def mult_cos(va, vb, vc, vd):
"""
Uses the following formula for scoring:
log(cos(vb, vd)) + log(cos(vc,vd)) - log(cos(va,vd))
"""
first = (1.0 + cosine(vb, vd)) / 2.0
second = (1.0 + cosine(vc, vd)) / 2.0
third = (1.0 + cosine(va, vd)) / 2.0
score = numpy.log(first) + numpy.log(second) - numpy.log(third)
return score
def elementwise_multiplication(va, vb, vc, vd):
"""
Represent the first word-pair by the elementwise multiplication of va and vb.
Do the same for vc and vd. Finally measure the cosine similarity between the
two resultant vectors.
"""
return cosine(va * vb, vc * vd)
def add_cos(va, vb, vc, vd):
"""
Uses the following formula for scoring:
cos(vb - va + vc, vd)
"""
x = normalize(vb - va + vc)
return cosine(x, vd)
def get_correlation(dataset_fname, vects, corr_measure):
"""
Measure the cosine similarities for words in the dataset using their representations
given in vects. Next, compute the correlation coefficient. Specify method form
spearman and pearson.
"""
ignore_missing = False
global VERBOSE
if VERBOSE:
if ignore_missing:
sys.stderr.write("Ignoring missing pairs\n")
else:
sys.stderr.write("Not ignoring missing pairs\n")
mcFile = open(dataset_fname)
mcPairs = {}
mcWords = set()
for line in mcFile:
p = line.strip().split()
mcPairs[(p[0], p[1])] = float(p[2])
mcWords.add(p[0])
mcWords.add(p[1])
mcFile.close()
#print "Total no. of unique words in the dataset =", len(mcWords)
found = mcWords.intersection(set(vects.keys()))
#print "Total no. of words found =", len(found)
missing = []
for x in mcWords:
if x not in vects:
missing.append(x)
human = []
computed = []
found_pairs = False
missing_count = 0
for wp in mcPairs:
(x, y) = wp
if (x in missing or y in missing):
missing_count += 1
if ignore_missing:
continue
else:
comp = 0
else:
found_pairs = True
comp = cosine(vects[x], vects[y])
rating = mcPairs[wp]
human.append(rating)
computed.append(comp)
#print "%s, %s, %f, %f" % (x, y, rating, comp)
if VERBOSE:
sys.stderr.write("Missing pairs = %d (out of %d)\n" %
(missing_count, len(mcPairs)))
if found_pairs is False:
#print "No pairs were scored!"
return (0, 0)
if corr_measure == "pearson":
return scipy.stats.pearsonr(computed, human)
elif corr_measure == "spearman":
return scipy.stats.spearmanr(computed, human)
else:
raise ValueError
pass
def PairDiff(va, vb, vc, vd):
"""
Uses the following formula for scoring:
cos(vd - vc, vb - va)
"""
return cosine(normalize(vd - vc), normalize(vb - va))
def subt_cos(va, vb, vc, vd):
"""
Uses the following formula for scoring:
cos(va - vc, vb - vd)
"""
return cosine(normalize(va - vc), normalize(vb - vd))
def domain_funct(va, vb, vc, vd):
"""
Uses the Formula proposed by Turney in Domain and Function paper.
"""
return numpy.sqrt(
(1.0 + cosine(va, vc)) / 2.0 * (1.0 + cosine(vb, vd)) / 2.0)