mosesdecoder/moses/FeatureFunction.h

169 lines
5.7 KiB
C
Raw Normal View History

#ifndef moses_FeatureFunction_h
#define moses_FeatureFunction_h
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
#include <vector>
#include "ScoreProducer.h"
namespace Moses
{
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
class TargetPhrase;
class TranslationOption;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
class Hypothesis;
class ChartHypothesis;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
class FFState;
class InputType;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
class ScoreComponentCollection;
class WordsBitmap;
class WordsRange;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
2012-09-21 14:56:01 +04:00
/**
* Contains all that a feature function can access without affecting recombination.
* For stateless features, this is all that it can access. Currently this is not
* used for stateful features, as it would need to be retro-fitted to the LM feature.
* TODO: Expose source segmentation,lattice path.
2012-09-21 14:56:01 +04:00
* XXX Don't add anything to the context that would break recombination XXX
**/
class PhraseBasedFeatureContext
{
// The context either has a hypothesis (during search), or a TranslationOption and
// source sentence (during pre-calculation).
const Hypothesis* m_hypothesis;
const TranslationOption& m_translationOption;
const InputType& m_source;
public:
PhraseBasedFeatureContext(const Hypothesis* hypothesis);
PhraseBasedFeatureContext(const TranslationOption& translationOption,
const InputType& source);
const TranslationOption& GetTranslationOption() const;
const InputType& GetSource() const;
const TargetPhrase& GetTargetPhrase() const; //convenience method
const WordsBitmap& GetWordsBitmap() const;
};
2012-09-21 14:56:01 +04:00
/**
* Same as PhraseBasedFeatureContext, but for chart-based Moses.
**/
class ChartBasedFeatureContext
{
//The context either has a hypothesis (during search) or a
//TargetPhrase and source sentence (during pre-calculation)
//TODO: should the context also include some info on where the TargetPhrase
//is anchored (assuming it's lexicalised), which is available at pre-calc?
const ChartHypothesis* m_hypothesis;
const TargetPhrase& m_targetPhrase;
const InputType& m_source;
public:
ChartBasedFeatureContext(const ChartHypothesis* hypothesis);
ChartBasedFeatureContext(const TargetPhrase& targetPhrase,
const InputType& source);
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
2012-09-21 14:56:01 +04:00
const InputType& GetSource() const;
const TargetPhrase& GetTargetPhrase() const;
};
/** base class for all feature functions.
* @todo is this for pb & hiero too?
* @todo what's the diff between FeatureFunction and ScoreProducer?
*/
class FeatureFunction: public ScoreProducer
{
/**< all the score producers in this run */
2013-01-18 22:22:06 +04:00
static std::vector<FeatureFunction*> m_producers;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
public:
2013-01-18 22:22:06 +04:00
static const std::vector<FeatureFunction*>& GetFeatureFunctions() { return m_producers; }
FeatureFunction(const std::string& description, const std::string &line);
FeatureFunction(const std::string& description, size_t numScoreComponents, const std::string &line);
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
virtual bool IsStateless() const = 0;
virtual ~FeatureFunction();
float GetSparseProducerWeight() const { return 1; }
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
};
/** base class for all stateless feature functions.
* eg. phrase table, word penalty, phrase penalty
*/
class StatelessFeatureFunction: public FeatureFunction
{
//All stateless FFs, except those that cache scores in T-Option
static std::vector<const StatelessFeatureFunction*> m_statelessFFs;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
public:
static const std::vector<const StatelessFeatureFunction*>& GetStatelessFeatureFunctions() {return m_statelessFFs;}
2013-02-04 05:09:15 +04:00
StatelessFeatureFunction(const std::string& description, const std::string &line);
StatelessFeatureFunction(const std::string& description, size_t numScoreComponents, const std::string &line);
/**
* This should be implemented for features that apply to phrase-based models.
**/
virtual void Evaluate(const PhraseBasedFeatureContext& context,
ScoreComponentCollection* accumulator) const = 0;
2012-09-21 14:56:01 +04:00
/**
* Same for chart-based features.
**/
virtual void EvaluateChart(const ChartBasedFeatureContext& context,
ScoreComponentCollection* accumulator) const = 0;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
//If true, then the feature is evaluated before search begins, and stored in
2012-09-21 18:00:24 +04:00
//the TranslationOptionCollection.
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
virtual bool ComputeValueInTranslationOption() const;
2012-09-21 18:00:24 +04:00
//!If true, the feature is stored in the ttable, so gets copied into the
//TargetPhrase and does not need cached in the TranslationOption
virtual bool ComputeValueInTranslationTable() const {return false;}
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
bool IsStateless() const;
};
/** base class for all stateful feature functions.
* eg. LM, distortion penalty
*/
class StatefulFeatureFunction: public FeatureFunction
{
//All statefull FFs
static std::vector<const StatefulFeatureFunction*> m_statefulFFs;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
public:
static const std::vector<const StatefulFeatureFunction*>& GetStatefulFeatureFunctions() {return m_statefulFFs;}
StatefulFeatureFunction(const std::string& description, const std::string &line);
StatefulFeatureFunction(const std::string& description, size_t numScoreComponents, const std::string &line);
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
/**
* \brief This interface should be implemented.
* Notes: When evaluating the value of this feature function, you should avoid
* calling hypo.GetPrevHypo(). If you need something from the "previous"
* hypothesis, you should store it in an FFState object which will be passed
* in as prev_state. If you don't do this, you will get in trouble.
*/
virtual FFState* Evaluate(
const Hypothesis& cur_hypo,
const FFState* prev_state,
ScoreComponentCollection* accumulator) const = 0;
virtual FFState* EvaluateChart(
const ChartHypothesis& /* cur_hypo */,
int /* featureID - used to index the state in the previous hypotheses */,
ScoreComponentCollection* accumulator) const = 0;
//! return the state associated with the empty hypothesis for a given sentence
virtual const FFState* EmptyHypothesisState(const InputType &input) const = 0;
Feature function overhaul. Each feature function is computed in one of three ways: 1) Stateless feature functions from the phrase table/generation table: these are computed when the TranslationOption is created. They become part of the ScoreBreakdown object contained in the TranslationOption and are added to the feature value vector when a hypothesis is extended. 2) Stateless feature functions that are computed during state exploration. Currently, only WordPenalty falls into this category, but these functions implement a method Evaluate which do does not receive a Hypothesis or any contextual information. 3) Stateful feature functions: these features receive the arc information (translation option), compute some value and then return some context information. The context information created by a particular feature function is passed back to it as the previous context when a hypothesis originating at the node where the previous edge terminates is created. States in the search space may be recombined if the context information is identical. The context information must be stored in an object implementing the FFState interface. TODO: 1) the command line interface / MERT interface needs to go to named parameters that are otherwise opaque 2) StatefulFeatureFunction's Evaluate method should just take a TranslationOption and a context object. It is not good that it takes a hypothesis, because then people may be tempted to access information about the "previous" hypothesis without "declaring" this dependency. 3) Future cost estimates should be handled using feature functions. All stateful feature functions need some kind of future cost estimate. 4) Philipp's poor-man's cube pruning is broken. git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@2087 1f5c12ca-751b-0410-a591-d2e778427230
2009-02-06 18:43:06 +03:00
bool IsStateless() const;
};
}
#endif