# Comparing machine learning models with a focus on tone in grooming chat logs

## TL;DR

This paper compares deep learning and traditional machine learning models for detecting grooming conversations online, finding that large language models perform better, especially with positive-toned chats.

## Contribution

The study introduces a novel comparison of deep learning and traditional models for grooming detection, emphasizing the role of tone in improving detection accuracy.

## Key findings

- Large language models outperform traditional machine learning in grooming detection.
- Positive tone improves detection accuracy, while negative tone is harder to distinguish.
- LLaMA 3.2 1B achieves high F1 and F0.5 scores in grooming author detection.

## Abstract

In online spaces, children are vulnerable to exploitation and sexual predators. Groomers contact minors in online chat rooms with the intent of sexual abuse. This study investigates how new deep learning models compare to traditional machine learning models in detecting grooming conversations and predatory authors. Furthermore, we detect the underlying tones used by predators and explore how these affect detection capabilities. Our goal is to better understand predator tactics and to advance automatic grooming detection in order to protect children in online spaces. The PAN12 chat logs, which contain grooming chat conversations, were used as the dataset for the research. These chat conversations were sorted into sentiments through the DistilBERT classifier based on the predator tone. SVMs and the LLaMA 3.2 1B large language model by Meta were then trained and fine-tuned on the different sentiments. The results measured through precision, recall and F1 score show that the large language model performs better in grooming detection than traditional machine learning. Moreover, performance differences between the positive and negative sentiment are captured and indicate that positive tone improves detection while negative toned grooming conversations have nuanced patterns that are harder to distinguish from non-grooming. This shows that groomers employ varying strategies to gain access to their victims. Lastly, with an F1 score of 0.99 and an F0.5 score of 0.99, the LLaMA 3.2 1B model outperforms both traditional machine learning, as well as previous versions of the large language model in grooming author detection.