Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data

arXiv:2601.09717 · cs.CL, cs.AI · Submitted 2025-12-25 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data".

Tom: This study introduces SALP-CG, a large language model-based extraction pipeline designed for classifying and grading privacy risks within large volumes of online conversational health data.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: To get into specifics, the title of this paper is "Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data," and that really sets the stage for what they are trying to achieve. Jane, how does that title reflect the core challenge they’re addressing?

Jane: It points directly to the difficulty in classifying information when it’s embedded within a conversation, which is different from just looking at a single text document. The context matters because health data is often nuanced and conversational, not just simple lists of facts.

Lu: Exactly; the context-aware part means their method isn't just looking for keywords; it’s understanding how those pieces fit together in the dialogue to determine what kind of risk we are dealing with, which is a much deeper level of analysis.

Meng: And that depth is where I see potential for real-world deployment. If a system can understand the context to grade risk accurately, it moves us closer to building automated systems that can actually enforce privacy rules in live interactions.

Lalam: For me, the implication here is about trust and safety; if we can reliably classify what's sensitive in health chats, it helps build better user interfaces and services that respect privacy from the start.

The paper's summary: Tom: So, let’s move into what they actually did in "Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data." Basically, what is the main finding or the core contribution they are presenting to us?

Jane: Their main contribution is SALP-CG, this large language model pipeline that classifies and grades privacy risks in those online health conversations. They achieved this by combining a few techniques: using specific rules, few-shot guidance, and a JSON Schema to make sure the output is structured correctly.

Lu: The summary highlights how they established classification and grading rules that align with national standards like GB/T thirty-nine thousand seven hundred twenty-five-two thousand twenty which is crucial because it brings structure to what was previously a lack of unified standards for this kind of data.

Meng: They also put forward three new metrics specifically to evaluate SALP-CG across different models, which solves a gap in how we measure sensitivity-aware extraction performance. That’s a solid piece of methodological work.

Lalam: The summary emphasizes that they successfully assessed the risks associated with sensitive information exposure from online conversational health data using the MedDialog-CN-two thousand twenty dataset, which gives us concrete evidence of the pipeline's capability.

The paper's improvements: Tom: Now that we know what they did, let’s talk about the specific improvements they suggest in their SALP-CG pipeline and what those enhancements mean for practical use. Jane, can you explain the technical upgrades?

Jane: They focus on making it standard-aligned by ensuring the pipeline is consistent across different models, which means it's backend-agnostic. Plus, they introduce a crucial rule layer that applies deterministic high-risk rules that override the model’s general output when certain patterns appear.

Lu: That rule layer is where I see a lot of creative potential for future work; it shows how we can inject hard, non-negotiable knowledge directly into the LLM process to handle specific high-stakes data points reliably.

Meng: From an engineer's view, this deterministic override is very practical because it means we don't have to rely entirely on the statistical likelihood of a general model; if a known high-risk pattern shows up, we get an accurate flag instantly.

Lalam: This ability to combine the LLM’s flexibility with strict, deterministic rules really addresses the need for reliability in sensitive tasks, which is exactly what we need when dealing with health information governance.

Conclusion: Tom: We’ve covered a lot about SALP-CG and how it tackles classification and grading using established standards. Jane, can you give us the final thoughts on the overall implications of this paper? What's the big picture here?

Jane: Overall, this research confirms that we can use LLMs in a structured way to automate sensitivity classification for health data, which is a huge step toward automated data governance in conversational settings. It validates using these models for tasks that require high reliability against specific regulatory frameworks.

Lu: The broader implication I see is that it opens the door for creating more intelligent, context-aware systems that respect complex privacy laws by building classification directly into the extraction pipeline itself rather than bolting on a simple filter later.

Meng: For practical impact, this means organizations can start deploying tools that can automatically tag and manage sensitive conversational data much more effectively than they could manually. It makes compliance scalable.

Lalam: I think the most exciting part is how this advance in structured extraction methodology helps improve our culture around data handling; it shows us how to build AI systems that are not just powerful, but also fundamentally responsible by design.

Tom: Fantastic points from everyone. So, to wrap up on "Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data," this paper shows a viable path forward for standardizing how we manage risk in health data using modern AI techniques. We’ll keep an eye on how SALP-CG evolves and what comes next in this space.

School of Computing, Macquarie University, Australia · Department of Biomedical Engineering, National University of Singapore · School of Mathematics and Statistics, Shandong University of Technology, China

cs.CL, cs.AI

Submitted: 2025-12-25

Updated: 2026-09-30

Code: https://github.com/dommii1218/SALP-CG

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: This study introduces SALP-CG, a large language model-based extraction pipeline designed for classifying and grading privacy risks within large volumes of online conversational health data.

Key concepts

Data Classification Categories: Health data is grouped into six broad categories: personal attributes, health status, medical applications, medical payment, health resources, and public health. These categories form the basis for understanding the type of information being processed in online conversations.
Sensitivity Levels (1-5): Information is assigned a sensitivity level from 1 to 5. Level 5 represents the highest risk, such as special diseases like STDs, while Level 4 includes direct identifiers like patient names and ID cards. This scale helps quantify the potential harm associated with different data types.
SALP-CG Pipeline: This is an LLM pipeline that extracts triples of (entity, category, level) from health text. It works by prompting the model with rules and examples, using JSON constraints to ensure structured output and deterministic rules to override the model for known high-risk patterns.
Deterministic High-Risk Rules: These are specific, hard-coded rules that force a certain classification regardless of what the general LLM might suggest. For example, if a specific list of HPV genotypes is found with 'positive' results, the system automatically assigns it the highest risk level (Level 5).

Terminology

Benchmark — 10× in this paper · explained in 2 episodes
Ablation — 6× in this paper · explained in 1 episode
Accuracy — 5× in this paper · explained in 1 episode
Ablation study — 3× in this paper · explained in 1 episode
F1 Score — 3× in this paper · explained in 3 episodes
Mean Count Inflation Factor — 3× in this paper
Core — 2× in this paper · explained in 2 episodes
Distillation — 2× in this paper · explained in 2 episodes
General Data Protection Regulation — 2× in this paper
Mean Category-Compatibility Rate — 2× in this paper
Accountability — 1× in this paper · explained in 2 episodes
Hallucination — 1× in this paper · explained in 3 episodes
Health Insurance Portability and Accountability Act — 1× in this paper
Hybrid Approach — 1× in this paper · explained in 4 episodes
Large Language Models (LLMs) — 1× in this paper · explained in 6 episodes
Machine Unlearning — 1× in this paper · explained in 2 episodes

Summary

This study introduces SALP-CG, a large language model-based extraction pipeline designed for classifying and grading privacy risks within large volumes of online conversational health data. This research is significant because existing methods lack unified standards for sensitivity classification in this domain, necessitating robust tools to ensure compliance with regulations like GB/T 39725-2020 and other privacy laws. By combining few-shot guidance, JSON Schema constrained decoding, and deterministic high-risk rules, SALP-CG provides a practical method for achieving strong category compliance and reliable sensitivity across diverse LLMs in health data governance.

Data Classification and Grading Rules

The paper establishes classification rules for online conversational health data aligned with national standards. Health data is divided into six categories: personal attributes, health status, medical applications, medical payment, health resources and public health. These are further mapped to five ascending sensitivity levels (1-5). For example, Level 5 is assigned to special diseases (e.g., STD), while Level 4 includes direct identifiers like patient name and ID card. The paper concludes that the category landscape shows that Level 2-3 items dominate, enabling re-identification when combined; Level 4-5 items are less frequent but carry outsize harm.

LLM-based Extraction Pipeline (SALP-CG)

The SALP-CG pipeline is a standard-aligned LLM pipeline designed to produce triples of (entity, category, level) from online conversational health data. The process involves several layers:

At the generation layer, the model is prompted with an instruction-tuned LLM with a task description, classification and grading rules, few-shot exemplars, and a JSON Scheme. It is instructed to output only JSON and use short Chinese spans as entities.
The schema constraint ensures field C must be drawn from a predefined set, yielding near-zero format errors.
The grading rules are concretely listed, and the level is constrained via the JSON schema to be in the range of [1, 5].
A crucial component is the rule layer, which applies deterministic high-risk rules that override the model when certain patterns are present. For instance, high-risk HPV genotypes (16/18/31/33/35/39/45/51/52/56/58/59) with 'positive' will be mapped to 'Sensitive test result' (level 5).

Evaluation Metrics and Performance

The performance of SALP-CG is assessed using four metrics:

Mean Count Inflation Factor (MCIF): Measures the predicted count of distinct entities, where an MCIF of 1 indicates a match to the gold standard.
Mean Category-Compatibility Rate (MCCR): Evaluates schema compliance, measuring whether the predicted category is compatible with the predefined set.
Mean Sensitivity Grading Quality (MSGR): Measures accuracy for sensitivity levels L ∈ [3, 4, 5].
Micro-F1: Evaluates correctness by casting it as a 5-class classification task (levels 1–5).

Experimental Results and Analysis

The study benchmarks SALP-CG across ten different LLMs on the MedDialog-CN benchmark using four metrics. The results demonstrate that SALP-CG is backend-agnostic and generalizes well across diverse LLMs. Specifically:

(i) Entity Count Inflation (MCIF):

(ii) Schema Category Compliance (MCCR):

The ablation study confirms the contribution of each component, showing that removing few-shot guidance yields the largest drops of 0.344 in MCIF and 0.560 in micro-F1 score, suggesting few-shot is the primary driver of recall and over-reliability. Similarly, removing rules affects high-risk grading with a lower MSGR and lower micro-F1. The category landscape analysis indicates that Level 2 categories are dominated by Chief complaint (30.75%) and Medication Name (12.05%), while Level 4 is most influenced by direct identifiers like Patient’s name (44.13%). The overall finding is that the design successfully balances precision, recall, and governance needs under real-world constraints.

Conclusion

The study concludes that SALP-CG provides a practical and standards-aligned pipeline for health data classification and grading in online conversational health data. It confirms the feasibility of using LLMs for automated sensitivity classification, emphasizing that while Level 4-5 items are less frequent, they pose outsized risks, and Level 2-3 items can enable re-identification when combined. Future work will focus on extending the benchmark to broader datasets and considering actual harms to personal life.

Improvements for AI systems

Here are the specific improvements to AI systems derived from the SALP-CG pipeline, and what these improved systems can achieve:

Improve health data classification accuracy for online medical consultations by achieving strong category compliance across diverse Large Language Models (LLMs). The improved system (SALP-CG) can reliably classify unstructured conversational health data into predefined categories (e.g., Personal Attribute Data, Health Status Data, Special Disease Data) with high fidelity, adhering to national standards like GB/T 39725-2020.
Enhance sensitivity grading reliability by assigning accurate risk levels (1–5) to extracted entities, even for complex or ambiguous mentions. The improved system can minimize level-assignment error by leveraging a combination of few-shot guidance and deterministic high-risk rules, ensuring that high-risk items (like specific positive test results or certain disease mentions) are correctly flagged as Level 5.
Develop a robust, backend-agnostic information extraction pipeline capable of processing conversational data from various LLMs (e.g., GPT series, Llama variants, Qwen). The improved system can integrate seamlessly with different model providers via HTTP API, allowing organizations to deploy classification and grading capabilities without vendor lock-in or reliance on a single proprietary model.
Quantify and manage privacy risks by providing metrics like Mean Count Inflation Factor (MCIF), Mean Category-Compatibility Rate (MCCR), and Micro-F1 score. The improved system allows researchers and data governance teams to objectively measure the precision, recall, schema adherence, and maximum risk prediction capability of any deployed extraction model against a gold standard benchmark.
Enable proactive risk mitigation by identifying patterns that indicate potential re-identification risks through category landscape analysis (e.g., recognizing that Level 2-3 items combined can lead to re-identification). The improved system provides actionable insights into which combinations of data elements pose the greatest threat, guiding targeted privacy policies and technical controls.
Reduce reliance on manual annotation for complex entities by utilizing LLMs in a structured extraction format (JSON Schema constrained decoding) augmented by deterministic rules for high-risk patterns (e.g., specific HPV genotypes). This allows AI systems to automate the identification of highly sensitive, specific medical information that would be difficult or impossible to capture reliably with traditional rule-based NER systems.