File size: 116,429 Bytes
6fa4bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:57:53.758286Z"
    },
    "title": "Joint Learning of the Graph and the Data Representation for Graph-Based Semi-Supervised Learning",
    "authors": [
        {
            "first": "Mariana",
            "middle": [],
            "last": "Vargas-Vieyra",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "CNRS",
                "location": {
                    "postCode": "9189 -CRIStAL 59650",
                    "settlement": "Villeneuve d'Ascq",
                    "region": "UMR",
                    "country": "France"
                }
            },
            "email": ""
        },
        {
            "first": "",
            "middle": [],
            "last": "Aur\u00e9lien Bellet",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Pascal",
            "middle": [],
            "last": "Denis",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Inria Lille -Nord Europe",
                "location": {}
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Graph-based semi-supervised learning is appealing when labels are scarce but large amounts of unlabeled data are available. These methods typically use a heuristic strategy to construct the graph based on some fixed data representation, independently of the available labels. In this paper, we propose to jointly learn a data representation and a graph from both labeled and unlabeled data such that (i) the learned representation indirectly encodes the label information injected into the graph, and (ii) the graph provides a smooth topology with respect to the transformed data. Plugging the resulting graph and representation into existing graph-based semi-supervised learning algorithms like label spreading and graph convolutional networks, we show that our approach outperforms standard graph construction methods on both synthetic data and real datasets.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Graph-based semi-supervised learning is appealing when labels are scarce but large amounts of unlabeled data are available. These methods typically use a heuristic strategy to construct the graph based on some fixed data representation, independently of the available labels. In this paper, we propose to jointly learn a data representation and a graph from both labeled and unlabeled data such that (i) the learned representation indirectly encodes the label information injected into the graph, and (ii) the graph provides a smooth topology with respect to the transformed data. Plugging the resulting graph and representation into existing graph-based semi-supervised learning algorithms like label spreading and graph convolutional networks, we show that our approach outperforms standard graph construction methods on both synthetic data and real datasets.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "An important bottleneck for the development of accurate Natural Language Processing (NLP) tools for many applications and languages is the lack of annotated data. A natural remedy to this issue lies in semi-supervised learning (SSL) methods, which are able to leverage smaller labeled datasets in combination with large amounts of unannotated text data (which are often more easily available). In particular, graph-based SSL algorithms (Subramanya and Talukdar, 2014) , among which variants of label propagation (Zhu and Ghahramani, 2002; Zhu et al., 2003; Zhou et al., 2004) and more recently graph convolutional networks (Kipf and Welling, 2017; Chen et al., 2018; Wu et al., 2019) , have attracted a lot of attention due to their interesting theoretical properties and good empirical performance. They have successfully been applied to several NLP problems, such as sentiment analysis (Goldberg and Zhu, 2006) , word sense disambiguation (Alexandrescu and Kirchhoff, 2007) , text categorization (Subramanya and Bilmes, 2008) , POS tagging (Subramanya et al., 2010) , semantic parsing (Das and Smith, 2011) , machine translation (Saluja et al., 2014) , or lexicon induction (Faruqui et al., 2016) . As their name suggests, graph-based SSL methods represent all data points (that is, labeled and unlabeled) as nodes in a graph with weighted edges encoding the similarity between pairs of points. This graph is then used as a propagation operator to transfer labels from labeled to unlabeled points. Despite differences in the way this propagation is achieved, graph-based SSL approaches all rely on two assumptions: (i) the graph representing the data provides a faithful approximation of the manifold on which the data actually live, and (ii) the underlying labels are smooth with respect to this manifold.",
                "cite_spans": [
                    {
                        "start": 436,
                        "end": 467,
                        "text": "(Subramanya and Talukdar, 2014)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 512,
                        "end": 538,
                        "text": "(Zhu and Ghahramani, 2002;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 539,
                        "end": 556,
                        "text": "Zhu et al., 2003;",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 557,
                        "end": 575,
                        "text": "Zhou et al., 2004)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 623,
                        "end": 647,
                        "text": "(Kipf and Welling, 2017;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 648,
                        "end": 666,
                        "text": "Chen et al., 2018;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 667,
                        "end": 683,
                        "text": "Wu et al., 2019)",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 888,
                        "end": 912,
                        "text": "(Goldberg and Zhu, 2006)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 941,
                        "end": 975,
                        "text": "(Alexandrescu and Kirchhoff, 2007)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 998,
                        "end": 1027,
                        "text": "(Subramanya and Bilmes, 2008)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 1042,
                        "end": 1067,
                        "text": "(Subramanya et al., 2010)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 1087,
                        "end": 1108,
                        "text": "(Das and Smith, 2011)",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 1131,
                        "end": 1152,
                        "text": "(Saluja et al., 2014)",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 1176,
                        "end": 1198,
                        "text": "(Faruqui et al., 2016)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In many NLP problems, there is often no a priori-known graph, which raises the question of how to best construct this graph over the dataset given some data representation. Most existing work rely on classic graph construction heuristics such as k-nn graphs, -graphs or radial kernel graphs (Subramanya and Talukdar, 2014) , which may poorly adapt to the intrinsic structure of the data manifold and hence violate assumption (i). More recently, in the machine learning and signal processing communities, some algorithms were proposed to learn more flexible graphs by enforcing the topology to be smooth with respect to the input data (Daitch et al., 2009; Kalofolias, 2016; Dong et al., 2016) . All these approaches heavily depend on the choice of data representation and disregard the label information, making them unable to adapt to the prediction task and therefore potentially violating assumption (ii). While supervised representation learning techniques such as metric learning (Bellet et al., 2015) could be used to adapt the representation to the task of interest, for instance by bringing closer points with the same label, the lack of labeled data in the semi-supervised learning scenario makes them prone to overfitting.",
                "cite_spans": [
                    {
                        "start": 291,
                        "end": 322,
                        "text": "(Subramanya and Talukdar, 2014)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 634,
                        "end": 655,
                        "text": "(Daitch et al., 2009;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 656,
                        "end": 673,
                        "text": "Kalofolias, 2016;",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 674,
                        "end": 692,
                        "text": "Dong et al., 2016)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 985,
                        "end": 1006,
                        "text": "(Bellet et al., 2015)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we propose an original semi-supervised algorithm for graph construction that adapts to both the data and the predictive task. Specifically, our approach leverages the labeled and unlabeled data to jointly learn a graph and a data representation. On the one hand, the graph is learned to provide a smooth topology with respect to the learned representation. On the other hand, the representation should bring closer (labeled and unlabeled) points that are neighbors in the graph as well as similarly labeled points, while pulling away points of different labels. A key feature of our approach is that the learned representation indirectly encodes and injects label information into the graph beyond the labeled points alone. We formulate our problem as a joint optimization problem over the representation and the graph weights, with a hyperparameter to easily control the sparsity of the resulting graph and thereby obtain a good approximation of the underlying manifold. We discuss some appropriate parameterizations for learning the representation, which revolve around adapting pre-trained embeddings so as to avoid overfitting. We then propose to solve our joint problem by alternating optimization on the representation and the graph. We validate our approach through several graph-based SSL experiments using label spreading (Zhou et al., 2004) and graph convolutional networks (GCN) (Kipf and Welling, 2017), both on synthetic and real text classification datasets. Incidentally, note that our approach is generic and could in principle be used in combination with any existing graph-based SSL framework. The results show that our approach outperforms previous methods which rely on heuristic graphs, generally by a considerable margin. Interestingly, we also observe that our approach effectively bridges the accuracy gap between a simple method like label spreading and a richer neural-based approach like GCN.",
                "cite_spans": [
                    {
                        "start": 1346,
                        "end": 1365,
                        "text": "(Zhou et al., 2004)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The rest of this paper is organized as follows. We introduce some notations and discuss the related work in Section 2. We then describe our approach and algorithm in Section 3. Our experimental results are presented in Section 4, and we conclude with future work directions in Section 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Our work lies in the intersection of two topics: graph-based semi-supervised learning and graph learning. In this section, we start by introducing some useful notations. We then briefly review some related work in both areas.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Notations and Related Work",
                "sec_num": "2"
            },
            {
                "text": "We consider a dataset consisting of l labeled points L = {(x i , y i )} l i=1 and u unlabeled points U = {x i } l+u i=l+1 , where data points x i lie in some space X (typically X \u2282 IR d ) and labels y i \u2208 {1, . . . , C} are discrete. We denote the combined data by X \u2208 IR n\u00d7d , where n = l+u. We place ourselves in the typical semi-supervised scenario where l u, and the goal is to predict y l+1 , . . . , y l+u . Let G = (V, W ) be a graph composed of a set of nodes V = {v 1 , . . . , v n } and a symmetric nonnegative weighted adjacency matrix W \u2208 IR n\u00d7n . We denote the set W of admissible weighted adjacency matrices by",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Notations",
                "sec_num": "2.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "W = {W : W \u2265 0, diag(W ) = 0, W = W }.",
                        "eq_num": "(1)"
                    }
                ],
                "section": "Notations",
                "sec_num": "2.1"
            },
            {
                "text": "Every observation x i (labeled or unlabeled) can be seen as a signal occurring at node v i , and W assigns a weight to each pair of nodes. We say that two nodes v i and v j are connected when W ij > 0.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Notations",
                "sec_num": "2.1"
            },
            {
                "text": "Graph-based SSL takes as input a graph G = (V, W ) over the labeled and unlabeled data, the known labels and optionally some feature representation X \u2208 IR n\u00d7d of the data, and aims to predict the labels of the unlabeled points. Many approaches exist for graph-based SSL, see (Subramanya and Talukdar, 2014) for a review of standard approaches. Classic algorithms take as input only the graph and use it to propagate the known labels to the unlabeled points (Zhu and Ghahramani, 2002; Zhou et al., 2004; Bengio et al., 2006) . A popular approach is label spreading (Zhou et al., 2004) , which can be formulated as a convex optimization problem implementing a trade-off between two terms. The first term is a graph Laplacian regularization term which encourages similar predictions for strongly connected points in the graph, while the second one tries to keep the predictions accurate for points with known labels. There also exist algorithms formalized as graph partitioning problems inspired from spectral clustering (Joachims, 2003) .",
                "cite_spans": [
                    {
                        "start": 275,
                        "end": 306,
                        "text": "(Subramanya and Talukdar, 2014)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 457,
                        "end": 483,
                        "text": "(Zhu and Ghahramani, 2002;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 484,
                        "end": 502,
                        "text": "Zhou et al., 2004;",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 503,
                        "end": 523,
                        "text": "Bengio et al., 2006)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 564,
                        "end": 583,
                        "text": "(Zhou et al., 2004)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 1018,
                        "end": 1034,
                        "text": "(Joachims, 2003)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Graph-based SSL",
                "sec_num": "2.2"
            },
            {
                "text": "Graph-based SSL methods can also leverage a feature representation of the data in addition to the graph. A generic strategy is to add the graph Laplacian regularization term to existing supervised learning algorithms such as SVM or neural networks (Belkin et al., 2006; Weston et al., 2012) . Following a different direction, (Yang et al., 2016) builds upon graph embeddings approaches to propose a method to predict labels based on the input representation as well as some embeddings learned from the input graph and known labels. Finally, there has been some recent interest in graph convolutional networks (GCN) (Kipf and Welling, 2017; Chen et al., 2018; Wu et al., 2019) . Much like CNNs for images, they rely on an (approximation of) a notion of graph convolution, allowing them to learn a nonlinear transformation of the input representation while encoding the graph structure when propagating inputs from a layer to the next.",
                "cite_spans": [
                    {
                        "start": 248,
                        "end": 269,
                        "text": "(Belkin et al., 2006;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 270,
                        "end": 290,
                        "text": "Weston et al., 2012)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 326,
                        "end": 345,
                        "text": "(Yang et al., 2016)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 615,
                        "end": 639,
                        "text": "(Kipf and Welling, 2017;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 640,
                        "end": 658,
                        "text": "Chen et al., 2018;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 659,
                        "end": 675,
                        "text": "Wu et al., 2019)",
                        "ref_id": "BIBREF29"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Graph-based SSL",
                "sec_num": "2.2"
            },
            {
                "text": "In all these approaches, the graph is fixed and given as input to the algorithm (for several methods, this is also true for the data representation). Their performance is thus very sensitive to the relevance of the graph for the task at hand: in particular, the underlying labeling should be smooth with respect to the graph. Our method is a principled approach to learn such a task-specific graph, and also infers a relevant data representation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Graph-based SSL",
                "sec_num": "2.2"
            },
            {
                "text": "As pointed out by (Subramanya and Talukdar, 2014) , graph-based SSL methods typically rely on graphs constructed from the input data representation with a simple heuristic strategy. Popular choices include k-nearest neighbor graphs (connecting pairs of points that are among the k-closest to each other), -graph (connecting points that are within distance ), and radial kernel graphs (building a fully connected graph with exponentially decreasing weights W ij = e \u2212\u03b3 x i \u2212x j 2 ). Recently, more sophisticated methods that learn the graph weights as the solution of an optimization problem have been introduced (Daitch et al., 2009; Kalofolias, 2016; Dong et al., 2016) . Essentially, the weights are learned to be smooth over the data representation (i.e., assigning large weights to nearby points) with some regularization to enforce or control some properties such as connectedness and sparsity. In any case, the graphs obtained with the above approaches are task-independent in the sense that they ignore the labels.",
                "cite_spans": [
                    {
                        "start": 18,
                        "end": 49,
                        "text": "(Subramanya and Talukdar, 2014)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 612,
                        "end": 633,
                        "text": "(Daitch et al., 2009;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 634,
                        "end": 651,
                        "text": "Kalofolias, 2016;",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 652,
                        "end": 670,
                        "text": "Dong et al., 2016)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Graph Construction and Learning",
                "sec_num": "2.3"
            },
            {
                "text": "To the best of our knowledge, there have been very few attempts to learn task-specific graphs for SSL. (Alexandrescu and Kirchhoff, 2007) propose to train a supervised classifier on labeled points and using the soft label predictions as the representation to build the graph. While this gives a way to incorporate label information, the supervised predictions are very dependent on the initial representation and the classifier itself can heavily overfit due to scarce labels.",
                "cite_spans": [
                    {
                        "start": 103,
                        "end": 137,
                        "text": "(Alexandrescu and Kirchhoff, 2007)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Graph Construction and Learning",
                "sec_num": "2.3"
            },
            {
                "text": "Our approach learns a graph and a data representation for use in downstream graph-based SSL algorithms. In this section, we start by introducing our formulation as a joint optimization problem over the representation and the graph. We then discuss some relevant choices for the parameterization of the learned representation, and finally present our alternating optimization scheme.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposed Model",
                "sec_num": "3"
            },
            {
                "text": "For the sake of generality, in this section we formulate our problem with respect to a generic representation function \u03c6 \u0398 : X \u2192 IR k , parameterized by \u0398, which represents any data point x \u2208 X as a k-dimensional vector \u03c6 \u0398 (x) \u2208 IR k . We discuss some relevant choices of representation functions in Section 3.2.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "We propose to learn a weighted adjacency matrix W * and a representation function \u03c6 \u0398 * by minimizing a joint objective function f that involves both the labeled and unlabeled data points:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "W * , \u0398 * = arg min W \u2208W,\u0398 f (W, \u0398).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "Once the above optimization problem has been solved, the learned graph W * (which is based on the learned representation function \u03c6 \u0398 * ) and possibly the representation \u03c6 \u0398 * can then be given as input to any graph-based SSL algorithm to obtain predictions for the unlabeled data.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "Our objective function f (W, \u0398) decomposes into three terms:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "f (W, \u0398) = f 1 (\u0398) + \u03b1[f 2 (W ) + f 3 (W, \u0398)]",
                        "eq_num": "(2)"
                    }
                ],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "where f 1 (\u0398) and f 2 (W ) are respectively the representation and graph specific terms, while f 3 (W, \u0398) is the joint term. Hyperparameter \u03b1 \u2265 0 controls the trade-off between the (supervised) representation learning term f 1 and the unsupervised part (f 2 and f 3 ). We now define these three terms. For notational convenience, let us denote by Z \u2208 IR n\u00d7n the matrix whose entries are the normalized squared Euclidean distances between data points in the transformed space",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": ", i.e. (Z \u0398 ) ij = ||\u03c6 \u0398 (x i )\u2212\u03c6 \u0398 (x j )|| 2 i<j ||\u03c6 \u0398 (x i )\u2212\u03c6 \u0398 (x j )|| 2 .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "The normalization conveniently removes the dependency on the scale of the data and \u0398. The representation term f 1 (\u0398) is defined on the labeled data points only and takes the following form:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "f 1 (\u0398) = x i ,x j ,x k \u2208L y i =y j ,y i =y k (Z \u0398 ) ij \u2212 (Z \u0398 ) ik + 1 + ,",
                        "eq_num": "(3)"
                    }
                ],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "where",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "[\u2022] + = max(0, \u2022)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": ". This is a large-margin triplet loss similar to those used in metric learning (Bellet et al., 2015) : it attempts to learn a representation function \u03c6 \u0398 that brings each point x i closer to points x j with the same label than to differently labeled points x k , with a safety margin of 1. In practice, we can subsample instead of summing over all possible triplets. The graph term f 2 (W ) is inspired from the (unsupervised) graph learning approach proposed by (Kalofolias, 2016) :",
                "cite_spans": [
                    {
                        "start": 79,
                        "end": 100,
                        "text": "(Bellet et al., 2015)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 463,
                        "end": 481,
                        "text": "(Kalofolias, 2016)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "f 2 (W ) = \u03b2||W || 2 F \u2212 1 log(1 W ),",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "The log-barrier term on the degrees prevents any node from being isolated in the graph, while the Frobenius norm is a shrinkage term over the graph weights. Combined with our joint term (6) defined below, hyperparameter \u03b2 \u2265 0 directly controls the sparsity of the learned graph: the smaller \u03b2, the more concentrated the weights of each point on its nearest neighbors in the learned representation (hence the sparser the graph). On the other hand, as \u03b2 \u2192 +\u221e, the graph becomes complete with uniform weights. Sparsity allows to enforce the locality property (only close points are connected in the graph) which is necessary to obtain a good approximation of the data manifold. It also reduces the computational cost in downstream graph-based SSL algorithms, whose complexity typically depends on the number of edges in the graph. Other options are possible for f 2 (W ) depending on the prior we want to have on the structure of the graph. For instance, one may use",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "f 2 (W ) = (1/\u03b3) i,j W ij [log(W ij ) \u2212 1],",
                        "eq_num": "(5)"
                    }
                ],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "where \u03b3 > 0 is a hyperparameter. This will force the graph to be fully connected. Finally, we introduce the joint term bringing together the graph and the representation:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "f 3 (W, \u0398) = tr(W Z \u0398 ) = i,j W ij (Z \u0398 ) ij .",
                        "eq_num": "(6)"
                    }
                ],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "This can be seen as a weighted L 1 norm term on W (which is why it induces sparsity), and equivalently written as a quadratic form of the Laplacian matrix of the graph encoded by the symmetric matrix W .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "It is also used in approaches based on graph Laplacian regularization (see Section 2), but in our case both the graph and the representation are learned in joint manner. This term makes the graph and the representation as smooth as possible with respect to each other on both labeled and unlabeled points. Overall, our joint objective function (2) is designed to produce a sparse topology that tends to be smooth with respect to the data manifold and the underlying labeling function through an appropriate representation. We now discuss the choice of representation function \u03c6 \u0398 .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "3.1"
            },
            {
                "text": "Many options are possible for the representation function \u03c6 \u0398 depending on the nature of the data and task at hand. However, it is important to keep in mind that the amount of labeled information is scarce, hence learning complex text representations from scratch is likely to lead to severe overfitting. We argue that it is preferable to adapt pre-trained representations, which generally requires to optimize much fewer parameters. We give some concrete examples below.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Choices of Representation Functions",
                "sec_num": "3.2"
            },
            {
                "text": "Linear transformation. Pre-trained word embeddings (Mikolov et al., 2013; Pennington et al., 2014) are commonly used to represent texts in a vectorial space, e.g. by averaging the embeddings of the words occurring in a document. In order to adapt the representation to the task, we can learn a simple linear mapping \u03c6 \u0398 (x) = \u0398x which transforms the initial d-dimensional representation into a k-dimensional one, with \u0398 \u2208 IR k\u00d7d and k \u2264 d. Such a strategy has been previously explored in the supervised setting to \"re-embed\" words in a task-specific manner (Denis and Ralaivola, 2017) . This is the representation function that we use in our experiments (see Section 4).",
                "cite_spans": [
                    {
                        "start": 51,
                        "end": 73,
                        "text": "(Mikolov et al., 2013;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 74,
                        "end": 98,
                        "text": "Pennington et al., 2014)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 557,
                        "end": 584,
                        "text": "(Denis and Ralaivola, 2017)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Choices of Representation Functions",
                "sec_num": "3.2"
            },
            {
                "text": "Weighted combination. Recent work in learning deep contextualized word representations such as ELMo (Peters et al., 2018) and BERT (Devlin et al., 2019 ) allows to learn a task-specific combination of the token representations obtained at the K layers of the model, which typically capture different aspects of tokens (from syntax to semantics). In this case, we have K initial d-dimensional representations x \u2208 IR K\u00d7d for each text x and we learn a weighted combination \u03c6 \u0398 (x) = \u0398x \u2208 IR d where \u0398 \u2208 IR K is simply a K-dimensional parameter vector.",
                "cite_spans": [
                    {
                        "start": 100,
                        "end": 121,
                        "text": "(Peters et al., 2018)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 131,
                        "end": 151,
                        "text": "(Devlin et al., 2019",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Choices of Representation Functions",
                "sec_num": "3.2"
            },
            {
                "text": "We propose to optimize the cost function f (W, \u0398) by alternating minimization over W and \u0398, which is guaranteed to converge to a local optimum. This is a natural approach: one step learns a smooth graph given the current representation \u0398, while the other learns a smooth representation with respect to the current graph (this can be seen as a regularizer for \u0398 based on unlabeled data) and also tries to keep labeled points of the same class closer than points of different class.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimization",
                "sec_num": "3.3"
            },
            {
                "text": "As the joint problem is nonconvex, initialization plays an important role. We propose to initialize the graph weights to zero and to start by optimizing \u0398 so that the initial representation focuses only on the (scarce) labeled data. The graph learned on this representation will thus strongly connect together the labeled points as well as unlabeled points that are very close to the labeled points and are thus likely to share the same label. At the next iteration, these unlabeled points will then contribute in learning a better representation and in turn a graph which strongly connects new unlabeled points. This process can be seen as a principled version of self-training heuristics popular in traditional (non-graph-based) semi-supervised learning (Triguero et al., 2015) .",
                "cite_spans": [
                    {
                        "start": 756,
                        "end": 779,
                        "text": "(Triguero et al., 2015)",
                        "ref_id": "BIBREF27"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimization",
                "sec_num": "3.3"
            },
            {
                "text": "The subproblem of optimizing W given \u0398 is convex regardless of whether we define f 2 (W ) as (4) or (5). Using (5) is computationally convenient as the subproblem has a closed-form solution: the weights are exponentially decreasing with the distance in the current representation \u03c6 \u0398 , as given by the radial kernel W ij = exp(\u2212\u03b3(Z \u0398 ) ij ) (Kalofolias, 2016) . Note that unlike the classic radial kernel baseline construction method mentioned in Section 2.3, our graph is computed based on the learned representation \u03c6 \u0398 by minimizing the joint objective function with respect to W . One drawback of using (5) is that the resulting graphs are always fully connected. Using (4) instead, we can obtain sparse graphs but the solution must be computed with an iterative algorithm. We found that the primal-dual algorithm introduced by (Kalofolias, 2016) converges slowly in practice -we instead optimize W by simple gradient descent over the \"effective\" n(n \u2212 1)/2 weights, adding a small positive constant inside the log term in (4) to make the objective function smooth.",
                "cite_spans": [
                    {
                        "start": 341,
                        "end": 359,
                        "text": "(Kalofolias, 2016)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 832,
                        "end": 850,
                        "text": "(Kalofolias, 2016)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimization",
                "sec_num": "3.3"
            },
            {
                "text": "As \u03c6 \u0398 is typically differentiable in \u0398 (as in the examples outlined in Section 3.2), we also solve the subproblem in \u0398 by (stochastic) gradient descent. Note that this subproblem is generally nonconvex due to the distance difference in f 1 (\u0398).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimization",
                "sec_num": "3.3"
            },
            {
                "text": "Remark. Updating W requires to optimize over O(n 2 ) variables, which was manageable for the datasets used in our experiments. To scale to larger datasets, one can restrict the optimization to the weights corresponding to pairs of points that are close enough in the learned representation space 1 (other weights are kept to 0). This has a negligible impact on the solution in sparse regimes (small \u03b2).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimization",
                "sec_num": "3.3"
            },
            {
                "text": "In this section, we study the practical behavior of our method by comparing the accuracy of downstream graph-based SSL algorithms when the graph (along with the underlying representation) is learned with our approach ( ours) rather than constructed with the following baseline strategies:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "\u2022 radial: Complete graph with weights W ij = exp(\u2212\u03b3 x i \u2212 x j",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "2 ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "\u2022 knn: W ij = 1 for x i in the k-neighborhood of x j (or vice versa), and W ij = 0 otherwise.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "\u2022 kalo: Unsupervised graph learning with the method of (Kalofolias, 2016) . This corresponds to our approach when using the graph term (4) and keeping the original representation fixed.",
                "cite_spans": [
                    {
                        "start": 55,
                        "end": 73,
                        "text": "(Kalofolias, 2016)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "In all cases the graph is constructed over the union of labeled (train set) and unlabeled data (validation and test sets). For experiments with our method, the learned representation is a linear transformation of the initial features as explained in Section 3.2.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "We perform experiments with two graph-based SSL approaches: Label Spreading (LS) (Zhou et al., 2004) and the Graph Convolutional Network (GCN) method of (Kipf and Welling, 2017). We used the scikit-learn (Pedregosa et al., 2011) implementation of LS. For GCN, we used the TensorFlow implementation provided by the authors 2 and follow the recommended architecture:",
                "cite_spans": [
                    {
                        "start": 81,
                        "end": 100,
                        "text": "(Zhou et al., 2004)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 204,
                        "end": 228,
                        "text": "(Pedregosa et al., 2011)",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Y = softmax(L max(0,LXH 0 )H 1 ),",
                        "eq_num": "(7)"
                    }
                ],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "whereL =D 1/2WD1/2 is the normalized Laplacian corresponding to the graphW = W +\u03bbI (the input graph augmented with self-loops),D is the diagonal degree matrix (D ii = jD ij ), and H 0 \u2208 IR d\u00d7h , H 1 \u2208 IR h\u00d7C are the parameters to be learned. We set the number of hidden units h to 16 and \u03bb to 1 as done in (Kipf and Welling, 2017).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "To illustrate the behavior of our approach, we first present some experiments on synthetic data. We then show some results on real text classification datasets.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "We generated a 3-dimensional dataset consisting of 100 points evenly distributed in two classes ( Figure  1 ). We have two clusters per class placed far from each other while keeping clusters from different classes closer. We randomly picked 60% of the points and removed their labels.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 98,
                        "end": 107,
                        "text": "Figure  1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Synthetic Data",
                "sec_num": "4.1"
            },
            {
                "text": "We compare the classification error of GCN and Label Spreading when the input graph is given by our approach instead of using baseline graph construction methods. For GCN, we also give as input the representation learned with our approach. For our approach, we use the graph term (4) and for each labeled point x i , we random sample 2 points x j of the same class and 3 points x k of different class and construct all combinations (x i , x j , x k ), leading to 6 triplets for each x i in the triplet loss (3). The results given in Table 1 show that our approach clearly and consistently outperforms all methods in both GCN and Label Spreading. 3 The improvements are especially large for Label Spreading, as LS makes predictions based on the graph only. In contrast, GCN learns its own (nonlinear) transformation of the representation given as input in an end-to-end manner. Still, our method is able to provide some gains for GCN as well, by providing it with a better graph. Note for instance the significant improvement compared to kalo, which learns the graph on the original representation. To visualize this difference, Figure 2a shows the graph learned by kalo. Although the graph is learned to minimize the smoothness criterion with respect to the data, it fails to accurately capture the label distribution due to the limitations of the initial representation. Our alternating optimization approach overcomes this issue by learning a task-specific graph through an appropriate representation. In Figure 2b-2c-2d, we can see how label information is gradually injected at each step: after the first iteration, the graph is already significantly more smooth with respect to the underlying labeling and the graph is also sparser, but some edges between differently labeled points as well as an overly connected point remain. The following iterations further improve the graph quality. This explains the better performance obtained in downstream semi-supervised algorithms.",
                "cite_spans": [
                    {
                        "start": 621,
                        "end": 647,
                        "text": "GCN and Label Spreading. 3",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 533,
                        "end": 540,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    },
                    {
                        "start": 1128,
                        "end": 1137,
                        "text": "Figure 2a",
                        "ref_id": null
                    },
                    {
                        "start": 1507,
                        "end": 1513,
                        "text": "Figure",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Synthetic Data",
                "sec_num": "4.1"
            },
            {
                "text": "We now evaluate our method on three text classification tasks derived from the 20NewsGroups dataset, 4 a collection of documents categorized into 20 topics, each one of which is partitioned into sub-topics. We chose the topics of computers with classes IBM and Mac (n = 1945 documents), religion with classes atheism and Christian (n = 1796), and sports with classes baseball and hockey (n = 1993).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Real Data",
                "sec_num": "4.2"
            },
            {
                "text": "For all datasets, we represent data points using the average token embedding based on word2vec (Mikolov et al., 2013) . These embeddings are of dimension d = 300 and were trained on a 100B word corpus of Google news data (vocabulary size is 3M). 5 We experiment with different proportions of unlabeled points in the training set (90%, 75%, 60% and 40%), while the rest of the data is evenly split into a validation and a test set. As commonly done in semisupervised learning, we train on the union of the (labeled) training set and the (unlabeled) validation and test sets, select the values of hyperparameters based on the accuracy on the validation set, and report the corresponding accuracy on the test set.",
                "cite_spans": [
                    {
                        "start": 95,
                        "end": 117,
                        "text": "(Mikolov et al., 2013)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 246,
                        "end": 247,
                        "text": "5",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Real Data",
                "sec_num": "4.2"
            },
            {
                "text": "To evaluate our approach we optimize the objective (2) as described in Section 3.3 with the graph term defined as in (4). To compute the representation term of our objective defined in (3), we construct triplets tained in LS, which is to be expected since GCN has the ability to learn nonlinear transformations of the data. Nevertheless, we do observe some performance gains, as our approach generally improves upon or closely matches the performance of the best baseline. An interesting finding is that our method tends to close the gap of performance between LS and the richer neural-based GCN model. This suggests that simple propagation approaches may be sufficient in practice for many datasets, if provided with the right graph.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Real Data",
                "sec_num": "4.2"
            },
            {
                "text": "Visualization. We provide visualizations of the representation and the graph learned with our approach on the rel dataset. Figure 3 shows 3D PCA visualizations of the original representation and the representation learned with our approach. We see that the two classes are quite mixed up in the original representation while the learned representation is much smoother with respect to the underlying labeling (even in this crude low-dimensional summary). Figure 4 gives a snapshot of the graph learned with our approach by showing a subgraph of 50 randomly sampled nodes (subsampling helps to avoid clutter). The graph is very smooth with respect to the underlying labeling, and suggests that the learned high-dimensional representation has a nice manifold structure, with some regions of higher densities.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 123,
                        "end": 131,
                        "text": "Figure 3",
                        "ref_id": "FIGREF2"
                    },
                    {
                        "start": 455,
                        "end": 463,
                        "text": "Figure 4",
                        "ref_id": "FIGREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Real Data",
                "sec_num": "4.2"
            },
            {
                "text": "We presented a novel method bringing together graph learning, representation learning and SSL by jointly inferring the graph and the representation from semi-supervised data. The output of our approach can then be plugged into any graph-based SSL algorithm in place of using common graph constructions. Our experimental results suggest that the gains are especially significant for graph-based SSL algorithms that are unable to adapt the data representation (like label spreading and its variants), although we observe some gains also for GCN. To further improve the performance with richer models like GCNs, a promising direction is to extend our approach to learn the graph, the representation and the classifier in an end-toend manner. We note that there has been very recent attempts in this general direction (Franceschi et al., 2019) , though specific to GCN and with completely different modeling and assumptions.",
                "cite_spans": [
                    {
                        "start": 814,
                        "end": 839,
                        "text": "(Franceschi et al., 2019)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion and Future Work",
                "sec_num": "5"
            },
            {
                "text": "The ideas underlying our approach could also be useful to tackle transfer learning settings and in particular domain adaptation (Ben-David et al., 2010) . The latter can be seen as a SSL problem where the distribution of the target (unlabeled) data follows a different distribution that the source (training) data. Our objective function could be modified to encourage the learned representation and graph to serve as a \"bridge\" between the source and target distributions.",
                "cite_spans": [
                    {
                        "start": 128,
                        "end": 152,
                        "text": "(Ben-David et al., 2010)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion and Future Work",
                "sec_num": "5"
            },
            {
                "text": "These can be identified in near-linear time using approximate nearest-neighbor techniques(Muja and Lowe, 2014).2 https://github.com/tkipf/gcn 3 For this illustrating experiment, we picked the values of hyperparameters giving the best results for each method.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "http://qwone.com/\u02dcjason/20Newsgroups/ 5 https://code.google.com/archive/p/word2vec/",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "The authors would like to thank the Agence Nationale de la Recherche for funding this project with the grant number ANR-16-CE33-0011, and the reviewers for their feedback and suggestions. We also thank Grid5000 team for providing the necessary resources and technical support for this project, and CPER Nord-Pas de Calais/FEDER DATA Advanced data science and technologies 2015-2020.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgements",
                "sec_num": null
            },
            {
                "text": "as follows: for each pair (x i , y i ) in the labeled set we obtain the closest points with labels other than y i (\"imposters\"), and the closest points with label y i (\"targets\"). We picked 8 imposters and 3 targets. We tune the hyperparameters \u03b1 from {0.001, 0.01, 1}, \u03b2 from {0.00001, 0.001, 0.1, 1}, the dimension k of the learned representation from {16, 32, 64}, and perform early stopping with respect to the number of alternating steps between learning the graph and learning the representation (up to 10 alternating steps). We also tuned the hyperparameters of each baseline method (\u03b3 for radial, k for knn and \u03b2 for kalo) and the trade-off hyperparameter of Label Spreading. Finally, we computed the McNemar test of significance (McNemar Quinn, 1947) to compare the performance of our method against the best baseline. Results marked with a dagger symbol \u2020 yield a statistically significant test for a significance level of 0.05.Label Spreading. Table 2a reports test classification accuracies obtained on the test set for each configuration of dataset and proportion of unlabeled data. Our approach clearly outperforms all baselines, most of the time by a large margin. Also, McNemar test indicates that we tend to be significantly better than the best baseline in the more challenging settings where labeled data is the most scarce. The results also show that learning the representation along with the graph makes a clear difference compared to learning the graph only (as seen by the superior performance of ours over kalo).As LS only uses the graph to make predictions, these results provide strong evidence of the superior quality of the graphs learned with our method.Graph Convolutional Networks. We now turn to the more complex GCN prediction model. We reuse the same setup as for LS and feed GCN with both the learned representation and the learned graph. Table 2b summarizes the results. The gains obtained with our approach are smaller than those ob-",
                "cite_spans": [
                    {
                        "start": 738,
                        "end": 759,
                        "text": "(McNemar Quinn, 1947)",
                        "ref_id": "BIBREF17"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 955,
                        "end": 963,
                        "text": "Table 2a",
                        "ref_id": null
                    },
                    {
                        "start": 1875,
                        "end": 1883,
                        "text": "Table 2b",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "annex",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Data-Driven Graph Construction for Semi-Supervised Graph-Based Learning in NLP",
                "authors": [
                    {
                        "first": "Andrei",
                        "middle": [],
                        "last": "Alexandrescu",
                        "suffix": ""
                    },
                    {
                        "first": "Katrin",
                        "middle": [],
                        "last": "Kirchhoff",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "HLT-NAACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andrei Alexandrescu and Katrin Kirchhoff. 2007. Data-Driven Graph Construction for Semi-Supervised Graph- Based Learning in NLP. In HLT-NAACL.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Manifold Regularization: A Geometric Framework for Learning from Labeled and Unlabeled Examples",
                "authors": [
                    {
                        "first": "Mikhail",
                        "middle": [],
                        "last": "Belkin",
                        "suffix": ""
                    },
                    {
                        "first": "Partha",
                        "middle": [],
                        "last": "Niyogi",
                        "suffix": ""
                    },
                    {
                        "first": "Vikas",
                        "middle": [],
                        "last": "Sindhwani",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Journal of Machine Learning Research",
                "volume": "7",
                "issue": "",
                "pages": "2399--2434",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mikhail Belkin, Partha Niyogi, and Vikas Sindhwani. 2006. Manifold Regularization: A Geometric Framework for Learning from Labeled and Unlabeled Examples. Journal of Machine Learning Research, 7:2399-2434.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Metric Learning",
                "authors": [
                    {
                        "first": "Aur\u00e9lien",
                        "middle": [],
                        "last": "Bellet",
                        "suffix": ""
                    },
                    {
                        "first": "Amaury",
                        "middle": [],
                        "last": "Habrard",
                        "suffix": ""
                    },
                    {
                        "first": "Marc",
                        "middle": [],
                        "last": "Sebban",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Aur\u00e9lien Bellet, Amaury Habrard, and Marc Sebban. 2015. Metric Learning. Morgan & Claypool Publishers.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Transductive Learning via Spectral Graph Partitioning",
                "authors": [
                    {
                        "first": "Shai",
                        "middle": [],
                        "last": "Ben-David",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Blitzer",
                        "suffix": ""
                    },
                    {
                        "first": "Koby",
                        "middle": [],
                        "last": "Crammer",
                        "suffix": ""
                    },
                    {
                        "first": "Alex",
                        "middle": [],
                        "last": "Kulesza",
                        "suffix": ""
                    },
                    {
                        "first": "Fernando",
                        "middle": [],
                        "last": "Pereira",
                        "suffix": ""
                    },
                    {
                        "first": "Jennifer",
                        "middle": [
                            "Wortman"
                        ],
                        "last": "Vaughan",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Machine Learning",
                "volume": "79",
                "issue": "",
                "pages": "151--175",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. 2010. Transductive Learning via Spectral Graph Partitioning. Machine Learning, 79(1-2):151-175.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Label Propagation and Quadratic Criterion",
                "authors": [
                    {
                        "first": "Yoshua",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    },
                    {
                        "first": "Olivier",
                        "middle": [],
                        "last": "Delalleau",
                        "suffix": ""
                    },
                    {
                        "first": "Nicolas",
                        "middle": [
                            "Le"
                        ],
                        "last": "Roux",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Semi-Supervised Learning",
                "volume": "",
                "issue": "",
                "pages": "193--216",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yoshua Bengio, Olivier Delalleau, and Nicolas Le Roux. 2006. Label Propagation and Quadratic Criterion. In Olivier Chapelle, Bernhard Sch\u00f6lkopf, and Alexander Zien, editors, Semi-Supervised Learning, pages 193-216. MIT Press.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling",
                "authors": [
                    {
                        "first": "Jie",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Tengfei",
                        "middle": [],
                        "last": "Ma",
                        "suffix": ""
                    },
                    {
                        "first": "Cao",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "ICLR",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jie Chen, Tengfei Ma, and Cao Xiao. 2018. FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling. In ICLR.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Fitting a graph to vector data",
                "authors": [
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Samuel",
                        "suffix": ""
                    },
                    {
                        "first": "Jonathan",
                        "middle": [
                            "A"
                        ],
                        "last": "Daitch",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [
                            "A"
                        ],
                        "last": "Kelner",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Spielman",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "ICML",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Samuel I. Daitch, Jonathan A. Kelner, and Daniel A. Spielman. 2009. Fitting a graph to vector data. In ICML.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Semi-supervised frame-semantic parsing for unknown predicates",
                "authors": [
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Noah",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dipanjan Das and Noah A Smith. 2011. Semi-supervised frame-semantic parsing for unknown predicates. In ACL.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Online Learning of Task-specific Word Representations with a JointBiconvex Passive-Aggressive Algorithm",
                "authors": [
                    {
                        "first": "Pascal",
                        "middle": [],
                        "last": "Denis",
                        "suffix": ""
                    },
                    {
                        "first": "Liva",
                        "middle": [],
                        "last": "Ralaivola",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "EACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Pascal Denis and Liva Ralaivola. 2017. Online Learning of Task-specific Word Representations with a JointBi- convex Passive-Aggressive Algorithm. In EACL.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Devlin",
                        "suffix": ""
                    },
                    {
                        "first": "Ming-Wei",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "NAACL-HLT",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirec- tional Transformers for Language Understanding. In NAACL-HLT.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Learning Laplacian Matrix in Smooth Graph Signal Representations",
                "authors": [
                    {
                        "first": "Xiaowen",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    },
                    {
                        "first": "Dorina",
                        "middle": [],
                        "last": "Thanou",
                        "suffix": ""
                    },
                    {
                        "first": "Pascal",
                        "middle": [],
                        "last": "Frossard",
                        "suffix": ""
                    },
                    {
                        "first": "Pierre",
                        "middle": [],
                        "last": "Vandergheynst",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "IEEE Transactions on Signal Processing",
                "volume": "64",
                "issue": "23",
                "pages": "6160--6173",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xiaowen Dong, Dorina Thanou, Pascal Frossard, and Pierre Vandergheynst. 2016. Learning Laplacian Matrix in Smooth Graph Signal Representations. IEEE Transactions on Signal Processing, 64(23):6160-6173.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Morpho-syntactic lexicon generation using graphbased semi-supervised learning",
                "authors": [
                    {
                        "first": "Manaal",
                        "middle": [],
                        "last": "Faruqui",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Mcdonald",
                        "suffix": ""
                    },
                    {
                        "first": "Radu",
                        "middle": [],
                        "last": "Soricut",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Transactions of the Association for Computational Linguistics",
                "volume": "4",
                "issue": "",
                "pages": "1--16",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Manaal Faruqui, Ryan McDonald, and Radu Soricut. 2016. Morpho-syntactic lexicon generation using graph- based semi-supervised learning. Transactions of the Association for Computational Linguistics, 4:1-16.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Learning Discrete Structures for Graph Neural Networks",
                "authors": [
                    {
                        "first": "Luca",
                        "middle": [],
                        "last": "Franceschi",
                        "suffix": ""
                    },
                    {
                        "first": "Mathias",
                        "middle": [],
                        "last": "Niepert",
                        "suffix": ""
                    },
                    {
                        "first": "Massimiliano",
                        "middle": [],
                        "last": "Pontil",
                        "suffix": ""
                    },
                    {
                        "first": "Xiao",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "ICML",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Luca Franceschi, Mathias Niepert, Massimiliano Pontil, and Xiao He. 2019. Learning Discrete Structures for Graph Neural Networks. In ICML.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Seeing stars when there aren't many stars: Graph-based semisupervised learning for sentiment categorization",
                "authors": [
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaojin",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "HLT-NAACL 2006 Workshop on Textgraph",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andrew Goldberg and Xiaojin Zhu. 2006. Seeing stars when there aren't many stars: Graph-based semi- supervised learning for sentiment categorization. In HLT-NAACL 2006 Workshop on Textgraph.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Transductive Learning via Spectral Graph Partitioning",
                "authors": [
                    {
                        "first": "Thorsten",
                        "middle": [],
                        "last": "Joachims",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Machine Learning",
                "volume": "20",
                "issue": "",
                "pages": "290--297",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Thorsten Joachims. 2003. Transductive Learning via Spectral Graph Partitioning. Machine Learning, 20(1):290- 297.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "How to learn a graph from smooth signals",
                "authors": [
                    {
                        "first": "Vassilis",
                        "middle": [],
                        "last": "Kalofolias",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "AISTATS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Vassilis Kalofolias. 2016. How to learn a graph from smooth signals. In AISTATS.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Semi-supervised Classification with Graph-convolutional Neural Networks",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Thomas",
                        "suffix": ""
                    },
                    {
                        "first": "Max",
                        "middle": [],
                        "last": "Kipf",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Welling",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "ICLR",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Thomas N Kipf and Max Welling. 2017. Semi-supervised Classification with Graph-convolutional Neural Net- works. In ICLR.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Note on the sampling error of the difference between correlated proportions or percentages",
                "authors": [
                    {
                        "first": "Mcnemar",
                        "middle": [],
                        "last": "Quinn",
                        "suffix": ""
                    }
                ],
                "year": 1947,
                "venue": "Psychometrika",
                "volume": "12",
                "issue": "2",
                "pages": "153--157",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "McNemar Quinn. 1947. Note on the sampling error of the difference between correlated proportions or percent- ages. Psychometrika, 12(2):153-157.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Efficient Estimation of Word Representations in Vector Space",
                "authors": [
                    {
                        "first": "Tomas",
                        "middle": [],
                        "last": "Mikolov",
                        "suffix": ""
                    },
                    {
                        "first": "Kai",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Greg",
                        "middle": [],
                        "last": "Corrado",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Dean",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1301.3781"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient Estimation of Word Representations in Vector Space. arXiv:1301.3781.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Scalable Nearest Neighbor Algorithms for High Dimensional Data",
                "authors": [
                    {
                        "first": "Marius",
                        "middle": [],
                        "last": "Muja",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [
                            "G"
                        ],
                        "last": "Lowe",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
                "volume": "36",
                "issue": "11",
                "pages": "2227--2240",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Marius Muja and David G. Lowe. 2014. Scalable Nearest Neighbor Algorithms for High Dimensional Data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11):2227-2240.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Scikitlearn: Machine Learning in Python",
                "authors": [
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Pedregosa",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Varoquaux",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Gramfort",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Michel",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Thirion",
                        "suffix": ""
                    },
                    {
                        "first": "O",
                        "middle": [],
                        "last": "Grisel",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Blondel",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Prettenhofer",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Weiss",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Dubourg",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Vanderplas",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Passos",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Cournapeau",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Brucher",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Perrot",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Duchesnay",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Journal of Machine Learning Research",
                "volume": "12",
                "issue": "",
                "pages": "2825--2830",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit- learn: Machine Learning in Python. Journal of Machine Learning Research, 12:2825-2830.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Glove: Global vectors for word representation",
                "authors": [
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Pennington",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global vectors for word representa- tion. In EMNLP.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Deep contextualized word representations",
                "authors": [
                    {
                        "first": "Matthew",
                        "middle": [
                            "E"
                        ],
                        "last": "Peters",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Neumann",
                        "suffix": ""
                    },
                    {
                        "first": "Mohit",
                        "middle": [],
                        "last": "Iyyer",
                        "suffix": ""
                    },
                    {
                        "first": "Matt",
                        "middle": [],
                        "last": "Gardner",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Clark",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "NAACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettle- moyer. 2018. Deep contextualized word representations. In NAACL.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Graph-based semi-supervised learning of translation models from monolingual data",
                "authors": [
                    {
                        "first": "Avneesh",
                        "middle": [],
                        "last": "Saluja",
                        "suffix": ""
                    },
                    {
                        "first": "Hany",
                        "middle": [],
                        "last": "Hassan",
                        "suffix": ""
                    },
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Quirk",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Avneesh Saluja, Hany Hassan, Kristina Toutanova, and Chris Quirk. 2014. Graph-based semi-supervised learning of translation models from monolingual data. In ACL.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Soft-supervised Learning for Text Classification",
                "authors": [
                    {
                        "first": "Amarnag",
                        "middle": [],
                        "last": "Subramanya",
                        "suffix": ""
                    },
                    {
                        "first": "Jeff",
                        "middle": [],
                        "last": "Bilmes",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Amarnag Subramanya and Jeff Bilmes. 2008. Soft-supervised Learning for Text Classification. In EMNLP.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Graph-Based Semi-Supervised Learning",
                "authors": [
                    {
                        "first": "Amarnag",
                        "middle": [],
                        "last": "Subramanya",
                        "suffix": ""
                    },
                    {
                        "first": "Partha",
                        "middle": [
                            "Pratim"
                        ],
                        "last": "Talukdar",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Amarnag Subramanya and Partha Pratim Talukdar. 2014. Graph-Based Semi-Supervised Learning. Morgan & Claypool Publishers.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Efficient Graph-based Semi-supervised Learning of Structured Tagging Models",
                "authors": [
                    {
                        "first": "Amarnag",
                        "middle": [],
                        "last": "Subramanya",
                        "suffix": ""
                    },
                    {
                        "first": "Slav",
                        "middle": [],
                        "last": "Petrov",
                        "suffix": ""
                    },
                    {
                        "first": "Fernando",
                        "middle": [],
                        "last": "Pereira",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Amarnag Subramanya, Slav Petrov, and Fernando Pereira. 2010. Efficient Graph-based Semi-supervised Learning of Structured Tagging Models. In EMNLP.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Self-labeled techniques for semi-supervised learning: taxonomy, software and empirical study",
                "authors": [
                    {
                        "first": "Isaac",
                        "middle": [],
                        "last": "Triguero",
                        "suffix": ""
                    },
                    {
                        "first": "Salvador",
                        "middle": [],
                        "last": "Garc\u00eda",
                        "suffix": ""
                    },
                    {
                        "first": "Francisco",
                        "middle": [],
                        "last": "Herrera",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Knowledge and Information Systems",
                "volume": "42",
                "issue": "2",
                "pages": "245--284",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Isaac Triguero, Salvador Garc\u00eda, and Francisco Herrera. 2015. Self-labeled techniques for semi-supervised learn- ing: taxonomy, software and empirical study. Knowledge and Information Systems, 42(2):245-284.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Deep Learning via Semi-supervised Embedding",
                "authors": [
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Weston",
                        "suffix": ""
                    },
                    {
                        "first": "Fr\u00e9d\u00e9ric",
                        "middle": [],
                        "last": "Ratle",
                        "suffix": ""
                    },
                    {
                        "first": "Hossein",
                        "middle": [],
                        "last": "Mobahi",
                        "suffix": ""
                    },
                    {
                        "first": "Ronan",
                        "middle": [],
                        "last": "Collobert",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Neural Networks: Tricks of the Trade -Second Edition",
                "volume": "",
                "issue": "",
                "pages": "639--655",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jason Weston, Fr\u00e9d\u00e9ric Ratle, Hossein Mobahi, and Ronan Collobert. 2012. Deep Learning via Semi-supervised Embedding. In Neural Networks: Tricks of the Trade -Second Edition, pages 639-655.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Simplifying Graph Convolutional Networks",
                "authors": [
                    {
                        "first": "Felix",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Tianyi",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Amauri",
                        "middle": [],
                        "last": "Holanda De Souza",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Fifty",
                        "suffix": ""
                    },
                    {
                        "first": "Tao",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    },
                    {
                        "first": "Kilian",
                        "middle": [
                            "Q"
                        ],
                        "last": "Weinberger",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1902.07153"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Felix Wu, Tianyi Zhang, Amauri Holanda de Souza Jr., Christopher Fifty, Tao Yu, and Kilian Q. Weinberger. 2019. Simplifying Graph Convolutional Networks. arXiv:1902.07153.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Revisiting semi-supervised learning with graph embeddings",
                "authors": [
                    {
                        "first": "Zhilin",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "William",
                        "suffix": ""
                    },
                    {
                        "first": "Ruslan",
                        "middle": [],
                        "last": "Cohen",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Salakhutdinov",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "ICML",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. 2016. Revisiting semi-supervised learning with graph embeddings. In ICML.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Learning with Local and Global Consistency",
                "authors": [
                    {
                        "first": "Dengyong",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Olivier",
                        "middle": [],
                        "last": "Bousquet",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Thomas Navin Lal",
                        "suffix": ""
                    },
                    {
                        "first": "Bernhard",
                        "middle": [],
                        "last": "Weston",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Sch\u00f6lkopf",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "NIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Sch\u00f6lkopf. 2004. Learning with Local and Global Consistency. In NIPS.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Learning from Labeled and Unlabeled Data with Label Propagation",
                "authors": [
                    {
                        "first": "Xiaojin",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    },
                    {
                        "first": "Zoubin",
                        "middle": [],
                        "last": "Ghahramani",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xiaojin Zhu and Zoubin Ghahramani. 2002. Learning from Labeled and Unlabeled Data with Label Propagation. Technical Report CMU-CALD-02-107, Carnegie Mellon University.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Semi-supervised learning using Gaussian fields and harmonic functions",
                "authors": [
                    {
                        "first": "Xiaojin",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Ghahramani",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Lafferty",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "ICML",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xiaojin Zhu, Z Ghahramani, and John Lafferty. 2003. Semi-supervised learning using Gaussian fields and har- monic functions. In ICML.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "text": "Original 3-dimensional synthetic data. Semi-transparent points are unlabeled.",
                "num": null,
                "type_str": "figure",
                "uris": null
            },
            "FIGREF2": {
                "text": "3D PCA visualization of the original representation (left) and the representation learned with our approach (right) on the rel dataset (%75 unlabeled). Transparent dots represent unlabeled documents.",
                "num": null,
                "type_str": "figure",
                "uris": null
            },
            "FIGREF3": {
                "text": "Force-directed drawing (spring layout) of a random 50-node subgraph of the graph learned with our approach on the rel dataset (%75 labeled).",
                "num": null,
                "type_str": "figure",
                "uris": null
            },
            "TABREF1": {
                "text": "Classification accuracy on the synthetic dataset.",
                "html": null,
                "content": "<table/>",
                "num": null,
                "type_str": "table"
            }
        }
    }
}