Pith. sign in

REVIEW 4 major objections 6 minor 33 references

Leveraging Taxonomy and LLMs for Improved Multimodal Hierarchical Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a plug-in taxonomy-aware output layer for multimodal large language models that classifies at every level of a product taxonomy at once, and reports that it improves prediction consistency and leaf-level accuracy for…

desk verdict TTC is a soft gating layer, not the taxonomy-enforcing constraint the paper claims, but the broad eight-backbone comparison makes it worth a serious referee. read the letter →

arxiv 2501.06827 v1 pith:IFGKDBC3 submitted 2025-01-12 cs.AI

Shijing Chen , Mohamed Reda Bouadjenek , Shoaib Jameel , Usman Naseem , Basem Suleiman , Flora D. Salim , Hakim Hacid , Imran Razzak This is my paper · ORCID
classification cs.AI
keywords hierarchicalclassificationmultimodallearninglargelanguagemodelstaxonomyembeddingtransitionmatrixe-commerceproductparameter-efficientfine-tuningpredictionconsistency
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 proposes a taxonomy-based transitional classifier (TTC), a plug-in output layer for multimodal large language models that classifies at every level of a product taxonomy at once. The authors claim that, on the MEP-3M food subset, adding TTC to a frozen low-rank-adapted backbone improves prediction consistency with the taxonomy and raises leaf-level (level-3) accuracy for most of the eight backbones tested, with a small drop in hierarchical F1 for some models. The intended payoff is a model-agnostic way to stop hierarchical predictors from returning labels that contradict their own parent classes, while also using coarse-level confidence to guide fine-grained decisions.

What carries the argument

The central object is the transition-matrix attention score $m[\ell_{i+1}] = \hat{y}[\ell_i] \times M[\ell_i,\ell_{i+1}]$, where $\hat{y}[\ell_i]$ is the temperature-scaled softmax output of the classifier for level $\ell_i$ and $M[\ell_i,\ell_{i+1}]$ is a binary matrix with 1 where the lower-level class is a subclass of the upper-level class. This score is multiplied element-wise into the next level's logits before softmax, so it acts as a soft, confidence-weighted mask that nudges child predictions toward the branch favoured by the parent level. The loss is a weighted sum of cross-entropy terms at each level with tunable importance factors $\pi[\ell_i]$. The backbone is first fine-tuned with low-rank adaptation and then frozen while TTC's linear heads are trained.

What would settle it

Permute the level-2 training labels so the parent head is unreliable, retrain only TTC's level-3 head, and check whether level-3 accuracy still exceeds the flat baseline; if it does, the top-down attention is not doing the work, and if it drops, the coarse-to-fine steering is the active mechanism.

Watch

Extended reading notes

Core claim

The central claim is that multiplying a lower level's logits by a top-down attention vector derived from the upper level's class probabilities and the taxonomy's transition matrix yields more coherent and often more accurate predictions than training independent flat heads. TTC attaches one linear head per hierarchy level to the frozen joint features of a multimodal LLM, then propagates information downward: the probability output at level i is multiplied by the binary transition matrix M to produce an attention score, and that score is multiplied element-wise into the next level's logits before a temperature-scaled softmax. Across the eight multimodal backbones tested on a 177,195-sample food subset of MEP-3M, TTC improves Consistency and level-3 Accuracy for most models, improves Exact Match in several cases, and trades off a little Hierarchical F1 in some cases. The discovery is that a frozen backbone's representation is enough for the hierarchical steering to help, provided the fine-tuned feature is linearly separable at each level.

Load-bearing premise

The method relies on the coarse-level classifier being trustworthy, because a wrong parent prediction steers all finer predictions into the wrong branch of the taxonomy.

Editorial extensions

If this is right

  • TTC is a drop-in output layer: any frozen multimodal LLM can be fitted with it, so hierarchical labels become usable without retraining the backbone.
  • For most evaluated backbones, leaf-level accuracy and taxonomy consistency improve, which matters when the taxonomy is deep and the final class is the one shown to customers.
  • A small Hierarchical F1 decrease can appear for some backbones, so applying TTC should be decided per backbone depending on whether coherence or global hierarchical F1 is the priority.
  • Because the loss is a weighted sum of per-level cross-entropies, the importance factors give an operator a single knob to shift emphasis between coarse and fine levels.
  • The observed positive correlation between Consistency and level-3 Accuracy is the paper's evidence that the hierarchical layer can be extended to flat tasks by constructing bottom-up label hierarchies and classifying top-down.

Reading between the lines

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

  • Editorial inference: the soft attention mask is not a hard constraint, so 'consistency' is a tendency, not a guarantee; making the transition matrix a hard zero-mask would be a direct test of how much of the gain comes from steering versus outright restriction.
  • Editorial inference: TTC's dependence on frozen features suggests the gains should transfer to any strong frozen encoder with a tree label set, not only LLMs; the paper's 'LLM-agnostic' framing is really 'encoder-agnostic'.
  • Editorial inference: the method is most likely to help when coarse classes are easy to separate and fine classes are hard, a regime the MEP-3M food subset exemplifies; on taxonomies with hard coarse levels, the top-down steering could amplify errors.
  • Editorial inference: a natural stress test is to move from a three-level tree with one root to a deeper or DAG-structured taxonomy, where the transition matrix grows and the sequential top-down pass becomes slower, a bottleneck the paper's own limitations section acknowledges.
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 / 6 minor

Summary. The paper introduces a Taxonomy-based Transitional Classifier (TTC), a model-agnostic output layer for multimodal hierarchical classification. TTC attaches independent linear classifiers at each hierarchy level to the frozen features of a multimodal LLM, and propagates parent-level predictions to child levels through a binary transition matrix that encodes the class taxonomy. The method is evaluated on a 177,195-instance food subset of MEP-3M with 1 root, 4 level-2 classes, and 40 level-3 classes, using eight multimodal LLMs as backbones. The reported results compare each backbone with and without TTC across hierarchical F1, exact match, consistency, and per-level accuracy, and the paper claims that TTC improves consistency and fine-grained accuracy.

Significance. If the mechanism worked as claimed, a cheap taxonomy-gating layer placed on frozen backbone features would be a practical and useful contribution for e-commerce and other hierarchical classification tasks. The paper has several strengths: it evaluates a broad range of backbones, the architecture is simple and clearly described, the transition matrix is derived from the external class hierarchy rather than from fitted values, and the limitations paragraph is candid about the top-down-only design and annotation cost. However, the current evidence does not support the central claim that TTC enforces taxonomic consistency, and the evaluation lacks comparisons to existing hierarchical classifiers and statistical grounding. The usefulness of the idea therefore remains plausible but unproven.

major comments (4)
  1. [§3.2, Eqs. (1)–(2)] The multiplicative mask in Eq. (1) does not enforce the taxonomic consistency claimed in the abstract and §3.2. When the transition matrix entry M[ℓi,ℓi+1] is 0 for a child, the element-wise product in Eq. (2) sets that child's effective logit to 0; softmax then assigns it probability exp(0)/Z > 0, so predictions can still violate the hierarchy. The Consistency scores in Table 2 (0.63–0.85 after TTC) confirm that violations remain. Moreover, because the mask is multiplicative, a parent-level error does not zero out the true child branch; it merely down-weights it, and an incorrect child of a wrongly predicted parent can dominate. To make the stated claim true, the mechanism must be changed to a hard mask (e.g., setting disallowed logits to -∞ conditional on the predicted parent) or the paper must be revised to describe soft gating that encourages, but does not guarantee, consistency.
  2. [§4.2, Table 2] The evaluation compares each backbone only against its own flat version. It does not compare TTC against previously proposed hierarchical classifiers, even though several are cited in §2, including C-HMCNN, HiAGM, and the mask-based output layer of Boone-Sifuentes et al. (2022a). The abstract's claim of a 'significant performance improvement compared to conventional LLM structures' is therefore not sufficient to establish that TTC advances the state of the art in hierarchical classification. The authors should add at least one or two representative hierarchical baselines, ideally trained on the same frozen features, to support the paper's central positioning.
  3. [§4.1, data sampling] The data sampling description is internally inconsistent. The paper states that the subset has 4 level-2 classes and 40 level-3 classes, yet it says 'we sampled 1,000 data entries from each ℓ2 label (40,000 in total)'. Four level-2 labels would give 4,000 samples, while 40,000 corresponds to 1,000 samples per level-3 class. This discrepancy affects the reproducibility of the fine-tuning protocol and should be corrected by specifying exactly which level was used for balanced sampling.
  4. [§4.2, Table 2] All experimental results are single-run point estimates with no variance, no multiple seeds, and no statistical significance testing. Many of the reported differences are small (e.g., Fuyu Accuracy@ℓ3 changes from 0.6803 to 0.6902; MiniGPT-4 HF1-Score changes from 0.8649 to 0.8652), so the repeated use of the word 'significant' is not empirically supported. I recommend reporting mean and standard deviation over at least three independent runs and, if feasible, paired significance tests for the key metrics.
minor comments (6)
  1. [Abstract, §1] There are several typos and grammatical issues, including 'an data entry' and 'n independent output layers' where 'n' is an artifact of the formatting; the paper would benefit from a careful copyedit.
  2. [§4.2, Figure 6] The correlation between Consistency and Accuracy@ℓ3 in Figure 6 is used in §5 to suggest that the TTC layer can be extended to traditional classification tasks, but with only 16 points across two settings the correlation is descriptive, not causal. I suggest presenting this as an observation rather than as evidence for the mechanism's transferability.
  3. [§4.1, Table 1] The hyperparameter entry 'Number of Epochs (3 tolerance on train acc)' is unclear; please specify the exact number of epochs or the early-stopping criterion precisely.
  4. [§3.1] The taxonomy definition states ℓ1 ∪ ℓ2 ∪ ⋯ ∪ ℓn = Y, but it is unclear whether classes at different levels are disjoint or how the union is intended when the same label name appears at more than one level; clarifying this would help the reader.
  5. [§4.1 and throughout] The dataset is referred to inconsistently as both MEP-3 and MEP-3M; please use one consistent name and verify the citation.
  6. [Appendix A, Figure 7] The heatmap labels use abbreviations such as 'LLAVA-1.5' while the main text uses 'LLaVA-1.5'; unify notation and define all acronyms in the figure caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the taxonomy mask is an external input and the headline metrics are held-out measurements, so the TTC derivation does not reduce to its own inputs.

full rationale

The paper's claimed derivation chain is self-contained. In Section 3.2, Eq. (1) defines the transition mask as m[ℓi+1] = yhat[ℓi] × M[ℓi,ℓi+1], where M is a fixed binary matrix encoding the external class taxonomy; Eq. (2) then applies this mask to the next-level logits. The taxonomy matrix is an input to the method, not a quantity fitted from the data, and no parameter is later renamed as a prediction. The main empirical claims (HF1-Score, Exact Match, Consistency, Accuracy@ℓ2, Accuracy@ℓ3 in Table 2 and Figures 4-5) are measured on a held-out 20% test split against ground-truth labels, so the results are not forced by construction. The self-citations in the paper (Boone-Sifuentes et al., 2022a/b; Boone Sifuentes et al., 2024) appear in the introduction and related work for motivation and do not carry the derivation; no uniqueness theorem from the authors is invoked to rule out alternatives. The Limitations section itself notes that the method only uses top-down transitions, which further confirms that the authors do not claim a derivation from first principles. Two evidence-quality issues are worth noting but they are not circularity: the Consistency vs ℓ3 Accuracy correlation in Figure 6 is a post-hoc correlation, and Equations (1)-(2) do not strictly force consistency because softmax assigns positive probability even to masked logits; however, neither issue makes the evaluation equivalent to the model's fitted inputs. The omission of a direct comparison to the same group's earlier mask-based output layer (Boone-Sifuentes et al., 2022a) is an attribution/novelty concern, not a circular dependency. Therefore the central derivation is independent of its inputs and the honest finding is no circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method introduces no new entities. It depends on a supplied tree taxonomy and learned per-level linear heads on frozen multimodal features. Two hyperparameters (π, τ) are mentioned but never specified, which adds unquantified degrees of freedom. The main untested assumption is that frozen fine-tuned features remain linearly separable for both levels, and that the single-parent tree assumption holds for MEP-3M.

free parameters (4)
  • W[ℓi], b[ℓi] per-level classifier weights = learned from training set
    Two linear heads on the frozen backbone feature a, one per hierarchy level; standard supervised parameters.
  • π[ℓi] loss importance factors = not reported (assumed 1)
    Equation 3 weights each level's cross-entropy; the paper says they 'can be tuned' but never reports values or an ablation.
  • Softmax temperature τ = not reported
    Used in softmax_τ in Equation 2; temperature affects mask sharpness but its value is not listed in Table 1.
  • LoRA rank / alpha / dropout = 8 / 16 / 0.1
    Backbone fine-tuning hyperparameters from Table 1; not the core claim but part of the pipeline.
assumptions (3)
  • domain assumption The taxonomy is a tree where every class at level ℓ_{i+1} has exactly one parent at level ℓ_i.
    Stated in Section 3.1; the transition matrix M and the consistency guarantees depend on single-parent trees.
  • domain assumption The frozen backbone feature a, after LoRA fine-tuning, is linearly separable at both ℓ2 and ℓ3 with the same representation.
    TTC classifiers are linear; if this fails, the method caps at the quality of the frozen features (Section 4.1).
  • domain assumption Higher-level classes are easier to classify than lower-level classes and therefore top-down guidance helps deeper levels.
    Motivated in Section 3.2 with a citation to Cortes 1995; used to justify why upper-level mistakes should bias lower-level predictions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Taxonomy and LLMs for Improved Multimodal Hierarchical Classification." pith.science (2026). https://pith.science/paper/IFGKDBC3

@misc{pith2026250106827,
  author       = {Pith},
  title        = {Pith review of: Leveraging Taxonomy and LLMs for Improved Multimodal Hierarchical Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IFGKDBC3}},
  note         = {Machine review of arXiv:2501.06827}
}
read the original abstract

Multi-level Hierarchical Classification (MLHC) tackles the challenge of categorizing items within a complex, multi-layered class structure. However, traditional MLHC classifiers often rely on a backbone model with independent output layers, which tend to ignore the hierarchical relationships between classes. This oversight can lead to inconsistent predictions that violate the underlying taxonomy. Leveraging Large Language Models (LLMs), we propose a novel taxonomy-embedded transitional LLM-agnostic framework for multimodality classification. The cornerstone of this advancement is the ability of models to enforce consistency across hierarchical levels. Our evaluations on the MEP-3M dataset - a multi-modal e-commerce product dataset with various hierarchical levels - demonstrated a significant performance improvement compared to conventional LLM structures.

Figures

Figures reproduced from arXiv: 2501.06827 by the authors.

Figure 1
Figure 1. (a) An data point of an “Apple” classified [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. illustrates the architecture of the pro￾posed TTC layer, an LLM-agnostic component de￾signed to leverage the taxonomy and ensure that predictions adhere to the hierarchical structure of Taxonomy-based transitional classifier (TTC) MultiModal BackBone LLM Flatten a Dense 1 W[ℓ1],b[ℓ1] Dense 2 W[ℓ2],b[ℓ2] Dense 3 W[ℓ3],b[ℓ3] Dense n W[ℓn],b[ℓn] … Softmaxτ Mask layer M[ℓ1, ℓ2] Element-wise product ⊙ Softmaxτ Element-wi… view at source ↗
Figure 3
Figure 3. The distribution of data across all classes at [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Spider diagram for all LLM backbones regarding 5 different metrics. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Group bar chart for detailed results for all LLM backbones on all 5 metrics. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Relationship Between Consistency and ℓ3 Ac￾curacy for Various Models with and without the TTC Layer. Orange (flat classifier), Blue (hierarchical classi￾fier). and-conquer approach has the potential to enhance performance in a wide range of classification prob￾lems. 5 …
Figure 7
Figure 7. Figure 7: The heatmap of improvements that models with TTC have over models without TTC [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 20 canonical work pages

  1. [1]

    Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. 2023. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390

    arXiv 2023

  2. [2]

    Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sağnak Taşırlar. 2023. https://www.adept.ai/blog/fuyu-8b/ Fuyu-8b: A multimodal architecture for ai agents

    work page 2023

  3. [3]

    Zineddine Bettouche, Anas Safi, and Andreas Fischer. 2024. Contextual categorization enhancement through llms latent-space. arXiv preprint arXiv:2404.16442

    work page Pith review arXiv 2024

  4. [4]

    Tanya Boone-Sifuentes, Mohamed Reda Bouadjenek, Imran Razzak, Hakim Hacid, and Asef Nazari. 2022 a . A mask-based output layer for multi-level hierarchical classification. In Proceedings of the 31st ACM international conference on information & knowledge management, pages 3833--3837

    work page 2022

  5. [5]

    Tanya Boone Sifuentes, Asef Nazari, Mohamed Reda Bouadjenek, and Imran Razzak. 2024. Mlt-trans: Multi-level token transformer for hierarchical image classification. In Pacific-Asia Conference on Knowledge Discovery and Data Mining, pages 385--396. Springer

    work page 2024

  6. [6]

    Tanya Boone-Sifuentes, Asef Nazari, Imran Razzak, Mohamed Reda Bouadjenek, Antonio Robles-Kelly, Daniel Ierodiaconou, and Elizabeth S Oh. 2022 b . Marine-tree: A large-scale marine organisms dataset for hierarchical image classification. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 3838--3842

    work page 2022

  7. [7]

    Corinna Cortes. 1995. Support-vector networks. Machine Learning

    1995

  8. [8]

    Eduardo P Costa, Ana C Lorena, Andr \'e CPLF Carvalho, Alex A Freitas, and Nicholas Holden. 2007. Comparing several approaches for hierarchical classification of proteins with decision trees. In Advances in Bioinformatics and Computational Biology: Second Brazilian Symposium on Bioinformatics, BSB 2007, Angra dos Reis, Brazil, August 29-31, 2007. Proceedi...

    work page 2007

Show all 33 references
  1. [9]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. 2024. Instructblip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36

    2024

  2. [10]

    Andr \'e CPLF de Carvalho and Alex A Freitas. 2009. A tutorial on multi-label classification techniques. Foundations of Computational Intelligence Volume 5: Function Approximation and Classification, pages 177--195

    2009

  3. [11]

    Susan Dumais and Hao Chen. 2000. Hierarchical classification of web content. In Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 256--263

    2000

  4. [12]

    Eleonora Giunchiglia and Thomas Lukasiewicz. 2020. Coherent hierarchical multi-label classification networks. Advances in neural information processing systems, 33:9662--9673

    2020

  5. [13]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  6. [14]

    Svetlana Kiritchenko, Stan Matwin, A Fazel Famili, et al. 2005. Functional annotation of genes using hierarchical text categorization. In Proc. of the ACL Workshop on Linking Biological Literature, Ontologies and Databases: Mining Biological Semantics

    2005

  7. [15]

    Daphne Koller and Mehran Sahami. 1997. Hierarchically classifying documents using very few words. Technical report, Stanford InfoLab

    1997

  8. [16]

    Aris Kosmopoulos, Ioannis Partalas, Eric Gaussier, Georgios Paliouras, and Ion Androutsopoulos. 2015. Evaluation measures for hierarchical classification: a unified view and novel approaches. Data Mining and Knowledge Discovery, 29:820--865

    2015

  9. [17]

    Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. 2019. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557

    2019 arXiv

  10. [18]

    Fan Liu, Delong Chen, Xiaoyu Du, Ruizhuo Gao, and Feng Xu. 2023 a . Mep-3m: A large-scale multi-modal e-commerce product dataset. Pattern Recognition, 140:109519

    2023

  11. [19]

    Hankai Liu, Xianying Huang, and Xiaoyang Liu. 2024 a . Improve label embedding quality through global sensitive gat for hierarchical text classification. Expert Systems with Applications, 238:122267

    2024

  12. [20]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024 b . Visual instruction tuning. Advances in neural information processing systems, 36

    2024

  13. [21]

    Ye Liu, Kai Zhang, Zhenya Huang, Kehang Wang, Yanghai Zhang, Qi Liu, and Enhong Chen. 2023 b . Enhancing hierarchical text classification through knowledge graph integration. In Findings of the Association for Computational Linguistics: ACL 2023, pages 5797--5810

    2023

  14. [22]

    Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. 2022. Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft

    2022

  15. [23]

    Andrew D Secker, Matthew N Davies, Alex A Freitas, Jon Timmis, Miguel Mendao, and Darren R Flower. 2007. An experimental comparison of classification algorithms for hierarchical prediction of protein function. Expert Update (Magazine of the British Computer Society's Specialis...

    2007

  16. [24]

    Carlos N Silla and Alex A Freitas. 2011. A survey of hierarchical classification across different application domains. Data mining and knowledge discovery, 22:31--72

    2011

  17. [25]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  18. [26]

    Giorgio Valentini. 2010. True path rule hierarchical ensembles for genome-wide gene function prediction. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 8(3):832--847

    2010

  19. [27]

    Celine Vens, Jan Struyf, Leander Schietgat, Sa s o D z eroski, and Hendrik Blockeel. 2008. Decision trees for hierarchical multi-label classification. Machine learning, 73:185--214

    2008

  20. [28]

    Feihong Wu, Jun Zhang, and Vasant Honavar. 2005. Learning classifiers using hierarchically structured class taxonomies. In Abstraction, Reformulation and Approximation: 6th International Symposium, SARA 2005, Airth Castle, Scotland, UK, July 26-29, 2005. Proceedings 6, pages 3...

    2005

  21. [29]

    Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. 2023. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178

    2023 arXiv

  22. [30]

    Yunyi Zhang, Ruozhen Yang, Xueqiang Xu, Jinfeng Xiao, Jiaming Shen, and Jiawei Han. 2024. Teleclass: Taxonomy enrichment and llm-enhanced hierarchical text classification with minimal supervision. arXiv preprint arXiv:2403.00165

    2024 arXiv

  23. [31]

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592

    2023 arXiv

  24. [32]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  25. [33]

    write newline

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

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.