File size: 124,987 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 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 |
{
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T02:11:57.217019Z"
},
"title": "Detecting East Asian Prejudice on Social Media",
"authors": [
{
"first": "Bertie",
"middle": [],
"last": "Vidgen",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The Alan Turing Institute",
"location": {}
},
"email": "bvidgen@turing.ac.uk"
},
{
"first": "Austin",
"middle": [],
"last": "Botelho",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Oxford",
"location": {}
},
"email": ""
},
{
"first": "David",
"middle": [],
"last": "Broniatowski",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The George Washington University",
"location": {}
},
"email": ""
},
{
"first": "Ella",
"middle": [],
"last": "Guest",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Matthew",
"middle": [],
"last": "Hall",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The University of Surrey",
"location": {}
},
"email": ""
},
{
"first": "Helen",
"middle": [],
"last": "Margetts",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Rebekah",
"middle": [],
"last": "Tromble",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The George Washington University",
"location": {}
},
"email": ""
},
{
"first": "Zeerak",
"middle": [],
"last": "Waseem",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Sheffield",
"location": {}
},
"email": ""
},
{
"first": "Scott",
"middle": [
"A"
],
"last": "Hale",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Oxford Meedan",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "During COVID-19 concerns have heightened about the spread of aggressive and hateful language online, especially hostility directed against East Asia and East Asian people. We report on a new dataset and the creation of a machine learning classifier that categorizes social media posts from Twitter into four classes: Hostility against East Asia, Criticism of East Asia, Meta-discussions of East Asian prejudice, and Non-related. The classifier achieves a macro-F1 score of 0.83. We then conduct an in-depth ground-up error analysis and show that the model struggles with edge cases and ambiguous content. We provide the 20,000 tweet training dataset (annotated by experienced analysts), which also contains several secondary categories and additional flags. We also provide the 40,000 original annotations (before adjudication), the full codebook, annotations for COVID-19 relevance and East Asian relevance and stance for 1,000 hashtags, and the final model.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "During COVID-19 concerns have heightened about the spread of aggressive and hateful language online, especially hostility directed against East Asia and East Asian people. We report on a new dataset and the creation of a machine learning classifier that categorizes social media posts from Twitter into four classes: Hostility against East Asia, Criticism of East Asia, Meta-discussions of East Asian prejudice, and Non-related. The classifier achieves a macro-F1 score of 0.83. We then conduct an in-depth ground-up error analysis and show that the model struggles with edge cases and ambiguous content. We provide the 20,000 tweet training dataset (annotated by experienced analysts), which also contains several secondary categories and additional flags. We also provide the 40,000 original annotations (before adjudication), the full codebook, annotations for COVID-19 relevance and East Asian relevance and stance for 1,000 hashtags, and the final model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The outbreak of COVID-19 has raised concerns about the spread of Sinophobia and other forms of East Asian prejudice across the world, with reports of online and offline abuse directed against East Asian people, including physical attacks (Flanagan, 2020; Wong, 2020; Liu, 2020; Walton, 2020; Solomon, 2020; Guy, 2020) . The United Nations High Commissioner for Human Rights has drawn attention to these issues, calling on UN member states to fight the 'tsunami' of hate and xenophobia (Shields, 2020) .",
"cite_spans": [
{
"start": 238,
"end": 254,
"text": "(Flanagan, 2020;",
"ref_id": "BIBREF9"
},
{
"start": 255,
"end": 266,
"text": "Wong, 2020;",
"ref_id": null
},
{
"start": 267,
"end": 277,
"text": "Liu, 2020;",
"ref_id": "BIBREF18"
},
{
"start": 278,
"end": 291,
"text": "Walton, 2020;",
"ref_id": "BIBREF36"
},
{
"start": 292,
"end": 306,
"text": "Solomon, 2020;",
"ref_id": "BIBREF27"
},
{
"start": 307,
"end": 317,
"text": "Guy, 2020)",
"ref_id": "BIBREF14"
},
{
"start": 485,
"end": 500,
"text": "(Shields, 2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "As digital technologies become even more important for maintaining social connections, it is crucial that online spaces remain safe, accessible and free from abuse (Cowls et al., 2020) -and that people's fears and distress are not exploited during the pandemic. Computational tools, including machine learning and natural language processing, offer powerful ways of creating scalable and robust systems for detecting and measuring prejudice. These, in turn, can assist with both online content moderation processes and further research into the dynamics, prevalence, causes, and impact of abuse.",
"cite_spans": [
{
"start": 164,
"end": 184,
"text": "(Cowls et al., 2020)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We report on the creation of a new dataset and classifier to detect East Asian prejudice in social media data. The classifier distinguishes between four primary categories: Hostility against East Asia, Criticism of East Asia, Meta-discussions of East Asian prejudice, and Non-related. It achieves a macro-F1 score of 0.83. The 20,000 tweet training dataset used to create the classifier and the annotation codebook are also provided. The dataset contains annotations for several secondary categories, including threatening language, interpersonal abuse, and dehumanization, which can be used for further research. In addition, we provide the 40,000 original annotations given by our experienced annotators, which can be used for further investigation of annotating prejudice. We also annotated 1,000 hashtags in our dataset for East Asian relevance and stance, as well as other attributes. These are also provided for other researchers. 1 To provide insight into what types of content causes the model to fail, we conduct a ground-up qualitative error analysis. We show that 17% of errors are due to annotation mistakes and 83% due to the machine learning model. Of these, 29% are clear errors (i.e. obvious mistakes) and 54% are edge cases (i.e. more complex and nuanced cases). In the machine learning edge cases, we show that the model struggles with lexically similar content (e.g. distinguishing Hostility against East Asia from Criticism of East Asia), as well as ambiguous content (i.e. where there is uncertainty among the annotators about the correct label).",
"cite_spans": [
{
"start": 937,
"end": 938,
"text": "1",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Finally, we analyze the hashtags most closely associated with the primary categories in the training dataset, identifying several terms which could guide future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "East Asian prejudice, such as Sinophobia, can be understood as fear or hatred of East Asia and East Asian people (Bill\u00e9, 2015) . This prejudice has a long history in the West: in the 19th century the term \"yellow peril\" was used to refer to Chinese immigrants who were stereotyped as dirty and diseased and considered akin to a plague (Goossen et al., 2004) . Often, the association of COVID-19 with China plays into these old stereotypes, as shown by derogatory references to 'bats' and 'monkeys' online (Zhang, 2020) . In 2017 a study found that 21% of Asian Americans had received threats based on their Asian identities, and 10% had been victims of violence (Neel, 2017) . Likewise, a 2009 report on the discrimination and experiences of East Asian people in the UK, described Sinophobia as a problem that was 'hidden from public view' (Adamson et al., 2009) .",
"cite_spans": [
{
"start": 113,
"end": 126,
"text": "(Bill\u00e9, 2015)",
"ref_id": "BIBREF1"
},
{
"start": 335,
"end": 357,
"text": "(Goossen et al., 2004)",
"ref_id": "BIBREF12"
},
{
"start": 505,
"end": 518,
"text": "(Zhang, 2020)",
"ref_id": "BIBREF41"
},
{
"start": 662,
"end": 674,
"text": "(Neel, 2017)",
"ref_id": "BIBREF21"
},
{
"start": 840,
"end": 862,
"text": "(Adamson et al., 2009)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "New research related to East Asian prejudice during COVID-19 has already provided insight into its nature, prevalence and dynamics with Schild et al. (2020) finding an increase in Sinophobic language on some social media platforms such as 4chan. Analysis by the company Moonshot CVE also suggests that the use of anti-Chinese hashtags has increased substantially (The New Statesman, 2020). They analysed more than 600 million tweets and found that 200,000 contained either Sinophobic hate speech or conspiracy theories, and identified a 300% increase in hashtags that supported or encouraged violence against China during a single week in March 2020. Similarly, Vel\u00e1squez et al. (2020) show the existence of a Sinophobic 'hate multiverse', with hateful content following contagion patterns and clusters which are similar to the epidemiological diffusion of COVID-19 itself. Ziems et al. (2020) argue that racism 'is a virus', and study a dataset of 30 million COVID-19 tweets using a classifier trained on 2,400 tweets. They identify 900,000 hateful tweets and 200,000 counter-hate, and find that 10% of users who share hate speech are bot accounts. Toney et al. (2020) used the Word Embedding Association Test in the context of COVID-19 to analyse anti-China sentiment on Twitter, finding substantial biases in how Asian people are viewed. East Asian prejudice has also been linked to the spread of COVID-19 healthrelated misinformation (Cinelli et al., 2020) and in March 2020 the polling company YouGov found that 1 in 5 Brits believed the conspiracy theory that the coronavirus was developed in a Chinese lab (Nolsoe, 2020) .",
"cite_spans": [
{
"start": 136,
"end": 156,
"text": "Schild et al. (2020)",
"ref_id": "BIBREF24"
},
{
"start": 662,
"end": 685,
"text": "Vel\u00e1squez et al. (2020)",
"ref_id": "BIBREF33"
},
{
"start": 874,
"end": 893,
"text": "Ziems et al. (2020)",
"ref_id": "BIBREF42"
},
{
"start": 1150,
"end": 1169,
"text": "Toney et al. (2020)",
"ref_id": "BIBREF30"
},
{
"start": 1438,
"end": 1460,
"text": "(Cinelli et al., 2020)",
"ref_id": "BIBREF3"
},
{
"start": 1613,
"end": 1627,
"text": "(Nolsoe, 2020)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Research into computational tools for detecting, categorizing, and measuring online hate has received substantial attention in recent years (Waseem et al., 2017) . However, a systematic review of hate speech training datasets conducted by Vidgen and Derczynski (2020) shows that classifiers and training datasets for East Asian prejudice are not currently available. Somewhat similar datasets are available, pertaining to racism (Waseem and Hovy, 2016) Islamophobia (Chung et al., 2019) and 'hate' in general de Gibert et al., 2018) but they cannot easily be re-purposed for East Asian prejudice detection. The absence of an appropriate training dataset (and automated detection tools) means that researchers have to rely instead on less precise ways of measuring East Asian prejudice, such as using keyword searches for slurs and other pejorative terms. These methods create substantial errors as more covert prejudice is missed because the content does not contain the target keywords, and non-hateful content is misclassified simply because it does contain a keyword.",
"cite_spans": [
{
"start": 140,
"end": 161,
"text": "(Waseem et al., 2017)",
"ref_id": "BIBREF37"
},
{
"start": 239,
"end": 267,
"text": "Vidgen and Derczynski (2020)",
"ref_id": "BIBREF34"
},
{
"start": 429,
"end": 452,
"text": "(Waseem and Hovy, 2016)",
"ref_id": "BIBREF38"
},
{
"start": 466,
"end": 486,
"text": "(Chung et al., 2019)",
"ref_id": "BIBREF2"
},
{
"start": 509,
"end": 532,
"text": "de Gibert et al., 2018)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Developing new detection tools is a complex and lengthy process. The field of hate speech detection sits at the intersection of social science and computer science and is fraught with not only technical challenges but also deep-routed ethical and theoretical considerations (Vidgen et al., 2019) . If machine learning tools are to be effective then they need to be developed with consideration of their social implications (Vidgen et al., 2019; Sap et al., 2019; Davidson et al., 2019; Garg et al., 2019) .",
"cite_spans": [
{
"start": 274,
"end": 295,
"text": "(Vidgen et al., 2019)",
"ref_id": "BIBREF35"
},
{
"start": 423,
"end": 444,
"text": "(Vidgen et al., 2019;",
"ref_id": "BIBREF35"
},
{
"start": 445,
"end": 462,
"text": "Sap et al., 2019;",
"ref_id": "BIBREF23"
},
{
"start": 463,
"end": 485,
"text": "Davidson et al., 2019;",
"ref_id": "BIBREF6"
},
{
"start": 486,
"end": 504,
"text": "Garg et al., 2019)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "To create our 20,000 tweet training dataset, we collected tweets from Twitter's Streaming API using 14 hashtags that relate to both East Asia and the novel coronavirus. 2 Some of these hashtags ex-press anti-East Asian sentiments (e.g. '#chinaflu') but others, such as '#wuhan' are more neutral, referring to the geographic origin of the virus. Data collection ran initially from 11 to 17 March 2020, returning 769,763 tweets, of which 96,283 were unique entries in English. To minimize biases that could emerge from collecting data over a relatively short period of time, we then collected tweets from 1 January to 10 March 2020 using the same keywords from a 10% random sample of Twitter (the 'Decahose'), provided by a third party. We identified a further 63,037 unique tweets in English, which we added to our dataset. The full dataset comprises 159,320 unique tweets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset Collection",
"sec_num": "3"
},
{
"text": "To create a training dataset for annotation we sampled from the full dataset. To guide the sampling process, we extracted the 1,000 most used hashtags from the 159,320 tweets. Three annotators independently marked them for: (1) whether they are East Asian relevant and, if so, 2what Asian entity is discussed (e.g., China, Xi Jinping, South Korea), 3what the stance is towards the Asian entity (Very Negative, Negative, Neutral, Positive, or Very Positive) and also (4) whether they relate to COVID-19. 97 hashtags were marked as either Negative or Very Negative toward East Asia by at least one annotator. All annotations for hashtags are available in our data repository.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset Collection",
"sec_num": "3"
},
{
"text": "We then sampled 10,000 tweets at random from the full dataset and a further 10,000 tweets which used one of the 97 hashtags identified as Negative or Very Negative towards East Asia, thereby increasing the likelihood that prejudicial tweets would be identified and ensuring that our dataset is suitable for training a classifier (Schmidt and Wiegand, 2017; Vidgen et al., 2019) . The training dataset comprises 20,000 tweets in total.",
"cite_spans": [
{
"start": 329,
"end": 356,
"text": "(Schmidt and Wiegand, 2017;",
"ref_id": "BIBREF25"
},
{
"start": 357,
"end": 377,
"text": "Vidgen et al., 2019)",
"ref_id": "BIBREF35"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset Collection",
"sec_num": "3"
},
{
"text": "Initial qualitative inspection of the dataset showed that hashtags played a key role in how COVID-19 was discussed and how hostility against East Asia was expressed. Hashtags often appeared in the middle of tweets, especially when they related to East Asia and/or COVID-19. For example: its wellknown #covid19 originated from #china. Instead of #Doingtherightthing they're blaming others, typical. You cant trust these #YellowFever to sort anything out.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "Without the hashtags it is difficult to discern whether this tweet expresses prejudice against East Asia. In this regard, it is important that they are seen by annotators to ensure high quality labels. However, in other cases, the inclusion of hashtags risked low quality labels. In a test round of annotation (not included in the dataset presented here) annotators over-relied on the prejudicial hashtags, marking up nearly all tweets which contained them as prejudiced against East Asia, even if they were otherwise neutral. This is problematic because we used hashtags to sample the training data so they are highly prevalent. If all of their uses were identified as prejudicial then any systems trained on the dataset would likely overfit to just a few keywords. This could severely constrain the system's generalizability, potentially leading to poor performance on new content.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "To address this challenge, we performed a hashtag replacement on all tweets prior to presenting them to the annotators. For the 1,000 most used hashtags (annotated as part of the data sampling phase), we had one annotator identify appropriate thematic replacement hashtags. We used five thematic replacements:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "\u2022 #EASTASIA: Relate only to an East Asian entity, e.g. #China or #Wuhan",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "\u2022 #VIRUS: Relate only to COVID-19, e.g. #coronavirus or #covid19.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "\u2022 #EASTASIAVIRUS: Relate to both an East Asian entity and COVID-19, e.g. #wuhanflu.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "\u2022 #OTHERCOUNTRYVIRUS: Relate to both a Country (which is not East Asian) and COVID-19, e.g. #coronacanada or #italycovid.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "\u2022 #HASHTAG: Not directly relevant to COVID-19 or East Asia, e.g. #maga or #spain.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "Annotators could still discern the meaning of tweets because they were presented with the hashtags' topics. However, they were not unduly biased by the substantive outlook, stance, and sentiment the hashtags express. All hashtags beyond our annotated list of 1,000 were replaced with a generic replacement, #HASHTAG. The 1,000 thematic hashtag replacements are available in our data repository. With this process, the quote above is transformed to: its wellknown #HASHTAGVIRUS originated from #HASHTAGEASTASIA. Instead of #HASHTAG they're blaming others, typical. You cant trust these #HASHTAGEASTASIAVIRUS to sort anything out. This process, although time consuming, strikes a balance between preserving the meaning of the tweet for annotation and minimizing the risk of overfitting. It means that the text the annotators are presented with is also the same text that is fed into our final models (i.e. both annotations and the model classifications are based on the replaced hashtags). In principle, it would be easy for anyone applying our final classification model to a new dataset to update the hashtag replacement list and then apply it to their data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data pre-processing for annotation",
"sec_num": "3.1"
},
{
"text": "Annotation was completed by a team of 26 annotators, all of whom had completed at least 4 weeks of training on a previous hate speech annotation project. The annotators were all aged between 18 and 35, spoke English fluently (50% were native speakers), were 75% female and were all educated to at least an undergraduate level. 25% were studying for higher degrees. Annotators came from the United Kingdom (60%), elsewhere in Europe (30%) and South America (10%). Information about their sexuality, religious and political affiliation is not available due to their sensitivity.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotators",
"sec_num": "4.1"
},
{
"text": "Two experts were used to adjudicate decisions on the primary categories. Both experts were final year PhD students working on extreme behaviour online. One was male; one was female. They were both aged between 25 and 35, and were native English speakers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotators",
"sec_num": "4.1"
},
{
"text": "Tweets were first annotated for the presence of two themes: (1) East Asia and (2) COVID-19. If a tweet was not East Asian relevant then no further annotations were required and it was automatically assigned to the Non-related class. Annotators then used an additional flag for how they marked up the two themes, which we call 'hashtag dependence.' For this label, annotators were asked whether they had used the hashtags to identify the themes or the themes were apparent without the hashtags.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Themes",
"sec_num": "4.2"
},
{
"text": "Our approach to annotating themes and the role of hashtags required substantial training for annotators (involving one-to-one on-boarding sessions). This detailed annotation process means that we can provide insight into not only what annotations were made but also how, which we anticipate will be of use to scholars working on online communications beyond online prejudice.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Themes",
"sec_num": "4.2"
},
{
"text": "Each tweet was assigned to one of five mutually exclusive primary categories.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "\u2022 Hostility against an East Asian (EA) entity:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "Express abuse or intense negativity against an East Asian entity, primarily by derogating/attacking them (e.g. \"Those oriental devils don't care about human life\" or \"Chinks will bring about the downfall of western civilization\"). It also includes: conspiracy theories, claiming East Asians are a threat, and expressing negative emotions about them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "\u2022 Criticism of an East Asian entity: Make a negative judgement/assessment of an East Asian entity, without being abusive. This includes commenting on perceived social, economic and political faults, including questioning their response to the pandemic and how they are governed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "The Hostility/Criticism distinction is crucial for addressing a core issue in online hate speech research, namely ensuring that freedom of speech is protected (Ullmann and Tomalin, 2020) . The Criticism category minimizes the chance that users who engage in what has been termed 'legitimate critique' (Imhoff and Recker, 2012) will have their comments erroneously labelled as hostile.",
"cite_spans": [
{
"start": 159,
"end": 186,
"text": "(Ullmann and Tomalin, 2020)",
"ref_id": "BIBREF31"
},
{
"start": 301,
"end": 326,
"text": "(Imhoff and Recker, 2012)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "\u2022 Counter speech: Explicitly challenge or condemn abuse against an East Asian entity. It includes rejecting the premise of abuse (e.g., \"it isn't right to blame China!\"), describing content as hateful or prejudicial (e.g., \"you shouldn't say that, it's derogatory\") or expressing solidarity with target entities (e.g., \"Stand with Chinatown against racists\").",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "\u2022 Discussion of East Asian prejudice Tweets that discuss prejudice related to East Asians but do not engage in, or counter, that prejudice (e.g., \"It's not racist to call it the Wuhan virus\"). It includes content which discusses whether East Asian prejudice has increased during COVID-19, the supposed media focus on prejudice, and/or free speech.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "\u2022 Non-related Do not fall into any of the other categories. Note that they could be abusive in other ways, such as expressing misogyny.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "The primary categories were annotated with a two step process. First, each tweet was annotated independently by two trained annotators. Second, one of two expert adjudicators reviewed cases where annotators disagreed about the primary category. Experts could decide an entirely new primary category if needed. Expert adjudication was not used for the themes and secondary categories.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "Agreement is reported for each pair of annotators in Table 3 , with the average, minimum, and maximum. Overall, agreement levels are moderate, with better results for the two most important and prevalent categories (Hostility and Non-related) but poorer on the less frequent and more nuanced categories (Counter Speech, Criticism and Discussion of EA prejudice). Note that if Counter Speech and Discussion of EA prejudice are combined then there is a marked improvement in overall agreement levels, with an average Kappa of 0.5 for the combined category.",
"cite_spans": [],
"ref_spans": [
{
"start": 53,
"end": 60,
"text": "Table 3",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "Experts adjudicated 4,478 cases (22%) where annotators did not agree. Experts tended to move tweets out of Non-related into other categories, primarily Hostility. Of the 8,956 original annotations given to the 4,478 tweets they adjudicated, 34% of them were in Non-related and yet only 29% of their adjudicated decisions were in this category. This was matched by an equivalent increase in the Hostility category, from 31.6% of the original annotations to 35% of the expert adjudications. The other three categories remained broadly stable. In 347 cases (7.7%), experts choose a category that was not selected by either annotator. Of the 694 original annotations given to these 347 cases, 18.7% were for Criticism compared with 39.4% of the expert adjudications for these entries (a similar decrease can be observed for the Non-related category). The most common decision made by experts for these 347 tweets was to label a tweet as Criticism when one annotator had selected Hostility and the other selected Non-related. These results shows the fundamental ambiguity of hate speech annotation and the need for expert adjudication. With complex and often-ambiguous content even well-trained annotators can make decisions which are inappropriate.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Primary categories",
"sec_num": "4.3"
},
{
"text": "For the Hostility and Criticism primary categories, annotators identified what East Asian entity was targeted (e.g., \"Hong Kongers\", \"CCP\", or \"Chinese scientists\"). Initially, annotators identified targets inductively, which resulted in several hundred unique values. We then implemented a reconciliation process in which the number of unique targets was reduced to 78, reflecting six geographical areas (China, Korea, Japan, Taiwan, Singapore and East Asia in general) and several specific entities, such as scientists, women and government, including intersectional identities. For tweets identified as Hostility annotators applied three additional flags.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Secondary categories",
"sec_num": "4.4"
},
{
"text": "\u2022 Interpersonal abuse: East Asian prejudice which is targeted against an individual. Whether the individual is East Asian was not considered. (Waseem et al., 2017) .",
"cite_spans": [
{
"start": 142,
"end": 163,
"text": "(Waseem et al., 2017)",
"ref_id": "BIBREF37"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Secondary categories",
"sec_num": "4.4"
},
{
"text": "\u2022 Use of threatening language: Content which makes a threat against an East Asian entity, which includes expressing a desire/willingness to inflict harm or inciting others (The Law Commission, 2018; Weber, 2009 ).",
"cite_spans": [
{
"start": 172,
"end": 198,
"text": "(The Law Commission, 2018;",
"ref_id": null
},
{
"start": 199,
"end": 210,
"text": "Weber, 2009",
"ref_id": "BIBREF39"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Secondary categories",
"sec_num": "4.4"
},
{
"text": "\u2022 Dehumanization: Content which describes, compares or suggests equivalences between East Asians and non-humans or subhumans, such as insects, weeds, or actual viruses (Leader Maynard and Benesch, 2016; Musolff, 2015) .",
"cite_spans": [
{
"start": 168,
"end": 202,
"text": "(Leader Maynard and Benesch, 2016;",
"ref_id": "BIBREF16"
},
{
"start": 203,
"end": 217,
"text": "Musolff, 2015)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Secondary categories",
"sec_num": "4.4"
},
{
"text": "Note that our expert adjudicators did not adjudicate for these secondary categories. In cases where experts decided a tweet is Hostile but neither of the original annotators had selected that category then none of the secondary categories are available. In other cases, experts decided a tweet was Hostile and so only one annotation for the secondary flags is available (as the other annotator selected a different category and did not provide these secondary annotations). Future researchers can decide how to use these secondary categories.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Secondary categories",
"sec_num": "4.4"
},
{
"text": "Due to their low prevalence and conceptual similarity, we combined the Counter Speech category with Discussion of East Asian Prejudice for classification. As such, the classification task was to distinguish between four primary categories: Hostility, Criticism, Discussion of East Asian Prejudice and Non-related. We implemented and fine-tuned several contextual embedding models as well as a one-hot LSTM model with a linear input layer, tanh activation, and a softmax output layer. We expect contextual embeddings to perform best as they take into account the context surrounding a token when generating each embedding (Vaswani et al., 2017) . We compared results against a one-hot LSTM model to test this expectation.",
"cite_spans": [
{
"start": 621,
"end": 643,
"text": "(Vaswani et al., 2017)",
"ref_id": "BIBREF32"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Classification results",
"sec_num": "5"
},
{
"text": "Models were developed with a stratified 80/10/10 training, testing, and validation split (maintaining the class distribution of the whole dataset). We processed all tweets by removing URLs and usernames, lower-casing, and replacing hashtags with either a generic hashtag-token or with the appropriate thematic hashtag-token from the annotation setup. Training was conducted using the same hyper-parameter sweep identified in Liu et al. (2019) as most effective for the GLUE benchmark tasks. This included testing across learning rates \u2208 {1e-5, 2e-5, 3e-5} and batch sizes \u2208 {32, 64} with an early stopping regime. Performance was optimized using the AdamW algorithm (Loshchilov and Hutter, 2019 ) and a scheduler that implements linear warmup and decay. For the LSTM baseline, we conduct a hyper-parameter search over batch sizes \u2208 {16, 32, 64} and learning rates {10 \u2212i , i \u2208 1, 5 increments}.",
"cite_spans": [
{
"start": 425,
"end": 442,
"text": "Liu et al. (2019)",
"ref_id": "BIBREF17"
},
{
"start": 666,
"end": 694,
"text": "(Loshchilov and Hutter, 2019",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Classification results",
"sec_num": "5"
},
{
"text": "All of the contextual embedding models outperformed the baseline in terms of macro F1. RoBERTa achieved the highest F1 score of the tested models (0.832), which is a 7-point improve- ment over the LSTM (0.76). This model harnesses the underlying bidirectional transformer architecture of BERT (Devlin et al., 2019) but alters the training hyperparameters and objectives to improve performance. Unexpectedly, the LSTM baseline outperforms all other models in terms of precision but has far lower recall. For the best performing model (RoBERTa), misclassifications are shown in the confusion matrix. The model performs well across all categories, with strongest performance in detecting tweets in the Non-related category (Recall of 91.6% and Precision of 93%). The model has few misclassifications between the most conceptually distinct categories (e.g. Hostility versus Non-related) but has far more errors between conceptually similar categories, such as Criticism and Hostility.",
"cite_spans": [
{
"start": 293,
"end": 314,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Classification results",
"sec_num": "5"
},
{
"text": "To better understand classification errors, we conducted a qualitative analysis of misclassified content (from the best performing model, RoBERTa), using a grounded theory approach (Corbin and Strauss, 1990) . This qualitative methodology is entirely inductive and data-driven. It involves systematically exploring themes as they emerge from the data and organizing them into a taxonomyrefining and collapsing the categories until 'saturation' is reached and the data fits neatly into a set of mutually exclusive and collectively exhaustive categories. Figure 2 shows the error categories within our sample of 340 misclassified tweets from the 2,000 (10%) validation split. The errors broadly fit within two branches: annotator errors (17%) and machine learning errors (83%). In future work, these errors could be addressed through creating a larger and more balanced dataset, more sophisticated machine learning architectures and reannotation of data.",
"cite_spans": [
{
"start": 181,
"end": 207,
"text": "(Corbin and Strauss, 1990)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 553,
"end": 561,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Error analysis",
"sec_num": "6"
},
{
"text": "Annotator errors are cases where the classification from the model better captures the tweets' content and is more consistent with our taxonomy and guidelines. In effect, we believe that the 'wrong' classification provided by the model is correctand that a mistake may have been made in the annotation process. Approximately 17% (N=58) of the errors were due to this. Note that this does not mean that 17% of the dataset is incorrectly annotated as this sample is biased by the fact that it has been selected precisely because the model made an 'error'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotator errors (17% of total)",
"sec_num": "6.1"
},
{
"text": "36 of the annotator errors were clear misapplications of primary categories. The other 22 were cases where annotators made detailed annotations for tweets which were incorrectly marked as East Asian relevant. These are path dependency errors and show the importance of annotators following the right instructions throughout the process. If an incorrect annotation is made early on then the subsequent annotations are likely to be flawed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotator errors (17% of total)",
"sec_num": "6.1"
},
{
"text": "83% of the total errors were due to errors from the model. We have separated these into clear errors and edge cases. Clear errors are where the model has made an error that is easily identified by humans (accounting for 29% of all errors). Edgecases are where the misclassified content contains some ambiguity and the model misclassification has some merit (accounting for 54% of all errors).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Prediction errors (83% of total)",
"sec_num": "6.2"
},
{
"text": "Clear error (Lexical similarity), 16% In several cases the misclassified tweets were clearly assigned to the wrong class. This suggests possible overfitting as the tweets were often lexically similar to tweets which did belong in the category (e.g., they contained phrases such as 'Made in China' and were mistaken for Hostility). This is most likely a learned over-sensitivity and could only be addressed through using a far larger dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Prediction errors (83% of total)",
"sec_num": "6.2"
},
{
"text": "Clear error (Target confusion), 13% The model sometimes identified tweets which were not East Asian relevant as Criticism, Hostility, or Discussion of East Asian prejudice. Aside from this, the classifications were correct, i.e. the tweets expressed hostility against another identity, such as women or gay people. Furthermore, in many cases, a relevant East Asian entity (e.g. China) was usually referred to but was not the object of the tweet, creating a mixed signal for the model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Prediction errors (83% of total)",
"sec_num": "6.2"
},
{
"text": "Edge case (Hostility vs. Criticism), 29% Misclassifying Hostility as Criticism (and vice versa) was the largest source of error. The model particularly struggled with cases where criticism was framed normatively or expressed with nuanced linguistic expressions, e.g., \"gee, china was lying to us. what a bloody shock\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Prediction errors (83% of total)",
"sec_num": "6.2"
},
{
"text": "Edge case (Non-related vs. Discussion of EA prejudice), 9% The model misclassified Nonrelated as Discussion of East Asian prejudice in several tweets. These were usually cases where the virus was named and discussed but prejudice was not discussed explicitly, e.g. \"corona just shows why you should blame all our problems with China on Trump\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Prediction errors (83% of total)",
"sec_num": "6.2"
},
{
"text": "In our taxonomy, annotators could assign each tweet to only one primary category. However, in some cases this was problematic and the model identified a co-present category rather than the primary category which had been annotated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Edge case (Co-present primary categories), 5%",
"sec_num": null
},
{
"text": "Edge case (Ambiguous content), 11% The model often misclassified content which was ambiguous. This is content where the true meaning is not immediately discernible to a human without careful analysis. For instance, positively framed criticism, e.g. \"so glad that china official finally admits the HASHTAGEASTASIA+VIRUS outbreaks\". In some tweets, complex forms of expression were used, such as innuendo or sarcasm, e.g. \"I think we owe you china, please accept our apologies to bring some virus into your great country\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Edge case (Co-present primary categories), 5%",
"sec_num": null
},
{
"text": "Classifying social media data is notoriously difficult. There are many different sources of error, each of which, in turn, require different remedies. Annotator errors, for example, illustrate the need for robust annotation processes and providing more support and training when taxonomies are applied. Removing such errors entirely is unlikely, but the number of obvious misclassifications could be minimized.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Addressing Classification Errors",
"sec_num": "6.3"
},
{
"text": "Machine learning errors are where the bulk of the errors fall (83%). Edge cases are a particularly difficult type of content for classification. They can be expected in any taxonomy that draws distinct lines between complex and non-mutually exclusive concepts, such as Hostility and Criticism. Nonetheless, larger and more balanced datasets (with more instances of content in each category) would help in reducing this source of error. Equally, the frequency of non-edge case machine learning errors (i.e. cases where the model made a very obvious mistake) could be addressed by larger datasets as well as more advanced machine learning architectures.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Addressing Classification Errors",
"sec_num": "6.3"
},
{
"text": "We analysed the 20,000 tweet training dataset to understand which hashtags were associated with which primary categories. For each category, we filtered the data so only hashtags which appeared in at least ten tweets assigned to that category were included. Then, we ranked the hashtags by the percentage of their uses which were in the primary category-our goal being to understand which hashtags are most closely associated with that category. For brevity, only the twenty hashtags most closely associated with the Hostility category are shown in Table 5 . This analysis is only possible because all hashtags were replaced in the tweets which were presented to annotators (either with a generic hashtag token or thematic replacement), letting us conduct meaningful analysis of the cooccurrence of hashtags with the annotated primary categories.",
"cite_spans": [],
"ref_spans": [
{
"start": 549,
"end": 556,
"text": "Table 5",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Hashtag analysis",
"sec_num": "7"
},
{
"text": "A small number of hashtags are highly likely to only appear in tweets that express Hostility against East Asian entities. These hashtags could be used to filter for prejudiced discourses online and, in some cases, their uses may intrinsically indicate prejudice. Surprisingly, many seemingly hostile hashtags against East Asia, such as #fuckchina and #blamechina are not always associated with hostile tweets (Hostility accounts for 67.5% and 60.5% of their total use, respectively). This shows the importance of having a purpose-built machine learning classifier for detecting East Asian hostility, rather than relying on hashtags and keywords alone.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Hashtag analysis",
"sec_num": "7"
},
{
"text": "Prejudice of all forms is a deeply concerning problem during COVID-19, reflecting the huge social costs and difficulties that the pandemic has inflicted. In this paper we have reported on development of several research artefacts that we hope will enable further research into East Asian prejudice, including a trained classifier, a training dataset (20,000 entries), annotations dataset (40,000 entries), 1,000 annotated hashtags, a list of hashtag replacements, a list of hashtags associated with hostility against East Asia, and the full codebook (with extensive guidelines, information and examples).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "One concern with any model is that it will not generalize to new settings and platforms, limiting its utility for real-world applications. Our hashtag replacement method was adopted to increase the generalizability of the final model, maximizing the likelihood that it can be applied to new contextsas it would pick up on the semantic features of hostile tweets rather than specific tokens. Nonetheless, we caution that any re-use of the model should be accompanied by additional testing to understand its performance on different data. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "We query for: #chinavirus, #wuhan, #wuhanvirus, #chinavirusoutbreak, #wuhancoronavirus, #wuhaninfluenza, #wuhansars, #chinacoronavirus, #wuhan2020, #chinaflu, #wuhanquarantine, #chinesepneumonia, #coronachina and #wohan.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This work was supported by Wave 1 of The UKRI Strategic Priorities Fund under the EPSRC Grant EP/T001569/1, particularly the \"Criminal Justice System\" theme within that grant, and The Alan Turing Institute. Scott Hale was supported in part by the Volkswagen Foundation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Hidden from public view? Racism against the UK Chinese population",
"authors": [
{
"first": "Sue",
"middle": [],
"last": "Adamson",
"suffix": ""
},
{
"first": "Bankole",
"middle": [],
"last": "Cole",
"suffix": ""
},
{
"first": "Gary",
"middle": [],
"last": "Craig",
"suffix": ""
},
{
"first": "Basharat",
"middle": [],
"last": "Hussain",
"suffix": ""
},
{
"first": "Luana",
"middle": [],
"last": "Smith",
"suffix": ""
},
{
"first": "Ian",
"middle": [],
"last": "Law",
"suffix": ""
},
{
"first": "Carmen",
"middle": [],
"last": "Lau",
"suffix": ""
},
{
"first": "Chak-Kwan",
"middle": [],
"last": "Chan",
"suffix": ""
},
{
"first": "Tom",
"middle": [],
"last": "Cheung",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sue Adamson, Bankole Cole, Gary Craig, Basharat Hussain, Luana Smith, Ian Law, Carmen Lau, Chak- Kwan Chan, and Tom Cheung. 2009. Hidden from public view? Racism against the UK Chinese popu- lation. Sheffield Hallam University.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Sinophobia: anxiety, violence, and the making of Mongolian identity",
"authors": [
{
"first": "Franck",
"middle": [],
"last": "Bill\u00e9",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franck Bill\u00e9. 2015. Sinophobia: anxiety, violence, and the making of Mongolian identity. University of Hawaii Press.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "CONAN -COunter NArratives through Nichesourcing: a Multilingual Dataset of Responses to Fight Online Hate Speech",
"authors": [
{
"first": "Yi-Ling",
"middle": [],
"last": "Chung",
"suffix": ""
},
{
"first": "Elizaveta",
"middle": [],
"last": "Kuzmenko",
"suffix": ""
},
{
"first": "Marco",
"middle": [],
"last": "Serra Sinem Tekiroglu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Guerini",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the ACL",
"volume": "",
"issue": "",
"pages": "2819--2829",
"other_ids": {
"DOI": [
"10.18653/v1/p19-1271"
]
},
"num": null,
"urls": [],
"raw_text": "Yi-Ling Chung, Elizaveta Kuzmenko, Serra Sinem Tekiroglu, and Marco Guerini. 2019. CONAN - COunter NArratives through Nichesourcing: a Mul- tilingual Dataset of Responses to Fight Online Hate Speech. In Proceedings of the 57th Annual Meeting of the ACL, pages 2819-2829.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "The COVID-19 Social Media Infodemic",
"authors": [
{
"first": "Matteo",
"middle": [],
"last": "Cinelli",
"suffix": ""
},
{
"first": "Walter",
"middle": [],
"last": "Quattrociocchi",
"suffix": ""
},
{
"first": "Alessandro",
"middle": [],
"last": "Galeazzi",
"suffix": ""
},
{
"first": "Carlo",
"middle": [
"Michele"
],
"last": "Valensise",
"suffix": ""
},
{
"first": "Emanuele",
"middle": [],
"last": "Brugnoli",
"suffix": ""
},
{
"first": "Ana",
"middle": [
"Lucia"
],
"last": "Schmidt",
"suffix": ""
},
{
"first": "Paola",
"middle": [],
"last": "Zola",
"suffix": ""
},
{
"first": "Fabiana",
"middle": [],
"last": "Zollo",
"suffix": ""
},
{
"first": "Antonio",
"middle": [],
"last": "Scala",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2003.05004"
]
},
"num": null,
"urls": [],
"raw_text": "Matteo Cinelli, Walter Quattrociocchi, Alessandro Galeazzi, Carlo Michele Valensise, Emanuele Brug- noli, Ana Lucia Schmidt, Paola Zola, Fabiana Zollo, and Antonio Scala. 2020. The COVID-19 Social Media Infodemic. arXiv:2003.05004.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Grounded Theory Research: Procedures, Canons and Evaluative Criteria",
"authors": [
{
"first": "Juliet",
"middle": [],
"last": "Corbin",
"suffix": ""
},
{
"first": "Anselm",
"middle": [],
"last": "Strauss",
"suffix": ""
}
],
"year": 1990,
"venue": "Qualitative Research",
"volume": "13",
"issue": "1",
"pages": "3--21",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Juliet Corbin and Anselm Strauss. 1990. Grounded Theory Research: Procedures, Canons and Evalua- tive Criteria. Qualitative Research, 13(1):3-21.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Why content moderators should be key workers protecting social media as critical infrastructure during covid-19. The Alan Turing Institute",
"authors": [
{
"first": "Josh",
"middle": [],
"last": "Cowls",
"suffix": ""
},
{
"first": "Bertie",
"middle": [],
"last": "Vidgen",
"suffix": ""
},
{
"first": "Helen",
"middle": [],
"last": "Margetts",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Josh Cowls, Bertie Vidgen, and Helen Margetts. 2020. Why content moderators should be key workers pro- tecting social media as critical infrastructure during covid-19. The Alan Turing Institute.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Racial Bias in Hate Speech and Abusive Language Detection Datasets",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "Davidson",
"suffix": ""
},
{
"first": "Debasmita",
"middle": [],
"last": "Bhattacharya",
"suffix": ""
},
{
"first": "Ingmar",
"middle": [],
"last": "Weber",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the ACL",
"volume": "",
"issue": "",
"pages": "1--11",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Thomas Davidson, Debasmita Bhattacharya, and Ing- mar Weber. 2019. Racial Bias in Hate Speech and Abusive Language Detection Datasets. In Proceed- ings of the 57th Annual Meeting of the ACL, pages 1-11.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Automated Hate Speech Detection and the Problem of Offensive Language",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "Davidson",
"suffix": ""
},
{
"first": "Dana",
"middle": [],
"last": "Warmsley",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Macy",
"suffix": ""
},
{
"first": "Ingmar",
"middle": [],
"last": "Weber",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 11th ICWSM",
"volume": "",
"issue": "",
"pages": "1--4",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. Automated Hate Speech Detection and the Problem of Offensive Language. In Proceedings of the 11th ICWSM, pages 1-4.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"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": 2019,
"venue": "Proceedings of NAACL-HLT 2019",
"volume": "",
"issue": "",
"pages": "4171--4186",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understand- ing. In Proceedings of NAACL-HLT 2019, page 4171-4186.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Canada's top doctor calls out 'racism and stigmatizing comments' over coronavirus",
"authors": [
{
"first": "Ryan",
"middle": [],
"last": "Flanagan",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ryan Flanagan. 2020. Canada's top doctor calls out 'racism and stigmatizing comments' over coro- navirus. CTVNews.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Counterfactual fairness in text classification through robustness",
"authors": [
{
"first": "Sahaj",
"middle": [],
"last": "Garg",
"suffix": ""
},
{
"first": "Ankur",
"middle": [],
"last": "Taly",
"suffix": ""
},
{
"first": "Vincent",
"middle": [],
"last": "Perot",
"suffix": ""
},
{
"first": "Ed",
"middle": [
"H"
],
"last": "Chi",
"suffix": ""
},
{
"first": "Nicole",
"middle": [],
"last": "Limtiaco",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Beutel",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society",
"volume": "",
"issue": "",
"pages": "219--226",
"other_ids": {
"DOI": [
"10.1145/3306618.3317950"
]
},
"num": null,
"urls": [],
"raw_text": "Sahaj Garg, Ankur Taly, Vincent Perot, Ed H. Chi, Nicole Limtiaco, and Alex Beutel. 2019. Counter- factual fairness in text classification through robust- ness. Proceedings of the 2019 AAAI/ACM Confer- ence on AI, Ethics, and Society, pages 219-226.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Hate Speech Dataset from a White Supremacy Forum",
"authors": [
{
"first": "Ona",
"middle": [],
"last": "De Gibert",
"suffix": ""
},
{
"first": "Naiara",
"middle": [],
"last": "Perez",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Second Workshop on Abusive Language Online (ACL)",
"volume": "",
"issue": "",
"pages": "11--20",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ona de Gibert, Naiara Perez, Aitor Garc\u00eda-Pablos, and Montse Cuadros. 2018. Hate Speech Dataset from a White Supremacy Forum. In Proceedings of the Sec- ond Workshop on Abusive Language Online (ACL), pages 11-20.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Yellow Peril Revisited: Impact of SARS on the Chinese and Southeast Asian Canadian Communities",
"authors": [
{
"first": "Tam",
"middle": [],
"last": "Goossen",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Guan",
"suffix": ""
},
{
"first": "Ito",
"middle": [],
"last": "Peng",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tam Goossen, Jian Guan, and Ito Peng. 2004. Yellow Peril Revisited: Impact of SARS on the Chinese and Southeast Asian Canadian Communities June, 2004",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Project Coordinator and Author: Carrianne Leung",
"authors": [],
"year": null,
"venue": "Resources for Feminist Research",
"volume": "33",
"issue": "1-2",
"pages": "135--150",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Project Coordinator and Author: Carrianne Leung. Resources for Feminist Research, 33(1-2):135-150.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "East Asian student assaulted in 'racist' coronavirus attack in London",
"authors": [
{
"first": "Jack",
"middle": [],
"last": "Guy",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jack Guy. 2020. East Asian student assaulted in 'racist' coronavirus attack in London. CNN.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Differentiating Islamophobia: Introducing a New Scale to Measure Islamoprejudice and Secular Islam Critique",
"authors": [
{
"first": "Roland",
"middle": [],
"last": "Imhoff",
"suffix": ""
},
{
"first": "Julia",
"middle": [],
"last": "Recker",
"suffix": ""
}
],
"year": 2012,
"venue": "Political Psychology",
"volume": "33",
"issue": "6",
"pages": "811--824",
"other_ids": {
"DOI": [
"10.1111/j.1467-9221.2012.00911.x"
]
},
"num": null,
"urls": [],
"raw_text": "Roland Imhoff and Julia Recker. 2012. Differentiating Islamophobia: Introducing a New Scale to Measure Islamoprejudice and Secular Islam Critique. Politi- cal Psychology, 33(6):811-824.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Dangerous Speech and Dangerous Ideology: An Integrated Model for Monitoring and Prevention",
"authors": [
{
"first": "Jonathan",
"middle": [],
"last": "Leader",
"suffix": ""
},
{
"first": "Maynard",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "Susan",
"middle": [],
"last": "Benesch",
"suffix": ""
}
],
"year": 2016,
"venue": "Genocide Studies and Prevention",
"volume": "9",
"issue": "3",
"pages": "70--95",
"other_ids": {
"DOI": [
"10.5038/1911-9933.9.3.1317"
]
},
"num": null,
"urls": [],
"raw_text": "Jonathan Leader Maynard and Susan Benesch. 2016. Dangerous Speech and Dangerous Ideology: An Integrated Model for Monitoring and Prevention. Genocide Studies and Prevention, 9(3):70-95.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Roberta: A robustly optimized bert pretraining approach",
"authors": [
{
"first": "Yinhan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Myle",
"middle": [],
"last": "Ott",
"suffix": ""
},
{
"first": "Naman",
"middle": [],
"last": "Goyal",
"suffix": ""
},
{
"first": "Jingfei",
"middle": [],
"last": "Du",
"suffix": ""
},
{
"first": "Mandar",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Danqi",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Lewis",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "Veselin",
"middle": [],
"last": "Stoyanov",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1907.11692"
]
},
"num": null,
"urls": [],
"raw_text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining ap- proach. arXiv:1907.11692.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Coronavirus prompts 'hysterical, shameful' Sinophobia in Italy",
"authors": [
{
"first": "Yuebai",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yuebai Liu. 2020. Coronavirus prompts 'hysterical, shameful' Sinophobia in Italy. Al Jazeera.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Decoupled weight decay regularization",
"authors": [
{
"first": "Ilya",
"middle": [],
"last": "Loshchilov",
"suffix": ""
},
{
"first": "Frank",
"middle": [],
"last": "Hutter",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of ICLR 2019",
"volume": "",
"issue": "",
"pages": "1--18",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In Proceedings of ICLR 2019, pages 1-18.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Dehumanizing metaphors in UK immigrant debates in press and online media",
"authors": [
{
"first": "Andreas",
"middle": [],
"last": "Musolff",
"suffix": ""
}
],
"year": 2015,
"venue": "Journal of Language Aggression and Conflict",
"volume": "3",
"issue": "1",
"pages": "41--56",
"other_ids": {
"DOI": [
"10.1075/jlac.3.1.02mus"
]
},
"num": null,
"urls": [],
"raw_text": "Andreas Musolff. 2015. Dehumanizing metaphors in UK immigrant debates in press and online me- dia. Journal of Language Aggression and Conflict, 3(1):41-56.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Poll: Asian-Americans See Individuals' Prejudice As Big Discrimination Problem",
"authors": [
{
"first": "Joe",
"middle": [],
"last": "Neel",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Joe Neel. 2017. Poll: Asian-Americans See Indi- viduals' Prejudice As Big Discrimination Problem. NPR.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "COVID-19: Bogus claims fool Britons",
"authors": [
{
"first": "Eir",
"middle": [],
"last": "Nolsoe",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eir Nolsoe. 2020. COVID-19: Bogus claims fool Britons. YouGov.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "The Risk of Racial Bias in Hate Speech Detection",
"authors": [
{
"first": "Maarten",
"middle": [],
"last": "Sap",
"suffix": ""
},
{
"first": "Dallas",
"middle": [],
"last": "Card",
"suffix": ""
},
{
"first": "Saadia",
"middle": [],
"last": "Gabriel",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Noah",
"suffix": ""
},
{
"first": "Paul G",
"middle": [],
"last": "Smith",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Allen",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the ACL",
"volume": "",
"issue": "",
"pages": "1668--1678",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, Noah A Smith, and Paul G Allen. 2019. The Risk of Racial Bias in Hate Speech Detection. In Proceed- ings of the 57th Annual Meeting of the ACL, pages 1668-1678.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Go eat a bat",
"authors": [
{
"first": "Leonard",
"middle": [],
"last": "Schild",
"suffix": ""
},
{
"first": "Chen",
"middle": [],
"last": "Ling",
"suffix": ""
},
{
"first": "Jeremy",
"middle": [],
"last": "Blackburn",
"suffix": ""
},
{
"first": "Gianluca",
"middle": [],
"last": "Stringhini",
"suffix": ""
},
{
"first": "Yang",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Savvas",
"middle": [],
"last": "Zannettou",
"suffix": ""
}
],
"year": 2020,
"venue": "An Early Look on the Emergence of Sinophobic Behavior on Web Communities in the Face of COVID-19",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2004.04046"
]
},
"num": null,
"urls": [],
"raw_text": "Leonard Schild, Chen Ling, Jeremy Blackburn, Gi- anluca Stringhini, Yang Zhang, and Savvas Zan- nettou. 2020. \"Go eat a bat, Chang!\": An Early Look on the Emergence of Sinophobic Behavior on Web Communities in the Face of COVID-19. arXiv:2004.04046.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "A survey on hate speech detection using natural language processing",
"authors": [
{
"first": "Anna",
"middle": [],
"last": "Schmidt",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Wiegand",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media. Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.18653/v1/W17-1101"
]
},
"num": null,
"urls": [],
"raw_text": "Anna Schmidt and Michael Wiegand. 2017. A survey on hate speech detection using natural language pro- cessing. In Proceedings of the Fifth International Workshop on Natural Language Processing for So- cial Media. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "2020. U.N. asks world to fight virusspawned discrimination",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Shields",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Shields. 2020. U.N. asks world to fight virus- spawned discrimination. Reuters.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Coronavirus Brings 'Sinophobia' to Africa",
"authors": [
{
"first": "Salem",
"middle": [],
"last": "Solomon",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Salem Solomon. 2020. Coronavirus Brings 'Sinopho- bia' to Africa. VOA News.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Abusive and Offensive Online Communications: A scoping report",
"authors": [],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "The Law Commission. 2018. Abusive and Offensive Online Communications: A scoping report. The Law Commission.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Covid-19 has caused a major spike in anti-Chinese and anti-Semitic hate speech. The New Statesman",
"authors": [],
"year": null,
"venue": "The New Statesman. 2020",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "The New Statesman. 2020. Covid-19 has caused a major spike in anti-Chinese and anti-Semitic hate speech. The New Statesman.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Pro-Russian Biases in Anti-Chinese Tweets about the Novel Coronavirus",
"authors": [
{
"first": "Autumn",
"middle": [],
"last": "Toney",
"suffix": ""
},
{
"first": "Akshat",
"middle": [],
"last": "Pandey",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Guo",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Broniatowski",
"suffix": ""
},
{
"first": "Aylin",
"middle": [],
"last": "Caliskan",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2004.08726"
]
},
"num": null,
"urls": [],
"raw_text": "Autumn Toney, Akshat Pandey, Wei Guo, David Broni- atowski, and Aylin Caliskan. 2020. Pro-Russian Bi- ases in Anti-Chinese Tweets about the Novel Coron- avirus. arXiv:2004.08726.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Quarantining online hate speech: technical and ethical perspectives",
"authors": [
{
"first": "Stefanie",
"middle": [],
"last": "Ullmann",
"suffix": ""
},
{
"first": "Marcus",
"middle": [],
"last": "Tomalin",
"suffix": ""
}
],
"year": 2020,
"venue": "Ethics and Information Technology",
"volume": "22",
"issue": "1",
"pages": "69--80",
"other_ids": {
"DOI": [
"10.1007/s10676-019-09516-z"
]
},
"num": null,
"urls": [],
"raw_text": "Stefanie Ullmann and Marcus Tomalin. 2020. Quar- antining online hate speech: technical and ethical perspectives. Ethics and Information Technology, 22(1):69-80.",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "Attention is all you need",
"authors": [
{
"first": "Ashish",
"middle": [],
"last": "Vaswani",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Niki",
"middle": [],
"last": "Parmar",
"suffix": ""
},
{
"first": "Jakob",
"middle": [],
"last": "Uszkoreit",
"suffix": ""
},
{
"first": "Llion",
"middle": [],
"last": "Jones",
"suffix": ""
},
{
"first": "Aidan",
"middle": [
"N"
],
"last": "Gomez",
"suffix": ""
},
{
"first": "Lukasz",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "Illia",
"middle": [],
"last": "Polosukhin",
"suffix": ""
}
],
"year": 2017,
"venue": "31st Conference on Neural Information Processing Systems (NIPS 2017)",
"volume": "",
"issue": "",
"pages": "1--11",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In 31st Conference on Neural Information Processing Systems (NIPS 2017), pages 1-11.",
"links": null
},
"BIBREF33": {
"ref_id": "b33",
"title": "Hate multiverse spreads malicious covid-19 content online beyond individual platform control",
"authors": [
{
"first": "N",
"middle": [],
"last": "Vel\u00e1squez",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Leahy",
"suffix": ""
},
{
"first": "N",
"middle": [
"Johnson"
],
"last": "Restrepo",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Lupu",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Sear",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Gabriel",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Jha",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Goldberg",
"suffix": ""
},
{
"first": "N",
"middle": [
"F"
],
"last": "Johnson",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2004.00673"
]
},
"num": null,
"urls": [],
"raw_text": "N. Vel\u00e1squez, R. Leahy, N. Johnson Restrepo, Y. Lupu, R. Sear, N. Gabriel, O. Jha, B. Goldberg, and N.F. Johnson. 2020. Hate multiverse spreads malicious covid-19 content online beyond individual platform control. arXiv:2004.00673.",
"links": null
},
"BIBREF34": {
"ref_id": "b34",
"title": "Directions in Abusive Language Training Data: Garbage In, Garbage Out",
"authors": [
{
"first": "Bertie",
"middle": [],
"last": "Vidgen",
"suffix": ""
},
{
"first": "Leon",
"middle": [],
"last": "Derczynski",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "1--26",
"other_ids": {
"arXiv": [
"arXiv:2004.01670"
]
},
"num": null,
"urls": [],
"raw_text": "Bertie Vidgen and Leon Derczynski. 2020. Directions in Abusive Language Training Data: Garbage In, Garbage Out. arXiv:2004.01670, pages 1-26.",
"links": null
},
"BIBREF35": {
"ref_id": "b35",
"title": "Challenges and frontiers in abusive content detection",
"authors": [
{
"first": "Bertie",
"middle": [],
"last": "Vidgen",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Harris",
"suffix": ""
},
{
"first": "Dong",
"middle": [],
"last": "Nguyen",
"suffix": ""
},
{
"first": "Rebekah",
"middle": [],
"last": "Tromble",
"suffix": ""
},
{
"first": "Scott",
"middle": [],
"last": "Hale",
"suffix": ""
},
{
"first": "Helen",
"middle": [],
"last": "Margetts",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the Third Workshop on Abusive Language Online (ACL)",
"volume": "",
"issue": "",
"pages": "80--93",
"other_ids": {
"DOI": [
"10.18653/v1/W19-3509"
]
},
"num": null,
"urls": [],
"raw_text": "Bertie Vidgen, Alex Harris, Dong Nguyen, Rebekah Tromble, Scott Hale, and Helen Margetts. 2019. Challenges and frontiers in abusive content detec- tion. In Proceedings of the Third Workshop on Abu- sive Language Online (ACL), pages 80-93.",
"links": null
},
"BIBREF36": {
"ref_id": "b36",
"title": "Wuhan Virus Boosts Indonesian Anti-Chinese Conspiracies",
"authors": [
{
"first": "Kate",
"middle": [],
"last": "Walton",
"suffix": ""
}
],
"year": 2020,
"venue": "Foreign Policy",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kate Walton. 2020. Wuhan Virus Boosts Indonesian Anti-Chinese Conspiracies. Foreign Policy.",
"links": null
},
"BIBREF37": {
"ref_id": "b37",
"title": "Understanding Abuse: A Typology of Abusive Language Detection Subtasks",
"authors": [
{
"first": "Zeerak",
"middle": [],
"last": "Waseem",
"suffix": ""
},
{
"first": "Thomas",
"middle": [],
"last": "Davidson",
"suffix": ""
},
{
"first": "Dana",
"middle": [],
"last": "Warmsley",
"suffix": ""
},
{
"first": "Ingmar",
"middle": [],
"last": "Weber",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the First Workshop on Abusive Language Online",
"volume": "",
"issue": "",
"pages": "78--84",
"other_ids": {
"DOI": [
"10.1080/17421770903114687"
]
},
"num": null,
"urls": [],
"raw_text": "Zeerak Waseem, Thomas Davidson, Dana Warmsley, and Ingmar Weber. 2017. Understanding Abuse: A Typology of Abusive Language Detection Subtasks. In Proceedings of the First Workshop on Abusive Language Online, pages 78-84.",
"links": null
},
"BIBREF38": {
"ref_id": "b38",
"title": "Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter",
"authors": [
{
"first": "Zeerak",
"middle": [],
"last": "Waseem",
"suffix": ""
},
{
"first": "Dirk",
"middle": [],
"last": "Hovy",
"suffix": ""
}
],
"year": 2016,
"venue": "NAACL-HLT",
"volume": "",
"issue": "",
"pages": "88--93",
"other_ids": {
"DOI": [
"10.18653/v1/n16-2013"
]
},
"num": null,
"urls": [],
"raw_text": "Zeerak Waseem and Dirk Hovy. 2016. Hateful Sym- bols or Hateful People? Predictive Features for Hate Speech Detection on Twitter. In NAACL-HLT, pages 88-93.",
"links": null
},
"BIBREF39": {
"ref_id": "b39",
"title": "Manual on Hate Speech",
"authors": [
{
"first": "Anne",
"middle": [],
"last": "Weber",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Anne Weber. 2009. Manual on Hate Speech. Council of Europe.",
"links": null
},
"BIBREF40": {
"ref_id": "b40",
"title": "2020. Sinophobia: How a virus reveals the many ways China is feared",
"authors": [
{
"first": "Tessa",
"middle": [],
"last": "Wong",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tessa Wong. 2020. Sinophobia: How a virus reveals the many ways China is feared. BBC News.",
"links": null
},
"BIBREF41": {
"ref_id": "b41",
"title": "Pinning coronavirus on how chinese people eat plays into racist assumptions",
"authors": [
{
"first": "",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhang. 2020. Pinning coronavirus on how chinese peo- ple eat plays into racist assumptions.",
"links": null
},
"BIBREF42": {
"ref_id": "b42",
"title": "Racism is a virus: Anti-asian hate and counterhate in social media during the covid-19 crisis",
"authors": [
{
"first": "Caleb",
"middle": [],
"last": "Ziems",
"suffix": ""
},
{
"first": "Bing",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Sandeep",
"middle": [],
"last": "Soni",
"suffix": ""
},
{
"first": "Srijan",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2005.12423"
]
},
"num": null,
"urls": [],
"raw_text": "Caleb Ziems, Bing He, Sandeep Soni, and Srijan Ku- mar. 2020. Racism is a virus: Anti-asian hate and counterhate in social media during the covid-19 cri- sis. arXiv:2005.12423.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"type_str": "figure",
"text": "Confusion matrix for RoBERTa classifications on the test set.",
"num": null
},
"FIGREF1": {
"uris": null,
"type_str": "figure",
"text": "Sources of classification error.",
"num": null
},
"TABREF0": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table><tr><td colspan=\"2\">Theme</td><td/><td>Number of Entries</td><td>Percentage</td></tr><tr><td colspan=\"3\">COVID-19 relevant / Both said No</td><td>2,940</td><td>14.7%</td></tr><tr><td colspan=\"3\">COVID-19 relevant / Both said Yes</td><td>12,255</td><td>61.3%</td></tr><tr><td colspan=\"3\">COVID-19 relevant / Disagreement</td><td>4,805</td><td>24.0%</td></tr><tr><td colspan=\"3\">East Asian relevant / Both said No</td><td>6,593</td><td>33.0%</td></tr><tr><td colspan=\"3\">East Asian relevant / Both said Yes</td><td>9,790</td><td>49.0%</td></tr><tr><td colspan=\"3\">East Asian relevant / Disagreement</td><td>3,617</td><td>18.0%</td></tr><tr><td>Category</td><td>Number of Entries</td><td>Percentage</td></tr><tr><td>Hostility</td><td>3,898</td><td>19.5%</td></tr><tr><td>Criticism</td><td>1,433</td><td>7.2%</td></tr><tr><td>Counter speech</td><td>116</td><td>0.6%</td></tr><tr><td>Discussion of EAP</td><td>1,029</td><td>5.1%</td></tr><tr><td>Non-related</td><td>13,528</td><td>67.6%</td></tr><tr><td>TOTAL</td><td>20,000</td><td>100%</td></tr></table>",
"text": "Prevalence of themes in the dataset."
},
"TABREF1": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table/>",
"text": ""
},
"TABREF3": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table/>",
"text": ""
},
"TABREF5": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table/>",
"text": "Classification performance of models on the test set."
},
"TABREF7": {
"type_str": "table",
"num": null,
"html": null,
"content": "<table/>",
"text": "Hashtags in Hostile tweets."
}
}
}
} |