File size: 87,222 Bytes
6fa4bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
{
    "paper_id": "2021",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T03:35:47.352492Z"
    },
    "title": "Using contextual and cross-lingual word embeddings to improve variety in template-based NLG for automated journalism",
    "authors": [
        {
            "first": "Miia",
            "middle": [],
            "last": "R\u00e4m\u00f6",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Helsinki",
                "location": {}
            },
            "email": "miia.ramo@helsinki.fi"
        },
        {
            "first": "Leo",
            "middle": [],
            "last": "Lepp\u00e4nen",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Helsinki",
                "location": {}
            },
            "email": "leo.leppanen@helsinki.fi"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "In this work, we describe our efforts in improving the variety of language generated from a rule-based NLG system for automated journalism. We present two approaches: one based on inserting completely new words into sentences generated from templates, and another based on replacing words with synonyms. Our initial results from a human evaluation conducted in English indicate that these approaches successfully improve the variety of the language without significantly modifying sentence meaning. We also present variations of the methods applicable to low-resource languages, simulated here using Finnish, where cross-lingual aligned embeddings are harnessed to make use of linguistic resources in a high-resource language. A human evaluation indicates that while proposed methods show potential in the low-resource case, additional work is needed to improve their performance.",
    "pdf_parse": {
        "paper_id": "2021",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "In this work, we describe our efforts in improving the variety of language generated from a rule-based NLG system for automated journalism. We present two approaches: one based on inserting completely new words into sentences generated from templates, and another based on replacing words with synonyms. Our initial results from a human evaluation conducted in English indicate that these approaches successfully improve the variety of the language without significantly modifying sentence meaning. We also present variations of the methods applicable to low-resource languages, simulated here using Finnish, where cross-lingual aligned embeddings are harnessed to make use of linguistic resources in a high-resource language. A human evaluation indicates that while proposed methods show potential in the low-resource case, additional work is needed to improve their performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "The use of automation to help journalists in news production is of great interest to many newsrooms across the world (Fanta, 2017; Sir\u00e9n-Heikel et al., 2019) . Natural Language Generation (NLG) methods have previously been employed, for example, to produce soccer reports (Chen and Mooney, 2008) , financial reports (Plachouras et al., 2016) and weather forecasts (Goldberg et al., 1994) . Such 'automated journalism' (Carlson, 2015; Graefe, 2016) or 'news automation' (Sir\u00e9n-Heikel et al., 2019 ) imposes restrictions on system aspects such as transparency, accuracy, modifiability, transferability and output's fluency (Lepp\u00e4nen et al., 2017) . Likely as a consequence of these requirements, news industry applications of NLG have traditionally employed the 'classical' rule-based approaches to NLG, rather than the more recent neural methods increasingly seen in recent academic literature (Sir\u00e9n-Heikel et al., 2019) . A major downside of these rule-based systems, however, is that their output often lacks variety. Adding variety by increasing the amount of templates is possible, but this would significantly increase the cost of system creation and limits reuse potential. As users of automated journalism already find the difficulty of reuse limiting (Linden, 2017) , this is not a sustainable solution.",
                "cite_spans": [
                    {
                        "start": 117,
                        "end": 130,
                        "text": "(Fanta, 2017;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 131,
                        "end": 157,
                        "text": "Sir\u00e9n-Heikel et al., 2019)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 272,
                        "end": 295,
                        "text": "(Chen and Mooney, 2008)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 316,
                        "end": 341,
                        "text": "(Plachouras et al., 2016)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 364,
                        "end": 387,
                        "text": "(Goldberg et al., 1994)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 418,
                        "end": 433,
                        "text": "(Carlson, 2015;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 434,
                        "end": 447,
                        "text": "Graefe, 2016)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 469,
                        "end": 495,
                        "text": "(Sir\u00e9n-Heikel et al., 2019",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 621,
                        "end": 644,
                        "text": "(Lepp\u00e4nen et al., 2017)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 893,
                        "end": 920,
                        "text": "(Sir\u00e9n-Heikel et al., 2019)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 1259,
                        "end": 1273,
                        "text": "(Linden, 2017)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we extend a modular rule-based NLG system -used for automated journalism in the domain of statistical news -with a dedicated component for varying the produced language in a controlled manner. The proposed extension enables two methods of inducing further variation: in insertion, new words are introduced into the generated text, whereas in replacement certain words in the original sentence are replaced with synonyms. To accomplish these tasks, we employ a combination of traditional language resources (e.g. synonym dictionaries) as well as recent neural processing models (i.e. word embeddings). These resources complement each other, enabling us to harness the power of statistical NLP tools while retaining control via the classical linguistic resources. We also experiment with using these methods in the context of a low-resource language which lacks linguistic resources such as synonym dictionaries. For this case, we propose to use cross-lingual aligned word embeddings to utilize a high-resource language's resources even within said low-resource language.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In the next section, we briefly describe some related previous works and further motivate our approach. Section 3 describes our proposed variation induction methods for both the high-resource and the low-resource contexts. Sections 4 and 5, respectively, introduce our human evaluation method and the results obtained. Sections 6 provides some additional thoughts on these results, while Section 7 concludes the paper.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Natural language generation has been associated with news production from the early years of the field, with some of the earliest industry applications of the NLG methods being in the domain of weather report production (Goldberg et al., 1994) . Interest in applying NLG to news production has only increased since, with many media houses experimenting with the technology (Fanta, 2017; Sir\u00e9n-Heikel et al., 2019) . Still, adoption of automated journalism methods has been slow. According to news media insiders, rule-based, classical, NLG system such as those described by Reiter and Dale (2000) , are costly to create and difficult to reuse (Linden, 2017) . At the same time, even the most recent neural (end-to-end) approaches to NLG are not fit for customer needs as they limit the ability to \"customise, configure, and control the content and terminology\" (Reiter, 2019) . Another major problem is the fact they suffer from a form of overfitting known as 'hallucination', where ungrounded output text is produced. This is catastrophic in automated journalism.",
                "cite_spans": [
                    {
                        "start": 220,
                        "end": 243,
                        "text": "(Goldberg et al., 1994)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 373,
                        "end": 386,
                        "text": "(Fanta, 2017;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 387,
                        "end": 413,
                        "text": "Sir\u00e9n-Heikel et al., 2019)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 574,
                        "end": 596,
                        "text": "Reiter and Dale (2000)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 643,
                        "end": 657,
                        "text": "(Linden, 2017)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 861,
                        "end": 875,
                        "text": "(Reiter, 2019)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background",
                "sec_num": "2"
            },
            {
                "text": "Concurrently with works on improved neural NLG methods, others have investigated increasingly modular rule-based approaches with the intent of addressing the reusability problem described by Linden (2017) . For example, Lepp\u00e4nen et al. (2017) describe a modular rule-based system for automated journalism that seeks to separate text domain specific processing from language specific processing to allow for easier transfer of the system to new text domains. While such rule-based approaches produce output that is grammatically and factually correct (Gatt and Krahmer, 2017) , they often suffer from a lack of variety in language. This is especially true for systems that are based on some type of templates, or fragmentary language resources that are combined to form larger segments of text and into which content dependent on system input is embedded. Using such templates (or hand-crafted grammars) is costly, especially when a large number is required for varied output.",
                "cite_spans": [
                    {
                        "start": 191,
                        "end": 204,
                        "text": "Linden (2017)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 220,
                        "end": 242,
                        "text": "Lepp\u00e4nen et al. (2017)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 550,
                        "end": 574,
                        "text": "(Gatt and Krahmer, 2017)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background",
                "sec_num": "2"
            },
            {
                "text": "As template (or grammar) production can be costly, automated variation induction methods that could be integrated into rule-based systems are very interesting. One trivial approach to inducing variation would be to employ a synonym dictionary, such as is available in WordNet (Miller, 1995) , to replace words within the generated text with their synonyms. This approach, however, suffers from some major problems. First, simply looking up all synonyms for all meanings of a token is not feasible due to polysemy and homonymy. At the same time, incorporating knowledge of which semantic meaning of a token is correct in each case significantly slows down template and grammar generation. Furthermore, even within a certain semantic meaning, the various (near) synonyms might not be equally suitable for a given context. Finally, such linguistic resources are not available for many low-resource languages.",
                "cite_spans": [
                    {
                        "start": 276,
                        "end": 290,
                        "text": "(Miller, 1995)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background",
                "sec_num": "2"
            },
            {
                "text": "An alternative approach, more suited to generation within medium and low-resource languages where there are no available synonym dictionaries, but large text corpora can be collected, would be to use word embeddings (E.g. Rumelhart et al., 1986; Bengio et al., 2003; Mikolov et al., 2013) to identify words that are semantically close to the words in the template. This approach, however, suffers from the fact that both synonyms and antonyms of a word reside close to it in the word embedding space. While potential solutions have been proposed (E.g. Nguyen et al., 2016), they are not foolproof.",
                "cite_spans": [
                    {
                        "start": 222,
                        "end": 245,
                        "text": "Rumelhart et al., 1986;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 246,
                        "end": 266,
                        "text": "Bengio et al., 2003;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 267,
                        "end": 288,
                        "text": "Mikolov et al., 2013)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background",
                "sec_num": "2"
            },
            {
                "text": "As described above, na\u00efve methods based on either classical linguistic resources or word embeddings alone are not suitable for variation induction. To this end, we are interested in identifying a simple variety induction method that combines the positive sides of both the classical linguistic resources (such as synonym dictionaries) with those of statistical resources such as word embeddings. Optimally, the method should also function for a wide variety of languages, including low-resource languages where costly resources such as comprehensive synonym dictionaries are not readily available.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Variety Induction Algorithms",
                "sec_num": "3"
            },
            {
                "text": "In this work, we introduce variety into the generated language using two distinct methods: by introducing completely new words into sentences, and by replacing existing words. We will use the terms insertion and replacement to distinguish between the two approaches, respectively.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Variety Induction Algorithms",
                "sec_num": "3"
            },
            {
                "text": "In our insertion method, new words are introduced to sentences at locations where placeholder tokens are defined in templates. We use a combination of a part-of-speech (POS) tagger and a contextual language model to control the process. A simplified Algorithm 1 Pseudocode describing the insertion approach. The parameters are a single sentence, a desired POS tag, some value of k, and finally min and max number of [MASK] tokens inserted. The approach is tailored for high-resource languages, such as English, and uses additional linguistic resources (here, a part of speech tagger) to conduct further filtering.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "function HIGHRESOURCEINSERTION(Sentence, P oS, k, minM asked, maxM asked)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "W ordsAndScores \u2190 \u2205 for n \u2208 [minM asked, maxM asked] do M askedSentence \u2190 Sentence with n [MASK] tokens inserted W ords, Scores \u2190 MASKEDLM.TOPKPREDICTIONS(M askedSentence, k) W ordsAndScores \u2190 W ordAndScores \u222a {(w, s)|w \u2208 W ords and s \u2208 Scores} end for return SAMPLE({w|(w, s) \u2208 W ordsAndScores, POSTAG(w) = P oS, s >= T hreshold}) end function",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "Step 1: In Austria in 2018 75 year old or older females {empty, pos=RB} received median equivalised net income of 22234 C.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "Step 2: In Austria in 2018 75 year old or older females still received median equivalised net income of 22234 C. Step 1 represents the intermediate step between a template and the final modified sentence presented in Step 2.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "example of the general idea is shown in Figure 1 . During variety induction, a contextual language model with a masked language modeling head (In this case, FinEstBert by Ul\u010dar and Robnik-Sikonja, 2020) is used to predict suitable content to replace the placeholder token. This is achieved by replacing the placeholder token with one or more [MASK] tokens in the sentence. Multiple [MASK] tokens are required where the language model uses subword tokens. The language model is then queried for the k most likely (subword) token sequences to replace the sequence of [MASK] tokens. This results in a selection of potential tokens ('proposals', each consisting of one or more subword tokens) to replace the original placeholder.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 40,
                        "end": 48,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "As an additional method for control, we associate the original placeholder token with a certain POS tag, and filter the generated proposals to those matching this POS tag. In addition, we use a threshold likelihood value so that each proposal has to reach a minimal language model score. This is re-quired for cases wherein a certain length sequence of mask tokens results in no believable proposals in the top-k selection. Finally, we sample one of the filtered proposals and replace the original placeholder token with it. In cases where there are no suitable proposals, the placeholder value is simply removed. This method is described in pseudocode in Algorithm 1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "Naturally, this approach is dependent on the availability of two linguistic resources: the contextual word embeddings and a POS tagging model. While word embeddings/language models are relatively easily trainable as long as there are any available text corpora, high-quality POS tagging models are less common outside of the most widely spoken languages. To extend this approach to such low-resource languages that have available corpora for training language models such as BERT, but lack POS tagging models, cross-lingual aligned word embeddings can be utilized.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "Once a low-resource language proposal has been obtained using the method described above, an aligned cross-lingual word embeddings model -in our case, FastText embeddings (Bojanowski et al., 2016) aligned using VecMap (Artetxe et al., 2018 )between the low-resource language and some highresource language (e.g. English) can be used to obtain the closest high-resource language token in the aligned embedding space. The retrieved highreasource language token is, in theory, the closest semantic high-resource language equivalent to the low-resource token. We then apply a POS tagging model for the high-resource language to the highresource 'translation', and use that POS tag as the low-resource token's POS tag for the purposes of filtering the proposals. This approach is described as pseudocode in Algorithm 2.",
                "cite_spans": [
                    {
                        "start": 171,
                        "end": 196,
                        "text": "(Bojanowski et al., 2016)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 218,
                        "end": 239,
                        "text": "(Artetxe et al., 2018",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "Algorithm 2 Pseudocode describing how the language resources, here a POS tagger, are utilized for a low-resource language with cross-lingual word embeddings. In other words, when working with a low-resource language, insertion is done as in Algorithm 1, but the POS tagging phase utilises this algorithm. The FINDVECTOR method finds the word embedding vector for the low resource word, and the CLOSESTWORD method is then used for finding the closest match for that vector from the aligned high-resource language embedding space. The algorithm parameters are the low-resource original word to be replaced, and the pairwise aligned low-and high-resource word embeddings. function POSTAGLOWRESOURCELANGUAGE(LowResW ord, LowResEmbeddings, HighResEmbeddings)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "LowResV ector \u2190 FINDVECTOR(LowResW ord, LowResEmbeddings) HighResW ord \u2190 CLOSESTWORD(LowResV ector, HighResEmbeddings) LowResT agged \u2190 (LowResW ord, POSTAG(HighResW ord)) return LowResT agged end function",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "Step 1: In Finland in 2016 households' total {expenditure, replace=True} on healthcare was 20.35 %.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "Step 2: In Finland in 2016 households' total spending on healthcare was 20.35 %. Step 1 represents the intermediate step between a template and the final modified sentence presented in Step 2.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introducing Variety with Insertion",
                "sec_num": "3.1"
            },
            {
                "text": "In addition to insertion of completely new words, variety can also be induced by replacing existing content, so that previously lexicalized words within the text are replaced by suitable alternatives. We propose to use a combination of a synonym dictionary and a contextual language model to do this in a controlled fashion. A simplified example of this approach is shown in Figure 2 . On a high level, we mark certain words within the template fragments used by our system as potential candidates for replacement. This provides us with further control, allowing us to limit the variety induction to relatively 'safe' words such as those not referring to values in the underlying data.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 375,
                        "end": 383,
                        "text": "Figure 2",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Inducing Variety with Replacement",
                "sec_num": "3.2"
            },
            {
                "text": "During variation induction, the synonym dictionary is first queried for synonyms of the marked word. To account for homonymy, polynymy, as well as the contextual fit of the proposed synonyms, we then use the contextual word embeddings (with a masked language model head) to score the proposed words. To score the word, it needs to be tokenized. In cases where the word is not part of BERT's fixed size vocabulary, it is tokenized as multiple subword tokens. To account for this we use the mean score of the (subword) tokens as the score of the complete word.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Inducing Variety with Replacement",
                "sec_num": "3.2"
            },
            {
                "text": "As above, a threshold is used to ensure that only candidates that are sufficiently good fits are retained in the pool of proposed replacements. The final word is sampled from the filtered pool of proposals. If the pool of proposed words is empty after filtering, the sentence is not modified. The original word is also explicitly retained in the proposals. This procedure is shown in Algorithm 3.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Inducing Variety with Replacement",
                "sec_num": "3.2"
            },
            {
                "text": "We emphasize that the use of the synonym dictionary is required to avoid predicting antonyms, as both antonyms and synonyms reside close to the original word in the word embedding space. While an antonym such as 'increase' for the verb 'decrease' would be a good replacement in terms of language modeling score, such antonymous replacement would change the sentence meaning tremendously and must be prevented.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Inducing Variety with Replacement",
                "sec_num": "3.2"
            },
            {
                "text": "The modification of the replacement approach for low-resource languages (where no synonym dictionary is available) is similar to that presented above for insertion: We conduct a round-trip via a high-resource language using the cross-lingual embeddings when retrieving synonyms. The lowresource language words are 'translated' to the high-resource language using the cross-lingual embeddings, after which synonyms for these translations are retrieved from the synonym dictionary available in the high-resource language. The synonyms are then 'translated' back to the lowresource language using the same cross-lingual embeddings. This approach is shown in Algorithm 4.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Inducing Variety with Replacement",
                "sec_num": "3.2"
            },
            {
                "text": "Algorithm 3 Pseudocode describing a method for replacement using a combination of a masked language model (based on contextual word embeddings) and a synonym dictionary, such as provided by WordNet. The parameters are the original word marked to be replaced in the input sentence ('expenditure' in Figure 2 ), and the input sentence for context. function HIGHRESOURCEREPLACEMENT(OriginalW ord, Sentence) W ordsAndScores \u2190 \u2205 Synonyms \u2190 GETSYNONYMS(OriginalW ord) for w \u2208 Synonyms do CandidateSentence \u2190 Sentence with w replacing the original word CandidateScore \u2190 MASKEDLM.SCORE(CandidateSentence, w) W ordsAndScores \u2190 W ordsAndScores \u222a (w, CandidateScore) end for return SAMPLE({w|(w, s) \u2208 W ordsAndScores, s >= T hreshold}) end function Algorithm 4 Pseudocode describing how synonyms are retrieved for a low-resource language by utilizing cross-lingual word embeddings. Low-resource variant of replacement is as Algorithm 3, but this algorithm is used to retrieve synonyms. The FINDVECTOR method finds the correct word embedding vector for the low resource word, and the CLOSESTWORD method is then used for finding the closest match for that vector from the aligned high-resource language embedding space. The algorithm parameters are the low-resource original word to be replaced, and the pairwise aligned low-and high-resource word embeddings.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 298,
                        "end": 306,
                        "text": "Figure 2",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Inducing Variety with Replacement",
                "sec_num": "3.2"
            },
            {
                "text": "function As we conduct our case study using Finnish as the (simulated) low-resource language, words need to be lemmatized before synonym lookup. We apply UralicNLP (H\u00e4m\u00e4l\u00e4inen, 2019) to analyze and lemmatize the original word and reinflect the retrieved synonyms after lookup. A difficulty is presented by the fact that oftentimes, a specific token can have multiple plausible grammatical analyses and lemmas. In our approach, synonyms are retrieved for all of the plausible lemmas, and the algorithm regenerates all morphologies proposed by UralicNLP for all synonyms. While this results in some ungrammatical or contextually incorrect tokens, we rely on the language model to score these as unlikely.",
                "cite_spans": [
                    {
                        "start": 164,
                        "end": 182,
                        "text": "(H\u00e4m\u00e4l\u00e4inen, 2019)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Inducing Variety with Replacement",
                "sec_num": "3.2"
            },
            {
                "text": "We have implemented the above algorithms within a multi-lingual (Finnish and English) natural language generation system that conducts automated journalism from time-series data provided by Eurostat (the statistical office of the European Union). The system is derived from the template-based modular architecture presented by Lepp\u00e4nen et al. (2017) . It produces text describing the most salient factors of the input data in several languages in a technically accurate manner using only a few templates, but the resulting language is very stiff, and the sentences are very alike. This makes the final report very repetitive and thus a good candidate for variety induction.",
                "cite_spans": [
                    {
                        "start": 327,
                        "end": 349,
                        "text": "Lepp\u00e4nen et al. (2017)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "For all of the algorithms described, we utilise the same trilingual BERT model: FinEst BERT (Ul\u010dar and Robnik-\u0160ikonja, 2020) . The FinEst BERT model is trained with monolingual corpora for English, Finnish and Estonian from a mixture of news articles and a general web crawl. In addition to the BERT model, the low-resource language variants of the algorithms utilize crosslingual pairwise aligned word embeddings for word 'translations'. We use monolingual FastText (Bojanowski et al., 2016) word embeddings mapped with VecMap (Artetxe et al., 2018) to form the cross-lingual embeddings. POS tagging is done with NLTK (Bird et al., 2009) and the lexical database used as a synonym dictionary is Word-Net (Miller, 1995) .",
                "cite_spans": [
                    {
                        "start": 92,
                        "end": 124,
                        "text": "(Ul\u010dar and Robnik-\u0160ikonja, 2020)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 467,
                        "end": 492,
                        "text": "(Bojanowski et al., 2016)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 528,
                        "end": 550,
                        "text": "(Artetxe et al., 2018)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 619,
                        "end": 638,
                        "text": "(Bird et al., 2009)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 696,
                        "end": 719,
                        "text": "Word-Net (Miller, 1995)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "A human evaluation of our methods was conducted following the best practices proposed by van der Lee et al. (2019) . In the evaluation setting, judges were first presented with three statements about a sentence pair. Sentence 1 of the pair was an original sentence, generated by the NLG system without variation induction. Sentence 2 of the pair was the same sentence with a variation induction procedure applied. Cases where the sentence would remain unchanged, or where no insertion/replacement candidates were identified, were ruled out from the evaluation set. The part of the sentence to be modified was marked in the original sentence and the inserted/replaced word highlighted.",
                "cite_spans": [
                    {
                        "start": 97,
                        "end": 114,
                        "text": "Lee et al. (2019)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "The judges were asked to evaluate the following statements on a Likert scale ranging from 1 ('Strongly Disagree') to 4 ('Neither Agree nor Disagree') to 7 ('Strongly Agree'):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "Q1: Sentence 1 is a good quality sentence in the target language.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "Q2: Sentence 2 is a good quality sentence in the target language.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "Q3: Sentences 1 and 2 have essentially the same meaning.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "In addition to the two sentences, the judges were presented with two groups of words to examine if using the scores by BERT would correctly distinguish suitable words from unsuitable words. Group 1 contained the words scored as acceptable by BERT while group 2 contained the words ruled out due to a low score. All words in both groups met the criteria of being synonyms (in the case of replacement) or being the correct POS (in the case of insertion). The judges were asked to evaluate the following questions on a 5-point Likert scale ranging from 1 ('None of the words') to 3 ('Half of the words') to 5 ('All of the words'):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "Q4: How many of the words in word group 1 could be used in the marked place in sentence 1 so that the meaning remains essentially the same?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "Q5: How many of the words in word group 2 could be used in the marked place in sentence 1 so that the meaning remains essentially the same?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "For the high-resource language results, we gathered 3 judgements each for 100 sentence pairs. The judges were recruited from an online crowdsourcing platform and they received a monetary reward for participating in the study. The judge recruitment was restricted to countries where majority of people are native speakers of English. For the low-resource language results, 21 judges evaluated 20 sentence pairs. The judges were recruited via student mailing lists of University of Helsinki in Finland and were not compensated monetarily. All but one of the participants in the low-resource evaluation were native speakers of the target language. The final participant self-identified as having a 'working proficiency.'",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "4"
            },
            {
                "text": "Table 1 presents our results in applying both the insertion and replacement methods to both a highresource language (English) and a low-resource language (Finnish).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "In the high-resource insertion case, the results indicate that inducing variation using the proposed method does not decrease output quality, as both the original sentences' qualities (Q1 mean 5.57) and modified sentences' qualities (Q2 mean 5.76) were similar. As the sentence meaning also remained largely unchanged (Q3 mean 5.54), we interpret this result as a success. The results for Q4 and Q5 indicate that our filtering method based on a threshold language model score can be improved: results for Q4 (mean 3.11 on a 5-point Likert scale) indicate that unsuitable words are left unfiltered, while Q5 (mean 3.03) indicates that some acceptable words are filtered out. (1-5 \u2193) 'How many of the words in word group 2 could be used in the marked place in sentence 1 so that the meaning remains essentially the same?'",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "3.03 (1.41) 1.46 (0.62) 3.21 (1.27) 1.62 (0.76) Table 1 : Evaluation results for the insertion and replacement approaches. English ('En') examples were generated using the high-resource variations, while the Finnish ('Fi') examples were generated using the low-resource variations. Arrows in the range column indicate whether higher (\u2191) or lower (\u2193) values indicate better performance.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 48,
                        "end": 55,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "Values are the mean evaluation result and the standard deviation (in parentheses). In the context of the statements, sentence 1 is the original, unmodified sentence, while sentence 2 is a sentence with added variety.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "In the low-resource case insertion, we observe some change in meaning (Q3 mean value 4.34) and a slight loss of quality, but even after variety induction the output quality is acceptable (Q1 mean 6.43 vs. Q2 mean 5.12). Interestingly, in the low-resource setting, we observe that the language model is slightly better at distinguishing between suitable and unsuitable candidates (Q4 and Q5 means 2.53 and 1.46, respectively) than in the high-resource case. We are, at this point, uncertain of the reason behind the difference in the ratios of Q4 and Q5 answers between the high-resource and the low-resource case. Notably, even this 'better' result is far from perfect.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "We also conducted POS tag specific analyses for both the high-resource and the low-resource insertion cases. In the high-resource case, no major differences were observed between various POS tags. In the low-resource (Finnish) case, however, we observed that with some POS tags, such as adverbs, the results are similar to those observed with English. Low-resource results for adverbs only are shown in Figure 3 . We emphasize that this is the best observed subresult and should be viewed as post-hoc analysis.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 403,
                        "end": 411,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "In the high-resource replacement case, we observe promising results. Inducing variation did not negatively affect sentence quality (Q1 mean 5.55 vs. Q2 mean 5.60) and concurrently retained meaning (Q3 mean 5.65). Results for Q4 and Q5 (means 3.39 and 3.21, respectively) indicate that, as above, the filtering method still has room for improvement, with poor quality options passing the filter and high-quality options being filtered out.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "However, in the low-resource case replacement case, we observe a significant drop in sentence quality after variation induction (Q1 mean 6.67 vs Q2 mean 3.89), as well as significant change in sentence meaning (Q3 mean 3.39). While Q5 results are relatively good (mean 1.62), as in very few if any good candidate words are filtered out, Q4 results (mean 1.76) indicate some fundamental problem in the candidate generation process: as there are few if any good candidates in either group, it seems that most of the proposed words are unsuitable. Original sentence is good Finnish strongly disagree quite strongly disagree somewhat disagree neither agree nor disagree somewhat agree quite strongly agree strongly agree Figure 3 : Quality of sentences with low-resource insertion in Finnish with English as the high-resource language, and preservation of sentence meaning. Results shown for adverbs only, representing the best observed performance across the various parts of speech generated. We emphasize that the graph shows only a subset of the complete results (See Table 1 ), identified as best-performing during post-hoc analysis.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 717,
                        "end": 725,
                        "text": "Figure 3",
                        "ref_id": null
                    },
                    {
                        "start": 1068,
                        "end": 1075,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5"
            },
            {
                "text": "Our high-resource results indicate that the proposed approach is suitable for inducing some light variation into automatically generated language. The use of synonym dictionaries removes the need to manually construct variants into the templates used in the generation, while the use of language models allows for contextual scoring of the proposed variants so that higher quality results are selected. We suspect that a major contributor to the low quality of the modified sentences in the lowresource scenarios was the complex morphology of the Finnish language. Especially in the case of Finnish, the process wherein the original word was grammatically analyzed and the replacement word reinflected into the same form would have likely resulted in cases where the resulting word is technically grammatically feasible in isolation, but not grammatical in the context of the rest of the sentence. Our post-hoc investigation also indicates that at least in some cases the resulting reinflected words were outright ungrammatical.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "6"
            },
            {
                "text": "In addition, it seems that the language model employed did not successfully distinguish these failure cases from plausible cases, which led to significant amounts of ungrammatical words populating the proposed set of replacement words. Our post-hoc analysis further indicates that the methods led to better results when use of compound words was avoided in the Finnish templates. We hypothesize that applying the method to a morphologically less complex language might yield significantly better results.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "6"
            },
            {
                "text": "At the same time, in the case of low-resource variation induction using insertion, our results indi-cate that some success could be found if the method is applied while restrained to certain pre-screened parts of speech, such as adverbs (See Figure 3) . This further indicates that the performance of the replacement approach might be improved significantly if the morphology issues were corrected.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 242,
                        "end": 251,
                        "text": "Figure 3)",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "6"
            },
            {
                "text": "Notably, our analysis of the results did not include an in-depth error analysis to determine what parts of the relatively complex procedure fundamentally caused the errors, i.e. were the errors introduced during POS-tagging, language model based scoring, or some other stage. Furthermore, we did not rigorously analyze whether the generation errors were semantic or grammatical in nature.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "6"
            },
            {
                "text": "As a final note, we emphasise that these results were evaluated on local (sentence) rather than on global (full news report) level. We anticipate that, for example, when inserting a word like 'still' in a sentence (see Figure 1) , the results might differ when evaluating on a global level.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 219,
                        "end": 228,
                        "text": "Figure 1)",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "6"
            },
            {
                "text": "ConclusionsIn this work, we proposed two approaches, with variations for both high-resource and low-resource languages, for increasing the variety of language in NLG system output in context of news, and presented empirical results obtained by human evaluation. The evaluation suggests that the high-resource variants of our approaches are promising: using them in the context of a case study did create variety, while preserving quality and meaning. The low-resource variants did not perform as well, but we show that there are some positive glimpses in these initial results, and suggest future improvements.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "This article is based on the Master's thesis of the first author. The work was supported by the European Union's Horizon 2020 research and innovation program under grant agreement No 825153, project EMBEDDIA (Cross-Lingual Embeddings for Less-Represented Languages in European News Media). We thank Matej Ul\u010dar and Marko Robnik-\u0160ikonja for the VecMap alignment of the FastText embeddings.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgements",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Generalizing and improving bilingual word embedding mappings with a multi-step framework of linear transformations",
                "authors": [
                    {
                        "first": "Mikel",
                        "middle": [],
                        "last": "Artetxe",
                        "suffix": ""
                    },
                    {
                        "first": "Gorka",
                        "middle": [],
                        "last": "Labaka",
                        "suffix": ""
                    },
                    {
                        "first": "Eneko",
                        "middle": [],
                        "last": "Agirre",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Thirty-Second AAAI Conference on Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2018. Generalizing and improving bilingual word embed- ding mappings with a multi-step framework of linear transformations. In Thirty-Second AAAI Conference on Artificial Intelligence.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "A neural probabilistic language model. The journal of machine learning research",
                "authors": [
                    {
                        "first": "Yoshua",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    },
                    {
                        "first": "R\u00e9jean",
                        "middle": [],
                        "last": "Ducharme",
                        "suffix": ""
                    },
                    {
                        "first": "Pascal",
                        "middle": [],
                        "last": "Vincent",
                        "suffix": ""
                    },
                    {
                        "first": "Christian",
                        "middle": [],
                        "last": "Janvin",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "",
                "volume": "3",
                "issue": "",
                "pages": "1137--1155",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yoshua Bengio, R\u00e9jean Ducharme, Pascal Vincent, and Christian Janvin. 2003. A neural probabilistic lan- guage model. The journal of machine learning re- search, 3:1137-1155.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Natural Language Processing with Python",
                "authors": [
                    {
                        "first": "Steven",
                        "middle": [],
                        "last": "Bird",
                        "suffix": ""
                    },
                    {
                        "first": "Ewan",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    },
                    {
                        "first": "Edward",
                        "middle": [],
                        "last": "Loper",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural Language Processing with Python, 1st edi- tion. O'Reilly Media, Inc.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Enriching Word Vectors with Subword Information",
                "authors": [
                    {
                        "first": "Piotr",
                        "middle": [],
                        "last": "Bojanowski",
                        "suffix": ""
                    },
                    {
                        "first": "Edouard",
                        "middle": [],
                        "last": "Grave",
                        "suffix": ""
                    },
                    {
                        "first": "Armand",
                        "middle": [],
                        "last": "Joulin",
                        "suffix": ""
                    },
                    {
                        "first": "Tomas",
                        "middle": [],
                        "last": "Mikolov",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2016. Enriching Word Vectors with Subword Information. CoRR, abs/1607.04606.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "The robotic reporter: Automated journalism and the redefinition of labor, compositional forms, and journalistic authority",
                "authors": [
                    {
                        "first": "Matt",
                        "middle": [],
                        "last": "Carlson",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Digital journalism",
                "volume": "3",
                "issue": "3",
                "pages": "416--431",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matt Carlson. 2015. The robotic reporter: Automated journalism and the redefinition of labor, composi- tional forms, and journalistic authority. Digital jour- nalism, 3(3):416-431.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Learning to sportscast: a test of grounded language acquisition",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "David",
                        "suffix": ""
                    },
                    {
                        "first": "Raymond J",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Mooney",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the 25th international conference on Machine learning",
                "volume": "",
                "issue": "",
                "pages": "128--135",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David L Chen and Raymond J Mooney. 2008. Learn- ing to sportscast: a test of grounded language acqui- sition. In Proceedings of the 25th international con- ference on Machine learning, pages 128-135.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Putting Europe's robots on the map: Automated journalism in news agencies",
                "authors": [
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Fanta",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Reuters Institute Fellowship Paper",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alexander Fanta. 2017. Putting Europe's robots on the map: Automated journalism in news agencies. Reuters Institute Fellowship Paper, 9.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Survey of the State of the Art in Natural Language Generation: Core tasks, applications and evaluation",
                "authors": [
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Gatt",
                        "suffix": ""
                    },
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Journal of Artificial Intelligence Research",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.1613/jair.5714"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Albert Gatt and Emiel Krahmer. 2017. Survey of the State of the Art in Natural Language Generation: Core tasks, applications and evaluation. Journal of Artificial Intelligence Research, 61.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Using natural-language processing to produce weather forecasts",
                "authors": [
                    {
                        "first": "Eli",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    },
                    {
                        "first": "Norbert",
                        "middle": [],
                        "last": "Driedger",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [
                            "I"
                        ],
                        "last": "Kittredge",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "IEEE Expert",
                "volume": "9",
                "issue": "2",
                "pages": "45--53",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Eli Goldberg, Norbert Driedger, and Richard I Kit- tredge. 1994. Using natural-language processing to produce weather forecasts. IEEE Expert, 9(2):45- 53.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Guide to automated journalism",
                "authors": [
                    {
                        "first": "Andreas",
                        "middle": [],
                        "last": "Graefe",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andreas Graefe. 2016. Guide to automated journalism.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "UralicNLP: An NLP library for Uralic Languages",
                "authors": [
                    {
                        "first": "Mika",
                        "middle": [],
                        "last": "H\u00e4m\u00e4l\u00e4inen",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Journal of Open Source Software",
                "volume": "4",
                "issue": "37",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.21105/joss.01345"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Mika H\u00e4m\u00e4l\u00e4inen. 2019. UralicNLP: An NLP library for Uralic Languages. Journal of Open Source Soft- ware, 4(37):1345.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Best practices for the human evaluation of automatically generated text",
                "authors": [
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Van Der Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Gatt",
                        "suffix": ""
                    },
                    {
                        "first": "Sander",
                        "middle": [],
                        "last": "Emiel Van Miltenburg",
                        "suffix": ""
                    },
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Wubben",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 12th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "355--368",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-8643"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Chris van der Lee, Albert Gatt, Emiel van Miltenburg, Sander Wubben, and Emiel Krahmer. 2019. Best practices for the human evaluation of automatically generated text. In Proceedings of the 12th Interna- tional Conference on Natural Language Generation, pages 355-368, Tokyo, Japan. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Data-Driven News Generation for Automated Journalism",
                "authors": [
                    {
                        "first": "Leo",
                        "middle": [],
                        "last": "Lepp\u00e4nen",
                        "suffix": ""
                    },
                    {
                        "first": "Myriam",
                        "middle": [],
                        "last": "Munezero",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Granroth-Wilding",
                        "suffix": ""
                    },
                    {
                        "first": "Hannu",
                        "middle": [],
                        "last": "Toivonen",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "The 10th International Natural Language Generation conference, Proceedings of the Conference",
                "volume": "",
                "issue": "",
                "pages": "188--197",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/w17-3528"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Leo Lepp\u00e4nen, Myriam Munezero, Mark Granroth- Wilding, and Hannu Toivonen. 2017. Data-Driven News Generation for Automated Journalism. In The 10th International Natural Language Gener- ation conference, Proceedings of the Conference, pages 188-197, United States. The Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Decades of automation in the newsroom: Why are there still so many jobs in journalism? Digital journalism",
                "authors": [
                    {
                        "first": "Carl-Gustav",
                        "middle": [],
                        "last": "Linden",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "5",
                "issue": "",
                "pages": "123--140",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Carl-Gustav Linden. 2017. Decades of automation in the newsroom: Why are there still so many jobs in journalism? Digital journalism, 5(2):123-140.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Efficient estimation of word representations in vector space",
                "authors": [
                    {
                        "first": "Tomas",
                        "middle": [],
                        "last": "Mikolov",
                        "suffix": ""
                    },
                    {
                        "first": "Kai",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Greg",
                        "middle": [],
                        "last": "Corrado",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Dean",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1301.3781"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jef- frey Dean. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "WordNet: a lexical database for English",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "George",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Miller",
                        "suffix": ""
                    }
                ],
                "year": 1995,
                "venue": "Communications of the ACM",
                "volume": "38",
                "issue": "11",
                "pages": "39--41",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "George A Miller. 1995. WordNet: a lexical database for English. Communications of the ACM, 38(11):39-41.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Integrating distributional lexical contrast into word embeddings for antonymsynonym distinction",
                "authors": [
                    {
                        "first": "Sabine",
                        "middle": [],
                        "last": "Kim Anh Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "Ngoc",
                        "middle": [
                            "Thang"
                        ],
                        "last": "Schulte Im Walde",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Vu",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
                "volume": "2",
                "issue": "",
                "pages": "454--459",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kim Anh Nguyen, Sabine Schulte im Walde, and Ngoc Thang Vu. 2016. Integrating distributional lexical contrast into word embeddings for antonym- synonym distinction. In Proceedings of the 54th An- nual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 454- 459.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Interacting with financial data using natural language",
                "authors": [
                    {
                        "first": "Vassilis",
                        "middle": [],
                        "last": "Plachouras",
                        "suffix": ""
                    },
                    {
                        "first": "Charese",
                        "middle": [],
                        "last": "Smiley",
                        "suffix": ""
                    },
                    {
                        "first": "Hiroko",
                        "middle": [],
                        "last": "Bretz",
                        "suffix": ""
                    },
                    {
                        "first": "Ola",
                        "middle": [],
                        "last": "Taylor",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Jochen",
                        "suffix": ""
                    },
                    {
                        "first": "Dezhao",
                        "middle": [],
                        "last": "Leidner",
                        "suffix": ""
                    },
                    {
                        "first": "Frank",
                        "middle": [],
                        "last": "Song",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Schilder",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval",
                "volume": "",
                "issue": "",
                "pages": "1121--1124",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Vassilis Plachouras, Charese Smiley, Hiroko Bretz, Ola Taylor, Jochen L Leidner, Dezhao Song, and Frank Schilder. 2016. Interacting with financial data using natural language. In Proceedings of the 39th Inter- national ACM SIGIR conference on Research and Development in Information Retrieval, pages 1121- 1124.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "ML is used more if it does not limit control",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "2020--2027",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2019. ML is used more if it does not limit control. https://ehudreiter.com/2019/ 08/15/ml-limits-control/. Accessed: 2020- 07-25.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Building natural language generation systems",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Dale",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter and Robert Dale. 2000. Building natural language generation systems. Cambridge university press.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Learning representations by backpropagating errors",
                "authors": [
                    {
                        "first": "Geoffrey",
                        "middle": [
                            "E"
                        ],
                        "last": "David E Rumelhart",
                        "suffix": ""
                    },
                    {
                        "first": "Ronald J",
                        "middle": [],
                        "last": "Hinton",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Williams",
                        "suffix": ""
                    }
                ],
                "year": 1986,
                "venue": "nature",
                "volume": "323",
                "issue": "6088",
                "pages": "533--536",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. 1986. Learning representations by back- propagating errors. nature, 323(6088):533-536.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Unboxing news automation: Exploring imagined affordances of automation in news journalism",
                "authors": [
                    {
                        "first": "Stefanie",
                        "middle": [],
                        "last": "Sir\u00e9n-Heikel",
                        "suffix": ""
                    },
                    {
                        "first": "Leo",
                        "middle": [],
                        "last": "Lepp\u00e4nen",
                        "suffix": ""
                    },
                    {
                        "first": "Carl-Gustav",
                        "middle": [],
                        "last": "Lind\u00e9n",
                        "suffix": ""
                    },
                    {
                        "first": "Asta",
                        "middle": [],
                        "last": "B\u00e4ck",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Nordic Journal of Media Studies",
                "volume": "1",
                "issue": "1",
                "pages": "47--66",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Stefanie Sir\u00e9n-Heikel, Leo Lepp\u00e4nen, Carl-Gustav Lind\u00e9n, and Asta B\u00e4ck. 2019. Unboxing news au- tomation: Exploring imagined affordances of au- tomation in news journalism. Nordic Journal of Me- dia Studies, 1(1):47-66.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "FinEst BERT and CroSloEngual BERT: less is more in multilingual models",
                "authors": [
                    {
                        "first": "Matej",
                        "middle": [],
                        "last": "Ul\u010dar",
                        "suffix": ""
                    },
                    {
                        "first": "Marko",
                        "middle": [],
                        "last": "Robnik-\u0160ikonja",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2006.07890"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Matej Ul\u010dar and Marko Robnik-\u0160ikonja. 2020. FinEst BERT and CroSloEngual BERT: less is more in mul- tilingual models. arXiv preprint arXiv:2006.07890.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "uris": null,
                "type_str": "figure",
                "text": "The general idea of sentence modification using the insertion method.",
                "num": null
            },
            "FIGREF1": {
                "uris": null,
                "type_str": "figure",
                "text": "The general idea of sentence modification using the replacement method.",
                "num": null
            }
        }
    }
}