Mismatched scores returned from AnalyzerEngine

Describe the bug For some inputs, the results returned by the analyzer have unexpected scores of 1.0 under the attribute , despite no context words being present in the input at all. These scores are different than the scores listed analysis_explanation when return_decision_process = True (these are the expected scores).

To Reproduce

An example:

from presidio_analyzer import AnalyzerEngine
analyzer = AnalyzerEngine()
text = 'You can call my phone [phone]'
results = analyzer.analyze(text , language='en', return_decision_process = True)
print(results)
> [type: UK_NHS, start: 22, end: 34, score: 1.0, type: PHONE_NUMBER, start: 22, end: 34, score: 0.75]
# UK_NHS has a score of 1.0 despite no context words in input, and default pattern score = 0.5

print([i.score for i in results])
> [1.0***, 0.75]

print([i.analysis_explanation.score for i in results])
> [0.5**, 0.75]
# this is the expected score for UK_NHS entity

Expected behavior Scores should match for both attributes.

Additional context

presidio-analyzer==2.2.27
presidio-anonymizer==2.2.27

[liaeh]

@omri374 sure, I’ll create a PR to update the docs.

I agree it would be nice to have the confidence more configurable. In this scenario, since the result validation boosts the score right away to the MAX_SCORE, then there’s no more room for context to be taken into consideration.

[omri374]

Since a checksum is a pretty strong guarantee of a number being a specific entity, we assign the confidence to 1.0. Having said that it could be that a phone number accidentally passes a checksum for another entity. One thing we can do is to have the confidence value more configurable.