def test(self, model):
self.model = model
self.beliefs = []
self.partitions = util.generatePossiblePartitions(self.word, self.pronunciation)
for p in self.partitions:
prop = 1.0
for k, f in p:
if k in model:
prop *= model[k].prob(f)
self.beliefs.append(prop)
util.normalize_vector(self.beliefs)
heuritics = util.omegaHeuristics(self.partitions)
beliefMatrix = sorted([(belief, heuritics[i], i) for i, belief in enumerate(self.beliefs)], reverse=True)
self.bestPartition = self.partitions[beliefMatrix[0][2]]
if len(beliefMatrix) == 1:
self.confidence = 10.0
else:
self.confidence = beliefMatrix[0][0] - beliefMatrix[1][0]
self.correctAnswer = True
for k, f in self.bestPartition:
if self.mapping[k] != f:
self.correctAnswer = False
break
def updateWeightVectorOmega(self, smoothing = 0.5):
self.omegas = []
for p in self.partitions:
prop = 1.0
for k, f in p:
prop *= (self._verticesMap[k].prob(f) + smoothing)
self.omegas.append(prop)
util.normalize_vector(self.omegas)
# find the most probable partition
_, index = max([(omega, i) for i, omega in enumerate(self.omegas)])
self.bestPartition = self.partitions[index]
def updateWeightVectorAlpha(self):
self.alphas = []
furiganas = []
for factor in self.factors:
furiganas.extend(factor.mostProbableFuriganas(self.kanji))
furigana_counter = Counter(furiganas)
for factor in self.factors:
alpha = 0.0
for f in factor.mostProbableFuriganas(self.kanji):
alpha += 1.0 / (furigana_counter[f] + 1)
self.alphas.append(alpha)
util.normalize_vector(self.alphas)
def baseline_test(self):
self.partitions = util.generatePossiblePartitions(self.word, self.pronunciation)
self.beliefs = util.omegaHeuristics(self.partitions)
util.normalize_vector(self.beliefs)
beliefMatrix = sorted([(belief, i) for i, belief in enumerate(self.beliefs)], reverse=True)
self.bestPartition = self.partitions[beliefMatrix[0][1]]
if len(beliefMatrix) == 1:
self.confidence = 10.0
else:
self.confidence = beliefMatrix[0][0] - beliefMatrix[1][0]
self.correctAnswer = True
for k, f in self.bestPartition:
if self.mapping[k] != f:
self.correctAnswer = False
break