def ScorePhonemes(self, source=[], target=[]):
"""Compare the phonemes of a source and target sentence and determine
which of the target items were correctly transcribed
Returns:
hits_phonemes (nested list): list of bools corresponding to the accuracy
of each phoneme in the target list for each sentence
Note:
This scoring method has no word accuracy awareness. Phonemes from correctly input
words may wind up as labeled wrong ( i.e. target:"with the" source: "with a" alignement: )
Modified from Eser Aygün (https://pypi.python.org/pypi/alignment/1.0.9)
"""
if not source:
source = self.source_phonemes
if not target:
target = self.target_phonemes
self.source_matched = []
hits = []
for x, ttup in enumerate(target):
tphon, twordnum, tword = zip(*ttup)
stup = source[x]
if not stup:
hitlist = [False] * len(tphon)
bPhonOut = ['-'] * len(tphon)
else:
sphon, swordnum, sword = zip(*stup)
# Create sequences to be aligned.
a = Sequence(tphon)
b = Sequence(sphon)
# Create a vocabulary and encode the sequences.
v = Vocabulary()
aEncoded = v.encodeSequence(a)
bEncoded = v.encodeSequence(b)
# Create a scoring and align the sequences using global aligner.
scoring = SimpleScoring(2, -1)
aligner = GlobalSequenceAligner(scoring, -2)
score, encodeds = aligner.align(aEncoded,
bEncoded,
backtrace=True)
encoded = encodeds[0]
#Score based only on hits vs misses, insertions are ignored
notInsert = encoded[:][0] != 0
nonInsertMatched = encoded[notInsert][:]
#Find the alignment in the target sequence
aSeq = nonInsertMatched[:][0]
bSeq = nonInsertMatched[:][1]
#Label all items not aligned to the target as false
hitlist = []
y = 0
for y in range(0, len(aEncoded) - len(aSeq) + 1):
aChunk = aEncoded[y:y + len(aSeq)]
#print aChunk
if sum(aChunk - aSeq) == 0:
break
hitlist.extend([False] * (y))
hitlist.extend(list(aSeq - bSeq == 0))
hitlist.extend([False] * (len(aEncoded) - y - len(aSeq)))
#Export the target aligned phonemes of the source sequence
bPhons = np.zeros(len(aEncoded), int)
bPhons[y:y + len(bSeq)] = bSeq
bPhonOut = np.array(v.elements())[bPhons].tolist()
hits.append(hitlist)
self.source_matched.append(bPhonOut)
self.hits_phonemes = hits
def ScoreWords(self):
"""Aligns the words of the source sentence to match the target sentence
to determine hit vs missed words
Returns:
hits (nested list): The target [0] and source [1] sentences in a nested list
Note:
Modified from Eser Aygün (https://pypi.python.org/pypi/alignment/1.0.9)
"""
target = self.target
source = self.source
self.source_matchWords = []
hits = []
wscore = np.empty(0)
for tnum, tsent in enumerate(target):
ssent = source[tnum]
# Create sequences to be aligned.
a = Sequence(tsent.split())
b = Sequence(ssent.split())
# Create a vocabulary and encode the sequences.
v = Vocabulary()
aEncoded = v.encodeSequence(a)
bEncoded = v.encodeSequence(b)
# Create a scoring and align the sequences using global aligner.
scoring = SimpleScoring(5, -1)
aligner = GlobalSequenceAligner(scoring, -1)
score, encodeds = aligner.align(aEncoded, bEncoded, backtrace=True)
encoded = encodeds[0]
#Score based only on hits vs misses, insertions are ignored
notInsert = encoded[:][0] != 0
nonInsertMatched = encoded[notInsert][:]
#Find the alignment in the target sequence
aSeq = nonInsertMatched[:][0]
bSeq = nonInsertMatched[:][1]
#Label all items not aligned to the target as false
hitlist = []
x = 0
for x in range(0, len(aEncoded) - len(aSeq) + 1):
aChunk = aEncoded[x:x + len(aSeq)]
#print aChunk
if sum(aChunk - aSeq) == 0:
break
hitlist.extend([False] * (x))
hitlist.extend(list(aSeq - bSeq == 0))
hitlist.extend([False] * (len(aEncoded) - x - len(aSeq)))
#Export the target aligned words of the source sequence
bWords = np.zeros(len(aEncoded), int)
bWords[x:x + len(bSeq)] = bSeq
bWordOut = np.array(v.elements())[bWords].tolist()
hits.append(hitlist)
iwscore = sum(hitlist) * 100 / float(len(hitlist))
wscore = np.hstack([wscore, iwscore])
print bWordOut
self.source_matchWords.append(bWordOut)
self.hits = hits
self.wscore = wscore