pavan
/
bitext-aligner
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/master'

Browse Source
master
Pavan Mandava 6 years ago
parent d7208088d4 5f26038de8
commit 4217754b84
1 changed files with 55 additions and 33 deletions
Show Stats Download Patch File Download Diff File

88
bitext_align.py
Unescape View File

@ -4,12 +4,13 @@
import os,sys
import re
import pandas as pd
import numpy as np
from numpy import cumsum
from pandas import DataFrame
from nltk import word_tokenize, sent_tokenize
import xml.etree.ElementTree as ET
#import xml.etree.ElementTree as ET
from jellyfish import levenshtein_distance as lev
import six
#import six
from google.cloud import translate_v2 as translate
from itertools import product as cp
@ -24,12 +25,12 @@ translate_client = translate.Client()
def master_align(text0, text1, lang0, lang1):
""" Takes two equivalent texts (original and trnslation) and returns
aligned texts. """
df0 = frame_from_text(text0, source=lang0, target=lang1)
df0 = frame_from_text(text0, lang0, lang1)
print('A')
df1 = frame_from_text(text1, source=lang1, target=lang0, is1=True)
df1 = frame_from_text(text1, lang1, lang0, is1=True)
print('B')
# returns dfs with ['sent', 'trans', 'rellen', 'relpos']
anchors = anchors_from_frames(df0, df1, score_funct, score_threshold, window=2)
anchors = anchors_from_frames(df0, df1, window=2)
print('C')
alignments = intermediate_align(df0, df1, anchors, lookahead=4)
print('D')
@ -38,9 +39,11 @@ def master_align(text0, text1, lang0, lang1):
return textdict0, textdict1
def frame_from_text(text, source='ru', target='en', is1=False): #
def frame_from_text(text, source, target, is1=False): #
""" """ #
#print(source, '-->', target)
cols = [c+str(int(is1)) for c in ['sent','trans','rellen','relpos']]
#print(cols)
frame = pd.DataFrame(columns=cols)
frame[cols[0]] = sent_tokenize(text)
frame[cols[1]] = frame[cols[0]].apply(lambda x: translate_client.translate(x, source_language=source, target_language=target, model='nmt')['translatedText'])
@ -49,35 +52,41 @@ def frame_from_text(text, source='ru', target='en', is1=False): #
cumul_b = list(np.cumsum(frame[cols[2]]))
cumul_a = [0]+cumul_b[:-1]
frame[cols[3]] = pd.Series(list(zip(cumul_a, cumul_b)))
#print(frame[[cols[0], cols[1]]])
return frame
def anchors_from_frames(frame0, frame1, window): #
""" """
pairdf = generate_pairdf(frame0, frame1, window)
pairdf['lev0'] = pairdf[['sent0', 'trans1']].apply(lambda x: trdist(x.sent0, x.trans1))
pairdf['lev1'] = pairdf[['sent1', 'trans0']].apply(lambda x: trdist(x.sent1, x.trans0))
frame0['index0'] = frame0.index
frame1['index1'] = frame1.index
pairdf = pairdf.merge(frame0, on='index0').merge(frame1, on='index1')
pairdf['lev0'] = pairdf.apply(lambda x: trdist(x.sent0, x.trans1), axis=1)
pairdf['lev1'] = pairdf.apply(lambda x: trdist(x.sent1, x.trans0), axis=1)
pairdf['rellen_ratio'] = (pairdf.rellen0/pairdf.rellen1).apply(gr1)
pairdf['minlev'] = pairdf[['lev0', 'lev1']].min(axis=1)
pairdf['maxlev'] = pairdf[['lev0', 'lev1']].min(axis=1)
pairdf['isanchor'] = pairdf.minlev<0.45 & pairdf.maxlev<0.6 & pairdf.rellen_ratio<1.3
return pairdf[pairdf.isanchor][['index0','index1']]
pairdf['isanchor'] = (pairdf.minlev<0.45) & (pairdf.maxlev<0.6) & (pairdf.rellen_ratio<1.3)
return list(pairdf[pairdf.isanchor][['index0','index1']].values)
def intermediate_align(frame0, frame1, anchs, lookahead): #
""" """
aligns = []
end0, end1 = frame0.shape[0], frame1.shape[0]
anchor_ranges = lis(zip([(-1,-1)]+anchs, anchs+[(end0, end1)]))
anchor_ranges = list(zip([(-1,-1)]+anchs, anchs+[(end0, end1)]))
for rang in anchor_ranges:
interaligns = get_interalign(frame0, frame1, *rang, lookahead)
aligns.append(rang[0])
a,b = rang[0]
aligns.append(((a,b),(a,b)))
aligns.extend(interaligns)
return aligns[1:] # format [((i_start, i_end),(j_start, j_end))]
def get_interalign(df0, df1, anchors_init, anchors_next, lookahead): #
""" """
print(anchors_init, anchors_next)
interaligns = []
i,j = anchors_init
i+=1
@ -85,19 +94,25 @@ def get_interalign(df0, df1, anchors_init, anchors_next, lookahead): #
end0, end1 = anchors_next
while i<end0 and j<end1:
room0, room1 = min(end0-i,lookahead), min(end1-j,lookahead)
best_alignment = min([(x,y) for x,y in cp(range(i,i+room0),range(j+room1)) if x==i or y==j], key=score(df0, df1, start0, start1, end0, end1))
interaligns.append((best_alignment))
lambdascore = lambda p,q: score(df0, df1, i, j, p, q)
i_,j_ = min([(x,y) for x,y in cp(range(i,i+room0),range(j,j+room1)) if x==i or y==j], key=lambda a: lambdascore(*a))
print((i,j), (i_,j_))
interaligns.append(((i,j),(i_,j_)))
i,j = i_+1,j_+1
return interaligns
def score(frame0, frame1, start0, start1, end0, end1): #
s0 = ' '.join(frame0.loc[start0:end0+1, 'sent0'])
s1 = ' '.join(frame0.loc[start1:end1+1, 'sent1'])
t0 = ' '.join(frame0.loc[start0:end0+1, 'trans0'])
t1 = ' '.join(frame0.loc[start1:end1+1, 'trans1'])
l0 = sum(frame0.loc[start0:end0+1, 'rellen0'])
l1 = sum(frame1.loc[start1:end1+1, 'rellen1'])
return (lev(s0,t1)+lev(s1,t0))*gr1(l0/l1)/2
#print(frame0.columns)
#print(frame1.columns)
s0 = ' '.join(frame0.loc[start0:end0, 'sent0'])
s1 = ' '.join(frame1.loc[start1:end1, 'sent1'])
t0 = ' '.join(frame0.loc[start0:end0, 'trans0'])
t1 = ' '.join(frame1.loc[start1:end1, 'trans1'])
l0 = sum(frame0.loc[start0:end0, 'rellen0'])
l1 = sum(frame1.loc[start1:end1, 'rellen1'])
#print(s0, s1, t0, t1, l0, l1)
return (trdist(s0,t1)+trdist(s1,t0))*gr1(l0/l1)/2
@ -105,9 +120,13 @@ def score(frame0, frame1, start0, start1, end0, end1): #
def textdicts_from_alignments(frame0, frame1, aligns): #
""" """
textdict0, textdict1 = {},{}
for i,((a0,b0),(a1,b1)) in enumerate(aligns):
t0 = ' '.join(frame0.loc[a0:b0+1, 'sent0'])
t1 = ' '.join(frame0.loc[a1:b1+1, 'sent0'])
for i,((a0,a1),(b0,b1)) in enumerate(aligns):
t0 = ' '.join(frame0.loc[a0:b0, 'sent0'])
t1 = ' '.join(frame1.loc[a1:b1, 'sent1'])
print('***************************')
print(aligns[i])
print(t0)
print(t1)
textdict0.update({i:t0})
textdict1.update({i:t1})
return textdict0, textdict1
@ -119,30 +138,33 @@ def generate_pairdf(frame0, frame1, window):
ranges0 = frame0.relpos0
ranges1 = frame1.relpos1
overlap = [(i,j) for (i,(a,b)),(j,(c,d)) in cp(enumerate(ranges0), enumerate(ranges1)) if get_overlap(a,b,c,d)>0]
len0 = frame0.shape[0]
len1 = frame1.shape[0]
allpairs = []
for i,j in overlap:
for k in range(-window, window+1):
for l in range(-window, window+1):
allpairs.append()
allpairs.append((i+k,j+l))
allpairs = [(a,b) for a,b in allpairs if min(a,b)>-1 and a<len0 and b<len1]
allpairs = sorted(list(set(allpairs)))
pairdf[pairdf.columns] = pd.DataFrame(allpairs).values
pairdf = pd.DataFrame(allpairs).rename(columns={0:'index0', 1:'index1'})
return pairdf
def get_overlap(a,b,c,d):
#print(a0,b0,a1,b1)
if b0>a1 and b0<=b1:
return b0-max(a0,a1)
elif a0>=a1 and a0<b1:
return min(b0,b1)-a0
elif a1>=a0 and a1<b0:
return b1-max(a0,a1)
if b>c and b<=d:
return b-max(a,c)
elif a>=c and a<d:
return min(b,d)-a
elif c>=a and c<b:
return d-max(a,c)
else:
return 0
gr1 = lambda x: 1/less1(x) #
less1 = lambda x: 1/x if abs(x)>1 else x #
trdist = lambda x,y: lev(x,y)/max(x,y) #
trdist = lambda x,y: lev(x,y)/max(len(x),len(y)) #

Write Preview
Loading…
Cancel
Save