This is very likely linked to this issue:
https://issues.apache.org/jira/browse/OPENNLP-154
We changed the way the perceptron is normalizing the probabilities.
Jörn
On 05/22/2012 04:46 PM, Jörn Kottmann wrote:
On 05/22/2012 04:34 PM, Jeyendran Balakrishnan wrote:
Hi Rupert, are you sure you are loading a model from
"en--pos-perceptron.bin", but passing it to a POSTaggerME tagger?
Unless it is a typo in your email, that could be the problem (ie
passing a perceptron model to a maxent classifier)...
OpenNLP is auto detecting the model which is passed in, but that would
easily explain
why the probs are different, e.g using maxent first and then the
perceptron model.
But he mentioned in an earlier mail that he used the *same* model on
both tests.
I will debug this tonight.
Jörn
