File size: 142,855 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 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T06:07:52.044663Z"
},
"title": "Disentangling Document Topic and Author Gender in Multiple Languages: Lessons for Adversarial Debiasing",
"authors": [
{
"first": "Erenay",
"middle": [],
"last": "Dayanik",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Stuttgart",
"location": {
"settlement": "Stuttgart",
"country": "Germany"
}
},
"email": "[email protected]"
},
{
"first": "Sebastian",
"middle": [],
"last": "Pad\u00f3",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Stuttgart",
"location": {
"settlement": "Stuttgart",
"country": "Germany"
}
},
"email": "[email protected]"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Text classification is a central tool in NLP, including social media analysis. However, when the target classes are strongly correlated with other textual attributes, text classification models can pick up \"wrong\" features, leading to bad generalization and biases. In social media analysis, this problem surfaces for demographic user classes such as language, topic, or gender, which influence how an author writes a text to a substantial extent. Adversarial training has been claimed to mitigate this problem, but a thorough evaluation is missing. In this paper, we experiment with text classification of the correlated attributes of document topic and author gender, using a novel multilingual parallel corpus of TED talk transcripts. Our findings are: (a) individual classifiers for topic and author gender are indeed biased; (b) debiasing with adversarial training works for topic, but breaks down for author gender; (c) gender debiasing results differ across languages. We interpret the result in terms of feature space overlap, highlighting the role of linguistic surface realization of the target classes.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "Text classification is a central tool in NLP, including social media analysis. However, when the target classes are strongly correlated with other textual attributes, text classification models can pick up \"wrong\" features, leading to bad generalization and biases. In social media analysis, this problem surfaces for demographic user classes such as language, topic, or gender, which influence how an author writes a text to a substantial extent. Adversarial training has been claimed to mitigate this problem, but a thorough evaluation is missing. In this paper, we experiment with text classification of the correlated attributes of document topic and author gender, using a novel multilingual parallel corpus of TED talk transcripts. Our findings are: (a) individual classifiers for topic and author gender are indeed biased; (b) debiasing with adversarial training works for topic, but breaks down for author gender; (c) gender debiasing results differ across languages. We interpret the result in terms of feature space overlap, highlighting the role of linguistic surface realization of the target classes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Natural language processing, and machine learning more generally, has recently received a significant deal of criticism because of the frequent presence of bias in the predictions, where we define bias as a systematic difference in system performance on one set of instances compared to another. Such biases have been identified in NLP tasks such as word representation (Bolukbasi et al., 2016) , textual inference (Rudinger et al., 2017) , coreference resolution (Zhao et al., 2018 ), text classification (Dixon et al., 2018) and emotion intensity prediction (Kiritchenko and Mohammad, 2018) .",
"cite_spans": [
{
"start": 370,
"end": 394,
"text": "(Bolukbasi et al., 2016)",
"ref_id": "BIBREF1"
},
{
"start": 415,
"end": 438,
"text": "(Rudinger et al., 2017)",
"ref_id": "BIBREF29"
},
{
"start": 464,
"end": 482,
"text": "(Zhao et al., 2018",
"ref_id": "BIBREF41"
},
{
"start": 506,
"end": 526,
"text": "(Dixon et al., 2018)",
"ref_id": "BIBREF8"
},
{
"start": 560,
"end": 592,
"text": "(Kiritchenko and Mohammad, 2018)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In text classification tasks, a principal source of such biases are demographic attributes of authors, such as gender, age, or race 1 . The reason is that these attributes shape speakers' language use substantially (Hovy, 2015) . NLP models are not only able in principle to pick up such cues, as studies on modeling demographic attributes show (Koppel et al., 2004) , but they actually have a motivation to do so whenever some demographic attribute is strongly correlated with the model's classification target and therefore supports its recognition. As an example, in social psychology, Gross et al. (1997) report that elderly people experience and express their emotions less intensely than younger people. Therefore, in a corpus of emotional expressions across age groups, it is reasonable for a model that predicts emotion intensity to look out for linguistic cues regarding author age, even if these cues are not really related to emotion intensity per se, such as typical markers of youth language (\"rad\", \"fam\", \"FTW\", etc.).",
"cite_spans": [
{
"start": 215,
"end": 227,
"text": "(Hovy, 2015)",
"ref_id": "BIBREF17"
},
{
"start": 345,
"end": 366,
"text": "(Koppel et al., 2004)",
"ref_id": "BIBREF21"
},
{
"start": 589,
"end": 608,
"text": "Gross et al. (1997)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This focus is arguably problematic, though, since it can give rise to a form of age bias -namely, overestimating emotion intensity for documents exhibiting youth language. More generally, the bias-inducing role of demographic attributes is dangerous for studies that use texts from a multitude of authors -often gathered from social mediato draw inferences about the authors (Sobkowicz et al., 2012; Cheng et al., 2015) . In such studies, demographic biases can lead to erroneous causal attributions (as our case will illustrate).",
"cite_spans": [
{
"start": 375,
"end": 399,
"text": "(Sobkowicz et al., 2012;",
"ref_id": "BIBREF31"
},
{
"start": 400,
"end": 419,
"text": "Cheng et al., 2015)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To counteract the presence of biases in NLP, researchers have devised debiasing methods. Due to its general applicability and high effectiveness, adversarial debiasing has become one of the most widely used methods for bias mitigation (Elazar and Goldberg, 2018; Zhang et al., 2018; Arduini et al., 2020) . Unfortunately, these advances are not accompanied by an analysis of the prerequisites that need to be satisfied for adversarial training to perform successfully. It has been established empirically is that adversarial training works well for many cases in NLP; nevertheless, we demonstrate that there are relatively simple setups where it can fail. We analyze what factors contribute to the failure.",
"cite_spans": [
{
"start": 235,
"end": 262,
"text": "(Elazar and Goldberg, 2018;",
"ref_id": "BIBREF10"
},
{
"start": 263,
"end": 282,
"text": "Zhang et al., 2018;",
"ref_id": "BIBREF38"
},
{
"start": 283,
"end": 304,
"text": "Arduini et al., 2020)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Concretely, we consider the correlated attributes of document topic (scientific / non-scientific) and author gender on a self-collected multilingual corpus of TED talk transcripts in French, German, Spanish, and Turkish. This setup enables us to observe the interplay between linguistic properties and adversarial debiasing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our investigation proceeds in three steps. First, we train independent classifiers for each attribute and evaluate them with regard to overall performance and with regard to the bias they exhibit. In the second step, we apply adversarial debiasing to the predicting of each attribute with respect to the other, and re-evaluate the debiased models. Finally, in the third step we discuss the differences observed in the previous step: (a), both document topic and author gender can be classified reasonably well by independent classifiers, but exhibit considerable bias; (b), author gender bias in topic classifiers can be reduced by adversarial training; however, adversarial debiasing in the opposite direction fails completely; (c) this effect is true for all languages except French. Our interpretation is that the failure of adversarial debiasing is due to the fact that feature space for author gender is subsumed the topic feature space for all languages except French, where gender is expressed overtly by morphological cues that can be picked up by the model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Regarding bias analysis at the representation level, the most important source of bias is arguably formed by corpus-derived embeddings which are used by virtually all current NLP systems. There have been several efforts to investigate the amount of bias within monolingual (Bolukbasi et al., 2016; Caliskan et al., 2017; Garg et al., 2018; Swinger et al., 2019) and multilingual embeddings. (Lauscher and Glava\u0161, 2019; Zhao et al., 2020) . Bias analysis at the system level has investigated a range of applications such as NER (Mehrabi et al., 2020) , Machine Translation (Stanovsky et al., 2019) , Natural Language Inference (Rudinger et al., 2017) , Emotion Intensity Prediction (Kiritchenko and Mohammad, 2018) , Coreference Resolution (Rudinger et al., 2018; Zhao et al., 2018) and Text Classification .",
"cite_spans": [
{
"start": 273,
"end": 297,
"text": "(Bolukbasi et al., 2016;",
"ref_id": "BIBREF1"
},
{
"start": 298,
"end": 320,
"text": "Caliskan et al., 2017;",
"ref_id": "BIBREF2"
},
{
"start": 321,
"end": 339,
"text": "Garg et al., 2018;",
"ref_id": "BIBREF12"
},
{
"start": 340,
"end": 361,
"text": "Swinger et al., 2019)",
"ref_id": "BIBREF33"
},
{
"start": 391,
"end": 418,
"text": "(Lauscher and Glava\u0161, 2019;",
"ref_id": "BIBREF22"
},
{
"start": 419,
"end": 437,
"text": "Zhao et al., 2020)",
"ref_id": "BIBREF40"
},
{
"start": 527,
"end": 549,
"text": "(Mehrabi et al., 2020)",
"ref_id": "BIBREF25"
},
{
"start": 572,
"end": 596,
"text": "(Stanovsky et al., 2019)",
"ref_id": "BIBREF32"
},
{
"start": 626,
"end": 649,
"text": "(Rudinger et al., 2017)",
"ref_id": "BIBREF29"
},
{
"start": 681,
"end": 713,
"text": "(Kiritchenko and Mohammad, 2018)",
"ref_id": "BIBREF19"
},
{
"start": 739,
"end": 762,
"text": "(Rudinger et al., 2018;",
"ref_id": "BIBREF30"
},
{
"start": 763,
"end": 781,
"text": "Zhao et al., 2018)",
"ref_id": "BIBREF41"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "A number of strategies have been explored for bias mitigation. One common strategy for reducing the bias is to target the representational level again, that is, corpora (Hall Maudslay et al., 2019; Zhao et al., 2018) and word embeddings (Kaneko and Bollegala, 2019; Bolukbasi et al., 2016) . Other methods target the model architecture in various ways. For example, Qian et al. (2019) introduce an additional term to be used in loss function of language generation model, seeking to reduce the gender bias exhibited by the model.",
"cite_spans": [
{
"start": 169,
"end": 197,
"text": "(Hall Maudslay et al., 2019;",
"ref_id": "BIBREF15"
},
{
"start": 198,
"end": 216,
"text": "Zhao et al., 2018)",
"ref_id": "BIBREF41"
},
{
"start": 237,
"end": 265,
"text": "(Kaneko and Bollegala, 2019;",
"ref_id": "BIBREF18"
},
{
"start": 266,
"end": 289,
"text": "Bolukbasi et al., 2016)",
"ref_id": "BIBREF1"
},
{
"start": 366,
"end": 384,
"text": "Qian et al. (2019)",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "A more fundamental idea is to adopt adversarial training (Goodfellow et al., 2014) to other tasks. For instance, Ganin and Lempitsky (2015) adapted adversarial training to the task of domain adaptation by introducing Gradient Reversal Layer (GRL) which acts as an identity function during forward pass and reverses the gradient by multiplying it by a negative scalar during the backward pass. Elazar and Goldberg (2018) apply the idea to the removal of demographic bias; McHardy et al. 2019remove publication source as a bias variable from a satire detection model. reported that adversarial training with GRL layer can remove unintended bias from the representations of POS tagging and Sentiment analysis models while maintaining task performance. Zhang et al. (2018) show that adversarial training mitigates the bias in word embeddings while maintaining its performance on word analogies task. Finally, Arduini et al. (2020) demonstrate how adversarial learning can be used for debiasing knowledge graph embeddings.",
"cite_spans": [
{
"start": 57,
"end": 82,
"text": "(Goodfellow et al., 2014)",
"ref_id": "BIBREF13"
},
{
"start": 113,
"end": 139,
"text": "Ganin and Lempitsky (2015)",
"ref_id": "BIBREF11"
},
{
"start": 393,
"end": 419,
"text": "Elazar and Goldberg (2018)",
"ref_id": "BIBREF10"
},
{
"start": 749,
"end": 768,
"text": "Zhang et al. (2018)",
"ref_id": "BIBREF38"
},
{
"start": 905,
"end": 926,
"text": "Arduini et al. (2020)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "In order to conduct a study on the relationship between topic and author gender in multiple languages, we require a multilingual comparable corpus for which topic and gender information are available. The corpus should be as parallel as possible so that any differences in outcome across languages are not simply due to differences in the evaluation data. Among the available multilingual parallel data sets, arguably the two most prominent ones are WIT 3 and OPUS. WIT 3 (Cettolo et al., 2012) consists of lecture translations automatically crawled from the TED talks in a variety of languages and was used in the evaluation campaigns IWSLT 2013 and 2014. OPUS (Tiedemann, 2012) is a collection of data from several sources which provides sentence alignments as well as linguistic markup (for some languages). Unfortunately, neither corpus provides topic or gender labels.",
"cite_spans": [
{
"start": 472,
"end": 494,
"text": "(Cettolo et al., 2012)",
"ref_id": "BIBREF3"
},
{
"start": 662,
"end": 679,
"text": "(Tiedemann, 2012)",
"ref_id": "BIBREF34"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "3"
},
{
"text": "For this reason, we create a new multilingual parallel dataset with these annotations, based on TED talks (http://ted.com/talks). A TED talk is a presentation at the TED conference or one of its international partner events. TED talks are limited to a maximum length of 18 minutes and may be on any topic. TED talks are rehearsed talks and at least semi-formal, while still definitely belonging to the category of spoken language. In this regard, they are comparable to the widely used Europarl corpus (Koehn, 2005) . The talks are divided according to the languages, topics and posted dates. All original talks are presented in English, but volunteers provide (and double check) translations into other languages. Authors are identified by name.",
"cite_spans": [
{
"start": 502,
"end": 515,
"text": "(Koehn, 2005)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "3"
},
{
"text": "Checking for which languages the TED webpage provided substantial numbers of transcripts (as of February 2020) led us to select German (DE), Spanish (ES), French (FR) and Turkish (TR) as target languages. We crawled all 1518 TED talks for which transcripts in all four target languages were available. We conducted some preprocessing: we cleaned transcripts by removing extra line breaks, extra spaces, and punctuation marks. Inspired by the work in open-domain Question Answering (Yang et al., 2019) , we then segmented the transcripts into a sequence of segments. Rather than using paragraphs or sentences as segments directly, we split articles into segments with the length of 60 words by sliding window as Wang et al. (2019) demonstrated that splitting articles into non-overlapping fixed-length segments leads to better results in Question Answering.",
"cite_spans": [
{
"start": 481,
"end": 500,
"text": "(Yang et al., 2019)",
"ref_id": "BIBREF37"
},
{
"start": 711,
"end": 729,
"text": "Wang et al. (2019)",
"ref_id": "BIBREF36"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "3"
},
{
"text": "Finally, we annotated the transcripts with topic and author gender information. For topic, we grouped transcripts into two classes according to the community-provided tags. The instances that have either Technology or Science tag were labeled as SciTech while the rest was labeled as Other. This grouping strategy led to a balanced dataset (53% Science, 47% Other). For author gender, we assume a binary gender classification (male/female) to be compatible with existing datasets Pardo et al., 2016) . This should not be understood as a rejection of non-binary gender. We manually determine the author's gender infor- mation on the basis of gender indicating pronouns such as he, she, his, her that are used to refer to the authors in their biographies published in the authors' TED Talks profile or on other websites, keeping only clear cases. The majority gender is male (69%). Table 1 describes the final dataset. 2 The corpus has very similar properties across languages. The main exception is the lower number of words in Turkish which is due to the agglutinative nature of Turkish morphology. For instance, the English sentence with four words \"I am at your house.\" is translated into a single word Turkish sentence \"Evinizdeyim.\"",
"cite_spans": [
{
"start": 480,
"end": 499,
"text": "Pardo et al., 2016)",
"ref_id": "BIBREF26"
},
{
"start": 917,
"end": 918,
"text": "2",
"ref_id": null
}
],
"ref_spans": [
{
"start": 880,
"end": 887,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Dataset",
"sec_num": "3"
},
{
"text": "4 Experimental Design Table 2 shows a correlation matrix for the two attributes of topic and author gender in our TED corpus. Indeed, the corpus shows a clear correlation between the two: while male authors are represented about equally in TED for scientific-technological topics and other topics, female authors are underrepresented for scientifictechnological topics. As motivated in the Introduction, this situation can lead to the model mistakenly picking up linguistic cues from one attribute to predict the other, leading to systematic biases. We therefore believe that this corpus can serve as a reasonable case study for correlated document attributes. We proceed as follows: Experiment 1: We learn individual neural models for topic and gender classification. We expect, for each attribute, that predictions are biased regarding the other attribute. Experiment 2: We debias these models by adversarial training. We expect the models to focus better on features that are predictive of the individual attributes, and to show less bias. We follow and Zhao et al. (2020) by measuring the amount of bias in the models as the average difference in classification performance between documents aggregated by author gender (Male vs Female) or aggregated by topic (SciTech vs Other).",
"cite_spans": [
{
"start": 1057,
"end": 1075,
"text": "Zhao et al. (2020)",
"ref_id": "BIBREF40"
}
],
"ref_spans": [
{
"start": 22,
"end": 29,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Dataset",
"sec_num": "3"
},
{
"text": "In our first experiment, we set up neural classification models for the two tasks of topic and gender classification individually and evaluate them for the presence of bias in their predictions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment 1: Simple Classification and Bias Analysis",
"sec_num": "5"
},
{
"text": "Figure 1 depicts the model architecture we use for both classification tasks. We use a neural text classifier based on the BERT Transformer (Devlin et al., 2019) with some adjustments. While transformers have shown good performance on many language tasks, most of them can only encode and generate representations for a fixed length token sequence -e.g., BERT implementations are often limited to 512 tokens per sequence. As the average token number per TED talk (cf. the global context vector for the input by summing paragraph representations element-wise. Finally, the global representation of the input is fed through a Multi-Layer Perceptron to a Softmax layer. Our model can be understood as an adaptation of standard transformer classifiers to longer texts.",
"cite_spans": [
{
"start": 140,
"end": 161,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Method",
"sec_num": "5.1"
},
{
"text": "We first set up the model for topic classification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Topic Classification",
"sec_num": "5.2"
},
{
"text": "We approach it as a document-level binary classification task. The input to the model is the full transcript, and the model labels each transcript either as \"SciTech\" or \"Other\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Topic Classification",
"sec_num": "5.2"
},
{
"text": "The topic classification results are shown in Table 3, using weighted F 1 score for evaluation. First, we compare the overall performance across languages. A majority baseline performs at 37.6% for all languages, due to the parallel design of the dataset. The neural topic classifiers do substantially better, all showing very similar results around 81% F-Score. Their similar performance may be expected from the parallel nature of the corpus, but it also provides support to our assumption that the texts and transformer models perform comparably across languages. When we break down these results by the other attribute we are interested, namely author gender (Male vs Female), we find that the prediction quality of the topic classifier is an average of 3.6 points lower for male than for female authors. In other words, the topic classifiers show a consistent gender bias across languages, presumably due to the higher-entropy (more equal) topic distribution for male authors (cf. Table 2) . While this bias is lower than the bias of a majority baseline (which directly reflects the correlation between the two attributes), it is still substantial and arguably worth mitigating. ",
"cite_spans": [],
"ref_spans": [
{
"start": 986,
"end": 994,
"text": "Table 2)",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Topic Classification",
"sec_num": "5.2"
},
{
"text": "We now address the opposite task, author gender classification, predicting the labels Male and Female, re-using the model architecture from before. Table 4 summarizes the results. We see a pattern that differs substantially from topic classification, with much larger cross-lingual differences in performance. The results are again substantially above the 57% baseline. We obtain the best result for French (82%), and the worst for Turkish (70%), with a difference of 12% F-Score. This indicates that gender classification builds much more on language-specific information than topic classification. Arguably, for a word piece-based neural model like BERT, a primary source of evidence on author gender are linguistically marked expressions in the text where the author refers to themselves. Thus, prediction of the author gender should be easiest if a language has a frequent and unambiguous mechanism for gender marking (Corbett, 1991; Zmigrod et al., 2019) . Table 5 shows a multilingual example where French marks gender inflectionally, while the other languages do not. This is indicative of the general case: The languages that we consider in our experiment provide gender marking to different degrees. At one extreme, French marks most adjectives and many nouns consistently for gender. In contrast, Spanish marks gender only for a subset of the lexicon, and morphologically inconsistently (Harris, 1991) ; German marks only (some) nouns, and marking is sometimes optional. At the other extreme, Turkish does not mark gender at all.",
"cite_spans": [
{
"start": 922,
"end": 937,
"text": "(Corbett, 1991;",
"ref_id": "BIBREF5"
},
{
"start": 938,
"end": 959,
"text": "Zmigrod et al., 2019)",
"ref_id": "BIBREF42"
},
{
"start": 1397,
"end": 1411,
"text": "(Harris, 1991)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [
{
"start": 148,
"end": 155,
"text": "Table 4",
"ref_id": "TABREF6"
},
{
"start": 962,
"end": 969,
"text": "Table 5",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "On this basis, we would expect French to perform best, and lower performance for the other three languages -exactly what we find. However, the performances for TR, DE, and ES are surprisingly close to one another, and substantially above the baseline: on the basis of what information in the texts do these classifiers base their predictions? DE Genau hier wurde ich geboren und verbrachte die ersten sieben Jahre meines Lebens. ES Esta es la tierra en la que nac\u00ed y pas\u00e9 los primeros siete a\u00f1os de mi vida. FR Je suis n\u00e9e ici m\u00eame, et j'y ai pass\u00e9 les sept premi\u00e8res ann\u00e9es de ma vie. TR Dogdugum yer buras\u0131 ve hayat\u0131m\u0131n ilk yedi y\u0131l\u0131n\u0131 burada ge\u00e7irdim. A look at the size of the biases suggests an explanation: The gender classifiers for DE, ES, and TR make substantial use of topic cues, which enables them to proceed to some extent due to the correlations between topic and gender, but also lead to biases of 6-9% (highest for Turkish, consistent with the analysis above). In contrast, the French classifier is least biased, indicating that its text contains enough cues for 'proper' gender classification. We illustrate this in Table 6 , where we report results on SciTech documents with female authors, that is, the smallest subcategory in our corpus. We find that the gender classifier for FR significantly outperforms the others, which provides additional evidence that the model relies less on the topic cues for gender classification.",
"cite_spans": [],
"ref_spans": [
{
"start": 1133,
"end": 1140,
"text": "Table 6",
"ref_id": "TABREF8"
}
],
"eq_spans": [],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "Summary. For both tasks, we find that the classification performance shows a bias with respect to the other attribute. The two tasks differ with respect to the cross-lingual component, though: Topic classification works about equally well in all languages. In contrast, author gender classification only works properly in the one language that has consistent linguistic marking of gender, while there is evidence that the other languages fall back on topic features also for this task, which directly leads to biased predictions. These observations motivate experiments into how well these models respond to debiasing. 6 Experiment 2: Adversarial Debiasing",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "Let P be some bias attribute (e.g., gender, race, age etc.) that we want our classifier to ignore while learning to solve another task T . Adversarial debiasing seeks to achieve this by constraining representations in a way so that representations do not rely on P in any substantial way. To this end, the model is trained to simultaneously predict the correct label for task T (\"main component\") and to prevent a jointly trained adversary (\"adversarial component\") from predicting P (McHardy et al., 2019) . We define the loss functions of the main (J M ) and adversarial (J A ) components as follows:",
"cite_spans": [
{
"start": 484,
"end": 506,
"text": "(McHardy et al., 2019)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "J A = \u2212E (x,y A )\u223cp data log P \u03b8 A \u222a\u03b8 F (y A , x)",
"eq_num": "(1)"
}
],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "J M = \u2212E (x,y M )\u223cp data log P \u03b8 M \u222a\u03b8 F (y M , x) (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "where \u03b8 A , \u03b8 M are the parameters of adversarial and main components; y A and y M are the gold labels for main and adversary tasks. Note that the adversarial and main components share the same feature extractor (i.e., BERT) whose parameters (\u03b8 F ) are therefore updated by the gradients coming through the objective functions of both model parts. Let \u03bb be the meta-parameter controlling the intensity of the adversarial training and \u03b7 the learning rate. Then the following equations describe update rules for each component in the model:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03b8 M := \u03b8 M \u2212 \u03b7 \u2202J M \u2202\u03b8 M (3) \u03b8 A := \u03b8 A \u2212 \u03b7 \u2202J A \u2202\u03b8 A (4) \u03b8 F := \u03b8 F \u2212 \u03b7 \u2202J M \u2202\u03b8 F \u2212 \u03bb \u2202J A \u2202\u03b8 F",
"eq_num": "(5)"
}
],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "Our application of this training method is shown in Figure 2 . We first debias the topic classifier by author gender (left-hand box); then we proceed to debias author gender classifier by topic (right-hand box). For example, to de-bias the topic classification, J M is the topic loss and J A the author gender loss; vice versa for author gender de-biasing.",
"cite_spans": [],
"ref_spans": [
{
"start": 52,
"end": 60,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Author Gender Classification",
"sec_num": "5.3"
},
{
"text": "First, we debias topic classification to reduce the gender bias. The left-hand side of Figure 3 compares overall results across a range of values of \u03bb between 0 (no adversarial training) and 1 (equal weight of main and adversarial loss). We find that, similar to Experiment 1, the results are essentially identical across languages. Furthermore, the choice of \u03bb hardly matters in this interval: adversarial training does not have a major impact on topic classification. We report detailed results for \u03bb=1 in the right-hand side of Figure 3 . The small differences between the Overall results of the Original and Debiased models show that topic classification overall does not lose much by debiasing for gender. 3 The breakdown by gender shows that gender bias is substantially reduced overall. However, there are noticeable differences among languages. For Spanish and German, we see no overall loss of performance in topic classification, and a substantial reduction in gender bias. For French and Turkish, in contrast, we see a decrease of about 1.5 points in topic classification. Gender bias is reduced for French but hardly for Turkish. This is a somewhat surprising result, given the typological differences between the two languages. Our explanation is that in French, as discussed above, many words are morphologically marked for gender. Due to the correlation between the two attributes, these can be re-used by the topic classifier, but when they are penalized through adversarial training, we see a mild decrease in topic classification accuracy. In Turkish, as we have argued in Experiment 1, gender classification depends almost entirely on topic features since there is no linguistic marking of referent gender. Consequently, the adversarial training works against itself to an extent, resulting in a mildly worse topic classification but hardly any decrease in gender bias.",
"cite_spans": [
{
"start": 711,
"end": 712,
"text": "3",
"ref_id": null
}
],
"ref_spans": [
{
"start": 87,
"end": 95,
"text": "Figure 3",
"ref_id": "FIGREF2"
},
{
"start": 531,
"end": 539,
"text": "Figure 3",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Topic-Debiased Gender Classification",
"sec_num": "6.1"
},
{
"text": "Now we swap the main and adversarial tasks again, debiasing author gender classification with regard to topic. We use the same setup as in Experiment 1. The results are shown in Figure 4 . The left-hand side shows that varying \u03bb has a substantial effect this time. If we set \u03bb to a value close to 1 -a good choice for gender-debiased topic classification, as we have established in the previous subsectionthis leads to a breakdown of the gender classification model. Performance for all languages drops to a F-Score of around 57, the level of the major-ity baseline (cf . Table 4 ). Apparently, debiasing author gender classification by adversarial training against topic breaks the author gender classifier for all but small values of \u03bb.",
"cite_spans": [],
"ref_spans": [
{
"start": 178,
"end": 186,
"text": "Figure 4",
"ref_id": "FIGREF3"
},
{
"start": 570,
"end": 579,
"text": ". Table 4",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Gender-Debiased Topic Classification",
"sec_num": "6.2"
},
{
"text": "As in the first experiment, we observe differences among languages: French stands out as the language for which the gender classification 'holds out' the longest for high values of \u03bb. Its ultimate failure indicates that even for French, gender marking on its own is not strong enough to support the author gender identification task -or at least our models are not powerful enough to pick up on these cues. The other languages, which, as we have argued in Experiment 1, make substantial use of topic cues for gender classification, fail even earlier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Gender-Debiased Topic Classification",
"sec_num": "6.2"
},
{
"text": "The right-hand side of Figure 4 reports detailed results for \u03bb=0.2. In line with our analyses above, debiasing works for French but not for the other languages: We find clear decreases in performance (up to 6.2 points, for Spanish), and inconclusive changes in bias (decrease for Turkish by 4.4 points, increase for German by 1.6 points). While the patterns for these languages are not straightforward to interpret, it seems safe to conclude that topicdebiasing author gender is a failure both with regard to model performance and reduction of bias.",
"cite_spans": [],
"ref_spans": [
{
"start": 23,
"end": 31,
"text": "Figure 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Gender-Debiased Topic Classification",
"sec_num": "6.2"
},
{
"text": "The results of our two experiments show an intriguing asymmetry between the two tasks of topic and author gender classification when debiased for the respective other attribute. Reducing author gender bias in topic classification with adversarial training proceeds as expected, is relatively robust to the choice of \u03bb in the interval between 0 and 1, and shows a consistent pattern across languages which can be explained by the properties of the languages involved. In contrast, reducing topic bias in author gender classification relies heavily on \u03bb, quickly deteriorating to baseline level for large values of \u03bb, and does not consistently manage to reduce bias in any case. This asymmetry cannot be an artifact of model architecture or data alone, since we use the same model architecture on the same data. Instead, we believe that these patterns result from an interaction between the representation learning of the model and the information that the model can draw from the data. They can be understood through the latent feature space of the final shared layer in our architecture below the two heads (cf. Figure 2) , where each class can be characterized by a region of informative features. Figure 5 shows Venn diagram-style depictions of the three possible cases for a pair of attributes. In the left-hand case, (a), there is no overlap between the latent features of the two attributes. That is, the two attributes are independent of one another, and so is learning. However, this is by definition the case without correlations among attributes that we do not consider. In the center case, (b), there is an intersection between the latent features of the two attributes. The classifiers' use of this overlap potentially creates biases, but adversarial training exactly punishes the use of this region of latent feature space. Thus, debiased classifiers can learn either attribute to the extent that the part of the feature space outside the intersection is still sufficiently informative. The right-hand case, (c), is the limit case when one of the two attributes does not have an independent standing, that is, the informative latent features of attribute 1 are completely contained in the informative feature space of attribute 2. This leads to biases in either classifier just as case (b), but also creates an asymmetry in the effect of adversarial debiasing: Attribute 2 can be debiased by simply 'cutting out' the informative space of attribute 1, but debiasing attribute 1 in the opposite manner results in an empty feature space for attribute 1, and we would expect the classifier to revert to baseline performance.",
"cite_spans": [],
"ref_spans": [
{
"start": 1112,
"end": 1121,
"text": "Figure 2)",
"ref_id": "FIGREF1"
},
{
"start": 1199,
"end": 1207,
"text": "Figure 5",
"ref_id": "FIGREF4"
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "7"
},
{
"text": "This set theoretic visualization is a major simplification of the latent feature space in neural models, where the three cases cannot apply categoriallythey rather represent different points on a continuum. Nevertheless, the predictions of the subsumption case, (c), match our experimental results well: Assuming that author gender features are included in topic features, we would expect to find successful debiasing of the topic classifier, but breakdown of the debiased author gender classifier. This is exactly the pattern of results that we have observed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "7"
},
{
"text": "Note that this analysis builds on the behavior of the features of the attributes in the training data, in particular in a representation learning approach like the one we have pursued. In other words, changes of the data -or differences within the data, such as between languages -are expected to influence the outcome. Again, this is what we see: French, due to its consistent morphological marking of gender, is closer to case (b), while the other languages are closer to case (c).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "7"
},
{
"text": "This paper was concerned with text classification for correlated attributes, which pose an important but often overlooked challenge to model fairness -in particular, as we have argued, in the case of demographic attributes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "We specifically analyzed the relationship between document topic and author gender. We established that topic classifiers exhibit gender bias and author gender classifiers show topic bias; that adversarial debiasing corrects gender bias in topic classification but breaks down in the opposite di-rection; and that this effect varies by language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "Beyond the concrete study, our contribution is to draw attention to the general question of prerequisites for successful adversarial debiasing, which, to our knowledge, has not received much attention. Our results indicate that when the target attribute and the bias attribute are too strongly correlatedor, indeed, when the target attribute is subsumed by the bias attribute -adversarial debiasing fails: with a small weight on the bias component, no debiasing takes place; with a large weight, target attribute classification deteriorates to baseline level.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "Furthermore, we find that the linguistic expression of the attributes matters greatly: the only language for which we achieved satisfactory results was French, due to the consistent morphological marking of gender which can be captured independently of topic (Zmigrod et al., 2019) . This highlights the importance of understanding the differences between languages regarding how they encode content (Dubossarsky et al., 2019) , and underscores the importance of cross-lingual methods.",
"cite_spans": [
{
"start": 259,
"end": 281,
"text": "(Zmigrod et al., 2019)",
"ref_id": "BIBREF42"
},
{
"start": 400,
"end": 426,
"text": "(Dubossarsky et al., 2019)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "In future work, we plan to develop a diagnostic to recognize potentially problematic constellations of correlated attributes and improve debiasing. Table 8 : Topic classification F-scores of topic-debiased gender classifier and baseline model. For main task evaluation, check Figure 4 .",
"cite_spans": [],
"ref_spans": [
{
"start": 148,
"end": 155,
"text": "Table 8",
"ref_id": null
},
{
"start": 276,
"end": 284,
"text": "Figure 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "train the gender classifier on top of it. Table 7 summarizes the results for gender classification. Significant drop in gender classification performance indicates effectiveness of adversarial training.",
"cite_spans": [],
"ref_spans": [
{
"start": 42,
"end": 49,
"text": "Table 7",
"ref_id": "TABREF13"
}
],
"eq_spans": [],
"section": "Overall",
"sec_num": null
},
{
"text": "As we swapped the main and adversarial tasks, we modify the baseline in the same way too. We start with training the feature extractor and gender classifier head on the topic classification task. Then,we freeze the feature extractor, remove the gender classifier and train the topic classifier on top of freezed feature extractor. Table 8 reports the results on topic classification. Similar to Table 7, adversarial debiasing drops the adversarial task (i.e. topic classification) performance significantly. However, as Figure 4 shows it leads to slight to moderate decrease in gender classification performance and inconclusive changes with regard to topic bias.",
"cite_spans": [],
"ref_spans": [
{
"start": 331,
"end": 338,
"text": "Table 8",
"ref_id": null
},
{
"start": 520,
"end": 528,
"text": "Figure 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Overall",
"sec_num": null
},
{
"text": "A subset of these has specific legal protection in many jurisdictions under the name of sensitive or protected attributes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Code and data are available at http://www.ims. uni-stuttgart.de/data/ted_wassa21. This includes the documents we based our gender determination on, along with the list of gendered pronouns we used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "See Appendix for performance on the adversarial task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Partial funding was provided by Deutsche Forschungsgemeinschaft (DFG) through project MARDY within SPP RATIO. We would like to thank Roman Klinger, Gabriella Lapesa, Vivi Nastase and Michael Roth for valuable comments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
},
{
"text": "In our experiments, for each language we consider we use a cased BERT variant that was trained specifically for the target language. 4 We use the Adam optimizer with learning rates of 5e-5, \u03b2 1 = 0.9, \u03b2 2 = 0.999, a batch size of 48, a gradient clip threshold of 1.0 and a dropout with p=0.5 on all layers. We train the model for 15 epochs. The Multi Layer Perceptron consists of a single hidden layer with 300 hidden units. We evaluate each classifier using weighted F1-Score which calculate metrics for each label, and find their average weighted by the number of true instances for each label. We repeat every experiment using 5 random train (80%) test (20%) splits and report average of these 5 experiments.",
"cite_spans": [
{
"start": 133,
"end": 134,
"text": "4",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Appendix A Training of BERT-based document classifiers",
"sec_num": null
},
{
"text": "In addition to majority class classifier and nonadversarial model, we use a third baseline model to analyze how adversarial debiasing effects the model's performance on the adversary task. First, we train feature extractor along with the topic classifier head on the topic classification task. Next, we freeze the weights of the feature extractor and 4 DE: https://deepset.ai/german-bert, ES:https://github.com/dccuchile/beto, FR:https://camembert-model.fr/, TR: https://github.com/dbmdz/berts",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "B Adversarial Debiasing: Performance on adversarial tasks",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Adversarial learning for debiasing knowledge graph embeddings",
"authors": [
{
"first": "Mario",
"middle": [],
"last": "Arduini",
"suffix": ""
},
{
"first": "Lorenzo",
"middle": [],
"last": "Noci",
"suffix": ""
},
{
"first": "Federico",
"middle": [],
"last": "Pirovano",
"suffix": ""
},
{
"first": "Ce",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 16th International Workshop on Mining and Learning with Graphs (MLG)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mario Arduini, Lorenzo Noci, Federico Pirovano, Ce Zhang, Yash Raj Shrestha, and Bibek Paudel. 2020. Adversarial learning for debiasing knowledge graph embeddings. In Proceedings of the 16th In- ternational Workshop on Mining and Learning with Graphs (MLG), San Diego, California.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Man is to computer programmer as woman is to homemaker? debiasing word embeddings",
"authors": [
{
"first": "Tolga",
"middle": [],
"last": "Bolukbasi",
"suffix": ""
},
{
"first": "Kai-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "James",
"suffix": ""
},
{
"first": "Venkatesh",
"middle": [],
"last": "Zou",
"suffix": ""
},
{
"first": "Adam",
"middle": [
"T"
],
"last": "Saligrama",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kalai",
"suffix": ""
}
],
"year": 2016,
"venue": "Advances in Neural Information Processing Systems",
"volume": "29",
"issue": "",
"pages": "4349--4357",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Ad- vances in Neural Information Processing Systems, volume 29, pages 4349-4357.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Semantics derived automatically from language corpora contain human-like biases",
"authors": [
{
"first": "Aylin",
"middle": [],
"last": "Caliskan",
"suffix": ""
},
{
"first": "Joanna",
"middle": [
"J"
],
"last": "Bryson",
"suffix": ""
},
{
"first": "Arvind",
"middle": [],
"last": "Narayanan",
"suffix": ""
}
],
"year": 2017,
"venue": "Science",
"volume": "356",
"issue": "6334",
"pages": "183--186",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183-186.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Wit3: Web inventory of transcribed and translated talks",
"authors": [
{
"first": "Mauro",
"middle": [],
"last": "Cettolo",
"suffix": ""
},
{
"first": "Christian",
"middle": [],
"last": "Girardi",
"suffix": ""
},
{
"first": "Marcello",
"middle": [],
"last": "Federico",
"suffix": ""
}
],
"year": 2012,
"venue": "Conference of European Association for Machine Translation",
"volume": "",
"issue": "",
"pages": "261--268",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mauro Cettolo, Christian Girardi, and Marcello Fed- erico. 2012. Wit3: Web inventory of transcribed and translated talks. In Conference of European As- sociation for Machine Translation, pages 261-268, Trento, Italy.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Antisocial behavior in online discussion communities",
"authors": [
{
"first": "Justin",
"middle": [],
"last": "Cheng",
"suffix": ""
},
{
"first": "Cristian",
"middle": [],
"last": "Danescu-Niculescu-Mizil",
"suffix": ""
},
{
"first": "Jure",
"middle": [],
"last": "Leskovec",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the International AAAI Conference on Web and Social Media",
"volume": "9",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Justin Cheng, Cristian Danescu-Niculescu-Mizil, and Jure Leskovec. 2015. Antisocial behavior in online discussion communities. In Proceedings of the Inter- national AAAI Conference on Web and Social Media, volume 9.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Gender. Cambridge Textbooks in Linguistics",
"authors": [
{
"first": "G",
"middle": [],
"last": "Greville",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Corbett",
"suffix": ""
}
],
"year": 1991,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1017/CBO9781139166119"
]
},
"num": null,
"urls": [],
"raw_text": "Greville G. Corbett. 1991. Gender. Cambridge Text- books in Linguistics. Cambridge University Press.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Bias in bios: A case study of semantic representation bias in a high-stakes setting",
"authors": [
{
"first": "Maria",
"middle": [],
"last": "De-Arteaga",
"suffix": ""
},
{
"first": "Alexey",
"middle": [],
"last": "Romanov",
"suffix": ""
},
{
"first": "Hanna",
"middle": [],
"last": "Wallach",
"suffix": ""
},
{
"first": "Jennifer",
"middle": [],
"last": "Chayes",
"suffix": ""
},
{
"first": "Christian",
"middle": [],
"last": "Borgs",
"suffix": ""
},
{
"first": "Alexandra",
"middle": [],
"last": "Chouldechova",
"suffix": ""
},
{
"first": "Sahin",
"middle": [],
"last": "Geyik",
"suffix": ""
},
{
"first": "Krishnaram",
"middle": [],
"last": "Kenthapadi",
"suffix": ""
},
{
"first": "Adam Tauman",
"middle": [],
"last": "Kalai",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the Conference on Fairness, Accountability, and Transparency, FAT* '19",
"volume": "",
"issue": "",
"pages": "120--128",
"other_ids": {
"DOI": [
"10.1145/3287560.3287572"
]
},
"num": null,
"urls": [],
"raw_text": "Maria De-Arteaga, Alexey Romanov, Hanna Wal- lach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnaram Kentha- padi, and Adam Tauman Kalai. 2019. Bias in bios: A case study of semantic representation bias in a high-stakes setting. In Proceedings of the Confer- ence on Fairness, Accountability, and Transparency, FAT* '19, page 120-128, New York, NY, USA. As- sociation for Computing Machinery.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"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 the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "4171--4186",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1423"
]
},
"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 under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Measuring and mitigating unintended bias in text classification",
"authors": [
{
"first": "Lucas",
"middle": [],
"last": "Dixon",
"suffix": ""
},
{
"first": "John",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Sorensen",
"suffix": ""
},
{
"first": "Nithum",
"middle": [],
"last": "Thain",
"suffix": ""
},
{
"first": "Lucy",
"middle": [],
"last": "Vasserman",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, AIES '18",
"volume": "",
"issue": "",
"pages": "67--73",
"other_ids": {
"DOI": [
"10.1145/3278721.3278729"
]
},
"num": null,
"urls": [],
"raw_text": "Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2018. Measuring and mitigat- ing unintended bias in text classification. In Pro- ceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, AIES '18, page 67-73, New York, NY, USA. Association for Computing Machin- ery.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Proceedings of TyP-NLP: The First Workshop on Typology for Polyglot NLP",
"authors": [
{
"first": "Haim",
"middle": [],
"last": "Dubossarsky",
"suffix": ""
},
{
"first": "Arya",
"middle": [
"D"
],
"last": "Mccarthy",
"suffix": ""
},
{
"first": "Edoardo",
"middle": [
"Maria"
],
"last": "Ponti",
"suffix": ""
},
{
"first": "Ivan",
"middle": [],
"last": "Vuli\u0107",
"suffix": ""
},
{
"first": "Ekaterina",
"middle": [],
"last": "Vylomova",
"suffix": ""
},
{
"first": "Yevgeni",
"middle": [],
"last": "Berzak",
"suffix": ""
},
{
"first": "Ryan",
"middle": [],
"last": "Cotterell",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Haim Dubossarsky, Arya D. McCarthy, Edoardo Maria Ponti, Ivan Vuli\u0107, Ekaterina Vylomova, Yevgeni Berzak, Ryan Cotterell, Manaal Faruqui, Anna Ko- rhonen, and Roi Reichart, editors. 2019. Proceed- ings of TyP-NLP: The First Workshop on Typology for Polyglot NLP. Association for Computational Linguistics, Florence, Italy.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Adversarial removal of demographic attributes from text data",
"authors": [
{
"first": "Yanai",
"middle": [],
"last": "Elazar",
"suffix": ""
},
{
"first": "Yoav",
"middle": [],
"last": "Goldberg",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "11--21",
"other_ids": {
"DOI": [
"10.18653/v1/D18-1002"
]
},
"num": null,
"urls": [],
"raw_text": "Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing, pages 11-21, Brussels, Belgium. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Unsupervised domain adaptation by backpropagation",
"authors": [
{
"first": "Yaroslav",
"middle": [],
"last": "Ganin",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Lempitsky",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 32nd International Conference on Machine Learning",
"volume": "37",
"issue": "",
"pages": "1180--1189",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yaroslav Ganin and Victor Lempitsky. 2015. Unsu- pervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 1180-1189, Lille, France. JMLR.org.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings of the National Academy of",
"authors": [
{
"first": "Nikhil",
"middle": [],
"last": "Garg",
"suffix": ""
},
{
"first": "Londa",
"middle": [],
"last": "Schiebinger",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Jurafsky",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Zou",
"suffix": ""
}
],
"year": 2018,
"venue": "Sciences",
"volume": "115",
"issue": "16",
"pages": "3635--3644",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Pro- ceedings of the National Academy of Sciences, 115(16):E3635-E3644.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Generative adversarial nets",
"authors": [
{
"first": "Ian",
"middle": [
"J"
],
"last": "Goodfellow",
"suffix": ""
},
{
"first": "Jean",
"middle": [],
"last": "Pouget-Abadie",
"suffix": ""
},
{
"first": "Mehdi",
"middle": [],
"last": "Mirza",
"suffix": ""
},
{
"first": "Bing",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Warde-Farley",
"suffix": ""
},
{
"first": "Sherjil",
"middle": [],
"last": "Ozair",
"suffix": ""
},
{
"first": "Aaron",
"middle": [],
"last": "Courville",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 27th International Conference on Neural Information Processing Systems",
"volume": "2",
"issue": "",
"pages": "2672--2680",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative ad- versarial nets. In Proceedings of the 27th Interna- tional Conference on Neural Information Processing Systems -Volume 2, page 2672-2680, Cambridge, MA, USA. MIT Press.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Emotion and aging: Experience, expression, and control",
"authors": [
{
"first": "J",
"middle": [],
"last": "James",
"suffix": ""
},
{
"first": "Laura",
"middle": [
"L"
],
"last": "Gross",
"suffix": ""
},
{
"first": "Monisha",
"middle": [],
"last": "Carstensen",
"suffix": ""
},
{
"first": "Jeanne",
"middle": [],
"last": "Pasupathi",
"suffix": ""
},
{
"first": "Carina",
"middle": [
"G\u00f6testam"
],
"last": "Tsai",
"suffix": ""
},
{
"first": "Angie Yc",
"middle": [],
"last": "Skorpen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hsu",
"suffix": ""
}
],
"year": 1997,
"venue": "Psychology and aging",
"volume": "12",
"issue": "4",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "James J Gross, Laura L Carstensen, Monisha Pasu- pathi, Jeanne Tsai, Carina G\u00f6testam Skorpen, and Angie YC Hsu. 1997. Emotion and aging: Experi- ence, expression, and control. Psychology and ag- ing, 12(4):590.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "It's all in the name: Mitigating gender bias with name-based counterfactual data substitution",
"authors": [
{
"first": "Hila",
"middle": [],
"last": "Rowan Hall Maudslay",
"suffix": ""
},
{
"first": "Ryan",
"middle": [],
"last": "Gonen",
"suffix": ""
},
{
"first": "Simone",
"middle": [],
"last": "Cotterell",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Teufel",
"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 (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "5267--5275",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1530"
]
},
"num": null,
"urls": [],
"raw_text": "Rowan Hall Maudslay, Hila Gonen, Ryan Cotterell, and Simone Teufel. 2019. It's all in the name: Mit- igating gender bias with name-based counterfactual data substitution. In Proceedings of the 2019 Con- ference on Empirical Methods in Natural Language Processing and the 9th International Joint Confer- ence on Natural Language Processing (EMNLP- IJCNLP), pages 5267-5275, Hong Kong, China. As- sociation for Computational Linguistics.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "The exponence of gender in spanish",
"authors": [
{
"first": "James",
"middle": [
"W"
],
"last": "Harris",
"suffix": ""
}
],
"year": 1991,
"venue": "Linguistic Inquiry",
"volume": "22",
"issue": "1",
"pages": "27--62",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "James W. Harris. 1991. The exponence of gender in spanish. Linguistic Inquiry, 22(1):27-62.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Demographic factors improve classification performance",
"authors": [
{
"first": "Dirk",
"middle": [],
"last": "Hovy",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing",
"volume": "1",
"issue": "",
"pages": "752--762",
"other_ids": {
"DOI": [
"10.3115/v1/P15-1073"
]
},
"num": null,
"urls": [],
"raw_text": "Dirk Hovy. 2015. Demographic factors improve classi- fication performance. In Proceedings of the 53rd An- nual Meeting of the Association for Computational Linguistics and the 7th International Joint Confer- ence on Natural Language Processing (Volume 1: Long Papers), pages 752-762, Beijing, China. As- sociation for Computational Linguistics.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Gender-preserving debiasing for pre-trained word embeddings",
"authors": [
{
"first": "Masahiro",
"middle": [],
"last": "Kaneko",
"suffix": ""
},
{
"first": "Danushka",
"middle": [],
"last": "Bollegala",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "1641--1650",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1160"
]
},
"num": null,
"urls": [],
"raw_text": "Masahiro Kaneko and Danushka Bollegala. 2019. Gender-preserving debiasing for pre-trained word embeddings. In Proceedings of the 57th Annual Meeting of the Association for Computational Lin- guistics, pages 1641-1650, Florence, Italy. Associa- tion for Computational Linguistics.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Examining gender and race bias in two hundred sentiment analysis systems",
"authors": [
{
"first": "Svetlana",
"middle": [],
"last": "Kiritchenko",
"suffix": ""
},
{
"first": "Saif",
"middle": [],
"last": "Mohammad",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics",
"volume": "",
"issue": "",
"pages": "43--53",
"other_ids": {
"DOI": [
"10.18653/v1/S18-2005"
]
},
"num": null,
"urls": [],
"raw_text": "Svetlana Kiritchenko and Saif Mohammad. 2018. Ex- amining gender and race bias in two hundred sen- timent analysis systems. In Proceedings of the Seventh Joint Conference on Lexical and Compu- tational Semantics, pages 43-53, New Orleans, Louisiana. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Europarl: A Parallel Corpus for Statistical Machine Translation",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2005,
"venue": "Conference Proceedings: the tenth Machine Translation Summit",
"volume": "",
"issue": "",
"pages": "79--86",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn. 2005. Europarl: A Parallel Corpus for Statistical Machine Translation. In Conference Proceedings: the tenth Machine Translation Summit, pages 79-86, Phuket, Thailand. AAMT.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Automatically categorizing written texts by author gender",
"authors": [
{
"first": "Moshe",
"middle": [],
"last": "Koppel",
"suffix": ""
},
{
"first": "Shlomo",
"middle": [],
"last": "Argamon",
"suffix": ""
},
{
"first": "Anat Rachel",
"middle": [],
"last": "Shimoni",
"suffix": ""
}
],
"year": 2004,
"venue": "Computing Reviews",
"volume": "45",
"issue": "1",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Moshe Koppel, Shlomo Argamon, and Anat Rachel Shimoni. 2004. Automatically categorizing writ- ten texts by author gender. Computing Reviews, 45(1):43.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Are we consistently biased? multidimensional analysis of biases in distributional word vectors",
"authors": [
{
"first": "Anne",
"middle": [],
"last": "Lauscher",
"suffix": ""
},
{
"first": "Goran",
"middle": [],
"last": "Glava\u0161",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics",
"volume": "",
"issue": "",
"pages": "85--91",
"other_ids": {
"DOI": [
"10.18653/v1/S19-1010"
]
},
"num": null,
"urls": [],
"raw_text": "Anne Lauscher and Goran Glava\u0161. 2019. Are we con- sistently biased? multidimensional analysis of bi- ases in distributional word vectors. In Proceed- ings of the Eighth Joint Conference on Lexical and Computational Semantics, pages 85-91, Minneapo- lis, Minnesota. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Towards robust and privacy-preserving text representations",
"authors": [
{
"first": "Yitong",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Timothy",
"middle": [],
"last": "Baldwin",
"suffix": ""
},
{
"first": "Trevor",
"middle": [],
"last": "Cohn",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "25--30",
"other_ids": {
"DOI": [
"10.18653/v1/P18-2005"
]
},
"num": null,
"urls": [],
"raw_text": "Yitong Li, Timothy Baldwin, and Trevor Cohn. 2018. Towards robust and privacy-preserving text represen- tations. In Proceedings of the 56th Annual Meet- ing of the Association for Computational Linguistics (Volume 2: Short Papers), pages 25-30, Melbourne, Australia. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Adversarial training for satire detection: Controlling for confounding variables",
"authors": [
{
"first": "Robert",
"middle": [],
"last": "Mchardy",
"suffix": ""
},
{
"first": "Heike",
"middle": [],
"last": "Adel",
"suffix": ""
},
{
"first": "Roman",
"middle": [],
"last": "Klinger",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "660--665",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1069"
]
},
"num": null,
"urls": [],
"raw_text": "Robert McHardy, Heike Adel, and Roman Klinger. 2019. Adversarial training for satire detection: Con- trolling for confounding variables. In Proceedings of the 2019 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 660-665, Minneapolis, Minnesota. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Man is to person as woman is to location: Measuring gender bias in named entity recognition",
"authors": [
{
"first": "Ninareh",
"middle": [],
"last": "Mehrabi",
"suffix": ""
},
{
"first": "Thamme",
"middle": [],
"last": "Gowda",
"suffix": ""
},
{
"first": "Fred",
"middle": [],
"last": "Morstatter",
"suffix": ""
},
{
"first": "Nanyun",
"middle": [],
"last": "Peng",
"suffix": ""
},
{
"first": "Aram",
"middle": [],
"last": "Galstyan",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 31st ACM Conference on Hypertext and Social Media",
"volume": "",
"issue": "",
"pages": "231--232",
"other_ids": {
"DOI": [
"10.1145/3372923.3404804"
]
},
"num": null,
"urls": [],
"raw_text": "Ninareh Mehrabi, Thamme Gowda, Fred Morstatter, Nanyun Peng, and Aram Galstyan. 2020. Man is to person as woman is to location: Measuring gender bias in named entity recognition. In Proceedings of the 31st ACM Conference on Hypertext and Social Media, page 231-232, New York, NY, USA. Associ- ation for Computing Machinery.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Overview of the 4th author profiling task at PAN 2016: Cross-genre evaluations",
"authors": [
{
"first": "Francisco Manuel Rangel",
"middle": [],
"last": "Pardo",
"suffix": ""
},
{
"first": "Paolo",
"middle": [],
"last": "Rosso",
"suffix": ""
},
{
"first": "Ben",
"middle": [],
"last": "Verhoeven",
"suffix": ""
},
{
"first": "Walter",
"middle": [],
"last": "Daelemans",
"suffix": ""
},
{
"first": "Martin",
"middle": [],
"last": "Potthast",
"suffix": ""
},
{
"first": "Benno",
"middle": [],
"last": "Stein",
"suffix": ""
}
],
"year": 2016,
"venue": "Working Notes of CLEF 2016 -Conference and Labs of the Evaluation forum",
"volume": "",
"issue": "",
"pages": "5--8",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Francisco Manuel Rangel Pardo, Paolo Rosso, Ben Verhoeven, Walter Daelemans, Martin Potthast, and Benno Stein. 2016. Overview of the 4th author pro- filing task at PAN 2016: Cross-genre evaluations. In Working Notes of CLEF 2016 -Conference and Labs of the Evaluation forum,\u00c9vora, Portugal, 5-8",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "CEUR Workshop Proceedings",
"authors": [],
"year": 2016,
"venue": "",
"volume": "1609",
"issue": "",
"pages": "750--784",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "September, 2016, volume 1609 of CEUR Workshop Proceedings, pages 750-784. CEUR-WS.org.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Reducing gender bias in word-level language models with a gender-equalizing loss function",
"authors": [
{
"first": "Yusu",
"middle": [],
"last": "Qian",
"suffix": ""
},
{
"first": "Urwa",
"middle": [],
"last": "Muaz",
"suffix": ""
},
{
"first": "Ben",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jae",
"middle": [
"Won"
],
"last": "Hyun",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop",
"volume": "",
"issue": "",
"pages": "223--228",
"other_ids": {
"DOI": [
"10.18653/v1/P19-2031"
]
},
"num": null,
"urls": [],
"raw_text": "Yusu Qian, Urwa Muaz, Ben Zhang, and Jae Won Hyun. 2019. Reducing gender bias in word-level language models with a gender-equalizing loss func- tion. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Stu- dent Research Workshop, pages 223-228, Florence, Italy. Association for Computational Linguistics.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Social bias in elicited natural language inferences",
"authors": [
{
"first": "Rachel",
"middle": [],
"last": "Rudinger",
"suffix": ""
},
{
"first": "Chandler",
"middle": [],
"last": "May",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Van Durme",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the First ACL Workshop on Ethics in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "74--79",
"other_ids": {
"DOI": [
"10.18653/v1/W17-1609"
]
},
"num": null,
"urls": [],
"raw_text": "Rachel Rudinger, Chandler May, and Benjamin Van Durme. 2017. Social bias in elicited natural lan- guage inferences. In Proceedings of the First ACL Workshop on Ethics in Natural Language Process- ing, pages 74-79, Valencia, Spain. Association for Computational Linguistics.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Gender bias in coreference resolution",
"authors": [
{
"first": "Rachel",
"middle": [],
"last": "Rudinger",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Naradowsky",
"suffix": ""
},
{
"first": "Brian",
"middle": [],
"last": "Leonard",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Van Durme",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "8--14",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 8-14, New Orleans, LA.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Opinion mining in social media: Modeling, simulating, and forecasting political opinions in the web",
"authors": [
{
"first": "Pawel",
"middle": [],
"last": "Sobkowicz",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Kaschesky",
"suffix": ""
},
{
"first": "Guillaume",
"middle": [],
"last": "Bouchard",
"suffix": ""
}
],
"year": 2012,
"venue": "Social Media in Government -Selections from the 12th Annual International Conference on Digital Government Research (dg.o2011)",
"volume": "29",
"issue": "",
"pages": "470--479",
"other_ids": {
"DOI": [
"10.1016/j.giq.2012.06.005"
]
},
"num": null,
"urls": [],
"raw_text": "Pawel Sobkowicz, Michael Kaschesky, and Guillaume Bouchard. 2012. Opinion mining in social media: Modeling, simulating, and forecasting political opin- ions in the web. Government Information Quarterly, 29(4):470-479. Social Media in Government -Se- lections from the 12th Annual International Confer- ence on Digital Government Research (dg.o2011).",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "Evaluating gender bias in machine translation",
"authors": [
{
"first": "Gabriel",
"middle": [],
"last": "Stanovsky",
"suffix": ""
},
{
"first": "Noah",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "1679--1684",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1164"
]
},
"num": null,
"urls": [],
"raw_text": "Gabriel Stanovsky, Noah A. Smith, and Luke Zettle- moyer. 2019. Evaluating gender bias in machine translation. In Proceedings of the 57th Annual Meet- ing of the Association for Computational Linguistics, pages 1679-1684, Florence, Italy. Association for Computational Linguistics.",
"links": null
},
"BIBREF33": {
"ref_id": "b33",
"title": "What are the biases in my word embedding?",
"authors": [
{
"first": "Nathaniel",
"middle": [],
"last": "Swinger",
"suffix": ""
},
{
"first": "Maria",
"middle": [],
"last": "De-Arteaga",
"suffix": ""
},
{
"first": "Neil",
"middle": [
"Thomas"
],
"last": "Heffernan",
"suffix": ""
},
{
"first": "I",
"middle": [
"V"
],
"last": "Mark",
"suffix": ""
},
{
"first": "D",
"middle": [
"M"
],
"last": "Leiserson",
"suffix": ""
},
{
"first": "Adam Tauman",
"middle": [],
"last": "Kalai",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, AIES '19",
"volume": "",
"issue": "",
"pages": "305--311",
"other_ids": {
"DOI": [
"10.1145/3306618.3314270"
]
},
"num": null,
"urls": [],
"raw_text": "Nathaniel Swinger, Maria De-Arteaga, Neil Thomas Heffernan IV, Mark DM Leiserson, and Adam Tau- man Kalai. 2019. What are the biases in my word embedding? In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, AIES '19, page 305-311, New York, NY, USA. Association for Computing Machinery.",
"links": null
},
"BIBREF34": {
"ref_id": "b34",
"title": "Parallel data, tools and interfaces in OPUS",
"authors": [
{
"first": "J\u00f6rg",
"middle": [],
"last": "Tiedemann",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12)",
"volume": "",
"issue": "",
"pages": "2214--2218",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J\u00f6rg Tiedemann. 2012. Parallel data, tools and inter- faces in OPUS. In Proceedings of the Eighth In- ternational Conference on Language Resources and Evaluation (LREC'12), pages 2214-2218, Istanbul, Turkey. European Language Resources Association (ELRA).",
"links": null
},
"BIBREF35": {
"ref_id": "b35",
"title": "TwiSty: A multilingual twitter stylometry corpus for gender and personality profiling",
"authors": [
{
"first": "Ben",
"middle": [],
"last": "Verhoeven",
"suffix": ""
},
{
"first": "Walter",
"middle": [],
"last": "Daelemans",
"suffix": ""
},
{
"first": "Barbara",
"middle": [],
"last": "Plank",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16)",
"volume": "",
"issue": "",
"pages": "1632--1637",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ben Verhoeven, Walter Daelemans, and Barbara Plank. 2016. TwiSty: A multilingual twitter stylometry cor- pus for gender and personality profiling. In Proceed- ings of the Tenth International Conference on Lan- guage Resources and Evaluation (LREC'16), pages 1632-1637, Portoro\u017e, Slovenia. European Language Resources Association (ELRA).",
"links": null
},
"BIBREF36": {
"ref_id": "b36",
"title": "Multi-passage BERT: A globally normalized BERT model for open-domain question answering",
"authors": [
{
"first": "Zhiguo",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Ng",
"suffix": ""
},
{
"first": "Xiaofei",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "Ramesh",
"middle": [],
"last": "Nallapati",
"suffix": ""
},
{
"first": "Bing",
"middle": [],
"last": "Xiang",
"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 (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "5878--5882",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1599"
]
},
"num": null,
"urls": [],
"raw_text": "Zhiguo Wang, Patrick Ng, Xiaofei Ma, Ramesh Nal- lapati, and Bing Xiang. 2019. Multi-passage BERT: A globally normalized BERT model for open-domain question answering. In Proceedings of the 2019 Conference on Empirical Methods in Nat- ural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5878-5882, Hong Kong, China. Association for Computational Linguistics.",
"links": null
},
"BIBREF37": {
"ref_id": "b37",
"title": "End-to-end open-domain question answering with BERTserini",
"authors": [
{
"first": "Wei",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Yuqing",
"middle": [],
"last": "Xie",
"suffix": ""
},
{
"first": "Aileen",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Xingyu",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Luchen",
"middle": [],
"last": "Tan",
"suffix": ""
},
{
"first": "Kun",
"middle": [],
"last": "Xiong",
"suffix": ""
},
{
"first": "Ming",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations)",
"volume": "",
"issue": "",
"pages": "72--77",
"other_ids": {
"DOI": [
"10.18653/v1/N19-4013"
]
},
"num": null,
"urls": [],
"raw_text": "Wei Yang, Yuqing Xie, Aileen Lin, Xingyu Li, Luchen Tan, Kun Xiong, Ming Li, and Jimmy Lin. 2019. End-to-end open-domain question answering with BERTserini. In Proceedings of the 2019 Confer- ence of the North American Chapter of the Asso- ciation for Computational Linguistics (Demonstra- tions), pages 72-77, Minneapolis, Minnesota. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF38": {
"ref_id": "b38",
"title": "Mitigating unwanted biases with adversarial learning",
"authors": [
{
"first": "Brian",
"middle": [],
"last": "Hu Zhang",
"suffix": ""
},
{
"first": "Blake",
"middle": [],
"last": "Lemoine",
"suffix": ""
},
{
"first": "Margaret",
"middle": [],
"last": "Mitchell",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1145/3278721.3278779"
]
},
"num": null,
"urls": [],
"raw_text": "Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. 2018. Mitigating unwanted biases with adversarial learning. In Proceedings of the 2018",
"links": null
},
"BIBREF39": {
"ref_id": "b39",
"title": "Association for Computing Machinery",
"authors": [],
"year": null,
"venue": "AAAI/ACM Conference on AI, Ethics, and Society, AIES '18",
"volume": "",
"issue": "",
"pages": "335--340",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "AAAI/ACM Conference on AI, Ethics, and Society, AIES '18, page 335-340, New York, NY, USA. As- sociation for Computing Machinery.",
"links": null
},
"BIBREF40": {
"ref_id": "b40",
"title": "Gender bias in multilingual embeddings and cross-lingual transfer",
"authors": [
{
"first": "Jieyu",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Subhabrata",
"middle": [],
"last": "Mukherjee",
"suffix": ""
},
{
"first": "Saghar",
"middle": [],
"last": "Hosseini",
"suffix": ""
},
{
"first": "Kai-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Ahmed",
"middle": [
"Hassan"
],
"last": "Awadallah",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "2896--2907",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.260"
]
},
"num": null,
"urls": [],
"raw_text": "Jieyu Zhao, Subhabrata Mukherjee, Saghar Hosseini, Kai-Wei Chang, and Ahmed Hassan Awadallah. 2020. Gender bias in multilingual embeddings and cross-lingual transfer. In Proceedings of the 58th Annual Meeting of the Association for Computa- tional Linguistics, pages 2896-2907, Online. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF41": {
"ref_id": "b41",
"title": "Gender bias in coreference resolution: Evaluation and debiasing methods",
"authors": [
{
"first": "Jieyu",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Tianlu",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Yatskar",
"suffix": ""
},
{
"first": "Vicente",
"middle": [],
"last": "Ordonez",
"suffix": ""
},
{
"first": "Kai-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "2",
"issue": "",
"pages": "15--20",
"other_ids": {
"DOI": [
"10.18653/v1/N18-2003"
]
},
"num": null,
"urls": [],
"raw_text": "Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Or- donez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15-20, New Orleans, Louisiana. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF42": {
"ref_id": "b42",
"title": "Counterfactual data augmentation for mitigating gender stereotypes in languages with rich morphology",
"authors": [
{
"first": "Ran",
"middle": [],
"last": "Zmigrod",
"suffix": ""
},
{
"first": "Sabrina",
"middle": [
"J"
],
"last": "Mielke",
"suffix": ""
},
{
"first": "Hanna",
"middle": [],
"last": "Wallach",
"suffix": ""
},
{
"first": "Ryan",
"middle": [],
"last": "Cotterell",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "1651--1661",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1161"
]
},
"num": null,
"urls": [],
"raw_text": "Ran Zmigrod, Sabrina J. Mielke, Hanna Wallach, and Ryan Cotterell. 2019. Counterfactual data augmen- tation for mitigating gender stereotypes in languages with rich morphology. In Proceedings of the 57th Annual Meeting of the Association for Computa- tional Linguistics, pages 1651-1661, Florence, Italy. Association for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"type_str": "figure",
"text": "Visualization of classification architecture for topic and author gender",
"num": null
},
"FIGREF1": {
"uris": null,
"type_str": "figure",
"text": "Visualization of debiasing by adversarial training. Left: Adversarial training of topic classifier on author gender, Right: Adversarial training of author gender classifier on topic.",
"num": null
},
"FIGREF2": {
"uris": null,
"type_str": "figure",
"text": "Results for topic classification with adversarial author gender training (F1 scores). Left: Overall results for different \u03bb values. Right: Detail results for \u03bb=1. Original: results from Experiment 1 (cf.",
"num": null
},
"FIGREF3": {
"uris": null,
"type_str": "figure",
"text": "Results for author gender classification with adversarial topic training (F1 scores). Left: Overall results for different \u03bb values. Right: Detail results for \u03bb=0.2. Original: results from Experiment 1 (cf.",
"num": null
},
"FIGREF4": {
"uris": null,
"type_str": "figure",
"text": "Three cases of latent feature space geometry for two attributes: (a) independent, (b) correlated, (c) subsumed",
"num": null
},
"TABREF1": {
"content": "<table><tr><td colspan=\"2\">Document Topic</td><td colspan=\"2\">SciTech Other</td></tr><tr><td>Author Gender</td><td>Male Female</td><td>524 180</td><td>518 296</td></tr></table>",
"html": null,
"num": null,
"type_str": "table",
"text": "Statistics of TED multilingual corpora."
},
"TABREF2": {
"content": "<table/>",
"html": null,
"num": null,
"type_str": "table",
"text": ""
},
"TABREF3": {
"content": "<table><tr><td>) is much larger.</td></tr><tr><td>To address this limitation, we encode the input at</td></tr><tr><td>the paragraph level (cf. Section 3). Specifically, we</td></tr><tr><td>use the final hidden state corresponding to a special</td></tr><tr><td>classification token, [CLS], as the representation</td></tr><tr><td>for the corresponding paragraph. We then obtain</td></tr></table>",
"html": null,
"num": null,
"type_str": "table",
"text": ""
},
"TABREF4": {
"content": "<table/>",
"html": null,
"num": null,
"type_str": "table",
"text": "F1 scores for topic classification (bottom line: majority baseline, identical for all languages)"
},
"TABREF6": {
"content": "<table/>",
"html": null,
"num": null,
"type_str": "table",
"text": "F1 scores for gender classification (bottom line: majority baseline, identical for all languages)"
},
"TABREF7": {
"content": "<table><tr><td colspan=\"4\">: Example of inflectional gender marking in dif-</td></tr><tr><td colspan=\"4\">ferent languages (marking only present in French)</td></tr><tr><td>DE</td><td>ES</td><td>FR</td><td>TR</td></tr><tr><td colspan=\"4\">SciTech/Female 75.0 75.8 83.0 74.4</td></tr></table>",
"html": null,
"num": null,
"type_str": "table",
"text": ""
},
"TABREF8": {
"content": "<table/>",
"html": null,
"num": null,
"type_str": "table",
"text": ""
},
"TABREF10": {
"content": "<table><tr><td>). Lower bias</td></tr></table>",
"html": null,
"num": null,
"type_str": "table",
"text": ""
},
"TABREF11": {
"content": "<table><tr><td>). Lower</td></tr></table>",
"html": null,
"num": null,
"type_str": "table",
"text": ""
},
"TABREF13": {
"content": "<table><tr><td colspan=\"2\">Overall</td></tr><tr><td colspan=\"2\">Adv Baseline</td></tr><tr><td>DE 34.0</td><td>74.1</td></tr><tr><td>ES 37.0</td><td>72.0</td></tr><tr><td>FR 34.6</td><td>78.2</td></tr><tr><td>TR 32.8</td><td>69.2</td></tr></table>",
"html": null,
"num": null,
"type_str": "table",
"text": "Gender classification F-scores of genderdebiased topic classifier and baseline model. For main task evaluation, seeFigure 3."
}
}
}
} |