When set to true
, the Crf will allow the state-space to be dynamically
sized - i.e.
When set to true
, the Crf will allow the state-space to be dynamically
sized - i.e. the number of states is dependent on each sequence
Alpha values.
Alpha values. Need values for each segment length for each label (in general, Semi-CRF case)
Beta values.
Beta values. Need values for each segment length for each label (in general, Semi-CRF case)
Current alpha values used for Forward-Backward computation
Current alpha values used for Forward-Backward computation
The Gaussian prior variance used as a regularizer
The Gaussian prior variance used as a regularizer
The value of the inverse square of the Gaussian prior
The value of the inverse square of the Gaussian prior
Parameter (lambda) vector
For each segment size, the mi matrix holds transition scores for adjacent labels
For each segment size, the mi matrix holds transition scores for adjacent labels
Number of neural gates per label (for NeuralCrf)
Number of neural gates per label (for NeuralCrf)
Number of neural gate input features (for NeuralCrf)
Number of neural gate input features (for NeuralCrf)
Alpha values at the next position used for Forward-Backward computation
Alpha values at the next position used for Forward-Backward computation
Number of features
Number of features
Number of labels/states
Number of labels/states
For each segment size (general case) the ri matrix holds state scores for each label
For each segment size (general case) the ri matrix holds state scores for each label
An array of scaling coefficients to avoid underflow without having to do computations in log space.
An array of scaling coefficients to avoid underflow without having to do computations in log space. See Manning and Schutze Chapter 9 for details (there in the context of HMMs)
The size of segments.
The size of segments. Sizes greater than 1 indicate the model is a semi-CRF