How to solve the problem of entity nesting, especially for digital types？

[ShelterWFF]

How to solve the problem of entity nesting, especially for digital types？
The replacement in LLM-based approaches is string-based replacement without positional information. It is a practical approximation, not a perfect solution.

[hush-cd]

Thanks for your question.

Currently, MemPrivacy adopts a minimum entity boundary extraction strategy.
For texts containing nested entities, the model will identify all distinguishable privacy-sensitive spans rather than only the outermost entity.

For example, in digital-related cases, different substrings may correspond to different privacy categories or privacy levels, so the model attempts to preserve fine-grained extraction results whenever possible.

This design is intentional for two main reasons:

1. Fine-grained privacy control
   Different nested spans may have different sensitivity levels (PL1–PL4). Extracting all identifiable privacy units enables more flexible downstream protection policies.

2. Semantic utility preservation
   MemPrivacy focuses on privacy protection without destroying task-relevant semantics. Retaining structured and typed privacy spans helps the cloud-side agent maintain contextual reasoning and memory quality.

In practice, this means that nested entities — especially numerical or digital types such as IDs, phone numbers, verification codes, dates, or account-related strings — may all be extracted independently if they are distinguishable privacy instances.

We may further optimize configurable boundary strategies in future releases depending on downstream application requirements.