You're right... I should have think a little bit more before bothering you. Giving the precision/recall for each tag has no real meaning.
Having a quick look at a recent paper on the topic, they give also the accuracy on unknown words. This is an idea. Thanks for your quick answer ! On Fri, Oct 7, 2011 at 4:08 PM, Jörn Kottmann <[email protected]> wrote: > On 10/7/11 3:58 PM, Nicolas Hernandez wrote: >> >> Dear Everyone, >> >> I have got a question: why do the PosTaggerEvaluator and the >> PosTaggerCrossValidator programs give an evaluation result in terms of >> Accuracy instead of Precision, Recall and F-measure like for the other >> programs (SentenceDetector, Tokenizer, Chunker) ? >> Does the Accuracy measure correspond to the Precision measure ? >> Personnaly, I ll be interested by having these measures for this program. > > We do you use Recall/Precision in OpenNLP only for components which output > spans, > but the POS Tagger just assigns a tag to a token, and that is either correct > or not. > In this case we just measure how often it was correct, and compute the > accuracy > score based on this. The same is actually done in the doccat component. > > How do you want to compute recall and precision for the POS Tagger? > > Jörn >
