Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Learning Robust Negation Text Representations

T0 review · 4 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 Fine-tuning on LLM-distilled triples with diverse negation and hedging patterns substantially improves text encoders' negation handling without sacrificing general performance.

desk verdict A practical data-centric recipe for negation robustness that mostly hits, with a confounded diversity-vs-quantity comparison and loose pair filtering; worth refereeing. read the letter →

arxiv 2507.12782 v1 pith:2DMUVHWU submitted 2025-07-17 cs.CL

Thinh Hung Truong , Karin Verspoor , Trevor Cohn , Timothy Baldwin This is my paper · ORCID
classification cs.CL
keywords negationhedgingtextembeddingscontrastivelearningsyntheticdatadistillationHedgeTripleBERTencodersbenchmarks
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

Negation is poorly captured by text embeddings: minimal pairs such as "I go to school" and "I do not go to school" end up with nearly identical vectors. This paper proposes to fix that by distilling synthetic data from an instruction-following LLM, using linguistic taxonomies to generate four types of negation and two broad families of hedging. The resulting HedgeTriple dataset gives each anchor sentence a hedged variant treated as a positive and a negated variant treated as a negative. Fine-tuning a BERT-based sentence encoder with contrastive loss on these triples produces large gains on negation-focused benchmarks while keeping performance on general text-understanding benchmarks nearly unchanged. The same triples also improve a decoder-only LLM under LoRA fine-tuning, at the cost of a small drop on arithmetic reasoning.

What carries the argument

The load-bearing object is the HedgeTriple dataset: each instance is a triple (anchor, hedged positive, negated negative) built from LLM rewrites of an affirmative sentence, paired with the semantic ordering premise that a hedged variant is closer in meaning to the anchor than a negated variant. That ordering is what the contrastive loss encodes: cosine similarity is maximized between anchor and hedged variant and minimized between anchor and negated variant, so the embedding space learns to separate affirmation, uncertainty, and denial. The diversity machinery is the taxonomy-guided prompt combined with explicit hedge-cue lists (134 single-word and 45 multi-word cues), which stops the LLM from collapsing to one negation or hedging pattern.

What would settle it

A direct check would be to have human raters order a random sample of HedgeTriple outputs by closeness to the anchor; if negated variants are often rated closer than hedged variants, the training signal is corrupted. A complementary test is to evaluate HedgeMPNet on a negation benchmark whose negation types are absent from the generation taxonomy — a collapse back to baseline would show the gains come from taxonomy overlap, not general negation competence.

Watch

Extended reading notes

Core claim

The paper's central claim is that diversity in the negation signal, rather than the volume of training data, is what makes a text encoder negation-aware. The authors prompt an LLM to rewrite 31K anchor sentences into four negated variants (verbal, absolute, affixal, lexical) and two hedged variants (single-word and phrase cues), using a linguistically grounded taxonomy for negation and a two-way simplified taxonomy for hedging, then filter rewrites by Levenshtein distance to keep minimal pairs. Training MPNet with Multiple Negative Ranking Loss on this HedgeTriple dataset raises average negation-benchmark performance from 32.13 for the base model to 54.17, while the MTEB general average moves only from 58.79 to 57.14. Ablations indicate negation data drives the gains, hedging data protects general capabilities, and an optimum is reached around 150K training samples.

Load-bearing premise

The load-bearing premise is that the LLM's negated rewrites really contradict the anchor and its hedged rewrites really express weakened certainty, so the contrastive loss encodes true semantic orderings rather than surface phrasing artifacts.

Editorial extensions

If this is right

  • A retrieval system can handle exclusion queries more reliably with a negation-aware encoder of MPNet size rather than a much larger generative model.
  • Data diversity should be treated as a first-class design axis: models trained on similar volumes of only verbal negation underperform the taxonomy-diverse model.
  • The same HedgeTriple data is reusable for decoder-only LLMs: LoRA fine-tuning of an 8B instruct model improves its negation ranking and similarity judgments.
  • The main measured cost of negation awareness is mathematical reasoning, not general language understanding, so the tradeoff is concentrated rather than diffuse.

Reading between the lines

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

  • The recipe suggests a general template: use taxonomy-guided LLM distillation to inject other semantic contrasts, such as antonymy or factivity, into embedding spaces without human-labeled contrastive pairs.
  • Because hedged positives appear to preserve general capabilities, one could decouple discrimination from retention by tuning the ratio of negation negatives to hedging positives, or by annealing it during fine-tuning.
  • The observed arithmetic drop opens a testable hypothesis: negation sensitivity and exact numerical reasoning may compete for shared representational resources, which a targeted probe on operand order and subtraction could confirm.
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. This paper addresses negation robustness in text encoders. The authors construct HedgeTriple, a synthetic dataset of contrastive triples generated by prompting GPT-3.5 with linguistic taxonomies for negation (verbal, absolute, affixal, lexical) and hedging (single-word and multi-word cues), followed by a Levenshtein-distance filter. They finetune MPNet with a contrastive loss (Eq. 1) to obtain HedgeMPNet, and also LoRA-finetune Llama-3-8B-Instruct on pairwise conversions of the same triples. Experiments report large gains on four negation benchmarks (NevIR, ExcluIR, Cannot, M3-Counterfactual) with a modest aggregate drop on MTEB, and improved LLM negation performance with a degradation on GSM8K.

Significance. If the results hold, the paper makes a practically useful contribution: a relatively simple data-distillation recipe, accompanied by a public dataset, can substantially improve the negation sensitivity of a widely used encoder while retaining most general benchmark performance. The work is not circular: the synthetic data come from an external LLM (GPT-3.5) and the main evaluations use external negation benchmarks. The taxonomy-grounded generation, the release of HedgeTriple, and the LLM adaptation study with an error analysis of GSM8K degradation are strengths. However, two load-bearing points need strengthening before the claims can be accepted: the semantic ordering of the synthetic triples is not verified, and the diversity-over-quantity claim is confounded by large differences in training-set size. The data-attribution section also contains an internal inconsistency with its own table.

major comments (4)
  1. [Section 3.1.2–3.1.3 and Eq. (1)] The core assumption that a hedged variant is semantically closer to the anchor than a negated variant is never validated. The Levenshtein filter with an upper threshold of 60 characters is not sufficient to guarantee 'minimal pairs': the phrase-hedge example in §3.1.2 adds the clause 'It’s not entirely clear what’s happening, but...', which introduces content absent from the anchor, and the examples in §3.1.3 show that the filter admits pairs that are topically similar but not semantically minimal or properly contradictory. Since the contrastive loss in Eq. (1) directly encodes the assumed anchor–hedge–negation ordering, a nontrivial fraction of misordered triples could contribute to the reported gains rather than to genuine negation robustness. Please report a sample-based audit (e.g., 100–200 random triples) of whether the hedged variant is indeed closer to the anchor than the negated variant, and, if possible, an ablation that retrains on only high-confidence triples to show the gains are not driven by label noise.
  2. [Section 5.1, 'Diversity vs. quantity'] The claim that data diversity is more important than quantity is confounded by sample-count differences. HedgeTriple contains 248K training samples, whereas Jina uses ~50K and NegMPNet uses ~80K, so the models are not trained on 'similar data sizes.' Figure 2 varies total dataset size while holding the data-generation method fixed, so it cannot separate diversity from quantity. Please provide a controlled comparison at matched sample sizes (e.g., downsample HedgeTriple to 50K or 80K) and clarify how the 248K count is derived from 31K anchors with 4 negation and 2 hedging outputs, since 31K × 6 = 186K.
  3. [Section 5.1, Table 2] The data-attribution summary is internally inconsistent with the reported numbers. In Table 2, 'Only negation' has a higher negation average (55.57 as printed; 54.97 if recomputed from the row values) than HedgeMPNet (54.17), and it outperforms HedgeMPNet on ExcluIR and M3-Counterfactual, yet the text states that 'combining both data types leads to the best performance.' Please correct the aggregation and either qualify the claim or provide the analysis that supports it.
  4. [Section 4.2.1, M3-Counterfactual] The reformatting of M3-Counterfactual assigns original–negation pairs a score of −1, original–'no evidence' pairs 0, and original–hedged pairs 1, which is exactly the ordering imposed by HedgeTriple training. Gains on this benchmark therefore partly reflect how well the model internalized the training prior rather than an independent semantic judgment. Please report the other negation benchmarks separately when interpreting the aggregate result, or add an evaluation mode that does not presuppose this ordering.
minor comments (5)
  1. [General] There are several typos: 'comparsion' in §4.1, 'prevalance' and 'unqiue' in §2, 'suchs' in §7, 'wether' and 'apppearence' in Appendix A, and 'used to expressed ambiguity' in §1.
  2. [Section 4.1 / Reproducibility] The encoder finetuning hyperparameters (learning rate, batch size, number of epochs, pooling strategy) are not reported; only the LLM LoRA details appear in Appendix B. Please add these details or release a configuration file so that HedgeMPNet can be reproduced.
  3. [Section 5.1, Figure 2] The text says finetuning data sizes range from 10K to 200K, but the x-axis of Figure 2 appears to start at 40K; please clarify whether the 10K point is omitted or the axis range is mislabeled.
  4. [Section 3.1.3] The statement that a Levenshtein distance threshold of 60 is 'equivalent to 10 words' is imprecise, because Levenshtein distance is measured in character edits and the mapping to words is not defined; please rephrase or provide the actual filtering criterion.
  5. [Section 7 (Limitations)] The claim that 'the findings of this work are generalizable to a multilingual setting' is an unsupported empirical claim; a small multilingual experiment or a more cautious wording is needed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the synthetic-data contrastive method is grounded in external LLM generation and evaluated on independent benchmarks.

full rationale

The paper's derivation chain is not circular. HedgeTriple is generated by prompting GPT-3.5 with negation taxonomies from Pullum and Huddleston (2002) and hedging taxonomies from Crompton (1997), with cue lists from HedgePeer (Ghosal et al., 2022); the anchors come from an external negation triples dataset used to train Jina Embedding, and the labels (which variant is hedged versus negated) are assigned by the prompting scheme rather than derived from the evaluation targets. The contrastive loss in Eq. (1) is the standard MNRL objective, and the model is evaluated on external negation benchmarks (NevIR, ExcluIR, Cannot, M3-Counterfactual) and general MTEB benchmarks that are not used to fit any parameter of the model. The self-citations (Truong et al., 2022a, 2022b, 2023) appear only as related work or as a caveat about LLM generalization and are not load-bearing for the main claim. The Levenshtein filter in Section 3.1.3 is a data-quality filter, not a fitted quantity renamed as a prediction. The assumption that hedged variants are semantically closer to anchors than negated variants is a linguistic premise of the training objective, but the reported gains are measured against external ground truth, so the claim does not reduce to its inputs. The concern raised in the skeptical reading about possible label noise in generated triples is a validity risk, not a circularity, because the training labels are not constructed from the benchmark labels or from the evaluation metrics.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

The central result rests on the quality of the synthetic HedgeTriple data and the semantic ordering of hedge versus negation. No new physical or mathematical entities are introduced.

free parameters (2)
  • Levenshtein distance threshold = 60 (equivalent to 10 words)
    Hand-set in Section 3.1.3 to filter generated texts that are 'too different' from anchors. Not tuned against any benchmark, but it directly controls which triples are kept.
  • Negation-to-hedging output ratio = 4 negation, 2 hedging per anchor
    Design choice in Section 3.1.3. No search over this ratio is reported, and it affects the balance of positive and negative examples in the contrastive loss.
assumptions (2)
  • domain assumption Hedged variants are semantically more similar to the anchor than negated variants
    Core ordering assumption behind the contrastive learning setup (Section 3.1): anchors are paired with hedges as positives and negations as negatives.
  • domain assumption GPT-3.5 generation with the provided taxonomy prompts yields correct linguistic negation and hedging labels
    The entire dataset depends on the LLM correctly instantiating verbal, absolute, affixal, and lexical negation, as well as hedge cues. Location: Section 3.1.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Robust Negation Text Representations." pith.science (2026). https://pith.science/paper/2DMUVHWU

@misc{pith2026250712782,
  author       = {Pith},
  title        = {Pith review of: Learning Robust Negation Text Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DMUVHWU}},
  note         = {Machine review of arXiv:2507.12782}
}
read the original abstract

Despite rapid adoption of autoregressive large language models, smaller text encoders still play an important role in text understanding tasks that require rich contextualized representations. Negation is an important semantic function that is still not properly captured by such methods, affecting many downstream applications relying on text embeddings. We propose a strategy to improve negation robustness of text encoders, by distilling data from large language models using diverse patterns of negation and hedging. We adopt a standard contrastive learning strategy to finetune a strong BERT-based model, and observe large improvement in negation understanding capabilities while maintaining competitive performance on general benchmarks. In addition, we also show that our method can be adapted to LLMs, leading to improved performance on negation benchmarks.

Figures

Figures reproduced from arXiv: 2507.12782 by the authors.

Figure 1
Figure 1. Cosine similarities between the sentence [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Negation–general performance tradeoff when [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Relative difference wrt. all-mpnet-base-v2 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: HedgeMPNet compared with similar negation [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

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. Position: Text Embeddings Should Capture Implicit Semantics, Not Just Surface Meaning

    cs.CL 2025-06 conditional novelty 5.0 of 10

    State-of-the-art text embeddings lag far behind on tasks requiring pragmatic inference, stance detection, and social meaning, relative to their strong performance on surface semantic benchmarks.

Reference graph

Works this paper leans on

46 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Miriam Ansch \"u tz, Diego Miguel Lozano, and Georg Groh. 2023. https://doi.org/10.18653/v1/2023.inlg-main.12 This is not correct! negation-aware evaluation of language generation systems . In Proceedings of the 16th International Natural Language Generation Conference, pages 163--175, Prague, Czechia. Association for Computational Linguistics

    work page doi:10.18653/v1/2023.inlg-main.12 2023

  2. [2]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. https://arxiv.org/abs/2404.05961 LLM2Vec : L arge language models are secretly powerful text encoders . arXiv preprint

    arXiv 2024

  3. [3]

    Bowman, Gabor Angeli, Christopher Potts, and Christopher D

    Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. https://doi.org/10.18653/v1/D15-1075 A large annotated corpus for learning natural language inference . In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632--642, Lisbon, Portugal. Association for Computational Linguistics

    doi:10.18653/v1/d15-1075 2015

  4. [4]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    arXiv 2021

  5. [5]

    William Coster and David Kauchak. 2011. https://aclanthology.org/P11-2117 S imple E nglish W ikipedia: A new text simplification task . In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 665--669, Portland, Oregon, USA. Association for Computational Linguistics

    work page 2011

  6. [6]

    Peter Crompton. 1997. https://doi.org/10.1016/S0889-4906(97)00007-0 Hedging in academic writing: Some theoretical problems . English for Specific Purposes, 16(4):271--287

    work page doi:10.1016/s0889-4906(97)00007-0 1997

  7. [7]

    Ronen Eldan and Yuanzhi Li. 2023. https://arxiv.org/abs/2305.07759 Tinystories: How small can language models be and still speak coherent english? Preprint, arXiv:2305.07759

    arXiv 2023

  8. [8]

    Allyson Ettinger. 2020. https://doi.org/10.1162/tacl_a_00298 What BERT is not: Lessons from a new suite of psycholinguistic diagnostics for language models . Transactions of the Association for Computational Linguistics, 8:34--48

    doi:10.1162/tacl_a_00298 2020

Show all 46 references
  1. [9]

    Katja Filippova and Yasemin Altun. 2013. https://aclanthology.org/D13-1155 Overcoming the lack of parallel data in sentence compression . In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1481--1491, Seattle, Washington, USA. Asso...

    2013

  2. [10]

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang S...

    2023 doi

  3. [11]

    Gemini Team . 2024. https://arxiv.org/abs/2312.11805 Gemini: A family of highly capable multimodal models . Preprint, arXiv:2312.11805

    2024 arXiv

  4. [12]

    Tirthankar Ghosal, Kamal Kaushik Varanasi, and Valia Kordoni. 2022. https://doi.org/10.1145/3529372.3533300 Hedgepeer: a dataset for uncertainty detection in peer reviews . In Proceedings of the 22nd ACM/IEEE Joint Conference on Digital Libraries, JCDL '22, New York, NY, USA. ...

    2022

  5. [13]

    Michael Günther, Louis Milliken, Jonathan Geuter, Georgios Mastrapas, Bo Wang, and Han Xiao. 2023. https://arxiv.org/abs/2307.11224 Jina embeddings: A novel set of high-performance sentence embedding models . Preprint, arXiv:2307.11224

    2023 arXiv

  6. [14]

    Zellig S Harris. 1954. Distributional structure. Word, 10(2-3):146--162

    1954

  7. [15]

    Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun hsuan Sung, Laszlo Lukacs, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. 2017. https://arxiv.org/abs/1705.00652 Efficient natural language response suggestion for smart reply . Preprint, arXiv:1705.00652

    2017 arXiv

  8. [16]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR)

    2021

  9. [17]

    Arian Hosseini, Siva Reddy, Dzmitry Bahdanau, R Devon Hjelm, Alessandro Sordoni, and Aaron Courville. 2021. https://doi.org/10.18653/v1/2021.naacl-main.102 Understanding by understanding not: Modeling negation in language models . In Proceedings of the 2021 Conference of the N...

    2021 doi

  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]

    Joel Jang, Seonghyeon Ye, and Minjoon Seo. 2023. Can large language models truly understand prompts? a case study with negated prompts. In Transfer learning for natural language processing workshop, pages 52--62. PMLR

    2023

  12. [20]

    Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts

    Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2023. Dspy: Compiling declarative language model calls into self...

    2023 arXiv

  13. [21]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

    2020

  14. [22]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C. Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision -- ECCV 2014, pages 740--755, Cham. Springer International Publishing

    2014

  15. [23]

    Felice Antonio Merra, Omar Zaidan, and Fabricio de Sousa Nascimento. 2023. Improving the relevance of product search for queries with negations. In Companion Proceedings of the ACM Web Conference 2023, pages 86--89

    2023

  16. [24]

    Nikola Mrk s i \'c , Diarmuid \'O S \'e aghdha, Blaise Thomson, Milica Gasic, Lina M Rojas Barahona, Pei-Hao Su, David Vandyke, Tsung-Hsien Wen, and Steve Young. 2016. Counter-fitting word vectors to linguistic constraints. In Proceedings of the 2016 Conference of the North Am...

    2016

  17. [25]

    Niklas Muennighoff, Nouamane Tazi, Lo \" c Magne, and Nils Reimers. 2022. https://doi.org/10.48550/ARXIV.2210.07316 Mteb: Massive text embedding benchmark . arXiv preprint arXiv:2210.07316

    2022 doi arXiv

  18. [26]

    OpenAI Team . 2024. https://arxiv.org/abs/2303.08774 GPT-4 technical report . Preprint, arXiv:2303.08774

    2024 arXiv

  19. [27]

    Yulia Otmakhova, Karin Verspoor, Timothy Baldwin, Antonio Jimeno Yepes, and Jey Han Lau. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.286 M 3: Multi-level dataset for multi-document summarisation of medical studies . In Findings of the Association for Computational Li...

    2022 doi

  20. [28]

    Jiaxin Pei and David Jurgens. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.784 Measuring sentence-level and aspect-level (un)certainty in science communications . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9959--10011,...

    2021 doi

  21. [29]

    Pullum and Rodney Huddleston

    Geoffrey K. Pullum and Rodney Huddleston. 2002. https://doi.org/10.1017/9781316423530.010 Negation , chapter 9. Cambridge University Press

    2002 doi

  22. [30]

    Abhilasha Ravichander, Matt Gardner, and Ana Marasovic. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.598 CONDAQA : A contrastive reading comprehension dataset for reasoning about negation . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Pro...

    2022 doi

  23. [31]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    2019 arXiv

  24. [32]

    MohammadHossein Rezaei and Eduardo Blanco. 2024. Paraphrasing in affirmative terms improves negation understanding. arXiv preprint arXiv:2406.07492

    2024 arXiv

  25. [33]

    Thinh Truong, Timothy Baldwin, Trevor Cohn, and Karin Verspoor. 2022 a . https://doi.org/10.18653/v1/2022.naacl-main.309 Improving negation detection with negation-focused pre-training . In Proceedings of the 2022 Conference of the North American Chapter of the Association for...

    2022 doi

  26. [34]

    Thinh Hung Truong, Timothy Baldwin, Karin Verspoor, and Trevor Cohn. 2023. https://doi.org/10.18653/v1/2023.starsem-1.10 Language models are not naysayers: an analysis of language models on negation benchmarks . In Proceedings of the 12th Joint Conference on Lexical and Comput...

    2023 doi

  27. [35]

    Thinh Hung Truong, Yulia Otmakhova, Timothy Baldwin, Trevor Cohn, Jey Han Lau, and Karin Verspoor. 2022 b . https://aclanthology.org/2022.aacl-main.65 Not another negation benchmark: The N a N - NLI test suite for sub-clausal negation . In Proceedings of the 2nd Conference of ...

    2022

  28. [36]

    o rgy Szarvas, Rich 'a rd Farkas, Gy

    Veronika Vincze, Gy "o rgy Szarvas, Rich 'a rd Farkas, Gy "o rgy M 'o ra, and J 'a nos Csirik. 2008. The bioscope corpus: biomedical texts annotated for uncertainty, negation and their scopes. BMC bioinformatics, 9(11):1--9

    2008

  29. [37]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533

    2022 arXiv

  30. [38]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. https://arxiv.org/abs/2401.00368 Improving text embeddings with large language models . Preprint, arXiv:2401.00368

    2024 arXiv

  31. [39]

    Smith, Daniel Khashabi, and Hannaneh Hajishirzi

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.acl-long.754 Self-instruct: Aligning language models with self-generated instructions . In Proceedings of the 61st Annual Mee...

    2023 doi

  32. [40]

    Orion Weller, Dawn Lawrie, and Benjamin Van Durme. 2024. https://aclanthology.org/2024.eacl-long.139 N ev IR : Negation in neural information retrieval . In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: L...

    2024

  33. [41]

    Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. https://doi.org/10.18653/v1/N18-1101 A broad-coverage challenge corpus for sentence understanding through inference . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computation...

    2018 doi

  34. [42]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. https://doi.org/10.18653/v1/D18-1259 H otpot QA : A dataset for diverse, explainable multi-hop question answering . In Proceedings of the 2018 Conference...

    2018 doi

  35. [43]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. https://doi.org/10.18653/v1/P19-1472 H ella S wag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791--4...

    2019 doi

  36. [44]

    Wenhao Zhang, Mengqi Zhang, Shiguang Wu, Jiahuan Pei, Zhaochun Ren, Maarten de Rijke, Zhumin Chen, and Pengjie Ren. 2024. https://arxiv.org/abs/2404.17288 Excluir: Exclusionary neural information retrieval . Preprint, arXiv:2404.17288

    2024 arXiv

  37. [45]

    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...

  38. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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