Datasets:
File size: 151,318 Bytes
f06250e | 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 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 | Towards Verifiable Generation: A Benchmark for Knowledge-aware
| | | | | Language | Model | Attribution | | | | | |
| --- | --- | --- | --- | -------- | ----- | ----------- | --- | --- | --- | --- | --- |
XinzeLi1,YixinCao2†,LiangmingPan3,YuboMa1,AixinSun1†
1 S-Lab,NanyangTechnologicalUniversity
2 SingaporeManagementUniversity3 UniversityofCalifornia,SantaBarbara
| | {xinze002, | | | yubo001}@e.ntu.edu.sg | | | | axsun@ntu.edu.sg | | | |
| --- | ---------- | --- | ---------------- | --------------------- | --- | --------------------- | --- | ---------------- | --- | --- | --- |
| | | | yxcao@smu.edu.sg | | | liangmingpan@ucsb.edu | | | | | |
Abstract
Question: How did Orazio Gentileschi's influence on Artemisia's life
and career shape her development as a Baroque painter?
| Althoughachievinggreatsuccess,LargeLan- | | | | | | | Retrieve | | | | |
| --------------------------------------- | --- | --- | --- | --- | --- | --- | -------- | --- | --- | --- | --- |
guageModels(LLMs)usuallysufferfromun-
| reliablehallucinations. | | | | Althoughlanguageat- | | Knowledge | | | | | |
| ----------------------- | --- | --- | --- | ------------------- | --- | --------- | --- | ------- | ------- | --- | --- |
| | | | | | | Graph | | Q 1 0 2 | 8 1 8 1 | | |
tributioncanbeapotentialsolution,thereare P a i n t e r Q37853
Baroque
nosuitablebenchmarksandevaluationmetrics
| toattributeLLMstostructuredknowledge. | | | | | In | | | | occupation | | |
| ------------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | ---------- | --- | --- |
style
| thispaper,wedefineanewtaskofKnowledge- | | | | | | | | o ccupation | | | |
| -------------------------------------- | --- | --- | --- | --- | --- | --- | --- | ----------- | --- | --- | --- |
movemen
| awareLanguageModelAttribution(KaLMA) | | | | | | | | t | child | | |
| ------------------------------------ | --- | --- | --- | --- | --- | --- | --- | --- | ----- | --- | --- |
place of birth
birth
| that | improves | | upon three | core concerns | with | | | | father | member | |
| -------------------------- | -------- | --- | ---------- | -------------- | ---- | --- | --- | ----------- | ------ | ----------- | --- |
| conventionalattributedLMs. | | | | First,weextend | | | | | | | |
| | | | | | | | | Orazio | | Artemisia | |
| | | | | | | | | Gentileschi | | Gentileschi | |
attribution source from unstructured texts to Citizen Citizen
| | | | | | | | | Q367560 | | Q212657 | |
| --- | --- | --- | --- | --- | --- | --- | --- | ------- | --- | ------- | --- |
KnowledgeGraph(KG),whoserichstructures
| benefit | both | the | attribution | performance | and | | | | | Italy | |
| ------- | ---- | --- | ----------- | ----------- | --- | --- | --- | --- | --- | ----- | --- |
Q38
| workingscenarios. | | | Second,weproposeanew | | | | Attribute | | | | |
| ----------------- | --- | --- | -------------------- | --- | --- | --- | --------- | --- | --- | --- | --- |
LLM
“ConsciousIncompetence"settingconsidering
Output: Artemisia Gentileschi was an Italian painter born on July 8,
the incomplete knowledge repository, where 1596 [NA] in Rome [Q212657, citizen: Italy, occupation: painter,
the model identifies the need for supporting place of birth: Rome]. She was a member of ...
knowledgebeyondtheprovidedKG.Third,we Figure1: Ademonstrationofourtasksetup. Givena
question,thesystemgeneratesanswersattributedfrom
proposeacomprehensiveautomaticevaluation
metricencompassingtextquality,citationqual- aretrievedknowledgegraph.Theunderlinesinquestion
ity,andtextcitationalignment. Toimplement aretheretrievedentities,andtheunderlinesinoutputs
| | | | | | | arethecitations. | | [NA]isthe“NotApplicableCitation”. | | | |
| --- | ----- | ------------ | --- | ------------------ | --- | ---------------- | --- | --------------------------------- | --- | --- | --- |
| the | above | innovations, | | we build a dataset | in | | | | | | |
biographydomainBioKaLMAviaevolution-
aryquestiongenerationstrategy,tocontrolthe
questioncomplexityandnecessaryknowledge triesthatrequireprecisionandfactualknowledge
to the answer. For evaluation, we develop a likefinance,law,andmedicaltreatment.
baselinesolutionanddemonstratetheroomfor To minimize the negative impacts, researchers
improvementinLLMs’citationgeneration,em-
| | | | | | | have | proposed | the | task | of language | attribu- |
| --- | --- | --- | --- | --- | --- | ---- | -------- | --- | ---- | ----------- | -------- |
phasizingtheimportanceofincorporatingthe
tion(Bohnetetal.,2023),whichnotonlyenables
"ConsciousIncompetence"setting,andthecrit-
userstoverifythegeneratedtextflexiblybutalso
icalroleofretrievalaccuracy.
| | | | | | | contributes | to | many | important | applications, | such |
| --- | ------------ | --- | --- | --- | --- | ------------------ | --------- | -------------------------- | ----------- | ------------- | ------- |
| 1 | Introduction | | | | | assituationreports | | (Reddyetal.,2023),academic | | | |
| | | | | | | papers | (Salvagno | et | al., 2023), | medical | diagno- |
Recently, Large Language Models (Brown et al., sis(ZucconandKoopman,2023).Existingworks
2020) (LLMs) have exhibited great capability in mainlyattributegeneratedoutputstounstructured
open-endedquestionanswering(Yangetal.,2019). documents like web pages (Nakano et al., 2021;
However,thegeneratedanswersmayincludefac- Menicketal.,2022)orpassages(Gaoetal.,2023).
tualerrorsandarenotalwaysreliable,andiscom- Toverifytheanswerquality,theytypicallycompare
monlyknownasthe“hallucination”(Shusteretal., withahumanannotatedreferenceanswerforauto-
2021;Jietal.,2023)problem. Forinstance,LLMs maticevaluationorconducthumanevaluation. We
maygivewrongdiagnosistopatient’ssymptoms. arguethatthereareseveralconcernsonsuchtask
Hallucinationhassevereharmsespeciallyonindus- definition. Firstly,aredocumentstheonlysource
database2
forattribution? Manyreal-worldapplicationshave ical (Plum et al., 2022) and WikiData,
theirownknowledgebasesorsemi-structuredre- BioKaLMAcontains1,085dataentries. Eachdata
ports. Secondly,doestheattributionsourcealways entryincludesquestionandknowledgerequiredto
includealltherequiredknowledge? Weconsider answerthequestion. Forevaluation,weseparately
thecoverageissuesincenoperfectrepositorycan evaluatethegeneratedtext,thegeneratedcitations,
containalltheinformationinthisworld. Thirdly, andthealignmentbetweentextsandcitations. We
howtosystematicallyevaluatetheattributedcon- useG-Eval(Liuetal.,2023b)toautomaticallyeval-
tentwithoutreferences? Foropen-endedquestions, uatethetextquality. Wealsodesignmeasurement
there are unlimited number of answers and it is forcorrectness,precision,andrecallforcitations.
difficulttodefineasinglegroundtruth. Lastly,wedeterminethealignmentbetweentexts
Toaddressthefirstchallenge,weutilizeknowl- andcitationsemployingNLI(Daganetal.,2005)
edgegraph(KG)asareliablesourceforattribution, Wesummarizeourcontributionsasfollows: 1)
WedefinethetaskofKnowledge-awareLanguage
namelyKnowledge-awareLanguageModelAttri-
bution(KaLMA).Weshowademonstrationoftask Model Attribution (KaLMA) that attributes lan-
inFigure1. KGsefficientlyorganizeworldknowl- guagemodelstostructuredknowledge. 2)Wede-
signacompletebenchmarkingpipeline,including
| edge in | a structured | | manner | and | has the | poten- | | | | | | | |
| -------------------------------- | ------------ | --- | ------ | --- | ----------- | ------ | -------- | --------- | --- | ---------- | --- | -------- | ----- |
| | | | | | | | dataset, | baseline, | and | evaluation | | metrics. | 3) We |
| tialtounifyvariousformatsofdata. | | | | | Forexample, | | | | | | | | |
databasescanbeeasilyconvertedintoKGs,or,pas- conductextensiveexperimentsandshowroomfor
sagesandwebpagescanberepresentedasanode improvementoftheLLMs’abilitytogenerateac-
| | | | | | | | curate | and thorough | | citations | based | on provided | |
| ------------------ | --- | --- | ---------------------- | --- | --- | --- | ------ | ------------ | --- | --------- | ----- | ----------- | --- |
| inKGlikeWikipedia. | | | KaLMAdiffersfromentity | | | | | | | | | | |
linking(Sevgilietal.,2022)sincethesentencesor knowledgegraphs. Ourexperimentson“Conscious
phrasesareattributedtoaknowledgetripletrather Incompetence”investigatethecapabilityofcurrent
LLMstoidentifyiftherearerequiredknowledge
| thanasingleentity. | | Forthesecondchallenge,we | | | | | | | | | | | |
| ------------------ | --- | ------------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
tacklethecoverageproblembymakingthemodel notinknowledgegraph. Wehighlightthenecessity
awareofitslimitations. Weintroduceanewsetting ofincorporatingthissettinginfuturelanguageat-
“ConsciousIncompetence”(CurtissandWarren, tributionworks. Furthermore,ourablationstudies
demonstratethecrucialroleofretrievalaccuracy
1974),whichisthepsychologicalstagethatoneis
aware of the knowledge gap. During generation, inachievingdesirablegenerationresults.
| LLMs | identify | sentences | that | require | supporting | | | | | | | | |
| --------- | -------- | --------- | ------------- | ------- | ---------- | --- | ---------------- | --- | --- | --- | --- | --- | --- |
| | | | | | | | 2 TaskandDataset | | | | | | |
| knowledge | absent | in | the knowledge | | graph. | Our | | | | | | | |
settingenablesanattributedLMtorecognizethe
| | | | | | | | 2.1 TaskFormulation | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | ------------------- | --- | --- | --- | --- | --- | --- |
knowledgegapsandallowsuserstoverifyuncer-
WeherebydefinethetaskKnowledge-awareLan-
| tainclaims, | whichenhancestrustworthiness. | | | | | For | | | | | | | |
| ----------- | ----------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
guageModelAttribution(KaLMA):Givenaques-
| the third | challenge, | we | propose | a | comprehensive | | | | | | | | |
| --------- | ---------- | --- | ------- | --- | ------------- | --- | --- | --- | --- | --- | --- | --- | --- |
automaticevaluationmetricincludingtextquality, tionq andtheknowledgegraphG,thesystemgen-
| | | | | | | | erates an | output | text | t that | answers | the question. | |
| ------------------------------------------ | -------- | -------- | -------- | ---------- | --- | --- | ----------------------------------------- | ------ | ---- | ------ | ------- | ------------- | --- |
| citation | quality, | and text | citation | alignment. | | The | | | | | | | |
| | | | | | | | Theoutputtextconsistsofalistofmsentencess | | | | | | , |
| entireevaluationprocessdoesnotrequirehuman | | | | | | | | | | | | | 1 |
annotatedgroundtruth. ...,s groundedwithalistofngroundedknowl-
m
| | | | | | | | edge k 1 | .. k n | where | {k 1 ..k | n } ∈ | G. Each | knowl- |
| ------------ | --- | --- | ----- | ------------ | --- | -------- | -------- | ---------------- | ----- | -------- | ---------------- | ------- | ------ |
| To implement | | the | above | innovations, | | we first | | | | | | | |
| | | | | | | | edgek | isasub-graphofG. | | | Eachsentencesmay | | |
designanautomaticdatasetconstructionpipeline.
begroundedbyzerouptomultipleknowledge.
| Usingthispipeline,weconstructadataset1 | | | | | | inthe | | | | | | | |
| -------------------------------------- | ---- | ----------- | ------ | --------- | --- | -------- | ------------------------------ | --- | --- | --- | --- | -------- | --- |
| biographical | | domain, | namely | BioKaLMA, | | for a | | | | | | | |
| | | | | | | | SettingofConsciousIncompetence | | | | | Weextend | |
| benchmark | with | all-rounded | | automatic | | measure- | | | | | | | |
thistasksettingtoincludeconsciousincompetence.
| ments. | Biography | forms | a good | test-set | | for attri- | | | | | | | |
| ------ | --------- | ----- | ------ | -------- | --- | ---------- | --- | --- | --- | --- | --- | --- | --- |
Giventhesameinput,eachsentencesintheoutput
| bution | due to | its practical | application | | and | conve- | | | | | | | |
| ------ | ------ | ------------- | ----------- | --- | --- | ------ | --- | --- | --- | --- | --- | --- | --- |
texttcanmaptoaNotApplicableCitation(weuse
nient evaluation. The availability of high-quality [NA]torepresentit)ifitincludessomeknowledge
knowledgegraphlikeWikiDataalsobenefitsour
| | | | | | | | to be verified, | | but the | knowledge | | is absent | in the |
| ------- | ------------- | ----------- | -------- | ---- | ------------- | -------- | ---------------- | ------ | ------------ | --------------------- | --------- | --------- | ------ |
| dataset | construction. | | Derived | from | the biograph- | | | | | | | | |
| | | | | | | | knowledgegraphG. | | | Asentencecanmaptoboth | | | |
| | | | | | | | [NA] and | a list | of sub-graph | | knowledge | if | it can |
| 1The | codes | and dataset | BioKaLMA | | are | publicly | | | | | | | |
available in https://github.com/lixinze777/ 2https://plumaj.github.io/
| Knowledge-aware-Language-Model-Attribution | | | | | | | biographical/ | | | | | | |
| ------------------------------------------ | --- | --- | --- | --- | --- | --- | ------------- | --- | --- | --- | --- | --- | --- |
| | | | | | | General | Question: | | | | | |
| ------------ | -------- | ------ | --------- | ----- | --- | -------- | --------- | --- | ------- | -------- | --- | --- |
| be partially | verified | by the | knowledge | graph | G. | | | | | | | |
| | | | | | | Who were | Oscar | and | Richard | Hertwig, | | |
[NA]isnotacitationonconventionalmeans,buta
| | | | | | | and what | were | their | contributions | | | to |
| ------------------------ | --- | --- | --- | --- | --- | ---------- | --------- | ------- | ------------- | -------- | --- | --- |
| indicatorofknowledgegap. | | | | | | the fields | of | anatomy | and | biology? | | |
| | | | | | | Specific | Question: | | | | | |
| | | | | | | What were | the | career | paths | and | | |
2.2 DatasetConstruction significant contributions of Oscar and
| | | | | | | Richard | Hertwig | in | the | fields | of | |
| --- | --- | --- | --- | --- | --- | ------- | ------- | --- | --- | ------ | --- | --- |
Each entry of dataset bioKaLMA includes two anatomy and biology, and who were
| | | | | | | their notable | | mentors | and | students? | | |
| --------------------------------- | --- | --- | --- | --- | ------ | ------------- | --- | ------- | --- | --------- | --- | --- |
| questionsandaminimumknowledgeset. | | | | | Thetwo | | | | | | | |
questionsenquireaboutthesamepeopleonsimilar Minimum Knowledge Set:
| | | | | | | [’Q85907’, | ’occupation’, | | | ’biologist’] | | |
| -------------------------- | --------------- | ---------------- | ------------ | --- | -------- | ----------- | ------------- | -------------- | --------- | ------------ | ------ | --- |
| aspectsoftheirlifestories. | | Theminimumknowl- | | | | | | | | | | |
| | | | | | | [’Q85907’, | ’doctoral | | student’, | | | |
| edge set | is the smallest | set | of knowledge | | that is | | | | | | | |
| | | | | | | ’Stanislaus | | von Prowazek’] | | | | |
| | | | | | | [’Q68753’, | ’doctoral | | advisor’, | | ’Ernst | |
| required | to answer each | question. | | One | question | | | | | | | |
Haeckel’]
isageneralversionandtheotherisspecific. The [’Q68753’, ’student of’, ’Ernst
| generalquestionsaremoreconciseandnaturalfor | | | | | | Haeckel’] | | | | | | |
| ------------------------------------------- | --- | --- | --- | --- | --- | ---------- | ---------- | --- | ----- | --- | ------ | --- |
| | | | | | | [’Q68753’, | ’nominated | | for’, | | ’Nobel | |
humanreaders,andthespecificversionquestions
| | | | | | | Prize in | Physiology | | or | Medicine’] | | |
| --- | --- | --- | --- | --- | --- | -------- | ---------- | --- | --- | ---------- | --- | --- |
haveatighterbondtotheminimumknowledgeset, Table 1: An example for generated data entry in
BioKaLMA.Q85907andQ68753areRichardHertwig
| and is hence | more accurate | | for evaluating | | LLMs. | | | | | | | |
| ------------ | ------------- | --- | -------------- | --- | ----- | --- | --- | --- | --- | --- | --- | --- |
AnexampledatapieceisshowninTable1. andOscarHertwig’sQIDsinWikiData
| We construct | the | dataset | using | an automatic | | | | | | | | |
| ------------------------------- | --- | ------- | ---------------- | ------------ | --- | ----------------- | --- | --- | ------- | --- | -------- | --- |
| | | | | | | Metric(fullscore) | | | General | | Specific | |
| pipelineconsistingofthreesteps: | | | PersonSelection, | | | | | | | | | |
| | | | | | | Authenticity(1) | | | 1.00 | | 1.00 | |
NameDisambiguation,andEvolutionaryQuestion
| | | | | | | Relevance(1) | | | 0.73 | | 0.84 | |
| ----------- | ------------------------------ | --- | --- | --- | --- | -------------- | --- | --- | ---- | --- | ---- | --- |
| Generation. | Inthefirsttwosteps,weuseSPARQL | | | | | | | | | | | |
| | | | | | | Naturalness(5) | | | 4.38 | | 3.52 | |
queriestoselectrelatedpeoplefromhumanwritten
| | | | | | | Significance(5) | | | 3.94 | | 3.68 | |
| --------- | ------------ | ----- | -------- | ------------ | --- | --------------- | --- | --- | ---- | --- | ---- | --- |
| sentences | and identify | their | identity | in WikiData. | | | | | | | | |
Inthethirdstep,weiterativelyconstructparagraph Table2: HumanEvaluationonBioKaLMAdataset.
| and question | about the | selected | people. | | The first | | | | | | | |
| ------------ | --------- | -------- | ------- | --- | --------- | --- | --- | --- | --- | --- | --- | --- |
iterationstartswithahumanwrittensentenceabout
| | | | | | | EvaluationofDatasetQuality | | | | Weevaluatethe | | |
| --- | --- | --- | --- | --- | --- | -------------------------- | --- | --- | --- | ------------- | --- | --- |
theselectedpeople. Ineachnextiteration,weapply BioKaLMAdatasetonthefollowingfourmetrics
adataselectionalgorithmtoselectanappropriate
| | | | | | | to ensure | the quality | of | the | dataset: | 1) | Authen- |
| --------- | ------------- | ----- | --- | --- | -------- | --------------------------------------------- | ----------- | --- | --- | -------- | --- | ------- |
| knowledge | from WikiData | based | on | the | existing | | | | | | | |
| | | | | | | ticity: Thegeneratedquestionsshouldaccurately | | | | | | |
paragraph, and extend the paragraph to include reflecttheobjectivefacts. 2)Relevance: Eachmin-
theadditionalknowledgeusingLLM.Then,LLM
imumknowledgesetshouldprovidesupporttothe
constructsthequestionsusingthefinalparagraph
| | | | | | | correspondingquestion. | | | Eachpieceofknowledge | | | |
| ------------- | ----------- | --- | -------- | --------- | --- | ---------------------- | --- | --- | -------------------- | --- | --- | --- |
| as an answer. | The general | and | specific | questions | | | | | | | | |
fromtheminimumknowledgesetisnotredundant.
are generated with different prompts and demon- 3)Naturalness: Thegeneratedquestionshouldbe
| strations. | All the selected | knowledge | | from | each | | | | | | | |
| ---------- | ----------------- | --------- | --------- | ---- | -------- | --------------------------------------- | --- | --------- | --- | -------- | ------ | --- |
| | | | | | | conciseandunderstandablebyhumanreaders. | | | | | | 4) |
| iteration | form the “minimum | | knowledge | | set” for | | | | | | | |
| | | | | | | Significance: | The | generated | | question | should | be |
thequestion. Whileweusethehumanbiography meaningfulandhelpfultousers.
| domain | as an example, | this | method | is applicable | | | | | | | | |
| ------------- | ---------------------------- | ---- | ------ | ------------- | --- | --------------------------------- | --------------- | --- | ----- | --- | ----------- | --- |
| | | | | | | To our | best knowledge, | | there | is | no perfect | au- |
| toalldomains. | Wepresentthedetailsofthedata | | | | | | | | | | | |
| | | | | | | tomaticevaluationforthesemetrics. | | | | | Naturalness | |
constructioninAppendixA.
| | | | | | | and significance | | are subjective. | | Hence, | | we apply |
| --- | --- | --- | --- | --- | --- | ---------------- | --- | --------------- | --- | ------ | --- | -------- |
humanevaluationtoensurethedatasetquality.
2.3 DatasetAnalysis
| | | | | | | We randomly | | sample | 50 | data | entries | from |
| --- | --- | --- | --- | --- | --- | ----------- | --- | ------ | --- | ---- | ------- | ---- |
Statistics There are 1,085 data entries in BioKaLMA and ask human annotators to evalu-
BioKalMA. On average, there are 6.8 pieces of atethedataentriesbasedonthefourmetrics. The
generalandspecificquestionsareevaluatedsepa-
| knowledge | in each | “minimum | knowledge | | set”. | | | | | | | |
| --------- | ------- | -------- | --------- | --- | ----- | --- | --- | --- | --- | --- | --- | --- |
BioKaLMA demonstrates a good demographic rately. MoredetailsaregiveninAppendixC.
variation. It includes a wide range of geograph- Thefinalresultforeachmetricistakenaverage
icaldistributionofpeoplefrom196countriesand andreportedinTable2. Forbothgeneralandspe-
949 cities, taking 279 kinds of different occupa- cificsettings,thequestionsfromsampleachievea
tions. Theerasofpeoplespanfrom1950B.C.to 100%authenticity,whichindicatesthattheoverall
| 2001A.D. | | | | | | authenticityofBioKaLMAdatasetishigh. | | | | | | Therel- |
| -------- | --- | --- | --- | --- | --- | ------------------------------------ | --- | --- | --- | --- | --- | ------- |
evanceongeneralandspecificsettingsare73%and to the same word (e.g., “Chinese” the language
84%respectively. Thespecificquestionnormally and “Chinese” the ethnic group). When multiple
consists of more parts and include more details entities are retrieved from the graph for a given
than its general version, and hence some knowl- entityname,werankthegraphsbasedontheExact
edgearenecessarytothespecificversionbutnot Match (EM) between the neighboring nodes and
to the general version. However, the general ver- thequestion. Weselecttheentitywiththehighest
sionquestionssacrificerelevancetoachievebetter numberofmatchedneighboringnodes.
naturalnessandsignificance.
| | | | | | | | 3.3 Generation | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | -------------- | --- | --- | --- | --- | --- | --- |
Inpractice,itisdifficulttodefineaprecise“min-
imum knowledge set” for a question unless it is Thegenerationcomponenteffectivelypromptthe
LLMswiththeretrievedknowledgegraphs(KGs)
| very specific. | | However, | a very | specific | | question | | | | | | | |
| -------------- | --- | -------- | ------ | -------- | --- | -------- | --- | --- | --- | --- | --- | --- | --- |
togenerateanswersthatattributetheKG.Toadapt
| tendstobeartificial. | | Therelevanceandnaturalness | | | | | | | | | | | |
| -------------------- | --- | -------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
ofaquestionhaveatrade-offrelationship. Itisyet totheinputformatoftheLLMs,wetransformthe
challenging to generate questions that have both structured KGs into flat texts. We preserve the
| | | | | | | | information | | of the | retrieved | sub-graphs | | by map- |
| --- | --- | --- | --- | --- | --- | --- | ----------- | --- | ------ | --------- | ---------- | --- | ------- |
highrelevanceandhighnaturalness,butourgener-
ationmethodallowsforacontrolonthegranularity pingeachsub-graphtoasetoftriples. Eachtriple
ofaquestiononwhetherittendstobemorenatural consists of two nodes and one edge, where one
| | | | | | | | node is | the centered | | entity, | the other | node | is its |
| --- | --- | --- | --- | --- | --- | --- | ------- | ------------ | --- | ------- | --------- | ---- | ------ |
ormorerelevant.
neighbor,andtheedgerepresentstherelationship
| 3 Method | | | | | | | betweenthem. | | Forexample,[Q212657-placeof | | | | |
| -------- | --- | --- | --- | --- | --- | --- | ------------ | --- | --------------------------- | --- | --- | --- | --- |
birth-Q220]canbetranslatedto[ArtemisiaGen-
| We build | a baseline | to | enable | LLMs | to | generate | | | | | | | |
| --------------- | ---------- | ---------- | --------- | --------- | --------- | -------- | ---------------------------- | --------- | --- | --- | ------------------ | ---------- | ---- |
| | | | | | | | tileschi-placeofbirth-Rome]. | | | | Inthistranslation, | | |
| knowledge-aware | | attributed | | answers. | Following | | | | | | | | |
| | | | | | | | we use | the names | of | the | entities | for better | com- |
| the approach | of | many | retrieval | augmented | | gener- | | | | | | | |
prehensionbyboththemodelsandhumans,since
| ation works | (Lee | et al., | 2022; | Izacard | and | Grave, | | | | | | | |
| ----------- | ---- | ------- | ----- | ------- | --- | ------ | -------- | -------- | ---- | ------ | ----- | --- | --------- |
| | | | | | | | WikiData | utilizes | QIDs | (e.g., | Q220) | to | represent |
2021),weutilizeapipelineconsistingofthreecom-
| | | | | | | | uniqueentities. | | Weconstructaprompt(Table13 | | | | |
| -------- | ----------------------------------- | --- | --- | --- | --- | --- | --------------- | --- | -------------------------- | -------- | --- | ----------- | --- |
| ponents: | retrieval,re-ranking,andgeneration. | | | | | | | | | | | | |
| | | | | | | | in appendix | | D) which | includes | 1) | instruction | to |
3.1 Retrieval the models to generate attributed answers. 2) re-
| | | | | | | | trievedknowledgegraph,and3)thequestion. | | | | | | We |
| --- | --- | --- | --- | --- | --- | --- | --------------------------------------- | --- | --- | --- | --- | --- | --- |
Ourbaselineretrievalprocessconsistsoftwoparts:
employone-shotin-contextlearning(Brownetal.,
| namedentityrecognitionandgraphretrieval. | | | | | | We | | | | | | | |
| ---------------------------------------- | ---------------- | ------------------------------ | ----- | ----- | --------- | --- | --------- | ---------- | ------------ | --- | -------------- | ------- | ------ |
| | | | | | | | 2020) by | prepending | | one | human | written | demon- |
| utilizespaCy3 | | toidentifythenamedentitiesmen- | | | | | | | | | | | |
| | | | | | | | stration. | In | the one-shot | | demonstration, | | we use |
| tioned | in the question. | | Using | these | entities, | we | | | | | | | |
thespecialtoken[NA]torepresentthe“NotAppli-
retrieveentity-centeredsub-graphsusingSPARQL.
| | | | | | | | cableCitations”forconsciousincompetence. | | | | | | We |
| --------- | --------- | --------- | -------- | ------ | --- | -------- | ---------------------------------------- | ---- | ---- | --------- | --- | ------ | ------ |
| For each | retrieved | entity, | we | search | for | nodes in | | | | | | | |
| | | | | | | | deliberately | omit | some | knowledge | | in the | demon- |
| the graph | that | match the | entity’s | name. | | We use | | | | | | | |
strationexampleknowledgegraph,andweinsert
| the named | entity | recognition | | (NER) | entity | type | | | | | | | |
| --------- | ------ | ----------- | --- | ----- | ------ | ---- | ----------- | --- | ------ | ------------- | --- | --------- | ---- |
| | | | | | | | [NA] tokens | | in the | corresponding | | sentences | that |
asasimplefilter(e.g.,theNERcategory“person”
usetheseknowledgewithintheexampleanswer.
| matchesthe“human”entitytypeinWikiData). | | | | | | Tak- | | | | | | | |
| --------------------------------------- | --- | --- | --- | --- | --- | ---- | --- | --- | --- | --- | --- | --- | --- |
ing each selected node as the center, we retrieve 4 EvaluationMetrics
one-hopsub-graphsthatcontainpropertiesassoci-
| | | | | | | | Our benchmark | | includes | | evaluation | metrics | for |
| --- | --- | --- | --- | --- | --- | --- | ------------- | --- | -------- | --- | ---------- | ------- | --- |
atedwiththeentity.
| | | | | | | | boththegeneratedtextandcitations. | | | | | Wealsoevalu- | |
| -------------- | --- | --- | --- | --- | --- | --- | ------------------------------------------- | --- | ----------------------------- | --- | --- | ------------ | --- |
| 3.2 Re-ranking | | | | | | | atethealignmentbetweenthetextandcorrespond- | | | | | | |
| | | | | | | | ingcitations. | | Weprovidemorediscussionsonthe | | | | |
There-rankingcomponentplaysacrucialrolein
designofevaluationmetricsinsubsection4.5.
| disambiguating | | retrieved | entities, | as | multiple | en- | | | | | | | |
| -------------- | ---------- | --------- | --------- | --- | ------------ | ----- | ------------------ | -------- | --- | ---------------- | --- | --- | -------- |
| tities may | share | the same | name | in | the WikiData | | | | | | | | |
| | | | | | | | 4.1 TextEvaluation | | | | | | |
| graph. | Two common | scenarios | | are | different | indi- | | | | | | | |
| | | | | | | | Since our | test-set | has | no human-written | | | gold an- |
vidualswiththesamename(e.g.,AnneHathaway
swersasreferences,wedonotutilizecomparison-
theAmericanactressandAnneHathawaythewife
| | | | | | | | based metrics | | such | as BERTScore | | (Zhang | et al., |
| --- | --- | --- | --- | --- | --- | --- | ------------- | --- | ---- | ------------ | --- | ------ | ------- |
ofWilliamShakespeare)anddifferentreferences
| | | | | | | | 2019a) | or MAUVE | | (Pillutla | et | al., 2021). | In- |
| --- | --- | --- | --- | --- | --- | --- | ------ | -------- | --- | --------- | --- | ----------- | --- |
3https://spacy.io/api/entityrecognizer stead,weemployreference-freeNLGevaluatorG-
| Model Output: | | | Mininum | | | Model Output: | | | | | |
| ------------- | --- | --- | --------- | --- | --- | ------------- | --- | --- | --- | --- | --- |
Absent
| Sentence1 [k1][k2]. | | | Knowledge Set: | | | Sentence1 [k1][NA]. | | | | | |
| -------------------- | --- | --- | --------------- | --- | --- | -------------------- | --- | --- | --- | --- | --- |
Knowledge Set:
| Sentence2 [k2][k6][NA]. | | | [k1] [k2] | | | Sentence2 [k3][k6][NA]. | | | | | |
| ------------------------ | --- | --- | ---------- | --- | --- | ------------------------ | --- | --- | --- | --- | --- |
[k2] [k4] [k5]
| Sentence3 [k6][k9]. | | | [k3] [k4] [k5] | | | Sentence3 [k6][k9]. | | | | | |
| -------------------- | --- | --- | --------------- | --- | --- | -------------------- | --- | --- | --- | --- | --- |
| Citation Precision | | | Citation Recall | | | | | NLI | | | |
P:Sentence1 H:[k2] P:Sentence2 H:[k2]
| Sentence1 [k1][k2]. | | | | Knowlege: | | | | | | | |
| ------------------- | --- | --- | --- | --------- | --- | --- | --- | --- | --- | --- | --- |
P:Sentence1 H:[k4] P:Sentence2 H:[k4]
| Sentence2 [k2][k6][NA]. | | | | [k1] [k2] | | | | | | | |
| ----------------------- | --- | --- | --- | --------- | --- | --- | --- | --- | --- | --- | --- |
P:Sentence1 H:[k5] P:Sentence2 H:[k5]
| Sentence3 [k6][k9]. | | | | [k3] [k4] [k5] | | | | | | | |
| -------------------- | --- | --- | --- | -------------- | --- | -------------- | --- | --- | ----------- | --- | --- |
| | | | | | | [NA] Precision | | | [NA] Recall | | |
correct = 3, all = 6 hit = 2, all = 5 correct = 1， all = 2 hit = 1， all = 3
precision = 3/6 = 0.5 recall = 2/5 = 0.4 precision = 1/2 = 0.5 recall = 1/3 = 0.33
Figure2: Anillustrationofhowweevaluatethepreci- Figure3: Anillustrationofhowweevaluatethepreci-
sionandrecallforgeneratedcitations. sionandrecallforconsciousincompetence([NA])
Eval(Liuetal.,2023b),whichdefinesthefollow- knowledgetripletfromminimumknowledgesetof
ingfourmetrics: 1)Coherence: whetherthegen- thequestion. (SeeFigure2.)
| erated | text is well-structured | | and | well-organized. | | | | | | | |
| ------ | ----------------------- | --- | --- | --------------- | --- | --- | --- | --- | --- | --- | --- |
2)Consistency: whetherthegeneratedtextiscon- Recall We calculate citation recall for each
| | | | | | | knowledge | (0 | or 1) in minimum | knowledge | | set, |
| -------------------------------- | --- | --- | --- | --- | ---------- | --------- | --- | ---------------- | --------- | --- | ---- |
| sistentwiththeknowledgeprovided. | | | | | 3)Fluency: | | | | | | |
andaverageoverallknowledgetogetmicrorecall.
whetherthegeneratedtextiswell-writtenandgram-
matical. 4)Relevance: howwellisthegenerated Recall=1ifandonlyiftheknowledgeifhitbya
| | | | | | | correctcitation. | | (SeeFigure2.) | | | |
| --- | --- | --- | --- | --- | --- | ---------------- | --- | ------------- | --- | --- | --- |
textrelevanttothequestion.
Weaverageoverallcitations/knowledgeinanan-
| We use | the model | text-davinci-003 | | | for evalua- | | | | | | |
| ------ | --------- | ---------------- | --- | --- | ----------- | --- | --- | --- | --- | --- | --- |
swer,andaverageallanswer-levelprecision/recall
| tion, which | assigns | an integer | score | of | 1 to 5 for | | | | | | |
| ----------- | ------- | ---------- | ----- | --- | ---------- | --- | --- | --- | --- | --- | --- |
eachmetric. WefollowthepromptprovidedinG- to get macro precision and recall. we calculate
microandmacroF1-Scorefromcorrespondingpre-
Eval(Liuetal.,2023b)andcustomizeitbasedon
cisionandrecall.
| ourtask. | ThefullpromptsaregiveninappendixD. | | | | | | | | | | |
| ---------------------- | ---------------------------------- | --- | --- | --- | --- | -------------------------- | --- | --- | --- | --- | --- |
| 4.2 CitationEvaluation | | | | | | 4.3 Text-CitationAlignment | | | | | |
We evaluate the citation qualities from three as- Otherthanthetextqualityandcitationquality,we
pects: 1)Correctness,whichmeasureswhetherthe measure whether the generated citations provide
generatedknowledgematchesthegivenknowledge supportforthecorrespondingsentences. Apiece
from the knowledge graph, 2) Precision, which of useful knowledge is not an ideal citation if it
| | | | | | | isirrelevanttothesentenceitlinksto. | | | | Therefore, | |
| ---------- | -------- | --- | ------------- | --- | --------- | ----------------------------------- | --- | --- | --- | ---------- | --- |
| determines | how much | of | the generated | | citations | | | | | | |
arehelpfultoanswerthequestion,and3)Recall, we propose the metric “Alignment” which deter-
whichmeasureshowmuchoftheminimumknowl- mineswhetherthegeneratedcitationsarealigned
edgesetarecoveredbythegeneratedcitations. We to the sentences to which they belong. We use a
alsocalculatetheF1-ScorebasedonthePrecision state-of-the-art natural language inference (NLI)
andRecalltoreflecttheoverallqualityofcitations. model TRUE (Honovich et al., 2022), which is
| | | | | | | a fine-tuned | T5-11B | (Raffel | et al., | 2020) | model, |
| ----------- | --- | --------- | --- | -------- | -------- | ------------ | ------- | ------------- | -------- | ----- | ------- |
| Correctness | We | calculate | the | citation | correct- | | | | | | |
| | | | | | | to check | whether | the generated | sentence | | entails |
nessforeachcitation(0or1)andaverageoverall
| | | | | | | the generated | | citation. Since | one sentence | | could |
| ---------- | ------------- | --------- | --- | --- | ------------- | ------------- | --- | --------------- | ------------ | --- | ----- |
| citations. | Each citation | comprises | | a | triplet of 1) | | | | | | |
havemultiplecitations,werunNLIonallsentence-
centerentityQID,2)relation3)neighbourentity
| | | | | | | citation | pairs and | report | the percentage | of | entail- |
| --- | --- | --- | --- | --- | --- | -------- | --------- | ------ | -------------- | --- | ------- |
value. Ifthegeneratedcitationiscompletewithall
ment. Additionally,weconducthumanevaluation
threeparts,andexactlymatchesatripletfromthe
in§5.4toshowcaseiftheautomaticevaluationis
question’sretrievedKG,correctness=1.
correlatedwithhumanjudgments.
| Precision | We calculate | | citation | precision | for | | | | | | |
| --------- | ------------ | --- | -------- | --------- | --- | --- | --- | --- | --- | --- | --- |
4.4 ConsciousIncompetenceEvaluation
eachcitation(0or1)andaverageoverallcitations
togetmicroprecision. Precision=1foracitation Theoretically, each [NA] mark should map to
ifandonlyif1)itiscorrect,and2)itmatchesone a piece of knowledge absent from the retrieved
knowledgegraph. However,itisdifficulttoidentify cases,citationsdonotgiveadviceonthecor-
if sentence requires any absent knowledge since rectnessoftheanswer.
| thereisnogroundtruth. | | | Therefore,weconducta | | | | | | | | | |
| --------------------- | --- | --- | -------------------- | --- | --- | --- | ------------------------------- | --- | --- | --- | --- | ---------- |
| | | | | | | | Bothunderstandingsareplausible, | | | | | andhencewe |
three-roundexperimenttomanuallycreateground
truthforabsentknowledge. Inround1,weselect haveconsideredbothofthemwhenwedesignmet-
| | | | | | | | rics. The | alignment | score | is | designed | based on |
| --- | --- | --- | --- | --- | --- | --- | --------- | --------- | ----- | --- | -------- | -------- |
oneknowledgefromtheminimumknowledgeset,
| | | | | | | | the first | understanding, | | which | measures | whether |
| ---------- | --- | ---- | ---------------- | --- | --------- | --- | ---------------------------------------- | -------------- | --- | ----- | -------- | ------- |
| and remove | it | from | the ground-truth | | knowledge | | | | | | | |
| | | | | | | | thecitationsarecloselylinkedtotheanswer. | | | | | The |
graph. WelettheLLMsattributetothisincomplete
| | | | | | | | precision | and recall | are | designed | for | the second |
| --- | --- | --- | --- | --- | --- | --- | --------- | ---------- | --- | -------- | --- | ---------- |
knowledgegraphtogenerateanswers,wherebythe
understanding,wherethecitationsarecompletely
removedknowledgeformsthe“absentknowledge
decoupledfromtheanswer,andarecorrectifthey
| ground | truth”. | In subsequent | | rounds, | we | each re- | | | | | | |
| ------ | ------- | ------------- | --- | ------- | --- | -------- | --- | --- | --- | --- | --- | --- |
moveoneadditionalknowledgefromtheminimum providesupportforthequestion.
| | | | | | | | In addition, | we | also | incorporate | | an edge case |
| --- | --- | --- | --- | --- | --- | --- | ------------ | --- | ---- | ----------- | --- | ------------ |
knowledgeset,simulatingaknowledgegraphwith
| | | | | | | | fordesignofthe[NA]precisioncalculation. | | | | | Ifan |
| --- | --- | --- | --- | --- | --- | --- | --------------------------------------- | --- | --- | --- | --- | ---- |
moreseriouscoverageproblem.
NA-markedsentencedoesnotanswerthequestion
| We employ | | the NLI | model | TRUE | (Honovich | | | | | | | |
| --------- | --- | ------- | ----- | ---- | --------- | --- | --- | --- | --- | --- | --- | --- |
atall,itisconsideredcorrectinthe[NA]precision
| et al., 2022) | to | measure | the | alignment | | between | | | | | | |
| ---------------------- | ---------- | ------- | --------- | ----------------- | --- | ------- | ----------------- | ------------------------------- | -------- | --- | ------- | ------------- |
| | | | | | | | calculation. | Inthiscase,theLLMcorrectlyiden- | | | | |
| sentencesandknowledge. | | | | Asentencewith[NA] | | | | | | | | |
| | | | | | | | tifies a sentence | that | requires | | further | verification. |
| should | be aligned | to | an absent | knowledge. | | We | | | | | | |
calculateprecisionandrecallfor[NA].
| | | | | | | | 5 Experiments | | | | | |
| ------------- | --- | --------------------------- | --- | --- | --- | --- | ------------- | --- | --- | --- | --- | --- |
| [NA]precision | | Wecalculate[NA]precisionfor | | | | | | | | | | |
eachsentencewith[NA](0or1)andaverageover Werunthroughthemethodpipelinedescribedin
| all sentences | with | [NA]. | Precision | | = 1 for | a sen- | | | | | | |
| ------------- | ---- | ----- | --------- | --- | ------- | ------ | ---------------- | ---- | --- | ----------- | --- | -------------- |
| | | | | | | | § 3 on different | LLMs | | and present | | the results in |
tenceifandonlyifitentailsoneknowledgetriplet
| | | | | | | | thissection. | Sinceweaimtoobtainamoreaccu- | | | | |
| ----------- | --------- | --- | --- | ------ | --------- | ---- | ------------ | ---------------------------- | --- | --- | --- | --- |
| from absent | knowledge | | set | of the | question. | (See | | | | | | |
rateevaluation,weconductourmainexperiments
| Figure3.) | | | | | | | on the specific | questions | | setting, | since | the mini- |
| --------- | --- | --- | --- | --- | --- | --- | --------------- | --------- | --- | -------- | ----- | --------- |
mumknowledgesethasahigherrelevanceonthe
| [NA] Recall | | We calculate | | [NA] | recall | for each | | | | | | |
| ----------- | ----- | ------------ | --------- | --------- | ------ | -------- | ------------------ | --- | ------------------------- | --- | --- | --- |
| | | | | | | | specificquestions. | | However,wewillalsoprovide | | | |
| knowledge | (0 or | 1) | in absent | knowledge | | set and | | | | | | |
evaluationresultsforthegeneralquestionsin§5.5
| average | over all | absent | knowledge. | | Recall | = 1 if | | | | | | |
| ------- | -------- | ------ | ---------- | --- | ------ | ------ | ----------- | -------- | --- | -------------- | --- | ------- |
| | | | | | | | as ablation | studies. | The | implementation | | details |
andonlyiftheknowledgeifentailedbyasentence
arereportedinappendixB.Wereportfivemodel
with[NA].(SeeFigure3.)
baselinesfrombothopenandclosedsourcemodel
families:
4.5 DiscussiononEvaluationmetrics
Inthissection,wediscussontheevaluationmetrics OpenAIModels WeuseGPT4(gpt-4-0314)and
ofbenchmarkBioBaLMA.Wedesigntheevalua- ChatGPT(gpt-3.5-turbo-0301)forourexperiments.
tionmetricsfrommultipledimensionstoincorpo- ForChatGPT,weexperimentontemperatureof0.1,
ratedifferentunderstandingsonwhatmakesahigh
0.5,and0.9toobtaindifferentlevelsofrandomness
| qualitycitation. | | | | | | | andcreativityingeneration. | | | | | |
| ---------------- | ------------- | --- | ------ | ---- | --- | ------ | -------------------------- | ------------------------------ | --- | --- | --- | --- |
| | | | | | | | LLaMA | WeconductexperimentswithLLaMA- | | | | |
| • One | understanding | | argues | when | the | answer | | | | | | |
7B(Touvronetal.,2023)andLLaMA-13Bsince
containsmistakes,evenifthecitationiscor-
rectlyansweringthequestions,itcannotrep- they are powerful open-source models that are
resentgoodLLMattributionperformance. In widely accessible. We have also conducted hu-
| | | | | | | | man instruction | | tuned | LLaMA | models, | includ- |
| ---- | ----- | -------- | ------- | --- | ---------- | ---- | --------------- | --- | ----- | ----- | ------- | ------- |
| this | case, | citation | quality | is | considered | as a | | | | | | |
measure of overall attribution performance, ing Alpaca-7B (Taori et al., 2023) and Vicuna-
| includingtheanswerquality. | | | | | | | 13B(Chiangetal.,2023). | | | | | |
| -------------------------- | --- | --- | --- | --- | --- | --- | ---------------------- | --- | --- | --- | --- | --- |
| | | | | | | | 5.1 MainResults | | | | | |
• Theotherunderstandingarguesforacomplete
decouplingofanswerandcitationquality. In Citation Quality Evaluation We present the
this scenario, even if the answer is wrong, mainresultsinTable3. Forcorrectness,wereport
thecitationisvaluableaslongasitprovides on a micro scale. For precision, recall, and F1-
reasonablesupportforthequestion. Insuch Score,wereportonbothmicroandmacroscales.
| | | | | | Micro | | | | | Macro | | |
| ------------ | --- | ------ | ----- | ----- | ----- | ----- | ---- | ---- | ----- | ----- | ----- | ---- |
| Model | | Align. | Corr. | | Prec. | Rec. | F1. | | Prec. | | Rec. | F1. |
| GPT-4(0.5) | | 92.0 | 97.6 | | 36.0 | 43.6 | 39.4 | 40.7 | | 43.9 | | 42.3 |
| | | | (1.5) | (0.1) | (0.6) | (1.0) | | | (1.1) | | (1.0) | |
| ChatGPT(0.1) | | 85.9 | 96.1 | | 29.0 | 50.8 | 36.9 | 32.7 | | 51.2 | | 39.9 |
| | | | (2.5) | (0.4) | (0.0) | (0.3) | | | (0.4) | | (0.3) | |
| ChatGPT(0.5) | | 84.5 | 94.8 | | 29.9 | 49.0 | 37.2 | 34.1 | | 49.4 | | 40.4 |
| | | | (1.1) | (0.2) | (0.2) | (0.8) | | | (0.5) | | (0.9) | |
| ChatGPT(0.9) | | 84.1 | 94.2 | | 28.7 | 49.0 | 36.2 | 32.5 | | 49.4 | | 39.2 |
| | | | (0.5) | (0.4) | (0.2) | (0.3) | | | (0.2) | | (0.3) | |
| Alpaca-7B | | 46.9 | 78.9 | | 14.9 | 19.4 | 16.8 | 19.8 | | 19.9 | | 19.8 |
| | | | (0.9) | (0.6) | (1.4) | (0.2) | | | (0.4) | | (0.3) | |
| LLaMA-7B | | 47.8 | 70.2 | | 7.7 | 41.1 | 13.0 | 11.0 | | 41.4 | | 17.4 |
| | | | (0.8) | (0.2) | (2.4) | (0.7) | | | (1.9) | | (0.7) | |
| LLaMA-13B | | 62.1 | 71.7 | | 10.5 | 43.7 | 16.9 | 13.8 | | 43.5 | | 20.9 |
| | | | (0.4) | (1.9) | (3.3) | (1.0) | | | (2.2) | | (1.0) | |
| Vicuna-13B | | 66.9 | 59.0 | | 14.9 | 16.8 | 15.8 | 15.1 | | 17.0 | | 16.0 |
| | | | (0.1) | (0.6) | (0.2) | (0.0) | | | (0.0) | | (0.0) | |
Table3: CitationQualityOpenAImodelsandLLaMAfamilymodels. ThefirstfivemetricsarereportedinMicro,
andthelastthreemetricsarereportedinMacro. Wealsoreporttextcitationalignment.
Theexperimentalresultsarethemeanofthreeruns, Model Coh. Con. Flu. Rel.
andthestandarddeviationisreportedinbrackets. GPT-4(0.5) 4.48 4.89 4.64 4.72
| | | | | | | ChatGPT(0.1) | | 4.57 | | 4.94 | 4.69 | 4.83 |
| --- | --- | --- | --- | --- | --- | ------------ | --- | ---- | --- | ---- | ---- | ---- |
Ingeneral,thereisaroomofimprovementfor
allmodelssincenomodelcanachieveamicroF1 ChatGPT(0.5) 4.57 4.94 4.71 4.81
Scoreofhigherthan40. TheOpenAImodelsout- ChatGPT(0.9) 4.52 4.91 4.67 4.79
perform the LLaMA family models in almost all Alpaca-7B 4.10 4.46 4.23 3.76
metrics. Thecorrectnessisabove94forOpenAI LLaMa-7B 3.06 3.79 3.62 2.96
models,butaround70forLLaMAbasedmodels. LLaMa-13B 3.60 4.23 3.94 3.56
ForChatGPT,temperaturedoesnotplayasignifi- Vicuna-13B 3.67 4.50 3.96 3.64
| | | | | | | Table4: | Evaluationongeneratedtextquality. | | | | | |
| --- | --- | --- | --- | --- | --- | ------- | --------------------------------- | --- | --- | --- | --- | --- |
cantrolesinceiteffectonF1Scoreisatmost1.2.
| The GPT-4 | model | achieves | the best | performance | | | | | | | | |
| --------- | ----- | -------- | -------- | ----------- | --- | --- | --- | --- | --- | --- | --- | --- |
across almost all metrics, except for recall, since Removed Corr. Prec. Rec. F1.
| | | | | | | 0(gold) | 95.5 | | 30.1 | 57.1 | 39.4 | |
| ----- | ------ | ------- | ---------------- | --- | ------- | ------- | ---- | --- | ---- | ---- | ---- | --- |
| GPT-4 | models | tend to | generate shorter | | answers | | | | | | | |
withfewercitations,resultinginhigherprecision. 1 94.1 26.1 42.5 32.3
While LLaMA is better at Recall by generating 2 94.0 21.0 31.4 25.2
| | | | | | | 3 | 93.9 | | 16.3 | 20.4 | 18.1 | |
| ----------------------------- | --- | --- | ------------- | --- | --- | ------- | ------------------------------------------ | --- | ---- | ---- | ---- | --- |
| longanswerswithmanycitations. | | | TheF1-Scoreof | | | | | | | | | |
| | | | | | | Table5: | Citationqualityevaluationforgeneratedtexts | | | | | |
modelsfromthesamefamilyareclosetoonean-
other,showingthatourautomaticevaluationmetric usingaKGwithNpiecesofknowledgeremoved.
designedisreliable.
| | | | | | | are capable | of | generating | | answers | that | are not |
| ---------------------- | -------- | --- | ------------------ | ---- | ---- | -------------- | -------------------------------- | ---------- | -------- | --------- | ---- | -------- |
| Text-CitationAlignment | | | FromTable3,similar | | | | | | | | | |
| | | | | | | contradictory | to | the | provided | knowledge | | or self- |
| to citation | quality, | the | OpenAI models | also | out- | | | | | | | |
| | | | | | | contradictory. | However,therelevanceisrelatively | | | | | |
performtheLLaMAbasedmodelsontext-citation
| | | | | | | low for | smaller | models, | indicating | | the | difficulty |
| ---------- | -------------------------------- | -------- | ------------ | ---------- | --- | ------------ | ------- | ------- | ---------- | --- | ------- | ---------- |
| alignment. | Inaddition,modelswith7B,13B,175B | | | | | | | | | | | |
| | | | | | | these models | face | in | generating | | answers | that are |
| (ChatGPT), | and | trillion | level (GPT4) | parameters | | | | | | | | |
relevanttothequestions.
haveanalignmentscoreof40+,60+,80+,and92
| respectively. | LLaMA-13Bmodelhasanimprove- | | | | | | | | | | | |
| ---------------------------------- | --------------------------- | --- | --- | --- | ---- | ------------------------- | --- | --- | --- | --- | --- | --- |
| | | | | | | 5.2 ConsciousIncompetence | | | | | | |
| mentof14.3comparedtoLLaMA-7Bmodel. | | | | | This | | | | | | | |
showsthatparametersizemayplayanimportant Wefirstevaluatecitationqualityofthegenerated
| | | | | | | text with | knowledge | | removed | using | | method de- |
| --- | --- | --- | --- | --- | --- | --------- | --------- | --- | ------- | ----- | --- | ---------- |
roleingeneratingsentencesandcitationswithgood
| | | | | | | scribed | in § 4.4. | From | Table | 5, | the | removal of |
| --- | --- | --- | --- | --- | --- | ------- | --------- | ---- | ----- | --- | --- | ---------- |
alignment.
requiredknowledgehasaminimalimpactoncor-
rectness,butsignificantlyaffectscitationprecision
| TextQualityEvaluation | | | Wepresenttheevalu- | | | | | | | | | |
| --------------------- | --- | --- | ------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- |
ation of generated text quality in Table 4. From andrecall. Withmoreknowledgeabsentfrompro-
theresults,wefindthatOpenAImodels,ingeneral, videdknowledgegraph,bothprecisionandrecall
have better text quality in all metrics compared dropsdrastically,demonstratingthatthecoverage
to LLaMA family models, which corresponds to issueposesaconsiderablechallengetogenerating
thecitationevaluationresults. Allmodelsexhibit answerswithhighqualitycitations.
ratherhighconsistency,indicatingthattheLLMs Next, we evaluate [NA] precision and recall.
Experiment Result on Conscious Incompetence Retrieval Analysis
precision
| 24 | recall | | | | | | | | | | |
| --- | ------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
F1_Score
80
22
precision
| erocS 20 | | | | | erocS 60 | | | | | | |
| -------- | --- | --- | --- | --- | -------- | --- | --- | --- | --- | --- | --- |
recall
F1_Score
| 18 | | | | | | | | | | correctness | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | ----------- | --- |
40
16
20
14
| | one | two | | three | | gold | 80 | 60 | | 40 20 | |
| --- | --------------------------- | --- | --- | ----- | --- | ---- | --- | ------------------ | --- | ----- | --- |
| | number of knowledge removed | | | | | | | Retrieval Accuracy | | | |
Figure4: Precision,Recall,andF1-Scorefor[NA]. Figure5: Citationevaluation(Micro)ofgeneratedtexts
| | | | | | using | knowledge | graphs | with | retrieval | accuracy | 100 |
| --- | --- | --- | --- | --- | ----- | --------- | ------ | ---- | --------- | -------- | --- |
(gold),80,60,40,and20.
| From Figure | 4, The | recall is | stable | at about 15 | | | | | | | |
| ----------- | ------ | --------- | ------ | ----------- | --- | --- | --- | --- | --- | --- | --- |
regardless of the number of absent knowledge. Alignment HumanAvg.
ThisindicatesthatthecurrentLLMs haveability ChatGPT(0.5) 84.5 82.0
to identify absent knowledge to a limited extent. LLaMA-7B 47.8 45.5
While precision and F1-Score exhibit a clear up- Vicuna-13B 66.9 64.5
ward trend, which shows that with more absent Table6: ResultofHumanEvaluationontext-citation
| knowledge | in KG, | [NA] enables | generated | out- | alignment | | | | | | |
| --------- | ------------- | ------------ | --------- | ----------- | --------- | --- | --- | --- | --- | --- | --- |
| puts to | locate absent | knowledge | more | accurately. | | | | | | | |
Therefore,the“ConsciousIncompetence”setting
| | | | | | (42.5), | demonstrating | | the | potential | of the | model |
| --- | --- | --- | --- | --- | ------- | ------------- | --- | --- | --------- | ------ | ----- |
playsanincreasinglycrucialrolewhenthecover-
togeneratehigh-qualitycitationsunderperfectre-
ageproblemofknowledgegraphismoreserious.
| | | | | | trievalconditions. | | | Inpractice,aretrievalaccuracy | | | |
| --- | --- | --- | --- | --- | ------------------ | --- | --- | ----------------------------- | --- | --- | --- |
of80isclosesttotheactualscenarioofourexper-
5.3 RetrievalAnalysis
| | | | | | iment | (our retrieval | | accuracy | is | 75.9). Therefore, | |
| --- | --- | --- | --- | --- | ----- | -------------- | --- | -------- | --- | ----------------- | --- |
Weconductanablationstudytoexaminetheimpact whenretrievalaccuracyisreasonablyhigh,thecor-
ofretrievalaccuracyonthemodel’soutput. Theex- rectnessofcitationsisnotthemostsignificantcon-
perimentsimulatesretrievalaccuracyfrom100to cerncomparedtorecall.
| 20atintervalsof20. | | Westartwiththegroundtruth | | | | | | | | | |
| ------------------ | --- | ------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
5.4 HumanEvaluation
| knowledge | graphs | that we used | for | question con- | | | | | | | |
| --------- | ------ | ------------ | --- | ------------- | --- | --- | --- | --- | --- | --- | --- |
struction. Ineachsubsequentrounds,werandomly Weconducthumanevaluationtoverifythecorrela-
replaceadditional20%knowledgegraphswithir- tionbetweenautomaticevaluationandhumanjudg-
relevant knowledge graphs to simulate retrieving ment. Werandomlysample100sentence-citation
wronggraphs. Theresultsforcitationqualityare pairs from each of the three baselines: ChatGPT
inFigure5. AnswersaregeneratedusingChatGPT (temperature 0.5), LLaMA-7B, and Vicuna-13B.
withatemperatureof0.5.
| | | | | | We request | two | proficient | | English | annotators | for |
| --- | --- | --- | --- | --- | ---------- | --- | ---------- | --- | ------- | ---------- | --- |
The results show clear downward trends in eachbaselinetodetermineifthecitationalignsto
all metrics as expected when retrieval accuracy thesentenceandprovidessupportforit. Thereason
dropped. Amongprecisionandrecall,theimpact wechoosemetricalignmenthereisinappendixC,
ofpoorretrievalqualityonrecall(green)ismuch withinstructiontoannotatorsandIAA.
moresignificantthanonprecision(yellow). This The comparison between automatically calcu-
indicatesthatthemodelhastheabilitytofilterout lated Alignment and human evaluation results is
incorrectknowledgetoacertainextent,resulting showninTable6. Forallthreebaselines,theauto-
in less noticeable impact on precision compared maticandhumanscoresareclosewithagapwithin
torecall. Thereductioninrecallwasnearlylinear 2.5,despitethesignificantdifferencesamongthe
as retrieval accuracy decreased, which is under- baselines. This indicates a strong correlation be-
standablesinceaknowledgecannotbecitedifitis tweentheautomaticallycalculatedalignmentand
notprovided. Thegreatestdropinrecalloccurred humanjudgments. Theexperimentresultsdemon-
between the ground truth (57.1) and 80 accuracy stratethattheautomaticevaluationservesasareli-
| | | | CitationEval. | | | | TextEval. | | | |
| --- | --- | --- | ------------- | --- | --- | --- | --------- | --- | --- | --- |
Setting Model Align. Corr. Prec. Rec. F1. Coh. Con. Flu. Rel.
General GPT-4(0.5) 90.9 97.6 30.8 42.1 35.6 4.38 4.77 4.48 4.48
| | ChatGPT(0.5) | 82.7 | 94.5 | 25.2 47.4 | 32.9 | 4.64 | 4.89 | 4.45 | 4.70 | |
| --- | ------------ | ---- | ---- | --------- | ---- | ---- | ---- | ---- | ---- | --- |
Specific GPT-4(0.5) 92.0 97.6 36.0 43.6 39.4 4.48 4.89 4.64 4.72
| | ChatGPT(0.5) | 84.5 | 94.8 | 29.9 49.0 | 37.2 | 4.57 | 4.94 | 4.71 | 4.81 | |
| --- | ------------ | ---- | ---- | --------- | ---- | ---- | ---- | ---- | ---- | --- |
Table7: ComparisonofevaluationresultsonGeneralandSpecificquestionsetting
ablemeasurementofthealignmentbetweengener- documentsbutnotknowledgegraph.
atedtextsandcitations.
| | | | | Evaluation | (Rashkinetal.,2021)definethe“At- | | | | | |
| --- | --- | --- | --- | ---------- | -------------------------------- | --- | --- | --- | --- | --- |
5.5 GeneralandSpecificQuestions
| | | | | tributable | to Identified | | Sources” | (AIS) | | to mea- |
| --- | --- | --- | --- | ---------- | ------------- | --- | -------- | ----- | --- | ------- |
We compare experiments results of text, citation surewhethermodel-generatedstatementsaresup-
(micro), and alignment between the general and portedbyunderlyingsources. (Bohnetetal.,2022)
specificquestionsinTable7. Theresultsshowthat studyanautomaticmetric(AutoAIS)thatformu-
the same model’s answers on specific questions lates evaluation of automated question answer-
outperformthoseongeneralquestionsinalmostall ing as a NLI task. (Yue et al., 2023) investigate
metrics. Thefindingisnotsurprisingbecausethe theautomaticevaluationofattributionbyprompt-
specific questions provide clearer instructions to ing LLMs and fine-tuning smaller LMs. (Liu
themodelsonwhichknowledgetouse. Inaddition, et al., 2023a) conduct human evaluation to audit
thegeneralquestionsinthedatasetareinherently generative search engines for their citation quali-
looselybondedtotheminimumknowledgeset,and ties. ALCE(Gaoetal.,2023)evaluatesgenerated
hencehaveimpactsontheevaluationresults. This answers by comparing with gold answers using
experimentshowsatrade-offbetweenhowexplic-
| | | | | MAUVE, | and calculates | | precision | and | recall | for |
| --- | --- | --- | --- | ------ | -------------- | --- | --------- | --- | ------ | --- |
itlythequestioncontextmentionstheknowledge, citationsusingNLI.Tothebestofourknowledge,
and how irreplaceably the knowledge is required ourevaluationmethodsarethefirstframeworkthat
bythequestion. Thespecificquestionstargetthe requiresnohumanannotateddata.
knowledgemoreexplicitlyinthequestioncontext,
andhencecoverthescopeoftheparagraphbetter.
7 Conclusion
Itstandsforanupperboundforknowledgecover-
ageandalowerboundforquestionnaturalness.The
WeproposeKaLMAthatcomprisesanewdataset
generalquestionsimplicitlytargettheknowledge
| | | | | BioKaLMA, | a | pipeline | for generating | | attributed | |
| --------------- | ------------ | ----- | ------------- | --------- | ------------- | -------- | -------------- | --- | ---------- | ------ |
| in the question | context, and | there | loosely cover | | | | | | | |
| | | | | answers | by retrieving | | from KGs, | and | a set | of au- |
thescopeoftheparagraph. Itstandsforanupper tomatic evaluation metrics to assess text quality,
boundforquestionnaturalnessandalowerbound
| | | | | citation | quality, | and text-citation | | alignment. | | We |
| --- | --- | --- | --- | -------- | -------- | ----------------- | --- | ---------- | --- | --- |
forknowledgecoverage.
| | | | | introduce | the “Conscious | | Incompetence” | | | setting, |
| --- | --- | --- | --- | --------- | -------------- | --- | ------------- | --- | --- | -------- |
enablingLLMstoidentifytheknowledgerequired
6 RelatedWork
tosupporttheanswersbutisabsentfromtheKG.
Retrieval-augmented LLMs KiC (Pan et al., Through this benchmark, we address three chal-
2022)empowermodelswithexternalmemoryof lenges: incorporating diverse attribution sources,
| | | | | limited | attribution | source | coverage, | | and | the ab- |
| ---------------- | --------- | --------- | ----- | ------- | ----------- | ------ | --------- | --- | --- | ------- |
| multiple formats | including | knowledge | graph | but | | | | | | |
does not explore attribution. WebGPT (Nakano sence of human annotated ground truth for auto-
| etal.,2021)outsourcesdocumentretrievaltoMi- | | | | maticevaluation. | | | | | | |
| ------------------------------------------- | --- | --- | --- | ---------------- | --- | --- | --- | --- | --- | --- |
crosoftBingandfine-tunesGPT3toanswerques- Ourextensiveexperimentalresultsdemonstrate
tions. GopherCite(Menicketal.,2022)fine-tunes that current LLMs still have room for improve-
aGopher(Raeetal.,2021)modeltogeneratetext ment when utilizing KGs as attribution sources.
alongside quotes extracted from Google search. We also highlight the increasing effectiveness of
ALCE(Gaoetal.,2023)retrievestop-kpassages “ConsciousIncompetence”settingasthecoverage
fromWikipediaandasksLLMstogenerateoutputs of attribution source becomes worse. Lastly, we
with citations to corresponding supporting docu- provethecrucialroleofretrievalaccuracyingen-
ments. TheseworksattributeLLMstounstructured eratinghigh-qualityattributedtexts.
Limitations
| | | | | | learners. | Advancesinneuralinformationprocessing | | | | |
| --- | --- | --- | --- | --- | --------- | ------------------------------------- | --- | --- | --- | --- |
systems,33:1877–1901.
Onelimitationisthatourworkonlyinvestigatesa
| | | | | | Wei-Lin Chiang, | | Zhuohan | Li, Zi | Lin, Ying | Sheng, |
| --- | --- | --- | --- | --- | --------------- | --- | ------- | ------ | --------- | ------ |
simpleformofknowledgegraph,whereeachnode
ZhanghaoWu,HaoZhang,LianminZheng,Siyuan
| is an entity, | and | each sub-graph | | is a knowledge | | | | | | |
| ------------- | --- | -------------- | --- | -------------- | --- | --- | --- | --- | --- | --- |
Zhuang,YonghaoZhuang,JosephE.Gonzalez,Ion
triple. Therearemorecomplicatedformsofknowl-
| | | | | | Stoica, and | Eric | P. Xing. | 2023. | Vicuna: | An open- |
| --- | --- | --- | --- | --- | ----------- | ---- | -------- | ----- | ------- | -------- |
edge graph, where each node is a document. We sourcechatbotimpressinggpt-4with90%*chatgpt
| willexplorethissettinginfutureworks. | | | | | quality. | | | | | |
| ------------------------------------ | ------------------------------ | ---- | ------ | ---------------- | --------- | ------ | ----------- | ----- | -------- | ------ |
| Another | limitation | lies | within | the text quality | | | | | | |
| | | | | | Elizabeth | Clark, | Tal August, | Sofia | Serrano, | Nikita |
| evaluation. | WeusesChatGPTasthemodeltoeval- | | | | | | | | | |
Haduong,SuchinGururangan,andNoahA.Smith.
uate texts, which could potentially have a bias if 2021. All that’s ‘human’ is not gold: Evaluating
themodelprefersthetextstylegeneratedbyitself. humanevaluationofgeneratedtext. InProceedings
| | | | | | of the 59th | Annual | Meeting | of | the Association | for |
| --- | --- | --- | --- | --- | ----------- | ------ | ------- | --- | --------------- | --- |
Suchbiascanbeobservedfromtheabnormalphe-
ComputationalLinguisticsandthe11thInternational
nomenon that the scores of ChatGPT generated JointConferenceonNaturalLanguageProcessing
| answers | are higher | than | that of | the GPT4 gener- | | | | | | |
| ------- | ---------- | ---- | ------- | --------------- | --------- | ---------------------------------- | --- | --- | --- | --- |
| | | | | | (Volume1: | LongPapers),pages7282–7296,Online. | | | | |
atedanswersforallfourdimensions. Duetocost AssociationforComputationalLinguistics.
| considerations, | | we do not | repeat | the text quality | | | | | | |
| --------------- | --- | --------- | ------ | ---------------- | -------------- | --- | ------- | --------- | ----- | ------- |
| | | | | | Paul R Curtiss | and | Phillip | W Warren. | 1974. | The dy- |
evaluationwithGPT-4. namicsoflifeskillscoaching.lifeskillsseries.
EthicalConsiderations Ido Dagan, Oren Glickman, and Bernardo Magnini.
2005. Thepascalrecognisingtextualentailmentchal-
The potential risk is when users leverage the au- lenge. In Machine learning challenges workshop,
tomatic dataset construction pipeline to generate pages177–190.Springer.
| massivehazardousdatasets. | | | Thiscanonlyhappen | | | | | | | |
| ------------------------- | --- | --- | ----------------- | --- | --- | --- | --- | --- | --- | --- |
TianyuGao,HowardYen,JiatongYu,andDanqiChen.
whenastructuredknowledgeofharmfulcontentis 2023. Enablinglargelanguagemodelstogenerate
| available. | Otherwisethereisnoriskaslongasthe | | | | | | | | | |
| ---------- | --------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
textwithcitations.
| benchmarkisusedcorrectly. | | | Alldataarecollected | | | | | | | |
| ------------------------------------- | --- | --- | ------------------- | ----- | --------------- | ---- | -------- | ---------- | ----- | ----------- |
| | | | | | Or Honovich, | Roee | Aharoni, | Jonathan | | Herzig, Ha- |
| fromWikiDatawhichispubliclyavailable. | | | | Hence | | | | | | |
| | | | | | gai Taitelbaum, | | Doron | Kukliansy, | Vered | Cohen, |
thereisnoprivacyissue. Wealsoconducthuman Thomas Scialom, Idan Szpektor, Avinatan Has-
checktoensurethereisnooffensivecontent. sidim,andYossiMatias.2022. True: Re-evaluating
| | | | | | factual | consistency | evaluation. | | arXiv | preprint |
| --- | --- | --- | --- | --- | ------- | ----------- | ----------- | --- | ----- | -------- |
arXiv:2204.04991.
References
| | | | | | GautierIzacardandEdouardGrave.2021. | | | | | Leveraging |
| --- | --- | --- | --- | --- | ----------------------------------- | --- | --- | --- | --- | ---------- |
Bernd Bohnet, Vinh Q. Tran, Pat Verga, Roee Aha- passageretrievalwithgenerativemodelsforopendo-
| | | | | | mainquestionanswering. | | | InProceedingsofthe16th | | |
| --- | --- | --- | --- | --- | ---------------------- | --- | --- | ---------------------- | --- | --- |
roni,DanielAndor,LivioBaldiniSoares,Massimil-
ConferenceoftheEuropeanChapteroftheAssoci-
ianoCiaramita,JacobEisenstein,KuzmanGanchev,
JonathanHerzig,KaiHui,TomKwiatkowski,JiMa, ationforComputationalLinguistics: MainVolume,
Jianmo Ni, Lierni Sestorain Saralegui, Tal Schus- pages 874–880, Online. Association for Computa-
| ter, William | W. | Cohen, Michael | | Collins, Dipanjan | tionalLinguistics. | | | | | |
| ------------ | --- | -------------- | --- | ----------------- | ------------------ | --- | --- | --- | --- | --- |
Das,DonaldMetzler,SlavPetrov,andKellieWebster.
ZiweiJi,NayeonLee,RitaFrieske,TiezhengYu,Dan
2023. Attributedquestionanswering:Evaluationand
modelingforattributedlargelanguagemodels. Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea
| | | | | | Madotto,andPascaleFung.2023. | | | | Surveyofhalluci- | |
| --- | --- | --- | --- | --- | ---------------------------- | --- | --- | --- | ---------------- | --- |
BerndBohnet,VinhQTran,PatVerga,RoeeAharoni, nationinnaturallanguagegeneration. ACMComput-
DanielAndor,LivioBaldiniSoares,JacobEisenstein, ingSurveys,55(12):1–38.
| KuzmanGanchev, | | JonathanHerzig, | | KaiHui, etal. | | | | | | |
| -------------- | --- | --------------- | --- | ------------- | --- | --- | --- | --- | --- | --- |
TusharKhot,AshishSabharwal,andPeterClark.2018.
2022. Attributedquestionanswering:Evaluationand
| | | | | | Scitail: | A textual | entailment | dataset | from | science |
| ----------------------------------------- | --- | --- | --- | ----- | ------------------ | --------- | ---------- | ------- | ---- | ------- |
| modelingforattributedlargelanguagemodels. | | | | arXiv | | | | | | |
| preprintarXiv:2212.08037. | | | | | questionanswering. | | volume32. | | | |
SamuelBowman,GaborAngeli,ChristopherPotts,and HaejunLee,AkhilKedia,JongwonLee,AshwinParan-
Christopher D Manning. 2015. A large annotated jape,ChristopherD.Manning,andKyoung-GuWoo.
corpusforlearningnaturallanguageinference. 2022. You only need one model for open-domain
questionanswering.
| Tom Brown, | Benjamin | Mann, | Nick | Ryder, Melanie | | | | | | |
| ---------- | -------- | ----- | ---- | -------------- | --- | --- | --- | --- | --- | --- |
Subbiah,JaredDKaplan,PrafullaDhariwal,Arvind NelsonFLiu,TianyiZhang,andPercyLiang.2023a.
Neelakantan,PranavShyam,GirishSastry,Amanda Evaluatingverifiabilityingenerativesearchengines.
Askell,etal.2020. Languagemodelsarefew-shot arXivpreprintarXiv:2304.09848.
Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Reitter. 2021. Measuring attribution in natu-
RuochenXu, andChenguangZhu.2023b. G-eval: ral language generation models. arXiv preprint
| Nlgevaluationusinggpt-4withbetterhumanalign- | | | | | | | arXiv:2112.12870. | | | | | |
| -------------------------------------------- | --- | --- | --- | --- | --- | --- | ----------------- | --- | --- | --- | --- | --- |
ment.
RevanthGangiReddy,YiRFung,QiZeng,ManlingLi,
| Jacob Menick, | | Maja | Trebacz, | Vladimir | Mikulik, | | | | | | | |
| ------------- | --- | ---- | -------- | -------- | -------- | --- | -------------------------------- | --- | --- | --- | ------------- | --- |
| | | | | | | | ZiqiWang,PaulSullivan,etal.2023. | | | | Smartbook:Ai- | |
John Aslanides, Francis Song, Martin Chadwick, assistedsituationreportgeneration. arXivpreprint
| Mia Glaese, | | Susannah | Young, | | Lucy Campbell- | | arXiv:2303.14337. | | | | | |
| ------------------------------------ | ------ | -------- | ------- | ------- | -------------- | --- | ----------------- | --- | --- | --- | --- | --- |
| Gillingham,GeoffreyIrving,etal.2022. | | | | | Teaching | | | | | | | |
| language | models | to | support | answers | with verified | | | | | | | |
MicheleSalvagno,FabioSilvioTaccone,AlbertoGio-
| quotes. | arXivpreprintarXiv:2203.11147. | | | | | | | | | | | |
| ------- | ------------------------------ | --- | --- | --- | --- | --- | ------------------------- | --- | --------- | ----------------------- | ----------------------- | --- |
| | | | | | | | vanni Gerli, | et | al. 2023. | Can | artificial intelligence | |
| | | | | | | | helpforscientificwriting? | | | Criticalcare,27(1):1–5. | | |
ReiichiroNakano,JacobHilton,SuchirBalaji,JeffWu,
| Long Ouyang, | | Christina | Kim, | Christopher | | Hesse, | | | | | | |
| ------------ | --- | --------- | ---- | ----------- | --- | ------ | --- | --- | --- | --- | --- | --- |
TalSchuster,AdamFisch,andReginaBarzilay.2021.
ShantanuJain,VineetKosaraju,WilliamSaunders,
| | | | | | | | Get your | vitamin | C! | robust | fact verification | with |
| ------------ | ---- | ------- | ---------------- | --- | -------------- | --- | ----------- | --------- | --- | -------------- | ----------------- | -------- |
| et al. 2021. | | Webgpt: | Browser-assisted | | question- | | | | | | | |
| | | | | | | | contrastive | evidence. | | In Proceedings | of | the 2021 |
| answering | with | human | feedback. | | arXiv preprint | | | | | | | |
arXiv:2112.09332. Conference of the North American Chapter of the
| | | | | | | | AssociationforComputationalLinguistics: | | | | | Human |
| --- | --- | --- | --- | --- | --- | --- | --------------------------------------- | --- | --- | --- | --- | ----- |
XiaomanPan,WenlinYao,HongmingZhang,DianYu, LanguageTechnologies,pages624–643,Online.As-
DongYu,andJianshuChen.2022. Knowledge-in- sociationforComputationalLinguistics.
| context: | Towards | knowledgeable | | | semi-parametric | | | | | | | |
| -------- | ------- | ------------- | --- | --- | --------------- | --- | --- | --- | --- | --- | --- | --- |
languagemodels. arXivpreprintarXiv:2210.16433. Özge Sevgili, Artem Shelmanov, Mikhail Arkhipov,
| | | | | | | | Alexander | Panchenko, | | and Chris | Biemann. | 2022. |
| --- | --- | --- | --- | --- | --- | --- | --------- | ---------- | --- | --------- | -------- | ----- |
Jorge Pérez, Marcelo Arenas, and Claudio Gutierrez. Neuralentitylinking: Asurveyofmodelsbasedon
2009. Semantics and complexity of sparql. ACM deeplearning. SemanticWeb,(Preprint):1–44.
TransactionsonDatabaseSystems(TODS),34(3):1–
45.
KurtShuster,SpencerPoff,MoyaChen,DouweKiela,
| | | | | | | | and Jason | Weston. | 2021. | Retrieval | augmentation | |
| --- | --- | --- | --- | --- | --- | --- | --------- | ------- | ----- | --------- | ------------ | --- |
KrishnaPillutla,SwabhaSwayamdipta,RowanZellers,
| | | | | | | | reduceshallucinationinconversation. | | | | arXivpreprint | |
| --- | --- | --- | --- | --- | --- | --- | ----------------------------------- | --- | --- | --- | ------------- | --- |
JohnThickstun,SeanWelleck,YejinChoi,andZaid
arXiv:2104.07567.
| Harchaoui. | 2021. | Mauve: | | Measuring | the gap | be- | | | | | | |
| ---------- | ----- | ------ | --- | --------- | ------- | --- | --- | --- | --- | --- | --- | --- |
tweenneuraltextandhumantextusingdivergence
| | | | | | | | Rohan Taori, | Ishaan | Gulrajani, | | Tianyi Zhang, | Yann |
| ---------- | ----------------------------------- | --- | --- | --- | --- | --- | ------------ | ------ | ---------- | --- | ------------- | ---- |
| frontiers. | AdvancesinNeuralInformationProcess- | | | | | | | | | | | |
Dubois,XuechenLi,CarlosGuestrin,PercyLiang,
ingSystems,34:4816–4828.
| | | | | | | | and Tatsunori | | B Hashimoto. | | 2023. Alpaca: | A |
| --- | --- | --- | --- | --- | --- | --- | ------------- | --- | ------------ | --- | ------------- | --- |
Alistair Plum, Tharindu Ranasinghe, Spencer Jones, strong,replicableinstruction-followingmodel. Stan-
ConstantinOrasan,andRuslanMitkov.2022. Bio- ford Center for Research on Foundation Models.
https://crfm.stanford.edu/2023/03/13/alpaca.html,
graphicalsemi-supervisedrelationextractiondataset.
3(6):7.
| In Proceedings | | of the | 45th | International | ACM | SI- | | | | | | |
| -------------- | --- | ------ | -------- | ------------- | ----------- | --- | --- | --- | --- | --- | --- | --- |
| GIR Conference | | on | Research | and | Development | in | | | | | | |
InformationRetrieval,pages3121–3130. James Thorne, Andreas Vlachos, Christos
| | | | | | | | Christodoulopoulos, | | | and Arpit | Mittal. | 2018. |
| --- | --- | --- | --- | --- | --- | --- | ------------------- | --- | --- | --------- | ------- | ----- |
HongjinQian,ZhichengDou,JiejunTan,HaonanChen, FEVER: a large-scale dataset for fact extraction
HaoqiGu,RuofeiLai,XinyuZhang,ZhaoCao,and and VERification. In Proceedings of the 2018
| Ji-RongWen.2023. | | | Optimizingfactualaccuracyin | | | | | | | | | |
| ---------------- | --- | --- | --------------------------- | --- | --- | --- | ---------- | --- | --- | -------------- | ------- | --- |
| | | | | | | | Conference | of | the | North American | Chapter | of |
text generation through dynamic knowledge selec- the Association for Computational Linguistics:
| tion. | | | | | | | Human | Language | Technologies, | | Volume | 1 (Long |
| ----- | --- | --- | --- | --- | --- | --- | -------- | -------- | ------------- | --- | -------- | ---------- |
| | | | | | | | Papers), | pages | 809–819, | New | Orleans, | Louisiana. |
Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie AssociationforComputationalLinguistics.
| Millican, | Jordan | Hoffmann, | | Francis | Song, | John | | | | | | |
| ---------- | ------ | ---------- | --- | ------- | ----- | ------ | --- | --- | --- | --- | --- | --- |
| Aslanides, | Sarah | Henderson, | | Roman | Ring, | Susan- | | | | | | |
HugoTouvron,ThibautLavril,GautierIzacard,Xavier
| nah Young, | et | al. 2021. | Scaling | language | models: | | | | | | | |
| ---------- | --- | --------- | ------- | -------- | ------- | --- | --- | --- | --- | --- | --- | --- |
Martinet,Marie-AnneLachaux,TimothéeLacroix,
Methods,analysis&insightsfromtraininggopher. BaptisteRozière,NamanGoyal,EricHambro,Faisal
arXivpreprintarXiv:2112.11446.
Azhar,AurelienRodriguez,ArmandJoulin,Edouard
| | | | | | | | Grave,andGuillaumeLample.2023. | | | | Llama: | Open |
| --- | --- | --- | --- | --- | --- | --- | ------------------------------ | --- | --- | --- | ------ | ---- |
ColinRaffel,NoamShazeer,AdamRoberts,Katherine
andefficientfoundationlanguagemodels.
Lee,SharanNarang,MichaelMatena,YanqiZhou,
| WeiLi,andPeterJLiu.2020. | | | | Exploringthelimits | | | | | | | | |
| ------------------------ | --- | --- | --- | ------------------ | --- | --- | --- | --- | --- | --- | --- | --- |
oftransferlearningwithaunifiedtext-to-texttrans- DennyVrandecˇic´ andMarkusKrötzsch.2014. Wiki-
former. TheJournalofMachineLearningResearch, data: afreecollaborativeknowledgebase. Communi-
cationsoftheACM,57(10):78–85.
21(1):5485–5551.
Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, AdinaWilliams,NikitaNangia,andSamuelBowman.
Lora Aroyo, Michael Collins, Dipanjan Das, Slav 2018. A broad-coverage challenge corpus for sen-
Petrov,GauravSinghTomar,IuliaTurc,andDavid tenceunderstandingthroughinference.
| | | | | | A DatasetConstruction | | | | | | |
| ---------------- | ---- | --- | ------------ | ----- | --------------------- | --- | --- | --- | --- | --- | --- |
| Can Xu, Qingfeng | Sun, | Kai | Zheng, Xiubo | Geng, | | | | | | | |
PuZhao,JiazhanFeng,ChongyangTao,andDaxin
Jiang.2023. Wizardlm: Empoweringlargelanguage Inthissection,wewillexplainthedetailedprocess
modelstofollowcomplexinstructions.
andalgorithmsfortheautomaticdatasetconstruc-
| | | | | | tion pipeline. | Using | | this pipeline, | | we are | able to |
| --- | --- | --- | --- | --- | -------------- | ----- | --- | -------------- | --- | ------ | ------- |
WeiYang,YuqingXie,AileenLin,XingyuLi,Luchen
constructdatasetswithagreaterscaleorinother
| Tan, Kun Xiong, | Ming | Li, | and Jimmy Lin. | 2019. | | | | | | | |
| --------------- | ------------------------------ | -------- | -------------- | ----- | ------------------- | --- | --- | --- | --- | --- | --- |
| End-to-end | open-domain | question | answering | with | domains. | | | | | | |
| bertserini. | arXivpreprintarXiv:1902.01718. | | | | | | | | | | |
| | | | | | A.1 PersonSelection | | | | | | |
XiangYue,BoshiWang,KaiZhang,ZiruChen,YuSu,
and Huan Sun. 2023. Automatic evaluation of at- Toimprovethecomplexityofthequestionsanddif-
| tributionbylargelanguagemodels. | | | arXivpreprint | | | | | | | | |
| ------------------------------- | --- | --- | ------------- | --- | --- | --- | --- | --- | --- | --- | --- |
ficultytoLLMs,weinvolvemorethanoneperson
arXiv:2305.06311.
| | | | | | in each question. | | In | addition, | we | need high | qual- |
| --- | --- | --- | --- | --- | ----------------- | --- | --- | --------- | --- | --------- | ----- |
TianyiZhang,VarshaKishore,FelixWu,KilianQWein- ity paragraphs for subsequent dataset generation
berger, and Yoav Artzi. 2019a. Bertscore: Eval- steps. Therefore, we utilize name pairs and para-
| uating text | generation | with | bert. arXiv | preprint | | | | | | | |
| ----------- | ---------- | ---- | ----------- | -------- | --- | --- | --- | --- | --- | --- | --- |
graphsfromthebiographicaldatabase,whichis
arXiv:1904.09675.
adatabasespecificallydesignedfortherelationex-
YuanZhang,JasonBaldridge,andLuhengHe.2019b. traction(RE)task. Eachpieceofdatafromthebio-
PAWS:Paraphraseadversariesfromwordscrambling.
graphicaldatabaseincludesashortparagraph,and
InProceedingsofthe2019ConferenceoftheNorth
| | | | | | arelationtripleextractedfromtheparagraph. | | | | | | The |
| --- | --- | --- | --- | --- | ----------------------------------------- | --- | --- | --- | --- | --- | --- |
AmericanChapteroftheAssociationforComputa-
relationtripleconsistsoftwopeopleandtheirre-
| tionalLinguistics: | | HumanLanguageTechnologies, | | | | | | | | | |
| ------------------ | --- | -------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
Volume1(LongandShortPapers),pages1298–1308, lationshipsuchas<WilliamShakespeare,Spouse,
Minneapolis,Minnesota.AssociationforComputa-
| | | | | | Anne Hathaway>. | | The | biographical | | database | in- |
| --- | --- | --- | --- | --- | --------------- | --- | --- | ------------ | --- | -------- | --- |
tionalLinguistics.
cludesanautomaticallyextractedsetandahuman
GuidoZucconandBevanKoopman.2023. Drchatgpt, annotated set. We specifically choose the human
tellmewhatiwanttohear: Howpromptknowledge annotated set from the database to ensure high-
| impacts health | answer | correctness. | arXiv | preprint | | | | | | | |
| -------------- | ------ | ------------ | ----- | -------- | ----------------- | --- | ---------------------------- | --- | --- | --- | --- |
| | | | | | qualitynamepairs. | | Toavoidpotentialambiguities, | | | | |
arXiv:2302.13793.
wefilteroutdataifanynameinthetripleisincom-
| | | | | | plete. Inpractice,weconsideranamecompleteif | | | | | | |
| --- | --- | --- | --- | --- | ------------------------------------------- | --- | --- | --- | --- | --- | --- |
ithasatleastafamilynameandasurname.
| | | | | | A.2 NameDisambiguation | | | | | | |
| --- | --- | --- | --- | --- | ---------------------- | --- | --- | --- | --- | --- | --- |
Duetothepresenceofduplicatenames(e.g.,Anne
| | | | | | Hathaway: | the | actress, | or | the wife | of | William |
| --- | --- | --- | --- | --- | --------- | --- | -------- | --- | -------- | --- | ------- |
Shakespeare),weperformnamedisambiguationto
mapeachnameinthetripletoauniqueentityfrom
WeutilizeWikiData4
| | | | | | theknowledgegraph. | | | | | | (Vran- |
| --- | --- | --- | --- | --- | ------------------------------------------- | --- | --- | --- | --- | --- | ------ |
| | | | | | decˇic´ andKrötzsch,2014)astheknowledgebase | | | | | | |
andemploySPARQL(Pérezetal.,2009)queriesto
| | | | | | retrieveallentitiesassociatedwiththename. | | | | | | Wiki- |
| --- | --- | --- | --- | --- | ----------------------------------------- | --- | ------ | --- | ------- | ------ | ----- |
| | | | | | Data assigns | a | unique | QID | to each | entity | which |
distinguishesbetweenentitieswiththesamename.
| | | | | | InWikiData, | eachentityrepresentsanodeinthe | | | | | |
| --- | --- | --- | --- | --- | --------------- | ------------------------------ | ---------------------------- | --- | --- | --- | --- |
| | | | | | knowledgegraph. | | Sinceeachtripleconsistsoftwo | | | | |
namesandonerelation,weselectthetwoentities
| | | | | | obtained | from | the query | if | they are | connected | to |
| --- | --- | --- | --- | --- | -------------------- | ------------- | --------- | ------------------------ | ------------ | --------- | --------- |
| | | | | | eachotheronWikiData. | | | Additionally,theconnect- | | | |
| | | | | | ing edge | should | align | with | the relation | | specified |
| | | | | | in the triple. | Subsequently, | | | we extract | | the one- |
hopsub-graphcenteredaroundeachpersonnode,
| | | | | | which provides | | properties | | related | to the | person, |
| --- | --- | --- | --- | --- | -------------- | --- | ---------- | --- | ------- | ------ | ------- |
4https://www.wikidata.org/wiki/
Wikidata:Main_Page
such as gender, birth date, occupation, and more. each relation r in the dataset with size N. The
Weconvertambiguouspersonnamesfromprevious coherence score is calculated through perplexity.
stepstouniqueQIDfromWikiData. Theextracted We convert each piece of knowledge to a simple
sub-graphscontainallknowledgefromWikiData sentencebyapplyingatemplate. Forinstance,the
about the selected people. We call the extracted knowledge [Artemisia, place of birth, Rome] is
graphs“knowledgepool”. convertedto“Artemisia’splaceofbirthisRome”.
TherearethreetemplatesdependingonthePOSof
A.3 EvolutionaryQuestionGeneration
| | | | | | | | therelation. | Weappendeachsentencetotheorig- | | | | |
| --- | --- | --- | --- | --- | --- | --- | ------------ | ------------------------------ | --- | --- | --- | --- |
Weemployan“evolutionaryquestiongeneration” inal paragraph and calculate normalized inverse
approachinspiredbyWizardLM(Xuetal.,2023) perplexitytoobtaincoherencescore. Theoverall
andDKGen(Qianetal.,2023),wherewegradually score is a weighted sum of specificity score and
| increasethesetofknowledgerequiredbyinjecting | | | | | | | coherencescore: | | | | | |
| -------------------------------------------- | ------- | ----------- | --- | --- | ---- | ---------- | --------------- | ---------------- | --- | --- | --- | --- |
| knowledge | through | iterations. | | In | each | iteration, | | | | | | |
| | | | | | | | Score | =α·log(2·N/Count | | | ) | |
| LLMsextendtheparagraphwithonesentenceby | | | | | | | | r | | | r | |
(1)
| | | | | | | | | | +(1−α)·softmax(1/perp | | | ) |
| ------------------------------------ | --- | --- | --- | --- | --- | -------- | --- | --- | --------------------- | --- | --- | --- |
| incorporatingtheadditionalknowledge. | | | | | | Afterthe | | | | | | r |
lastiteration,LLMsproposetwoquestionsaccord-
ing to the extended paragraph, one is a general In each iteration, we leverage the “text-davinci-
| | | | | | | | 003” | model for | annotation | or | generation | with in- |
| -------- | ------- | ----- | ---- | -------- | -------- | --- | ---- | --------- | ---------- | --- | ---------- | -------- |
| version, | and the | other | is a | specific | version. | The | | | | | | |
generalquestionismoreconcise,andthespecific contextlearning. Weprovideseparateinstructions
question is more detailed. Both questions target anddemonstrationsforgeneralandspecificques-
tions. Thedetailedprompttemplatesusedispro-
| the same | set | of knowledge. | | All | injected | knowl- | | | | | | |
| --------- | ---------- | ------------- | --------- | --- | -------- | ------ | ----- | --------------- | --- | ----- | ------- | --------- |
| | | | | | | | vided | in the appendix | | D. We | provide | one human |
| edge form | a “minimum | | knowledge | | set”, | which | | | | | | |
includes the least knowledge required to answer writtendemonstration. Someexamplesoffullques-
theproposedquestion(Table1). Wedonotthrow tionevolutionprocessareprovidedinappendixE.
Inpractice,weemployfiveiterationstoensuresuf-
allknowledgetoLLMatoncetoformaparagraph
because extending the paragraph and knowledge ficientcomplexityinthequestionswithoutmaking
| setincrementallyallowustoselecttheappropriate | | | | | | | themoverlytedious. | | | | | |
| --------------------------------------------- | --- | --- | --- | --- | --- | --- | ------------------ | --- | --- | --- | --- | --- |
knowledgeaftereachiteration.
B ExperimentDetails
| In the | first | iteration, | LLMs | annotate | | the orig- | | | | | | |
| -------------- | ----- | ---------- | ------------ | -------- | -------- | --------- | --- | --- | --- | --- | --- | --- |
| inal paragraph | | from | Biographical | | Database | with | | | | | | |
B.1 MainExperiment
| theknowledgefromthe“knowledgepool”. | | | | | | Forin- | | | | | | |
| ----------------------------------- | --- | --- | --- | --- | --- | ------ | --- | --- | --- | --- | --- | --- |
Forthemainexperiments,werunreachmodelwith
stance,thesentence“ArtemisiawasborninRome.”
isannotatedwithknowledge[Artemisia,placeof differentseedsforthreetimes. TheOpenAIfam-
birth,Rome]. Ineachsubsequentiteration,wese- ily models are implemented using OpenAI APIs.
RunningoneroundofexperimentwithChatGPT
lectapieceofappropriateknowledgeaccordingto
theexistingparagraph. Asentencewithappropriate modeltakesapproximately1hour,andcostsabout
knowledgeshouldhavegoodspecificityandcoher- 3 USD. Running one round of experiment with
| | | | | | | | GPT4 | model | takes approximately | | 1.5 | to 2 hours, |
| ----------------- | --- | ------ | --- | ---------------- | --- | ------ | --------- | ----- | ------------------- | ---- | ----- | ----------- |
| ence. Specificity | | refers | to | the significance | | of the | | | | | | |
| | | | | | | | and costs | about | 60 USD. | Each | LLaMA | family |
knowledge,suchthatitisnottoogeneralortrivial.
Coherence refers to the naturalness of the addi- model is run on one TESLA V100 GPU, where
eachruntakesabout6to8hoursforAlpaca-7Band
| tional knowledge. | | The | added | knowledge | | should | | | | | | |
| ----------------- | --- | --- | ----- | --------- | --- | ------ | --- | --- | --- | --- | --- | --- |
Vicuna-13B,andabout12-16hoursforLLaMA-7B
notdeviatefromtheexistingparagraphandshould
| becoherentwhenreading. | | | | Duringknowledgese- | | | andLLaMA-13B. | | | | | |
| ---------------------- | --- | --- | --- | ------------------ | --- | --- | ------------- | --- | --- | --- | --- | --- |
lection,eachpieceofknowledgeisassignedascore
B.2 TextQualityEvaluation
byaddingthespecificityscoreandcoherencescore.
The specificity score measures the uniqueness of Fortextqualityevaluation,weusethemodel“text-
the knowledge. We discourage the system from davinci-003”withtemperature0toensurestability
selectingtoofrequentrelationtypeslike“gender” and reproducibility of the results. We randomly
or "date of birth" which may be less informative. sample 100 outputs from each baseline and take
A less frequent relation tend to provide a knowl- three runs to report mean. We do not report stan-
edgespecifictotheperson. DerivedfromIDF,we dard deviation since most of them are mostly in-
calculatethenumberofoccurrencesCount for significantlysmall(below0.1).
r
| B.3 NLI | | | | | | 5 means | | the annotator | | feels | that he | or she |
| ------- | --- | --- | --- | --- | --- | ------- | --- | ------------- | --- | ----- | ------- | ------ |
maybeinterestedinthisquestionundersome
Fortheautomaticevaluationoftextcitationalign-
circumstances,and1meanstheopposite.
| ment and | evaluation | | of the | known unknown | cita- | | | | | | | |
| -------- | ---------- | --- | ------ | ------------- | ----- | --- | --- | --- | --- | --- | --- | --- |
tions,weimplementtheTRUEmodelfromHug-
| gingFace5, | which | was | trained | on | SNLI (Bow- | | | | | | | |
| ---------- | ----- | --- | ------- | --- | ---------- | --- | --- | --- | --- | --- | --- | --- |
Theagreementbetweenthetwoannotatorsareas
man et al., 2015), MNLI (Williams et al., 2018), follow: theagreementbetweenthemis100%for
Fever (Thorne et al., 2018), Scitail (Khot et al., authenticityand86%forrelevance. Sincetheeval-
| 2018), | PAWS | (Zhang | et | al., 2019b), | and Vita- | | | | | | | |
| ------ | ---- | ------ | --- | ------------ | --------- | ---------- | ----------- | --- | ---------------- | --- | --- | ----- |
| | | | | | | uation for | naturalness | | and significance | | are | score |
minC (Schuster et al., 2021). The model uses based, in 92% and 90% of the evaluated datasets
thepromptof“premise: {PREMISE}hypothesis: respectively,thescoredifferencebetweenthetwo
{HYPOTHESIS}”. Foreachsentencecitationpair, annotatorsisnolargerthan1.
weplacethesentenceinthe“PREMISE”,andthe
citationtothe“HYPOTHESIS”,likethefollowing:
C.2 GeneratedTextEvaluation
| “premise: | {HertwigservedasaprofessorattheUni- | | | | | | | | | | | |
| --------- | ----------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
versityofJenaforthelast40yearsofhiscareer.}
Amongtextqualityevaluation,citationqualityeval-
hypothesis: {employer: UniversityofJena}” uation,andtext-citationalignment,weconducthu-
| | | | | | | man evaluation | | on text-citation | | alignment. | | Text |
| --- | --- | --- | --- | --- | --- | -------------- | --- | ---------------- | --- | ---------- | --- | ---- |
C HumanEvaluation
| | | | | | | qualityevaluationisconductedusingG-Eval. | | | | | | We |
| --- | --- | --- | --- | --- | --- | ---------------------------------------- | --- | --- | --- | --- | --- | --- |
acknowledgethisisnotaperfectmetric,butthehu-
C.1 DatasetEvaluation
manevaluationisconductedin(Liuetal.,2023b).
| Toevaluatethedatasetquality, | | | | wehavetwoindi- | | | | | | | | |
| ---------------------------- | --- | --- | --- | -------------- | --- | --------- | ------- | ----- | ------ | ---------- | --- | ------- |
| | | | | | | The focus | is this | paper | is not | to improve | | G-Eval. |
vidualannotatorswhoareproficientintheEnglish
Citationqualityevaluationisconductedwithlook-
| language. | Belowaretheexactmethodforevaluat- | | | | | | | | | | | |
| --------- | --------------------------------- | --- | --- | --- | --- | ------- | ----------- | ------- | --- | --------- | --- | --------- |
| | | | | | | ing for | exact match | between | | generated | | citations |
ingeachmetric:
andminimumknowledgeset,whichisanobjective
| • Authenticity. | | Weasktheannotatorstocheck | | | | | | | | | | |
| --------------- | --- | ------------------------- | --- | --- | --- | ----------- | ----------------------------------- | --- | --- | --- | --- | --- |
| | | | | | | evaluation. | Thetext-citationalignmentevaluation | | | | | |
from WikiPedia and understand the back- isconductedusingNLI,whichwearenotcertain
groundstoriesofthementionedpeople,and ifentailmentmeansprovidingsupport. Inaddition,
decideifthegeneratedquestionmatchesthe
| | | | | | | whether | a knowledge | | supports | a sentence | | can be |
| --------------------------------------- | --- | --- | ---------------------- | --- | --- | ---------------- | ------------------------------- | --- | -------- | ---------- | --- | ------ |
| backgroundstory. | | | Eachquestionisassigned | | | | | | | | | |
| | | | | | | subjective. | Therefore,weconducthumanevalua- | | | | | |
| score1ifitmatchesthebackgroundstory,and | | | | | | tiononalignment. | | | | | | |
score0ifthereiscontradiction.
WepresenttheHumanEvaluationInstructions
| | | | | | | provided | to the | annotators | in | Table | 8. We | follow |
| ------------ | --- | ----- | ------------- | --- | --------- | -------- | ------ | ---------- | --- | ----- | ----- | ------ |
| • Relevance. | | After | understanding | | the back- | | | | | | | |
theimplementationfrom(Clarketal.,2021),and
groundstories,weasktheannotatorstolabel
eachpieceofknowledgefromtheminimum providedetailedinstructionsandexamplestoim-
| | | | | | | proveevaluationaccuracy. | | | Forthishumanevalu- | | | |
| --------- | --- | ---- | ------- | ------------ | ------ | ------------------------ | --- | --- | ------------------ | --- | --- | --- |
| knowledge | | set. | A piece | of knowledge | is la- | | | | | | | |
ation,therearefourindividualannotatorsintotal.
beled1iftheannotatorthinksitisnecessary
to answer the question, and 0 if it is redun- Wearrangedifferentannotatorsfordifferentbase-
| | | | | | | lines, and | each | baseline | has | two annotators. | | The |
| ----- | --------------------------------- | --- | --- | --- | --- | ---------- | ---- | -------- | --- | --------------- | --- | --- |
| dant. | Therelevancescoreofaquestionisthe | | | | | | | | | | | |
Inter-AnnotatorAgreementforChatGPT,LLaMA-
ratioofnumberofnecessaryknowledgetothe
numberofknowledgeintheminimumset. 7B,andVicuna-13Barereportedasfollows: 90%,
97%,and89%respectively.
| • Naturalness. | | We | ask | the annotators | to give | | | | | | | |
| -------------- | --- | --- | --- | -------------- | ------- | --- | --- | --- | --- | --- | --- | --- |
anintegerscore1to5tolabeleachquestion.
D Prompts
5meansthequestioncanbeeasilyunderstand-
| able,andisconcise. | | | 1meansthequestionis | | | | | | | | | |
| ------------------ | --- | --- | ------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
Wepresentthepromptsandinstructionsweused
notwritteninnaturalEnglishlanguageoris
| | | | | | | in this section. | | We present | | the prompts | | for the |
| --- | --- | --- | --- | --- | --- | ---------------- | --- | ---------- | --- | ----------- | --- | ------- |
extremelytedious.
evolutionaryquestionconstructioninTable9,10,
• Significance. We ask the annotators to give 11,and12. Wepresentthepromptfortheanswer
anintegerscore1to5tolabeleachquestion. generationinTable13. Wepresentthepromptswe
usefortextevaluationwithG-EvalinTable14,15,
5https://huggingface.co/google/t5_xxl_
| true_nli_mixture | | | | | | 16,and17. | | | | | | |
| ---------------- | --- | --- | --- | --- | --- | --------- | --- | --- | --- | --- | --- | --- |
Annotation Method:
Each evaluation content includes a
sentence and a piece of knowledge.
Our task is to determine whether this
sentence contains the given knowledge,
i.e., whether this knowledge provides
support for the sentence. If the
sentence does not mention the given
knowledge or if the content of the
sentence does not align with the
knowledge, it is considered
unsupported. We use 1 to indicate
support and 0 to indicate lack of
support.
Here are some examples:
Sentence: Stephen Crane was an
American writer born on November 1,
1871, in Newark, and died on June 5,
1900, in Badenweiler.
Knowledge: date of birth:
1871-11-01
Result: 1, because the sentence’s
date of birth matches the knowledge’s
date of birth.
Sentence: Merton died on December 10,
1968, in Bangkok, Thailand.
Knowledge: country of citizenship:
United States of America
Result: 0, because the sentence does
not mention Merton’s nationality.
Table8:Instructionweprovidetothehumanannotators.
E EvolutionaryQuestionGeneration
We provide an example of evolutionary question
generationinTable18.
F Examples
Weshowexamplesoftheattributedanswersgener-
atedbytheLLMsinTable19and20.
Instruction: Your objective is to select relevant knowledge to
label the sentence and generate a question
sentence: Artemisia Gentileschi was born Artemisia Gentileschi
Lomi in Rome on July 8 1593 although her birth certificate from the
Archivio di Stato indicated she was born in 1590 the eldest child
of the Tuscan painter Orazio Gentileschi and Prudenzia di Ottaviano
Montoni.
knowledge: {qid: Q367360, name: Orazio Gentileschi, sex or
gender: male, place of birth: Pisa, place of death: London,
instance of: human, occupation: painter, child: Artemisia
Gentileschi, described by source: The Great Theatre of Dutch
Painters, notable works: Diana the Huntress, given name: Orazio,
topic’s main category: Category:Orazio Gentileschi, surname:
Gentileschi, genre: portrait, languages spoken: Italian,
movement: mannerism, work location: Rome, ethnic group:
Italians, date of birth: 1563-07-19, date of death: 1639-02-07}
{qid: Q212657, name: Artemisia Gentileschi, sex or gender:
female, place of birth: Rome, place of death: Naples, instance
of: human, occupation: painter, member of: Accademia delle Arti
del Disegno, father: Orazio Gentileschi, described by source: The
Great Theatre of Dutch Painters, notable works: Judith Slaying
Holofernes, topic’s main category: Category:Artemisia Gentileschi,
movement: Caravaggisti, ethnic group: Italians, work location:
Florence, depicted by: Artemisia, field of work: painting,
surname: Gentileschi, genre: portrait, languages spoken:
Italian, position held: court painter, student of: Orazio
Gentileschi, spouse: Pierantonio Stiattesi, given name:
Artemisia, mother: Prudenzia di Ottaviano Montoni, date of birth:
1596-07-08, date of death: 1654-01-01}
Generated Answer: Artemisia Gentileschi [qid: Q212657, name:
Artemisia Gentileschi] was born Artemisia Gentileschi Lomi in Rome
[qid: Q212657, place of birth: Rome] on July 8 1593 [qid:
Q212657, date of birth: 1596-07-08] although her birth certificate
from the Archivio di Stato indicated she was born in 1590 the
eldest child of the Tuscan painter Orazio Gentileschi [qid:
Q212657, father: Orazio Gentileschi] [qid: Q367360, name: Orazio
Gentileschi, occupation: painter] and Prudenzia di Ottaviano
Montoni.
Table9: Instructionanddemonstrationforinitialroundofevolutionaryquestionconstruction. Weusebrowncolor
fortheknowledge,andbluecolorforthedemonstrationofexpectedoutput.
Instruction: Your objective is to extend the original paragraph by
| adding | one | sentence | | that | includes | | the | given | knowledge | |
| ------- | --------- | -------- | ----------- | ---- | -------- | ----- | --- | -------- | --------- | --------- |
| answer: | Artemisia | | Gentileschi | | | [qid: | | Q212657, | name: | Artemisia |
Gentileschi] was born Artemisia Gentileschi Lomi in Rome [qid:
Q212657, place of birth: Rome] on July 8 1593 [qid: Q212657, date
of birth: 1596-07-08] although her birth certificate from the
Archivio di Stato indicated she was born in 1590 the eldest child
of the Tuscan painter Orazio Gentileschi [qid: Q212657, father:
| Orazio | Gentileschi] | | | [qid: | Q367360, | | name: | Orazio | Gentileschi, | |
| ----------- | ------------ | -------- | --- | ----- | --------- | --- | ----- | --------- | ------------ | --- |
| occupation: | | painter] | | and | Prudenzia | | di | Ottaviano | Montoni. | |
knowledge: {qid: Q212657, name: Artemisia Gentileschi, notable
| works: | Judith | Slaying | | Holofernes} | | | | | | |
| --------- | ------ | ------- | --------- | ----------- | --- | ----------- | --- | ----- | -------- | ----- |
| Generated | | Answer: | Artemisia | | | Gentileschi | | [qid: | Q212657, | name: |
Artemisia Gentileschi] was born Artemisia Gentileschi Lomi in Rome
| [qid: | Q212657, | | place | of | birth: | | Rome] | on July | 8 1593 | [qid: |
| ----- | -------- | --- | ----- | --- | ------ | --- | ----- | ------- | ------ | ----- |
Q212657, date of birth: 1596-07-08] although her birth certificate
from the Archivio di Stato indicated she was born in 1590 the
| eldest | child | of | the | Tuscan | painter | | Orazio | Gentileschi | | [qid: |
| ------ | ----- | --- | --- | ------ | ------- | --- | ------ | ----------- | --- | ----- |
Q212657, father: Orazio Gentileschi] [qid: Q367360, name: Orazio
| Gentileschi, | | occupation: | | | painter] | | and | Prudenzia | di | Ottaviano |
| ------------ | --- | ----------- | --- | --- | -------- | --- | --- | --------- | --- | --------- |
Montoni. Under the influence of her father, Artemisia Gentileschi
| created | her | iconic | painting | | | Judith | Slaying | Holofernes | | [qid: |
| ------- | --- | ------ | -------- | --- | --- | ------ | ------- | ---------- | --- | ----- |
Q212657, notable works: Judith Slaying Holofernes] when she was
| around | twenty | years | | old. | | | | | | |
| ------ | ------ | ----- | --- | ---- | --- | --- | --- | --- | --- | --- |
Table10: Instructionanddemonstrationforevolutionaryquestionconstructionforround2to5. Weusebrown
colorfortheknowledge,andbluecolorforthedemonstrationofexpectedoutput.
Instruction: Your objective is to ask a question whose answer is
| the given | | paragraph. | | There | should | | be only | | one question | | when |
| --------- | --- | ---------- | --- | ----- | ------ | --- | ------- | --- | ------------ | --- | ---- |
possible, if not, make sure the question is as concise as possible.
Paragraph: Artemisia Gentileschi was born Artemisia Gentileschi
Lomi in Rome on July 8 1593 although her birth certificate from the
Archivio di Stato indicated she was born in 1590 the eldest child
of the Tuscan painter Orazio Gentileschi and Prudenzia di Ottaviano
| Montoni. | | Her life | and | work | were | | later | depicted | | in the | film |
| ---------- | --- | -------- | ------- | ---- | ---- | ----- | ----- | -------- | --------- | ------ | ---- |
| Ärtemisia, | | which | brought | | her | story | to a | wider | audience. | | Her |
¨
| father, | Orazio, | | was a | prominent | | figure | | in the | Mannerism | | art |
| ------- | ------- | --- | ----- | --------- | --- | ------ | --- | ------ | --------- | --- | --- |
movement, which likely influenced Artemisia’s own artistic style.
However, Artemisia herself was a part of the Caravaggisti movement,
a group of artists who followed the style of Caravaggio. She was
also a student of her father, Orazio Gentileschi, which further
shaped her artistic development. Orazio’s influence on Artemisia’s
development as a prominent Baroque painter can be seen in her
highly naturalistic portrayal of figures, dramatic scenes and the
| use of | chiaroscuro | | technique | | | | | | | | |
| ---------- | ----------- | --------- | --------- | -------- | --- | --------- | --- | ----------- | --- | --- | ---- |
| Generated | | Question: | | Who | was | Artemisia | | Gentileschi | | and | what |
| influences | | shaped | her | artistic | | style? | | | | | |
Table11: Instructionanddemonstrationforgeneralquestiongeneration. Weusebluecolorforthedemonstrationof
expectedoutput.
Instruction: Your objective is to ask a question whose answer is
| the given | | paragraph. | | The | question | | should | not | be | too | tedious. |
| --------- | --- | ---------- | --- | --- | -------- | --- | ------ | --- | --- | --- | -------- |
Paragraph: Artemisia Gentileschi was born Artemisia Gentileschi
Lomi in Rome on July 8 1593 although her birth certificate from the
Archivio di Stato indicated she was born in 1590 the eldest child
of the Tuscan painter Orazio Gentileschi and Prudenzia di Ottaviano
| Montoni. | | Her life | and | work | were | | later | depicted | | in the | film |
| ---------- | --- | -------- | ------- | ---- | ---- | ----- | ----- | -------- | --------- | ------ | ---- |
| Ärtemisia, | | which | brought | | her | story | to a | wider | audience. | | Her |
¨
| father, | Orazio, | | was a | prominent | | figure | | in the | Mannerism | | art |
| ------- | ------- | --- | ----- | --------- | --- | ------ | --- | ------ | --------- | --- | --- |
movement, which likely influenced Artemisia’s own artistic style.
However, Artemisia herself was a part of the Caravaggisti movement,
a group of artists who followed the style of Caravaggio. She was
also a student of her father, Orazio Gentileschi, which further
shaped her artistic development. Orazio’s influence on Artemisia’s
development as a prominent Baroque painter can be seen in her
highly naturalistic portrayal of figures, dramatic scenes and the
| use of | chiaroscuro | | technique | | | | | | | | |
| ------ | ----------- | --- | --------- | --- | --- | --- | --- | --- | --- | --- | --- |
Generated Question: What were the key artistic influences and
characteristics that shaped Artemisia Gentileschi’s unique Baroque
| style, | and | how did | her | relationship | | | with | her | father, | | Orazio |
| ------------ | --- | ------- | --- | ------------ | --- | --- | ------------ | --- | ------- | --- | ------ |
| Gentileschi, | | impact | her | artistic | | | development? | | | | |
Table12: Instructionanddemonstrationforspecificquestiongeneration. Weusebluecolorforthedemonstrationof
expectedoutput.
Instruction: You answer the question based on your knowledge, with
the given information for annotation, following the given format.
Use [NA] for claims that need annotation but is unprovided.
Question: Considering the information:
{name: Orazio Gentileschi, place of death: London, child:
Artemisia Gentileschi, notable works: Diana the Huntress, given
name: Orazio, surname: Gentileschi, languages spoken: Italian,
movement: mannerism, work location: Rome, ethnic group:
Italians, date of birth: 1563-07-19, date of death: 1639-02-07,
qid: Q367360}
{name: Artemisia Gentileschi, place of birth: Rome, place of
death: Naples, occupation: painter, member of: Accademia delle
Arti del Disegno, father: Orazio Gentileschi, notable works:
Judith Slaying Holofernes, movement: Caravaggisti, ethnic group:
Italians, work location: Florence, depicted by: Artemisia,
surname: Gentileschi, languages spoken: Italian, student of:
Orazio Gentileschi, given name: Artemisia, mother: Prudenzia di
Ottaviano Montoni, date of death: 1654-01-01, qid: Q212657}
How did Orazio Gentileschi’s influence on Artemisia’s life and
career shape her development as a prominent Baroque painter,
despite facing significant challenges as a female artist in a
male-dominated field?
Answer: Artemisia Gentileschi was an Italian painter born on July
8, 1596 [NA] in Rome [Q212657, ethnic group: Italians, occupation:
painter, place of birth: Rome]. She was a member of the Accademia
delle Arti del Disegno and is best known for her work Judith
Slaying Holofernes [Q212657, member of: Accademia delle Arti del
Disegno, notable works: Judith Slaying Holofernes]. She was the
eldest child of the Tuscan painter Orazio Gentileschi [Q212657,
father: Orazio Gentileschi]. Orazio Gentileschi was an Italian
painter [NA] born in 1563 and died in 1639 [Q367360, ethnic group:
Italians, date of birth: 1563-07-19, date of death: 1639-02-07].
He was born in Pisa and died in London [Q367360, place of death:
London]. Orazio’s influence on Artemisia’s development as a
prominent Baroque [NA] painter can be seen in her highly
naturalistic portrayal of figures, dramatic scenes and the use of
chiaroscuro technique [NA]. He also provided her with the
opportunity to study with him and learn from his experience and
expertise. She became an important second-generation proponent of
Caravaggio’s dramatic realism [Q212657, movement: Caravaggisti].
Table 13: Full instruction and demonstration for answer generation with citaion. We use brown color for the
knowledgepool,andbluecolorfortheexpectedoutcomeprovidedbythedemonstration.
Instruction: You will be given one question and answer. Your task
is to rate the answer on one metric. Please make sure you read and
understand these instructions carefully. Please keep this document
| open while | | reviewing, | | and | refer | to | it | as | needed. | |
| ---------- | --- | ---------- | --- | --- | ----- | --- | --- | --- | ------- | --- |
| Evaluation | | Criteria: | | | | | | | | |
Coherence (1-5) - the collective quality of all sentences. We
align this dimension with the DUC quality question of structure and
| coherence | | whereby | | the answer | | should | be | well-structured | | and |
| --------- | --- | ------- | --- | ---------- | --- | ------ | --- | --------------- | --- | --- |
well-organized. The answer should not just be a heap of related
| information, | | but | should | build | | from | sentence | | to sentence | to a |
| ------------ | ---- | ------ | -------------- | ----- | --- | ----- | -------- | ------ | ----------- | ---- |
| coherent | body | | of information | | | about | a | topic. | | |
| Evaluation | | Steps: | | | | | | | | |
1. Read the questions carefully and identify the main topic and
key points.
2. Read the answer and compare it to the question. Check if the
answer covers the main topic and key points of the question, and if
| it presents | | them | in | a clear | and | logical | | order. | | |
| ----------- | --- | ---- | --- | ------- | --- | ------- | --- | ------ | --- | --- |
3. Assign a score for coherence on a scale of 1 to 5, where 1 is
the lowest and 5 is the highest based on the Evaluation Criteria.
| | | Table14: | | InstructionfortextevaluationwithGPT-EVAL-Coherence | | | | | | |
| --- | --- | -------- | --- | -------------------------------------------------- | --- | --- | --- | --- | --- | --- |
Instruction: You will be given one question and answer. Your task
is to rate the answer on one metric. Please make sure you read and
understand these instructions carefully. Please keep this document
| open while | | reviewing, | | and | refer | to | it | as | needed. | |
| ---------- | --- | ---------- | --- | --- | ----- | --- | --- | --- | ------- | --- |
| Evaluation | | Criteria: | | | | | | | | |
Consistency (1-5) - the answer should be consistent with the given
knowledge. The answer should also be self-consistent, without any
| contradiction | | | to itself. | | | | | | | |
| ------------- | --- | -------- | ---------- | --- | --------- | --- | ---------- | --- | --- | --- |
| Evaluation | | Steps: | | | | | | | | |
| 1. Read | the | question | | and | knowledge | | carefully. | | | |
2. Read the answer and compare it to the knowledge. Check if the
| answer | is | consistent | | with | the | give | knowledge. | | | |
| ------ | --- | ---------- | --- | ---- | --- | ---- | ---------- | --- | --- | --- |
3. Assign a score for consistency on a scale of 1 to 5, where 1 is
the lowest and 5 is the highest based on the Evaluation Criteria.
| | | Table15: | | InstructionfortextevaluationwithGPT-EVAL-Consistency | | | | | | |
| --- | --- | -------- | --- | ---------------------------------------------------- | --- | --- | --- | --- | --- | --- |
Instruction: You will be given one question and answer. Your task
is to rate the answer on one metric. Please make sure you read and
understand these instructions carefully. Please keep this document
| open while | | reviewing, | and refer | to it | as needed. |
| ---------- | --- | ---------- | --------- | ----- | ---------- |
| Evaluation | | Criteria: | | | |
Fluency (1-5) - the answer should be written in fluent language.
The answer should use appropriate vocabulary, grammar, and sentence
structures that enable readers or listeners to comprehend the
| content | effortlessly. | | | | |
| ---------- | ------------- | -------- | ---------- | --- | --- |
| Evaluation | | Steps: | | | |
| 1. Read | the | question | carefully. | | |
2. Read the answer and check if the language in the answer is
fluent.
3. Assign a score for fluency on a scale of 1 to 5, where 1 is the
lowest and 5 is the highest based on the Evaluation Criteria.
| | | Table16: | InstructionfortextevaluationwithGPT-EVAL-Fluency | | |
| --- | --- | -------- | ------------------------------------------------ | --- | --- |
Instruction: You will be given one question and answer. Your task
is to rate the answer on one metric. Please make sure you read and
understand these instructions carefully. Please keep this document
| open while | | reviewing, | and refer | to it | as needed. |
| ---------- | --- | ---------- | --------- | ----- | ---------- |
| Evaluation | | Criteria: | | | |
Relevance (1-5) - the answer should be relevant to the question.
The answer should directly answers the question, without providing
| any irrelevant | | information. | | | |
| -------------- | --- | ------------ | ---------- | --- | --- |
| Evaluation | | Steps: | | | |
| 1. Read | the | question | carefully. | | |
2. Read the answer and compare with the question to check if it
| fully | answers | the question | and | have no | redundancies. |
| ----- | ------- | ------------ | --- | ------- | ------------- |
3. Assign a score for relevance on a scale of 1 to 5, where 1 is
the lowest and 5 is the highest based on the Evaluation Criteria.
| | | Table17: InstructionfortextevaluationwithGPT-EVAL-Relevance | | | |
| --- | --- | ----------------------------------------------------------- | --- | --- | --- |
Round 1:
Annotated Knowledge:
[qid: Q258115, name: Diego Simeone, occupation: association football player]
[qid: Q6439494, name: Giovanni Simeone, father: Diego Simeone]
[qid: Q6439494, name: Giovanni Simeone, place of birth: Madrid]
[qid: Q258115, name: Diego Simeone, member of sports team: Atletico Madrid]
Paragraph: The son of former Argentine international footballer Diego Simeone
and Carolina Baldini Giovanni Simeone was born in Madrid while his father was
playing for Atletico Madrid.
Round 2:
Input Knowledge: [qid: Q258115, name: Diego Simeone, position played on team:
midfielder]
Paragraph: The son of former Argentine international footballer Diego Simeone
and Carolina Baldini Giovanni Simeone was born in Madrid while his father was
playing as a midfielder for Atletico Madrid.
Round 3:
Input Knowledge: [qid: Q258115, Diego Simeone, member of sports team: S.S.
Lazio]
Paragraph: The son of former Argentine international footballer Diego Simeone
and Carolina Baldini Giovanni Simeone was born in Madrid while his father was
playing as a midfielder for Atletico Madrid. Before his stint at Atletico Madrid,
Diego Simeone had also played for S.S. Lazio.
Round 4:
Input Knowledge: [qid: Q258115, Diego Simeone, medical condition: COVID-19]
Paragraph: The son of former Argentine international footballer Diego Simeone
and Carolina Baldini Giovanni Simeone was born in Madrid while his father was
playing as a midfielder for Atletico Madrid. Before his stint at Atletico Madrid,
Diego Simeone had also played for S.S. Lazio. Later in life,
Diego Simeone contracted the COVID-19 virus, but managed to recover.
Round 5:
Input Knowledge: [qid: Q6439494, Giovanni Simeone, position played on team:
forward]
Paragraph: The son of former Argentine international footballer Diego Simeone
and Carolina Baldini Giovanni Simeone was born in Madrid while his father was
playing as a midfielder for Atletico Madrid. Before his stint at Atletico Madrid,
Diego Simeone had also played for S.S. Lazio. Later in life, Diego Simeone
contracted the COVID-19 virus, but managed to recover. Following in his father’s
footsteps but with a different role on the pitch,
Giovanni Simeone grew up to be a forward in the game of football.
General Question:
How did the career of Diego Simeone influence his son’s path in football?
Specific Question:
Who are Diego Simeone and Giovanni Simeone, and how are there career path in
football different from each other?
Table18: Anexampleofthegeneratedquestionsaftereachroundinevolutionaryquestionconstruction. Weuse
browncolorfortheknowledge,andbluecolorforthegeneratedquestions. Ineachroundweunderlinethetextsto
thenewlyinjectedknowledge.
Question:
What were some of Stephen Crane’s notable works that reflected his
views on religion and the harsh realities of life, and how did his
experiences at Syracuse University influence his writing? Also,
what was the name of the sports team he was a member of at Syracuse
University?
| Retrieved | Knowledge: | | | | | | | |
| --------- | ---------- | --- | --- | --- | --- | --- | --- | --- |
[qid: Q206534, name: Stephen Crane, sex or gender: male, place
| of birth: | | Newark, | place | of death: | Badenweiler, | | country | of |
| ------------ | --- | ------- | ------ | --------- | ------------ | ----- | --------- | --- |
| citizenship: | | United | States | of | America, | cause | of death: | |
Tuberculosis, place of burial: Evergreen Cemetery, alma mater:
Syracuse University, instance of: human, given name: Stephen,
medical condition: Tuberculosis, occupation: writer, described by
source: Great Soviet Encyclopedia (1969âC“1978), manner of death:
natural causes, movement: literary realism, notable works: The
Red Badge of Courage, surname: Crane’, "topic’s main category":
’Category:Stephen Crane, religion: atheism, languages spoken:
English, mother: Mary Helen Peck Crane, sport: baseball, member
| of sports | team: | | Syracuse | Orange | baseball, | father: | Jonathan | |
| --------- | ----- | --- | -------- | ------ | --------- | ------- | -------- | --- |
Townley Crane, name: Stephen Crane, date of birth: 1871-11-01,
| date | of death: | | 1900-06-05] | | | | | |
| --------- | --------- | --- | ----------- | ------ | ---- | ----------- | -------- | --- |
| Generated | Output: | | | | | | | |
| Stephen | Crane | was | a male | writer | born | on November | 1, 1871, | in |
Newark, United States of America, and died on June 5, 1900, in
Badenweiler due to tuberculosis [Q206534, sex or gender: male,
place of birth: Newark, place of death: Badenweiler, country of
| citizenship: | | United | States | of | America, | cause | of death: | |
| ------------ | --- | ------ | ------ | --- | -------- | ----- | --------- | --- |
Tuberculosis, medical condition: Tuberculosis, date of birth:
1871-11-01, date of death: 1900-06-05]. Crane was a prominent
literary realist writer [Q206534, movement: literary realism] who
wrote about the harsh realities of life and his atheistic views on
religion [Q206534, religion: atheism]. His notable works include
"The Red Badge of Courage" [Q206534, notable works: The Red Badge
of Courage].
Crane attended Syracuse University and was a member of the Syracuse
Orange baseball team [Q206534, alma mater: Syracuse University,
| sport: | baseball, | | member | of sports | team: | Syracuse | Orange | |
| ------ | --------- | --- | ------ | --------- | ----- | -------- | ------ | --- |
baseball]. His experiences at Syracuse University influenced his
writing, and he wrote several stories about college life, including
"The Eighty-Yard Run," which was based on his own experiences as a
| football | player | | at Syracuse | [NA]. | | | | |
| -------- | ------ | --- | ----------- | ----- | --- | --- | --- | --- |
Table19: AnexampleoftheattributedanswergeneratedbyChatGPT(temperature0.5). Weusebluecolorforthe
questionandbrowncolorfortheretrievedknowledge.
Question:
What were some of Stephen Crane’s notable works that reflected his
views on religion and the harsh realities of life, and how did his
experiences at Syracuse University influence his writing? Also,
what was the name of the sports team he was a member of at Syracuse
University?
| Retrieved | Knowledge: | | | | | | | | | |
| --------- | ---------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
[qid: Q206534, name: Stephen Crane, sex or gender: male, place
| of birth: | | Newark, | place | | of death: | Badenweiler, | | | country | of |
| ------------ | --- | ------- | ----- | ------ | --------- | ------------ | ----- | --- | --------- | --- |
| citizenship: | | United | | States | of | America, | cause | | of death: | |
Tuberculosis, place of burial: Evergreen Cemetery, alma mater:
Syracuse University, instance of: human, given name: Stephen,
medical condition: Tuberculosis, occupation: writer, described by
source: Great Soviet Encyclopedia (1969âC“1978), manner of death:
natural causes, movement: literary realism, notable works: The
Red Badge of Courage, surname: Crane’, "topic’s main category":
’Category:Stephen Crane, religion: atheism, languages spoken:
English, mother: Mary Helen Peck Crane, sport: baseball, member
| of sports | team: | | Syracuse | | Orange | baseball, | | father: | Jonathan | |
| --------- | ----- | --- | -------- | --- | ------ | --------- | --- | ------- | -------- | --- |
Townley Crane, name: Stephen Crane, date of birth: 1871-11-01,
| date | of death: | | 1900-06-05] | | | | | | | |
| --------- | --------- | --- | ----------- | --- | --- | --- | --- | --- | --- | --- |
| Generated | Output: | | | | | | | | | |
Stephen Crane was an American writer born on November 1, 1871, in
Newark, and died on June 5, 1900, in Badenweiler [Q206534, date of
birth: 1871-11-01, date of death: 1900-06-05, place of birth:
| Newark, | place | of | death: | | Badenweiler]. | | He | was | a part | of the |
| ------- | ----- | --- | ------ | --- | ------------- | --- | --- | --- | ------ | ------ |
literary realism movement and is best known for his notable work,
The Red Badge of Courage [Q206534, movement: literary realism,
| notable | works: | | The Red | Badge | | of Courage]. | | Crane’s | views | on |
| ------- | ------ | --- | ------- | ----- | --- | ------------ | --- | ------- | ----- | --- |
religion were atheistic [Q206534, religion: atheism]. His works
often reflected the harsh realities of life, such as war, poverty,
| and human | suffering | | [NA]. | | | | | | | |
| --------- | --------- | --- | ----- | --- | --- | --- | --- | --- | --- | --- |
Crane attended Syracuse University [Q206534, alma mater: Syracuse
University], where his experiences likely influenced his writing.
He was exposed to a diverse range of ideas and perspectives, which
may have contributed to his realistic approach to literature [NA].
While at Syracuse University, Stephen Crane was a member of the
Syracuse Orange baseball team [Q206534, member of sports team:
| Syracuse | Orange | | baseball]. | | | | | | | |
| -------- | ------ | --- | ---------- | --- | --- | --- | --- | --- | --- | --- |
Table20: AnexampleoftheattributedanswergeneratedbyGPT-4. Weusebluecolorforthequestionandbrown
colorfortheretrievedknowledge. |