Interpretation of High and Low scores for correct entity recognitions

I observed two groups of scores while training named entity recognizer in order to recognize Last Name and First Name in one entity from resume.

Group 1 consist of high scores for correct predictions
Examples:


My interpretation: The high scores mean that the entity recognizer was able to learn from suffficient amount of named entities in such context. I mean that the entity recognizer is familiar enough with context from the above examples.

Group 2 consist of low scores for correct predictions
Examples:



My interpretation: The low scores mean that the entity recognizer was not able to learn from suffficient amount of named entities in such context. I mean that the entity recognizer is not familiar enough with context from the above examples. So I need to show the model more entities in such context in order to increase scores and to make the model more curtain about entities in such context.

Question: Are my interpretations correct?

1 Like

It’s true that we want to see high scores for the true entities, and if we’re seeing low scores, that means the entity might not be returned by the current model. During training the model batches up some examples and sorts them, so that you get runs of likely-correct entities and then runs of likely-incorrect entities. This makes annotation a bit faster, because you have to think a little bit less.