{"id":"7937e752-d78d-4b9f-a2c7-68e5388b43a0","arxiv_id":"2412.08489","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-module denoising model for multimodal aspect-based sentiment analysis reports slight F1 gains on Twitter-15/17, with the most relevant baseline missing from the experiments.","lead":"This paper proposes DualDe, a model that filters unhelpful images in two stages while analyzing opinions about specific topics in tweets. It reports small performance gains on two Twitter benchmarks, but it omits the closest prior method and tunes a key setting on the test data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Novelty and SOTA claims rest on the unverified exclusion of M2DF (Zhao et al., 2023), cited as the closest curriculum-denoiser baseline but absent from all comparisons; if M2DF already handles aspect-image noise or matches DualDe's F1, both central claims collapse.","rationale":"After reading the paper, the DualDe architecture is clearly described and the ablation study shows that each component contributes to internal performance; I credit the authors for detailing HCD and AED and for reporting the alpha ratio analysis. The central risk is not internal inconsistency but the paper's unsupported exclusion of the closest prior work. The reader's weakest assumption and my concern are the same: the paper asserts that M2DF does not handle aspect-image noise and then omits M2DF from all experiments, despite the cited title suggesting multi-grained denoising. Without checking M2DF's actual method or including it in Table 2, neither the 'first' claim nor the 'highest scores' claim is established. The missing baseline is a genuine correctness risk, and the absence of variance reporting makes the small F1 gains additionally fragile. The concrete test proposed above would settle whether the concern actually lands. Therefore I would not change the reader's REJECT verdict.","tokens_in":10702,"tokens_out":5036,"duration_ms":55231,"concrete_test":"Obtain Zhao et al. (2023, EMNLP) and inspect the M2DF architecture for an explicit component that filters image regions by aspect relevance. Then add M2DF as a required baseline to Table 2 using the same Twitter-15/Twitter-17 splits, with at least five random seeds and mean/std reported. If M2DF contains an aspect-image denoising component, or if its reported F1 on either dataset is within or above DualDe's 67.1/67.3, the novelty and SOTA claims fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claims are (1) DualDe is the first model to address sentence-image and aspect-image noise jointly, and (2) it achieves state-of-the-art P/R/F1 on Twitter2015 and Twitter2017. Both claims depend on how M2DF (Zhao et al., 2023) is treated. The paper itself identifies M2DF as the closest curriculum-learning denoising framework, yet the only support for excluding it is the assertion in Section 2.2 that 'they do not account for aspect-image noise.' The reference list entry describes M2DF as a 'multi-grained multi-curriculum denoising framework,' and no M2DF result appears in Tables 2-4. If M2DF already combines curriculum denoising with aspect-image denoising, the novelty claim is factually wrong. Even if it does not, the SOTA claim is unsupported because the single most relevant baseline is omitted from all comparisons and no error bars or significance tests are provided. This is an evidentiary gap in the paper's own argument, not a disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DualDe, a framework for Multimodal Aspect-Based Sentiment Analysis (MABSA) that targets two types of visual noise: sentence-image noise and aspect-image noise. It combines a Hybrid Curriculum Denoising Module (HCD), which uses a weighted mixture of a CLIP-based similarity difficulty and a model-loss difficulty to schedule training, with an Aspect-Enhance Denoising Module (AED) built on a BART encoder-decoder with aspect-guided attention, SenticNet features, and a GCN over a dependency-based association matrix. Experiments on Twitter-15 and Twitter-17 report improved F1 over the listed baselines for JMASA, MASC, and MATE, and an ablation indicates that both main modules contribute. The central claims are that DualDe is the first model to concurrently address both noise types and that it achieves state-of-the-art precision, recall, and F1 on both benchmark datasets.","tokens_in":10845,"tokens_out":6832,"duration_ms":70574,"significance":"If the results were fully substantiated, the paper would make a useful incremental contribution by combining two types of denoising and by adapting curriculum learning with a composite difficulty metric. The ablation study and the positive F1 deltas on three tasks are encouraging, and the use of CLIP as an external similarity signal for curriculum difficulty is a reasonable design choice. However, the significance is currently capped by two load-bearing deficiencies: the most relevant prior framework, M2DF, is cited but entirely absent from the experiments, and the main curriculum weight α is selected on the test set. Because the paper's headline novelty and SOTA claims rest on these points, the significance of the reported numbers cannot be assessed as stated.","major_comments":[{"comment":"M2DF (Zhao et al., 2023), which the paper itself identifies in Section 2.2 as the closest curriculum-learning denoising framework for MABSA, never appears in any comparison table, and the only justification for its exclusion is the unsupported sentence 'they do not account for aspect-image noise.' The reference entry describes M2DF as a 'multi-grained multi-curriculum denoising framework,' which prima facie suggests it may already address both sentence-level and aspect-level noise. Without a direct comparison or at least a detailed, evidence-based discussion of M2DF's scope, the two central claims in Section 1 ('first to concurrently address both' noise types) and Section 4.3 ('highest scores across Precision, Recall, and F1') are unsubstantiated.","section":"§2.2, Tables 2–4"},{"comment":"The comprehensive difficulty weight α is selected by tuning on the test set: Section 3.1.3 states 'Empirical results indicate that setting α = 0.8 yields optimal performance,' and Section 4.4.2 evaluates candidate ratios directly by their test-set F1 scores. Because the reported JMASA F1 of 67.1 is the very number used to pick α, the result is an in-sample statistic rather than an honest evaluation. The paper does not describe a validation-based selection procedure or a nested resampling approach, so the reported improvements are optimistically biased and the SOTA claim is unsupported.","section":"§3.1.3, §4.4.2, Fig. 4"},{"comment":"No standard deviations, confidence intervals, or significance tests are reported for any result, and many of the claimed improvements are smaller than 0.25 F1 points (e.g., 0.24 on Twitter-17 JMASA F1, 0.15 on Twitter-17 MATE F1). With single-run evaluations, such small deltas are within typical run-to-run variation for transformer-based models. The paper should report multiple runs with standard deviations and, ideally, statistical significance tests before claiming to outperform the previous state of the art.","section":"Tables 2–4"},{"comment":"The GCN update rule is mathematically inconsistent as written. The summation runs over j = 1 to n, where n is the number of text tokens, but the association matrix A is defined in Section 3.2.2 as an (m+n) × (m+n) matrix over both image and text nodes. Additionally, the right-hand side uses h_i,l-1 inside the sum rather than h_j,l-1, which is the standard form. These issues make the graph convolution and its region-specific mixing ambiguous, and the implementation cannot be reproduced from the equations alone.","section":"§3.2.3, Eq. (16)"},{"comment":"Several critical hyperparameters are either not reported or not specified: the initial competence λ_init and the duration T in p(t) of Eq. (6) are never given numeric values; the fusion weights α1 and α2 in Eq. (17) are not stated to be learned parameters or fixed constants; and the dependency distance threshold in Section 3.2.2 is set to 2 without justification. Without these values, the HCD curriculum schedule and the AED fusion cannot be replicated, and the sensitivity of the results to these choices is unknown.","section":"§3.1.4, §3.2.4, §4.1"}],"minor_comments":[{"comment":"The introduction contains a nearly verbatim duplicated sentence: 'In real-world scenarios, not all images are relevant to the accompanying text; some even mislead the contextual and emotional understanding of the sentence.' followed two sentences later by 'In real-world scenarios, images accompanying text may not always be relevant and can sometimes mislead the interpretation of the sentence’s context and emotion.' One of these should be removed.","section":"§1"},{"comment":"The rewritten matrix condition in Eq. (14) says A_ij = cos(h_i,h_j) under '(II) and (III)', but conditions (II) and (III) describe disjoint regions (text-to-image vs. text-to-text). The intended construction of the association matrix is unclear, in particular whether the aspect-driven entries apply to image nodes only or to both text and image neighbors.","section":"§3.2.2, Eq. (14)"},{"comment":"The notation H^T = W + ~H^T_2 introduces a subscript '2' that is undefined. It is also not clear whether W is the embedding matrix or a learned projection. Please clarify the dimensions and the intended role of each term.","section":"§3.2.4, Eq. (19)"},{"comment":"The list of compared models omits M2DF despite it being described as the closest related work; also, for the JMASA task, the table lists 'CLIP' as a baseline, but CLIP is primarily a feature extractor rather than a full MABSA model, and the manner in which it is applied to the joint task is not described.","section":"§4.2"},{"comment":"There are numerous typographical and stylistic errors, such as 'Diffculty' in the Section 3.1.2 heading, 'Aspect-Enhance' vs. 'Aspect-Enhanced' in module names, 'which unrelated to the specific aspects' in the abstract, and inconsistent capitalization of Twitter15/Twitter17. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The missing M2DF comparison is a serious evidentiary gap that should be treated as load-bearing. If M2DF indeed already addresses aspect-image noise (as its title suggests), the paper's novelty claim collapses, and the authors would need to reposition the contribution. I would encourage requiring the authors to either run M2DF or provide a detailed, evidence-based analysis of its scope, and to move all hyperparameter selection (especially α) to a validation set with error reporting. The paper is not ready for acceptance in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: DualDe is a sensible combination of two known ideas—curriculum denoising with a CLIP-similarity plus model-loss difficulty mix, and an A3M-style aspect attention denoiser with SenticNet and GCN. The paper is clearly written and the ablation suggests both modules do something. The problem is that the two central claims—'first to handle both noise types' and 'state of the art on Twitter-15/17'—are both unsupported because M2DF (Zhao et al., 2023) is cited as the closest curriculum denoiser but never appears in any table.\n\nCredit where it is due: the HCD twist is a genuine variation on prior curriculum denoising. Mixing a fixed CLIP similarity with a per-epoch model loss is a reasonable way to make the curriculum less brittle, and the competence schedule from Platanios is applied cleanly. The AED module is mostly a reuse of A3M, but the paper discloses that, and the SenticNet and GCN additions are modest extensions. The ablation results are real evidence that both modules contribute; removing HCD or AESA costs a few F1 points, which is enough to believe the mechanism is not inert.\n\nThe soft spots are load-bearing. First, the missing M2DF baseline. The paper asserts that M2DF does not account for aspect-image noise, but no comparison with M2DF appears in Tables 2–4, and the reference list describes M2DF as a multi-grained multi-curriculum denoising framework. If M2DF already handles both noise types or lands within a point of DualDe's F1, both the novelty claim and the SOTA claim collapse. Second, the curriculum weight alpha is selected from Figure 4, which plots F1 against alpha values on what looks like the same test set later reported as the result. That is test-set tuning with no held-out validation. Third, the reported gains are small—0.75 and 0.24 F1 on JMASA, 0.08 on MATE—and there are no error bars or significance tests. For a 0.08-point difference, that is noise. None of this kills the engineering contribution, but it does kill the claims as written.\n\nThis paper is for MABSA practitioners who want a working recipe and a clear description of a combined denoising pipeline. It deserves a serious referee, not a desk reject, but the referee should demand a real M2DF comparison, a proper split for alpha selection, and error bars. Without those, it should not be accepted.","headline":"A competent engineering combo whose novelty and SOTA claims both hinge on one missing baseline, M2DF.","tokens_in":11510,"tokens_out":1630,"would_cite":false,"duration_ms":17938,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a dual-module model, DualDe, is the first to concurrently denoise sentence-image and aspect-image noise in multimodal aspect-based sentiment analysis, and reports state-of-the-art F1 scores on Twitter-15 and…","keywords":["multimodal aspect-based sentiment analysis","sentence-image denoising","aspect-image denoising","curriculum learning","aspect-guided attention","joint multimodal aspect-sentiment analysis","Twitter-15","Twitter-17"],"falsifier":"Run DualDe and the closest prior curriculum denoising framework on the same Twitter-15 and Twitter-17 splits and compare JMASA F1; if the prior framework already filters aspect-specific visual regions, or if it matches or beats the reported F1 of 67.1 and 67.3, the paper's novelty and state-of-the-art claims are refuted. As a second check, replace the fixed similarity score $d^s_i$ with random scores while keeping $d^l_i$; if the F1 does not drop, the claimed benefit of the hybrid difficulty metric is not causal.","tokens_in":10400,"feed_emoji":"🧹","tokens_out":12306,"duration_ms":105200,"temperature":0.7,"pith_summary":"The paper proposes DualDe, a model for multimodal aspect-based sentiment analysis that separates image noise into two kinds: noise at the level of the whole sentence-image pair, and noise at the level of individual image regions as they relate to a specific aspect. It claims to be the first to handle both kinds in a single system. A curriculum-learning module (HCD) sorts training samples by a composite difficulty score that mixes a fixed text-image similarity from a pretrained image-text model (CLIP) with a model-loss score refreshed each epoch, then feeds samples to the model in increasing difficulty. An aspect-enhance denoising module (AED) uses aspect-guided attention to suppress image regions and text tokens irrelevant to the aspect, and injects affective knowledge from SenticNet through a graph convolutional network. On the Twitter-15 and Twitter-17 benchmarks, the paper reports the highest precision, recall, and F1 among the compared models for the joint task, with JMASA F1 of 67.1 and 67.3.","feed_headline":"Dual-module denoising lifts Twitter sentiment F1 to 67.3","feed_subtitle":"The model pairs curriculum learning with aspect-guided attention and outperforms every baseline on Twitter-15 and Twitter-17.","key_machinery":"The central mechanism is the composite difficulty score $d^c_i = \\alpha d^l_i + (1-\\alpha) d^s_i$ with $\\alpha = 0.8$, fused with a competence-based curriculum schedule $p(t)$ that admits samples only when $d^c_i < p(t)$. This carries the sentence-image denoising half. The aspect-image half is carried by the AESA and GCN sub-modules: an aspect-aware attention mechanism produces aspect-relevant hidden states, SenticNet values sharpen sentiment polarity, and a graph convolutional network over a weighted association matrix propagates affective dependencies across text and image nodes while suppressing unrelated visual blocks. Together these two sub-systems implement the paper's two-source denoising claim.","core_discovery":"The paper's central claim is that sentence-image noise and aspect-image noise are distinct failure modes in multimodal aspect-based sentiment analysis, and that a model can address both simultaneously. DualDe combines the Hybrid Curriculum Denoising Module (HCD) and the Aspect-Enhance Denoising Module (AED). HCD orders training examples by a composite difficulty score $d^c_i = \\alpha \\cdot d^l_i + (1-\\alpha) \\cdot d^s_i$ with $\\alpha = 0.8$, where $d^s_i$ is a fixed cosine similarity between CLIP text and image features and $d^l_i$ is a normalized per-sample generation loss refreshed each epoch; a competence function $p(t)$ admits only samples easier than the current competence. AED takes an aspect-aware attention output, adds SenticNet affective values, and runs a graph convolutional network over a weighted association matrix that encodes aspect-to-image and dependency-tree relations, so noisy visual blocks are filtered before the BART decoder predicts aspect spans and polarities. On Twitter-15 and Twitter-17, the paper reports JMASA F1 of 67.1 and 67.3 and states these are the highest scores across precision, recall, and F1 among the compared models.","pith_inferences":["A testable extension outside the paper: the same composite difficulty score could be applied to other multimodal sequence-to-sequence tasks, such as multimodal named entity recognition, where sentence-image noise is implicitly present.","Because the $\\alpha=0.8$ weighting was selected by a search over ratios, a natural follow-up would be to anneal $\\alpha$ during training, since the model-loss difficulty $d^l_i$ should become more reliable as training progresses.","The paper's distinction between the two noise types implies that region-level noise annotations would let the two modules be evaluated separately rather than only through ablations; without such labels, the relative contribution of each module is inferred rather than directly measured."],"forward_implications":["If the two-module design is correct, adding a sentence-level curriculum on top of aspect-level attention should beat either denoising strategy alone; the paper's ablation table supports this, with removal of HCD and of AESA both reducing F1.","The reported gains on two different tweet corpora imply the approach transfers across datasets with different noise profiles, not just a single benchmark.","The ratio test, which found $\\alpha=0.8$ optimal, implies the model's own training signal is more informative than the fixed similarity score for ordering samples by difficulty.","The subtask results indicate the joint-task gains carry over to classification of known aspects (MASC) and extraction of aspect terms (MATE), although the MATE F1 gains are small."],"supporting_citations":[{"why":"Introduces curriculum learning, the training strategy that the HCD module adapts.","marker":"(Bengio et al., 2009)"},{"why":"Supplies the competence-based curriculum schedule $p(t)$ used to admit samples by difficulty.","marker":"(Platanios et al., 2019)"},{"why":"Provides the aspect-aware attention module (A3M) that the AED module builds on, and serves as the strongest previous SOTA baseline.","marker":"(Zhou et al., 2023)"},{"why":"The closest prior curriculum denoising framework; the paper defines its novelty claim against this work but does not compare with it.","marker":"(Zhao et al., 2023)"},{"why":"The CLIP model whose text-image cosine similarity defines the static difficulty metric $d^s_i$.","marker":"(Radford et al., 2021)"},{"why":"The BART backbone that carries the encoder-decoder architecture and loss of the AED module.","marker":"(Lewis et al., 2019)"},{"why":"Provides the Twitter-15 and Twitter-17 datasets and the TomBERT baseline.","marker":"(Yu and Jiang, 2019)"},{"why":"Defines the JMASA task formulation and the JML baseline with auxiliary cross-modal relation detection.","marker":"(Ju et al., 2021)"},{"why":"A vision-language pre-training baseline (VLP-MABSA) that the paper must beat on JMASA and subtasks.","marker":"(Ling et al., 2022)"},{"why":"Supplies SenticNet affective commonsense values that enhance sentiment features in AESA.","marker":"(Cambria et al., 2016)"}],"fun_headline_variants":["DualDe filters both sentence and aspect noise for top MABSA scores","Curriculum learning plus aspect attention beat MABSA baselines on Twitter","Two-step denoising pushes multimodal sentiment F1 past 67 on Twitter","New hybrid denoising modules set benchmark on aspect-based sentiment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's claims of novelty and state-of-the-art performance rest on the assumption that the closest prior curriculum denoising framework does not already handle aspect-image noise and therefore could be omitted from all comparisons; if that assumption is wrong, the claim to be the first to address both noise types and the claim of highest scores are unsupported.","fun_headline_variants_meta":{"raw":{"variants":["DualDe filters both sentence and aspect noise for top MABSA scores","Curriculum learning plus aspect attention beat MABSA baselines on Twitter","Two-step denoising pushes multimodal sentiment F1 past 67 on Twitter","New hybrid denoising modules set benchmark on aspect-based sentiment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1492,"prompt_tokens":967,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":447}},"tokens_in":583,"tokens_out":525,"duration_ms":6278,"temperature":1.0,"reasoning_tokens":447,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:44:37.098860+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DualDe and the closest prior curriculum denoising framework on the same Twitter-15 and Twitter-17 splits and compare JMASA F1; if the prior framework already filters aspect-specific visual regions, or if it matches or beats the reported F1 of 67.1 and 67.3, the paper's novelty and state-of-the-art claims are refuted. As a second check, replace the fixed similarity score $d^s_i$ with random scores while keeping $d^l_i$; if the F1 does not drop, the claimed benefit of the hybrid difficulty metric is not causal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the aspect-aware attention module (A3M) that the AED module builds on, and serves as the strongest previous SOTA baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The closest prior curriculum denoising framework; the paper defines its novelty claim against this work but does not compare with it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Twitter-15 and Twitter-17 datasets and the TomBERT baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A vision-language pre-training baseline (VLP-MABSA) that the paper must beat on JMASA and subtasks."}],"review_version":1}