Anonymization matching strategy

[maks-operlejn-ds]

No description provided.

[mateusz-wosinski-ds]

I'm not sure about current implementation. Shouldn't this method used rather for anonymize than denanonymize ? I guess we want to catch inaccuracies before anonymization will happen, so that the instance matching will be more precise 🤔

e.g. there may be both John Kennedy and John F. Kenedy in the input and we want to anonymize them to the same value. Am I missing something?

[maks-operlejn-ds]

I'm not sure about current implementation. Shouldn't this method used rather for anonymize than denanonymize ? I guess we want to catch inaccuracies before anonymization will happen, so that the instance matching will be more precise 🤔

e.g. there may be both John Kennedy and John F. Kenedy in the input and we want to anonymize them to the same value. Am I missing something?

@mateusz-wosinski-ds I would say we should focus on deanonymize, as in general, we want to pass some anonymized string to LLM and then, match entities of the response, that can be slightly different

But maybe we should think about doing it also in anonymization phase - this would require to rebuild anonymizing part from scratch and I guess at some point I will do it, but for now I would leave it like that. But we can discuss during daily 😄