File size: 44,527 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
{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T12:14:06.782062Z"
    },
    "title": "ReINTEL Challenge 2020: Vietnamese Fake News Detection using Ensemble Model with PhoBERT embeddings",
    "authors": [
        {
            "first": "Cao",
            "middle": [],
            "last": "Nguyen",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "VNG Corporation",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Nguyen",
            "middle": [],
            "last": "Hieu",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "VNG Corporation",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "",
            "middle": [],
            "last": "Thuan",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "VNG Corporation",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Vo",
            "middle": [],
            "last": "Quoc",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "VNG Corporation",
                "location": {}
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Along with the increasing traffic of social networks in Vietnam in recent years, the number of unreliable news has also grown rapidly. As we make decisions based on the information we come across daily, fake news, depending on the severity of the matter, can lead to disastrous consequences. This paper presents our approach for the Fake News Detection on Social Network Sites (SNSs), using an ensemble method with linguistic features extracted using PhoBERT (Nguyen and Nguyen, 2020). Our method achieves AUC score of 0.9521 and got 1 st place on the private test at the 7 th International Workshop on Vietnamese Language and Speech Processing (VLSP). For reproducing the result, the code can be found at https://gitlab.com/thuan.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Along with the increasing traffic of social networks in Vietnam in recent years, the number of unreliable news has also grown rapidly. As we make decisions based on the information we come across daily, fake news, depending on the severity of the matter, can lead to disastrous consequences. This paper presents our approach for the Fake News Detection on Social Network Sites (SNSs), using an ensemble method with linguistic features extracted using PhoBERT (Nguyen and Nguyen, 2020). Our method achieves AUC score of 0.9521 and got 1 st place on the private test at the 7 th International Workshop on Vietnamese Language and Speech Processing (VLSP). For reproducing the result, the code can be found at https://gitlab.com/thuan.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Social network sites have become a very influential part of Vietnamese people's daily life. We use them to connect with each other, and get access to the latest information. However, such advances in large scale communication also bring their problems, one of which is fake news. It can be seen as information which is altered, manipulated, misguiding users to achieve personal gains, such as increase advertisement interaction, political power gain, or even terrorism. Without proper censoring, they can spread fear in the public community, causing panic and invoking violence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Due to such dire consequences, a lot of researches have been done to prevent this type of harmful information. However, there has been little effort put in for the Vietnamese language. This is a challenging task, due to a lack of quality humanverified data, and the difficult nature of the fake contents. Fake news may have:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "\u2022 Similar contents to the real ones, however some key information is twisted (figures, celebrities, locations, ...) in order to capture the attention of readers.",
                "cite_spans": [
                    {
                        "start": 77,
                        "end": 115,
                        "text": "(figures, celebrities, locations, ...)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "\u2022 Contents encapsulated inside images, which requires human verification",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "\u2022 Special slangs, acronyms, misspellings which makes it difficult for machine to automate the process",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "\u2022 Unseen information that can take times before it is verified, which then might be too late",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we present our approach to the problem of fake news detection presented at the VLSP 2020, shared-task Reliable Intelligence Identification on Vietnamese SNSs (ReINTEL) (Le et al., 2020) . We experimented with 3 types of features: the time the news is posted, the community interaction to its (through number of share, like, comment) and, most importantly, the content of the news. After much preprocessing and exploration had been done, we combined the strength of basic handcrafted linguistic cues in the training data with term frequency encoding (TF-IDF) and PhoBERT as context embedding. These features are combined and used as input for an ensemble model using StackNet 1 . Our model achieved the AUC score of 0.9521, ranked first place on the private leader board of ReINTEL.",
                "cite_spans": [
                    {
                        "start": 183,
                        "end": 200,
                        "text": "(Le et al., 2020)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We discuss related work and previous approaches in section 2. We then describe our method workflow in section 3, starting with data cleaning and preprocessing, how we extracted the features we used, and the ensemble of models for our final result. Experiment's results and detailed description of parameters are shown in section 4. We conclude our report and discuss what could be improved in section 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "For the linguistic-based features, some approaches focus on extract special discriminative features such as acronymns, pronoun, special characters (Shu et al., 2017; Gupta et al., 2014) . However, these features are not well understood, as well as require extensive labour for validation and can be domain specific. Ruchansky et al. extend the method by using doc2vec embeddings, which learn semantic representation of the posts. Recent advancement in Natural Language Processing, and most importantly BERT (Devlin et al., 2018) , has helped to advance the research on this topic. Bhatt et al. combine the context generated by using LSTM and CNN, in combination with statistically handcrafted features to perform the final prediction.The work by Yang et al. use a combination of multiple Recurrent Neural Network (RNN) architectures as a natural language inference (NLI) mechanism, combining with BERT to make the final prediction. Research done by Huang and Chen focuses more on ensembling multiple deep learning architectures to achieve State Of The Art result for Fake News Detection. Ahmad et al. also shows that ensembling methods help achieve better performance on the current task.",
                "cite_spans": [
                    {
                        "start": 147,
                        "end": 165,
                        "text": "(Shu et al., 2017;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 166,
                        "end": 185,
                        "text": "Gupta et al., 2014)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 507,
                        "end": 528,
                        "text": "(Devlin et al., 2018)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related works",
                "sec_num": "2"
            },
            {
                "text": "In this section, we will describe our approach to solve the problem. Linguistic features extracted with PhoBERT and tf-idf, in conjunction with metadata provided, are used as input to an ensemble of models to achieve the best result in the private dataset. Using models that don't require much computation power not only helps us to tune each model quickly, but also enable us to analyze the impact of each feature on the fake news detection problem as a whole.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methodology",
                "sec_num": "3"
            },
            {
                "text": "To extract valuable features, we started with some preprocessing steps, which is described as follow:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "3.1"
            },
            {
                "text": "1. Convert numeric-like features to numeric type if possible, null value otherwise;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "3.1"
            },
            {
                "text": "2. Remove rows having null or empty content;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "3.1"
            },
            {
                "text": "3. Deduplicated rows having the same content and interactions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "3.1"
            },
            {
                "text": "The first step were applied on both training and test set, while the remain ones were done only on training set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing",
                "sec_num": "3.1"
            },
            {
                "text": "We considered all features except the content of the posts are metadata features.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metadata",
                "sec_num": "3.2.1"
            },
            {
                "text": "Number of likes, comments, and shares: We first transformed these 3 features to log scale for normalization. Then for each of them, a is_null feature were generated, equaling to 0 if the corresponding value is presented, and 1 otherwise.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metadata",
                "sec_num": "3.2.1"
            },
            {
                "text": "We extracted the hour and the day of week from the timestamp of posts.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Timestamp of posts:",
                "sec_num": null
            },
            {
                "text": "Combinations: We tried to generate some combinations of the above numeric features. Particularly, we computed the divisions of the number of likes, comments, and shartes to each other and obtained 3 new numeric features.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Timestamp of posts:",
                "sec_num": null
            },
            {
                "text": "Finally, any not-a-number value was filled by -1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Timestamp of posts:",
                "sec_num": null
            },
            {
                "text": "Term Frequency -Inverse Document Frequency (TF-IDF): TF-IDF is a simple but strong feature extraction technique for text data. We fitted a TF-IDF vectorizer from 1-gram to 3-gram on post contents of our training data, followed by a Single Value Decomposition (SVD) model to reduce the dimension of transformed TF-IDF features. A 300dimensional vector of latent features was obtained for each post at the end of this step.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Post content",
                "sec_num": "3.2.2"
            },
            {
                "text": "PhoBERT Embedding: BERT (Devlin et al., 2018) is a robust language model recently boosting many NLP tasks to a new level of achievement.",
                "cite_spans": [
                    {
                        "start": 24,
                        "end": 45,
                        "text": "(Devlin et al., 2018)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Post content",
                "sec_num": "3.2.2"
            },
            {
                "text": "PhoBERT (Nguyen and Nguyen, 2020), in our knowledge, is the best pre-trained BERT model for Vietnamese. In our solution, we leveraged PhoBERT to extract document embeddings from the posts. Notably, to receive more meaningful contextual embedding, some cleaning operations were applied to the contents before feeding into PhoBERT, consisting of word tokenization, special characters removal, redundant content removal. Moreover, another SVD model was fitted on top of those embedding to map 768-d output vectors of the BERT model to 100-dimensional space.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Post content",
                "sec_num": "3.2.2"
            },
            {
                "text": "Characters Counting: After extensive exploratory analysis, it turned out that the occurrence of some special characters and patterns have impact on the performance of our model, such as question mark, exclamation mark, triple dot, link, and so on. Thus, we created a list of those characters and created corresponding features which present the number of each of them in the posts.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Post content",
                "sec_num": "3.2.2"
            },
            {
                "text": "Tree-based models are the first choice when dealing with tabular data, thanks to their strength in both predictability and explainability. Furthermore, ensemble learning, especially stacking, is a good way to prevent overfitting and improve the performance of the overall system. Pursuing these observations, we designed our modeling phase as an ensemble system including 25 different base models and 5 stacked models on top of them. Precisely, the base models are from 5 different kinds: 5 Random Forests, 5 LightGBM Gradient Boosting Trees (GDBTs), 5 CatBoost GDBTs, 5 shallow Neural Networks, and 5 Naive Bayes classifiers; and the stacked models are 5 CatBoost GDBTs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Modelling",
                "sec_num": "3.3"
            },
            {
                "text": "Training phase: we formulate our training data in a 5-folds cross-validation manner. In each fold, 5 different-kind models were trained. After these training finished, 5 probability vectors were predicted and treated as 5 features, combined with the original features to form a new training set to train the corresponding stacked model of that fold.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Modelling",
                "sec_num": "3.3"
            },
            {
                "text": "Inference phase: probabilities from 5 trained stacked models are averaged to get final scores.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Modelling",
                "sec_num": "3.3"
            },
            {
                "text": "We evaluated our methods on the datasets provided by the 2020 VLSP competition, which contain totally about 6000 training and 2000 testing examples, divided into multiple sets described in table 1. The manually annotated labels equal to 1 if the news as potentially unreliable, and 0 otherwise. Our training set is composed of the public training and the warmup training set. Table 2 is a statistic summarization of our training set. After the feature engineering steps, our final training set consisted of 420 features and 4956 examples, 831 (16.8%) of which are label 1.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 376,
                        "end": 383,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Datasets",
                "sec_num": "4.1"
            },
            {
                "text": "It should be noted that, although only the 2 training sets contain labels, we still leveraged the con- tent of posts from all datasets except the private one to extract features described in section 3.2.2. This way of making full use of unlabeled data help the model generalize well and result in better performance. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Datasets",
                "sec_num": "4.1"
            },
            {
                "text": "All steps were executed on the same machine with the following specs: 4 Intel Xeon CPUs 2.20GHz, 1 16GB RAM, and 1 Tesla T4 16GB GPU. The step that occupied the most amount of RAM (~10GB) is fitting SVD on vectorized TF-IDF features. Only the training step of ensemble model used all of CPU cores, the others only used one core at a time. GPU was only used for extracting document embeddings from PhoBERT model. Table 4 summarizes approximate time of some time-consuming steps of the proposed method on our training set. We use Area Under the Curve (AUC) score as our evaluation metric and a 5-folds cross-validation scheme to evaluate our models. Though lots of experiments were made, we only shows the main versions that improve the performance significantly. All versions before ensemble were trained with a tuned CatBoost classifier. Comparison to top teams in the competition are shown in table 5. Our experiments were conducted as follow:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 412,
                        "end": 419,
                        "text": "Table 4",
                        "ref_id": "TABREF6"
                    }
                ],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4.3"
            },
            {
                "text": "\u2022 Version 1: no embedding, no combination features (described in section 3.2.1).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4.3"
            },
            {
                "text": "\u2022 Version 2: add PhoBERT embedding.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4.3"
            },
            {
                "text": "\u2022 Version 3: add ensemble learning manner.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4.3"
            },
            {
                "text": "\u2022 Version 4: add combination features",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4.3"
            },
            {
                "text": "\u2022 Final version: leverage unlabeled data.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4.3"
            },
            {
                "text": "We list out some remarkable insights that we discovered in this task:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion 5.1 Summary",
                "sec_num": "5"
            },
            {
                "text": "\u2022 Combining high-importance features is a good way of feature generation",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion 5.1 Summary",
                "sec_num": "5"
            },
            {
                "text": "\u2022 TF-IDF should be applied on raw contents to capture their original form, while document embedding should be applied on cleaned ones to obtain contextual features. Table 5 : AUC scores of proposed method and other teams on different datasets.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 165,
                        "end": 172,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Conclusion 5.1 Summary",
                "sec_num": "5"
            },
            {
                "text": "\u2022 The more the content the model learnt, the better the performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion 5.1 Summary",
                "sec_num": "5"
            },
            {
                "text": "\u2022 Stacking with complementary bagging is very powerful.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion 5.1 Summary",
                "sec_num": "5"
            },
            {
                "text": "Due to the time limit, a lot of methods we tried still need more validation and tuning, therefore were left out of the final submission. Other information, such as post images, can also give a boost in performance, due to the content is embedded in the images, or special information such as watermarks. Other Natural Language Processing features like sentiment of the comments, Part Of Speech tagging, bias, although tried, but haven't tuned carefully to produce good result, could be helpful. We also believe the URL, if provided, could also help improve the performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Future work",
                "sec_num": "5.2"
            },
            {
                "text": "A framework using stacked generalization to combine results of different models https://github.com/ kaz-Anova/StackNet.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Fake news detection using machine learning ensemble methods",
                "authors": [
                    {
                        "first": "Iftikhar",
                        "middle": [],
                        "last": "Ahmad",
                        "suffix": ""
                    },
                    {
                        "first": "Muhammad",
                        "middle": [],
                        "last": "Yousaf",
                        "suffix": ""
                    },
                    {
                        "first": "Suhail",
                        "middle": [],
                        "last": "Yousaf",
                        "suffix": ""
                    },
                    {
                        "first": "Muhammad Ovais",
                        "middle": [],
                        "last": "Ahmad",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Complexity",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Iftikhar Ahmad, Muhammad Yousaf, Suhail Yousaf, and Muhammad Ovais Ahmad. 2020. Fake news detection using machine learning ensemble methods. Complexity, 2020.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "On the benefit of combining neural, statistical and external features for fake news identification",
                "authors": [
                    {
                        "first": "Gaurav",
                        "middle": [],
                        "last": "Bhatt",
                        "suffix": ""
                    },
                    {
                        "first": "Aman",
                        "middle": [],
                        "last": "Sharma",
                        "suffix": ""
                    },
                    {
                        "first": "Shivam",
                        "middle": [],
                        "last": "Sharma",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1712.03935"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Gaurav Bhatt, Aman Sharma, Shivam Sharma, Ankush Nagpal, Balasubramanian Raman, and Ankush Mit- tal. 2017. On the benefit of combining neural, sta- tistical and external features for fake news identifica- tion. arXiv preprint arXiv:1712.03935.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "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": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1810.04805"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understand- ing. arXiv preprint arXiv:1810.04805.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Tweetcred: Realtime credibility assessment of content on twitter",
                "authors": [
                    {
                        "first": "Aditi",
                        "middle": [],
                        "last": "Gupta",
                        "suffix": ""
                    },
                    {
                        "first": "Ponnurangam",
                        "middle": [],
                        "last": "Kumaraguru",
                        "suffix": ""
                    },
                    {
                        "first": "Carlos",
                        "middle": [],
                        "last": "Castillo",
                        "suffix": ""
                    },
                    {
                        "first": "Patrick",
                        "middle": [],
                        "last": "Meier",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "International Conference on Social Informatics",
                "volume": "",
                "issue": "",
                "pages": "228--243",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Aditi Gupta, Ponnurangam Kumaraguru, Carlos Castillo, and Patrick Meier. 2014. Tweetcred: Real- time credibility assessment of content on twitter. In International Conference on Social Informatics, pages 228-243. Springer.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Fake news detection using an ensemble learning model based on self-adaptive harmony search algorithms",
                "authors": [
                    {
                        "first": "Yin-Fu",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Po-Hong",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Expert Systems with Applications",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yin-Fu Huang and Po-Hong Chen. 2020. Fake news detection using an ensemble learning model based on self-adaptive harmony search algorithms. Expert Systems with Applications, page 113584.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Reintel: A multimodal data challenge for responsible information identification on social network sites",
                "authors": [
                    {
                        "first": "Duc-Trong",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Xuan-Son",
                        "middle": [],
                        "last": "Vu",
                        "suffix": ""
                    },
                    {
                        "first": "Nhu-Dung",
                        "middle": [],
                        "last": "To",
                        "suffix": ""
                    },
                    {
                        "first": "Huu-Quang",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "Thuy-Trinh",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "Linh",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Anh-Tuan",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "Minh-Duc",
                        "middle": [],
                        "last": "Hoang",
                        "suffix": ""
                    },
                    {
                        "first": "Nghia",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Huyen",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "Hoang",
                        "middle": [
                            "D"
                        ],
                        "last": "Nguyen",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Duc-Trong Le, Xuan-Son Vu, Nhu-Dung To, Huu- Quang Nguyen, Thuy-Trinh Nguyen, Linh Le, Anh- Tuan Nguyen, Minh-Duc Hoang, Nghia Le, Huyen Nguyen, and Hoang D. Nguyen. 2020. Reintel: A multimodal data challenge for responsible informa- tion identification on social network sites.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "PhoBERT: Pre-trained language models for Vietnamese",
                "authors": [
                    {
                        "first": "Anh",
                        "middle": [
                            "Tuan"
                        ],
                        "last": "Dat Quoc Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Findings of the Association for Computational Linguistics: EMNLP 2020",
                "volume": "",
                "issue": "",
                "pages": "1037--1042",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dat Quoc Nguyen and Anh Tuan Nguyen. 2020. PhoBERT: Pre-trained language models for Viet- namese. In Findings of the Association for Computa- tional Linguistics: EMNLP 2020, pages 1037-1042.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Csi: A hybrid deep model for fake news detection",
                "authors": [
                    {
                        "first": "Natali",
                        "middle": [],
                        "last": "Ruchansky",
                        "suffix": ""
                    },
                    {
                        "first": "Sungyong",
                        "middle": [],
                        "last": "Seo",
                        "suffix": ""
                    },
                    {
                        "first": "Yan",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 2017 ACM on Conference on Information and Knowledge Management",
                "volume": "",
                "issue": "",
                "pages": "797--806",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Natali Ruchansky, Sungyong Seo, and Yan Liu. 2017. Csi: A hybrid deep model for fake news detection. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 797-806.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Fake news detection on social media: A data mining perspective",
                "authors": [
                    {
                        "first": "Kai",
                        "middle": [],
                        "last": "Shu",
                        "suffix": ""
                    },
                    {
                        "first": "Amy",
                        "middle": [],
                        "last": "Sliva",
                        "suffix": ""
                    },
                    {
                        "first": "Suhang",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Jiliang",
                        "middle": [],
                        "last": "Tang",
                        "suffix": ""
                    },
                    {
                        "first": "Huan",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "ACM SIGKDD explorations newsletter",
                "volume": "19",
                "issue": "1",
                "pages": "22--36",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kai Shu, Amy Sliva, Suhang Wang, Jiliang Tang, and Huan Liu. 2017. Fake news detection on social me- dia: A data mining perspective. ACM SIGKDD ex- plorations newsletter, 19(1):22-36.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Fake news detection as natural language inference",
                "authors": [
                    {
                        "first": "Kai-Chou",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Timothy",
                        "middle": [],
                        "last": "Niven",
                        "suffix": ""
                    },
                    {
                        "first": "Hung-Yu",
                        "middle": [],
                        "last": "Kao",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1907.07347"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kai-Chou Yang, Timothy Niven, and Hung-Yu Kao. 2019. Fake news detection as natural language in- ference. arXiv preprint arXiv:1907.07347.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF1": {
                "html": null,
                "type_str": "table",
                "text": "Datasets.",
                "content": "<table><tr><td># rows</td><td>5172</td></tr><tr><td># label 1</td><td>934</td></tr><tr><td># user_name</td><td>3706</td></tr><tr><td># unique post_message</td><td>4868</td></tr><tr><td>latest timestamp_post</td><td>Jan 2, 2014</td></tr><tr><td colspan=\"2\">nearest timestamp_post Sep 28, 2020</td></tr></table>",
                "num": null
            },
            "TABREF2": {
                "html": null,
                "type_str": "table",
                "text": "Statistic summarization of our training set.",
                "content": "<table/>",
                "num": null
            },
            "TABREF4": {
                "html": null,
                "type_str": "table",
                "text": "Model hyper-parameters.",
                "content": "<table/>",
                "num": null
            },
            "TABREF5": {
                "html": null,
                "type_str": "table",
                "text": "",
                "content": "<table><tr><td>shows the tuned hyper-parameters we used</td></tr><tr><td>for each model described in Section 3.3. All classi-</td></tr><tr><td>fiers except Naive Bayes used our predefined class</td></tr><tr><td>weights of 0.15 for class 0 and 0.75 for class 1.</td></tr></table>",
                "num": null
            },
            "TABREF6": {
                "html": null,
                "type_str": "table",
                "text": "Approx. run time of proposed method.",
                "content": "<table/>",
                "num": null
            }
        }
    }
}