File size: 115,445 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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T14:31:57.130072Z"
},
"title": "Large-Scale Contextualised Language Modelling for Norwegian",
"authors": [
{
"first": "Andrey",
"middle": [],
"last": "Kutuzov",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Oslo",
"location": {}
},
"email": ""
},
{
"first": "Jeremy",
"middle": [],
"last": "Barnes",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Oslo",
"location": {}
},
"email": ""
},
{
"first": "Erik",
"middle": [],
"last": "Velldal",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Oslo",
"location": {}
},
"email": ""
},
{
"first": "Lilja",
"middle": [],
"last": "\u00d8vrelid",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Oslo",
"location": {}
},
"email": ""
},
{
"first": "Stephan",
"middle": [],
"last": "Oepen",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Oslo",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We present the ongoing NorLM initiative to support the creation and use of very large contextualised language models for Norwegian (and in principle other Nordic languages), including a ready-to-use software environment, as well as an experience report for data preparation and training. This paper introduces the first largescale monolingual language models for Norwegian, based on both the ELMo and BERT frameworks. In addition to detailing the training process, we present contrastive benchmark results on a suite of NLP tasks for Norwegian.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "We present the ongoing NorLM initiative to support the creation and use of very large contextualised language models for Norwegian (and in principle other Nordic languages), including a ready-to-use software environment, as well as an experience report for data preparation and training. This paper introduces the first largescale monolingual language models for Norwegian, based on both the ELMo and BERT frameworks. In addition to detailing the training process, we present contrastive benchmark results on a suite of NLP tasks for Norwegian.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "In this work, we present NorLM, an ongoing community initiative and emerging collection of largescale contextualised language models for Norwegian. We here introduce the NorELMo and Nor-BERT models, that have been trained on around two billion tokens of running Norwegian text. We describe the training procedure and compare these models with the multilingual mBERT model (Devlin et al., 2019) , as well as an additional Norwegian BERT model developed contemporaneously, with some interesting differences in training data and setup. We report results over a number of Norwegian benchmark datasets, addressing a broad range of diverse NLP tasks: part-of-speech tagging, negation resolution, sentence-level and fine-grained sentiment analysis and named entity recognition (NER).",
"cite_spans": [
{
"start": 372,
"end": 393,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "All the models are publicly available for download from the Nordic Language Processing Lab-oratory (NLPL) Vectors Repository 1 with a CC BY 4.0 license. They are also accessible locally, together with the training and supporting software, on the two national superclusters Puhti and Saga, in Finland and Norway, respectively, which are available to university NLP research groups in Northern Europe through the Nordic Language Processing Laboratory (NLPL). 2 The NorBERT model is in addition served via the Huggingface Transformers model hub. 3 NorLM is a joint effort of the projects EOSC-Nordic (European Open Science Cloud) and SANT (Sentiment Analysis for Norwegian), coordinated by the Language Technology Group (LTG) at the University of Oslo. The goal of this work is to provide these models and supporting tools for researchers and developers in Natural Language Processing (NLP) for the Norwegian language. We do so in the hope of facilitating scientific experimentation with and practical applications of state-of-the-art NLP architectures, as well as to enable others to develop their own large-scale models, for example for domain-or applicationspecific tasks, language variants, or even other languages than Norwegian. Under the auspices of the NLPL use case in EOSC-Nordic, we are also coordinating with colleagues in Denmark, Finland, and Sweden on a collection of large contextualised language models for the Nordic languages, including language variants or related groups of languages, as linguistically or technologically appropriate.",
"cite_spans": [
{
"start": 457,
"end": 458,
"text": "2",
"ref_id": null
},
{
"start": 543,
"end": 544,
"text": "3",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Bokm\u00e5l and Nynorsk There are two official standards for written Norwegian; Bokm\u00e5l, the main variety, and Nynorsk, used by 10-15% of the Norwegian population. Norwegian language legislation specifies that minimally 25% of the written public service information should be in Nynorsk. While the two varieties are closely related, there can also be relatively large differences lexically (though often with a large degree of overlap on the character-level still). Several previous studies have indicated that joint modeling of Bokm\u00e5l and Nynorsk works well for many NLP tasks, like tagging and parsing and NER (J\u00f8rgensen et al., 2020) . The contextualised language models presented in this paper are therefore trained jointly on both varieties, but with the minority variant Nynorsk represented by comparatively less data than Bokm\u00e5l (reflecting the natural usage).",
"cite_spans": [
{
"start": 606,
"end": 630,
"text": "(J\u00f8rgensen et al., 2020)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Datasets For all our models presented below, we used the following training corpora:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "1. Norsk Aviskorpus (NAK), a collection of Norwegian news texts 4 (both Bokm\u00e5l and Nynorsk) from 1998 to 2019; 1.7 billion words;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "2. Bokm\u00e5l Wikipedia dump from September 2020; 160 million words;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "3. Nynorsk Wikipedia dump from September 2020; 40 million words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "The corpora contain ordered sentences (which is important for BERT-like models, because one of their training tasks is next sentence prediction). In total, our training corpus comprises about two billion (1,907,072,909) word tokens in 203 million (202, 802, 665) sentences.",
"cite_spans": [
{
"start": 247,
"end": 252,
"text": "(202,",
"ref_id": null
},
{
"start": 253,
"end": 257,
"text": "802,",
"ref_id": null
},
{
"start": 258,
"end": 262,
"text": "665)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "We conducted the following pre-processing steps:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "1. Wikipedia texts were extracted from the dumps using the segment wiki script from the Gensim project (\u0158eh\u016f\u0159ek and Sojka, 2010).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "2. For the news texts from Norwegian Aviskorpus, we performed de-tokenization and conversion to UTF-8 encoding, where required.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "3. The resulting corpus was sentencesegmented using Stanza (Qi et al., 2020) . We left blank lines between documents (and 3 Prerequisites: software and computing",
"cite_spans": [
{
"start": 59,
"end": 76,
"text": "(Qi et al., 2020)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Developing very large contextualised language models is no small challenge, both in terms of engineering sophistication and computing demands. Training ELMo-and in particular BERT-like models presupposes access to specialised hardwaregraphical processing units (GPUs) -over extended periods of time. Compared to the original work at Google or to our sister initiative at the National Library of Norway (see below), our two billion tokens in Norwegian training data can be characterised as moderate in size. Nevertheless, training a single NorBERT model requires close to one full year of GPU utilisation, which through parallelization over multiple compute nodes, each featuring four GPUs, could be completed in about three weeks of wall clock time. At this scale, premium software efficiency and effective parallelization are prerequisites, not only to allow repeated incremental training and evaluation cycles to complete in practical intervals, but equally so for cost-efficient utilisation of scarce, shared computing resources and, ultimately, a shred of environmental sustainability.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "To prepare the NorLM software environment, we have teamed up with support staff at the Norwegian national e-infrastructure provider, Uninett Sigma2, and developed a fully automated and modularised installation procedure using the Easy-Build framework (https://easybuild.io). All necessary tools are compiled from source with the right set of hardware-specific optimizations and platform-specific optimised libraries for basic linear algebra ('math kernels') and communication across multiple compute nodes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "This approach to software provisioning makes it possible to (largely) automatically create fully parallel training and experimentation environments on multiple computing infrastructures -in our work to date two national HPC superclusters, in Norway and Finland, but in principle just as much any suitable local GPU cluster. In our view, making available both a ready-to-run software environment on Nordic national e-infrastructures, where university research groups typically can gain nocost access, coupled with the recipe for recreating the environment on other HPC systems, may contribute to 'democratising' large-scale NLP research; if nothing else, it eliminates dependency on commercial cloud computing services.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Large-scale deep learning language models (LM) are important components of current NLP systems. They are often based on BERT (Bidirectional Encoder Representations from Transformers) (Devlin et al., 2019) and other contextualised architectures. A number of language-specific initiatives have in recent years released monolingual versions of these models for a number of languages (Fares et al., 2017; Kutuzov and Kuzmenko, 2017; Virtanen et al., 2019; de Vries et al., 2019; Ul\u010dar and Robnik-\u0160ikonja, 2020; Koutsikakis et al., 2020; Nguyen and Nguyen, 2020; Farahani et al., 2020; Malmsten et al., 2020) . For our purposes, the most important such previous training effort is that of Virtanen et al. (2019) on creating a BERT model for Finnish -FinBERT 5as our training setup for creating NorBERT builds heavily on this; see Section 6 for more details.",
"cite_spans": [
{
"start": 183,
"end": 204,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF1"
},
{
"start": 380,
"end": 400,
"text": "(Fares et al., 2017;",
"ref_id": "BIBREF4"
},
{
"start": 401,
"end": 428,
"text": "Kutuzov and Kuzmenko, 2017;",
"ref_id": "BIBREF10"
},
{
"start": 429,
"end": 451,
"text": "Virtanen et al., 2019;",
"ref_id": "BIBREF25"
},
{
"start": 452,
"end": 474,
"text": "de Vries et al., 2019;",
"ref_id": "BIBREF26"
},
{
"start": 475,
"end": 506,
"text": "Ul\u010dar and Robnik-\u0160ikonja, 2020;",
"ref_id": "BIBREF22"
},
{
"start": 507,
"end": 532,
"text": "Koutsikakis et al., 2020;",
"ref_id": "BIBREF6"
},
{
"start": 533,
"end": 557,
"text": "Nguyen and Nguyen, 2020;",
"ref_id": "BIBREF14"
},
{
"start": 558,
"end": 580,
"text": "Farahani et al., 2020;",
"ref_id": "BIBREF3"
},
{
"start": 581,
"end": 603,
"text": "Malmsten et al., 2020)",
"ref_id": null
},
{
"start": 684,
"end": 706,
"text": "Virtanen et al. (2019)",
"ref_id": "BIBREF25"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "4"
},
{
"text": "Many low-resource languages do not have dedicated monolingual large-scale language models, and instead resort to using a multilingual model, such as Google's multilingual BERT model -mBERT -which was trained on data that also included Norwegian. Up until the release of the models described in the current paper, mBERT was the only BERT-instance that could be used for Norwegian. 6 Another widely used architecture for contextualised LMs is Embeddings From Language Models or ELMo (Peters et al., 2018) . The ElmoFor-ManyLangs initiative (Che et al., 2018) trained and released monolingual ELMo models for a wide range of different languages, including Norwegian (with separate models for Bokm\u00e5l and Nynorsk). However, these models were trained on very modestly sized corpora of 20 million words for each language (randomly sampled from Wikipedia dumps and Common Crawl data).",
"cite_spans": [
{
"start": 380,
"end": 381,
"text": "6",
"ref_id": null
},
{
"start": 481,
"end": 502,
"text": "(Peters et al., 2018)",
"ref_id": "BIBREF17"
},
{
"start": 538,
"end": 556,
"text": "(Che et al., 2018)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "4"
},
{
"text": "In a parallel effort to that of the current paper, the AI Lab of the National Library of Norway, through their Norwegian Transformer Model (No-TraM) project, has released a Norwegian BERT (Base, cased) model dubbed NB-BERT (Kummervold et al., 2021 ). 7 The model is trained on the Colossal Norwegian Corpus, reported to comprise close to 18,5 billion words (109.1 GB of text).",
"cite_spans": [
{
"start": 223,
"end": 247,
"text": "(Kummervold et al., 2021",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "4"
},
{
"text": "In raw numbers, this is about ten times more than the corpus we use for training the NorLM models. However, the vast majority of this is from OCR'ed historical sources, which is bound to introduce at least some noise. In Section 7 below, we demonstrate that in some NLP tasks, a language model trained on less (but arguably cleaner) data can outperform a model trained on larger but noisy corpora.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "4"
},
{
"text": "NorELMo is a set of bidirectional recurrent ELMo language models trained from scratch on the Norwegian corpus described in Section 1. They can be used as a source of contextualised token representations for various Norwegian natural language processing tasks. As we show below, in many cases, they present a viable alternative to Transformer-based models like BERT. Their performance is often only marginally lower, while the compute time required to adapt the model to the task at hand can be an order of magnitude less on identical hardware.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorELMo",
"sec_num": "5"
},
{
"text": "Currently we present two models, with more following in the future:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorELMo",
"sec_num": "5"
},
{
"text": "1. NorELMo 30 : 30,000 most frequent words in the vocabulary 2. NorELMo 100 : 100,000 most frequent words in the vocabulary Note that independent of the vocabulary size, both NorELMo 30 and NorELMo 100 can process arbitrary word tokens, due to the ELMo architecture (where the first CNN layer converts input strings to non-contextual word embeddings). Thus, the size of the vocabulary controls only the number of words used as targets for the language modelling task in the course of training. Supposedly, the model with a larger vocabulary is more effective in treating less frequent words at the cost of being less effective with more frequent words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorELMo",
"sec_num": "5"
},
{
"text": "Each model was trained for 3 epochs with batch size 192. We employed a version of the original ELMo training code from Peters et al. (2018) updated to work better with the recent TensorFlow versions. All the hyperparameters were left at their default values, except the LSTM dimensionality reduced to 2,048 from the default 4,096 (in our experience, this rarely influences performance). Training of each model took about 100 hours on four NVIDIA P100 GPUs.",
"cite_spans": [
{
"start": 119,
"end": 139,
"text": "Peters et al. (2018)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "NorELMo",
"sec_num": "5"
},
{
"text": "These are the first ELMo models for Norwegian trained on a large corpus. As has already been mentioned, the Norwegian ELMo models from the ElmoForManyLangs project (Che et al., 2018) were trained on very small corpora samples and seriously under-perform on semanticrelated NLP tasks, although they can yield impressive results on POS tagging and syntactic parsing (Zeman et al., 2018) . In addition, they were trained with custom code modifications and can be used only with the custom ElmoForManyLangs library. On the other hand, our NorELMo models are fully compatible both with the original ELMo implementation by Peters et al. (2018) and with the more modern simple elmo Python library provided by us. 8 The vocabularies are published together with the models. For different tasks, different models can be better, as we show below. The published packages contain both TensorFlow checkpoints (for possible fine-tuning, if need be) and model files in the standard Hierarchical Data Format (HDF5) for easier inference usage. In addition, we have setup ELMoViz, a demo web service to explore Norwegian ELMo models. 9",
"cite_spans": [
{
"start": 164,
"end": 182,
"text": "(Che et al., 2018)",
"ref_id": "BIBREF0"
},
{
"start": 364,
"end": 384,
"text": "(Zeman et al., 2018)",
"ref_id": "BIBREF28"
},
{
"start": 617,
"end": 637,
"text": "Peters et al. (2018)",
"ref_id": "BIBREF17"
},
{
"start": 706,
"end": 707,
"text": "8",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "NorELMo",
"sec_num": "5"
},
{
"text": "Our NorBERT model is trained from scratch for Norwegian, and can be used in exactly the same way as any other BERT-like model. The NorBERT training setup heavily builds on prior work on Fin-BERT conducted at the University of Turku (Virtanen et al., 2019).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorBERT",
"sec_num": "6"
},
{
"text": "NorBERT features a custom WordPiece vocabulary which is case-sensitive and includes accented characters. It has much better coverage of Norwegian words than the mBERT model or NB-BERT (which uses the same vocabulary as mBERT). This is clearly seen on the example of the tokenization performed by both for the Norwe-gian sentence 'Denne gjengen h\u00e5per at de sammen skal bidra til\u00e5 gi kvinnefotballen i Kristiansand et lenge etterlengtet l\u00f8ft'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorBERT",
"sec_num": "6"
},
{
"text": "\u2022 mBERT/NB-BERT: 'Denne g ##jeng ##en h ##\u00e5 ##per at de sammen skal bid ##ra til a gi k ##vinne ##fo ##t ##ball ##en i Kristiansand et lenge etter ##len ##gte ##t l ##\u00f8 ##ft'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorBERT",
"sec_num": "6"
},
{
"text": "\u2022 NorBERT: 'Denne gjengen h\u00e5per at de sammen skal bidra til\u00e5 gi kvinne ##fotball ##en i Kristiansand et lenge etterl ##engt ##et l\u00f8ft'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorBERT",
"sec_num": "6"
},
{
"text": "NorBERT tokenization splits the sentence into pieces which much better reflect the real Norwegian words and morphemes (cf. 'k vinne fo t ball en' versus 'kvinne fotball en'). We believe this to be extremely important for more linguistically-oriented studies, where it is critical to deal with words, not with arbitrarily fragmented pieces (even if they are well-performing in practical tasks).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorBERT",
"sec_num": "6"
},
{
"text": "The vocabulary for the model is of size 30,000. It is much less than the 120,000 of mBERT, but it is compensated by these entities being almost exclusively Norwegian. The vocabulary was generated from raw text, without, e.g., separating punctuation from word tokens. This means one can feed raw text into NorBERT.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NorBERT",
"sec_num": "6"
},
{
"text": "For the vocabulary generation, we used the Sen-tencePiece algorithm (Kudo, 2018) and Tokenizers library. 10 The resulting Tokenizers model was converted to the standard BERT WordPiece format. The final vocabulary contains several thousand unused wordpiece slots which can be filled in with task-specific lexical entries for further finetuning by future NorBERT users.",
"cite_spans": [
{
"start": 68,
"end": 80,
"text": "(Kudo, 2018)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "NorBERT",
"sec_num": "6"
},
{
"text": "NorBERT corresponds in its configuration to the Google's Bert-Base Cased for English, with 12 layers and hidden size 768 (Devlin et al., 2019) . We used the standard masked language modeling and next sentence prediction losses with the LAMB optimizer (You et al., 2020) . The model was trained on the Norwegian academic HPC system called Saga. Most of the time the training process was distributed across 4 compute nodes and 16 NVIDIA P100 GPUs. Overall, it took approximately 3 weeks (more than 500 hours). 2019, we employed the BERT implementation by NVIDIA 11 , which allows fast multi-node and multi-GPU training.",
"cite_spans": [
{
"start": 121,
"end": 142,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF1"
},
{
"start": 251,
"end": 269,
"text": "(You et al., 2020)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training technicalities",
"sec_num": "6.1"
},
{
"text": "We made minor changes to this code, mostly to adapt it to the newer TensorFlow versions. All these patches and the utilities we used at the preprocessing, training and evaluation stages are published in our GitHub repository. 12 Instructions to reproduce the training setup with the EasyBuild software build and installation framework are also available. 13",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training technicalities",
"sec_num": "6.1"
},
{
"text": "Phase 1 (training with maximum sequence length of 128) was done with batch size 48 and global batch size 48*16=768. Since one global batch contains 768 sentences, approximately 265,000 training steps constitute 1 epoch (one pass over the whole corpus). We have done 3 epochs: 795,000 training steps.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training workflow",
"sec_num": "6.2"
},
{
"text": "Phase 2 (training with maximum sequence length of 512) was done with batch size 8 and global batch size 8*16=128. We aimed at mimicking the original BERT in that at Phase 2 the model should see about 1/9 of the number of sentences seen during Phase 1. Thus, we needed about 68 million sentences, which at the global batch size of 128 boils down to 531,000 training steps more.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training workflow",
"sec_num": "6.2"
},
{
"text": "The loss plots are shown in Figure 1 (the training was on pause on December 25 and 26, since we were solving problems with mixed precision 11 https://github.com/NVIDIA/ DeepLearningExamples/tree/master/ TensorFlow/LanguageModeling/BERT, version 20.06.08",
"cite_spans": [],
"ref_spans": [
{
"start": 28,
"end": 36,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Training workflow",
"sec_num": "6.2"
},
{
"text": "12 https://github.com/ltgoslo/NorBERT 13 http://wiki.nlpl.eu/index.php/Eosc/ pretraining/nvidia ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training workflow",
"sec_num": "6.2"
},
{
"text": "This section presents benchmark results across a range of different tasks. We compare NorELMO and NorBERT to both mBERT and to the recently released NB-BERT model described in Section 4. Where applicable, we show separate evaluation results for Bokm\u00e5l and Nynorsk. Below we first provide an overview of the different tasks and the corresponding classifiers that we train, before turning to discuss the results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "7"
},
{
"text": "We start by briefly describing each task and associated dataset, in addition to the architectures we use. The sentence counts for the different datasets and their train, dev. and test splits are provided in Table 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 207,
"end": 214,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Task descriptions",
"sec_num": "7.1"
},
{
"text": "Part-of-speech tagging The Norwegian Dependency Treebank (NDT) (Solberg et al., 2014) in-cludes annotation of POS tags for both Bokm\u00e5l and Nynorsk. NDT has also been converted to the Universal Dependencies format (\u00d8vrelid and Hohle, 2016; and this is the version we are using here (for UD 2.7) for predicting UPOS tags. We use a typical sequence labelling approach with the BERT models, adding a linear layer after the final token representations and taking the softmax to get token predictions. We fine-tune all parameters for 20 epochs, using a learning rate of 2e-5, a training batch size of 8, max length of 256, and keep the best model on the development set. ELMo models were not fine-tuned, following the recommendations from Peters et al. (2019) . Instead we trained a simple neural classifier (a feed forward network with one hidden layer of size 128, ReLU non-linear activation function and dropout), using ELMo token embeddings as features. The random seed has been kept fixed all the time. Models are evaluated on accuracy.",
"cite_spans": [
{
"start": 213,
"end": 238,
"text": "(\u00d8vrelid and Hohle, 2016;",
"ref_id": "BIBREF15"
},
{
"start": 733,
"end": 753,
"text": "Peters et al. (2019)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Task descriptions",
"sec_num": "7.1"
},
{
"text": "Named entity recognition The NorNE 14 dataset annotates the UD-version of NDT with a rich set of entity types (J\u00f8rgensen et al., 2020) . The evaluation metrics here is 'strict' micro F 1 , requiring both the correct entity type and exact match of boundary surface string. We predict 8 entity types: Person (PER), Organisation (ORG), Location (LOC), Geo-political entity, with a locative sense (GPE-LOC), Geo-political entity, with an organisation sense (GPE-ORG), Product (PROD), Event (EVT), Nominals derived from names (DRV). The evaluation is done using the code for the SemEval'13 Task 9 15 .",
"cite_spans": [
{
"start": 110,
"end": 134,
"text": "(J\u00f8rgensen et al., 2020)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Task descriptions",
"sec_num": "7.1"
},
{
"text": "We cast the named entity recognition problem as a sequence labelling task, using a BIO label encoding. For the BERT-based models, we solve it by fine-tuning the pre-trained model on the NorNE dataset for 20 epochs with early stopping and batch size 32. The resulting model is applied to the test set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Task descriptions",
"sec_num": "7.1"
},
{
"text": "For ELMo models, we infer contextualised token embeddings (averaged representations across all 3 layers) for all words. Then, these token embeddings are fed to a neural classifier with dropout, identical to the one we used for POS tagging earlier. This classifier is also trained for 20 epochs with early stopping and batch size 32.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Task descriptions",
"sec_num": "7.1"
},
{
"text": "Fine-grained sentiment analysis NoReC fine is a dataset 16 comprising a subset of the Norwegian Review Corpus (NoReC; Velldal et al., 2018) annotated for sentiment holders, targets, expressions, and polarity, as well as the relationships between them (\u00d8vrelid et al., 2020) . We here cast the problem as a graph prediction task and train a graph parser (Dozat and Manning, 2018; Kurtz et al., 2020) to predict sentiment graphs. The parser creates token-level representations which is the concatenation of a word embedding, POS tag embedding, lemma embedding, and character embedding created by a character-based LSTM. We further augment these representations with contextualised embeddings from each model. Models are trained for 100 epochs, keeping the best model on development F 1 . For span extraction (holders, targets, expressions), we evaluate token-level F 1 , and the common Targeted F 1 metric, which requires correctly extracting a target (strict) and its polarity. We also evaluate Labelled and Unlabelled F 1 , which correspond to Labelled and Unlabelled Attachment in dependency parsing. Finally, we evaluate on Sentiment Graph F 1 (SF 1 ) and Non-polar Sentiment Graph F 1 (NSF 1 . SF 1 requires predicting all elements (holder, target, expression, polarity) and their relationships (NSF 1 removes the polarity). A true positive is defined as an exact match at graph-level, weighting the overlap in predicted and gold spans for each element, averaged across all three spans. For precision we weight the number of correctly predicted tokens divided by the total number of predicted tokens (for recall, we divide instead by the number of gold tokens). We allow for empty holders and targets.",
"cite_spans": [
{
"start": 118,
"end": 139,
"text": "Velldal et al., 2018)",
"ref_id": "BIBREF23"
},
{
"start": 251,
"end": 273,
"text": "(\u00d8vrelid et al., 2020)",
"ref_id": null
},
{
"start": 353,
"end": 378,
"text": "(Dozat and Manning, 2018;",
"ref_id": "BIBREF2"
},
{
"start": 379,
"end": 398,
"text": "Kurtz et al., 2020)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Task descriptions",
"sec_num": "7.1"
},
{
"text": "We further evaluate on the task of sentence-level binary (positive or negative) polarity classification, using labels that we derive from NoReC fine described above. We create the dataset for this by aggregating the fine-grained annotations to the sentence-level, removing sentences with mixed or no sentiment. The resulting dataset, NoReC sentence , is made publicly available. 17 For the BERT models, we use the [CLS] embedding of the last layer as a representation for the sentence and pass this to a softmax layer for classification. We fine-tune the models in the same way as for the POS tagging task, training the models for 20 epochs and keeping the model that performs best on the development data. For ELMo models, we used a BiLSTM with global max pooling, taking ELMo token embeddings from the top layer as an input. The evaluation metric is macro F 1 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence-level binary sentiment classification",
"sec_num": null
},
{
"text": "Negation detection Finally, the NoReC fine dataset has recently been annotated with negation cues and their corresponding in-sentence scopes (Maehlum et al., 2021) . The resulting dataset is dubbed NoReC neg . 18 We use the same graph-based modeling approach as described for fine-grained sentiment above. We evaluate on the same metrics as in the *SEM 2012 shared task (Morante and Blanco, 2012) : cue-level F 1 (CUE), scope token F 1 over individual tokens (ST), and the combined full negation F 1 (FN).",
"cite_spans": [
{
"start": 141,
"end": 163,
"text": "(Maehlum et al., 2021)",
"ref_id": null
},
{
"start": 210,
"end": 212,
"text": "18",
"ref_id": null
},
{
"start": 370,
"end": 396,
"text": "(Morante and Blanco, 2012)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence-level binary sentiment classification",
"sec_num": null
},
{
"text": "We present the results for the various benchmarking tasks below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "7.2"
},
{
"text": "POS tagging As can be seen from Table 2 , Nor-BERT outperforms mBERT on both tasks: on POS tagging for Bokm\u00e5l by 5 percentage points and 1 percentage point for Nynorsk. NorBERT is almost on par with NB-BERT on POS tagging. NorELMo models are outperformed by NB-BERT and Nor-BERT, but are on par with mBERT in POS tagging. Note that their adaptation to the tasks (extracting token embeddings and learning a classifier) takes 30x less time than with the BERT models.",
"cite_spans": [],
"ref_spans": [
{
"start": 32,
"end": 39,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "7.2"
},
{
"text": "18 https://github.com/ltgoslo/norec_neg Table 3 : NER evaluation scores (micro F 1 ) of the NorLM models on the NorNE test set in comparison with other large pre-trained models for Norwegian. Running time is given in minutes for the Bokm\u00e5l part (on 1 NVIDIA P100 GPU).",
"cite_spans": [],
"ref_spans": [
{
"start": 40,
"end": 47,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "7.2"
},
{
"text": "See Figure 2 for the examples of training dynamics of the Nynorsk model. Table 3 shows the performance on the NER task. NB-BERT is the best on both Bokm\u00e5l and Nynorsk, closely followed by NorBERT. Unsurprisingly, mBERT falls behind all the models trained for Norwegian, when evaluated on Bokm\u00e5l data. With Nynorsk, it manages to outperform NorELMo. Bokm\u00e5l is presumably dominant in the training corpora of both. However, in the course of fine-tuning, mBERT seems to be able to adapt to the specifics of Nynorsk. Since our ELMo setup did not include the fine-tuning step, the NorELMo models' adaptation abilities were limited by what can be learned from contextualised token embeddings produced by a frozen model. Still, when used on the data more similar to the training corpus (Bokm\u00e5l), ELMo achieves competitive results even without any fine-tuning.",
"cite_spans": [],
"ref_spans": [
{
"start": 4,
"end": 12,
"text": "Figure 2",
"ref_id": null
},
{
"start": 73,
"end": 80,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "7.2"
},
{
"text": "In terms of computational efficiency, the adaptation of ELMo models to this task requires 6x less time than mBERT or NB-BERT and 4x less time than NorBERT. Note also that the NorBERT model takes less time to fine-tune than the NB-BERT model (although the number of epochs was exactly the same), because of a smaller vocabulary, and thus less parameters in the model. Again, in this case an NLP practitioner has a rich spectrum of tools to choose from, depending on whether speed or performance on the downstream task is prioritised. Table 4 shows that NorBERT outperforms mBERT on all metrics and NB-BERT on all but SF 1 , although the differences between NorBERT and NB-BERT are gen- Table 4 : Average score of NorLM models on fine-grained sentiment (5 runs with set random seeds).",
"cite_spans": [],
"ref_spans": [
{
"start": 533,
"end": 540,
"text": "Table 4",
"ref_id": null
},
{
"start": 685,
"end": 692,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Named entity recognition",
"sec_num": null
},
{
"text": "Bold denotes the best result on each metric.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fine-grained sentiment analysis",
"sec_num": null
},
{
"text": "[1] Span extraction baseline from \u00d8vrelid et al. (2020) , which uses a BiLSTM CRF with pretrained fastText embeddings.",
"cite_spans": [
{
"start": 34,
"end": 55,
"text": "\u00d8vrelid et al. (2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fine-grained sentiment analysis",
"sec_num": null
},
{
"text": "NorELMo 30 75.0 NorELMo 100 75.0 mBERT 67.7 NB-BERT 83.9 NorBERT 77.1 erally small. On this task the NorELMo models generally outperform mBERT as well. However, unlike in the previous tasks, the running times here are similar for BERT and ELMo models, since no finetuning was applied (the same is true for negation detection). We furthermore compare with the previous best model (\u00d8vrelid et al., 2020) , a span extraction model which uses a single-layer Bidirectional LSTM with Conditional Random Field inference, and an embedding layer initialized with fastText vectors trained on the NoWaC corpus. All approaches using language models outperform the previous baseline by a large margin on the span extraction tasks. 19 NorBERT, in particular, achieves improvements of 20.6 percentage points on Holder F 1 (24.9 and 25.8 on Target and Exp. F 1 , respectively). Table 5 shows that NorBERT outperforms mBERT by 9.4 percentage points on sentiment analysis. However, it seems that in binary sentiment classification the sheer amount of training data starts to show its benefits, and NB-BERT outperforms NorBERT by 6.8 points. NorELMo models outperform mBERT by 7.3 points. Figure 2 shows the training dynamics of the models. Table 6 we can see that mBERT gives the best overall results, followed by NorBERT and NorELMo 100 . NB-BERT and NorELMo 30 perform worse than the others on Scope token F 1 (ST) and full negation F 1 (FN), while all models perform similarly at cue-level F 1 (CUE). We hypothesise that the structural similarity of negation across many of the pretraining languages gives mBERT an advantage, but it is still surprising that it outperforms NB-BERT and Nor-BERT.",
"cite_spans": [
{
"start": 379,
"end": 401,
"text": "(\u00d8vrelid et al., 2020)",
"ref_id": null
},
{
"start": 718,
"end": 720,
"text": "19",
"ref_id": null
}
],
"ref_spans": [
{
"start": 862,
"end": 869,
"text": "Table 5",
"ref_id": "TABREF6"
},
{
"start": 1170,
"end": 1178,
"text": "Figure 2",
"ref_id": null
},
{
"start": 1222,
"end": 1229,
"text": "Table 6",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Model F 1",
"sec_num": null
},
{
"text": "In the future, separate Bokm\u00e5l and Nynorsk BERT models are planned, and we further expect to train and evaluate models with a higher number of epochs over the training corpus. While we plan to develop additional monolingual Norwegian models based on other contextualised LM architectures Figure 2 : Per-epoch performance on training and development data for two of the tasks. Left: accuracy for POS tagging (Norwegian Nynorsk). Right: F 1 for binary sentiment classification.",
"cite_spans": [],
"ref_spans": [
{
"start": 288,
"end": 296,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Future plans",
"sec_num": "8"
},
{
"text": "beyond BERT and ELMo, we would also be interested to explore the usefulness of multilingual models restricted to Scandinavian languages. Further streamlining of the benchmarking process, in terms of both data access and computation of metrics, is something we also want to address in future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Future plans",
"sec_num": "8"
},
{
"text": "In addition, the ready availability of a highly optimised software stack on multiple HPC systems (published as part of NorLM) may contribute to other researchers developing very large contextualised language models for additional languages or language variants, e.g. domain-or application-specific sub-corpora. We hope that more pre-trained NLP models for Norwegian from both academy and industry will be openly released, making it possible to study the interplay between training corpora sizes, hyperparameters, pre-preprocessing decisions and performance in different tasks. At the same time, given the resource demands and sustainability issues related to training such models, we believe it will be important to coordinate efforts and we hope to collaborate closely with other players moving forward.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Future plans",
"sec_num": "8"
},
{
"text": "This paper has described the first outcomes of NorLM, an initiative coordinated by the Language Technology Group at the University of Oslo seeking to provide Norwegian (and Nordic) largescale contextualised language models, while simultaneously focusing on maintaining a re-usable software environment for model development on national and Nordic HPC infrastructure. We have here described the training and testing of NorELMo and NorBERT -the first large-scale monolingual LMs for Norwegian. We have benchmarked the models across a wide array of Norwegian NLP tasks, also comparing to the multilingual mBERT model and another large-scale LM for Norwegian developed in parallel work, NB-BERT, trained on large amounts of text from historical sources. The results show that while the monolingual models tend to yield better results, which particular model ranks first varies across tasks. This underscores the importance of building an ecosystem of diversified models, accompanied by systematic benchmarking.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summary",
"sec_num": "9"
},
{
"text": "http://vectors.nlpl.eu/repository 2 http://www.nlpl.eu 3 https://huggingface.co/ltgoslo/ norbert",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://www.nb.no/sprakbanken/ ressurskatalog/oai-nb-no-sbr-4/ sections in the case of Wikipedia) so that the 'next sentence prediction' task of BERT does not span between documents.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/TurkuNLP/FinBERT 6 A BERT model trained on Norwegian data was published at https://github.com/botxo/nordic_bert in the beginning of 2020. However, the vocabulary of this model seems to be broken, and to the best of our knowledge nobody has achieved any meaningful results with it.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/NBAiLab/notram",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://pypi.org/project/simple-elmo/ 9 http://vectors.nlpl.eu/explore/ embeddings/en/contextual/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/huggingface/ tokenizers",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/ltgoslo/norne 15 https://github.com/davidsbatista/ NER-Evaluation",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/ltgoslo/norec_fine 17 https://github.com/ltgoslo/norec_ sentence",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "\u00d8vrelid et al. (2020) only perform span extraction. Therefore, it is not possible to compare the other metrics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "The NorLM resources are being developed on the Norwegian national super-computing services operated by UNINETT Sigma2, the National Infrastructure for High Performance Computing and Data Storage in Norway, as well as on the Finnish national supercomputing facilities operated by the CSC IT Center for Science. Software provisioning was financially supported through the European EOSC-Nordic project; data preparation and evaluation were supported by the SANT project (Sentiment Analysis for Norwegian Text), funded by the Research Council of Norway (grant number 270908). We are indebted to all funding agencies involved, the University of Oslo, and the Norwegian tax payer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Towards better UD parsing: Deep contextualized word embeddings, ensemble, and treebank concatenation",
"authors": [
{
"first": "Wanxiang",
"middle": [],
"last": "Che",
"suffix": ""
},
{
"first": "Yijia",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Yuxuan",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Bo",
"middle": [],
"last": "Zheng",
"suffix": ""
},
{
"first": "Ting",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies",
"volume": "",
"issue": "",
"pages": "55--64",
"other_ids": {
"DOI": [
"10.18653/v1/K18-2005"
]
},
"num": null,
"urls": [],
"raw_text": "Wanxiang Che, Yijia Liu, Yuxuan Wang, Bo Zheng, and Ting Liu. 2018. Towards better UD parsing: Deep contextualized word embeddings, ensemble, and treebank concatenation. In Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, pages 55-64, Brussels, Belgium. Association for Compu- tational Linguistics.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"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
},
"BIBREF2": {
"ref_id": "b2",
"title": "Simpler but more accurate semantic dependency parsing",
"authors": [
{
"first": "Timothy",
"middle": [],
"last": "Dozat",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "484--490",
"other_ids": {
"DOI": [
"10.18653/v1/P18-2077"
]
},
"num": null,
"urls": [],
"raw_text": "Timothy Dozat and Christopher D. Manning. 2018. Simpler but more accurate semantic dependency parsing. In Proceedings of the 56th Annual Meet- ing of the Association for Computational Linguis- tics (Volume 2: Short Papers), pages 484-490, Mel- bourne, Australia. Association for Computational Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Parsbert: Transformer-based model for Persian language understanding",
"authors": [
{
"first": "Mehrdad",
"middle": [],
"last": "Farahani",
"suffix": ""
},
{
"first": "Mohammad",
"middle": [],
"last": "Gharachorloo",
"suffix": ""
},
{
"first": "Marzieh",
"middle": [],
"last": "Farahani",
"suffix": ""
},
{
"first": "Mohammad",
"middle": [],
"last": "Manthouri",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2005.12515"
]
},
"num": null,
"urls": [],
"raw_text": "Mehrdad Farahani, Mohammad Gharachorloo, Marzieh Farahani, and Mohammad Manthouri. 2020. Parsbert: Transformer-based model for Persian language understanding. arXiv preprint arXiv:2005.12515.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Word vectors, reuse, and replicability: Towards a community repository of large-text resources",
"authors": [
{
"first": "Murhaf",
"middle": [],
"last": "Fares",
"suffix": ""
},
{
"first": "Andrey",
"middle": [],
"last": "Kutuzov",
"suffix": ""
},
{
"first": "Stephan",
"middle": [],
"last": "Oepen",
"suffix": ""
},
{
"first": "Erik",
"middle": [],
"last": "Velldal",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 21st Nordic Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "271--276",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Murhaf Fares, Andrey Kutuzov, Stephan Oepen, and Erik Velldal. 2017. Word vectors, reuse, and replica- bility: Towards a community repository of large-text resources. In Proceedings of the 21st Nordic Con- ference on Computational Linguistics, pages 271- 276, Gothenburg, Sweden. Association for Compu- tational Linguistics.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "NorNE: Annotating Named Entities for Norwegian",
"authors": [
{
"first": "Fredrik",
"middle": [],
"last": "J\u00f8rgensen",
"suffix": ""
},
{
"first": "Tobias",
"middle": [],
"last": "Aasmoe",
"suffix": ""
},
{
"first": "Anne-Stine Ruud",
"middle": [],
"last": "Husev\u00e5g",
"suffix": ""
},
{
"first": "Lilja",
"middle": [],
"last": "\u00d8vrelid",
"suffix": ""
},
{
"first": "Erik",
"middle": [],
"last": "Velldal",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 12th Edition of the Language Resources and Evaluation Conference",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fredrik J\u00f8rgensen, Tobias Aasmoe, Anne-Stine Ruud Husev\u00e5g, Lilja \u00d8vrelid, and Erik Velldal. 2020. NorNE: Annotating Named Entities for Norwegian. In Proceedings of the 12th Edition of the Language Resources and Evaluation Conference, Marseille, France, 2020.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Prodromos Malakasiotis, and Ion Androutsopoulos",
"authors": [
{
"first": "John",
"middle": [],
"last": "Koutsikakis",
"suffix": ""
},
{
"first": "Ilias",
"middle": [],
"last": "Chalkidis",
"suffix": ""
}
],
"year": 2020,
"venue": "11th Hellenic Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "110--117",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John Koutsikakis, Ilias Chalkidis, Prodromos Malaka- siotis, and Ion Androutsopoulos. 2020. Greek-Bert: The Greeks visiting Sesame street. In 11th Hellenic Conference on Artificial Intelligence, pages 110- 117.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Subword regularization: Improving neural network translation models with multiple subword candidates",
"authors": [
{
"first": "Taku",
"middle": [],
"last": "Kudo",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "66--75",
"other_ids": {
"DOI": [
"10.18653/v1/P18-1007"
]
},
"num": null,
"urls": [],
"raw_text": "Taku Kudo. 2018. Subword regularization: Improv- ing neural network translation models with multiple subword candidates. In Proceedings of the 56th An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 66- 75, Melbourne, Australia. Association for Compu- tational Linguistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Operationalizing a National Digital Library: The Case for a Norwegian Transformer Model",
"authors": [
{
"first": "E",
"middle": [],
"last": "Per",
"suffix": ""
},
{
"first": "Javier",
"middle": [],
"last": "Kummervold",
"suffix": ""
},
{
"first": "Freddy",
"middle": [],
"last": "De La Rosa",
"suffix": ""
},
{
"first": "Svein Arne",
"middle": [],
"last": "Wetjen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Brygfjeld",
"suffix": ""
}
],
"year": 2021,
"venue": "Proceedings of the 23rd Nordic Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Per E. Kummervold, Javier De la Rosa, Freddy Wetjen, and Svein Arne Brygfjeld. 2021. Operationalizing a National Digital Library: The Case for a Norwe- gian Transformer Model. In Proceedings of the 23rd Nordic Conference on Computational Linguistics.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "End-to-end negation resolution as graph parsing",
"authors": [
{
"first": "Robin",
"middle": [],
"last": "Kurtz",
"suffix": ""
},
{
"first": "Stephan",
"middle": [],
"last": "Oepen",
"suffix": ""
},
{
"first": "Marco",
"middle": [],
"last": "Kuhlmann",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 16th International Conference on Parsing Technologies and the IWPT 2020",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.18653/v1/2020.iwpt-1.3"
]
},
"num": null,
"urls": [],
"raw_text": "Robin Kurtz, Stephan Oepen, and Marco Kuhlmann. 2020. End-to-end negation resolution as graph pars- ing. In Proceedings of the 16th International Con- ference on Parsing Technologies and the IWPT 2020",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Building web-interfaces for vector semantic models with the WebVectors toolkit",
"authors": [
{
"first": "Andrey",
"middle": [],
"last": "Kutuzov",
"suffix": ""
},
{
"first": "Elizaveta",
"middle": [],
"last": "Kuzmenko",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the Software Demonstrations of the 15th Conference of the European Chapter of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "99--103",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andrey Kutuzov and Elizaveta Kuzmenko. 2017. Building web-interfaces for vector semantic models with the WebVectors toolkit. In Proceedings of the Software Demonstrations of the 15th Conference of the European Chapter of the Association for Compu- tational Linguistics, pages 99-103, Valencia, Spain. Association for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Lilja \u00d8vrelid, and Erik Velldal. 2021. Negation in Norwegian: an annotated dataset",
"authors": [
{
"first": "Petter",
"middle": [],
"last": "Maehlum",
"suffix": ""
},
{
"first": "Jeremy",
"middle": [],
"last": "Barnes",
"suffix": ""
},
{
"first": "Robin",
"middle": [],
"last": "Kurtz",
"suffix": ""
}
],
"year": null,
"venue": "Proceedings of the 23rd Nordic Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Petter Maehlum, Jeremy Barnes, Robin Kurtz, Lilja \u00d8vrelid, and Erik Velldal. 2021. Negation in Nor- wegian: an annotated dataset. In Proceedings of the 23rd Nordic Conference on Computational Linguis- tics.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Love B\u00f6rjeson, and Chris Haffenden. 2020. Playing with words at the National Library of Sweden -making a Swedish BERT",
"authors": [
{
"first": "Martin",
"middle": [],
"last": "Malmsten",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2007.01658"
]
},
"num": null,
"urls": [],
"raw_text": "Martin Malmsten, Love B\u00f6rjeson, and Chris Haf- fenden. 2020. Playing with words at the National Li- brary of Sweden -making a Swedish BERT. arXiv preprint arXiv:2007.01658.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "*SEM 2012 shared task: Resolving the scope and focus of negation",
"authors": [
{
"first": "Roser",
"middle": [],
"last": "Morante",
"suffix": ""
},
{
"first": "Eduardo",
"middle": [],
"last": "Blanco",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the First Joint Conference on Lexical and Computational Semantics (*SEM)",
"volume": "",
"issue": "",
"pages": "265--274",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Roser Morante and Eduardo Blanco. 2012. *SEM 2012 shared task: Resolving the scope and focus of negation. In Proceedings of the First Joint Con- ference on Lexical and Computational Semantics (*SEM), pages 265-274, Montr\u00e9al, Canada.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Phobert: Pre-trained language models for vietnamese",
"authors": [
{
"first": "Anh",
"middle": [
"Tuan"
],
"last": "Dat Quoc Nguyen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Nguyen",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2003.00744"
]
},
"num": null,
"urls": [],
"raw_text": "Dat Quoc Nguyen and Anh Tuan Nguyen. 2020. Phobert: Pre-trained language models for viet- namese. arXiv preprint arXiv:2003.00744.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Universal Dependencies for Norwegian",
"authors": [
{
"first": "Lilja",
"middle": [],
"last": "\u00d8vrelid",
"suffix": ""
},
{
"first": "Petter",
"middle": [],
"last": "Hohle",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 10th International Conference on Language Resources and Evaluation (LREC'16)",
"volume": "",
"issue": "",
"pages": "1579--1585",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lilja \u00d8vrelid and Petter Hohle. 2016. Universal De- pendencies for Norwegian. In Proceedings of the 10th International Conference on Language Re- sources and Evaluation (LREC'16), pages 1579- 1585, Portoro\u017e, Slovenia.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "A fine-grained sentiment dataset for Norwegian",
"authors": [
{
"first": "Lilja",
"middle": [],
"last": "\u00d8vrelid",
"suffix": ""
},
{
"first": "Petter",
"middle": [],
"last": "Maehlum",
"suffix": ""
},
{
"first": "Jeremy",
"middle": [],
"last": "Barnes",
"suffix": ""
},
{
"first": "Erik",
"middle": [],
"last": "Velldal",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 12th Language Resources and Evaluation Conference",
"volume": "",
"issue": "",
"pages": "5025--5033",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lilja \u00d8vrelid, Petter Maehlum, Jeremy Barnes, and Erik Velldal. 2020. A fine-grained sentiment dataset for Norwegian. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 5025- 5033, Marseille, France. European Language Re- sources Association.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Deep contextualized word representations",
"authors": [
{
"first": "Matthew",
"middle": [],
"last": "Peters",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Neumann",
"suffix": ""
},
{
"first": "Mohit",
"middle": [],
"last": "Iyyer",
"suffix": ""
},
{
"first": "Matt",
"middle": [],
"last": "Gardner",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Clark",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "2227--2237",
"other_ids": {
"DOI": [
"10.18653/v1/N18-1202"
]
},
"num": null,
"urls": [],
"raw_text": "Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word rep- resentations. In Proceedings of the 2018 Confer- ence of the North American Chapter of the Associ- ation for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long Papers), pages 2227-2237, New Orleans, Louisiana. Association for Computational Linguistics.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "To tune or not to tune? adapting pretrained representations to diverse tasks",
"authors": [
{
"first": "Matthew",
"middle": [
"E"
],
"last": "Peters",
"suffix": ""
},
{
"first": "Sebastian",
"middle": [],
"last": "Ruder",
"suffix": ""
},
{
"first": "Noah",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019)",
"volume": "",
"issue": "",
"pages": "7--14",
"other_ids": {
"DOI": [
"10.18653/v1/W19-4302"
]
},
"num": null,
"urls": [],
"raw_text": "Matthew E. Peters, Sebastian Ruder, and Noah A. Smith. 2019. To tune or not to tune? adapting pre- trained representations to diverse tasks. In Proceed- ings of the 4th Workshop on Representation Learn- ing for NLP (RepL4NLP-2019), pages 7-14, Flo- rence, Italy. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Stanza: A python natural language processing toolkit for many human languages",
"authors": [
{
"first": "Peng",
"middle": [],
"last": "Qi",
"suffix": ""
},
{
"first": "Yuhao",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Yuhui",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Bolton",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
"volume": "",
"issue": "",
"pages": "101--108",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-demos.14"
]
},
"num": null,
"urls": [],
"raw_text": "Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A python natural language processing toolkit for many human languages. In Proceedings of the 58th An- nual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 101- 108, Online. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Software Framework for Topic Modelling with Large Corpora",
"authors": [
{
"first": "Petr",
"middle": [],
"last": "Radim\u0159eh\u016f\u0159ek",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Sojka",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks",
"volume": "",
"issue": "",
"pages": "45--50",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Radim\u0158eh\u016f\u0159ek and Petr Sojka. 2010. Software Frame- work for Topic Modelling with Large Corpora. In Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks, pages 45-50, Val- letta, Malta.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "The Norwegian Dependency Treebank",
"authors": [
{
"first": "Erik",
"middle": [],
"last": "Per",
"suffix": ""
},
{
"first": "Arne",
"middle": [],
"last": "Solberg",
"suffix": ""
},
{
"first": "Lilja",
"middle": [],
"last": "Skjaerholt",
"suffix": ""
},
{
"first": "Kristin",
"middle": [],
"last": "\u00d8vrelid",
"suffix": ""
},
{
"first": "Janne Bondi",
"middle": [],
"last": "Hagen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Johannessen",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the Ninth International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Per Erik Solberg, Arne Skjaerholt, Lilja \u00d8vrelid, Kristin Hagen, and Janne Bondi Johannessen. 2014. The Norwegian Dependency Treebank. In Proceedings of the Ninth International Conference on Language Resources and Evaluation, Reykjavik, Iceland.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "High quality ELMo embeddings for seven less-resourced languages",
"authors": [
{
"first": "Matej",
"middle": [],
"last": "Ul\u010dar",
"suffix": ""
},
{
"first": "Marko",
"middle": [],
"last": "Robnik-\u0160ikonja",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 12th Language Resources and Evaluation Conference",
"volume": "",
"issue": "",
"pages": "4731--4738",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matej Ul\u010dar and Marko Robnik-\u0160ikonja. 2020. High quality ELMo embeddings for seven less-resourced languages. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 4731- 4738, Marseille, France.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "NoReC: The Norwegian Review Corpus",
"authors": [
{
"first": "Erik",
"middle": [],
"last": "Velldal",
"suffix": ""
},
{
"first": "Lilja",
"middle": [],
"last": "\u00d8vrelid",
"suffix": ""
},
{
"first": "Eivind",
"middle": [
"Alexander"
],
"last": "Bergem",
"suffix": ""
},
{
"first": "Cathrine",
"middle": [],
"last": "Stadsnes",
"suffix": ""
},
{
"first": "Samia",
"middle": [],
"last": "Touileb",
"suffix": ""
},
{
"first": "Fredrik",
"middle": [],
"last": "J\u00f8rgensen",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 11th edition of the Language Resources and Evaluation Conference",
"volume": "",
"issue": "",
"pages": "4186--4191",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Erik Velldal, Lilja \u00d8vrelid, Eivind Alexander Bergem, Cathrine Stadsnes, Samia Touileb, and Fredrik J\u00f8rgensen. 2018. NoReC: The Norwegian Review Corpus. In Proceedings of the 11th edition of the Language Resources and Evaluation Conference, pages 4186-4191, Miyazaki, Japan.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Joint UD parsing of Norwegian Bokm\u00e5l and Nynorsk",
"authors": [
{
"first": "Erik",
"middle": [],
"last": "Velldal",
"suffix": ""
},
{
"first": "Lilja",
"middle": [],
"last": "\u00d8vrelid",
"suffix": ""
},
{
"first": "Petter",
"middle": [],
"last": "Hohle",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 21st Nordic Conference of Computational Linguistics",
"volume": "",
"issue": "",
"pages": "1--10",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Erik Velldal, Lilja \u00d8vrelid, and Petter Hohle. 2017. Joint UD parsing of Norwegian Bokm\u00e5l and Nynorsk. In Proceedings of the 21st Nordic Con- ference of Computational Linguistics, pages 1-10, Gothenburg, Sweden.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Multilingual is not enough: Bert for finnish",
"authors": [
{
"first": "Antti",
"middle": [],
"last": "Virtanen",
"suffix": ""
},
{
"first": "Jenna",
"middle": [],
"last": "Kanerva",
"suffix": ""
},
{
"first": "Rami",
"middle": [],
"last": "Ilo",
"suffix": ""
},
{
"first": "Jouni",
"middle": [],
"last": "Luoma",
"suffix": ""
},
{
"first": "Juhani",
"middle": [],
"last": "Luotolahti",
"suffix": ""
},
{
"first": "Tapio",
"middle": [],
"last": "Salakoski",
"suffix": ""
},
{
"first": "Filip",
"middle": [],
"last": "Ginter",
"suffix": ""
},
{
"first": "Sampo",
"middle": [],
"last": "Pyysalo",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1912.07076"
]
},
"num": null,
"urls": [],
"raw_text": "Antti Virtanen, Jenna Kanerva, Rami Ilo, Jouni Luoma, Juhani Luotolahti, Tapio Salakoski, Filip Ginter, and Sampo Pyysalo. 2019. Multilingual is not enough: Bert for finnish. arXiv preprint arXiv:1912.07076.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Bertje: A Dutch Bert model",
"authors": [
{
"first": "Andreas",
"middle": [],
"last": "Wietse De Vries",
"suffix": ""
},
{
"first": "Arianna",
"middle": [],
"last": "Van Cranenburgh",
"suffix": ""
},
{
"first": "Tommaso",
"middle": [],
"last": "Bisazza",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Caselli",
"suffix": ""
},
{
"first": "Malvina",
"middle": [],
"last": "Gertjan Van Noord",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Nissim",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1912.09582"
]
},
"num": null,
"urls": [],
"raw_text": "Wietse de Vries, Andreas van Cranenburgh, Arianna Bisazza, Tommaso Caselli, Gertjan van Noord, and Malvina Nissim. 2019. Bertje: A Dutch Bert model. arXiv preprint arXiv:1912.09582.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Large batch optimization for deep learning: Training BERT in 76 minutes",
"authors": [
{
"first": "Yang",
"middle": [],
"last": "You",
"suffix": ""
},
{
"first": "Jing",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Sashank",
"middle": [],
"last": "Reddi",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [],
"last": "Hseu",
"suffix": ""
},
{
"first": "Sanjiv",
"middle": [],
"last": "Kumar",
"suffix": ""
},
{
"first": "Srinadh",
"middle": [],
"last": "Bhojanapalli",
"suffix": ""
},
{
"first": "Xiaodan",
"middle": [],
"last": "Song",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Demmel",
"suffix": ""
},
{
"first": "Kurt",
"middle": [],
"last": "Keutzer",
"suffix": ""
},
{
"first": "Cho-Jui",
"middle": [],
"last": "Hsieh",
"suffix": ""
}
],
"year": 2020,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. 2020. Large batch optimization for deep learning: Training BERT in 76 minutes. In International Con- ference on Learning Representations.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "CoNLL 2018 shared task: Multilingual parsing from raw text to Universal Dependencies",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Zeman",
"suffix": ""
},
{
"first": "Jan",
"middle": [],
"last": "Haji\u010d",
"suffix": ""
},
{
"first": "Martin",
"middle": [],
"last": "Popel",
"suffix": ""
},
{
"first": "Martin",
"middle": [],
"last": "Potthast",
"suffix": ""
},
{
"first": "Milan",
"middle": [],
"last": "Straka",
"suffix": ""
},
{
"first": "Filip",
"middle": [],
"last": "Ginter",
"suffix": ""
},
{
"first": "Joakim",
"middle": [],
"last": "Nivre",
"suffix": ""
},
{
"first": "Slav",
"middle": [],
"last": "Petrov",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies",
"volume": "",
"issue": "",
"pages": "1--21",
"other_ids": {
"DOI": [
"10.18653/v1/K18-2001"
]
},
"num": null,
"urls": [],
"raw_text": "Daniel Zeman, Jan Haji\u010d, Martin Popel, Martin Pot- thast, Milan Straka, Filip Ginter, Joakim Nivre, and Slav Petrov. 2018. CoNLL 2018 shared task: Mul- tilingual parsing from raw text to Universal Depen- dencies. In Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Univer- sal Dependencies, pages 1-21, Brussels, Belgium. Association for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "NorBERT loss plots at the Phase 1 (left) and Phase 2 (right).Similar to Virtanen et al."
},
"TABREF1": {
"num": null,
"content": "<table/>",
"type_str": "table",
"html": null,
"text": "Number of sentences in the training, development, and test splits in the datasets used for the evaluation tasks."
},
"TABREF3": {
"num": null,
"content": "<table><tr><td>: Evaluation scores of the NorLM models</td></tr><tr><td>on the POS tagging of Bokm\u00e5l (BM) and Nynorsk</td></tr><tr><td>(NN) test sets in comparison with other large pre-</td></tr><tr><td>trained models for Norwegian. Running times in</td></tr><tr><td>minutes are given for Bokm\u00e5l.</td></tr></table>",
"type_str": "table",
"html": null,
"text": ""
},
"TABREF6": {
"num": null,
"content": "<table><tr><td>Model</td><td>CUE</td><td>ST</td><td colspan=\"2\">FN Time</td></tr><tr><td>NorELMo 30</td><td colspan=\"3\">91.7 80.6 63.8</td><td>428</td></tr><tr><td colspan=\"4\">NorELMo 100 92.2 81.3 65.5</td><td>407</td></tr><tr><td>mBERT</td><td colspan=\"3\">92.8 84.0 65.9</td><td>353</td></tr><tr><td>NB-BERT</td><td colspan=\"3\">92.4 83.1 63.5</td><td>342</td></tr><tr><td>NorBERT</td><td colspan=\"3\">92.1 83.6 65.5</td><td>426</td></tr></table>",
"type_str": "table",
"html": null,
"text": "F 1 scores for the different LMs models on the binary sentiment classification test set."
},
"TABREF7": {
"num": null,
"content": "<table><tr><td>: Results of our negation parser, augment-</td></tr><tr><td>ing the features with token representations from</td></tr><tr><td>each language model. The results are averaged</td></tr><tr><td>over 5 runs.</td></tr></table>",
"type_str": "table",
"html": null,
"text": ""
}
}
}
} |