Through some internet sources, I read that partial matching of classes is also important for finding the precision and recall of a Named entity system. Why is it so?
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
If you haven't yet found your answer, I suggest you consider the following articles:
Essentially, we like partial matches because we want to assess the performance of NER, not as an all-or-nothing system, but as a best-effort system to help other systems. The second paper above will make this quite clear. Giving partial credit allows us to measure the performance of the segmentation aspects of NER when entities have multiple tokens.
E.g. you don't get a miss (FN or FP) when your system labels as a VEHICLE "the blue car driving", including the verb. Instead you get 3 TP and 1 FP (if counting only tokens, but not separators). This makes use of F-measure as a metric of overall performance much more "fair", evaluating both the segmentation and the tagging fairly.
