-
Notifications
You must be signed in to change notification settings - Fork 5
/
takahe.py
1488 lines (1119 loc) · 58.6 KB
/
takahe.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#!/usr/bin/python
# -*- coding: utf-8 -*-
"""
:Name:
takahe
:Authors:
Florian Boudin ([email protected])
:Version:
0.4
:Date:
Mar. 2013
:Description:
takahe is a multi-sentence compression module. Given a set of redundant
sentences, a word-graph is constructed by iteratively adding sentences to
it. The best compression is obtained by finding the shortest path in the
word graph. The original algorithm was published and described in
[filippova:2010:COLING]_. A keyphrase-based reranking method, described in
[boudin-morin:2013:NAACL]_ can be applied to generate more informative
compressions.
.. [filippova:2010:COLING] Katja Filippova, Multi-Sentence Compression:
Finding Shortest Paths in Word Graphs, *Proceedings of the 23rd
International Conference on Computational Linguistics (Coling 2010)*,
pages 322-330, 2010.
.. [boudin-morin:2013:NAACL] Florian Boudin and Emmanuel Morin, Keyphrase
Extraction for N-best Reranking in Multi-Sentence Compression,
*Proceedings of the 2013 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies
(NAACL-HLT 2013)*, 2013.
:History:
Development history of the takahe module:
- 0.4 (Mar. 2013) adding the keyphrase-based nbest reranking algorithm
- 0.33 (Feb. 2013), bug fixes and better code documentation
- 0.32 (Jun. 2012), Punctuation marks are now considered within the
graph, compressions are then punctuated
- 0.31 (Nov. 2011), modified context function (uses the left and right
contexts), improved docstring documentation, bug fixes
- 0.3 (Oct. 2011), improved K-shortest paths algorithm including
verb/size constraints and ordered lists for performance
- 0.2 (Dec. 2010), removed dependencies from nltk (i.e. POS-tagging,
tokenization and stopwords removal)
- 0.1 (Nov. 2010), first version
:Dependencies:
The following Python modules are required:
- `networkx <http://networkx.github.com/>`_ for the graph construction
(v1.2+)
:Usage:
A typical usage of this module is::
import takahe
# A list of tokenized and POS-tagged sentences
sentences = ['Hillary/NNP Clinton/NNP wanted/VBD to/stop visit/VB ...']
# Create a word graph from the set of sentences with parameters :
# - minimal number of words in the compression : 6
# - language of the input sentences : en (english)
# - POS tag for punctuation marks : PUNCT
compresser = takahe.word_graph( sentences,
nb_words = 6,
lang = 'en',
punct_tag = "PUNCT" )
# Get the 50 best paths
candidates = compresser.get_compression(50)
# 1. Rerank compressions by path length (Filippova's method)
for cummulative_score, path in candidates:
# Normalize path score by path length
normalized_score = cummulative_score / len(path)
# Print normalized score and compression
print round(normalized_score, 3), ' '.join([u[0] for u in path])
# Write the word graph in the dot format
compresser.write_dot('test.dot')
# 2. Rerank compressions by keyphrases (Boudin and Morin's method)
reranker = takahe.keyphrase_reranker( sentences,
candidates,
lang = 'en' )
reranked_candidates = reranker.rerank_nbest_compressions()
# Loop over the best reranked candidates
for score, path in reranked_candidates:
# Print the best reranked candidates
print round(score, 3), ' '.join([u[0] for u in path])
:Misc:
The Takahe is a flightless bird indigenous to New Zealand. It was thought to
be extinct after the last four known specimens were taken in 1898. However,
after a carefully planned search effort the bird was rediscovered by on
November 20, 1948. (Wikipedia, http://en.wikipedia.org/wiki/takahe)
"""
import math
import codecs
import os
import re
import sys
import bisect
import networkx as nx
from networkx.drawing.nx_pydot import write_dot
#import matplotlib.pyplot as plt
#~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
# [ Class word_graph
#~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
class word_graph:
"""
The word_graph class constructs a word graph from the set of sentences given
as input. The set of sentences is a list of strings, sentences are tokenized
and words are POS-tagged (e.g. ``"Saturn/NNP is/VBZ the/DT sixth/JJ
planet/NN from/IN the/DT Sun/NNP in/IN the/DT Solar/NNP System/NNP"``).
Four optional parameters can be specified:
- nb_words is is the minimal number of words for the best compression
(default value is 8).
- lang is the language parameter and is used for selecting the correct
stopwords list (default is "en" for english, stopword lists are localized
in /resources/ directory).
- punct_tag is the punctuation mark tag used during graph construction
(default is PUNCT).
"""
#-T-----------------------------------------------------------------------T-
def __init__(self, sentence_list, nb_words=8, lang="en", punct_tag="PUNCT", pos_separator='/'):
self.sentence = list(sentence_list)
""" A list of sentences provided by the user. """
self.length = len(sentence_list)
""" The number of sentences given for fusion. """
self.nb_words = nb_words
""" The minimal number of words in the compression. """
self.resources = os.path.dirname(__file__) + '/resources/'
""" The path of the resources folder. """
self.stopword_path = self.resources+'stopwords.'+lang+'.dat'
""" The path of the stopword list, e.g. stopwords.[lang].dat. """
self.stopwords = self.load_stopwords(self.stopword_path)
""" The set of stopwords loaded from stopwords.[lang].dat. """
self.punct_tag = punct_tag
""" The stopword tag used in the graph. """
self.pos_separator = pos_separator
""" The character (or string) used to separate a word and its Part of Speech tag """
self.graph = nx.DiGraph()
""" The directed graph used for fusion. """
self.start = '-start-'
""" The start token in the graph. """
self.stop = '-end-'
""" The end token in the graph. """
self.sep = '/-/'
""" The separator used between a word and its POS in the graph. """
self.term_freq = {}
""" The frequency of a given term. """
self.verbs = set(['VB', 'VBD', 'VBP', 'VBZ', 'VH', 'VHD', 'VHP', 'VBZ',
'VV', 'VVD', 'VVP', 'VVZ'])
"""
The list of verb POS tags required in the compression. At least *one*
verb must occur in the candidate compressions.
"""
# Replacing default values for French
if lang == "fr":
self.verbs = set(['V', 'VPP', 'VINF'])
# 1. Pre-process the sentences
self.pre_process_sentences()
# 2. Compute term statistics
self.compute_statistics()
# 3. Build the word graph
self.build_graph()
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def pre_process_sentences(self):
"""
Pre-process the list of sentences given as input. Split sentences using
whitespaces and convert each sentence to a list of (word, POS) tuples.
"""
for i in range(self.length):
# Normalise extra white spaces
self.sentence[i] = re.sub(' +', ' ', self.sentence[i])
self.sentence[i] = self.sentence[i].strip()
# Tokenize the current sentence in word/POS
sentence = self.sentence[i].split(' ')
# Creating an empty container for the cleaned up sentence
container = [(self.start, self.start)]
# Looping over the words
for w in sentence:
# Splitting word, POS
pos_separator_re = re.escape(self.pos_separator)
m = re.match("^(.+)" +pos_separator_re +"(.+)$", w)
if m is not None:
# Extract the word information
token, POS = m.group(1), m.group(2)
# Add the token/POS to the sentence container
container.append((token.lower(), POS))
# Add the stop token at the end of the container
container.append((self.stop, self.stop))
# Recopy the container into the current sentence
self.sentence[i] = container
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def build_graph(self):
"""
Constructs a directed word graph from the list of input sentences. Each
sentence is iteratively added to the directed graph according to the
following algorithm:
- Word mapping/creation is done in four steps:
1. non-stopwords for which no candidate exists in the graph or for
which an unambiguous mapping is possible or which occur more than
once in the sentence
2. non-stopwords for which there are either several possible
candidates in the graph
3. stopwords
4. punctuation marks
For the last three groups of words where mapping is ambiguous we check
the immediate context (the preceding and following words in the sentence
and the neighboring nodes in the graph) and select the candidate which
has larger overlap in the context, or the one with a greater frequency
(i.e. the one which has more words mapped onto it). Stopwords are mapped
only if there is some overlap in non-stopwords neighbors, otherwise a
new node is created. Punctuation marks are mapped only if the preceding
and following words in the sentence and the neighboring nodes are the
same.
- Edges are then computed and added between mapped words.
Each node in the graph is represented as a tuple ('word/POS', id) and
possesses an info list containing (sentence_id, position_in_sentence)
tuples.
"""
# Iteratively add each sentence in the graph ---------------------------
for i in range(self.length):
# Compute the sentence length
sentence_len = len(self.sentence[i])
# Create the mapping container
mapping = [0] * sentence_len
#-------------------------------------------------------------------
# 1. non-stopwords for which no candidate exists in the graph or for
# which an unambiguous mapping is possible or which occur more
# than once in the sentence.
#-------------------------------------------------------------------
for j in range(sentence_len):
# Get the word and tag
token, POS = self.sentence[i][j]
# If stopword or punctuation mark, continues
if token in self.stopwords or re.search('(?u)^\W$', token):
continue
# Create the node identifier
node = token.lower() + self.sep + POS
# Find the number of ambiguous nodes in the graph
k = self.ambiguous_nodes(node)
# If there is no node in the graph, create one with id = 0
if k == 0:
# Add the node in the graph
self.graph.add_node( (node, 0), info=[(i, j)],
label=token.lower() )
# Mark the word as mapped to k
mapping[j] = (node, 0)
# If there is only one matching node in the graph (id is 0)
elif k == 1:
# Get the sentences id of this node
ids = []
for sid, pos_s in self.graph.nodes[(node, 0)]['info']:
ids.append(sid)
# Update the node in the graph if not same sentence
if not i in ids:
self.graph.nodes[(node, 0)]['info'].append((i, j))
mapping[j] = (node, 0)
# Else Create new node for redundant word
else:
self.graph.add_node( (node, 1), info=[(i, j)],
label=token.lower() )
mapping[j] = (node, 1)
#-------------------------------------------------------------------
# 2. non-stopwords for which there are either several possible
# candidates in the graph.
#-------------------------------------------------------------------
for j in range(sentence_len):
# Get the word and tag
token, POS = self.sentence[i][j]
# If stopword or punctuation mark, continues
if token in self.stopwords or re.search('(?u)^\W$', token):
continue
# If word is not already mapped to a node
if mapping[j] == 0:
# Create the node identifier
node = token.lower() + self.sep + POS
# Create the neighboring nodes identifiers
prev_token, prev_POS = self.sentence[i][j-1]
next_token, next_POS = self.sentence[i][j+1]
prev_node = prev_token.lower() + self.sep + prev_POS
next_node = next_token.lower() + self.sep + next_POS
# Find the number of ambiguous nodes in the graph
k = self.ambiguous_nodes(node)
# Search for the ambiguous node with the larger overlap in
# context or the greater frequency.
ambinode_overlap = []
ambinode_frequency = []
# For each ambiguous node
for l in range(k):
# Get the immediate context words of the nodes
l_context = self.get_directed_context(node, l, 'left')
r_context = self.get_directed_context(node, l, 'right')
# Compute the (directed) context sum
val = l_context.count(prev_node)
val += r_context.count(next_node)
# Add the count of the overlapping words
ambinode_overlap.append(val)
# Add the frequency of the ambiguous node
ambinode_frequency.append(
len( self.graph.nodes[(node, l)]['info'] )
)
# Search for the best candidate while avoiding a loop
found = False
selected = 0
while not found:
# Select the ambiguous node
selected = self.max_index(ambinode_overlap)
if ambinode_overlap[selected] == 0:
selected = self.max_index(ambinode_frequency)
# Get the sentences id of this node
ids = []
for sid, p in self.graph.nodes[(node, selected)]['info']:
ids.append(sid)
# Test if there is no loop
if i not in ids:
found = True
break
# Remove the candidate from the lists
else:
del ambinode_overlap[selected]
del ambinode_frequency[selected]
# Avoid endless loops
if len(ambinode_overlap) == 0:
break
# Update the node in the graph if not same sentence
if found:
self.graph.nodes[(node, selected)]['info'].append((i, j))
mapping[j] = (node, selected)
# Else create new node for redundant word
else:
self.graph.add_node( (node, k), info=[(i, j)],
label=token.lower() )
mapping[j] = (node, k)
#-------------------------------------------------------------------
# 3. map the stopwords to the nodes
#-------------------------------------------------------------------
for j in range(sentence_len):
# Get the word and tag
token, POS = self.sentence[i][j]
# If *NOT* stopword, continues
if not token in self.stopwords :
continue
# Create the node identifier
node = token.lower() + self.sep + POS
# Find the number of ambiguous nodes in the graph
k = self.ambiguous_nodes(node)
# If there is no node in the graph, create one with id = 0
if k == 0:
# Add the node in the graph
self.graph.add_node( (node, 0), info=[(i, j)],
label=token.lower() )
# Mark the word as mapped to k
mapping[j] = (node, 0)
# Else find the node with overlap in context or create one
else:
# Create the neighboring nodes identifiers
prev_token, prev_POS = self.sentence[i][j-1]
next_token, next_POS = self.sentence[i][j+1]
prev_node = prev_token.lower() + self.sep + prev_POS
next_node = next_token.lower() + self.sep + next_POS
ambinode_overlap = []
# For each ambiguous node
for l in range(k):
# Get the immediate context words of the nodes, the
# boolean indicates to consider only non stopwords
l_context = self.get_directed_context(node, l, 'left',\
True)
r_context = self.get_directed_context(node, l, 'right',\
True)
# Compute the (directed) context sum
val = l_context.count(prev_node)
val += r_context.count(next_node)
# Add the count of the overlapping words
ambinode_overlap.append(val)
# Get best overlap candidate
selected = self.max_index(ambinode_overlap)
# Get the sentences id of the best candidate node
ids = []
for sid, pos_s in self.graph.nodes[(node, selected)]['info']:
ids.append(sid)
# Update the node in the graph if not same sentence and
# there is at least one overlap in context
if i not in ids and ambinode_overlap[selected] > 0:
# if i not in ids and \
# (ambinode_overlap[selected] > 1 and POS==self.punct_tag) or\
# (ambinode_overlap[selected] > 0 and POS!=self.punct_tag) :
# Update the node in the graph
self.graph.nodes[(node, selected)]['info'].append((i, j))
# Mark the word as mapped to k
mapping[j] = (node, selected)
# Else create a new node
else:
# Add the node in the graph
self.graph.add_node( (node, k) , info=[(i, j)],
label=token.lower() )
# Mark the word as mapped to k
mapping[j] = (node, k)
#-------------------------------------------------------------------
# 4. lasty map the punctuation marks to the nodes
#-------------------------------------------------------------------
for j in range(sentence_len):
# Get the word and tag
token, POS = self.sentence[i][j]
# If *NOT* punctuation mark, continues
if not re.search('(?u)^\W$', token):
continue
# Create the node identifier
node = token.lower() + self.sep + POS
# Find the number of ambiguous nodes in the graph
k = self.ambiguous_nodes(node)
# If there is no node in the graph, create one with id = 0
if k == 0:
# Add the node in the graph
self.graph.add_node( (node, 0), info=[(i, j)],
label=token.lower() )
# Mark the word as mapped to k
mapping[j] = (node, 0)
# Else find the node with overlap in context or create one
else:
# Create the neighboring nodes identifiers
prev_token, prev_POS = self.sentence[i][j-1]
next_token, next_POS = self.sentence[i][j+1]
prev_node = prev_token.lower() + self.sep + prev_POS
next_node = next_token.lower() + self.sep + next_POS
ambinode_overlap = []
# For each ambiguous node
for l in range(k):
# Get the immediate context words of the nodes
l_context = self.get_directed_context(node, l, 'left')
r_context = self.get_directed_context(node, l, 'right')
# Compute the (directed) context sum
val = l_context.count(prev_node)
val += r_context.count(next_node)
# Add the count of the overlapping words
ambinode_overlap.append(val)
# Get best overlap candidate
selected = self.max_index(ambinode_overlap)
# Get the sentences id of the best candidate node
ids = []
for sid, pos_s in self.graph.nodes[(node, selected)]['info']:
ids.append(sid)
# Update the node in the graph if not same sentence and
# there is at least one overlap in context
if i not in ids and ambinode_overlap[selected] > 1:
# Update the node in the graph
self.graph.nodes[(node, selected)]['info'].append((i, j))
# Mark the word as mapped to k
mapping[j] = (node, selected)
# Else create a new node
else:
# Add the node in the graph
self.graph.add_node( (node, k), info=[(i, j)],
label=token.lower() )
# Mark the word as mapped to k
mapping[j] = (node, k)
#-------------------------------------------------------------------
# 4. Connects the mapped words with directed edges
#-------------------------------------------------------------------
for j in range(1, len(mapping)):
self.graph.add_edge(mapping[j-1], mapping[j])
# Assigns a weight to each node in the graph ---------------------------
for node1, node2 in self.graph.edges():
edge_weight = self.get_edge_weight(node1, node2)
self.graph.add_edge(node1, node2, weight=edge_weight)
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def ambiguous_nodes(self, node):
"""
Takes a node in parameter and returns the number of possible candidate
(ambiguous) nodes in the graph.
"""
k = 0
while(self.graph.has_node((node, k))):
k += 1
return k
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def get_directed_context(self, node, k, dir='all', non_pos=False):
"""
Returns the directed context of a given node, i.e. a list of word/POS of
the left or right neighboring nodes in the graph. The function takes
four parameters :
- node is the word/POS tuple
- k is the node identifier used when multiple nodes refer to the same
word/POS (e.g. k=0 for (the/DET, 0), k=1 for (the/DET, 1), etc.)
- dir is the parameter that controls the directed context calculation,
it can be set to left, right or all (default)
- non_pos is a boolean allowing to remove stopwords from the context
(default is false)
"""
# Define the context containers
l_context = []
r_context = []
# For all the sentence/position tuples
for sid, off in self.graph.nodes[(node, k)]['info']:
prev = self.sentence[sid][off-1][0].lower() + self.sep +\
self.sentence[sid][off-1][1]
next = self.sentence[sid][off+1][0].lower() + self.sep +\
self.sentence[sid][off+1][1]
if non_pos:
if self.sentence[sid][off-1][0] not in self.stopwords:
l_context.append(prev)
if self.sentence[sid][off+1][0] not in self.stopwords:
r_context.append(next)
else:
l_context.append(prev)
r_context.append(next)
# Returns the left (previous) context
if dir == 'left':
return l_context
# Returns the right (next) context
elif dir == 'right':
return r_context
# Returns the whole context
else:
l_context.extend(r_context)
return l_context
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def get_edge_weight(self, node1, node2):
"""
Compute the weight of an edge *e* between nodes *node1* and *node2*. It
is computed as e_ij = (A / B) / C with:
- A = freq(i) + freq(j),
- B = Sum (s in S) 1 / diff(s, i, j)
- C = freq(i) * freq(j)
A node is a tuple of ('word/POS', unique_id).
"""
# Get the list of (sentence_id, pos_in_sentence) for node1
info1 = self.graph.nodes[node1]['info']
# Get the list of (sentence_id, pos_in_sentence) for node2
info2 = self.graph.nodes[node2]['info']
# Get the frequency of node1 in the graph
# freq1 = self.graph.degree(node1)
freq1 = len(info1)
# Get the frequency of node2 in cluster
# freq2 = self.graph.degree(node2)
freq2 = len(info2)
# Initializing the diff function list container
diff = []
# For each sentence of the cluster (for s in S)
for s in range(self.length):
# Compute diff(s, i, j) which is calculated as
# pos(s, i) - pos(s, j) if pos(s, i) < pos(s, j)
# O otherwise
# Get the positions of i and j in s, named pos(s, i) and pos(s, j)
# As a word can appear at multiple positions in a sentence, a list
# of positions is used
pos_i_in_s = []
pos_j_in_s = []
# For each (sentence_id, pos_in_sentence) of node1
for sentence_id, pos_in_sentence in info1:
# If the sentence_id is s
if sentence_id == s:
# Add the position in s
pos_i_in_s.append(pos_in_sentence)
# For each (sentence_id, pos_in_sentence) of node2
for sentence_id, pos_in_sentence in info2:
# If the sentence_id is s
if sentence_id == s:
# Add the position in s
pos_j_in_s.append(pos_in_sentence)
# Container for all the diff(s, i, j) for i and j
all_diff_pos_i_j = []
# Loop over all the i, j couples
for x in range(len(pos_i_in_s)):
for y in range(len(pos_j_in_s)):
diff_i_j = pos_i_in_s[x] - pos_j_in_s[y]
# Test if word i appears *BEFORE* word j in s
if diff_i_j < 0:
all_diff_pos_i_j.append(-1.0*diff_i_j)
# Add the mininum distance to diff (i.e. in case of multiple
# occurrencies of i or/and j in sentence s), 0 otherwise.
if len(all_diff_pos_i_j) > 0:
diff.append(1.0/min(all_diff_pos_i_j))
else:
diff.append(0.0)
weight1 = freq1
weight2 = freq2
return ( (freq1 + freq2) / sum(diff) ) / (weight1 * weight2)
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def k_shortest_paths(self, start, end, k=10):
"""
Simple implementation of a k-shortest paths algorithms. Takes three
parameters: the starting node, the ending node and the number of
shortest paths desired. Returns a list of k tuples (path, weight).
"""
# Initialize the list of shortest paths
kshortestpaths = []
# Initializing the label container
orderedX = []
orderedX.append((0, start, 0))
# Initializing the path container
paths = {}
paths[(0, start, 0)] = [start]
# Initialize the visited container
visited = {}
visited[start] = 0
# Initialize the sentence container that will be used to remove
# duplicate sentences passing throught different nodes
sentence_container = {}
# While the number of shortest paths isn't reached or all paths explored
while len(kshortestpaths) < k and len(orderedX) > 0:
# Searching for the shortest distance in orderedX
shortest = orderedX.pop(0)
shortestpath = paths[shortest]
# Removing the shortest node from X and paths
del paths[shortest]
# Iterating over the accessible nodes
for node in self.graph.neighbors(shortest[1]):
# To avoid loops
if node in shortestpath:
continue
# Compute the weight to node
w = shortest[0] + self.graph[shortest[1]][node]['weight']
# If found the end, adds to k-shortest paths
if node == end:
#-T-------------------------------------------------------T-
# --- Constraints on the shortest paths
# 1. Check if path contains at least one werb
# 2. Check the length of the shortest path, without
# considering punctuation marks and starting node (-1 in
# the range loop, because nodes are reversed)
# 3. Check the paired parentheses and quotation marks
# 4. Check if sentence is not redundant
nb_verbs = 0
length = 0
paired_parentheses = 0
quotation_mark_number = 0
raw_sentence = ''
for i in range(len(shortestpath) - 1):
word, tag = shortestpath[i][0].split(self.sep)
# 1.
if tag in self.verbs:
nb_verbs += 1
# 2.
if not re.search('(?u)^\W$', word):
length += 1
# 3.
else:
if word == '(':
paired_parentheses -= 1
elif word == ')':
paired_parentheses += 1
elif word == '"':
quotation_mark_number += 1
# 4.
raw_sentence += word + ' '
# Remove extra space from sentence
raw_sentence = raw_sentence.strip()
if nb_verbs >0 and \
length >= self.nb_words and \
paired_parentheses == 0 and \
(quotation_mark_number%2) == 0 \
and not sentence_container.__contains__(raw_sentence):
path = [node]
path.extend(shortestpath)
path.reverse()
weight = float(w) #/ float(length)
kshortestpaths.append((path, weight))
sentence_container[raw_sentence] = 1
#-B-------------------------------------------------------B-
else:
# test if node has already been visited
if visited.__contains__(node):
visited[node] += 1
else:
visited[node] = 0
id = visited[node]
# Add the node to orderedX
bisect.insort(orderedX, (w, node, id))
# Add the node to paths
paths[(w, node, id)] = [node]
paths[(w, node, id)].extend(shortestpath)
# Returns the list of shortest paths
return kshortestpaths
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def get_compression(self, nb_candidates=50):
"""
Searches all possible paths from **start** to **end** in the word graph,
removes paths containing no verb or shorter than *n* words. Returns an
ordered list (smaller first) of nb (default value is 50) (cummulative
score, path) tuples. The score is not normalized with the sentence
length.
"""
# Search for the k-shortest paths in the graph
self.paths = self.k_shortest_paths((self.start+self.sep+self.start, 0),
(self.stop+self.sep+self.stop, 0),
nb_candidates)
# Initialize the fusion container
fusions = []
# Test if there are some paths
if len(self.paths) > 0:
# For nb candidates
for i in range(min(nb_candidates, len(self.paths))):
nodes = self.paths[i][0]
sentence = []
for j in range(1, len(nodes)-1):
word, tag = nodes[j][0].split(self.sep)
sentence.append((word, tag))
bisect.insort(fusions, (self.paths[i][1], sentence))
return fusions
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def max_index(self, l):
""" Returns the index of the maximum value of a given list. """
ll = len(l)
if ll < 0:
return None
elif ll == 1:
return 0
max_val = l[0]
max_ind = 0
for z in range(1, ll):
if l[z] > max_val:
max_val = l[z]
max_ind = z
return max_ind
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def compute_statistics(self):
"""
This function iterates over the cluster's sentences and computes the
following statistics about each word:
- term frequency (self.term_freq)
"""
# Structure for containing the list of sentences in which a term occurs
terms = {}
# Loop over the sentences
for i in range(self.length):
# For each tuple (token, POS) of sentence i
for token, POS in self.sentence[i]:
# generate the word/POS token
node = token.lower() + self.sep + POS
# Add the token to the terms list ,
#in python3, has_key(key) is replaced by __contains__(key)
if not terms.__contains__(node):
terms[node] = [i]
else:
terms[node].append(i)
# Loop over the terms
for w in terms:
# Compute the term frequency
self.term_freq[w] = len(terms[w])
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def load_stopwords(self, path):
"""
This function loads a stopword list from the *path* file and returns a
set of words. Lines begining by '#' are ignored.
"""
# Set of stopwords
stopwords = set([])
# For each line in the file
for line in codecs.open(path, 'r', 'utf-8'):
if not re.search('^#', line) and len(line.strip()) > 0:
stopwords.add(line.strip().lower())
# Return the set of stopwords
return stopwords
#-B-----------------------------------------------------------------------B-
#-T-----------------------------------------------------------------------T-
def write_into_dot(self, dotfile):
""" Outputs the word graph in dot format in the specified file. """
write_dot(self.graph, dotfile)
#-B-----------------------------------------------------------------------B-
#~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
# ] Ending word_graph class
#~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
#~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~