Pith. sign in

REVIEW 4 major objections 5 minor 40 references

SynthCTI: LLM-Driven Synthetic CTI Generation to enhance MITRE Technique Mapping

T0 review · 4 major / 5 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 The paper aims to establish that LLM-generated synthetic threat sentences, produced from cluster-derived semantic prompts, consistently improve macro-F1 for MITRE ATT&CK technique mapping and let small classifiers beat larger ones trained…

desk verdict Useful, honest paper on LLM-based data augmentation for MITRE mapping, but the missing random-oversampling control means the headline claim about guided generation is not yet supported. read the letter →

arxiv 2507.16852 v1 pith:423OX5M2 submitted 2025-07-21 cs.CR cs.AIcs.LG

Álvaro Ruiz-Ródenas , Jaime Pujante Sáez , Daniel García-Algora , Mario Rodríguez Béjar , Jorge Blasco , José Luis Hernández-Ramos This is my paper · ORCID
classification cs.CRcs.AIcs.LG
keywords cyberthreatintelligenceMITREATT&CKdataaugmentationsyntheticgenerationlargelanguagemodelstextclassificationclassimbalancemacro-F1
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the main bottleneck in automated mapping of threat descriptions to MITRE ATT&CK techniques is data, not architecture: synthetic examples generated by an LLM can meaningfully improve classifiers. It reports consistent macro-F1 gains across four models and two datasets, with ALBERT rising from 0.3496 to 0.5256 and SecureBERT from 0.4412 to 0.6558 on CTI-to-MITRE. If true, an 11M-parameter model trained with augmented data outperforms 110M-parameter models trained on the original data alone, making accurate CTI mapping cheaper and more practical. The paper also documents that gains are uneven: techniques with fewer than about ten original examples generate noisy, entity-overfit text.

What carries the argument

The load-bearing mechanism is cluster-conditioned prompting. Each technique class is embedded, clustered with HDBSCAN, a density-based clustering algorithm that finds semantic subgroups without a preset cluster count, and each cluster yields a prompt containing the top two representative sentences, LDA topic terms, KeyBERT keyphrases, WordNet-expanded synonyms ranked by embedding similarity and usage frequency, a readability-derived tone label, and average text length. Gemma-3 4B then generates sentences per class, with the count set to raise each class to the dataset average, and those sentences are appended to the training split. The prompt structure carries the argument: it tries to keep generated sentences inside the class's semantic region while forcing lexical variation.

What would settle it

Take a technique class with fewer than ten original examples, generate SynthCTI sentences, train classifiers with and without them, then strip entity names such as 'DarkVishnya' from both training and test text and measure per-class F1 again: if the gain disappears, the improvement is carried by entity memorization rather than by semantic technique labels.

Watch

Extended reading notes

Core claim

On the paper's own terms, SynthCTI is a data augmentation method that balances long-tailed CTI corpora by generating synthetic sentences for underrepresented techniques. The central discovery is that semantically guided LLM generation—prompts built from cluster-derived topics, keyphrases, synonyms, tone, and few-shot examples—improves classifier F1-macro beyond no augmentation and beyond standard methods such as synonym replacement, random swap, MixUp, back-translation, and character noise. The paper interprets this as evidence that high-quality synthetic data can substitute for model capacity: small augmented models exceed larger unaugmented ones, and augmentation accelerates convergence during fine-tuning. It also reports boundary conditions: generation quality degrades sharply when a class has fewer than ten original sentences, with over-represented entities, hallucinated technical details, and semantic drift in exactly the classes the method targets.

Load-bearing premise

The load-bearing premise is that the LLM-generated sentences carry the correct MITRE ATT&CK technique label with enough semantic fidelity to improve training, even though the paper's own analysis finds noticeable label noise, entity overfitting, and hallucinated details in exactly the small classes the method targets.

Editorial extensions

If this is right

  • If the central claim holds, data augmentation becomes a direct substitute for model capacity in CTI classification: small augmented models can beat large unaugmented ones.
  • Augmented training reaches high F1-macro in fewer epochs, so regular model retraining to keep up with new threats costs less.
  • Because gains are concentrated in macro-F1 rather than accuracy, the benefit is specifically better detection of rare techniques.
  • Techniques with fewer than ten examples remain a known failure region, so the next step for the method is to reduce label noise in exactly those classes.

Reading between the lines

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

  • Editorial inference: applying entity normalization before generation—replacing threat-actor and tool names with generic placeholders—could reduce the reported entity overfitting in tiny classes, potentially extending gains to the techniques where SynthCTI currently struggles.
  • Editorial inference: the cluster-then-prompt recipe could transfer to other long-tail expert-labeling tasks, since it only requires sentence embeddings and a generative model.
  • Editorial inference: the per-subtechnique result suggests an adaptive generator that augments less, or filters more aggressively, when a technique covers many subtechniques could improve efficiency.
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 proposes SynthCTI, a data augmentation framework for CTI sentence classification into MITRE ATT&CK techniques. SynthCTI clusters training sentences per technique with HDBSCAN, extracts features (LDA topics, KeyBERT keyphrases, WordNet synonyms, tone, and text type) from each cluster, and builds prompts that guide Gemma-3 to generate synthetic sentences for underrepresented classes. The synthetic data are added to the training split up to the average class size (Gi = max(0, mu - Ni)), and ALBERT, DistilBERT, BERT, and SecureBERT are fine-tuned on the augmented data. The evaluation on CTI-to-MITRE and TRAM reports macro-F1 improvements for all models, with the largest gains on small models (e.g., ALBERT from 0.3496 to 0.5256 on CTI-to-MITRE), and the authors claim that augmented smaller models outperform larger non-augmented models. The paper also includes a comparison with five classic augmentation baselines and a qualitative/quantitative analysis of generated data quality.

Significance. If the reported gains are robust, SynthCTI would provide a practical way to compensate for limited model capacity in CTI classification, which is relevant for privacy-preserving and low-resource deployments. The paper has notable strengths: it uses two public datasets and four classifier families, compares against five augmentation baselines, and includes both quantitative latent-space metrics and qualitative inspection of generated sentences. The central claim, however, depends on isolating the effect of the generation pipeline from the effect of simple class rebalancing; the current experimental design does not include the critical control for that, and the statistical stability of the single-run results is not established. The qualitative analysis itself documents label noise and entity overfitting in the very classes the method targets.

major comments (4)
  1. [Section 4.2, Figures 6-9] The experimental design does not include a random oversampling control. Because Gi = max(0, mu - Ni) pads every underrepresented class to the mean and the evaluation metric is macro-F1, any augmentation that adds minority-class examples will tend to improve the score, even if the generated sentences are duplicates of existing ones. The paper compares SynthCTI against synonym replacement, random swap, MixUp, back-translation, and character noise, but never against random oversampling (duplicating existing minority sentences) or a class-weighted loss. Without this control, the reported relative gains (e.g., ALBERT +50.36% in Section 4.3) do not isolate whether the clustering-guided generation, rather than the act of rebalancing, drives the improvements. The authors should add a random oversampling baseline with the same target counts, and ideally a class-weighted loss, and show that SynthCTI outperforms them.
  2. [Section 4.3, Figures 8-9] No variance or significance information is reported: there are no multiple seeds, error bars, confidence intervals, or significance tests. The improvement on TRAM is roughly 15% in macro-F1, and fine-tuning small transformers is known to be sensitive to random initialization. The 'consistent improvements' claim is therefore not statistically supported as stated. The authors should run at least 3-5 seeds per condition and report means with standard deviations or confidence intervals, and use a paired test (e.g., paired bootstrap or Wilcoxon) to support the claim of consistent gains.
  3. [Section 4.4.1 and 4.4.2] The paper's own quality analysis weakens the premise that the generated data are semantically reliable for the classes the method targets. Section 4.4.1 defines 'strong' and 'weak' classes using thresholds that were 'empirically selected based on the observed distributions' (Section 4.4.1), which is a form of post-hoc threshold fitting, and the qualitative analysis in Section 4.4.2 documents severe entity overfitting (DarkVishnya in 23% of T1200 sentences), technical inaccuracies (McAfee described as antivirus in T1072), and semantic drift (T1189, T1557). Since the augmentation focuses on underrepresented classes and the 'weak' classes are precisely those with few original examples, the reported macro-F1 gains rest on the assumption that the label noise introduced in these classes does not dominate. The missing random oversampling control is essential to determine whether the gains stem from the quantity of added examples rather than their semantic quality.
  4. [Abstract and Section 4.3] The abstract states that ALBERT improves from 0.35 to 0.52 with a relative gain of 48.6%, while Section 4.3 reports an improvement from 0.3496 to 0.5256 with a relative gain of 50.36%. These numbers should be made internally consistent, and the abstract should cite the same values as the body.
minor comments (5)
  1. [Section 4.1] The description of the datasets would benefit from reporting the exact number of classes, total sentences, and the degree of imbalance (e.g., min/median/max class size) for both CTI-to-MITRE and TRAM; this would help readers gauge the effect of the Gi balancing rule.
  2. [Section 4.4.1] The phrase 'predefined thresholds' is misleading because the thresholds (Silhouette >= 0.17, DB < 2.0, cosine >= 0.10, etc.) are described as empirically selected after observing the data. Please report how these thresholds were chosen and whether the qualitative conclusions are sensitive to their exact values.
  3. [Figure 10] Figure 10 is described in the text as comparing accuracy and F1-macro across datasets and models, but the caption and axes are not fully self-explanatory; please specify which panel corresponds to which dataset and which metric, and whether the curves represent single runs or averages.
  4. [Section 3.1.1] The synonym scoring formula uses a parameter alpha whose value is not reported in the experimental setup. Please state the value of alpha used in the evaluations, and ideally show a small sensitivity analysis.
  5. [References] Some references are incomplete or inconsistently formatted (e.g., [15] includes a CoRR arXiv identifier but no publisher; [37] lacks the venue page numbers). A careful reference cleanup would improve the presentation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: synthetic data are generated only from the training split and evaluated on a held-out test split, so the reported F1-macro gains are empirical; the only mild self-reference is post-hoc thresholding in the secondary quality analysis.

full rationale

The central derivation is self-contained. Section 4.1 states that the synthetic data produced by SynthCTI are "added exclusively to the training split," while the held-out test split (20%, stratified) is never used for generation. The main claims, such as ALBERT improving from 0.3496 to 0.5256 macro-F1 on CTI-to-MITRE and SecureBERT reaching 0.6558, are measured out-of-sample and are not forced by construction. The augmentation budget Gi = max(0, mu - Ni) in Section 4.2 is a count of examples to generate, not a fitted label or a prediction, so it does not define the evaluation outcome. The comparison against synonym replacement, random swap, MixUp, back-translation, and character noise is empirical; the absence of a random-oversampling control is a potential confound for the claim that guided generation quality, rather than class rebalancing, drives the gains, but a missing baseline is a correctness/validity concern, not circularity. The authors do not rely on their own prior work: the reference list contains no overlap with the present author list, and no uniqueness theorem or prior ansatz is imported from the authors. The only candidate for circularity is in Section 4.4.1, where the strong/weak quality thresholds are "empirically selected based on the observed distributions of these metrics across all classes in the datasets." This makes the strong/weak categorization post hoc and somewhat self-referential, and the subsequent observation that weak classes tend to have fewer than 10 examples is partly a description of the same data used to set the thresholds. However, this thresholding appears in the auxiliary data-quality analysis, not in the main evaluation, and the correlation with class size is not definitionally entailed. It is a mild, non-load-bearing issue and does not rise to structural circularity.

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

The central method introduces several hand-chosen hyperparameters (alpha, HDBSCAN settings, LDA topic count) and empirically selected thresholds for the quality categorization, none of which are grounded in external benchmarks. The main dependence is on the LLM's fidelity in preserving the label of the source technique, which the paper itself shows is unreliable for classes with fewer than 10 examples.

free parameters (4)
  • alpha in synonym scoring
    Weight in score(w,s) = cos(ew,es) + alpha*f(s) balancing semantic similarity against Zipf frequency; value not reported in the paper (Section 3.1.1).
  • strong/weak class thresholds = Silhouette >= 0.17, DB < 2.0, cosine >= 0.10 for strong; Silhouette < 0.05, DB >= 7.0, cosine <= 0.03 for weak
    Empirically selected based on observed metric distributions in the same datasets (Section 4.4.1); used to categorize augmentation quality and support qualitative conclusions.
  • HDBSCAN hyperparameters
    min_cluster_size and related parameters are not reported; the clustering output directly controls which sentences become few-shot examples (Section 3.1.1).
  • LDA topic number
    Number of topics per cluster is not specified; LDA-derived topics are used as prompt hints (Section 3.1.1).
assumptions (4)
  • domain assumption Ground-truth labels in CTI-to-MITRE and TRAM are correct
    The evaluation treats human-annotated technique labels as ground truth (Section 4.1); any annotation noise propagates into both training and evaluation.
  • domain assumption all-MiniLM-L6-v2 embeddings capture semantic similarity relevant to ATT&CK technique distinctions
    The clustering step relies on these embeddings to define coherent subgroups within each technique (Section 3.1.1).
  • domain assumption Gemma-3 generated sentences inherit the prompt's technique label with sufficient fidelity
    The method adds generated sentences to the training split with the source technique label; the paper's own analysis shows this assumption is violated in low-resource and ambiguous classes (Sections 4.4.1, 4.4.2).
  • standard math F1-macro on the reported stratified 80/20 split is the appropriate measure of improvement
    The paper uses F1-macro to account for class imbalance (Section 4); this is a reasonable evaluation choice but assumes the split is representative and stable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SynthCTI: LLM-Driven Synthetic CTI Generation to enhance MITRE Technique Mapping." pith.science (2026). https://pith.science/paper/423OX5M2

@misc{pith2026250716852,
  author       = {Pith},
  title        = {Pith review of: SynthCTI: LLM-Driven Synthetic CTI Generation to enhance MITRE Technique Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/423OX5M2}},
  note         = {Machine review of arXiv:2507.16852}
}
read the original abstract

Cyber Threat Intelligence (CTI) mining involves extracting structured insights from unstructured threat data, enabling organizations to understand and respond to evolving adversarial behavior. A key task in CTI mining is mapping threat descriptions to MITRE ATT\&CK techniques. However, this process is often performed manually, requiring expert knowledge and substantial effort. Automated approaches face two major challenges: the scarcity of high-quality labeled CTI data and class imbalance, where many techniques have very few examples. While domain-specific Large Language Models (LLMs) such as SecureBERT have shown improved performance, most recent work focuses on model architecture rather than addressing the data limitations. In this work, we present SynthCTI, a data augmentation framework designed to generate high-quality synthetic CTI sentences for underrepresented MITRE ATT\&CK techniques. Our method uses a clustering-based strategy to extract semantic context from training data and guide an LLM in producing synthetic CTI sentences that are lexically diverse and semantically faithful. We evaluate SynthCTI on two publicly available CTI datasets, CTI-to-MITRE and TRAM, using LLMs with different capacity. Incorporating synthetic data leads to consistent macro-F1 improvements: for example, ALBERT improves from 0.35 to 0.52 (a relative gain of 48.6\%), and SecureBERT reaches 0.6558 (up from 0.4412). Notably, smaller models augmented with SynthCTI outperform larger models trained without augmentation, demonstrating the value of data generation methods for building efficient and effective CTI classification systems.

Figures

Figures reproduced from arXiv: 2507.16852 by the authors.

Figure 1
Figure 1. General overview of SynthCTI. dreds of samples, others have only a few due to lim￾ited reporting or their niche application. This imbalance negatively impacts the classifier’s ability to generalize across all techniques. To mitigate this, we propose a data augmentation strategy tailored to enrich sparsely represented techniques with realistic synthetic samples. This strategy is further described below. 3.1.1. Data a… view at source ↗
Figure 2
Figure 2. General view of the Data Augmentation process. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. illustrates a representative example of the prompt generated for label T1006 belonging to the CTI-to-MITRE dataset[5], using the method described above. In particular, the Examples section includes rep￾resentative cluster sentences that mention specific tech￾niques and tools (e.g., "This technique bypasses Win￾dows file access controls..." and "Utilities, such as Nin￾jaCopy..."), offering grounded context to guide g… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Generated text in response to a prompt about file access [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Distribution of MITRE IDs from the CTI-to-MITRE dataset [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the F1-macro obtained for the di [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Comparison of the F1-macro obtained for the di [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Results for CTI-to-MITRE dataset. (a) Accuracy results for TRAM. (b) F1-macro results for TRAM [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Results for TRAM dataset. the CTI domain, synthetic data enables faster conver￾gence, reducing training time while improving perfor￾mance. This is especially relevant for security teams that need to retrain models regularly to keep up with new threats but operate under…
Figure 10
Figure 10. Figure 10: Comparison of accuracy and F1-macro obtained for the di [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Analysis of synthetic data quality and embedding structure. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Relationship between lexical novelty and semantic fidelity [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Classification performance per number of subtechniques. [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 26 canonical work pages

  1. [1]

    Alaeifar, S

    P. Alaeifar, S. Pal, Z. Jadidi, M. Hussain, E. Foo, Current approaches and future directions for cyber threat intelligence sharing: A survey, Journal of Information Security and Applications 83 (2024) 103786

    work page 2024

  2. [2]

    Brown, K

    R. Brown, K. Nickels, Sans 2023 cti survey: Keep- ing up with a changing threat landscape, in: Tech. Rep, SANS Institute, 2023

    work page 2023

  3. [3]

    N. Sun, M. Ding, J. Jiang, W. Xu, X. Mo, Y . Tai, J. Zhang, Cyber threat intelligence mining for proactive cybersecurity defense: A survey and new perspectives, IEEE Communications Surveys & Tutorials 25 (3) (2023) 1748–1774

    work page 2023

  4. [4]

    Della Penna, R

    S. Della Penna, R. Natella, V . Orbinato, L. Par- racino, L. Pianese, Cti-hal: A human-annotated dataset for cyber threat intelligence analysis, arXiv preprint arXiv:2504.05866 (2025)

    arXiv 2025

  5. [5]

    Dessert Lab, GitHub - dessertlab /cti-to-mitre- with-nlp, https://github.com/dessertlab/ cti-to-mitre-with-nlp

    work page

  6. [6]

    Center for Threat-Informed Defense, center-for-threat-informed-defense/tram - GitHub, https://github.com/ center-for-threat-informed-defense/ tram

    work page

  7. [7]

    Aghaei, X

    E. Aghaei, X. Niu, W. Shadid, E. Al-Shaer, Se- curebert: A domain-specific language model for cybersecurity, in: Security and Privacy in Commu- nication Networks: 18th EAI International Con- ference, SecureComm 2022, Virtual Event, Octo- ber 2022, Proceedings, Springer, 2023, pp. 39–56

    work page 2022

  8. [8]

    Orbinato, M

    V . Orbinato, M. Barbaraci, R. Natella, D. Cotro- neo, Automatic mapping of unstructured cyber threat intelligence: an experimental study:(practical experience report), in: 2022 IEEE 33rd International Symposium on Software Reliability Engineering (ISSRE), IEEE, 2022, pp. 181–192. 15

    work page 2022

Show all 40 references
  1. [9]

    L. Li, C. Huang, J. Chen, Automated discovery and mapping att&ck tactics and techniques for un- structured cyber threat intelligence, Computers & Security 140 (2024) 103815

    2024

  2. [10]

    W. You, Y . Park, Cyber-attack technique clas- sification using two-stage trained large language models, arXiv preprint arXiv:2411.18755 (2024)

    2024 arXiv

  3. [11]

    Y . Li, K. Ding, J. Wang, K. Lee, Empowering large language models for textual data augmenta- tion, arXiv preprint arXiv:2404.17642 (2024)

    2024 arXiv

  4. [12]

    H. Dai, Z. Liu, W. Liao, X. Huang, Y . Cao, Z. Wu, L. Zhao, S. Xu, F. Zeng, W. Liu, et al., Aug- gpt: Leveraging chatgpt for text data augmenta- tion, IEEE Transactions on Big Data (2025)

    2025

  5. [13]

    Bayer, T

    M. Bayer, T. Frey, C. Reuter, Multi-level fine- tuning, data augmentation, and few-shot learning for specialized cyber threat intelligence, Comput- ers & Security 134 (2023) 103430

    2023

  6. [14]

    R. J. G. B. Campello, D. Moulavi, A. Zimek, J. Sander, Hierarchical density estimates for data clustering, visualization, and outlier detection, ACM Transactions on Knowledge Discovery from Data 10 (1) (2015) 5:1–5:51. URL https://dl.acm.org/doi/10.1145/ 2733381

    2015

  7. [15]

    Z. Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, R. Soricut, ALBERT: A lite BERT for self-supervised learning of language represen- tations, CoRR abs /1909.11942 (2019). arXiv: 1909.11942. URL http://arxiv.org/abs/1909.11942

    2019 arXiv

  8. [16]

    Ayoade, S

    G. Ayoade, S. Chandra, L. Khan, K. Hamlen, B. Thuraisingham, Automated threat report clas- sification over multi-source data, in: 2018 IEEE 4th International Conference on Collaboration and Internet Computing (CIC), IEEE, 2018, pp. 236– 245

    2018

  9. [17]

    Legoy, M

    V . Legoy, M. Caselli, C. Seifert, A. Peter, Au- tomated retrieval of att&ck tactics and tech- niques for cyber threat reports, arXiv preprint arXiv:2004.14322 (2020)

    2020 arXiv

  10. [18]

    Zhang, M

    B. Zhang, M. Takeuchi, R. Kawahara, S. Asthana, M. M. Hossain, G.-J. Ren, K. Soule, Y . Mai, Y . Zhu, Evaluating large language models with enterprise benchmarks, in: W. Chen, Y . Yang, M. Kachuee, X.-Y . Fu (Eds.), Proceedings of the 2025 Conference of the Nations of the Amer...

    2025

  11. [19]

    S. Y . Feng, V . Gangal, J. Wei, S. Chandar, S. V osoughi, T. Mitamura, E. Hovy, A survey of data augmentation approaches for nlp, in: Find- ings of ACL-IJCNLP, Association for Computa- tional Linguistics, 2021, pp. 968–988

    2021

  12. [20]

    Cuong Nguyen, S

    H. Cuong Nguyen, S. Tariq, M. Baruwal Chhetri, B. Quoc V o, Towards effective identification of at- tack techniques in cyber threat intelligence reports using large language models, in: Companion Pro- ceedings of the ACM on Web Conference 2025, 2025, pp. 942–946

    2025

  13. [21]

    Beltagy, K

    I. Beltagy, K. Lo, A. Cohan, Scibert: A pretrained language model for scientific text, arXiv preprint arXiv:1903.10676 (2019)

    2019 arXiv

  14. [22]

    Kamath, J

    A. Kamath, J. Ferret, S. Pathak, N. Vieillard, R. Merhej, S. Perrin, T. Matejovicova, A. Ramé, M. Rivière, L. Rouillard, et al., Gemma 3 techni- cal report, CoRR abs/2503.19786 (March 2025). URL https://doi.org/10.48550/arXiv. 2503.19786

    2025 doi arXiv

  15. [23]

    W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, M. Zhou, Minilm: deep self-attention distillation for task-agnostic compression of pre-trained trans- formers (2020)

    2020

  16. [24]

    Reimers, I

    N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks, Pro- ceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (2019). URL https://doi.org/10.18653/v1/ D19-1410

    2019 doi

  17. [25]

    Ester, H.-P

    M. Ester, H.-P. Kriegel, J. Sander, X. Xu, A density-based algorithm for discovering clusters in large spatial databases with noise, in: Pro- ceedings of the 2nd International Conference on Knowledge Discovery and Data Mining, 1996, pp. 226–231. URL https://dl.acm.org/doi/10.5...

    1996

  18. [26]

    D. M. Blei, A. Y . Ng, M. I. Jordan, Latent dirichlet allocation, Journal of Machine Learning Research 3 (2003) 993–1022. URL http://www.jmlr.org/papers/v3/ blei03a.html

    2003

  19. [27]

    Grootendorst, Keybert: Minimal keyword extraction with bert

    M. Grootendorst, Keybert: Minimal keyword extraction with bert. (2020). doi:10.5281/zenodo.4461265. URL https://doi.org/10.5281/zenodo. 4461265

    2020 doi

  20. [28]

    G. A. Miller, Wordnet: A lexical database for english, Communications of the ACM 38 (11) (1995) 39–41. doi:10.1145/219717.219748. URL https://dl.acm.org/doi/10.1145/ 219717.219748

    1995

  21. [29]

    Moreno-Sánchez, F

    I. Moreno-Sánchez, F. Font-Clos, Á. Corral, Large-scale analysis of zipf’s law in english texts, PloS one 11 (1) (2016) e0147073

    2016

  22. [30]

    Flesch, A new readability yardstick, Journal of Applied Psychology 32 (3) (1948) 221–233

    R. Flesch, A new readability yardstick, Journal of Applied Psychology 32 (3) (1948) 221–233. doi: 10.1037/h0057532

    1948 doi

  23. [31]

    Gunning, The Technique of Clear Writing, re- vised Edition, McGraw-Hill, New York, 1952, in- troduces the Gunning Fog Index for measuring readability

    R. Gunning, The Technique of Clear Writing, re- vised Edition, McGraw-Hill, New York, 1952, in- troduces the Gunning Fog Index for measuring readability

    1952

  24. [32]

    V . Sanh, L. Debut, J. Chaumond, T. Wolf, Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter, ArXiv abs/1910.01108 (2019)

    2019 arXiv

  25. [33]

    Devlin, M

    J. Devlin, M. Chang, K. Lee, K. Toutanova, BERT: pre-training of deep bidirectional transformers for language understanding, CoRR abs /1810.04805 (2018). arXiv:1810.04805. URL http://arxiv.org/abs/1810.04805

    2018 arXiv

  26. [34]

    J. Wei, K. Zou, EDA: Easy data augmenta- tion techniques for boosting performance on text classification tasks, in: K. Inui, J. Jiang, V . Ng, X. Wan (Eds.), Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint ...

    2019 doi

  27. [35]

    Zhang, M

    H. Zhang, M. Cisse, Y . N. Dauphin, D. Lopez-Paz, mixup: Beyond empirical risk minimization, in: International Conference on Learning Represen- tations (ICLR), 2018. URL https://openreview.net/forum?id= r1Ddp1-Rb

    2018

  28. [36]

    Sennrich, B

    R. Sennrich, B. Haddow, A. Birch, Improving neu- ral machine translation models with monolingual data, in: K. Erk, N. A. Smith (Eds.), Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Pa- pers), Association for Computatio...

    2016 doi

  29. [37]

    Belinkov, Y

    Y . Belinkov, Y . Bisk, Synthetic and natural noise both break neural machine translation, in: International Conference on Learning Represen- tations(ICLR), 2018. URL https://openreview.net/forum?id= BJ8vJebC-

    2018

  30. [38]

    McInnes, J

    L. McInnes, J. Healy, N. Saul, L. Großberger, Umap: Uniform manifold approximation and projection, Journal of Open Source Software 3 (29) (2018) 861. doi:10.21105/joss.00861. URL https://doi.org/10.21105/joss. 00861

    2018 doi

  31. [39]

    Fayyazi, S

    R. Fayyazi, S. J. Yang, On the uses of large lan- guage models to interpret ambiguous cyberattack descriptions, arXiv preprint arXiv:2306.14062 (2023)

    2023 arXiv

  32. [40]

    A. Rege, J. Williams, R. Bleiman, K. Williams, Students’ application of the MITRE ATT&CK ® framework via a real-time cybersecurity exercise, in: Proceedings of the 22nd European Confer- ence on Cyber Warfare and Security (ECCWS), V ol. 22, Academic Conferences International Li...

    2023 doi

Pith tools

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