File size: 88,405 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 | {
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T03:35:40.880795Z"
},
"title": "Exploring Neural Language Models via Analysis of Local and Global Self-Attention Spaces",
"authors": [
{
"first": "Shane",
"middle": [],
"last": "Sheehan",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Nika",
"middle": [],
"last": "Er\u017een",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Marko",
"middle": [],
"last": "Robnik-\u0160ikonja",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Saturnino",
"middle": [],
"last": "Luz",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Senja",
"middle": [],
"last": "Pollak",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Large pretrained language models using the transformer neural network architecture are becoming a dominant methodology for many natural language processing tasks, such as question answering, text classification, word sense disambiguation, text completion and machine translation. Commonly comprising hundreds of millions of parameters, these models offer state-of-the-art performance, but at the expense of interpretability. The attention mechanism is the main component of transformer networks. We present AttViz, a method for exploration of self-attention in transformer networks, which can help in explanation and debugging of the trained models by showing associations between text tokens in an input sequence. We show that existing deep learning pipelines can be explored with AttViz, which offers novel visualizations of the attention heads and their aggregations. We implemented the proposed methods in an online toolkit and an offline library. Using examples from news analysis, we demonstrate how AttViz can be used to inspect and potentially better understand what a model has learned.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "Large pretrained language models using the transformer neural network architecture are becoming a dominant methodology for many natural language processing tasks, such as question answering, text classification, word sense disambiguation, text completion and machine translation. Commonly comprising hundreds of millions of parameters, these models offer state-of-the-art performance, but at the expense of interpretability. The attention mechanism is the main component of transformer networks. We present AttViz, a method for exploration of self-attention in transformer networks, which can help in explanation and debugging of the trained models by showing associations between text tokens in an input sequence. We show that existing deep learning pipelines can be explored with AttViz, which offers novel visualizations of the attention heads and their aggregations. We implemented the proposed methods in an online toolkit and an offline library. Using examples from news analysis, we demonstrate how AttViz can be used to inspect and potentially better understand what a model has learned.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Currently the most successful machine learning approaches for text-related tasks predominantly use large language models. They are implemented with transformer neural network architecture (Vaswani et al., 2017) , extensively pretrained on large text corpora to capture context-dependent meanings of individual tokens (Devlin et al., 2019; Liu et al., 2019; Yang et al., 2019) . Even though training of such neural networks with hundreds of millions of The online visualization (http://attviz.ijs.si; first part of the paper) offers direct exploration of token attention across the space of input documents; its purpose is anomaly detection and general inspection of the attention space (of trained models). The offline part of AttViz (second part of the paper) is a Python library that offers computationally more demanding statistical analyses, ranging from visualization of key tokens for each attention head, comparison of the attention head properties via FUJI integrals, and inspection of the attention distribution per-token basis.",
"cite_spans": [
{
"start": 188,
"end": 210,
"text": "(Vaswani et al., 2017)",
"ref_id": "BIBREF13"
},
{
"start": 317,
"end": 338,
"text": "(Devlin et al., 2019;",
"ref_id": "BIBREF1"
},
{
"start": 339,
"end": 356,
"text": "Liu et al., 2019;",
"ref_id": "BIBREF5"
},
{
"start": 357,
"end": 375,
"text": "Yang et al., 2019)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "parameters is long and expensive (Radford et al., 2019) , many pre-trained models have been made freely available. This has created an opportunity to explore how, and why these models perform well on many tasks. One of the main problems with neural network models is their lack of interpretability. Even though the models learn the given task well, understanding the reasons behind the predictions, and assessing whether the model is susceptible to undue biases or spurious correlations is a non-trivial task. Approaches to understanding black-box (noninterpretable) models include post-hoc perturbation methods, such as IME (\u0160trumbelj and Kononenko, 2010) and SHAP (Lundberg and Lee, 2017) . These methods explain a given decision by assigning a credit to inputs (i.e. attributes or tokens) that contributed to it. These methods are not internal to the model itself and are not well adapted to the sequential nature of text-based inputs. Another way of extracting token relevance is the attention mechanism (Bahdanau et al., 2015; Luong et al., 2015) that learns token pair-value mappings, potentially encoding relations between token pairs. The attention of a token with respect to itself (called self-attention due its position on diagonal of the token attention matrix) offers certain insight into the importance of the token. Typically, a trained transformer network contains several attention heads, each bringing a different focus to the final decision of the network. Exploration of attention can be analytically and numerically cumbersome task, resulting in development of several approaches aimed at attention visualization collection.",
"cite_spans": [
{
"start": 33,
"end": 55,
"text": "(Radford et al., 2019)",
"ref_id": "BIBREF8"
},
{
"start": 625,
"end": 656,
"text": "(\u0160trumbelj and Kononenko, 2010)",
"ref_id": null
},
{
"start": 661,
"end": 690,
"text": "SHAP (Lundberg and Lee, 2017)",
"ref_id": null
},
{
"start": 1008,
"end": 1031,
"text": "(Bahdanau et al., 2015;",
"ref_id": "BIBREF0"
},
{
"start": 1032,
"end": 1051,
"text": "Luong et al., 2015)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "As neural networks require numerical input, words are first transformed into a high dimensional numeric vector space, in a process called embedding that aims to preserve similarities and relations between words. Visualizations of embedding spaces is becoming ubiquitous in contemporary natural language processing. For example, Google's online Embedding Projector 1 offers numerous visualizations for technically non-savvy users, by projecting word vectors to low dimensional (humanunderstandable) spaces. While visualization of embedding spaces is already accessible, visualization of internal workings of complex transformer neural networks (e.g.,their self-attention mechanism) is a challenging task. The works of (Liu et al., 2018) and (Yanagimto et al., 2018) attempt to unveil the workings of black-box attention layers and offer an interface for human researches to learn and inspect their models. Liu et al. (2018) visualize the attention space by coloring it, and Yanagimto et al. (2018) visualize the self-attention with examples from a sentiment analysis.",
"cite_spans": [
{
"start": 717,
"end": 735,
"text": "(Liu et al., 2018)",
"ref_id": "BIBREF4"
},
{
"start": 740,
"end": 764,
"text": "(Yanagimto et al., 2018)",
"ref_id": "BIBREF16"
},
{
"start": 905,
"end": 922,
"text": "Liu et al. (2018)",
"ref_id": "BIBREF4"
},
{
"start": 973,
"end": 996,
"text": "Yanagimto et al. (2018)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this work, we present AttViz, an online system that focuses exclusively on self-attention and introduces two novel ways of visualizing this prop-erty. The tool serves as an additional tool in the toolbox of a language model researcher, offering exploration of the learned models with minimal effort. AttViz can interactively aggregate the attention vectors and offers simultaneous exploration of the output probability space, as well as the attention space. A schematic overview of the proposed work is shown in Figure 1 , and the main contributions are summarised as follows:",
"cite_spans": [],
"ref_spans": [
{
"start": 515,
"end": 523,
"text": "Figure 1",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "1. We present and describe AttViz, an interactive, online toolkit for visualization of the attention space of trained transformer neural language models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "2. We demonstrate the capabilities of AttViz on three problems: news classification, hate speech detection, and insults detection.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "3. AttViz includes a stand-alone python library for offline analysis of the attention space, with the key focus on the relations between the attention heads.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The remainder of the paper is structured as follows. In Section 2, we discuss works related to the proposed AttViz approach. In Section 3, we present the key ideas and technical implementation of the online part of the AttViz system, including a use case on news classification. In Section 4, we discuss the capabilities of the AttViz library, available in an offline mode, and showcase its use on additional two datasets. In Section 5 we discuss capabilities and limitations of AttViz, present conclusions, and propose ideas for future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Neural language models are becoming the prevailing methodology for solving various text-related tasks, from entity recognition to classification. Visualization of the attention mechanism that is the key component of such models has recently emerged as an active research area due to an increased popularity of attention based methods in natural language processing. Recent deep neural network language models such as BERT (Devlin et al., 2019) , XLNet (Yang et al., 2019) , and RoBERTa (Liu et al., 2019) consist of multiple attention heads-separate weight spaces each associated with the input sequence in a unique way. These transformer language models consist of multiple attention matrices, all contributing to the final prediction. Visualising the attention weights from each of the attention matrices is an important component in understanding and interpreting these models.",
"cite_spans": [
{
"start": 422,
"end": 443,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF1"
},
{
"start": 452,
"end": 471,
"text": "(Yang et al., 2019)",
"ref_id": "BIBREF18"
},
{
"start": 486,
"end": 504,
"text": "(Liu et al., 2019)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background and related work on attention visualization",
"sec_num": "2"
},
{
"text": "The attention mechanism, which originated in the neural machine translation, lends itself naturally to visualisation. Bahdanau et al. (2015) used heat maps to display the attention weights between input and output text. This visualisation technique was first applied in machine translation but found its use in many other tasks. Rush et al. (2015) visualized an input sentence and the output abstractive summary, while Rockt\u00e4schel et al. (2016) showed an association between an input document and a textual entailment hypothesis on the output. In these heat map visualisations, a matrix is used to represent the token-token pairs and color intensity illustrates attention weights. This provides a summary of the attention patterns describing how they map the input to the output. For classification tasks, a similar visualisation approach can be used to display the attention weights between the classified document and the predicted label (Yang et al., 2016; Tsaptsinos, 2017) . Here, the visualisation of attention often displays the input document with the attention weights superimposed onto individual words. The superimposed attention weights are represented similarly to heat map visualisations, using the color saturation to encode attention value. The neat-vision tool 2 encodes attention weights associated with input text in this manner. Similarly, the Text Attention Heatmap Visualization (TAHV 3 ) which is included in the NCRF++ toolkit (Yang and Zhang, 2018) can be used to generate weighted sequences which are visualised using superimposed attention scores.",
"cite_spans": [
{
"start": 118,
"end": 140,
"text": "Bahdanau et al. (2015)",
"ref_id": "BIBREF0"
},
{
"start": 329,
"end": 347,
"text": "Rush et al. (2015)",
"ref_id": "BIBREF10"
},
{
"start": 419,
"end": 444,
"text": "Rockt\u00e4schel et al. (2016)",
"ref_id": "BIBREF9"
},
{
"start": 940,
"end": 959,
"text": "(Yang et al., 2016;",
"ref_id": "BIBREF19"
},
{
"start": 960,
"end": 977,
"text": "Tsaptsinos, 2017)",
"ref_id": "BIBREF12"
},
{
"start": 1451,
"end": 1473,
"text": "(Yang and Zhang, 2018)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background and related work on attention visualization",
"sec_num": "2"
},
{
"text": "The purpose of the proposed AttViz is to unveil the attention layer space to human explorers in an intuitive manner. The tool emphasizes selfattention, that is, the diagonal of the token-token attention matrix which possibly corresponds to the relevance of individual tokens. Using different encoding techniques, attention weights across the layers and attention heads can be explored dynamically to investigate the interactions between the model and the input data. The AttViz tool differs from other tools in that if focuses on self-attention, thus allowing visualization of (attention-annotated) input token sequences to be carried out directly.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background and related work on attention visualization",
"sec_num": "2"
},
{
"text": "2 https://github.com/cbaziotis/ neat-vision 3 https://github.com/jiesutd/ Text-Attention-Heatmap-Visualization 3 AttViz: An online toolkit for visualization of self-attention",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background and related work on attention visualization",
"sec_num": "2"
},
{
"text": "AttViz is an online visualization tool that can visualize neural language models from the PyTorchtransformers library 4 -one of the most widely used resources for natural language modeling. The idea behind AttViz is that it is simple to use and lightweight, therefore it does not offer computationally expensive (online) neural model training, but facilitates the exploration of trained models. Along with AttViz, we provide a set of Python scripts that take as an input a trained neural language model and output a JSON file to be used by the AttViz visualisation tool. A common pipeline for using AttViz is outlined in Figure 1 . First, a transformerbased trained neural network model is chosen to obtain predictions on a desired set of instances (documents or some other texts). The predictions are converted into the JSON format suitable for use with the AttViz tool, along with the attention space of the language model. The JSON file is loaded into the AttViz tool (on the user's machine, i.e. on the client side), where its visualization and exploration is possible. In Sections 3.1 and 3.3, we present the proposed self-attention visualizations, followed by an example of their use on the news classification task in Section 3.4.",
"cite_spans": [],
"ref_spans": [
{
"start": 621,
"end": 629,
"text": "Figure 1",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Background and related work on attention visualization",
"sec_num": "2"
},
{
"text": "We discuss the proposed visualization schemes that emphasize different aspects of self-attention. Following the first row that represents the input text, consequent rows correspond to attention values that represent the importance of a given token with respect to a given attention head. As discussed in the empirical part of the paper (Section 3.4), the rationale for this display is that typically only a certain number of attention heads are activated (colored fields). Thus, the visualization has to entail both the whole attention space, as well as emphasize individual heads (and tokens). The initial AttViz view offers sequence-level visualization, where each (byte-pair encoded) token is equipped with a self-attention value based on a given attention head (see Figure 4 ; central text space). The same document can also be viewed in the \"aggregation\" mode ( Figure 2) , where the attention sequence is shown across the token space. The user can interactively explore how the self-attention varies for individ-ual input tokens, by changing the scale, as well as the type of the aggregation. The visualization can emphasize various aspects of the self-attention space.",
"cite_spans": [],
"ref_spans": [
{
"start": 770,
"end": 778,
"text": "Figure 4",
"ref_id": null
},
{
"start": 867,
"end": 876,
"text": "Figure 2)",
"ref_id": null
}
],
"eq_spans": [],
"section": "Visualization of self-attention heads",
"sec_num": "3.1"
},
{
"text": "The third proposed visualization (Figure 3) is the overall distribution of attention values across the whole token space. For each consequent token, the attention values are plotted separately, resembling a time series. This visualization offers an insight into self-attention peaks, i.e. parts of the attention space around certain tokens that potentially impact the performance and decision making process of a given neural network. This view can emphasize different aggregations of the attention vector space for a single token (e.g., mean, entropy, and maximum). The visualization, apart from the mean self-attention (per token), offers the information on maximum and minimum attention values (red dots), as well as the remainder of the self-attention values (gray dots). In this way, a user can explore both the self-attention peaks, as well as the overall spread.",
"cite_spans": [],
"ref_spans": [
{
"start": 33,
"end": 43,
"text": "(Figure 3)",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Visualization of self-attention heads",
"sec_num": "3.1"
},
{
"text": "In the following section, we discuss similarities and differences between AttViz and other state-ofthe-art visualization approaches. Comparisons are summarized in Table 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 163,
"end": 170,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Comparison with state-of-the-art",
"sec_num": "3.2"
},
{
"text": "Novel functionality introduced by AttViz include the capability to aggregate the attention vectors with four different aggregation schemes, offering insights both into the average attention but also its dispersion around a given token. The neat-vision project 5 is the closest to AttViz in terms of functionality. However, a few differences should be noted. First, neat-vision is not directly bound to the PyTorch transformers library, requiring additional pre-processing on the user-side. Second, switching between the sequence and aggregate view is faster and more emphasized in AttViz, as it offers a more general overview of the attention space.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with state-of-the-art",
"sec_num": "3.2"
},
{
"text": "The self-attention is captured in the matrix A \u2208 R h\u00d7t , where h is the number of attention vectors and t the number of tokens. Aggregation operators are applied the second dimension of the attention matrix A (index j). We denote with P ij the probability of observing A ij in the j-th column. The m j 5 Available at https://github.com/cbaziotis/ neat-vision corresponds to the number of unique values in that column. The proposed schemes are summarized in Table 2 . The attention aggregates are visualized as part of the the aggregate view (see Figure 4) . For example, the mean attention is plotted as a line along with the attention space for each token, depicting the dispersion around certain parts of the input text. Figure 2 : Visualization of aggregations. The document was classified as a politics-related topic; the aggregations emphasize tokens such as \"development\",\"uk\" and \"poorer\". The user can highlight desired head information -in this example the maximum attention (purple) is highlighted. ",
"cite_spans": [],
"ref_spans": [
{
"start": 457,
"end": 464,
"text": "Table 2",
"ref_id": "TABREF2"
},
{
"start": 546,
"end": 555,
"text": "Figure 4)",
"ref_id": null
},
{
"start": 723,
"end": 731,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Aggregation of self-attention",
"sec_num": "3.3"
},
{
"text": "In this section, we present a step-by-step use of the AttViz system along with potential insights a user can obtain.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Example: News visualization",
"sec_num": "3.4"
},
{
"text": "The examples are based on the BBC news data set 6 (Greene and Cunningham, 2006) that contains 2,225 news articles on five different topics (business, entertainment, politics, sport, tech). The documents from the dataset were split into short segments. The splits allow easier training (manage- (Wolf et al., 2020) token-to-token attention ",
"cite_spans": [
{
"start": 50,
"end": 79,
"text": "(Greene and Cunningham, 2006)",
"ref_id": "BIBREF2"
},
{
"start": 294,
"end": 313,
"text": "(Wolf et al., 2020)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Example: News visualization",
"sec_num": "3.4"
},
{
"text": "1 h i A ij Entropy(j) (ent) \u2212 1 m j h i=0 A ij log A ij Standard deviation(j) (std) 1 h\u22121 i (A ij \u2212 A ij ) 2 Elementwise Max(j) (max) max i (A ij ) Elementwise Min(j) (min) min i (A ij )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Example: News visualization",
"sec_num": "3.4"
},
{
"text": "Figure 4: Visualization of all attention heads. The sixth heads's self attention is used to highlight the text. The document was classified as a business-related, which can be linked to high self attention at the \"trillion\" and \"uk\" tokens. Compared to the first two examples (Figures 2 and 3) , the network is less certain -in this example, the business (orange) and politics (red) classes were predicted with similar probabilities (orange and red parts of the bar above visualized text).",
"cite_spans": [],
"ref_spans": [
{
"start": 276,
"end": 293,
"text": "(Figures 2 and 3)",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Example: News visualization",
"sec_num": "3.4"
},
{
"text": "able sequence lengths), as well as easier inspection of the models. We split the dataset into 60% of the documents that were used to fine-tune the BERT-base (Devlin et al., 2019) model, 20% for validation and 20% for testing. The Nvidia Tesla V100 GPU processor was used for these experiments. The resulting model classified the whole documents into five categories with 96% accuracy, which is comparable with the state-of-the-art performance (Trieu et al., 2017) . For prediction and visualisation, we used only short segments. The fine-tuning of the BERT model follows examples given in the PyTorch-Transformers library (Wolf et al., 2020) . The best-performing hyper parameter combination used 3 epochs with the sequence length of 512 (other hyper parameters were left at their default values). While we have used BERT, similar explorations could be made for more recent larger models such as XLNet (Yang et al., 2019) that might could produce better classification accuracy.",
"cite_spans": [
{
"start": 157,
"end": 178,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF1"
},
{
"start": 443,
"end": 463,
"text": "(Trieu et al., 2017)",
"ref_id": "BIBREF11"
},
{
"start": 622,
"end": 641,
"text": "(Wolf et al., 2020)",
"ref_id": "BIBREF15"
},
{
"start": 902,
"end": 921,
"text": "(Yang et al., 2019)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Example: News visualization",
"sec_num": "3.4"
},
{
"text": "The user interface of AttViz is displayed in Figures 2, 3, and 4 . In the first example (Figure 3) , the user can observe the main view that consists of two parts. The leftmost part shows (by id) individual self-attention vectors, along with visualization, aggregation and file selection options. The file selection indexes all examples contained in the input (JSON) file. Attention vectors can be colored with custom colors, as shown in the central (tokenvalue view). The user can observe that, for example, the violet attention head (no. 5) is active, and emphasizes tokens such as \"development\", which indicates a politics-related topic (as correctly classified). Here, the token (byte-pair encoded) space is shown along with self-attention values for each token. The attention vectors are shown below the token space and aligned for direct inspection (and correspondence).",
"cite_spans": [],
"ref_spans": [
{
"start": 45,
"end": 64,
"text": "Figures 2, 3, and 4",
"ref_id": "FIGREF2"
},
{
"start": 88,
"end": 98,
"text": "(Figure 3)",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Example: News visualization",
"sec_num": "3.4"
},
{
"text": "In Figure 4 , the user can observe the same text segment as an attention series spanning the input token space. Again, note that tokens, such as \"trillion\" and \"uk\" correspond to high values in a subset of the attention heads, indicating their potential importance for the obtained classification. However, we observed that only a few attention heads activate with respect to individual tokens, indicating that other attention heads are not focusing on the tokens themselves, but possibly on relations between them. This is possible, and the attention matrices contain such information (Vig, 2019) . However, as mentioned earlier, the study of token relations is not the focus of this work. As self-attention in-formation can be mapped across token sequences, emphasizing tokens that are of relevance to the classification task at hand, we see AttViz as being the most useful when exploring models used for text classification tasks, such as hate speech detection and sentiment analysis, where individual tokens contain the key information for classification.",
"cite_spans": [
{
"start": 586,
"end": 597,
"text": "(Vig, 2019)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [
{
"start": 3,
"end": 11,
"text": "Figure 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Example: News visualization",
"sec_num": "3.4"
},
{
"text": "The example above shows how different attention heads detect different aspects of the sentence, even at the single token (self-attention) level. The user can observe that the next most probable category for this topic was politics (red color), which is indeed a more sensible classification than, for instance, sports. The example shows how interpretation of the attention can be coupled with the model's output for increased interpretability.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Example: News visualization",
"sec_num": "3.4"
},
{
"text": "In Section 3 we presented how the online version of AttViz can be used for direct analysis of model output (in the JSON format). Albeit suitable for quick inspections, the online system has its limitations such as poor support for computationally more intensive types of analysis (in terms of waiting times), and the lack of customized visualization tools accessible in the Python ecosystem. To address these aspects, we developed AttViz library that offers more detailed analysis of a given neural language model's properties. The library operates on the same JSON structures as the online version and is compatible with the initial user input. We demonstrate the analytical capabilities of our visualization tools on three datasets. The BBC news classification was already presented in Section 3.4.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "AttViz library: statistical analysis of the attention space",
"sec_num": "4"
},
{
"text": "The first offline functionality is a barplot visualization that offers insight into relevant aspects of the attention distribution at token level. Whilst understanding the attention peaks is relevant for direct inspections (Section 3), the attention space of a given token can be contextualized on the dataset level as well. The AttViz library offers fast visualization of the mean and spread of attention distributions, simultaneously showing the attention peaks for individual tokens. We visualized the distribution for three classification datasets ( Figure 5 ): BBC news (5a), insults 7 (5b), and hate speech comments (5c) 8 . The proposed visualizations present top k tokens according to their mean attention throughout the whole dataset. It is interesting to observe, that the insults and hate speech data sets are not completely characterized by swear words or similar singletoken-like features. This potentially indicates that the attention tries to detect interactions between the byte-pair encoded tokens, even for data sets where the attention could be focused on single tokens. It is interesting to observe that the terms with the highest attention are not necessarily keywords or other tokens carrying large semantic meaning. Similarly, the high maxima indicate that the emphasis of the tokens is very contextual, and potentially not as informative for global aggregation. Contemporary neural language model architectures comprise multiple attention heads. These separate weight spaces capture distinct aspects of the considered learning task. Even though the weight spaces are easily accessible, it is not trivial to convert the large amount of information into a quickto-inspect visualization. With the proposed visualization, shown in Figure 6 , we leverage word clouds (Kaser and Lemire, 2007) to reveal humanunderstandable patterns captured by separate attention heads and display this information in a compact way.",
"cite_spans": [
{
"start": 1786,
"end": 1810,
"text": "(Kaser and Lemire, 2007)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [
{
"start": 554,
"end": 562,
"text": "Figure 5",
"ref_id": "FIGREF4"
},
{
"start": 1751,
"end": 1759,
"text": "Figure 6",
"ref_id": "FIGREF6"
}
],
"eq_spans": [],
"section": "Dissecting the token space",
"sec_num": "4.1"
},
{
"text": "As AttViz is an online and offline toolkit for attention exploration, we discuss possible concerns regarding its use, namely: privacy, memory and performance overheads, and coverage. Privacy is a potential concern for most web-based systems. As currently AttViz does not employ any anonymization strategy, private processing of the input data is not guaranteed. While we intend to address this issue in furture work, a private installation of the tool can be done to get around this current limitation. AttViz uses the users' computing capabilities, which means that large data sets may cause memory overheads when a large number of instances is loaded (typically several million). Such situa-tions are difficult to address with AttViz and similar web-based tool, but users can filter instances before using them in AttViz and explore a subset of the data (e.g., only (in)correctly predicted instances, or certain time slot of instances). Finally, AttViz is focused on the exploration of self-attention. This is not the only important aspect of a transformer neural network, but it is the one, where visualisation techniques have not yet been sufficiently explored. Similarly to the work of (Liu et al., 2018) , we plan to further explore potentially interesting relations emerging from the attention matrices.",
"cite_spans": [
{
"start": 1191,
"end": 1209,
"text": "(Liu et al., 2018)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and conclusions",
"sec_num": "5"
},
{
"text": "The software is available at https://github.com/ SkBlaz/attviz.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Availability",
"sec_num": "6"
},
{
"text": "https://projector.tensorflow.org/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/huggingface/ transformers",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/suraj-deshmukh/ BBC-Dataset-News-Classification/blob/ master/dataset/dataset.csv",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://www.kaggle.com/c/detecting-insults-in-socialcommentary/overview 8 https://github.com/aitor-garcia-p/hate-speech-dataset",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We acknowledge European Union's Horizon 2020 research and innovation programme under grant agreement No 825153, project EMBEDDIA (Cross-Lingual Embeddings). The first author was also funded by Slovenian Research Agency as a young researcher.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Neural machine translation by jointly learning to align and translate",
"authors": [
{
"first": "Dzmitry",
"middle": [],
"last": "Bahdanau",
"suffix": ""
},
{
"first": "Kyunghyun",
"middle": [],
"last": "Cho",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2015,
"venue": "3rd International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- gio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd Inter- national Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "4171--4186",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1423"
]
},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Practical solutions to the problem of diagonal dominance in kernel document clustering",
"authors": [
{
"first": "Derek",
"middle": [],
"last": "Greene",
"suffix": ""
},
{
"first": "Padraig",
"middle": [],
"last": "Cunningham",
"suffix": ""
}
],
"year": 2006,
"venue": "Machine Learning, Proceedings of the Twenty-Third International Conference (ICML 2006)",
"volume": "148",
"issue": "",
"pages": "377--384",
"other_ids": {
"DOI": [
"10.1145/1143844.1143892"
]
},
"num": null,
"urls": [],
"raw_text": "Derek Greene and Padraig Cunningham. 2006. Practi- cal solutions to the problem of diagonal dominance in kernel document clustering. In Machine Learn- ing, Proceedings of the Twenty-Third International Conference (ICML 2006), Pittsburgh, Pennsylvania, USA, June 25-29, 2006, volume 148 of ACM Inter- national Conference Proceeding Series, pages 377- 384. ACM.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Tag-cloud drawing: Algorithms for cloud visualization",
"authors": [
{
"first": "Owen",
"middle": [],
"last": "Kaser",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Lemire",
"suffix": ""
}
],
"year": 2007,
"venue": "Procedings of WWW Workshop on Tagging and Metadata for Social Information Organization",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Owen Kaser and Daniel Lemire. 2007. Tag-cloud drawing: Algorithms for cloud visualization. In Pro- cedings of WWW Workshop on Tagging and Meta- data for Social Information Organization.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Visual interrogation of attention-based models for natural language inference and machine comprehension",
"authors": [
{
"first": "Shusen",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Tao",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Zhimin",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Vivek",
"middle": [],
"last": "Srikumar",
"suffix": ""
},
{
"first": "Valerio",
"middle": [],
"last": "Pascucci",
"suffix": ""
},
{
"first": "Peer-Timo",
"middle": [],
"last": "Bremer",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
"volume": "",
"issue": "",
"pages": "36--41",
"other_ids": {
"DOI": [
"10.18653/v1/D18-2007"
]
},
"num": null,
"urls": [],
"raw_text": "Shusen Liu, Tao Li, Zhimin Li, Vivek Srikumar, Vale- rio Pascucci, and Peer-Timo Bremer. 2018. Visual interrogation of attention-based models for natural language inference and machine comprehension. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 36-41, Brussels, Belgium. Association for Computational Linguistics.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"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
},
"BIBREF6": {
"ref_id": "b6",
"title": "A unified approach to interpreting model predictions",
"authors": [
{
"first": "M",
"middle": [],
"last": "Scott",
"suffix": ""
},
{
"first": "Su-In",
"middle": [],
"last": "Lundberg",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2017,
"venue": "Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "4765--4774",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Scott M. Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions. In Ad- vances in Neural Information Processing Systems 30: Annual Conference on Neural Information Pro- cessing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 4765-4774.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Effective approaches to attention-based neural machine translation",
"authors": [
{
"first": "Thang",
"middle": [],
"last": "Luong",
"suffix": ""
},
{
"first": "Hieu",
"middle": [],
"last": "Pham",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1412--1421",
"other_ids": {
"DOI": [
"10.18653/v1/D15-1166"
]
},
"num": null,
"urls": [],
"raw_text": "Thang Luong, Hieu Pham, and Christopher D. Man- ning. 2015. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natu- ral Language Processing, pages 1412-1421, Lis- bon, Portugal. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Language models are unsupervised multitask learners",
"authors": [
{
"first": "Alec",
"middle": [],
"last": "Radford",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Rewon",
"middle": [],
"last": "Child",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Luan",
"suffix": ""
},
{
"first": "Dario",
"middle": [],
"last": "Amodei",
"suffix": ""
},
{
"first": "Ilya",
"middle": [],
"last": "Sutskever",
"suffix": ""
}
],
"year": 2019,
"venue": "OpenAI Blog",
"volume": "",
"issue": "8",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1(8).",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Reasoning about entailment with neural attention",
"authors": [
{
"first": "Tim",
"middle": [],
"last": "Rockt\u00e4schel",
"suffix": ""
},
{
"first": "Edward",
"middle": [],
"last": "Grefenstette",
"suffix": ""
},
{
"first": "Karl",
"middle": [
"Moritz"
],
"last": "Hermann",
"suffix": ""
},
{
"first": "Tom\u00e1s",
"middle": [],
"last": "Kocisk\u00fd",
"suffix": ""
},
{
"first": "Phil",
"middle": [],
"last": "Blunsom",
"suffix": ""
}
],
"year": 2016,
"venue": "4th International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tim Rockt\u00e4schel, Edward Grefenstette, Karl Moritz Hermann, Tom\u00e1s Kocisk\u00fd, and Phil Blunsom. 2016. Reasoning about entailment with neural attention. In 4th International Conference on Learning Repre- sentations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A neural attention model for abstractive sentence summarization",
"authors": [
{
"first": "Alexander",
"middle": [
"M"
],
"last": "Rush",
"suffix": ""
},
{
"first": "Sumit",
"middle": [],
"last": "Chopra",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Weston",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "379--389",
"other_ids": {
"DOI": [
"10.18653/v1/D15-1044"
]
},
"num": null,
"urls": [],
"raw_text": "Alexander M. Rush, Sumit Chopra, and Jason Weston. 2015. A neural attention model for abstractive sen- tence summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Lan- guage Processing, pages 379-389, Lisbon, Portugal. Association for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "News classification from social media using twitter-based doc2vec model and automatic query expansion",
"authors": [
{
"first": "Q",
"middle": [],
"last": "Lap",
"suffix": ""
},
{
"first": "Huy",
"middle": [
"Q"
],
"last": "Trieu",
"suffix": ""
},
{
"first": "Minh-Triet",
"middle": [],
"last": "Tran",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Tran",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the Eighth International Symposium on Information and Communication Technology",
"volume": "",
"issue": "",
"pages": "460--467",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lap Q. Trieu, Huy Q. Tran, and Minh-Triet Tran. 2017. News classification from social media using twitter-based doc2vec model and automatic query expansion. In Proceedings of the Eighth Interna- tional Symposium on Information and Communica- tion Technology, SoICT 2017, pages 460-467.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Lyrics-based music genre classification using a hierarchical attention network",
"authors": [
{
"first": "Alexandros",
"middle": [],
"last": "Tsaptsinos",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexandros Tsaptsinos. 2017. Lyrics-based music genre classification using a hierarchical attention net- work. CoRR, abs/1707.04678.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Attention is all you need",
"authors": [
{
"first": "Ashish",
"middle": [],
"last": "Vaswani",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Niki",
"middle": [],
"last": "Parmar",
"suffix": ""
},
{
"first": "Jakob",
"middle": [],
"last": "Uszkoreit",
"suffix": ""
},
{
"first": "Llion",
"middle": [],
"last": "Jones",
"suffix": ""
},
{
"first": "Aidan",
"middle": [
"N"
],
"last": "Gomez",
"suffix": ""
},
{
"first": "Lukasz",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "Illia",
"middle": [],
"last": "Polosukhin",
"suffix": ""
}
],
"year": 2017,
"venue": "Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "5998--6008",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Pro- cessing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4- 9, 2017, Long Beach, CA, USA, pages 5998-6008.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Visualizing attention in transformerbased language representation models",
"authors": [
{
"first": "Jesse",
"middle": [],
"last": "Vig",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jesse Vig. 2019. Visualizing attention in transformer- based language representation models. CoRR, abs/1904.02679.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "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\u00e9mi",
"middle": [],
"last": "Louf",
"suffix": ""
},
{
"first": "Morgan",
"middle": [],
"last": "Funtowicz",
"suffix": ""
},
{
"first": "Joe",
"middle": [],
"last": "Davison",
"suffix": ""
},
{
"first": "Sam",
"middle": [],
"last": "Shleifer",
"suffix": ""
},
{
"first": "Clara",
"middle": [],
"last": "Patrick Von Platen",
"suffix": ""
},
{
"first": "Yacine",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Jernite",
"suffix": ""
},
{
"first": "Canwen",
"middle": [],
"last": "Plu",
"suffix": ""
},
{
"first": "Teven",
"middle": [
"Le"
],
"last": "Xu",
"suffix": ""
},
{
"first": "Sylvain",
"middle": [],
"last": "Scao",
"suffix": ""
},
{
"first": "Mariama",
"middle": [],
"last": "Gugger",
"suffix": ""
},
{
"first": "Quentin",
"middle": [],
"last": "Drame",
"suffix": ""
},
{
"first": "Alexander",
"middle": [
"M"
],
"last": "Lhoest",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Rush",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
"volume": "",
"issue": "",
"pages": "38--45",
"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\u00e9mi Louf, Morgan Funtow- icz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language pro- cessing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Attention visualization of gated convolutional neural networks with self attention in sentiment analysis",
"authors": [
{
"first": "H",
"middle": [],
"last": "Yanagimto",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Hashimoto",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Okada",
"suffix": ""
}
],
"year": 2018,
"venue": "2018 International Conference on Machine Learning and Data Engineering (iCMLDE)",
"volume": "",
"issue": "",
"pages": "77--82",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Yanagimto, K. Hashimoto, and M. Okada. 2018. Attention visualization of gated convolutional neu- ral networks with self attention in sentiment analy- sis. In 2018 International Conference on Machine Learning and Data Engineering (iCMLDE), pages 77-82.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "NCRF++: An opensource neural sequence labeling toolkit",
"authors": [
{
"first": "Jie",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Yue",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of ACL 2018, System Demonstrations",
"volume": "",
"issue": "",
"pages": "74--79",
"other_ids": {
"DOI": [
"10.18653/v1/P18-4013"
]
},
"num": null,
"urls": [],
"raw_text": "Jie Yang and Yue Zhang. 2018. NCRF++: An open- source neural sequence labeling toolkit. In Proceed- ings of ACL 2018, System Demonstrations, pages 74-79, Melbourne, Australia. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Xlnet: Generalized autoregressive pretraining for language understanding",
"authors": [
{
"first": "Zhilin",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Zihang",
"middle": [],
"last": "Dai",
"suffix": ""
},
{
"first": "Yiming",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Jaime",
"middle": [],
"last": "Carbonell",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Russ",
"suffix": ""
},
{
"first": "Quoc V",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Le",
"suffix": ""
}
],
"year": 2019,
"venue": "Advances in neural information processing systems",
"volume": "",
"issue": "",
"pages": "5754--5764",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Car- bonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural in- formation processing systems, pages 5754-5764.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Hierarchical attention networks for document classification",
"authors": [
{
"first": "Zichao",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Diyi",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Dyer",
"suffix": ""
},
{
"first": "Xiaodong",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Smola",
"suffix": ""
},
{
"first": "Eduard",
"middle": [],
"last": "Hovy",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "1480--1489",
"other_ids": {
"DOI": [
"10.18653/v1/N16-1174"
]
},
"num": null,
"urls": [],
"raw_text": "Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. 2016. Hierarchical attention networks for document classification. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, pages 1480-1489, San Diego, California. Associa- tion for Computational Linguistics.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "An efficient explanation of individual classifications using game theory",
"authors": [
{
"first": "Igor",
"middle": [],
"last": "Erik\u0161trumbelj",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kononenko",
"suffix": ""
}
],
"year": 2010,
"venue": "Journal of Machine Learning Research",
"volume": "11",
"issue": "",
"pages": "1--18",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Erik\u0160trumbelj and Igor Kononenko. 2010. An ef- ficient explanation of individual classifications us- ing game theory. Journal of Machine Learning Re- search, 11:1-18.",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"uris": null,
"type_str": "figure",
"text": "An overview of AttViz suite. The system consists of two main functional modules supporting online and offline visualizations.",
"num": null
},
"FIGREF2": {
"uris": null,
"type_str": "figure",
"text": "The interactive series view. The user can, by hoovering over the desired part of the sequence, inspect the attention values and their aggregations. The text above the visualization is highlighted automatically.",
"num": null
},
"FIGREF4": {
"uris": null,
"type_str": "figure",
"text": "Visualization of the 35 most attended-to tokens for the three inspected data sets. Interestingly, the attention peaks of tokens (maximum, in the background) all take high values, albeit lower-ranked tokens are on average characterized by lower mean attention values.",
"num": null
},
"FIGREF6": {
"uris": null,
"type_str": "figure",
"text": "The distribution of tokens over individual attention heads for the three datasets summarised with word clouds.",
"num": null
},
"TABREF1": {
"html": null,
"num": null,
"type_str": "table",
"text": "Comparison of different aspects of the attention visualization approaches.",
"content": "<table/>"
},
"TABREF2": {
"html": null,
"num": null,
"type_str": "table",
"text": "Aggregation schemes used in AttViz. The A represents a real valued (attention) matrix.",
"content": "<table><tr><td>Aggregate name</td><td>Definition</td></tr><tr><td>Mean(j) (mean)</td><td/></tr></table>"
}
}
}
} |