File size: 73,439 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
{
    "paper_id": "A88-1002",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T02:04:09.755324Z"
    },
    "title": "A News Story Categorization System",
    "authors": [
        {
            "first": "Philip",
            "middle": [
                "J"
            ],
            "last": "Hayes",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Carnegie Group Inc 650 Commerce Court at Station Square Pittsburgh",
                "location": {
                    "postCode": "15219",
                    "region": "PA"
                }
            },
            "email": ""
        },
        {
            "first": "Laura",
            "middle": [
                "E"
            ],
            "last": "Knecht",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Carnegie Group Inc 650 Commerce Court at Station Square Pittsburgh",
                "location": {
                    "postCode": "15219",
                    "region": "PA"
                }
            },
            "email": ""
        },
        {
            "first": "Monica",
            "middle": [
                "J"
            ],
            "last": "Cellio",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Carnegie Group Inc 650 Commerce Court at Station Square Pittsburgh",
                "location": {
                    "postCode": "15219",
                    "region": "PA"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This paper describes a pilot version of a commercial application of natural language processing techniques to the problem of categorizing news stories into broad topic categories. The system does not perform a complete semantic or syntactic analyses of the input stories. Its categorizations are dependent on fragmentary recognition using patternmatching techniques. The fragments it looks for are determined by a set of knowledge-based rules. The accuracy of the system is only slightly lower than that of human categorizers.",
    "pdf_parse": {
        "paper_id": "A88-1002",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This paper describes a pilot version of a commercial application of natural language processing techniques to the problem of categorizing news stories into broad topic categories. The system does not perform a complete semantic or syntactic analyses of the input stories. Its categorizations are dependent on fragmentary recognition using patternmatching techniques. The fragments it looks for are determined by a set of knowledge-based rules. The accuracy of the system is only slightly lower than that of human categorizers.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "The large economic potential of automatic text processing is leading to an increasing interest in its commercial applications. This paper describes a pilot version of a commercial application of natural language processing techniques to the problem of categorizing news stories into broad topic categories.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "The conventional way to process natural language texts is to have people read them and perform some action based on what they have read. People, for instance, currently categorize news stories for routing purposes and extract information from banking payment telexes so that transactions can be executed.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "Unfortunately, using people tends to be:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "\u2022 slow -people read text slowly;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "\u2022 expensive -if the volume of text is high, processing it requires the efforts of many people;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "\u2022 inconsistent -it is very hard to get a group of people to make consistent decisions about text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "In many cases, the proper processing of text is central to a company's revenue stream, so that improvements in the processing can provide major leverage and justify major contract system expenditures.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "Automatic text processing offers the possibility of such improvements in all three areas. A single text processing machine can potentially do the job of several people faster, cheaper, and more consistently.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "This paper describes an implementation of a system to do text categorization. The texts it operates on are news stories, but similar techniques could be employed on electronic mail messages, telex traffic, technical abstracts, etc.. Once categorization has been accomplished, the results can be used to route the texts involved to interested parties or to facilitate later retrieval of the texts from an archival database.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "The system described here uses the wellestablished natural language processing technique of pattern-matching [1, 5] . Since the input to the system is an arbitrary news story on any topic whatsoever, no attempt is made to perform a complete syntactic or semantic analysis. Instead, categorization is based on the presence of particular words and phrases in particular lexical contexts. As the more detailed description in Section 3 will make clear, however, the approach used goes well beyond the keyword approaches used in information retrieval (e.g. [6] ). In particular, the words and phrases the system looks for and the context in which it looks for them are specified through a modified version of the powerful pattern matching language used in Carnegie Group's Language Craft TM product I [3] . Moreover, the system determines which words and phrases to search for in a given story and how to interpret the presence of these words and phrases according to knowledgebased rules.",
                "cite_spans": [
                    {
                        "start": 109,
                        "end": 112,
                        "text": "[1,",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 113,
                        "end": 115,
                        "text": "5]",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 552,
                        "end": 555,
                        "text": "[6]",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 796,
                        "end": 799,
                        "text": "[3]",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "As simple as these techniques are by current natural language processing standards, the accuracy of the system is high. As described in more detail in Section 4, the system had an average accuracy of 1Language Craft also uses caseframe parsing techniques for complete linguistic analyses. 93% 2 on a sample of 500 random stories that had not been previously processed by the system or seen by its developers. Moreover, this accuracy was obtained without sacrificing computational efficiency. The average processing time was 15 seconds per story 3 on a Symbolics 3640, a figure which we believe could be considerably improved through a detailed performance optimization which we have not performed.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "The remaining sections of the paper describe in more detail: the problem tackled by the system, the approach used, and the results obtained.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "The primary goal in developing the system described in this paper was to demonstrate the feasibility of categorizing news stories by computer in small amounts of time (a few seconds) using natural language processing techniques. The specific task chosen to do this was emulation of the performance of a group of human categorizers. Our raw material was a data base containing many thousands of news stories that had been hand-categorized 4 for any of 72 categories. Our system was required to assign 6 of the 72 categories: acquisitions/mergers, metals, shipping, bonds, war, and disorders. A story could be assigned one or more of these codes, or no code at all if none of the chosen six was appropriate. The restriction to six codes was imposed to keep the effort required to build the system within certain budgetary limits. As Section 3 will show, the approach taken is equally applicable to the larger set of categories.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "Modelling the categorizations produced by human beings presented some difficulties. To summarize:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "\u2022 The text processing techniques used in the system were oriented to identifying concepts explicitly mentioned in a story. They were not 2More pmdsely, its average recall was 93% (i.e. it made 93% of the topic assignments it should have made) and its average precision was also 93% (i.e. 93% of the topics it did assign were correct).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "~The average story length was 250 words; stories varied from about a 100 to about 3000 words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "4The hand-categorizations were done by a group of people who had no involvement with or knowledge of the system we developed.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "well suited to identifying the class of people that a story might be of interest to. The human categorizers of the stories in our data base used both these kinds of considerations when they assigned topic codes to stories. e Some topic codes had relatively vague, subjective definitions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "o The human categorizers were not always consistent in the way they made their topic assignments.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "The news stories themselves posed another challenge. Though the set of topics to be assigned by the system was narrowed from 72 to 6, there was no parallel narrowing of the stream of stories that would serve as input to the system. The full range of story types found in a newspaper occurred in the data base of news stories. As a consequence, our task was not the relatively simple one of, for instance, distinguishing a story about war from one about bonds. War stories also had to be distinguished from military, disaster, crime, diplomacy, politics, and sports stories, to name just a few.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "It was often the case that we could characterize the kind of stories that might mislead the system. We were prepared for sports stories that looked like metals stories (\"...captured the gold medal at the summer Olympics...\") or like war/disorders stories (\"...the battle on center court at Wimbledon...'3. A more difficult challenge was posed by words and phrases that were good predictors of a particular topic but occurred randomly across all story types, sometimes with the same meaning, sometimes not. For instance, the noun note, in the sense of financial instrument, was useful for finding stories about bonds; however, numerous, random stories used that word in a different sense. Metaphorical language was also a problem --not use of fixed phrases (we had no trouble failing to assign the category metals to a story that contained the phrase like a lead balloon) --but rather creative metaphorical language. So, a story about a series of battles in the continuing disposable diaper war between Proctor and Gamble and its competitors was assigned to the disorders category.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Problem",
                "sec_num": "2."
            },
            {
                "text": "The system tackles story categorization in two distinct phases:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1."
            },
            {
                "text": "\u2022 hypothesization: an attempt to pick out all categories into which the story might fall on the basis of the words and phrases it contains; if particular words and phrases suggest more than one category, they will contribute to the hypothesization of each of these categories;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1."
            },
            {
                "text": "\u2022 conf'wmation: an attempt to find additional evidence in support of a hypothesized topic or to determine whether or not the language that led to the topic's being hypothesized was used in a way that misled the system; it is this phase, for instance, that would detect that conflict vocabulary was being used in the context of a sports story and disconf'trm the war and disorders categories for that story. This phase thus has an expert system flavor to it.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1."
            },
            {
                "text": "Both phases use the same basic kind of processing: a contextually limited search for words and phrases using pattern-matching techniques. They are also both organized (conceptually) as a collection of knowledge-based rules. The phases differ only in the directedness with which the search is conducted. Hypothesization always looks for the same words and phrases. Confirmation looks for different words and phrases using specific knowledge-based rules associated with each of the topics that have been hypothesized.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1."
            },
            {
                "text": "The search for words and phrases in both phases is organized around patternsets. A patternset represents a collection of words and phrases that are associated with a given concept, such as conflict. The concepts associated with patternsets sometimes correspond to the topics we are trying to categorize stories into, but they may also be more specific or may span several topics.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1."
            },
            {
                "text": "The basic operation on a patternset is to determine how many of the words and phrases it represents appear in a story. System actions are taken when the number of matches crosses a threshold, at which point we say that the patternset has matched. The thresholds are empirically determined and differ from patternset to patternset and even from use to use of the patternset.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1."
            },
            {
                "text": "Hypothesization is typically performed on the basis of matches of single patternsets. Confirmation rules typically involve branching conditions depending on the results of multiple patternset matches. Individual patternsets may be involved in both hypothesization and confirmation phases.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1."
            },
            {
                "text": "The remainder of this section describes the operation of the system in greater technical detail.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1."
            },
            {
                "text": "Patternsets are collections of patterns. A pattern is an expression in a pattern-matching language that corresponds to one or more words and phrases that might appear in a story. A pattern is said to match the story if any of the words or phrases that it specifies appear in the story. Each pattern has a weight, either probable or possible, with matches of probables counting more than matches of possibles,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Patterns and Patternsets",
                "sec_num": "3.2."
            },
            {
                "text": "according to a scheme explained below. Patterns also have names.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Patterns and Patternsets",
                "sec_num": "3.2."
            },
            {
                "text": "The following pattern, called \"titanium\", will match the word titanium and assign the match a weight of \"probable\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Patterns and Patternsets",
                "sec_num": "3.2."
            },
            {
                "text": "Eight operators are available to allow individual patterns to specify several words and phrases. They are:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "\u2022 ?: specifies an optional subpattern;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "\u2022 ! and,,!!:,, specify alternatives (i.e. they both mean or );~",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "\u2022 ~ and &not: specify a subpattern that should not be matched; ~",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "\u2022 &skip: specifies the maximum number of words to skip over,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "\u2022 +N: specifies that a word is a noun and can therefore be pluralized;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "\u2022 +V: specifies that a word is a verb and can therefore occur with the full range of verbal inflections.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "The following examples illustrate how these operators are used.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "5The operator !! is more efficient than !, but there are some situations where it cannot be used.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "~hc operator &not filters out a subpaRcrn to the left of the subpattem to be matched; ~Fdters out a subpattem to the right. The pattern operator &skip deserves special comment. It allows us to find key expressions even when it is impossible to predict exactly what extraneous words they will contain. Consider, for example, the phrases sell the business and sell the unit; these phrases must be matched if the system is to detect stories about acquisitions. The problem is that expressions like sell the business are rare. Examples of the sorts of phrases that we actually find in acquisitions stories are given below: sell the Seven-Up business sell the ailing Seven-Up unit sell its Seven-Up soft drink business sell 51 pct of the common stock of its unit sell the worldwide franchise beverage business sell about 5 mln dlrs worth of shares in the company With &skip, we can look for the verb sell followed by company, unit, or business without having specify what the intervening words might be.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "In addition to pattern operators, a set of wildeards is also available to rule-writers for matching words that cannot be specified in advance. $ is the general wildcard: it matches any single word or other symbol. $d matches any determiner (a, the, this, etc.); Sn matches any number; $q matches any quantifier (much, many, some, etc.); and $p matches any punctuation mark.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "(titanium) -> probable = titanium",
                "sec_num": null
            },
            {
                "text": "After a story has been read in, the system begins the process of topic determination by applying its hypothesization rules. A hypothesization rule tells the system to hypothesize one or more specified topics if a given pattemset matches the story with a strength greater than a given threshold.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "For example, one of the system's hypothesization rules specifies that the topics war and disorders should be hypothesized if the score for matches in the \"conflict\" patternset is 4 or greater; another rule specifies that the metals topic be hypothesized if the \"metals\" patternset matches with a score greater than 2. The thresholds for each rule are determined empirically based on the rule developer's observation of the performance of the system when different thresholds are used. Note also that there is not necessarily a direct correlation between topics and patternsets; some patternsets could provide evidence for more than one topic, and some topics could make use of more than one patternset.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "The scores for patternset matches are calculated according to the formula:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "(2 x probables) + possibles i.e. a match with a \"probable\" pattern has a weight of 2 while a match with a \"possible\" pattern has a weight of 1. In the course of establishing this weighting system, we experimented with several more complex and finely-grained schemes, but found that they provided no significant advantage in practice.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "After the hypothesization phase comes confirmation. This involves more detailed topicspecific processing to determine whether or not the vocabulary used in hypothesizing the topic was used in a misleading way. The confirmation phase uses topic-specific knowledge-based rules which may try to match additional patterns or pattemsets.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "The most complex confirmation rules in the system are those for the war and disorders topics. These topics were difficult to tell apart, so considerable additional processing was involved. The rules use additional specialized patternsets: one patternset looked specifically for words (including proper names) that occur in war but not disorders stories and another looked for vocabulary that occurs in stories that are both war and disorders stories. There are also patternsets for sports, crime, and disaster vocabulary. The confirmation rules associated with war and disorders attempt to match these rules according to a branching set of conditions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "Consider the following story, for example. The words and phrases in boldface match patterns in the \"conflict\" patternset; the total value of matches is great enough to get the story hypothesized as war and disorders. In the confirmation phase, additional patternsets are run against the story. As soon as Iran and lraq are matched, the topic war is confirmed and the topic disorders is disconfirmed.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "[ [..3",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "The statement by the Iranian High Command appeared to herald the close of an assault on the port city of Basra in southern Iraq.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "[..3",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "It said 81 Iraqi brigades and battalions were totally destroyed, along with 700 tanks and 1,500 other vehicles. The victory list also included 80 warplanes downed, 250 anti.aircraft guns and 400 pieces of military hardware destroyed and the seizure of 220 tanks and armored personnel carriers.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "For the story that follows, the topics war and disorders are also originally hypothesized. In the confirmation phase, two things are discovered: the story mentions no wars by name nor contains any references to countries or organizations involved in conflicts that are classified as wars; and there is nothing in the story that suggests that the topic disorders should be disconfirrned. Hence war is disconf'm-ned and disorders is confirmed.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hypothesization and Confirmation",
                "sec_num": "3.3."
            },
            {
                "text": "Seoul, July 5 -Twelve South Korean women detainees refused food for the fifth consecutive day today after a riot against their maltreatment in a Seoul prison was put down, dissident sources said.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "RIOT REPORTED IN SOUTH KOREAN PRISON",
                "sec_num": null
            },
            {
                "text": "The 12, detained for anti-government protests and awaiting trial, pushed away prison officials, smashed windows and occupied a prison building on Tuesday as a protest against what they called \"suppression of prisoners' human rights\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "RIOT REPORTED IN SOUTH KOREAN PRISON",
                "sec_num": null
            },
            {
                "text": "After two hours, about 40 riot police, firing tear gas, stormed the building and overpowered the protesters, the sources said. Some protesters were injured, they added.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "RIOT REPORTED IN SOUTH KOREAN PRISON",
                "sec_num": null
            },
            {
                "text": "For the story below, both war and disorders are hypothesized and then disconfirmed because tennis is matched during the disconf'u'mation phase.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "RIOT REPORTED IN SOUTH KOREAN PRISON",
                "sec_num": null
            },
            {
                "text": "LONDON, July 2 -Czechoslovak top seed Ivan Lendl served warning that he may finally have come of age on grass when he emerged victorious from a pitched battle with one of the finest exponents of the fast court game at Wimbledon today.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LENDL DEMONSTRATES GRASS COURT MATURITY",
                "sec_num": null
            },
            {
                "text": "The U.S. and French Open tennis champion has never won a title on grass but he outlasted American 10th seed Tim Mayotte 6-4 4-6 6-4 3-6 9-7 over three and a half hours to join Boris Becker, Henri Leconte and Slobodan Zivojinovic in Friday's semifinals.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LENDL DEMONSTRATES GRASS COURT MATURITY",
                "sec_num": null
            },
            {
                "text": "The titanic struggle on court one upstaged the centre court clash between seventh seed Leconte and the remarkable Australian Pat Cash, which had been billed as the day's main attraction [...] The story below is the rare sports story which is also a disorders story. Even though the name of a sporting event, Asian Games, occurs in the text, the topic disorders is not disconfirmed. The reason is that the confirmation patternsets match words and phrases in the story (e.g. radicals and violent protests) that very strongly suggest that real disorders are being described.",
                "cite_spans": [
                    {
                        "start": 186,
                        "end": 191,
                        "text": "[...]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LENDL DEMONSTRATES GRASS COURT MATURITY",
                "sec_num": null
            },
            {
                "text": "SEOUL, July 2 -Police said today they wanted to detain 160 South Koreans to stop sabotage attempts during September's Asian Games in Seoul.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "POLICE SEEK 160 SOUTH KOREAN RADICALS",
                "sec_num": null
            },
            {
                "text": "The 160, mostly students and workers, masterminded various violent protests against the government and the United States in the past months but managed to escape arrest, police said.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "POLICE SEEK 160 SOUTH KOREAN RADICALS",
                "sec_num": null
            },
            {
                "text": "They had been tipped that the radicals were trying to organise big demonstrations against the government during the Asiad, which is to run from September 20 to October 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "POLICE SEEK 160 SOUTH KOREAN RADICALS",
                "sec_num": null
            },
            {
                "text": "\"It is highly probable that they will form radical underground groups to step up their anti-government and anti-U.S, protests and may disrupt the Asian Games in an attempt to defame the government,\" a senior police officer told reporters.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "POLICE SEEK 160 SOUTH KOREAN RADICALS",
                "sec_num": null
            },
            {
                "text": "[...]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "POLICE SEEK 160 SOUTH KOREAN RADICALS",
                "sec_num": null
            },
            {
                "text": "Rather than being expressed in a formal rule language, topic hypothesization and confirmation rules are specified through a lisp program. Having a program allows for fine-grained control by the rule developer.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Flow of Control",
                "sec_num": "3.4."
            },
            {
                "text": "Rather than having a set of hypothesization and conf'u'mation rules which are processed in a fixed order, we allow the rule developer to specify the order and manner of processing in a topic-dependent manner. The major kinds of activities available to rule developers for incorporation into the control code are the following: running one or more patternsets, applying evaluation functions to the resulting matches, and confirming or disconfirming topics. In developing the system, we observed many regularities in the lisp code which controls the flow of processing and we believe it would be possible and profitable to provide rule developers with a more restricted control language which embodies many of these regularities in its primitives.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Flow of Control",
                "sec_num": "3.4."
            },
            {
                "text": "The process of formulating the rulebase of the system, i.e. the collection of patterns, patternsets, and hypothesization and confirmation rules it uses, is an empirical one. It requires human rule developers to examine many stories, create rulebase components according to their intuitions, run stories through the system, observe the results, and modify the system to avoid any miscategorizations that have occurred without introducing new miscategorizations. This task is time-consuming and sometimes tedious. Nevertheless, our experience with the system suggests that it does tend to converge without undue oscillation at an accuracy level that while far from perfect is adequate for many tasks of practical importance (see Section 4). The rule development effort on this system took approximately six person months.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rulebase Development",
                "sec_num": "3.5."
            },
            {
                "text": "An important factor in the success of the rulebase development effort was the separation of the vocabulary the system looks for into a collection of abstract concepts represented by patternsets. The patternsets provide rule developers with a way of thinking about the themes they are looking for in a story when they write the hypothesization and confh'mation rules without becoming mired in questions about which specific words and phrases indicate those themes. In designing the system, we also considered a different approach in which the selection of words and phrases to look for would be determined automatically by a statistical method. Since we did not adopt this approach, we have no direct evidence that it would not have worked as well as the laborintensive method chosen. However, our choice was influenced by a belief that a statistical method would not provide us with a choice of words and phrases that could be used to make distinctions as precisely as the patterns of the kind described above that were chosen by humans.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rulebase Development",
                "sec_num": "3.5."
            },
            {
                "text": "As shown in [2] , accuracy is particularly problematic with a traditional keyword approach regardless of whether the keywords are selected by humans or statistically. And if we had adopted a statistical approach, it would have been computationally expensive to vary the length of the phrases chosen as much as human rule developers do. It would also have been difficult to establish the contextual restrictions that human rule developers establish (e.g. this word, so long as it is not followed by one of these four others). Rules of the complexity of the confirmation rule for war and disorders described in Section 3.3 are of course essentially impossible to establish by statistical means.",
                "cite_spans": [
                    {
                        "start": 12,
                        "end": 15,
                        "text": "[2]",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rulebase Development",
                "sec_num": "3.5."
            },
            {
                "text": "Some interesting possibilities for a statistical approach to defining keywords have appeared recently in conjunction with semantic information about potential keywords [7] and in conjunction with very powerful parallel hardware devices [4] . However, given the current state of the art, we continue to believe that our decision to use rules formulated and refined by human developers was a sound one from the point of view of the accuracy of the resulting system.",
                "cite_spans": [
                    {
                        "start": 168,
                        "end": 171,
                        "text": "[7]",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 236,
                        "end": 239,
                        "text": "[4]",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rulebase Development",
                "sec_num": "3.5."
            },
            {
                "text": "The accuracy of the system for topic assignments was measured through two percentages for each of the six topics:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Measuring Performance",
                "sec_num": "4.1."
            },
            {
                "text": "\u2022 recall: the percentage of stories assigned the topic code by human categorizers that were also assigned that code by the system;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Measuring Performance",
                "sec_num": "4.1."
            },
            {
                "text": "\u2022 precision: the percentage of stories assigned the topic code by the system that actually carried the topic code assigned by the human categorizers.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Measuring Performance",
                "sec_num": "4.1."
            },
            {
                "text": "The recall rate serves as a measure of the number of stories for which the system misses an appropriate topic code; a high recall percentage will therefore mean few such false negatives. The precision rate, on the other hand, measures the number of stories for which the system chooses an incorrect topic. A high precision percentage means few such false positives. We emphasized high recall over high precision.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Measuring Performance",
                "sec_num": "4.1."
            },
            {
                "text": "The results obtained from the system were very promising.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "After certain necessary adjustments (described below) to the raw results, the system had an average recall rate of 93% (i.e. it made 93% of the topic assignments it should have made and missed only 7%) and an average precision rate also of 93% (i.e. 93% of the topics it did assign were correct). Another way of expressing this is that it had on average only 7% false negatives and 7% false positives in its topic assignments. This level of accuracy was achieved in an average of around 15 seconds per story on a Symbolics 3640 in Common Lisp.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "Little effort was spent to optimize the execution time and we believe that a substantial improvement in speed is possible.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "Adjustments to the raw recall and precision figures produced by the system were necessary because, as described in Section 2, we discovered three problematic features of the hand-categorizations against which the system was being evaluated: they were not always content-based; they were not always consistent; and some topic definitions were vague. Given this, it was clear that raw performance scores would not give a meaningful picture of how well the system worked, so we devised a score-adjustment procedure to provide results that would reflect system performance more accurately. The remainder of this section describes that procedure and presents the raw and adjusted results we obtained.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "We used an adjustment procedure that was based on the assumption that there are three explanations for disagreements between the system and the human categorizers about the assignment of a topic to a story:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "\u2022 The human categorizer is clearly wrong.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "\u2022 The system is clearly wrong.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "\u2022 The topic assignment is debatable. This case can typically be attributed to one of the three sources of difficulty described above.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "A set of 500 stories was run through the system. These stories had never before been processed, and no hypothesization or conf'trmation rules had ever been based on them. A Carnegie Group employee who was not involved with the system produced score adjustments for each topic disagreement between the system and the human categorizers. The employee was presented with a story and told that there was a disgreement on a specific named topic;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "she was not told which choice the system or the human categorizers had made. The employee was asked to decide whether the topic was appropriate for the story, inappropriate, or debatable. Debatable cases counted in favor of the system.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "The results of this experiment before and after adjustment of the system's scores were as follows (where acq is acquisitions/mergers, mtl is metals, shp is shipping, bnd is bonds, and dis is disorders). Recall is 92% or higher, except in the case of the shipping code. This is not surprising because it turned out that shipping was a strongly interest-based category, as far as the human categorizers were concerned. So, stories about rough weather in the St. Lawrence seaway (but not the Rhine) and the devaluation of the rupee (but not the Turkish lira) were classified as shipping stories because human categorizers possessed the expert knowledge that shippers are interested in that particular waterway and that particular currency.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4.2."
            },
            {
                "text": "The precision scores are actually higher than the corresponding recall scores in the case of war and bonds. Since we have found that precision can be traded off against recall by appropriate manipulation of thresholds associated with our rulebase, this suggests that the recall rate for those two topics could be further improved while still maintaining an acceptable precision rate.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Raw",
                "sec_num": null
            },
            {
                "text": "The adjustment procedure also allowed us to measure the performance of the hand-categorizers. While adjusted precision scores were perfect for all six topics, adjusted recall scores ranged from 81% to 100%, with an average of 94%. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Raw",
                "sec_num": null
            },
            {
                "text": "While human performance on precision is clearly superior to that of the system, the average recall rates of human categorizers and of the system are very similar (94% v. 93%). Closer examination of the results, however, shows that the kind of errors made are quite different. Human errors stem mainly from inconsistent application of categories, especially the categories with the vaguest definitions, and from failing to specify all the categories when several should have been assigned to a story. System errors on the other hand stem largely from misinterpretation of the way in which language is being used. This sometimes results in ridiculous categorizations of a kind that humans never produce.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "Out of 500 stories, the system produced a total of 28 \"lemons\" (stories that were clearly assigned the wrong categories). We analyzed these stories and discovered six sources of errors:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "\u2022 The system did not match useful words or phrases, or the disconfirmation rules were too powerful.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "\u2022 The topic vocabulary was not much used in the story.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "\u2022 The system used the story background to derive the topic.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "\u2022 The topic vocabulary came too late in the story.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "\u2022 The topic vocabulary was used with different meanings.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "\u2022 The topic vocabulary was used with the same meaning, but different focus.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "Examples and further discussion follow.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "100%",
                "sec_num": null
            },
            {
                "text": "Some stories did not use the topic vocabulary more than one or two times. Setting thresholds very low would catch these stories, but generate many false positives as well. Most stories that had this problem were also very short, so we added length-dependent thresholds to address the problem. This technique worked for metals stories, where the vocabulary is somewhat distinctive, but would not work for acquisitions stories, where the vocabulary consists of very common words like buy and sell.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Topic Vocabulary Not Much Used In Story",
                "sec_num": "4.2.1."
            },
            {
                "text": "News stories sometimes have background information included which does not have much to do with the main point of the story. For example, the following story, about the Pope's visit to Colombia, was miscategorized as a metals story because of the background information about the country. Solving this problem requires a deep understanding of the structure of the story.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Story Background Used To Derive Topic",
                "sec_num": "4.2.2."
            },
            {
                "text": "Chiquinquira, Colombia, July 3 -Security precautions for Pope John Paul II were tightened today, with hundreds of troops making thorough body searches of visitors to this colonial town high in the Andes mountains.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "SECURITY FOR POPE TIGHTENS",
                "sec_num": null
            },
            {
                "text": "Chiquinquira has been spared the guerilla warfare which has torn much of Colombia over the past three decades. But the nearby Muzo emerald mines, the country's biggest, have attracted adventurers who often feud violently in the town.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "[...]",
                "sec_num": null
            },
            {
                "text": "Some Muzo miners have moved on to the more lucrative drug traffic [...].",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "[...]",
                "sec_num": null
            },
            {
                "text": "Sometimes stories are miscategorized because of the metaphorical language they use. For example, in one story the word revolution appeared numerous times: the British government was calling for a revolution in broadcasting. Another contained the phrases ready to go to war, make peace, make war, target, and heavy losses; the subject of the story was labor negotiations in the automobile industry. Since the system does not really understand the texts it processes, it is inevitable that it will be fooled from time to time by such usage.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Topic Vocabulary Used With Different Meaning",
                "sec_num": "4.2.3."
            },
            {
                "text": "But Different Focus The following story illustrates another problem for which there is no obvious solution. The word army occurs four times (not all shown), and the sense of the word in this military story is exactly the sense it might have in an actual war or disorders story.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Topic Vocabulary Used with Same Meaning",
                "sec_num": "4.2.4."
            },
            {
                "text": "Peking, July 4 -The Chinese army will allow non-commissioned officer ranks for the first time as part of its reform program, the New China News Agency said today.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "CHINESE ARMY TO HAVE NCOS FOR FIRST TIME",
                "sec_num": null
            },
            {
                "text": "It said soldiers who have been in the army for one year and had a good record would, after training at two special schools, serve as NCOS.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "CHINESE ARMY TO HAVE NCOS FOR FIRST TIME",
                "sec_num": null
            },
            {
                "text": "[...]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "CHINESE ARMY TO HAVE NCOS FOR FIRST TIME",
                "sec_num": null
            },
            {
                "text": "This paper has shown that high accuracy automatic text categorization is feasible for texts covering a diverse set of topics, using the well-established natural language processing technique of pattern matching applied according to knowledge-based rules, but without requiring a complete syntactic or semantic analysis of the texts. Automatic text processing of this kind has many potential applications with high economic paybacks in the routing and archiving of news stories, electronic messages, or other forms of on-line text. We expect that many such systems will be in commercial use within the next few years.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "5."
            }
        ],
        "back_matter": [
            {
                "text": "Peter Neuss and Scott Sailer contributed substantially to the design and implementation of the system described in this paper.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgements",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Semantic Case Frame Parsing and Syntactic Generality",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "J"
                        ],
                        "last": "Hayes",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Andersen",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Safier",
                        "suffix": ""
                    }
                ],
                "year": 1985,
                "venue": "Proc. of 23rd Annual Meeting of the Assoc. for Comput. Ling",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hayes, P. J., Andersen, P., Safier, S. Semantic Case Frame Parsing and Syntactic Generality. Proc. of 23rd Annual Meeting of the Assoc. for Comput. Ling., Chicago, June, 1985.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "The Connection Machine",
                "authors": [
                    {
                        "first": "W",
                        "middle": [
                            "D"
                        ],
                        "last": "Hillis",
                        "suffix": ""
                    }
                ],
                "year": 1985,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hillis, W. D.. The Connection Machine. MIT Press, Cambridge, Mass., 1985.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Conversational Language Comprehension Using Integrated Pattern-Matching and Parsing",
                "authors": [
                    {
                        "first": "R",
                        "middle": [
                            "C"
                        ],
                        "last": "Parkison",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [
                            "M"
                        ],
                        "last": "Colby",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [
                            "S"
                        ],
                        "last": "Faught",
                        "suffix": ""
                    }
                ],
                "year": 1977,
                "venue": "Artificial Intelligence",
                "volume": "9",
                "issue": "",
                "pages": "111--134",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Parkison, R. C., Colby, K. M., and Faught, W. S. \"Conversational Language Comprehension Using Integrated Pattern-Matching and Parsing\". Artificial Intelligence 9 (1977), 111-134.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Introduction to Modern Information Retrieval",
                "authors": [
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Salton",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [
                            "J"
                        ],
                        "last": "Mcgill",
                        "suffix": ""
                    }
                ],
                "year": 1983,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Salton, G. and McGill, M. J.. Introduction to Modern Information Retrieval. McGraw-Hill, New York, 1983.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "The Use of Machine-Readable Dictionaries in Sublanguage Analysis",
                "authors": [
                    {
                        "first": "D",
                        "middle": [
                            "E"
                        ],
                        "last": "Walker",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Amsler",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Ri A ; R. Grishman",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kittredge",
                        "suffix": ""
                    }
                ],
                "year": 1986,
                "venue": "Analyzing Language in Restricted Domains: Sublanguage Description and Processing",
                "volume": "",
                "issue": "",
                "pages": "69--83",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Walker, D. E. and Amsler, RI A. The Use of Machine-Readable Dictionaries in Sublanguage Analysis. In Analyzing Language in Restricted Domains: Sublanguage Description and Processing, R. Grishman and R. Kittredge, Ed., Lawrence Erlbaum Associates, Hillsdale, New Jersey, 1986, pp. 69-83.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "The XCALIBUR Project: A Natural Language Interface to Expert Systems",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "G"
                        ],
                        "last": "Carbonell",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [
                            "M"
                        ],
                        "last": "Boggs",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [
                            "L"
                        ],
                        "last": "Mauldin",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [
                            "G"
                        ],
                        "last": "Anick",
                        "suffix": ""
                    }
                ],
                "year": 1983,
                "venue": "Proc. Eighth Int. JL Conf. on Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Carbonell, J. G., Boggs, W. M., Mauldin, M. L., and Anick, P. G. The XCALIBUR Project: A Natural Language Interface to Expert Systems. Proc. Eighth Int. JL Conf. on Artificial Intelligence, Karlsruhe, August, 1983.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Statistical semantics: Analysis of the potential performance of keyword information systems",
                "authors": [
                    {
                        "first": "G",
                        "middle": [
                            "W"
                        ],
                        "last": "Furnas",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [
                            "K"
                        ],
                        "last": "Landauer",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [
                            "T"
                        ],
                        "last": "Dumais",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Gomez",
                        "suffix": ""
                    }
                ],
                "year": 1983,
                "venue": "Bell System Technical Journal",
                "volume": "62",
                "issue": "6",
                "pages": "1753--1806",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Furnas, G. W., Landauer, T. K., Dumais, S. T., and Gomez, L. M. \"Statistical semantics: Analysis of the potential performance of keyword information systems\". Bell System Technical Journal 62, 6 (1983), 1753-1806.",
                "links": null
            }
        },
        "ref_entries": {}
    }
}