Collaboration diagram for joshua.decoder.phrase.Candidate:

[legend]

Public Member Functions
boolean	equals (Object obj)
int	hashCode ()
String	toString ()
	Candidate (List< Hypothesis > hypotheses, TargetPhrases phrases, Span span, float delta)
	Candidate (List< Hypothesis > hypotheses, TargetPhrases phrases, Span span, float delta, int[] ranks)
Candidate[]	extend ()
Candidate	extendHypothesis ()
Candidate	extendPhrase ()
int	compareTo (Candidate other)
Span	getSpan ()
Hypothesis	getHypothesis ()
Rule	getRule ()
List< HGNode >	getTailNodes ()
Coverage	getCoverage ()
void	setResult (ComputeNodeResult result)
float	score ()
float	getFutureEstimate ()
List< DPState >	getStates ()
ComputeNodeResult	getResult ()
Public Attributes
Span	span
Package Attributes
float	future_delta
Private Attributes
List< Hypothesis >	hypotheses
TargetPhrases	phrases
int[]	ranks
ComputeNodeResult	result

Constructor & Destructor Documentation

joshua.decoder.phrase.Candidate.Candidate	(	List< Hypothesis >	hypotheses,
		TargetPhrases	phrases,
		Span	span,
		float	delta
	)

Here is the caller graph for this function:

joshua.decoder.phrase.Candidate.Candidate	(	List< Hypothesis >	hypotheses,
		TargetPhrases	phrases,
		Span	span,
		float	delta,
		int[]	ranks
	)

Member Function Documentation

int joshua.decoder.phrase.Candidate.compareTo ( Candidate other )

Here is the call graph for this function:

boolean joshua.decoder.phrase.Candidate.equals ( Object obj )

When candidate objects are extended, the new one is initialized with the same underlying "phrases" and "hypotheses" and "span" objects. So these all have to be equal, as well as the ranks.

This is used to prevent cube pruning from adding the same candidate twice, having reached a point in the cube via different paths.

Here is the call graph for this function:

Candidate [] joshua.decoder.phrase.Candidate.extend ( )

Extends the cube pruning dot in both directions and returns the resulting set. Either of the results can be null if the end of their respective lists is reached.

Returns:: The neighboring candidates (possibly null)

Here is the call graph for this function:

Here is the caller graph for this function:

Candidate joshua.decoder.phrase.Candidate.extendHypothesis ( )

Extends the cube pruning dot along the dimension of existing hypotheses.

Returns:: the next candidate, or null if none

Here is the call graph for this function:

Here is the caller graph for this function:

Candidate joshua.decoder.phrase.Candidate.extendPhrase ( )

Extends the cube pruning dot along the dimension of candidate target sides.

Returns:: the next Candidate, or null if none

Here is the call graph for this function:

Here is the caller graph for this function:

Coverage joshua.decoder.phrase.Candidate.getCoverage ( )

Returns the bit vector of this hypothesis. The bit vector is computed by ORing the coverage vector of the tail node (hypothesis) and the source span of phrases in this candidate.

Returns:

Here is the call graph for this function:

Here is the caller graph for this function:

float joshua.decoder.phrase.Candidate.getFutureEstimate ( )

Here is the call graph for this function:

Here is the caller graph for this function:

Hypothesis joshua.decoder.phrase.Candidate.getHypothesis ( )

A candidate is a (hypothesis, target phrase) pairing. The hypothesis and target phrase are drawn from a list that is indexed by (ranks[0], ranks[1]), respectively. This is a shortcut to return the hypothesis of the candidate pair.

Returns:: the hypothesis at position ranks[0]

Here is the caller graph for this function:

ComputeNodeResult joshua.decoder.phrase.Candidate.getResult ( )

Here is the caller graph for this function:

Rule joshua.decoder.phrase.Candidate.getRule ( )

This returns the target side Phrase, which is a Rule object. This is just a convenience function that works by returning the phrase indexed in ranks[1].

Returns:: the phrase at position ranks[1]

Here is the caller graph for this function:

Span joshua.decoder.phrase.Candidate.getSpan ( )

Returns the input span from which the phrases for this candidates were gathered.

Returns:: the span object

Here is the caller graph for this function:

List<DPState> joshua.decoder.phrase.Candidate.getStates ( )

Here is the call graph for this function:

Here is the caller graph for this function:

List<HGNode> joshua.decoder.phrase.Candidate.getTailNodes ( )

The hypotheses list is a list of tail pointers. This function returns the tail pointer currently selected by the value in ranks.

Returns:: a list of size one, wrapping the tail node pointer

Here is the call graph for this function:

Here is the caller graph for this function:

int joshua.decoder.phrase.Candidate.hashCode ( )

Here is the call graph for this function:

float joshua.decoder.phrase.Candidate.score ( )

This returns the sum of two costs: the HypoState cost + the transition cost. The HypoState cost is in turn the sum of two costs: the Viterbi cost of the underlying hypothesis, and the adjustment to the future score incurred by translating the words under the source phrase being added. The transition cost is the sum of new features incurred along the transition (mostly, the language model costs).

The Future Cost item should probably just be implemented as another kind of feature function, but it would require some reworking of that interface, which isn't worth it.