvin Johnson

    work page

  3. [7]

    Ayyoob Imani, Peiqin Lin, Amir Hossein Kargaran, Silvia Severini, Masoud Jalili Sabet, Nora Kass- ner, Chunlan Ma, Helmut Schmid, André Martins, FrançoisYvon,andHinrichSchütze.2023

    Not all languages are created equal in llms: Improving multilingual capability by cross-lingual- thought prompting. Ayyoob Imani, Peiqin Lin, Amir Hossein Kargaran, Silvia Severini, Masoud Jalili Sabet, Nora Kass- ner, Chunlan Ma, Helmut Schmid, André Martins, FrançoisYvon,andHinrichSchütze.2023. Glot500: Scaling multilingual corpora and language models t...

    work page 2023

  4. [11]

    InProceedings of the 7th Workshop onBalto-SlavicNaturalLanguageProcessing ,pages 114–119, Florence, Italy

    Improving sentiment classification in Slo- vak language. InProceedings of the 7th Workshop onBalto-SlavicNaturalLanguageProcessing ,pages 114–119, Florence, Italy. Association for Computa- tional Linguistics. Branislav Pecher, Ivan Srba, and Maria Bielikova

    work page

  5. [13]

    In2019 International Conference of Advanced Informat- ics: Concepts, Theory and Applications (ICAICTA), pages 1–5

    Improvingbi-lstmperformanceforindonesian sentiment analysis using paragraph vector. In2019 International Conference of Advanced Informat- ics: Concepts, Theory and Applications (ICAICTA), pages 1–5. IEEE. Alec Radford. 2018. Improving language understand- ing by generative pre-training. Surangika Ranathunga and Isuru Udara Liyanage

    work page 2018

  6. [14]

    Transactions on Asian and Low-Resource Language Information Processing, 20(4):1–23

    Sentimentanalysisofsinhalanewscomments. Transactions on Asian and Low-Resource Language Information Processing, 20(4):1–23. Nathaniel R. Robinson, Perez Ogayo, David R. Mortensen, and Graham Neubig. 2023. Chatgpt mt: Competitive for high- (but not low-) resource lan- guages. Julian Salazar, Davis Liang, Toan Q Nguyen, and Ka- trin Kirchhoff. 2019. Masked ...

    arXiv 2023

  7. [15]

    InProceedings of the 16th Conference of the European Chapter of the As- sociation for Computational Linguistics: Main Vol- ume, pages 949–956, Online

    Clustering word embeddings with self- organizingmaps.applicationonLaRoSeDa-alarge Romanian sentiment data set. InProceedings of the 16th Conference of the European Chapter of the As- sociation for Computational Linguistics: Main Vol- ume, pages 949–956, Online. Association for Com- putational Linguistics. Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-...

    arXiv 2019

  8. [2011]

    Multi-Task Text Classification using Graph Convolutional Networks for Large-Scale Low Resource Language

    Learning word vectors for sentiment analy- sis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Hu- man Language Technologies, pages 142–150, Port- land, Oregon, USA. Association for Computational Linguistics. Mounika Marreddy, Subba Reddy Oota, Lak- shmi Sireesha Vakada, Venkata Charan Chinni, and Radhika Mamid...

    work page Pith review arXiv 2021

Show all 15 references
  1. [2015]

    In Proceedings of the International Confer- enceRecentAdvancesinNaturalLanguageProcess- ing,pages249–257,Hissar,Bulgaria.INCOMALtd

    Sentiment analysis in Twitter for Macedo- nian. In Proceedings of the International Confer- enceRecentAdvancesinNaturalLanguageProcess- ing,pages249–257,Hissar,Bulgaria.INCOMALtd. Shoumen, BULGARIA. Georgios Kalamatianos, Dimitrios Mallis, Symeon Symeonidis, and Avi Arampatzis...

    2015 arXiv

  2. [2018]

    In Proceedingsofthe27thInternationalConferenceon ComputationalLinguistics,pages2242–2252,Santa Fe, New Mexico, USA

    Representations and architectures in neu- ral sentiment analysis for morphologically rich lan- guages: A case study from Modern Hebrew. In Proceedingsofthe27thInternationalConferenceon ComputationalLinguistics,pages2242–2252,Santa Fe, New Mexico, USA. Association for Computa- ...

    2021 arXiv

  3. [2019]

    Fabio Barth, Manuel Brack, Maurice Kraus, Pedro Or- tiz Suarez, Malte Ostendorf, Patrick Schramowski, and Georg Rehm

    Simple, scalable adaptation for neural ma- chine translation. Fabio Barth, Manuel Brack, Maurice Kraus, Pedro Or- tiz Suarez, Malte Ostendorf, Patrick Schramowski, and Georg Rehm. 2024. Occiglot euro llm leader- board. Steven Bird. 2022. Local languages, third spaces, and othe...

    2024

  4. [2020]

    In International Conference on Machine Learning, pages 4411–4421

    Xtreme: A massively multilingual multi- task benchmark for evaluating cross-lingual gener- alisation. In International Conference on Machine Learning, pages 4411–4421. PMLR. Haoyang Huang, Tianyi Tang, Dongdong Zhang, WayneXinZhao,TingSong,YanXia,andFuruWei

  5. [2021]

    ShaoxiongJi,ZihaoLi,IndraneilPaul,JaakkoPaavola, PeiqinLin,PinzhenChen,DayyánO’Brien,Hengyu Luo, Hinrich Schütze, Jörg Tiedemann, and Barry Haddow

    Should we stop training more monolingual models,andsimplyusemachinetranslationinstead? In Proceedings of the 23rd Nordic Conference on ComputationalLinguistics(NoDaLiDa) ,pages385– 390,Reykjavik,Iceland(Online).LinköpingUniver- sity Electronic Press, Sweden. ShaoxiongJi,ZihaoL...

    2024

  6. [2023]

    NusaWrites: Constructing high-quality corpora for underrepresented and extremely low- resourcelanguages. InProceedingsofthe13thInter- nationalJointConferenceonNaturalLanguagePro- cessing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Li...

    2020 arXiv

  7. [2024]

    MatthewE.Peters,MarkNeumann,RobertLogan,Roy Schwartz, Vidur Joshi, Sameer Singh, and Noah A

    Comparing specialised small and general large language models on text classification: 100 la- belled samples to achieve break-even performance. MatthewE.Peters,MarkNeumann,RobertLogan,Roy Schwartz, Vidur Joshi, Sameer Singh, and Noah A. Smith. 2019. Knowledge enhanced contextu...

    2019 arXiv

Pith tools

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