CoChat
Sign Up
Cited by 12Hate Speech and Cyberbullying DetectionSpam and Phishing DetectionText and Document Classification Technologies

Detecting Online Grooming By Simple Contrastive Chat Embeddings

Open access
Rezaee Borj, Parisa, Raja, Kiran, Bours, Patrick·2023·Proceedings of the 9th ACM International Workshop on Security and Privacy Analytics
DOI 10.1145/3579987.3586564
Open this paper in CoChat — free accountOpen PDF

Abstract

Protecting children against predators that use the internet as a medium to find their victims is essential, where the difficulty of monitoring online messaging platforms to avoid potential threats toward underage users is alarming. Online grooming detection requires a deep knowledge of predatory behaviour where an online conversation's positive or negative connotations rely on the context where it can change rapidly in an online chat. Therefore, it is essential to use robust feature vectors for predatory conversation detection. This research paper proposes a contrastive learning framework for feature extraction in a sentence-based manner where it can assign a feature vector to the conversation with misspellings using subword information. Also, it is vital to have a high detection rate of true positives to avoid potential non-wanted consequences. We propose a configuration of RoBERTa encoders and supervised SimCSE for training the SVM model, leading to a high rate of detecting relevant samples (predatory samples). Our proposed approach gains an F0.5-score of 0.96, an F1-score of 0.96, and an accuracy of 0.99 for predatory conversation detection that benchmarks the state-of-the-art. In order to improve the performance, we also perform experiments with various fusion approaches, where the sum fusion of all configurations obtains an accuracy of 0.99, an F_1-score of 0.97, and an F0.5-score of 0.98.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes supervised SimCSE sentence embeddings from RoBERTa encoders with an SVM classifier to detect online grooming in chat conversations, evaluated on PAN 2012. It reports high accuracy and F-scores, with fusion of multiple embeddings and classifiers improving F0.5 performance.

Key findings

  • RoBERTa encoders + supervised SimCSE + SVM achieved accuracy 0.99, F1-score 0.96, and F0.5-score 0.96 for predatory conversation detection.
  • Sum fusion of all configurations (16 scores) achieved accuracy 0.99, F1-score 0.97, and F0.5-score 0.98; the weighted average result from this sum rule was accuracy 0.99, F1-score 0.99, and F0.5-score 0.99.
  • SimCSE-large BERT embeddings with sum fusion at the classifier level achieved accuracy 0.99, F1-score 0.96, and F0.5-score 0.97; its weighted average gave accuracy 0.99, F1-score 0.99, and F0.5-score 0.99.
  • Compared with Borj et al. (accuracy 0.99, F0.5-score 0.94), the proposed RoBERTa + supervised SimCSE + SVM improved F0.5 to 0.96, though its F1-score of 0.96 was lower than Borj et al.’s.

Methods

  • Data: PAN 2012 conversations, including ordinary chats, adult sexual chats, and predator–minor/predator–police chats. Preprocessing removed conversations with more than two or only one user, fewer than seven messages, and non-English words; no stemming or lemmatization was performed. Training set: 951 predatory and 8,477 non-predatory conversations; test set: 1,697 predatory and 19,922 non-predatory conversations.
  • Model: sentence-based embeddings extracted using SimCSE, specifically supervised SimCSE with RoBERTa encoders, then classified with an SVM. The paper also tested SimCSE-base/large BERT and RoBERTa configurations with Random Forest, Naive Bayes, and SGD classifiers, plus majority voting, product rule, and sum rule fusion.
  • Evaluation: accuracy, precision, recall, F1-score, and F0.5-score on the highly imbalanced PAN 2012 dataset.

Limitations

  • Stated: no predatory detection data is open and massively available, and accessing victim–predator conversations is challenging due to security and privacy issues. The paper also states future work should explore cultural differences, since the same sentence can vary in meaning by culture.
  • Evident from the text: evaluation is limited to PAN 2012 and its preprocessing choices, such as removing conversations with fewer than seven messages and non-English words, which may affect generalizability.
  • Evident from the text: the paper reports “F0-score” in some places and “F0.5-score” elsewhere, making the exact metric naming ambiguous in parts of the reported results.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3579987.3586564

Cites

21 of 21 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 20 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The training data contains 951 predatory and 8477 non-predatory conversations.

    “After the pre-processing step, the training data contains 951 predatory and 8477 non-predatory conversations, and the test set has 1697 predatory samples and 19922 non-predatory conversations (see Table 1).”
    Quote in the paperSubstantial claim
  • Findingown data

    The configuration of sum fusion models for SimCSE-large Bert embeddings gives the best performance with an accuracy of 0.99, an F1-score of 0.96, and an F0.5-score of 0.97.

    “The configuration of sum fusion models for SimCSE-large Bert embeddings gives the best performance with an accuracy of 0.99, an 𝐹 1 -score of 0.96, and an 𝐹 0.5 -score of 0.97.”
    • Also stated in: § 5.5 Comparison of the Results with the state-of-the art · plain
      “The configuration of SimCSE-Large Bert embeddings and sum fusion at the classifier level gives an accuracy of 0.99, an 𝐹 1 -score of 0.96, and 𝐹 0.5 -score of 0.97.”
    Quote in the paperSubstantial claim
  • Findingown data

    The proposed approach detects a higher number of true positives (predatory chats) than seven state-of-the-art approaches.

    “In addition, we benchmark our approach against seven different state-of-the-art (SOTA) approaches to show the benefit of our proposed approach in detecting a higher number of true positives (i.e., predatory chats) than SOTA approaches.”
    Quote in the paperSubstantial claim
  • Findingown data

    The number of predatory conversations is much lower than the number of non-predatory conversations.

    “The number of predatory conversations is much lower than the non-predatory ones, which satisfies the requirements for having an imbalanced dataset to be close to the real-world situation.”
    • Also stated in: § 5 EXPERIMENTAL RESULTS · plain
      “It should be mentioned that the dataset is highly imbalanced, where the number of positive cases (grooming conversations) is much lower than the number of negative ones (ordinary conversations).”
    Quote in the paperSubstantial claim
  • Findingown data

    The data covering conversations between minor victims and predators contains conversations between online predators and police officers who pretend to be minors to trap the predator.

    “This data contains the conversations between online predators and police officers who pretend to be minors to trap the predator.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 22
  • Classified as probably not a substantial claim. 15

  • May combine two claims.

    Findingown data

    The proposed configuration of RoBERTa encoders and supervised SimCSE for training the SVM model achieves an F0-score of 0.96, an F1-score of 0.96, and an accuracy of 0.99 for predatory conversation detection, benchmarking the state-of-the-art.

    “Our proposed approach gains an 𝐹 0 -score of 0.96, an 𝐹 1 -score of 0.96, and an accuracy of 0.99 for predatory conversation detection that benchmarks the stateof-the-art.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The sum fusion of all configurations obtains an accuracy of 0.99, an F1-score of 0.97, and an F0.5-score of 0.98.

    “In order to improve the performance, we also perform experiments with various fusion approaches, where the sum fusion of all configurations obtains an accuracy of 0.99, an 𝐹 1 -score of 0.97, and an 𝐹 0.5 -score of 0.98.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    Table 5 reports results based on the macro average of classifiers where the best performance is gained by product and sum.

    “Table 5 reports the results based on the macro average of the classifiers where the best performance is gained by product and sum.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The configuration of Naive Bayes and RoBERTa-base results in an accuracy of 0.97 and an F0-score of 0.91.

    “The configuration of Naive Bayes and RoBERTa-base results in an accuracy of 0.97 and an 𝐹 0 -score of 0.91”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The RF classifier gives the best accuracy of 0.97, an F1-score of 0.90, and an F0-score of 0.94 when applied on RoBERTa base embeddings.

    “RF classifier gives the best accuracy of 0.97, an 𝐹 1 -score of 0.90, and an 𝐹 0 -score of 0.94 when applied on RoBERTa base embeddings.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Our configuration of RoBERTa encoders and supervised SimCSE for training the SVM model has gained an F0-score of 0.96, an F1-score of 0.96, and an accuracy of 0.99 for predatory conversation detection that benchmarks the state-of-the-art.

    “Our configuration of RoBERTa encoders and supervised Sim-CSE for training the SVM model has gained an 𝐹 0 -score of 0.96, an 𝐹 1 -score of 0.96, and an accuracy of 0.99 for predatory conversation detection that benchmarks the state-of-the-art.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The sum fusion of all configurations gives an accuracy of 0.99, an F1-score of 0.97, and an F0.5-score of 0.98.

    “The sum fusion of all configurations gives an accuracy of 0.99, an 𝐹 1 -score of 0.97, and an 𝐹 0.5 -score of 0.98.”
    • Also stated in: § 5.5 Comparison of the Results with the state-of-the art · plain
      “The fusion of all embeddings and classifiers (sum fusion of 16 scores) gives the best result, where the accuracy is 0.99, the 𝐹 1 -score is 0.97, and the 𝐹 0.5 -score is 0.98.”
    • Also stated in: § 6 CONCLUSION · plain
      “The sum fusion of all configurations gives an accuracy of 0.99, an 𝐹 1 -score of 0.97, and an 𝐹 0.5 -score of 0.98.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The configuration of SimCSE-base Bert and SGD gains the best accuracy of 0.99, an F1-score of 0.96, and an F0-score of 0.95.

    “The configuration of SimCSE-base Bert and SGD gains the best accuracy of 0.99, an 𝐹 1 -score of 0.96, and an 𝐹 0 -score of 0.95.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The test set contains 1697 predatory samples and 19922 non-predatory conversations.

    “After the pre-processing step, the training data contains 951 predatory and 8477 non-predatory conversations, and the test set has 1697 predatory samples and 19922 non-predatory conversations (see Table 1).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    Each sample includes conversations between two users that can be predatory or non-predatory.

    “Each sample includes conversations between two users that can be predatory or non-predatory.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The applied data contains three main types of conversations: typical non-sexual conversations, sexual conversations between adults, and conversations between a child molester and a minor victim.

    “The applied data has various conversations from online platforms.Mainly, three different conversations are • a typical conversation that does not have any sexual topic, • a conversation on sexual topics between adults, and • a conversation between a child molester and a minor victim.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The weighted average result from the sum rule for all configurations gives an accuracy of 0.99, an F1-score of 0.99, and an F0.5-score of 0.99.

    “However, it should be mentioned that the weighted average result from the sum rule gives an accuracy of 0.99, an 𝐹 1 -score of 0.99, and an 𝐹 0.5 -score of 0.99.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    Bag of Words feature vectors do not capture the relationship between words, resulting in the same feature space for words in different contexts and a drop in predatory conversation detection performance.

    “It should be mentioned that BoW feature vectors do not capture the relationship between the words, which results in the same feature space for words in different contexts leading to a drop in predatory conversation detection performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Conventional distributed representations such as word2vec and GloVe create low-level representations of the input and cannot adequately capture the semantics behind the word.

    “Although, conventional distributed representations, such as word2vec and GloVe, create low-level representations of the input and cannot adequately capture the semantics behind the word [9].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [9]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The PAN 2012 competition collected its conversations from several resources to cover all characteristics of real-world data.

    “The PAN 2012 [24] competition on identifying online child groomers collected these conversations from several resources to cover all characteristics of real-world data.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Supervised RoBERTa has shown to be promising for various semantic textual similarity tasks.

    “Supervised RoBERTa [23] has shown to be promising for various semantic textual similarity tasks [23].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [23]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The PAN 2012 dataset is publicly available.

    “We demonstrate the applicability of the proposed approach on the publicly available PAN 2012 dataset [24]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Word2vec and GloVe encode the semantics of linguistic units such as words or word-like entities but ignore sentence-level semantics.

    “While word2vec [10] and GloVe [11] can encode the semantics of linguistic units such as words or word-like entities, they ignore sentence-level semantics.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [10]

      Not checked in the free run.

    • cites [11] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation

    Having a high F0.5-score is essential in the case of grooming detection.

    “It was mentioned that having a high 𝐹 0.5 -score is essential in the case of grooming detection [24].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    The primary idea of the supervised SimCSE technique is that the distance between embeddings of two paired instances, such as two sentences of a grooming conversation, should be relatively shorter than the distance between a predatory sentence and a non-predatory sentence.

    “The primary idea of the supervised SimCSE technique is that the distance between embeddings of two paired instances, such as two sentences of a grooming conversation, should be relatively shorter than the distance between a predatory sentence and a non-predatory sentence.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [23]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Using subword information, the proposed model can create a feature vector for a conversation with misspellings.

    “Using subword information, our proposed model can create a feature vector for the conversation with misspellings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Premisenone given

    Online grooming detection requires deep knowledge of predatory behaviour because an online conversation's positive or negative connotations depend on context that can change rapidly in an online chat.

    “Online grooming detection requires a deep knowledge of predatory behaviour where an online conversation's positive or negative connotations rely on the context where it can change rapidly in an online chat.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisereasoning

    Pre-trained models have knowledge of a larger vocabulary than chat conversation alone and can assign a feature vector to a conversation with misspellings using subword information.

    “The employed pre-trained models inherently have the advantage of having knowledge of a larger vocabulary than chat conversation alone and, at the same time, have the capability of assigning a feature vector to the conversation with misspellings using subword information.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: A large corpus of predatory versus non-predatory conversations is not well available.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisereasoning

    People's behaviour is understood based on what they say in a sentence or phrase, not only one or two words.

    “We understand people's behaviour based on what they say in a sentence or phrase, not only one or two words.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Premisecitation

    Sentences and phrases always represent a person's knowledge and semantic intuitions.

    “However, sentences and phrases always represent a person's knowledge, and semantic intuitions [16].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [16]

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    Applying different fusion models improves performance and provides a more reliable decision for predatory conversation detection.

    “We also show an improvement in performance by applying different fusion models in order to gain a more reliable decision for predatory conversation detection.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Fauzi and Bours [29] proposed an ensemble model that applies various fusions.

    “Fauzi and Bours [29] proposed an ensemble model that applies various fusions.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [29] source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    In NLP, contrastive learning can be done by putting semantically close neighbors together and pushing apart non-neighbors.

    “In the case of NLP, it can be done by putting semantically close neighbors together and pushing apart non-neighbors [28].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [28] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    While the F1-score of 0.96 is relatively lower than Borj et al. [8], the F0.5-score is higher, making the approach more applicable in law-enforcement cases.

    “While the 𝐹 1 -score of 0.96 is relatively lower than Borj et al. [8], we note that 𝐹 0.5 -score is higher, making the approach more applicable in law-enforcement cases.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Borj et al. [8] tried to handle the imbalanced nature of grooming data by hybrid sampling and class re-distribution and gained an accuracy of 0.99 and an F0.5-score of 0.94.

    “For instance, Borj et al. [8] tried to handle the imbalanced nature of grooming data by hybrid sampling and class re-distribution and gained an accuracy of 0.99 and an 𝐹 0.5 -score of 0.94.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Dictionary-based approaches have a good success rate but often perform sub-optimally when the chat conversation contains many unseen words such as slang words.

    “While dictionary-based approaches have a good success rate, they often perform sub-optimally when the chat conversation contains many unseen words [9], such as slang words.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [9]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisenone given

    The primary goal of contrastive learning is to learn effective representation by contrasting positive and negative examples.

    “The primary goal of contrastive learning is to learn effective representation by contrasting positive and negative examples.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    RoBERTa encoders and supervised SimCSE gained 83.76% Spearman's correlation, which was higher than other state-of-the-art approaches.

    “Specifically, RoBERTa encoders and supervised SimCSE gained 83.76% Spearman's correlation which was higher than other state-of-the-art approaches.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [23]

      Not checked in the free run.

33 of 33

  • The paper relies on this without support.

    A large corpus of predatory versus non-predatory conversations is not well available.

    “However, training such a network would need a large corpus of predatory v/s non-predatory conversations, which, unfortunately, is not well available.”
1 of 1

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Most prior work has focused on detecting perpetrators' behaviors, leaving a limited understanding of how teens might respond to unwanted advances.

    “Yet, most prior work has focused on detecting perpetrators' behaviors, leaving a limited understanding of how teens might respond to such unwanted advances.”

    From From Vulnerable to Resilient: Examining Parent and Teen Perceptions on How to Respond to Unwanted Cybergrooming Advances (2026) · cites this paper as [77]

    The Introduction and State of the Art describe prior work primarily as detecting predatory messages, conversations, or users. The paper does not assess how teens respond to unwanted advances or explicitly conclude that understanding of those responses is limited. Its Data section also says the purported minor-side conversations involve police officers posing as minors. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • F-score weighting is described in opposite ways

    Major

    The paper says setting β to 0.5 weights precision lower than recall, but later says a β below 1 gives more importance to precision. These statements cannot both describe the same F-score.

  • Different F0 scores for the RoBERTa-SimCSE SVM configuration

    Minor

    The abstract reports an F0-score of 0.96 for the proposed configuration, while Section 5.4 reports 0.95 for the SVM trained on RoBERTa encoders and supervised SimCSE.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 8

  1. W. H. Organization , , What works to prevent violence against children online? ( 2022 ). W. H. Organization, et al., What works to prevent violence against children online? (2022).
  2. 10.5555/2392963.2392986 · DOI
  3. E. Villatoro-Tello , A. Juárez-González , H. J. Escalante , M. Montes-y Gómez , L. V. Pineda , A two-step approach for effective detection of misbehaving users in chats ., in: CLEF (Online Working Notes/Labs/Workshop) , Vol. 1178 , 2012 . E. Villatoro-Tello, A. Juárez-González, H. J. Escalante, M. Montes-y Gómez, L. V. Pineda, A two-step approach for effective detection of misbehaving users in chats., in: CLEF (Online Working Notes/Labs/Workshop), Vol. 1178, 2012.
  4. B. Myagmar , J. Li , S. Kimura, Transferable high-level representations of bert for cross-domain sentiment classification , in: Proceedings on the International Conference on Artificial Intelligence (ICAI), The Steering Committee of The World Congress in Computer Science , Computer , 2019 , pp. 135 -- 141 . B. Myagmar, J. Li, S. Kimura, Transferable high-level representations of bert for cross-domain sentiment classification, in: Proceedings on the International Conference on Artificial Intelligence (ICAI), The Steering Committee of The World Congress in Computer Science, Computer, 2019, pp. 135--141.
  5. F. Hill , K. Cho , A. Korhonen , Learning distributed representations of sentences from unlabelled data, arXiv preprint arXiv:1602.03483 ( 2016 ). F. Hill, K. Cho, A. Korhonen, Learning distributed representations of sentences from unlabelled data, arXiv preprint arXiv:1602.03483 (2016). · arXiv
  6. Y. Xu , K. Raja , M. Pedersen, Supervised contrastive learning for generalizable and explainable deepfakes detection , in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2022 , pp. 379 -- 389 . Y. Xu, K. Raja, M. Pedersen, Supervised contrastive learning for generalizable and explainable deepfakes detection, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 379--389.
  7. G. Inches , F. Crestani , Overview of the international sexual predator identification competition at pan-2012 ., in: CLEF (Onlineworking notes/labs/workshop) , Vol. 30 , 2012 . G. Inches, F. Crestani, Overview of the international sexual predator identification competition at pan-2012., in: CLEF (Onlineworking notes/labs/workshop), Vol. 30, 2012.
  8. P. R. Borj , K. Raja , P. Bours, On preprocessing the data for improving sexual predator detection: Anonymous for review , in: 2020 15th InternationalWorkshop on Semantic and Social Media Adaptation and Personalization (SMA, IEEE , 2020 , pp. 1 -- 6 . P. R. Borj, K. Raja, P. Bours, On preprocessing the data for improving sexual predator detection: Anonymous for review, in: 2020 15th InternationalWorkshop on Semantic and Social Media Adaptation and Personalization (SMA, IEEE, 2020, pp. 1--6.

Cited by

Sorted by citations

Related

Authors
Rezaee Borj, P., Raja, K., & Bours, P.
Affiliations
  • Norwegian University of Science and Technology (NTNU), Gjøvik, Norway
  • Norwegian University of Science and Technology (NO)
ORCID
Venue
Proceedings of the 9th ACM International Workshop on Security and Privacy Analytics
Publisher
ACM
Published
2023-04-24
Pages
57-65
DOI
10.1145/3579987.3586564
OpenAlex
W4362660680
Semantic Scholar
387616f4e385082bccd9400483c75caf3188914a
Cited by
11
References
16
Citations by year
2026: 3 · 2025: 4 · 2024: 2 · 2023: 2
FWCI
1.95
Type
conference paper
Language
en
Open access
Licence
Locations
ACM International Workshop on Security and Privacy Analytics (IWSPA) (publishedVersion, cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Hate Speech and Cyberbullying Detection
  • Physical Sciences › Computer Science › Information Systems › Spam and Phishing Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Text and Document Classification Technologies
Keywords
child sexual exploitation, contrastive learning, SimCSE, RoBERTa, child online safety

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review