From: Amit J. <ami...@gm...> - 2007-02-08 04:21:06
|
Hi, I will take an example to explain the difference between feature label and value. Suppose, you have 4 labels (starting from 0) - NN, J, VB, DT for doing POS tagging. Now, a sequence can start from only some specific labels and you want to encode this fact as a feature. (Lets call it StartingLabelFeatureType) There will be in all four features for this feature type (one for each class) During training of this feature, you try to determine that which of the labels occur at pos = 0, by looking at the actual labels in training data. Suppose it turns out that only label 0 and 3 occur at the starting position (i.e. NN and DT) Now, while firing features, in startScanFeaturesAt(DataSequence dataSeq, int prev, int pos) if (pos != 0) then you will not fire any features, and will return false. Implicity, that means that all the feature values for pos != 0 will be 0. Now, for pos = 0, you will fire two features, one for label = 0 and one for label = 3 (and setting the feature value = 1) For the rest of the labels, no feature is fired so that implicity means that their feature value is 0. You can achieve the same by setting the feature values like this: i) yend = 0, val = 1, ystart = -1, id = 0 ii) yend = 1, val = 0, ystart = -1, id = 1 iii) yend = 2, val = 0, ystart = -1, id = 2 iv) yend = 3, val = 1, ystart = -1, id = 3 And just to clarify, the actual label of the data is seen only during the training. (And in a FeatureType, the actual label is seen only in the train() method, and not while firing the features) When you are firing features, you fire it for all possible labels for which the feature will hold true. If you search the previous posting on this mailing list, you will get some explanation on this. -amit |