Pith. sign in

REVIEW 4 major objections 5 minor 47 references

A Transformer-based Autoregressive Decoder Architecture for Hierarchical Text Classification

T0 review · 4 major / 5 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 RADAr matches state-of-the-art hierarchical text classification without graph encoders or label semantics, using a RoBERTa encoder and an autoregressive decoder.

desk verdict A solid empirical HTC paper whose main claim—no graph encoder or label semantics needed if you order symbolic labels child-to-parent—is plausible and worth reviewing, but the baseline comparability needs tightening. read the letter →

arxiv 2501.13598 v1 pith:DUJEY4VH submitted 2025-01-23 cs.LG

Younes Yousef , Lukas Galke , Ansgar Scherp This is my paper · ORCID
classification cs.LG
keywords hierarchicaltextclassificationautoregressivedecodersymboliclabelslabelhierarchysemanticsRoBERTafocallosssequencegeneration
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

RADAr is a hierarchical text classifier that treats label prediction as sequence generation: a RoBERTa encoder reads the text, and a two-layer transformer decoder emits the document's labels one token at a time using only symbolic label identifiers. The paper's central claim is that this simple setup is competitive with state-of-the-art HTC models that explicitly encode the label hierarchy with graph networks and exploit label-name semantics. On WOS, NYT, and RCV1-V2, RADAr lands within 0 to 1.55 F1 points of HBGL, the strongest baseline, while cutting inference time roughly in half. The authors argue this shows neither label semantics nor a graph encoder is necessary, and that ordering labels from children to parents is consistently better than parents to children.

What carries the argument

The load-bearing mechanism is the autoregressive decoder over a vocabulary of symbolic labels, where each hierarchy level is marked by the `<unk>` token and sequences are ordered from the most specific child labels up to the root. This ordering turns hierarchical classification into a constrained generation task: the decoder conditions each next label on the encoder's text representation and on the labels already generated, so the hierarchy is captured as sequential dependency rather than as a graph. The decoder has two layers with eight attention heads, and a linear head produces logits with batch-level focal loss scaling $(1-e^{-\text{loss}})^\gamma$ that up-weights hard samples. This component is what makes label semantics and graph encoders dispensable.

What would settle it

Re-run RADAr, HBGL, and HGCLR under one identical protocol, using the same train/dev/test splits, tokenizer truncation length, label ordering, random seeds, and micro/macro averaging, and check whether HBGL's margin on every dataset stays within the reported 0 to 1.55 points. A consistently larger margin on any dataset would falsify the claim that RADAr is competitive without graph encoders or label semantics.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that the label hierarchy can be learned implicitly from the order of symbolic labels in training sequences. RADAr replaces graph encoders and label-name embeddings with an autoregressive decoder that generates child-to-parent label sequences separated by level markers, trained with batch-level focal loss ($\gamma=2$). The reported scores (Micro-F1 87.17/79.84/87.23 and Macro-F1 81.84/68.64/69.64 on WOS, NYT, and RCV1-V2) are within 0 to 1.55 points of HBGL, and ablations show child-to-parent ordering, level separators, and organized sequences each contribute. The paper concludes that HTC's usual ingredients, label semantics and explicit graph encoding, are not required for competitive performance.

Load-bearing premise

The competitive claim rests on comparing RADAr's numbers with baseline scores taken from different papers and from runs whose protocol is not fully specified; if those baselines were tuned under different evaluation settings, input truncations, or splitting conventions, the reported parity could shift.

Editorial extensions

If this is right

  • If RADAr's claim holds, new HTC deployments can skip label-name embeddings and graph encoders, reducing model requirements and simplifying training pipelines.
  • Inference throughput roughly doubles relative to HBGL, making the architecture attractive for very large test sets such as RCV1-V2's 781,265 documents.
  • Child-to-parent label order should become the default ordering in sequence-generation HTC models, since it consistently outperforms parent-to-child order across all three datasets.
  • Because the decoder is trained from scratch on symbolic labels, the encoder can be swapped for newer text encoders without redesigning the decoder.
  • The ablation showing degraded performance on shuffled label sequences implies that the model genuinely uses order as a structural signal, not just as a fixed permutation.

Reading between the lines

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

  • Beyond the paper, the child-to-parent result suggests a general design principle for label-sequence models: place the most informative, fine-grained label first and let coarser ancestors be inferred, which may transfer to other taxonomy-constrained generation tasks.
  • The strong effect of `<unk>` level separators implies the decoder uses level boundaries as structural cues; a testable extension is whether learned level embeddings or level-specific positional encodings could make the separator unnecessary.
  • The reported speed-up compares wall-clock inference against one baseline; a fairer cross-model comparison would require latency measurements under identical batching and hardware, ideally with variance across seeds.
  • The exposure-bias error pattern on WOS suggests that training-time correction methods, such as scheduled sampling, could improve child-label accuracy even though greedy decoding already matches the baseline.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces RADAr, an encoder-decoder model for hierarchical text classification (HTC) that uses an off-the-shelf RoBERTa encoder and a custom two-layer autoregressive decoder. The decoder operates on symbolic label identifiers rather than label-name text, and labels are presented in child-to-parent order with a level separator token. Training uses a batch-level focal loss. On WOS, NYT, and RCV1-V2, RADAr is reported to be competitive with the HBGL state-of-the-art system while using no graph encoder and no label semantics, and to cut inference time roughly in half. The paper includes an ablation study on label ordering, separator tokens, focal loss, label semantics, and encoder choice, plus an error analysis.

Significance. If the claims hold, the paper has a useful practical message: competitive HTC performance can be achieved without graph encoders or label-name embeddings, using a simple autoregressive decoder over symbolic labels. The paper is also honest in reporting that RADAr trails HBGL on most metrics by small margins, and it reports five-run means with standard deviations for the main results. The release of source code and the explicit description of hyperparameter tuning on validation sets are strengths. The significance is tempered by the fact that the main comparative claim rests on baseline numbers whose protocols are not fully documented, and by the absence of multiple-seed variability in the ablation study.

major comments (4)
  1. [Table 3 and Section 5] The central 'competitive to the state of the art' claim depends on comparing RADAr with HBGL entries whose provenance is heterogeneous. The HBGL+BERT row is imported from [14], while the HBGL+RoBERTa row is labeled 'our experiment' with no protocol listing (e.g., truncation length, dev split, training budget, or number of runs). The margins are small: RADAr is below HBGL+BERT by 0.19, 0.16, 0.63, 1.55, 0.0, and 1.43 points across the six reported metrics. Since the conclusion that 'neither label semantics nor an explicit graph encoder is needed' is essentially a parity claim, the authors should document the exact evaluation protocol for the HBGL+RoBERTa runs and, where possible, re-run the literature baselines under the same conditions, or at least report which baseline settings were used.
  2. [Section 6.1, Table 4] The ablation study is reported without standard deviations or any indication of the number of runs, and the text in Table 4 gives point estimates (e.g., 87.19/81.84/69.09/87.22/69.48 for the RADAr row) that differ slightly from the five-run means in Table 3. Because the claims about child-to-parent ordering, the role of the <unk> separator, and the ineffectiveness of label semantics all rest on these ablation comparisons, the authors should report multiple-seed mean and variance for at least the key ablation rows, or clearly state which rows are single runs and why those single runs are sufficient.
  3. [Section 3 and Section 6.1] The paper's title claim is that a graph encoder is not needed, but the experiments do not include a controlled variant of RADAr with a graph encoder added to the same encoder-decoder architecture. The only controlled component comparison is 'with labels semantics', which is compared against the no-semantics RADAr. The 'graph encoder is dispensable' conclusion is therefore inferred from comparisons with different architectures (HBGL, HGCLR, HiMatch) rather than from an ablation. The authors should either add a graph-encoder variant of RADAr or explicitly delimit the claim to say that the compared systems perform comparably without a graph encoder.
  4. [Section 6] The inference-time speed-up claim is based on separate timings of HBGL and RADAr, but the text does not specify the hardware, batch sizes, implementation details, or whether both models were timed under identical conditions. Since the abstract and conclusions emphasize a factor-of-two inference speed-up, the authors should provide the measurement protocol and, ideally, report timings from the same environment with a small number of repetitions to rule out implementation-induced differences.
minor comments (5)
  1. [Title page] The affiliation 'Max Planck Institute for Psycholiguistics' appears to contain a typo; it should likely be 'Psycholinguistics'.
  2. [Section 4] The statement that 'the tokenizer uses the <unk> token to separate hierarchy levels' is clear, but the choice of reusing <unk> (rather than a dedicated separator token) is unusual and could be clarified: <unk> is normally an unknown-token placeholder, and its use as a structural separator may confuse readers.
  3. [Table 4] The first row of Table 4 shows scores that are not identical to the RADAr row in Table 3 (e.g., WOS Micro-F1 87.19 vs 87.17). The difference is presumably due to Table 4 being a single run while Table 3 reports averages; this should be stated explicitly.
  4. [Section 6.2] The error analysis is informative, but the statement 'the model always predicts the parent label correctly' for WOS is stronger than the preceding numbers warrant; the text says 852 cases had correct parent with wrong child and 778 had both wrong, so 'always' appears to be an overstatement that should be rephrased.
  5. [Limitations] The Limitations section is brief but appropriately notes that the model is English-only; it could also mention the restriction to tree-structured hierarchies, since RCV1-V2 and NYT are single-parent hierarchies, and the child-to-parent ordering assumption may not transfer to DAG-structured label sets.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RADAr's claims are empirical, held-out comparisons with validation-based hyperparameter tuning, not derivations that reduce to their inputs.

full rationale

The paper's central claim is that an off-the-shelf RoBERTa encoder plus a two-layer autoregressive decoder over symbolic labels, without label semantics or a graph encoder, achieves competitive hierarchical text classification. This claim is established by held-out test-set comparisons on three benchmarks, with RADAr results reported as means over five seeds and hyperparameters selected on validation sets via random search and grid search. There is no fitted parameter that is then relabeled as a prediction, and no equation in the paper defines the model's output in terms of the benchmark numbers. The ablations are genuine comparisons of alternative design choices (parent-to-child order, no level separator, shuffled labels, label-level focal loss, label-semantics initialization, alternative encoders), and the results show the reported configuration is not forced by construction; for example, label semantics actually slightly hurt performance. The few self-citations, such as [10] for baseline RoBERTa results and learning-rate ranges, are not load-bearing for the main conclusion, because the competitive claim rests on externally sourced HBGL numbers from [14] and [37] plus the paper's own runs. Concerns about whether literature baselines were tuned under identical protocols are a measurement and reporting risk, not circularity: nothing in the paper's derivation reduces to its own inputs, and no self-citation chain is invoked to forbid alternative interpretations. The paper is self-contained as an empirical architecture study, so the honest finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The paper's empirical claims rest on standard transformer machinery and pretrained RoBERTa; no invented theoretical entities are introduced. The most important fitted choices are hyperparameters tuned on the validation sets (decoder depth, focal-loss gamma, label smoothing, learning rates, batch size, dropout). The hierarchy-as-tree assumption enters in Section 1 and is used in the children-only ablation. The evaluation also assumes literature-reported baselines are comparable, which is a self-imposed methodological premise rather than a physical or mathematical axiom.

free parameters (7)
  • Number of decoder layers = 2
    Chosen from pre-experiments on WOS; affects model capacity and inference speed.
  • Focal loss focusing parameter gamma = 2
    Selected per dataset via wandb/grid search over the validation set; controls reweighting of hard samples.
  • Label smoothing = 0.1
    Selected by validation-based hyperparameter search; used inside cross-entropy before focal scaling.
  • Encoder learning rate = 5e-5 (per dataset, tuned)
    Tuned to validation performance; differs from decoder LR because encoder is pretrained.
  • Decoder learning rate = 3e-4 (per dataset, tuned)
    Tuned to validation performance; decoder is trained from scratch.
  • Batch size and gradient accumulation = 32 examples with accumulation over 2 batches
    Chosen via validation-based search; affects optimization dynamics.
  • Decoder dropout = 0.2
    Set during pre-experiments; affects regularization and final scores.
assumptions (5)
  • domain assumption The label hierarchy is a tree in which every non-root label has exactly one parent.
    Stated in Section 1; the decoder's child-to-parent sequences and the 'children only + hierarchy' reconstruction both rely on unique parent paths.
  • domain assumption Pre-trained RoBERTa provides a text representation that is sufficient for downstream label generation.
    Used throughout as the fixed-capacity encoder; no evidence is given that a weaker encoder would preserve the results.
  • domain assumption Ordered symbolic labels, without their textual names, carry enough signal for the decoder to learn the hierarchy.
    This is the central design premise behind replacing label semantics with symbolic identifiers; tested only on the three datasets.
  • standard math Transformer cross-attention can transfer encoder text information into the decoder's label generation.
    The decoder uses standard multi-head attention and TransformerBlock cross-attention per Vaswani et al.; treated as a given architectural capacity.
  • domain assumption Hyperparameters tuned on validation sets generalize to the test sets.
    All hyperparameters were selected with wandb/grid search on validation, and test results are reported from the best validation configuration.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Transformer-based Autoregressive Decoder Architecture for Hierarchical Text Classification." pith.science (2026). https://pith.science/paper/DUJEY4VH

@misc{pith2026250113598,
  author       = {Pith},
  title        = {Pith review of: A Transformer-based Autoregressive Decoder Architecture for Hierarchical Text Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUJEY4VH}},
  note         = {Machine review of arXiv:2501.13598}
}
read the original abstract

Recent approaches in hierarchical text classification (HTC) rely on the capabilities of a pre-trained transformer model and exploit the label semantics and a graph encoder for the label hierarchy. In this paper, we introduce an effective hierarchical text classifier RADAr (Transformer-based Autoregressive Decoder Architecture) that is based only on an off-the-shelf RoBERTa transformer to process the input and a custom autoregressive decoder with two decoder layers for generating the classification output. Thus, unlike existing approaches for HTC, the encoder of RADAr has no explicit encoding of the label hierarchy and the decoder solely relies on the label sequences of the samples observed during training. We demonstrate on three benchmark datasets that RADAr achieves results competitive to the state of the art with less training and inference time. Our model consistently performs better when organizing the label sequences from children to parents versus the inverse, as done in existing HTC approaches. Our experiments show that neither the label semantics nor an explicit graph encoder for the hierarchy is needed. This has strong practical implications for HTC as the architecture has fewer requirements and provides a speed-up by a factor of 2 at inference time. Moreover, training a separate decoder from scratch in conjunction with fine-tuning the encoder allows future researchers and practitioners to exchange the encoder part as new models arise. The source code is available at https://github.com/yousef-younes/RADAr.

Figures

Figures reproduced from arXiv: 2501.13598 by the authors.

Figure 1
Figure 1. The RADAr Model Architecture Encoder The model uses the RoBERTa base as its encoder part. The encoder is responsible for producing a fixed-size context tensor that captures the linguistic information found in the input text. To obtain the context tensor, the text is fed into the RoBERTa tokenizer to produce the token indices and attention masks. These are then fed into the RoBERTa model to produce an output. We use … view at source ↗
Figure 2
Figure 2. Label preprocessing and tokenization. Line a) contains the original labels. Line b) maps the original labels to the symbolic labels. Line c) adds the level separator token <unk>. Line d) organizes the labels level-wise from children to parents. Line e) contains the padded tokenizer output. tained query and the encoder mask m to a transformer block to per￾form cross attention [36] and produce the output of the decode… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 24 canonical work pages

  1. [14]

    Jiang, D

    T. Jiang, D. Wang, L. Sun, Z. Chen, F. Zhuang, and Q. Yang. Exploit- ing global and local hierarchies for hierarchical text classification. In EMNLP, pages 4030–4039, Abu Dhabi, United Arab Emirates, Dec

    work page

  2. [1]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. GPT-4 tech- nical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  3. [2]

    Al-Kababji, F

    A. Al-Kababji, F. Bensaali, and S. P. Dakua. Scheduling techniques for liver segmentation: ReduceLRonPlateau vs OneCycleLR. In Intel- ligent Systems and Pattern Recognition , pages 204–212, Cham, 2022. Springer

    work page 2022

  4. [3]

    J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016

    arXiv 2016

  5. [4]

    Banerjee, C

    S. Banerjee, C. Akkaya, F. Perez-Sorrosal, and K. Tsioutsiouliklis. Hi- erarchical transfer learning for multi-label text classification. In ACL, pages 6295–6300, 2019

    work page 2019

  6. [5]

    L. Biewald. Experiment tracking with weights and biases, 2020. URL https://www.wandb.com/. Software available from wandb.com

    2020

  7. [6]

    H. Chen, Q. Ma, Z. Lin, and J. Yan. Hierarchy-aware label semantics matching network for hierarchical text classification. In ACL-IJCNLP, pages 4370–4379. ACL, 2021. doi: 10.18653/v1/2021.acl-long.337

    doi:10.18653/v1/2021.acl-long.337 2021

  8. [7]

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple frame- work for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020

    work page 2020

Show all 47 references
  1. [8]

    De Silva, J

    A. De Silva, J. L. Wijekoon, R. Liyanarachchi, R. Panchendrarajan, and W. Rajapaksha. AI insights: A case study on utilizing chatgpt intelli- gence for research paper analysis. arXiv preprint arXiv:2403.03293 , 2024

    2024 arXiv

  2. [9]

    Z. Deng, H. Peng, D. He, J. Li, and P. Yu. HTCInfoMax: A global model for hierarchical text classification via information maximization. In K. Toutanova, A. Rumshisky, L. Zettlemoyer, D. Hakkani-Tur, I. Belt- agy, S. Bethard, R. Cotterell, T. Chakraborty, and Y . Zhou, editors...

    2021 doi

  3. [10]

    Galke, A

    L. Galke, A. Diera, B. X. Lin, B. Khera, T. Meuser, T. Singhal, F. Karl, and A. Scherp. Are we really making much progress in text classifica- tion? A comparative review. arXiv preprint arXiv:2204.03954, 2023

    2023 arXiv

  4. [11]

    Gopal and Y

    S. Gopal and Y . Yang. Recursive regularization for large-scale classifi- cation with hierarchical and graphical dependencies. InSIGKDD, pages 257–265, 2013

    2013

  5. [12]

    Große-Bölting, C

    G. Große-Bölting, C. Nishioka, and A. Scherp. A comparison of differ- ent strategies for automated semantic document annotation. In K-CAP 2015, pages 8:1–8:8. ACM, 2015. doi: 10.1145/2815833.2815838

    2015

  6. [13]

    G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012

    2012 arXiv

  7. [15]

    J. D. M.-W. C. Kenton and L. K. Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019

    2019

  8. [16]

    Kowsari, D

    K. Kowsari, D. E. Brown, M. Heidarysafa, K. J. Meimandi, M. S. Ger- ber, and L. E. Barnes. HDLTex: Hierarchical deep learning for text classification. In 2017 16th IEEE international conference on machine learning and applications (ICMLA) , pages 364–371. IEEE, 2017

    2017

  9. [17]

    D. D. Lewis, Y . Yang, T. Russell-Rose, and F. Li. RCV1: A new bench- mark collection for text categorization research. Journal of machine learning research, 5(Apr):361–397, 2004

    2004

  10. [18]

    Lewis, Y

    M. Lewis, Y . Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V . Stoyanov, and L. Zettlemoyer. BART: Denoising sequence-to- sequence pre-training for natural language generation, translation, and comprehension. In ACL, pages 7871–7880, Online, July 2020. ACL. doi: 10.18...

    2020 doi

  11. [19]

    Q. Li, H. Peng, J. Li, C. Xia, R. Yang, L. Sun, P. S. Yu, and L. He. A survey on text classification: From traditional to deep learning. ACM Transactions on Intelligent Systems and Technology (TIST) , 13:1 – 41,

  12. [20]

    X. Li, S. Chan, X. Zhu, Y . Pei, Z. Ma, X. Liu, and S. Shah. Are ChatGPT and GPT-4 general-purpose solvers for financial text analytics? a study on several typical tasks. InConference on Empirical Methods in Natural Language Processing, 2023

    2023

  13. [21]

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Dollár. Focal loss for dense object detection. In ICCV, pages 2980–2988, 2017

    2017

  14. [22]

    Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V . Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692, 2019

    1907 arXiv

  15. [23]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  16. [24]

    MacCartney

    B. MacCartney. Natural language inference. Stanford University, 2009

    2009

  17. [25]

    F. Mai, L. Galke, and A. Scherp. Using deep learning for title-based semantic subject indexing to reach competitive performance to full-text. In JCDL, pages 169–178. ACM, 2018. doi: 10.1145/3197026.3197039

    2018

  18. [26]

    Radford, K

    A. Radford, K. Narasimhan, T. Salimans, I. Sutskever, et al. Improving language understanding by generative pre-training. 2018

    2018

  19. [27]

    Raffel, N

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learn- ing Research, 21(1):5485–5551, 2020

    2020

  20. [28]

    Sebastiani

    F. Sebastiani. Machine learning in automated text categorization. ACM computing surveys (CSUR), 34(1):1–47, 2002

    2002

  21. [29]

    Shimura, J

    K. Shimura, J. Li, and F. Fukumoto. HFT-CNN: Learning hierar- chical category structure for multi-label short text categorization. In EMNLP, pages 811–816, Brussels, Belgium, Oct.-Nov. 2018. ACL. doi: 10.18653/v1/D18-1093

    2018 doi

  22. [30]

    C. N. Silla and A. A. Freitas. A survey of hierarchical classification across different application domains. Data mining and knowledge dis- covery, 22:31–72, 2011

    2011

  23. [31]

    Sun and E.-P

    A. Sun and E.-P. Lim. Hierarchical text classification and evaluation. In Proceedings 2001 IEEE ICDM, pages 521–528. IEEE, 2001

    2001

  24. [32]

    X. Sun, X. Li, J. Li, F. Wu, S. Guo, T. Zhang, and G. Wang. Text clas- sification via large language models. arXiv preprint arXiv:2305.08377, 2023

    2023 arXiv

  25. [33]

    R. Tarjan. Depth-first search and linear graph algorithms. SIAM Journal on Computing, 1(2):146–160, 1972. doi: 10.1137/0201010

    1972 doi

  26. [34]

    Touvron, L

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  27. [35]

    Vagliano, L

    I. Vagliano, L. Galke, and A. Scherp. Recommendations for item set completion: on the semantics of item co-occurrence with data sparsity, input size, and input modalities. Inf. Retr . J., 25(3):269–305, 2022. doi: 10.1007/S10791-022-09408-9

    2022 doi

  28. [36]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. Ad- vances in neural information processing systems (NeurIPS) , 30, 2017

    2017

  29. [37]

    Z. Wang, P. Wang, L. Huang, X. Sun, and H. Wang. Incorporating hierarchy into text encoder: a contrastive learning approach for hierar- chical text classification. In ACL, pages 7109–7119, Dublin, Ireland, May 2022. ACL. doi: 10.18653/v1/2022.acl-long.491

    2022 doi

  30. [38]

    R. J. Williams and D. Zipser. A learning algorithm for continually run- ning fully recurrent neural networks. Neural Computation, 1(2):270– 280, 1989. doi: 10.1162/neco.1989.1.2.270

    1989 doi

  31. [39]

    C. Wilt, J. Thayer, and W. Ruml. A comparison of greedy search algo- rithms. In Proceedings of the International Symposium on Combinato- rial Search, volume 1, pages 129–136, 2010

    2010

  32. [40]

    P. Yang, X. Sun, W. Li, S. Ma, W. Wu, and H. Wang. SGM: Se- quence generation model for multi-label classification. In COLING, pages 3915–3926, Santa Fe, New Mexico, USA, Aug. 2018. ACL

    2018

  33. [41]

    C. Ying, T. Cai, S. Luo, S. Zheng, G. Ke, D. He, Y . Shen, and T.-Y . Liu. Do transformers really perform badly for graph representation? In NeurIPS, volume 34, pages 28877–28888. Curran Associates, Inc., 2021

    2021

  34. [42]

    C. Yu, Y . Shen, and Y . Mao. Constrained sequence-to-tree generation for hierarchical text classification. In Proceedings of the 45th Interna- tional ACM SIGIR Conference on Research and Development in Infor- mation Retrieval, pages 1865–1869, 2022

    2022

  35. [43]

    H. Yu, Z. Yang, K. Pelrine, J. F. Godbout, and R. Rabbany. Open, closed, or small language models for text classification? arXiv preprint arXiv:2308.10092, 2023

    2023 arXiv

  36. [44]

    L. Yuan, Y . Chen, G. Cui, H. Gao, F. Zou, X. Cheng, H. Ji, Z. Liu, and M. Sun. Revisiting out-of-distribution robustness in NLP: Benchmark, analysis, and LLMs evaluations. arXiv preprint arXiv:2306.04618 , 2023

    2023 arXiv

  37. [45]

    Zhang, M

    Y . Zhang, M. Wang, C. Ren, Q. Li, P. Tiwari, B. Wang, and J. Qin. Pushing the limit of LLM capacity for text classification. CoRR, abs/2402.07470, 2024. doi: 10.48550/ARXIV .2402.07470

    2024 doi arXiv

  38. [46]

    J. Zhou, C. Ma, D. Long, G. Xu, N. Ding, H. Zhang, P. Xie, and G. Liu. Hierarchy-aware global model for hierarchical text classification. In ACL, pages 1106–1117, 2020. doi: 10.18653/v1/2020.acl-main.104

    2020 doi

  39. [2020]

    URL https://api.semanticscholar.org/CorpusID:220961531

Pith tools

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