Datasets:
File size: 164,596 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 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 | Fact-Checking Complex Claims with Program-Guided Reasoning
LiangmingPan1,2 XiaobaoWu3 XinyuanLu4 AnhTuanLuu3
WilliamYangWang1 Min-YenKan4 PreslavNakov2
1 UniversityofCalifornia,SantaBarbara 2 MBZUAI
3 NanyangTechnologicalUniversity 4 NationalUniversityofSingapore
liangmingpan@ucsb.edu xiaobao002@e.ntu.edu.sg luxinyuan@u.nus.edu
anhtuan.luu@ntu.edu.sg william@cs.ucsb.edu
kanmy@comp.nus.edu.sg preslav.nakov@mbzuai.ac.ae
Abstract Instead,ahumanfact-checkerneedstodecom-
posetheclaim,gathermultiplepiecesofevidence,
Fact-checking real-world claims often re-
andperformstep-by-stepreasoning(Nakovetal.,
quires collecting multiple pieces of evidence
2021a),asillustratedinFigure1. Thismakesveri-
and applying complex multi-step reasoning.
fyingcomplexclaimsmuchmorechallengingthan
In this paper, we present Program-Guided
Fact-Checking (PROGRAMFC), a novel fact- thetypicalsettingexploredinpreviouswork,where
checking model that decomposes complex informationfromasinglearticleissufficienttosup-
claimsintosimplersub-tasksthatcanbesolved port/refutetheclaim(Thorneetal.,2018;Saakyan
usingasharedlibraryofspecializedfunctions.
etal.,2021;Schusteretal.,2021;Panetal.,2021;
Wefirstleveragethein-contextlearningability
Waddenetal.,2022a;Krishnaetal.,2022).
of large language models to generate reason-
ingprogramstoguidetheverificationprocess. Besides multi-step reasoning, we still need to
Afterward,weexecutetheprogrambydelegat- considertwokeyaspectsfordevelopingareliable
ingeachsub-tasktothecorrespondingsub-task fact-checkingsystem: (i)Explanability: Themodel
handler. This process makes our model both should not only predict the veracity of the claim,
explanatoryanddata-efficient,providingclear
butitshouldalsoprovideaclearexplanationofits
explanationsofitsreasoningprocessandrequir-
reasoningprocesstohelpusersunderstandandtrust
ingminimaltrainingdata. Weevaluate PRO-
theresults. (ii)Dataefficiency: Humanannotation
GRAMFC on two challenging fact-checking
is often time-consuming, costly, and potentially
datasets and show that it outperforms seven
fact-checkingbaselinesacrossdifferentsettings biased,makingitdifficulttocollectsufficienthigh-
of evidence availability, with explicit output qualitylabeleddataformodeltraining,particularly
programsthatbenefithumandebugging.1 for complex claims. Therefore, it is desirable to
buildamodelthatcanperformwellwithminimal
1 Introduction
or no training data. Despite a few models (Zhou
Theproliferationofdisinformation,e.g.,insocial etal.,2019;Zhongetal.,2020;AlyandVlachos,
media,hasmadeautomatedfact-checkingacrucial 2022) being proposed to facilitate multi-step rea-
applicationofnaturallanguageprocessing(NLP). soning in fact-checking, they either lack explain-
Given a claim, the goal is to find evidence and abilityintheirreasoningprocessorrequirealarge
then to make a verdict about the claim’s veracity numberoftask-specifictrainingexamples.
basedonthatevidence(ThorneandVlachos,2018; Inthispaper,wepresentProgram-GuidedFact-
Glockneretal.,2022;Guoetal.,2022). Checking (PROGRAMFC), a novel fact-checking
Evaluatingtheveracityofreal-worldclaimsof- framework that is both explanatory and data-
teninvolvescollectingmultiplepiecesofevidence efficient. Figure 1 illustrates our approach. To
andapplyingcomplexreasoning(Jiangetal.,2020; verifycomplexclaims, PROGRAMFC decomposes
Nguyenetal.,2020;AlyandVlachos,2022;Chen themintosimplersub-tasksthatcanbesolvedus-
et al., 2022a). For instance, consider the claim ing a shared library of specialized sub-task func-
“BothJamesCameronandthedirectorofthefilm tions. Tobespecific, PROGRAMFCbeginsbygen-
InterstellarwereborninCanada”. Itmaybechal- erating a reasoning program for the input claim,
lenging to find direct evidence on the web that which is a sequence of sub-tasks (e.g., S1-S4 in
refutesorsupportsthisclaim. Figure 1) in the form of ACTION[ARGUMENT],
1Theprogramcodeandthedataarepubliclyavailableat
where ACTION and ARGUMENT define the type
https://github.com/mbzuai-nlp/ProgramFC andthecontentofthesub-task,respectively.
6981
Proceedingsofthe61stAnnualMeetingoftheAssociationforComputationalLinguistics
Volume1:LongPapers,pages6981–7004
July9-14,2023©2023AssociationforComputationalLinguistics
Claim: Both James Cameron and the director of the film Interstellar were born in Canada.
Knowledge
| | | | Reasoning Program | | | Functions | | Source |
| --- | --- | --- | ----------------- | --- | --- | --------- | --- | ------ |
Language
| Models | | S1 Verify | [James Cameron was born in Canada.] | | | | Fact | |
| ---------------- | --- | --------- | ----------------------------------- | --- | --- | ------- | ----- | --- |
| (Codex, GPT3, …) | | | | | | Checker | | |
FACT_1 = TRUE
Gold Evidence
| | | S2 Question[Who is the director of the film Interstellar?] | | | | | | |
| --- | --- | ---------------------------------------------------------- | --- | --- | --- | --- | --- | --- |
QA
| | | | | ANSWER_1 = Christopher Nolan | | Model | | |
| --- | --- | --------- | --------------------------------- | ---------------------------- | --- | ----- | --- | --------- |
| | | S3 Verify | [ {ANSWER_1} was born in Canada.] | | | | | Open-book |
Fact
Claim: ⋯
FACT_2 = FALSE Checker
ProCglariamm: :⋯ ⋯
ProCglariamm: :⋯ ⋯
| | | S4 Predict | [ {FACT_1} AND{FACT_2}] | | | | | |
| ---------- | --- | ---------- | ----------------------- | --- | --- | -------- | --- | ----------- |
| Program: ⋯ | | | | | | Logical | | Closed-book |
Reasoner
| | | | | PREDICTED_LABEL = | REFUTES | | | |
| --- | --- | --- | --- | ----------------- | ------- | --- | --- | --- |
Exemplars
Figure1: OverviewofourPROGRAMFCmodel,whichconsistsoftwomodules: (i)ProgramGenerationgenerates
areasoningprogramfortheinputclaimusingCodexwithin-contextlearning,andthen(ii)ProgramExecution
sequentiallyinterpretstheprogrambydelegatingeachsteptothecorrespondingsub-taskfunction.
The generated reasoning program serves as a The strategy of program-guided reasoning be-
step-by-step guide for verifying the claim. We comesincreasinglyeffectiveastherequiredreason-
thenexecutetheprogrambysequentiallydelegat- ingdepthincreases(§4.1). Intheopen-domainset-
ing each sub-task to the corresponding sub-task ting,wefindthatreasoningprogramscanenhance
handler,asshowninthefunctionscolumnsinFig- theretrievalofrelevantevidencefromknowledge
ure 1. These sub-tasks may include answering sources(§4.2). Moreover,PROGRAMFCisrobust
questions,verifyingsimpleclaims,orconducting evenwhenweuseweakmodelsassub-tasksolvers
logicalreasoning. (§4.2). Wealsoevaluatetheinterpretabilityofthe
reasoningprogramsthroughhumanevaluationand
PROGRAMFCcombinesexplainabilitywithdata
efficiency. Itusesreasoningprogramstoprovide erroranalysis(§4.3).
| clear explanations | | of its reasoning | process. | For | | | | |
| ------------------ | --- | ---------------- | -------- | --- | --- | --- | --- | --- |
2 RelatedWork
| data efficiency, | | Large Language | Models (LLMs) | | | | | |
| ---------------- | --- | -------------- | ------------- | --- | --- | --- | --- | --- |
can solve various tasks given only a few exam- Fact-Checking. Automated fact-checking has
plesasprompts,e.g.,in-contextlearning(Brown
| | | | | gained significant | attention | | in the | NLP research |
| ------------ | ----------------------------- | --- | --- | ------------------ | --------- | --- | ------ | ------------ |
| etal.,2020). | WeleveragethisabilityofLLMsto | | | | | | | |
communityinrecentyearsasameansofcombat-
generatereasoningprogramsforagivenclaimby ing misinformation and disinformation. Various
showingthemodeljustafewdozenof(claim,pro-
datasetshavebeenproposedthatenablethedevel-
| gram) pairs | as | demonstrations. | PROGRAMFC | is | | | | |
| ----------- | --- | --------------- | --------- | --- | --- | --- | --- | --- |
opmentandtheevaluationofsystemsforautomatic
alsoflexibleasitallowsforeasyswappingofsub- fact-checking,themostpopularonesbeingbased
task function implementations to work under dif- on human-crafted claims from Wikipedia con-
ferentsettingsoffact-checking,withoutaffecting
tent(Thorneetal.,2018;Satheetal.,2020;Schus-
| therestofthesystem. | | Wecanallowthefunctions | | | | | | |
| ------------------- | --- | ---------------------- | --- | ----------------- | --- | --------- | --------- | ------ |
| | | | | ter et al., 2021) | and | naturally | occurring | claims |
to retrieve information from external sources (in inthepoliticalorinthescientificdomain(Wang,
anopen-booksetting)orwecanaskthemtogen-
2017;Nakovetal.,2021b,2022;Augensteinetal.,
erateanswersbasedsolelyontheLLM’sinternal
| | | | | 2019; Saakyan | et al., | 2021; | Gupta | and Srikumar, |
| --- | --- | --- | --- | ------------- | ------- | ----- | ----- | ------------- |
parametricknowledge(inaclosed-booksetting). 2021;Waddenetal.,2020,2022a). Notably,most
We evaluate PROGRAMFC on two challeng- of these datasets are constructed in a way that
ing datasets designed for fact-checking complex the evidence to support or to refute a claim can
single
claims: HOVER(Jiangetal.,2020)andFEVER- be found in a document. For example, in
OUS(Alyetal.,2021),andweshowthatitoutper- FEVER (Thorne et al., 2018), more than 87% of
forms seven few-shot fact-checking baselines on theclaimsonlyrequireinformationfromasingle
| bothdatasets(§4.1). | | | | Wikipediaarticle(Jiangetal.,2020). | | | | |
| ------------------- | --- | --- | --- | ---------------------------------- | --- | --- | --- | --- |
6982
Tobridgethisgap,datasetshavebeenproposed This approach reduces the burden on the lan-
tostudyfact-checkingcomplexclaimsthatrequire guage model and allows for more flexibility
multi-stepreasoning(Jiangetal.,2020;Alyetal., in incorporating necessary components for fact-
2021). Graph-based models (Zhou et al., 2019; checkingsuchasanevidenceretriever. Thestrat-
Liuetal.,2020;Zhongetal.,2020;Nguyenetal., egy of program-guided reasoning is also in line
2020; Barnabò et al., 2022, 2023) are used to fa- withtherecenttrendoftool-augmentedlanguage
cilitate the reasoning over multiple pieces of evi- models(Mialon et al., 2023; Schicket al.,2023),
dence. Althoughsuchmodelsachievesizableper- i.e., augmenting language models with access to
formance gains, they lack explanability and thet externaltoolsandresources.
| relyonlargeamountsoftrainingdata. | | | | | Toaddress | | | | | | | |
| --------------------------------- | --- | --- | --- | --- | --------- | --- | --- | --- | --- | --- | --- | --- |
3 PROGRAMFC
| the above | problems, | we | propose | an explainable, | | | | | | | | |
| --------- | --------- | --- | ------- | --------------- | --- | --- | --- | --- | --- | --- | --- | --- |
flexible, and data-efficient model that generates Wefirstformulatetheproblemoffact-checkingand
| reasoning | graphs | as explanations | | and | utilizes in- | | | | | | | |
| --------- | ------ | --------------- | --- | --- | ------------ | --- | --- | --- | --- | --- | --- | --- |
thenweintroduceourproposedmodelforProgram-
contextlearningtoenablefew-shotlearning.
GuidedFact-Checking(PROGRAMFC).
ExplanationGeneration. Facingthecomplex- 3.1 ProblemFormulation
| ities of real-world | | claims, | simply | giving | a final | | | | | | | |
| ------------------- | --- | ------- | ------ | ------ | ------- | -------------------------------- | --- | --- | --- | --- | ------ | --- |
| | | | | | | GivenaclaimC,afact-checkingmodel | | | | | aimsto | |
F
veracitytoaclaimoftenfailstobepersuasive(Guo predictalabelY toevaluatetheclaimasTRUEor
| et al., 2022). | | Previous | research | has | proposed | | | | | | | |
| ------------------ | ----- | ------------ | -------- | -------- | -------- | ------------------------------------ | --------- | -------- | ------ | --- | ---------- | ------ |
| | | | | | | FALSE,basedonaknowledgesource | | | | | . Themodel | |
| various approaches | | to | provide | post-hoc | explana- | | | | | | K | |
| | | | | | | isalsorequiredtooutputanexplanationE | | | | | | tojus- |
| tions for | model | predictions, | such | as using | atten- | | | | | | | |
| | | | | | | tify the | predicted | veracity | label. | We | summarize | |
tion weights to highlight relevant parts of the ev- threedifferentsettingsoffact-checkingdepending
| idence (Popat | et | al., 2017; | Cui | et al., 2019; | Yang | | | | | | | |
| ------------- | --- | ---------- | --- | ------------- | ---- | -------------------------- | --- | --- | --- | --- | --- | --- |
| | | | | | | onthetypeofknowledgesource | | | | . | | |
K
| et al., 2019; | Lu | and Li, | 2020), | generating | justifi- | | | | | | | |
| ------------------------------------------ | --- | ------- | ------ | ---------- | -------- | ------- | --------- | --------- | ---- | -------- | ------- | --- |
| | | | | | | Gold | evidence: | For | each | claim, | is the | set |
| cationswithlogic-basedsystemsbasedonknowl- | | | | | | • | | | | | K | |
| | | | | | | of gold | evidence | documents | | that can | support | or |
edgegraphs(Gad-Elrabetal.,2019;Ahmadietal.,
| | | | | | | refutetheclaim. | | Thissettingisalsocalledclaim | | | | |
| --- | --- | --- | --- | --- | --- | --------------- | --- | ---------------------------- | --- | --- | --- | --- |
2019),andgeneratingasummaryoftheretrieved
verification(Panetal.,2021;Wrightetal.,2022).
relevantevidence(Atanasovaetal.,2020;Kotonya
| | | | | | | Open-booksetting: | | | isalargetextualcorpus | | | |
| ------------------------------ | --- | --- | --- | ------------- | --- | ----------------- | --- | --------------------------- | --------------------- | --- | --- | --- |
| andToni,2020;Jollyetal.,2022). | | | | Incontrast,we | | | | | | | | |
| | | | | | | • | | | K | | | |
| | | | | | | suchasWikipedia. | | Themodelfirstretrievesrele- | | | | |
proposetousereasoningprogramstoprovideex-
vantevidencefromthecorpusandthenpredictsthe
planationsthatconsistofsub-tasksdescribedina
| | | | | | | veracity | label | based | on the evidence | | (Jiang | et al., |
| ---------------------------- | --- | --- | --- | ----------------- | --- | -------- | ----- | ----- | --------------- | --- | ------ | ------- |
| program-likenaturallanguage. | | | | Thisoffersseveral | | | | | | | | |
2021;Waddenetal.,2022b).
| advantages: | itallowsforexplanationsthatarenot | | | | | | | | | | | |
| --------------------------------------------- | --------------------------------- | --- | --- | --- | --- | ----------------------------------- | ----------- | --- | ------------------- | --- | ------ | ------ |
| | | | | | | Closed-booksetting: | | | Themodeldoesnothave | | | |
| confinedtotheevidence,likeattentionweights,it | | | | | | • | | | | | | |
| | | | | | | accesstoanyexternalknowledgesource( | | | | | | = ). |
| ismoreflexiblethanlogic-basedexplanations,and | | | | | | | | | | | K | ∅ |
| | | | | | | It needs | to leverage | | the knowledge | | stored | in its |
itismoreconcisethanfree-formsummarization.
parameters(acquiredduringpre-trainingandfine-
| | | | | | | tuning) | to verify | the | claim. This | setting | was | ex- |
| ---------------- | --- | ---------- | --- | --------- | --- | ------- | --------- | --- | ----------- | ------- | --- | --- |
| Chain-of-Thought | | Reasoning. | | Moreover, | un- | | | | | | | |
ploredinworkthatapplieslargelanguagemodels
likepreviousworkthatgeneratespost-hocexplana-
forfact-checking(Leeetal.,2020,2021).
tions,wealsousereasoningprogramsasguidance
forpredictingtheveracityoftheclaim. Thisismo- 3.2 Program-GuidedReasoning
| tivated by | the recent | success | of | chain-of-thought | | | | | | | | |
| ---------- | ---------- | ------- | --- | ---------------- | --- | ----------------------------------- | --- | --- | --- | --- | --- | ---- |
| | | | | | | Ourgoalistofact-checkacomplexclaimC | | | | | | that |
prompting(CoT)(Weietal.,2022;Kojimaetal.,
Wefocusonthefew-
requiresmulti-stepreasoning.
2022;Wangetal.,2022),whichgeneratesstep-by-
| | | | | | | shot setting, | whereonlyasmallsetofin-domain | | | | | |
| --- | --- | --- | --- | --- | --- | ------------- | ----------------------------- | --- | --- | --- | --- | --- |
stepnaturallanguagereasoningstepstoguidethe
| | | | | | | examplesareavailabletoteachthemodel. | | | | | Tosolve | |
| --------------------------------- | --- | --- | --- | --- | ------- | ------------------------------------ | --- | --- | --- | --- | ------- | --- |
| modelinansweringcomplexquestions. | | | | | Weadopt | | | | | | | |
this,PROGRAMFCfollowsaprogramgeneration-
| thisideatofact-checkingcomplexclaims. | | | | | Unlike | | | | | | | |
| ------------------------------------- | --- | --- | --- | --- | ------ | --- | --- | --- | --- | --- | --- | --- |
and-executionparadigm,asshowninFigure1.
theoriginalCoT,whichusesasingleLLMforboth
decompositionandquestionanswering,weusethe Program Generation. At this stage, given the
language model only to generate reasoning pro- inputclaimC,aplanner generatesareasoning
P
gramsastheblueprintforproblem-solving,andwe program P = [S , ,S ] for it, which consists
| | | | | | | | | 1 ··· | n | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | ----- | --- | --- | --- | --- |
delegateeachsub-tasktospecializedfunctions. ofnsequentiallyorderedreasoningstepsS i .
6983
Each reasoning step S i P is an instruction However,thegrammarofareasoningprogram
∈
S
in controlled natural language that directs i to is different from the grammar of a programming
a function in an auxiliary set of sub-task func- language. WetakeadvantageofCodex’sfew-shot
tions available to the system. To be specific, generalizationabilityandwefindthatitcanlearn
F
we define S i = (f i ,A i ,V i ), where f i specifies effectivelyfromonlyasmallnumberofin-context
thesub-taskfunctionf , A istheargument examples = d , ,d . Each example d
| | | | i | i | | | | | | 1 | D | | | i |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | ---- | --- | --- | --- |
| | | | ∈ F | | | | | D | { | ··· | | |} | | | |
passed to the function f , and V is the variable consistsofaclaimandaprogram. Theprogramhas
| | | | i | i | | | | | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
thatstoresthereturnedresultfromthefunctioncall aPython-likegrammar,whereeachreasoningstep
f (A ). Foravalidreasoningprogram,thereturn iswrittenintheformatV = f (A ). Atinference
| i i | | | | | | | | | | | i i | i | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
valueofthelastreasoningstepmustbeaBoolean time,wepromptCodexwithaninstructionofthe
valueindicatingtheveracitylabeloftheclaimC, task,K in-contextexamples,andtheinputclaim
i.e.,V n TRUE,FALSE . C. Codex then attempts to complete the follow-
| | ∈ { | | } | | | | | | | | | | | |
| ----------------- | -------- | ----- | ----------------------- | --------- | ----------- | ------- | ----------------------------------------- | -------------------------------- | --- | --- | --- | --- | ----- | --- |
| | | | | | | | ingtexts, | andtherebygeneratesaprogramforC. | | | | | | |
| ProgramExecution. | | | Intheexecutionstage,the | | | | | | | | | | | |
| | | | | | | | TheprompttemplateisshowninFigure2. | | | | | | Weuse | |
| reasoning | program | P | is run | by an | interpreter | to | | | | | | | | |
| | | | | | | | K = 20tomaintainatradeoffbetweenthediver- | | | | | | | |
| derive the | veracity | label | of | the claim | C. | The in- | | | | | | | | |
sityofreasoningtypesandthemodel’smaximum
terpretersequentiallyparsesthereasoningstepsin
| | | | | | | | input | capacity. | Weuse | sampling-based | | | decoding | |
| --------------- | --- | --- | ------- | ------------------- | --- | --- | ----- | --------- | ----- | -------------- | --- | --- | -------- | --- |
| P. ForeachstepS | | | = (f ,A | ,V ),itcallsthecor- | | | | | | | | | | |
| | | i | i | i i | | | | | | | | | | |
(temperatureof0.7)togeneratedifferentreasoning
| responding | off-the-shelf | | sub-task | function | | f and | | | | | | | | |
| ---------- | ------------- | --- | -------- | -------- | --- | ----- | --- | --- | --- | --- | --- | --- | --- | --- |
i
programsformultipleruns.
| passestheargumentA | | | toit. | TheargumentA | | isei- | | | | | | | | |
| ------------------ | --- | --- | ----- | ------------ | --- | ----- | --- | --- | --- | --- | --- | --- | --- | --- |
| | | | i | | | i | | | | | | | | |
theralogicalexpressionoranaturallanguagesen- 3.4 Sub-TaskFunctions
| tence,e.g.,aquestionorasimpleclaim. | | | | | Theresult | | | | | | | | | |
| ------------------------------------------- | ------ | ---- | ------------ | ------ | --------- | ------ | ------------------------------------- | --- | -------------------------------- | -------- | --- | --------- | --- | --- |
| | | | | | | | We implement | | three | sub-task | | functions | for | the |
| ofthefunctioncallisthenstoredinthevariableV | | | | | | | . | | | | | | | |
| | | | | | | i | modeltocallduringtheprogramexecution. | | | | | | | |
| As it is | common | for | a subsequent | step | to | depend | | | | | | | | |
| | | | | | | | QUESTION: | | Thissub-taskfunctionisaquestion- | | | | | |
| on the results | | from | previous | steps, | we allow | the | | | | | | | | |
•
| | | | | | | | answering | module | | that takes | a | question | Q as | the |
| -------------- | --------------------------------- | ----------------------------- | ------------ | --- | ---- | --- | --------- | -------------------------------- | --- | ---------- | --- | -------- | ---- | --- |
| argument | A to | refer | to variables | V | , ,V | in | | | | | | | | |
| | i | | | | 1 | i 1 | | | | | | | | |
| | | | | | ··· | − | input | argument | and | returns | the | answer | A to | the |
| previoussteps. | | Forexample,inFigure1,theargu- | | | | | | | | | | | | |
| | | | | | | | question. | WeuseFLAN-T5(Chungetal.,2022),an | | | | | | |
| mentinS | 3 is“{ANSWER_1}wasborninCanada.”, | | | | | | | | | | | | | |
improvedT5model(Raffeletal.,2020)pretrained
| which refers | to | the return | variable | | {ANSWER_1} | | | | | | | | | |
| ------------ | ------ | ---------- | -------- | ------- | ---------- | ------ | ------- | ---- | ---- | ----- | ---------------- | --- | ------- | --- |
| | | | | | | | on more | than | 1.8K | tasks | with instruction | | tuning, | |
| from S | . When | executing | | S , the | variable | is re- | | | | | | | | |
| 2 | | | | 3 | | | | | | | | | | |
whichhasachievedstate-of-the-artzero/few-shot
| placed | by its | actual | value, | and the | argument | be- | | | | | | | | |
| ------ | ------ | ------ | ------ | ------- | -------- | --- | ------------------------------ | --- | --- | --- | --- | --- | ------- | --- |
| | | | | | | | performanceonmanyQAbenchmarks. | | | | | | Asshown | |
comes“ChristopherNolanwasborninCanada”.
| | | | | | | | in Figure | 3, | we prompt | | the model | differently | | de- |
| --- | --- | --- | --- | --- | --- | --- | --------- | --- | --------- | --- | --------- | ----------- | --- | --- |
Afterexecutingthelaststep,thereturnvalueisthe
| | | | | | | | pendingonthesettingsdefinedinSection3.1. | | | | | | | For |
| --- | --- | --- | --- | --- | --- | --- | ---------------------------------------- | --- | --- | --- | --- | --- | --- | --- |
predictedveracityoftheclaimC.
theclosed-booksetting,theinputpromptis
| AggregatingReasoningPaths. | | | | Notethatthere | | | | | | | | | | |
| -------------------------- | -------- | --------- | --- | ------------- | -------- | ----- | --- | --------- | --- | --- | ------ | --- | --- | --- |
| | | | | | | | Q: | QUESTION? | | The | answer | is: | | |
| might be | multiple | reasoning | | paths | that can | reach | | | | | | | | |
Fortheothertwosettings,theinputpromptis
| the final | veracity | label. | Therefore, | | we generate | | | | | | | | | |
| --------- | -------- | ------ | ---------- | --- | ----------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
a diverse set of N candidate reasoning programs EVIDENCE Q: QUESTION?
| | | | | | | | The | answer | | is: | | | | |
| ------------------- | ------ | ------------------- | ---------------------- | --- | --------- | --- | ------- | ------ | --------------------------------- | --- | --- | --- | --- | --- |
| = P | 1 , ,P | N fortheinputclaim. | | | Afterexe- | | | | | | | | | |
| P { | ··· | } | | | | | | | | | | | | |
| cutingallprogramsin | | | ,wetakethemajorityvote | | | | | | | | | | | |
| | | | | | | | VERIFY: | | Thisisafactverificationmodulethat | | | | | |
P
| overallN | predictedlabelsasthefinallabel. | | | | | This | • | | | | | | | |
| -------- | ------------------------------- | --- | --- | --- | --- | ---- | ------------ | --- | ---------------------------- | --- | --- | --- | --- | --- |
| | | | | | | | takesaclaimC | | astheinputargumentandreturns | | | | | |
approachissimilartohowhumansrelyonmultiple
| | | | | | | | a label | of either | TRUE | | or FALSE. | We | also | use |
| --- | --- | --- | --- | --- | --- | --- | ------- | --------- | ---- | --- | --------- | --- | ---- | --- |
methodsofvalidationtoincreasetheirconfidence FLAN-T5forthismodule,bypromptingthemodel
| infact-checking. | | Italsomakesthemodellesssus- | | | | | | | | | | | | |
| ---------------- | --- | --------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
withthefollowingquestion-answeringformat.
ceptibletoerrorsinindividualreasoningprograms.
EVIDENCE
| 3.3 ReasoningProgramGeneration | | | | | | | Q: | Is | it true | that | CLAIM? | | | |
| ------------------------------ | ----------- | --- | --------- | --- | ----- | ----- | ---- | --- | ------- | ---- | ------ | --- | --- | --- |
| | | | | | | | True | or | False? | The | answer | is: | | |
| We base | our program | | generator | on | Codex | (Chen | | | | | | | | |
et al., 2021), a code-pretrained LLM, which can PREDICT: Thismoduletakesasinputalogical
•
parse natural language into symbolic representa- expressionthatperformsAND,OR,NOToperations
tionssuchasSQL(Chengetal.,2022)orPython overthevariablesintheprevioussteps. Itsoutput
programs(Gaoetal.,2022;Chenetal.,2022b). isreturnedasthepredictedveracitylabel.
6984
'''Generate a python-like program that describes the reasoning steps
required to verify the claim step-by-step. You can call three functions
in the program: 1. Question() to answer a question; 2. Verify() to
verify a simple claim; 3. Predict() to predict the veracity label.'''
# The claim is that Both James Cameron and the director of the film
| | Interstellar | | were | born | in | Canada. | | | | | | |
| --- | ------------ | --- | ------------- | ---- | ------- | ------- | ---- | --- | --------- | --- | --- | --- |
| def | program(): | | | | | | | | | | | |
| | fact_1 | = | Verify("James | | Cameron | was | born | in | Canada.") | | | |
Answer_1 = Question("Who is the director of the film Interstellar?")
| | fact_2 | = | Verify("{Answer_1} | | | was | born | in Canada.") | | | | |
| --- | ---------- | ---------------- | ------------------ | ------------- | ---- | ------- | ---- | ------------ | --- | --- | --- | --- |
| | label | = Predict(fact_1 | | | and | fact_2) | | | | | | |
| ( | more | in-context | examples | | here | ) | | | | | | |
| | ··· | | | | | ··· | | | | | | |
| # | The claim | is | that | <input_claim> | | | | | | | | |
| def | program(): | | | | | | | | | | | |
Figure 2: The Codex prompt template used to generate reasoning programs, consisting of a task instruction,
in-contextexamples,andapromptforthe<input_claim>. ThefulltemplatesaregiveninAppendixD.
<Gold Evidence>
Forevaluationintheopen-booksetting,weuse
| Gold Evidence | | Q: <Question> | | | | | | | | | | |
| ------------- | --- | ------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
The answer is: the corresponding Wikipedia corpus constructed
| | Open-book | | | | | | for | these | two datasets | as the knowledge | sources. | |
| --- | --------- | --- | --- | --- | --- | --- | --- | ----- | ------------ | ---------------- | -------- | --- |
<Retrieved Evidence>
<Question> Q: <Question> FLAN-T5 Ans HOVER uses the October 2017 Wikipedia dump
The answer is:
| | Retriever | | | | | | processed | | by Yang | et al. (2018), | consisting | of |
| --- | --------- | --- | --- | --- | --- | --- | --------- | --- | ------- | -------------- | ---------- | --- |
theintroductorysectionsof5.2millionWikipedia
| Closed-book | | Q: <Question> | | | | | | | | | | |
| ----------- | --- | ------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
The answer is:
| | | | | | | | pages. | FEVEROUSusestheDecember2020dump, | | | | |
| --- | --- | --- | --- | --- | --- | --- | ------ | -------------------------------- | --- | --- | --- | --- |
Figure 3: Implementation of the question-answering including5.4millionfullWikipediaarticles.
sub-taskfunctionforthreedifferentsettings.
| | | | | | | | Baselines. | | Wecompare | PROGRAMFC | toseven | |
| ------------- | ------------------------------------ | --- | --- | --- | --- | --- | -------------- | -------- | ---------------------------- | --------------- | ------------- | -------- |
| | | | | | | | baselines, | | categorized | into three | groups. | (i) Pre- |
| 4 Experiments | | | | | | | trainedmodels: | | BERT-FC(Soleimanietal.,2020) | | | |
| | | | | | | | and | LisT5 | (Jiang | et al., 2021) | are two | models |
| Datasets. | Mostfact-checkingdatasetsconsistpri- | | | | | | | | | | | |
| | | | | | | | that | leverage | BERT | and T5 for fact | verification, | |
marilyofsimpleclaimsthatcanbesubstantiated
| | | | | | | | respectively. | | (ii) FC/NLI | fine-tuned | models: | we |
| ------------------------------ | --- | --- | --- | --- | ------------ | --- | ------------- | --- | ----------- | ---------- | ------- | --- |
| throughasinglepieceofevidence. | | | | | However,here | | | | | | | |
choosethreepretrainedmodelsthatarefine-tuned
| wefocusoncomplexclaimsthatneedmulti-step | | | | | | | | | | 1 | | |
| ---------------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
onotherfact-checkingdatasetsornaturallanguage
| reasoning. | Given | this | context, | we | opt | to evalu- | | | | | | |
| ---------- | ----- | ---- | -------- | --- | --- | --------- | ----------------------- | --- | --- | --------------------- | --- | --- |
| | | | | | | | inference(NLI)datasets. | | | RoBERTa-NLI(Nieetal., | | |
ate our model on the only two datasets that, to 2020)usesfine-tunedRoBERTa-largeonfourNLI
| the best | of our | knowledge, | | fulfill | these | criteria: | | | | | | |
| -------- | ------ | ---------- | --- | ------- | ----- | --------- | --------- | --- | ------------- | ------ | ---------- | ----- |
| | | | | | | | datasets; | | DeBERTaV3-NLI | (He et | al., 2021) | fine- |
HOVER(Jiangetal.,2020)andFEVEROUS(Aly
tunestheDeBERTaV3modelon885,242(claim,
| etal.,2021). | Weusethevalidationsetsforevalu- | | | | | | | | | | | |
| ------------ | ------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
evidence,label)annotationsfromFEVERandfour
| ation | since the | test sets | are | not publicly | | released. | | | | | | |
| ----- | --------- | --------- | ---- | ------------ | ----------- | --------- | --- | --------- | --------- | ------- | ------- | ------ |
| | | | | | | | NLI | datasets. | MULTIVERS | (Wadden | et al., | 2022b) |
| HOVER | contains | claims | that | require | integration | | | | | | | |
isaLongFormer(Beltagyetal.,2020)modelfine-
| andreasoningovermultipleWikipediaarticles. | | | | | | We | | | | | | |
| ------------------------------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
tunedonFEVER.(iii)In-contextlearningmodels:
divideitsvalidationsetintothreesubsetsbasedon one baseline is that we directly use the FLAN-T5
thenumberof“hops”requiredtoverifytheclaim:
| | | | | | | | modelinourVERIFY | | | moduleforfact-checking. | | |
| --- | --- | --- | --- | --- | --- | --- | ---------------- | --- | --- | ----------------------- | --- | --- |
1,126two-hopclaims,1,835three-hopclaims,and
Theotherbaselineusesthein-contextlearningof
1,039 four-hop claims. FEVEROUS focuses on Codexforfew-shotfact-checking.
Theimplemen-
| fact-checking | | complex | claims | over | unstructured | | | | | | | |
| ------------- | --- | ------- | ------ | ---- | ------------ | --- | --- | --- | --- | --- | --- | --- |
tationdetailsaregiveninAppendixA.
andstructureddata,whereeachclaimisannotated
withevidenceintheformofsentencesand/orcells Few-ShotLearning. Westudyfew-shotlearning
fromtablesinWikipedia. Sincewefocusontextual whereonlyafewin-domainexamplesareavailable.
fact-checking,weonlyselectedclaimsthatrequire Therefore, for a fair comparison, we restrict all
exclusivelysentenceevidence,constituting2,962 models to have access to only 20 examples from
| claims. | WecallthissubsetFEVEROUS-S. | | | | | | HOVERorFEVEROUS-S. | | | | | |
| ------- | --------------------------- | --- | --- | --- | --- | --- | ------------------ | --- | --- | --- | --- | --- |
6985
| | | HOVER(2-hop) | HOVER(3-hop) | HOVER(4-hop) | FEVEROUS-S | |
| --- | --- | ------------ | ------------ | ------------ | ---------- | --- |
Few-shotlearningmodels
| | | Gold Open | Gold | Open Gold | Open Gold | Open |
| --- | --- | --------- | ---- | --------- | --------- | ---- |
BERT-FC(Soleimanietal.,2020) 53.40 50.68 50.90 49.86 50.86 48.57 74.71 51.67
I
LisT5(Jiangetal.,2021) 56.15 52.56 53.76 51.89 51.67 50.46 77.88 54.15
RoBERTa-NLI(Nieetal.,2020) 74.62 63.62 62.23 53.99 57.98 52.40 88.28 57.80
II DeBERTaV3-NLI(Heetal.,2021) 77.22 68.72 65.98 60.76 60.49 56.00 91.98 58.81
MULTIVERS(Waddenetal.,2022b) 68.86 60.17 59.87 52.55 55.67 51.86 86.03 56.61
Codex(Chenetal.,2021)
| | | 70.63 65.07 | 66.46 | 56.63 63.49 | 57.27 89.77 | 62.58 |
| --- | --- | ----------- | ----- | ----------- | ----------- | ----- |
III
FLAN-T5(Chungetal.,2022) 73.69 69.02 65.66 60.23 58.08 55.42 90.81 63.73
| ProgramFC | (N=1) | | | | | |
| --------- | ----- | ----------- | ----- | ----------- | ----------- | ----- |
| | | 74.10 69.36 | 66.13 | 60.63 65.69 | 59.16 91.77 | 67.80 |
IV
ProgramFC (N=5) 75.65 70.30 68.48 63.43 66.75 57.74 92.69 68.06
Table 1: Macro-F1 scores of PROGRAMFC (IV) and baselines (I-III) on the evaluation set of HOVER and
FEVEROUS-Sforfew-shotfact-checking. GoldandOpenrepresentthegoldevidencesettingandtheopenbook
setting,respectively. I:pretrainedTransformers;II:FC/NLIfine-tunedmodels;III:in-contextlearningmodels.
We use these examples either for fine-tuning However,thisgeneralizationbecomesmorechal-
pre-trainedmodels(BERT-FCandLisT5),forcon- lengingasthecomplexityoftheclaimsincreases.
tinuousfine-tuningtheFC/NLIfine-tunedmodels, OnHOVER,theF1scoreofDeBERTaV3-NLIdrops
orasin-contextexamplesforFLAN-T5andCodex.
| | | | from 77.22 | for 2-hop claims | to 60.49 | for 4-hop |
| --- | --- | --- | ---------- | ---------------- | -------- | --------- |
For PROGRAMFC,weusethemasin-contextex- claims,whichisadecreaseof21.7%. Incontrast,
amplesforreasoningprogramgeneration. theperformancedropforProgramFC,whichuses
Weevaluateboththegoldevidencesettingand thestrategyofprogram-guidedreasoning,ismuch
smaller: just11.7%.
theopen-booksetting. Thebaselinemodelsarethe
| same for | both settings. However, | during testing | | | | |
| -------- | ----------------------- | -------------- | ------------- | ------- | -------------- | -------- |
| | | | Decomposition | is more | effective than | one-step |
intheopen-booksetting,themodelsaregiventhe
| | | | prediction. | The ProgramFC | model, which | uses |
| --- | --- | --- | ----------- | ------------- | ------------ | ---- |
retrievedevidenceratherthantheground-truthev-
| | | | the same | FLAN-T5 model | as the sub-task | func- |
| ------- | ---------------------- | ------------- | -------- | ------------- | --------------- | ----- |
| idence. | We use BM25 (Robertson | and Zaragoza, | | | | |
tions, outperformsthebaselineofdirectlyverify-
2009)implementedwiththePyserinitoolkit(Lin
| | | | ingclaimswithFLAN-T5onallfourdatasets. | | | On |
| --- | --- | --- | -------------------------------------- | --- | --- | --- |
etal.,2021)astheretrieverforbothPROGRAMFC
average,thereisa6.0%improvementinthegold
| andthebaselines. | Weuseasevidencethetop-10 | | | | | |
| ---------------- | ------------------------ | --- | -------- | ------------------ | ----------- | ------ |
| | | | evidence | setting and a 4.5% | improvement | in the |
paragraphsretrievedfromtheknowledgecorpus.
| | | | open-booksetting. | Thissuggeststhatdecomposing | | |
| --- | --- | --- | ----------------- | --------------------------- | --- | --- |
acomplexclaimintosimplerstepswithaprogram
4.1 MainResults
| | | | canfacilitatemoreaccuratereasoning. | | | Thisises- |
| --- | --- | --- | ----------------------------------- | --- | --- | --------- |
WereporttheoverallresultsforPROGRAMFCand pecially evident when the required reasoning is
forthebaselinesforfew-shotfact-checkinginTa- complex: thereisa14.9%improvementinthegold
ble1. PROGRAMFCachievesthebestperformance
| | | | evidence | setting and a 6.7% | improvement | in the |
| --- | --- | --- | -------- | ------------------ | ----------- | ------ |
on7outof8evaluations,demonstratingitseffec- open-booksettingfor4-hopclaims.
| tiveness. | Wehavethreemorespecificobservations. | | | | | |
| --------- | ------------------------------------ | --- | --- | --- | --- | --- |
Aggregatingreasoningprogramsishelpful.
ProgramFCismoreeffectiveondeeperclaims.
On the HOVER dataset, ProgramFC (N=5) out- WefindthataggregatingthepredictionsofN = 5
performs the baselines on average by 10.38%, reasoningprogramsimprovestheperformanceover
11.37%, and 14.77% on two-hop, three-hop, and using a single program by an average of 1.5%.
four-hop claims, respectively. This suggests that ThisalignswiththefindingsofWangetal.(2022),
ProgramFCbecomesincreasinglyeffectiveasthe wheretheideawasappliedforquestionanswering:
required reasoning depth increases. Among the if multiple different ways of thinking lead to the
DeBERTaV3-NLI
baselines, performs comparably sameanswer,wecanhavegreaterconfidencethat
to ProgramFC on two-hop claims, indicating that thefinalansweriscorrect. Thisintuitionalsoap-
large-scalepre-trainingonsimplerclaimscanhelp pliestofact-checking,aseachprogramrepresents
themodelgeneralizetomorecomplexclaims. auniquereasoningchaintoverifytheclaim.
6986
FLAN-T5 ProgramFC HOVER (2-hop) FLAN-T5 ProgramFC HOVER (3-hop) FLAN-T5 ProgramFC HOVER (4-hop)
| 80 | | 76.11 | 75.65 | 77.62 | 80 | | | 80 | | | |
| --- | --- | ----- | ----- | ----- | --- | --- | --- | --- | --- | --- | --- |
72.56
77.07
| 70 | | | 73.69 | | 70 67.88 | 68.55 68.48 | 69.56 | 70 | 68.37 | 68.56 | 68.18 |
| ----- | ----- | ----- | ----- | --- | -------- | ----------- | ----- | ----- | ----- | ----- | ----- |
| 64.35 | | 71.69 | | | | | | | | 66.75 | |
| | 68.24 | | | | 62.23 | | | 62.46 | | | |
| | | | | | | 65.07 65.66 | 66.89 | | | | |
| 60 | | | | | 60 63.05 | | | 60 | | | 63.39 |
61.36
58.08
56.58
| 50 | | | | | 50 | | | 50 | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
49.29
| 47.75 | | | | | | | | 48.59 | | | |
| ----- | --- | --- | --- | --- | --- | --- | --- | ----- | --- | --- | --- |
| 40 | | | | | 40 | | | 40 | | | |
80M 250M 780M 3B 11B 80M 250M 780M 3B 11B 80M 250M 780M 3B 11B
F1scoreforfact-checkingwithgoldevidenceusingFLAN-T5(blueline)andPROGRAMFC(greenline)
Figure4:
forlanguagemodelsofincreasingsizes: FLAN-T5-small(80M),FLAN-T5-base(250M),FLAN-large(780M),
FLAN-T5-XL(3B),andFLAN-T5-XXL(11B)onHOVER2-hop(left),3-hop(middle),and4-hop(right).
1
| 90 | | | | | | We | measure | the recall | of the | gold paragraphs | |
| --- | ----- | ------------------ | --- | --------- | ----- | ------- | ------- | ---------- | ---------- | --------------- | --- |
| | | One-step Retrieval | | ProgramFC | | | | | | | |
| 80 | | | | | 85.65 | | | | | | |
| | | | | | | for the | top-10 | retrieved | paragraphs | (recall@10). | |
| | 77.13 | | | | 76.25 | | | | | | |
70 73.18
For PROGRAMFC,wecombinetheretrievedpara-
60
graphsofallstepsandweconsiderthetop-10re-
59.17
50
| | | 51.33 | | 49.93 | | sults. | We can | see in | Figure 5 that | PROGRAMFC | |
| --- | --- | ----- | --- | ----- | --- | ------ | ------ | ------ | ------------- | --------- | --- |
40
outperformsone-stepretrievalonalldatasets,with
| 30 | | | | 36.43 | | | | | | | |
| --- | --- | --- | --- | ----- | --- | ----------- | ----------- | --- | -------- | -------- | --- |
| | | | | | | the largest | improvement | | of 37.1% | on HOVER | 4- |
20
HOVER (2-hop) HOVER (3-hop) HOVER (4-hop) FEVEROUS-S hop. Thisisbecausesomeinformationmaynotbe
presentintheoriginalclaim,butisonlyrevealed
Figure5: Retrievalrecall@10fortheone-stepretrieval
andtheiterativeretrievalinPROGRAMFC. during the reasoning process (e.g., “Christopher
| | | | | | | Nolan”inFigure1). | | Thus,iterativeretrievalguided | | | |
| --- | --- | --- | --- | --- | --- | ----------------- | --- | ----------------------------- | --- | --- | --- |
bythereasoningprogramyieldsbetterresults.
4.2 HowDoestheReasoningProgramHelp?
4.3 InterpretabilityofReasoningPrograms
| To further | understand | | how | reasoning | programs | | | | | | |
| ---------- | ---------- | --- | --- | --------- | -------- | --- | --- | --- | --- | --- | --- |
facilitate fact-checking, we compare the perfor- Anadvantageof PROGRAMFC isthatitimproves
FLAN-T5
mance of PROGRAMFC with using dif- the interpretability of fact-checking compared to
| ferentlanguagemodelsizes: | | | | small,base,large, | | | | | | | |
| ------------------------- | --- | --- | --- | ----------------- | --- | --- | --- | --- | --- | --- | --- |
end-to-endmodels,astheexplicitprogramcanaid
XL, and XXL. The results are shown in Figure 4 human understanding and debugging. Examples
andindicatethatprogram-guidedreasoningispar- ofgeneratedreasoningprogramscanbefoundin
| ticularly | effective | when | the | model | size is small. | | | | | | |
| --------- | --------- | ---- | --- | ----- | -------------- | ------ | ------------- | --- | ------------ | ----------- | --- |
| | | | | | | Figure | 7 of Appendix | | B. To assess | the quality | of |
As smaller models have less capacity for com- thegeneratedreasoningprograms,wesampled300
plexreasoning,theperformanceoftheend-to-end claimswhere PROGRAMFC incorrectlypredicted
FLAN-T5
model decreases significantly with de- the final veracity labels from the HOVER 2-hop,
creasing model size. However, this trend is less 3-hop,and4-hopdatasets,with100examplesper
notableforPROGRAMFC.Thehigh-levelreason- dataset. Weaskedhumanannotatorstoanalyzethe
ing plan offered by reasoning programs substan- errortypesandweclassifiedtheresultsintothree
tiallyalleviatesthedemandsonthesubsequentsub-
| | | | | | | categories: | (i)Syntacticerrors,wheretheprogram | | | | |
| --- | --- | --- | --- | --- | --- | ----------- | ---------------------------------- | --- | --- | --- | --- |
task solvers. Our results show that the program- doesnotconformtothedefinedgrammarandcan-
guidedmodelusingFLAN-T5-small(80Mparam- notbeparsed,(ii)Semanticerrors,whichinclude
eters)assub-tasksolverscanachievecomparable
incorrectormissingarguments/variables(Token),
performancetothe137xlargerFLAN-T5-XXL(11B) incorrectprogramstructure(Structure),andincor-
modelwithend-to-endreasoningfor4-hopclaims. rect sub-task calls (Subtask), and (iii) Incorrect
Intheopen-domainsetting,wefindthatreason- execution,wheretheprogramiscorrect,butwhere
ingprogramscanenhancetheretrievalofrelevant theincorrectpredictionisaresultofitsexecution.
evidence from the knowledge source. Figure 5 We show the error analysis in Table 2. First,
comparestheretrievalperformanceoftheone-step nosyntaxerrorswerefoundinoursamples,indi-
BM25retrieverusedinthebaselinestotheiterative catingthatCodexeffectivelygeneratesexecutable
step-by-stepBM25retrieverin PROGRAMFC. programsthroughfew-shotin-contextlearning.
6987
Claim:
Emery, located in the same state as Edison Local School District, is a ghost town. It is near the
city that lies close to the Ohio Turnpike, a 241.26 mi highway.
Predicted Program:
answer_1 = Question("Which state is Emery located in?")
answer_2 = Question("Which state is Edison Local School District located in?")
fact_1 = Verify("{answer_1} and {answer_2} are the same state.")
fact_2 = Verify("Emery is a ghost town.")
answer_3 = Question("Which city is near Emery?")
answer_4 = Question("Which city lies close to the Ohio Turnpike, a 241.26 mi highway?")
fact_3 = Verify("{answer_3} is near {answer_4}.") fact_3 = Verify(“Emery is near {answer_4}.”)
label = Predict(fact_1 and fact_2 and fact_3)
Figure6: AnerrorcasefromtheHOVER4-hopdatasetwherethegeneratedreasoningprogramhasanincorrect
programstructure. Theincorrectsegment(s)aremarkedinred,andthecorrectrevisionsaremarkedingreen.
1
| | | | | Proportion(%) | | | | | HOVER | | |
| --- | --- | --- | --- | ------------- | --- | --- | --- | --- | ----- | --- | --- |
ErrorType
| | | | 2-hop | 3-hop | 4-hop | | Model | | | | FEVEROUS |
| --- | ------------- | --- | ----- | ----- | ----- | --- | ----------- | --- | ----------- | ----- | -------- |
| | | | | | | | | | 2-hop 3-hop | 4-hop | |
| | Syntaxerror | | 0% | 0% | 0% | | InstructGPT | | | | |
| | Semanticerror | | 29% | 38% | 77% | | | | | | |
| | | | | | | | - Direct | | 56.51 51.75 | 49.68 | 60.13 |
| | Token | | 8% | 20% | | 18% | | | | | |
| | | | | | | | - ZS-CoT | | 50.30 52.30 | 51.58 | 54.78 |
| | Structure | | 19% | 13% | | 57% | | | | | |
- CoT
| | | | | | | | | | 57.20 53.66 | 51.83 | 61.05 |
| --- | ------------------ | --- | --- | --- | --- | --- | ---------- | --- | ----------- | ----- | ----- |
| | Subtask | | 2% | 5% | | 2% | | | | | |
| | | | | | | | - Self-Ask | | 51.54 51.47 | 52.45 | 56.82 |
| | Incorrectexecution | | 71% | 62% | 23% | | | | | | |
Codex
| | | | | | | | | | 55.57 53.42 | 45.59 | 57.85 |
| --- | --- | --- | --- | --- | --- | --- | ------- | --- | ----------- | ----- | ----- |
| | | | | | | | FLAN-T5 | | 48.27 52.11 | 51.13 | 55.16 |
Table2: Reasoningprogramevaluationforincorrectly-
predictedexamplesfromeachhoplengthinHOVER. ProgramFC 54.27 59.66
| | | | | | | | | | 54.18 | 52.88 | |
| --- | ------- | --------- | ------- | ------- | ---- | ------ | --------------------------- | --- | ----- | --------------------- | --- |
| | | | | | | | Table3: Closed-booksetting: | | | macro-F1scoresforPRO- | |
| | Second, | for 2-hop | claims, | we find | that | 71% of | | | | | |
GRAMFCandforthebaselines.
| the | programs | are | correct. | The majority | of | the er- | | | | | |
| --- | -------- | --- | -------- | ------------ | --- | ------- | --- | --- | --- | --- | --- |
rorsaretheresultofincorrectprogramexecution,
| | | | | | | | We also | include | the 175B-parameter | | Instruct- |
| --- | --- | --- | --- | --- | --- | --- | ------- | ------- | ------------------ | --- | --------- |
wherethequestionansweringorthefact-checking
modulesfailedtoreturnthecorrectanswer. GPT (text-davinci-002) (Ouyang et al., 2022)
Third,asthecomplexityoftheclaimsincreases, with four different prompts: (i) direct prompt-
| | | | | | | | ing with | the claim, | (ii) | CoT (Wei | et al., 2022) or |
| --- | --- | --- | --- | --- | --- | --- | -------- | ---------- | ---- | -------- | ---------------- |
theproportionofsemanticerrorsintheprograms
chain-of-thoughtpromptingwithdemonstrations,
alsoincreases,withstructuralerrorsbecomingpar-
ticularlyprevalent. Thishighlightsthedifficultyof (iii) ZS-CoT (Kojima et al., 2022) or zero-shot
generatingtheappropriatestep-by-stepreasoning chain-of-thoughtwiththeprompt“let’sthinkstep
| | | | | | | | by step”, | and (iv) | Self-Ask | (Press | et al., 2022), |
| ---------- | --- | ---------- | ---- | ------------------ | --- | ---- | --------- | -------- | -------- | ------ | -------------- |
| strategies | | for claims | that | require long-chain | | rea- | | | | | |
soning. An example structural error is shown in whichisavariantofCoTthatguidesthemodelrea-
Figure6,wherethemodelfailstoparsethesecond soningbyaskingaseriesofquestions. Thedetailed
promptingtemplatesaregiveninAppendixE.
sentenceoftheclaimintocorrectprograminstruc-
Ourresults,presentedinTable3,showthatmost
| tions. | Additional | | errorexamples | can | befound | in | | | | | |
| ---------- | ---------- | --- | ------------- | --- | ------- | --- | -------------- | -------- | ---------- | ------------ | ------------- |
| AppendixC. | | | | | | | models achieve | | a Macro-F1 | score | only slightly |
| | | | | | | | above random | guessing | | on the HOVER | dataset, |
4.4 Closed-BookFact-Checking
indicatingthedifficultyofsolelyrelyingonpara-
Finally,weevaluatetheclosed-booksetting,where metric knowledge of large language models for
themodeldoesnothaveaccesstoanyknowledge fact-checkingcomplexclaims. Similartotheobser-
sourceandneedstorelyonitsparametricknowl- vationsinSection4.1,weseeatrendofimproved
edgeonly. ThebaselinemodelsfromgroupsIand performance as the number of the required rea-
II in Table 1 are trained with (evidence, claim) soninghopsincreases. Chain-of-thoughtprompt-
pairs and thus are not applicable in this setting. ingscoresanaverage2.7pointshigherthandirect
Wecompareourmethodtothebaselinesthatuse prompting, highlighting the importance of step-
largelanguagemodelsforin-contextlearning,in- by-step reasoning for complex fact-checking. It
cludingCodex(code-davinci-002)andFLAN-T5 outperformsourPROGRAMFConHOVER2-hop
| fromTable1. | | | | | | | andFEVEROUSbutperformsworseonHOVER | | | | |
| ----------- | --- | --- | --- | --- | --- | --- | ---------------------------------- | --- | --- | --- | --- |
6988
3-hopand4-hop. Generating reasoning programs for such implicit
ThiscanbeduetoCoTgeneratingfree-formex- complex claims requires a deeper understanding
planations,whichcanleadtounpredictableerrors of the claim and also access to world and com-
inlongreasoningchains. Incontrast,ourprogram monsenseknowledge. Weconductedpreliminary
generation-and-execution strategy is more stable experimentsonthesetypesofclaims,butwefound
forlongerreasoningchains. that our Codex-based generator struggled to pro-
| | | | | | | | duceacorrectreasoningprogram. | | | | Thishighlights |
| --- | --- | --- | --- | --- | --- | --- | ----------------------------- | --- | --- | --- | -------------- |
5 ConclusionandFutureWork thegapinapplyingourPROGRAMFCtofact-check
| | | | | | | | real-worldclaims. | | Addressingthesechallengesis | | |
| ----------- | --- | ---------- | --- | ---------- | --- | ------ | ----------------- | --- | --------------------------- | --- | --- |
| We proposed | | PROGRAMFC, | | a few-shot | | neuro- | | | | | |
animportantdirectionforfuturework.
| symbolic | model | for | fact-checking | | that | learns to | | | | | |
| -------- | ----- | --- | ------------- | --- | ---- | --------- | ------- | --------- | --- | --------------------- | --- |
| | | | | | | | Second, | PROGRAMFC | | incursahighercomputa- | |
mapinputclaimstoareasoningprogramconsisting
tionalcostthanbaselineend-to-endfact-checking
ofasequenceofsub-taskfunctioncallsforanswer-
| | | | | | | | models. | Itrequirescallinglargelanguagemodels | | | |
| --------------- | --- | ----------------- | --- | --- | ------ | ------ | ------- | ------------------------------------ | --- | --- | --- |
| ing a question, | | for fact-checking | | a | simple | claim, | | | | | |
forprogramgenerationandfurthercallingmultiple
| andforcomputingalogicalexpression. | | | | | Thenfact- | | | | | | |
| ---------------------------------- | --- | --- | --- | --- | --------- | --- | --------------- | --- | ---------------------------- | --- | --- |
| | | | | | | | sub-taskmodels. | | Thisresultsintheactualcompu- | | |
checkingisperformedbyexecutingthatprogram.
| | | | | | | | tationaltimethatis | | | 4–5 higherthanforanend- | |
| --------------- | --- | -------- | ------------------ | ---------- | --- | -------- | ------------------- | --- | --- | ----------------------- | --- |
| PROGRAMFC | | combines | the | advantages | | of sym- | | | ∼ | × | |
| | | | | | | | to-endFLAN-T5model. | | | Developingmoreefficient | |
| bolic programs, | | such | as explainability, | | | with the | | | | | |
methodsforprogramgenerationandexecutionis
| flexibility | of | end-to-end | neural | models. | | Using | | | | | |
| ----------- | --- | ---------- | ------ | ------- | --- | ----- | --- | --- | --- | --- | --- |
animportantdirectionforfuturework.
| Codex as | the | program | generator, | | PROGRAMFC | | | | | | |
| --------------------------------------- | --- | ------- | ---------- | --- | --------- | --- | --------------- | --- | --- | --- | --- |
| demonstratespromisingperformanceonHOVER | | | | | | | EthicsStatement | | | | |
andFEVEROUSwithonlyasmallnumberofin-
contextdemonstrationsandnoadditionaltraining. Biases. Wenotethattheremightbesomebiases
| | | | | | | | in the data | used | to train | the LLMs, | as well as in |
| --- | --- | --- | --- | --- | --- | --- | ----------- | ---- | -------- | --------- | ------------- |
Wealsoinvestigatedtheimpactofmodelsizeand
| | | | | | | | factualityjudgments. | | | Botharebeyondourcontrol. | |
| ---------------- | --- | ----------------------- | --- | ---------- | --- | ------ | -------------------- | --- | --- | ------------------------ | --- |
| the benefits | of | programs | for | retrieval, | and | we an- | | | | | |
| alyzedtheerrors. | | Theresultsindicatedthat | | | | PRO- | | | | | |
Ourmod-
IntendedUseandMisusePotential.
| GRAMFC | effectively | | balances | model | capability, | | | | | | |
| ------ | ----------- | --- | -------- | ----- | ----------- | --- | ------- | -------------- | --- | -------------- | ---------- |
| | | | | | | | els can | be of interest | | to the general | public and |
learningefficiency,andinterpretability.
couldalsosavealotoftimetohumanfact-checkers.
Infuturework,wewanttoadaptPROGRAMFC
However,theycouldalsobemisusedbymalicious
tomorereal-worldfact-checkingscenarios,suchas
| | | | | | | | actors. | Weaskresearcherstoexercisecaution. | | | |
| --- | --- | --- | --- | --- | --- | --- | ------- | ---------------------------------- | --- | --- | --- |
fakenewsdetectionandmulti-modalfact-checking,
| | | | | | | | | | | The | use of large lan- |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | ----------------- |
withadvancedreasoningprogramdesignandsub- Environmental Impact.
taskfunctionalities. guage models requires a significant amount of
| | | | | | | | energy | for computation | | for training, | which con- |
| --- | --- | --- | --- | --- | --- | --- | ------ | --------------- | --- | ------------- | ---------- |
Limitations
| | | | | | | | tributestoglobalwarming. | | | Ourworkperformsfew- | |
| --- | --- | --- | --- | --- | --- | --- | ------------------------ | --- | --- | ------------------- | --- |
shotin-contextlearninginsteadoftrainingmodels
| Weidentifytwomainlimitationsof | | | | | PROGRAMFC. | | | | | | |
| ------------------------------ | --- | --- | --- | --- | ---------- | --- | --- | --- | --- | --- | --- |
fromscratch,sotheenergyfootprintofourworkis
First,despitebeingcomplexintheirsurfaceform,
| | | | | | | | less. Thelargelanguagemodel(Codex)whoseAPI | | | | |
| --- | --- | --- | --- | --- | --- | --- | ------------------------------------------ | --- | --- | --- | --- |
theclaimsintheHOVERandFEVEROUSdatasets
weuseforinferenceconsumessignificantenergy.
| mostlyrequireonlyexplicit | | | | multi-stepreasoning, | | | | | | | |
| ------------------------- | --- | --- | --- | -------------------- | --- | --- | --- | --- | --- | --- | --- |
i.e., the decomposition can be derived from the Acknowledgements
| claim’s | syntactic | structure | | or how | the | claim is | | | | | |
| ------- | --------- | --------- | --- | ------ | --- | -------- | --- | --- | --- | --- | --- |
framed. Thislowersthedifficultyofgeneratingrea- This work was supported in part by the National
ScienceFoundationaward#2048122andbySin-
| soning programs. | | However, | | for many | real-world | | | | | | |
| ---------------- | --- | -------- | --- | -------- | ---------- | --- | --- | --- | --- | --- | --- |
gapore’sMinistryofEducationTier3grant“Dig-
| complex | claims, | the | reasoning | is | often | implicit. | | | | | |
| ------- | ------- | --- | --------- | --- | ----- | --------- | --- | --- | --- | --- | --- |
Forexample,fortheclaim“Aristotlecouldn’thave ital Information Resilience: Restoring Trust and
| | | | | | | | NudgingBehavioursinDigitalisation”. | | | | Theviews |
| --- | --- | --- | --- | --- | --- | --- | ----------------------------------- | --- | --- | --- | -------- |
usedalaptop”,thereasoningprogramis:
expressedarethoseoftheauthorsanddonotreflect
answer_1=Question(“WhendidAristotlelive?”);
theofficialpolicyorpositionoftheUSgovernment.
| answer_2 | = Question(“When | | | was | the laptop | in- | | | | | |
| -------- | ---------------- | --- | --- | --- | ---------- | --- | --- | --- | --- | --- | --- |
WethankAlexMei,XinyiWang,DanqingWang,
vented?”);
SharonLevy,GyuwanKim,andothermembersof
fact_1=Verify(“answer_1isbeforeanswer_2.”);
theUCSBNLPgroupfortheirvaluablefeedback.
label=Predict(fact_1)
6989
| References | | | | | | MethodsinNaturalLanguageProcessing(EMNLP), | | | | | | |
| ---------- | --- | --- | --- | --- | --- | ------------------------------------------ | --- | --- | --- | --- | --- | --- |
pages632–642,Lisbon,Portugal.
NaserAhmadi,JoohyungLee,PaoloPapotti,andMo-
hammed Saeed. 2019. Explainable fact checking TomB.Brown,BenjaminMann,NickRyder,Melanie
withprobabilisticanswersetprogramming. InPro- Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind
ceedingsoftheTruthandTrustOnlineConference Neelakantan,PranavShyam,GirishSastry,Amanda
(TTO),London,UK.
| | | | | | | Askell, | Sandhini | Agarwal, | | Ariel | Herbert-Voss, | |
| --- | --- | --- | --- | --- | --- | -------- | -------- | -------- | --------- | ----- | ------------- | ------ |
| | | | | | | Gretchen | Krueger, | Tom | Henighan, | | Rewon | Child, |
Rami Aly, Zhijiang Guo, Michael Sejr Schlichtkrull, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu,
James Thorne, Andreas Vlachos, Christos ClemensWinter,ChristopherHesse,MarkChen,Eric
Christodoulopoulos, Oana Cocarascu, and Arpit Sigler,MateuszLitwin,ScottGray,BenjaminChess,
| Mittal. 2021. | | FEVEROUS: | | Fact Extraction | and | | | | | | | |
| ------------- | --- | --------- | --- | --------------- | --- | ----------- | ----------- | --- | ------- | --- | ----------- | --- |
| | | | | | | Jack Clark, | Christopher | | Berner, | Sam | McCandlish, | |
VERificationOverUnstructuredandStructuredin-
| | | | | | | Alec Radford, | | Ilya Sutskever, | | and Dario | Amodei. | |
| ---------- | ----------------------------------- | --- | --- | --- | --- | -------------- | --- | --------------- | ------------ | --------- | --------- | --- |
| formation. | InProceedingsoftheNeuralInformation | | | | | | | | | | | |
| | | | | | | 2020. Language | | models | are few-shot | | learners. | In |
Processing Systems (NeurIPS) Track on Datasets Proceedings of the Annual Conference on Neural
andBenchmarks,Online. InformationProcessingSystems(NeurIPS),Online.
RamiAlyandAndreasVlachos.2022. Naturallogic- JifanChen,AniruddhSriram,EunsolChoi,andGreg
guidedautoregressivemulti-hopdocumentretrieval
| | | | | | | Durrett.2022a. | | Generatingliteralandimpliedsub- | | | | |
| -------------------- | --- | -------------------------- | --- | --- | --- | ----------------------------------- | --- | ------------------------------- | --- | --- | ---------- | --- |
| forfactverification. | | InProceedingsofthe2022Con- | | | | | | | | | | |
| | | | | | | questionstofact-checkcomplexclaims. | | | | | InProceed- | |
ferenceonEmpiricalMethodsinNaturalLanguage ingsofthe2022ConferenceonEmpiricalMethods
Processing(EMNLP),pages6123–6135,AbuDhabi, in Natural Language Processing (EMNLP), pages
| UnitedArabEmirates. | | | | | | 3495–3516,AbuDhabi,UnitedArabEmirates. | | | | | | |
| ------------------- | --- | --- | --- | --- | --- | -------------------------------------- | --- | --- | --- | --- | --- | --- |
Pepa Atanasova, Jakob Grue Simonsen, Christina Li- MarkChen,JerryTworek,HeewooJun,QimingYuan,
| oma,andIsabelleAugenstein.2020. | | | | Generatingfact | | | | | | | | |
| ------------------------------- | --- | --- | --- | -------------- | --- | -------- | ----- | ----------- | ------ | --- | ------------- | --- |
| | | | | | | Henrique | Ponde | de Oliveira | Pinto, | | Jared Kaplan, | |
checkingexplanations. InProceedingsofthe58th Harrison Edwards, Yuri Burda, Nicholas Joseph,
AnnualMeetingoftheAssociationforComputational Greg Brockman, Alex Ray, Raul Puri, Gretchen
Linguistics(ACL),pages7352–7364,Online. Krueger,MichaelPetrov,HeidyKhlaaf,GirishSas-
| | | | | | | try, Pamela | Mishkin, | | Brooke | Chan, | Scott | Gray, |
| --- | --- | --- | --- | --- | --- | ----------- | -------- | --- | ------ | ----- | ----- | ----- |
Isabelle Augenstein, Christina Lioma, Dongsheng NickRyder,MikhailPavlov,AletheaPower,Lukasz
| Wang, Lucas | Chaves | Lima, | Casper | Hansen, | Chris- | | | | | | | |
| ------------------------------------- | ------ | ----- | ------ | ------- | ------ | -------- | -------- | --------- | -------- | ------- | ------- | ---- |
| | | | | | | Kaiser, | Mohammad | Bavarian, | | Clemens | Winter, | |
| tianHansen,andJakobGrueSimonsen.2019. | | | | | Mul- | | | | | | | |
| | | | | | | Philippe | Tillet, | Felipe | Petroski | Such, | Dave | Cum- |
tiFC:Areal-worldmulti-domaindatasetforevidence- mings, Matthias Plappert, Fotios Chantzis, Eliza-
based fact checking of claims. In Proceedings of beth Barnes, Ariel Herbert-Voss, William Hebgen
the2019ConferenceonEmpiricalMethodsinNatu- Guss,AlexNichol,AlexPaino,NikolasTezak,Jie
ralLanguageProcessingandthe9thInternational
Tang,IgorBabuschkin,SuchirBalaji,ShantanuJain,
JointConferenceonNaturalLanguageProcessing
| | | | | | | William | Saunders, | Christopher | | Hesse, | Andrew | N. |
| --- | --- | --- | --- | --- | --- | ------- | --------- | ----------- | --- | ------ | ------ | --- |
(EMNLP-IJCNLP),pages4685–4697,HongKong, Carr,JanLeike,JoshuaAchiam,VedantMisra,Evan
| China. | | | | | | Morikawa, | Alec | Radford, | Matthew | | Knight, | Miles |
| ------ | --- | --- | --- | --- | --- | --------- | ---- | -------- | ------- | --- | ------- | ----- |
Brundage,MiraMurati,KatieMayer,PeterWelinder,
| Giorgio Barnabò, | | Federico | Siciliano, | Carlos | Castillo, | | | | | | | |
| ---------------- | --- | -------- | ---------- | ------ | --------- | --- | --- | --- | --- | --- | --- | --- |
BobMcGrew,DarioAmodei,SamMcCandlish,Ilya
| Stefano | Leonardi, | Preslav | | Nakov, | Giovanni | | | | | | | |
| ------- | --------- | ------- | -------- | ---------- | -------- | ---------- | -------- | -------- | -------- | ----- | -------- | ----- |
| | | | | | | Sutskever, | and | Wojciech | Zaremba. | 2021. | Evaluat- | |
| Da San | Martino, | and | Fabrizio | Silvestri. | 2022. | | | | | | | |
| | | | | | | ing large | language | models | trained | on | code. | ArXiv |
FbMultiLingMisinfo: Challenginglarge-scalemul- preprint,abs/2107.03374.
| tilingual | benchmark | for | misinformation | | detection. | | | | | | | |
| --------- | --------- | --- | -------------- | --- | ---------- | --- | --- | --- | --- | --- | --- | --- |
In Proceedings of the 2022 International Joint Wenhu Chen, Xueguang Ma, Xinyi Wang, and
ConferenceonNeuralNetworks(IJCNN),pages1–8, William W. Cohen. 2022b. Program of thoughts
Padova,Italy.
| | | | | | | prompting: | Disentangling | | computation | | from | rea- |
| ---------------- | --- | -------- | ---------- | ------ | --------- | --------------- | ------------- | --- | ----------- | ------ | ---- | ----- |
| | | | | | | soning | for numerical | | reasoning | tasks. | | CoRR, |
| Giorgio Barnabò, | | Federico | Siciliano, | Carlos | Castillo, | abs/2211.12588. | | | | | | |
StefanoLeonardi,PreslavNakov,GiovanniDaSan
Martino,andFabrizioSilvestri.2023. Deepactive Zhoujun Cheng, Tianbao Xie, Peng Shi, Chengzu
learningformisinformationdetectionusinggeomet- Li, Rahul Nadkarni, Yushi Hu, Caiming Xiong,
ricdeeplearning. OnlineSocialNetworksandMedia, DragomirRadev,MariOstendorf,LukeZettlemoyer,
| 33:100244. | | | | | | Noah A. | Smith, | and | Tao Yu. | 2022. | Binding | |
| --------------------------------------------- | ---------------------------- | --- | --- | --- | ----- | --------------- | ------ | ----------- | ------- | ---------- | ------- | ----- |
| | | | | | | language | models | in symbolic | | languages. | | CoRR, |
| IzBeltagy,MatthewE.Peters,andArmanCohan.2020. | | | | | | abs/2210.02875. | | | | | | |
| Longformer: | Thelong-documenttransformer. | | | | ArXiv | | | | | | | |
preprint,abs/2004.05150. HyungWonChung,LeHou,ShayneLongpre,Barret
Zoph,YiTay,WilliamFedus,EricLi,XuezhiWang,
SamuelR.Bowman,GaborAngeli,ChristopherPotts, MostafaDehghani,SiddharthaBrahma,AlbertWeb-
and Christopher D. Manning. 2015. A large anno- son, Shixiang Shane Gu, Zhuyun Dai, Mirac Suz-
tatedcorpusforlearningnaturallanguageinference. gun,XinyunChen,AakankshaChowdhery,Sharan
InProceedingsofthe2015ConferenceonEmpirical Narang,GauravMishra,AdamsYu,VincentY.Zhao,
6990
YanpingHuang,AndrewM.Dai,HongkunYu,Slav Yichen Jiang, Shikha Bordia, Zheng Zhong, Charles
Petrov, EdH.Chi, JeffDean, JacobDevlin, Adam Dognin, Maneesh Singh, and Mohit Bansal. 2020.
Roberts, DennyZhou, QuocV.Le, andJasonWei. HoVer: Adatasetformany-hopfactextractionand
| | | | | | | | Findings | of the | Association |
| ----- | ------------------------------------------- | --- | --- | --- | ------------------- | --- | -------- | ------ | ----------- |
| 2022. | Scalinginstruction-finetunedlanguagemodels. | | | | claim verification. | | In | | |
CoRR,abs/2210.11416. forComputationalLinguistics: EMNLP2020,pages
3441–3460,Online.
| Limeng Cui, | Kai Shu, | Suhang Wang, | Dongwon | Lee, | | | | | |
| ----------- | -------- | ------------ | ------- | ---- | --- | --- | --- | --- | --- |
andHuanLiu.2019. dEFEND:Asystemforexplain- Shailza Jolly, Pepa Atanasova, and Isabelle Augen-
ablefakenewsdetection. InProceedingsofthe28th stein.2022. Generatingfluentfactcheckingexpla-
ACMInternationalConferenceonInformationand nationswithunsupervisedpost-editing. Information,
| KnowledgeManagement(CIKM),pages2961–2964, | | | | | 13(10):500. | | | | |
| ----------------------------------------- | --- | --- | --- | --- | ----------- | --- | --- | --- | --- |
Beijing,China.
TakeshiKojima,ShixiangShaneGu,MachelReid,Yu-
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and taka Matsuo, and Yusuke Iwasawa. 2022. Large
Kristina Toutanova. 2019. BERT: Pre-training of language models are zero-shot reasoners. CoRR,
| deepbidirectionaltransformersforlanguageunder- | | | | | abs/2205.11916. | | | | |
| ---------------------------------------------- | -------------- | ------ | ---- | ---------- | --------------- | --- | --- | --- | --- |
| standing. | In Proceedings | of the | 2019 | Conference | | | | | |
of the North American Chapter of the Association NeemaKotonyaandFrancescaToni.2020. Explainable
for Computational Linguistics: Human Language automatedfact-checkingforpublichealthclaims. In
| | | | | | Proceedings | of | the 2020 Conference | | on Empirical |
| --- | --- | --- | --- | --- | ----------- | --- | ------------------- | --- | ------------ |
Technologies(NAACL-HLT),pages4171–4186,Min-
MethodsinNaturalLanguageProcessing(EMNLP),
neapolis,Minnesota,USA.
pages7740–7754,Online.
MohamedH.Gad-Elrab,DariaStepanova,JacopoUr-
AmrithKrishna,SebastianRiedel,andAndreasVlachos.
| bani,andGerhardWeikum.2019. | | | Exfakt: | Aframe- | | | | | |
| --------------------------- | --- | --- | ------- | ------- | --- | --- | --- | --- | --- |
workforexplainingfactsoverknowledgegraphsand 2022. ProoFVer: Naturallogictheoremprovingfor
text. In Proceedings of the Twelfth ACM Interna- factverification. TransactionsoftheAssociationfor
tionalConferenceonWebSearchandDataMining ComputationalLinguistics(TACL),10:1013–1030.
(WSDM),pages87–95,Melbourne,Australia.
NayeonLee,YejinBang,AndreaMadotto,andPascale
Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Fung.2021. Towardsfew-shotfact-checkingviaper-
plexity. InProceedingsofthe2021Conferenceof
| PengfeiLiu, | YimingYang, | JamieCallan, | | andGra- | | | | | |
| ----------- | ----------- | ------------ | --- | ------- | --- | --- | --- | --- | --- |
hamNeubig.2022. PAL:program-aidedlanguage theNorthAmericanChapteroftheAssociationfor
models. CoRR,abs/2211.10435. ComputationalLinguistics: HumanLanguageTech-
nologies(NAACL-HLT),pages1971–1981,Online.
MaxGlockner,YufangHou,andIrynaGurevych.2022.
Missingcounter-evidencerendersNLPfact-checking Nayeon Lee, Belinda Z. Li, Sinong Wang, Wen-tau
unrealisticformisinformation. InProceedingsofthe Yih,HaoMa,andMadianKhabsa.2020. Language
2022ConferenceonEmpiricalMethodsinNatural models as fact checkers? In Proceedings of the
ThirdWorkshoponFactExtractionandVERification
LanguageProcessing(EMNLP),pages5916–5936,
| AbuDhabi,UnitedArabEmirates. | | | | | (FEVER),pages36–41,Online. | | | | |
| ---------------------------- | --- | --- | --- | --- | -------------------------- | --- | --- | --- | --- |
ZhijiangGuo,MichaelSchlichtkrull,andAndreasVla- Jimmy Lin, Xueguang Ma, Sheng-Chieh Lin, Jheng-
chos. 2022. A survey on automated fact-checking. HongYang,RonakPradeep,andRodrigoNogueira.
TransactionsoftheAssociationforComputational 2021. Pyserini: A Python toolkit for reproducible
Linguistics,10:178–206. informationretrievalresearchwithsparseanddense
| | | | | | representations. | | In Proceedings | of | the 44th Inter- |
| --- | --- | --- | --- | --- | ---------------- | --- | -------------- | --- | --------------- |
AshimGuptaandVivekSrikumar.2021. X-Fact:Anew nationalACMSIGIRConferenceonResearchand
| benchmarkdatasetformultilingualfactchecking. | | | | In | | | | | |
| -------------------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
DevelopmentinInformationRetrieval(SIGIR),pages
| Proceedingsofthe59thAnnualMeetingoftheAsso- | | | | | 2356–2362,Online. | | | | |
| ------------------------------------------- | --- | --- | --- | --- | ----------------- | --- | --- | --- | --- |
ciationforComputationalLinguisticsandthe11th
InternationalJointConferenceonNaturalLanguage AlisaLiu,SwabhaSwayamdipta,NoahA.Smith,and
Processing(ACL-IJCNLP),pages675–682,Online. YejinChoi.2022. WANLI:WorkerandAIcollabora-
tionfornaturallanguageinferencedatasetcreation.
Pengcheng He, Jianfeng Gao, and Weizhu Chen. In Findings of the Association for Computational
| 2021. | DeBERTaV3: | Improving | DeBERTa | us- | | | | | |
| ----------------- | ---------- | ------------ | ------- | --------- | ------------ | ---------------------------- | --- | --- | --- |
| | | | | | Linguistics: | EMNLP2022,pages6826–6847,Abu | | | |
| ing ELECTRA-style | | pre-training | with | gradient- | | | | | |
Dhabi,UnitedArabEmirates.
| disentangled | embedding | sharing. | ArXiv | preprint, | | | | | |
| --------------- | --------- | -------- | ----- | --------- | ------------------------------------------- | ----- | ---------- | ----- | ----------- |
| abs/2111.09543. | | | | | YinhanLiu,MyleOtt,NamanGoyal,JingfeiDu,Man- | | | | |
| | | | | | dar Joshi, | Danqi | Chen, Omer | Levy, | Mike Lewis, |
KelvinJiang,RonakPradeep,andJimmyLin.2021. Ex- Luke Zettlemoyer, and Veselin Stoyanov. 2019.
ploringlistwiseevidencereasoningwithT5forfact RoBERTa: ArobustlyoptimizedBERTpretraining
verification. InProceedingsofthe59thAnnualMeet- approach. ArXivpreprint,abs/1907.11692.
| ing of | the Association | for Computational | | Linguis- | | | | | |
| ------ | --------------- | ----------------- | --- | -------- | --- | --- | --- | --- | --- |
ticsandthe11thInternationalJointConferenceon Zhenghao Liu, Chenyan Xiong, Maosong Sun, and
NaturalLanguageProcessing(ACL-IJCNLP),pages Zhiyuan Liu. 2020. Fine-grained fact verification
402–410,Online. withkernelgraphattentionnetwork. InProceedings
6991
of the 58th Annual Meeting of the Association for NLI:Anewbenchmarkfornaturallanguageunder-
ComputationalLinguistics(ACL),pages7342–7351, standing. InProceedingsofthe58thAnnualMeet-
| Online. | | | | | | | ingoftheAssociationforComputationalLinguistics | | | | | |
| ------- | --- | --- | --- | --- | --- | --- | ---------------------------------------------- | --- | --- | --- | --- | --- |
(ACL),pages4885–4901,Online.
| Yi-JuLuandCheng-TeLi.2020. | | | | GCAN:Graph-aware | | | | | | | | |
| -------------------------- | --- | --- | --- | ---------------- | --- | --- | --- | --- | --- | --- | --- | --- |
co-attentionnetworksforexplainablefakenewsde- LongOuyang,JeffWu,XuJiang,DiogoAlmeida,Car-
tectiononsocialmedia. InProceedingsofthe58th rollL.Wainwright,PamelaMishkin,ChongZhang,
AnnualMeetingoftheAssociationforComputational SandhiniAgarwal,KatarinaSlama,AlexRay,John
Linguistics(ACL),pages505–514,Online. Schulman,JacobHilton,FraserKelton,LukeMiller,
| | | | | | | | Maddie | Simens, | Amanda | Askell, | Peter | Welinder, |
| --- | --- | --- | --- | --- | --- | --- | ------ | ------- | ------ | ------- | ----- | --------- |
GrégoireMialon,RobertoDessì,MariaLomeli,Christo- PaulF.Christiano,JanLeike,andRyanLowe.2022.
foros Nalmpantis, Ramakanth Pasunuru, Roberta Traininglanguagemodelstofollowinstructionswith
Raileanu, Baptiste Rozière, Timo Schick, Jane humanfeedback. CoRR,abs/2203.02155.
Dwivedi-Yu,AsliCelikyilmaz,EdouardGrave,Yann
| | | | | | | | Liangming | Pan, Wenhu | Chen, | Wenhan | Xiong, | Min- |
| --------------- | --- | -------- | -------- | -------------------- | --- | --------- | ---------------------------------- | ---------- | ----- | ------ | ------------- | --------- |
| LeCun, | and | Thomas | Scialom. | 2023. | | Augmented | | | | | | |
| | | | | | | | YenKan,andWilliamYangWang.2021. | | | | | Zero-shot |
| languagemodels: | | asurvey. | | CoRR,abs/2302.07842. | | | | | | | | |
| | | | | | | | factverificationbyclaimgeneration. | | | | InProceedings | |
Preslav Nakov, Alberto Barrón-Cedeño, Giovanni of the 59th Annual Meeting of the Association for
Da San Martino, Firoj Alam, Julia Maria ComputationalLinguisticsandthe11thInternational
Struß, Thomas Mandl, Rubén Míguez, Tom- JointConferenceonNaturalLanguageProcessing
(ACL-IJCNLP),pages476–483,Online.
| maso | Caselli, | Mucahid | | Kutlu, | Wajdi | Zaghouani, | | | | | | |
| ---- | -------- | ------- | --- | ------ | ----- | ---------- | --- | --- | --- | --- | --- | --- |
ChengkaiLi,ShadenShaar,GautamKishoreShahi,
AliciaParrish,WilliamHuang,OmarAgha,Soo-Hwan
HamdyMubarak,AlexNikolov,NikolayBabulkov,
YavuzSelimKartal,andJavierBeltrán.2022. The Lee,NikitaNangia,AlexiaWarstadt,KarmanyaAg-
CLEF-2022CheckThat! labonfightingtheCOVID- garwal,EmilyAllaway,TalLinzen,andSamuelR.
| | | | | | | | Bowman.2021. | | Doesputtingalinguistintheloop | | | |
| -------------------------------- | ------------------------------------- | ------------- | --- | ---------- | --- | ---------- | --------------------------------------- | -------- | ----------------------------- | --- | ----------- | ------ |
| 19infodemicandfakenewsdetection. | | | | | | InProceed- | | | | | | |
| | | | | | | | improve | NLU data | collection? | | In Findings | of the |
| ings of | the | 44th European | | Conference | | on IR Re- | | | | | | |
| | | | | | | | AssociationforComputationalLinguistics: | | | | | EMNLP |
| search: | AdvancesinInformationRetrieval(ECIR), | | | | | | | | | | | |
2021,pages4886–4901,PuntaCana,DominicanRe-
pages416–428,Berlin,Heidelberg.
public.
PreslavNakov,DavidCorney,MaramHasanain,Firoj
| | | | | | | | Kashyap Popat, | Subhabrata | | Mukherjee, | Jannik | Ströt- |
| --- | --- | --- | --- | --- | --- | --- | -------------- | ---------- | --- | ---------- | ------ | ------ |
Alam,TamerElsayed,AlbertoBarrón-Cedeño,Paolo
| | | | | | | | gen, and | Gerhard | Weikum. | 2017. | Where | the truth |
| -------- | ------ | ------ | --- | -------- | --- | -------- | -------- | ------- | ------- | ----- | ----- | --------- |
| Papotti, | Shaden | Shaar, | and | Giovanni | Da | San Mar- | | | | | | |
lies: Explainingthecredibilityofemergingclaims
| tino.2021a. | | Automatedfact-checkingforassisting | | | | | | | | | | |
| ----------- | -------------- | ---------------------------------- | --- | ----------- | --- | ------------ | ----------------------- | --- | --- | ----------------- | --- | --- |
| | | | | | | | onthewebandsocialmedia. | | | InProceedngsofthe | | |
| human | fact-checkers. | | In | Proceedings | | of the Joint | | | | | | |
InternationalWorldWideWebConference(WWW),
ConferenceonArtificialIntelligence(IJCAI),pages
pages1003–1012.
4551–4558,Online.
OfirPress,MuruZhang,SewonMin,LudwigSchmidt,
| Preslav | Nakov, | Giovanni | | Da San | Martino, | Tamer | | | | | | |
| -------- | ------- | -------- | -------------- | ------ | -------- | ------- | ------------------------------ | --- | --- | --- | --- | --------- |
| | | | | | | | NoahA.Smith,andMikeLewis.2022. | | | | | Measuring |
| Elsayed, | Alberto | | Barrón-Cedeño, | | Rubén | Míguez, | | | | | | |
andnarrowingthecompositionalitygapinlanguage
ShadenShaar,FirojAlam,FatimaHaouari,Maram
| | | | | | | | models. | CoRR,abs/2210.03350. | | | | |
| --------- | ------- | --- | --------- | ---- | -------- | ---- | ------- | -------------------- | --- | --- | --- | --- |
| Hasanain, | Nikolay | | Babulkov, | Alex | Nikolov, | Gau- | | | | | | |
tamKishoreShahi,JuliaMariaStruß,andThomas
ColinRaffel,NoamShazeer,AdamRoberts,Katherine
| Mandl. | 2021b. | The | CLEF-2021 | | CheckThat! | lab | | | | | | |
| ------ | ------ | --- | --------- | --- | ---------- | --- | --- | --- | --- | --- | --- | --- |
Lee,SharanNarang,MichaelMatena,YanqiZhou,
| on detecting | | check-worthy | | claims, | previously | fact- | | | | | | |
| -------------------------- | --- | ------------ | --- | ------------------ | ---------- | ----- | ------------------------- | --- | --- | ------------------ | --- | --- |
| | | | | | | | WeiLi,andPeterJ.Liu.2020. | | | Exploringthelimits | | |
| checkedclaims,andfakenews. | | | | InProceedingsofthe | | | | | | | | |
oftransferlearningwithaunifiedtext-to-texttrans-
43rdEuropeanConferenceonInformationRetrieval former. J.Mach.Learn.Res.,21:140:1–140:67.
(ECIR),pages639–649,Lucca,Italy.
| | | | | | | | StephenE.RobertsonandHugoZaragoza.2009. | | | | | The |
| ------------------------- | ------- | ------------------------------- | --------- | --------- | --------------- | ----------- | --------------------------------------- | --------- | ---------- | ------ | -------------- | ------- |
| Van-Hoang | Nguyen, | | Kazunari | Sugiyama, | | Preslav | | | | | | |
| | | | | | | | probabilistic | relevance | framework: | | BM25 | and be- |
| Nakov,andMin-YenKan.2020. | | | | | FANG:leveraging | | | | | | | |
| | | | | | | | yond. Foundations | | and | Trends | in Information | Re- |
| social | context | for | fake news | detection | | using graph | trieval,3(4):333–389. | | | | | |
| representation. | | InProceedingsofthe29thACMInter- | | | | | | | | | | |
nationalConferenceonInformationandKnowledge Arkadiy Saakyan, Tuhin Chakrabarty, and Smaranda
Management(CIKM),pages1165–1174. Muresan. 2021. COVID-fact: Fact extraction and
verificationofreal-worldclaimsonCOVID-19pan-
| Yixin Nie, | Haonan | Chen, | | and Mohit | Bansal. | 2019. | | | | | | |
| ---------- | ------ | ----- | --- | --------- | ------- | ----- | ------------------------------------------ | --- | --- | --- | --- | --- |
| | | | | | | | demic. InProceedingsofthe59thAnnualMeeting | | | | | |
Combiningfactextractionandverificationwithneu- oftheAssociationforComputationalLinguisticsand
ralsemanticmatchingnetworks. InProceedingsof the11thInternationalJointConferenceonNatural
the33rdAAAIConferenceonArtificialIntelligence LanguageProcessing(ACL-IJCNLP),pages2116–
| (AAAI),pages6859–6866,Honolulu,Hawaii,USA. | | | | | | | 2129,Online. | | | | | |
| ------------------------------------------ | --- | --- | --- | --- | --- | --- | ------------ | --- | --- | --- | --- | --- |
YixinNie,AdinaWilliams,EmilyDinan,MohitBansal, AalokSathe,SalarAther,TuanManhLe,NathanPerry,
JasonWeston,andDouweKiela.2020. Adversarial andJoonsukPark.2020. Automatedfact-checking
6992
of claims from Wikipedia. In Proceedings of the WilliamYangWang.2017. “Liar,liarpantsonfire”: A
TwelfthLanguageResourcesandEvaluationConfer- newbenchmarkdatasetforfakenewsdetection. In
ence(LREC),pages6874–6882,Marseille,France. Proceedingsofthe55thAnnualMeetingoftheAsso-
ciationforComputationalLinguistics(ACL),pages
TimoSchick,JaneDwivedi-Yu,RobertoDessì,Roberta
422–426,Vancouver,Canada.
Raileanu,MariaLomeli,LukeZettlemoyer,Nicola
| Cancedda,andThomasScialom.2023. | | | | | Toolformer: | | | | | |
| ------------------------------- | --- | --- | --- | --- | ----------- | ----------- | --------- | --------------- | --- | ------ |
| | | | | | | XuezhiWang, | JasonWei, | DaleSchuurmans, | | QuocV. |
Languagemodelscanteachthemselvestousetools. Le, Ed H. Chi, and Denny Zhou. 2022. Self-
CoRR,abs/2302.04761. consistencyimproveschainofthoughtreasoningin
| | | | | | | languagemodels. | CoRR,abs/2203.11171. | | | |
| --- | --- | --- | --- | --- | --- | --------------- | -------------------- | --- | --- | --- |
TalSchuster,AdamFisch,andReginaBarzilay.2021.
Get your vitamin C! robust fact verification with JasonWei,XuezhiWang,DaleSchuurmans,Maarten
contrastive evidence. In Proceedings of the 2021 Bosma,EdH.Chi,QuocLe,andDennyZhou.2022.
| Conference | of the | North | American | Chapter | of the | | | | | |
| ---------- | ------ | ----- | -------- | ------- | ------ | --- | --- | --- | --- | --- |
Chainofthoughtpromptingelicitsreasoninginlarge
| AssociationforComputationalLinguistics: | | | | | Human | | | | | |
| --------------------------------------- | --- | --- | --- | --- | ----- | --------------- | ----------------------------- | --- | --- | --- |
| | | | | | | languagemodels. | ArXivpreprint,abs/2201.11903. | | | |
LanguageTechnologies(NAACL-HLT),pages624–
643,Online.
AdinaWilliams,NikitaNangia,andSamuelBowman.
| | | | | | | 2018. A | broad-coverage | challenge | corpus | for sen- |
| --- | --- | --- | --- | --- | --- | ------- | -------------- | --------- | ------ | -------- |
AmirSoleimani,ChristofMonz,andMarcelWorring.
| | | | | | | tenceunderstandingthroughinference. | | | InProceed- | |
| --- | --- | --- | --- | --- | --- | ----------------------------------- | --- | --- | ---------- | --- |
2020. BERTforevidenceretrievalandclaimverifi-
ingsofthe2018ConferenceoftheNorthAmerican
| cation. | InAdvancesinInformationRetrieval(ECIR), | | | | | | | | | |
| ------- | --------------------------------------- | --- | --- | --- | --- | ------- | ------------------ | ----------------- | --- | ---- |
| | | | | | | Chapter | of the Association | for Computational | | Lin- |
volume12036,pages359–366.
| | | | | | | guistics: | Human Language | Technologies | | (NAACL- |
| --- | --- | --- | --- | --- | --- | --------- | -------------- | ------------ | --- | ------- |
JamesThorneandAndreasVlachos.2018. Automated HLT), pages 1112–1122, New Orleans, Louisiana,
USA.
factchecking:Taskformulations,methodsandfuture
| directions. | InProceedingsofthe27thInternational | | | | | | | | | |
| ----------- | ----------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
DustinWright,DavidWadden,KyleLo,BaileyKuehl,
ConferenceonComputationalLinguistics(COLING),
ArmanCohan,IsabelleAugenstein,andLucyWang.
pages3346–3359,SantaFe,NewMexico,USA.
2022. Generatingscientificclaimsforzero-shotsci-
James Thorne, Andreas Vlachos, Christos entificfactchecking. InProceedingsofthe60thAn-
Christodoulopoulos, and Arpit Mittal. 2018. nualMeetingoftheAssociationforComputational
FEVER: a large-scale dataset for fact extraction Linguistics(ACL),pages2448–2460,Dublin,Ireland.
| and VERification. | | In | Proceedings | of | the 2018 | | | | | |
| ----------------- | --- | --- | ----------- | --- | -------- | --- | --- | --- | --- | --- |
Conference of the North American Chapter of the Fan Yang, Shiva K. Pentyala, Sina Mohseni, Meng-
| | | | | | | nan Du, | Hao Yuan, | Rhema Linder, | Eric | D. Ragan, |
| --------------------------------------- | ------------ | --- | ------------ | --- | ----- | ------------------------------ | --------- | ------------- | ------ | --------- |
| AssociationforComputationalLinguistics: | | | | | Human | | | | | |
| | | | | | | ShuiwangJi,andXia(Ben)Hu.2019. | | | XFake: | Ex- |
| Language | Technologies | | (NAACL-HLT), | | pages | | | | | |
809–819,NewOrleans,Louisiana. plainablefakenewsdetectorwithvisualizations. In
ProceedingsoftheTheWorldWideWebConference
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob (WWW),pages3600–3604,SanFrancisco,California,
| Uszkoreit, | Llion | Jones, | Aidan | N. Gomez, | Lukasz | USA. | | | | |
| ------------------------------- | ----- | ------ | ----- | -------------- | ------ | ---- | --- | --- | --- | --- |
| Kaiser,andIlliaPolosukhin.2017. | | | | Attentionisall | | | | | | |
youneed. InAdvancesinNeuralInformationPro- ZhilinYang,PengQi,SaizhengZhang,YoshuaBengio,
cessingSystems30: AnnualConferenceonNeural WilliamCohen,RuslanSalakhutdinov,andChristo-
Information Processing Systems (NeurIPS), pages pher D. Manning. 2018. HotpotQA: A dataset for
5998–6008,LongBeach,California,USA. diverse, explainablemulti-hopquestionanswering.
InProceedingsofthe2018ConferenceonEmpirical
| David Wadden, | Shanchuan | | Lin, | Kyle Lo, | Lucy Lu | | | | | |
| ------------- | --------- | --- | ---- | -------- | ------- | --- | --- | --- | --- | --- |
MethodsinNaturalLanguageProcessing(EMNLP),
Wang, Madeleine van Zuylen, Arman Cohan, and pages2369–2380,Brussels,Belgium.
| HannanehHajishirzi.2020. | | | Factorfiction:Verifying | | | | | | | |
| ------------------------ | --- | --- | ----------------------- | --- | --- | --- | --- | --- | --- | --- |
scientific claims. In Proceedings of the 2020 Con- Wanjun Zhong, Jingjing Xu, Duyu Tang, Zenan Xu,
ferenceonEmpiricalMethodsinNaturalLanguage
| | | | | | | NanDuan, | MingZhou, | JiahaiWang, | andJianYin. | |
| --- | --- | --- | --- | --- | --- | -------- | --------- | ----------- | ----------- | --- |
Processing(EMNLP),pages7534–7550,Online.
2020. Reasoningoversemantic-levelgraphforfact
| | | | | | | checking. | InProceedingsofthe58thAnnualMeet- | | | |
| --- | --- | --- | --- | --- | --- | --------- | --------------------------------- | --- | --- | --- |
DavidWadden,KyleLo,BaileyKuehl,ArmanCohan,
ingoftheAssociationforComputationalLinguistics
IzBeltagy,LucyLuWang,andHannanehHajishirzi.
(ACL),pages6170–6180,Online.
| 2022a. | SciFact-open: | Towardsopen-domainscien- | | | | | | | | |
| ------ | ------------- | ------------------------ | --- | --- | --- | --- | --- | --- | --- | --- |
tificclaimverification. InFindingsoftheAssociation JieZhou,XuHan,ChengYang,ZhiyuanLiu,Lifeng
forComputationalLinguistics: EMNLP2022,pages Wang, Changcheng Li, and Maosong Sun. 2019.
4719–4734,AbuDhabi,UnitedArabEmirates.
GEAR:Graph-basedevidenceaggregatingandrea-
| | | | | | | soning for | fact verification. | In Proceedings | | of the |
| ------------- | ---- | -------- | ----- | ----- | ------ | ---------- | ------------------ | -------------- | --- | ------ |
| David Wadden, | Kyle | Lo, Lucy | Wang, | Arman | Cohan, | | | | | |
57thAnnualMeetingoftheAssociationforCompu-
| Iz Beltagy, | and Hannaneh | | Hajishirzi. | 2022b. | Mul- | | | | | |
| ----------- | ------------ | --- | ----------- | ------ | ---- | --- | --- | --- | --- | --- |
tationalLinguistics(ACL),pages892–901,Florence,
tiVerS:Improvingscientificclaimverificationwith
Italy.
| weaksupervisionandfull-documentcontext. | | | | | InFind- | | | | | |
| --------------------------------------- | --- | --- | --- | --- | ------- | --- | --- | --- | --- | --- |
ingsoftheAssociationforComputationalLinguis-
tics:NAACL2022,pages61–76,Seattle,Washington,
USA.
6993
A ImplementationDetailsaboutthe In this category, we selected the following three
| Baselines | | | | | | fine-tunedmodels: | | | | | | | |
| --------- | --- | --- | --- | --- | --- | ----------------- | --- | --- | --- | --- | --- | --- | --- |
Inthissection,wegivetheimplementationdetails • RoBERTa-NLI (Nie et al., 2020) fine-tunes
| forthesevenbaselinesweusedinourwork. | | | | | | Typ- | | | | | | | |
| ------------------------------------ | --- | --- | --- | --- | --- | ---- | ------------- | --- | ---- | --- | ---------- | --- | ------ |
| | | | | | | | RoBERTa-large | | (Liu | et | al., 2019) | on | a com- |
ical ways to perform few-shot fact-checking us- bination of four well-known NLI datasets:
inglargelanguagemodelsarefine-tuningandin- SNLI(Bowmanetal.,2015),MNLI(Williams
| contextlearning. | | Thus,wecategorizethebaselines | | | | | | | | | | | |
| ---------------- | --- | ----------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
etal.,2018),FEVER-NLI(Nieetal.,2019),
| intothreecategories. | | | | | | | ANLI(R1,R2,R3)(Nieetal.,2020). | | | | | | Weused |
| -------------------- | --- | --- | --- | --- | --- | --- | ------------------------------ | --- | --- | --- | --- | --- | ------ |
thepublicmodelcheckpointavailableatHug-
A.1 Pre-trainedModels
| | | | | | | | gingFace4 | | and we | further | fine-tuned | | it with |
| ----------- | ------ | --- | ---------- | --- | ---------- | --- | --------- | --- | ------ | ------- | ---------- | --- | ------- |
| Pre-trained | models | use | pretrained | | Transform- | | | | | | | | |
20randomexamplesfromHOVER/FEVER-
| ers (Vaswani | et | al., 2017) | such | as | BERT | (Devlin | OUS. | | | | | | |
| ------------- | --- | ---------- | ---- | ---- | ----- | --------- | ---- | --- | --- | --- | --- | --- | --- |
| et al., 2019) | and | T5 (Raffel | et | al., | 2020) | for fact- | | | | | | | |
DeBERTaV3-NLI
checking. Forfew-shotlearning,wefine-tunethem • (He et al., 2021) fine-
| | | | | | | | tunes | the | DeBERTaV3-large | | | model | on |
| --- | --- | --- | --- | --- | --- | --- | ----- | --- | --------------- | --- | --- | ----- | --- |
using20randomlysampledtrainingexamplesfrom
885,242NLIhypothesis–premisepairsfrom
| HOVER | or FEVEROUS. | | We | ran the | training | 10 | | | | | | | |
| ----- | ------------ | --- | --- | ------- | -------- | --- | --- | --- | --- | --- | --- | --- | --- |
timeswithdifferentrandomseedsandreporttheav- FEVER and on four NLI datasets: MNLI,
erageperformanceonthevalidationset. Wechose ANLI, LingNLI (Parrish et al., 2021), and
| | | | | | | | WANLI | (Liu | et al., | 2022). | This | is | the best- |
| --- | --- | --- | --- | --- | --- | --- | ----- | ---- | ------- | ------ | ---- | --- | --------- |
twomodels:
performingNLImodelonHuggingFaceasof
| • BERT-FC | | (Soleimani | et | al., 2020): | | It uses | | | | | | | |
| --------- | --- | ---------- | --- | ----------- | --- | ------- | --- | --- | --- | --- | --- | --- | --- |
06/06/2022.5
| BERT | for | claim | verification. | | The | claim | | | | | | | |
| ---- | --- | ----- | ------------- | --- | --- | ----- | --- | --- | --- | --- | --- | --- | --- |
and the evidence are concatenated ([CLS] • MULTIVERS(Waddenetal.,2022b),formerly
| claim | [SEP] | evidence) | | | | | | | | | | | |
| ----- | ----- | --------- | --- | --- | ---- | ------ | ----- | --- | ------------ | --- | ---- | --- | ----- |
| | | | | and | used | as in- | known | as | LongChecker, | | uses | the | Long- |
put for a binary classification task to pre- Former (Beltagy et al., 2020) for claim ver-
dict the veracity label of the claim. We use ification to address the long input evidence
thebert-large-uncased(345Mparameters)
| | | | | | | | problem. | | We use | a model | checkpoint | | fine- |
| ---------------------------- | --- | --- | --- | --- | --- | --- | -------------- | --- | ------ | ------- | ---------- | --- | ----- |
| modelprovidedinHuggingFace.2 | | | | | | | tunedonFEVER.6 | | | | | | |
LisT5
• (Jiang et al., 2021): This is a fact- A.3 In-ContextLearningModels
| checking | | framework | built | with | a pretrained | | | | | | | | |
| -------- | --- | --------- | ----- | ---- | ------------ | --- | --- | --- | --- | --- | --- | --- | --- |
Thesemodelshaverecentlyshownstrongfew-shot
| sequence-to-sequence | | | transformer, | | | namely | | | | | | | |
| ----------------------------------- | --- | --- | ------------ | --- | --- | --------------------------------- | --- | --- | --- | --- | --- | --------- | --- |
| | | | | | | learningabilityinvariousNLPtasks. | | | | | | Byprompt- | |
| T5(Raffeletal.,2020),asitsbackbone. | | | | | | We | | | | | | | |
ingalargelanguagemodelwithafewin-context
adoptthe“listwiseconcatenation”proposedin
examples,themodelcanquicklylearnataskfrom
thepaperforlabelprediction,whichconcate-
| | | | | | | demonstrations. | | | Tomakeafaircomparisontoour | | | | |
| --- | --- | --- | --- | --- | --- | --------------- | --- | --- | -------------------------- | --- | --- | --- | --- |
natesallcandidateevidencesentencesintoa
model,wechoosetwoin-contextlearningbaselines
singleinputandwetrainthet5-largemodel
asfollows.
todirectlyclassifytheclaimasSupportedor
Refuted. Weusetheoriginalimplementation • Codex (Chen et al., 2021) is used in
ofthismodel.3
| | | | | | | | our | model | to generate | | reasoning | programs. | |
| --- | --- | --- | --- | --- | --- | --- | --- | --------------- | ----------- | -------- | --------- | --------- | ---- |
| | | | | | | | One | straightforward | | baseline | | directly | uses |
A.2 FC/NLIFine-TunedModels
| | | | | | | | it for | fact-checking. | | | To | this | end, we |
| --- | --- | --- | --- | --- | --- | --- | ------ | -------------- | --- | --- | --- | ---- | ------- |
ThesemodelsarepretrainedTransformermodels
| | | | | | | | prompt | Codex | (code-davinci-002) | | | | as fol- |
| ----------------- | -------- | ------------ | ---------- | --------- | ---- | -------- | ------------ | ----------- | ------------------ | ---------- | -------- | ------------- | ------- |
| that have | been | specifically | fine-tuned | | on | single- | | | | | | | |
| | | | | | | | lows: | “<Evidence> | | Based | | on the | above |
| hop fact-checking | | datasets | (e.g., | FEVER) | | or nat- | | | | | | | |
| | | | | | | | information, | | is | it | true | that <Claim>? | |
| ural language | | inference | (NLI) | datasets. | This | ad- | | | | | | | |
| | | | | | | | True | or False? | | The answer | | is:”. | Wepre- |
| ditional | training | allows | these | models | to | excel at | | | | | | | |
| | | | | | | | fix the | same | 20 in-context | | examples | | for our |
fact-checkingsimpleclaims,andthustheycangen-
modelbeforethepromptasdemonstrations.
eralizebettertocomplexclaimsthatrequiremulti-
4https://huggingface.co/ynie/
hopreasoningduringfurtherfew-shotfine-tuning.
roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli
2https://huggingface.co/ 5https://huggingface.co/MoritzLaurer/
3https://github.com/castorini/pygaggle/tree/ DeBERTa-v3-large-mnli-fever-anli-ling-wanli
master/experiments/list5 6https://github.com/dwadden/multivers
6994
• FLAN-T5(Chungetal.,2022)isanimproved D ProgramGenerationPrompts
| version | of | T5, which | is fine-tuned | | on 1.8K | | | | | |
| ------- | --- | --------- | ------------- | --- | ------- | --- | --- | --- | --- | --- |
OurmanuallywrittenpromptsfortheHOVERand
tasksphrasedasinstructions,withandwithout
theFEVEROUS-SdatasetsaregiveninListings1
| exemplars,i.e.,zero-shotandfew-shot. | | | | | The | | | | | |
| ------------------------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
and2,respectively.
modelhasshownstrongperformanceinvar-
iousin-contextfew-shotlearningNLPtasks,
E PromptsforClosed-Book
| such | as reasoning, | | and question-answering. | | | | | | | |
| ---- | ------------- | --- | ----------------------- | --- | --- | --- | --- | --- | --- | --- |
Fact-Checking
| We | prompt | the model | with the | same | format | | | | | |
| --- | ------ | --------- | -------- | ---- | ------ | --- | --- | --- | --- | --- |
as we used in Section 3.4: “<Evidence> Q: Belowweshowthetemplatesforthefourprompt-
<Claim> Is it true that <Claim>? True ing methods used for InstructGPT for the closed-
or False? The answer is:”,prefixingwith bookfact-checkingsettinginSection4.4.
| thesame20in-contextexamples. | | | | Wealsouse | | DirectPrompting | | | | |
| ---------------------------- | --- | --- | --- | --------- | --- | --------------- | --- | --- | --- | --- |
thesamemodelsize(FLAN-T5-XXL3B)with
| | | | | | | # Answer | the following | true/false | questions: | |
| --- | --- | --- | --- | --- | --- | -------- | ------------- | ---------- | ---------- | --- |
ourmodelforfaircomparison.
| | | | | | | Is it true | that The | woman | the story behind | Girl Crazy |
| ------------------------------ | --- | --- | --- | --- | --- | ----------- | ------------------- | ----- | ---------------- | ---------- |
| | | | | | | is credited | to is older | than | Ted Kotcheff? | |
| B ExamplesofGeneratedReasoning | | | | | | The answer | is: False | | | |
| Programs | | | | | | ( more | in-context examples | here | ) | |
| | | | | | | ··· | | | ··· | |
| | | | | | | Is it true | that <input_claim>? | | | |
Figure7showssixexamplesofgeneratedreason-
| | | | | | | The answer | is: | | | |
| ---------------- | --- | --------- | ---------------- | --- | --- | --------------- | ------------- | ---------- | --------- | --- |
| ingprogramsby | | PROGRAMFC | thatcoverdiverse | | | | | | | |
| reasoningchains. | | | | | | ZS-CoTPrompting | | | | |
| | | | | | | # Answer | the following | true/false | question: | |
C ErrorAnalysisforReasoning
| | | | | | | Is it true | that <input_claim>? | | True or False? | |
| -------- | ----- | ------------- | --- | --------- | ----- | ------------ | ------------------- | --- | -------------- | --- |
| Programs | | | | | | Let us think | step-by-step. | | The answer is: | |
| Figure 8 | shows | five examples | of | erroneous | cases | CoTPrompting | | | | |
wherethegeneratedreasoningprogramsareincor-
| | | | | | | # Answer | the following | true/false | questions: | |
| --- | --- | --- | --- | --- | --- | -------- | ------------- | ---------- | ---------- | --- |
rect. Weprovideexplanationsforeachoftheerror
| | | | | | | Is it true | that The | woman | the story behind | Girl Crazy |
| --- | --- | --- | --- | --- | --- | ----------- | ----------- | ----- | ---------------- | ---------- |
| | | | | | | is credited | to is older | than | Ted Kotcheff? | |
casesbelow:
| | | | | | | Let's think | step by | step. | | |
| --- | --- | --- | --- | --- | --- | ------------ | -------- | -------- | ---------- | --------- |
| | | | | | | Girl Crazy's | story is | credited | to Hampton | Del Ruth. |
Example1 Itgeneratesawronglogicalreason- Hampton Del Ruth was born on September 7, 1879.
| | | | | | | Ted Kotcheff | was born | on April | 7, 1931. | |
| --- | --- | --- | --- | --- | --- | ------------ | -------- | -------- | -------- | --- |
ing operator for the final step. The correct logic Therefore, the answer is: False.
| should be | “not | (fact_1 | and fact_2)” | | instead | | | | | |
| --------- | ---- | -------- | ------------ | --- | ------- | ----------- | ------------------- | ----- | --- | --- |
| | | | | | | ( more | in-context examples | here | ) | |
| of“fact_1 | and | fact_2”. | | | | ··· | | | ··· | |
| | | | | | | Is it true | that <input_claim>? | | | |
| | | | | | | Let's think | step by | step. | | |
Itfailstoperformco-referencereso-
Example2
lutionfortheargumentsinthethirdandthefourth
Self-AskPrompting
| reasoningsteps. | | “Thisalbum”shouldbereplaced | | | | | | | | |
| --------------- | --- | --------------------------- | --- | --- | --- | -------- | ------------- | ---------- | ---------- | --- |
| | | | | | | # Answer | the following | true/false | questions: | |
with“Thebluegrass”tomakethesub-taskcontext-
| | | | | | | Is it true | that The | woman | the story behind | Girl Crazy |
| ------------ | ----- | -------- | ------ | --- | -------- | ------------ | ----------- | ---------- | ---------------- | ---------- |
| independent. | “This | musical” | should | be | replaced | | | | | |
| | | | | | | is credited | to is older | than | Ted Kotcheff? | |
| | | | | | | Q: The story | behind | Girl Crazy | is credited | to whom? |
withthevariable“answer_1”fromthefirststep.
| | | | | | | A: Hampton | Del Ruth | | | |
| -------------------------------------------- | ---------------------------------- | ------- | ----------------- | --------- | --------- | ------------- | ------------------- | ---------- | ------------------ | --- |
| | | | | | | Q: Is Hampton | Del Ruth | older | than Ted Kotcheff? | |
| Example3 | Itfailstocreateameaningfulprob- | | | | | A: No | | | | |
| | | | | | | So the final | answer | is: False. | | |
| lemdecompositionfortheclaim. | | | Itgeneratesatriv- | | | | | | | |
| | | | | | | ( more | in-context examples | here | ) | |
| ialprogramthatsimplyrepeatstheoriginalclaim. | | | | | | ··· | | | ··· | |
| | | | | | | Is it true | that <input_claim>? | | | |
| Example4 | Itfailstogenerateafine-grainedrea- | | | | | | | | | |
| soning structure | | for the | input claim. | It | also gen- | | | | | |
| erates a | trivial program | | that simply | separates | the | | | | | |
claimintosentences.
| Example | 5 It | generates | a redundant | | reason- | | | | | |
| --------- | --------------- | --------- | ----------- | ------------ | -------- | --- | --- | --- | --- | --- |
| ing step | “Question("When | | was | the musician | | | | | | |
| born?")”, | which | does not | add any | new | informa- | | | | | |
tiontothereasoningchain.
6995
Claim:
Tritonia and Phyteuma are both names for a plant genus.
Predicted Program:
fact_1 = Verify("Tritonia is a name for a plant genus.")
fact_2 = Verify("Phyteuma is a name for a plant genus.")
label = Predict(fact_1 and fact_2)
Claim:
The country that Fujairah College is located in had a 2013 population of 9.2 million until it was
hit by the plague in 1483 when the population was halved.
Predicted Program:
answer_1 = Question("Which country is Fujairah College located in?")
fact_1 = Verify("{answer_1} had a 2013 population of 9.2 million.")
fact_2 = Verify("{answer_1} was hit by the plague in 1483.")
fact_3 = Verify("The population of {answer_1} was halved in 1483.")
label = Predict(fact_1 and fact_2 and fact_3)
Claim:
The first female board member for the Liberal Party, she was born in Vestfold county in Norway.
Predicted Program:
answer_1 = Question("Who was the first female board member for the Liberal Party?")
fact_1 = Verify("{answer_1} was born in Vestfold county in Norway.")
label = Predict(fact_1)
Claim:
The solicitor who won the show Back to Reality ahead of Maureen Rees and Craig Phillips is
English. The solicitor that was a chair of Global Witness is also English.
Predicted Program:
answer_1 = Question("Which solicitor won the show Back to Reality ahead of Maureen Rees and Craig
Phillips?")
answer_2 = Question("Which solicitor was a chair of Global Witness?")
fact_1 = Verify("{answer_1} is English.")
fact_2 = Verify("{answer_2} is English.")
label = Predict(fact_1 and fact_2)
Claim:
Anthony Burgess addressed the novelist and essayist, the author of Grimus, in a lengthy love
letter. The author is of the same nationality as Raj Koothrappali.
Predicted Program:
answer_1 = Question("Which novelist and essayist did Anthony Burgess address in a lengthy love
letter?")
fact_1 = Verify("{answer_1} is the author of Grimus.")
answer_2 = Question("What is the nationality of Raj Koothrappali?")
fact_2 = Verify("{answer_1} is of the same nationality as {answer_2}.")
label = Predict(fact_1 and fact_2)
Claim:
The critically acclaimed film, that Buddy Baker scored in 1975, is a Walt Disney film. It was
produced first before the film that featured Bruce M. Fischer as Mr. Coogar.
Predicted Program:
answer_1 = Question("Which critically acclaimed film did Buddy Baker score in 1975?")
fact_1 = Verify("{answer_1} is a Walt Disney film.")
answer_2 = Question("Which film featured Bruce M. Fischer as Mr. Coogar?")
fact_2 = Verify("{answer_1} was produced first before {answer_2}.")
label = Predict(fact_1 and fact_2)
Figure7: ExamplesofgeneratedreasoningprogramsbyPROGRAMFC.
6996
1
Semantic Error —Token: incorrect or missing arguments/variables
Example 1:
Bitter Jester and The Future of Food are not both documentaries about food.
Predicted Program:
fact_1 = Verify("Bitter Jester is a documentary about food.")
fact_2 = Verify("The Future of Food is a documentary about food.")
label = Predict(fact_1 and fact_2) label = Predict(not (fact_1 and fact_2))
Example 2:
The record producer that produced the bluegrass album was born on 22 June, 1944. This album
inspired a Tony award winning musical. This musical had a character that was originated by Carmen
Cusack.
Predicted Program:
answer_1 = Question("Which record producer produced the bluegrass album?")
fact_1 = Verify("{answer_1} was born on 22 June, 1944.")
fact_2 = Verify("This album inspired a Tony award winning musical.")
fact_3 = Verify("This musical had a character that was originated by Carmen Cusack.")
label = Predict(fact_1 and fact_2 and fact_3)
answer_1 = Question("Which record producer produced the bluegrass album?")
fact_1 = Verify("{answer_1} was born on 22 June, 1944.")
fact_2 = Verify("The bluegrass album inspired a Tony award winning musical.")
fact_3 = Verify("{answer_1} had a character that was originated by Carmen Cusack.")
label = Predict(fact_1 and fact_2 and fact_3)
Semantic Error —Structure: incorrect program structure
Example 3:
Richard Strauss, German composer of Die Nacht, composed another well-established opera called Der
Rosenkavalier.
Predicted Program:
fact_1 = Verify("Richard Strauss, German composer of Die Nacht, composed another well-established
opera called Der Rosenkavalier.")
label = Predict(fact_1)
fact_1 = Verify("Richard Strauss is a German composer of Die Nacht.”)
fact_2 = Verify("Richard Strauss composed a well-established opera called Der Rosenkavalier.")
label = Predict(fact_1 and fact_2)
Example 4:
The film Deanna Oliver produced in 1999 grossed $36.8 million domestically. The musical film
based on coach Herman Boone, did not.
Predicted Program:
fact_1 = Verify("The film Deanna Oliver produced in 1999 grossed $36.8 million domestically.")
fact_2 = Verify("The musical film based on coach Herman Boone, did not.")
label = Predict(fact_1 and fact_2)
answer_1 = Question("What film did Deanna Oliver produce in 1999?")
fact_1 = Verify("{answer_1} grossed $36.8 million domestically.")
answer_2 = Question("What musical film was based on coach Herman Boone?")
fact_2 = Verify("{answer_2} grossed less than $36.8 million domestically.")
label = Predict(fact_1 and fact_2)
Semantic Error —Subtask: missing / redundant / incorrect sub-task calls
Example 5:
The musician, who founded Morningwood with Max Green, is older than Max Green.
Predicted Program:
answer_1 = Question("Who founded Morningwood with Max Green?")
answer_2 = Question("When was Max Green born?")
answer_3 = Question("When was the musician born?")
fact_1 = Verify("{answer_3} is older than {answer_2}.") {answer_1} is older than {answer_2}.
label = Verify(fact_1)
Figure8: Exampleerrorcaseswherethegeneratedreasoningprogramsareincorrect. Theincorrectsegment(s)are
markedinred,andthecorrectrevisionsaremarkedingreen.
6997
1
'''Generate a python-like program that describes the reasoning steps required to
verify the claim step-by-step. You can call three functions in the program: 1.
Question () to answer a question; 2. Verify () to verify a simple claim; 3.
| Predict() | to | predict | the | veracity | | label.''' | | |
| --------- | --- | ------- | --- | -------- | --- | --------- | --- | --- |
# The claim is that Howard University Hospital and Providence Hospital are both
| located | in Washington, | | | D.C. | | | | |
| ------- | -------------- | --- | --- | ---- | --- | --- | --- | --- |
def program():
fact_1 = Verify("Howard University Hospital is located in Washington, D.C.")
fact_2 = Verify("Providence Hospital is located in Washington, D.C.")
| label = | Predict(fact_1 | | | and | fact_2) | | | |
| ------- | -------------- | --- | --- | --- | ------- | --- | --- | --- |
# The claim is that WWE Super Tuesday took place at an arena that currently goes by
| the name | TD Garden. | | | | | | | |
| -------- | ---------- | --- | --- | --- | --- | --- | --- | --- |
def program():
answer_1 = Question("Which arena the WWE Super Tuesday took place?")
fact_1 = Verify(f"{answer_1} currently goes by the name TD Garden.")
| label = | Predict(fact_1) | | | | | | | |
| ------- | --------------- | --- | --- | --- | --- | --- | --- | --- |
# The claim is that Talking Heads, an American rock band that was "one of the most
critically acclaimed bands of the 80's" is featured in KSPN's AAA format.
def program():
fact_1 = Verify("Talking Heads is an American rock band that was 'one of the
| most critically | | acclaimed | | bands | of | the | 80's'.") | |
| --------------- | --- | --------- | --- | ----- | --- | --- | -------- | --- |
fact_2 = Verify("Talking Heads is featured in KSPN's AAA format.")
| label = | Predict(fact_1 | | | and | fact_2) | | | |
| ------- | -------------- | --- | --- | --- | ------- | --- | --- | --- |
# The claim is that An IndyCar race driver drove a Formula 1 car designed by Peter
| McCool | during | the | 2007 | Formula | One | season. | | |
| ------ | ------ | --- | ---- | ------- | --- | ------- | --- | --- |
def program():
answer_1 = Question("Which Formula 1 car was designed by Peter McCool during the
| 2007 Formula | | One | season?") | | | | | |
| ------------ | --- | --- | --------- | --- | --- | --- | --- | --- |
fact_1 = Verify(f"An IndyCar race driver drove the car {answer_1}.")
| label = | Predict(fact_1) | | | | | | | |
| ------- | --------------- | --- | --- | --- | --- | --- | --- | --- |
# The claim is that Gina Bramhill was born in a village. The 2011 population of the
| area that | includes | | this | village | was | 167,446. | | |
| --------- | -------- | --- | ---- | ------- | --- | -------- | --- | --- |
def program():
answer_1 = Question("Which village was Gina Bramhill born in?")
fact_1 = Verify(f"The 2011 population of the area that includes {answer_1} was
167,446.")
| label = | Predict(fact_1) | | | | | | | |
| ------- | --------------- | --- | --- | --- | --- | --- | --- | --- |
# The claim is that Don Ashley Turlington graduated from Saint Joseph's College, a
| private | Catholic | liberal | | arts | college | | in Standish. | |
| ------- | -------- | ------- | --- | ---- | ------- | --- | ------------ | --- |
def program():
fact_1 = Verify("Saint Joseph's College is a private Catholic liberal arts
| college | is located | | in Standish.") | | | | | |
| ------- | ---------- | --- | -------------- | --- | --- | --- | --- | --- |
fact_2 = Verify(f"Don Ashley Turlington graduated from Saint Joseph's College.")
| label = | Predict(fact_1 | | | and | fact_2) | | | |
| ------- | -------------- | --- | --- | --- | ------- | --- | --- | --- |
# The claim is that Gael and Fitness are not published in the same country.
def program():
| answer_1 | = Question("Which | | | | country | was | Gael published | in?") |
| -------- | ----------------- | --- | --- | --- | ------- | --- | -------------- | ----- |
answer_2 = Question("Which country was Fitness published in?")
fact_1 = Verify(f"{answer_1} and {answer_2} are not the same country.")
| label = | Predict(fact_1) | | | | | | | |
| ------- | --------------- | --- | --- | --- | --- | --- | --- | --- |
# The claim is that Blackstar is the name of the album released by David Bowie that
| was recorded | | in secret. | | | | | | |
| ------------ | --- | ---------- | --- | --- | --- | --- | --- | --- |
def program():
fact_1 = Verify("David Bowie released an album called Blackstar.")
| fact_2 | = Verify("David | | | Bowie | recorded | | an album | in secret.") |
| ------- | --------------- | --- | --- | ----- | -------- | --- | -------- | ------------ |
| label = | Predict(fact_1 | | | and | fact_2) | | | |
# The claim is that In the 2004 Hockey film produced by a former major league
| baseball | pitcher | Kurt | Russell | | played | the | USA coach. | |
| -------- | ------- | ---- | ------- | --- | ------ | --- | ---------- | --- |
def program():
answer_1 = Question("Which 2004 Hockey film was produced a former major league
6998
| baseball | pitcher?") | | | | | |
| -------- | ---------- | --- | --- | --- | --- | --- |
fact_1 = Verify("Kurt Russell played the USA coach in the film {answer_1}.")
| label = | Predict(fact_1) | | | | | |
| ------- | --------------- | --- | --- | --- | --- | --- |
# The claim is that Along with the New York Islanders and the New York Rangers, the
New Jersey Devils NFL franchise is popular in the New York metropolitan area.
def program():
fact_1 = Verify("The New York Islanders and the New York Rangers are popular in
| the New | York metropolitan | | area.") | | | |
| ------- | ----------------- | --- | ------- | --- | --- | --- |
fact_2 = Verify("The New Jersey Devils NFL franchise is popular in the New York
| metropolitan | area.") | | | | | |
| ------------ | -------------- | --- | --- | ------- | --- | --- |
| label = | Predict(fact_1 | | and | fact_2) | | |
# The claim is that Jack McFarland is the best known role of the host of the 64th
| Annual Tony | Awards. | | | | | |
| ----------- | ------- | --- | --- | --- | --- | --- |
def program():
answer_1 = Question("Who is the host of the 64th Annual Tony Awards?")
fact_1 = Verify(f\"Jack McFarland is the best known role of {answer_1}.")
| label = | Predict(fact_1) | | | | | |
| ------- | --------------- | --- | --- | --- | --- | --- |
# The claim is that The song recorded by Fergie that was produced by Polow da Don
| and was | followed | by | Life Goes | On | was M.I.L.F.$. | |
| ------- | -------- | --- | --------- | --- | -------------- | --- |
def program():
fact_1 = Verify("M.I.L.F.$ was recorded by Fergie that was produced by Polow da
Don.")
fact_2 = Verify("M.I.L.F.$ was was followed by Life Goes On.")
| label = | Predict(fact_1 | | and | fact_2) | | |
| ------- | -------------- | --- | --- | ------- | --- | --- |
# The claim is that Eatza Pizza and Your Pie were not founded in the same state.
def program():
answer_1 = Question("Which state was Eatza Pizza founded in?")
| answer_2 | = Question("Which | | | state | was Your Pie founded | in?") |
| -------- | ----------------- | --- | --- | ----- | -------------------- | ----- |
fact_1 = Verify(f"{answer_1} and {answer_2} are not the same state.")
| label = | Predict(fact_1) | | | | | |
| ------- | --------------- | --- | --- | --- | --- | --- |
# The claim is that Gregg Rolie and Rob Tyner, are not a keyboardist.
def program():
| fact_1 = | Verify("Gregg | | Rolie | is not | a keyboardist.") | |
| -------- | -------------- | --- | ----- | ------- | ---------------- | --- |
| fact_2 = | Verify("Rob | | Tyner | is not | a keyboardist.") | |
| label = | Predict(fact_1 | | and | fact_2) | | |
# The claim is that Maria Esther Andion Bueno, not Jimmy Connors, is the player that
| is from | Brazil. | | | | | |
| ------- | ------- | --- | --- | --- | --- | --- |
def program():
| fact_1 = | Verify("Maria | | Esther | Andion | Bueno is from | Brazil.") |
| -------- | -------------- | --- | ------- | ------- | ------------------ | --------- |
| fact_2 = | Verify("Jimmy | | Connors | is | not from Brazil.") | |
| label = | Predict(fact_1 | | and | fact_2) | | |
# The claim is that Vladimir Igorevich Arnold died after Georg Cantor.
def program():
answer_1 = Question("When did Vladimir Igorevich Arnold die?")
| answer_2 | = Question("When | | | did Georg | Cantor die?") | |
| -------- | ------------------- | --- | --- | --------- | ------------- | --- |
| fact_1 = | Verify(f"{answer_1} | | | is after | {answer_2}.") | |
| label = | Predict(fact_1) | | | | | |
# The claim is that Barton Mine was halted by a natural disaster not Camlaren Mine.
def program():
fact_1 = Verify("Barton Mine was halted by a natural disaster.")
fact_2 = Verify("Camlaren Mine was not halted by a natural disaster.")
| label = | Predict(fact_1 | | and | fact_2) | | |
| ------- | -------------- | --- | --- | ------- | --- | --- |
# The claim is that John O'Hara and Rabindranath Tagore are not the same nationality
.
def program():
answer_1 = Question("What is the nationality of John O'Hara?")
answer_2 = Question("What is the nationality of Rabindranath Tagore?")
fact_1 = Verify(f"{answer_1} and {answer_2} are not the same nationality.")
| label = | Predict(fact_1) | | | | | |
| ------- | --------------- | --- | --- | --- | --- | --- |
6999
# The claim is that Thomas Loren Friedman has won more Pulitzer Prizes than Colson
Whitehead.
def program():
answer_1 = Question("How many Pulitzer Prizes has Thomas Loren Friedman won?")
answer_2 = Question("How many Pulitzer Prizes has Colson Whitehead won?")
fact_1 = Verify(f"{answer_1} is more than {answer_2}.")
label = Predict(fact_1)
# The claim is that The model of car Trevor Bayne drives was introduced for model
year 2006. The Rookie of The Year in the 1997 CART season drives it in the
NASCAR Sprint Cup.
def program():
answer_1 = Question("Which model of car is drived by Trevor Bayne?")
fact_1 = Verify(f"{answer_1} was introduced for model year 2006.")
answer_2 = Question("Who is the Rookie of The Year in the 1997 CART season?")
fact_2 = Verify(f"{answer_2} drives the model of car Trevor Bayne drives in the
NASCAR Sprint Cup.")
label = predict(fact_1 and fact_2)
# The claim is that <input_claim>
def program():
Listing1: ThepromptusedforProgramGenerationforHOVER.
7000
'''Generate a python-like program that describes the reasoning steps required to
verify the claim step-by-step. You can call three functions in the program: 1.
Question () to answer a question; 2. Verify () to verify a simple claim; 3.
| Predict() | to predict | the | veracity | label.''' | | |
| --------- | ---------- | --- | -------- | --------- | --- | --- |
# The claim is that In 1959, former Chilean boxer Alfredo Cornejo Cuevas (born June
6, 1933) won the gold medal in the welterweight division at the Pan American
Games (held in Chicago, United States, from August 27 to September 7) in Chicago
, United States, and the world amateur welterweight title in Mexico City.
def program():
fact_1 = Verify("Alfredo Cornejo Cuevas was born in June 6, 1933.")
fact_2 = Verify("Alfredo Cornejo Cuevas won the gold medal in the welterweight
| division | at the Pan | American | Games | in | 1959.") | |
| -------- | ---------- | -------- | ----- | --- | ------- | --- |
fact_3 = Verify("The Pan American Games in 1959 was held in Chicago, United
| States, | from August | 27 to | September | 7.") | | |
| ------- | ----------- | ----- | --------- | ---- | --- | --- |
fact_4 = Verify("Alfredo Cornejo Cuevas won the world amateur welterweight title
| in Mexico | City.") | | | | | |
| --------- | -------------- | --- | ------ | --- | ------ | ----------- |
| label = | Predict(fact_1 | and | fact_2 | and | fact_3 | and fact_4) |
# The claim is that The Footwork FA12, which was intended to start the season,
finally debuted at the San Marino Grand Prix, a Formula One motor race held at
| Imola on | 28 April | 1991. | | | | |
| -------- | -------- | ----- | --- | --- | --- | --- |
def program():
fact_1 = Verify("The Footwork FA12, which was intended to start the season.")
fact_2 = Verify("The Footwork FA12 finally debuted at the San Marino Grand Prix.
")
fact_3 = Verify("The San Marino Grand Prix was a Formula One motor race held at
| Imola on | 28 April | 1991.") | | | | |
| -------- | -------------- | ------- | ------ | --- | ------- | --- |
| label = | Predict(fact_1 | and | fact_2 | and | fact_3) | |
# The claim is that SkyHigh Mount Dandenong (formerly Mount Dandenong Observatory)
is a restaurant located on top of Mount Dandenong, Victoria, Australia.
def program():
fact_1 = Verify("SkyHigh Mount Dandenong is a restaurant located on top of Mount
| Dandenong, | Victoria, | Australia.") | | | | |
| ---------- | --------- | ------------ | --- | --- | --- | --- |
fact_2 = Verify("SkyHigh Mount Dandenong is formerly known as Mount Dandenong
Observatory.")
| label = | Predict(fact_1 | and | fact_2) | | | |
| ------- | -------------- | --- | ------- | --- | --- | --- |
# The claim is that Before the first Europeans arrived or copra companies leased it,
| Maupihaa | was home | to Inca's | in | ancient | times. | |
| -------- | -------- | --------- | --- | ------- | ------ | --- |
def program():
fact_1 = Verify("Maupihaa was home to Inca's in ancient times.")
fact_2 = Verify("Maupihaa was home to Inca's before the first Europeans arrived
| or copra | companies | leased | it.") | | | |
| -------- | -------------- | ------ | ------- | --- | --- | --- |
| label = | Predict(fact_1 | and | fact_2) | | | |
# The claim is that Shulin, a 33.1288 km (12.7911 sq mi) land located in New Taipei
City, China, a country in East Asia, has a total population of 183,946 in
| December | 2018. | | | | | |
| -------- | ----- | --- | --- | --- | --- | --- |
def program():
fact_1 = Verify("Shulin is a 33.1288 km (12.7911 sq mi) land located in New
| Taipei | City, China.") | | | | | |
| ------ | -------------- | --- | --- | --- | --- | --- |
fact_2 = Verify("Shulin has a total population of 183,946 in December 2018.")
| label = | Predict(fact_1 | and | fact_2) | | | |
| ------- | -------------- | --- | ------- | --- | --- | --- |
# The claim is that Sumo wrestler Toyozakura Toshiaki committed match-fixing, ending
| his career | in 2011 | that | started | in 1989. | | |
| ---------- | ------- | ---- | ------- | -------- | --- | --- |
def program():
fact_1 = Verify("Toyozakura Toshiaki ended his career in 2011 that started in
1989.")
| fact_2 | = Verify("Toyozakura | | Toshiaki | | is a | Sumo wrestler.") |
| ------ | -------------------- | --- | -------- | --- | ---- | ---------------- |
fact_3 = Verify("Toyozakura Toshiaki committed match-fixing.")
| label = | Predict(fact_1 | and | fact_2 | and | fact_3) | |
| ------- | -------------- | --- | ------ | --- | ------- | --- |
# The claim is that In 1959, former Chilean boxer Alfredo Cornejo Cuevas (born June
6, 1933) won the gold medal in the welterweight division at the Pan American
Games (held in Chicago, United States, from August 27 to September 7) in Chicago
7001
, United States, and the world amateur welterweight title in Mexico City.
def program():
fact_1 = Verify("Alfredo Cornejo Cuevas is a former Chilean boxer.")
fact_2 = Verify("Alfredo Cornejo won the gold medal in the welterweight division
| at the | Pan American | | Games.") | | |
| ------ | ------------ | --- | -------- | --- | --- |
fact_3 = Verify("The Pan American Games was held in Chicago, United States, from
| August | 27 to | September | 7.") | | |
| ------ | ----- | --------- | ---- | --- | --- |
fact_4 = Verify("Alfredo Cornejo won the world amateur welterweight title in
Mexico City.")
| label = | Predict(fact_1 | | and fact_2 | and fact_3 | and fact_4) |
| ------- | -------------- | --- | ---------- | ---------- | ----------- |
# The claim is that Adductor hiatus is associated with nine structures, seven of
| which enter | and | leave | through hiatus. | | |
| ----------- | --- | ----- | --------------- | --- | --- |
def program():
fact_1 = Verify("Adductor hiatus is associated with nine structures.")
fact_2 = Verify("Seven of the nine structures associated with Adductor hiatus
| enter and | leave | through | hiatus.") | | |
| --------- | -------------- | ------- | ----------- | --- | --- |
| label = | Predict(fact_1 | | and fact_2) | | |
# The claim is that Ifor Bowen Lloyd was educated at Winchester (an independent
boarding school for boys in the British public school tradition) and Exeter
College, Oxford where he was a member of the Library Committee of the Oxford
Union Society, as well as, received a BA in Modern History in 1924.
def program():
fact_1 = Verify("Ifor Bowen Lloyd was educated at Winchester and Exeter College,
Oxford.")
fact_2 = Verify("Winchester is an independent boarding school for boys in the
| British | public | school | tradition.") | | |
| ------- | ------ | ------ | ------------ | --- | --- |
fact_3 = Verify("While at Oxford, Ifor Bowen Lloyd was a member of the Library
| Committee | of the | Oxford | Union Society.") | | |
| --------- | ------ | ------ | ---------------- | --- | --- |
fact_4 = Verify("Ifor Bowen Lloyd received a BA in Modern History in 1924 at
Oxford.")
| label = | Predict(fact_1 | | and fact_2 | and fact_3 | and fact_4) |
| ------- | -------------- | --- | ---------- | ---------- | ----------- |
# The claim is that In the 2001 Stanley Cup playoffs Eastern Conference Semifinals
Devils' Elias scored and Maple Leafs' left Devils player Scott Neidermayer hurt.
def program():
fact_1 = Verify("In the 2001 Stanley Cup playoffs Eastern Conference Semifinals
| Devils' | Elias | scored.") | | | |
| ------- | ----- | --------- | --- | --- | --- |
fact_2 = Verify("Maple Leafs' left Devils player Scott Neidermayer hurt.")
| label = | Predict(fact_1 | | and fact_2) | | |
| ------- | -------------- | --- | ----------- | --- | --- |
# The claim is that Teldenia helena is a moth first described in 1967 by Wilkinson.
def program():
| fact_1 | = Verify("Teldenia | | helena | is a moth.") | |
| ------ | ------------------ | --- | ------ | ------------ | --- |
fact_2 = Verify("Teldenia helena was first described by Wilkinson in 1967.")
| label = | Predict(fact_1 | | and fact_2) | | |
| ------- | -------------- | --- | ----------- | --- | --- |
# The claim is that Born December 30, 1974, William Frick was a dark horse candidate
| in the | Maryland | House | of Delegates | appointment | process. |
| ------ | -------- | ----- | ------------ | ----------- | -------- |
def program():
fact_1 = Verify("William Frick was born in December 30, 1974.")
fact_2 = Verify("William Frick was a dark horse candidate in the Maryland House
| of Delegates | appointment | | process.") | | |
| ------------ | -------------- | ------------- | ----------- | --- | --- |
| label = | Predict(fact_1 | | and fact_2) | | |
| # The claim | is that | <input_claim> | | | |
def program():
| | Listing2: | ThepromptusedforProgramGenerationforFEVEROUS-S. | | | |
| --- | --------- | ----------------------------------------------- | --- | --- | --- |
7002
ACL2023ResponsibleNLPChecklist
A Foreverysubmission:
3
(cid:3) A1. Didyoudescribethelimitationsofyourwork?
Line587-620
3
(cid:3) A2. Didyoudiscussanypotentialrisksofyourwork?
Line626-630
3
(cid:3) A3. Dotheabstractandintroductionsummarizethepaper’smainclaims?
Line67-86
(cid:3) 7 A4. HaveyouusedAIwritingassistantswhenworkingonthispaper?
Leftblank.
3
B (cid:3) Didyouuseorcreatescientificartifacts?
Line327-352
3
(cid:3) B1. Didyoucitethecreatorsofartifactsyouused?
Line328-329
(cid:3) B2. Didyoudiscussthelicenseortermsforuseand/ordistributionofanyartifacts?
Notapplicable. Thedatasetsusedinthispaperarepubliclyavailabledatasetsfromexistingworks.
3
(cid:3) B3. Didyoudiscussifyouruseofexistingartifact(s)wasconsistentwiththeirintendeduse,provided
thatitwasspecified? Fortheartifactsyoucreate,doyouspecifyintendeduseandwhetherthatis
compatiblewiththeoriginalaccessconditions(inparticular,derivativesofdataaccessedforresearch
purposesshouldnotbeusedoutsideofresearchcontexts)?
Line327-344
(cid:3) B4. Didyoudiscussthestepstakentocheckwhetherthedatathatwascollected/usedcontainsany
informationthatnamesoruniquelyidentifiesindividualpeopleoroffensivecontent,andthesteps
takentoprotect/anonymizeit?
Notapplicable. Leftblank.
(cid:3) B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and
linguisticphenomena,demographicgroupsrepresented,etc.?
Notapplicable. Leftblank.
3
(cid:3) B6. Didyoureportrelevantstatisticslikethenumberofexamples,detailsoftrain/test/devsplits,
etc. forthedatathatyouused/created? Evenforcommonly-usedbenchmarkdatasets,includethe
numberofexamplesintrain/validation/testsplits,astheseprovidenecessarycontextforareader
tounderstandexperimentalresults. Forexample,smalldifferencesinaccuracyonlargetestsetsmay
besignificant,whileonsmalltestsetstheymaynotbe.
Line327-344
3
C (cid:3) Didyouruncomputationalexperiments?
Section4
3
(cid:3) C1. Didyoureportthenumberofparametersinthemodelsused, thetotalcomputationalbudget
(e.g.,GPUhours),andcomputinginfrastructureused?
Figure4;AppendixA
TheResponsibleNLPChecklistusedatACL2023isadoptedfromNAACL2022,withtheadditionofaquestiononAIwriting
assistance.
7003
3
(cid:3) C2. Did you discuss the experimental setup, including hyperparameter search and best-found
hyperparametervalues?
AppendixA
3
(cid:3) C3. Didyoureportdescriptivestatisticsaboutyourresults(e.g.,errorbarsaroundresults,summary
statisticsfromsetsofexperiments),andisittransparentwhetheryouarereportingthemax,mean,
etc. orjustasinglerun?
AppendixA
(cid:3) C4. Ifyouusedexistingpackages(e.g.,forpreprocessing,fornormalization,orforevaluation),did
youreporttheimplementation,model,andparametersettingsused(e.g.,NLTK,Spacy,ROUGE,
etc.)?
Notapplicable. Leftblank.
7
D (cid:3) Didyouusehumanannotators(e.g.,crowdworkers)orresearchwithhumanparticipants?
Leftblank.
(cid:3) D1. Did you report the full text of instructions given to participants, including e.g., screenshots,
disclaimersofanyriskstoparticipantsorannotators,etc.?
Notapplicable. Leftblank.
(cid:3) D2. Didyoureportinformationabouthowyourecruited(e.g.,crowdsourcingplatform,students)
andpaidparticipants,anddiscussifsuchpaymentisadequategiventheparticipants’demographic
(e.g.,countryofresidence)?
Notapplicable. Leftblank.
(cid:3) D3. Did you discuss whether and how consent was obtained from people whose data you’re
using/curating? For example, if you collected data via crowdsourcing, did your instructions to
crowdworkersexplainhowthedatawouldbeused?
Notapplicable. Leftblank.
(cid:3) D4. Wasthedatacollectionprotocolapproved(ordeterminedexempt)byanethicsreviewboard?
Notapplicable. Leftblank.
(cid:3) D5. Didyoureportthebasicdemographicandgeographiccharacteristicsoftheannotatorpopulation
thatisthesourceofthedata?
Notapplicable. Leftblank.
7004 |