File size: 57,590 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
{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:58:10.548059Z"
    },
    "title": "Contextual BERT: Conditioning the Language Model Using a Global State",
    "authors": [
        {
            "first": "Timo",
            "middle": [
                "I"
            ],
            "last": "Denk",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Zalando SE",
                "location": {
                    "settlement": "Berlin",
                    "country": "Germany"
                }
            },
            "email": "timo.denk@zalando.de"
        },
        {
            "first": "Ana",
            "middle": [
                "Peleteiro"
            ],
            "last": "Ramallo",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Zalando SE",
                "location": {
                    "settlement": "Berlin",
                    "country": "Germany"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "BERT is a popular language model whose main pre-training task is to fill in the blank, i.e., predicting a word that was masked out of a sentence, based on the remaining words. In some applications, however, having an additional context can help the model make the right prediction, e.g., by taking the domain or the time of writing into account. This motivates us to advance the BERT architecture by adding a global state for conditioning on a fixed-sized context. We present our two novel approaches and apply them to an industry use-case, where we complete fashion outfits with missing articles, conditioned on a specific customer. An experimental comparison to other methods from the literature shows that our methods improve personalization significantly.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "BERT is a popular language model whose main pre-training task is to fill in the blank, i.e., predicting a word that was masked out of a sentence, based on the remaining words. In some applications, however, having an additional context can help the model make the right prediction, e.g., by taking the domain or the time of writing into account. This motivates us to advance the BERT architecture by adding a global state for conditioning on a fixed-sized context. We present our two novel approaches and apply them to an industry use-case, where we complete fashion outfits with missing articles, conditioned on a specific customer. An experimental comparison to other methods from the literature shows that our methods improve personalization significantly.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Since its publication, the BERT model by Devlin et al. (2019) has enjoyed great popularity in the natural language processing (NLP) community. To apply the model to a specific problem, it is commonly pretrained on large amounts of unlabeled data, and subsequently fine-tuned on a target task. During both stages, the model's only input is a variably-sized sequence of words.",
                "cite_spans": [
                    {
                        "start": 41,
                        "end": 61,
                        "text": "Devlin et al. (2019)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "There are use-cases, however, where having an additional context can help the model. Consider a query intent classifier whose sole input is a user's text query. Under the assumption that users from different age groups and professions express the same intent in different ways, the classifier would benefit from having access to that user context in addition to the query. Alternatively, one might consider training multiple models on separate, age group-and profession-specific samples. However, this approach does not scale well, requires more training data, and does not share knowledge between the models.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "To the best of our knowledge, there is a shortcoming in effective methods for conditioning BERT on a fixed-sized context. Motivated by this, and inspired by the graph-networks perspective on self-attention models (Battaglia et al., 2018) , we advance BERT's architecture by adding a global state that enables conditioning. With our proposed methods [GS] and [GSU], we combine two previously independent streams of work. The first is centered around the idea of explicitly adding a global state to BERT, albeit without using it for conditioning. The second is focused on injecting additional knowledge into the BERT model. By using a global state for conditioning, we enable the application of BERT in a range of use-cases that require the model to make context-based predictions.",
                "cite_spans": [
                    {
                        "start": 213,
                        "end": 237,
                        "text": "(Battaglia et al., 2018)",
                        "ref_id": null
                    },
                    {
                        "start": 349,
                        "end": 353,
                        "text": "[GS]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We use the outfit completion problem to test the performance of our new methods: The model predicts fashion items to complete an outfit and has to account for both style coherence and personalization. For the latter, we condition on a fixed-sized customer representation containing information such as customer age, style preferences, hair color, and body type. We compare our methods against two others from the literature and observe that ours are able to provide more personalized predictions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "BERT's Global State In the original BERT paper, Devlin et al. (2019) use a [CLS] token which is prepended to the input sequence (e.g., a sentence of natural language). The assumption is that the model aggregates sentence-wide, global knowledge at the position of the [CLS] token. This intuition was confirmed through attention score analysis (Clark et al., 2019) , however, the BERT architecture does not have an inductive bias that aids it. Recent work therefore treats the [CLS] token differently. Zaheer et al. (2020) constrain their BERT variant Big Bird to local attention only, with the exception that every position may always attend to [CLS] regardless of its spatial proximity. Ke et al. (2020) also observe that the [CLS] attention exhibits peculiar patterns. This motivates them to introduce a separate set of weights for attending to and from [CLS] . The authors thereby explicitly encode into the architecture that the sequence's first position has a special role and different modality than the other positions. The result is an increased performance on downstream GLUE tasks.",
                "cite_spans": [
                    {
                        "start": 48,
                        "end": 68,
                        "text": "Devlin et al. (2019)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 342,
                        "end": 362,
                        "text": "(Clark et al., 2019)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 500,
                        "end": 520,
                        "text": "Zaheer et al. (2020)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 687,
                        "end": 703,
                        "text": "Ke et al. (2020)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 855,
                        "end": 860,
                        "text": "[CLS]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "It is important to note that all related work on BERT's global state does not use the global state for conditioning. Instead, the architectural changes are solely being introduced to improve the performance on non-contextual NLP benchmarks.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Conditioning on a Context To the best of our knowledge, Wu et al. (2018) are the first to provide sentence-wide information to the model to ease the masked language model (MLM) pre-training task. The authors inject the target label (e.g., positive or negative review) of sentiment data by adding it to the [CLS] token embedding. In a similar application, Li et al. (2020) process the context separately and subsequently combine it with the model output to make a sentiment prediction.",
                "cite_spans": [
                    {
                        "start": 56,
                        "end": 72,
                        "text": "Wu et al. (2018)",
                        "ref_id": null
                    },
                    {
                        "start": 355,
                        "end": 371,
                        "text": "Li et al. (2020)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Xia et al. (2020) condition on richer information, namely an intent, which can be thought of as a task descriptor given to the model. The intent is represented in text form, is variably sized, and prepended to the sequence. This is very similar to a wide range of GPT (Radford et al., 2019) applications.",
                "cite_spans": [
                    {
                        "start": 268,
                        "end": 290,
                        "text": "(Radford et al., 2019)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Chen et al. 2019condition on a customer's variably-sized click history using a Transformer (Vaswani et al., 2017) . The most similar to our work are Wu et al. (2020) who personalize by concatenating every position in the input sequence with a user embedding -method [C] from Section 3. Their approach, however, lacks an architectural bias that makes the model treat the user embedding as global information.",
                "cite_spans": [
                    {
                        "start": 91,
                        "end": 113,
                        "text": "(Vaswani et al., 2017)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 149,
                        "end": 165,
                        "text": "Wu et al. (2020)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "BERT as a Graph Neural Network (GNN) Battaglia et al. 2018introduce a framework that unites several lines of research on GNNs. In the Appendix, the authors show that -within their framework -the Transformer architecture is a type of GNN; Joshi (2020) supports this finding. In both cases the observation is that a sentence can be seen as a graph, where words correspond to nodes and the computation of an attention score is the assignment of a weight to an edge between two words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "In the GNN framework, a global state is accessible from every transfer function and can be individually updated from layer to layer. 1 Neither Transformer nor BERT, however, have a global state in that sense. Inspired by this observation, we introduce a global state and use it for conditioning. We explain our two novel methods in the following section, alongside with two that are derived from the literature.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Let w denote a sequence of n words w i \u2208 V from a fixed-sized vocabulary V. Further, let w \u2212i be the sequence without the ith word. Recall that a vanilla BERT model (Devlin et al., 2019) can predict the probability Pr(M = w i | w \u2212i ) that a word w i is masked-out in a sequence (M = w i being the masking event), conditioned on the other words in the sequence. Next, we introduce four methods to additionally condition BERT on a context vector c \u2208 R dcontext , which allows it to predict Pr( Figure 1 : A BERT block with global state. Boxes have learned parameters; our additions are bold.",
                "cite_spans": [
                    {
                        "start": 165,
                        "end": 186,
                        "text": "(Devlin et al., 2019)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 493,
                        "end": 501,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Conditioning BERT With a Global State",
                "sec_num": "3"
            },
            {
                "text": "M = w i | w \u2212i , c).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditioning BERT With a Global State",
                "sec_num": "3"
            },
            {
                "text": ":= W x 1 . . . x n c . . . c . W \u2208 R d model \u00d7(d model +dcontext",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conditioning BERT With a Global State",
                "sec_num": "3"
            },
            {
                "text": "Customer Context female, 28 y/o, 58kg, 172cm, Swiss, colors gray/red/black, slim, no heels, no leather, . .. Figure 2 : The fill-in-the-blank task on fashion outfits. Given a set of articles (left-hand side) and customer context, the model makes several predictions (right-hand side) for a masked out item (here: the coat). The predictions are personalized, because the model is utilizing the customer context. ",
                "cite_spans": [
                    {
                        "start": 17,
                        "end": 24,
                        "text": "female,",
                        "ref_id": null
                    },
                    {
                        "start": 25,
                        "end": 32,
                        "text": "28 y/o,",
                        "ref_id": null
                    },
                    {
                        "start": 33,
                        "end": 38,
                        "text": "58kg,",
                        "ref_id": null
                    },
                    {
                        "start": 39,
                        "end": 45,
                        "text": "172cm,",
                        "ref_id": null
                    },
                    {
                        "start": 46,
                        "end": 52,
                        "text": "Swiss,",
                        "ref_id": null
                    },
                    {
                        "start": 53,
                        "end": 75,
                        "text": "colors gray/red/black,",
                        "ref_id": null
                    },
                    {
                        "start": 76,
                        "end": 81,
                        "text": "slim,",
                        "ref_id": null
                    },
                    {
                        "start": 82,
                        "end": 91,
                        "text": "no heels,",
                        "ref_id": null
                    },
                    {
                        "start": 92,
                        "end": 103,
                        "text": "no leather,",
                        "ref_id": null
                    },
                    {
                        "start": 104,
                        "end": 105,
                        "text": ".",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 109,
                        "end": 117,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Masked",
                "sec_num": null
            },
            {
                "text": "W \u2208 R d model",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Masked",
                "sec_num": null
            },
            {
                "text": "Our method is inspired by the GNN perspective on BERT. Its implementation is similar to the way the Transformer (Vaswani et al., 2017) decoder attends to the encoder. [GS] treats the context as a read-only global state from which the internal representations can be updated. In order to adjust the architecture of BERT accordingly, we insert a global state attention layer between the intrasequence attention and the (originally) subsequent feed-forward neural network (FNN). Figure 1 shows how the inserted elements fit into the vanilla BERT block.",
                "cite_spans": [
                    {
                        "start": 112,
                        "end": 134,
                        "text": "(Vaswani et al., 2017)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 476,
                        "end": 484,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Global State [GS]",
                "sec_num": null
            },
            {
                "text": "More formally, let X (l) \u2208 R n\u00d7d model be the output of the lth BERT block (of which there are N ); let X (0) := x 1 . . . x n be the model input; and letc := FNN(c) be the global state derived from the context vector using a non-linear transformation FNN(x) := W 2 max(0, W 1 x + b 1 ) + b 2 . With our modification, X (l) is defined by first performing the normal intra-sequence attention as in BERT 1) ; multi-head attention can be used here as well. Then, also unchanged,\u00c2 := LayerNorm Dropout(A) + X (l\u22121) . The internal representation is then updated once more by reading 2 from the global statec with ",
                "cite_spans": [
                    {
                        "start": 505,
                        "end": 510,
                        "text": "(l\u22121)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 402,
                        "end": 404,
                        "text": "1)",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Global State [GS]",
                "sec_num": null
            },
            {
                "text": "A := Attention W Q X (l\u22121) , W K X (l\u22121) , W V X (l\u2212",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Global State [GS]",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "B := Attention V Q\u00c2 , V K c , V V c = V V c ,",
                        "eq_num": "(1)"
                    }
                ],
                "section": "Global State [GS]",
                "sec_num": null
            },
            {
                "text": "Note that in the [GS] method the global statec, which is being added to the internal representation in Equation 1, is the same for all blocks. With [GSU], a global state transfer functionc (l+1) := FNN c (l) updates the global state, with its initial value being derived from the context:c (1) := FNN(c). The weights of FNN are not shared between layers. Equation 1 is updated to read from the global state belonging to the lth layer: ",
                "cite_spans": [
                    {
                        "start": 204,
                        "end": 207,
                        "text": "(l)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Global State With Update [GSU]",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "B (l) [GSU] := V V c (l) .",
                        "eq_num": "(2"
                    }
                ],
                "section": "Global State With Update [GSU]",
                "sec_num": null
            },
            {
                "text": "We evaluate the performance of our proposed methods on a real-world industry problem: personalized fashion outfit completion (see Figure 2 ) for Europe's largest fashion platform. Our proprietary dataset consists of 380k outfits, created by professional stylists for individual customers. When styling a customer, i.e., putting together an outfit, the stylist has access to all customer features that we later use to condition the model. Therefore, customer data and outfit are statistically dependent. The customer features are individually embedded using trainable, randomly initialized embedding spaces. The per-feature embedding vectors are subsequently concatenated, yielding a context vector of d context = 736. Features include the customer's age, gender, country, preferred brands/colors/styles, nogo types, clothing sizes, price preferences, and the occasion for which the outfit is needed. The second model input, namely the outfit itself, is constructed from learned embeddings for every individual fashion article, with d model = 128. We stack N = 4 BERT blocks with multi-head attention (eight heads). For a masked-out item, the model predicts a probability distribution over |V| = 30 000 articles.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 130,
                        "end": 138,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Empirical Evaluation and Discussion",
                "sec_num": "4"
            },
            {
                "text": "While not being an NLP dataset, our data resembles many of the important traits of a textual corpus: the vocabulary size is comparable to the one of word-piece vocabularies commonly used with BERT models. Fashion outfits are similar to sentences in that some articles appear often together (match stylewise) and others do not. Different is the typical sequence length which ranges from four to eight fashion articles, with an average length of exactly five. In contrast to sentences, outfits do not have an inherent order. To account for that we remove the positional encoding from BERT so it treats its input as a set. Table 1 shows the results of evaluating the four different methods. We compare cross-entropy and recall@rank (r@r for short) on a randomly selected validation dataset consisting of 17k outfits that are held-out during training. The r@r is defined as the percentage of cases in which the masked-out item is among the top-r most probable items, according to the model. Model parameters are counted without embedding spaces for customer features and the last dense layer.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 620,
                        "end": 627,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Empirical Evaluation and Discussion",
                "sec_num": "4"
            },
            {
                "text": "The empirical evaluation reveals the effectiveness of using a context for making predictions. The model's ability to replicate the stylist behavior better, i.e., achieve a higher r@r, improves substantially with the addition of a context. On r@1 we see a relative improvement of +43% by using the [GSU] method for conditioning over using no customer context at all ([None]) and +16% compared to [NP] (the best method without global state).",
                "cite_spans": [
                    {
                        "start": 395,
                        "end": 399,
                        "text": "[NP]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Empirical Evaluation and Discussion",
                "sec_num": "4"
            },
            {
                "text": "A comparison of the four different conditioning methods shows [GSU] to be most effective, followed by [GS] , [NP] , and [C] . The methods [C] and [NP] do not have any bias towards treating the context vector specially. They attend to other positions in the sequence the same way they attend to the context. The superiority of [GS] and [GSU] can presumably be explained by their explicit architectural ability to retrieve information from the global state and therefore effectively utilize the context for their prediction.",
                "cite_spans": [
                    {
                        "start": 102,
                        "end": 106,
                        "text": "[GS]",
                        "ref_id": null
                    },
                    {
                        "start": 109,
                        "end": 113,
                        "text": "[NP]",
                        "ref_id": null
                    },
                    {
                        "start": 120,
                        "end": 123,
                        "text": "[C]",
                        "ref_id": null
                    },
                    {
                        "start": 138,
                        "end": 141,
                        "text": "[C]",
                        "ref_id": null
                    },
                    {
                        "start": 146,
                        "end": 150,
                        "text": "[NP]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Empirical Evaluation and Discussion",
                "sec_num": "4"
            },
            {
                "text": "We acknowledge the differences between our outfits dataset and typical NLP benchmarks. Nonetheless we hypothesize that the effectiveness of our method translates to NLP. In particular when applied to usecases in which the modality of context and sequence differ, e.g., for contexts comprised of numerical or categorical meta data about the text. That is because the model's freedom to read from the context separately allows it to process the different modalities of context and input sequence adequately.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Empirical Evaluation and Discussion",
                "sec_num": "4"
            },
            {
                "text": "With Contextual BERT, we presented novel ways of conditioning the BERT model. The strong performance on a real-world use-case provides evidence for the superiority of using a global state to inject context into the Transformer-based architecture. Our proposal enables the effective conditioning of BERT, potentially leading to improvements in a range of applications where contextual information is relevant.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions and Future Work",
                "sec_num": "5"
            },
            {
                "text": "A promising idea for follow-up work is to allow for information to flow from the sequence to the global state. Further, it would be desirable to establish a contextual NLP benchmark for the research community to compete on. This benchmark would task competitors with contextualized NLP problems, e.g., social media platform-dependent text generation or named entity recognition for multiple domains.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions and Future Work",
                "sec_num": "5"
            },
            {
                "text": "For details on the GNN definition of a global state we refer the reader to Section 3.2 inBattaglia et al. (2018).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Note that the global state sequence c does not admit for attention in the sense of selecting a weighted average of multiple vectors, because it consists only of a single vector. Instead, Equation 1 is reduced to the possibility for the lth layer to transform the context and update its internal state based on it.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF1": {
                "ref_id": "b1",
                "title": "POG: personalized outfit generation for fashion recommendation at alibaba ifashion",
                "authors": [
                    {
                        "first": "Wen",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Pipei",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Jiaming",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Xin",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    },
                    {
                        "first": "Cheng",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    },
                    {
                        "first": "Fei",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Chao",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Andreas",
                        "middle": [],
                        "last": "Pfadler",
                        "suffix": ""
                    },
                    {
                        "first": "Huan",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    },
                    {
                        "first": "Binqiang",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining",
                "volume": "",
                "issue": "",
                "pages": "2662--2670",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wen Chen, Pipei Huang, Jiaming Xu, Xin Guo, Cheng Guo, Fei Sun, Chao Li, Andreas Pfadler, Huan Zhao, and Binqiang Zhao. 2019. POG: personalized outfit generation for fashion recommendation at alibaba ifashion. In Ankur Teredesai, Vipin Kumar, Ying Li, R\u00f3mer Rosales, Evimaria Terzi, and George Karypis, editors, Proceed- ings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD 2019, Anchorage, AK, USA, August 4-8, 2019, pages 2662-2670. ACM.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "What does BERT look at? an analysis of bert's attention. CoRR",
                "authors": [
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Clark",
                        "suffix": ""
                    },
                    {
                        "first": "Urvashi",
                        "middle": [],
                        "last": "Khandelwal",
                        "suffix": ""
                    },
                    {
                        "first": "Omer",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of bert's attention. CoRR, abs/1906.04341.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "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": "2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "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 Jill Burstein, Christy Doran, and Thamar Solorio, editors, 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Transformers are graph neural networks. The Gradient",
                "authors": [
                    {
                        "first": "Chaitanya",
                        "middle": [],
                        "last": "Joshi",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chaitanya Joshi. 2020. Transformers are graph neural networks. The Gradient.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Rethinking positional encoding in language pre-training. ArXiv, abs",
                "authors": [
                    {
                        "first": "Guolin",
                        "middle": [],
                        "last": "Ke",
                        "suffix": ""
                    },
                    {
                        "first": "Di",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Tie-Yan",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Guolin Ke, Di He, and Tie-Yan Liu. 2020. Rethinking positional encoding in language pre-training. ArXiv, abs/2006.15595.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Enhancing BERT representation with context-aware embedding for aspect-based sentiment analysis",
                "authors": [
                    {
                        "first": "Xinlong",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Xingyu",
                        "middle": [],
                        "last": "Fu",
                        "suffix": ""
                    },
                    {
                        "first": "Guangluan",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Jiuniu",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Jin",
                        "suffix": ""
                    },
                    {
                        "first": "Qing",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Tianyuan",
                        "middle": [],
                        "last": "Xiang",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "IEEE Access",
                "volume": "8",
                "issue": "",
                "pages": "46868--46876",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xinlong Li, Xingyu Fu, Guangluan Xu, Yang Yang, Jiuniu Wang, Li Jin, Qing Liu, and Tianyuan Xiang. 2020. Enhancing BERT representation with context-aware embedding for aspect-based sentiment analysis. IEEE Access, 8:46868-46876.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Language models are unsupervised multitask learners",
                "authors": [
                    {
                        "first": "Alec",
                        "middle": [],
                        "last": "Radford",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Rewon",
                        "middle": [],
                        "last": "Child",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Luan",
                        "suffix": ""
                    },
                    {
                        "first": "Dario",
                        "middle": [],
                        "last": "Amodei",
                        "suffix": ""
                    },
                    {
                        "first": "Ilya",
                        "middle": [],
                        "last": "Sutskever",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "OpenAI Blog",
                "volume": "1",
                "issue": "8",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1(8):9.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Attention is all you need",
                "authors": [
                    {
                        "first": "Ashish",
                        "middle": [],
                        "last": "Vaswani",
                        "suffix": ""
                    },
                    {
                        "first": "Noam",
                        "middle": [],
                        "last": "Shazeer",
                        "suffix": ""
                    },
                    {
                        "first": "Niki",
                        "middle": [],
                        "last": "Parmar",
                        "suffix": ""
                    },
                    {
                        "first": "Jakob",
                        "middle": [],
                        "last": "Uszkoreit",
                        "suffix": ""
                    },
                    {
                        "first": "Llion",
                        "middle": [],
                        "last": "Jones",
                        "suffix": ""
                    },
                    {
                        "first": "Aidan",
                        "middle": [
                            "N"
                        ],
                        "last": "Gomez",
                        "suffix": ""
                    },
                    {
                        "first": "Lukasz",
                        "middle": [],
                        "last": "Kaiser",
                        "suffix": ""
                    },
                    {
                        "first": "Illia",
                        "middle": [],
                        "last": "Polosukhin",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "5998--6008",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pages 5998-6008.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "SSE-PT: Sequential recommendation via personalized transformer",
                "authors": [
                    {
                        "first": "Liwei",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Shuqing",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Cho-Jui",
                        "middle": [],
                        "last": "Hsieh",
                        "suffix": ""
                    },
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Sharpnack",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Liwei Wu, Shuqing Li, Cho-Jui Hsieh, and James Sharpnack. 2020. SSE-PT: Sequential recommendation via personalized transformer.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "CG-BERT: conditional text generation with BERT for generalized few-shot intent detection. CoRR, abs",
                "authors": [
                    {
                        "first": "Congying",
                        "middle": [],
                        "last": "Xia",
                        "suffix": ""
                    },
                    {
                        "first": "Chenwei",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Hoang",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "Jiawei",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Philip",
                        "middle": [
                            "S"
                        ],
                        "last": "Yu",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Congying Xia, Chenwei Zhang, Hoang Nguyen, Jiawei Zhang, and Philip S. Yu. 2020. CG-BERT: conditional text generation with BERT for generalized few-shot intent detection. CoRR, abs/2004.01881.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Big bird: Transformers for longer sequences. CoRR, abs",
                "authors": [
                    {
                        "first": "Manzil",
                        "middle": [],
                        "last": "Zaheer",
                        "suffix": ""
                    },
                    {
                        "first": "Guru",
                        "middle": [],
                        "last": "Guruganesh",
                        "suffix": ""
                    },
                    {
                        "first": "Avinava",
                        "middle": [],
                        "last": "Dubey",
                        "suffix": ""
                    },
                    {
                        "first": "Joshua",
                        "middle": [],
                        "last": "Ainslie",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Alberti",
                        "suffix": ""
                    },
                    {
                        "first": "Santiago",
                        "middle": [],
                        "last": "Onta\u00f1\u00f3n",
                        "suffix": ""
                    },
                    {
                        "first": "Philip",
                        "middle": [],
                        "last": "Pham",
                        "suffix": ""
                    },
                    {
                        "first": "Anirudh",
                        "middle": [],
                        "last": "Ravula",
                        "suffix": ""
                    },
                    {
                        "first": "Qifan",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Onta\u00f1\u00f3n, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2020. Big bird: Transformers for longer sequences. CoRR, abs/2007.14062.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "type_str": "figure",
                "uris": null,
                "num": null,
                "text": "Concat [C] Similar toWu et al. (2020), we concatenate the context vector with every position in the input sequence. Let x i \u2208 R d model denote the embedding of word w i at position i. The resulting input matrix is I[C]"
            },
            "FIGREF1": {
                "type_str": "figure",
                "uris": null,
                "num": null,
                "text": "\u00d7dcontext is a trainable weight matrix. The resulting input is I [NP] := W c x 1 . . . x n . The model's attention masks are adjusted, such that every position can attend to the new first position."
            },
            "FIGREF2": {
                "type_str": "figure",
                "uris": null,
                "num": null,
                "text": "and computingB := LayerNorm Dropout(B) +\u00c2 . Lastly, the BERT layer output is computed as X (l) := LayerNorm FNN B +B . The definitions of Attention(\u2022), LayerNorm(\u2022), and Dropout(\u2022) are identical to Vaswani et al. (2017). The weight matrices W and V are not shared between layers. The layer indices (l) of A, B, and the weight matrices are omitted to aid readability."
            },
            "TABREF0": {
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table><tr><td/><td>Layer Output</td></tr><tr><td>[GS]</td><td/></tr><tr><td>BERT Block</td><td>Add &amp; Norm</td></tr><tr><td>N\u00d7</td><td>FNN</td></tr><tr><td/><td>GS Attention</td></tr><tr><td/><td>Add &amp; Norm</td></tr><tr><td>Global State</td><td/></tr><tr><td/><td>Attention</td></tr><tr><td>FNN</td><td/></tr><tr><td>Context</td><td>Layer Input</td></tr></table>",
                "text": ") is a trainable weight matrix that reduces the input dimensionality."
            },
            "TABREF1": {
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table><tr><td>)</td></tr></table>",
                "text": "53% \u00b1 0.06 21.44% \u00b1 0.04 87.14% \u00b1 0.09 546 432 [C] 4.9428 \u00b1 0.0047 10.26% \u00b1 0.16 25.75% \u00b1 0.09 90.76% \u00b1 0.10 673 664 [NP] 4.9260 \u00b1 0.0078 10.53% \u00b1 0.06 26.27% \u00b1 0.25 90.80% \u00b1 0.11 640 768 [GS] 4.8542 \u00b1 0.0084 11.19% \u00b1 0.06 27.56% \u00b1 0.28 91.35% \u00b1 0.13 723 328 [GSU] 4.7459 \u00b1 0.0043 12.21% \u00b1 0.17 29.40% \u00b1 0.06 92.19% \u00b1 0.08 921 856 Comparison of the different conditioning methods on validation data. [None] provides no customer context to the model. Values are means across three training runs reported with standard error."
            }
        }
    }
}