Pith. sign in

REVIEW 3 major objections 5 minor 96 references

Text-ADBench: Text Anomaly Detection Benchmark based on LLMs Embedding

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper establishes that in text anomaly detection, the choice of embedding and pooling governs performance while deep learning detectors add no advantage over shallow algorithms once LLM embeddings are used.

desk verdict Useful benchmark and released artifacts, but the headline claim about deep detectors is contradicted by the paper's own averaged tables and needs re-scoping. read the letter →

arxiv 2507.12295 v1 pith:UFWCLNCA submitted 2025-07-16 cs.CL cs.AIcs.LG

classification cs.CLcs.AIcs.LG MSC 68T5015A8362H30
keywords textanomalydetectionLLMembeddingsbenchmarklow-rankmatrixcompletionshallowversusdeeppoolingstrategiesAUROCandAUPRCevaluationtwo-stage
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 introduces Text-ADBench, a benchmark for unsupervised text anomaly detection built from 12 embedding models, three pooling strategies, and ten anomaly detectors applied to eight text corpora recast as 14 anomaly-detection datasets. Its central claim is that the embedding, not the detector, determines performance: deep detectors such as AutoEncoder, Deep SVDD, and DPAD show no advantage over shallow algorithms such as KNN, Isolation Forest, OCSVM, LOF, and KDE once LLM embeddings are in place. The paper further argues that the resulting per-dataset AUROC matrices are strongly low-rank, with two singular values carrying over 90% of the energy on most datasets, so unmeasured (embedding, method) entries can be predicted by rank-1 matrix completion. A sympathetic reader would care because, if true, practitioners should spend their effort on embedding selection and pooling rather than on detector sophistication, and can cheaply estimate how a new embedding or dataset will perform.

What carries the argument

The central object is the per-dataset performance matrix P whose rows are embeddings (model times pooling strategy) and whose columns are anomaly detectors, together with its singular value decomposition. The mechanism is the two-stage composition f(s) = h_UAD(Pooling(M_emb(s))) that maps raw text to an anomaly score, first aggregating token embeddings with mean, weighted-mean, or EOS pooling and then applying an unsupervised detector. The load-bearing identity is that P is effectively rank-1, ccr(2) > 0.90 on most datasets, so the rank-constrained completion problem min ||P_Omega(UV^T - P)||$_F^{2}$ with r = 1 predicts unmeasured entries; this is what licenses the paper's practical claim that querying a handful of (embedding, detector) pairs on a new dataset suffices to estimate the full table.

What would settle it

Run the identical two-stage pipeline (same LLM embeddings, same pooling strategies, same ten detectors) on a text corpus with genuine anomaly labels, such as expert-flagged misinformation or real-world failure reports, and check whether shallow detectors still match deep detectors on AUROC and whether LLM embeddings still clearly beat GloVe and BERT; a reversed ranking or a collapsed embedding gap would refute the claim that embedding quality alone governs performance. Separately, compute the cumulative contribution ratio ccr(2) on the performance matrix of a new independent dataset, since a value well below 0.90 would show that the rank-1 completion assumption is not universal.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that two-stage text anomaly detection — embed texts with a pretrained language model, pool token embeddings, then run an unsupervised detector — is governed primarily by the quality of the embedding and the pooling strategy. Across the benchmark, the best configurations all use LLM-derived embeddings, and end-of-sequence (EOS) pooling of LLM2Vec-fine-tuned models is the consistently strong choice. The decisive empirical finding is that standard shallow detectors match or beat deep detectors in this setting: the best AUROC of KNN, Isolation Forest, OCSVM, LOF, KDE, and PCA is not improved upon by AutoEncoder, Deep SVDD, or DPAD. A second structural claim is that the embedding-by-method AUROC matrix on each dataset is nearly rank-1, with cumulative contribution ratio ccr(2) exceeding 0.90 on most datasets, and matrix completion with rank 1 recovers missing entries with mean absolute percentage error mostly below 0.1, enabling rapid evaluation of new embeddings or methods from a few measured combinations.

Load-bearing premise

Every Text-AD dataset in the benchmark is made by declaring one class of a text-classification corpus (for example, all negative IMDB reviews) to be the anomaly class, so the results hold only if class or topic separation behaves like the fraud, spam, and misinformation anomalies the paper motivates.

Editorial extensions

If this is right

  • Practical text anomaly detection pipelines should spend compute on better embeddings and pooling rather than on deep detectors, since KNN, IForest, OCSVM, LOF, and KDE suffice once embeddings are strong.
  • EOS pooling of supervised-LLM2Vec-tuned models is a strong default, while mean and weighted-mean pooling are nearly interchangeable and usually worse.
  • The low-rank structure means a new embedding or a new dataset can be evaluated by running only a subset of (embedding, method) combinations and completing the rest, making model selection substantially cheaper.
  • The open-sourced benchmark resources let a new embedding be slotted in and compared without re-running every detector, supporting reproducible comparison across future models.
  • LLM embeddings give clear AUROC gains over GloVe and BERT on topic- and sentiment-separated datasets, supporting them as the representation of choice in two-stage text anomaly detection.

Reading between the lines

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

  • Because the benchmark declares one classification class anomalous, its anomalies are semantically or topically distinct; the shallow-vs-deep conclusion may not transfer to subtle real-world anomalies such as tone shifts, deception, or rare factual errors, and a test would be to run the same ten detectors on a corpus with genuine anomaly labels.
  • The near-rank-1 structure suggests a single latent factor, likely embedding quality, dominates the performance matrix, but the same low rank could also arise from correlated detector behaviors; distinguishing those would tell practitioners whether completion is predicting embedding ranking or method choice.
  • The high anomaly ratios in datasets like IMDB and SST2 make AUROC a near-balanced-class measure, so the large spread between embedding models on those datasets is itself a practical signal about which representation captures class structure.
  • A testable extension is that the EOS-pooling advantage may carry over to newer decoder-based embedders, and rank-1 completion could be used to pick the best embedding for a new dataset from a few probe runs.
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. This paper presents Text-ADBench, an empirical benchmark for embedding-based text anomaly detection. It evaluates ten two-stage detectors (OCSVM, IForest, LOF, PCA, KNN, KDE, ECOD, AE, DSVDD, DPAD) plus two text-specific methods (CVDD, DATE) on twelve text-AD datasets derived from eight classification corpora, under 33 embedding configurations spanning GloVe, BERT, LLM2Vec-fine-tuned LLaMA-2/LLaMA-3/Mistral with three pooling strategies, and three OpenAI embedding models, reporting AUROC and AUPRC. The paper's central claims are (i) that deep detectors show no advantage over shallow algorithms given LLM-derived embeddings, and (ii) that the cross-embedding performance matrices are strongly low-rank (ccr(2) > 0.90 on most datasets), enabling rank-1 matrix completion to predict unmeasured (embedding, method) entries with MAPE mostly below 0.1. All embeddings, datasets, and code are released.

Significance. The benchmark is a genuine community resource: it is broader than prior benchmarks (NLP-ADBench, TAD-Bench) in the number of embedding configurations, it is the first to systematically vary pooling strategies, and it ships all precomputed embeddings and code, which directly supports reproducibility and downstream use. The low-rank finding is practically interesting and is accompanied by a concrete speed demonstration (1.88 s to predict the 20Newsgroups matrix at 50% missing versus 5099.50 s to evaluate it). However, the headline deep-versus-shallow conclusion is contradicted by the paper's own Table 6 mean rows as currently worded, and the dataset construction limits external validity to class-separability tasks; both issues require re-scoping before the claims can be accepted.

major comments (3)
  1. [Abstract; §5 (Table 6); §7] The abstract and Section 7 state that 'deep learning based detectors (AE, DSVDD, DPAD) exhibit no advantage over conventional shallow algorithms (OCSVM, IForest, KNN, LOF, KDE) when using LLM-derived embeddings.' This is contradicted by Table 6's own mean rows: AE (mean) = 78.88 AUROC is the highest of all ten methods, exceeding LOF (78.28), KNN (77.74), and OCSVM (74.69). The related statement in Section 5 that 'the average performance of KNN outperforms all others methods' is likewise inconsistent with the same table (KNN mean 77.74 < AE mean 78.88) and appears to refer to the 'best' column without saying so. The claim is only supportable under oracle embedding selection (the 'best' column), and even there AE (89.45) is within 1.3 points of KNN (90.71) and above KDE (89.31). The abstract and conclusion must be re-scoped to the oracle-selection version and must report the mean-row outcome as a caveat; as written, the headline claim is internally contradicted by the paper's own results.
  2. [§4.1 (Table 2)] All twelve Text-AD datasets are built by declaring one class of a classification corpus to be the anomaly class and the remaining classes normal (e.g., IMDB: every negative review is an anomaly at a 62.5% anomaly ratio; DBpedia: one Wikipedia topic is normal against four others, where several configurations exceed 99 AUROC). The tasks are therefore dominated by topic or sentiment separability, whereas the applications motivating the paper (fraud, misinformation, spam, content moderation, Section 1) involve anomalies that are typically not class-separable. Prior benchmarks use the same construction, and the paper says so, but the consequence is never tested: no dataset with genuine anomaly labels is included. This affects the transferability of both the detector ranking and the 'embedding quality governs efficacy' conclusion. The paper should prominently scope the conclusions to class-separability benchmarks, or add at least one dataset whose anomalies are not defined by class membership.
  3. [§6 (Eq. 5, Table 7)] The contribution statements (Introduction, Section 6, Conclusion) claim that the low-rank property enables reliable prediction of the detection performance of 'novel text datasets or AD methods' from a subset of measurements. What is actually tested is rank-1 matrix completion of randomly missing entries within the twelve already-measured matrices under MCAR rates 0.5-0.7. This validates within-matrix recovery, not the extrapolation to an unmeasured embedding, method, or dataset; for a genuinely new row or column the rank-1 model would require at least some observations of that row or column to estimate its factor, and the paper provides no leave-one-row-out or leave-one-column-out experiment. The claim should be re-scoped to missing-entry recovery within the benchmark, or the extrapolation experiment should be added.
minor comments (5)
  1. [§1] The attributions for LSA and LDA appear swapped: Latent Semantic Analysis is Dumais [2004] and Latent Dirichlet Allocation is Blei et al. [2003].
  2. [§4.1] The text says the benchmark constructs '14 specialized Text-AD datasets', but Table 2 lists 12 rows (7 single-corpus datasets plus 5 DBpedia splits); please reconcile the count.
  3. [§4.3, §5, §6, Table 7] Several typos: 'an anomaly score is tailed for sequential data' should read 'tailored'; 'Futhermore' should be 'Furthermore'; 'Thant means' should be 'That means'; 'statistic, context-free embeddings' should be 'static, context-free embeddings'; and the Table 7 column header 'Enton' should be 'Enron'.
  4. [§6 (Eq. 6)] MAPE is sensitive to entries with small denominators, and several AUROC values are in the 30-50 range (e.g., ECOD on IMDB); reporting MAE or RMSE alongside MAPE would make the recovery errors in Table 7 easier to interpret.
  5. [§4.4] The paper reports averages over five runs but no standard deviations; since the mean-row comparison that will now carry the re-scoped claim (AE 78.88 vs LOF 78.28 vs KNN 77.74) involves differences of about one point, reporting variances or significance tests for the mean rows would materially strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark study with held-out matrix-completion evaluation; self-citations are not load-bearing.

full rationale

This is an empirical benchmarking paper with no derivation chain that reduces to its own inputs. The main analytical claim is the low-rank structure of AUROC matrices (Section 6), established by SVD cumulative contribution ratios on measured performance matrices, and then tested by MCAR matrix completion with missing rates 0.5-0.7; the recovered entries are compared against withheld ground-truth entries, so the 'prediction' is not the fit by construction. The shallow-versus-deep conclusion (Section 7, Table 6) is an empirical summary of measured AUROC values; even if the 'no advantage' wording is not fully supported by the mean rows (AE mean 78.88 appears above KNN mean 77.74), an internal inconsistency or over-generalization is a correctness concern, not circularity. Self-citations to DPAD (Fu et al. 2024) and to Fan et al. 2019 for matrix recovery are citations of methods or standard tools that are evaluated or used, not unverified premises invoked to force the conclusion. Dataset construction by declaring one classification class anomalous is an external-validity limitation, not a circular step. No step was found where an output equals an input by definition or where a fitted parameter is renamed as a prediction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The benchmark postulates no new physical or theoretical entities; the only constructs are the empirical performance matrices, SVD components, and matrix completion outputs, all computed from the collected data. The load-bearing inputs are the dataset construction choices (classification classes relabeled as anomalies, Section 4.1), the two-stage pipeline decomposition (Eq. 2), contaminated training sets, and four hand-chosen hyperparameters.

free parameters (4)
  • KNN neighborhood size k = 3
    Hand-chosen in Section 4.3 and used for every embedding and dataset; no sensitivity analysis is given, and KNN is the method the paper highlights as best on average, so the choice matters.
  • LOF number of neighbors = 30
    Hand-chosen in Section 4.3; LOF ranks high in Table 6 (best 88.23), and no sensitivity analysis is reported.
  • Matrix completion rank r = 1
    Section 6 sets r = 1 in Eq. (5) for the recovery experiments, while the low-rank evidence is presented as ccr(2) > 0.90; the reported MAPE values depend on this rank choice.
  • Weighted-mean pooling weights = not specified in paper
    Table 3 lists 'weighted mean' as a pooling strategy for all nine LLaMA and Mistral variants (27 of 33 embedding configurations), but the weighting scheme is never defined, so the embeddings cannot be exactly reproduced from the paper's description.
assumptions (5)
  • domain assumption Re-labeling one class of a classification dataset as 'anomalous' and the remaining classes as 'normal' produces valid text anomaly detection tasks.
    Section 4.1 constructs all Text-AD datasets this way (e.g., IMDB negative reviews are anomalous at a 62.5% anomaly ratio; DBpedia uses one Wikipedia topic as normal against four others). The resulting anomalies are topic or sentiment differences, which are far easier to separate than the subtle fraud and misinformation the introduction motivates.
  • domain assumption The unsupervised training splits may contain anomalous samples, and the reported anomaly ratios in Table 2 apply to those training splits.
    Table 2 reports anomaly ratios for the training sets (e.g., 20Newsgroups training set has a 0.12 anomaly ratio). The paper neither removes anomalies from training nor analyzes the effect of contamination on detector rankings.
  • domain assumption The two-stage decomposition f(s) = h_UAD(Pooling(M_emb(s))) fully represents the text anomaly detection problem.
    Section 3, Eq. (2): the entire benchmark evaluates this decomposition; end-to-end methods (DATE, CVDD) are acknowledged but evaluated only partially (DATE cannot consume embeddings, CVDD runs only on GloVe and BERT), so the benchmark does not fully cover them.
  • standard math SVD and low-rank matrix completion theory (Candes and Recht 2012) apply to the empirical AUROC matrices.
    Section 6 relies on SVD and the rank-constrained problem in Eqs. (3)-(5); this is standard mathematics, with empirical validity tested by MCAR-held-out completion.
  • standard math The Nemenyi post-hoc test assumptions are satisfied by the method rankings in Figure 4.
    Figure 4 reports Nemenyi p-values to compare AD methods; the test is invoked as standard statistics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text-ADBench: Text Anomaly Detection Benchmark based on LLMs Embedding." pith.science (2026). https://pith.science/paper/UFWCLNCA

@misc{pith2026250712295,
  author       = {Pith},
  title        = {Pith review of: Text-ADBench: Text Anomaly Detection Benchmark based on LLMs Embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UFWCLNCA}},
  note         = {Machine review of arXiv:2507.12295}
}
read the original abstract

Text anomaly detection is a critical task in natural language processing (NLP), with applications spanning fraud detection, misinformation identification, spam detection and content moderation, etc. Despite significant advances in large language models (LLMs) and anomaly detection algorithms, the absence of standardized and comprehensive benchmarks for evaluating the existing anomaly detection methods on text data limits rigorous comparison and development of innovative approaches. This work performs a comprehensive empirical study and introduces a benchmark for text anomaly detection, leveraging embeddings from diverse pre-trained language models across a wide array of text datasets. Our work systematically evaluates the effectiveness of embedding-based text anomaly detection by incorporating (1) early language models (GloVe, BERT); (2) multiple LLMs (LLaMa-2, LLama-3, Mistral, OpenAI (small, ada, large)); (3) multi-domain text datasets (news, social media, scientific publications); (4) comprehensive evaluation metrics (AUROC, AUPRC). Our experiments reveal a critical empirical insight: embedding quality significantly governs anomaly detection efficacy, and deep learning-based approaches demonstrate no performance advantage over conventional shallow algorithms (e.g., KNN, Isolation Forest) when leveraging LLM-derived embeddings.In addition, we observe strongly low-rank characteristics in cross-model performance matrices, which enables an efficient strategy for rapid model evaluation (or embedding evaluation) and selection in practical applications. Furthermore, by open-sourcing our benchmark toolkit that includes all embeddings from different models and code at https://github.com/jicongfan/Text-Anomaly-Detection-Benchmark, this work provides a foundation for future research in robust and scalable text anomaly detection systems.

Figures

Figures reproduced from arXiv: 2507.12295 by the authors.

Figure 1
Figure 1. Flowchart of Text-ADBench. This work serves as a foundational resource for both researchers and practitioners, advancing the text AD field through its methodological rigor and holistic evaluation. By open-sourcing our benchmark framework, precomputed embeddings, we aim to catalyze further research into hybrid pooling strategies, domain adaptation, and efficient LLM utilization for text anomaly detection. The remaind… view at source ↗
Figure 2
Figure 2. The average AUROC(%) across all datasets. The “unsup.” and “super.” correspond to the embedding models [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison of different embeddings from LLMs. Note that “llama2-1, llama3-1, mistral-1” [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance comparison of different AD methods. In plot (a), a pair of methods with p-value > 0.05 indicates [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: The heatmap of AUROC(%) of each two-stage AD method with different text embeddings. Note that [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: The cumulation contribution ratio of singular value of detection performance (AUROC) across all datasets. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The visualization of AUROC performance prediction on the 20Newsgroups dataset, where the missing rate is [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

96 extracted references · 51 canonical work pages

  1. [1]

    Anomaly detection: A survey

    Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomaly detection: A survey. ACM Comput. Surv., 41 0 (3), July 2009. ISSN 0360-0300. doi:10.1145/1541880.1541882. URL https://doi.org/10.1145/1541880.1541882

  2. [2]

    An introduction to outlier analysis

    Charu C Aggarwal. An introduction to outlier analysis. In Outlier analysis, pages 1--34. Springer, 2016

  3. [3]

    A unifying review of deep and shallow anomaly detection

    Lukas Ruff, Jacob R Kauffmann, Robert A Vandermeulen, Gr \'e goire Montavon, Wojciech Samek, Marius Kloft, Thomas G Dietterich, and Klaus-Robert M \"u ller. A unifying review of deep and shallow anomaly detection. Proceedings of the IEEE, 109 0 (5): 0 756--795, 2021

  4. [4]

    Machine learning for medical diagnosis: history, state of the art and perspective

    Igor Kononenko. Machine learning for medical diagnosis: history, state of the art and perspective. Artificial Intelligence in medicine, 23 0 (1): 0 89--109, 2001

  5. [5]

    Artificial neural networks in medical diagnosis, 2013

    Filippo Amato, Alberto L \'o pez, Eladia Mar \' a Pe \ n a-M \'e ndez, Petr Va n hara, Ale s Hampl, and Josef Havel. Artificial neural networks in medical diagnosis, 2013

  6. [6]

    Deep learning and medical diagnosis: A review of literature

    Mihalj Bakator and Dragica Radosav. Deep learning and medical diagnosis: A review of literature. Multimodal Technologies and Interaction, 2 0 (3): 0 47, 2018

  7. [7]

    Improving the accuracy of medical diagnosis with causal machine learning

    Jonathan G Richens, Ciar \'a n M Lee, and Saurabh Johri. Improving the accuracy of medical diagnosis with causal machine learning. Nature communications, 11 0 (1): 0 3923, 2020

  8. [8]

    Network intrusion detection

    Biswanath Mukherjee, L Todd Heberlein, and Karl N Levitt. Network intrusion detection. IEEE network, 8 0 (3): 0 26--41, 2002

Show all 96 references
  1. [9]

    Intrusion detection by machine learning: A review

    Chih-Fong Tsai, Yu-Feng Hsu, Chia-Ying Lin, and Wei-Yang Lin. Intrusion detection by machine learning: A review. expert systems with applications, 36 0 (10): 0 11994--12000, 2009

  2. [10]

    Intrusion detection system: A comprehensive review

    Hung-Jen Liao, Chun-Hung Richard Lin, Ying-Chih Lin, and Kuang-Yuan Tung. Intrusion detection system: A comprehensive review. Journal of network and computer applications, 36 0 (1): 0 16--24, 2013

  3. [11]

    Survey of intrusion detection systems: techniques, datasets and challenges

    Ansam Khraisat, Iqbal Gondal, Peter Vamplew, and Joarder Kamruzzaman. Survey of intrusion detection systems: techniques, datasets and challenges. Cybersecurity, 2 0 (1): 0 1--22, 2019

  4. [12]

    Statistical fraud detection: A review

    Richard J Bolton and David J Hand. Statistical fraud detection: A review. Statistical science, 17 0 (3): 0 235--255, 2002

  5. [13]

    Fraud detection system: A survey

    Aisha Abdallah, Mohd Aizaini Maarof, and Anazida Zainal. Fraud detection system: A survey. Journal of Network and Computer Applications, 68: 0 90--113, 2016

  6. [14]

    Financial fraud detection using graph neural networks: A systematic review

    Soroor Motie and Bijan Raahemi. Financial fraud detection using graph neural networks: A systematic review. Expert Systems with Applications, 240: 0 122156, 2024

  7. [15]

    Model-based fault-detection and diagnosis--status and applications

    Rolf Isermann. Model-based fault-detection and diagnosis--status and applications. Annual Reviews in control, 29 0 (1): 0 71--85, 2005

  8. [16]

    Autoencoder based high-dimensional data fault detection system

    Jicong Fan, Wei Wang, and Haijun Zhang. Autoencoder based high-dimensional data fault detection system. In 2017 ieee 15th international conference on industrial informatics (indin), pages 1001--1006. IEEE, 2017

  9. [17]

    Jicong Fan, Tommy W. S. Chow, and S. Joe Qin. Kernel-based statistical process monitoring and fault detection in the presence of missing data. IEEE Transactions on Industrial Informatics, 18 0 (7): 0 4477--4487, 2022

  10. [18]

    Webb, Shirui Pan, Charu Aggarwal, and Mahsa Salehi

    Zahra Zamanzadeh Darban, Geoffrey I. Webb, Shirui Pan, Charu Aggarwal, and Mahsa Salehi. Deep learning for time series anomaly detection: A survey. ACM Comput. Surv., 57 0 (1), October 2024. ISSN 0360-0300. doi:10.1145/3691338. URL https://doi.org/10.1145/3691338

  11. [19]

    Deep learning for medical anomaly detection – a survey

    Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection – a survey. ACM Comput. Surv., 54 0 (7), July 2021. ISSN 0360-0300. doi:10.1145/3464423. URL https://doi.org/10.1145/3464423

  12. [20]

    Jing Liu, Yang Liu, Jieyu Lin, Jielin Li, Liang Cao, Peng Sun, Bo Hu, Liang Song, Azzedine Boukerche, and Victor C.M. Leung. Networking systems for video anomaly detection: A tutorial and survey. ACM Comput. Surv., 57 0 (10), May 2025. ISSN 0360-0300. doi:10.1145/3729222. URL ...

  13. [21]

    Distributional structure

    Zellig S Harris. Distributional structure. Word, 10 0 (2-3): 0 146--162, 1954

  14. [22]

    A statistical interpretation of term specificity and its application in retrieval

    Karen Sparck Jones. A statistical interpretation of term specificity and its application in retrieval. Journal of documentation, 28 0 (1): 0 11--21, 1972

  15. [23]

    An information-theoretic perspective of tf--idf measures

    Akiko Aizawa. An information-theoretic perspective of tf--idf measures. Information Processing & Management, 39 0 (1): 0 45--65, 2003

  16. [24]

    Latent dirichlet allocation

    David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. Journal of machine Learning research, 3 0 (Jan): 0 993--1022, 2003

  17. [25]

    Latent semantic analysis

    Susan T Dumais. Latent semantic analysis. Annual Review of Information Science and Technology (ARIST), 38: 0 189--230, 2004

  18. [26]

    Efficient estimation of word representations in vector space

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013

  19. [27]

    G lo V e: Global vectors for word representation

    Jeffrey Pennington, Richard Socher, and Christopher Manning. G lo V e: Global vectors for word representation. In Alessandro Moschitti, Bo Pang, and Walter Daelemans, editors, Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing ( EMNLP ) , pa...

  20. [28]

    Bag of tricks for efficient text classification

    Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016

  21. [29]

    Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer

    Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations, 2018. URL https://arxiv.org/abs/1802.05365

  22. [30]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human langu...

  23. [31]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018

  24. [32]

    A comprehensive overview of large language models

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435, 2023

  25. [33]

    A review of current trends, techniques, and challenges in large language models (llms)

    Rajvardhan Patil and Venkat Gudivada. A review of current trends, techniques, and challenges in large language models (llms). Applied Sciences, 14 0 (5): 0 2074, 2024

  26. [34]

    Support vector method for novelty detection

    Bernhard Sch \"o lkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999

  27. [35]

    Lof: identifying density-based local outliers

    Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and J \"o rg Sander. Lof: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 93--104, 2000

  28. [36]

    Isolation forest

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 eighth ieee international conference on data mining, pages 413--422. IEEE, 2008

  29. [37]

    Deep one-class classification

    Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel M \"u ller, and Marius Kloft. Deep one-class classification. In International conference on machine learning, pages 4393--4402. PMLR, 2018

  30. [38]

    Dense projection for anomaly detection

    Dazhi Fu, Zhao Zhang, and Jicong Fan. Dense projection for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 8398--8408, 2024

  31. [39]

    Self-attentive, multi-context one-class classification for unsupervised anomaly detection on text

    Lukas Ruff, Yury Zemlyanskiy, Robert Vandermeulen, Thomas Schnake, and Marius Kloft. Self-attentive, multi-context one-class classification for unsupervised anomaly detection on text. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, p...

  32. [40]

    Date: Detecting anomalies in text via self-supervision of transformers

    Andrei Manolache, Florin Brad, and Elena Burceanu. Date: Detecting anomalies in text via self-supervision of transformers. arXiv preprint arXiv:2104.05591, 2021

  33. [41]

    Nlp-adbench: Nlp anomaly detection benchmark

    Yuangang Li, Jiaqi Li, Zhuo Xiao, Tiankai Yang, Yi Nian, Xiyang Hu, and Yue Zhao. Nlp-adbench: Nlp anomaly detection benchmark. arXiv preprint arXiv:2412.04784, 2024 a

  34. [42]

    Tad-bench: A comprehensive benchmark for embedding-based text anomaly detection

    Yang Cao, Sikun Yang, Chen Li, Haolong Xiang, Lianyong Qi, Bo Liu, Rongsheng Li, and Ming Liu. Tad-bench: A comprehensive benchmark for embedding-based text anomaly detection. arXiv preprint arXiv:2501.11960, 2025

  35. [43]

    Image anomaly detection with generative adversarial networks

    Lucas Deecke, Robert Vandermeulen, Lukas Ruff, Stephan Mandt, and Marius Kloft. Image anomaly detection with generative adversarial networks. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2018, Dublin, Ireland, September 10--14, 2018,...

  36. [44]

    Neural transformation learning for deep anomaly detection beyond images

    Chen Qiu, Timo Pfrommer, Marius Kloft, Stephan Mandt, and Maja Rudolph. Neural transformation learning for deep anomaly detection beyond images. In International conference on machine learning, pages 8703--8714. PMLR, 2021

  37. [45]

    Unsupervised anomaly detection using inverse generative adversarial networks

    Feng Xiao, Jianfeng Zhou, Kunpeng Han, Haoyuan Hu, and Jicong Fan. Unsupervised anomaly detection using inverse generative adversarial networks. Information Sciences, 689: 0 121435, 2025 a

  38. [46]

    Auto UAD : Hyper-parameter optimization for unsupervised anomaly detection

    Wei Dai and Jicong Fan. Auto UAD : Hyper-parameter optimization for unsupervised anomaly detection. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=ErQPdaD5wJ

  39. [47]

    Deep learning for anomaly detection: A review

    Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR), 54 0 (2): 0 1--38, 2021

  40. [48]

    Ecod: Unsupervised outlier detection using empirical cumulative distribution functions

    Zheng Li, Yue Zhao, Xiyang Hu, Nicola Botta, Cezar Ionescu, and George H Chen. Ecod: Unsupervised outlier detection using empirical cumulative distribution functions. IEEE Transactions on Knowledge and Data Engineering, 35 0 (12): 0 12181--12193, 2022

  41. [49]

    Efficient algorithms for mining outliers from large data sets

    Sridhar Ramaswamy, Rajeev Rastogi, and Kyuseok Shim. Efficient algorithms for mining outliers from large data sets. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 427--438, 2000

  42. [50]

    Unsupervised anomaly detection with generative adversarial networks to guide marker discovery

    Thomas Schlegl, Philipp Seeb \"o ck, Sebastian M Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In International conference on information processing in medical imaging, pages 1...

  43. [51]

    Deep autoencoding gaussian mixture model for unsupervised anomaly detection

    Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International conference on learning representations, 2018

  44. [52]

    Deep anomaly detection with deviation networks

    Guansong Pang, Chunhua Shen, and Anton Van Den Hengel. Deep anomaly detection with deviation networks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 353--362, 2019

  45. [53]

    Ocgan: One-class novelty detection using gans with constrained latent representations

    Pramuditha Perera, Ramesh Nallapati, and Bing Xiang. Ocgan: One-class novelty detection using gans with constrained latent representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2898--2906, 2019

  46. [54]

    Drocc: Deep robust one-class classification

    Sachin Goyal, Aditi Raghunathan, Moksh Jain, Harsha Vardhan Simhadri, and Prateek Jain. Drocc: Deep robust one-class classification. In International conference on machine learning, pages 3711--3721. PMLR, 2020

  47. [55]

    Perturbation learning based anomaly detection

    Jinyu Cai and Jicong Fan. Perturbation learning based anomaly detection. Advances in Neural Information Processing Systems, 35: 0 14317--14330, 2022

  48. [56]

    Deep isolation forest for anomaly detection

    Hongzuo Xu, Guansong Pang, Yijie Wang, and Yongjun Wang. Deep isolation forest for anomaly detection. IEEE Transactions on Knowledge and Data Engineering, 35 0 (12): 0 12591--12604, 2023 a

  49. [57]

    Drl: Decomposed representation learning for tabular anomaly detection

    Hangting Ye, He Zhao, Wei Fan, Mingyuan Zhou, Dan dan Guo, and Yi Chang. Drl: Decomposed representation learning for tabular anomaly detection. In The Thirteenth International Conference on Learning Representations, 2025

  50. [58]

    Reducing the dimensionality of data with neural networks

    Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313 0 (5786): 0 504--507, 2006

  51. [59]

    Deep anomaly detection with outlier exposure

    Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606, 2018

  52. [60]

    Learning representations of ultrahigh-dimensional data for random distance-based outlier detection

    Guansong Pang, Longbing Cao, Ling Chen, and Huan Liu. Learning representations of ultrahigh-dimensional data for random distance-based outlier detection. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2041--2050, 2018

  53. [61]

    Ganomaly: Semi-supervised anomaly detection via adversarial training

    Samet Akcay, Amir Atapour-Abarghouei, and Toby P Breckon. Ganomaly: Semi-supervised anomaly detection via adversarial training. In Computer Vision--ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2--6, 2018, Revised Selected Papers, Part III 14,...

  54. [62]

    o rnitz, Alexander Binder, Emmanuel M \

    Lukas Ruff, Robert A Vandermeulen, Nico G \"o rnitz, Alexander Binder, Emmanuel M \"u ller, Klaus-Robert M \"u ller, and Marius Kloft. Deep semi-supervised anomaly detection. In Proceedings of the International Conference on Learning Representations, 2020

  55. [63]

    Feature encoding with autoencoders for weakly supervised anomaly detection

    Yingjie Zhou, Xucheng Song, Yanru Zhang, Fanxing Liu, Ce Zhu, and Lingqiao Liu. Feature encoding with autoencoders for weakly supervised anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 33 0 (6): 0 2454--2465, 2021

  56. [64]

    Deep weakly-supervised anomaly detection

    Guansong Pang, Chunhua Shen, Huidong Jin, and Anton van den Hengel. Deep weakly-supervised anomaly detection. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1795--1807, 2023

  57. [65]

    Semi-supervised anomaly detection using restricted distribution transformation

    Feng Xiao, Youqing Wang, S Joe Qin, and Jicong Fan. Semi-supervised anomaly detection using restricted distribution transformation. IEEE Transactions on Neural Networks and Learning Systems, 2025 b

  58. [66]

    Glove: Global vectors for word representation

    Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532--1543, 2014 b

  59. [67]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

  60. [68]

    Improving text embeddings with large language models

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023

  61. [69]

    Llm2vec: Large language models are secretly powerful text encoders

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961, 2024 a

  62. [70]

    Comparative analysis of anomaly detection algorithms in text data

    Yizhou Xu, J \'e r \^o me Milleret, and Fr \'e d \'e rique Segond. Comparative analysis of anomaly detection algorithms in text data. In Proceedings of the 14th International Conference on Recent Advances in Natural Language Processing, pages 1234--1245, 2023 b

  63. [71]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019

  64. [72]

    Few-shot anomaly detection in text with deviation learning

    Anindya Sundar Das, Aravind Ajay, Sriparna Saha, and Monowar Bhuyan. Few-shot anomaly detection in text with deviation learning. In International Conference on Neural Information Processing, pages 425--438. Springer, 2023

  65. [73]

    A robust autoencoder ensemble-based approach for anomaly detection in text

    Jeremie Pantin and Christophe Marsala. A robust autoencoder ensemble-based approach for anomaly detection in text. arXiv preprint arXiv:2405.13031, 2024

  66. [74]

    Large language models for forecasting and anomaly detection: A systematic literature review

    Jing Su, Chufeng Jiang, Xin Jin, Yuxin Qiao, Tingsong Xiao, Hongda Ma, Rong Wei, Zhi Jing, Jiajun Xu, and Junhong Lin. Large language models for forecasting and anomaly detection: A systematic literature review. arXiv preprint arXiv:2402.10350, 2024

  67. [75]

    Ad-llm: Benchmarking large language models for anomaly detection

    Tiankai Yang, Yi Nian, Shawn Li, Ruiyao Xu, Yuangang Li, Jiaqi Li, Zhuo Xiao, Xiyang Hu, Ryan Rossi, Kaize Ding, et al. Ad-llm: Benchmarking large language models for anomaly detection. arXiv preprint arXiv:2412.11142, 2024

  68. [76]

    Can llms serve as time series anomaly detectors? arXiv preprint arXiv:2408.03475, 2024

    Manqing Dong, Hao Huang, and Longbing Cao. Can llms serve as time series anomaly detectors? arXiv preprint arXiv:2408.03475, 2024

  69. [77]

    Anomalygpt: Detecting industrial anomalies using large vision-language models

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 1932--1940, 2024

  70. [78]

    Anomaly detection of tabular data using llms

    Aodong Li, Yunhan Zhao, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, and Stephan Mandt. Anomaly detection of tabular data using llms. arXiv preprint arXiv:2406.16308, 2024 b

  71. [79]

    Ad-nlp: A benchmark for anomaly detection in natural language processing

    Matei Bejan, Andrei Manolache, and Marius Popescu. Ad-nlp: A benchmark for anomaly detection in natural language processing. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10766--10778, 2023

  72. [80]

    Newsweeder: Learning to filter netnews

    Ken Lang. Newsweeder: Learning to filter netnews. In Proceedings of the Twelfth International Conference on Machine Learning, pages 331--339, 1995

  73. [81]

    David D. Lewis. Reuters-21578 text categorization collection data set, 1997. URL https://kdd.ics.uci.edu/databases/reuters21578/reuters21578.html

  74. [82]

    Maas, Raymond E

    Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142--150...

  75. [83]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Proces...

  76. [84]

    Almeida, Jose Maria Gomez Hidalgo, and Akebo Yamakami

    Tiago A. Almeida, Jose Maria Gomez Hidalgo, and Akebo Yamakami. Contributions to the study of sms spam filtering: New collection and results. In Proceedings of the 2011 ACM Symposium on Document Engineering (DOCENG'11), 2011

  77. [85]

    Hdltex: Hierarchical deep learning for text classification

    Kamran Kowsari, Donald E Brown, Mojtaba Heidarysafa, Kiana Jafari Meimandi, , Matthew S Gerber, and Laura E Barnes. Hdltex: Hierarchical deep learning for text classification. In Machine Learning and Applications (ICMLA), 2017 16th IEEE International Conference on. IEEE, 2017

  78. [86]

    Character-level convolutional networks for text classification

    Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. URL ht...

  79. [87]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

  80. [88]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

  81. [89]

    Llama 3 model card

    AI@Meta. Llama 3 model card. 2024. URL https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md

  82. [90]

    LLM2V ec: Large language models are secretly powerful text encoders

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. LLM2V ec: Large language models are secretly powerful text encoders. In First Conference on Language Modeling, 2024 b . URL https://openreview.net/forum?id=IW1PR7vEBf

  83. [91]

    A novel anomaly detection scheme based on principal component classifier

    Mei-Ling Shyu, Shu-Ching Chen, Kanoksri Sarinnapakorn, and LiWu Chang. A novel anomaly detection scheme based on principal component classifier. In Proceedings of the IEEE foundations and new directions of data mining workshop, pages 172--179. IEEE Press Piscataway, NJ, USA, 2003

  84. [92]

    Robust kernel density estimation

    JooSeuk Kim and Clayton D Scott. Robust kernel density estimation. The Journal of Machine Learning Research, 13 0 (1): 0 2529--2565, 2012

  85. [93]

    Electra: Pre-training text encoders as discriminators rather than generators

    Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020

  86. [94]

    Exact matrix completion via convex optimization

    Emmanuel Candes and Benjamin Recht. Exact matrix completion via convex optimization. Communications of the ACM, 55 0 (6): 0 111--119, 2012

  87. [95]

    Factor group-sparse regularization for efficient low-rank matrix recovery

    Jicong Fan, Lijun Ding, Yudong Chen, and Madeleine Udell. Factor group-sparse regularization for efficient low-rank matrix recovery. Advances in neural information processing Systems, 32, 2019

  88. [96]

    Low-rank matrix completion [lecture notes]

    Yuejie Chi. Low-rank matrix completion [lecture notes]. IEEE Signal Processing Magazine, 35 0 (5): 0 178--181, 2018

Pith tools

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