Implications of Using Internet Sting Corpora to Approximate Underage Victims

Tatiana Ringenberg, Kathryn Seigfried-Spellar, Julia RayzACL 2021

TL;DR

This paper investigates whether internet sting operations, commonly used to train NLP models for child exploitation detection, accurately represent real underage victim conversations. By annotating a corpus of victim, vigilante, and law enforcement officer (LEO) chats, the authors found significant differences in grooming tactics and stages, suggesting that current NLP models might be biased due to reliance on potentially unrepresentative sting data.

What Problem Does This Paper Address?

Law enforcement agencies are increasingly overwhelmed by online child exploitation cases. Natural Language Processing (NLP) models are being developed to help triage these cases by analyzing textual communications between predators and victims. However, training these models requires large, annotated datasets. Due to the sensitive and emotionally taxing nature of real child exploitation data, researchers often use transcripts from internet sting operations. In these stings, either law enforcement officers (LEOs) or adult vigilantes impersonate minors to elicit incriminating evidence from predators. The core problem this paper addresses is the implicit assumption that these sting operation transcripts accurately reflect the dynamics of actual child victimization. Psychology research suggests that the goals and methods of predators interacting with real underage victims may differ significantly from those interacting with LEOs or vigilantes, potentially leading to biased NLP models.

Methods

The researchers developed a methodology for annotating online grooming conversations. They created a codebook based on established psychological research on online grooming stages and tactics. This codebook was then used to annotate a corpus of chat transcripts. The annotation process involved:

  • Corpus Construction: A corpus of 60 chat transcripts was compiled, with 20 each from underage victims, LEOs, and vigilantes, all interacting with convicted predators.
  • Annotation Protocol: A two-annotator system was employed. One annotator was an expert in child exploitation, and the other was an undergraduate with limited experience. This approach aimed to ensure the corpus was accessible to researchers with diverse backgrounds.
  • Training and Refinement: Annotators underwent training, reviewed the codebook, and annotated practice chats. Disagreements were resolved through discussion. The codebook itself was iteratively refined (a process called "neutering") to combine similar tactics and stages that were difficult to distinguish.
  • Quantitative Analysis: Statistical methods, including ANOVA and chi-squared tests, were used to compare chat length and the presence of specific grooming tactics across the three groups (victim, vigilante, LEO).
  • Qualitative Analysis: Annotator observations on language and conversational differences were qualitatively analyzed to provide deeper insights into the nuances of each group's interactions.

Datasets

The paper introduces a newly curated corpus for research on online child exploitation. This corpus consists of:

  • 60 chat transcripts:
    • 20 transcripts from interactions involving underage victims and predators.
  • 20 transcripts from interactions involving law enforcement officers (LEOs) impersonating minors and predators.
  • 20 transcripts from interactions involving adult vigilantes impersonating minors and predators.

This dataset is distinct from many existing corpora in that it explicitly includes a comparison between real victim interactions and those conducted by LEOs and vigilantes, addressing a known limitation in prior research.

Key Insights and Ideas

  • Sting Data May Not Be Representative: The central finding is that internet sting operations, while useful, do not perfectly mirror real underage victimizations. This has direct implications for NLP models trained on such data.
  • Differing Grooming Tactics: Predators employ different tactics and exhibit different conversational patterns when interacting with real victims versus LEOs/vigilantes. This suggests that NLP models might learn to identify predator behavior based on characteristics that are more prevalent in sting operations than in actual criminal activity.
  • LEO/Vigilante "Priming" Techniques: LEOs and vigilantes, in their attempts to elicit evidence, sometimes use specific techniques (like describing clothing when asked for pictures) that real victims would not typically employ. This can influence predator behavior and the conversation's trajectory.
  • "Reverse Power" as a Strategy: The concept of "reverse power" (giving control to the other participant) was observed as a tactic that LEOs and vigilantes might use to influence conversations without explicitly initiating certain discussions, potentially to avoid detection.
  • Age and Sexual History Differences: The paper suggests that predators might use "age difference" and "sexual history" tactics differently depending on whether they are interacting with a real victim (who might already know the predator) or an LEO/vigilante.

Key Results

  • Chat Length Differences: A significant difference was found in chat length across the three groups, with vigilante chats being significantly different from both LEO and victim chats.
  • Tactical Usage Variations: Statistical tests revealed significant differences in the use of several grooming tactics, including:
  • Sexual History
    • Willingness
    • Discussion of Phone Calls
    • Acknowledgement of Age Difference
    • Personal Compliments
    • Reverse Power
  • Vigilante Conversations Most Active: Vigilante conversations generally showed the highest occurrence of most grooming tactics, often including all tactics from the analyzed set.
  • LEO Conversations Shortest and Vague: LEO conversations tended to be shorter and more vague, particularly in responses to direct questions about meetings or sexual activities.
  • Victim Conversations Less Direct: Victim conversations showed less frequent use of tactics like "sexual history" and "age difference," possibly because the predator already had this information.
  • Avoidance Tactics: Real victims rarely used avoidance tactics (e.g., citing family plans) to deflect unwanted interactions, unlike LEOs and vigilantes.

Limitations and Future Work

  • Corpus Size: The authors acknowledge that their corpus of 60 chats is small and not fully representative ("saturated") of all possible interactions in the domain.
  • Single Annotator for Remaining Data: While the initial annotation involved two annotators, the majority of the corpus was annotated by a single expert annotator. Future work should ideally use multiple annotators for the entire dataset.
  • Partial Membership: The study focused on "full membership" annotations, excluding "partial membership" annotations which were used for discussion. Incorporating these could enrich future analysis.
  • Implicit Denial of Advance: The authors found it difficult to consistently distinguish between "sexual" and "non-consensual sexual" stages, leading them to "neuter" these into a single stage. Future research could investigate the presence of a distinct non-consensual sexual stage.
  • Cross-Country Comparisons: The study focused on U.S. transcripts to minimize confounding variables. Future work could explore cross-country differences in grooming tactics and legal frameworks.
  • Dynamic Analysis: The paper notes that the intensity and graphic content of sexual tactics increased over time in longer conversations. Future research should examine these changes dynamically.

Summary updated

BibTeX
@inproceedings{ringenberg2021implications,
  author = {Tatiana Ringenberg and Kathryn Seigfried-Spellar and Julia Rayz},
  title = {{Implications of Using Internet Sting Corpora to Approximate Underage Victims}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure