Pith. sign in

REVIEW 4 major objections 5 minor 29 references

HNCSE: Advancing Sentence Embeddings via Hybrid Contrastive Learning with Hard Negatives

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read Hard-negative mixing into positives and negatives gives sentence embeddings that beat SimCSE

desk verdict Promising new Positive Mixing idea, but Eq. (2) is broken and the method is not reproducible from the text; desk reject. read the letter →

arxiv 2411.12156 v1 pith:GZLTGIMW submitted 2024-11-19 cs.CL cs.AI

classification cs.CLcs.AI
keywords unsupervisedsentencerepresentationlearningcontrastivehardnegativesamplesSimCSEpositivemixingsemantictextualsimilaritytransferembeddings
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 tries to establish that unsupervised sentence representations improve when contrastive training deliberately exploits hard negatives—samples that sit close to the query in embedding space and are therefore easy to confuse with the positive. The authors extend SimCSE with two mechanisms. Positive Mixing interpolates the positive target with the hardest in-batch negative when the positive is unexpectedly farther from the query than that negative, pulling the positive closer. Hard Negative Mixing linearly blends embeddings of the top-k nearest negatives to synthesize new, more challenging negatives. If the reported results hold, the claim is that a simple geometry-based hard-negative recipe gives consistent gains over SimCSE on semantic similarity and transfer tasks, and also outperforms prompted LLaMA2-7B variants on STS.

What carries the argument

Two named mechanisms carry the argument. Positive Mixing (PM) is the conditional interpolation rule of Eq. (2): it compares the query–positive similarity d1 with the query–hardest-negative similarity d2 and, depending on the gap, replaces the positive target s+1 with a weighted mixture w1s+1+w2s−1 (or with a ratio-weighted mixture when d1<d2), so the positive moves toward the query and away from the confusing negative. Hard Negative Mixing (HNM) is the linear-combination generator of Eq. (7): from the top-k nearest neighbours of the query in a previous minibatch it forms m new negatives uo = αGi + (1−α)Gj, with α adapted to the similarity of Gi and Gj, L2-normalizes them, and adds them to the negative bank of the InfoNCE loss. The paper's Appendix A offers additional formalizations (hard-negative threshold, loss, gradients, embedding-space distance change) meant to show why mixing should improve discriminability.

What would settle it

Retrain HNCSE with the same hyperparameters but with the 'hardest negative' replaced by a random in-batch sentence; if the average STS Spearman score stays near 78, the claimed effect is not due to hardness. Alternatively, compare against a variant that defines hardness by a paraphrase model instead of cosine distance: if the paraphrase-based hard negatives do not improve on the cosine-based ones, the geometric assumption is the active ingredient; if they do, the paper's mechanism is a proxy rather than the cause.

Watch

Extended reading notes

Core claim

The central discovery is that hard negatives can be used in both directions: to repair weak positives and to create stronger negatives. For each query, the method computes cosine distances d1 to its dropout positive and d2 to the most similar in-batch sentence. When d1 and d2 are close or d1 < d2, Eq. (2) mixes the positive embedding with the hardest negative to produce a new positive target. Separately, Eq. (7) takes the top-k nearest neighbours of the query from a previous minibatch and creates new hard negatives by random or adaptive linear combination αGi+(1−α)Gj followed by L2 normalization. The paper reports Spearman averages of 78.38 (positive mixing) and 78.27 (hard negative mixing) on seven STS benchmarks with BERT-base, against 76.16–76.25 for SimCSE baselines, with larger BERT variants reaching 79.13 and 79.20.

Load-bearing premise

The load-bearing premise is that the nearest in-batch sentence by cosine similarity is genuinely the semantically hardest negative, and that linear interpolation in embedding space produces a semantically meaningful positive or negative; if geometric closeness does not track semantic difficulty, both mixup rules could still help for unrelated reasons.

Editorial extensions

If this is right

  • If HNCSE is right, unsupervised sentence embedding can be improved by a cheap, data-free extension of SimCSE: no new data or supervision, only interpolation rules applied during training.
  • Both HNCSE-PM and HNCSE-HNM beat the SimCSE baselines they reproduce on the average of seven STS benchmarks, and on the majority of individual STS and transfer tasks.
  • With BERT-large, the same mechanisms push average STS scores to about 79.2, suggesting the gains do not vanish as model capacity grows.
  • The reported results also imply that a small contrastively trained encoder can outperform LLaMA2-7B with prompt-engineering strategies on STS, i.e., task-specific contrastive training can beat generic generative prompting for similarity measurement.

Reading between the lines

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

  • A reader-level implication the paper does not state: the same cosine-proximity rule could be used to filter or re-weight negatives in other contrastive text objectives, not just SimCSE, because the mechanisms only require an embedding space and a batch.
  • The positive-mixing branch may function partly as a regularizer that smooths the embedding geometry, so part of the gain could survive even if nearest-by-cosine negatives are not semantically hard; a test would compare against mixing with random negatives or with true paraphrase negatives.
  • A natural extension is to define hardness by semantic similarity (e.g., entailment or paraphrase scores) rather than by cosine distance, and then measure whether the gains increase; the paper's threshold-based mixing rule supports such a plug-in change.
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 proposes HNCSE, an extension of unsupervised SimCSE for sentence embeddings, with two components: Positive Mixing (PM), which interpolates the positive embedding with the hardest in-batch negative, and Hard Negative Mixing (HNM), which generates synthetic hard negatives by linearly combining top-k in-batch negatives. The authors report Spearman correlations on seven STS benchmarks and accuracies on seven transfer tasks, claiming consistent improvements over SimCSE. They also include an appendix intended as a theoretical analysis of HNM.

Significance. Hard negatives are known to improve contrastive learning in vision, so adapting this idea to sentence embeddings is a plausible and potentially useful direction; the experimental study covers standard benchmarks. However, the central contribution is not implementable from the text: the positive-mixing rule is mathematically undefined or invalid, the HNM hyperparameters and selection function are unspecified, and the baselines are inconsistently chosen. The appendix does not provide a derivation from stated assumptions to the claimed improvements. Unless these gaps are closed and code, seeds, and variance estimates are supplied, the reported gains cannot be verified or attributed to the proposed mechanism.

major comments (4)
  1. [Section 3.1, Eq (2)] The positive-mixing update rule is not well-defined. The weights w1 and w2 are introduced but never given a value or constraint. In the third branch (d1 < d2), both coefficients (d1-d2)/d1 and (d1-d2)/d2 are negative, producing a non-convex, potentially non-normalized combination of s+_1 and s-_1; no renormalization is specified. Since s+_2 is used as the training target, this branch is mathematically invalid as written and prevents faithful reproduction.
  2. [Section 3.2.1, Eqs (7)-(8)] The hard-negative generation procedure is under-specified. The set G is first called 'top k vectors' but defined as G={G1,...,Gm}; the roles of k and m are unclear. The mixing coefficient α is given only as α=f(sim(Gi,Gj)) with f unspecified, and no values for α, k, or m are reported in the experiments. Appendix A.2 defines a different mixup with λ∈[0,1] and never connects it to Eq (7).
  3. [Section 4.2 and Tables 1-2] The comparisons are not apples-to-apples. Implementation details state that HNCSE uses a BERT model ('Utilizing a BERT model that's been finely optimized'), and Table 1 lists 'SIMCSE' without architecture, while Table 2 uses 'SimCSE-RoBERTa' as the baseline. If HNCSE is BERT-based, comparing against SimCSE-RoBERTa in Table 2 is favorable to HNCSE and does not isolate the proposed components. The text also overclaims: Section 4.3.1 says HNCSE surpasses SimCSE across all STS benchmarks, but Table 1 shows HNCSE-HNM-Base underperforms SIMCSE(reproduce) on STS12 (69.76 vs 70.82).
  4. [Section 4.3, Table 3, and Appendix A] No code, data, seeds, or variance estimates are provided, so the reported improvements over SIMCSE(reproduce) (e.g., 78.38 vs 76.16 averaged over seven STS tasks) cannot be distinguished from run-to-run noise. The so-called theoretical analysis in Appendix A is largely a restatement of the contrastive loss and generic claims; A.3 and A.4 are identical, and parameters β, λ, γ in Eqs (16), (17), (24) are never used in experiments. This does not provide the announced 'theoretical analysis of HNM'.
minor comments (5)
  1. [Table 1, Base section] The row 'SIMCSE(reproduce)' has STS-B 77.24, while the original SimCSE paper reports 76.85 for STS-B; the authors should clarify whether the reproduced baseline matches the original hyperparameters and report the variance across runs.
  2. [Section 4.3.1] The sentence 'In the study by Zhang et al. (2022), we benchmarked...' is grammatically awkward; the citation appears misplaced and should be rewritten.
  3. [Figure 3] Figure 3 contains garbled placeholder text (e.g., '???1', '?????????') and should be redrawn with proper labels.
  4. [Eq (7)] Equation (7) uses the same symbol uo for the unnormalized and normalized vector; the notation should distinguish them.
  5. [Section 4.3.4] The description of HNCSE-PMsingle ('only optimize the case where d1 is greater than d2') conflicts with the three-branch structure of Eq (2); clarify which branch is used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HNCSE's central results are benchmarked against external SimCSE baselines and standard STS/transfer tasks, and the hard-negative mixing idea is explicitly credited to prior external work.

full rationale

The main empirical claim of the paper, that HNCSE-PM and HNCSE-HNM improve over unsupervised SimCSE, is tested on standard external benchmarks (STS 2012-2016, STS-B, SICK-R, and SentEval transfer tasks), with SimCSE reproduced from the original external work. The proposed components add hard-negative-guided positive mixing and hard-negative mixup, neither of which is fit to the evaluation data and then renamed as a prediction. The HNM mixup idea is explicitly attributed to Kalantidis et al. (2020), an external source, rather than to an unverified self-citation chain. The appendix contains definitions and restatements of contrastive objectives, distance-change measures, and margin measures, but these are not load-bearing for the empirical comparison and do not constitute a derivation that reduces to its own inputs. The under-specification of Eq. (2), Eq. (7), and Eq. (8) is a reproducibility and correctness concern, not a circularity concern: the method as written may be impossible to instantiate, but the paper does not claim to predict a quantity that is already an input to the construction. No fitted input is relabeled as a prediction, no uniqueness theorem is imported from the authors' own prior work, and no ansatz is smuggled in through a self-citation. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 2 invented entities

The method rests on several unstated or unspecified choices: the weights and thresholds in Positive Mixing, the number and selection of hard negatives, and the validity of interpolating BERT embeddings as if they were semantically meaningful sentences. The appendix adds parameters (theta, beta, lambda, gamma) that are never used in experiments. These are not pure math constants or inherited SimCSE settings; they are free choices that the paper does not justify.

free parameters (7)
  • w1, w2 in Positive Mixing (Eq 2)
    Weights of the positive and the hardest negative in the mixup for the new positive; no definitions or values are given in the paper.
  • Threshold 0.1 in Eq 2 = 0.1
    Hand-chosen cutoff deciding whether to mix or keep the original positive; no sensitivity analysis is reported for this value.
  • k (top-k hard negatives)
    Number of nearest vectors selected from the previous minibatch in HNM; never specified.
  • m (number of generated hard negatives)
    Number of synthetic hard negatives produced per query in HNM; never specified.
  • alpha mixing coefficient in Eq 7
    Adaptive selection is defined as f(sim(Gi,Gj)) in Eq 8, but the function f is never specified.
  • theta threshold in Appendix Eq 9
    Threshold defining hard negatives in the theoretical analysis; not used in the experiments.
  • beta, lambda, gamma in Appendix A.8-A.16
    Diversity-difficulty tradeoff, regularization, and learning-rate decay parameters introduced in the appendix but never used or set in the experiments.
assumptions (5)
  • domain assumption InfoNCE contrastive loss is an effective objective for learning sentence embeddings (Eq 3, inherited from SimCSE).
    The entire framework optimizes a contrastive loss; no independent justification is given within this paper.
  • ad hoc to paper The nearest in-batch negative by cosine similarity in [CLS] embedding space is semantically the hardest negative and is a suitable mixup partner for the positive (Section 3.1, Figure 2).
    This is the load-bearing premise of Positive Mixing and is not validated with examples or ablations.
  • ad hoc to paper Linear interpolation of two hard-negative embeddings yields a valid, semantically meaningful hard negative (Eq 7, Appendix A.2).
    Embedding-space interpolation is not shown to correspond to any real sentence or semantic content.
  • ad hoc to paper A hard negative can be characterized by a similarity threshold theta (Eq 9), and the appendix analysis applies to the experimental setup.
    Theta is never specified or used in experiments; the appendix analysis is disconnected from the implementation.
  • domain assumption Dropout on the same sentence creates a valid positive pair (SimCSE assumption, Section 3).
    Borrowed from SimCSE; the paper does not re-examine it.
invented entities (2)
  • Synthetic hard negative embeddings u_o generated by mixing top-k negatives (Eq 7)
    purpose: Augment the negative bank with harder negatives during contrastive training.
    No external validation that interpolated vectors correspond to meaningful semantic negatives; they exist only as training-time vectors.
  • Mixed positive embedding s+_2 produced by interpolating the positive and the hardest negative (Eq 2)
    purpose: Replace the original positive to pull the query closer to a refined positive direction.
    No evidence that this interpolated point is a better positive target; the equation contains undefined weights.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HNCSE: Advancing Sentence Embeddings via Hybrid Contrastive Learning with Hard Negatives." pith.science (2026). https://pith.science/paper/GZLTGIMW

@misc{pith2026241112156,
  author       = {Pith},
  title        = {Pith review of: HNCSE: Advancing Sentence Embeddings via Hybrid Contrastive Learning with Hard Negatives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZLTGIMW}},
  note         = {Machine review of arXiv:2411.12156}
}
read the original abstract

Unsupervised sentence representation learning remains a critical challenge in modern natural language processing (NLP) research. Recently, contrastive learning techniques have achieved significant success in addressing this issue by effectively capturing textual semantics. Many such approaches prioritize the optimization using negative samples. In fields such as computer vision, hard negative samples (samples that are close to the decision boundary and thus more difficult to distinguish) have been shown to enhance representation learning. However, adapting hard negatives to contrastive sentence learning is complex due to the intricate syntactic and semantic details of text. To address this problem, we propose HNCSE, a novel contrastive learning framework that extends the leading SimCSE approach. The hallmark of HNCSE is its innovative use of hard negative samples to enhance the learning of both positive and negative samples, thereby achieving a deeper semantic understanding. Empirical tests on semantic textual similarity and transfer task datasets validate the superiority of HNCSE.

Figures

Figures reproduced from arXiv: 2411.12156 by the authors.

Figure 1
Figure 1. An example of a query with its positive, over-easy, over-hard and missing hard negative [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Schematic diagram of Positive Mixing. where fθ is the sentence encoder parameterized by θ, si and s˜i are two views of the i-th sentence, sim is a cosine similarity function, τ is a temperature parameter, and N is the batch size. This objective encourages the embeddings of the same sentence to be close to each other, while pushing away the embeddings of different sentences. This is a simple but effective method to l… view at source ↗
Figure 3
Figure 3. Schematic diagram of Hard Negative Mixing. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The influence of different hyperparameters on HNCSE. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 11 canonical work pages

  1. [1]

    Representation learning using multi-task deep neural networks for semantic classification and information retrieval

    Xiaodong Liu, Jianfeng Gao, Xiaodong He, Li Deng, Kevin Duh, and Ye-Yi Wang. Representation learning using multi-task deep neural networks for semantic classification and information retrieval. 2015

  2. [2]

    Self-supervised Contrastive Cross-Modality Representation Learning for Spoken Question Answering

    Chenyu You, Nuo Chen, and Yuexian Zou. Self-supervised contrastive cross-modality representation learning for spoken question answering. arXiv preprint arXiv:2109.03381, 2021

  3. [3]

    Bootstrapped unsupervised sentence representation learning

    Yan Zhang, Ruidan He, Zuozhu Liu, Lidong Bing, and Haizhou Li. Bootstrapped unsupervised sentence representation learning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5168--5180, 2021

  4. [4]

    Forward models: Supervised learning with a distal teacher

    Michael I Jordan and David E Rumelhart. Forward models: Supervised learning with a distal teacher. In Backpropagation, pages 189--236. Psychology Press, 2013

  5. [5]

    Simcse: Simple contrastive learning of sentence embeddings

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021

  6. [6]

    Diffcse: Difference-based contrastive learning for sentence embeddings

    Yung-Sung Chuang, Rumen Dangovski, Hongyin Luo, Yang Zhang, Shiyu Chang, Marin Solja c i \'c , Shang-Wen Li, Wen-tau Yih, Yoon Kim, and James Glass. Diffcse: Difference-based contrastive learning for sentence embeddings. arXiv preprint arXiv:2204.10298, 2022

  7. [7]

    Vision-language pre-training with triple contrastive learning

    Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liqun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671--15680, 2022

  8. [8]

    Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm

    Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021

Show all 29 references
  1. [9]

    A simple but tough-to-beat baseline for sentence embeddings

    Sanjeev Arora, Yingyu Liang, and Tengyu Ma. A simple but tough-to-beat baseline for sentence embeddings. In International conference on learning representations, 2017

  2. [10]

    How contextual are contextualized word representations? comparing the geometry of bert, elmo, and gpt-2 embeddings

    Kawin Ethayarajh. How contextual are contextualized word representations? comparing the geometry of bert, elmo, and gpt-2 embeddings. arXiv preprint arXiv:1909.00512, 2019

  3. [11]

    Skip-thought vectors

    Ryan Kiros, Yukun Zhu, Russ R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Skip-thought vectors. Advances in neural information processing systems, 28, 2015

  4. [12]

    Understanding the behaviors of bert in ranking

    Yifan Qiao, Chenyan Xiong, Zhenghao Liu, and Zhiyuan Liu. Understanding the behaviors of bert in ranking. arXiv preprint arXiv:1904.07531, 2019

  5. [13]

    On the sentence embeddings from pre-trained language models

    Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. On the sentence embeddings from pre-trained language models. arXiv preprint arXiv:2011.05864, 2022

  6. [14]

    Nice: Non-linear independent components estimation

    Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014

  7. [15]

    Whitening sentence representations for better semantics and faster retrieval

    Jianlin Su, Jiarun Cao, Weijie Liu, and Yangyiwen Ou. Whitening sentence representations for better semantics and faster retrieval. arXiv preprint arXiv:2103.15316, 2023

  8. [16]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597--1607. PMLR, 2020

  9. [17]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729--9738, 2020

  10. [18]

    Hard negative mixing for contrastive learning

    Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33: 0 21798--21809, 2020

  11. [19]

    Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation

    Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055, 2017

  12. [20]

    Sicknl: a dataset for dutch natural language inference

    Gijs Wijnholds and Michael Moortgat. Sicknl: a dataset for dutch natural language inference. arXiv preprint arXiv:2101.05716, 2021

  13. [21]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

  14. [22]

    Unsupervised sentence representation via contrastive learning with mixing negatives

    Yanzhao Zhang, Richong Zhang, Samuel Mensah, Xudong Liu, and Yongyi Mao. Unsupervised sentence representation via contrastive learning with mixing negatives. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11730--11738, 2022

  15. [23]

    Glove: Global vectors for word representation

    Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532--1543, 2014

  16. [24]

    An unsupervised sentence embedding method by mutual information maximization

    Yan Zhang, Ruidan He, Zuozhu Liu, Kwan Hui Lim, and Lidong Bing. An unsupervised sentence embedding method by mutual information maximization. arXiv preprint arXiv:2009.12061, 2020

  17. [25]

    Declutr: Deep contrastive learning for unsupervised textual representations

    John Giorgi, Osvald Nitski, Bo Wang, and Gary Bader. Declutr: Deep contrastive learning for unsupervised textual representations. arXiv preprint arXiv:2006.03659, 2020

  18. [26]

    Semantic re-tuning with contrastive tension

    Fredrik Carlsson, Amaru Cuba Gyllensten, Evangelia Gogoulou, Erik Ylip \"a \"a Hellqvist, and Magnus Sahlgren. Semantic re-tuning with contrastive tension. In International conference on learning representations, 2020

  19. [27]

    Angle-optimized text embeddings

    Xianming Li and Jing Li. Angle-optimized text embeddings. arXiv preprint arXiv:2309.12871, 2023

  20. [28]

    Simple techniques for enhancing sentence embeddings in generative language models

    Bowen Zhang, Kehua Chang, and Chunping Li. Simple techniques for enhancing sentence embeddings in generative language models. arXiv preprint arXiv:2404.03921, 2024

  21. [29]

    Senteval: An evaluation toolkit for universal sentence representations

    Alexis Conneau and Douwe Kiela. Senteval: An evaluation toolkit for universal sentence representations. arXiv preprint arXiv:1803.05449, 2018

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.