File size: 44,967 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
{
    "paper_id": "2021",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:13:01.154517Z"
    },
    "title": "Comparing Euclidean and Hyperbolic Embeddings on the WordNet Nouns Hypernymy Graph",
    "authors": [
        {
            "first": "Sameer",
            "middle": [],
            "last": "Bansal",
            "suffix": "",
            "affiliation": {},
            "email": "sbansal70@bloomberg.net"
        },
        {
            "first": "Adrian",
            "middle": [],
            "last": "Benton",
            "suffix": "",
            "affiliation": {},
            "email": "abenton10@bloomberg.net"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Nickel and Kiela (2017) present a new method for embedding tree nodes in the Poincar\u00e9 ball, and suggest that these hyperbolic embeddings are far more effective than Euclidean embeddings at embedding nodes in large, hierarchically structured graphs like the WordNet nouns hypernymy tree. This is especially true in low dimensions (Nickel and Kiela, 2017, Table 1). In this work, we seek to reproduce their experiments on embedding and reconstructing the WordNet nouns hypernymy graph. Counter to what they report, we find that Euclidean embeddings are able to represent this tree at least as well as Poincar\u00e9 embeddings, when allowed at least 50 dimensions. We note that this does not diminish the significance of their work given the impressive performance of hyperbolic embeddings in very low-dimensional settings. However, given the wide influence of their work, our aim here is to present an updated and more accurate comparison between the Euclidean and hyperbolic embeddings.",
    "pdf_parse": {
        "paper_id": "2021",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Nickel and Kiela (2017) present a new method for embedding tree nodes in the Poincar\u00e9 ball, and suggest that these hyperbolic embeddings are far more effective than Euclidean embeddings at embedding nodes in large, hierarchically structured graphs like the WordNet nouns hypernymy tree. This is especially true in low dimensions (Nickel and Kiela, 2017, Table 1). In this work, we seek to reproduce their experiments on embedding and reconstructing the WordNet nouns hypernymy graph. Counter to what they report, we find that Euclidean embeddings are able to represent this tree at least as well as Poincar\u00e9 embeddings, when allowed at least 50 dimensions. We note that this does not diminish the significance of their work given the impressive performance of hyperbolic embeddings in very low-dimensional settings. However, given the wide influence of their work, our aim here is to present an updated and more accurate comparison between the Euclidean and hyperbolic embeddings.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Nickel and Kiela (2017) introduced a method for learning embeddings in hyperbolic space for large, hierarchically structured objects like the WordNet nouns hypernymy graph. This work convincingly shows that across a range of embedding dimensions, from as low as 5 to as high as 200, hyperbolic embeddings consistently outperformed their Euclidean counterparts (Nickel and Kiela, 2017, Table 1 ). Illustrating the difference in performance at the highest experimental setting of 200 dimensions, the mean average precision (MAP) score for hyperbolic embeddings was shown to be around 5 times that of Euclidean for embedding nouns in the WordNet hypernymy graph. 1 These experiments have been extremely influential, with the results on embedding the WordNet nouns hypernymy graph baselines often cited in later works on enhanced hyperbolic embeddings (De Sa et al., 2018; Ganea et al., 2018; Dhingra et al., 2018; L\u00f3pez et al., 1 The embeddings were evaluated on a reconstruction task where a MAP score closer to 1 indicates better performance. 2019 ; Balazevic et al., 2019; Feyisetan et al., 2019; Chami et al., 2020). 2 In this work, we reproduce the reconstruction error experiments on the WordNet noun hypernymy graph from Nickel and Kiela (2017) . Counter to what they report, we find that Euclidean word embeddings are as effective at encoding the WordNet nouns graph as hyperbolic embeddings when given at least 50 dimensions. In fact, Euclidean embeddings with \u2265 100 dimensions achieve lower reconstruction error over embeddings in the Lorentz model, an improved hyperbolic embedding method, which was published the following year (Nickel and Kiela, 2018) .",
                "cite_spans": [
                    {
                        "start": 360,
                        "end": 384,
                        "text": "(Nickel and Kiela, 2017,",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 660,
                        "end": 661,
                        "text": "1",
                        "ref_id": null
                    },
                    {
                        "start": 848,
                        "end": 868,
                        "text": "(De Sa et al., 2018;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 869,
                        "end": 888,
                        "text": "Ganea et al., 2018;",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 889,
                        "end": 910,
                        "text": "Dhingra et al., 2018;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 911,
                        "end": 926,
                        "text": "L\u00f3pez et al., 1",
                        "ref_id": null
                    },
                    {
                        "start": 1042,
                        "end": 1046,
                        "text": "2019",
                        "ref_id": null
                    },
                    {
                        "start": 1049,
                        "end": 1072,
                        "text": "Balazevic et al., 2019;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 1073,
                        "end": 1096,
                        "text": "Feyisetan et al., 2019;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 1097,
                        "end": 1119,
                        "text": "Chami et al., 2020). 2",
                        "ref_id": null
                    },
                    {
                        "start": 1225,
                        "end": 1248,
                        "text": "Nickel and Kiela (2017)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 1637,
                        "end": 1661,
                        "text": "(Nickel and Kiela, 2018)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 385,
                        "end": 392,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The inability to reproduce the reported Euclidean experiments has been raised in several issues in the associated GitHub repository. 3 This has also been acknowledged by the authors of the original study, who suggest that the original Euclidean embeddings were regularized in a way that may have hurt performance. 4 However, the published manuscript has not been updated to reflect these problems with reproducing the Euclidean embedding baselines. As such, we hope that our reproduction will serve as a useful reference for those who are interested in exploring hyperbolic embeddings.",
                "cite_spans": [
                    {
                        "start": 133,
                        "end": 134,
                        "text": "3",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We use the source code released by the authors to carry out all experiments in this study, reusing the data processing, model training, and evaluation pipelines.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Setup",
                "sec_num": "2"
            },
            {
                "text": "Dataset. Following Nickel and Kiela (2017), we embed the WordNet noun hierarchy (Fellbaum, 1998, WordNet-Nouns) in both Euclidean and hyperbolic space. Though the original study also published results on additional datasets, we restrict our focus to WordNet-Nouns, which exhibited a considerable gap in performance between embeddings.",
                "cite_spans": [
                    {
                        "start": 80,
                        "end": 111,
                        "text": "(Fellbaum, 1998, WordNet-Nouns)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Setup",
                "sec_num": "2"
            },
            {
                "text": "Model training. Other than learning rate, we retain the default hyperparameters specified in the released source code, and train embeddings for 1,500 epochs. The learning rate is tuned in the range 10 [\u22122,3] independently for each class of embeddings ( Figure 1 ) with dimensionality fixed to 20, and selection based on loss after 200 epochs. We selected a learning rate of 0.5 for Euclidean and 5.0 for both Poincar\u00e9 and Lorentz embeddings. In our initial experiments, we found that other hyperparameters, such as number of negative samples, also affected embeddings performance, but were less influential than learning rate. See Appendix A for details on the exact version of the codebase used in our experiments and how it was called. See the original study (Nickel and Kiela, 2017) for additional details on model training and evaluation.",
                "cite_spans": [
                    {
                        "start": 761,
                        "end": 785,
                        "text": "(Nickel and Kiela, 2017)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 253,
                        "end": 261,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Experimental Setup",
                "sec_num": "2"
            },
            {
                "text": "Evaluation. Embeddings are evaluated under the original reconstruction error setting. For each hypernym pair < u, v > in the tree, rank all non-hypernyms along with v by distance from u in the embedded space. The fidelity to which a set of embeddings represents the tree is evaluated according to mean average precision (MAP) and mean rank (MR) of the positive example, v, averaged across rankings. We focus on reconstruction error experiments as they are meant to highlight the capacity of each embedding space. The ability to generalize out of sample is an orthogonal question, however which we don't address in this work. This is mostly since the source code to reproduce these results from Nickel and Kiela (2017) has not yet been released, and the particular folds of heldout edges are also not provided. As such, we leave reproduction of the link prediction evaluation of Euclidean vs. hyperbolic embeddings to future work. 5 3 Results Table 1 shows the MAP and MR results for the WordNet-Nouns hierarchy reconstruction task. There are clear differences between the reproduced and reported performance for Euclidean embeddings. In the 50 dimensions setting, the reproduced Euclidean embeddings achieve a MAP score of 88.9 compared to 14 in the original study, and an MR score of 1.8 compared to 1,281. In fact, this MR score for Euclidean embed- 5 At the time of writing, there is an open GitHub issue to provide more details on the out of sample, link prediction evaluation https://github.com/ facebookresearch/poincare-embeddings/ issues/10. dings at 50 dimensions is better than that achieved by Poincar\u00e9 and Lorentz embeddings even with 200 dimensions. With greater than 50 dimensions, Euclidean embeddings outperform both Lorentz and Poincar\u00e9 embeddings according to MR and MAP. In contrast, the performance of hyperbolic embeddings remain stable across dimensionality and are similar between reproduced and reported results. The improved reconstruction error of Lorentz embeddings in prior work is likely due to a more comprehensive hyperparameter search.",
                "cite_spans": [
                    {
                        "start": 694,
                        "end": 717,
                        "text": "Nickel and Kiela (2017)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 930,
                        "end": 931,
                        "text": "5",
                        "ref_id": null
                    },
                    {
                        "start": 1352,
                        "end": 1353,
                        "text": "5",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 942,
                        "end": 949,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Experimental Setup",
                "sec_num": "2"
            },
            {
                "text": "We also found that performance is robust to random seed for all methods, with a standard deviation of less than a point for both MAP and MR score when training 50-dimensional Euclidean, Poincar\u00e9, or Lorentz embeddings (Table 2) Kiela, 2017, 2018) and our reproduction with Euclidean embeddings constrained to unit norm are similar.",
                "cite_spans": [
                    {
                        "start": 228,
                        "end": 246,
                        "text": "Kiela, 2017, 2018)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 218,
                        "end": 227,
                        "text": "(Table 2)",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Experimental Setup",
                "sec_num": "2"
            },
            {
                "text": "faithfully than similarly trained hyperbolic embeddings with at least 50 dimensions. In this section, we discuss possible reasons for the discrepancy between reported and reproduced Euclidean embeddings performance. We first posed this question to the authors themselves (Nickel and Kiela, 2017) who clarified on GitHub that the difference in performance for Euclidean embeddings in their published manuscript was due to a regularization method used at the time. They further added that in the released code they \"disabled this regularization by default and it turned out to work better\". 6 Follow-up questions regarding the details of this regularization method have yet to be addressed, at the time of writing. 7",
                "cite_spans": [
                    {
                        "start": 271,
                        "end": 295,
                        "text": "(Nickel and Kiela, 2017)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Setup",
                "sec_num": "2"
            },
            {
                "text": "We speculate that the authors may have normalized the Euclidean embeddings to constrain them to lie within a unit 2-norm ball, similar to how Poincar\u00e9 embeddings are trained (Nickel and Kiela, 2017, Section 3.1) . Note that while projection into the unit ball is necessary to learn valid Poincar\u00e9 embeddings, Euclidean embeddings require no such constraint. The released source code actually supports projecting Euclidean embeddings into the unit ball after each iteration, but this is disabled by default, with the argument max norm set to None. 8 To test whether this constraint has an effect on embedding quality, we train Euclidean embeddings constrained to the unit ball by setting max norm to 1. Table 3 shows that reconstruction scores for Euclidean embeddings constrained to the unit ball are much closer to those published in Nickel 6 Author response regarding difference in Euclidean performance https://github.com/ facebookresearch/poincare-embeddings/ issues/35#issuecomment-685261354",
                "cite_spans": [
                    {
                        "start": 174,
                        "end": 211,
                        "text": "(Nickel and Kiela, 2017, Section 3.1)",
                        "ref_id": null
                    },
                    {
                        "start": 842,
                        "end": 843,
                        "text": "6",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 702,
                        "end": 709,
                        "text": "Table 3",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Constraining Euclidean Embedding Norm",
                "sec_num": null
            },
            {
                "text": "7 Question seeking further information regarding regularization https://github.com/ facebookresearch/poincare-embeddings/ issues/35#issuecomment-685261354 and a relevant comment in issuecomment-735209781.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constraining Euclidean Embedding Norm",
                "sec_num": null
            },
            {
                "text": "8 L2 normalization disabled by default: https://github.com/facebookresearch/ poincare-embeddings/blob/4c7316b/hype/ manifolds/euclidean.py#L16.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constraining Euclidean Embedding Norm",
                "sec_num": null
            },
            {
                "text": "and Kiela (2017, 2018) than unconstrained Euclidean embeddings.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constraining Euclidean Embedding Norm",
                "sec_num": null
            },
            {
                "text": "Varying Norm Constraints To explore the impact of norm constraints, we conduct further experiments on training 100 dimensional Euclidean embeddings. We vary the max norm setting between 1 and 10, allowing the Euclidean embeddings to grow larger during training. As an alternative method for controlling the norm of the embeddings, we also vary the strength of an L2 penalty as an additional term in the loss. MAP and MR scores improve as the max norm setting is increased, the norm constraint is relaxed (Table 4 ). In fact, setting max norm to 5 yields Euclidean embeddings that achieve similar reconstruction performance as unconstrained embeddings.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 504,
                        "end": 512,
                        "text": "(Table 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Constraining Euclidean Embedding Norm",
                "sec_num": null
            },
            {
                "text": "We found that including an L2 regularization penalty in the loss has little effect on the final reconstruction scores. Thus, although we suspect unnecessary renormalization of the Euclidean embeddings may have caused the poor reconstruction performance reported in prior work, we defer to the authors of the original study for confirmation. Table 4 : MAP and MR for 100 dimensional embeddings. We include results with no regularization or constraints (row 1) and with L2 regularization of varying degree. max norm = k means that embeddings are projected back into a radius k 2-norm ball after each iteration. N&K refers to best published results from Kiela (2017, 2018) .",
                "cite_spans": [
                    {
                        "start": 651,
                        "end": 669,
                        "text": "Kiela (2017, 2018)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 341,
                        "end": 348,
                        "text": "Table 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Constraining Euclidean Embedding Norm",
                "sec_num": null
            },
            {
                "text": "In our reproduction of the experiments in Table 1 of Nickel and Kiela (2017) , we find that Euclidean actually outperform Poincar\u00e9 embeddings when allowed a moderate number of dimensions. This is a realistic number of dimensions for typical non-contextual word embeddings, and a far lower dimensionality than subword token embeddings used in pretrained transformer language models. For example, released GloVe embeddings range from 50 to 300 dimensions (Pennington et al., 2014) and BERT base uses 768-dimensional subword embeddings (Devlin et al., 2019) . Nevertheless, the strong performance of hyperbolic embeddings in very low dimensions (less than 20) highlights their main strength: succinctly embedding nodes in hierarchically structured graphs with tight limitations on embedding size.",
                "cite_spans": [
                    {
                        "start": 53,
                        "end": 76,
                        "text": "Nickel and Kiela (2017)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 453,
                        "end": 478,
                        "text": "(Pennington et al., 2014)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 533,
                        "end": 554,
                        "text": "(Devlin et al., 2019)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 42,
                        "end": 49,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "5"
            },
            {
                "text": "However, part of Nickel and Kiela (2017)'s impact came from the astounding gains from hyperbolic embeddings over Euclidean across a wide range of embedding widths. Subsequent application of Poincar\u00e9 embeddings often report mixed results when using non-Euclidean vs. Euclidean embeddings in downstream tasks (Dhingra et al., 2018; L\u00f3pez et al., 2019) . We hope that this reproduction will serve as a valuable reference for others who are just beginning to explore hyperbolic embeddings.",
                "cite_spans": [
                    {
                        "start": 307,
                        "end": 329,
                        "text": "(Dhingra et al., 2018;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 330,
                        "end": 349,
                        "text": "L\u00f3pez et al., 2019)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "5"
            },
            {
                "text": "Nickel and Kiela (2017) has been cited over 500 times (source: www.semanticscholar.org).3 https://github.com/facebookresearch/ poincare-embeddings/issues/35 ; 68 ; 72 4 Author response on GitHub github.com/ facebookresearch/poincare-embeddings/ issues/35#issuecomment-685174866",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "AnalysisIn Section 3, we show it is possible to learn Euclidean embeddings that can reconstruct WordNet-Nouns more",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We thank Maximilian Nickel and Douwe Kiela for their helpful feedback while carrying out this study and for making the source code available. We also thank the workshop organizers for providing an avenue for such work to be published and the anonymous reviewers whose feedback helped us improve this work. Thanks also to the GitHub users who reported similar issues in the source code repository, which motivated us to publish this study.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            },
            {
                "text": "Sample call to train embeddings and run reconstruction evaluation:# ## P a r a m e t e r s LR= 0 . 5 DIM=20 MANIFOLD=\" e u c l i d e a n \" We use the poincare-embeddings implementation https://github. com/facebookresearch/ poincare-embeddings at commit 4c7316b14dce3b89e6a2d0c7994d418dffb42c94.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Call to Poincar\u00e9 Embedding Trainer",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Multi-relational poincar\u00e9 graph embeddings",
                "authors": [
                    {
                        "first": "Ivana",
                        "middle": [],
                        "last": "Balazevic",
                        "suffix": ""
                    },
                    {
                        "first": "Carl",
                        "middle": [],
                        "last": "Allen",
                        "suffix": ""
                    },
                    {
                        "first": "Timothy",
                        "middle": [],
                        "last": "Hospedales",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proc. NeurIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ivana Balazevic, Carl Allen, and Timothy Hospedales. 2019. Multi-relational poincar\u00e9 graph embeddings. In Proc. NeurIPS.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Lowdimensional hyperbolic knowledge graph embeddings",
                "authors": [
                    {
                        "first": "Ines",
                        "middle": [],
                        "last": "Chami",
                        "suffix": ""
                    },
                    {
                        "first": "Adva",
                        "middle": [],
                        "last": "Wolf",
                        "suffix": ""
                    },
                    {
                        "first": "Da-Cheng",
                        "middle": [],
                        "last": "Juan",
                        "suffix": ""
                    },
                    {
                        "first": "Frederic",
                        "middle": [],
                        "last": "Sala",
                        "suffix": ""
                    },
                    {
                        "first": "Sujith",
                        "middle": [],
                        "last": "Ravi",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "R\u00e9",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proc. ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ines Chami, Adva Wolf, Da-Cheng Juan, Frederic Sala, Sujith Ravi, and Christopher R\u00e9. 2020. Low- dimensional hyperbolic knowledge graph embed- dings. In Proc. ACL.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Representation tradeoffs for hyperbolic embeddings",
                "authors": [
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Christopher De Sa",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Gu",
                        "suffix": ""
                    },
                    {
                        "first": "Frederic",
                        "middle": [],
                        "last": "R\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Sala",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proc. MLR",
                "volume": "80",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Christopher De Sa, Albert Gu, Christopher R\u00e9, and Frederic Sala. 2018. Representation tradeoffs for hy- perbolic embeddings. Proc. MLR, 80:4460.",
                "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": "Proc. NAACL",
                "volume": "",
                "issue": "",
                "pages": "4171--4186",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understand- ing. In Proc. NAACL, pages 4171-4186.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Embedding text in hyperbolic spaces",
                "authors": [
                    {
                        "first": "Bhuwan",
                        "middle": [],
                        "last": "Dhingra",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Shallue",
                        "suffix": ""
                    },
                    {
                        "first": "Mohammad",
                        "middle": [],
                        "last": "Norouzi",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Dai",
                        "suffix": ""
                    },
                    {
                        "first": "George",
                        "middle": [],
                        "last": "Dahl",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proc. (TextGraphs-12)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W18-1708"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Bhuwan Dhingra, Christopher Shallue, Mohammad Norouzi, Andrew Dai, and George Dahl. 2018. Embedding text in hyperbolic spaces. In Proc. (TextGraphs-12).",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "WordNet: An Electronic Lexical Database",
                "authors": [
                    {
                        "first": "Christiane",
                        "middle": [],
                        "last": "Fellbaum",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Leveraging hierarchical representations for preserving privacy and utility in text",
                "authors": [
                    {
                        "first": "Oluwaseyi",
                        "middle": [],
                        "last": "Feyisetan",
                        "suffix": ""
                    },
                    {
                        "first": "Tom",
                        "middle": [],
                        "last": "Diethe",
                        "suffix": ""
                    },
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Drake",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proc. ICDM",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Oluwaseyi Feyisetan, Tom Diethe, and Thomas Drake. 2019. Leveraging hierarchical representations for preserving privacy and utility in text. In Proc. ICDM.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Hyperbolic neural networks",
                "authors": [
                    {
                        "first": "Octavian",
                        "middle": [],
                        "last": "Ganea",
                        "suffix": ""
                    },
                    {
                        "first": "Gary",
                        "middle": [],
                        "last": "B\u00e9cigneul",
                        "suffix": ""
                    },
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Hofmann",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proc. NeurIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Octavian Ganea, Gary B\u00e9cigneul, and Thomas Hof- mann. 2018. Hyperbolic neural networks. In Proc. NeurIPS.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Fine-grained entity typing in hyperbolic space",
                "authors": [
                    {
                        "first": "Federico",
                        "middle": [],
                        "last": "L\u00f3pez",
                        "suffix": ""
                    },
                    {
                        "first": "Benjamin",
                        "middle": [],
                        "last": "Heinzerling",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Strube",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proc",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-4319"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Federico L\u00f3pez, Benjamin Heinzerling, and Michael Strube. 2019. Fine-grained entity typing in hyper- bolic space. In Proc. (RepL4NLP-2019).",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Poincar\u00e9 embeddings for learning hierarchical representations",
                "authors": [
                    {
                        "first": "Maximillian",
                        "middle": [],
                        "last": "Nickel",
                        "suffix": ""
                    },
                    {
                        "first": "Douwe",
                        "middle": [],
                        "last": "Kiela",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proc. NeurIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Maximillian Nickel and Douwe Kiela. 2017. Poincar\u00e9 embeddings for learning hierarchical representa- tions. In Proc. NeurIPS.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Learning continuous hierarchies in the lorentz model of hyperbolic geometry",
                "authors": [
                    {
                        "first": "Maximillian",
                        "middle": [],
                        "last": "Nickel",
                        "suffix": ""
                    },
                    {
                        "first": "Douwe",
                        "middle": [],
                        "last": "Kiela",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proc. ICML",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Maximillian Nickel and Douwe Kiela. 2018. Learning continuous hierarchies in the lorentz model of hyper- bolic geometry. In Proc. ICML.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Glove: Global vectors for word representation",
                "authors": [
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Pennington",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher D",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proc. EMNLP",
                "volume": "",
                "issue": "",
                "pages": "1532--1543",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word rep- resentation. In Proc. EMNLP, pages 1532-1543.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "text": "MAP of Euclidean, Poincar\u00e9, and Lorentz 20dimensional embeddings as a function of learning rate. MAP scores for Poincar\u00e9 and Lorentz embeddings are very similar up to a learning rate of 10.0.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "TABREF1": {
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table/>",
                "text": "Mean and standard deviation of MAP and MR for 50-dimensional embeddings across three different random restarts."
            },
            "TABREF3": {
                "html": null,
                "type_str": "table",
                "num": null,
                "content": "<table/>",
                "text": "MAP and MR for reconstructing the WordNet-Nouns hypernymy graph. Results from N&K"
            }
        }
    }
}