class SentencePairScore(ScoreFunction):
"""
This class provides a score of how close two sentences are to being translations
of each other.
"""
SCORE_MULTIPLIER = 3
def __init__(self):
super(SentencePairScore, self).__init__(0, 1)
self.classifier = None
self.sign = 1
def train(self, pairs, word_score_function):
"""
Trains the sentence pair likelihood score using examples.
`pairs` is an interable of `SentencePair` instances.
`word_score_function` is an instance of ScoreFunction, perhaps even an
instance of `WordPairScore`.
"""
pairs = list(pairs)
self.problem = SentencePairScoreProblem(word_score_function)
self.classifier = SVMClassifier(pairs, self.problem)
class_ = None
for a, b in pairs:
sent = SentencePair(a, b)
score = self.classifier.score(sent)
if score != 0:
class_ = bool(self.classifier.classify(sent)[0])
if (score > 0 and class_ is True) or \
(score < 0 and class_ is False):
self.sign = -1
break
if class_ is None:
raise ValueError("Cannot infer sign with this data")
def __call__(self, a, b):
"""
Returns a score representing how good a
translation sentence b is of sentence a.
"""
if self.classifier is None:
raise LookupError("Score not trained or loaded yet")
a = SentencePair(a, b)
score = self.classifier.score(a) * self.sign
result = self.logistic_function(score * SentencePairScore.SCORE_MULTIPLIER)
assert self.min_bound <= result <= self.max_bound
return result
def logistic_function(self, x):
""" See: http://en.wikipedia.org/wiki/Logistic_function"""
return 1 / (1 + math.e ** (-x))
@property
def word_pair_score(self):
return self.classifier.problem.word_pair_score
def train(self, pairs, word_score_function):
"""
Trains the sentence pair likelihood score using examples.
`pairs` is an interable of `SentencePair` instances.
`word_score_function` is an instance of ScoreFunction, perhaps even an
instance of `WordPairScore`.
"""
pairs = list(pairs)
self.problem = SentencePairScoreProblem(word_score_function)
self.classifier = SVMClassifier(pairs, self.problem)
class_ = None
for a, b in pairs:
sent = SentencePair(a, b)
score = self.classifier.score(sent)
if score != 0:
class_ = bool(self.classifier.classify(sent)[0])
if (score > 0 and class_ is True) or \
(score < 0 and class_ is False):
self.sign = -1
break
if class_ is None:
raise ValueError("Cannot infer sign with this data")
