Anomaly Detection in Human Language via Meta-Learning: A Few-Shot Approach

Aarav Singla; Advik Gupta; Parnika Gupta; Saurav Singla

REVIEW 4 major objections 5 minor 26 references

Anomaly Detection in Human Language via Meta-Learning: A Few-Shot Approach

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 A prototypical network trained on cross-domain episodes detects spam, fake news, and hate speech with far fewer labels than fine-tuned BERT, reaching ROC-AUC up to 0.972.

desk verdict The cross-domain episode sampling idea is worth a look, but the paper's headline few-shot result is not actually measured: the evaluation computes prototypes from the full training set, so Table 2 is about full-data transfer, not 5–10 shot adaptation. read the letter →

arxiv 2507.20019 v1 pith:TOOICX2P submitted 2025-07-26 cs.CL cs.AI

Saurav Singla , Aarav Singla , Advik Gupta , Parnika Gupta This is my paper · ORCID

classification cs.CLcs.AI

keywords few-shotanomalydetectionmeta-learningprototypicalnetworkscross-domainsamplingtextspamfakenewshatespeech

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 tries to establish that meta-learning can make few-shot anomaly detection in human language work with scarce anomaly labels: training on episodes drawn from several text domains (spam, fake news, hate speech) lets a model adapt to a new anomaly-detection task better than fine-tuning a pretrained BERT classifier on the same few examples. Its central proposal is cross-domain episode sampling, in which some training episodes pair normal texts from one domain with anomalies from another, forcing the model to learn anomaly signals that are not tied to one topic. The paper reports that a prototypical network with this sampling reaches ROC-AUC of 0.972 on SMS spam, 0.853 on COVID-19 fake news, and 0.901 on hate speech, beating fine-tuned BERT by 4.5 to 8.9 points. If the claim holds, rare and evolving text anomalies can be detected with a handful of labeled examples, which matters for content moderation and misinformation detection.

What carries the argument

The mechanism is the prototypical network with cross-domain episode sampling. In each episode the model embeds a small support set containing a few normal and a few anomalous texts, computes the mean embedding (prototype) of each class, and classifies query texts by squared Euclidean distance to the two prototypes. The proposed cross-domain variant builds about a quarter of training episodes from normal texts of one domain and anomaly texts of a different domain (e.g., normal tweets versus spam messages), so the encoder is trained to separate anomalies from normals without leaning on topic-specific cues. This cross-domain mixing is the load-bearing component that the paper credits for the extra generalization.

What would settle it

Train the same ProtoNet with cross-domain sampling and the fine-tuned BERT baseline, then at test time give both methods only k = 5, 10, and 20 labeled anomalies (with the same number of normals) from each target domain's training split and measure test-set AUC; if the reported 5- to 9-point gains shrink or reverse, the central claim is falsified. A second check: run the leave-one-out experiment from Section 5.3 with the same small-support protocol and see whether the meta-learned prior still beats per-task fine-tuning.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that a metric-based meta-learner, prototypical networks trained over multi-domain episodes, with a fraction of episodes constructed by mixing normal examples from one domain with anomalous examples from another, is an effective few-shot anomaly detector for natural language. The paper further claims that this cross-domain mixing is what delivers the larger gains: without it, ProtoNet already beats fine-tuned BERT, and adding it lifts AUC by roughly 2 to 5 more points on the three datasets. The paper interprets the mechanism as the model being unable to rely on domain-specific cues in mixed episodes, so it learns more general features of anomalous text such as stylistic inconsistency or semantic outlierness.

Load-bearing premise

The paper's headline few-shot claim assumes that forming prototypes or fine-tuning on the entire training set of a target task (e.g., all 144 spam messages) is still a fair demonstration of adapting from as few as 5 to 10 anomaly examples; if the method's advantage disappears when the test-time support set is truly limited to a few examples, the central claim would not hold.

Editorial extensions

If this is right

A few-shot text anomaly detector can be trained once over several source domains and then adapted to a new anomaly type with a handful of labels, outperforming per-task fine-tuning of a pretrained transformer.
Cross-domain episode mixing is a reusable data-augmentation idea for meta-learning; the paper's ablations show it helps most on tasks with novel anomaly content.
Leave-one-out results in Section 5.3 suggest the model transfers to a domain never seen in meta-training, reaching near in-training performance on fake news after a small support-set fine-tune.
ROC-AUC gains of roughly 5 to 9 points over a fine-tuned BERT baseline would make the approach attractive for deployment settings where labeled anomalies are scarce but expensive to collect.

Reading between the lines

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

The strength on fake news and hate speech suggests the cross-domain prior may be learning a generic textual-deviance representation; a testable extension is whether the same encoder helps on unrelated anomaly types such as phishing emails or propaganda without retraining.
The paper reports AUC with prototypes built from the full training set; a stricter test of the few-shot claim would adapt from exactly k=5-10 anomalies at test time and compare to the same budget for fine-tuning, clarifying whether the advantage is a meta-learned prior or mostly access to the training distribution.
Because all three datasets are short informal texts, a natural next test is whether the gains survive on long-form documents such as news articles or clinical notes, where the notion of anomalous is less tied to lexical outliers.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The cross-domain episode sampling idea is the real news here: pairing normal examples from one domain with anomalies from another during meta-training is a clean, sensible extension of outlier exposure, and it's not in the cited literature. The paper also reports consistent gains over baselines on all three datasets, and the code is released, which is more than many papers do.

The soft spots, though, are not minor. The central claim—detecting anomalies with as few as 5–10 labeled examples—is never actually tested. Section 4.3 says ProtoNet prototypes are computed from the full training set (all 144 spam etc.), and evaluate.py does exactly that. So the abstract's "5–10 points AUC improvement in few-shot settings" is really a full-data comparison. The fine-tuned BERT baseline also trains on the full split, so the comparison is fair as a full-data transfer result, but it is not few-shot adaptation. The leave-one-out experiments that would actually demonstrate few-shot transfer are only described in prose, with no table; that is a real omission.

The MAML implementation in the released code is also not MAML. In train_meta.py, the inner loop updates the model parameters in place, and the outer loop then steps the optimizer on the same already-updated parameters, without restoring the original initialization. That is sequential fine-tuning, not meta-learning. So the MAML rows in Table 2 should be treated with caution until the code is fixed. The test sets also have roughly 20% anomaly prevalence, which is far from the claimed ~3% setting; this does not invalidate the AUC comparisons, but it does make the F1 numbers harder to interpret as an operating point for real deployments.

None of this means the cross-domain idea is wrong. The full-data results suggest a meta-learned encoder plus prototypes can beat a fine-tuned BERT, which is interesting and worth following up. But the paper needs to be reframed—or better, re-run with genuine few-shot support sets at test time, and with the leave-one-out experiments reported in a table. As submitted, it does not support the headline few-shot claim.

Who should read it: people working on few-shot anomaly detection in NLP might get a useful idea from the sampling strategy, and the paper is a good teaching example of how an evaluation protocol can drift from the stated claim. I would not desk-reject it, but I would send it back for major revision with the few-shot evaluation as a hard requirement. The idea is new enough and the flaws are fixable enough that a serious referee could turn it into something solid.

Referee Report

4 major / 5 minor

Summary. The paper proposes a meta-learning framework for few-shot anomaly detection in text, combining MAML and Prototypical Networks with a cross-domain episode sampling strategy, and evaluates on SMS spam, COVID-19 fake news, and hate speech datasets. The central claim is that the proposed methods detect anomalies with as few as 5–10 labeled examples and outperform baselines in ROC-AUC and F1. The manuscript includes an extensive code appendix for reproducibility.

Significance. If the few-shot claims were properly validated, this work would be a useful contribution to a practical and under-explored problem: detecting rare, evolving text anomalies with minimal labels. The cross-domain episode sampling idea is sensible and potentially transferable. Credit is due for providing a complete PyTorch codebase in the manuscript and for reporting multiple metrics. However, the current evaluation does not test the stated few-shot setting: test-time prototypes are computed from the full training sets, not from few-shot supports, so Table 2 demonstrates full-data transfer rather than few-shot adaptation. The MAML implementation in the code is also not a correct meta-learning procedure. These issues are load-bearing because the paper's novelty and headline results rest on the few-shot framing.

major comments (4)

[§4.3 and Code (evaluate.py)] The evaluation protocol contradicts the paper's few-shot claim. Section 4.3 states that 'For ProtoNet, we did not fine-tune further; we simply computed prototypes from the full training set (e.g. using all 144 spam and a sample of ham for SMS) and evaluated on test,' and the released evaluate.py reads the full train CSV and embeds all normal and all anomaly training texts to form prototypes. This means the Table 2 results for ProtoNet and ProtoNet+X-domain are obtained with 144, 53, and 124 anomaly examples per domain, not the '5–10 anomaly examples' claimed in Contribution 3 and the abstract. Since §3.1 defines the support set as containing 'only a handful of anomaly examples,' the reported F1/AUC numbers measure full-data transfer, not few-shot adaptation. The few-shot premise is load-bearing: the method's advantage over fine-tuning BERT is presented as few-shot advantage, but both methods receive the full training set at test time. The paper would need to re-run the evaluation with support sets of 5–10 anomalies (and a modest number of normals) to support the headline claim.
[§5.3 (Generalization to Unseen Task)] The leave-one-out experiments that would demonstrate true transfer to unseen anomaly types are reported only in prose, without a table, exact numbers, standard deviations, or the number of support examples used for adaptation. For example, 'ProtoNet (no X-domain) gave AUC ~0.78 on Fake,' and with cross-domain episodes '~0.82,' and after fine-tuning '~0.85.' These are central to the claim that the method generalizes to unseen anomaly types, yet no numeric table is provided and the protocol (how many shots, which hyperparameters, how the threshold was chosen) is not specified. This evidence cannot be checked or compared against Table 2. A table with AUC/F1 (and standard deviations) for each held-out domain, with a clear statement of support sizes, is required.
[Code (train_meta.py, MAML branch)] The MAML implementation in the provided code does not implement MAML or FOMAML. In the MAML branch, the inner-loop optimizer performs a gradient step on the support loss directly on the model parameters, then the query loss is backpropagated and applied to the same parameters via the outer optimizer. The original parameters are never restored, and the query-loss gradient is not propagated through the inner-loop update. This is equivalent to two sequential gradient updates, not a meta-learning update that optimizes the initial parameters for fast adaptation. The paper states 'we use the first-order approximation (FOMAML) for efficiency,' but the code does not implement FOMAML either, as it does not compute the meta-gradient with respect to the pre-adaptation parameters. Consequently, the MAML rows in Table 2 (and MAML+X-domain) do not reflect the claimed algorithm. The authors should either fix the implementation (e.g., using torchmeta or properly restoring parameters before the outer update) or substantially weaken the MAML claims.
[Table 1 and §4.1] The test sets are constructed with roughly 20% anomaly prevalence (Table 1: SMS 20.3%, COVID 20.0%, Hate 20.0%), contrary to the stated 'realistic class imbalance (anomaly rate ~3%)' in the abstract and to §4.1's claim that test anomaly rates are 'around 5–10%.' The paper reports F1 and AP at this 20% test prevalence, which can substantially inflate F1 relative to the 3% deployment setting; the threshold is also chosen to maximize validation F1 without reporting the validation prevalence. The authors should either evaluate at the natural test prevalence (or report precision-recall curves / operating points for a range of prevalences) and reconcile the prevalence description with Table 1.

minor comments (5)

[§3.2 and throughout] The manuscript contains numerous corrupted LaTeX fragments (e.g., the MAML objective and ProtoNet distance formula in §3.2) that obscure the mathematical definitions and should be cleaned up.
[References] The reference list is incomplete: many in-text citations such as [4], [7], [8], [10]–[13], [15], [16], [19]–[29], [31]–[57] are cited but not listed in the References section, which prevents verification of the related-work claims.
[Table 2] The table reports standard deviations only for AUC, not for AP or F1, despite the text stating that results are averaged over 5 runs with standard deviations in parentheses; the missing deviations for the headline F1 numbers make it hard to judge the significance of the claimed improvements.
[§4.3] The validation-set threshold selection for F1 is described but no validation F1 values or validation anomaly rates are reported, so it is unclear whether the operating point is overfit to the validation set.
[Abstract] The phrase 'few-shot settings' in the abstract is not defined; given that the evaluation uses the full training set, the term as used is misleading even if the training episodes are few-shot.

Circularity Check

1 steps flagged · score 6.0 of 10

Few-shot results in Table 2 are computed from full-training-set prototypes, so the claimed 5-10-shot prediction reduces by construction to full-data prototype classification.

fitted input called prediction [Section 4.3 (Evaluation); Table 2; evaluate.py]
"For ProtoNet, we did not fine-tune further; we simply computed prototypes from the full training set (e.g. using all 144 spam and a sample of ham for SMS) and evaluated on test."

The paper's headline claim is that meta-learning improves detection 'with as few as 5-10 anomaly examples' (Contribution 3), and Section 3.1 defines the support set as containing 'only a handful of anomaly examples'. The evaluation, however, forms ProtoNet prototypes from every anomaly in the training split: 144 spam, 53 fake, and 124 hate examples. The released evaluate.py code does the same by averaging embeddings of all rows in the train CSV. Table 2's ProtoNet+X-domain AUC/F1 values are therefore exactly the scores of a full-data prototype classifier, with anomaly scores computed as distance differences to full-data prototypes.

full rationale

No self-citation chain, uniqueness import, or ansatz-smuggling was found; the meta-learning algorithms and cross-domain sampling are implemented as standard episodic training. The central circularity is the evaluation-protocol mismatch. The paper explicitly says ProtoNet 'simply computed prototypes from the full training set' and the code confirms that the prototype means use all training anomalies. Consequently, the reported ROC-AUC/F1 improvements do not measure adaptation from 5-10 shots, and the headline few-shot claim reduces to a full-data classifier comparison. The 'generalization to unseen tasks' claim is also not independently grounded: the three test sets are drawn from the same three domains used in meta-training, and the leave-one-out transfer experiments are described in prose without a numeric table. These are evaluation-validity concerns that make the central claim partially circular, but they do not force the empirical numbers by an equation, so the score is moderate rather than extreme.

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

The paper introduces no new theoretical objects, but it does rest on several modeling choices and hyperparameters that are not derived from first principles. The most important is the cross-domain sampling assumption, which is a hypothesis about machine learning rather than a demonstrated law. The free parameters listed above are all chosen by hand or tuned on validation, and none is justified by a theoretical argument.

free parameters (6)

Cross-domain episode probability p = 0.25
The fraction of meta-training episodes that are cross-domain; selected by ablation in Section 5.3 to maximize validation performance.
Support anomaly count n_a = 5
Fixed number of anomaly examples in each support set; chosen by hand in Section 4.3 to maintain a stable inner-loop update.
Support normal count n_n = 20 for MAML, 5 for ProtoNet
Number of normal examples in support sets; set in Section 4.3 and code, balancing the class ratio.
Inner and outer learning rates = alpha=5e-5/5e-4, beta=1e-5
Hyperparameters tuned on validation, stated in Section 4.3, with no sensitivity analysis.
F1 threshold = chosen to maximize validation F1
Operating threshold selected on validation data per method, as stated in Section 4.3 and Table 2 note.
Training anomaly rate target = ~3%
The paper downsamples anomalies in training to about 3% (e.g., 144 spam out of 4,969), a design choice stated in Section 4.1.

assumptions (4)

domain assumption Language anomaly detection can be treated as binary normal-vs-anomalous classification with extreme class imbalance.
Section 3.1 formalizes each task as one-vs-rest binary classification; this framing presupposes that a single 'anomaly' class is well defined per domain.
ad hoc to paper Episodic meta-training on three text domains transfers to new anomaly types.
Section 3.2 assumes that the model can acquire cross-task knowledge from only three source domains, which is a strong assumption given the limited diversity.
ad hoc to paper Cross-domain episodes, pairing normals from one domain with anomalies from another, teach general anomaly features.
Section 3.3 hypothesizes that mixing domains forces the model to learn domain-independent anomaly indicators; this is the central assumption behind the proposed method.
domain assumption Pre-trained BERT embeddings are a sufficient feature space for all three anomaly detection tasks.
Section 4.2 uses BERT as the shared encoder without comparing to other representations; the results depend on BERT's transfer quality.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Anomaly Detection in Human Language via Meta-Learning: A Few-Shot Approach." pith.science (2026). https://pith.science/paper/TOOICX2P

@misc{pith2026250720019,
  author       = {Pith},
  title        = {Pith review of: Anomaly Detection in Human Language via Meta-Learning: A Few-Shot Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TOOICX2P}},
  note         = {Machine review of arXiv:2507.20019}
}

read the original abstract

We propose a meta learning framework for detecting anomalies in human language across diverse domains with limited labeled data. Anomalies in language ranging from spam and fake news to hate speech pose a major challenge due to their sparsity and variability. We treat anomaly detection as a few shot binary classification problem and leverage meta-learning to train models that generalize across tasks. Using datasets from domains such as SMS spam, COVID-19 fake news, and hate speech, we evaluate model generalization on unseen tasks with minimal labeled anomalies. Our method combines episodic training with prototypical networks and domain resampling to adapt quickly to new anomaly detection tasks. Empirical results show that our method outperforms strong baselines in F1 and AUC scores. We also release the code and benchmarks to facilitate further research in few-shot text anomaly detection.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

[1]

$p=0.25$), sample two different domains $A$ and $B$

With probability $p$ (e.g. $p=0.25$), sample two different domains $A$ and $B$. Otherwise, sample one domain $A = B$

work page
[2]

From domain $A$, sample a support set of $n_n$ normal texts and a query set of additional normal texts

work page
[3]

If $A=B$ (intra-domain case), ensure that support+query sampling includes at most $n_a$ anomalies and majority normal, as per the anomaly rate

From domain $B$, sample $n_a$ anomalous texts for support and some anomalous texts for query. If $A=B$ (intra-domain case), ensure that support+query sampling includes at most $n_a$ anomalies and majority normal, as per the anomaly rate

work page
[4]

For each, shuffle the normal and anomaly examples (the model doesn’t know which is which a priori)

Form combined support and query sets. For each, shuffle the normal and anomaly examples (the model doesn’t know which is which a priori)

work page
[5]

Compute adaptation (for MAML) or prototypes (for ProtoNet) on the support set and then evaluate loss on the query set

work page
[6]

potential anomalies

Update meta-parameters $\theta$ to minimize the query loss (via backpropagation through adaptation in MAML, or directly in ProtoNet). Over many such episodes, the training optimizes $\theta$ to perform well on both intra- domain and cross-domain detection tasks. The hope is that $\theta$ will settle on an embedding space and decision function that capture...

work page
[7]

see you at noon at the cafe

Experimental Setup 4.1 Datasets and Preparation. We evaluate on three public anomaly detection datasets in NLP, chosen to cover distinct domains: (1) SMS-Spam, (2) COVID-Fake, (3) HateSpeech. These were introduced in Section 1; here we provide details and how we prepared them for our experiments. Table 1 summarizes dataset statistics after processing. SMS...

work page 2011
[8]

Sunshine Holla! Get cash now

Results 5.1 Overall Detection Performance. We first present the main results on the test sets of the three datasets. Table 2 summarizes the performance of each method. We report ROC-AUC and Average Precision (AP) for each dataset (higher is better), as well as the F1-score for the anomaly class at an operating threshold chosen to maximize F1 on the valida...

work page

Show all 26 references

[9]

This is a promising outcome for real-world applications

Discussion Our results demonstrate that meta-learning can substantially improve few-shot anomaly detection in language, thanks to its ability to transfer knowledge across tasks. This is a promising outcome for real-world applications. For instance, consider detecting new forms...
[10]

learn to detect anomalies

Conclusion We presented a comprehensive study on anomaly detection in human language using meta-learning. By leveraging three public datasets (spam, fake news, hate speech) and state-of-the-art meta-learning methods (MAML and prototypical networks), we showed that a model can ...
[11]

A., Hidalgo, J

Almeida, T. A., Hidalgo, J. M. G., & Yamakami, A. (2011). Contributions to the study of SMS spam filtering: new collection and results. Proceedings of the 11th ACM Symposium on Document Engineering, 259–262. (SMS Spam dataset source.)

2011
[12]

D., Basak, A., & Dutta, S

Das, S. D., Basak, A., & Dutta, S. (2021). A heuristic-driven ensemble framework for COVID-19 fake news detection. In Combating Online Hostile Posts in Regional Languages during Emergency Situation (CONSTRAINT 2021), Lecture Notes in Computer Science, Springer. (COVID-19 Fake ...

2021
[13]

Davidson, T., Warmsley, D., Macy, M., & Weber, I. (2017). Automated hate speech detection and the problem of offensive language. Proceedings of ICWSM, 11(1), 512–515. (Hate Speech dataset source.)

2017
[14]

Finn, C., Abbeel, P., & Levine, S. (2017). Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. Proceedings of ICML. (MAML algorithm – meta- learning for few-shot adaptation.)

2017
[15]

Snell, J., Swersky, K., & Zemel, R. (2017). Prototypical Networks for Few-shot Learning. Proceedings of NeurIPS. (Prototypical Networks – metric-based few-shot classification.)

2017
[16]

Cao, Y., et al. (2025). TAD-Bench: A Comprehensive Benchmark for Embedding-Based Text Anomaly Detection. arXiv preprint arXiv:2501.11960. (Introduces multiple NLP anomaly datasets and highlights the need for generalizable detection methods.)

2025 arXiv
[17]

Xu, Y., et al. (2023). Comparative Analysis of Anomaly Detection Algorithms in Text Data. Proceedings of RANLP 2023. (Showed semi-supervised (few labels) > unsupervised for text anomalies.)

2023
[18]

Sun, H., Huang, Y., Han, L., & Zhou, C. (2023). Few-shot Detection of Anomalies in Industrial Cyber-Physical Systems via Prototypical Network and Contrastive Learning. arXiv:2302.10601. (Few-shot anomaly detection in ICS, using ProtoNet + contrastive loss.)

2023 arXiv
[19]

Hendrycks, D., Mazeika, M., & Dietterich, T. (2019). Deep Anomaly Detection with Outlier Exposure. Proceedings of ICLR. (Outlier exposure technique for improving anomaly detection by using auxiliary out-of-distribution data.)

2019
[23]

Preprocess the datasets using the provided scripts
[24]

Train the ProtoNet model with cross-domain sampling: python train_meta.py --method prototypical --episodes 3000 -- use_cross_domain This configuration corresponds to our best model (Prototypical Network with cross-domain episodes)
[25]

These should closely match the values in Table 2 of the paper (allowing for minor randomness)

The script will print the final ROC-AUC and F1 for each dataset's test set. These should closely match the values in Table 2 of the paper (allowing for minor randomness)
[26]

label" (with values

For additional experiments (e.g., training MAML or disabling cross-domain sampling), adjust the --method or omit --use_cross_domain accordingly. All code is documented with comments for clarity. Please refer to the paper for detailed explanations of the algorithms and design c...

2000
[30]

Ma, Z., et al. (2024). A Biological Immunity-Based Neuro Prototype for Few- Shot Anomaly Detection with Character Embedding. Cyborg and Bionic Systems, 5:0086. (Few-shot text anomaly detection method using immune-inspired prototype refinement.)

2024
[58]

Li, C., et al. (2024). NLP-ADBench: Anomaly Detection Benchmark for NLP. arXiv preprint arXiv:2401.xxxxx. (Expanded algorithms and datasets for NLP anomaly detection.)

2024
[59]

Anomaly Detection in Human Language via Meta-Learning: A Few-Shot Approach

Zhou, K., et al. (2022). Few-Shot Fast-Adaptive Anomaly Detection. Proceedings of NeurIPS. (Proposes an energy-based model with meta-learning for few- shot anomaly detection in images, emphasizing quick adaptation without retraining.) PyTorch Implementation Project Structure: ...

2022

Pith tools

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

[1] [1]

$p=0.25$), sample two different domains $A$ and $B$

With probability $p$ (e.g. $p=0.25$), sample two different domains $A$ and $B$. Otherwise, sample one domain $A = B$

work page

[2] [2]

From domain $A$, sample a support set of $n_n$ normal texts and a query set of additional normal texts

work page

[3] [3]

If $A=B$ (intra-domain case), ensure that support+query sampling includes at most $n_a$ anomalies and majority normal, as per the anomaly rate

From domain $B$, sample $n_a$ anomalous texts for support and some anomalous texts for query. If $A=B$ (intra-domain case), ensure that support+query sampling includes at most $n_a$ anomalies and majority normal, as per the anomaly rate

work page

[4] [4]

For each, shuffle the normal and anomaly examples (the model doesn’t know which is which a priori)

Form combined support and query sets. For each, shuffle the normal and anomaly examples (the model doesn’t know which is which a priori)

work page

[5] [5]

Compute adaptation (for MAML) or prototypes (for ProtoNet) on the support set and then evaluate loss on the query set

work page

[6] [6]

potential anomalies

Update meta-parameters $\theta$ to minimize the query loss (via backpropagation through adaptation in MAML, or directly in ProtoNet). Over many such episodes, the training optimizes $\theta$ to perform well on both intra- domain and cross-domain detection tasks. The hope is that $\theta$ will settle on an embedding space and decision function that capture...

work page

[7] [7]

see you at noon at the cafe

Experimental Setup 4.1 Datasets and Preparation. We evaluate on three public anomaly detection datasets in NLP, chosen to cover distinct domains: (1) SMS-Spam, (2) COVID-Fake, (3) HateSpeech. These were introduced in Section 1; here we provide details and how we prepared them for our experiments. Table 1 summarizes dataset statistics after processing. SMS...

work page 2011

[8] [8]

Sunshine Holla! Get cash now

Results 5.1 Overall Detection Performance. We first present the main results on the test sets of the three datasets. Table 2 summarizes the performance of each method. We report ROC-AUC and Average Precision (AP) for each dataset (higher is better), as well as the F1-score for the anomaly class at an operating threshold chosen to maximize F1 on the valida...

work page