File size: 129,859 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 | {
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T13:05:04.239402Z"
},
"title": "Which Dependency Parser to Use for Distributional Semantics in a Specialized Domain?",
"authors": [
{
"first": "Pauline",
"middle": [],
"last": "Brunet",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LIST",
"location": {
"postCode": "F-91191",
"settlement": "Gif-sur-Yvette",
"country": "France"
}
},
"email": "pauline.brunet@cea.fr"
},
{
"first": "Olivier",
"middle": [],
"last": "Ferret",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LIST",
"location": {
"postCode": "F-91191",
"settlement": "Gif-sur-Yvette",
"country": "France"
}
},
"email": "olivier.ferret@cea.fr"
},
{
"first": "Ludovic",
"middle": [],
"last": "Tanguy",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Toulouse",
"location": {
"country": "France"
}
},
"email": "ludovic.tanguy@univ-tlse2.fr"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We present a study whose objective is to compare several dependency parsers for English applied to a specialized corpus for building distributional count-based models from syntactic dependencies. One of the particularities of this study is to focus on the concepts of the target domain, which mainly occur in documents as multi-terms and must be aligned with the outputs of the parsers. We compare a set of ten parsers in terms of syntactic triplets but also in terms of distributional neighbors extracted from the models built from these triplets, both with and without an external reference concerning the semantic relations between concepts. We show more particularly that some patterns of proximity between these parsers can be observed across our different evaluations, which could give insights for anticipating the performance of a parser for building distributional models from a given corpus.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "We present a study whose objective is to compare several dependency parsers for English applied to a specialized corpus for building distributional count-based models from syntactic dependencies. One of the particularities of this study is to focus on the concepts of the target domain, which mainly occur in documents as multi-terms and must be aligned with the outputs of the parsers. We compare a set of ten parsers in terms of syntactic triplets but also in terms of distributional neighbors extracted from the models built from these triplets, both with and without an external reference concerning the semantic relations between concepts. We show more particularly that some patterns of proximity between these parsers can be observed across our different evaluations, which could give insights for anticipating the performance of a parser for building distributional models from a given corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "This work takes place in the broader context of studying distributional semantic analysis methods for specialized corpora. This type of corpora are usually small-sized (a few million words or less), which poses a challenge for distributional methods, and contain specific, highly technical vocabulary, meaning that adapting methods based on large generic corpora might be difficult. We make the hypothesis, supported by the work of (Tanguy et al., 2015) , that the small amount of data may be circumvented by a method based on syntactic contexts. Such methods have already been investigated by a large body of work. The largest part of it is dedicated to count-based approaches (Grefenstette, 1994; Habert et al., 1996; Lin, 1998; Curran and Moens, 2002; Pad\u00f3 and Lapata, 2007 ; Baroni and Lenci, 2010) but it also includes work adding dimensionality reduction methods (Lapesa and Evert, 2017) or more recently, work about word embeddings (Levy and Goldberg, 2014) . One of our focuses is to select the best-suited tools for semantic analysis of specialized corpora. In particular, given that syntactic contexts will be a building block for the task, which syntactic parser should be used to extract these contexts? The goal of this article is, thus, to study the impact of the choice of parser on the construction of a distributional model with a frequency-based method. Our work is not the first work on comparing different parsers. Several evaluation campaigns were previously organized for various languages: the Easy (Paroubek et al., 2008) , Passage (De La Clergerie et al., 2008) , SPMRL (Seddah et al., 2013) and CoNLL (Zeman et al., 2018) campaigns as well as more focused studies like (Candito et al., 2010) or (De La Clergerie, 2014) . However, the benchmarks used in these studies, adopting the kind of diverse, generic corpora on which the tools have been trained, might not be the most relevant option for specialized corpus parsing. Moreover, even though some of these campaigns are recent, the main tools available have not been compared on the same evaluation sets. We previously performed a first study (Tanguy et al., 2020) , comparing 11 different versions of parsers on a small specialized corpus made up of Natural Language Processing papers for French. However, we lacked a reliable external reference to measure the results of the parsers against. So our evaluation was only a qualitative comparison.",
"cite_spans": [
{
"start": 432,
"end": 453,
"text": "(Tanguy et al., 2015)",
"ref_id": "BIBREF36"
},
{
"start": 678,
"end": 698,
"text": "(Grefenstette, 1994;",
"ref_id": "BIBREF12"
},
{
"start": 699,
"end": 719,
"text": "Habert et al., 1996;",
"ref_id": "BIBREF13"
},
{
"start": 720,
"end": 730,
"text": "Lin, 1998;",
"ref_id": "BIBREF21"
},
{
"start": 731,
"end": 754,
"text": "Curran and Moens, 2002;",
"ref_id": "BIBREF8"
},
{
"start": 755,
"end": 776,
"text": "Pad\u00f3 and Lapata, 2007",
"ref_id": "BIBREF25"
},
{
"start": 779,
"end": 802,
"text": "Baroni and Lenci, 2010)",
"ref_id": "BIBREF3"
},
{
"start": 869,
"end": 893,
"text": "(Lapesa and Evert, 2017)",
"ref_id": "BIBREF18"
},
{
"start": 939,
"end": 964,
"text": "(Levy and Goldberg, 2014)",
"ref_id": "BIBREF19"
},
{
"start": 1522,
"end": 1545,
"text": "(Paroubek et al., 2008)",
"ref_id": null
},
{
"start": 1548,
"end": 1586,
"text": "Passage (De La Clergerie et al., 2008)",
"ref_id": null
},
{
"start": 1595,
"end": 1616,
"text": "(Seddah et al., 2013)",
"ref_id": "BIBREF32"
},
{
"start": 1627,
"end": 1647,
"text": "(Zeman et al., 2018)",
"ref_id": "BIBREF41"
},
{
"start": 1695,
"end": 1717,
"text": "(Candito et al., 2010)",
"ref_id": "BIBREF7"
},
{
"start": 1721,
"end": 1744,
"text": "(De La Clergerie, 2014)",
"ref_id": "BIBREF10"
},
{
"start": 2121,
"end": 2142,
"text": "(Tanguy et al., 2020)",
"ref_id": "BIBREF37"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "To go beyond the limitation in (Tanguy et al., 2020) , we have chosen, in the work we present in this article, to run a new evaluation on a small, specialized biomedical corpus, whose building is described in Section 3.1 and for which we may compare the relations implied by the extracted syntactic contexts against an external resource, the Unified Medical Language System (UMLS) (Bodenreider, 2004) , which contains relations between medical and biomedical concepts (see Section 3.3). More precisely, we defined the following process: we applied each of the 10 studied parsers we present in Section 3.2 to the corpus, outputting morphological, syntactic and grammatical information. In parallel, we ran MetaMap (Aronson and Lang, 2010) , a biomedical entity linker, to identify biomedical concepts as defined and recorded in the UMLS. Then, we aligned these concepts with the tokens outputted by the parsers (see Section 4.1). From this alignment, we were able to extract grammatical relations between concept-mapped tokens and other tokens, which gave us syntactic contexts for concept-mapped tokens, and, therefore, for biomedical concepts themselves (see Section 4.2). We then built distributional thesauri for each of the parsers (see Section 5.1), leading to a large set of distributional similarity relations between biomedical concepts. Finally, we compared these similarity relations to the relations between biomedical concepts given by the UMLS (see Section 5.3) and used this comparison for characterizing our studied parsers.",
"cite_spans": [
{
"start": 31,
"end": 52,
"text": "(Tanguy et al., 2020)",
"ref_id": "BIBREF37"
},
{
"start": 381,
"end": 400,
"text": "(Bodenreider, 2004)",
"ref_id": "BIBREF5"
},
{
"start": 713,
"end": 737,
"text": "(Aronson and Lang, 2010)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Overview",
"sec_num": "2."
},
{
"text": "For this experiment, we used a small part of the Open Access subset of the PubMed Central corpus (PMC) 1 , a collection of more than 5 million full-text articles from thousands of biomedical and life science journals. This corpus, originally in a very rich XML format, was cleaned up by removing a lot of non-parsable content like tables, formulas, links, then converted to raw text for parsing. We chose a subset based on a specialty domain centered on stem cells. Articles in PMC OA are indexed by the MeSH index, which tags each article with their themes (or subject headings), with an indication of whether the theme is a main theme of the article or not. To obtain a corpus that was the right size for our purposes, we chose to include any article that was tagged with a heading containing the words \"Stem Cells\", which includes headings such as \"Stem Cells\", \"Adult Stem Cells\", \"Totipotent Stem Cells\", \"Mouse Embryonic Stem Cells\", and others. This was done regardless of whether the heading was indicated as a main theme of the article or not. The resulting corpus is comprised of 23,094 articles, and 104 million words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Corpus",
"sec_num": "3.1."
},
{
"text": "We selected 5 tools able to perform dependency parsing in English, focusing on easily available and ready-to-use parsers, i.e. those that take in charge the whole processing chain, from raw text to dependencies. These tools were applied with their default options. All these tools use statistical models trained on annotated corpora. Their differences concern implementation choices like parsing techniques (graph-or transition-based, for instance), learning models (SVM, maximal entropy or more recently, recurrent neural networks), and upstream or side processing (segmentation, lemmatization). There is much less choice among the training corpora, given the high cost of the annotation and validation processes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Parsers",
"sec_num": "3.2."
},
{
"text": "CoreNLP (Manning et al., 2014) , the main tool of the Stanford team, implements a maximum entropy tagger, which uses the Penn Treebank tagset (Marcus et al., 1993) , and a transition-based parser.",
"cite_spans": [
{
"start": 8,
"end": 30,
"text": "(Manning et al., 2014)",
"ref_id": "BIBREF22"
},
{
"start": 142,
"end": 163,
"text": "(Marcus et al., 1993)",
"ref_id": "BIBREF23"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Parsers",
"sec_num": "3.2."
},
{
"text": "StanfordNLP (Qi et al., 2018) is a tool that, on top of giving access to the CoreNLP chain in Python, implements an entirely different parsing chain. Its graphbased parser relies on a LSTM neural network. Stan-fordNLP offers 3 English models, trained on the UD EWT (Silveira et al., 2014) , LinES (Ahrenberg, 2015) and ParTUT (Bosco et al., 2012) corpora. We used all three of these models.",
"cite_spans": [
{
"start": 12,
"end": 29,
"text": "(Qi et al., 2018)",
"ref_id": "BIBREF29"
},
{
"start": 265,
"end": 288,
"text": "(Silveira et al., 2014)",
"ref_id": "BIBREF33"
},
{
"start": 297,
"end": 314,
"text": "(Ahrenberg, 2015)",
"ref_id": "BIBREF0"
},
{
"start": 326,
"end": 346,
"text": "(Bosco et al., 2012)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Parsers",
"sec_num": "3.2."
},
{
"text": "Spacy is an industry-targeting tool whose main characteristic is its speed compared to most other parsers. The tagger is based on a perceptron, with attributes based on Brown clusters, following (Koo et al., 2008) . It implements a non-monotonous transition-based parser which can revise previous decisions (Honnibal and Johnson, 2015) . The default model we used was 1 http://www.ncbi.nlm.nih.gov/pmc trained on OntoNotes (Hovy et al., 2006) and uses the ClearNLP dependency labels 2 .",
"cite_spans": [
{
"start": 195,
"end": 213,
"text": "(Koo et al., 2008)",
"ref_id": "BIBREF17"
},
{
"start": 307,
"end": 335,
"text": "(Honnibal and Johnson, 2015)",
"ref_id": "BIBREF14"
},
{
"start": 423,
"end": 442,
"text": "(Hovy et al., 2006)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Parsers",
"sec_num": "3.2."
},
{
"text": "UDPipe (Straka and Strakov\u00e1, 2017) uses a neural network with a Gated Recurrent Unit mechanism to do both tokenization and segmentation at once. For PoS tagging, it generates possible tags for words from their suffix and performs disambiguation with a perceptron. The transition-based parsing relies on a simple one-layer neural network. UDPipe includes four English models. We used all of them, trained on the UD GUM (Zeldes, 2017) , EWT, LinES and ParTUT corpora.",
"cite_spans": [
{
"start": 7,
"end": 34,
"text": "(Straka and Strakov\u00e1, 2017)",
"ref_id": "BIBREF35"
},
{
"start": 418,
"end": 432,
"text": "(Zeldes, 2017)",
"ref_id": "BIBREF40"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Parsers",
"sec_num": "3.2."
},
{
"text": "Talismane (Urieli and Tanguy, 2013) uses a mix of statistic models and language-specific features and rules incorporating linguistic knowledge. It was trained on the Penn Treebank.",
"cite_spans": [
{
"start": 10,
"end": 35,
"text": "(Urieli and Tanguy, 2013)",
"ref_id": "BIBREF38"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Parsers",
"sec_num": "3.2."
},
{
"text": "We are fully aware that these parsers can only be compared on a practical level since the technologies used, their goals, their training data, and even the times at which they were created can scarcely be compared.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Parsers",
"sec_num": "3.2."
},
{
"text": "The UMLS is a set of knowledge sources related to biomedical sciences. The main part of the system is the UMLS Metathesaurus, which aggregates nearly 200 biomedical controlled vocabularies in an attempt to provide a reference frame for medical and biomedical concepts and links the different names under which they are known in different vocabularies as synonyms. The Metathesaurus is organized around these concepts, which, in theory, have only one meaning, and are unique in the Metathesaurus. Each concept has a unique identifier called CUI and is linked to one or more names, in specific vocabularies, for this concept, which have identifiers called AUI. For example, the concept \"Headache\" (CUI: C0018681) can be found as the following variations (among others): in vocabulary SNOMED, \"Headache\" (AUI: A2882187), in vocabulary MeSH, \"Headache\" (AUI: A0066000) and \"Cranial Pain\" (AUI: A1641293), and in vocabulary DxP, \"HEAD PAIN CEPHALGIA\" (AUI: A0418053). On top of these concepts, the Metathesaurus provides some relations between concepts 3 . Most of these relations come from individual source vocabularies; some of these are added by the Metathesaurus maintainers and the others by the users. All relations have general REL labels, which specify the type of relations: synonym, parent, child, sibling, broader, narrower, qualifier, qualified by, or unspecified (several degrees). There are 14 possible REL labels. Around one-fourth of relations also have a RELA label, which further specifies the relation, like is_a, has_ingredient, property_of. . . These labels come from the source vocabularies. As such, they are more diversified than the REL labels, with nearly 900 different labels in total, and we cannot assume that they are coherently used throughout the Metathesaurus. We observed that the distribution of relations among concepts (CUI) is not very well balanced: The first step in our study is to match tokens, as segmented by various parsers, to biomedical concepts, as recognized by a biomedical entity linking tool. This task was greatly impeded by various alignment issues. There are some available tools for biomedical UMLS concepts extraction from text. After testing several of them, among which cTakes (Savova et al., 2010) and Quick-UMLS (Soldaini and Goharian, 2016) , we decided to use MetaMap (Aronson, 2001) , the reference tool for this task, because it had the clear advantage of providing disambiguation between possible candidate concepts for a phrase. MetaMap splits the input documents into sentences, which are further split into phrases. It analyzes these phrases individually and outputs candidate mappings of UMLS concepts to the phrase. These mappings are given an evaluation score based on 4 metrics: centrality, variation, coverage, and cohesiveness. The mapping with the highest score may be selected as the most likely to be correct but MetaMap also provides a disambiguation module based on context. We exploited the possibility of MetaMap to output only the most likely mapping based on score and context disambiguation. For instance, the phrase \"Generation of single-copy transgenic mouse embryos\" is linked to UMLS concepts \"Generations\" (C0079411), \"Singular\" (C0205171), \"Copy\" (C1948062), \"Mice, Transgenic\" (C0025936) and \"Embryo\" (C0013935).",
"cite_spans": [
{
"start": 2232,
"end": 2253,
"text": "(Savova et al., 2010)",
"ref_id": null
},
{
"start": 2269,
"end": 2298,
"text": "(Soldaini and Goharian, 2016)",
"ref_id": "BIBREF34"
},
{
"start": 2327,
"end": 2342,
"text": "(Aronson, 2001)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Terminological Reference Resource",
"sec_num": "3.3."
},
{
"text": "number",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Terminological Reference Resource",
"sec_num": "3.3."
},
{
"text": "The linguistic analysis performed by MetaMap for identifying concepts in documents is, of course, different from the analysis performed by our target parsers. More precisely, the tokenization step is particularly important for aligning the concepts it identifies with the tokens issued from the various tokenizations of our parsers. MetaMap gives two position data for each match. The first one gathers the character offset in the phrase and the length of the matched words. This information is highly difficult to match with parsers' offsets because of imprecisions and different counting conventions from both MetaMap and the various parsers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Terminological Reference Resource",
"sec_num": "3.3."
},
{
"text": "The second position information is the rank of the matched words in the phrase. For instance, in the above example, the concept \"Generations\" has both a TextMatch-Start and TextMatchEnd attributes equal to 1 while \"Mice, Transgenic\" has a TextMatchStart attribute of 5 and a TextMatchEnd attribute of 6. However, this information cannot be directly matched with the tokenization of a parser since it depends on both the tokenization and MetaMap's phrase splitting. The first step is then to associate each concept identified by MetaMap with its own tokenization, which we later align with each parser's tokenization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Terminological Reference Resource",
"sec_num": "3.3."
},
{
"text": "4.1.1. Matching MetaMap's tokens with MetaMap's concepts This first step is not trivial as the tokenization performed by MetaMap is not directly accessible in its output. However, each phrase in this output is segmented into syntactic units and each of these units is associated with a list of tokens. For example, \"single-copy\" is a syntactic unit associated with the token list [\"single\", \"copy\"]. From these syntactic units, we can collect the associated tokens and number them according to their order, which gives us something close to the behind-the-scene MetaMap tokenization. This numbering can be used to match the tokens with the biomedical concepts, but with the necessity to take two additional problems into account. First, the punctuation is not considered in MetaMap's numbering, but must obviously be recorded for the later alignment with the parsers' tokenization. Second, MetaMap's numbering sometimes skips the first or the first few tokens in a phrase if they are not associated with a concept, which is more troublesome. For example, in the phrase \"from cells\", MetaMap may declare that the \"Cells\" concept starts from 1 or 2. We were not able to determine the cause of this behavior but we found a workaround for the problem by comparing the offset of the matched start-ing position to the offset of the phrase starting position. If a discrepancy was found, the character count of each word was added until the discrepancy was filled to compute the number of skipped words. With this process, we were able to match MetaMap's concepts to its tokenization with very good accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Terminological Reference Resource",
"sec_num": "3.3."
},
{
"text": "tokens The next step matches MetaMap's tokenization with each parser's tokenization. Another solution could have consisted in feeding MetaMap's tokenization to the parsers, as most of them are modular enough to allow it. We rejected the idea for two reasons. First, the tokens we retrieved from MetaMap could be different from the initial text: for example, by modifying some punctuations, destroying case information and even expanding acronyms. Second, we wanted to use the parsers out of the box, with their own tokenization suited to their own tagging and parsing processes. The algorithm for matching these different tokenizations is based on the fact that the tokenizations are essentially similar. The majority of words are tokenized similarly by MetaMap and the parsers. Thus, for each document, we can align the outputs of MetaMap and the considered parser by relying on their common tokens and use a small set of heuristics to deal with discrepancies. The discrepancies we handle are of several types. They may come from parserspecific issues, such as Spacy inserting \"SPACE\" tokens when confronted with large breaks in the text. One of the two tokenizations may have inserted a sentence break while the other may not, in which case the sentence break is skipped. One of the tokenizations may have split a token while the other may not, such as \"single-copy\" on one side and \"single\" followed by \"copy\" on the other side. In such cases, we add the next tokens on the shorter side, separating them with both spaces and \"-\" until they stop matching the longer side or the whole split token is covered. If the process has been successful, the longer token is matched with all the smaller ones. If it has failed, we skip tokens on both sides and see if the next ones match. This is especially useful for cases where MetaMap or the parser modifies the tokens in some way, like \"99%\" becoming \"99\". Failing that, we are only concerned with finding some part of the text where the tokens match again, ideally as close as possible to the failure point. We implement this strategy by recording, from the failure point, the list of tokens from both MetaMap and the target parser and checking at each step if the last two tokens seen on one side can be found in the list of the other side 4 . If so, we skip up to this part and start matching from there. This algorithm works fairly well and a very large percentage of tokens are matched.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Matching MetaMap's tokens with parsers'",
"sec_num": "4.1.2."
},
{
"text": "The next step is to extract dependency relations between words to build the contexts that will be used for distributional analysis. This follows a similar process to the work of (Lin, 1998) and produces, from the dependency relations outputted by syntactic parsers, typically illustrated by Figure 1, Not all relations provide useful context information. Generally, relations including closed-class words (determiners, conjunctions, pronouns, etc.) are not considered for building distributional contexts. For this study, we performed our selection not on the PoS tag of the governor and dependent, but on the dependency relation itself, choosing to exclude some of them. For parsers following the Universal Dependency scheme, the excluded relations were root, cc, cc:preconj, punct, case, mark, det, det:predet, cop, neg, aux, and nmod:poss. Typically, relations such as neg, aux or det include negation markers (not. . . ), auxiliary verbs (have, be, can..) or determiners (the, a. . . ) that we don't want to see in distributional contexts. For Spacy, it was root, ccc, case, prep, det, neg, expl, predet, aux, auxpass, and mark. For Talismane, given its specific dependency scheme, we had to rely on PoS tags for achieving the same kind of filtering, excluding IN, DT, MD, CC, EX, PDT, PRP, PRP$, TO and, RP. Relations with prepositions had to be modified to link the actual related words and include the preposition in the relation. We illustrate the different ways this kind of grammatical constructions are parsed with the phrase \"region within the cluster\". Figure 2 shows the output produced by UD dependency scheme-following parsers.",
"cite_spans": [
{
"start": 178,
"end": 189,
"text": "(Lin, 1998)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [
{
"start": 291,
"end": 300,
"text": "Figure 1,",
"ref_id": "FIGREF0"
},
{
"start": 1566,
"end": 1574,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "region within the cluster case nmod Figure 2 : Dependencies identified by Universal Dependencies scheme-type parsers for the phrase \"region within the cluster\". variants with a prep/within dependency relation, as illustrated in Figure 5 .",
"cite_spans": [],
"ref_spans": [
{
"start": 36,
"end": 44,
"text": "Figure 2",
"ref_id": null
},
{
"start": 228,
"end": 236,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "region within the cluster prep/within Figure 5 : Dependency constructed from the existing dependencies for the phrase \"region within the cluster\".",
"cite_spans": [],
"ref_spans": [
{
"start": 38,
"end": 46,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "For the example of Figure 1 , the resulting list of triplets is:",
"cite_spans": [],
"ref_spans": [
{
"start": 19,
"end": 27,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "(embryos, prep/of, generation) (generation, prep/of-1, embryos) (single-copy, amod, transgenic) (transgenic, amod-1, single-copy) (transgenic, compound, mouse) (mouse, compound-1, transgenic) (mouse, compound, embryo) (embryos, compound-1, mouse)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "However, we need to adapt this representation of context to our task, which is specifically to extract the contexts of biomedical concepts. Thus, we only extract the relations where at least one side, the dependent or the governor, is a token that is part of a concept and the other side is not part of the same concept. Moreover, we only consider nominal concepts, which we define here as concepts where at least one word was tagged as a noun by the MetaMap tagger. Furthermore, for each side of a triplet, we include the following data:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "\u2022 CUI: the unique UMLS id if it is a concept, _ otherwise;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "\u2022 PREF: the preferred form of the concept in the UMLS if it is a concept, _ otherwise;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "\u2022 NORM: the normalized form of the concept as it occurs in the text if it is a concept, _ otherwise. Concretely, it corresponds to the concatenation of all the lemmas of the concept;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "\u2022 LEMMA of the token actually part of the relation;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "\u2022 PoS of the token actually part of the relation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "In the above phrase, five concepts were recognized by MetaMap: \"Generation\" (noun), \"Singular\" (adj), \"Copy (object)\" (adj), \"Mice, Transgenic\" (noun), and \"Embryo\" (noun). The corresponding triplets are given in Table 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 213,
"end": 220,
"text": "Table 1",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Dependency Triple Extraction",
"sec_num": "4.2."
},
{
"text": "Several previous studies (starting with (Grefenstette, 1994) and (Lin, 1998) ) have considered subsets of syntactic relations for distributional models. More recent works (Pad\u00f3 and Lapata, 2007; Baroni and Lenci, 2010) have selected a short list of core relations, and we decided to limit our experiment to these, which we regrouped in the categories described below. They follow the main syntactic relations identified by dependency parsers and correspond to the minimal configuration of (Pad\u00f3 and Lapata, 2007) , to which we added the last one from (Baroni and Lenci, 2010) , that consider the prepositions themselves as relations between a head and a dependent word, as described in section 4.2. ADV mod ADJ/V : adverb modifying an adjective or a verb;",
"cite_spans": [
{
"start": 40,
"end": 60,
"text": "(Grefenstette, 1994)",
"ref_id": "BIBREF12"
},
{
"start": 65,
"end": 76,
"text": "(Lin, 1998)",
"ref_id": "BIBREF21"
},
{
"start": 171,
"end": 194,
"text": "(Pad\u00f3 and Lapata, 2007;",
"ref_id": "BIBREF25"
},
{
"start": 195,
"end": 218,
"text": "Baroni and Lenci, 2010)",
"ref_id": "BIBREF3"
},
{
"start": 489,
"end": 512,
"text": "(Pad\u00f3 and Lapata, 2007)",
"ref_id": "BIBREF25"
},
{
"start": 551,
"end": 575,
"text": "(Baroni and Lenci, 2010)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison of Parsers in Terms of Dependency Triples",
"sec_num": "4.3."
},
{
"text": "X coord X : coordination between two nouns, adjectives, adverbs, or verbs (note: the conjunction itself is not considered); X prep_P X: prepositional binding between nouns, adjectives, or verbs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison of Parsers in Terms of Dependency Triples",
"sec_num": "4.3."
},
{
"text": "This brings down the number of triplets (occurrences) extracted from each parser from around 60M to around 40M, with SpaCy having the least (38.3M) and the version of UDPipe trained on ParTUT having the most (52.4M, far ahead of the others). We compare the triplets' coverage between parsers but first, we reduce the triplets to some core elements: the CUI of the left-hand side, the CUI or lemma (depending on whether it is a concept or not) and the PoS tag of the right-hand side, and the relation between the two. Our triplets now look like the following: C0040648 N:C0237753_prep/of C1883221 ADJ:distinct_mod-1 We also limit ourselves to triplets in which the left-hand side (CUI) appears at least 10 times in each parser's output, and in which the right-hand side (CUI/lemma and PoS) appears at least twice in each parser's output. This results in 39M unique triplets, of which 21M appear for at least two parsers and 3.5M appear for all parsers. Among these 39M triplets, each parser has found around 13M of them, with the least being CoreNLP with 12.2M and SpaCy with 12.5M. The parser with the highest number of common triplets is the ParTUT version of UDPipe, with 16.8M triplets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison of Parsers in Terms of Dependency Triples",
"sec_num": "4.3."
},
{
"text": "Relation Governor We computed the agreement of the parsers about the triplets they produced by computing Spearman's correlation coefficient (\u03c1) both for the 21M triplets shared by at least two parsers and the 3.5M triplets common to all parsers. The first measure focuses on the differences between parsers in terms of diversity of triplets while the second measure looks more precisely at their differences in terms of frequency for the common triplets. We did not include Talismane as it was difficult to adapt its PoS tags and dependency labels to our normalization, as was done for CoreNLP and SpaCy. Figure 6 shows the hierarchical clustering of the parsers according to these correlations (more precisely, 1 -\u03c1 for having a distance). Globally, we can observe that the type of parser has a significant impact on the triplets, which is not a surprise: the UDPipe parsers are particularly close to each other but most of the StanfordNLP models are also grouped. However, the training corpus can also have an impact when we consider the triplets shared by all parsers, with StanfordNLP-Lines much closer to UDpipe-Lines than to the two other StanfordNLP models. This is why the clusterings built for the two sets of triplets are a little bit different, even if they also share some patterns: for instance, SpaCy is close to CoreNLP, which is close to StanfordNLP-Ewt while UDpipe-Gum and UDpipe-Ewt form a group for the two sets.",
"cite_spans": [],
"ref_spans": [
{
"start": 605,
"end": 613,
"text": "Figure 6",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Dependent",
"sec_num": null
},
{
"text": "CUI 1 PREF 1 NORM 1 LEMMA 1 PoS 1 CUI 2 PREF 2 NORM 2 LEMMA 2 PoS 2 C0013935",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dependent",
"sec_num": null
},
{
"text": "Following the distinction made in Baroni et al. (2014) , we built our distributional models according to a countbased approach, such as in (Lin, 1998) , rather than according to a predictive approach such as in (Mikolov et al., 2013) . The first justification of this choice is that, except for (Levy and Goldberg, 2014) , the number of studies relying on dependency relations is very limited among predictive approaches. More importantly, some recent studies (Pierrejean and Tanguy, 2018) have shown that predictive approaches are unstable to some extent concerning the search of the nearest distributional neighbors of a word. Since we want specifically to concentrate on the effects resulting from the use of different syntactic parsers, we adopted a count-based approach. We implemented this approach by building on the findings of recent studies in the field (Kiela and Clark, 2014; Baroni et al., 2014; Levy et al., 2015) and more particularly took up two main options from (Ferret, 2010) : the use of Positive Pointwise Mutual Information (PPMI) for weighting the context elements and the application of very loose filtering that removes the elements of these contexts with only one occurrence. The second choice is justified by both the fairly small size of our target corpus and the experiments of (Ferret, 2010) with linear co-occurrents. The main particularity of our work is the fact that the entries of our distri- butional models are not words but UMLS concepts. More precisely, each entry is made of the triple (CUI, PREF, PoS) under the form PREF_CUI#PoS. The elements of contexts can be either words or concepts since dependency triples can include concepts or words. However, this particularity did not influence on the way we built our distributional models and we classically computed the similarity of two concepts by measuring the Cosine similarity score between their contexts vectors. For a given model, this computation was done for each pair of concepts with contexts sharing at least one element. The results of this process can also be viewed as a distributional thesaurus in which each entry corresponds to a concept of the considered vocabulary and is associated with the list of all other concepts of this vocabulary, sorted in descending order of their similarity value with the entry. In practice, only the nearest 100 distributional neighbors are kept, which is a fairly large number compared to the average number of relations by concept -24.6 -but is justified by the fact that some concepts may have a much higher number of relations.",
"cite_spans": [
{
"start": 34,
"end": 54,
"text": "Baroni et al. (2014)",
"ref_id": "BIBREF4"
},
{
"start": 139,
"end": 150,
"text": "(Lin, 1998)",
"ref_id": "BIBREF21"
},
{
"start": 211,
"end": 233,
"text": "(Mikolov et al., 2013)",
"ref_id": "BIBREF24"
},
{
"start": 295,
"end": 320,
"text": "(Levy and Goldberg, 2014)",
"ref_id": "BIBREF19"
},
{
"start": 864,
"end": 887,
"text": "(Kiela and Clark, 2014;",
"ref_id": "BIBREF16"
},
{
"start": 888,
"end": 908,
"text": "Baroni et al., 2014;",
"ref_id": "BIBREF4"
},
{
"start": 909,
"end": 927,
"text": "Levy et al., 2015)",
"ref_id": "BIBREF20"
},
{
"start": 980,
"end": 994,
"text": "(Ferret, 2010)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Building of Distributional Models",
"sec_num": "5.1."
},
{
"text": "The first step for comparing our distributional models, and more indirectly the parsers used for extracting the distributional data they rely on, was to compute their agreement of our models on the nearest neighbors retrieved for each word. Among the concepts shared by all models, 47,647 concepts had at least one distributional neighbor. For each pair of models, the agreement on the nearest neighbor retrieved for each concept was computed 7 and used for building a similarity matrix. Hierarchical clustering was performed from this matrix, which leads to the left side of Figure 7 . First, we can observe that the model built from Tal-ismane is clearly aside from the others. The second main trend is that the training corpus of the parsers can be more important than the type of parser. For instance, the Stan-fordNLP and UDPipe parsers trained on the LinES corpus are grouped together and fairly distant from the same parsers trained on the GUM and EWT corpora. However, among the parsers trained on these two corpora, which are fairly heterogeneous compared to the LinES corpus, the proximity between the models they contributed to build is guided by the type of parser. Even if the overall aspect of the dendrogram is a little bit different due to the position of the model built from Spacy, these trends are globally confirmed by comparing the neighbors of concepts by the means of the Rank-Biased Overlap measure (Webber et al., 2010) , as illustrated by the right side of Figure 7 . This measure is applied to all neighbors of our thesaurus' entries (100 neighbors in practice) and extends the notion of average overlap -the average of the overlap between two lists at different ranksby decreasing the importance of overlap as the rank of the considered neighbors increases. As a consequence, nearest neighbors are given greater importance. This importance is defined by the p parameter, which can be interpreted as the probability, starting from the beginning of the list of neighbors, to continue to consider the following neighbors in the list. The value p = 0.98 used in our case means that the first 50 nearest neighbors of an entry account for around 85% of the evaluation. Figure 7 is based on the distance 1 \u2212 RBO, which can be considered as a metric. The clusterings of Figure 7 can also be compared to those of Figure 6 : Talismane is absent from Figure 6 but has a very limited impact in Figure 7 since it is clearly distant from the other parsers. This comparison shows that the clustering based on the distributional neighbors is much closer to the clustering based on the triplets shared by all parsers than to the clustering based on the triplets shared by only two parsers. This suggests that the triplets of the first set are Model #concepts #eval. concepts #rel./ concept Recall R prec MAP P@1 P@5 P@10 P@100 globally more frequent than the triplets of the second set and can be used for having a first indication of the proximity of the distributional models built from them.",
"cite_spans": [
{
"start": 1423,
"end": 1444,
"text": "(Webber et al., 2010)",
"ref_id": "BIBREF39"
}
],
"ref_spans": [
{
"start": 576,
"end": 584,
"text": "Figure 7",
"ref_id": "FIGREF4"
},
{
"start": 1483,
"end": 1491,
"text": "Figure 7",
"ref_id": "FIGREF4"
},
{
"start": 2191,
"end": 2199,
"text": "Figure 7",
"ref_id": "FIGREF4"
},
{
"start": 2290,
"end": 2298,
"text": "Figure 7",
"ref_id": "FIGREF4"
},
{
"start": 2332,
"end": 2340,
"text": "Figure 6",
"ref_id": "FIGREF3"
},
{
"start": 2368,
"end": 2376,
"text": "Figure 6",
"ref_id": "FIGREF3"
},
{
"start": 2410,
"end": 2418,
"text": "Figure 7",
"ref_id": "FIGREF4"
}
],
"eq_spans": [],
"section": "Comparison of Distributional Models",
"sec_num": "5.2."
},
{
"text": "The comparison of our distributional models according to the neighbors of their entries gives some insights about their proximity but no information about their relevance for representing the semantic relations in the target domain. This second type of evaluation has to rely on a reference resource accounting for these relations, which can be done in our case by exploiting the UMLS relations we have presented in Section 3.3. More precisely, we adopted the evaluation framework proposed in (Ferret, 2010) , based on the Information Retrieval paradigm: each entry of our models is considered as a query and the sorted list of its distributional neighbors as the list of retrieved documents. In this context, a neighbor is considered as relevant if the pair (entry, neighbor) corresponds to a UMLS relation 8 . As mentioned in Section 3.3, no restrictions are applied to the type of these reference relations for two main reasons. First, we wanted to have a large enough set of relations for making our evaluation as reliable as possible. Second, even at the first level, with the REL labels, the relation types are fairly fuzzy in their definition, which makes the selection of a specific type of relations difficult in practice. For measuring the relevance of the neighbors of an entry according to the UMLS relations, we adopted the classical measures used in the Information Retrieval field: Rprecision, MAP (Mean Average Precision) and precision at various ranks (P@r). R-precision (R prec ) is the precision after the first R neighbors were retrieved, R being the number of reference relations while Mean Average Precision (MAP) is the average of the precision values calculated each time a reference relation is found. All these measures are given for each of our distributional models with a scaling factor equal to 100 by the six last columns of Table 2. The second column of this table corresponds to the number of concepts in each model while the third column is the number of these concepts with at least one UMLS 8 More precisely, it means that the neighbor is part of a UMLS relation including both the entry and the neighbor.",
"cite_spans": [
{
"start": 493,
"end": 507,
"text": "(Ferret, 2010)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation of Distributional Models",
"sec_num": "5.3."
},
{
"text": "relation. The fourth column gives the average number of UMLS relations for a concept in a model and the fifth column provides the average percentage of these relations that are present in the first 100 neighbors of each concept. The results of this evaluation lead to several observations. First, their overall level seems to be fairly low. However, this is not abnormal given the size of our corpus. For instance, Ferret (2010) reports a value of 7.7 for R prec with his most complete reference (38.7 reference relations by entry on average) but with a corpus nearly four times the size of ours. We can also observe from the second column that using different syntactic parsers has a limited but not negligible influence on the number of concepts extracted from the corpus: the model based on UDpipe-Lines has 5% fewer concepts than the model based on StanfordNLP-Ewt. In terms of global trends, the first two models, StanfordNLP-Ewt and CoreNLP, are slightly better than a group of seven models with fairly close results while the last model is more clearly distant in terms of performance. This last observation is fully consistent with the separate position of the corresponding model in the dendrograms of Figure 7 . More globally, similarities between models in Table 2 are consistent with their similarities in Figure 7 , which suggests that even without an external reference, the distributional models can be compared in terms of semantic relevance by focusing on the neighbors retrieved for their entries. For instance, StanfordNLP-Ewt, CoreNLP, and StanfordNLP-Gum are close to each other in the two evaluations. This is also the case for UDpipe-Ewt, Spacy, and UDpipe-Gum. The main difference between the two evaluations concerns the relative importance of the training corpus and the type of the parser: in Table 2 , the type of the parser seems to be the main factor while in Figure 7 , the two factors are more intertwined. Figure 8 gives a more global view of similarities between models according to the UMLS relations they retrieve by reporting the same type of analysis as Figure 7 but restricted to neighbors having a UMLS relation with their entry. This view confirms the main observations resulting from the analysis of Table 2 . The model built with Talismane is significantly different from the others and the main patterns in terms of clustering are present, with a group made up of CoreNLP, StanfordNLP-Gum, and StanfordNLP-Ewt and a group with UDpipe-Gum and UDpipe-Ewt. As a con- Figure 8 : Hierarchical clustering of models according to their agreement on the first nearest neighbor with an UMLS relation with its entry (left side) and according to the RBO measure for all the neighbors having a UMLS relation with their entry (right side).",
"cite_spans": [],
"ref_spans": [
{
"start": 1211,
"end": 1219,
"text": "Figure 7",
"ref_id": "FIGREF4"
},
{
"start": 1268,
"end": 1275,
"text": "Table 2",
"ref_id": "TABREF4"
},
{
"start": 1318,
"end": 1326,
"text": "Figure 7",
"ref_id": "FIGREF4"
},
{
"start": 1820,
"end": 1827,
"text": "Table 2",
"ref_id": "TABREF4"
},
{
"start": 1890,
"end": 1898,
"text": "Figure 7",
"ref_id": "FIGREF4"
},
{
"start": 1939,
"end": 1947,
"text": "Figure 8",
"ref_id": null
},
{
"start": 2092,
"end": 2100,
"text": "Figure 7",
"ref_id": "FIGREF4"
},
{
"start": 2242,
"end": 2249,
"text": "Table 2",
"ref_id": "TABREF4"
},
{
"start": 2508,
"end": 2516,
"text": "Figure 8",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation of Distributional Models",
"sec_num": "5.3."
},
{
"text": "sequence, this evaluation emphasizes that the type of the parser used for extracting dependency triplets is the first criterion in terms of impact on the distributional models built from them but it also shows that in this context, the corpora used for training these parsers also have an influence and that heterogeneous corpora such as GUM and EWT are probably better for this training than a much more homogeneous corpus such as LinES.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation of Distributional Models",
"sec_num": "5.3."
},
{
"text": "In this article, we have investigated the influence of syntactic parsers on the distributional count-based models built from syntactic dependencies. More precisely, we have performed this study in the context of a specialized domain in the biomedical area with a moderate-size corpus made of scientific articles. One particularity of this study is to focus on the concepts of a reference ontology in the medical and biomedical areas. These concepts are mainly present in documents through multi-terms and identified by a reference tool, MetaMap, which requires aligning MetaMap's results with the results of the considered parsers. We have investigated the differences between parsers in terms of syntactic triplets but also in terms of distributional neighbors extracted from the models built from these triplets, both with and without an external reference concerning the semantic relations between concepts. We have more particularly shown the influence of the type of parser in these different evaluations but also the impact of the corpus used for training the parsers. Finally, we have found that some patterns of proximity between parsers are stable across our evaluations, which means that some measures applied to the output of syntactic parsers may perhaps be used to anticipate the performance of a parser for building distributional models from a given corpus. This will be the focus of our future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Perspectives",
"sec_num": "6."
},
{
"text": "This work has been funded by French National Research Agency (ANR) under project ADDICTE (ANR-17-CE23-0001).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": "7."
},
{
"text": "https://github.com/clir/ clearnlp-guidelines/blob/master/md/ specifications/dependency_labels.md 3 https://www.ncbi.nlm.nih.gov/books/ NBK9684/#_ch02_sec2_4_",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We tested the use of one token instead of two but found better results with two tokens.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://www.mathcs.emory.edu/~choi/doc/ cu-2012-choi.pdf",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Ratio of the number of words sharing the same nearest neighbor to the size of the considered vocabulary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Converting an english-swedish parallel treebank to universal dependencies",
"authors": [
{
"first": "L",
"middle": [],
"last": "Ahrenberg",
"suffix": ""
}
],
"year": 2015,
"venue": "Third International Conference on Dependency Linguistics",
"volume": "",
"issue": "",
"pages": "10--19",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ahrenberg, L. (2015). Converting an english-swedish parallel treebank to universal dependencies. In Third International Conference on Dependency Linguistics (DepLing 2015), Uppsala, Sweden, August 24-26, 2015, pages 10-19. Association for Computational Linguistics.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "An overview of metamap: historical perspective and recent advances",
"authors": [
{
"first": "A",
"middle": [
"R"
],
"last": "Aronson",
"suffix": ""
},
{
"first": "F.-M",
"middle": [],
"last": "Lang",
"suffix": ""
}
],
"year": 2010,
"venue": "Journal of the American Medical Informatics Association (JAMIA)",
"volume": "17",
"issue": "3",
"pages": "229--236",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aronson, A. R. and Lang, F.-M. (2010). An overview of metamap: historical perspective and recent advances. Journal of the American Medical Informatics Associa- tion (JAMIA), 17(3):229-236.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Effective mapping of biomedical text to the umls metathesaurus: the metamap program",
"authors": [
{
"first": "A",
"middle": [
"R"
],
"last": "Aronson",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of the AMIA Symposium",
"volume": "17",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aronson, A. R. (2001). Effective mapping of biomedical text to the umls metathesaurus: the metamap program. In Proceedings of the AMIA Symposium, page 17. Amer- ican Medical Informatics Association.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Distributional memory: A general framework for corpus-based semantics",
"authors": [
{
"first": "M",
"middle": [],
"last": "Baroni",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Lenci",
"suffix": ""
}
],
"year": 2010,
"venue": "Computational Linguistics",
"volume": "36",
"issue": "4",
"pages": "673--721",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Baroni, M. and Lenci, A. (2010). Distributional memory: A general framework for corpus-based semantics. Com- putational Linguistics, 36(4):673-721.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Don't count, predict! A systematic comparison of contextcounting vs. context-predicting semantic vectors",
"authors": [
{
"first": "M",
"middle": [],
"last": "Baroni",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Dinu",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Kruszewski",
"suffix": ""
}
],
"year": 2014,
"venue": "52nd Annual Meeting of the Association for Computational Linguistics (ACL 2014)",
"volume": "",
"issue": "",
"pages": "238--247",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Baroni, M., Dinu, G., and Kruszewski, G. (2014). Don't count, predict! A systematic comparison of context- counting vs. context-predicting semantic vectors. In 52nd Annual Meeting of the Association for Computa- tional Linguistics (ACL 2014), pages 238-247, Balti- more, Maryland.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "The Unified Medical Language System (UMLS): integrating biomedical terminology",
"authors": [
{
"first": "O",
"middle": [],
"last": "Bodenreider",
"suffix": ""
}
],
"year": 2004,
"venue": "Nucleic acids research",
"volume": "32",
"issue": "1",
"pages": "267--270",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bodenreider, O. (2004). The Unified Medical Language System (UMLS): integrating biomedical terminology. Nucleic acids research, 32(suppl 1):D267-D270.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "The parallel-TUT: a multilingual and multiformat treebank",
"authors": [
{
"first": "C",
"middle": [],
"last": "Bosco",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Sanguinetti",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Lesmo",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of LREC",
"volume": "",
"issue": "",
"pages": "1932--1938",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bosco, C., Sanguinetti, M., and Lesmo, L. (2012). The parallel-TUT: a multilingual and multiformat treebank. In Proceedings of LREC, pages 1932-1938. European Language Resources Association (ELRA).",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Benchmarking of statistical dependency parsers for French",
"authors": [
{
"first": "M",
"middle": [],
"last": "Candito",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Nivre",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Denis",
"suffix": ""
},
{
"first": "E",
"middle": [
"H"
],
"last": "Anguiano",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the 23rd International Conference on Computational Linguistics: Posters",
"volume": "",
"issue": "",
"pages": "108--116",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Candito, M., Nivre, J., Denis, P., and Anguiano, E. H. (2010). Benchmarking of statistical dependency parsers for French. In Proceedings of the 23rd International Conference on Computational Linguistics: Posters, pages 108-116. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Improvements in automatic thesaurus extraction",
"authors": [
{
"first": "J",
"middle": [
"R"
],
"last": "Curran",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Moens",
"suffix": ""
}
],
"year": 2002,
"venue": "Workshop of the ACL Special Interest Group on the Lexicon (SIGLEX)",
"volume": "",
"issue": "",
"pages": "59--66",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Curran, J. R. and Moens, M. (2002). Improvements in au- tomatic thesaurus extraction. In Workshop of the ACL Special Interest Group on the Lexicon (SIGLEX), pages 59-66, Philadelphia, USA.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Passage: from French parser evaluation to large sized treebank",
"authors": [
{
"first": "E",
"middle": [
"V"
],
"last": "De La Clergerie",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Hamon",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Mostefa",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Ayache",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Paroubek",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Vilnat",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of LREC",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "De La Clergerie, E. V., Hamon, O., Mostefa, D., Ayache, C., Paroubek, P., and Vilnat, A. (2008). Passage: from French parser evaluation to large sized treebank. In Pro- ceedings of LREC.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Jouer avec des analyseurs syntaxiques",
"authors": [
{
"first": "\u00c9",
"middle": [
"V"
],
"last": "De La Clergerie",
"suffix": ""
}
],
"year": 2014,
"venue": "Actes de TALN",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "De La Clergerie, \u00c9. V. (2014). Jouer avec des analyseurs syntaxiques. In Actes de TALN.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Testing semantic similarity measures for extracting synonyms from a corpus",
"authors": [
{
"first": "O",
"middle": [],
"last": "Ferret",
"suffix": ""
}
],
"year": 2010,
"venue": "th International Conference on Language Resources and Evaluation (LREC'10)",
"volume": "7",
"issue": "",
"pages": "3338--3343",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ferret, O. (2010). Testing semantic similarity measures for extracting synonyms from a corpus. In 7 th Interna- tional Conference on Language Resources and Evalua- tion (LREC'10), pages 3338-3343, Valletta, Malta.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Explorations in automatic thesaurus discovery",
"authors": [
{
"first": "G",
"middle": [],
"last": "Grefenstette",
"suffix": ""
}
],
"year": 1994,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Grefenstette, G. (1994). Explorations in automatic the- saurus discovery. Kluwer.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Symbolic word clustering for medium-size corpora",
"authors": [
{
"first": "B",
"middle": [],
"last": "Habert",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Naulleau",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Nazarenko",
"suffix": ""
}
],
"year": 1996,
"venue": "16th International Conference on Computational Linguistics (COLING 1996)",
"volume": "",
"issue": "",
"pages": "490--495",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Habert, B., Naulleau, E., and Nazarenko, A. (1996). Sym- bolic word clustering for medium-size corpora. In 16th International Conference on Computational Linguistics (COLING 1996), pages 490-495, Copenhagen, Den- mark.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "An Improved Non-monotonic Transition System for Dependency Parsing",
"authors": [
{
"first": "M",
"middle": [],
"last": "Honnibal",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Johnson",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1373--1378",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Honnibal, M. and Johnson, M. (2015). An Improved Non-monotonic Transition System for Dependency Pars- ing. In Proceedings of the 2015 Conference on Empir- ical Methods in Natural Language Processing, pages 1373-1378, Lisbon, Portugal, September. Association for Computational Linguistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Ontonotes: The 90% solution",
"authors": [
{
"first": "E",
"middle": [],
"last": "Hovy",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Marcus",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Palmer",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Ramshaw",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Weischedel",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Human Language Technology Conference of the NAACL (HLT-NAACL 2006)",
"volume": "",
"issue": "",
"pages": "57--60",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hovy, E., Marcus, M., Palmer, M., Ramshaw, L., and Weischedel, R. (2006). Ontonotes: The 90% solu- tion. In Proceedings of the Human Language Technology Conference of the NAACL (HLT-NAACL 2006), Short Pa- pers, page 57-60, USA. Association for Computational Linguistics.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "A Systematic Study of Semantic Vector Space Model Parameters",
"authors": [
{
"first": "D",
"middle": [],
"last": "Kiela",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Clark",
"suffix": ""
}
],
"year": 2014,
"venue": "2nd Workshop on Continuous Vector Space Models and their Compositionality (CVSC)",
"volume": "",
"issue": "",
"pages": "21--30",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kiela, D. and Clark, S. (2014). A Systematic Study of Se- mantic Vector Space Model Parameters. In 2nd Work- shop on Continuous Vector Space Models and their Com- positionality (CVSC), pages 21-30, Gothenburg, Swe- den.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Simple semi-supervised dependency parsing",
"authors": [
{
"first": "T",
"middle": [],
"last": "Koo",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Carreras",
"suffix": ""
},
{
"first": "Collins",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "595--603",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Koo, T., Carreras, X., and Collins, M. (2008). Sim- ple semi-supervised dependency parsing. Proceedings of ACL-08: HLT, pages 595-603.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Large-scale evaluation of dependency-based DSMs: Are they worth the effort?",
"authors": [
{
"first": "G",
"middle": [],
"last": "Lapesa",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Evert",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "394--400",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lapesa, G. and Evert, S. (2017). Large-scale evaluation of dependency-based DSMs: Are they worth the effort? In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguis- tics (EACL 2017), Short Papers, volume 2, pages 394- 400.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Dependency-Based Word Embeddings",
"authors": [
{
"first": "O",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Goldberg",
"suffix": ""
}
],
"year": 2014,
"venue": "52nd Annual Meeting of the Association for Computational Linguistics (ACL 2014)",
"volume": "",
"issue": "",
"pages": "302--308",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Levy, O. and Goldberg, Y. (2014). Dependency-Based Word Embeddings. In 52nd Annual Meeting of the Asso- ciation for Computational Linguistics (ACL 2014), pages 302-308, Baltimore, Maryland, June.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Improving Distributional Similarity with Lessons Learned from Word Embeddings",
"authors": [
{
"first": "O",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Goldberg",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Dagan",
"suffix": ""
}
],
"year": 2015,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "3",
"issue": "",
"pages": "211--225",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Levy, O., Goldberg, Y., and Dagan, I. (2015). Improv- ing Distributional Similarity with Lessons Learned from Word Embeddings. Transactions of the Association for Computational Linguistics, 3:211-225.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Automatic retrieval and clustering of similar words",
"authors": [
{
"first": "D",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 1998,
"venue": "Proceedings of the 17th international conference on Computational linguistics",
"volume": "",
"issue": "",
"pages": "768--774",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lin, D. (1998). Automatic retrieval and clustering of simi- lar words. In Proceedings of the 17th international con- ference on Computational linguistics, pages 768-774. Association for Computational Linguistics.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "The Stanford CoreNLP natural language processing toolkit",
"authors": [
{
"first": "C",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Surdeanu",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Bauer",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Finkel",
"suffix": ""
},
{
"first": "S",
"middle": [
"J"
],
"last": "Bethard",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Mcclosky",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
"volume": "",
"issue": "",
"pages": "55--60",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J., Bethard, S. J., and McClosky, D. (2014). The Stanford CoreNLP natural language processing toolkit. In Pro- ceedings of the 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 55-60.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Building a large annotated corpus of English: The Penn Treebank",
"authors": [
{
"first": "M",
"middle": [],
"last": "Marcus",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Santorini",
"suffix": ""
},
{
"first": "M",
"middle": [
"A"
],
"last": "Marcinkiewicz",
"suffix": ""
}
],
"year": 1993,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marcus, M., Santorini, B., and Marcinkiewicz, M. A. (1993). Building a large annotated corpus of English: The Penn Treebank.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Efficient estimation of word representations in vector space",
"authors": [
{
"first": "T",
"middle": [],
"last": "Mikolov",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Corrado",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Dean",
"suffix": ""
}
],
"year": 2013,
"venue": "ICLR 2013, workshop track",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013). Efficient estimation of word representations in vector space. In ICLR 2013, workshop track.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Dependency-based construction of semantic space models",
"authors": [
{
"first": "S",
"middle": [],
"last": "Pad\u00f3",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Lapata",
"suffix": ""
}
],
"year": 2007,
"venue": "Computational Linguistics",
"volume": "33",
"issue": "2",
"pages": "161--199",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pad\u00f3, S. and Lapata, M. (2007). Dependency-based con- struction of semantic space models. Computational Lin- guistics, 33(2):161-199.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Evaluation of Parsers of French: what are the results?",
"authors": [],
"year": null,
"venue": "Proceedings of LREC",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "EASY, Evaluation of Parsers of French: what are the re- sults? In Proceedings of LREC.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Towards qualitative word embeddings evaluation: Measuring neighbors variation",
"authors": [
{
"first": "B",
"middle": [],
"last": "Pierrejean",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Tanguy",
"suffix": ""
}
],
"year": 2018,
"venue": "Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop",
"volume": "",
"issue": "",
"pages": "32--39",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pierrejean, B. and Tanguy, L. (2018). Towards qualitative word embeddings evaluation: Measuring neighbors vari- ation. In Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 32-39.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Universal Dependency Parsing from Scratch",
"authors": [
{
"first": "P",
"middle": [],
"last": "Qi",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Dozat",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "C",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies",
"volume": "",
"issue": "",
"pages": "160--170",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Qi, P., Dozat, T., Zhang, Y., and Manning, C. D. (2018). Universal Dependency Parsing from Scratch. In Pro- ceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, pages 160-170, Brussels, Belgium, October. Association for Computational Linguistics.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Mayo clinical Text Analysis and Knowledge Extraction System (cTAKES): architecture, component evaluation and applications",
"authors": [],
"year": null,
"venue": "Journal of the American Medical Informatics Association",
"volume": "17",
"issue": "5",
"pages": "507--513",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mayo clinical Text Analysis and Knowledge Extraction System (cTAKES): architecture, component evaluation and applications. Journal of the American Medical In- formatics Association, 17(5):507-513.",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "Overview of the SPMRL 2013 shared task: A cross-framework evaluation of parsing morphologically rich languages",
"authors": [
{
"first": "D",
"middle": [],
"last": "Seddah",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Tsarfaty",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "K\u00fcbler",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Candito",
"suffix": ""
},
{
"first": "J",
"middle": [
"D"
],
"last": "Choi",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Farkas",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Foster",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Goenaga",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Gojenola Galletebeitia",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Goldberg",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Green",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Habash",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Kuhlmann",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Maier",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Nivre",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Przepi\u00f3rkowski",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Roth",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Seeker",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Versley",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Vincze",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Woli\u0144ski",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Wr\u00f3blewska",
"suffix": ""
},
{
"first": "Villemonte",
"middle": [],
"last": "De La Clergerie",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of the Fourth Workshop on Statistical Parsing of Morphologically-Rich Languages",
"volume": "",
"issue": "",
"pages": "146--182",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Seddah, D., Tsarfaty, R., K\u00fcbler, S., Candito, M., Choi, J. D., Farkas, R., Foster, J., Goenaga, I., Gojenola Gal- letebeitia, K., Goldberg, Y., Green, S., Habash, N., Kuhlmann, M., Maier, W., Nivre, J., Przepi\u00f3rkowski, A., Roth, R., Seeker, W., Versley, Y., Vincze, V., Woli\u0144ski, M., Wr\u00f3blewska, A., and Villemonte de la Clergerie, E. (2013). Overview of the SPMRL 2013 shared task: A cross-framework evaluation of parsing morpholog- ically rich languages. In Proceedings of the Fourth Workshop on Statistical Parsing of Morphologically- Rich Languages, pages 146-182, Seattle, Washington, USA, October. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF33": {
"ref_id": "b33",
"title": "A gold standard dependency corpus for English",
"authors": [
{
"first": "N",
"middle": [],
"last": "Silveira",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Dozat",
"suffix": ""
},
{
"first": "M.-C",
"middle": [],
"last": "De Marneffe",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Bowman",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Connor",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Bauer",
"suffix": ""
},
{
"first": "C",
"middle": [
"D"
],
"last": "Manning",
"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": "Silveira, N., Dozat, T., de Marneffe, M.-C., Bowman, S., Connor, M., Bauer, J., and Manning, C. D. (2014). A gold standard dependency corpus for English. In Pro- ceedings of the Ninth International Conference on Lan- guage Resources and Evaluation (LREC-2014).",
"links": null
},
"BIBREF34": {
"ref_id": "b34",
"title": "Quickumls: a fast, unsupervised approach for medical concept extraction",
"authors": [
{
"first": "L",
"middle": [],
"last": "Soldaini",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Goharian",
"suffix": ""
}
],
"year": 2016,
"venue": "MedIR workshop, sigir",
"volume": "",
"issue": "",
"pages": "1--4",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Soldaini, L. and Goharian, N. (2016). Quickumls: a fast, unsupervised approach for medical concept extraction. In MedIR workshop, sigir, pages 1-4.",
"links": null
},
"BIBREF35": {
"ref_id": "b35",
"title": "Tokenizing, POS tagging, lemmatizing and parsing UD 2.0 with UDPipe",
"authors": [
{
"first": "M",
"middle": [],
"last": "Straka",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Strakov\u00e1",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies",
"volume": "",
"issue": "",
"pages": "88--99",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Straka, M. and Strakov\u00e1, J. (2017). Tokenizing, POS tag- ging, lemmatizing and parsing UD 2.0 with UDPipe. In Proceedings of the CoNLL 2017 Shared Task: Multilin- gual Parsing from Raw Text to Universal Dependencies, pages 88-99, Vancouver, Canada, August. Association for Computational Linguistics.",
"links": null
},
"BIBREF36": {
"ref_id": "b36",
"title": "\u00c9valuation sur mesure de mod\u00e8les distributionnels sur un corpus sp\u00e9cialis\u00e9: comparaison des approches par contextes syntaxiques et par fen\u00eatres graphiques",
"authors": [
{
"first": "L",
"middle": [],
"last": "Tanguy",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Sajous",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Hathout",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "56",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tanguy, L., Sajous, F., and Hathout, N. (2015). \u00c9valua- tion sur mesure de mod\u00e8les distributionnels sur un cor- pus sp\u00e9cialis\u00e9: comparaison des approches par contextes syntaxiques et par fen\u00eatres graphiques. Traitement au- tomatique des langues, 56(2).",
"links": null
},
"BIBREF37": {
"ref_id": "b37",
"title": "Extrinsic evaluation of french dependency parsers on a specialised corpus: comparison of distributional thesauri",
"authors": [
{
"first": "L",
"middle": [],
"last": "Tanguy",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Brunet",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Ferret",
"suffix": ""
}
],
"year": 2020,
"venue": "12th International Conference on Language Resources and Evaluation (LREC 2020)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tanguy, L., Brunet, P., and Ferret, O. (2020). Extrinsic evaluation of french dependency parsers on a specialised corpus: comparison of distributional thesauri. In 12th International Conference on Language Resources and Evaluation (LREC 2020), Marseille, France.",
"links": null
},
"BIBREF38": {
"ref_id": "b38",
"title": "L'apport du faisceau dans l'analyse syntaxique en d\u00e9pendances par transitions : \u00e9tudes de cas avec l'analyseur Talismane",
"authors": [
{
"first": "A",
"middle": [],
"last": "Urieli",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Tanguy",
"suffix": ""
}
],
"year": 2013,
"venue": "Actes de TALN",
"volume": "",
"issue": "",
"pages": "188--201",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Urieli, A. and Tanguy, L. (2013). L'apport du faisceau dans l'analyse syntaxique en d\u00e9pendances par transitions : \u00e9tudes de cas avec l'analyseur Talismane. In Actes de TALN, pages 188-201, Les Sables d'Olonne, France.",
"links": null
},
"BIBREF39": {
"ref_id": "b39",
"title": "A similarity measure for indefinite rankings",
"authors": [
{
"first": "W",
"middle": [],
"last": "Webber",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Moffat",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Zobel",
"suffix": ""
}
],
"year": 2010,
"venue": "ACM Transactions on Information Systems (TOIS)",
"volume": "28",
"issue": "4",
"pages": "1--38",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Webber, W., Moffat, A., and Zobel, J. (2010). A similarity measure for indefinite rankings. ACM Transactions on Information Systems (TOIS), 28(4):1-38, November.",
"links": null
},
"BIBREF40": {
"ref_id": "b40",
"title": "The GUM corpus: Creating multilayer resources in the classroom",
"authors": [
{
"first": "A",
"middle": [],
"last": "Zeldes",
"suffix": ""
}
],
"year": 2017,
"venue": "Language Resources and Evaluation",
"volume": "51",
"issue": "3",
"pages": "581--612",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zeldes, A. (2017). The GUM corpus: Creating multilayer resources in the classroom. Language Resources and Evaluation, 51(3):581-612.",
"links": null
},
"BIBREF41": {
"ref_id": "b41",
"title": "CoNLL 2018 shared task: Multilingual parsing from raw text to universal dependencies",
"authors": [
{
"first": "D",
"middle": [],
"last": "Zeman",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Haji",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Popel",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Potthast",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Straka",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Ginter",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Nivre",
"suffix": ""
},
{
"first": "S",
"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": {},
"num": null,
"urls": [],
"raw_text": "Zeman, D., Haji, J., Popel, M., Potthast, M., Straka, M., Ginter, F., Nivre, J., and Petrov, S. (2018). CoNLL 2018 shared task: Multilingual parsing from raw text to uni- versal dependencies. Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Uni- versal Dependencies, pages 1-21.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "the representation of the contexts of a word in a corpus under the form of syntactic triplets (dependent, relation, Dependencies identified by the Stanford NLP parser trained on the LinES corpus for the phrase \"Generation of single-copy transgenic mouse embryos\".",
"uris": null,
"type_str": "figure"
},
"FIGREF1": {
"num": null,
"text": "gives the result of the parsing by Spacy. Finally,Figure 3presents the output of Talismane for this phrase. These three cases are the three basic patterns of how prepositions are managed in each scheme. We normalize all these region within the cluster prep pobj Dependencies identified by SpaCy (Clear Style dependencies) 6 for the phrase \"region within the cluster\". region within the cluster NMOD POBJ Dependencies identified by Talismane (Penn Treebank dependencies) for the phrase \"region within the cluster\".",
"uris": null,
"type_str": "figure"
},
"FIGREF2": {
"num": null,
"text": "suj V : nominal subject of a verb; N obj V : nominal direct object of a verb; ADJ mod N : adjective modifying a noun;",
"uris": null,
"type_str": "figure"
},
"FIGREF3": {
"num": null,
"text": "Hierarchical clustering of parsers according to their correlation on the triplets found by at least two parsers (right side) and all parsers (left side).",
"uris": null,
"type_str": "figure"
},
"FIGREF4": {
"num": null,
"text": "Hierarchical clustering of models according to their agreement on the first nearest neighbor (left side) and according to the RBO measure (right side).",
"uris": null,
"type_str": "figure"
},
"TABREF2": {
"html": null,
"text": "Syntactic triplets with concepts identified by MetaMap",
"type_str": "table",
"num": null,
"content": "<table><tr><td>0.40</td><td>0.45</td><td>0.50</td><td>0.55</td><td>0.60</td><td>0.65</td><td>0.70</td><td>0.08</td><td>0.10</td><td>0.12</td><td>0.14</td><td>0.16</td></tr><tr><td/><td/><td>UDpipe-Gum</td><td/><td/><td/><td/><td/><td>Spacy</td><td/><td/><td/></tr><tr><td/><td/><td>UDpipe-Ewt</td><td/><td/><td/><td colspan=\"2\">CoreNLP</td><td/><td/><td/><td/></tr><tr><td/><td/><td colspan=\"2\">UDpipe-Lines</td><td/><td/><td>StanfordnNLP-Gum</td><td/><td/><td/><td/><td/></tr><tr><td/><td/><td colspan=\"2\">UDpipe-Partut</td><td/><td/><td>StanfordNLP-Ewt</td><td/><td/><td/><td/><td/></tr><tr><td colspan=\"2\">StanfordnNLP-Gum</td><td/><td/><td/><td/><td/><td/><td>UDpipe-Gum</td><td/><td/><td/></tr><tr><td colspan=\"2\">StanfordNLP-Lines</td><td/><td/><td/><td/><td/><td/><td>UDpipe-Ewt</td><td/><td/><td/></tr><tr><td/><td/><td>Spacy</td><td/><td/><td/><td/><td/><td/><td colspan=\"2\">UDpipe-Partut</td><td/></tr><tr><td>StanfordNLP-Ewt</td><td/><td/><td/><td/><td/><td/><td/><td colspan=\"2\">UDpipe-Lines</td><td/><td/></tr><tr><td>CoreNLP</td><td/><td/><td/><td/><td/><td/><td/><td colspan=\"2\">StanfordNLP-Lines</td><td/><td/></tr></table>"
},
"TABREF4": {
"html": null,
"text": "Evaluation of our distributional models with UMLS relations as reference (measures x 100).",
"type_str": "table",
"num": null,
"content": "<table/>"
}
}
}
} |