edu.stanford.nlp.sequences

Interface SequenceModel

All Known Subinterfaces:

ListeningSequenceModel

All Known Implementing Classes:

CRFCliqueTree, EmpiricalNERPriorBIO, EntityCachingAbstractSequencePriorBIO, FactoredSequenceModel, TestSentence, TestSequenceModel

public interface SequenceModel

Interface for scoring the labeling of sequences of a fixed length. Each label is represented by integers, typically generated using an Index. Usually, labeling is done via a probability distribution over labels for sequences.

Author:

Teg Grenager (grenager@stanford.edu)

Method Summary

Modifier and Type	Method and Description
int[]	getPossibleValues(int position) Return the valid sequence labels (as integer indices) for a particular position in the sequence.
int	leftWindow() How many label positions to the left influence the label assignment at a particular position.
int	length()
int	rightWindow() How many label positions to the right influence the label assignment at a particular position.
double	scoreOf(int[] sequence) Computes the score assigned by this model to the whole sequence.
double	scoreOf(int[] sequence, int position) Computes the score of the element at the given position in the sequence, conditioned on the values of the elements in all other positions of the provided sequence.
double[]	scoresOf(int[] sequence, int position) Computes the scores of labels for the element at the given position in the sequence, conditioned on the values of the labels at all other positions of the provided sequence.

Method Detail

length

int length()

Returns:

The length of the sequences modeled by this SequenceModel. This is its real length, not including the leftWindow() + rightWindow() of extra padding positions.

leftWindow

int leftWindow()

How many label positions to the left influence the label assignment at a particular position.

Returns:

The size of the left window used by this sequence model

rightWindow

int rightWindow()

How many label positions to the right influence the label assignment at a particular position.

Returns:

the size of the right window used by this sequence model

getPossibleValues

int[] getPossibleValues(int position)

Return the valid sequence labels (as integer indices) for a particular position in the sequence. Since the sequence is padded at each end, typically sequence items 0...leftWindow-1 are null, leftWindow...length+leftWindow-1 are words, length+leftWindow...length+leftWindow+rightWindow-1 are null.

Parameters:

position - The position

Returns:

The set of possible int values at this position, as an int array

scoreOf

double scoreOf(int[] sequence)

Computes the score assigned by this model to the whole sequence. Typically this will be an unnormalized probability in log space (since the probabilities are small).

Parameters:

sequence - The sequence of labels to compute a score for

Returns:

The score for the entire sequence

scoreOf

double scoreOf(int[] sequence,
               int position)

Computes the score of the element at the given position in the sequence, conditioned on the values of the elements in all other positions of the provided sequence. Typically, this is an unnormalized log conditional probability of the label at the given position in the sequence, given the input data nd the other labels.

Parameters:

sequence - The sequence containing the prediction and the rest of the labels to condition on

position - The position of the element to give a score for

Returns:

The score of the label at the specified position in the sequence

scoresOf

double[] scoresOf(int[] sequence,
                  int position)

Computes the scores of labels for the element at the given position in the sequence, conditioned on the values of the labels at all other positions of the provided sequence. The returned array elements correspond index-by-index to the possible values returned by getPossibleValues(int). The label at sequence[position] is ignored in the calculations. The scores are often given as an unnormalized log conditional distribution over possible labels. Otherwise, they may be probabilities. If it is a probability distribution, conceptually, the scores should sum to 1 (perhaps after transforming them). Note that implementations often alter sequence[position] to do their calculations, and so access to the array isn't threadsafe, but should restore its value afterwards.

Parameters:

sequence - The sequence containing the rest of the values to condition on

position - The position of the element to give a distribution for

Returns:

The scores of the possible tokens at the specified position in the sequence