On 13/02/14 07:42, Amir Pouya Agha Sadeghi wrote:
Hello Guys,
Ok, I changed the feature generation file and improved my result.Now I want
to combine my open nlp name finder result with my Stanford NER result. My
first idea was a simple voting system but since I have only two system and
it can be a tie, I can not do that. Any one ever did that sort of things
before?
You need to resolve ties somehow...If both your models are probabilistic
you can perhaps look at their prediction confidence values...If you have
a mixture of probabilistic and deterministic models you have to come up
with something else. I've done something similar in the past where I
combined a perceptron model, a maxent model, 3 dictionaries and 11
regex...what I did was to assign 100% confidence to the dictionaries and
regexes...
HTH,
Jim
ps: have a look at my paper for more details "Dealing with data-sparsity
in drug NER"
