BGM-HAN: A Hierarchical Attention Network for Accurate and Fair Decision Assessment on Semi-Structured Profiles

Junhua Liu; Kwan Hui Lim; Roy Ka-Wei Lee

REVIEW 4 major objections 6 minor 28 references

BGM-HAN: A Hierarchical Attention Network for Accurate and Fair Decision Assessment on Semi-Structured Profiles

T0 review · 4 major / 6 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 BGM-HAN reports 85.06% accuracy and beats GPT-4o on semi-structured admissions profiles.

desk verdict A clearly written architecture paper whose headline claim of significant superiority is not backed by any significance testing on a 154-sample test set; the idea is worth reviewing, but the evidence needs major reinforcement. read the letter →

arxiv 2507.17472 v1 pith:YSGS5W5P submitted 2025-07-23 cs.LG cs.AIcs.IR

Junhua Liu , Roy Ka-Wei Lee , Kwan Hui Lim This is my paper

classification cs.LGcs.AIcs.IR

keywords hierarchicalattentionnetworkbyte-pairencodingmulti-headself-attentiongatedresidualconnectionsuniversityadmissionssemi-structureddatadecisionsupportLLMbenchmark

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

BGM-HAN is a neural network built to score university applicant profiles made of structured grades, semi-structured leadership entries, and free-form essays. The paper claims that by encoding each field with byte-pair encoding and passing the text through three levels of attention — tokens, sentences, fields — the model captures the hierarchy of a profile and makes better admission predictions than standard machine-learning models and GPT-4o. On a proprietary dataset of 3,083 profiles from a single admissions cycle, BGM-HAN reports 85.06% accuracy and a macro-averaged F1-score of 0.8453, beating XGBoost, BiLSTM variants, HAN, and zero-shot and retrieval-augmented GPT-4o. The paper frames the result as evidence that domain-aware hierarchical models can support high-stakes decisions with better accuracy and interpretability than general-purpose LLMs.

What carries the argument

The load-bearing object is the three-level hierarchy token → sentence → field, implemented in the BGM-HAN network. Each level applies layer normalization, multi-head self-attention, a gated residual connection, and mean pooling, so local word information is composed into sentence embeddings and then field embeddings before a final binary classifier. Byte-pair encoding supplies a 5,000-token subword vocabulary that keeps rare and out-of-vocabulary terms representable; the gated residual term $\gamma \odot \mathrm{FFN}(X) + X$ lets the network decide how much nonlinear transformation to pass forward; and a class-weighted binary cross-entropy loss counteracts the imbalance between offers and rejections.

What would settle it

Resample the 3,083 profiles repeatedly (e.g., bootstrap or repeated stratified 90/5/5 splits), retrain or refit the strongest baselines, and compute confidence intervals for accuracy and F1; if the interval for BGM-HAN's margin over BiLSTM-Concat and XGBoost frequently includes zero, the reported superiority is not established.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the multi-level structure of an admissions profile is learnable and predictive. BGM-HAN scores each applicant by first tokenizing every text field with a learned BPE subword vocabulary, then composing word-level representations into sentence-level and field-level representations with layer normalization, multi-head self-attention, gated residual connections, and mean pooling. Applied to 3,083 real applicant profiles, this architecture reaches a macro-averaged F1-score of 0.8453 and an accuracy of 0.8506, the highest values in the paper's table, ahead of XGBoost (F1 0.7878), BiLSTM-Concat (F1 0.8176), and GPT-4o-RA (F1 0.7352). The paper concludes that hierarchical modeling of semi-structured profiles beats both traditional discriminative models and out-of-the-box LLMs for this decision task.

Load-bearing premise

The performance claim depends on a single stratified 5% test split — roughly 154 profiles from one admission year — with no confidence intervals or significance tests, so a few percentage points of separation between models could be sampling noise.

Editorial extensions

If this is right

A comparatively small trained classifier can beat zero-shot and five-shot GPT-4o on this admissions task, weakening the case for using general-purpose LLMs without domain adaptation.
The hierarchy itself contributes: plain HAN trails BGM-HAN by 7.4 F1 points and 9.6 accuracy points in the paper's reported results.
Each added mechanism improves the base model in the ablation, with multi-head attention listed as the largest single gain.
The model's field-level attention could be shown to reviewers as evidence behind an admission decision, supporting the paper's interpretability goal.

Reading between the lines

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

Editorial inference: the 2.8-point F1 margin over BiLSTM-Concat and the 5.8-point margin over XGBoost come from one test split; a bootstrap analysis would tell whether the hierarchy is actually buying the stated gains.
Editorial inference: the fairness framing is not backed by outcome metrics; group-wise error rates, calibration, or counterfactual tests would be needed before calling the model fair.
Editorial inference: if the hierarchy generalizes, the same token-to-sentence-to-field pattern could transfer to other semi-structured evaluation tasks such as resume screening and loan applications, but that transfer requires separate validation on each new domain.
Editorial inference: attention weights are presented as interpretability, yet the paper does not test whether the attended sentences are the ones that actually drive the prediction; an attention-faithfulness perturbation test would settle that.

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

Colleague,

Here's the short version: BGM-HAN is a sensible engineering combination—BPE tokenization, multi-head attention, gated residual connections, and a hierarchical attention backbone—applied to semi-structured university admissions profiles. The paper is clearly written, the code and hyperparameters are provided, and the ablation study gives each component credit. That part is fine. The problem is the central empirical claim: the abstract and contribution list say BGM-HAN "significantly outperforms" all baselines, including LLMs and human evaluators, but the reported evidence does not support that.

The test set is about 154 profiles (5% of 3,083). The accuracy gap over the best baseline, BiLSTM-Concat, is 0.023 (0.8506 vs 0.8276). With n=154, that's roughly three to four correct predictions, and the standard error of an accuracy difference is on the order of 0.04. No confidence intervals, no significance tests, no multiple-seed variance are reported. The word "significantly" is doing work the numbers cannot bear. This is not a nitpick; it is the load-bearing condition for the paper's main contribution. The hyperparameter selection on validation accuracy also introduces optimism bias that is never quantified.

Other soft spots are more minor but real. The title and abstract promise fairness, yet there is no fairness evaluation—no demographic breakdowns, no group-wise metrics, no bias analysis. The class-weighted loss is not a fairness measure. The contributions also claim comparison with human evaluators, but no human evaluation appears in Table 2 or anywhere in Section 4. The ablation percentages in Section 4.7 also don't quite add up (the component gains sum to 9.6% F1, but the total is reported as 7.4%), which suggests sloppy arithmetic or reporting. The dataset is proprietary and unavailable, so replication is blocked regardless of the code.

Citation pattern: fine. Self-citations appear in related work but don't carry the main claim. No obvious citation stuffing. The writing is honest about the data limitation, which I appreciate.

Who is this for? Researchers working on hierarchical models for structured applicant review might find the architecture a useful baseline. But the paper, as it stands, overclaims. It deserves a serious referee because the problem is legitimate and the architecture is plausible, and the flaws are fixable with more careful statistics and a revised abstract. I would not cite the performance numbers yet, but I would engage with the design.

Recommendation: send to peer review, but the review should push hard on the statistical substantiation and the missing fairness/human-evaluation components.

Best, [You]

Referee Report

4 major / 6 minor

Summary. The paper proposes BGM-HAN, a hierarchical attention network with byte-pair encoding, gated residual connections, and multi-head attention, for binary admission decisions on semi-structured applicant profiles. The model is evaluated on a proprietary dataset of 3,083 profiles against traditional ML, neural, and zero-/few-shot GPT-4o baselines, reporting the best accuracy (0.8506) and macro-F1 (0.8453). The authors claim significant superiority over all baselines, interpretability, and fairness, and they release source code. The architecture and training details are clearly specified, but the central empirical claims are undermined by the small test set, the absence of any statistical uncertainty quantification, and the lack of a fairness evaluation.

Significance. If the empirical claims were established, BGM-HAN would be a useful contribution to modeling hierarchical semi-structured data in high-stakes decision settings, and the code release would support follow-up work. The paper's strengths include a clearly specified architecture with algorithmic descriptions, a real-world admissions dataset, a broad baseline suite spanning traditional ML to LLMs, and the use of a class-weighted loss to address imbalance. However, the significance is currently limited by three major gaps: (1) the claim of significant superiority is not supported by any confidence intervals, significance tests, or repeated-seed variance; (2) the fairness promised in the title and introduction is never measured; and (3) the ablation study's component-wise gains are asserted without reporting the intermediate model results. These gaps are load-bearing for the paper's main claims, so the contribution remains promising but not yet established.

major comments (4)

[Abstract; §4.4; Table 2] The central claim that BGM-HAN 'significantly outperforms' all baselines is not supported by the reported statistics. The test split is 5% of 3,083 profiles, i.e., approximately 154 samples. The margin over BiLSTM-Concat is 0.0230 in accuracy and 0.0277 in macro-F1, which corresponds to only a handful of test-set predictions. No confidence intervals, significance tests, or repeated-seed variance are reported anywhere in Section 4, so the observed margins are within typical sampling noise for this test size. Please provide uncertainty quantification (e.g., bootstrap confidence intervals, paired significance tests, or multiple-seed runs) and revise the wording 'significantly' in the abstract and Section 4.6 to match the statistical evidence.
[§3.5; §4.6; Title] The fairness claim is unevaluated. The weighted binary cross-entropy loss in Section 3.5 compensates for class imbalance, but it does not by itself ensure or measure fairness. No protected attributes, fairness metrics (e.g., demographic parity or equalized odds), or fairness audits are presented in Section 4, yet the title and contribution 3 promise a 'fair' decision assessment. The manuscript should either add a concrete fairness evaluation or explicitly reframe the fairness contribution as a design aspiration rather than an empirically supported property.
[§4.7; Table 2] The ablation study reports component-wise gains (BPE +1.8%, multi-head attention +5.2%, gated residual connections +2.6%) and a total F1 gain of 7.4% over the base HAN, but Table 2 contains only the base HAN and the full BGM-HAN rows. The intermediate configurations with each enhancement added are never reported, making the claimed component contributions unverifiable and preventing readers from checking whether the gains are additive or interact. Please provide a complete ablation table with all component combinations, including uncertainty estimates, or remove the component-wise percentages.
[§1, Contribution 3; §4.6] The introduction claims that BGM-HAN outperforms 'human evaluators,' but no human evaluation results are presented in Table 2 or anywhere in Section 4. This unsupported claim should be either substantiated with data or removed. The text around Table 2 also refers to results 'across proposed models, human evaluation, and different categories of baseline models,' yet no human row exists in the table; please reconcile the text and the table.

minor comments (6)

[§4.2; §4.5] Section 4.2 reports that BGM-HAN was trained with the AdamW optimizer, while Section 4.5 states that 'neural models' use Adam; please clarify which optimizer applies to BGM-HAN and to each neural baseline to avoid ambiguity.
[§3.3] There is a typo in Section 3.3: 'attend to different different positions' should read 'attend to different positions.'
[References] References [24] and [25] both cite Vaswani et al., and references [27] and [28] both cite Yang et al. with overlapping titles; please consolidate the duplicates.
[§4.4] The phrase 'missing values in text fields are replaced with NaN tokens' should be made more precise, for example by specifying a dedicated '[PAD]' or '[MASK]' token and describing how the model treats it.
[§2] The mapping in Section 2 is written as D : P → 0, 1; it should be D : P → {0, 1} to denote the binary label set.
[§4.5; Table 2] The GPT-4o baselines are zero-shot and retrieval-augmented few-shot without fine-tuning; the text should consistently say so when claiming superiority over 'large language models,' since a fine-tuned LLM would be a stronger comparison point.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BGM-HAN's performance claim rests on a standard train/validation/test evaluation, and the self-citations in related work are not load-bearing.

full rationale

The paper's central claim is an empirical performance comparison: BGM-HAN is trained on a 90% split and evaluated on a held-out 5% test split. Section 4.4 states: 'We split the dataset into training (90%), validation (5%), and test (5%) subsets using stratified sampling to preserve the class distribution across splits.' The reported F1-score and accuracy in Table 2 are therefore genuine held-out predictions, not quantities defined in terms of the target result. Hyperparameters were selected on validation accuracy (Section 4.2), not on test labels, so the test numbers are not fitted inputs renamed as predictions. The ablation discussion compares BGM-HAN to the base HAN empirically; no component is claimed to derive from the final metric by construction. The only self-citations are [16] and [17], which appear in related-work and baseline contexts and do not justify the central performance claim. There is no uniqueness theorem imported from the authors, no ansatz smuggled in via self-citation, and no renaming of a known result as a new contribution. The main weakness is statistical: with roughly 154 test profiles and no significance tests or confidence intervals, the word 'significantly' in the Abstract is not established. That is a statistical-inference concern, not a circularity concern. Accordingly, no circular step is identified.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central experimental claim rests on a number of free hyperparameters chosen by grid search on validation accuracy, and several domain assumptions about the dataset and problem setup. No new physical entities are introduced. The model architecture is a combination of existing components, so the main burden is on the availability and representativeness of the data.

free parameters (8)

hidden dimension = 1024
Selected by grid search on validation accuracy (Section 4.2).
number of attention heads = 8
Selected by grid search on validation accuracy (Section 4.2).
dropout rate = 0.6
Selected by grid search on validation accuracy (Section 4.2).
learning rate = 1e-5
Selected by grid search on validation accuracy (Section 4.2).
batch size = 32
Selected by grid search on validation accuracy (Section 4.2).
BPE vocabulary size = 5000
Set in Section 3.2 without a search or justification; affects tokenization and model capacity.
hierarchical dimensions (s,w,d) = (10, 50, 768)
Set in Section 3.2 as maximum sentences, words per sentence, embedding dimension; chosen by hand.
L2 regularization weight decay lambda
Mentioned in Section 3.5 but its value is not reported, so the exact training objective is incompletely specified.

assumptions (5)

domain assumption The four profile components (GCE A-Level results, GCE O-Level results, leadership records, PIQ responses) and the binary admission label are sufficient to model the admission decision.
Section 2 defines the problem this way; if other factors such as interviews or recommendations determine admissions, the model's input space is incomplete.
domain assumption The 3,083 profiles from a single year at one engineering university are representative of the admissions decision task in general.
Section 4.3 describes the dataset; no multi-institution or multi-year validation is provided.
standard math Standard neural network training, including backpropagation and AdamW optimization, converges to a model whose performance generalizes to the held-out test set.
Invoked in the training setup in Sections 3.5 and 4.1-4.2; no theoretical guarantees are given.
ad hoc to paper Class-weighted binary cross-entropy loss is an adequate mechanism for achieving fairness.
Section 3.5 introduces class weights; no fairness metric is evaluated, so the connection between the loss and fairness is asserted rather than demonstrated.
domain assumption The BPE vocabulary learned on the corpus does not leak test-set information, even though the train/test split for vocabulary learning is not specified.
Section 3.2 describes BPE vocabulary construction over the corpus; the paper does not state whether the split is respected.

how reviews work

0 comments

Cite this review

Pith. "Pith review of BGM-HAN: A Hierarchical Attention Network for Accurate and Fair Decision Assessment on Semi-Structured Profiles." pith.science (2026). https://pith.science/paper/YSGS5W5P

@misc{pith2026250717472,
  author       = {Pith},
  title        = {Pith review of: BGM-HAN: A Hierarchical Attention Network for Accurate and Fair Decision Assessment on Semi-Structured Profiles},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YSGS5W5P}},
  note         = {Machine review of arXiv:2507.17472}
}

read the original abstract

Human decision-making in high-stakes domains often relies on expertise and heuristics, but is vulnerable to hard-to-detect cognitive biases that threaten fairness and long-term outcomes. This work presents a novel approach to enhancing complex decision-making workflows through the integration of hierarchical learning alongside various enhancements. Focusing on university admissions as a representative high-stakes domain, we propose BGM-HAN, an enhanced Byte-Pair Encoded, Gated Multi-head Hierarchical Attention Network, designed to effectively model semi-structured applicant data. BGM-HAN captures multi-level representations that are crucial for nuanced assessment, improving both interpretability and predictive performance. Experimental results on real admissions data demonstrate that our proposed model significantly outperforms both state-of-the-art baselines from traditional machine learning to large language models, offering a promising framework for augmenting decision-making in domains where structure, context, and fairness matter. Source code is available at: https://github.com/junhua/bgm-han.

Figures

Figures reproduced from arXiv: 2507.17472 by the authors.

**Figure 1.** Architecture of the proposed BGM-HAN model. The multi-level model learns features from token to sentence to field. At each level, the data will go through layer normalisation, multi-head self-attention (MHA), gated residual connection (GRC), mean pooling to form the higher level embeddings. The embeddings are then concatenated and reshaped into 3D tensors to continue with the next level processing. automated system… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 20 canonical work pages

[1]

Advances in Neural Information Processing Systems36 (2024)

Alur, R., Laine, L., Li, D., Raghavan, M., Shah, D., Shung, D.: Auditing for human expertise. Advances in Neural Information Processing Systems36 (2024)

work page 2024
[2]

and it’s bi- ased against blacks

Angwin, J., Larson, J., Mattu, S., Kirchner, L.: Machine bias: There’s software used across the country to predict future criminals. and it’s bi- ased against blacks. ProPublica (2016), https://www.propublica.org/article/ machine-bias-risk-assessments-in-criminal-sentencing

work page 2016
[3]

Anthropic: The claude 3 model family: Opus, sonnet, haiku (2024), https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/ Model_Card_Claude_3.pdf

work page 2024
[4]

In: Proceedings of the 40th International Conference on Machine Learning

Basu, S., Rawat, A.S., Zaheer, M.: A statistical perspective on retrieval-based models. In: Proceedings of the 40th International Conference on Machine Learning. pp. 1852–1886 (2023)

work page 2023
[5]

In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)

Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder–decoder for sta- tistical machine translation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 1724–1734 (2014)

work page 2014
[6]

In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 4171–4186 (2019)

work page 2019
[7]

Science Advances 4(1), eaao5580 (2018)

Dressel, J., Farid, H.: The accuracy, fairness, and limits of predicting recidivism. Science Advances 4(1), eaao5580 (2018). https://doi.org/10.1126/sciadv.aao5580, https://www.science.org/doi/10.1126/sciadv.aao5580

doi:10.1126/sciadv.aao5580 2018
[8]

arXiv preprint arXiv:2407.21783 (2024)

Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

arXiv 2024

Show all 28 references

[9]

arXiv preprint arXiv:2403.00811 (2024), https://arxiv.org/abs/2403.00811

Echterhoff, J., et al.: Cognitive bias in high-stakes decision-making with llms. arXiv preprint arXiv:2403.00811 (2024), https://arxiv.org/abs/2403.00811

2024 arXiv
[10]

arXiv preprint arXiv:2208.05126 (2022), https://arxiv

Ghai, B., Mueller, K.: D-bias: A causality-based human-in-the-loop system for tackling algorithmic bias. arXiv preprint arXiv:2208.05126 (2022), https://arxiv. org/abs/2208.05126

2022 arXiv
[11]

Haag, F., Stingl, C., Zerfass, K., Hopf, K., Staake, T.: Overcoming anchoring bias: Thepotentialofaiandxai-baseddecisionsupport.arXivpreprintarXiv:2405.04972 (2024), https://arxiv.org/abs/2405.04972

2024 arXiv
[12]

Neural Computation 9(8), 1735–1780 (1997) Byte-Pair Encoded, Gated Multi-head Hierarchical Attention Network 15

Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735–1780 (1997) Byte-Pair Encoded, Gated Multi-head Hierarchical Attention Network 15

1997
[13]

In: 2021 IEEE International Conference on Data Mining (ICDM)

Iyer, R.G., Wang, W., Sun, Y.: Bi-level attention graph neural networks. In: 2021 IEEE International Conference on Data Mining (ICDM). pp. 1126–1131 (2021)

2021
[14]

Frontiers in Psychology14, 10071311 (2023), https://pmc.ncbi.nlm.nih

Kahneman,D.,Tversky,A.:Cognitivebiasandhowtoimprovesustainabledecision making. Frontiers in Psychology14, 10071311 (2023), https://pmc.ncbi.nlm.nih. gov/articles/PMC10071311/

2023
[15]

In: Advances in Neural Information Processing Systems

Lample, G., Conneau, A.: Cross-lingual language model pretraining. In: Advances in Neural Information Processing Systems. vol. 32 (2019)

2019
[16]

Liu, J., Lee, R.K.W., Lim, K.H.: Understanding fairness-accuracy trade-offs in ma- chine learning models: Does promoting fairness undermine performance? In: Pro- ceedings of the International Conference on Advances in Social Networks Analysis and Mining (ASONAM’25) (2025)

2025
[17]

Journal of Big Data9(1), 1–16 (2022)

Liu, J., Ng, Y.C., Gui, Z., Singhal, T., Blessing, L.T.M., Wood, K.L., Lim, K.H.: Title2vec: a contextual job title embedding for occupational named entity recog- nition and other applications. Journal of Big Data9(1), 1–16 (2022)

2022
[18]

arXiv preprint arXiv:1907.11692 (2019)

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

2019 arXiv
[19]

OpenAI: Gpt-4 technical report (2024)

2024
[20]

Decision Support Systems63, 63–66 (2019)

Phillips-Wren, G., Power, D.J., Mora, M.: Cognitive bias, decision styles, and risk attitudes in decision making and dss. Decision Support Systems63, 63–66 (2019)

2019
[21]

arXiv preprint arXiv:2004.04343 (2020), https: //arxiv.org/abs/2004.04343

Ribeiro, J.G., Felisberto, F.S., Neto, I.C.: Pruning and sparsemax methods for hierarchical attention networks. arXiv preprint arXiv:2004.04343 (2020), https: //arxiv.org/abs/2004.04343

2020 arXiv
[22]

arXiv preprint arXiv:1611.01260 (2016)

Savarese, P.H., Mazza, L.O., Figueiredo, D.R.: Learning identity mappings with residual gates. arXiv preprint arXiv:1611.01260 (2016)

2016 arXiv
[23]

Ethics and Information Technology 26, 85–102 (2024)

Smith, J., Doe, J., Lee, A.: Bias and fairness in high-stakes ai: Challenges of data sensitivity and access. Ethics and Information Technology 26, 85–102 (2024). https://doi.org/10.1007/s10676-024-09746-w, https://link.springer.com/ article/10.1007/s10676-024-09746-w

2024 doi
[24]

Advances in Neural Information Processing Systems (2017)

Vaswani, A.: Attention is all you need. Advances in Neural Information Processing Systems (2017)

2017
[25]

In: Advances in Neural Information Processing Systems

Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30 (2017)

2017
[26]

arXiv preprint arXiv:2404.05228 (2024), https://arxiv.org/abs/2404.05228

Yang, M., et al.: Fair machine guidance to enhance fair decision making in biased people. arXiv preprint arXiv:2404.05228 (2024), https://arxiv.org/abs/2404.05228

2024 arXiv
[27]

In: Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies

Yang, Z., Yang, D., Dyer, C., He, X., Smola, A., Hovy, E.: Hierarchical attention networks for document classification. In: Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 1480–...

2016
[28]

In: Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies

Yang, Z., Yang, D., Dyer, C., He, X., Smola, A., Hovy, E.: Hierarchical attention networks for document classification. In: Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies. pp. 1480–...

2016

Pith tools

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

[1] [1]

Advances in Neural Information Processing Systems36 (2024)

Alur, R., Laine, L., Li, D., Raghavan, M., Shah, D., Shung, D.: Auditing for human expertise. Advances in Neural Information Processing Systems36 (2024)

work page 2024

[2] [2]

and it’s bi- ased against blacks

Angwin, J., Larson, J., Mattu, S., Kirchner, L.: Machine bias: There’s software used across the country to predict future criminals. and it’s bi- ased against blacks. ProPublica (2016), https://www.propublica.org/article/ machine-bias-risk-assessments-in-criminal-sentencing

work page 2016

[3] [3]

Anthropic: The claude 3 model family: Opus, sonnet, haiku (2024), https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/ Model_Card_Claude_3.pdf

work page 2024

[4] [4]

In: Proceedings of the 40th International Conference on Machine Learning

Basu, S., Rawat, A.S., Zaheer, M.: A statistical perspective on retrieval-based models. In: Proceedings of the 40th International Conference on Machine Learning. pp. 1852–1886 (2023)

work page 2023

[5] [5]

In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)

Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder–decoder for sta- tistical machine translation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 1724–1734 (2014)

work page 2014

[6] [6]

In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 4171–4186 (2019)

work page 2019

[7] [7]

Science Advances 4(1), eaao5580 (2018)

Dressel, J., Farid, H.: The accuracy, fairness, and limits of predicting recidivism. Science Advances 4(1), eaao5580 (2018). https://doi.org/10.1126/sciadv.aao5580, https://www.science.org/doi/10.1126/sciadv.aao5580

doi:10.1126/sciadv.aao5580 2018

[8] [8]

arXiv preprint arXiv:2407.21783 (2024)

Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

arXiv 2024