DocHRL: A Hierarchical Reinforcement Learning Framework for Cost-Optimised Document Classification

Arjun Pankajakshan; Madhu Reddiboina; Mohammed Yousif; Prabhjot Singh

REVIEW 4 major objections 5 minor 28 references

DocHRL: A Hierarchical Reinforcement Learning Framework for Cost-Optimised Document Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read DocHRL claims that a learned hierarchical RL policy can route each document to the cheapest sufficient mix of classifiers, OCR, LLMs, or human review, producing higher accuracy at far lower cost than any fixed pipeline.

desk verdict Plausible framework, but the headline result isn't established: the cost metric is dominated by misclassification, and an unspecified frozen vision encoder may be doing the work, not the learned routing. read the letter →

arxiv 2607.22644 v1 pith:JIKGDQ77 submitted 2026-06-24 cs.AI cs.CVcs.LG

Mohammed Yousif , Prabhjot Singh , Arjun Pankajakshan , Madhu Reddiboina This is my paper

classification cs.AIcs.CVcs.LG

keywords documentclassificationhierarchicalreinforcementlearningcost-sensitiveadaptiveroutinghuman-in-the-loopPPORVL-CDIPresourceallocation

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

The paper sets out to show that document classification pipelines need not spend the same compute on every document. It introduces a hierarchical reinforcement learning agent that chooses, per document, among cheap vision classifiers, OCR, LLMs, and human review, with a reward equal to negative expected total cost (inference plus misclassification plus labor). On the RVL-CDIP benchmark the learned router reports macro F1 of 0.973 at an average cost of 2.74 normalized units, beating the best standalone classifier (0.913 accuracy, 8.74 cost), the best ensemble (0.923, 7.74), and the best static human-in-the-loop optimum (3.085 cost). The broader message is that accuracy and cost can improve together when routing is learned rather than fixed.

What carries the argument

The central object is a Markov decision process over a heterogeneous pool of document-understanding tools, with a two-level policy hierarchy: πtop picks an option, and a dedicated sub-policy πθ picks the exact model or annotator. The state combines a frozen vision embedding, an accumulating grid of soft-label predictions, an OCR embedding grid, and a binary mask of already-used actions, which prevents redundant queries. Learning uses Proximal Policy Optimisation with a tanh-gated trust region (a smooth replacement for hard ratio clipping) and AdaPolyTan, an adaptive return-scaling transform matched to a bounded value head. The cost model Ctotal = Σ κt sums inference fees, execution costs, an

What would settle it

Run the trained policy with real human annotators instead of the stochastic oracle that emits correct labels with probability 0.96, and measure true per-document cost on the RVL-CDIP test set; if the realized average cost is not below the 3.085 normalized-unit optimum of the best static entropy-threshold HIL system at comparable accuracy, the central cost-superiority claim is falsified. Alternatively, re-evaluate under Cf = 10 and Cf = 1000 and check whether the router still beats an optimally thresholded fixed cascade.

Watch

Extended reading notes

Core claim

The paper's central claim is that a two-level policy trained end-to-end with a cost-aware reward can dominate fixed document-classification pipelines on both accuracy and cost. The top-level policy selects among options—classify, llm, ocr, human, or a direct class prediction—and option-specific sub-policies choose the concrete model or tool. The reward is the negative discounted sum of action costs plus a terminal misclassification penalty, with failure cost Cf = 100 and human review at Ch = 2.25. On the RVL-CDIP test set the paper reports macro F1 of 0.973 at an average per-document cost of 2.74, which is 11.1% below the cheapest optimal static human-in-the-loop configuration and 69% below

Load-bearing premise

All reported savings are evaluated under the paper's own cost model with hand-set constants (Cf = 100, human review $2.25, human error 0.04), so the claimed 2.74 average cost and 69–88% reductions would change under different misclassification losses, human behavior, or API pricing; the paper itself flags this as a limitation.

Editorial extensions

If this is right

Fixed pipelines are dominated: if the reported results hold, a learned router beats every static standalone classifier, ensemble, and threshold-based human-in-the-loop system on both accuracy and cost simultaneously.
Escalation to humans can be made cheaper than the optimal static entropy threshold: 2.74 average cost versus 3.085 for the best HIL configuration, while also raising accuracy.
Accuracy gains concentrate on ambiguous classes (form, scientific report, presentation), where extra OCR or LLM evidence is invoked; easy classes like email, questionnaire, and resume terminate after one cheap classifier call.
The framework is re-optimizable by re-tuning the cost constants: changing Cf, Ch, or human error rate changes the routing policy, so the same machinery can be recalibrated for different deployment economics.
At test time the policy runs greedily with no reward computation, so the adaptive routing adds no training-time machinery at inference.

Reading between the lines

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

A simpler baseline the paper leaves untested is a learned confidence-threshold cascade from one cheap classifier to a strong one; comparing DocHRL against such a cascade would isolate how much of the gain comes from the hierarchy versus from adaptive routing in general.
Under more expensive human review or higher misclassification costs, the optimal policy would likely escalate more or less aggressively; the paper's 2.74 figure is specific to its chosen constants, not a universal number.
The balanced RVL-CDIP benchmark may favor the router's difficulty-aware behavior; real-world class imbalance and distribution shift could change both which actions get selected and the magnitude of cost savings.
If class-specific failure costs were available, the same framework could learn to be aggressive on low-stakes classes and cautious on high-stakes ones, which the paper lists as future work for regulated domains.

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.

Desk Editor's Note

read the letter

Quick take: this is a reasonable framework paper, but the headline result—that hierarchical RL routing beats static pipelines on both accuracy and cost—is not actually established by the evidence presented. The cost metric is dominated by misclassification cost, and the accuracy may be coming from an unspecified frozen VisionEncoder rather than from learned routing.

What's genuinely useful: the cost model in Eqs. (1)–(3) is coherent and cleanly separates inference, failure, and human costs. Formulating document classification as a hierarchical MDP over heterogeneous tools (vision, OCR, LLM, human) is a sensible extension of early-exit and cascade ideas. The paper is honest about its limitations—it explicitly lists the cascade baseline and realistic human annotators as future work. The numbers are internally consistent: with macro F1 = 0.973 and Cf = 100, the expected failure cost is 2.70, and the reported total cost is 2.74, so secondary tool costs are nearly negligible.

That last point is the soft spot. The claimed 69–88% cost reductions over static classifiers are essentially the accuracy advantage: static classifiers have error rates around 0.09, costing ~9 per document, while DocHRL's error rate is 0.027, costing ~2.7. The routing itself contributes at most ~0.04 per document. If a simple supervised head on the frozen vision embedding can reach 0.973 macro F1—and we are not told what that encoder is—then the entire benefit could come from the encoder, not the RL. The paper reports no action distribution, no episode-length statistics, and no ablation where the policy can only take terminal actions. The deferred cascade baseline is exactly the comparison that could undercut the need for hierarchical RL.

There are also standard reproducibility gaps: no code, no training details, no error bars, a single unseeded run, and the HIL baselines are tuned on the test-set entropy distribution. These are addressable, but they matter for a paper whose central claim is empirical.

Who should read it: people working on cost-sensitive adaptive inference and document understanding pipelines. The formal setup is clean and the limitations section is candid. But the empirical section needs major work before the central claim is convincing. I would send it to a serious referee, with the explicit expectation that the authors provide the missing encoder specifications, ablations (including a terminal-only policy and a cascade), and error bars. Without those, the result is a research teaser, not a demonstrated system.

Referee Report

4 major / 5 minor

Summary. The paper introduces DocHRL, a hierarchical reinforcement learning framework for document classification that adaptively selects among vision classifiers, LLMs, OCR, and human review on a per-document basis, with the reward being negative total expected cost. It defines cost models for standalone, ensemble, and human-in-the-loop pipelines (Eqs. 1–3), trains a two-level policy hierarchy with a PPO variant and AdaPolyTan return scaling, and evaluates on RVL-CDIP. The reported results show macro F1 = 0.973 and average per-document cost 2.74 normalized units, outperforming all static baselines (best standalone: 0.913 accuracy / 8.74 cost; CNN ensemble: 0.923 / 7.74; best HIL: 3.085) in both accuracy and cost.

Significance. If the results are reproducible and the ablations confirm that learned routing is the active ingredient, DocHRL would be a meaningful contribution to cost-sensitive document understanding: it formalizes a practical cost model, treats routing as an MDP, and demonstrates simultaneous accuracy and cost gains on an external benchmark. The paper also brings a useful inventory of cost estimates for heterogeneous tools and formulates a coherent PPO variant. However, the empirical core currently lacks the ablations and statistical support needed to establish that the hierarchical routing, rather than an unspecified frozen visual encoder with a terminal classifier head, is responsible for the gains.

major comments (4)

[Section 3.3 and Table 3] The central claim that learned hierarchical routing drives the gains is confounded by the terminal actions c1..cK and the unspecified frozen VisionEncoder. Because the policy can classify directly from v, and the reported average cost 2.74 is almost exactly (1−0.973)·100 = 2.70 (misclassification penalty) plus ~0.04 for all sub-policy costs, the cost advantage over static classifiers is numerically almost the accuracy advantage. Without (i) identifying the VisionEncoder, (ii) reporting action distributions and episode lengths, and (iii) ablating a terminal-only policy and a supervised head on v, the results do not show that adaptive routing (rather than the frozen features) is responsible. The Limitations section explicitly defers a simple cascade baseline, which is exactly the comparison needed.
[Section 4 (Evaluation Protocol / Results)] All empirical claims rest on a single unseeded run: no number of seeds, no standard deviations, no PPO hyperparameters (learning rate, number of episodes, discount factor, network widths, batch size), and no code release. With the cost metric being the training objective, the reported 2.74 could be within noise of alternative configurations; error bars over at least 5 seeds are required to support the superiority claims in Table 3.
[Section 3.1 / Table 2] The HIL baselines choose the escalation threshold Tmin by minimizing Eq. (3) using the test-set entropy distribution. This tunes the baseline on the test set, giving it an information advantage over DocHRL, which only uses training episodes. For a fair comparison, thresholds should be selected on a validation set (or reported as oracle upper bounds). Also, the Tmax columns are unused in the cost comparison; their presence confuses the presentation.
[Related Work / Limitations] The paper argues DocHRL generalises cascades and early-exit networks, but it provides no comparison to a simple confidence-threshold cascade (e.g., a cheap classifier routing to a stronger one), which is the natural non-RL baseline for cost-adaptive inference. Since the Limitations section defers this to future work, the incremental benefit of the RL hierarchy over threshold-based routing is not established.

minor comments (5)

[Eq. (11)] The text says β=0.2 controls an 'asymptotic bound [1−β,1+β]', but the linear leak term α(r−1) makes the multiplier unbounded as r→∞. Rephrase as a 'soft bound' or clarify the role of the leak.
[Eqs. (13)–(17)] The AdaPolyTan parameter ranges are given as p∈[100,1000] and βb∈[0.2,0.5], but Eq. (14) uses pmax/pmin without defining them; align notation.
[Table 2] The maximum-cost threshold Tmax and Cmax are never used in the comparison. Either remove them or explicitly justify why they are included.
[Section 4.3.2] Macro accuracy and macro F1 are reported as identical (0.973). In a balanced single-label setting this can happen but is not automatic; clarify whether per-class precision equals per-class recall or report both separately.
[References] Several references use 'et al.' without full author lists (e.g., Kim et al., Li et al., Su et al., McInnis et al.). Provide complete citations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DocHRL is benchmarked on an external dataset and its cost figure is the RL reward objective, which is standard evaluation rather than a derivation.

full rationale

DocHRL's central accuracy claim (macro F1 = 0.973) is measured on the fixed RVL-CDIP test split, an external benchmark, and is not derived from the paper's cost equations. The reported average cost of 2.74 normalized units is the value of the PPO cost objective on test episodes; evaluating a learned policy by the same reward it was trained to optimize is standard RL practice and does not constitute circular inference. The constants Cf = 100, Ch = $2.25, and Eh = 0.04 are stated hyperparameters, not parameters fitted to the reported outcome, and the Limitations section explicitly lists varying Cf as future work. No load-bearing self-citations or imported uniqueness theorems appear: the related-work references are contextual, and the novel components (tanh-gated PPO, AdaPolyTan) are defined in this paper rather than borrowed from the authors' prior results. The skeptic concerns about the unspecified frozen VisionEncoder and the missing terminal-only baseline are genuine evidence gaps and are partly conceded in the Limitations section (a simple cascade is deferred to future work), but a missing ablation is not an equation-level reduction of a prediction to an input. The human-in-the-loop baselines' thresholds are tuned on the test-set entropy distribution, which is a fairness concern, but it makes the comparison harder for DocHRL and is not circular. No step exhibits X defined in terms of Y, a fitted parameter renamed as a prediction, or a self-citation chain supporting the main conclusion. Therefore no significant circularity is found.

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

The central contribution is empirical, so the main ledger items are the hand-set cost constants and domain assumptions about the RVL-CDIP benchmark and the cost model. No new physical entities are introduced.

free parameters (4)

Cf (cost of failure) = 100.0 normalized units
Fixed hyperparameter across all classes/experiments; all cost comparisons and the reward function scale with it. The average cost 2.74 is in these units; without sensitivity analysis the cost claim is tied to this choice (Section 3.2).
Ch (human labeller cost) = $2.25/doc
Derived from assumed $27/hr and 5 minutes/doc; drives human-escalation cost in reward and HIL baseline.
Eh (human error rate) = 0.04
Chosen human error probability in the stochastic human model; drives HIL thresholds and human-action costs.
PPO hyperparameters (beta, alpha, c2 range) = beta=0.2, alpha=0.01, c2 from 0.2 to 0.02
Hand-chosen for the smooth trust region; not fitted to data but required for the method to work.

assumptions (4)

domain assumption RVL-CDIP ground-truth labels are correct and the 80/10/10 split is a valid evaluation of generalization.
Used as reward signal and evaluation; the original RVL-CDIP benchmark provides a standard split, and the paper adopts a similar ratio.
domain assumption The cost model (Eqs. 1-3) with independent cost components accurately represents real pipeline economics.
All cost comparisons use this model; if true costs differ (e.g., Cf varies by class), the claimed savings change. Paper admits Cf is domain-independent (Limitations).
domain assumption Classifier softmax probabilities are informative enough for the probability grid to serve as useful evidence for the policy's terminal classification.
The state's prob grid is the main evidence for the policy's final prediction; no calibration analysis is reported.
domain assumption Throughput scaling by pixel-count ratio (7.0 from 1024x1024 to 384x384) is a valid approximation for inference cost.
Used to compute Ci values in Table 1; Ci is negligible in the overall cost, so this is low impact.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DocHRL: A Hierarchical Reinforcement Learning Framework for Cost-Optimised Document Classification." pith.science (2026). https://pith.science/paper/JIKGDQ77

@misc{pith2026260722644,
  author       = {Pith},
  title        = {Pith review of: DocHRL: A Hierarchical Reinforcement Learning Framework for Cost-Optimised Document Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JIKGDQ77}},
  note         = {Machine review of arXiv:2607.22644}
}

read the original abstract

Real-world document classification pipelines typically apply the same sequence of models to every incoming document, regardless of its complexity or type. This leads to inefficient use of compute and human resources: simple documents are over-processed while difficult ones may not receive enough scrutiny. We introduce DocHRL, a hierarchical reinforcement learning framework that learns to adaptively and dynamically select the most cost-effective classification policy on a per-document basis. DocHRL formulates document classification as a sequential decision problem with a two-level policy hierarchy: a top-level policy selects among broad options (vision classifiers, LLMs, OCR, and human-in-the-loop review), while option-specific sub-policies choose the concrete model or tool to invoke. The reward signal is the negative total expected cost, which captures inference cost, cost of misclassification, and cost of human labelling. Trained with Proximal Policy Optimisation on the RVL-CDIP benchmark, DocHRL achieves a macro F1 of 0.973 across 16 document classes while reducing average per-document cost to 2.74 normalised units compared to substantially higher costs incurred by fixed standalone classifiers. Our results demonstrate that cost-aware reinforcement learning can simultaneously improve classification performance and operational efficiency in document understanding systems.

Figures

Figures reproduced from arXiv: 2607.22644 by the authors.

**Figure 1.** Human-in-the-loop pipeline with entropy-based escalation. The classifier processes an input; if the normalized Shannon entropy H exceeds threshold Th, the sample is escalated to a human operator. Otherwise, the classifier makes the final decision. Entropy-based escalation. Using the normalised Shannon entropy of the classifier’s output distribution, H = − K X−1 i=0 P(yi) log P(yi) log K , (4) the escalation probab… view at source ↗

**Figure 2.** Overview of the DocHRL framework. (a) Training: the agent interacts with each document for at most Nmax steps, accumulates cost Ct, and updates both policies via a custom PPO variant with tanh-gated trust region and AdaPolyTan reward scaling. (b) Inference: the same hierarchy runs greedily (argmax) with no reward computation. with value loss coefficient c1 = 0.5 and an entropy coefficient c2(t) that decays linearly… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 4 linked inside Pith

[1]

2015 13th international conference on document analysis and recognition (ICDAR) , pages=

Deepdocclassifier: Document classification with deep convolutional neural network , author=. 2015 13th international conference on document analysis and recognition (ICDAR) , pages=. 2015 , organization=

2015
[2]

2015 13th international conference on document analysis and recognition (ICDAR) , pages=

Evaluation of deep convolutional nets for document image classification and retrieval , author=. 2015 13th international conference on document analysis and recognition (ICDAR) , pages=. 2015 , organization=

2015
[3]

2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume=

Real-time document image classification using deep CNN and extreme learning machines , author=. 2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume=. 2017 , organization=

2017
[4]

Joint European conference on machine learning and knowledge discovery in databases , pages=

Multimodal deep networks for text and image-based document classification , author=. Joint European conference on machine learning and knowledge discovery in databases , pages=. 2019 , organization=

2019
[5]

Ninth international conference on document analysis and recognition (ICDAR 2007) , volume=

An overview of the Tesseract OCR engine , author=. Ninth international conference on document analysis and recognition (ICDAR 2007) , volume=. 2007 , organization=

2007
[6]

ACM Computing Surveys (CSUR) , volume=

Document layout analysis: a comprehensive survey , author=. ACM Computing Surveys (CSUR) , volume=. 2019 , publisher=

2019
[7]

arXiv preprint arXiv:1904.08398 , year=

Docbert: Bert for document classification , author=. arXiv preprint arXiv:1904.08398 , year=

arXiv 1904
[8]

International conference on machine learning , pages=

Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

2021

Show all 28 references

[9]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=
[10]

International Conference on Computational Science , pages=

Improving accuracy and speeding up document image classification through parallel systems , author=. International Conference on Computational Science , pages=. 2020 , organization=

2020
[11]

Layoutlmv2: Multi-modal pre-training for visually-rich document understanding , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=
[12]

2022 26th International Conference on Pattern Recognition (ICPR) , pages=

Are deep models robust against real distortions? A case study on document image classification , author=. 2022 26th International Conference on Pattern Recognition (ICPR) , pages=. 2022 , organization=

2022
[13]

Proceedings of the European Conference on Computer Vision (ECCV) , year =

Geewook Kim and others , title =. Proceedings of the European Conference on Computer Vision (ECCV) , year =
[14]

Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI) , year =

Charles Elkan , title =. Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI) , year =
[15]

Barto and Sridhar Mahadevan , title =

Andrew G. Barto and Sridhar Mahadevan , title =. Discrete Event Dynamic Systems , volume =
[16]

Sutton and Doina Precup and Satinder Singh , title =

Richard S. Sutton and Doina Precup and Satinder Singh , title =. Artificial Intelligence , volume =
[17]

Surat Teerapittayanon and Bradley McDanel and H. T. Kung , title =. Proceedings of the International Conference on Pattern Recognition (ICPR) , year =
[18]

CVPR , year =

Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun , title =. CVPR , year =
[19]

Le , title =

Mingxing Tan and Quoc V. Le , title =. ICML , year =
[20]

ICCV , year =

Ze Liu and others , title =. ICCV , year =
[21]

arXiv preprint arXiv:1707.06347 , year =

John Schulman and Filip Wolski and Prafulla Dhariwal and Alec Radford and Oleg Klimov , title =. arXiv preprint arXiv:1707.06347 , year =

arXiv
[22]

Proximal Policy Optimization Smoothed Algorithm , journal =

Wangshu Zhu and Andr. Proximal Policy Optimization Smoothed Algorithm , journal =
[23]

Neural Processing Letters , volume =

Binglin Su and others , title =. Neural Processing Letters , volume =
[24]

arXiv preprint arXiv:2511.20347 , year =

Yuan Li and others , title =. arXiv preprint arXiv:2511.20347 , year =

arXiv
[25]

arXiv preprint arXiv:2509.21282 , year =

Madeleine Dwyer and Adam Sobey and Adriane Chapman , title =. arXiv preprint arXiv:2509.21282 , year =
[26]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Hado van Hasselt and Arthur Guez and Matteo Hessel and Volodymyr Mnih and David Silver , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[27]

arXiv preprint arXiv:2301.04104 , year =

Danijar Hafner and Jurgis Pasukonis and Jimmy Ba and Timothy Lillicrap , title =. arXiv preprint arXiv:2301.04104 , year =

arXiv
[28]

OpenReview , year =

Benjamin McInnis and others , title =. OpenReview , year =

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

2015 13th international conference on document analysis and recognition (ICDAR) , pages=

Deepdocclassifier: Document classification with deep convolutional neural network , author=. 2015 13th international conference on document analysis and recognition (ICDAR) , pages=. 2015 , organization=

2015

[2] [2]

2015 13th international conference on document analysis and recognition (ICDAR) , pages=

Evaluation of deep convolutional nets for document image classification and retrieval , author=. 2015 13th international conference on document analysis and recognition (ICDAR) , pages=. 2015 , organization=

2015

[3] [3]

2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume=

Real-time document image classification using deep CNN and extreme learning machines , author=. 2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume=. 2017 , organization=

2017

[4] [4]

Joint European conference on machine learning and knowledge discovery in databases , pages=

Multimodal deep networks for text and image-based document classification , author=. Joint European conference on machine learning and knowledge discovery in databases , pages=. 2019 , organization=

2019

[5] [5]

Ninth international conference on document analysis and recognition (ICDAR 2007) , volume=

An overview of the Tesseract OCR engine , author=. Ninth international conference on document analysis and recognition (ICDAR 2007) , volume=. 2007 , organization=

2007

[6] [6]

ACM Computing Surveys (CSUR) , volume=

Document layout analysis: a comprehensive survey , author=. ACM Computing Surveys (CSUR) , volume=. 2019 , publisher=

2019

[7] [7]

arXiv preprint arXiv:1904.08398 , year=

Docbert: Bert for document classification , author=. arXiv preprint arXiv:1904.08398 , year=

arXiv 1904

[8] [8]

International conference on machine learning , pages=

Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

2021