File size: 121,197 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T02:09:51.619801Z"
},
"title": "M-Arg: Multimodal Argument Mining Dataset for Political Debates with Audio and Transcripts",
"authors": [
{
"first": "Rafael",
"middle": [],
"last": "Mestre",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Southampton",
"location": {}
},
"email": "r.mestre@soton.ac.uk"
},
{
"first": "Razvan",
"middle": [],
"last": "Milicin",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Southampton",
"location": {}
},
"email": ""
},
{
"first": "Stuart",
"middle": [
"E"
],
"last": "Middleton",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Southampton",
"location": {}
},
"email": ""
},
{
"first": "Matt",
"middle": [],
"last": "Ryan",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Southampton",
"location": {}
},
"email": "m.ryan@soton.ac.uk"
},
{
"first": "Jiatong",
"middle": [],
"last": "Zhu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Southampton",
"location": {}
},
"email": ""
},
{
"first": "Timothy",
"middle": [
"J"
],
"last": "Norman",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Southampton",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Argumentation mining aims at extracting, analysing and modelling people's arguments, but large, high-quality annotated datasets are limited, and no multimodal datasets exist for this task. In this paper, we present M-Arg, a multimodal argument mining dataset with a corpus of US 2020 presidential debates, annotated through crowd-sourced annotations. This dataset allows models to be trained to extract arguments from natural dialogue such as debates using information like the intonation and rhythm of the speaker. Our dataset contains 7 hours of annotated US presidential debates, 6527 utterances and 4104 relation labels, and we report results from different baseline models, namely a text-only model, an audio-only model and multimodal models that extract features from both text and audio. With accuracy reaching 0.86 in multimodal models, we find that audio features provide added value with respect to text-only models.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "Argumentation mining aims at extracting, analysing and modelling people's arguments, but large, high-quality annotated datasets are limited, and no multimodal datasets exist for this task. In this paper, we present M-Arg, a multimodal argument mining dataset with a corpus of US 2020 presidential debates, annotated through crowd-sourced annotations. This dataset allows models to be trained to extract arguments from natural dialogue such as debates using information like the intonation and rhythm of the speaker. Our dataset contains 7 hours of annotated US presidential debates, 6527 utterances and 4104 relation labels, and we report results from different baseline models, namely a text-only model, an audio-only model and multimodal models that extract features from both text and audio. With accuracy reaching 0.86 in multimodal models, we find that audio features provide added value with respect to text-only models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Understanding and modelling argumentation is a recent and key challenge in Natural Language Processing (NLP). Most work addressing this task has focused on extracting arguments from argumentative essays (Stab and Gurevych, 2017) , social networks like Twitter (Bosc et al., 2016) or online reviews (Cocarascu and Toni, 2018) and discussions (Habernal and Gurevych, 2017) , and not much attention has been paid to mining arguments in natural dialogue. The two most common research questions consider how argumentative relations between units (e.g. support or attack) are annotated or how claims and/or premises are identified (Lawrence and Reed, 2019) . We offer, to the best of our knowledge, the first multimodal argumentation mining dataset (M-Arg) of political debates annotated for such argumentative relations of support and attack, using crowd-sourcing techniques. Our contributions are: i) to provide a high quality annotated dataset of political debates with audio and time-stamped transcripts for multimodal argumentation mining; ii) to offer benchmark model results for the research community; and iii) a comparative analysis of the value that multi-modal models bring compared to text-only and audio-only models (Section 5).",
"cite_spans": [
{
"start": 203,
"end": 228,
"text": "(Stab and Gurevych, 2017)",
"ref_id": "BIBREF28"
},
{
"start": 260,
"end": 279,
"text": "(Bosc et al., 2016)",
"ref_id": "BIBREF2"
},
{
"start": 298,
"end": 324,
"text": "(Cocarascu and Toni, 2018)",
"ref_id": "BIBREF8"
},
{
"start": 341,
"end": 370,
"text": "(Habernal and Gurevych, 2017)",
"ref_id": "BIBREF12"
},
{
"start": 625,
"end": 650,
"text": "(Lawrence and Reed, 2019)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The dataset is derived from a collection of US 2020 presidential debates. Five debates were used with the principal speakers being Donald Trump, Joe Biden, Mike Pence and Kamala Harris, and a moderator (Table 1) . In three of the debates the candidates spoke only with each other and the moderator, while in the remaining two they interacted with the audience in so-called Town Hall events. The lengths of the audio files ranged from approximately 1 hour to 1 hour 35 minutes. The debates were tokenised by sentences or utterances, with 6527 in total. The relationship between pairs of sentences were then classified by crowd-workers as support, attack or neither using the annotation scheme proposed by Carstens and Toni (2015) (Section 3) . The crowd-workers were presented with the sentence pair along with a small extract from the debate to provide context. The resulting dataset consists of 4104 pairs of sentences with the argumentative relationship between them classified, along with features such as the trustworthiness of the crowd-workers, the level of agreement between crowd-workers, and their self-confidence scores (Section 4). Prior to giving details of our methodology, the dataset and comparative analysis, we provide a brief review of related research.",
"cite_spans": [
{
"start": 704,
"end": 728,
"text": "Carstens and Toni (2015)",
"ref_id": "BIBREF6"
},
{
"start": 729,
"end": 740,
"text": "(Section 3)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 202,
"end": 211,
"text": "(Table 1)",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Much of the research in argumentation mining has been dedicated to the identification of argumentative discourse units (ADUs) like claims, major claims and premises. For instance, in a first iteration, Stab and Gurevych (2014) Table 1 : Description of the five debates used in the dataset. The column \"Split\" indicates the number of sub-files in which the audio was split.",
"cite_spans": [
{
"start": 202,
"end": 226,
"text": "Stab and Gurevych (2014)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [
{
"start": 227,
"end": 234,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "persuasive essays with 1673 sentences and 1552 argumentative units. Then, they extended their dataset to 402 essays, achieving a total of 7116 sentences and 6089 argumentative components (Stab and Gurevych, 2017) . Carstens and Toni (2015) advocate a relation-based approach towards argumentation mining. Instead of separating the issue of identifying argumentative units and their relation, they reconstitute the task as one of classifying the relationship between sentence pairs as support, attack or neither. They argue that this relation depends upon the context of the discussion. We take the same approach. There are, however, few datasets for relation-based argumentation mining (Paul et al., 2020) . Carstens and Toni (2015) , for example, annotate 854 pairs of sentences for support/attack without identifying the arguments first. Likewise, the DART dataset (Bosc et al., 2016) consists of 4000 tweets, 446 support relations and 112 attack relations, and Stab and Gurevych (2017) annotate 3616 supports and 219 attack relations in their second version of their essay dataset. While certain tasks in argument mining have been applied in other disciplines, interdisciplinary approaches are important for the impact of these methods to be fully realised. Some research in political science has started to bridge the gap in tasks like identifying emotion rhetoric (Osnabr\u00fcge et al., 2021) , gender and emotional expression in politics (Boussalis et al., 2021) , emotional mining in political campaigns (Greco and Polli, 2020) , lexicometrics of Euromanifestos (Jadot and Kelbel, 2017) , and, from the AI perspective, ethos mining (Duthie and Budzynska, 2018) using Hansard as a dataset. Argument mining in political debate is, however, still largely to be explored, although Visser et al. (2021) provide an annotation of 2016 US presidential debates with argument types. Benoit et al. (2016) have advocated for the use of crowd-sourced text analysis for political science, finding high levels of agreement and reproducibility between crowd-workers and experts. However, in subjective tasks like identifying support/attack relations, lower levels of agreement are expected. For instance, Faulkner (2014) used Amazon Mechanical Turk (AMT) to annotate 8176 sentences with \"for\", \"against\" or \"neutral\", achieving about 66% of neutral cases and a Cohen's \u03ba of 0.70. Al-Khatib et al. (2020) also used AMT to obtain 16429 labels of different types, including 1736 \"relation\" labels, defined in their case as \"positive\", \"negative\" or \"no-argument\", with \u03ba = 0.51.",
"cite_spans": [
{
"start": 187,
"end": 212,
"text": "(Stab and Gurevych, 2017)",
"ref_id": "BIBREF28"
},
{
"start": 215,
"end": 239,
"text": "Carstens and Toni (2015)",
"ref_id": "BIBREF6"
},
{
"start": 686,
"end": 705,
"text": "(Paul et al., 2020)",
"ref_id": "BIBREF26"
},
{
"start": 708,
"end": 732,
"text": "Carstens and Toni (2015)",
"ref_id": "BIBREF6"
},
{
"start": 867,
"end": 886,
"text": "(Bosc et al., 2016)",
"ref_id": "BIBREF2"
},
{
"start": 964,
"end": 988,
"text": "Stab and Gurevych (2017)",
"ref_id": "BIBREF28"
},
{
"start": 1369,
"end": 1393,
"text": "(Osnabr\u00fcge et al., 2021)",
"ref_id": "BIBREF25"
},
{
"start": 1440,
"end": 1464,
"text": "(Boussalis et al., 2021)",
"ref_id": "BIBREF3"
},
{
"start": 1507,
"end": 1530,
"text": "(Greco and Polli, 2020)",
"ref_id": "BIBREF11"
},
{
"start": 1565,
"end": 1589,
"text": "(Jadot and Kelbel, 2017)",
"ref_id": "BIBREF13"
},
{
"start": 1635,
"end": 1663,
"text": "(Duthie and Budzynska, 2018)",
"ref_id": "BIBREF9"
},
{
"start": 1780,
"end": 1800,
"text": "Visser et al. (2021)",
"ref_id": "BIBREF30"
},
{
"start": 1876,
"end": 1896,
"text": "Benoit et al. (2016)",
"ref_id": "BIBREF1"
},
{
"start": 2367,
"end": 2390,
"text": "Al-Khatib et al. (2020)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "These datasets focus exclusively on text, and, as far as we can tell, there is not much argument mining research using multiple modalities such as both text and audio, particularly focusing on identifying support or attack relations between ADUs. There are, however, some datasets that could be used by the community for this task. For instance, Mirkin et al. (2019) and Orbach et al. (2020b) provide datasets of debate speeches with transcriptions that could help in the extraction of arguments. Likewise, Mirkin et al. (2020) and Orbach et al. (2020a) come closer to argumentation mining research offering datasets of argumentative content and general-purpose rebuttal in speeches. Also, Kopev et al. (2019) use audio and transcripts of political debates to detect deception. Other research explores emotion recognition or sentiment analysis using the IEMOCAP dataset, which contains text, audio and video with emotion annotations (Busso et al., 2008; Cai et al., 2019) . Classic NLP models for relation classification have relied on bag of words (BoW) approaches with common classifiers like random forests, support vector machines or na\u00efve Bayes (Carstens and Toni, 2017) , although more recently LSTMs and Bi-LSTMs have been used with good results (Cocarascu and Toni, 2018) . Some efforts are being devoted to the use of background knowledge or context. For instance, Paul et al. (2020) proposed Bi-LSTM encoders with self-attention, together with commonsense knowledge extraction. The use of both textual and audio features for the identification of argumentative relations, with approaches similar to those used in multimodal emotion recognition, seems to be mostly unexplored.",
"cite_spans": [
{
"start": 346,
"end": 366,
"text": "Mirkin et al. (2019)",
"ref_id": "BIBREF21"
},
{
"start": 371,
"end": 392,
"text": "Orbach et al. (2020b)",
"ref_id": "BIBREF24"
},
{
"start": 507,
"end": 527,
"text": "Mirkin et al. (2020)",
"ref_id": "BIBREF22"
},
{
"start": 532,
"end": 553,
"text": "Orbach et al. (2020a)",
"ref_id": "BIBREF23"
},
{
"start": 690,
"end": 709,
"text": "Kopev et al. (2019)",
"ref_id": "BIBREF17"
},
{
"start": 933,
"end": 953,
"text": "(Busso et al., 2008;",
"ref_id": "BIBREF4"
},
{
"start": 954,
"end": 971,
"text": "Cai et al., 2019)",
"ref_id": "BIBREF5"
},
{
"start": 1150,
"end": 1175,
"text": "(Carstens and Toni, 2017)",
"ref_id": "BIBREF7"
},
{
"start": 1253,
"end": 1279,
"text": "(Cocarascu and Toni, 2018)",
"ref_id": "BIBREF8"
},
{
"start": 1374,
"end": 1392,
"text": "Paul et al. (2020)",
"ref_id": "BIBREF26"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "Argument mining of political debates can be seen as a long conversation text classification problem where context matters. Unlike the well studied a) b) c) problem area of single and short utterance classification (e.g. 2-3 utterances), dialogue modelling and classification of longer conversations has received little attention to date (Xu et al., 2021) . Recent approaches to handle long sequence classification include augmented transformer models with information retrieval (IR) or summarisation models (Xu et al. 2021; Tigunova et al. 2020) . We constrain ourselves in this paper to providing results for a set of short conversation classification baseline models as we want to focus on showing the value of using multimodal data. However, we expect recent advances in long conversation classification models to yield good results with our dataset in the future.",
"cite_spans": [
{
"start": 337,
"end": 354,
"text": "(Xu et al., 2021)",
"ref_id": "BIBREF31"
},
{
"start": 507,
"end": 523,
"text": "(Xu et al. 2021;",
"ref_id": "BIBREF31"
},
{
"start": 524,
"end": 545,
"text": "Tigunova et al. 2020)",
"ref_id": "BIBREF29"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "The original source of the M-Arg dataset was available as audio tracks with transcripts from a Kaggle competition. 1 This public-domain dataset was originally constructed by downloading audio from YouTube and transcripts from Rev, 2 as explained in the source metadata. The M-Arg dataset with annotations, full transcripts and audio files, source code and model checkpoints for reproducibility is available online in our GitHub repository. 3",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "1 The source materials can be found in https://www.kaggle.com/headsortails/ us-election-2020-presidential-debates as of August 9th, 2021. The version used was v. 7.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "2 https://www.rev.com 3 https://github.com/rafamestre/m-arg_ multimodal-argumentation-dataset",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "The original data was presented as audio in .mp3 files and transcriptions in both .txt and .csv files. The .csv files contained three columns: speaker, minute, and text. Since the timestamps did not align perfectly to the audio clips, we performed our own tokenisation and text-audio alignment. The M-Arg dataset associates each sentence with a matched timestamp in the corresponding debate audio file. To do this, each text was split into utterances, defined as single sentences 4 . Visual inspection revealed the transcriptions to be grammatically correct, with no apparent typos and proper use of punctuation, and so automatic sentence-level tokenization performed well. The utterances were then force-aligned to the audio using the web application of the aeneas tool 5 , obtaining new timestamps. The source audio files were split into different files to comply with the file size limit for the force alignment and to avoid segments where the debate was starting, finishing or going to a break, applause, music, etc. Table 1 summarises the datasets.",
"cite_spans": [],
"ref_spans": [
{
"start": 1021,
"end": 1028,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data overview",
"sec_num": "3.1"
},
{
"text": "Across the five debates, Donald Trump and Joe Biden spoke the most, as can be seen in Figure 1 together, spoke a roughly similar number of utterances. Figure 1 (b), however, indicates that the main participants in the debate did not speak in a similar manner. The violin plots show the probability density of the average number of words per sentence, and we can observe that Trump spoke sentences of a smaller average length than the rest of the participants. Finally, 1(c) shows the most common words (after removing stop-words) in a stacked barplot according to the speaker (removing moderators and audience members), with certain differences in the usage of words. These results indicate potentially important differences in communication strategies and styles.",
"cite_spans": [],
"ref_spans": [
{
"start": 86,
"end": 95,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 152,
"end": 160,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Data overview",
"sec_num": "3.1"
},
{
"text": "The M-Arg dataset consists of 4104 labelled pairs of sentences selected from the debates. Sections of the debates were manually labelled by the authors for their \"topic\", following the explanations of the moderator introducing each section, obtaining high level classifications like \"foreign policy\". Excerpts of 15 sentences were randomly selected (the \"context\") and a pair of sentences within the context were chosen to classify their relation (with their distance weighted by a Gaussian distribution to ensure they were close enough). Approximately 1500 sentences were forced to be from different speakers, to balance the dataset by increasing the possibility of finding attack relations. More details on the pair generation strategy and codes can be found in the repository alongside the dataset 6 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairs creation",
"sec_num": "3.2"
},
{
"text": "The annotation scheme was based on the relationbased argumentation scheme from Carstens and Toni (2015) . They argue that an argumentative relation of support or attack is highly dependent on the context. Carstens and Toni (2015) suggest starting from a root claim to construct pairs or match sentences containing the same entities, but we chose to divide them into topics and weight them by distance as explained in Section 3.2. We presented the crowd-workers with a pair of sentences along with the labelled topic of discussion (e.g. \"families\" or \"climate change\"), as well as a short 15-sentence extract of the dialogue surrounding these sentences as context. The crowd-workers are asked to use this context, as well as any personal knowledge, to classify the argumentative relation as support, attack or neither, to the best of their ability. By not relying only on the surface meaning of the sentences, we open the way for the use of this dataset in more complex scenarios. For instance, it could be applied together with long-or short-text summarisation to take into account the context in a dialogue (Xu et al., 2021) or knowledge-based models linked to databases or fact-checking websites (Paul et al., 2020) . Consider the following pair:",
"cite_spans": [
{
"start": 79,
"end": 103,
"text": "Carstens and Toni (2015)",
"ref_id": "BIBREF6"
},
{
"start": 1108,
"end": 1125,
"text": "(Xu et al., 2021)",
"ref_id": "BIBREF31"
},
{
"start": 1198,
"end": 1217,
"text": "(Paul et al., 2020)",
"ref_id": "BIBREF26"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Annotation scheme",
"sec_num": "3.3"
},
{
"text": "\u2022 Joe Biden: It's criminal.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotation scheme",
"sec_num": "3.3"
},
{
"text": "\u2022 Donald Trump: They are so well taken care of.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotation scheme",
"sec_num": "3.3"
},
{
"text": "At a first glance, it is not possible to know what Biden and Trump are talking about. We might assume that the relationship is attack, but this would be a big assumption only based on the fact that they are opposing candidates. Reading the context, we a) b) c) d) might find out that they are actually talking about the infamous controversy of the U.S. Immigration and Customs Enforcement's (ICE) camps, where children were put in cages and separated from their parents. Joe Biden claims that what the Trump administration has done is criminal. Trump answers by saying that they (the children) are \"so well taken care of\" because reporters went there and saw that the facilities were very clean. The argumentative relation between these two sentences is clearly an attack. Indeed, out of the 81 annotators who classified this test question, 84% of them agreed that this was an attack relation. Work on context summarisation or knowledge extraction could help train models to understand why this was an attack. In our dataset, ADUs consist of sentences fully delimited by periods, but in many cases they will not (Lawrence and Reed, 2019) . They might span several sentences or even be one of the clauses within a sentence. Likewise, during a heated debate, the structure of the argument might not be easily identifiable:",
"cite_spans": [
{
"start": 1112,
"end": 1137,
"text": "(Lawrence and Reed, 2019)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Annotation scheme",
"sec_num": "3.3"
},
{
"text": "\u2022 Joe Biden: We learned that this president paid 50 times the tax in China as a secret bank account with China, does business in China, and in fact, is talking about me taking money?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotation scheme",
"sec_num": "3.3"
},
{
"text": "\u2022 Joe Biden: What are you hiding?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotation scheme",
"sec_num": "3.3"
},
{
"text": "This pair might be interpreted together as a legitimate question being raised to attack the ethos of Donald Trump. However, in and of itself, the second sentence can be taken as a rhetorical way of claiming that Trump is hiding something. In this context, the first sentence is supporting this claim. Annotators did well in this task, with 83% of judgements of this test question correctly labelling it as support.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Annotation scheme",
"sec_num": "3.3"
},
{
"text": "For the annotation task, the platform Appen was used. 7 Crowd-workers were presented with a pair of sentences, topic and context. They were asked to classify the argumentative relation and report their confidence on a Likert scale, ranging from 1: \"not confident at all\", to 5: \"very confident\". Each worker was then paid per \"page\" of work completed, with each page containing between 4 and 6 tasks. To ensure accuracy in the annotations, the contributors were quizzed at the beginning and during the annotations (once per page) with test (or gold) questions. The trust in the annotator was thus defined as the percentage of test questions that they answered correctly, and we set an accuracy threshold of 81%. Other quality settings were enabled, such as: minimum time spent per page to 90 seconds and no more than 60% of supports and 35% of attacks classified. If the annotators did not meet any of these standards, their judgements were not used. Dynamic judgements, that could range from 3 to 7 annotations if the agreement in the annotation was below 70%, were also enabled to improve the agreement of each annotation. A total of 101 test questions were used in this annotation and 104 reliable workers participated, out of 287 that attempted it. Overall, considering the quality settings (e.g. dynamic judgements, tainted answers), 21646 reliable annotations were collected (5746 belonging to gold questions and 15900 to random pairs), and a separate 1663 annotations were rejected. 8",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Crowd-worker annotation",
"sec_num": "3.4"
},
{
"text": "The M-Arg dataset consists of a total of 4104 pairs of sentences (including golden ones), of which 384 are support relations, 120 are attack relations and 3600 are neither support nor attack, as shown in Figure 2(a) . Despite efforts to increase the number of support/attack relations, as explained in Section 3.2, the dataset is imbalanced towards the neither side. This is nevertheless expected, as most of the utterances during a debate are not argumentative in nature. Eighteen different topics were identified in the debates, with the most common ones being \"COVID\", \"Racism\", \"Climate change\" and \"Economy\". Figure 2b shows the total number of utterances from each topic throughout the whole dataset. 9 Some topics, such as \"LGBTQ\" or \"Leadership\" had very few instances, since they were only discussed briefly in one of the debates. Many of these topics, however, could be combined, such as \"Taxes\" and \"Economy\", as desired. For each topic we can see in Figure 2 (c) the distribution of argumentative relations that were annotated. Topics such as \"Foreign Policy\" or \"Taxes\" did not contain attack relations, most likely due to the fact that those sections were small.",
"cite_spans": [],
"ref_spans": [
{
"start": 204,
"end": 215,
"text": "Figure 2(a)",
"ref_id": "FIGREF1"
},
{
"start": 614,
"end": 623,
"text": "Figure 2b",
"ref_id": "FIGREF1"
},
{
"start": 962,
"end": 970,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Description and relevant examples",
"sec_num": "4.1"
},
{
"text": "Whether an argument is supporting or attacking a claim is a subjective matter. Philosophy of argumentation has attempted to establish more or less general argumentation frameworks with different categorisations. However, it is almost certain that thresholds for what quality of information supports or attacks an argument, or judgements on whether such argument is sufficiently valid or not vary by person and context. Our annotated dataset, thus, provides a collective representation of how people reason and understand arguments, and a large number of disagreements are expected. Indeed, the prevalence of fake news or fallacies, or even reasonable disagreements over interpretation of values and inferences in everyday political discourse, has shown us that the same premises can be deemed supports or attacks in different contexts. As we cannot expect that everyone thinks of arguments or fallacies in the same way, the annotation task needs to be accessible and understandable but still closely guided and validated by the theoretical frameworks to reflect informed but real interpre-tations of support and attack in open dialogue in political domains. In the instructions for the annotation task, the contributors were asked to focus on whether a sentence provided a reason that supported or attacked its counterpart, in order to avoid confusing an attack towards something/someone with an attack towards a claim. Consider the following example in which Biden is not providing any reason to support his claim, although it was meant to attack Trump. Many people would interpret this as attack (as many annotators did):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Description and relevant examples",
"sec_num": "4.1"
},
{
"text": "\u2022 Donald Trump: There's abuse, tremendous abuse.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Description and relevant examples",
"sec_num": "4.1"
},
{
"text": "\u2022 Joe Biden: Simply not true.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Description and relevant examples",
"sec_num": "4.1"
},
{
"text": "This case was correctly labeled as neither, but only with an agreement score of 56%. In other cases, support was simply confused with repetition:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Description and relevant examples",
"sec_num": "4.1"
},
{
"text": "\u2022 Joe Biden And what's happening is too many transgender women of color are being murdered.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Description and relevant examples",
"sec_num": "4.1"
},
{
"text": "\u2022 Joe Biden: They're being murdered.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Description and relevant examples",
"sec_num": "4.1"
},
{
"text": "This was annotated as support by three crowdworkers, but this is simply coherence between the sentences or a simple reiteration of a claim. This subjectivity is summarised in Figure 3(a)-(c) , which shows the agreement score, the trust in the annotator and the self-confidence score for each label. In general, crowd-workers labeled relations independently of their trust score and their selfconfidence score. However, attack relations were more controversial with 25% of annotations above 0.87, whereas for support and neither at least 25% of annotations had an agreement of 1. Overall, the average agreement was 0.87 and the median agreement 1.",
"cite_spans": [],
"ref_spans": [
{
"start": 175,
"end": 190,
"text": "Figure 3(a)-(c)",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Description and relevant examples",
"sec_num": "4.1"
},
{
"text": "The presence of subjectivity leads us to evaluate the agreement among crowd-workers (also known as intern-annotator agreement) using Krippendorf's \u03b1 (Krippendorff, 1980) . This agreement score allows for a variable number of annotations in each instance, with an unspecified number of crowdworkers that do not necessarily need to annotate every single instance, making it suitable for our case. Considering all the annotations, we obtained \u03b1 = 0.43. Considering that the distribution of annotations show that some contributors annotated many sentences, while others very few ( Figure 3(d) ), we filtered by the most diligent workers but found no significant change in \u03b1. 10 However, given that annotators are assigned a trust score and they annotated with different self-confidence, we calculated different \u03b1's by filtering by these values. Table 2 shows the \u03b1 scores when we filter by crowd-worker trust rating for all annotations (left) and only for those that were annotated with the maximum confidence (right). We can see that the crowd-worker agreement increases up to 0.53 when we filter annotators with higher trust (although it drops again for the maximum trust), however at the cost of decreasing the number of workers and annotations in the dataset. Likewise, if we only consider those annotations that were provided with 10 Krippendorff's \u03b1 was calculated with the nltk.metrics.agreement module (v. 3.6). Results were doublechecked with the krippendorff module, https://github. com/pln-fing-udelar/fast-krippendorff, which yielded almost identical results. high certainty, we see an overall large value of 0.57, going up to 0.79 if we also filter for high trust workers. Other reports in the literature classifying argumentative relations have yielded Cohen's \u03ba = 0.70 (Faulkner, 2014) , \u03ba = 0.51 (Al-Khatib et al., 2020) , \u03b1 = 0.67 (Bosc et al., 2016) , \u03b1 = 0.81 (Stab and Gurevych, 2014) .",
"cite_spans": [
{
"start": 149,
"end": 169,
"text": "(Krippendorff, 1980)",
"ref_id": "BIBREF18"
},
{
"start": 1333,
"end": 1335,
"text": "10",
"ref_id": null
},
{
"start": 1781,
"end": 1797,
"text": "(Faulkner, 2014)",
"ref_id": "BIBREF10"
},
{
"start": 1809,
"end": 1833,
"text": "(Al-Khatib et al., 2020)",
"ref_id": "BIBREF0"
},
{
"start": 1845,
"end": 1864,
"text": "(Bosc et al., 2016)",
"ref_id": "BIBREF2"
},
{
"start": 1876,
"end": 1901,
"text": "(Stab and Gurevych, 2014)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [
{
"start": 577,
"end": 589,
"text": "Figure 3(d)",
"ref_id": "FIGREF2"
},
{
"start": 842,
"end": 849,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Crowd-worker agreement",
"sec_num": "5.1"
},
{
"text": "It might seem surprising to obtain a low crowdworker agreement, given that the average agreement in the annotation was 85%, as mentioned above. These numbers, however, need to be considered with care. Krippendorff's \u03b1 measures disagreement beyond that expected by chance, but our data is not balanced, so our labels are not equally probable. It has been observed that Krippendorff's \u03b1 can be heavily attenuated in imbalanced datasets (Jeni. et al., 2013) . Indeed, if we sub-sample our dataset for 100 attack, support and neither relations, we find \u03b1 = 0.540 \u00b1 0.015 (standard deviation after 10 trials). If we sub-sample it unbalanced, with 10 attack, 10 support and 1000 neither, we find \u03b1 = 0.170 \u00b1 0.047 (standard deviation after 10 trials). This is a big difference in values, even though the source data is the same. In any case, given the subjectivity of the task, we do not believe a small \u03b1 to be necessarily a bad result, since many judgements might not lend an obvious collective answer and, most importantly, people might believe one instance is a supportive argument, while others believe it is not an argument at all. We believe there is significant value in these unclear annotations, as they give insight into how people understand the arguments put forward in political debate.",
"cite_spans": [
{
"start": 434,
"end": 454,
"text": "(Jeni. et al., 2013)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Crowd-worker agreement",
"sec_num": "5.1"
},
{
"text": "To measure the quality of our corpus and study the potential added value of audio features in argumentation mining, we evaluated the performance of different classification models based on a multimodal model, as well as text-only and audio-only models (Figure 4(a) ). First, the input pair of sentences were split into audio and text. In the multimodal model, each sentence pair was passed through an audio and text module and their outputs concatenated, passed through a 100-unit middle layer and a 3-output classification layer. In the text-only and audio-only models, the sentences were only passed through the text or audio module, respectively, and the middle and classification layers were the same. The audio module shown in Figure 4 (b) was based on a previous model by Cai et al. (2019) for multimodal emotion recognition and consisted of a feature extraction module followed by a CNN in parallel with a Bi-LSTM, chosen to maximise the extraction of local and global features. The textonly module Figure 4 (c) consisted of a BERT preprocessor and a BERT encoding of L=12 hidden layers (i.e., Transformer blocks), a hidden size of H=768, and A=12 attention heads. The missing dropout rates can be found in Table 3 . Audio feature extraction was performed using the Python module \"librosa\" (McFee et al., 2015) . The features were: Mel-frequency cepstral coeffi-cients (MFCCs), which are widely used features for characterising and detecting voice signals (Klapuri and Davy, 2006) ; several spectral features like spectral centroids (Klapuri and Davy, 2006) , spectral bandwidth (Klapuri and Davy, 2006), spectral roll-off (McFee et al., 2015) and spectral contrast (Jiang et al., 2002) ; and a 12-bit chroma vector (McFee et al., 2015) . For each sentence, we used the timestamp to clip the audio file with a buffer of \u00b12 s to ensure the full audio of the utterance was captured.",
"cite_spans": [
{
"start": 778,
"end": 795,
"text": "Cai et al. (2019)",
"ref_id": "BIBREF5"
},
{
"start": 1297,
"end": 1317,
"text": "(McFee et al., 2015)",
"ref_id": "BIBREF20"
},
{
"start": 1463,
"end": 1487,
"text": "(Klapuri and Davy, 2006)",
"ref_id": "BIBREF16"
},
{
"start": 1540,
"end": 1564,
"text": "(Klapuri and Davy, 2006)",
"ref_id": "BIBREF16"
},
{
"start": 1586,
"end": 1598,
"text": "(Klapuri and",
"ref_id": "BIBREF16"
},
{
"start": 1599,
"end": 1650,
"text": "Davy, 2006), spectral roll-off (McFee et al., 2015)",
"ref_id": null
},
{
"start": 1673,
"end": 1693,
"text": "(Jiang et al., 2002)",
"ref_id": "BIBREF15"
},
{
"start": 1723,
"end": 1743,
"text": "(McFee et al., 2015)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [
{
"start": 252,
"end": 264,
"text": "(Figure 4(a)",
"ref_id": "FIGREF3"
},
{
"start": 732,
"end": 740,
"text": "Figure 4",
"ref_id": "FIGREF3"
},
{
"start": 1006,
"end": 1014,
"text": "Figure 4",
"ref_id": "FIGREF3"
},
{
"start": 1214,
"end": 1221,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Argumentative relation classification",
"sec_num": "5.2"
},
{
"text": "To train the three models, we used the Adam optimiser with a learning rate of 0.00005, a batch size of 16 and 50 epochs. A time-based learning rate schedule function was used with a decay rate of 0.0000002 and the loss function was categorical cross-entropy. Table 3 shows the evaluation metrics of these models. The values in \"text dropout\" and \"audio dropout\" are the rates of all the dropout layers of the respective models (Figure 4 ). All three models perform well identifying neither labels, but they do not perform so well identifying attacks or supports, with the highest F 1 values being 0.24 and 0.21, respectively. The text-only model fails to identify attack relations to the same level of the audio-only and multimodal models, most likely due to the imbalance of the data. A multimodal model with a dropout rate of 0.2, however, increases the F 1 for attacks from 0.06 in the text-only model to 0.24, and for supports from 0.14 to 0.21. Surprisingly, the audio-only model performs better than the text-only model in identifying attacks and neither, and closely matches the multimodal model. As proof of concept, we filtered the annotations by their agreement (\u22650.85) and we assessed the best performing multimodal model. We obtained an even higher accuracy value of 0.91, especially for neither labels, with m-F1 of 0.95, although identification of support and attack relations was worse, most likely due to a decrease of useful labels. Over-all, we believe that audio provides relevant features for the identification of argumentative relations and its added value with respect to text helps recapitulate the complexity of this type of data in heavily unbalanced datasets.",
"cite_spans": [],
"ref_spans": [
{
"start": 259,
"end": 266,
"text": "Table 3",
"ref_id": "TABREF4"
},
{
"start": 427,
"end": 436,
"text": "(Figure 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Argumentative relation classification",
"sec_num": "5.2"
},
{
"text": "In this paper, we have presented a multimodal argumentation mining dataset (M-Arg) for political debates based on a corpus of the US 2020 presidential debates with audio and transcripts. The dataset was annotated using crowd-sourcing techniques and we present descriptive statistics of the dataset itself, as well as of the annotations, with discussion of some interesting examples. As a baseline for future research, we evaluated the classification performance of audio-only, text-only and multimodal models. We found that the audio-only and multimodal models could perform with high levels of accuracy and F 1 , although they encountered problems classifying support and attack relations very efficiently. The text-only model performed similarly, but its accuracy in attack classifications was low due to the imbalance of the data. Adding the audio to the text, however, in a multimodal model, helped increase the metrics of both support and attacks, although they still remained quite low. We believe these to be encouraging results, as improvements like reinforcement learning to tackle data imbalance, optimised extraction of audio features, addition of (cross-)attention layers, summarisation of the surrounding context or use of background knowledge databases, could greatly improve these performance metrics. Moreover, the data can be filtered according to annotation agreement, the annotator's trust and self-confidence in the annotation, potentially training models with higher precision and/or recall, although with less data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and future work",
"sec_num": "6"
},
{
"text": "One limitation of our dataset is that ADUs are defined in a very simple manner (by a period with tokeniser). ADUs might be full sentences on certain occasions, but they might encompass several sentences or simply a clause within one. Further work to improve this dataset would include the identification of ADUs (without necessarily labelling them as claim or premise). Likewise, even if a sentence contains a full ADU, in natural dialogue it might not present as a clearly stated claim or premise, but might contain irony or rhetorical questions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and future work",
"sec_num": "6"
},
{
"text": "As already discussed, whether a pair of sentences are showing support or attack is a somewhat sub-jective matter, and for that reason we obtain Krippendorff's \u03b1 = 0.43. One cannot expect crowdworkers to identify, or even easily understand, all types of arguments or what constitutes a fallacy; philosophers continue to disagree. Yet with some instruction and information these annotations can better reflect real-world judgements about support and attack arguments. For certain applications, especially where including marginalised voices, AI systems will need to understand and detect how people argue, even if they do not follow the dictates of argumentation theory (Young, 2000) . We believe our dataset will be of interest for understanding what people think a proper argument is.",
"cite_spans": [
{
"start": 668,
"end": 681,
"text": "(Young, 2000)",
"ref_id": "BIBREF32"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and future work",
"sec_num": "6"
},
{
"text": "Using the sentence tokenizer PunktSentenceTokenizer from www.nltk.org.5 The website of the web application is https:// aeneasweb.org/help and their GitHub https:// github.com/readbeyond/aeneas/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "GitHub: https://github.com/rafamestre/ m-arg_multimodal-argumentation-dataset",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://appen.com/. 8 Extensive details of the annotation scheme and quality settings can be found in the GitHub of the project: https://github.com/rafamestre/m-arg_ multimodal-argumentation-dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Due to the constraints in random pair generation, the topic distribution in the annotated dataset differs slightly, but the distribution closely resembles that of the original dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This work has been funded by the Web Science Institute of the University of Southampton. The authors would also like to acknowledge the support of UK Research and Innovation (UKRI) funding (grant ref MR/S032711/1).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
},
{
"text": "Ethics approval for this research was received from the University of Southampton's Faculty of Social Science Ethics and Research Governance committee, Ref: 66226, Date 22/07/2021. The original dataset in which we build the M-Arg dataset was available under license CC0: Public Domain. Fair treatment of the workers involved in the annotation of the dataset was ensured by Appen's code of ethics (https://appen.com/ crowd-wellness/). We aimed at providing a fair wage for the work provided and, according to the platform statistics, the workers received an hourly compensation with median $7.69 and interquartile mean $7.57. Before sharing our annotated dataset, we have stripped all information that could be potentially sensitive, such as IP's or locations, and we have re-anonymised the anonymous worker ID's that were provided.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Ethical considerations",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "End-to-end argumentation knowledge graph construction",
"authors": [
{
"first": "K",
"middle": [],
"last": "Al-Khatib",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Hou",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Wachsmuth",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Jochim",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Bonin",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Stein",
"suffix": ""
}
],
"year": 2020,
"venue": "Proc. AAAI Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "7367--7374",
"other_ids": {
"DOI": [
"10.1609/aaai.v34i05.6231"
]
},
"num": null,
"urls": [],
"raw_text": "K. Al-Khatib, Y. Hou, H. Wachsmuth, C. Jochim, F. Bonin, and B. Stein. 2020. End-to-end argumen- tation knowledge graph construction. Proc. AAAI Conference on Artificial Intelligence, pages 7367- 7374.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Crowd-sourced text analysis: Reproducible and agile production of political data",
"authors": [
{
"first": "K",
"middle": [],
"last": "Benoit",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Conway",
"suffix": ""
},
{
"first": "B",
"middle": [
"E"
],
"last": "Lauderdale",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Laver",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Mikhaylov",
"suffix": ""
}
],
"year": 2016,
"venue": "American Political Science Review",
"volume": "110",
"issue": "2",
"pages": "278--295",
"other_ids": {
"DOI": [
"10.1017/S0003055416000058"
]
},
"num": null,
"urls": [],
"raw_text": "K. Benoit, D. Conway, B. E. Lauderdale, M. Laver, and S. Mikhaylov. 2016. Crowd-sourced text anal- ysis: Reproducible and agile production of po- litical data. American Political Science Review, 110(2):278-295.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "DART: A dataset of arguments and their relations on twitter",
"authors": [
{
"first": "T",
"middle": [],
"last": "Bosc",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Cabrio",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Villata",
"suffix": ""
}
],
"year": 2016,
"venue": "Proc. 10th Int. Conf. on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "1258--1263",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Bosc, E. Cabrio, and S. Villata. 2016. DART: A dataset of arguments and their relations on twitter. Proc. 10th Int. Conf. on Language Resources and Evaluation, pages 1258-1263.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Gender, candidate emotional expression, and voter reactions during televised debates",
"authors": [
{
"first": "C",
"middle": [],
"last": "Boussalis",
"suffix": ""
},
{
"first": "T",
"middle": [
"G"
],
"last": "Coan",
"suffix": ""
},
{
"first": "M",
"middle": [
"R"
],
"last": "Holman",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "M\u00fcller",
"suffix": ""
}
],
"year": 2021,
"venue": "American Political Science Review",
"volume": "",
"issue": "",
"pages": "1--16",
"other_ids": {
"DOI": [
"10.1017/S0003055421000666"
]
},
"num": null,
"urls": [],
"raw_text": "C. Boussalis, T. G. Coan, M. R. Holman, and S. M\u00fcller. 2021. Gender, candidate emotional expression, and voter reactions during televised debates. American Political Science Review, pages 1-16.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "IEMOCAP: Interactive emotional dyadic motion capture database. Language Resources and Evaluation",
"authors": [
{
"first": "C",
"middle": [],
"last": "Busso",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Bulut",
"suffix": ""
},
{
"first": "C",
"middle": [
"C"
],
"last": "Lee",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Kazemzadeh",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Mower",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "J",
"middle": [
"N"
],
"last": "Chang",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "S",
"middle": [
"S"
],
"last": "Narayanan",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "42",
"issue": "",
"pages": "335--359",
"other_ids": {
"DOI": [
"10.1007/s10579-008-9076-6"
]
},
"num": null,
"urls": [],
"raw_text": "C. Busso, M. Bulut, C. C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan. 2008. IEMOCAP: Interactive emotional dyadic motion capture database. Language Re- sources and Evaluation, 42(4):335-359.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Audiotextual emotion recognition based on improved neural networks",
"authors": [
{
"first": "L",
"middle": [],
"last": "Cai",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Hu",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Dong",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Zhou",
"suffix": ""
}
],
"year": 2019,
"venue": "Mathematical Problems in Engineering",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1155/2019/2593036"
]
},
"num": null,
"urls": [],
"raw_text": "L. Cai, Y. Hu, J. Dong, and S. Zhou. 2019. Audio- textual emotion recognition based on improved neu- ral networks. Mathematical Problems in Engineer- ing, 2019.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Towards relation based argumentation mining",
"authors": [
{
"first": "L",
"middle": [],
"last": "Carstens",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Toni",
"suffix": ""
}
],
"year": 2015,
"venue": "Proc. 2nd Workshop on Argumentation Mining",
"volume": "",
"issue": "",
"pages": "29--34",
"other_ids": {
"DOI": [
"10.3115/v1/w15-0504"
]
},
"num": null,
"urls": [],
"raw_text": "L. Carstens and F. Toni. 2015. Towards relation based argumentation mining. In Proc. 2nd Workshop on Argumentation Mining, pages 29-34.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Using argumentation to improve classification in natural language problems",
"authors": [
{
"first": "L",
"middle": [],
"last": "Carstens",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Toni",
"suffix": ""
}
],
"year": 2017,
"venue": "ACM Trans. on Internet Technology",
"volume": "17",
"issue": "3",
"pages": "1--23",
"other_ids": {
"DOI": [
"10.1145/3017679"
]
},
"num": null,
"urls": [],
"raw_text": "L. Carstens and F. Toni. 2017. Using argumentation to improve classification in natural language problems. ACM Trans. on Internet Technology, 17(3):1-23.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Combining deep learning and argumentative reasoning for the analysis of social media textual content using small data sets",
"authors": [
{
"first": "O",
"middle": [],
"last": "Cocarascu",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Toni",
"suffix": ""
}
],
"year": 2018,
"venue": "Computational Linguistics",
"volume": "44",
"issue": "4",
"pages": "833--858",
"other_ids": {
"DOI": [
"10.1162/colia00338"
]
},
"num": null,
"urls": [],
"raw_text": "O. Cocarascu and F. Toni. 2018. Combining deep learn- ing and argumentative reasoning for the analysis of social media textual content using small data sets. Computational Linguistics, 44(4):833-858.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A deep modular RNN approach for ethos mining",
"authors": [
{
"first": "R",
"middle": [],
"last": "Duthie",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Budzynska",
"suffix": ""
}
],
"year": 2018,
"venue": "Proc. Int. Joint Conf. on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "4041--4047",
"other_ids": {
"DOI": [
"10.24963/ijcai.2018/562"
]
},
"num": null,
"urls": [],
"raw_text": "R. Duthie and K. Budzynska. 2018. A deep modular RNN approach for ethos mining. Proc. Int. Joint Conf. on Artificial Intelligence, pages 4041-4047.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Automated Classification of Argument Stance in Student Essays: A Linguistically Motivated Approach with an Application for Supporting Argument Summarization",
"authors": [
{
"first": "A",
"middle": [
"R"
],
"last": "Faulkner",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. R. Faulkner. 2014. Automated Classification of Argument Stance in Student Essays: A Linguisti- cally Motivated Approach with an Application for Supporting Argument Summarization. Ph.D. thesis, CUNY.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "The political debate on immigration in the election campaigns in Europe",
"authors": [
{
"first": "F",
"middle": [],
"last": "Greco",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Polli",
"suffix": ""
}
],
"year": 2020,
"venue": "Springer Proceedings in Complexity",
"volume": "",
"issue": "",
"pages": "111--123",
"other_ids": {
"DOI": [
"10.1007/978-3-030-48993-9_9"
]
},
"num": null,
"urls": [],
"raw_text": "F. Greco and A. Polli. 2020. The political debate on immigration in the election campaigns in Europe. In Springer Proceedings in Complexity, pages 111-123. Springer, Cham.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Argumentation mining in user-generated Web discourse",
"authors": [
{
"first": "I",
"middle": [],
"last": "Habernal",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Gurevych",
"suffix": ""
}
],
"year": 2017,
"venue": "Cmputational Linguistics",
"volume": "43",
"issue": "1",
"pages": "125--179",
"other_ids": {
"DOI": [
"10.1162/COLI"
]
},
"num": null,
"urls": [],
"raw_text": "I. Habernal and I. Gurevych. 2017. Argumentation mining in user-generated Web discourse. Cmputa- tional Linguistics, 43(1):125-179.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Assessing the politicisation of the European debate using a lexicometric study of the 2014 Euromanifestos",
"authors": [
{
"first": "C",
"middle": [],
"last": "Jadot",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Kelbel",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "55",
"issue": "",
"pages": "60--85",
"other_ids": {
"DOI": [
"10.3917/poeu.055.0060"
]
},
"num": null,
"urls": [],
"raw_text": "C. Jadot and C. Kelbel. 2017. 'Same, same, but dif- ferent.' Assessing the politicisation of the European debate using a lexicometric study of the 2014 Euro- manifestos. Politique Europeenne, 55(1):60-85.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Facing imbalanced data -Recommendations for the use of performance metrics",
"authors": [
{
"first": "L",
"middle": [
"A"
],
"last": "Jeni",
"suffix": ""
},
{
"first": "J",
"middle": [
"F"
],
"last": "Cohn",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "De La Torre",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings -2013 Humaine Association Conference on Affective Computing and Intelligent Interaction, ACII 2013",
"volume": "",
"issue": "",
"pages": "245--251",
"other_ids": {
"DOI": [
"10.1109/ACII.2013.47"
]
},
"num": null,
"urls": [],
"raw_text": "L. A. Jeni., J. F. Cohn, and F. De La Torre. 2013. Fac- ing imbalanced data -Recommendations for the use of performance metrics. Proceedings -2013 Hu- maine Association Conference on Affective Comput- ing and Intelligent Interaction, ACII 2013, pages 245-251.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Music type classification by spectral contrast feature",
"authors": [
{
"first": "D",
"middle": [
"N"
],
"last": "Jiang",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Lu",
"suffix": ""
},
{
"first": "H",
"middle": [
"J"
],
"last": "Zhang",
"suffix": ""
},
{
"first": "J",
"middle": [
"H"
],
"last": "Tao",
"suffix": ""
},
{
"first": "L",
"middle": [
"H"
],
"last": "Cai",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. IEEE Int. Conf. on Multimedia and Expo",
"volume": "",
"issue": "",
"pages": "113--116",
"other_ids": {
"DOI": [
"10.1109/ICME.2002.1035731"
]
},
"num": null,
"urls": [],
"raw_text": "D. N. Jiang, L. Lu, H. J. Zhang, J. H. Tao, and L. H. Cai. 2002. Music type classification by spectral contrast feature. Proc. IEEE Int. Conf. on Multimedia and Expo, pages 113-116.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Signal processing methods for music transcription",
"authors": [
{
"first": "A",
"middle": [],
"last": "Klapuri",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Davy",
"suffix": ""
}
],
"year": 2006,
"venue": "Signal Processing Methods for Music Transcription",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1007/0-387-32845-9"
]
},
"num": null,
"urls": [],
"raw_text": "A. Klapuri and M. Davy. 2006. Signal processing methods for music transcription. In Signal Pro- cessing Methods for Music Transcription, chapter 5. Springer Science & Business Media.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Detecting deception in political debates using acoustic and textual features",
"authors": [
{
"first": "D",
"middle": [],
"last": "Kopev",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Ali",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Koychev",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Nakov",
"suffix": ""
}
],
"year": 2019,
"venue": "Proc. IEEE Automatic Speech Recognition and Understanding Workshop",
"volume": "",
"issue": "",
"pages": "652--659",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Kopev, A. Ali, I. Koychev, and P. Nakov. 2019. De- tecting deception in political debates using acous- tic and textual features. In Proc. IEEE Automatic Speech Recognition and Understanding Workshop, pages 652-659.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Content Analysis: An Introduction to its Methodology",
"authors": [
{
"first": "K",
"middle": [],
"last": "Krippendorff",
"suffix": ""
}
],
"year": 1980,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Krippendorff. 1980. Content Analysis: An Intro- duction to its Methodology. SAGE Publications, Inc, Thousand Oaks, CA.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Argument mining: A survey",
"authors": [
{
"first": "J",
"middle": [],
"last": "Lawrence",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Reed",
"suffix": ""
}
],
"year": 2019,
"venue": "Computational Linguistics",
"volume": "45",
"issue": "4",
"pages": "765--818",
"other_ids": {
"DOI": [
"10.1162/COLIa00364"
]
},
"num": null,
"urls": [],
"raw_text": "J. Lawrence and C. Reed. 2019. Argument mining: A survey. Computational Linguistics, 45(4):765-818.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "librosa: Audio and music signal analysis in Python",
"authors": [
{
"first": "B",
"middle": [],
"last": "Mcfee",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Raffel",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Ellis",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Mcvicar",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Battenberg",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Nieto",
"suffix": ""
}
],
"year": 2015,
"venue": "Proc. 14th Python in Science Conference",
"volume": "",
"issue": "",
"pages": "18--24",
"other_ids": {
"DOI": [
"10.25080/majora-7b98e3ed-003"
]
},
"num": null,
"urls": [],
"raw_text": "B. McFee, C. Raffel, D. Liang, D. Ellis, M. McVicar, E. Battenberg, and O. Nieto. 2015. librosa: Au- dio and music signal analysis in Python. Proc. 14th Python in Science Conference, pages 18-24.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "A recorded debating dataset",
"authors": [
{
"first": "S",
"middle": [],
"last": "Mirkin",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Jacovi",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Lavee",
"suffix": ""
},
{
"first": "H",
"middle": [
"K"
],
"last": "Kuo",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Thomas",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Sager",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Kotlerman",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Venezian",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Slonim",
"suffix": ""
}
],
"year": 2019,
"venue": "LREC 2018 -11th International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "250--254",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Mirkin, M. Jacovi, T. Lavee, H. K. Kuo, S. Thomas, L. Sager, L. Kotlerman, E. Venezian, and N. Slonim. 2019. A recorded debating dataset. In LREC 2018 -11th International Conference on Language Re- sources and Evaluation, pages 250-254.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Listening comprehension over argumentative content",
"authors": [
{
"first": "S",
"middle": [],
"last": "Mirkin",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Moshkowich",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Orbach",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Kotlerman",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Kantor",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Lavee",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Jacovi",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Bilu",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Aharonov",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Slonim",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "719--724",
"other_ids": {
"DOI": [
"10.18653/v1/d18-1078"
]
},
"num": null,
"urls": [],
"raw_text": "S. Mirkin, G. Moshkowich, M. Orbach, L. Kotlerman, Y. Kantor, T. Lavee, M. Jacovi, Y. Bilu, R. Aharonov, and N. Slonim. 2020. Listening comprehension over argumentative content. In Proceedings of the 2018 Conference on Empirical Methods in Natural Lan- guage Processing, EMNLP 2018, pages 719-724.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "A dataset of general-purpose rebuttal",
"authors": [
{
"first": "M",
"middle": [],
"last": "Orbach",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Bilu",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Gera",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Kantor",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Dankin",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Lavee",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Kotlerman",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Mirkin",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Jacovi",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Aharonov",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Slonim",
"suffix": ""
}
],
"year": 2020,
"venue": "EMNLP-IJCNLP 2019 -2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "5591--5601",
"other_ids": {
"DOI": [
"10.18653/v1/d19-1561"
]
},
"num": null,
"urls": [],
"raw_text": "M. Orbach, Y. Bilu, A. Gera, Y. Kantor, L. Dankin, T. Lavee, L. Kotlerman, S. Mirkin, M. Jacovi, R. Aharonov, and N. Slonim. 2020a. A dataset of general-purpose rebuttal. In EMNLP-IJCNLP 2019 -2019 Conference on Empirical Methods in Natu- ral Language Processing and 9th International Joint Conference on Natural Language Processing, Pro- ceedings of the Conference, pages 5591-5601. As- sociation for Computational Linguistics.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Out of the Echo Chamber: Detecting Countering Debate Speeches",
"authors": [
{
"first": "M",
"middle": [],
"last": "Orbach",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Bilu",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Toledo",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Lahav",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Jacovi",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Aharonov",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Slonim",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "7073--7086",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.633"
]
},
"num": null,
"urls": [],
"raw_text": "M. Orbach, Y. Bilu, A. Toledo, D. Lahav, M. Ja- covi, R. Aharonov, and N. Slonim. 2020b. Out of the Echo Chamber: Detecting Countering Debate Speeches. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 7073-7086.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Playing to the gallery: Emotive rhetoric in parliaments",
"authors": [
{
"first": "M",
"middle": [],
"last": "Osnabr\u00fcge",
"suffix": ""
},
{
"first": "S",
"middle": [
"B"
],
"last": "Hobolt",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Rodon",
"suffix": ""
}
],
"year": 2021,
"venue": "American Political Science Review",
"volume": "",
"issue": "",
"pages": "1--15",
"other_ids": {
"DOI": [
"10.1017/s0003055421000356"
]
},
"num": null,
"urls": [],
"raw_text": "M. Osnabr\u00fcge, S. B. Hobolt, and T. Rodon. 2021. Play- ing to the gallery: Emotive rhetoric in parliaments. American Political Science Review, pages 1-15.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Argumentative relation classification with background knowledge",
"authors": [
{
"first": "D",
"middle": [],
"last": "Paul",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Opitz",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Becker",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Kobbe",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Hirst",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Frank",
"suffix": ""
}
],
"year": 2020,
"venue": "Frontiers in Artificial Intelligence and Applications",
"volume": "326",
"issue": "",
"pages": "319--330",
"other_ids": {
"DOI": [
"10.3233/FAIA200515"
]
},
"num": null,
"urls": [],
"raw_text": "D. Paul, J. Opitz, M. Becker, J. Kobbe, G. Hirst, and A. Frank. 2020. Argumentative relation classifica- tion with background knowledge. Frontiers in Arti- ficial Intelligence and Applications, 326:319-330.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Annotating argument components and relations in persuasive essays",
"authors": [
{
"first": "C",
"middle": [],
"last": "Stab",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Gurevych",
"suffix": ""
}
],
"year": 2014,
"venue": "Proc. 25th Int. Conf. on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "1501--1510",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Stab and I. Gurevych. 2014. Annotating argument components and relations in persuasive essays. In Proc. 25th Int. Conf. on Computational Linguistics, pages 1501-1510.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Parsing argumentation structures in persuasive essays",
"authors": [
{
"first": "C",
"middle": [],
"last": "Stab",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Gurevych",
"suffix": ""
}
],
"year": 2017,
"venue": "Computational Linguistics",
"volume": "43",
"issue": "3",
"pages": "619--659",
"other_ids": {
"DOI": [
"10.1162/COLI_a_00295"
]
},
"num": null,
"urls": [],
"raw_text": "C. Stab and I. Gurevych. 2017. Parsing argumentation structures in persuasive essays. Computational Lin- guistics, 43(3):619-659.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "CHARM: Inferring personal attributes from conversations",
"authors": [
{
"first": "A",
"middle": [],
"last": "Tigunova",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Yates",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Mirza",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Weikum",
"suffix": ""
}
],
"year": 2020,
"venue": "Proc. 2020 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "5391--5404",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.434"
]
},
"num": null,
"urls": [],
"raw_text": "A. Tigunova, A. Yates, P. Mirza, and G. Weikum. 2020. CHARM: Inferring personal attributes from conversations. In Proc. 2020 Conference on Empiri- cal Methods in Natural Language Processing, pages 5391-5404.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Annotating argument schemes. Argumentation",
"authors": [
{
"first": "J",
"middle": [],
"last": "Visser",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Lawrence",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Reed",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Wagemans",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Walton",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "35",
"issue": "",
"pages": "101--139",
"other_ids": {
"DOI": [
"10.1007/s10503-020-09519-x"
]
},
"num": null,
"urls": [],
"raw_text": "J. Visser, J. Lawrence, C. Reed, J. Wagemans, and D. Walton. 2021. Annotating argument schemes. Argumentation, 35(1):101-139.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Beyond goldfish memory: Long-term open-domain conversation",
"authors": [
{
"first": "J",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Szlam",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Weston",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2107.07567"
]
},
"num": null,
"urls": [],
"raw_text": "J. Xu, A. Szlam, and J. Weston. 2021. Beyond gold- fish memory: Long-term open-domain conversation. arXiv:2107.07567.",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "Inclusion and Democracy",
"authors": [
{
"first": "I",
"middle": [
"M"
],
"last": "Young",
"suffix": ""
}
],
"year": 2000,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1002/9781119084679.ch19"
]
},
"num": null,
"urls": [],
"raw_text": "I. M. Young. 2000. Inclusion and Democracy. Oxford University Press.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"num": null,
"type_str": "figure",
"text": "Descriptive visualisation of the original dataset. a) Number of utterances (after sentence tokenisation) by each person across all the debates. Audience members were all aggregated in the label \"Audience Members\". b) Average number of words per sentence of the four main participants of the debate, showing their density distribution. c) Number of sentences in which the most common used words appeared in by their speaker."
},
"FIGREF1": {
"uris": null,
"num": null,
"type_str": "figure",
"text": "(a). This is expected, as they were both present in three of the five debates. Mike Pence and Kamala Harris, who only participated in one of the debates Descriptive visualisation of the annotated dataset, M-Arg. a) Number of pairs annotated for support, attack or neither. b) Total number of sentences in the original dataset labeled as one of the topics in the y-axis. c) Percentage of argumentative relations of pairs of sentences belonging to each one of the topics."
},
"FIGREF2": {
"uris": null,
"num": null,
"type_str": "figure",
"text": "Relationship between annotations and confidence parameters. a) Distribution of annotations according to the annotation agreement; b) to the trust in the annotator; and c) to the self-confidence score given by the annotator. d) Distribution of the annotators with respect to the number of annotations they provided."
},
"FIGREF3": {
"uris": null,
"num": null,
"type_str": "figure",
"text": "Schematic of the relation classification models. a) In the multimodal model each sentences is passed in parallel through an audio module and a text module. b) The text module consists only on a BERT encoding layer with dropout. c) The audio module is based on parallel CNN and Bi-LSTM."
},
"TABREF2": {
"num": null,
"type_str": "table",
"content": "<table><tr><td>Multimodal Text-only</td><td>Sentence 1 Sentence 2 Sent. 1 Sent. 2</td><td>Audio module Text module Audio module Text module Text module Text module</td><td>Concatenations Batch Norm. Dense layer 100 ReLU Act. Dropout 0.1</td><td>Classification layer 3 SoftMax Output Output</td><td>Conv layer + Batch Norm. + ReLU Act. + Dropout (8x7x7) kernel b) Audio module Feature extraction</td><td>(1x1) stride</td><td>Max Pool (2x2) stride (2x2) size Bi-LSTM (64x7x7) kernel L2 regularisation (0.5 rate) 128 neurons (1x1) stride</td><td colspan=\"2\">Max Pool Flatten (2x2) size Dropout (2x2) stride</td><td>Concatenate</td></tr><tr><td>Audio-only</td><td>Sent. 1 Sent. 2</td><td>Audio module Audio module</td><td/><td>Output</td><td>Text</td><td/><td>pre-processor</td><td>BERT encoding</td><td colspan=\"2\">Dropout Pooled</td></tr></table>",
"html": null,
"text": "Krippendorff's \u03b1 values for different filterings of the data. Notice the value in bold corresponds to the overall \u03b1 from the whole dataset, since our trust threshold was \u2265 0.81. With decreasing number of annotations, high fluctuations in \u03b1 are to be expected, hence the smaller value of 0.44 for the highest trust."
},
"TABREF4": {
"num": null,
"type_str": "table",
"content": "<table/>",
"html": null,
"text": "Models' performance. M-F 1 stands for macro-averaged F 1 , w-F 1 for weighted-F 1 , and m-F 1 for microaveraged F 1 ."
}
}
}
} |