File size: 108,915 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 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 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 |
{
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:27:44.085616Z"
},
"title": "Reducing Non-Normative Text Generation from Language Models",
"authors": [
{
"first": "Xiangyu",
"middle": [],
"last": "Peng",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Georgia Institute of Technology Atlanta",
"location": {
"postCode": "30332",
"region": "GA"
}
},
"email": ""
},
{
"first": "Siyan",
"middle": [],
"last": "Li",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Georgia Institute of Technology Atlanta",
"location": {
"postCode": "30332",
"region": "GA"
}
},
"email": ""
},
{
"first": "Spencer",
"middle": [],
"last": "Frazier",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Georgia Institute of Technology Atlanta",
"location": {
"postCode": "30332",
"region": "GA"
}
},
"email": "sfrazier7@gatech.edu"
},
{
"first": "Mark",
"middle": [],
"last": "Riedl",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Georgia Institute of Technology Atlanta",
"location": {
"postCode": "30332",
"region": "GA"
}
},
"email": "riedl@gatech.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Large-scale, transformer-based language models such as GPT-2 are pretrained on diverse corpora scraped from the internet. Consequently, they are prone to generating non-normative text (i.e. in violation of social norms). We introduce a technique for fine-tuning GPT-2, using a policy gradient reinforcement learning technique and a normative text classifier to produce reward and punishment values. We evaluate our technique on five data sets using automated and human participant experiments. The normative text classifier is 81-90% accurate when compared to gold-standard human judgements of normative and non-normative generated text. Our normative fine-tuning technique is able to reduce non-normative text by 27-61%, depending on the data set.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "Large-scale, transformer-based language models such as GPT-2 are pretrained on diverse corpora scraped from the internet. Consequently, they are prone to generating non-normative text (i.e. in violation of social norms). We introduce a technique for fine-tuning GPT-2, using a policy gradient reinforcement learning technique and a normative text classifier to produce reward and punishment values. We evaluate our technique on five data sets using automated and human participant experiments. The normative text classifier is 81-90% accurate when compared to gold-standard human judgements of normative and non-normative generated text. Our normative fine-tuning technique is able to reduce non-normative text by 27-61%, depending on the data set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Human societies implicitly establish codes of acceptable behavior in social contexts. Normativity is behavior that conforms to expected societal norms and contracts, whereas non-normative behavior aligns to values that deviate from these expected norms. Sumner (1967) defines norms as: \"...informal rules that are not written, but, when violated, result in severe punishments and social sanction upon the individuals, such as social and religious exclusions.\" Non-normativity does not connote behavior devoid of value or immoral, but behavior that fails to conform to social standards shared by other individuals in the relevant group, organization or society. Norms can also be thought of as actions taken by an entity which conform to an identity (Katzenstein, 1996) , thus allowing others to categorize behavior as in-group or out-group. Different societies and groups collectively have different ideals about what actions constitute normative behavior; group members use these ideals * Equal contributions to heuristically guide their actions to avoid social ostracization. For example, many societies have norms against violence, or certain behaviors being conducted in public. Conflicts between individuals can arise when enacting non-normative behaviors or uttering non-normative speech.",
"cite_spans": [
{
"start": 254,
"end": 267,
"text": "Sumner (1967)",
"ref_id": "BIBREF21"
},
{
"start": 749,
"end": 768,
"text": "(Katzenstein, 1996)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This paper examines generative language models and the frequency at which they generate descriptions of non-normative behavior. Large-scale, transformer-based neural language models such as ELMo (Peters et al., 2018) , BERT (Devlin et al., 2018) , GPT-2 (Radford et al., 2019) , GPT-3, Grover (Zellers et al., 2019) , CTRL (Keskar et al., 2019) , T5 (Raffel et al., 2019) , and XLNet (Yang et al., 2019) are trained on very large corpora such as text scraped from the internet, books, or both.",
"cite_spans": [
{
"start": 195,
"end": 216,
"text": "(Peters et al., 2018)",
"ref_id": "BIBREF14"
},
{
"start": 224,
"end": 245,
"text": "(Devlin et al., 2018)",
"ref_id": "BIBREF6"
},
{
"start": 248,
"end": 276,
"text": "GPT-2 (Radford et al., 2019)",
"ref_id": null
},
{
"start": 293,
"end": 315,
"text": "(Zellers et al., 2019)",
"ref_id": "BIBREF24"
},
{
"start": 318,
"end": 344,
"text": "CTRL (Keskar et al., 2019)",
"ref_id": null
},
{
"start": 350,
"end": 371,
"text": "(Raffel et al., 2019)",
"ref_id": "BIBREF16"
},
{
"start": 384,
"end": 403,
"text": "(Yang et al., 2019)",
"ref_id": "BIBREF23"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "These language models generate text that is statistically representative of the corpora they were trained on. As such, text scraped from the internet co-mingles text produced by many groups with differing norms, as well as text produced by people intentionally using non-normative speech, like \"trolling\" language. Models trained on these data can then produce undesirable, harmful output. Stories from the internet and books also contain normative and non-normative situations (e.g., antagonists, as well as protagonists conducting conventionally non-normative behaviors). Consequently, it is possible, and often likely, for language models to generate non-normative descriptions of behavior (murder, crime, suicide, racist actions, rude behavior, etc.), exhibit biases against certain demographics groups (Sheng et al., 2019; Solaiman et al., 2019) , stereotypical biases (Nadeem et al., 2020) or racist text when prompted with trigger phrases (Wallace et al., 2019) .",
"cite_spans": [
{
"start": 807,
"end": 827,
"text": "(Sheng et al., 2019;",
"ref_id": "BIBREF18"
},
{
"start": 828,
"end": 850,
"text": "Solaiman et al., 2019)",
"ref_id": "BIBREF20"
},
{
"start": 874,
"end": 895,
"text": "(Nadeem et al., 2020)",
"ref_id": "BIBREF13"
},
{
"start": 946,
"end": 968,
"text": "(Wallace et al., 2019)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Value alignment (Russell et al., 2015) is the concept that an agent is unable to perform actions that cause harm to humans. Harmful behavior is not limited to physical actions by robots, the focus of some AI value alignment research. We recognize that natural language communication can also cause harm. For example, Amazon Alexa, a virtual assistant AI, was reported to suggest a user commit suicide. 1 Frazier et al. (2019) developed a classifier for normative behavior which exhibits strong zeroshot and few-shot transfer across a variety of text corpora. The authors speculate that their modelwhich they call a value-aligned prior-can bias model output perceived as more normative. In this paper we ask a different question: whether a valuealigned prior can be used to reduce the generation of descriptions of non-normative behavior by neural language models.",
"cite_spans": [
{
"start": 16,
"end": 38,
"text": "(Russell et al., 2015)",
"ref_id": "BIBREF17"
},
{
"start": 402,
"end": 403,
"text": "1",
"ref_id": null
},
{
"start": 404,
"end": 425,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The common approach to fine-tuning language models is to provide additional corpora of exemplars. If a corpus of exemplars is normative, the language model can be trained to emulate this over time. Generally, in the absence of very large normative corpora, we need an alternative approach to fine-tuning language models. We use a reinforcement learning approach to fine-tuning language models, using the normative behavior classifier of Frazier et al. (2019) as a non-differentiable reward function. Our method back-propagates reward relative to the degree of non-normativity of text generated by the language model. We evaluate our reinforcement learning finetuning technique with three sets of experiments. First, we replicate the experiments by Frazier et al. (2019) on text generated by a language model instead of originally held-out corpus text. Second, we show with automated and human participant experiments that fine-tuning on reward generated by a normative classifier model can reduce the generation of non-normative text by 27 \u2212 61%. Third, we ablate our technique and show with automated and human participant experiments that the fine-tuning technique works with classifiers other than the normative classifier-specifically models trained to classify negative-sentiment and toxic language.",
"cite_spans": [
{
"start": 437,
"end": 458,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
},
{
"start": 748,
"end": 769,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Humans have expectations that -just like other humans -agents will avoid harmful actions, conform to personal values and to social norms (Bicchieri, 2005) , even when not explicitly communicated. This is referred to as the value alignment problem (Soares and Fallenstein, 2014; Russell et al., 2015; Arnold et al., 2017; Abel et al., 2016) . Harmful agent behavior can theoretically be mitigated by casting values as preferences over action sequences. For example Christiano et al. (2017) collected human preferences to shape rewards for game-playing agents in reinforcement learning.",
"cite_spans": [
{
"start": 137,
"end": 154,
"text": "(Bicchieri, 2005)",
"ref_id": "BIBREF3"
},
{
"start": 247,
"end": 277,
"text": "(Soares and Fallenstein, 2014;",
"ref_id": "BIBREF19"
},
{
"start": 278,
"end": 299,
"text": "Russell et al., 2015;",
"ref_id": "BIBREF17"
},
{
"start": 300,
"end": 320,
"text": "Arnold et al., 2017;",
"ref_id": "BIBREF2"
},
{
"start": 321,
"end": 339,
"text": "Abel et al., 2016)",
"ref_id": "BIBREF0"
},
{
"start": 464,
"end": 488,
"text": "Christiano et al. (2017)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Value Alignment and Normative Priors",
"sec_num": "2.1"
},
{
"text": "Instead of preference learning, Frazier et al. (2019) used the BERT (Devlin et al., 2018) language model's token embeddings to train a binary classifier. This model is used to differentiate between normative and non-normative natural language sentences containing events, utterances and descriptions of behavior. They obtained training data from Goofus & Gallant (G&G), a children's educational comic strip featuring two characters of the same names. Goofus always deviates from the \"proper\" way to behave, while Gallant always performs the behavior of an exemplary child in western society at the time the comics were created. As a result, G&G is a naturally labeled source of normative and non-normative text, for the specific society it represents. Frazier et al. (2019) demonstrated this method could accurately classify descriptions of behavior as normative or non-normative. Furthermore, this classifier retained high performance in zeroshot and few-shot transfer tasks. For example, they show that their classifier, trained on G&G comics, can classify normative event descriptions in contemporary collections of popular plot points and science fiction plot summaries, instances of medium-and far-transfer, respectively. The authors speculate that their classifier model can bias agent behavior toward normative courses of action in other contexts. However, this was not directly shown. We ask whether a normative classifier can be used to fine-tune the \"behavior\" of a large-scale transformer-based language model.",
"cite_spans": [
{
"start": 32,
"end": 53,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
},
{
"start": 68,
"end": 89,
"text": "(Devlin et al., 2018)",
"ref_id": "BIBREF6"
},
{
"start": 752,
"end": 773,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Value Alignment and Normative Priors",
"sec_num": "2.1"
},
{
"text": "Large-scale transformer-based neural language models such as BERT and GPT-2 are trained on large corpora of text scraped from the web and books. They can be fine-tuned to a specific domain of interest, commonly accomplished by providing a corpus of exemplars from that domain. Over time, the weights of the pre-trained model will shift and increasingly generate passages which better emulate the corpus of exemplars. If the fine-tuning corpus of exemplars is normative, the language model will, in theory, learn to prefer normative language over time. Goofus & Gallant is one such normative corpus, and if a language model is fine-tuned on it then it may prefer to generate normative language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Language Model Training & Fine-Tuning",
"sec_num": "2.2"
},
{
"text": "GPT-2 , in particular, is a large-scale transformer-based language model trained on a large corpus of text scraped from web pages and social media. Applying the concept of value alignment as preference learning, Ziegler et al. (2019) use a reinforcement learning method on the 774M-parameter version of GPT-2 to favor human-preferred text. Crowd workers were asked to select generated text completions from a set of given prompts that had positive sentiment. These preference values were used to fine-tune GPT-2. This is one possible technique for reinforcementbased fine-tuning; sentiment is, however, not necessarily a good measure of adherence to norms. We replace the linear reward model for fine-tuning GPT-2 (Ziegler et al., 2019) with a pre-trained normative text classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Language Model Training & Fine-Tuning",
"sec_num": "2.2"
},
{
"text": "The Plug and Play Language Models (PPLM) (Dathathri et al., 2019 ) also apply attribute classifiers to fine-tune language models; the technique is demonstrated via generating text with a target sentiment and also decreasing the frequency of toxic language. There are two limitations: (a) PPLM trains a model to operate on a fixed set of prefix input, and (b) the classification must be done on a word-by-word basis and thus cannot easily be applied to problems where the normative valence of individual words relies on a single or multiple sentence context (e.g. quoting and admonishing toxic speech). Our fine-tuning technique, in contrast, works on arbitrary prefixes and assesses the the normativity of entire sentences.",
"cite_spans": [
{
"start": 41,
"end": 64,
"text": "(Dathathri et al., 2019",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Language Model Training & Fine-Tuning",
"sec_num": "2.2"
},
{
"text": "We make use of five datasets, chosen to represent a diverse set of domains. The normative text classifier by Frazier et al. (2019) was tested on a corpus of science fiction plot summaries (Ammanabrolu et al., 2019) as well as a new Plotto dataset, based on a book by the same name that catalogues plot points for scaffolding fictional story-writing. Story corpora are particularly good for testing problems pertaining to textual descriptions of normative and non-normative behavior. Stories contain antagonists that frequently violate societal norms and protagonists who are more likely to exemplify contemporary social norms. We recognize many sto-ries require protagonists to perform non-normative behaviors like violence against others to achieve normative ends, further indicating the importance of accounting for a broader frame of context when determining normativity.",
"cite_spans": [
{
"start": 109,
"end": 130,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
},
{
"start": 188,
"end": 214,
"text": "(Ammanabrolu et al., 2019)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets",
"sec_num": "2.3"
},
{
"text": "The science-fiction plot summary corpus (Ammanabrolu et al., 2019) is a collection of 2,276 stories scraped from crowd-sourced plot summaries on fan sites. These stories have an average length of 89.23 sentences. Sentences in this corpus tend to give high-level overviews of the actions that characters are performing (e.g. \"Lyta accuses Sinclair of attempting to murder the ambassador\"). The sci-fi corpus also presents a transfer challenge because it involves a lot of novel entities-aliens, spaceships, laser weapons, etc.-that do not exist in Goofus & Gallant. It is notable that a normative text classifier trained on G&G would do well on zero-shot transfer to the sci-fi corpus. This makes it an attractive dataset for our experiments for the same reasons.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets",
"sec_num": "2.3"
},
{
"text": "The Plotto dataset consists of 900 sentences extracted from a book, which catalogues plot points used in popular fiction. Frazier et al. 2019pruned some exceptionally anachronistic and misogynistic sentences from the corpora. These sentences approximate the level of abstraction in the sci-fi corpus but have more contemporary narratives.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets",
"sec_num": "2.3"
},
{
"text": "The ROCstories (Mostafazadeh et al., 2016) corpus contains 52,666 five-sentence stories, often about everyday life situations (e.g. going for a jog, taking a test in school, etc.). Unlike the previous two corpora, it covers a different space of more common, mundane events which usually do not have strong normativity connotations.",
"cite_spans": [
{
"start": 15,
"end": 42,
"text": "(Mostafazadeh et al., 2016)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets",
"sec_num": "2.3"
},
{
"text": "Sentiment is often used as a surrogate for normativity under the belief that non-normative behavior would be associated with negative sentiment. The relationship between normativity and sentiment is not that simple, as we will show in Section 4.3. We include sentiment experiments using large review datasets from IMDb, Yelp, and Amazon (Kotzias et al., 2015) because (1) previous value alignment research has incorporated sentiment analysis, and (2) we want to test our techniques on classifiers other than the normative text classifier.",
"cite_spans": [
{
"start": 337,
"end": 359,
"text": "(Kotzias et al., 2015)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets",
"sec_num": "2.3"
},
{
"text": "Non-normativity is a superset of toxic language in the sense that toxic language is non-normative, but not all non-normative descriptions are toxic. We also conduct experiments using toxic language classifiers -fine-tuned on sentiment corpora like the dataset from the Toxic Comment Classification Challenge 2 as an alternative to the normative text classifier fine-tuned on G&G.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets",
"sec_num": "2.3"
},
{
"text": "The GPT-2 model is trained by minimizing its cross-entropy loss given by :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Normative Fine-Tuning",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "loss w (X, y) = \u2212 log exp(X[y]) i\u2208V exp(X[i]) = \u2212 log (\u03c3(X) y )",
"eq_num": "(1)"
}
],
"section": "Normative Fine-Tuning",
"sec_num": "3"
},
{
"text": "where X is a vector containing output logits and y is the index of the word from the ground truth in X. V is the model's vocabulary, \u03c3 is the softmax function, and \u03c3(X) y is the ground truth probability of the word. To punish GPT-2 for producing non-normative text, we use a normative text classifier to evaluate the model's performance and produce a reward value, which is applied to the loss and backpropagated through GPT-2. Given that the normativity of a sentence can only be determined by reading the entire sentence, the classifier must therefore produce a single numeric value per sentence. Specifically, we augment the cross-entropy loss computation with predictions from the pre-trained classifier. We define the sentence loss as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Normative Fine-Tuning",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "loss s (s) = 1 n j\u2208s loss w (X j , y j ) + u(s)",
"eq_num": "(2)"
}
],
"section": "Normative Fine-Tuning",
"sec_num": "3"
},
{
"text": "where s is the continuation sentence generated by the neural language model, n = |s| \u2212 1 is the number of the words in the continuation sentence, X j is the j th logit vector, and y j is the ground truth index for the j th word. u(s) is a function of the output of the classifier converted into a punishment value; a value of zero indicates no punishment, and higher positive values indicating increasingly nonnormative sentences. The loss w counter-balances the reward and prevents the generated texts from descending into incoherent fragments. The punishment function u(s) generates a value proportional to the average word loss so that it does not become overwhelmed by loss w :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Normative Fine-Tuning",
"sec_num": "3"
},
{
"text": "u(s) = \u03c1\u03b2(1 \u2212 C(s))( 1 n j\u2208s loss w (X j , y j )) (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Normative Fine-Tuning",
"sec_num": "3"
},
{
"text": "where s is a continuation sentence, C(s) is the binary {0, 1} label given by the normative classifier, \u03c1 is a hyper-parameter to control the strength of the penalty, and \u03b2 = (1 \u2212 i \u00d7 0.05) decreases the penalty as the number of fine-tuning iterations i increases. That is, if the generated sentence is classified as normative, a loss close to zero will be applied to each logit generated by the language model. If the generated sentence is non-normative, a higher total sentence loss will be applied to each logit. \u03b2 decreases the step size during back-propagation to avoid over-shooting the local minima. loss w acts as a cycle loss component, punishing the sentences with undesirable characteristics. The fine-tuning process is as follows: given a set of input sentences from a corpus, GPT-2 is used to generate successor sentences. We generate 60 tokens and truncate at the first punctuation mark (e.g. periods, question marks). These continuation sentences are fed through a classifier, which outputs the binary label we treat as a reward C(s) \u2208 {0, 1}. Sentences labeled as 0 are those with undesirable characteristic. As per Equation 3, the reward is used to calculate the punishment score by subtracting from 1.0 and scaling by the average word loss of the sentence. This value is then used to compute a sentence loss as in Equation (2). The sentence loss is averaged to obtain the token-level loss, which is then added to each logit from the continuation sentence and the loss is back-propagated into GPT-2. The process is illustrated in Figure 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 1546,
"end": 1554,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Normative Fine-Tuning",
"sec_num": "3"
},
{
"text": "To prevent the model from deviating too much from the language in the original dataset, we feed the fine-tuned model with the same set of input sentences at every loop and use the output sentences to even further fine-tune the model. As the model is trained, the output sentences will differ, and the reward value C(s) may change after every iteration as the model shifts its distribution.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Normative Fine-Tuning",
"sec_num": "3"
},
{
"text": "We conduct three sets of experiments to (1) verify the normative text classifier on generated continuations, (2) evaluate our reward-based fine-tuning with the normative text classifier, (3) evaluate our reward-based fine-tuning on other classifiers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "The normative text classifier by Frazier et al. (2019) was evaluated on original sentences from a number of corpora, including the science fiction story corpus (sci-fi) we use in subsequent evaluation experiments. Generated text potentially constitutes a shift in the text distribution. Hence, the accuracy of classifiers on generated continuations must be validated. The 117M parameter GPT-2 is fine-tuned with training sets from Plotto, ROCstories, sci-fi, Toxic and Sentiment datasets, separately, in order to shift the output probability distribution of GPT-2 and to make it generate text similar to the corpus we used for training. The sci-fi and Plotto corpora were used for fine-tuning two different versions of the normative text classifier, starting with the classifier by Frazier et al. (2019) . Thus the original classifier originally trained on G&G was updated to the respective domains; a few-shot transfer paradigm. We fine-tuned the classifiers for 2-5 iterations.",
"cite_spans": [
{
"start": 33,
"end": 54,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
},
{
"start": 782,
"end": 803,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 1: Replication of the Normative Classifier",
"sec_num": "4.1"
},
{
"text": "For the ROCstories, Toxic and Sentiment datasets, we directly train a BERT-based classifier on the given labels instead of fine-tuning the classifier that was first trained on G&G. We found the G&G-trained classifier did not transfer well to ROCstories and thus collected our own normative and non-normative labels. Toxic and Sentiment experiments do not look at normativity so we did not use the normative classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 1: Replication of the Normative Classifier",
"sec_num": "4.1"
},
{
"text": "A human participant study was then conducted to validate the fine-tuned classifier's accuracy. Sentences from each corpus test set were randomly chosen and used as prompts for GPT-2 to generate continuation sentences. 70 crowd workers on Mechanical Turk labeled those generated sentences as normative or non-normative (or positive or negative sentiment, or toxic or non-toxic). Each sentence received at least three labels. We treat the majority label from humans participants as the ground-truth. Table 1 shows the accuracy of classifiers on generated continuations and on sentences directly from the test sets. Accuracy decreases on generated continuations, but are on par with accuracy on sentences taken directly from the test corpora, and on par with the results from Frazier et al. (2019) . This indicates that any distributional shift during generation is likely inconsequential and the classifer achieves good zero-shot transfer to more datasets.",
"cite_spans": [
{
"start": 773,
"end": 794,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 498,
"end": 505,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Experiment 1: Replication of the Normative Classifier",
"sec_num": "4.1"
},
{
"text": "Non-Normative Generation",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "In this set of experiments, we seek to determine if, and by how much, the amount of non-normative behavior descriptions generated by GPT-2 decreases when fine-tuned with the normative text classifier. We emphasis the decrease of non-normative language because both normative and neutral languages are acceptable. Consistent with Experiment 1, we first fine-tune the 117M parameter version of GPT-2 with sentences sampled from three datasets: ROCstories, sci-fi, and Plotto. This gives us three versions of GPT-2: GPT-ROCstories, GPT-scifi and GPTplotto, respectively. The 117M GPT-2 model is fine-tuned for two, three and five iterations separately on ROCstories, Plotto and sci-fi to avoid overfitting. We then fine-tune each of these models a second time using the reinforcement learning, reward-based technique in Section 3. We refer to these models as GPT-ROCstories-norm, GPT-sci-finorm, and GPT-Plotto-norm, respectively. Due to the small size of the datasets, GPT-2 easily overfits during training. Therefore, we only fine-tune one of its 12 attention heads to avoid overfitting.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "We evaluate the performance of our fine-tuned GPT-X-norm models (X=plotto, scifi, ROCstories) by analyzing the change in the proportion of generated text that is non-normative. We measure the ratio of non-normative to normative text in two ways. First, we use the normative text classifier on continuations generated by baselines and fine-tuned models. This is an automated evaluation; Experiment 1 suggests the normative text classifier has high accuracy on the continuations. However, the gold standard is the human participant labels. For our second evaluation metric, we hired 70 Mechanical Turk workers to label generated continuation sentences as normative (including neutral) or nonnormative. At least 3 crowd workers labeled each sentence and the majority vote is considered as the ground truth label. Table 2 shows the proportions of non-normative sentence continuations for both the baseline and the fine-tuned models. The results are summarized below. We note percentage decreases, which are the relative percentage decrease compared to the original statistics. 3",
"cite_spans": [],
"ref_spans": [
{
"start": 810,
"end": 817,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "\u2022 GPT-ROCstories generates non-normative continuations 64% of the time according to the normative text classifier, which reduces to 26% after further fine-tuning, a 59% decrease. Humans label GPT-ROCstories continuations as non-normative 58% of the time, which drops to 22%, a 61% decrease.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "\u2022 GPT-Plotto generates non-normative continuations 81% of the time according to the normative text classifier, which reduces to 59% after further fine-tuning, a 27% decrease. Humans label GPT-Plotto continuations as nonnormative 72% of the time, which drops to Label Sentence Non-norm. Mollari now refuses to pay the two parents' expenses and lives. Non-norm. Garibaldi slaps the door behind them and locks it behind them. Non-norm. He considers himself morally superior to his family because he is wealthy. Norm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "Nathaniel repays his debt through an honest act of honest enterprise. Norm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "He then makes a generous and appropriate sacrifice. Norm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "He returns home to support his country. 53%, also a 27% decrease.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "\u2022 GPT-scifi generates non-normative continuations 35% of the time according to the normative text classifier, which reduces to 15% after further fine-tuning, a 55% decrease. Humans label GPT-scifi continuations as nonnormative 26% of the time, which drops to 18%, a 31% decrease.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "We observe that the classifier results are generally in line with the human evaluation results. The models fine-tuned on the Plotto dataset generally generate more non-normative continuation sentences and are more difficult to induce normativity. The models fine-tuned on the sci-fi dataset have the lowest frequency of non-normative generations, but can still be induced to produce lower frequencies with this method. The perplexity remains steady after fine-tuning with the normative text classifier, indicating that the GPT-X-norm models are not overfitting nor losing their fluency. Table 3 shows some examples of sentences generated by the fine-tuned GPT-2 models for the sci-fi and Plotto domains.",
"cite_spans": [],
"ref_spans": [
{
"start": 587,
"end": 594,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Experiment 2: Decreasing",
"sec_num": "4.2"
},
{
"text": "In the previous sections we evaluate how the normative text classifier and our fine-tuning technique work together to decrease the generation of nonnormative text. In this section, we ablate our system and evaluate our fine-tuning method independently of the normative text classifier. We seek to determine whether the fine-tuning technique is general enough to work with other classifiers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 3: Other Classifiers",
"sec_num": "4.3"
},
{
"text": "We replicate the experimental methodolgy in Section 4.2 but with the Toxic Comment Classification dataset and the Sentiment dataset. Sentiment is often used as a surrogate for normativity because non-normative behavior might be inferred to be perceived with negative sentiment, and because labeled sentiment data is more readily available. Toxic language is a subclass of non-normative behavior. First, we train two classifiers using the same technique as Frazier et al. (2019) . Specifically, we fine-tune BERT on the datasets with ground-truth sentiment and toxicity labels. Both classifiers are fine-tuned 5 times on training set of corpora and tested on test sets (see Table 1 ).",
"cite_spans": [
{
"start": 456,
"end": 477,
"text": "Frazier et al. (2019)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 673,
"end": 680,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Experiment 3: Other Classifiers",
"sec_num": "4.3"
},
{
"text": "We follow the methodology in Section 4.2 to produce new GPT-2 baseline models. To make the improvement from applying our technique more visible, we fine-tuned the 117M GPT-2 with only toxic or negative sentences when producing the baseline models, and obtained GPT-senti-ext and GPT-toxic-ext, which frequently produce negative or toxic generated continuations. We also finetuned the 117M GPT-2 with balanced datasets (half negative texts and half toxic texts, respectively), and refer to these two models as GPT-senti-bal and GPT-toxic-bal. We then further fine-tune these models using their respective classifiers per the technique in Section 4.2. Table 4 shows the percentage of textual continuations that are either toxic or contain negative sentiment. Fine-tuning GPT-senti-ext with the sentiment classifier can reduce negative sentiment from 76% to 42%, a 45% reduction. Fine-tuning GPTsenti-bal with the sentiment classifier can reduce negative sentiment from 36% to 29%, a 17% reduction. Fine-tuning GPT-toxic-ext with the toxic classifier can reduce toxic language from 57% to 30%, a 46% reduction. Fine-tuning GPT-toxic-bal with the toxic classifier can reduce toxic language from 33% to 28%, a 13% reduction. This shows that the fine-tuning technique working on sentence loss is agnostic to which classifier is used.",
"cite_spans": [],
"ref_spans": [
{
"start": 650,
"end": 657,
"text": "Table 4",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Experiment 3: Other Classifiers",
"sec_num": "4.3"
},
{
"text": "We did not compare our results directly to the Plug and Play Langauge Models (PPLM) work (Dathathri et al., 2019) , which also uses a toxic word classifier to fine-tune a language model. PPLM fine-tunes on a word-by-word basis instead of at the sentence unit, making it difficult to account for the context needed for determining normativity. For toxic language reduction, their model is trained to operate on a pre-given set of prompts such as \"black\" or \"asian\". For these prompts, given during training, they can reduce GPT-2's toxic language frequency from \u223c10% to \u223c6%. This is non-significant (p < 0.23) though it is challenging to reduce a number that is already close to zero. When we prompt our GPT-toxic-bal and GPT-toxicbal-norm with \"asian\" and \"black\", we see reductions from 52% to 32% and from 70% to 50%, respectively. Our classifier has only ever seen the word \"black\" once and has never seen the word \"asian\". We see higher occurrences of toxicity because GPT-2 is fine-tuned on the equal numbers of toxic and non-toxic sentences from the Toxic dataset (whereas PPLM compares against a nonfine-tuned version of GPT-2) and because GPT-2 has a pre-existing unjust bias toward these words.",
"cite_spans": [
{
"start": 89,
"end": 113,
"text": "(Dathathri et al., 2019)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 3: Other Classifiers",
"sec_num": "4.3"
},
{
"text": "To address the relationship between sentiment and normativity, we sample 300 sentences from the sci-fi corpus and 300 sentences from the generation results of the trained GPT-2 model and classify them using the normative text classifier and Sen-tiWordNet (Esuli and Sebastiani, 2006) . Figure 2 shows the percentage of sentences were classified as (a) both normative and positive/neutral sentiment (orange), (b) both non-normative and negative sentiment (blue), (c) normative but negative sentiment (green), and (d) non-normative but positive/neutral sentiment (brown). Only about half the sentences tested (53.08%) had sentiment and normativity labels that matched, whereas 46.92% of sentences have conflicting labels.",
"cite_spans": [
{
"start": 255,
"end": 283,
"text": "(Esuli and Sebastiani, 2006)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [
{
"start": 286,
"end": 294,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Experiment 3: Other Classifiers",
"sec_num": "4.3"
},
{
"text": "We demonstrate how value-aligned priors such as normative text classifiers can act as a reward provider to nudge the GPT-2 language model towards producing more normative and neutral descriptions of behaviors and events. Applying this reward-based fine-tuning technique reduces the likelihood of generating sentences containing non-normative behavior by approximately \u223c27-61%, depending on the dataset. Some datasets are more non-normative and more resistant to reduction of non-normativity. Datasets with low nonnormativity to begin with are, naturally, more resistant.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "Beyond the numerical results, this shows evidence that policy-gradient based reinforcement learning approaches to fine-tuning can be an effective means for reducing the generation of nonnormative descriptions. By using a normative text classifier-a value aligned prior-one can finetune a language model to the desired domain and then fine-tune using the prior again to reduce nonnormative generation that stems either from GPT-2 or from the domain corpus. We provide this approach as an alternative-or in complement todebiasing techniques that attempt to correct for prejudicial bias in datasets prior to training. Our approach is roughly equivalent to teaching a language model to censor itself.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "The policy-gradient based reinforcement learning technique using a value-aligned prior can be even more valuable with GPT-3, where fine-tuning to a domain is less necessary. GPT-3 has been demonstrated to be capable of non-normative, toxic, and prejudicially biased generation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "By using the normative text classifier trained on Goofus & Gallant comics, our results are limited to Western-and in particular Americanmainstream ideals of normative behavior. We acknowledge that culture is not monolithic, even within the United States of America, and this represents only one of many possible sources of normativity. We cannot conclusively say that our results will hold if we had normative text classifiers trained from different source materials. Because general sources of normative behavior are currently hard to come by, we attempt to show generalization of our technique with experiments using sentiment and toxic language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "One limitation of our work is that the normative text classifier can only classify individual sentences without context. Given the context-dependent nature of normativity, the normative text classifier may overlook non-normative sentences that may appear normative out-of-context. This may lead to GPT-2 still producing this sentence in its nonnormative context. Another limitation is that finetuning GPT-2 on Plotto, ROCstories and sci-fi datasets leads to it generating both neutral and normative sentences. If a model that generates solely normative sentences is desired, one can substitute the normative classifier with a ternary classifier with labels for normative, non-normative, and neutral sentences, and adjust the reward signals accordingly. Furthermore, fine-tuning classifiers requires datasets with labeled exemplars, hence, in order to replicate our work to generate texts with some other desirable characteristics, datasets with labels would be prerequisites. The motivation of our work is to show how those who are concerned with generating undesirable text language models can obtain some control over the generation process. We look at normativity, but also show how other criteria can be applied. However, as is true for most algorithms, those with malicious intent can find ways to corrupt the intentions of the work. For example, equation 3can be trivially modified to punish normative text instead of non-normative text.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "We have shown that large-scale transformer-based neural language models can be made to generate text containing fewer descriptions of nonnormative behavior by applying data-efficient, policy-gradient reinforcement learning. As most large-scale language models are trained on datasets from the internet and from books, the potential for intentional or unintentional non-normative language persists. We see this as a first step toward decreasing the potential for unintended, unacceptable, anachronistic or harmful language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "While our primary result is to show that we can decrease the generation of non-normative behavior descriptions, our normative classifier of choice is rooted in Western/American norms and values. Normative classifiers are rare and datasets containing normative or preference learning examples are difficult to obtain. However, our results show that even small datasets of normative examples can be converted into few-shot classifiers and applied to new domains. By replicating our results with sentiment and toxic classifier, we show that our technique is not specific to any one classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "6"
},
{
"text": "https://www.newsweek.com/amazon-echo% 2Dtells-uk-woman-stab-herself-1479074/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Percentage decrease is calculated by (p \u2212p)/p, where p andp are the proportion of non-normative behavior (% Nonnorm) generated by GPT-X and GPT-X-norm, respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Reinforcement learning as a framework for ethical decision making",
"authors": [
{
"first": "David",
"middle": [],
"last": "Abel",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Macglashan",
"suffix": ""
},
{
"first": "Michael",
"middle": [
"L"
],
"last": "Littman",
"suffix": ""
}
],
"year": 2016,
"venue": "Workshops at the Thirtieth AAAI Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Abel, James MacGlashan, and Michael L Littman. 2016. Reinforcement learning as a frame- work for ethical decision making. In Workshops at the Thirtieth AAAI Conference on Artificial Intelli- gence.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Story realization: Expanding plot events into sentences",
"authors": [
{
"first": "Prithviraj",
"middle": [],
"last": "Ammanabrolu",
"suffix": ""
},
{
"first": "Ethan",
"middle": [],
"last": "Tien",
"suffix": ""
},
{
"first": "Wesley",
"middle": [],
"last": "Cheung",
"suffix": ""
},
{
"first": "Zhaochen",
"middle": [],
"last": "Luo",
"suffix": ""
},
{
"first": "William",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Lara",
"suffix": ""
},
{
"first": "Mark",
"middle": [
"O"
],
"last": "Martin",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Riedl",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1909.03480"
]
},
"num": null,
"urls": [],
"raw_text": "Prithviraj Ammanabrolu, Ethan Tien, Wesley Che- ung, Zhaochen Luo, William Ma, Lara J Martin, and Mark O Riedl. 2019. Story realization: Ex- panding plot events into sentences. arXiv preprint arXiv:1909.03480.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Value alignment or misalignmentwhat will keep systems accountable?",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "Arnold",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Kasenberg",
"suffix": ""
},
{
"first": "Matthias",
"middle": [],
"last": "Scheutz",
"suffix": ""
}
],
"year": 2017,
"venue": "Workshops at the Thirty-First AAAI Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Thomas Arnold, Daniel Kasenberg, and Matthias Scheutz. 2017. Value alignment or misalignment- what will keep systems accountable? In Workshops at the Thirty-First AAAI Conference on Artificial In- telligence.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "The grammar of society: The nature and dynamics of social norms",
"authors": [
{
"first": "Cristina",
"middle": [],
"last": "Bicchieri",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cristina Bicchieri. 2005. The grammar of society: The nature and dynamics of social norms. Cambridge University Press.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Deep reinforcement learning from human preferences",
"authors": [
{
"first": "Jan",
"middle": [],
"last": "Paul F Christiano",
"suffix": ""
},
{
"first": "Tom",
"middle": [],
"last": "Leike",
"suffix": ""
},
{
"first": "Miljan",
"middle": [],
"last": "Brown",
"suffix": ""
},
{
"first": "Shane",
"middle": [],
"last": "Martic",
"suffix": ""
},
{
"first": "Dario",
"middle": [],
"last": "Legg",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Amodei",
"suffix": ""
}
],
"year": 2017,
"venue": "Advances in Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "4299--4307",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Paul F Christiano, Jan Leike, Tom Brown, Miljan Mar- tic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems, pages 4299-4307.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Plug and play language models: a simple approach to controlled text generation",
"authors": [
{
"first": "Sumanth",
"middle": [],
"last": "Dathathri",
"suffix": ""
},
{
"first": "Andrea",
"middle": [],
"last": "Madotto",
"suffix": ""
},
{
"first": "Janice",
"middle": [],
"last": "Lan",
"suffix": ""
},
{
"first": "Jane",
"middle": [],
"last": "Hung",
"suffix": ""
},
{
"first": "Eric",
"middle": [],
"last": "Frank",
"suffix": ""
},
{
"first": "Piero",
"middle": [],
"last": "Molino",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Yosinski",
"suffix": ""
},
{
"first": "Rosanne",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1912.02164"
]
},
"num": null,
"urls": [],
"raw_text": "Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2019. Plug and play language mod- els: a simple approach to controlled text generation. arXiv preprint arXiv:1912.02164.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "BERT: pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language under- standing. CoRR, abs/1810.04805.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Sentiwordnet: A publicly available lexical resource for opinion mining",
"authors": [
{
"first": "Andrea",
"middle": [],
"last": "Esuli",
"suffix": ""
},
{
"first": "Fabrizio",
"middle": [],
"last": "Sebastiani",
"suffix": ""
}
],
"year": 2006,
"venue": "LREC",
"volume": "6",
"issue": "",
"pages": "417--422",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andrea Esuli and Fabrizio Sebastiani. 2006. Senti- wordnet: A publicly available lexical resource for opinion mining. In LREC, volume 6, pages 417-422. Citeseer.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Learning norms from stories: A prior for value aligned agents",
"authors": [
{
"first": "Spencer",
"middle": [],
"last": "Frazier",
"suffix": ""
},
{
"first": "Md Sultan Al",
"middle": [],
"last": "Nahian",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Riedl",
"suffix": ""
},
{
"first": "Brent",
"middle": [],
"last": "Harrison",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1912.03553"
]
},
"num": null,
"urls": [],
"raw_text": "Spencer Frazier, Md Sultan Al Nahian, Mark Riedl, and Brent Harrison. 2019. Learning norms from sto- ries: A prior for value aligned agents. arXiv preprint arXiv:1912.03553.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The culture of national security: Norms and identity in world politics",
"authors": [
{
"first": "Mary",
"middle": [],
"last": "Fainsod Katzenstein",
"suffix": ""
}
],
"year": 1996,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mary Fainsod Katzenstein. 1996. The culture of na- tional security: Norms and identity in world politics. Columbia University Press.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Ctrl: A conditional transformer language model for controllable generation",
"authors": [
{
"first": "Bryan",
"middle": [],
"last": "Nitish Shirish Keskar",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Mccann",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Lav",
"suffix": ""
},
{
"first": "Caiming",
"middle": [],
"last": "Varshney",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Xiong",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Socher",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1909.05858"
]
},
"num": null,
"urls": [],
"raw_text": "Nitish Shirish Keskar, Bryan McCann, Lav R Varshney, Caiming Xiong, and Richard Socher. 2019. Ctrl: A conditional transformer language model for control- lable generation. arXiv preprint arXiv:1909.05858.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "From group to individual labels using deep features",
"authors": [
{
"first": "Dimitrios",
"middle": [],
"last": "Kotzias",
"suffix": ""
},
{
"first": "Misha",
"middle": [],
"last": "Denil",
"suffix": ""
},
{
"first": "Padhraic",
"middle": [],
"last": "Nando De Freitas",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Smyth",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining",
"volume": "",
"issue": "",
"pages": "597--606",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dimitrios Kotzias, Misha Denil, Nando De Freitas, and Padhraic Smyth. 2015. From group to individual la- bels using deep features. In Proceedings of the 21th ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining, pages 597-606.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "A corpus and cloze evaluation for deeper understanding of commonsense stories",
"authors": [
{
"first": "Nasrin",
"middle": [],
"last": "Mostafazadeh",
"suffix": ""
},
{
"first": "Nathanael",
"middle": [],
"last": "Chambers",
"suffix": ""
},
{
"first": "Xiaodong",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Devi",
"middle": [],
"last": "Parikh",
"suffix": ""
},
{
"first": "Dhruv",
"middle": [],
"last": "Batra",
"suffix": ""
},
{
"first": "Lucy",
"middle": [],
"last": "Vanderwende",
"suffix": ""
},
{
"first": "Pushmeet",
"middle": [],
"last": "Kohli",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Allen",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "839--849",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A cor- pus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 839-849.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Stereoset: Measuring stereotypical bias in pretrained language models",
"authors": [
{
"first": "Moin",
"middle": [],
"last": "Nadeem",
"suffix": ""
},
{
"first": "Anna",
"middle": [],
"last": "Bethke",
"suffix": ""
},
{
"first": "Siva",
"middle": [],
"last": "Reddy",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2004.09456"
]
},
"num": null,
"urls": [],
"raw_text": "Moin Nadeem, Anna Bethke, and Siva Reddy. 2020. Stereoset: Measuring stereotypical bias in pretrained language models. arXiv preprint arXiv:2004.09456.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Deep contextualized word representations",
"authors": [
{
"first": "Matthew",
"middle": [
"E"
],
"last": "Peters",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Neumann",
"suffix": ""
},
{
"first": "Mohit",
"middle": [],
"last": "Iyyer",
"suffix": ""
},
{
"first": "Matt",
"middle": [],
"last": "Gardner",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Clark",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2018,
"venue": "Proc. of NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word repre- sentations. In Proc. of NAACL.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Language models are unsupervised multitask learners",
"authors": [
{
"first": "Alec",
"middle": [],
"last": "Radford",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Rewon",
"middle": [],
"last": "Child",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Luan",
"suffix": ""
},
{
"first": "Dario",
"middle": [],
"last": "Amodei",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
}
],
"year": 2019,
"venue": "OpenAI Blog",
"volume": "",
"issue": "8",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1(8).",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"authors": [
{
"first": "Colin",
"middle": [],
"last": "Raffel",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Adam",
"middle": [],
"last": "Roberts",
"suffix": ""
},
{
"first": "Katherine",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Sharan",
"middle": [],
"last": "Narang",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Matena",
"suffix": ""
},
{
"first": "Yanqi",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Peter J",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1910.10683"
]
},
"num": null,
"urls": [],
"raw_text": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text trans- former. arXiv preprint arXiv:1910.10683.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Research priorities for robust and beneficial artificial intelligence",
"authors": [
{
"first": "Stuart",
"middle": [],
"last": "Russell",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Dewey",
"suffix": ""
},
{
"first": "Max",
"middle": [],
"last": "Tegmark",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "36",
"issue": "",
"pages": "105--114",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Stuart Russell, Daniel Dewey, and Max Tegmark. 2015. Research priorities for robust and beneficial artificial intelligence. Ai Magazine, 36(4):105-114.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "The woman worked as a babysitter: On biases in language generation",
"authors": [
{
"first": "Emily",
"middle": [],
"last": "Sheng",
"suffix": ""
},
{
"first": "Kai-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Premkumar",
"middle": [],
"last": "Natarajan",
"suffix": ""
},
{
"first": "Nanyun",
"middle": [],
"last": "Peng",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1909.01326"
]
},
"num": null,
"urls": [],
"raw_text": "Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2019. The woman worked as a babysitter: On biases in language generation. arXiv preprint arXiv:1909.01326.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Aligning superintelligence with human interests: A technical research agenda",
"authors": [
{
"first": "Nate",
"middle": [],
"last": "Soares",
"suffix": ""
},
{
"first": "Benja",
"middle": [],
"last": "Fallenstein",
"suffix": ""
}
],
"year": 2014,
"venue": "Machine Intelligence Research Institute (MIRI) technical report",
"volume": "8",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nate Soares and Benja Fallenstein. 2014. Aligning su- perintelligence with human interests: A technical re- search agenda. Machine Intelligence Research Insti- tute (MIRI) technical report, 8.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Release strategies and the social impacts of language models",
"authors": [
{
"first": "Irene",
"middle": [],
"last": "Solaiman",
"suffix": ""
},
{
"first": "Miles",
"middle": [],
"last": "Brundage",
"suffix": ""
},
{
"first": "Jack",
"middle": [],
"last": "Clark",
"suffix": ""
},
{
"first": "Amanda",
"middle": [],
"last": "Askell",
"suffix": ""
},
{
"first": "Ariel",
"middle": [],
"last": "Herbert-Voss",
"suffix": ""
},
{
"first": "Jeff",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Alec",
"middle": [],
"last": "Radford",
"suffix": ""
},
{
"first": "Jasmine",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1908.09203"
]
},
"num": null,
"urls": [],
"raw_text": "Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, and Jasmine Wang. 2019. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Normative ethics and metaethics",
"authors": [
{
"first": "Leonard Wayne",
"middle": [],
"last": "Sumner",
"suffix": ""
}
],
"year": 1967,
"venue": "Ethics",
"volume": "77",
"issue": "2",
"pages": "95--106",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Leonard Wayne Sumner. 1967. Normative ethics and metaethics. Ethics, 77(2):95-106.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Universal adversarial triggers for attacking and analyzing nlp",
"authors": [
{
"first": "Eric",
"middle": [],
"last": "Wallace",
"suffix": ""
},
{
"first": "Shi",
"middle": [],
"last": "Feng",
"suffix": ""
},
{
"first": "Nikhil",
"middle": [],
"last": "Kandpal",
"suffix": ""
},
{
"first": "Matt",
"middle": [],
"last": "Gardner",
"suffix": ""
},
{
"first": "Sameer",
"middle": [],
"last": "Singh",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1908.07125"
]
},
"num": null,
"urls": [],
"raw_text": "Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial trig- gers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Xlnet: Generalized autoregressive pretraining for language understanding",
"authors": [
{
"first": "Zhilin",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Zihang",
"middle": [],
"last": "Dai",
"suffix": ""
},
{
"first": "Yiming",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Jaime",
"middle": [],
"last": "Carbonell",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Russ",
"suffix": ""
},
{
"first": "Quoc V",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Le",
"suffix": ""
}
],
"year": 2019,
"venue": "Advances in neural information processing systems",
"volume": "",
"issue": "",
"pages": "5753--5763",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Car- bonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural in- formation processing systems, pages 5753-5763.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Defending against neural fake news. CoRR, abs",
"authors": [
{
"first": "Rowan",
"middle": [],
"last": "Zellers",
"suffix": ""
},
{
"first": "Ari",
"middle": [],
"last": "Holtzman",
"suffix": ""
},
{
"first": "Hannah",
"middle": [],
"last": "Rashkin",
"suffix": ""
},
{
"first": "Yonatan",
"middle": [],
"last": "Bisk",
"suffix": ""
},
{
"first": "Ali",
"middle": [],
"last": "Farhadi",
"suffix": ""
},
{
"first": "Franziska",
"middle": [],
"last": "Roesner",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 1905,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019. Defending against neural fake news. CoRR, abs/1905.12616.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Fine-tuning language models from human preferences",
"authors": [
{
"first": "M",
"middle": [],
"last": "Daniel",
"suffix": ""
},
{
"first": "Nisan",
"middle": [],
"last": "Ziegler",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Stiennon",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Tom",
"suffix": ""
},
{
"first": "Alec",
"middle": [],
"last": "Brown",
"suffix": ""
},
{
"first": "Dario",
"middle": [],
"last": "Radford",
"suffix": ""
},
{
"first": "Paul",
"middle": [],
"last": "Amodei",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [],
"last": "Christiano",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Irving",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1909.08593"
]
},
"num": null,
"urls": [],
"raw_text": "Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Chris- tiano, and Geoffrey Irving. 2019. Fine-tuning lan- guage models from human preferences. arXiv preprint arXiv:1909.08593.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "https://www.kaggle.com/c/jigsaw% 2Dtoxic-comment-classification-challenge/ Pipeline for fine-tuning GPT-2 with the classifier. Loss is backpropagated through the output logits to GPT-2."
},
"FIGREF1": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Differences between normative classification and sentiment classification."
},
"TABREF1": {
"text": "Results of Mechanical Turk study. Accuracy on generated continuations equals to the percentage of Mechanical Turk worker labels equivalent to labels produced by the normative classifier when classifying generated sentences, since Mechanical Turk worker labels are considered as ground truth label of generated continuations. Accuracy on original corpora is measured by the classifier on the held-out test sets of corpora sentences.",
"content": "<table/>",
"html": null,
"num": null,
"type_str": "table"
},
"TABREF3": {
"text": "",
"content": "<table><tr><td>: The proportion of non-normative behavior and</td></tr><tr><td>events (% Non-norm) generated by different fine-tuned</td></tr><tr><td>models on different datasets. Ratios are measured us-</td></tr><tr><td>ing the normative text classifier (automated) and Me-</td></tr><tr><td>chanical Turk studies (human labeling).</td></tr></table>",
"html": null,
"num": null,
"type_str": "table"
},
"TABREF4": {
"text": "Examples of generated normative and nonnormative sentences from GPT-Plotto and GPT-scifi.",
"content": "<table/>",
"html": null,
"num": null,
"type_str": "table"
},
"TABREF6": {
"text": "",
"content": "<table><tr><td>: The proportion of non-normative behavior and</td></tr><tr><td>events (% non-norm.) generated by different fine-tuned</td></tr><tr><td>models on different datasets.</td></tr></table>",
"html": null,
"num": null,
"type_str": "table"
}
}
}
} |