Pith. sign in

REVIEW 3 major objections 5 minor 35 references

Negation is not an embedding-level signal: it emerges only when text and video are fused, according to this study of political video-text pairs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 17:11 UTC pith:GPBKENJ2

load-bearing objection The paper's central asymmetry claim rests on an annotator the authors themselves show is blind to visual negation, and the key table doesn't add up; the cross-modal attention results are real but don't save the argument. the 3 major comments →

arxiv 2607.17712 v1 pith:GPBKENJ2 submitted 2026-07-20 cs.AI

Learning to Detect Cross-Modal Negation: An Analysis of Latent Representations and an Attention-Based Solution

Ali AbuSaleh , Leon Hammerla , Alexander Mehler This is my paper
classification cs.AI
keywords negation detectionvision-language modelscross-modal attentionmultimodal fusionlatent representationspolitical discourseLLM annotationvideo-text alignment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that negation — words like 'not' or 'never' and visual signals like contradiction or absence — is not stored as a distinct, separable pattern in the latent spaces of standard vision-language models. Instead, the authors argue, negation is a cross-modal construct: textual negation can often stand alone, but visual negation only becomes interpretable when aligned with linguistic context. To support this, they run classifier batteries on pooled embeddings (which perform at chance), then show that a cross-modal attention mechanism that fuses text and video embeddings outperforms unimodal baselines by up to 7% F1. They also analyze 3,222 political video-text pairs annotated by an LLM, and report that visual negation appears only when the text and video are aligned. If true, this means robust negation understanding requires architectures that explicitly model inter-modal dependencies, not just bigger or better unimodal encoders.

Core claim

The paper's central claim is that negation is not a unimodal, embedding-level signal. Using 26 classifiers on Qwen2.5-VL and JEPA2 embeddings, they find balanced accuracy and ROC AUC near 0.5 — chance level — even when nominal accuracy reaches 81%, and the DummyClassifier matches sophisticated models. PCA shows the dominant variance is modality-specific, not negation-related. In contrast, a simple cross-modal attention model that concatenates text and video embeddings and applies multi-head attention achieves up to +7.03% F1 over unimodal baselines, and the gain persists when both modalities use the same FLAVA embeddings, isolating the fusion mechanism as the cause. The paper further claims

What carries the argument

The central mechanism is a cross-modal attention block: text and video embeddings are concatenated into a single sequence and processed by multi-head self-attention, so every token can attend to tokens of the other modality. Four attention heads are optimal. This is supported by a secondary instrument, an LLM-based annotation pipeline (Qwen2.5-VL) that classifies negation type, confidence, and modality alignment (full, partial, single-modality, unrelated) for each video-text pair; its outputs are the ground truth for the asymmetry statistics.

Load-bearing premise

The statistical claims about visual negation depend on LLM-generated annotations whose own visual-negation ability is near-chance relative to humans, so the asymmetry may be an artifact of a text-biased annotator.

What would settle it

If a human-annotated dataset of visual negation (e.g., images or videos reliably labeled by humans as depicting absence or contradiction) shows that pooled VLM embeddings separate negated from non-negated samples above chance under the same 26-classifier battery, the paper's central claim would be falsified. Alternatively, if a text-only annotation of the same 3,222 pairs reproduces the full asymmetry (visual negation only in aligned pairs), that would indicate the annotator, not the data, is driving the effect.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If negation is not separable in pooled embeddings, then fine-tuning or better encoders alone will not fix negation understanding in multimodal systems; the fusion architecture must change.
  • Cross-modal attention provides a simple, effective baseline for multimodal negation tasks, and the gain over identical-feature unimodal baselines shows fusion itself matters.
  • The asymmetry result implies that evaluations of visual negation should include textual context; unimodal visual negation benchmarks may measure the wrong thing.
  • LLM annotations, despite the model's poor agreement with humans on visual negation, can serve as weak supervision for training fusion models, as the paper demonstrates.
  • The eight-class taxonomy offers a finer-grained target for future negation datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's strongest claim — that visual negation is semantically dependent on text — rests on labels produced by a model that itself cannot detect visual negation without text (kappa ≈ 0 with humans when visual input is present). A conservative reading is that the asymmetry may reflect the annotator's bias rather than a property of negation; the co-occurrence table could be re-analyzed with huma
  • An immediate testable extension: take a dataset where visual negation is independently verified (e.g., human captions of absence), then repeat the classifier battery — if the paper is right, pooled embeddings will still perform at chance; if wrong, a separable signal will appear.
  • The attention-weight analysis (negated samples show 29.4% lower cross-modal attention) hints that the model learns to suppress alignment on contradiction; this suppression pattern could be used as a weak but interpretable negation cue in its own right.
  • The authors do not test sequence-aware transformers on the same embeddings; the claim that 'no separable signal exists' is limited to pooled representations, so a sequence-aware model over token embeddings remains an open alternative.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies negation in multimodal political discourse. It claims that (i) negation is not linearly or non-linearly separable in the latent spaces of standard vision-language models (VLMs), (ii) visual negation is semantically dependent on linguistic context while textual negation is often independent, and (iii) a proposed cross-modal attention architecture improves F1 by up to +7.03% over unimodal baselines. The evidence comes from classifier experiments on text/video/photo embeddings, attention-weight analyses, and a statistical analysis of 3,222 video-text pairs annotated by Qwen2.5-VL.

Significance. If the asymmetry claim were valid, it would be a useful contribution to multimodal semantics: it would indicate that negation is not a unimodal embedding-level property and that cross-modal fusion is necessary. The paper also provides a scalable LLM-based annotation pipeline and a cross-modal attention architecture. However, the main statistical evidence for the asymmetry is invalidated by the paper's own annotation-reliability results and by arithmetic inconsistencies in the central contingency tables. The non-separability claim, as stated in the abstract, is also contradicted by matched-modality results in the paper. These problems are load-bearing rather than cosmetic, so the significance of the contribution is not established.

major comments (3)
  1. [§V.A.7 and Tables VI–VII] The central asymmetry claim ('visual negation is semantically dependent on linguistic context') is based on Qwen2.5-VL annotations in Tables VI and VII. However, §V.A.7 shows that the same model with full multimodal input agrees with human experts at κ≈0.07/−0.0, while text-only input yields κ≈0.53/0.47. This means the annotator cannot reliably detect visual negation and is text-biased. The finding that video negation occurs only in FULL/PARTIAL ALIGNMENT with text (Table VII) is therefore an artifact of the measurement instrument, not a property of visual negation. The paper's own limitation section acknowledges annotation biases but does not connect this to the validity of the headline asymmetry.
  2. [Table VII] Table VII is arithmetically inconsistent with Table VI. For 'Text negation No', the row entries sum to 2253, not 2547; for 'Text negation Yes' they sum to 2928, not 675; for 'Video negation No' the entries sum to 23, not the 2850 shown in Table VI; and the 'Video negation Yes' row sums to 2928 instead of 372. These totals cannot be reconciled with the stated N=3222. Since Table VII is the direct evidence for the 'video negation only appears in aligned pairs' claim, the statistical foundation of the asymmetry is broken.
  3. [Abstract and §IV.B vs. Table I] The abstract claims that 'negation does not form a linearly or non-linearly separable class' in VLM latent spaces. Table I shows matched-modality Photo Model achieving accuracy 0.9441 and ROC AUC 0.8632, and Text Model achieving 0.8037/0.8569—well above chance. The paper dismisses these as 'dataset-specific correlations or labeling artifacts' in §VI.b, but provides no evidence for that dismissal. At minimum, the abstract's unqualified claim overstates the finding; the matched-modality results are more naturally interpreted as showing that negation is separable when labels and embeddings come from the same modality.
minor comments (5)
  1. [§III] Typo: 'RWe then perform' should read 'We then perform'. Also, the dataset size appears twice: '13370 tweets' and '3 222 containing one media item'; clarify whether 3,222 is the subset used for the LLM analysis.
  2. [§IV.F] Inconsistent decimal formatting: '0.834 2' vs '0.8342', and 'FLA V A' should be 'FLAVA'.
  3. [Table I] The checkmark notation is used to indicate input and label modalities, but the table is not self-explanatory; consider a separate legend or clearer column headers. The row labels 'Text Model(video-contains)' and 'Text-photo Model' are ambiguous.
  4. [§V.A.2] The statement that 'over 75 percent' of data contain no detectable negation is imprecise; give the exact count or percentage.
  5. [§IV.E] The attention-weight distribution analysis reports p-values (e.g., p<0.0001, p=0.817) without specifying the test used, sample sizes, or effect sizes. Add these details or remove the inferential statistics.

Circularity Check

1 steps flagged

Key asymmetry claim is an artifact of the paper's own text-biased annotator: Qwen2.5-VL's 'video negation' labels are essentially textual, so the finding that visual negation depends on linguistic context is built into the measurement.

specific steps
  1. self definitional [Section V.A.6–V.A.7 (Tables VI–VII); Appendix A prompt; abstract claim]
    "Video negation occurs exclusively in aligned examples (FULL-ALIGNMENT or PARTIAL-ALIGNMENT), and is completely absent in unrelated or unimodal pairs. This provides strong evidence for the hypothesis that video-based negation is semantically dependent on textual context ... The original multimodal model yields near-zero agreement with humans (κ=0.07 with BU; κ=−0.0 with MO). When the same 3B model is given only tweet text, agreement rises dramatically to moderate levels (κ=0.53 with BU; κ=0.47 with MO)."

    The paper's headline asymmetry is derived from Tables VI/VII, whose video-negation labels are produced by Qwen2.5-VL. The same section reports that this model cannot detect visual negation in multimodal input (near-zero human agreement) and that its negation judgments are largely a function of the text stream (text-only agreement 0.53/0.47). Thus 'video negation present' is effectively a text-driven projection onto the video channel. Concluding from these labels that 'visual negation is semantically dependent on linguistic context' restates the annotator's bias rather than testing the semantic property. Table VII also fails arithmetically (rows sum to 23 and 2,928, not 3,222; 2,447 video-negation cases are labeled UNRELATED, contradicting the exclusivity claim), so the statistical 'validat

full rationale

The non-separability and cross-modal-attention results are not circular: they are controlled comparisons on fixed embeddings with fixed labels (e.g., chance-level balanced accuracy across 26 classifiers, cross-modal vs. self-attention on identical features), and the F1 gains are internally benchmarked. Those claims may be overgeneralized relative to Table I's matched-modality results, but that is a scope/correctness issue, not circularity. Self-citations ([1], [28]) are not load-bearing for the representation-learning or attention claims. The circularity is concentrated in the LLM-annotation analysis used for the 'key asymmetry': Qwen2.5-VL's video-negation labels are shown in V.A.7 to be text-driven and at chance against human visual judgment, and the conclusion that visual negation depends on linguistic context is drawn from those very labels. The prompt itself asks for separate text/video negation and alignment; because the annotator's video channel contributes little beyond text, the observed alignment pattern is an artifact of the measurement. Therefore the paper's central asymmetry claim reduces to its annotation instrument, while the rest of the paper stands independently.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities. Its load-bearing assumptions are the adequacy of pooled embeddings for separability claims, the reliability of rule-based labels, and the validity of Qwen2.5-VL annotations for visual negation—the last of which fails against the paper's own human-validation results. The only fitted hyperparameters reported are the attention head count and training epochs.

free parameters (2)
  • number of attention heads = 4 (vs 8)
    Head count selected by comparing 4 vs 8 configurations in Table II; no principled criterion is given, and the selection is made on the same evaluation setup used to report the headline gains.
  • training epochs = 30
    All attention experiments use 30 epochs (Tables II, III, V, VIII); no early stopping, learning-rate schedule, or epoch sensitivity analysis is reported.
axioms (4)
  • domain assumption Pooled 2048-dimensional static embeddings are adequate to test whether negation is separable in VLM latent spaces.
    Section IV.B runs classifiers on pooled embeddings and generalizes to all latent-space separability; Section VI.e itself acknowledges sequence-aware models might behave differently.
  • domain assumption Rule-based and lexical negation detection on tweet text and transcriptions provides reliable labels for the embedding and attention experiments.
    Section III says negation detection uses rule-based and lexical pattern matching; no validation of these labels is reported for the classifier experiments.
  • ad hoc to paper Qwen2.5-VL annotations of visual negation are treated as ground truth for the 3,222-pair statistical analysis.
    Section V uses LLM outputs for Tables VI and VII, but Section V.A.7 reports multimodal LLM-human agreement kappa ≈ 0.07/−0.0, so this assumption fails.
  • standard math The chi-squared independence test is applicable to the LLM-annotated contingency table.
    Appendix C applies the chi-squared test to Table VI; the test is standard, but it inherits the validity of the LLM annotations.

pith-pipeline@v1.3.0-alltime-deepseek · 15400 in / 14245 out tokens · 149018 ms · 2026-08-01T17:11:58.906011+00:00 · methodology

0 comments
read the original abstract

Detecting high-level semantic concepts like negation across modalities remains a challenge for current multimodal systems. We analyze this as a fundamental representation learning problem, providing the first evidence that negation does not form a linearly or non-linearly separable class in the latent spaces of standard vision-language models (VLMs). We demonstrate that pretrained embeddings primarily encode modality-specific features, lacking a generalizable negation signal. To overcome this, we propose a novel cross-modal attention architecture that explicitly models inter-modal dependencies, achieving performance gains of up to +7.03% F1 over unimodal baselines. Our analysis reveals a key asymmetry: while textual negation often appears independently, visual negation is semantically dependent on linguistic context, a finding validated through our statistical analysis of 3,222 political video-text pairs automatically annotated via \textsc{Qwen2.5-VL}. By combining this analysis with self-supervised video representations (JEPA2), we advance the modeling of temporal negation. This work provides new methods and insights for learning robust, semantically-aligned representations in multimodal systems.

Figures

Figures reproduced from arXiv: 2607.17712 by Alexander Mehler, Ali Abusaleh, Leon Hammerla.

Figure 1
Figure 1. Figure 1: Overall pipeline we implement cross-modal attention mechanisms to fuse text and video embeddings, enabling joint modeling of negation across modalities. A. Negation Classes We define eight classes based on possible negation patterns in tweets. Each tweet includes either an image or a video, and all contain one of these media types. Classes are defined by the presence or absence of negation in the tweet tex… view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE visualization of LDA features for media, tweet [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of negation types across modalities. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of negation confidence across modalities. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Parallel Coordinates [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 3 canonical work pages

  1. [1]

    Learning to detect cross- modal negation: An analysis of latent representations and an attention- based solution,

    A. AbuSaleh, L. Hammerla, and A. Mehler, “Learning to detect cross- modal negation: An analysis of latent representations and an attention- based solution,” in2026 8th International Conference on Natural Language Processing (ICNLP), 2026, pp. 613–622

    2026

  2. [2]

    Negation,

    L. R. Horn and H. Wansing, “Negation,” inThe Stanford Encyclopedia of Philosophy, Spring 2025 ed., E. N. Zalta and U. Nodelman, Eds. Metaphysics Research Lab, Stanford University, 2025

    2025

  3. [3]

    Disentangling inhibition and prediction in negation processing,

    V . Haase, “Disentangling inhibition and prediction in negation processing,” Frontiers in Language Sciences, vol. V olume 3 - 2024, 2024. [Online]. Available: https://www.frontiersin.org/journals/language-sciences/articles/ 10.3389/flang.2024.1304613

    arXiv 2024

  4. [4]

    Vision-language models do not understand negation,

    K. Alhamoud, S. Alshammari, Y . Tian, G. Li, P. Torr, Y . Kim, and M. Ghassemi, “Vision-language models do not understand negation,”

  5. [5]

    How does “not left

    C. Dudschig and B. Kaup, “How does “not left” become “right”? electrophysiological evidence for a dynamic conflict-bound negation processing account,”J Exp Psychol Hum Percept Perform, vol. 44, no. 5, pp. 716–728, Nov. 2017

    2017

  6. [6]

    Negation as conflict: Conflict adaptation following negating vertical spatial words,

    D. Carolin and K. Barbara, “Negation as conflict: Conflict adaptation following negating vertical spatial words,”Brain Lang, vol. 210, p. 104842, Sep. 2020

    2020

  7. [7]

    Negation – an overview of typological research,

    M. Miestamo, “Negation – an overview of typological research,” Language and Linguistics Compass, vol. 1, no. 5, pp. 552–570, 2007. [Online]. Available: https://compass.onlinelibrary.wiley.com/doi/abs/10. 1111/j.1749-818X.2007.00026.x

    Pith/arXiv arXiv 2007

  8. [8]

    V-jepa 2: Self-supervised video models enable understanding, prediction and planning,

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, Mojtaba, Komeili, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholus, S. Arnaud, A. Gejji, A. Martin, F. R. Hogan, D. Dugas, P. Bojanowski, V . Khalidov, P. Labatut, F. Massa, M. Szafraniec, K. Krishnakumar, Y . Li, X. Ma, S. Chandar, F. Meier, Y . LeCun, M. Rabbat, and N. Ballas, “V-jepa 2: Self-supe...

    Pith/arXiv arXiv 2025

  9. [9]

    Qwen2.5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y . Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y . Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin, “Qwen2.5-vl technical report,” 2025. [Online]. Available: https://arxiv.org/abs/2502.13923

    Pith/arXiv arXiv 2025

  10. [10]

    A simple algorithm for identifying negated findings and diseases in discharge summaries,

    W. W. Chapman, W. Bridewell, P. Hanbury, G. F. Cooper, and B. G. Buchanan, “A simple algorithm for identifying negated findings and diseases in discharge summaries,”Journal of Biomedical Informatics, vol. 34, no. 5, p. 301–310, Oct. 2001. [Online]. Available: http://dx.doi.org/10.1006/jbin.2001.1029

    arXiv 2001

  11. [11]

    Deepen: A negation detection system for clinical text incorporating dependency relation into negex,

    S. Mehrabi, A. Krishnan, S. Sohn, A. M. Roch, H. Schmidt, J. Kesterson, C. Beesley, P. Dexter, C. Max Schmidt, H. Liu, and M. Palakal, “Deepen: A negation detection system for clinical text incorporating dependency relation into negex,”Journal of Biomedical Informatics, vol. 54, pp. 213–219, 2015. [Online]. Available: https://www.sciencedirect.com/science...

    2015

  12. [12]

    Negbio: a high-performance tool for negation and uncertainty detection in radiology reports,

    Y . Peng, X. Wang, L. Lu, M. Bagheri, R. Summers, and Z. Lu, “Negbio: a high-performance tool for negation and uncertainty detection in radiology reports,” 2017. [Online]. Available: https://arxiv.org/abs/1712.05898

    Pith/arXiv arXiv 2017

  13. [13]

    UiO1: Constituent-based discriminative ranking for negation resolution,

    J. Read, E. Velldal, L. Øvrelid, and S. Oepen, “UiO1: Constituent-based discriminative ranking for negation resolution,” in*SEM 2012: The First Joint Conference on Lexical and Computational Semantics – Volume 1: Proceedings of the main conference and the shared task, and Volume 2: Proceedings of the Sixth International Workshop on Semantic Evaluation (Sem...

    2012

  14. [14]

    A machine-learning approach to negation and speculation detection for sentiment analysis,

    N. P. Cruz, M. Taboada, and R. Mitkov, “A machine-learning approach to negation and speculation detection for sentiment analysis,”Journal of the Association for Information Science and Technology, vol. 67, no. 9, pp. 2118–2136, 2016. [Online]. Available: https://asistdl.onlinelibrary.wiley.com/doi/abs/10.1002/asi.23533

    work page doi:10.1002/asi.23533 2016

  15. [15]

    Automatic negation detection in narrative pathology reports,

    Y . Ou and J. Patrick, “Automatic negation detection in narrative pathology reports,”Artificial Intelligence in Medicine, vol. 64, no. 1, pp. 41–50, 2015. [Online]. Available: https://www.sciencedirect.com/science/ article/pii/S0933365715000159

    2015

  16. [16]

    Negbert: A transfer learning approach for negation detection and scope resolution,

    A. Khandelwal and S. Sawant, “Negbert: A transfer learning approach for negation detection and scope resolution,” 2020. [Online]. Available: https://arxiv.org/abs/1911.04211

    Pith/arXiv arXiv 2020

  17. [17]

    Improving negation detection with negation-focused pre-training,

    T. H. Truong, T. Baldwin, T. Cohn, and K. Verspoor, “Improving negation detection with negation-focused pre-training,” 2022. [Online]. Available: https://arxiv.org/abs/2205.04012

    Pith/arXiv arXiv 2022

  18. [18]

    Speculation and negation identification via unified machine reading comprehension frameworks with lexical and syntactic data augmentation,

    Z. Qian, T. Zou, Z. Zhang, P. Li, Q. Zhu, and G. Zhou, “Speculation and negation identification via unified machine reading comprehension frameworks with lexical and syntactic data augmentation,”Engineering Applications of Artificial Intelligence, vol. 131, p. 107806, 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0952197623019905

    2024

  19. [19]

    Thunder- nubench: A benchmark for llms’ sentence-level negation understanding,

    Y . So, G. Lee, S. Jung, J. Lee, J. Kang, S. Kim, and J. Lee, “Thunder- nubench: A benchmark for llms’ sentence-level negation understanding,”

  20. [20]

    Towards the roots of the negation problem: A multilingual NLI dataset and model scaling analysis,

    T. Vrabcová, M. Kadl ˇcík, P. Sojka, M. Štefánik, and M. Spiegel, “Towards the roots of the negation problem: A multilingual NLI dataset and model scaling analysis,” inFindings of the Association for Computational Linguistics: EMNLP 2025, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V . Peng, Eds. Suzhou, China: Association for Computational Lingui...

    2025

  21. [21]

    Available: https://arxiv.org/abs/2506.14397

    [Online]. Available: https://arxiv.org/abs/2506.14397

    Pith/arXiv arXiv

  22. [22]

    A review on negation role in twitter sentiment analysis,

    I. Gupta and N. Joshi, “A review on negation role in twitter sentiment analysis,”International Journal of Healthcare Information Systems and Informatics, vol. 16, no. 4, p. 1–19, Jul. 2021. [Online]. Available: http://dx.doi.org/10.4018/IJHISI.20211001.oa14

    work page doi:10.4018/ijhisi.20211001.oa14 2021

  23. [23]

    Negation scope detection for Twitter sentiment analysis,

    J. Reitan, J. Faret, B. Gambäck, and L. Bungum, “Negation scope detection for Twitter sentiment analysis,” inProceedings of the 6th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, A. Balahur, E. van der Goot, P. V ossen, and A. Montoyo, Eds. Lisboa, Portugal: Association for Computational Linguistics, Sep. 2015, ...

    2015

  24. [24]

    Context matters: A pragmatic study of PLMs’ negation understanding,

    R. Gubelmann and S. Handschuh, “Context matters: A pragmatic study of PLMs’ negation understanding,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), S. Muresan, P. Nakov, and A. Villavicencio, Eds. Dublin, Ireland: Association for Computational Linguistics, May 2022, pp. 4602–

    2022

  25. [25]

    Can negation be depicted? comparing human and machine understanding of visual representations,

    Y . Sato, K. Mineshima, and K. Ueda, “Can negation be depicted? comparing human and machine understanding of visual representations,” Cognitive Science, vol. 47, no. 3, p. e13258, 2023. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1111/cogs.13258

    work page doi:10.1111/cogs.13258 2023

  26. [26]

    X-pool: Cross-modal language-video attention for text-video retrieval,

    S. Gorti, N. V ouitsis, J. Ma, K. Golestan, M. V olkovs, A. Garg, and G. Yu, “X-pool: Cross-modal language-video attention for text-video retrieval,” 06 2022, pp. 4996–5005

    2022

  27. [27]

    LINKED: Eliciting, filtering and integrating knowledge in large language model for commonsense reasoning,

    J. Li, P. Cao, C. Wang, Z. Jin, Y . Chen, K. Liu, X. Jiang, J. Xu, and J. Zhao, “LINKED: Eliciting, filtering and integrating knowledge in large language model for commonsense reasoning,” inFindings of the Association for Computational Linguistics: EMNLP 2024, Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, Eds. Miami, Florida, USA: Association for Computatio...

    2024

  28. [28]

    Self-supervised learning from images with a joint-embedding predictive architecture,

    M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y . LeCun, and N. Ballas, “Self-supervised learning from images with a joint-embedding predictive architecture,” 2023. [Online]. Available: https://arxiv.org/abs/2301.08243

    Pith/arXiv arXiv 2023

  29. [29]

    The "something something

    R. Goyal, S. E. Kahou, V . Michalski, J. Materzy ´nska, S. Westphal, H. Kim, V . Haenel, I. Fruend, P. Yianilos, M. Mueller-Freitag, F. Hoppe, C. Thurau, I. Bax, and R. Memisevic, “The "something something" video database for learning and evaluating visual common sense,” 2017. [Online]. Available: https://arxiv.org/abs/1706.04261

    Pith/arXiv arXiv 2017

  30. [30]

    Flava: A foundational language and vision alignment model,

    A. Singh, R. Hu, V . Goswami, G. Couairon, W. Galuba, M. Rohrbach, and D. Kiela, “Flava: A foundational language and vision alignment model,” 2022. [Online]. Available: https://arxiv.org/abs/2112.04482 APPENDIX This appendix presents methodological and analytical details supporting the LLM based negation detection framework. It includes the structured pro...

    Pith/arXiv arXiv 2022

  31. [31]

    Extending a parliamentary corpus with mps’ tweets: Automatic annotation and evaluation using multipartweet,

    M. Bagci, A. Abusaleh, D. Baumartz, G. Abrami, M. Konca, and A. Mehler, “Extending a parliamentary corpus with mps’ tweets: Automatic annotation and evaluation using multipartweet,” 2025. [Online]. Available: https://arxiv.org/abs/2512.11567

    arXiv 2025

  32. [32]

    text_negation

    If negations align between text and video 4Inputs Provided: 5- Political text excerpt: [TEXT ADDED HERE] 6- Video :Attached 7Analysis Framework 8A. TEXT NEGATION VERIFICATION: 9[ ] Check for explicit negation markers: 10* Direct negatives (not, never, no) 11* Implicit denial terms (deny, reject, oppose) 12[ ] Identify political negation type: 13* Policy r...

  33. [34]

    Whether negation exists in EACH modality separately

  34. [2025]

    Available: https://arxiv.org/abs/2501.09425

    [Online]. Available: https://arxiv.org/abs/2501.09425

    Pith/arXiv arXiv

  35. [4621]

    Available: https://aclanthology.org/2022.acl-long.315/

    [Online]. Available: https://aclanthology.org/2022.acl-long.315/

    2022