Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Adversarial Topic-aware Prompt-tuning for Cross-topic Automated Essay Scoring

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Cross-topic essay scoring improves by jointly learning shared and topic-specific prompts.

desk verdict The ATOP architecture is a genuine new combination for cross-topic AES, but the evaluation protocol in Appendix B selects the best epoch on test-set QWK, so the reported gains are not trustworthy as-is. read the letter →

arxiv 2508.05987 v1 pith:UIWFXXCB submitted 2025-08-08 cs.CL

Chunyun Zhang , Hongyan Zhao , Chaoran Cui , Qilong Song , Zhiqing Lu , Shuai Gong , Kailin Liu This is my paper · ORCID
classification cs.CL
keywords automatedessayscoringcross-topicgeneralizationprompt-tuningsoftpromptsadversarialtrainingpseudo-labelingmulti-traitASAP++
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 claims that cross-topic automated essay scoring works better when a model explicitly captures both what all essay topics share and what each topic needs. It introduces ATOP, a prompt-tuning method that prepends a soft prompt made of shared and topic-specific vectors to every essay and feeds it into a frozen pre-trained language model. A topic discriminator and a pseudo-labeling loop train the two prompt parts: adversarial training pushes the shared prompt toward topic-invariant features, while neighbor-based pseudo-labels give the target topic's specific prompt a learning signal without human scores. On the ASAP++ benchmark, ATOP reports higher average QWK than nine baselines, with the largest gains on topic-sensitive traits such as word choice and topic adherence.

What carries the argument

The topic-aware soft prompt: a sequence of learnable embeddings split into a topic-shared part and per-topic parts, prepended to the essay before a frozen BERT and optimized to elicit task knowledge. Adversarial training through a gradient reversal layer is confined to the classification branch so topic alignment does not distort regression feature scales; a neighbor-based classifier over a memory bank generates target pseudo-labels that supervise the topic-specific prompt.

What would settle it

Replace the neighbor-based pseudo-labels with random labels for the target topic and retrain: if the topic-specific prompt still improves QWK by the same margin, the pseudo-label signal is not the active ingredient. A more direct check is to measure pseudo-label accuracy against true target-topic labels on ASAP++; on topics where accuracy is near chance, the paper's supervised-learning story would predict a drop in topic-specific traits such as topic adherence.

Watch

Extended reading notes

Core claim

ATOP's central claim is that the bottleneck in cross-topic AES is not a lack of shared features but the neglect of topic-specific ones, and that soft prompts are the right mechanism to provide both. The model freezes BERT and optimizes a topic-aware prompt consisting of a shared prompt plus one prompt per topic; the [CLS] representation of the prompt-augmented essay feeds both a regression scorer for holistic and trait scores and a classifier. Adversarial training between the shared prompt and per-source-target discriminators is applied only in the classification branch, avoiding the feature-scale distortion that pure regression alignment causes. Target-topic essays receive pseudo-labels fro

Load-bearing premise

The pseudo-labels assigned to target-topic essays by the neighbor-based classifier are accurate enough to supervise the topic-specific prompt; if they are noisy or biased, the target prompt learns from the model's own errors with no independent correction.

Editorial extensions

If this is right

  • If ATOP's central claim is right, a frozen pre-trained language model with learnable prompts can outperform full fine-tuning on cross-topic scoring, using only about 0.4% of BERT's trainable parameters.
  • Topic-specific traits such as word choice and topic adherence improve most, meaning the decomposition into shared and specific prompts is what carries the gain on content fidelity.
  • Ablation results attribute about 1.8-1.9 average QWK points to adversarial training and about 1.0-1.1 points to the topic-specific prompt, so both components earn their place in the framework.
  • Performance stays relatively stable across prompt lengths from 2 to 32, so short prompts suffice and the method is not sensitive to this hyperparameter.
  • The pseudo-labeling loop provides a way to adapt the target topic's prompt without any human-labeled target essays, which is the practical setting that motivates cross-topic AES.

Reading between the lines

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

  • A natural testable extension is to apply the shared/specific prompt decomposition to other domain-adaptation regression tasks without target labels, such as scoring open-ended responses in other languages or subjects.
  • Because the pseudo-label loop is self-supervised, adding confidence thresholds or consistency regularization could reduce error accumulation; the paper does not report such a safeguard.
  • The very small trainable parameter count suggests the same two-part prompt could be tested with much larger frozen language models, a regime the paper does not explore.
  • ATOP's known weaker result on Topic 1, an argumentative essay, hints that explicit syntactic features still matter for rigid genres; a follow-up could combine the prompt decomposition with POS-style features rather than raw tokens only.
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

3 major / 5 minor

Summary. The paper proposes ATOP, a prompt-tuning method for cross-topic automated essay scoring. A frozen BERT is guided by a topic-shared soft prompt and per-topic soft prompts; the model jointly optimizes regression and classification losses, uses adversarial training with a gradient reversal layer on the shared prompt, and pseudo-labels target-topic essays via a memory-bank neighbor classifier so that a target-specific prompt can be learned. Experiments on ASAP++ across eight topics report average QWK gains over nine baselines for holistic regression/classification and multi-trait scoring, supported by ablations and a parameter-efficiency comparison.

Significance. If the reported results are valid, the paper would make a useful empirical contribution: it introduces prompt-tuning to cross-topic AES in a parameter-efficient way, releases its code, and its ablation study gives some insight into the role of each component. However, the evaluation protocol in Appendix B selects the best epoch using target-topic test QWK, so the reported numbers are not unbiased estimates of cross-topic generalization. The central empirical claim therefore needs to be re-validated with a proper model-selection protocol before the contribution can be accepted.

major comments (3)
  1. [Appendix B; Tables 1-3, §4.4] The protocol states: "For the multi-trait AES task, we treat all traits as equally important and take the epoch with the highest average QWK for all traits as the best result." Since QWK is computed against the target-topic human scores, this is test-set checkpoint selection. The same issue likely affects the holistic results in Appendix D, although the protocol is not stated for those tasks. This inflates the reported improvements (e.g., the 1.9% gain in §4.4) and makes the comparison to baselines unfair if the baselines were selected under their own published protocols. No error bars or significance tests are given. The authors should re-run with a held-out validation split (e.g., a source-topic split or leave-one-source-out), report mean and standard deviation over multiple seeds, and apply the identical selection protocol to all methods.
  2. [§4.7] The hyperparameter analysis reports QWK for grid searches over α, β, and prompt length, apparently measured on the target-test set. The final settings (α=10, β=1, m=n=8) are then read off these curves. This constitutes another form of test-set peeking. A separate validation split or nested model selection is required before the hyperparameter choices and the resulting test numbers can be interpreted.
  3. [§3.5, Eq. (15)] The pseudo-labeling loop supervises the target-topic prompt with labels produced by the model's own memory bank and neighbor classifier. If these pseudo-labels are noisy, the target prompt can reinforce systematic errors with no independent corrective signal. The ASAP++ dataset contains human labels for all topics, so the paper should quantify pseudo-label quality during development (e.g., accuracy against human labels) and ideally include an ablation with oracle labels. This is not by itself a rejection-level flaw, but it is needed to support the claimed benefit of Eq. (15).
minor comments (5)
  1. [§3, Conclusion] The sentence "Adversarial training between the topic-shared prompt and the topic discriminator is employed to learn the topic-specific prompt" is inconsistent with the method, where adversarial training is used for the topic-shared prompt. This wording appears in the introduction of §3 and in the conclusion; it should be corrected.
  2. [Appendix C vs §4.3] Appendix C says eight baselines, while §4.3 says nine and includes PT-V2. Reconcile the count and list.
  3. [References] References [6] and [18] appear to be the same paper with different author ordering; please merge or clarify why both are cited.
  4. [Eq. (4)] The notation F-j,i in the trait-attention formula is not clearly defined. It should be explicit about which columns/entries are excluded and what the summation index ranges over.
  5. [Table 5] The parameter count for ATOP should specify which components are included (prompts, classifiers, discriminators, trait heads) so the comparison with PT-V2 and Hi-att is meaningful.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline results are test-set fits: the best epoch and hyperparameters are selected on target-test QWK and then reported as ATOP's performance; otherwise the method is an empirical training recipe with no equation-level circularity.

  1. fitted input called prediction [Appendix B (Implementation Details) and Section 4.4 (Results and Analysis)]
    "For the multi-trait AES task, we treat all traits as equally important and take the epoch with the highest average QWK for all traits as the best result. ... our method ATOP improves the average QWK score across different topics by 1.9% compared to the best baseline"

    The epoch is selected by maximizing QWK computed against the target-topic human scores (the test labels). The reported QWK is therefore not an unbiased estimate of ATOP's cross-topic performance; it is the maximum over epochs of the test-set QWK by construction. The 'prediction' that ATOP improves over baselines by 1.9% is a value fitted to the evaluation labels used for model selection, so the headline comparison is statistically forced rather than an independent forecast.

  2. fitted input called prediction [Section 4.7 (Hyperparameter Analysis)]
    "we employ a grid search strategy, sampling α from {1, 5, 10, 50, 100} and β from {0.1, 0.5, 1, 5, 10}. As shown in Figure 3 (a), the model's performance initially improves with increasing α and β, peaking at α = 10 and β = 1"

    The grid search is evaluated on the same target-topic QWK used in the final tables, and the selected α=10, β=1 are then used for the reported results. This is a second fit of model configuration to the test labels. Together with Appendix B's epoch selection, the central empirical claim (ATOP outperforms nine baselines) rests on numbers that are selected, not predicted, on the target test set.

full rationale

ATOP is an empirical training recipe rather than a mathematical derivation, so there is no equation-to-equation circularity in the loss definitions. The pseudo-labeling loop (Section 3.5, Eq. 15-16) is self-referential in that target-topic prompts are supervised by labels generated from the model's own memory-bank representations, but this is a standard self-training mechanism and is not presented as an independent ground-truth signal; it does not by itself make the evaluation circular. No load-bearing self-citation or imported uniqueness theorem appears: the authors' prior work [39] is cited only as one of several related methods. The genuine circularity is in the evaluation protocol: Appendix B selects the epoch with the highest average QWK on the target-topic test labels, and Section 4.7 selects hyperparameters using the same target QWK. The headline gains in Tables 1-3 and the 1.9% improvement claim are therefore max-over-test-selection values rather than unbiased estimates of cross-topic generalization. This is a fitted-input-called-prediction pattern and it is load-bearing because the paper's contribution is entirely empirical. The score is set to 6 rather than higher because the method itself is not derived from the test labels; only the reported performance numbers are inflated by test-set model selection.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central contribution rests on standard domain-adaptation and self-training assumptions plus a set of tuned hyperparameters. No physically invented entities are introduced.

free parameters (6)
  • alpha (regression loss weight) = 10
    Set by grid search over {1,5,10,50,100} in Section 4.7; no validation split is described, so it is tuned on target-topic performance.
  • beta (adversarial loss weight) = 1
    Grid-searched over {0.1,0.5,1,5,10} in Section 4.7; same test-based tuning concern.
  • Prompt lengths m and n = 8
    Chosen from {2,4,8,16,32}; the paper reports performance is relatively insensitive to length.
  • Memory bank temperature tau = unspecified
    Appears in Eq. 13 for sharpening pseudo-label soft labels; no value is given in the main text or appendix.
  • Memory bank EMA factor lambda = unspecified
    Appears in Eq. 14 for exponential moving averaging of the memory bank; no value is given.
  • Number of neighbors m in pseudo-labeling = unspecified
    Used in Eq. 15 for aggregating neighbor soft labels; no value is specified despite sharing a symbol with the prompt length.
assumptions (4)
  • domain assumption A frozen BERT with learned soft prompts can encode both topic-shared and topic-specific essay features sufficient for scoring.
    Central premise of the prompt-tuning approach, introduced in Section 3.2.
  • domain assumption Pseudo-labels generated by the neighbor-based classifier over a memory bank are accurate enough to supervise the target topic-specific prompt.
    Section 3.5, Eq. 15; if violated, the target prompt learns from the model's own noise.
  • domain assumption Adversarial alignment in the classification branch transfers to scoring without harming regression feature scales.
    Section 3.4, cites [4] as acknowledgment that domain alignment can affect regression feature scales.
  • domain assumption The ASAP++ trait annotations and topic splits are a valid benchmark for cross-topic generalization.
    Section 4.1; the paper relies on the benchmark without independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adversarial Topic-aware Prompt-tuning for Cross-topic Automated Essay Scoring." pith.science (2026). https://pith.science/paper/UIWFXXCB

@misc{pith2026250805987,
  author       = {Pith},
  title        = {Pith review of: Adversarial Topic-aware Prompt-tuning for Cross-topic Automated Essay Scoring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UIWFXXCB}},
  note         = {Machine review of arXiv:2508.05987}
}
read the original abstract

Cross-topic automated essay scoring (AES) aims to develop a transferable model capable of effectively evaluating essays on a target topic. A significant challenge in this domain arises from the inherent discrepancies between topics. While existing methods predominantly focus on extracting topic-shared features through distribution alignment of source and target topics, they often neglect topic-specific features, limiting their ability to assess critical traits such as topic adherence. To address this limitation, we propose an Adversarial TOpic-aware Prompt-tuning (ATOP), a novel method that jointly learns topic-shared and topic-specific features to improve cross-topic AES. ATOP achieves this by optimizing a learnable topic-aware prompt--comprising both shared and specific components--to elicit relevant knowledge from pre-trained language models (PLMs). To enhance the robustness of topic-shared prompt learning and mitigate feature scale sensitivity introduced by topic alignment, we incorporate adversarial training within a unified regression and classification framework. In addition, we employ a neighbor-based classifier to model the local structure of essay representations and generate pseudo-labels for target-topic essays. These pseudo-labels are then used to guide the supervised learning of topic-specific prompts tailored to the target topic. Extensive experiments on the publicly available ASAP++ dataset demonstrate that ATOP significantly outperforms existing state-of-the-art methods in both holistic and multi-trait essay scoring. The implementation of our method is publicly available at: https://anonymous.4open.science/r/ATOP-A271.

Figures

Figures reproduced from arXiv: 2508.05987 by the authors.

Figure 1
Figure 1. The framework of ATOP. ATOP optimizes topic-shared and topic-specific prompts within a joint classification and regression framework. 3.2 Topic-aware Prompt Enhanced Essay Encoding In this paper, we introduce prompt-tuning into cross-topic AES by conceptualizing it as a task of eliciting relevant knowledge from PLMs. Unlike existing prompt-tuning methods that only utilize domain-shared prompts [37], we incorporate t… view at source ↗
Figure 2
Figure 2. Visualization of features for Topics 1, 3 and 5. Red represents the features of the source topic, green represents the features of the target topic. without adversarial training module, the feature distributions of the source and target topics become more separated, reflecting a reduc￾tion in cross-topic consistency. Similarly, removing the topic-specific prompts leads to unconstrained learning of the target topic’s… view at source ↗
Figure 3
Figure 3. The impact of hyperparameters on model performance. We investigated the impact of different hyperparameters on model performance, including loss weights and prompt length. α and β are trade-off parameters that balance the weights of the different losses. Given that regression and classification losses differ in magnitude, we employ a grid search strategy, sampling α from {1, 5, 10, 50, 100} and β from {0.1, 0.5, 1, … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: These figures show the visualization results for each topic, respectively. Red represents the features of the source topic, green represents the features of the target topic [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

  1. [1]

    Attali and J

    Y . Attali and J. Burstein. Automated essay scoring with e-rater® v. 2. The Journal of Technology, Learning and Assessment, 4(3), 2006

    work page 2006

  2. [2]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language mod- els are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  3. [3]

    Y . Cao, H. Jin, X. Wan, and Z. Yu. Domain-adaptive neural automated essay scoring. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval, pages 1011–1020, 2020

    work page 2020

  4. [4]

    X. Chen, S. Wang, J. Wang, and M. Long. Representation subspace distance for domain adaptation regression. In ICML, pages 1749–1759, 2021

    work page 2021

  5. [5]

    Chen and X

    Y . Chen and X. Li. Pmaes: Prompt-mapping contrastive learning for cross-prompt automated essay scoring. In Proceedings of the 61st An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1489–1503, 2023

    work page 2023

  6. [6]

    Chen and X

    Y . Chen and X. Li. Plaes: Prompt-generalized and level-aware learn- ing framework for cross-prompt automated essay scoring. In Proceed- ings of the 2024 Joint International Conference on Computational Lin- guistics, Language Resources and Evaluation (LREC-COLING 2024) , pages 12775–12786, 2024

    work page 2024

  7. [7]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Pro- ceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019

    work page 2019

  8. [8]

    H. Do, Y . Kim, and G. G. Lee. Prompt-and trait relation-aware cross- prompt essay trait scoring. In Findings of the Association for Computa- tional Linguistics: ACL 2023, pages 1538–1551, 2023

    work page 2023

Show all 39 references
  1. [9]

    Donahue, Y

    J. Donahue, Y . Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng, and T. Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In International conference on machine learning , pages 647–655. PMLR, 2014

    2014

  2. [10]

    F. Dong, Y . Zhang, and J. Yang. Attention-based recurrent convolutional neural network for automatic essay scoring. In Proceedings of the 21st conference on computational natural language learning (CoNLL 2017), pages 153–162, 2017

    2017

  3. [11]

    Ganin and V

    Y . Ganin and V . Lempitsky. Unsupervised domain adaptation by back- propagation. In International conference on machine learning , pages 1180–1189. PMLR, 2015

    2015

  4. [12]

    M. A. Hussein, H. A. Hassan, and M. Nassef. A trait-based deep learn- ing automated essay scoring system with adaptive feedback. Interna- tional Journal of Advanced Computer Science and Applications, 11(5), 2020

    2020

  5. [13]

    Jiang, T

    Z. Jiang, T. Gao, Y . Yin, M. Liu, H. Yu, Z. Cheng, and Q. Gu. Improving domain generalization for prompt-aware essay scoring via disentangled representation learning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa...

    2023

  6. [14]

    C. Jin, B. He, K. Hui, and L. Sun. Tdnn: a two-stage deep neural net- work for prompt-independent automated essay scoring. In Proceedings of the 56th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 1088–1097, 2018

    2018

  7. [15]

    Ke and V

    Z. Ke and V . Ng. Automated essay scoring: A survey of the state of the art. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, volume 19, pages 6300–6308, 2019

    2019

  8. [16]

    B. B. Klebanov, N. Madnani, and J. Burstein. Using pivot-based para- phrasing and sentiment profiles to improve a subjectivity lexicon for essay data. Transactions of the Association for Computational Linguis- tics, 1:99–110, 2013

    2013

  9. [17]

    Lester, R

    B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Con- ference on Empirical Methods in Natural Language Processing , pages 3045–3059, 2021

    2021

  10. [18]

    Li and Y

    X. Li and Y . Chen. Plaes: Prompt-generalized and level-aware learn- ing framework for cross-prompt automated essay scoring. In Proceed- ings of the 2024 Joint International Conference on Computational Lin- guistics, Language Resources and Evaluation (LREC-COLING 2024) , pages...

    2024

  11. [19]

    X. Li, M. Chen, and J.-Y . Nie. Sednn: Shared and enhanced deep neural network model for cross-prompt automated essay scoring. Knowledge- Based Systems, 210:106491, 2020

    2020

  12. [20]

    X. L. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), ...

    2021

  13. [21]

    Liang, D

    J. Liang, D. Hu, and J. Feng. Domain adaptation with auxiliary target domain-oriented classifier. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16632–16642, 2021

    2021

  14. [22]

    X. Liu, K. Ji, Y . Fu, W. L. Tam, Z. Du, Z. Yang, and J. Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021

    2021 arXiv

  15. [23]

    Mathias and P

    S. Mathias and P. Bhattacharyya. Asap++: Enriching the asap auto- mated essay grading dataset with essay attribute scores. In Proceed- ings of the eleventh international conference on language resources and evaluation (LREC 2018), 2018

    2018

  16. [24]

    Mayfield and A

    E. Mayfield and A. W. Black. Should you fine-tune bert for automated essay scoring? In Proceedings of the Fifteenth Workshop on Innova- tive Use of NLP for Building Educational Applications, pages 151–162, 2020

    2020

  17. [25]

    Ramesh and S

    D. Ramesh and S. K. Sanampudi. An automated essay scoring sys- tems: a systematic literature review. Artificial Intelligence Review, 55 (3):2495–2527, 2022

    2022

  18. [26]

    Ridley, L

    R. Ridley, L. He, X. Dai, S. Huang, and J. Chen. Prompt agnostic essay scorer: a domain generalization approach to cross-prompt automated essay scoring. arXiv preprint arXiv:2008.01441, 2020

    2008 arXiv

  19. [27]

    Ridley, L

    R. Ridley, L. He, X.-y. Dai, S. Huang, and J. Chen. Automated cross- prompt scoring of essay traits. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 13745–13753, 2021

    2021

  20. [28]

    P. U. Rodriguez, A. Jafari, and C. M. Ormerod. Language models and automated essay scoring. arXiv preprint arXiv:1909.09482, 2019

    1909 arXiv

  21. [29]

    Schick and H

    T. Schick and H. Schütze. Exploiting cloze-questions for few-shot text classification and natural language inference. InProceedings of the 16th Conference of the European Chapter of the Association for Computa- tional Linguistics: Main Volume, pages 255–269, 2021

    2021

  22. [30]

    Sethi and K

    A. Sethi and K. Singh. Natural language processing based automated essay scoring with parameter-efficient transformer approach. In 2022 6th International Conference on Computing Methodologies and Com- munication (ICCMC), pages 749–756. IEEE, 2022

    2022

  23. [31]

    Stahl, L

    M. Stahl, L. Biermann, A. Nehring, and H. Wachsmuth. Exploring llm prompting strategies for joint essay scoring and feedback generation. In Proceedings of the 19th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2024), pages 283–298, 2024

    2024

  24. [32]

    Taghipour and H

    K. Taghipour and H. T. Ng. A neural approach to automated essay scoring. In Proceedings of the 2016 conference on empirical methods in natural language processing, pages 1882–1891, 2016

    2016

  25. [33]

    Y . Tay, M. Phan, L. A. Tuan, and S. C. Hui. Skipflow: Incorporating neural coherence features for end-to-end automatic text scoring. InPro- ceedings of the AAAI conference on artificial intelligence , volume 32, 2018

    2018

  26. [34]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  27. [35]

    S. Vajjala. Automated assessment of non-native learner essays: Investi- gating the role of linguistic features. International Journal of Artificial Intelligence in Education, 28:79–105, 2018

    2018

  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, 30, 2017

    2017

  29. [37]

    Wu and X

    H. Wu and X. Shi. Adversarial soft prompt tuning for cross-domain sentiment analysis. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 2438–2447, 2022

    2022

  30. [38]

    R. Yang, J. Cao, Z. Wen, Y . Wu, and X. He. Enhancing automated essay scoring performance via fine-tuning pre-trained language models with combination of regression and ranking. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1560–1569, 2020

    2020

  31. [39]

    Zhang, J

    C. Zhang, J. Deng, X. Dong, H. Zhao, K. Liu, and C. Cui. Pairwise dual-level alignment for cross-prompt automated essay scoring. Expert Systems with Applications, 265:125924, 2025. Table 1. Detailed statistics of the ASAP++ datasets. Cont: Content, Org: Organization, WC: Word ...

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.