def length_cost(sx, sy, mean_xy, variance_xy):
"""
Calculate length cost given 2 sentence. Lower cost = higher prob.
The original Gale-Church (1993:pp. 81) paper considers l2/l1 = 1 hence:
delta = (l2-l1*c)/math.sqrt(l1*s2)
If l2/l1 != 1 then the following should be considered:
delta = (l2-l1*c)/math.sqrt((l1+l2*c)/2 * s2)
substituting c = 1 and c = l2/l1, gives the original cost function.
"""
lx, ly = sum(sx), sum(sy)
m = (lx + ly * mean_xy) / 2
try:
delta = (lx - ly * mean_xy) / math.sqrt(m * variance_xy)
except ZeroDivisionError:
return float('-inf')
return - 100 * (LOG2 + norm_logsf(abs(delta)))
def length_cost(sx, sy, mean_xy, variance_xy):
"""
Calculate length cost given 2 sentence. Lower cost = higher prob.
The original Gale-Church (1993:pp. 81) paper considers l2/l1 = 1 hence:
delta = (l2-l1*c)/math.sqrt(l1*s2)
If l2/l1 != 1 then the following should be considered:
delta = (l2-l1*c)/math.sqrt((l1+l2*c)/2 * s2)
substituting c = 1 and c = l2/l1, gives the original cost function.
"""
lx, ly = sum(sx), sum(sy)
m = old_div((lx + ly * mean_xy), 2)
try:
delta = old_div((lx - ly * mean_xy), math.sqrt(m * variance_xy))
except ZeroDivisionError:
return float('-inf')
return -100 * (LOG2 + norm_logsf(abs(delta)))