I'm new to machine learning so I apologize if this question is silly. I'm using Stanford NER's english 4class classifier with good results. However, since my dataset is mostly focused on organizations, I think the results could be improved if I could boost the probability for an entity to be an organization to the detriment of other classes.(Ex: I would prefer "Carl Zeis" to be identified as an organization rather than a person). Is my supposition correct? If so, can it be achieved in an easier way than retraining the model?
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Your data set will influence the labeling results. If it is focused on organizations, the NER should favor them simply by virtue of the data it's fed. So you might need to do anything. But if you do observe undesirable behavior in the resulting NER, you can adjust the weights.
It's the same with any machine learning algorithm. A humorous demonstration was provided by Google's Deep Dream: it kept seeing dogs every where. Why? Because the data set they used for training had an abundance of dogs.
(And Carl Zeis should be labeled as a person. The company is Carl Zeiss.)
