Pith. sign in

REVIEW 5 major objections 7 minor 58 references

Multimodal Fine-grained Reasoning for Post Quality Evaluation

T0 review · 5 major / 7 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 Post-quality assessment is better framed as a multimodal ranking problem than as unimodal classification, and the proposed MFTRR framework is reported to rank posts more accurately than six baselines, with up to 9.52% NDCG@3 improvement…

desk verdict Plausible architecture and useful new datasets, but the experimental record has inconsistencies that undermine the central claim. read the letter →

arxiv 2507.17934 v1 pith:HV7APHC2 submitted 2025-07-21 cs.LG cs.AI

Xiaoxu Guo , Siyan Liang , Yachao Cui , Juxiang Zhou , Lei Wang , Han Cao This is my paper
classification cs.LGcs.AI
keywords multimodaldataanalysisforumpostspostqualityevaluationlearningtorankfusionrelationalreasoningMOOCdiscussionforumsnoisefiltering
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 post-quality evaluation in online discussion forums should be treated as a fine-grained multimodal ranking problem rather than a unimodal classification problem. It argues that the subtle judgments teachers make—whether a post covers all parts of a topic and answers each question accurately—can be modeled by reasoning about topic-post relations at two complementary scales: local versus global semantics, and macro versus micro evidence. The proposed MFTRR framework does this with a noise-filtering gated fusion of text and image features plus graph-based evidential reasoning, and it is trained with a pairwise ranking loss. On three newly constructed Chinese MOOC datasets and the public Lazada-Home dataset, MFTRR is reported to outperform six baselines, with the largest margin on the Art History dataset: NDCG@3 of 93.62 versus 84.10 for the best text-only method, a 9.52 percentage point improvement.

What carries the argument

The machinery is a two-module reasoning stack. The Local-Global Semantic Correlation Reasoning Module computes scaled dot-product attention between post and topic features at four local modality pairings plus a global whole-to-whole pairing, then combines them through five topic-conditioned gates designed to suppress noise and preserve the most topic-relevant semantic signal. The Multi-Level Evidential Relational Reasoning Module builds graph neural networks over the topic and post features, one graph for macro-level significant-information evidence retrieved by attention and one for micro-level internal-logic evidence weighted by coherence with the topic graph, and concatenates the two evidence features before a linear scoring head. The scoring head is trained with a pairwise margin ranking loss so that the outputs order posts by quality.

What would settle it

Take a random sample of posts from the three new datasets, have independent judges rank them by quality under a written rubric without seeing the model scores, and measure agreement with the paper's ground-truth scores; if agreement is near chance, every reported NDCG and MAP gap is unsupported. A weaker check is to retrain MFTRR on only the human-proofread labels and see whether the ranking gains persist.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that reframing post-quality assessment as ranking and reasoning over the topic-post pair in two complementary ways is enough to beat all tested baselines. MFTRR first projects text and image features into a shared latent space, computes four local attention relations (text-text, text-image, image-text, image-image) plus a global post-as-a-whole versus topic-as-a-whole relation, and then uses five topic-conditioned gates to filter noise and keep the maximally relevant semantic signal. In parallel, it constructs two graph-based evidence reasoners: a macro-level graph that retrieves salient relationship information via attention over a topic-post relation graph, and a micro-level graph that weights post nodes by their coherence with the topic graph. The concatenated semantic and evidence features feed a linear scorer optimized by a pairwise margin ranking loss. The paper reports consistently higher MAP, NDCG@3, and NDCG@5 than all six baselines on all four datasets, including 93.62 NDCG@3 on Art History versus 84.10 for the leading text-only method.

Load-bearing premise

The load-bearing premise is that the ground-truth quality scores are reliable: Section 4.1 reports that scores were pre-annotated by a multimodal correlation analysis model and then proofread by five people, but it does not report agreement between those five proofreaders or any adjudication rule, so if their scores disagreed the ranking labels and hence all reported NDCG and MAP values would be arbitrary.

Editorial extensions

If this is right

  • Multimodal post evaluation does not just add more features; it changes the task, since ranking with fine-grained topic-post reasoning outperforms text-only classifiers and rankers on all four datasets.
  • Noise in multimodal fusion is a first-class problem: the topic-based gating mechanism contributes to the reported gains, and ablations that remove it hurt performance.
  • Multi-level evidence reasoning matters: removing the evidential relational reasoning module causes the largest performance drops in the ablation study.
  • The three newly constructed multimodal topic-post datasets, together with Lazada-Home, offer a testbed for later work on multimodal educational post evaluation.
  • Because the model outputs a ranked list, instructors could use it to review the top posts, which is the review scenario reflected in the NDCG@3 and NDCG@5 metrics.

Reading between the lines

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

  • Editorial inference: because the topic-post pair structure is generic, the same ranking-plus-reasoning recipe should transfer to short-answer grading, where the topic becomes the question and the post becomes the student answer; an experiment on a public short-answer dataset would test that.
  • Editorial inference: the gated fusion idea suggests a testable extension in which noisy modalities are deliberately added, such as low-resolution images or misaligned captions, and the five gates should be expected to down-weight them; if the gates do not, the noise-filtering story is weakened.
  • Editorial inference: the 9.52-point NDCG@3 gain over text-only methods could partly reflect properties of the annotation process rather than the model alone, since the labels were pre-annotated by a multimodal model; re-annotating a subset with human-only scores and re-running the comparison would separate those effects.
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

5 major / 7 minor

Summary. The paper proposes MFTRR, a multimodal framework that casts post-quality assessment as a ranking task. It combines a local-global semantic correlation reasoning module, which models topic-post interactions at local and global scales with a gating-based noise filter, and a multi-level evidential relational reasoning module, which builds evidence graphs at macro and micro levels. The authors construct three Chinese MOOC datasets with multimodal topic-post pairs and evaluate on these plus the public Lazada-Home dataset, reporting that MFTRR outperforms six baselines and achieves up to a 9.52% NDCG@3 improvement over the best text-only method on the Art History dataset.

Significance. If the empirical claims held, the paper would contribute a new ranking-based formulation for post-quality evaluation, a substantial new set of multimodal educational datasets, and an architecture that combines semantic and evidential reasoning. The manuscript is commendable in scope: it reports extensive ablations (Tables 9 and 12) and constructs three datasets that could be useful to the community. However, the central empirical claim is not currently supported: experiments are single-run with no variance or significance testing, label reliability is not demonstrated, and several reported tables contain internal inconsistencies. The architecture description also has equation-level problems. The paper would be more valuable after a thorough revision that addresses these load-bearing issues.

major comments (5)
  1. [Section 4.1 (data annotation); Eq. (34); Table 7] Section 4.1 describes the only supervision signal: posts receive a 0-4 quality score, pre-annotated by a multimodal correlation model and then proofread by five individuals. No inter-annotator agreement, adjudication rule, or label-quality analysis is reported. Since every MAP/NDCG value and the pairwise ranking loss in Eq. (34) are computed from these scores, the headline comparison in Table 7 is unsupported without evidence that the scores are reliable. Please report Fleiss' kappa or an equivalent agreement measure, the adjudication procedure, and the score distribution for each dataset.
  2. [Section 5.1; Section 5.4; Tables 5-7 and 10] All reported metrics appear to be single runs with no error bars, confidence intervals, or significance tests. Several headline differences are small in absolute terms (e.g., Table 7: MFTRR NDCG@3 93.62 vs G2SAM 93.20), so the claim of significant outperformance needs a paired test across topic-level rankings. Moreover, Table 10 lists identical MFTRR(GCN) scores (68.41, 42.11, 47.18) for the Education and Excellence datasets; this strongly suggests an error in the experimental record and must be corrected before the results can be trusted.
  3. [Section 4.1; Table 4] Table 4 contradicts the statement that the test set is 20% of the entire dataset. For Art History, test topics are 282 of 548 (51.5%) and test posts 2468 of 8896 (27.7%); for Education, 240 of 766 (31.3%) topics and 5000 of 24142 (20.7%) posts; for Excellence, 428 of 1485 (28.8%) topics and 5000 of 24775 (20.2%) posts. There is no consistent 20% split for either topics or posts. Please reconcile the split definition or correct the table.
  4. [Section 3.3.2; Eq. (17)] Equation (17) omits the M_{V-W} ⊗ g_{V-W} term defined in Eq. (14) and uses M_{S-S} twice, once as an additive constant and once gated. In addition, Eq. (15) reuses the weight W_{V-W} for the g_{V-V} gate. As written, the fusion formula does not implement the five-gate mechanism described in the text, making the module unverifiable.
  5. [Section 3.4.1; Eqs. (23)-(24)] The retrieval attention formula is underspecified: the index set C in Eq. (24) is not defined, and the denominator in Eq. (23) uses S as a summation index that is not bound to the query position. Please clarify the notation and specify the index ranges.
minor comments (7)
  1. [Section 4.1] The score notation S_{i,j} = {0, ...S} leaves S undefined; the text says there are 5 levels, so the notation should be {0, 1, 2, 3, 4}.
  2. [Section 3.5; Eq. (34)] In Eq. (34), 'p+, p+ ∈ Pi' should read 'p+, p− ∈ Pi'.
  3. [Section 4.3] The text says the model is compared to 'six' state-of-the-art methods, but seven baselines are listed (BiMPM, Conv-KNRM, CDAP, SSE-Cross, MCR, DIP, G2SAM). Please correct the count.
  4. [Section 5.5; Abstract] The public dataset is called Lazada-MRHP in Section 5.5 but Lazada-Home in the abstract and Table 11; please unify the name.
  5. [Tables 5-7] The proposed method is labeled 'The current method' in Tables 5-7 but 'MFTRR' elsewhere; please use the model name consistently.
  6. [Section 5.4] The sentence explaining why MFTRR outperforms MFTRR(GCN) attributes the gain to multi-level analysis, but the GCN replacement already uses graph-based reasoning; the attribution should be supported by the ablation rather than asserted.
  7. [General] No data or code availability statement is provided. Since the paper's contributions include three new datasets, a public release or at least a clear availability statement is important for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is an empirical benchmark comparison with labels external to the model's own equations.

full rationale

MFTRR is presented as a supervised ranking model: Eq. 33 maps post/topic features to a scalar quality score, Eq. 34 is a pairwise margin ranking loss defined against annotated quality labels, and Tables 5-7 report MAP and NDCG on test splits held out from training. The abstract's 9.52% NDCG@3 improvement is a comparison between the proposed model and baseline systems on those held-out rankings, not a quantity that is fitted and then renamed as a prediction. The only input-derived element is dataset construction: Section 4.1 states that data were pre-annotated using an external multimodal correlation analysis model and then proofread by five individuals. This bears on label reliability but does not make the evaluation circular, because the supervision and evaluation labels are not generated by MFTRR's own equations and the proofreading step is an independent human check. No load-bearing self-citation or uniqueness theorem is invoked; references [5]-[12], which include the co-author's own work, are contextual safety-motivation citations and are not used to justify the framework's predictions. The paper also benchmarks against public Lazada-Home data in Section 5.5, providing an external check. Suspicious internal inconsistencies, such as identical MFTRR(GCN) rows in Table 10 and split counts in Table 4 that do not match the stated 20% test proportion, are correctness and reporting concerns, not circularity of the derivation chain. Overall, the paper's central empirical claim is self-contained with respect to circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The model's output depends on the reliability of human-annotated scores, the adequacy of pretrained text and image features, and the assumption that filtering to fully multimodal pairs does not bias the evaluation. Hyperparameters (kernel sizes, hidden dimensions, graph layers, batch size, margin) are chosen by hand; none are fitted to the central claim. No new physical entities are introduced.

free parameters (5)
  • n-gram kernel sizes = [1, 3, 5]
    Text CNN kernel sizes chosen by hand in Section 4.2.
  • text hidden dimension = 128
    Dimension of text hidden layer and image feature encoding in Section 4.2.
  • graph inference layers = 2
    Number of graph inference layers in the evidence reasoning module in Section 4.2.
  • batch size = 4
    Training batch size in Section 4.2.
  • ranking margin gamma = 1
    Margin in the pairwise ranking loss, Eq. (34) and Section 4.2.
assumptions (6)
  • domain assumption Post quality is determined by whether a post comprehensively and accurately answers the questions in the discussion topic.
    Stated in Section 3.1; the entire ranking label is defined this way.
  • domain assumption Multimodal (text plus image) information is necessary for accurate post quality assessment; text alone is inadequate.
    Motivated by examples in Figures 1-3 in the Introduction; not empirically tested on a text-only control.
  • domain assumption The manually proofread quality scores are reliable ground truth.
    Annotation process in Section 4.1 reports no inter-annotator agreement, so label reliability is assumed.
  • domain assumption Pre-trained fastText embeddings and CSPDarkNet visual features provide adequate input representations.
    Used in Section 3.2; no comparison with other backbones in the main model.
  • domain assumption Graph-based evidence reasoning can model the subtle topic-post relationships needed for ranking.
    The design of Section 3.4 rests on this; the ablation shows it matters empirically but the mechanism is not validated.
  • ad hoc to paper The dataset construction's filtering to 100% multimodal pairs does not introduce a selection bias that invalidates the evaluation.
    Section 4.1 keeps only topic-post pairs containing both text and images, so the evaluation set is not representative of all forum posts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Fine-grained Reasoning for Post Quality Evaluation." pith.science (2026). https://pith.science/paper/HV7APHC2

@misc{pith2026250717934,
  author       = {Pith},
  title        = {Pith review of: Multimodal Fine-grained Reasoning for Post Quality Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HV7APHC2}},
  note         = {Machine review of arXiv:2507.17934}
}
read the original abstract

Accurately assessing post quality requires complex relational reasoning to capture nuanced topic-post relationships. However, existing studies face three major limitations: (1) treating the task as unimodal categorization, which fails to leverage multimodal cues and fine-grained quality distinctions; (2) introducing noise during deep multimodal fusion, leading to misleading signals; and (3) lacking the ability to capture complex semantic relationships like relevance and comprehensiveness. To address these issues, we propose the Multimodal Fine-grained Topic-post Relational Reasoning (MFTRR) framework, which mimics human cognitive processes. MFTRR reframes post-quality assessment as a ranking task and incorporates multimodal data to better capture quality variations. It consists of two key modules: (1) the Local-Global Semantic Correlation Reasoning Module, which models fine-grained semantic interactions between posts and topics at both local and global levels, enhanced by a maximum information fusion mechanism to suppress noise; and (2) the Multi-Level Evidential Relational Reasoning Module, which explores macro- and micro-level relational cues to strengthen evidence-based reasoning. We evaluate MFTRR on three newly constructed multimodal topic-post datasets and the public Lazada-Home dataset. Experimental results demonstrate that MFTRR significantly outperforms state-of-the-art baselines, achieving up to 9.52% NDCG@3 improvement over the best unimodal method on the Art History dataset.

Figures

Figures reproduced from arXiv: 2507.17934 by the authors.

Figure 1
Figure 1. Multimode post examples under the topic ”Thinking about high-order matrix [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Multimode post examples under the topic ”Buddhist statue”. Using text alone [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Multimode post examples under the topic ”Greece and China”. Analyzing posts [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Model overview of the end-to-end MFTRR method, which consists of two compo [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: The Excellent Course dataset [PITH_FULL_IMAGE:figures/full_fig_p027_5.png]
Figure 6
Figure 6. Figure 6: The Education Course dataset [PITH_FULL_IMAGE:figures/full_fig_p027_6.png]
Figure 7
Figure 7. Figure 7: The Art History Course dataset 25 [PITH_FULL_IMAGE:figures/full_fig_p027_7.png]
Figure 8
Figure 8. Figure 8: The structure of MFTRR(I) 31 [PITH_FULL_IMAGE:figures/full_fig_p033_8.png]
Figure 9
Figure 9. Figure 9: The structure of MFTRR(II) [PITH_FULL_IMAGE:figures/full_fig_p034_9.png]
Figure 10
Figure 10. Figure 10: The structure of MFTRR(III) [PITH_FULL_IMAGE:figures/full_fig_p034_10.png]
Figure 11
Figure 11. Figure 11: The structure of MFTRR(IV) 32 [PITH_FULL_IMAGE:figures/full_fig_p034_11.png]
Figure 12
Figure 12. Figure 12: An example of discussion topic and posts. ( [PITH_FULL_IMAGE:figures/full_fig_p039_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 48 canonical work pages

  1. [1]

    Lalingkar, V

    A. Lalingkar, V. Audichya, P. Mishra, S. Mandyam, S. Srinivasa, Models for finding quality of affirmation and points of intervention in an aca- demic discussion forum, Computers and Education: Artificial Intelligence 3 (2022) 100046

    work page 2022

  2. [2]

    Ntourmas, S

    A. Ntourmas, S. Daskalaki, Y. Dimitriadis, N. Avouris, Classifying mooc forum posts using corpora semantic similarities: a study on transferability across different courses, Neural Computing and Applications (2023) 1–15

    work page 2023

  3. [3]

    X. Tao, A. Shannon-Honson, P. Delaney, C. Dann, H. Xie, Y. Li, S. O’Neill, Towards an understanding of the engagement and emotional behaviour of mooc students using sentiment and semantic features, Com- puters and Education: Artificial Intelligence 4 (2023) 100116

    work page 2023

  4. [4]

    M. A. El-Rashidy, N. A. Khodeir, A. Farouk, H. K. Aslan, N. A. El- Fishawy, Attention-based contextual local and global features for urgent posts classification in moocs discussion forums, Ain Shams Engineering Journal 15 (4) (2024) 102605

    work page 2024

  5. [5]

    M. Liu, S. Liang, K. Howlader, L. Wang, D. Tao, W. Zhang, Natural re- flection backdoor attack on vision language model for autonomous driving, arXiv preprint arXiv:2505.06413 (2025)

    arXiv 2025

  6. [6]

    Liang, K

    S. Liang, K. Liu, J. Gong, J. Liang, Y. Xun, E.-C. Chang, X. Cao, Unlearning backdoor threats: Enhancing backdoor defense in multi- modal contrastive learning via local token unlearning, arXiv preprint arXiv:2403.16257 (2024)

    arXiv 2024

  7. [7]

    Kuang, S

    J. Kuang, S. Liang, J. Liang, K. Liu, X. Cao, Adversarial backdoor defense in clip, arXiv preprint arXiv:2409.15968 (2024)

    arXiv 2024

  8. [8]

    L. Lu, S. Pang, S. Liang, H. Zhu, X. Zeng, A. Liu, Y. Liu, Y. Zhou, Ad- versarial training for multimodal large language models against jailbreak attacks, arXiv preprint arXiv:2503.04833 (2025)

    arXiv 2025

Show all 58 references
  1. [9]

    L. Wang, Z. Ying, T. Zhang, S. Liang, S. Hu, M. Zhang, A. Liu, X. Liu, Manipulating multimodal agents via cross-modal prompt injection, arXiv preprint arXiv:2504.14348 (2025). 40

    2025 arXiv

  2. [10]

    Z. Ying, D. Zhang, Z. Jing, Y. Xiao, Q. Zou, A. Liu, S. Liang, X. Zhang, X. Liu, D. Tao, Reasoning-augmented conversation for multi-turn jail- break attacks on large language models, arXiv preprint arXiv:2502.11054 (2025)

    2025 arXiv

  3. [11]

    Liang, M

    S. Liang, M. Zhu, A. Liu, B. Wu, X. Cao, E.-C. Chang, Badclip: Dual- embedding guided backdoor attack on multimodal contrastive learning, arXiv preprint arXiv:2311.12075 (2023)

    2023 arXiv

  4. [12]

    Liang, J

    S. Liang, J. Liang, T. Pang, C. Du, A. Liu, E.-C. Chang, X. Cao, Revisit- ing backdoor attacks against large vision-language models, arXiv preprint arXiv:2406.18844 (2024)

    2024 arXiv

  5. [13]

    A. Chanaa, et al., Sentiment analysis on massive open online courses (moocs): multi-factor analysis, and machine learning approach, Interna- tional Journal of Information and Communication Technology Education (IJICTE) 18 (1) (2022) 1–22

    2022

  6. [14]

    Almatrafi, A

    O. Almatrafi, A. Johri, H. Rangwala, Needle in a haystack: Identifying learner posts that require urgent response in mooc discussion forums, Computers & Education 118 (2018) 1–9

    2018

  7. [15]

    J. Moon, F. Ke, Z. Sokolikj, S. Chakraborty, Applying multimodal data fusion to track autistic adolescents’ representational flexibility develop- ment during virtual reality-based training, Computers & Education: X Reality 4 (2024) 100063

    2024

  8. [16]

    S. Liu, S. Liu, Z. Liu, X. Peng, Z. Yang, Automated detection of emo- tional and cognitive engagement in mooc discussions to predict learning achievement, Computers & Education 181 (2022) 104461

    2022

  9. [17]

    Munigadiapa, T

    P. Munigadiapa, T. Adilakshmi, Mooc-lstm: The lstm architecture for sentiment analysis on moocs forum posts, in: Computational Intelligence and Data Analytics: Proceedings of ICCIDA 2022, Springer, 2022, pp. 283–293

    2022

  10. [18]

    Ouhaichi, D

    H. Ouhaichi, D. Spikol, B. Vogel, Research trends in multimodal learn- ing analytics: A systematic mapping study, Computers and Education: Artificial Intelligence (2023) 100136. 41

    2023

  11. [19]

    Y. Wang, J. He, D. Wang, Q. Wang, B. Wan, X. Luo, Multimodal trans- former with adaptive modality weighting for multimodal sentiment anal- ysis, Neurocomputing 572 (2024) 127181

    2024

  12. [20]

    G. Zhao, Y. Zhang, J. Chu, A multimodal teacher speech emotion recogni- tion method in the smart classroom, Internet of Things 25 (2024) 101069

    2024

  13. [21]

    M. Li, X. Zhuang, L. Bai, W. Ding, Multimodal graph learning based on 3d haar semi-tight framelet for student engagement prediction, Informa- tion Fusion 105 (2024) 102224

    2024

  14. [22]

    Z. Liu, Q. Tang, F. Ouyang, T. Long, S. Liu, Profiling students’ learning engagement in mooc discussions to identify learning achievement: An automated configurational approach, Computers & Education 219 (2024) 105109

    2024

  15. [23]

    Z. Luo, C. Jingying, W. Guangshuai, L. Mengyi, A three-dimensional model of student interest during learning using multimodal fusion with natural sensing technology, Interactive Learning Environments 30 (6) (2022) 1117–1130

    2022

  16. [24]

    D. Wu, J. Chen, W. Deng, Y. Wei, H. Luo, Y. Wei, The recognition of teacher behavior based on multimodal information fusion, Mathematical Problems in Engineering 2020 (2020) 1–8

    2020

  17. [25]

    S. D. Bhattacharjee, J. S. A. V. Gokaraju, J. Yuan, A. Kalwa, Multi- view knowledge graph for explainable course content recommendation in course discussion posts, in: 2022 26th International Conference on Pattern Recognition (ICPR), IEEE, 2022, pp. 2785–2791

    2022

  18. [26]

    Z. Liu, W. Kong, X. Peng, Z. Yang, S. Liu, S. Liu, C. Wen, Dual- feature-embeddings-based semi-supervised learning for cognitive engage- ment classification in online course discussions, Knowledge-Based Systems 259 (2023) 110053

    2023

  19. [27]

    Chango, R

    W. Chango, R. Cerezo, M. Sanchez-Santillan, R. Azevedo, C. Romero, Improving prediction of students’ performance in intelligent tutoring sys- tems using attribute selection and ensembles of different multimodal data sources, Journal of Computing in Higher Education 33 (2021) 614–634

    2021

  20. [28]

    J. Song, H. Chen, C. Li, K. Xie, Mifm: Multimodal information fusion model for educational exercises, Electronics 12 (18) (2023) 3909. 42

    2023

  21. [29]

    Y. Tian, W. Yang, Q. Liu, Q. Yang, Deep supervised multimodal semantic autoencoder for cross-modal retrieval, Computer Animation and Virtual Worlds 31 (4-5) (2020) e1962

    2020

  22. [30]

    X. Wang, P. Hu, L. Zhen, D. Peng, Drsl: Deep relational similarity learn- ing for cross-modal retrieval, Information Sciences 546 (2021) 298–311

    2021

  23. [31]

    H. Lu, N. Fei, Y. Huo, Y. Gao, Z. Lu, J.-R. Wen, Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval, in: Proceedings of the IEEE/CVF conference on computer Vision and pattern recognition, 2022, pp. 15692–15701

    2022

  24. [32]

    M. Meng, H. Wang, J. Yu, H. Chen, J. Wu, Asymmetric supervised con- sistent and specific hashing for cross-modal retrieval, IEEE Transactions on Image Processing 30 (2020) 986–1000

    2020

  25. [33]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language supervision, in: International con- ference on machine learning, PMLR, 2021, pp. 8748–8763

    2021

  26. [34]

    J. Gu, J. Cai, S. R. Joty, L. Niu, G. Wang, Look, imagine and match: Improving textual-visual cross-modal retrieval with generative models, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7181–7189

    2018

  27. [35]

    H. Abdi, V. Guillemot, A. Eslami, D. Beaton, Canonical correlation anal- ysis. (2018)

    2018

  28. [36]

    K. Luo, C. Zhang, H. Li, X. Jia, C. Chen, Adaptive marginalized se- mantic hashing for unpaired cross-modal retrieval, IEEE Transactions on Multimedia (2023)

    2023

  29. [37]

    L. Li, Z. Shu, Z. Yu, X.-J. Wu, Robust online hashing with label semantic enhancement for cross-modal retrieval, Pattern Recognition 145 (2024) 109972

    2024

  30. [38]

    C. Jia, Y. Yang, Y. Xia, Y.-T. Chen, Z. Parekh, H. Pham, Q. Le, Y.-H. Sung, Z. Li, T. Duerig, Scaling up visual and vision-language representa- tion learning with noisy text supervision, in: International conference on machine learning, PMLR, 2021, pp. 4904–4916. 43

    2021

  31. [39]

    A. Khan, I. Ibrahim, M. I. Uddin, M. Zubair, S. Ahmad, M. D. Al Fir- dausi, M. Zaindin, Machine learning approach for answer detection in discussion forums: an application of big data analytics, Scientific Pro- gramming 2020 (2020) 1–10

    2020

  32. [40]

    Z. Liu, X. Kong, H. Chen, S. Liu, Z. Yang, Mooc-bert: Automatically identifying learner cognitive presence from mooc discussion data, IEEE Transactions on Learning Technologies (2023)

    2023

  33. [41]

    Capuano, S

    N. Capuano, S. Caball´ e, J. Conesa, A. Greco, Attention-based hierarchi- cal recurrent neural networks for mooc forum posts analysis, Journal of Ambient Intelligence and Humanized Computing 12 (2021) 9977–9989

    2021

  34. [42]

    Zhang, X

    Z. Zhang, X. Zhu, Q. He, L. Zhang, Bert-based global semantic refinement and local semantic extraction for distinguishing urgent posts in mooc forums, IEEE Access (2024)

    2024

  35. [43]

    A. F. Wise, Y. Cui, W. Jin, J. Vytasek, Mining for gold: Identifying content-related mooc discussion threads across domains through linguistic modeling, The Internet and Higher Education 32 (2017) 11–28

    2017

  36. [44]

    W. Zou, X. Hu, Z. Pan, C. Li, Y. Cai, M. Liu, Exploring the relationship between social presence and learners’ prestige in mooc discussion forums using automated content analysis and social network analysis, Computers in Human Behavior 115 (2021) 106582

    2021

  37. [45]

    J. Jin, Y. Dong, B. He, N. Zhou, L. Yan, Sentiment classification of mooc courses by merging local context focus and bi-directional gated recurrent unit, in: Fourth International Conference on Computer Vision, Applica- tion, and Algorithm (CV AA 2024), Vol. 13486, SPIE, 2025, ...

    2024

  38. [46]

    Z. Dai, C. Xiong, J. Callan, Z. Liu, Convolutional neural networks for soft-matching n-grams in ad-hoc search, in: Proceedings of the eleventh ACM international conference on web search and data mining, 2018, pp. 126–134

    2018

  39. [47]

    Wang, H.-Y

    C.-Y. Wang, H.-Y. M. Liao, Y.-H. Wu, P.-Y. Chen, J.-W. Hsieh, I.-H. Yeh, Cspnet: A new backbone that can enhance learning capability of cnn, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 390–391. 44

    2020

  40. [48]

    Kumar, A

    S. Kumar, A. Kulkarni, M. S. Akhtar, T. Chakraborty, When did you be- come so smart, oh wise one?! sarcasm explanation in multi-modal multi- party dialogues, arXiv preprint arXiv:2203.06419 (2022)

    2022 arXiv

  41. [49]

    M. Fan, C. Feng, L. Guo, M. Sun, P. Li, Product-aware helpfulness pre- diction of online reviews, in: The world wide web conference, 2019, pp. 2715–2721

    2019

  42. [50]

    Z. Wang, W. Hamza, R. Florian, Bilateral multi-perspective matching for natural language sentences, arXiv preprint arXiv:1702.03814 (2017)

    2017 arXiv

  43. [51]

    N. Capuano, Transfer learning techniques for cross-domain analysis of posts in massive educational forums, in: Intelligent Systems and Learning Data Analytics in Online Education, Elsevier, 2021, pp. 133–152

    2021

  44. [52]

    Abavisani, L

    M. Abavisani, L. Wu, S. Hu, J. Tetreault, A. Jaimes, Multimodal catego- rization of crisis events in social media, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 14679–14689

    2020

  45. [53]

    J. Liu, Z. Hai, M. Yang, L. Bing, Multi-perspective coherent reasoning for helpfulness prediction of multimodal reviews, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language ...

    2021

  46. [54]

    C. Wen, G. Jia, J. Yang, Dip: Dual incongruity perceiving network for sarcasm detection, in: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2023, pp. 2540–2550

    2023

  47. [55]

    Y. Wei, S. Yuan, H. Zhou, L. Wang, Z. Yan, R. Yang, M. Chen, Gˆ 2sam: Graph-based global semantic awareness method for multimodal sarcasm detection, in: Proceedings of the AAAI Conference on Artificial Intelligence, 2024, pp. 9151–9159

    2024

  48. [56]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  49. [57]

    Y. Cui, W. Che, T. Liu, B. Qin, Z. Yang, Pre-training with whole word masking for chinese bert, IEEE/ACM Transactions on Audio, Speech, and Language Processing 29 (2021) 3504–3514. 45

    2021

  50. [58]

    Liang, C

    B. Liang, C. Lou, X. Li, M. Yang, L. Gui, Y. He, W. Pei, R. Xu, Multi- modal sarcasm detection via cross-modal graph convolutional network, in: Proceedings of the 60th Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), Vol. 1, Associatio...

    2022

Pith tools

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