def train(self, bitextGen):
self.frenchAlphabet = Alphabet.from_iterable(
word for frSent, enSent in bitextGen(desc='French Alphabet')
for word in frSent)
self.englishAlphabet = Alphabet.from_iterable(
word for frSent, enSent in bitextGen(desc='English Alphabet')
for word in enSent)
self.frenchAlphabet.freeze()
self.englishAlphabet.freeze()
vF = len(self.frenchAlphabet)
vE = len(self.englishAlphabet)
tOfEGivenF = np.ones((vE, vF)) / vF
aOfIJGivenLenELenF = AlignmentDict()
for ep in tqdm(range(self.epochs), desc='Epoch'):
countOfEGivenF = np.zeros((vE, vF))
totalOfF = np.zeros(vF)
countOfIGivenJ = AlignmentDict()
totalOfJ = CountDict()
for frSent, enSent in bitextGen('Training'):
# Compute Normalization stuff
lenF = len(frSent)
frMask = self.frenchAlphabet.map(frSent)
lenE = len(enSent)
enMask = self.englishAlphabet.map(enSent)
aOfIJ = aOfIJGivenLenELenF[lenE, lenF]
# total probability of each english word being translated from the french ones
# has size of {len(enSent) x 1}
sTotalOfE = np.sum(tOfEGivenF[np.ix_(enMask, frMask)] * aOfIJ,
axis=1,
keepdims=True)
# calculate counts
delta = tOfEGivenF[np.ix_(enMask, frMask)] * aOfIJ / sTotalOfE
deltaSummedOverE = np.sum(delta, axis=0)
countOfEGivenF[np.ix_(enMask, frMask)] += delta
totalOfF[frMask] += deltaSummedOverE
countOfIGivenJ[lenE, lenF] += delta
totalOfJ[lenE, lenF] += deltaSummedOverE
# estimate probabilities
tOfEGivenF = countOfEGivenF / totalOfF
for lenE, lenF in aOfIJGivenLenELenF:
aOfIJGivenLenELenF[
lenE,
lenF] = countOfIGivenJ[lenE, lenF] / totalOfJ[lenE, lenF]
self.tOfEGivenF = tOfEGivenF
self.aOfIJGivenLenELenF = aOfIJGivenLenELenF
def train(self, bitextGen):
self.frenchAlphabet = Alphabet.from_iterable(
word for frSent, enSent in bitextGen(desc='French Alphabet')
for word in frSent)
self.englishAlphabet = Alphabet.from_iterable(
word for frSent, enSent in bitextGen(desc='English Alphabet')
for word in enSent)
self.frenchAlphabet.freeze()
self.englishAlphabet.freeze()
vF = len(self.frenchAlphabet)
vE = len(self.englishAlphabet)
tOfEGivenF = np.ones((vE, vF)) / vF
for ep in tqdm(range(self.epochs), desc='Epoch'):
countOfEGivenF = np.zeros((vE, vF))
totalOfF = np.zeros(vF)
for frSent, enSent in bitextGen('Training'):
# Compute Normalization stuff
frMask = self.frenchAlphabet.map(frSent)
enMask = self.englishAlphabet.map(enSent)
# total probability of each english word being translated from the french ones
# has size of {len(enSent) x 1}
sTotalOfE = np.sum(tOfEGivenF[np.ix_(enMask, frMask)],
axis=1,
keepdims=True)
# calculate counts
delta = tOfEGivenF[np.ix_(enMask, frMask)] / sTotalOfE
countOfEGivenF[np.ix_(enMask, frMask)] += delta
totalOfF[frMask] += np.sum(delta, axis=0)
# estimate probabilities
tOfEGivenF = countOfEGivenF / totalOfF
self.tOfEGivenF = tOfEGivenF
