File size: 105,988 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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T14:05:52.021100Z"
},
"title": "Plug-and-Blend: A Framework for Controllable Story Generation with Blended Control Codes",
"authors": [
{
"first": "Zhiyu",
"middle": [],
"last": "Lin",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Georgia Institute of Technology North Ave NW",
"location": {
"postCode": "30332",
"settlement": "Atlanta",
"region": "GA"
}
},
"email": "zhiyulin@gatech.edu"
},
{
"first": "Mark",
"middle": [
"O"
],
"last": "Riedl",
"suffix": "",
"affiliation": {},
"email": "riedl@cc.gatech.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We describe a Plug-and-Play controllable language generation framework, Plug-and-Blend, that allows a human user to input multiple control codes (topics). In the context of automated story generation, this allows a human user loose or fine grained control of the topics that will appear in the generated story, and can even allow for overlapping, blended topics. We show that our framework, working with different generation models, controls the generation towards given continuous-weighted control codes while keeping the generated sentences fluent, demonstrating strong blending capability. Blending generative model Planner Decoder Generative Language Model Control Model John realized that basketballs fall to the ground like apples Apple Line Context Control codes 3 John was playing basketball 70% sports 30% science",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "We describe a Plug-and-Play controllable language generation framework, Plug-and-Blend, that allows a human user to input multiple control codes (topics). In the context of automated story generation, this allows a human user loose or fine grained control of the topics that will appear in the generated story, and can even allow for overlapping, blended topics. We show that our framework, working with different generation models, controls the generation towards given continuous-weighted control codes while keeping the generated sentences fluent, demonstrating strong blending capability. Blending generative model Planner Decoder Generative Language Model Control Model John realized that basketballs fall to the ground like apples Apple Line Context Control codes 3 John was playing basketball 70% sports 30% science",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Recent advancement in very large pre-trained neural language models (e.g. (Radford et al., 2019; Brown et al., 2020) ) have enabled a new generation of applications that make use of the text generation capability they provide, ranging from autocompletion of e-mails to solving complicated math equations. However these very large pre-trained neural language models are also difficult to control beyond providing a prompt for a generated continuation. This makes very large language models ill-suited for co-creative tasks wherein a human works with a language model in an iterative fashion to produce novel content, such as stories or poems. Co-creative tasks require an ability to not only prompt the language model but to guide the generation with, for example, style, context, or topic constraints.",
"cite_spans": [
{
"start": 74,
"end": 96,
"text": "(Radford et al., 2019;",
"ref_id": "BIBREF18"
},
{
"start": 97,
"end": 116,
"text": "Brown et al., 2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Conditional generation is a family of text generation methods that attempt to provide controllability by either directly modifying the model to accept control signals or posing constraints in the generation process. Conditional text generation techniques add an extra input feature (Ficler and Goldberg, 2017) and fine-tuning with additional information embedded (Fang et al., 2021; Hosseini-Asl et al., 2020; Keskar et al., 2019; Khalifa et al., 2020; Hu et al., 2017; Wu et al., 2020; Ficler and Goldberg, 2017; Chan et al., 2020) , or by sideloading additional discriminators along with a pre-trained model, without changing base model parameters holisticly (Dathathri et al., 2020; Madotto et al., 2020; Duan et al., 2020; Mai et al., 2020) .",
"cite_spans": [
{
"start": 282,
"end": 309,
"text": "(Ficler and Goldberg, 2017)",
"ref_id": "BIBREF6"
},
{
"start": 363,
"end": 382,
"text": "(Fang et al., 2021;",
"ref_id": "BIBREF5"
},
{
"start": 383,
"end": 409,
"text": "Hosseini-Asl et al., 2020;",
"ref_id": "BIBREF8"
},
{
"start": 410,
"end": 430,
"text": "Keskar et al., 2019;",
"ref_id": "BIBREF10"
},
{
"start": 431,
"end": 452,
"text": "Khalifa et al., 2020;",
"ref_id": null
},
{
"start": 453,
"end": 469,
"text": "Hu et al., 2017;",
"ref_id": "BIBREF9"
},
{
"start": 470,
"end": 486,
"text": "Wu et al., 2020;",
"ref_id": "BIBREF8"
},
{
"start": 487,
"end": 513,
"text": "Ficler and Goldberg, 2017;",
"ref_id": "BIBREF6"
},
{
"start": 514,
"end": 532,
"text": "Chan et al., 2020)",
"ref_id": "BIBREF2"
},
{
"start": 661,
"end": 685,
"text": "(Dathathri et al., 2020;",
"ref_id": "BIBREF3"
},
{
"start": 686,
"end": 707,
"text": "Madotto et al., 2020;",
"ref_id": "BIBREF3"
},
{
"start": 708,
"end": 726,
"text": "Duan et al., 2020;",
"ref_id": "BIBREF4"
},
{
"start": 727,
"end": 744,
"text": "Mai et al., 2020)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We seek \"plug-and-play\" approaches to controllable text generation wherein new language models can be slotted into existing generative systems; new language models are being developed and it becomes intractable to update and retrain controlled generation architectures. Plug-and-play techniques such as (Krause et al., 2020; Pascual et al., 2020) aim to only intervene with the outputs-a vector of logits-of a generative language model. This becomes especially important as the latest iteration of very large pre-trained language models such as GPT-3 (Brown et al., 2020) restrict access to the hidden states and layer weights of models. As language models improve, they can be easily incorporated into existing, controllable generation frameworks.",
"cite_spans": [
{
"start": 303,
"end": 324,
"text": "(Krause et al., 2020;",
"ref_id": null
},
{
"start": 325,
"end": 346,
"text": "Pascual et al., 2020)",
"ref_id": "BIBREF17"
},
{
"start": 551,
"end": 571,
"text": "(Brown et al., 2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We present Plug-and-Blend 1 , an efficient plugand-play generative framework for controllable text generation that (a) works with the logit outputs of any language model; (b) facilitates fine control of generated sentences by allowing continuous bias towards specific control codes; and (c) allows multiple control codes representing style and topic constraints to be provided in overlapping contexts. These control codes can be blended together to generate content that meets multiple style or topic constraints. We describe that these key capabilities empower latent space walking in the hyperspace of generated sentences, and show a simple content planning technique that utilizes this feature to generate paragraphs regarding user intentions in a co-authoring. We present our work in the context 1 Code available at https://github.com/ xxbidiao/plug-and-blend 10 sentence story Topic: sports, lines 1-5 Topic: science, lines 5-10 User Figure 1 : Illustration of overall architecture of our framework of automated story generation wherein a human author provides a prompt as well as a high-level control specification for topics.",
"cite_spans": [
{
"start": 800,
"end": 801,
"text": "1",
"ref_id": null
}
],
"ref_spans": [
{
"start": 939,
"end": 947,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Researchers aim for \"plug-and-play\" (PnP) frameworks (Dathathri et al., 2020) which can be used along an existing generative LM (referred to as the \"base LM\") with minimum or no interference between the PnP components and the base LM.",
"cite_spans": [
{
"start": 53,
"end": 77,
"text": "(Dathathri et al., 2020)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Plug-and-Play Conditional Generation",
"sec_num": "2.1"
},
{
"text": "Comparing to non-plug-and-play methods (\"white-box\" approaches), these frameworks can be roughly classified into three categories. Graybox approaches access and modify some non-inputoutput layer computations, usually the hidden representation, hence \"plugging\" an additional model in the middle of the base LM (Dathathri et al., 2020; Madotto et al., 2020; Duan et al., 2020; Mai et al., 2020) . Black-box approaches including \"Prompt Engineering\" that aim to change the prompts fed into the base LM at inference time (Wallace et al., 2019; Li and Liang, 2021) . Guided generation targets at building a controllable \"guiding\" model that shifts the output from base LM at inference time (Krause et al., 2020; Pascual et al., 2020) .",
"cite_spans": [
{
"start": 310,
"end": 334,
"text": "(Dathathri et al., 2020;",
"ref_id": "BIBREF3"
},
{
"start": 335,
"end": 356,
"text": "Madotto et al., 2020;",
"ref_id": "BIBREF3"
},
{
"start": 357,
"end": 375,
"text": "Duan et al., 2020;",
"ref_id": "BIBREF4"
},
{
"start": 376,
"end": 393,
"text": "Mai et al., 2020)",
"ref_id": "BIBREF15"
},
{
"start": 518,
"end": 540,
"text": "(Wallace et al., 2019;",
"ref_id": "BIBREF20"
},
{
"start": 541,
"end": 560,
"text": "Li and Liang, 2021)",
"ref_id": "BIBREF13"
},
{
"start": 686,
"end": 707,
"text": "(Krause et al., 2020;",
"ref_id": null
},
{
"start": 708,
"end": 729,
"text": "Pascual et al., 2020)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Plug-and-Play Conditional Generation",
"sec_num": "2.1"
},
{
"text": "The generation model we propose is an extension of GeDi (Krause et al., 2020) . Adding to the complete decoupling of generation and controlling, we enhanced it with additional capabilities to support multi-topic generation with continuous weighting, supporting the downstreaming applications while keeping its capability to transfer to different base LMs.",
"cite_spans": [
{
"start": 56,
"end": 77,
"text": "(Krause et al., 2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Plug-and-Play Conditional Generation",
"sec_num": "2.1"
},
{
"text": "Neural story generation systems train or fine-tune a language model on story data. Sampling from a language model trained on story data tends to result in text output that looks like stories as well. However, sampling from P \u03b8 (x t |x <t ) (See Section 3) is uncontrolled in the sense that one does not have any influence over the output after the initial context prompt.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Controllable Story Generation",
"sec_num": "2.2"
},
{
"text": "A number of story generation systems have attempted to condition the generation with some form of high-level plan. Storytelling systems such as (Akoury et al., 2020; Yao et al., 2019) embeds topic constraints directly into the model. These system extract a set of topics from a dataset that must be incorporated into the story. PlotMachines (Rashkin et al., 2020) allows a human user to specify topics that can be incorporated into a story in any order. generate a story by interpolating between a start event and an end event in a slot filling fashion, targeted the same goal. Our work differs in two ways. First, we allow blending of topics such that a single line in a story can meet more than one topic provided by a human user. Second, we have developed a black-box plug-and-play system that works with different LMs.",
"cite_spans": [
{
"start": 144,
"end": 165,
"text": "(Akoury et al., 2020;",
"ref_id": "BIBREF0"
},
{
"start": 166,
"end": 183,
"text": "Yao et al., 2019)",
"ref_id": "BIBREF25"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Controllable Story Generation",
"sec_num": "2.2"
},
{
"text": "Generative Language Models (LMs), specifically continuation models, take a context (\"prompt\") and generate a continuation by predicting the next tokens. This is achieved by optimizing the model parameters \u03b8 that best estimates the probability density of a sequence of word tokens x 1:T = {x 1 , . . . , x T } represented as an auto-regressive factorization",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P \u03b8 (x 1:T ) = T t=1 P \u03b8 (x t | x <t ) .",
"eq_num": "(1)"
}
],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "By iteratively predicting a distribution on the next token given the previous tokens, a continuation can be generated by repeatedly sampling",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "P \u03b8 (x t | x <t )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "and attach the selected token back to the \"previous\" tokens for the next step. Sequences generated this way are not controlled; To control the generated sequence, an attribute represented as a class variable (Keskar et al., 2019) that could describe sentiment or topics can be introduced to equation 1to form a Class-Conditional Language Model (CC-LM):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P \u03b8 (x 1:T | c) = T t=1 P \u03b8 (x t | x <t , c)",
"eq_num": "(2)"
}
],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "where c represents the class variable, or \"control code\", that describes an attribute of the sequence x 1:T . However, since c and x 1:T are entangled in equation 2, naively optimizing P \u03b8 requires a new CC-LM to be trained.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "To decouple the conditional generation component, c, from the unconditional part, P LM (x 1:T ), (Krause et al., 2020) proposed the GeDi framework and an algorithm to enable a separate controlling model to guide the generation process of a base language model. Instead of tackling P \u03b8 (x 1:T | c) directly, they train a contrastive discriminator model on the side to estimate",
"cite_spans": [
{
"start": 97,
"end": 118,
"text": "(Krause et al., 2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "P \u03b8 (c | x 1:t ) = \u03b1P (c) t j=1 P \u03b8 (x j | x <j , c) (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "where \u03b1 is the normalization constant \u03b1 = 1/( c \u2208{c,c}",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "t j=1 P (c ) P \u03b8 (x j | x <j , c ))",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": ", and c and c are contrastive control codes (c and not-c). At the decoding stage of the generation process, one can guide the generation by using P \u03b8 (c | x 1:t ) as a posterior to the output probability distribution of the base LM:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "P (x t | x <t , c) \u221d P LM (x t | x <t ) P \u03b8 (c | x t , x <t ) \u03c9 (4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "where \u03c9 is a parameter for control strength, with larger values biasing generation more strongly towards c. CC-LMs trained this way do not require access to any internal data of the base LM, and works independently of it.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preliminaries",
"sec_num": "3"
},
{
"text": "Our Plug-and-Blend framework consists of two components (See figure 1): (1) a blending generative Model that is responsible for plug-and-play controlled continuations using the control specifications; and (2) a planner that plans and assigns control specifications based on control sketches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Plug-and-Blend Framework",
"sec_num": "4"
},
{
"text": "A control sketch is a high-level specification of what topics should be present in the story and what portions of the story each topic should approximately appear in. This provides a human co-creator the ability to guide the generator loosely, with a broad range per topic, or tightly, with a narrow range per topic. We envision a co-creative loop wherein the human user provides a control sketch and iteratively updates the control sketch based on generation results, refining the topics and refining the ranges for the topics. The user interface for eliciting control sketches from a human is outside the scope of this paper and experiments about the co-creative loop are left for future work. The next sections provide the algorithmic support for control sketches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Plug-and-Blend Framework",
"sec_num": "4"
},
{
"text": "The blending generative model generates the sentence continuation. It consists of two parts, a (1) plug-and-play language model and (2) a control model. Given a prompt x <t , the plug-andplay language model produces a vector of logits P LM (x t | x <t ). The control model biases the output of the language model toward particular tokens associated with the topics of the control codes c \u2208 C based on the desired strengths of each topic \u03c9 * c\u2208C \u2208 \u2126. Together the two models iteratively find the best token x t that reflects both natural language composition and control bias presented by c and \u03c9. A larger \u03c9 * c means more steering towards the topic represented by control code c.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Blending Generative Model",
"sec_num": "4.1"
},
{
"text": "Inspired by the application of generative adversarial networks to latent space walking, we treat P \u03b8 (c | x t , x <t ) (described in section 3) as a heuristic of direction that increases P (x t | x <t , c) in a |V |-dimensional latent space, where V is the language model's vocabulary. For example, consider two different control codes c 1 and c 2 instantiating equation (4). To apply both control codes in the generation process, we use the heuristic",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Blending Generative Model",
"sec_num": "4.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (x t | x <t , c 1 , c 2 ) \u221d P LM (x t | x <t ) \u00d7 P \u03b8 (c 1 | x t , x <t ) \u03c9 1 P \u03b8 (c 2 | x t , x <t ) \u03c9 2",
"eq_num": "(5)"
}
],
"section": "Blending Generative Model",
"sec_num": "4.1"
},
{
"text": "to combine the effect of both posterior distributions into one universal posterior. \u03c9 1 and \u03c9 2 in this case represents control strength for each control code, c 1 and c 2 respectively, and can be different, enabling continuous blending between topics. This process can be repeated with a set of control codes C = {c 1 , . . . , c n } with weights \u2126 = {\u03c9 1 , . . . , \u03c9 n }.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Blending Generative Model",
"sec_num": "4.1"
},
{
"text": "Formally, at the decoding stage of the generation process, a control model compute controlled probability using the following equation:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Blending Generative Model",
"sec_num": "4.1"
},
{
"text": "P (x t | x <t , C) = P LM (x t | x <t ) c * \u2208C P \u03b8 (c * | x t , x <t ) \u03c9 * c (6)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Blending Generative Model",
"sec_num": "4.1"
},
{
"text": "where the control strengths of individual control codes are normalized with c \u03c9 * c = \u03c9, where \u03c9 is total control strength. 2 This can be efficiently computed by batching input sequences appended by different control codes, with little overhead compared to the original GeDi (Krause et al., 2020) framework.",
"cite_spans": [
{
"start": 275,
"end": 296,
"text": "(Krause et al., 2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Blending Generative Model",
"sec_num": "4.1"
},
{
"text": "The human user provides a high-level control sketch of the story, consisting of the number of sentences, N , a set of topics, C, and a range of lines to which to apply the topic, r := (s, e) where s \u2264 e. See figure 2 for example sketches. Sketches can have their range r overlap such that multiple topics can be applied to the same lines of the story.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Planner",
"sec_num": "4.2"
},
{
"text": "Given the control sketch, the planner produces a control configuration C n , \u2126 n for each sentence position n = {0, . . . , N \u2212 1}. The control configuration for each sentence is passed to the blending generative model along with previous generated sentences as prompt.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Planner",
"sec_num": "4.2"
},
{
"text": "We interpret a control sketch as story arc on a specific topic, which typically contains a transition, an engagement and a phase-out, the planner should give highest control strength to the midpoint of the area, m := (s + e)/2, and lower strength towards the start and end of the span of the area; We capture this as a Gaussian distribution.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Planner",
"sec_num": "4.2"
},
{
"text": "Formally, the following equation translates the sketch into a control configuration for each position n \u2208 N :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Planner",
"sec_num": "4.2"
},
{
"text": "\u03c9 + c,n = f (N (m, (\u03c3/(e \u2212 s + ) 2 ))(n \u2212 m) (7) where f (\u2022) indicates probability density function,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Planner",
"sec_num": "4.2"
},
{
"text": "is an infinitesimal, and \u03c3 is a tunable parameter representing overall transition smoothness, where higher \u03c3 grants smoother transitions in the cost of reduced topic engagement for midpoint. Since there can be multiple control sketches and they can be of the same control code, we apply each individual sketch in the order they are presented and normalize after each application so that \u03a3 n \u03c9 c,n = 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Planner",
"sec_num": "4.2"
},
{
"text": "For our experiments, we use the GPT2-large model fine-tuned on ROCStories (Mostafazadeh et al., 2016) as our base language model. Fine-tuning GPT2 on ROCStories results in a model that generates short stories about common everyday situations. We pair the language model with a pretrained GeDi (which in turn is based on GPT-2-medium) trained on AG-news 3 as the guiding model. Across all setups, at generation time, we use greedy decoding with repetition penalty described in Keskar et al. 2019, and only use the first sentence generated as the output, discarding every token after it if any.",
"cite_spans": [
{
"start": 74,
"end": 101,
"text": "(Mostafazadeh et al., 2016)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "Since there is no ground truth for any generated sequence, metrics such as BLEU and other n-grambased metrics are not applicable. This poses a unique challenge in evaluating our system, limiting us to unsupervised metrics. In this section, we report evaluation of our blending generative model from two aspects:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "\u2022 Fluency: measuring how our generated sequence forms natural language; and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "\u2022 Control fidelity: measuring how our generated sequence respects the requested control codes and strength.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "To evaluate fluency of sequences generated by our blending generation model, we use perplexity of base language model. The intuition is that if generated sentences have low average perplexity when evaluated by the base LM then they are consistent with sentences we would find in the English language, as represented by the data used to train the base LM. This in turn results in fluent-appearing sentences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Blending Fluency",
"sec_num": "5.1"
},
{
"text": "To generate sequences from our model, we used 100 sentences from a held-out evaluation set of ROCStories not seen at fine-tuning time. ROC-Stories contains five-sentence stories; we always pick the first sentence. That sentence becomes our prompt and is paired with all possible combinations of two topic choices chosen from \"Business\", \"Science\", \"Sports\", or \"World\". These are the topics that the GeDi model are optimized for. Our control sketch gives equal blending weighting for all topics. We vary the control strength using the following Figure 2 : Perplexity (lower is better) of generated sequences with 2 topics. Baseline performance set at 1x of (Krause et al., 2020)-suggested control strength. increments: [0, 0.5, 1, 1.5, 2, 3, 4]x, where 0 represents an uncontrolled base LM and 4x represents 400% of the control strength hyperparameter used by Krause et al. (2020) . Figure 2 shows the average perplexity of generated sequences, measured by the Base LM. We observe that average perplexity increases with stronger control, signaling a departure of generated sequences from what the base LM would generate, and a potential decrease in fluency. This is to be expected as the control is biasing the generated text more and more toward the use of words that are consistent with a particular topic and away from general word frequency. While perplexity increase is more or less linear in the range of 0 to 2x strength, once above 2x strength, it can be better described as exponential, hinting a stabler capability to generate fluent sentences in the region of 0 to 2x control strength.",
"cite_spans": [
{
"start": 860,
"end": 880,
"text": "Krause et al. (2020)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 545,
"end": 553,
"text": "Figure 2",
"ref_id": null
},
{
"start": 883,
"end": 891,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Blending Fluency",
"sec_num": "5.1"
},
{
"text": "Control fidelity is how well the generator responds to multiple control codes applied at once (see Krause et al. (2020) for experiments applying one control code at a time; we do not replicate them in this paper). For story generation, multiple control codes can be applied to the same sentence in a story at different weights. We perform experiments in a latent space walking setting, to measure content changes of generated sentences under the same prompt, same control codes but different relative control strength, in an unsupervised way.",
"cite_spans": [
{
"start": 99,
"end": 119,
"text": "Krause et al. (2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Control Fidelity",
"sec_num": "5.2"
},
{
"text": "Given a particular prompt line in a story and two control topics c 1 and c 2 , we re-generate the same line multiple times under different control strengths for each topic. Specifically we set \u03c9 c 1 to 0%, 25%, 50%, 75% or 100% and \u03c9 c 2 = 1 \u2212 \u03c9 c 1 to represent a range of different possible blends of topics in the same line. See table 1 for an example. Since we know the control parameters used to generate these sentences, in which c 1 receives more and more control strength relative to c 2 , we expect to see sentences that are increasingly about topic c 1 and decreasingly about topic c 2 . These sentences do not comprise a story sequence, but are different alternative sentences for the same line in a story under different topic control specifications.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Control Fidelity",
"sec_num": "5.2"
},
{
"text": "To determine whether a given generated sentence was representative of a topic, we score each generated sentence with an off-the-shelf BART-based zero-shot classifier (Wolf et al., 2020) 4 with c 1 and c 2 , in raw text form, as possible classes. We then compare the order of the sentences as determined by the classifier to the ground truth order of increasing control strength of c 1 . We report the correlation of order between these two sequences using Kendall's \u03c4 -a metric. A perfectly strictly increasing classifier score will grant a \u03c4 -a score of 1 for a sequence. If the sentences have some reordering based on classification score, \u03c4 -a is reduced. A score of 0 indicates a random ordering and and a score of \u22121 indicates a sequence that is exactly in opposite order. Table 1 shows the classifier scores for the possible next sentences under different control strengths; the classifier scores are not monotonically decreasing, resulting in a \u03c4 -a score of 0.8. Figure 3 shows a heat-map of the average \u03c4 -a score of sequences of sentences generated with different control code pairs and different total control strength (percentages). For each combination of parameters, 100 sequences of 5 sentences are generated and evaluated. Comparing to the baseline, which is the evaluation metric applied to orderrandomized stories in ROCStories dataset, we observe universal statistical significance (p < .01) in improvement in \u03c4 -a metric. That is, without a control bias, rank ordering is random. As we increase the total control strength, the rank order of generated sentences more closely matches the ground truth order.",
"cite_spans": [
{
"start": 166,
"end": 185,
"text": "(Wolf et al., 2020)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 778,
"end": 785,
"text": "Table 1",
"ref_id": null
},
{
"start": 971,
"end": 979,
"text": "Figure 3",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Control Fidelity",
"sec_num": "5.2"
},
{
"text": "Some topic combinations (For example, Science-Sports) work better than others (For example, Science-World); the \"World\" category appears to include a lot of overlapping vocabulary usage with Prompt: The people gathered to protest the court's ruling last week. c1 = Sports c2 = Business Generated Sentence Classifier score \u03c9c 1 \u03c9c 2 c1 c2 100% 0% Coach Leeman was in a wheelchair and had been taken to hospital for treatment. 86% 14%",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Control Fidelity",
"sec_num": "5.2"
},
{
"text": "Coach Reebok was one of them. 65% 35% 50% 50% The players were joined by a few of them. 84% 16%",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "75% 25%",
"sec_num": null
},
{
"text": "The company that owns the team was fined $1,000 for violating a rule prohibiting employees from using their own equipment. 37% 63%",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "25% 75%",
"sec_num": null
},
{
"text": "Bankruptcy Judge William H. said that the bank had failed to pay its creditors and was in default on $1 billion of loans it owed them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "0% 100%",
"sec_num": null
},
{
"text": "Comparing column 1 with column 4, Kendall's \u03c4 -a = 0.8 for this generated sequence. Table 1 : An example sequence of sentences generated for evaluation of control fidelity. The first two columns indicate the requested control strengths for two topics, sports and business. The generated sentence results from the prompt and the control weights (all numbers are 2x the default control strength). The last two columns indicate the probability that each line is either Sports or Business based on a BART-based topic classifier. We expect to see the classifier score for c 1 decrease as the classifier score for c 2 increases.",
"cite_spans": [],
"ref_spans": [
{
"start": 84,
"end": 91,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "24% 76%",
"sec_num": null
},
{
"text": "the other categories. Note that a perfect Kendall's \u03c4 -a of 1.0 is likely impossible because our zeroshot topic classifier will introduce some noise to the ranking. However, the results show us that the plug-and-blend technique (a) significantly increases the likelihood that topics will be incorporated into sentences, and (b) is sensitive to blended topics. Figure 4 shows the same experiment as above, but with a non-fine-tuned version of GPT2-large. This shows that the plug-and-blend technique works on language models that haven't been finetuned on ROCStories. The prompts are still selected from ROCStories, however, for comparison, but are not as representative of the untuned model. In this condition, the text generated will not read as sentences in stories. We observe similar improvements over the baseline, demonstrating the ability of our method in keeping the strong adaptation capability.",
"cite_spans": [],
"ref_spans": [
{
"start": 360,
"end": 368,
"text": "Figure 4",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "24% 76%",
"sec_num": null
},
{
"text": "In this section, we qualitatively demonstrate the capability of our pipeline by analyzing the generated paragraphs using simulated user inputs described as sets of control sketches. Table 2 (left column) shows three sets of control sketches with overlapping topic ranges. For example, sketch 1 requests a 10-line story that covers the topic of sports for the first 6 lines and covers the topic of science for the last 6 lines (topics overlap in the middle). For each control sketch we generate 10-line stories (N = 10) using the hyper-parameter \u03c3 = 1 (see Equation 7). We use a neutral prompt consisting of only the word \"Recently\" as the con-text to generate the first line or if the generator ever generates an empty line. The remainder of lines use up to 2 sentences generated for the previous context. Table 2 (right column) shows the generated stories for each control sketch. We bold the sentence where it is most clear that the topic has changed. Figure 5 shows how the heuristic transforms each control sketch into bias weights. The figure shows \u03c9 c 1 for c 1 = Sports showing how the planner decreases the probability density bias for the topic (the probability density for the second topic, \u03c9 c 2 , is the mirror image).",
"cite_spans": [],
"ref_spans": [
{
"start": 182,
"end": 189,
"text": "Table 2",
"ref_id": null
},
{
"start": 806,
"end": 828,
"text": "Table 2 (right column)",
"ref_id": null
},
{
"start": 954,
"end": 962,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Planner Experiments",
"sec_num": "5.3"
},
{
"text": "With slight differences in the input control sketches, we observe very different generated stories, with the transition between sports and science happening later. One can see from Figure 5 why this would be the case: the probability density for the first topic becomes increasingly stronger for the first lines of the story as the control sketch requests the second topic later.",
"cite_spans": [],
"ref_spans": [
{
"start": 181,
"end": 189,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Planner Experiments",
"sec_num": "5.3"
},
{
"text": "Because each sentence is biased by the previous sentences in addition to the control sketch, the sentence where the topic appears to switch often comes later than the point of earliest topic overlap. The requirement that each sentence continue the previous context creates a sense of momentum from the previous context and thus from the previous topic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Planner Experiments",
"sec_num": "5.3"
},
{
"text": "Incoherent transitions may still happen. In the story in Table 2 for sketch 3 shows one such incoherent transition due to the generation of an endof-text token. Our implementation uses the initial prompt in this case, causing a portion of the story to not be contextualized by the earlier story sentences. Our ROCStories-tuned language model, based on 5-sentence stories, tends to predict end-of-text earlier than models trained on longer stories.",
"cite_spans": [],
"ref_spans": [
{
"start": 57,
"end": 64,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Planner Experiments",
"sec_num": "5.3"
},
{
"text": "Our experiments suggest that there is a trade-off between control fidelity and fluency. As Figures 2 and 3 show, a higher total control strength results in overall better \u03c4 -a scores, meaning more sensitivity and ability to correctly differentiate between topic blends, but worse perplexity, risking less fluent language. In practice, an iterative deepening algorithm where multiple control strengths are used to generate multiple candidate sentences per line, can be used. Control strength modifiers of 1x, 2x, 3x, 4x, etc. can be tried and the best generated sentence, as measured by perplexity (or any other task-specific metric), is selected. This can, just like how multiple control codes are handled, be implemented very efficiently. The current planner is heuristic. Empirically we find the heuristic to create good blends. We envision a planner that can be parameterized and learn from demonstrations. Reinforcement learning, in which the context and control sketches work as world states, can choose control configurations as actions. Feedback (reward) from the user would be necessary. This would incorporate the plug-andblend technique into a human-in-the-loop creative Figure 5 : Control strength generated by the planner for the first control code used for each control sketch in Table 2 . The control strength for the second control code is the mirror of each. process wherein the generator learns blending preferences from the human creator (Guzdial et al., 2018) .",
"cite_spans": [
{
"start": 1457,
"end": 1479,
"text": "(Guzdial et al., 2018)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [
{
"start": 91,
"end": 107,
"text": "Figures 2 and 3",
"ref_id": "FIGREF1"
},
{
"start": 1182,
"end": 1190,
"text": "Figure 5",
"ref_id": null
},
{
"start": 1294,
"end": 1301,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "In this paper we have made no assumptions about how control sketches are acquired from the human user. We envision a processes whereby the human co-creator has an idea of what the final product-in this case a story-should be and can express that vision at the level of how topics manifest through the story. Supporting a co-creative",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "This is not the only way to formalize this heuristic; We found this to be effective and efficient.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://groups.di.unipi.it/~gulli/AG_ corpus_of_news_articles.html",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "pipeline(\"zero-shot-classifier\")",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Generated Paragraph Sketch 1 c1 = Sports, {0 \u2192 5} c2 = Science, {4 \u2192 10} Jackie Robinson was playing in the NBA. He had just been drafted by the Cleveland Cavaliers. He decided to play for a team that was trying to win a championship. \"I'm gonna go out there and beat them,\" he said \"They're going to be great.\" \"Yeah, they're good,\" he said \"they'll be great.\" He added that his wife had been working on the project for about a year. He was happy to see her progress. She was able to get some work done and she started getting more comfortable with it. She had a lot of fun doing it! She also got her own computer, which is now in the living room.Sketch 2 c1 = Sports, 0 \u2192 5 c2 = Science, 5 \u2192 10 Jackie Robinson was playing in the NBA. He had just played basketball for a few years. He was looking forward to his first game of the season. He decided to play with his friends and play against them in the court. He got a lot of feedback from everyone who played against him, including some that were very excited about it! I was really happy when I saw how he played. I also had to admit that my favorite player was the guy who beat me in the finals. The computer game Super Mario 64 is a great game, but it's not perfect. I played it on my laptop and found that I couldn't play it properly because of some bugs. The problem was that the graphics were bad, so I had to use an emulator instead of playing the game.Sketch 3 c1 = Sports, 0 \u2192 5 c2 = Science, 6 \u2192 10 Jackie Robinson was playing in the NBA. He had just played basketball for a few years. He was looking forward to his first game of the season. He decided to play with his friends and play against them in the court. He had a lot of fun playing against them, but he didn't want to lose any time. So he played with his friends for about an hour before going home and playing again. He was very happy when they got home and started playing again! I think it's a good idea to have some fun with your kids, especially if you're not too busy. I'm sure that you'll enjoy this post as much as I did! my daughter was diagnosed with a rare form of cancer. human-AI interaction, the human user can update the control sketch and re-generate parts (or all) of the story by changing the range of topics or choosing different topics. The control model will need to support different topics at different levels of granularity; currently the control model only supports four topics, which is sufficient for conducting experiments to characterize the plug-and-blend technique but not for full co-creativity.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Control Sketches",
"sec_num": null
},
{
"text": "In this paper, we present Plug-and-Blend, a plugand-play framework that enhances a base LM, enables controllable generation with continuousweighted control codes, along with capability of generating paragraphs based on control sketches, all without access to internal knowledge of this base LM. These capabilities will fuel a new generation of controllable generation applications with the key assets of decoupling between the controllable component and the generative component, and easiness of adapting to new advancements in the field of generative LMs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "This material is based upon work supported by the Office of Naval Research (ONR) under Grant #N00014-14-1-0003.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgment",
"sec_num": "8"
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "STO-RIUM: A Dataset and Evaluation Platform for Machine-in-the-Loop Story Generation",
"authors": [
{
"first": "Nader",
"middle": [],
"last": "Akoury",
"suffix": ""
},
{
"first": "Shufan",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Josh",
"middle": [],
"last": "Whiting",
"suffix": ""
},
{
"first": "Stephen",
"middle": [],
"last": "Hood",
"suffix": ""
},
{
"first": "Nanyun",
"middle": [],
"last": "Peng",
"suffix": ""
},
{
"first": "Mohit",
"middle": [],
"last": "Iyyer",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "6470--6484",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.525"
]
},
"num": null,
"urls": [],
"raw_text": "Nader Akoury, Shufan Wang, Josh Whiting, Stephen Hood, Nanyun Peng, and Mohit Iyyer. 2020. STO- RIUM: A Dataset and Evaluation Platform for Machine-in-the-Loop Story Generation. In Proceed- ings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6470-6484, Online. Association for Computational Linguistics.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "CoCon: A Self-Supervised Approach for Controlled Text Generation",
"authors": [
{
"first": "Alvin",
"middle": [],
"last": "Chan",
"suffix": ""
},
{
"first": "Yew-Soon",
"middle": [],
"last": "Ong",
"suffix": ""
},
{
"first": "Bill",
"middle": [],
"last": "Pung",
"suffix": ""
},
{
"first": "Aston",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jie",
"middle": [],
"last": "Fu",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2006.03535[cs].ArXiv:2006.03535"
]
},
"num": null,
"urls": [],
"raw_text": "Alvin Chan, Yew-Soon Ong, Bill Pung, Aston Zhang, and Jie Fu. 2020. CoCon: A Self- Supervised Approach for Controlled Text Genera- tion. arXiv:2006.03535 [cs]. ArXiv: 2006.03535.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"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": 2020,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2020. Plug and Play Language Mod- els: A Simple Approach to Controlled Text Genera- tion. International Conference on Learning Repre- sentations, (2020). ArXiv: 1912.02164.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Pre-train and Plug-in: Flexible Conditional Text Generation with Variational Auto-Encoders",
"authors": [
{
"first": "Yu",
"middle": [],
"last": "Duan",
"suffix": ""
},
{
"first": "Canwen",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Jiaxin",
"middle": [],
"last": "Pei",
"suffix": ""
},
{
"first": "Jialong",
"middle": [],
"last": "Han",
"suffix": ""
},
{
"first": "Chenliang",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "253--262",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yu Duan, Canwen Xu, Jiaxin Pei, Jialong Han, and Chenliang Li. 2020. Pre-train and Plug-in: Flexible Conditional Text Generation with Variational Auto- Encoders. Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, (2020):253-262. ArXiv: 1911.03882.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Transformerbased Conditional Variational Autoencoder for Controllable Story Generation",
"authors": [
{
"first": "Le",
"middle": [],
"last": "Fang",
"suffix": ""
},
{
"first": "Tao",
"middle": [],
"last": "Zeng",
"suffix": ""
},
{
"first": "Chaochun",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Liefeng",
"middle": [],
"last": "Bo",
"suffix": ""
},
{
"first": "Wen",
"middle": [],
"last": "Dong",
"suffix": ""
},
{
"first": "Changyou",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2101.00828"
]
},
"num": null,
"urls": [],
"raw_text": "Le Fang, Tao Zeng, Chaochun Liu, Liefeng Bo, Wen Dong, and Changyou Chen. 2021. Transformer- based Conditional Variational Autoencoder for Con- trollable Story Generation. arXiv:2101.00828 [cs].",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Controlling Linguistic Style Aspects in Neural Language Generation",
"authors": [
{
"first": "Jessica",
"middle": [],
"last": "Ficler",
"suffix": ""
},
{
"first": "Yoav",
"middle": [],
"last": "Goldberg",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the Workshop on Stylistic Variation",
"volume": "",
"issue": "",
"pages": "94--104",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jessica Ficler and Yoav Goldberg. 2017. Controlling Linguistic Style Aspects in Neural Language Gen- eration. Proceedings of the Workshop on Stylistic Variation, (2017):94-104. ArXiv: 1707.02633.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Co-Creative Level Design via Machine Learning",
"authors": [
{
"first": "Matthew",
"middle": [],
"last": "Guzdial",
"suffix": ""
},
{
"first": "Nicholas",
"middle": [],
"last": "Liao",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Riedl",
"suffix": ""
}
],
"year": 2018,
"venue": "Fifth Experimental AI in Games Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matthew Guzdial, Nicholas Liao, and Mark Riedl. 2018. Co-Creative Level Design via Machine Learn- ing. Fifth Experimental AI in Games Workshop. ArXiv: 1809.09420.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "A Simple Language Model for Task-Oriented Dialogue",
"authors": [
{
"first": "Ehsan",
"middle": [],
"last": "Hosseini-Asl",
"suffix": ""
},
{
"first": "Bryan",
"middle": [],
"last": "Mccann",
"suffix": ""
},
{
"first": "Chien-Sheng",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Semih",
"middle": [],
"last": "Yavuz",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
}
],
"year": 2020,
"venue": "Advances in Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. 2020. A Simple Language Model for Task-Oriented Dialogue. Ad- vances in Neural Information Processing Systems, 33.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Toward Controlled Generation of Text",
"authors": [
{
"first": "Zhiting",
"middle": [],
"last": "Hu",
"suffix": ""
},
{
"first": "Zichao",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Xiaodan",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "Eric",
"middle": [
"P"
],
"last": "Xing",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 34th International Conference on Machine Learning",
"volume": "70",
"issue": "",
"pages": "1587--1596",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P. Xing. 2017. Toward Con- trolled Generation of Text. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Re- search, pages 1587-1596, International Convention Centre, Sydney, Australia. PMLR.",
"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": "Lav",
"middle": [
"R"
],
"last": "Mccann",
"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. Varsh- ney, Caiming Xiong, and Richard Socher. 2019. CTRL: A Conditional Transformer Language Model for Controllable Generation. arXiv:1909.05858 [cs].",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Hady Elsahar, and Marc Dymetman. 2020. A Distributional Approach to Controlled Text Generation",
"authors": [
{
"first": "Muhammad",
"middle": [],
"last": "Khalifa",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2012.11635[cs].ArXiv:2012.11635"
]
},
"num": null,
"urls": [],
"raw_text": "Muhammad Khalifa, Hady Elsahar, and Marc Dymet- man. 2020. A Distributional Approach to Controlled Text Generation. arXiv:2012.11635 [cs]. ArXiv: 2012.11635.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Shafiq Joty, Richard Socher, and Nazneen Fatema Rajani. 2020. GeDi: Generative Discriminator Guided Sequence Generation",
"authors": [
{
"first": "Ben",
"middle": [],
"last": "Krause",
"suffix": ""
},
{
"first": "Akhilesh",
"middle": [],
"last": "Deepak Gotmare",
"suffix": ""
},
{
"first": "Bryan",
"middle": [],
"last": "Mc-Cann",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2009.06367[cs].ArXiv:2009.06367"
]
},
"num": null,
"urls": [],
"raw_text": "Ben Krause, Akhilesh Deepak Gotmare, Bryan Mc- Cann, Nitish Shirish Keskar, Shafiq Joty, Richard Socher, and Nazneen Fatema Rajani. 2020. GeDi: Generative Discriminator Guided Sequence Genera- tion. arXiv:2009.06367 [cs]. ArXiv: 2009.06367.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Prefix-Tuning: Optimizing Continuous Prompts for Generation",
"authors": [
{
"first": "Lisa",
"middle": [],
"last": "Xiang",
"suffix": ""
},
{
"first": "Percy",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Liang",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2101.00190[cs].ArXiv:2101.00190"
]
},
"num": null,
"urls": [],
"raw_text": "Xiang Lisa Li and Percy Liang. 2021. Prefix-Tuning: Optimizing Continuous Prompts for Generation. arXiv:2101.00190 [cs]. ArXiv: 2101.00190.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Sumanth Dathathri, and Pascale Fung. 2020. Plug-and-Play Conversational Models",
"authors": [
{
"first": "Andrea",
"middle": [],
"last": "Madotto",
"suffix": ""
},
{
"first": "Etsuko",
"middle": [],
"last": "Ishii",
"suffix": ""
},
{
"first": "Zhaojiang",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2010.04344"
]
},
"num": null,
"urls": [],
"raw_text": "Andrea Madotto, Etsuko Ishii, Zhaojiang Lin, Sumanth Dathathri, and Pascale Fung. 2020. Plug-and-Play Conversational Models. arXiv:2010.04344 [cs].",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Plug and Play Autoencoders for Conditional Text Generation",
"authors": [
{
"first": "Florian",
"middle": [],
"last": "Mai",
"suffix": ""
},
{
"first": "Nikolaos",
"middle": [],
"last": "Pappas",
"suffix": ""
},
{
"first": "Ivan",
"middle": [],
"last": "Montero",
"suffix": ""
},
{
"first": "Noah",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Henderson",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "6076--6092",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.491"
]
},
"num": null,
"urls": [],
"raw_text": "Florian Mai, Nikolaos Pappas, Ivan Montero, Noah A. Smith, and James Henderson. 2020. Plug and Play Autoencoders for Conditional Text Generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6076-6092, Online. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "A Corpus and Evaluation Framework 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 {A}merican 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 Evaluation Framework for Deeper Under- standing of Commonsense Stories. Proceedings of the 2016 Conference of the North {A}merican Chap- ter of the Association for Computational Linguis- tics: Human Language Technologies, pages 839- 849. ArXiv: 1604.01696.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Directed Beam Search: Plug-and-Play Lexically Constrained Language Generation",
"authors": [
{
"first": "Damian",
"middle": [],
"last": "Pascual",
"suffix": ""
},
{
"first": "Beni",
"middle": [],
"last": "Egressy",
"suffix": ""
},
{
"first": "Florian",
"middle": [],
"last": "Bolli",
"suffix": ""
},
{
"first": "Roger",
"middle": [],
"last": "Wattenhofer",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2012.15416"
]
},
"num": null,
"urls": [],
"raw_text": "Damian Pascual, Beni Egressy, Florian Bolli, and Roger Wattenhofer. 2020. Directed Beam Search: Plug-and-Play Lexically Constrained Language Generation. arXiv:2012.15416 [cs].",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"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": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Lan- guage Models are Unsupervised Multitask Learners. page 24.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "PlotMachines: Outlineconditioned generation with dynamic plot state tracking",
"authors": [
{
"first": "Asli",
"middle": [],
"last": "Hannah Rashkin",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Celikyilmaz",
"suffix": ""
},
{
"first": "Jianfeng",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Gao",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "4274--4295",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.349"
]
},
"num": null,
"urls": [],
"raw_text": "Hannah Rashkin, Asli Celikyilmaz, Yejin Choi, and Jianfeng Gao. 2020. PlotMachines: Outline- conditioned generation with dynamic plot state tracking. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 4274-4295, Online. Associa- tion for Computational Linguistics.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"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": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "2153--2162",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gard- ner, and Sameer Singh. 2019. Universal Adver- sarial Triggers for Attacking and Analyzing NLP. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lan- guage Processing (EMNLP-IJCNLP), (2019):2153- 2162. ArXiv: 1908.07125.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Narrative Interpolation for Generating and Understanding Stories",
"authors": [
{
"first": "Su",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Greg",
"middle": [],
"last": "Durrett",
"suffix": ""
},
{
"first": "Katrin",
"middle": [],
"last": "Erk",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2008.07466"
]
},
"num": null,
"urls": [],
"raw_text": "Su Wang, Greg Durrett, and Katrin Erk. 2020. Nar- rative Interpolation for Generating and Understand- ing Stories. arXiv:2008.07466 [cs].",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "2020. HuggingFace's Transformers: State-of-the-art Natural Language Processing",
"authors": [
{
"first": "Sylvain",
"middle": [],
"last": "Teven Le Scao",
"suffix": ""
},
{
"first": "Mariama",
"middle": [],
"last": "Gugger",
"suffix": ""
},
{
"first": "Quentin",
"middle": [],
"last": "Drame",
"suffix": ""
},
{
"first": "Alexander",
"middle": [
"M"
],
"last": "Lhoest",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Rush",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1910.03771"
]
},
"num": null,
"urls": [],
"raw_text": "Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. HuggingFace's Transformers: State-of-the-art Nat- ural Language Processing. arXiv:1910.03771 [cs].",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Hannaneh Hajishirzi, Mari Ostendorf, and Bill Dolan. 2020. A Controllable Model of Grounded Response Generation",
"authors": [
{
"first": "Zeqiu",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Michel",
"middle": [],
"last": "Galley",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Brockett",
"suffix": ""
},
{
"first": "Yizhe",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Xiang",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Quirk",
"suffix": ""
},
{
"first": "Rik",
"middle": [],
"last": "Koncel-Kedziorski",
"suffix": ""
},
{
"first": "Jianfeng",
"middle": [],
"last": "Gao",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2005.00613[cs].ArXiv:2005.00613"
]
},
"num": null,
"urls": [],
"raw_text": "Zeqiu Wu, Michel Galley, Chris Brockett, Yizhe Zhang, Xiang Gao, Chris Quirk, Rik Koncel-Kedziorski, Jianfeng Gao, Hannaneh Hajishirzi, Mari Ostendorf, and Bill Dolan. 2020. A Controllable Model of Grounded Response Generation. arXiv:2005.00613 [cs]. ArXiv: 2005.00613.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Plan-And-Write: Towards Better Automatic Storytelling",
"authors": [
{
"first": "Lili",
"middle": [],
"last": "Yao",
"suffix": ""
},
{
"first": "Nanyun",
"middle": [],
"last": "Peng",
"suffix": ""
},
{
"first": "Ralph",
"middle": [],
"last": "Weischedel",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
},
{
"first": "Dongyan",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Rui",
"middle": [],
"last": "Yan",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the AAAI Conference on Artificial Intelligence",
"volume": "33",
"issue": "",
"pages": "7378--7385",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lili Yao, Nanyun Peng, Ralph Weischedel, Kevin Knight, Dongyan Zhao, and Rui Yan. 2019. Plan- And-Write: Towards Better Automatic Storytelling. Proceedings of the AAAI Conference on Artificial In- telligence, 33(1):7378-7385. ArXiv: 1811.05701.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"type_str": "figure",
"text": "(a) Baseline on order-shuffled stories in ROCStories dataset. (b) Total control strength 1x. (c) Total control strength 2x. (d) Total control strength 4x.",
"num": null
},
"FIGREF1": {
"uris": null,
"type_str": "figure",
"text": "average \u03c4 -a (higher meaning better control fidelity) under different Total control strength for the tuned model with topics: (c1) Business, (c2) Science, (c3) Sports, (c4) World, comparing to uncontrolled baseline. Heat map strength is given as percentages (\u2212100% . . . 100%).(a) Perplexity of generated sequences. (b) Total control strength 1x. (c) Total control strength 2x. (d) Total control strength 4x.",
"num": null
},
"FIGREF2": {
"uris": null,
"type_str": "figure",
"text": "Experiment results for the untuned model. Refer to Figure 3a for baseline comparison.",
"num": null
}
}
}
} |