Abstract
To optimize network monitoring and identify high-impact illegal trends, this paper proposes a method for assessing the influence of pornographic cryptic expressions using ChatGPT-4o. By leveraging knowledge graphs to structure cryptic text, the system quantifies influence based on frequency, transmission paths, and the number of individuals involved. ChatGPT-4o is employed for data preprocessing and multi-turn knowledge extraction, enabling the construction of a domain-specific knowledge graph in a low-resource, timely manner. Experimental results demonstrate that this approach effectively captures propagation trends, providing critical support for network regulation and the protection of social morality.
Introduction
Internet regulation faces a persistent challenge in the form of "cryptograms"—highly specialized euphemisms and coded language used to evade detection. These expressions evolve rapidly, rendering traditional manual empirical analysis slow and subjective. From a technical perspective, automated detection is hindered by the scarcity of labeled datasets (low-resource environments). Standard NLP models, such as BERT or Bi-LSTM-CRF, typically require massive annotated corpora to grasp the nuances of underground slang. This research bridges the gap by utilizing the zero-shot and few-shot capabilities of Large Language Models (LLMs) to structure cryptic data and quantify its social threat level.
Methodology
The proposed pipeline for influence assessment consists of four primary stages:
- Data Preprocessing: ChatGPT-4o standardizes vague temporal information (e.g., converting "a few days later" into specific ISO date formats) to establish a clear timeline of events.
- Knowledge Extraction via Multi-turn Q&A: Rather than relying on a single prompt, the system uses a multi-turn Q&A template. The model first identifies entities (users, locations, cryptic types) and subsequently defines the relationships between them. This iterative decomposition significantly improves logical consistency and extraction accuracy.
- Knowledge Graph Construction: Extracted entities and relationships are stored in a Neo4j graph database, facilitating complex path analysis and the visualization of how cryptic phrases propagate.
- Influence Assessment Model: A mathematical formula calculates a comprehensive influence factor ($\omega$) by combining three sub-factors:
- $\omega_1$ (Reach): The total number of people involved.
- $\omega_2$ (Frequency): The recurrence rate of the cryptic phrase.
- $\omega_3$ (Transmission Path/Time): A decay function based on the temporal distance between the occurrence and the assessment point. The model employs exponential saturation ($1 - e^{-x}$) to ensure that outliers do not cause influence scores to explode, maintaining a manageable range for regulatory action.
Datasets
The study utilizes the Yellow Hint Corpus (YHCs), comprising 1,680 cases. This includes 840 public platform entries crawled from Telegram and online forums, and 840 regulatory entries from standardized case files. A subset of 200 entries was manually annotated by experts to serve as the gold standard for evaluation. The corpus spans seven major categories, including "Flight" (referring to aircraft-themed platforms), "Brothel," "Escort," and "Student."
Key Results
ChatGPT-4o demonstrated a significant performance advantage over traditional deep learning models and other domestic LLMs. On judicial datasets, it achieved an F1-score of 71.25% for entity extraction and 67.72% for relationship extraction.
Key findings from the performance analysis include:
- Prompt Engineering Impact: The transition from simple Q&A to multi-turn logic (Template 3) resulted in a massive jump in relationship extraction F1-scores (from ~14% to ~68%).
- Robustness in Informal Contexts: ChatGPT-4o excelled in "Field Cases" (informal Telegram conversations) where traditional models often fail due to the lack of formal linguistic structure.
- Real-world Correlation: The influence factors successfully mirrored real-world events. For instance, "Welfare" cases involving fraudulent sales of protective gear spiked during the COVID-19 pandemic. Additionally, while "Flight" cases were the most frequent, "Escort" scams were found to target high-profile individuals with significantly higher monetary impact.
Conclusion
Combining ChatGPT-4o with Knowledge Graph technology allows regulators to grasp the changing trends of cryptic communication in a timely and objective manner. Future work will focus on expanding the corpus size, exploring cross-lingual cryptogram detection, and integrating the pipeline into live network monitoring systems for real-time early warning capabilities.
Summary updated