File size: 91,678 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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T06:07:32.854803Z"
},
"title": "Emotional RobBERT and Insensitive BERTje: Combining Transformers and Affect Lexica for Dutch Emotion Detection",
"authors": [
{
"first": "Luna",
"middle": [],
"last": "De Bruyne",
"suffix": "",
"affiliation": {
"laboratory": "Language and Translation Technology Team Ghent University",
"institution": "",
"location": {}
},
"email": "[email protected]"
},
{
"first": "Orph\u00e9e",
"middle": [],
"last": "De Clercq",
"suffix": "",
"affiliation": {
"laboratory": "Language and Translation Technology Team Ghent University",
"institution": "",
"location": {}
},
"email": "[email protected]"
},
{
"first": "V\u00e9ronique",
"middle": [
"Hoste"
],
"last": "Lt",
"suffix": "",
"affiliation": {
"laboratory": "Language and Translation Technology Team Ghent University",
"institution": "",
"location": {}
},
"email": "[email protected]"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In a first step towards improving Dutch emotion detection, we try to combine the Dutch transformer models BERTje and RobBERT with lexicon-based methods. We propose two architectures: one in which lexicon information is directly injected into the transformer model and a meta-learning approach where predictions from transformers are combined with lexicon features. The models are tested on 1,000 Dutch tweets and 1,000 captions from TV-shows which have been manually annotated with emotion categories and dimensions. We find that RobBERT clearly outperforms BERTje, but that directly adding lexicon information to transformers does not improve performance. In the meta-learning approach, lexicon information does have a positive effect on BERTje, but not on RobBERT. This suggests that more emotional information is already contained within this latter language model.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "In a first step towards improving Dutch emotion detection, we try to combine the Dutch transformer models BERTje and RobBERT with lexicon-based methods. We propose two architectures: one in which lexicon information is directly injected into the transformer model and a meta-learning approach where predictions from transformers are combined with lexicon features. The models are tested on 1,000 Dutch tweets and 1,000 captions from TV-shows which have been manually annotated with emotion categories and dimensions. We find that RobBERT clearly outperforms BERTje, but that directly adding lexicon information to transformers does not improve performance. In the meta-learning approach, lexicon information does have a positive effect on BERTje, but not on RobBERT. This suggests that more emotional information is already contained within this latter language model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Computational analysis of affect in Dutch texts is mostly restricted to polarity analysis (negative/positive/neutral), for which we know a tradition of lexicon-based approaches. Recently, a BERT-based model, BERTje (de Vries et al., 2019) , and a RoBERTa-based model, RobBERT (Delobelle et al., 2020) , have been created for Dutch, and they have achieved promising results on the task of sentiment analysis. For emotion detection, however, these models have not yet been evaluated.",
"cite_spans": [
{
"start": 215,
"end": 238,
"text": "(de Vries et al., 2019)",
"ref_id": "BIBREF30"
},
{
"start": 276,
"end": 300,
"text": "(Delobelle et al., 2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In a first step towards improving emotion detection for Dutch, we will evaluate BERTje and RobBERT on the task of emotion detection. Instead of casting aside the many efforts that have been made in the creation of Dutch sentiment and emotion lexica, we will investigate whether transformers and affect lexica can complement each other. Attempts of combining BERT models with additional features have already been successful for tasks like abusive language and sarcasm detection (Koufakou et al., 2020; Kumar et al., 2021) .",
"cite_spans": [
{
"start": 478,
"end": 501,
"text": "(Koufakou et al., 2020;",
"ref_id": "BIBREF13"
},
{
"start": 502,
"end": 521,
"text": "Kumar et al., 2021)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We consider two architectures. In the first one, we inject lexicon information in the transformer model before the prediction layer. We do this by concatenating the [CLS] token of the target sentence (which BERT and RoBERTa models use as input for prediction) with a lexicon vector obtained from seven Dutch affect lexica. In the second approach, we employ a meta-learning architecture and use a support vector machine (SVM) that learns from the transformer model's output. The predictions from the transformer are concatenated with the lexicon vector and used as input for the SVM.",
"cite_spans": [
{
"start": 165,
"end": 170,
"text": "[CLS]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We evaluate our models on 1,000 Dutch Tweets and 1,000 transcribed utterances from Flemish TVshows. As multiple researchers have emphasized the need of studying emotions not only in terms of basic emotions, but based on dimensions like valence, arousal and dominance as well (Buechel and Hahn, 2016; , the data has been annotated in a bi-representational design: both with categorical annotations for anger, joy, fear, love, sadness or neutral, and scores for the dimensions valence, arousal and dominance. First we will discuss related work on Dutch emotion detection and other attempts on combining transformer models with additional features in Section 2. In Section 3, we will describe the methodology of our experiments and in Section 4, we report the results. We end with a conclusion in Section 5.",
"cite_spans": [
{
"start": 275,
"end": 299,
"text": "(Buechel and Hahn, 2016;",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Although most emotion detection research deals with the English language, recent studies have shown interest in other languages as well, e.g. the recent work of Ahmad et al. (2020) SemEval-2018 task on Affect in Tweets, for which in addition to English data, Arabic and Spanish datasets were released . Moreover, research on multilingual emotion detection (Buechel and Hahn, 2018; \u00d6hman et al., 2018) and sentiment analysis (Lo et al., 2017; Vilares et al., 2018) is emerging.",
"cite_spans": [
{
"start": 161,
"end": 180,
"text": "Ahmad et al. (2020)",
"ref_id": "BIBREF0"
},
{
"start": 356,
"end": 380,
"text": "(Buechel and Hahn, 2018;",
"ref_id": "BIBREF3"
},
{
"start": 381,
"end": 400,
"text": "\u00d6hman et al., 2018)",
"ref_id": null
},
{
"start": 424,
"end": 441,
"text": "(Lo et al., 2017;",
"ref_id": "BIBREF17"
},
{
"start": 442,
"end": 463,
"text": "Vilares et al., 2018)",
"ref_id": "BIBREF29"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Concerning the automatic modelling of affect in Dutch, the main focus still is on sentiment analysis instead of fine-grained emotions. Many studies have used sentiment lexica for this purpose (e.g. Van de Kauter et al., 2015; De Clercq et al., 2017; Wang et al., 2020) . Recently, transformer models like BERTje (de Vries et al., 2019) and RobBERT (Delobelle et al., 2020) have been used to classify reviews from the Dutch Book Reviews dataset as either positive or negative, in which RobBERT came out as best (accuracy of 95% versus 93%).",
"cite_spans": [
{
"start": 205,
"end": 225,
"text": "Kauter et al., 2015;",
"ref_id": "BIBREF12"
},
{
"start": 226,
"end": 249,
"text": "De Clercq et al., 2017;",
"ref_id": "BIBREF6"
},
{
"start": 250,
"end": 268,
"text": "Wang et al., 2020)",
"ref_id": "BIBREF31"
},
{
"start": 348,
"end": 372,
"text": "(Delobelle et al., 2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "The only publicly available dataset for Dutch emotion detection is deLearyous (Vaassen and Daelemans, 2011) . It consists of 740 Dutch sentences from conversations, annotated according to Leary's Rose (Leary, 1957) . Vaassen and Daelemans (2011) found that classifying sentences from deLearyous into the quadrants or octants of Leary's Rose was difficult for machine learners, most likely because of the sparseness of the data and low interannotator agreement. However, after this study, no further studies on Dutch emotion detection were published. This clearly shows the need of new data and methods for this task.",
"cite_spans": [
{
"start": 78,
"end": 107,
"text": "(Vaassen and Daelemans, 2011)",
"ref_id": "BIBREF28"
},
{
"start": 201,
"end": 214,
"text": "(Leary, 1957)",
"ref_id": "BIBREF15"
},
{
"start": 217,
"end": 245,
"text": "Vaassen and Daelemans (2011)",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Some researchers already revealed that combining BERT models with handcrafted features can have a positive effect on performance, e.g. for the task of essay scoring (accuracy of 71% versus 80%) (Uto et al., 2020) and sarcasm detection (F1score of 78% versus 80%) (Kumar et al., 2021) . Lexicon features have been combined with BERT as well, e.g. for abusive language detection (Koufakou et al., 2020) , for which the authors found an improvement on four out of six datasets compared to a plain BERT model (maximum improve-ment of 3%). In all of these studies, the handcrafted features or lexicon features were injected into the transformer architecture by concatenating them with the BERT representation before the prediction layer.",
"cite_spans": [
{
"start": 194,
"end": 212,
"text": "(Uto et al., 2020)",
"ref_id": "BIBREF27"
},
{
"start": 263,
"end": 283,
"text": "(Kumar et al., 2021)",
"ref_id": null
},
{
"start": 377,
"end": 400,
"text": "(Koufakou et al., 2020)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "We collect data from two domains: Twitter posts (Tweets subcorpus) and utterances from reality TVshows (Captions). For the Tweets subcorpus, a list of 72 emojis was used as query in the Dutch tweets database Twiqs.nl, with as search period 1-1-2017 to 31-12-2017. Based on this one-year datadump we sampled a random subset of 1,000 tweets, but made sure that no duplicates or non-Dutch tweets were present in the dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method 3.1 Data",
"sec_num": "3"
},
{
"text": "For Captions, episodes of three emotionally loaded Flemish reality TV-shows (Blind getrouwd; Bloed, zweet en luxeproblemen and Ooit vrij) were used. Three episodes per show were transcribed using a literal transcription method (without correcting colloquial elements). 1,000 utterances (sentences or short sequences of sentences) were selected from these transcripts, based on a rough screening of emotional content and more or less equally distributed over the shows (335 instances from Blind getrouwd, 331 from Bloed, zweet en luxeproblemen and 334 from Ooit vrij).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method 3.1 Data",
"sec_num": "3"
},
{
"text": "All data was annotated with both categorical labels and dimensions. For the categorical annotation, the instances were labeled with one out of six labels: joy, love, anger, fear, sadness, or neutral. For annotating the dimensions valence, arousal and dominance, best-worst scaling was employed as this was shown to be a reliable annotation method (De Bruyne et al., 2021) . Per subcorpus, the 1,000 instances were converted into 2,000 4-tuples and distributed among the annotators. For each trial, the annotator had to indicate the best and worst example for each dimension: highest and lowest valence, highest and lowest arousal, and highest and lowest dominance. Best-worst counts were then converted to scores from 0 to 1 with the Rescorla-Wagner update rule (Rescorla et al., 1972) . See Table 1 for an annotated example of an instance in each domain. Table 2 lists the number of instances per emotion category in each domain. For the valence, arousal and dominance annotations, the mean ranges between 0.46 and 0.52 for all dimensions in both subsets, the standard deviation ranges between 0.18 and 0.22, the minimum between 0.05 and 0.07 and the maximum between 0.96 and 0.97. For the experiments, both datasets were split in 800 instances for training, 100 for validating and 100 for testing (same splits for all models/tasks).",
"cite_spans": [
{
"start": 347,
"end": 371,
"text": "(De Bruyne et al., 2021)",
"ref_id": "BIBREF5"
},
{
"start": 762,
"end": 785,
"text": "(Rescorla et al., 1972)",
"ref_id": "BIBREF25"
}
],
"ref_spans": [
{
"start": 792,
"end": 799,
"text": "Table 1",
"ref_id": "TABREF1"
},
{
"start": 856,
"end": 863,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Method 3.1 Data",
"sec_num": "3"
},
{
"text": "We investigate whether lexicon information and transformer models can be complementary either by injecting lexicon information directly into the transformer architecture or by using a metalearning approach in which predictions from transformer models are combined with lexicon features. Both models require the creation of a lexicon vector per target sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lexicon information",
"sec_num": "3.2"
},
{
"text": "For the creation of this vector, we combine seven existing open-source Dutch sentiment and emotion lexica, namely Pattern (De Smedt and Daelemans, 2012) , Duoman (Jijkoun and Hofmann, 2009) , LIWC (Boot et al., 2017) , NRC Emotion (Mohammad and Turney, 2013) , NRC VAD (Mohammad, 2018), Memolon (Buechel et al., 2020) and the VAD norms by Moors et al. (2013) . For each word in the target sentence, lexicon values are obtained through a lookup in each affect lexicon. These values are then averaged over the words in the target sentence. The vector is 33-dimensional, as all lexica include values for multiple emotion categories or dimensions which add up to 33 in total. For lexica that do not have entries for any of the words in the sentence, the respective value in the lexicon vector is 0.",
"cite_spans": [
{
"start": 122,
"end": 152,
"text": "(De Smedt and Daelemans, 2012)",
"ref_id": "BIBREF7"
},
{
"start": 162,
"end": 189,
"text": "(Jijkoun and Hofmann, 2009)",
"ref_id": "BIBREF11"
},
{
"start": 197,
"end": 216,
"text": "(Boot et al., 2017)",
"ref_id": "BIBREF1"
},
{
"start": 231,
"end": 258,
"text": "(Mohammad and Turney, 2013)",
"ref_id": "BIBREF22"
},
{
"start": 295,
"end": 317,
"text": "(Buechel et al., 2020)",
"ref_id": "BIBREF4"
},
{
"start": 339,
"end": 358,
"text": "Moors et al. (2013)",
"ref_id": "BIBREF23"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Lexicon information",
"sec_num": "3.2"
},
{
"text": "In this architecture, we inject the lexicon information into the transformer model while fine-tuning the model on the emotion detection tasks: emotion classification and emotion regression with the dimensions valence, arousal and dominance (VAD), and this in both domains. This injection occurs just before the prediction layer by concatenating the [CLS] token, which is normally used on its own as input for the classification, with the lexicon vector. This concatenated vector goes through a pre-classifier (linear layer with 2,048 nodes) and then to the prediction layer with Sigmoid activation function. The model architecture is shown in Figure 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 643,
"end": 651,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Transformer model",
"sec_num": "3.3"
},
{
"text": "Two Dutch transformer models are investigated: BERTje (de Vries et al., 2019), based on BERT by Devlin et al. (2019) and RobBERT (Delobelle et al., 2020) , the Dutch version of the robustly optimized RoBERTa (Liu et al., 2019) . RobBERT is trained on 39GB of common crawl data (Su\u00e1rez et al., 2019) , while BERTje is trained on only 12GB (including multiple genres).",
"cite_spans": [
{
"start": 96,
"end": 116,
"text": "Devlin et al. (2019)",
"ref_id": "BIBREF9"
},
{
"start": 129,
"end": 153,
"text": "(Delobelle et al., 2020)",
"ref_id": "BIBREF8"
},
{
"start": 208,
"end": 226,
"text": "(Liu et al., 2019)",
"ref_id": "BIBREF16"
},
{
"start": 277,
"end": 298,
"text": "(Su\u00e1rez et al., 2019)",
"ref_id": "BIBREF26"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Transformer model",
"sec_num": "3.3"
},
{
"text": "Both models are implemented with Hugging-Face's Transformers library (Wolf et al., 2019) . We use AdamW optimizer (Loshchilov and Hutter, 2017) and the ReduceLROnPlateau learning rate scheduler with l r = 5e \u2212 5. The loss function is Binary Cross Entropy for the classification task and Mean Squared Error loss for regression. The maximum sequence length is 64 tokens, batch size is 16 for Tweets and 64 for Captions. We set dropout to 0.2 and use GELU as activation function in the implementation of Hendrycks and Gimpel (2016) . The [CLS] token based on the concatenation of the last four layers of the model is used for prediction (and concatenated with the lexicon vector). The maximum number of epochs is set to 100 with a patience of 5 for early stopping.",
"cite_spans": [
{
"start": 69,
"end": 88,
"text": "(Wolf et al., 2019)",
"ref_id": null
},
{
"start": 114,
"end": 143,
"text": "(Loshchilov and Hutter, 2017)",
"ref_id": null
},
{
"start": 501,
"end": 528,
"text": "Hendrycks and Gimpel (2016)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Transformer model",
"sec_num": "3.3"
},
{
"text": "In the second approach, we use a support vector machine as meta-learner that learns from the predictions of a transformer model. We apply the transformer model on the training set and the test set to extract probabilities and predictions for sentences. For the classification task we use the probabilities of the emotion classes as features (6 features) and for the regression task the predicted scores (3 features). During training, the output on the training set is accompanied by lexicon features (33 features) as input for the SVM. During testing, the output on the test set is combined with the lexicon features to feed the SVM. The model architecture is shown in Figure 2 .",
"cite_spans": [],
"ref_spans": [
{
"start": 669,
"end": 677,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Meta-learner",
"sec_num": "3.4"
},
{
"text": "For all subtasks, we use a linear kernel and 1.0 as regularization parameter C. We use hinge loss and L2 penalty for classification and epsilon insensitive loss (L1 loss) for regression.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Meta-learner",
"sec_num": "3.4"
},
{
"text": "We evaluate BERTje and RobBERT on the emotion classification (macro F1) and regression tasks (Pearson's r) for Tweets and Captions. We compare the results of the plain models and the models where the lexicon vector was injected.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Transformer model",
"sec_num": "4.1"
},
{
"text": "Models are run ten times to account for variability and to be able to statistically compare the performance of different methods. The mean and standard deviation of these runs are reported in Table 3 . Except for Captions classification, we observe that RobBERT is better for all datasets, with a notable margin: for Tweets classification and Captions regression there is an improvement of around 10% (from .16 to .26 and .60 to .68 respectively), and for Tweets regression the score almost doubled (from .36 to .70). This is in line with previous findings of RobBERT being more accurate in predicting sentiment (Delobelle et al., 2020) , although our results are even more distinct. The robustly optimised RoBERTa framework and the larger corpus on which RobBERT was pre-trained clearly show their effect, although scores are in general still rather low. The results for the regression tasks seem more promising than for the classification task, although scores are difficult to compare as different metrics are used. We further notice that there is quite some variation between the runs, especially for the classification tasks.",
"cite_spans": [
{
"start": 612,
"end": 636,
"text": "(Delobelle et al., 2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 192,
"end": 199,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Transformer model",
"sec_num": "4.1"
},
{
"text": "Independent two-tailed t-tests were used for assessing whether the means of the models with and without lexicon vector are statistically different on a 5% significance level. For none of the datasets adding the lexicon vector was beneficial, as there was no significant difference in mean for any of the tasks (Tweets classification with BERTje: t(9) = 1.3, p = 0.22; Tweets classification with Rob-BERT: t(9) = 0.6, p = 0.54; Tweets regression with BERTje: t(9) = 0.7, p = 0.50; Tweets regression with RobBERT: t(9) = -0.6, p = 0.52; Captions classification with BERTje: t(9) = 0.8, p = 0.42; Captions classification with RobBERT: t(9) = -1.0, p = 0.31; Captions regression with BERTje: t(9) = 0.5, p = 0.60; Captions regression with RobBERT: t(9) = -0.6, p = 0.58). In contrast to previous studies regarding sarcasm and abusive language detection, we must thus conclude that this method of injecting lexicon information into the transformer model does not lead to higher performance on the task of emotion detection. We see two possible reasons for this: a) the lexica have no added value compared to the information that is already present in the language models, and/or b) the lexica do not have enough weight compared to the high-dimensional CLS token.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Transformer model",
"sec_num": "4.1"
},
{
"text": "In the meta-learning approach, we identify the best run (lowest loss on validation set) among the plain transformer models (i.e. without lexica), and apply that model on the training and test sets of the corresponding datasets. The probabilities (for classification) and predictions (for regression) are then Table 3 : Results (macro F1-score and Pearson's r) of the plain transformer models (BERTje and RobBERT) versus the results with lexicon features added (BERTje/RobBERT + lex). F1-score and r is the average of 10 runs, standard deviation is shown between brackets. Table 4 : Results (macro F1-score and Pearson's r) of the plain transformer model versus the meta-learner. The results of an SVM model with only lexicon features are given in the first line as reference. The best result is shown in bold or bold italics if the best model is the meta-learner approach.",
"cite_spans": [],
"ref_spans": [
{
"start": 309,
"end": 316,
"text": "Table 3",
"ref_id": null
},
{
"start": 572,
"end": 579,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Meta-learner",
"sec_num": "4.2"
},
{
"text": "fed into an SVM together with the lexicon features. The results of the original plain transformer models and meta-learning models are shown in Table 4 .",
"cite_spans": [],
"ref_spans": [
{
"start": 143,
"end": 150,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Tweets Captions",
"sec_num": null
},
{
"text": "We see that a meta-learner using the output of a transformer model combined with lexicon features outperforms the plain transformer model in the case of BERTje for three out of four tasks (all but Captions classification). In the case of RobBERT, the meta-learner outperforms the plain transformer model for only one dataset (Captions regression), but only to a minor extent. In contrast to the previous approach, where lexicon information was injected directly into the transformer model, the meta-learner does seem to be able to improve performance. Where the previous approach might have failed because of the lexica not having enough weight, the meta-learner seems to be better in exploiting the lexicon information, however only for BERTje. Possibly, emotions are more contained in the RobBERT language model because of the improved training framework and notably larger training corpus, making RobBERT a more emotional language model. Therefore, RobBERT benefits less from the added lexicon information in contrast to BERTje.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tweets Captions",
"sec_num": null
},
{
"text": "As the recently developed transformer models for Dutch, BERTje and RobBERT, have not yet been tested on emotion detection tasks, we evaluated them on 1,000 Dutch Tweets and 1,000 captions for an emotion classification and regression task. We found that RobBERT outperformed BERTje in almost all cases. Further, we investigated whether these models could be enhanced with lexicon features and proposed two methods: one in which a lexicon vector was concatenated with the transformer's [CLS] token before prediction, and a meta-learning approach. In the first method, adding lexicon information did not seem beneficial. For the second approach, however, we found that the meta-learner had a positive effect in half of the cases, and especially on the models relying on BERTje, whereas the meta-learner had almost no impact on RobBERT. This is probably because more emotional information is already contained within this language model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
}
],
"back_matter": [
{
"text": "This research was carried out with the support of the Research Foundation -Flanders under a Strategic Basic Research fellowship.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Borrow from rich cousin: transfer learning for emotion detection using cross lingual embedding",
"authors": [
{
"first": "Zishan",
"middle": [],
"last": "Ahmad",
"suffix": ""
},
{
"first": "Raghav",
"middle": [],
"last": "Jindal",
"suffix": ""
},
{
"first": "Asif",
"middle": [],
"last": "Ekbal",
"suffix": ""
},
{
"first": "Pushpak",
"middle": [],
"last": "Bhattachharyya",
"suffix": ""
}
],
"year": 2020,
"venue": "Expert Systems with Applications",
"volume": "139",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zishan Ahmad, Raghav Jindal, Asif Ekbal, and Push- pak Bhattachharyya. 2020. Borrow from rich cousin: transfer learning for emotion detection us- ing cross lingual embedding. Expert Systems with Applications, 139:112851.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "The dutch translation of the linguistic inquiry and word count (liwc) 2007 dictionary",
"authors": [
{
"first": "Peter",
"middle": [],
"last": "Boot",
"suffix": ""
},
{
"first": "Hanna",
"middle": [],
"last": "Zijlstra",
"suffix": ""
},
{
"first": "Rinie",
"middle": [],
"last": "Geenen",
"suffix": ""
}
],
"year": 2017,
"venue": "Dutch Journal of Applied Linguistics",
"volume": "6",
"issue": "1",
"pages": "65--76",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Peter Boot, Hanna Zijlstra, and Rinie Geenen. 2017. The dutch translation of the linguistic inquiry and word count (liwc) 2007 dictionary. Dutch Journal of Applied Linguistics, 6(1):65-76.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Emotion analysis as a regression problem -dimensional models and their implications on emotion representation and metrical evaluation",
"authors": [
{
"first": "Sven",
"middle": [],
"last": "Buechel",
"suffix": ""
},
{
"first": "Udo",
"middle": [],
"last": "Hahn",
"suffix": ""
}
],
"year": 2016,
"venue": "ECAI",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sven Buechel and Udo Hahn. 2016. Emotion analy- sis as a regression problem -dimensional models and their implications on emotion representation and metrical evaluation. In ECAI.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Word emotion induction for multiple languages as a deep multi-task learning problem",
"authors": [
{
"first": "Sven",
"middle": [],
"last": "Buechel",
"suffix": ""
},
{
"first": "Udo",
"middle": [],
"last": "Hahn",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "1907--1918",
"other_ids": {
"DOI": [
"10.18653/v1/N18-1173"
]
},
"num": null,
"urls": [],
"raw_text": "Sven Buechel and Udo Hahn. 2018. Word emotion in- duction for multiple languages as a deep multi-task learning problem. In Proceedings of the 2018 Con- ference of the North American Chapter of the Asso- ciation for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long Papers), pages 1907-1918, New Orleans, Louisiana. Association for Computational Linguistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Learning and evaluating emotion lexicons for 91 languages",
"authors": [
{
"first": "Sven",
"middle": [],
"last": "Buechel",
"suffix": ""
},
{
"first": "Susanna",
"middle": [],
"last": "R\u00fccker",
"suffix": ""
},
{
"first": "Udo",
"middle": [],
"last": "Hahn",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "1202--1217",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.112"
]
},
"num": null,
"urls": [],
"raw_text": "Sven Buechel, Susanna R\u00fccker, and Udo Hahn. 2020. Learning and evaluating emotion lexicons for 91 lan- guages. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 1202-1217, Online. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Annotating affective dimensions in user-generated content",
"authors": [
{
"first": "Orph\u00e9e",
"middle": [],
"last": "Luna De Bruyne",
"suffix": ""
},
{
"first": "V\u00e9ronique",
"middle": [],
"last": "De Clercq",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hoste",
"suffix": ""
}
],
"year": 2021,
"venue": "Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "1--29",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Luna De Bruyne, Orph\u00e9e De Clercq, and V\u00e9ronique Hoste. 2021. Annotating affective dimensions in user-generated content. Language Resources and Evaluation, pages 1-29.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Towards an integrated pipeline for aspect-based sentiment analysis in various domains",
"authors": [
{
"first": "Els",
"middle": [],
"last": "Orph\u00e9e De Clercq",
"suffix": ""
},
{
"first": "Gilles",
"middle": [],
"last": "Lefever",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Jacobs",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis",
"volume": "",
"issue": "",
"pages": "136--142",
"other_ids": {
"DOI": [
"10.18653/v1/W17-5218"
]
},
"num": null,
"urls": [],
"raw_text": "Orph\u00e9e De Clercq, Els Lefever, Gilles Jacobs, Tijl Carpels, and V\u00e9ronique Hoste. 2017. Towards an in- tegrated pipeline for aspect-based sentiment analysis in various domains. In Proceedings of the 8th Work- shop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 136- 142, Copenhagen, Denmark. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "vreselijk mooi!\"(terribly beautiful): A subjectivity lexicon for dutch adjectives",
"authors": [
{
"first": "Tom",
"middle": [],
"last": "De",
"suffix": ""
},
{
"first": "Smedt",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "Walter",
"middle": [],
"last": "Daelemans",
"suffix": ""
}
],
"year": 2012,
"venue": "LREC",
"volume": "",
"issue": "",
"pages": "3568--3572",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tom De Smedt and Walter Daelemans. 2012. \" vre- selijk mooi!\"(terribly beautiful): A subjectivity lex- icon for dutch adjectives. In LREC, pages 3568- 3572.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Robbert: a dutch roberta-based language model",
"authors": [
{
"first": "Pieter",
"middle": [],
"last": "Delobelle",
"suffix": ""
},
{
"first": "Thomas",
"middle": [],
"last": "Winters",
"suffix": ""
},
{
"first": "Bettina",
"middle": [],
"last": "Berendt",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2001.06286"
]
},
"num": null,
"urls": [],
"raw_text": "Pieter Delobelle, Thomas Winters, and Bettina Berendt. 2020. Robbert: a dutch roberta-based language model. arXiv preprint arXiv:2001.06286.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "4171--4186",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1423"
]
},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Gaussian error linear units (gelus)",
"authors": [
{
"first": "Dan",
"middle": [],
"last": "Hendrycks",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Gimpel",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1606.08415"
]
},
"num": null,
"urls": [],
"raw_text": "Dan Hendrycks and Kevin Gimpel. 2016. Gaus- sian error linear units (gelus). arXiv preprint arXiv:1606.08415.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Generating a non-english subjectivity lexicon: Relations that matter",
"authors": [
{
"first": "Valentin",
"middle": [],
"last": "Jijkoun",
"suffix": ""
},
{
"first": "Katja",
"middle": [],
"last": "Hofmann",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the 12th Conference of the European Chapter",
"volume": "",
"issue": "",
"pages": "398--405",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Valentin Jijkoun and Katja Hofmann. 2009. Generat- ing a non-english subjectivity lexicon: Relations that matter. In Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pages 398-405.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Fine-grained analysis of explicit and implicit sentiment in financial news articles",
"authors": [
{
"first": "Marjan",
"middle": [],
"last": "Van De Kauter",
"suffix": ""
},
{
"first": "Diane",
"middle": [],
"last": "Breesch",
"suffix": ""
},
{
"first": "Veronique",
"middle": [],
"last": "Hoste",
"suffix": ""
}
],
"year": 2015,
"venue": "EXPERT SYSTEMS WITH APPLICATIONS",
"volume": "42",
"issue": "11",
"pages": "4999--5010",
"other_ids": {
"DOI": [
"10.1016/j.eswa.2015.02.007"
]
},
"num": null,
"urls": [],
"raw_text": "Marjan Van de Kauter, Diane Breesch, and Veronique Hoste. 2015. Fine-grained analysis of explicit and implicit sentiment in financial news arti- cles. EXPERT SYSTEMS WITH APPLICATIONS, 42(11):4999-5010.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "HurtBERT: Incorporating lexical features with BERT for the detection of abusive language",
"authors": [
{
"first": "Anna",
"middle": [],
"last": "Koufakou",
"suffix": ""
},
{
"first": "Valerio",
"middle": [],
"last": "Endang Wahyu Pamungkas",
"suffix": ""
},
{
"first": "Viviana",
"middle": [],
"last": "Basile",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Patti",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the Fourth Workshop on Online Abuse and Harms",
"volume": "",
"issue": "",
"pages": "34--43",
"other_ids": {
"DOI": [
"10.18653/v1/2020.alw-1.5"
]
},
"num": null,
"urls": [],
"raw_text": "Anna Koufakou, Endang Wahyu Pamungkas, Valerio Basile, and Viviana Patti. 2020. HurtBERT: Incorpo- rating lexical features with BERT for the detection of abusive language. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 34-43, Online. Association for Computational Linguistics.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Adversarial and auxiliary features-aware bert for sarcasm detection",
"authors": [
{
"first": "Avinash",
"middle": [],
"last": "Kumar",
"suffix": ""
},
{
"first": "Pranjal",
"middle": [],
"last": "Vishnu Teja Narapareddy",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Gupta",
"suffix": ""
}
],
"year": null,
"venue": "Veerubhotla Aditya Srikanth, Lalita Bhanu Murthy Neti, and Aruna Malapati. 2021",
"volume": "",
"issue": "",
"pages": "163--170",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Avinash Kumar, Vishnu Teja Narapareddy, Pran- jal Gupta, Veerubhotla Aditya Srikanth, Lalita Bhanu Murthy Neti, and Aruna Malapati. 2021. Ad- versarial and auxiliary features-aware bert for sar- casm detection. In 8th ACM IKDD CODS and 26th COMAD, CODS COMAD 2021, page 163-170, New York, NY, USA. Association for Computing Machinery.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Interpersonal diagnosis of personality: A functional theory and methodology for personality evaluation",
"authors": [
{
"first": "Timothy",
"middle": [],
"last": "Leary",
"suffix": ""
}
],
"year": 1957,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Timothy Leary. 1957. Interpersonal diagnosis of per- sonality: A functional theory and methodology for personality evaluation. Ronald Press Company.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Roberta: A robustly optimized bert pretraining approach",
"authors": [
{
"first": "Yinhan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Myle",
"middle": [],
"last": "Ott",
"suffix": ""
},
{
"first": "Naman",
"middle": [],
"last": "Goyal",
"suffix": ""
},
{
"first": "Jingfei",
"middle": [],
"last": "Du",
"suffix": ""
},
{
"first": "Mandar",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Danqi",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Lewis",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "Veselin",
"middle": [],
"last": "Stoyanov",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1907.11692"
]
},
"num": null,
"urls": [],
"raw_text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining ap- proach. arXiv preprint arXiv:1907.11692.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Multilingual sentiment analysis: from formal to informal and scarce resource languages",
"authors": [
{
"first": "Erik",
"middle": [],
"last": "Siaw Ling Lo",
"suffix": ""
},
{
"first": "Raymond",
"middle": [],
"last": "Cambria",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Chiong",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Cornforth",
"suffix": ""
}
],
"year": 2017,
"venue": "Artificial Intelligence Review",
"volume": "48",
"issue": "4",
"pages": "499--527",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Siaw Ling Lo, Erik Cambria, Raymond Chiong, and David Cornforth. 2017. Multilingual sentiment analysis: from formal to informal and scarce re- source languages. Artificial Intelligence Review, 48(4):499-527.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "SemEval-2018 task 1: Affect in tweets",
"authors": [
{
"first": "Saif",
"middle": [],
"last": "Mohammad",
"suffix": ""
},
{
"first": "Felipe",
"middle": [],
"last": "Bravo-Marquez",
"suffix": ""
},
{
"first": "Mohammad",
"middle": [],
"last": "Salameh",
"suffix": ""
},
{
"first": "Svetlana",
"middle": [],
"last": "Kiritchenko",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of The 12th International Workshop on Semantic Evaluation",
"volume": "",
"issue": "",
"pages": "1--17",
"other_ids": {
"DOI": [
"10.18653/v1/S18-1001"
]
},
"num": null,
"urls": [],
"raw_text": "Saif Mohammad, Felipe Bravo-Marquez, Mohammad Salameh, and Svetlana Kiritchenko. 2018. SemEval- 2018 task 1: Affect in tweets. In Proceedings of The 12th International Workshop on Semantic Eval- uation, pages 1-17, New Orleans, Louisiana. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Understanding emotions: A dataset of tweets to study interactions between affect categories",
"authors": [
{
"first": "Saif",
"middle": [],
"last": "Mohammad",
"suffix": ""
},
{
"first": "Svetlana",
"middle": [],
"last": "Kiritchenko",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Saif Mohammad and Svetlana Kiritchenko. 2018. Un- derstanding emotions: A dataset of tweets to study interactions between affect categories. In Proceed- ings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources As- sociation (ELRA).",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Obtaining reliable human ratings of valence, arousal, and dominance for 20,000 english words",
"authors": [
{
"first": "M",
"middle": [],
"last": "Saif",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Mohammad",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of The Annual Conference of the Association for Computational Linguistics (ACL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Saif M. Mohammad. 2018. Obtaining reliable hu- man ratings of valence, arousal, and dominance for 20,000 english words. In Proceedings of The An- nual Conference of the Association for Computa- tional Linguistics (ACL), Melbourne, Australia.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Crowdsourcing a word-emotion association lexicon",
"authors": [
{
"first": "M",
"middle": [],
"last": "Saif",
"suffix": ""
},
{
"first": "Peter",
"middle": [
"D"
],
"last": "Mohammad",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Turney",
"suffix": ""
}
],
"year": 2013,
"venue": "Computational Intelligence",
"volume": "29",
"issue": "3",
"pages": "436--465",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Saif M. Mohammad and Peter D. Turney. 2013. Crowd- sourcing a word-emotion association lexicon. Com- putational Intelligence, 29(3):436-465.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Norms of valence, arousal, dominance, and age of acquisition for 4,300 dutch words. Behavior research methods",
"authors": [
{
"first": "Agnes",
"middle": [],
"last": "Moors",
"suffix": ""
},
{
"first": "Jan",
"middle": [
"De"
],
"last": "Houwer",
"suffix": ""
},
{
"first": "Dirk",
"middle": [],
"last": "Hermans",
"suffix": ""
},
{
"first": "Sabine",
"middle": [],
"last": "Wanmaker",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Van Schie",
"suffix": ""
},
{
"first": "Anne-Laura",
"middle": [],
"last": "Van Harmelen",
"suffix": ""
},
{
"first": "Maarten",
"middle": [
"De"
],
"last": "Schryver",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "De Winne",
"suffix": ""
},
{
"first": "Marc",
"middle": [],
"last": "Brysbaert",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "45",
"issue": "",
"pages": "169--177",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Agnes Moors, Jan De Houwer, Dirk Hermans, Sabine Wanmaker, Kevin Van Schie, Anne-Laura Van Harmelen, Maarten De Schryver, Jeffrey De Winne, and Marc Brysbaert. 2013. Norms of valence, arousal, dominance, and age of acquisition for 4,300 dutch words. Behavior research methods, 45(1):169-177.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Creating a dataset for multilingual fine-grained emotion-detection using gamification-based annotation",
"authors": [
{
"first": "Kaisla",
"middle": [],
"last": "Emily\u00f6hman",
"suffix": ""
},
{
"first": "J\u00f6rg",
"middle": [],
"last": "Kajava",
"suffix": ""
},
{
"first": "Timo",
"middle": [],
"last": "Tiedemann",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Honkela",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 9th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis",
"volume": "",
"issue": "",
"pages": "24--30",
"other_ids": {
"DOI": [
"10.18653/v1/W18-6205"
]
},
"num": null,
"urls": [],
"raw_text": "Emily\u00d6hman, Kaisla Kajava, J\u00f6rg Tiedemann, and Timo Honkela. 2018. Creating a dataset for multilingual fine-grained emotion-detection using gamification-based annotation. In Proceedings of the 9th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 24-30, Brussels, Belgium. Association for Computational Linguistics.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "A theory of pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement",
"authors": [
{
"first": "Allan",
"middle": [
"R"
],
"last": "Robert A Rescorla",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wagner",
"suffix": ""
}
],
"year": 1972,
"venue": "Classical conditioning II: Current research and theory",
"volume": "2",
"issue": "",
"pages": "64--99",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Robert A Rescorla, Allan R Wagner, et al. 1972. A the- ory of pavlovian conditioning: Variations in the ef- fectiveness of reinforcement and nonreinforcement. Classical conditioning II: Current research and the- ory, 2:64-99.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Asynchronous pipeline for processing huge corpora on medium to low resource infrastructures",
"authors": [
{
"first": "Pedro Javier Ortiz",
"middle": [],
"last": "Su\u00e1rez",
"suffix": ""
},
{
"first": "Beno\u00eet",
"middle": [],
"last": "Sagot",
"suffix": ""
},
{
"first": "Laurent",
"middle": [],
"last": "Romary",
"suffix": ""
}
],
"year": 2019,
"venue": "7th Workshop on the Challenges in the Management of Large Corpora (CMLC-7). Leibniz-Institut f\u00fcr Deutsche Sprache",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pedro Javier Ortiz Su\u00e1rez, Beno\u00eet Sagot, and Laurent Romary. 2019. Asynchronous pipeline for process- ing huge corpora on medium to low resource infras- tructures. In 7th Workshop on the Challenges in the Management of Large Corpora (CMLC-7). Leibniz- Institut f\u00fcr Deutsche Sprache.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Neural automated essay scoring incorporating handcrafted features",
"authors": [
{
"first": "Masaki",
"middle": [],
"last": "Uto",
"suffix": ""
},
{
"first": "Yikuan",
"middle": [],
"last": "Xie",
"suffix": ""
},
{
"first": "Maomi",
"middle": [],
"last": "Ueno",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 28th International Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Masaki Uto, Yikuan Xie, and Maomi Ueno. 2020. Neural automated essay scoring incorporating hand- crafted features. In Proceedings of the 28th Inter- national Conference on Computational Linguistics, Barcelona, Spain (Online). International Committee on Computational Linguistics.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Automatic emotion classification for interpersonal communication",
"authors": [
{
"first": "Frederik",
"middle": [],
"last": "Vaassen",
"suffix": ""
},
{
"first": "Walter",
"middle": [],
"last": "Daelemans",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the 2nd Workshop on Computational Approaches to Subjectivity and Sentiment Analysis (WASSA 2.011)",
"volume": "",
"issue": "",
"pages": "104--110",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Frederik Vaassen and Walter Daelemans. 2011. Auto- matic emotion classification for interpersonal com- munication. In Proceedings of the 2nd Workshop on Computational Approaches to Subjectivity and Sen- timent Analysis (WASSA 2.011), pages 104-110.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Babelsenticnet: a commonsense reasoning framework for multilingual sentiment analysis",
"authors": [
{
"first": "David",
"middle": [],
"last": "Vilares",
"suffix": ""
},
{
"first": "Haiyun",
"middle": [],
"last": "Peng",
"suffix": ""
},
{
"first": "Ranjan",
"middle": [],
"last": "Satapathy",
"suffix": ""
},
{
"first": "Erik",
"middle": [],
"last": "Cambria",
"suffix": ""
}
],
"year": 2018,
"venue": "2018 IEEE Symposium Series on Computational Intelligence (SSCI)",
"volume": "",
"issue": "",
"pages": "1292--1298",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Vilares, Haiyun Peng, Ranjan Satapathy, and Erik Cambria. 2018. Babelsenticnet: a common- sense reasoning framework for multilingual senti- ment analysis. In 2018 IEEE Symposium Series on Computational Intelligence (SSCI), pages 1292- 1298. IEEE.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Bertje: A dutch bert model",
"authors": [
{
"first": "Andreas",
"middle": [],
"last": "Wietse De Vries",
"suffix": ""
},
{
"first": "Arianna",
"middle": [],
"last": "Van Cranenburgh",
"suffix": ""
},
{
"first": "Tommaso",
"middle": [],
"last": "Bisazza",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Caselli",
"suffix": ""
},
{
"first": "Malvina",
"middle": [],
"last": "Gertjan Van Noord",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Nissim",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1912.09582"
]
},
"num": null,
"urls": [],
"raw_text": "Wietse de Vries, Andreas van Cranenburgh, Arianna Bisazza, Tommaso Caselli, Gertjan van Noord, and Malvina Nissim. 2019. Bertje: A dutch bert model. arXiv preprint arXiv:1912.09582.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Dutch general public reaction on governmental covid-19 measures and announcements in twitter data",
"authors": [
{
"first": "Shihan",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Marijn",
"middle": [],
"last": "Schraagen",
"suffix": ""
},
{
"first": "Erik",
"middle": [
"Tjong"
],
"last": "",
"suffix": ""
},
{
"first": "Kim",
"middle": [],
"last": "Sang",
"suffix": ""
},
{
"first": "Mehdi",
"middle": [],
"last": "Dastani",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2006.07283"
]
},
"num": null,
"urls": [],
"raw_text": "Shihan Wang, Marijn Schraagen, Erik Tjong Kim Sang, and Mehdi Dastani. 2020. Dutch general public reaction on governmental covid-19 measures and announcements in twitter data. arXiv preprint arXiv:2006.07283.",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "Morgan Funtowicz, and Jamie Brew. 2019. HuggingFace's Transformers: State-of-the-art Natural Language Processing",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "Wolf",
"suffix": ""
},
{
"first": "Lysandre",
"middle": [],
"last": "Debut",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Sanh",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Chaumond",
"suffix": ""
},
{
"first": "Clement",
"middle": [],
"last": "Delangue",
"suffix": ""
},
{
"first": "Anthony",
"middle": [],
"last": "Moi",
"suffix": ""
},
{
"first": "Pierric",
"middle": [],
"last": "Cistac",
"suffix": ""
},
{
"first": "Tim",
"middle": [],
"last": "Rault",
"suffix": ""
},
{
"first": "R'emi",
"middle": [],
"last": "Louf",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier- ric Cistac, Tim Rault, R'emi Louf, Morgan Funtow- icz, and Jamie Brew. 2019. HuggingFace's Trans- formers: State-of-the-art Natural Language Process- ing. ArXiv, abs/1910.03771.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"uris": null,
"num": null,
"text": "Transformer model with injection of lexicon features."
},
"FIGREF1": {
"type_str": "figure",
"uris": null,
"num": null,
"text": "Meta-learner (SVM) with predictions from transformer and lexicon values as input features."
},
"TABREF0": {
"num": null,
"html": null,
"type_str": "table",
"content": "<table><tr><td>Corpus</td><td>Text example</td><td/><td>categorical</td><td>dimensional</td></tr><tr><td/><td/><td/><td/><td>V</td><td>A</td><td>D</td></tr><tr><td>Tweets</td><td colspan=\"2\">@transavia Jaaah volgende vakantie Barcelona en na het zomerseizoen</td><td>joy</td><td>0.689 0.491 0.622</td></tr><tr><td/><td>naar de Algarve</td><td/><td/></tr><tr><td/><td>EN: @transavia Yeah</td><td>next holiday</td><td/></tr><tr><td/><td/><td/><td/><td>for Hindi or the</td></tr></table>",
"text": "Barcelona and after summer season to the Algarve Captions Ik zou liever sterven dan hier te wonen, denk ik. sadness 0.156 0.384 0.301 EN: I'd rather die than live here, I think."
},
"TABREF1": {
"num": null,
"html": null,
"type_str": "table",
"content": "<table><tr><td>valence, A = arousal, D = dominance).</td></tr></table>",
"text": "Text examples from the Tweets and Captions subcorpora with their assigned categorical and dimensional label (V ="
},
"TABREF3": {
"num": null,
"html": null,
"type_str": "table",
"content": "<table/>",
"text": "Number of instances in each emotion category per subdataset. A = anger, F = fear, J = joy, L = love, S = sadness, N = neutral."
},
"TABREF4": {
"num": null,
"html": null,
"type_str": "table",
"content": "<table><tr><td>Transformers</td><td colspan=\"2\">Tweets</td><td colspan=\"2\">Captions</td></tr><tr><td>Model</td><td>F1 (SD)</td><td>r (SD)</td><td>F1 (SD)</td><td>r (SD)</td></tr><tr><td>plain BERTje</td><td>.16 (</td><td/><td/><td/></tr></table>",
"text": ".05) .36 (.03) .26 (.04) .60 (.02) BERTje + lex .14 (.02) .35 (.03) .24 (.04) .59 (.02) plain RobBERT .26 (.09) .69 (.03) .22 (.08) .68 (.02) RobBERT + lex .23 (.07) .70 (.01) .25 (.04) .68 (.02)"
}
}
}
} |