File size: 101,135 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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:24:35.708808Z"
},
"title": "Builder, we have done it: Evaluating & Extending Dialogue-AMR NLU Pipeline for Two Collaborative Domains",
"authors": [
{
"first": "Claire",
"middle": [],
"last": "Bonial",
"suffix": "",
"affiliation": {
"laboratory": "Army Research Laboratory",
"institution": "U.S",
"location": {
"postCode": "20783",
"settlement": "Adelphi",
"region": "MD"
}
},
"email": "claire.n.bonial.civ@mail.mil"
},
{
"first": "Mitchell",
"middle": [],
"last": "Abrams",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Institute for Human and Machine Cognition",
"location": {
"postCode": "32502",
"settlement": "Pensacola",
"region": "FL"
}
},
"email": ""
},
{
"first": "David",
"middle": [],
"last": "Traum",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "USC Institute for Creative Technologies",
"location": {
"postCode": "90094",
"settlement": "Playa Vista",
"region": "CA"
}
},
"email": ""
},
{
"first": "Clare",
"middle": [
"R"
],
"last": "Voss",
"suffix": "",
"affiliation": {
"laboratory": "Army Research Laboratory",
"institution": "U.S",
"location": {
"postCode": "20783",
"settlement": "Adelphi",
"region": "MD"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We adopt, evaluate, and improve upon a twostep natural language understanding (NLU) pipeline that incrementally tames the variation of unconstrained natural language input and maps to executable robot behaviors. The pipeline first leverages Abstract Meaning Representation (AMR) parsing to capture the propositional content of the utterance, and second converts this into \"Dialogue-AMR,\" which augments standard AMR with information on tense, aspect, and speech acts. Several alternative approaches and training data sets are evaluated for both steps and corresponding components of the pipeline, some of which outperform the original. We extend the Dialogue-AMR annotation schema to cover a different collaborative instruction domain and evaluate on both domains. With very little training data, we achieve promising performance in the new domain, demonstrating the scalability of this approach.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "We adopt, evaluate, and improve upon a twostep natural language understanding (NLU) pipeline that incrementally tames the variation of unconstrained natural language input and maps to executable robot behaviors. The pipeline first leverages Abstract Meaning Representation (AMR) parsing to capture the propositional content of the utterance, and second converts this into \"Dialogue-AMR,\" which augments standard AMR with information on tense, aspect, and speech acts. Several alternative approaches and training data sets are evaluated for both steps and corresponding components of the pipeline, some of which outperform the original. We extend the Dialogue-AMR annotation schema to cover a different collaborative instruction domain and evaluate on both domains. With very little training data, we achieve promising performance in the new domain, demonstrating the scalability of this approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "We adopt, evaluate, and improve upon the two-step NLU pipeline, described in , which aims to incrementally tame the variation of incoming natural language that the robot must interpret before responding. For each domain in which it operates, the robot must determine whether or not the commands it receives correspond to one of its executable behaviors, such as MOVEMENT (along a front-back axis) and ROTATION. The NLU pipeline leverages AMR to capture the basic content of the input language, and then a conversion system adds behavior time, completion status and speech act information to the original \"Standard-AMR,\" and updates the main action relation of the input AMR to a relation consistently representing an executable robot behavior (see Fig. 1 for a Standard and Dialogue-AMR example comparison). There are two high-level components of the NLU pipeline: a Standard-AMR parser and a graphto-graph conversion system to convert the Standard-AMR into Dialogue-AMR. Here, we offer the first evaluation of both the Dialogue-AMR annotation schema itself and the components of the pipeline used to automatically obtain the Dialogue-AMR. We test not only in the human-robot, search-andnavigation dialogue domain for which the schema and pipeline was developed, but also in a somewhat similar, yet challenging domain: human-human communication collaboratively building structures in the virtual gaming environment, \"Minecraft.\" In this way, we address the question of what would happen if we wanted our robot to collaborate on a new and different task. We refer to this challenge as \"domain extension,\" instead of \"domain adaptation,\" as we aim to maintain the coverage of our original domain while extending to a new one. After providing background on AMR and Dialogue-AMR ( \u00a72) and detailing our approach ( \u00a73), we report on the human-robot evaluation ( \u00a74), followed by the Minecraft evaluation ( \u00a75), and domain extension of the conversion system and subsequent evaluation ( \u00a76). Our contributions include: i. Retraining existing Standard-AMR parsers (3.1) and evaluating on the human-robot (4.1) and Minecraft domains (5.1); ii. Evaluating and improving a conversion system for automatically obtaining Dialogue-AMR (3.2) in both the robot (4.2) and Minecraft (5.2) domains; iii. Extending the coverage of the Dialogue-AMR annotation schema (2.1) to a new domain (6.1) and evaluation after domain extension (6.3).",
"cite_spans": [],
"ref_spans": [
{
"start": 748,
"end": 754,
"text": "Fig. 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To summarize where this work is situated with respect to the past research on this topic-while details the Dialogue-AMR annotation schema and proposes the two-step pipeline as one way of automatically obtaining Dialogue-AMR, the technical details of an implementation of the pipeline itself are not described and no evaluation is given. Subsequent research from does provide an initial evaluation of a baseline version of the graph-to-graph conversion component of the proposed two-step pipeline; we adopt and evaluate an updated version of this component (described in greater detail in \u00a73.2), however, our evaluation is not directly comparable to the evaluation given in , since the earlier version of the component was tested on only a limited subset of the annotation categories of Dialogue-AMR. Thus, the current paper constitutes the first evaluation of the proposed two-step pipeline and its components, as well as an evaluation of the extensibility of those components and the Dialogue-AMR schema itself to a new domain.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "The two-step NLU pipeline of leverages AMR, as it abstracts away from some idiosyncratic surface variation in favor of a more consistent representation for the same concept. This serves the purposes of a dialogue system well: AMR smooths over the nuances of language that may be unimportant for mapping a particular input to one of the robot's behaviors. Nonetheless, \"Standard-AMR\" does not represent some aspects of meaning that are critical for the human-robot dialogue domain, where the robot must be cued as to what the current dialogue state is, as well as what the current time and completion status of various instructions are. To capture this information, the NLU pipeline uses the \"Dialogue-AMR\" formalism , which adds action time, completion status (i.e., limited tense, aspect) and speech act information to the Standard-AMR. Additionally, to facilitate the final step of mapping to one of the robot's behaviors, Dialogue-AMR further generalizes from the input language, converting a variety of surface realizations (e.g., turn, rotate, pivot) of a particular action relation into a single canonical numbered relation (e.g., turn-01) to represent one of the robot's behaviors (e.g., RO-TATION). Standard-AMR and Dialogue-AMR are contrasted in Figs. 1 and 2 .",
"cite_spans": [],
"ref_spans": [
{
"start": 1255,
"end": 1268,
"text": "Figs. 1 and 2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "AMR & Dialogue-AMR",
"sec_num": "2.1"
},
{
"text": "In Dialogue-AMR, the content of the Standard-AMR is nested in a structure that adds the speech act information as the root predicate (e.g., command-SA in Figs. 1, 2) . Additionally, the main action from the Standard-AMR (e.g., move-01) is converted to one of the action relations (e.g., go-02), termed the \"robot-concept relation\" that maps to an executable robot behavior. Information about the time of that behavior is added (in Fig. 2 , the motion event will happen in the future, after the speaking time of the command; thus, it is represented as :time after-now). 1 Finally, the behavior completion status, a type of aspect information, is added-whether or not the instructed behavior is telic or contains a clear end point (in Fig. 2 , indicated by completable +). 2 Dialogue-AMR draws upon an inventory of 13 speech acts and 26 robot behaviors or \"robotconcept relations.\" Action time and completion status are integrated into Dialogue-AMR by adopting the annotation schema of Donatelli et al. (2018) , which categorizes the robot behavior as past, present, or future, and categorizes 4 aspectual labels: :stable +/-, :ongoing +/-, :complete +/-, and :habitual +/-. Dialogue-AMR uses the added category :completable +/-to signal whether or not a hypothetical event has an end-goal achievable for the robot.",
"cite_spans": [
{
"start": 771,
"end": 772,
"text": "2",
"ref_id": null
},
{
"start": 984,
"end": 1007,
"text": "Donatelli et al. (2018)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 154,
"end": 165,
"text": "Figs. 1, 2)",
"ref_id": "FIGREF0"
},
{
"start": 431,
"end": 437,
"text": "Fig. 2",
"ref_id": "FIGREF1"
},
{
"start": 733,
"end": 739,
"text": "Fig. 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "AMR & Dialogue-AMR",
"sec_num": "2.1"
},
{
"text": "We draw from two datasets with Standard-AMR annotations, collected with the aim of developing an interactive agent for collaboration in grounded scenarios. We leverage the DialAMR corpus (Bonial et al., 2020) as training and evaluation data for the NLU pipeline within the human-robot dialogue domain. DialAMR encompasses 1122 instances of The Situated Corpus of Understanding Trans- 1 In ongoing work to extend the Dialogue-AMR schema, we plan to refine the :time annotations to better capture the possibility that an instructed action could already be underway at speaking time, given that we observed that in highly collaborative dialogue, utterances often overlap with actions.",
"cite_spans": [
{
"start": 384,
"end": 385,
"text": "1",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Annotated Corpora",
"sec_num": "2.2"
},
{
"text": "2 End-point information is needed by a robot to execute a behavior in a low-bandwidth environment where there is a communications lag, precluding real-time voice teleoperation. What constitutes a fully specified behavior is somewhat task and robot-specific; for example, a robot with a static, front-facing camera can assume, as a default, that a picture taken for a user will be from this perspective unless the user specifies otherwise, but a robot with a movable, 360-degree view camera may need to ask the user to provide information on the desired camera angle. actions (SCOUT), annotated with both Standard-AMR and Dialogue-AMR. SCOUT is comprised of over 80 hours of dialogues from the robot navigation domain (Marge et al., 2016 (Marge et al., , 2017 , collected via a \"Wizard-of-Oz\" experimental design (Riek, 2012) , in which participants directed what they believed to be an autonomous robot to complete search and navigation tasks. The DialAMR corpus was used in the development of the Dialogue-AMR schema, as well as training and testing of the components of the conversion system of Abrams et al. 2020, which we initially adopt, described in \u00a73.2. The data from SCOUT selected for the Di-alAMR corpus includes a randomly selected, continuous 20-minute experimental trial, which contains 304 utterances (called the Continuous-Trial subset). This is the held-out test set that we use throughout our \"in-domain\" evaluation, as it is representative of an ongoing human-robot interaction.",
"cite_spans": [
{
"start": 717,
"end": 736,
"text": "(Marge et al., 2016",
"ref_id": "BIBREF13"
},
{
"start": 737,
"end": 758,
"text": "(Marge et al., , 2017",
"ref_id": "BIBREF14"
},
{
"start": 812,
"end": 824,
"text": "(Riek, 2012)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Annotated Corpora",
"sec_num": "2.2"
},
{
"text": "In addition to in-domain evaluation, we extend evaluation of the Dialogue-AMR schema and NLU pipeline by annotating and testing on the Minecraft Dialogue Corpus (Narayan-Chen et al., 2019) . This corpus consists of 509 conversations and game logs, in which two humans communicate via the Minecraft gaming interface chat window while collaboratively building blocks structures. Standard-AMR annotations for the Minecraft corpus (Bonn et al., 2020) were obtained from the developers via a private data-sharing agreement. Our addition of Dialogue-AMR annotations to this corpus is described in \u00a76.1.",
"cite_spans": [
{
"start": 161,
"end": 188,
"text": "(Narayan-Chen et al., 2019)",
"ref_id": "BIBREF15"
},
{
"start": 427,
"end": 446,
"text": "(Bonn et al., 2020)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Annotated Corpora",
"sec_num": "2.2"
},
{
"text": "We adopt and evaluate the two-step NLU pipeline described in and Bonial et al. (2019) , including both a Standard-AMR parser and a system for converting this into Dialogue-AMR. We describe our selection of an initial Standard-AMR parser and conversion system, both of which we retrain and improve upon, below.",
"cite_spans": [
{
"start": 65,
"end": 85,
"text": "Bonial et al. (2019)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Approach: Two-Step NLU Pipeline",
"sec_num": "3"
},
{
"text": "Standard-AMR provides an initial interpretation of an utterance to be transferred to the Dialogue-AMR. Therefore, an effective Standard-AMR parser is critical for the overall success of the NLU pipeline.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Standard-AMR Retrained Parser",
"sec_num": "3.1"
},
{
"text": "We considered several open-source AMR parsers as candidates, and selected two recent releases, the parsers described in Zhang et al. (2019) and Lindemann et al. (2019) , which both make use of BERT embeddings (Devlin et al., 2019) and were evaluated on AMR releases, thus providing us with baselines to compare them to each other and to assess our retrained models against their reported performances.",
"cite_spans": [
{
"start": 120,
"end": 139,
"text": "Zhang et al. (2019)",
"ref_id": "BIBREF22"
},
{
"start": 144,
"end": 167,
"text": "Lindemann et al. (2019)",
"ref_id": "BIBREF11"
},
{
"start": 209,
"end": 230,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Standard-AMR Retrained Parser",
"sec_num": "3.1"
},
{
"text": "We were able to retrain both of these state-of-theart AMR parsers on the AMR 2.0 corpus and the recently released AMR 3.0 corpus (a larger corpus including the 2.0 data), and then also retrain them on each of these individual releases of Standard-AMR together with the Standard-AMR subset of the DialAMR corpus of over 800 Standard-AMRs, to adapt them to our human-robot dialogue domain. We evaluated these particular combinations of training data because we wanted to explore whether or not the larger set of data in the AMR 3.0 corpus improved performance on the human-robot dialogue domain, or if it further washed out the distinctions from our smaller in-domain corpus. This yielded a total of eight parsers (see Table 1 ) for us to evaluate and select from for the purpose of then including in the full NLU parsing pipeline.",
"cite_spans": [],
"ref_spans": [
{
"start": 717,
"end": 724,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Standard-AMR Retrained Parser",
"sec_num": "3.1"
},
{
"text": "The next step in the NLU pipeline is a graph-tograph conversion system that uses the input of the utterance text and the Standard-AMR graph to create a Dialogue-AMR graph. We leverage an existing conversion system, \"Abrams+\", and experiment with improvements to how it classifies the robotconcept relation in our own updated graph-to-graph conversion system, \"G2G\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conversion System",
"sec_num": "3.2"
},
{
"text": "We obtained a version of the conversion system described in , which had been updated by that author in two ways: i. expanded to handle the additional speech acts and robotconcept relation categories of the full Dialogue-AMR schema outlined in , not all of which were present during the original development, and ii. shifted from a Na\u00efve Bayes to a SVM model for speech act classification. We refer to this system as \"Abrams+\". This graph-to-graph conversion system implements both rule-based and classifier-based methods in converting a Standard-AMR graph into a Dialogue-AMR graph, and leverages the original utterance and the structure of the Standard-AMR to produce the final Dialogue-AMR, which includes the speech act, tense and aspect information, and a designation of the robotconcept relation. As we use this system as our starting point for improvement, we will briefly describe how each of these additions are made in the order just listed, but refer the reader to Abrams et al. (2020) for full details.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abrams+ Conversion",
"sec_num": "3.2.1"
},
{
"text": "Following the numbering of the example in Fig. 2 , the first step in the transformation process employs a SVM model with token unigrams features to predict the speech act from the original utterance-critical information for human-robot communication that cannot be gleaned from the Standard-AMR graphs alone. 3 After classification, the speech act label is then stored as a slot to be added to the Dialogue-AMR graph and referenced for decision-making processes downstream. Second, to add behavior time, another classifiera Na\u00efve Bayes model using token unigrams as features-determines if the event corresponding to the robot behavior pertains to a past, present, or future action. Third, designation of the robot behavior is implemented through a keyword-based approach, which extracts the top root relation (keyword) in the Standard-AMR and checks it against a keyword dictionary of similar actions, and maps it to a single robot-concept relation. Fourth, particu- 3 We acknowledge that the interpretation of speech acts, and indirect speech acts in particular, can be affected by context. Following (Hinkelman and Allen, 1989) , we start with only the linguistic signal in the first phase. Since the restricted domain is predictable, it is usually sufficient, but further research aims to leverage situational information and dialogue context where necessary, e.g., to disambiguate an ability question from an indirect instruction. lar combinations of speech act, tense, and the presence or absence of certain arguments of the robotconcept relation trigger an aspectual label that corresponds to an action's completion status. In the final step of transformation process, the system's rule-based methods use pattern matching techniques to serve multiple functions, including slot filling and slot changing (e.g., transforming mentions of you to the fixed role of the addressee in Dialogue-AMR).",
"cite_spans": [
{
"start": 967,
"end": 968,
"text": "3",
"ref_id": null
},
{
"start": 1102,
"end": 1129,
"text": "(Hinkelman and Allen, 1989)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 42,
"end": 48,
"text": "Fig. 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Abrams+ Conversion",
"sec_num": "3.2.1"
},
{
"text": "While we hypothesize speech act, tense, and aspect classification may be fairly robust to language in a new domain, we readily acknowledge that new domains will require the robot to engage in novel behaviors, for example, BUILDING in the Minecraft domain. Thus, although there are many different aspects of the conversion system that we could attempt to improve upon (e.g., classifier types, ordering of components), we saw an opportunity to have the most impact on system performance in multiple domains by focusing on varying the robotconcept relation classification approach. We describe three variants (one keyword-based and two classifier-based) of our updated G2G conversion system below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "G2G: Our Updated Conversion System",
"sec_num": "3.2.2"
},
{
"text": "G2G Expanded Keyword-Based Variant We expanded upon the keyword approach of the Abrams+ system, which was restricted to searching for keyword matches with the top, root relation of the Standard-AMR. We found that this restriction was problematic because the same root relation in the Standard-AMR could correspond to multiple robot-concept relations. Move and go, generally parsed as move-01 and go-02, are particularly prevalent and could correspond to either front-back MOVEMENT or a ROTATION behavior; both of these were keywords triggering front-back movement in Abrams+, which therefore incorrectly categorized utterances like Move right 45 degrees (a ROTATION behavior). In our expansion, the G2G keyword variant searches for matches within all utterance tokens, AMR relations, and arguments. Furthermore, the keyword dictionary was informed by a data-driven analysis in which we created histograms of all utterance tokens and Standard-AMR relations within an instance mapped to a particular robot-concept relation in the manual Dialogue-AMR annotations. In this way, we could see which words and relations occurred with multiple robotconcept relations, like move-01, and therefore remove these from our keyword dictionaries, while adding keywords that are unique to a particular robot-concept relation in the data, such as degrees, which consistently cues a ROTATION behavior.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "G2G: Our Updated Conversion System",
"sec_num": "3.2.2"
},
{
"text": "G2G One-Hot and GloVe Variants We also experimented with classifier-based approaches to robot-behavior classification, which we hypothesized may be more efficient to extend to a new domain than a keyword-based approach. The classifiers are Support Vector Machines with different vectorization methods including one-hot encoding and word embeddings from GloVe. Training data for the robot-concept relation classifier comes from examples of each robot-concept category in Bonial et al. 2020, gold-standard labels from the Continuous-Trial subset utterances 101-305 (those not used in a held-out test set), and examples pulled from speech act classifier training bins. There are a total of 26 labels for this task, and while many of the movement actions were abundant from these other sources, some of the minority labels (e.g., equip-01, wait-01, clarify-10) required up-sampling to balance training proportions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "G2G: Our Updated Conversion System",
"sec_num": "3.2.2"
},
{
"text": "We evaluated the retrained parsers on the SCOUT Continuous-trial dataset. We note substantial improvement in Standard-AMR parsing Smatch scores on this set when training with DialAMR in addition to the base training sets (AMR 2.0 and 3.0). 4 Results for the AMR parsing models are presented in Table 1 . The noticeably high scores on the parsers retrained on the AMR 3.0 + DialAMR is due in large part to the nature of the speakers' language in the SCOUT corpus and the high levels of similarity in participants' instructions to the robot. This underscores how critical evaluation in another dialogue domain is. We note that, at the segment level as well as can be seen in the Table 1, the Lindemann et al. (2019) parser retrained with DialAMR data evaluated across-the-board to higher scores than the comparably retrained Zhang et al. (2019) parser. Of those two Lindemann et al. (2019) parsers whose Smatch scores did not differ significantly, we selected the one trained with the larger 3.0 dataset with its larger language model as the first component in the full parsing pipeline. ",
"cite_spans": [
{
"start": 823,
"end": 842,
"text": "Zhang et al. (2019)",
"ref_id": "BIBREF22"
}
],
"ref_spans": [
{
"start": 294,
"end": 301,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "In-Domain Standard-AMR Parsing",
"sec_num": "4.1"
},
{
"text": "To pinpoint the performance of the conversion system alone (without error introduced by the automatic Standard-AMR parsing), we report results with gold-standard, manually assigned input Standard-AMR parses. Results are summarized in Evaluation Domain A of Table 2 . Focusing initially on the overall Smatch Precision, Recall, and F-scores of the conversion system, our updated system, G2G, leveraging the classifier with one-hot vectorization achieves the highest precision (.85) and F-score (.83) in our domain. All approaches perform comparably overall, especially given that Smatch scores can vary slightly (Opitz et al., 2020) because Smatch is a non-deterministic, greedy hillclimbing algorithm with a preset, default number of random restarts (Cai and Knight, 2013) .",
"cite_spans": [
{
"start": 611,
"end": 631,
"text": "(Opitz et al., 2020)",
"ref_id": "BIBREF16"
},
{
"start": 750,
"end": 772,
"text": "(Cai and Knight, 2013)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [
{
"start": 257,
"end": 264,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "In-Domain Conversion to Dialogue-AMR",
"sec_num": "4.2"
},
{
"text": "Drilling down into the accuracy of the individual component classification tasks, we find accuracy scores of 1.00 for speech acts, .93 for tense, and .93 for aspect across all system variants, as these components are unchanged, and we only alter the robot-concept classification. Again, we note that these accuracy scores are extremely high, given the repetitive nature of the language and prevalence of certain types of commands and feedback assertions. For robot-concept classification, the G2G expanded keyword approach (.97 accuracy) does outperform the Abrams+ baseline keyword method (.94 accuracy). Both keyword approaches outperform the G2G classifier-based approaches: one-hot vectorization achieves an accuracy of .90 and GloVe an accuracy of .84. Notably, higher accuracy on the robot-concept classification task does not necessarily translate to higher Smatch Fscores overall. High component accuracy but lower overall F-Score generally indicates that while the system is correctly determining all of the information being added to the Dialogue-AMR, it is not always putting these pieces together correctly. In other words, the final step in the conversion system, where slots are captured and changed from the original Standard-AMR structure to the structure of the Dialogue-AMR, is where some of the error reflected in Smatch scores stems from.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "In-Domain Conversion to Dialogue-AMR",
"sec_num": "4.2"
},
{
"text": "In this section, we report on the Minecraft domain performance of the NLU pipeline with the retrained Standard-AMR parser, the Abrams+ conversion system, and our updated G2G system variants prior to any domain adaptation in order to determine how vital domain extension really is in somewhat similar instruction-giving domains. Given that theoretically speech acts, tense and aspect are somewhat consistent in language regardless of the domain, we hypothesize that these features of our annotation schema and the components of the conversion system capturing them will perform reasonably well on the new Minecraft dialogue domain. However, the main actions or behaviors involved in the collaboration of interlocutors in the original search and navigation domain are quite different from those of building virtual structures from blocks in the new Minecraft domain. We therefore expect that the conversion system will fail to correctly map many of the main action predicates in the Minecraft dialogues to an executable robot behavior. However, we accept this as an interesting question of domain extension for moving our robot to a new task: Is it more efficient to expand a rule-based approach for capturing these new behaviors, or to use a classifierbased approach?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Minecraft Domain Evaluation",
"sec_num": "5"
},
{
"text": "We test the parser selected as the first pipeline component (described in \u00a74.1) on Minecraft data, scor-ing the parser output on 100 sequential instances of Minecraft dialogue against manually assigned Standard-AMR annotations. 5 The overall Smatch F-score is .57, with a Precision of .63 and Recall of .52. Thus, despite the potential similarity in the two instruction-giving dialogue domains, it is clear that the automatic parsing performance is significantly worse for the Minecraft data than our original domain (where the best Smatch F-score was .93). Error analysis reveals some extremely complicated language phenomena, including dimensions and frequency expressions capturing, for example, the repetition of a placement action: For the four squares that come out from the middle blocks, add two blue blocks on. Although this indicates that the parser would benefit from retraining with Minecraft data, 6 in our immediate research we focus on domain extension of the conversion system in order to explore how robust the conversion system might be to noise in the parser input.",
"cite_spans": [
{
"start": 228,
"end": 229,
"text": "5",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Minecraft Standard-AMR Parsing",
"sec_num": "5.1"
},
{
"text": "This evaluation compares the conversion system output against manually assigned Dialogue-AMRs for the same 100-instance, sequential subset of utterances from the Minecraft corpus used as the test set for the Standard-AMR parser (see \u00a76.1 for Dialogue-AMR annotation details); again, we use gold-standard, manually assigned Standard-AMR parses as input to the conversion system. Results are summarized in Evaluation Domain B of Ta-ble 2. Focusing first on overall Smatch scores, our updated system variant leveraging the expanded keyword approach performs slightly better (.68 Fscore) than both the baseline Abrams+ (.67 F-score) and the classifier-based approaches (.67 F-scores).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Minecraft Conversion to Dialogue-AMR",
"sec_num": "5.2"
},
{
"text": "Although the scores have dropped about 15 points from the original domain, they remain comparable across variants. When drilling down into the accuracy of the individual components of the conversion system, we find that robot concept classification yields the lowest accuracy scores, with a range of .20-.32. Among the variant approaches to robot-concept classification explored, the expanded keyword approach achieves the highest accuracy. The speech act and tense have the same accuracy scores across all versions, .44 and .56, respectively, since these classifiers are stable within the system variants. In this evaluation, aspect varies slightly across approaches as it depends on combinations of speech act and robot-concept relation slot values-its accuracy ranges from .25-.49, with the Abrams+ variant obtaining the highest result. Thus, we see that our hypothesis that speech act, tense, and aspect classification may be fairly robust to a new domain is partially confirmed: robot-concept classification is certainly the most challenging with the lowest accuracy, but the performance of all components is significantly worse than the original domain, suggesting more widespread differences in the language of the two domains.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Minecraft Conversion to Dialogue-AMR",
"sec_num": "5.2"
},
{
"text": "Here, we describe the small amount of domain extension done to tailor our G2G conversion system to the Minecraft domain, beginning with extensions of the annotation schema itself.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Domain Extension",
"sec_num": "6"
},
{
"text": "One expert Standard-AMR and Dialogue-AMR annotator provided manual Dialogue-AMR annotations to a continuous 100-instance subset of the Minecraft corpus to serve as a test set. This was done by manually augmenting the Standard-AMR release of the Minecraft corpus, maintaining all of the Standard-AMR annotation choices. Additionally, a separate, continuous 200-instance subset of the data was annotated with speech acts and the corresponding robot-concept relations of Dialogue-AMR to serve as training data for the speech act classifier and robot-concept relation classification. 7 In providing the manual Dialogue-AMR annotation of the Minecraft data, we noted several changes and additions that needed to be made to the annotation schema to account for novel concepts arising in the collaborative building domain, as well as novel dialogue phenomena. First, as expected, we added agent behaviors that would be needed for this domain: BUILDING, represented with the relation build-01 (e.g., What are we building this time?), and PLACING, represented with the relation move-01 (e.g., Please place two red blocks on top of each side...).",
"cite_spans": [
{
"start": 580,
"end": 581,
"text": "7",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Extending Dialogue-AMR Schema",
"sec_num": "6.1"
},
{
"text": "Second, we noted novel dialogue phenomena that we had not observed in the SCOUT data. Speech acts were often nested in this data, such that the content of one speech act was not a typical agent behavior (e.g., a speech act of commanding a ROTATION behavior), but instead another speech act. For example, there were frequent requests for evaluation, often after each building step was completed: How's this? and Is this good? 8 As a result, we had to shift our annotation schema and conversion system in order to allow for speech act relations to sit where we would normally expect the robot-concept relation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extending Dialogue-AMR Schema",
"sec_num": "6.1"
},
{
"text": "Finally, we noted frequent use of the verb need as an indicator of a less direct command in the Minecraft data: This will need to be placed as far right as you can.... This was interpreted by the interlocutor as a command, i.e., Place this as far right as you can. Thus, the need relation that roots the Standard-AMR ultimately mapped to the command-SA relation of the Dialogue AMR. This phenomenon has significant ramifications for the conversion system, as it was generally assumed, for the SCOUT data, that the utterance and Standard-AMR provides propositional content cuing the robot-concept relation, but we did not expect AMR relations corresponding to the speech act in our 7 Contact the first author for Minecraft Dialogue-AMR annotations used for train/test.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extending Dialogue-AMR Schema",
"sec_num": "6.1"
},
{
"text": "8 Following Bunt et al. (2012) , Dialogue-AMR speech acts are distinguished between Information Transfer Functions and Action Discussion Functions. Thus, while syntactically questions, cases such as How's this? are not annotated using the Dialogue-AMR Question speech act, which is reserved for questions that obligate the addressee to introduce new information content into the conversation and demonstrate a commitment to the answer assertion (Traum, 2003) . In contrast, these cases obligate the addressee to evaluate the current state of play while simultaneously providing feedback that common conversational ground has been achieved with respect to the desired structure. Indeed, common responses such as Excellent, Builder do not fit with a question interpretation. domain, although plausible (e.g., I command you to move forward).",
"cite_spans": [
{
"start": 12,
"end": 30,
"text": "Bunt et al. (2012)",
"ref_id": "BIBREF5"
},
{
"start": 445,
"end": 458,
"text": "(Traum, 2003)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Extending Dialogue-AMR Schema",
"sec_num": "6.1"
},
{
"text": "We added to our expanded keyword dictionary to test the effectiveness of a rule-based approach in domain extension. Only two additional concepts were required, build-01 and move-01, but these robot concepts are extremely prevalent in the data. Additionally, in order to test how well a classifier-based approach would capture new behaviors and extend the conversion system to a new domain, we retrained the robot-concept classifier on 166 new manually-annotated training instances of robot concepts from the Minecraft domain. Domain extension also included retraining the speech act classifier on 224 speech acts found in 200 instances of manually annotated Minecraft data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extending Robot-Concept Classification",
"sec_num": "6.2"
},
{
"text": "After domain extension, the G2G variant leveraging the one-hot classifier (.71 F-score) very slightly outperforms the keyword (.70 F-score) and GloVe variants (.70 F-score) (again, comparing system output against manually assigned Dialogue-AMRs for the continuous, 100-instance Minecraft test set).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Domain-Extended G2G Evaluation",
"sec_num": "6.3"
},
{
"text": "Results are summarized in the bottom three rows of Evaluation Domain B of Table 2 . The scores remain comparable across all three variants, but we do see improvement overall when comparing against system variants prior to domain extension.",
"cite_spans": [],
"ref_spans": [
{
"start": 74,
"end": 81,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Domain-Extended G2G Evaluation",
"sec_num": "6.3"
},
{
"text": "Turning to analysis of the accuracy of individual components of the conversion system, the additional training instances improve speech act classification (from .44 prior to retraining to .57 after) and robot-concept classification for the Minecraft domain. Prior to domain extension, the expanded keyword variant achieved the highest accuracy for robot-concept classification (.32), but classifierbased methods with more training data outperform even a domain-extended, data-driven keyword approach, which achieves an accuracy of .41, while one-hot vectorization achieves an accuracy of .54 and GloVe .45. Error analysis reveals that the keyword-based approach struggles to classify robot concepts in this domain, in part, because of language that contains vocatives (e.g. Excellent, builder)-which triggers a top say-01 relation in the Standard-AMR graph-and various uses of need, which trigger a need-01 relation. As noted in the discussion of domain extension of the annotation schema ( \u00a76.1), both of these root relations do not cue any domain robot concept, but rather provide information about speech acts and speaker/listener roles, which were consistently implicit in our original domain. Thus, we are currently updating the system to allow for certain relations in the Standard-AMR (e.g., need-01) to cue for or map to particular speech acts (e.g., command-SA).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Domain-Extended G2G Evaluation",
"sec_num": "6.3"
},
{
"text": "This demonstrates a weakness of the keywordbased approach in general: unforeseen linguistic phenomena such as vocatives can strongly affect the accuracy of this approach, while the classifier approach is more robust to these differences since it considers all tokens in the utterance for robot-concept relation prediction, thereby avoiding mis-classification due to this kind of \"noise\" in the data. When considering our earlier hypothesis that the classifier-based approach to robot-concept classification would be more efficient to extend to a new domain than the keyword-based approach, the results and error analysis here provide modest support for this hypothesis. Both approaches are similarly time-efficient as far as the initial extension efforts are concerned: the keyword approach requires manual observation of the data and subsequent selection and addition of keywords to the dictionaries associated with certain robot-concept relations, while the classifier approach requires some additional manual annotation in the new domain. However, empirically the classifier-based approach slightly outperforms the keyword-based approach in the Minecraft domain, and extending the keyword-based approach requires additional changes in traversal of the graph in order to find the appropriate concept to serve as the keyword for matching, so the effort necessarily goes beyond merely selecting and adding keywords.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Domain-Extended G2G Evaluation",
"sec_num": "6.3"
},
{
"text": "Turning back to our original SCOUT test set after Minecraft domain extension (results summarized in the bottom three rows of Evaluation Domain A in Table 2 ), we find that tailoring the conversion system to Minecraft and expanding the coverage of language that the system can handle has little negative effect on performance in our original domain. We see comparable results for the classifier-based model using one-hot vectorization, maintaining an F-score of .83, which was also the best-performing model for the original domain.",
"cite_spans": [],
"ref_spans": [
{
"start": 148,
"end": 155,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Domain-Extended G2G Evaluation",
"sec_num": "6.3"
},
{
"text": "In order to scale up to real-time use, the two-step NLU pipeline will leverage the retrained automatic Standard-AMR parser described in \u00a73.1; however, up to this point we have reported conversion system results using manually obtained, gold-standard Standard-AMR parses in order to explore the validity of our conversion system approaches without the noise from parsing. Table 3 summarizes the performance of the overall best-performing (across both Smatch scores and component accuracy) expanded keyword and one-hot vectorization classifier G2G variants, after domain extension, given Standard-AMR input from the parser. The expanded keyword variant is the best-performing model with automatic input, but the scores are close. Although the Smatch F-score has dropped from .71 (with gold-standard input) to .59, we still find this to be very encouraging performance, given the challenges of semantic parsing in a new domain. ",
"cite_spans": [],
"ref_spans": [
{
"start": 371,
"end": 378,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Full Automatic Pipeline Evaluation",
"sec_num": "6.4"
},
{
"text": "This research is part of a growing body of work in representing various levels of interpretation in existing meaning representation frameworks, and in AMR in particular. We briefly note especially relevant work here. Bastianelli et al. (2014) present their Human Robot Interaction Corpus (HuRIC) following the same Penman Notation (Penman Natural Language Group, 1989) syntax of AMR, but significantly altering AMR to use the sense distinctions and semantic role labels of FrameNet (Fillmore et al., 2012), thereby rendering the use of automatic parsers trained on AMR data challenging. Shen (2018) presents a small corpus (266 instances) of manually annotated AMRs for spoken language to explore the validity of using AMR for spoken language understanding, with promising results but noting that additional data is needed. There is also a neural AMR graph converter for abstractive summarization (producing summary graphs from source graphs) (Liu et al., 2015) ; however, neural approaches require substantial training data in the form of annotated input and output graphs. The current motivation for the multi-step approach explored here is to handle a low resource problem, as we lack sufficient data to experiment with employing a neural network.",
"cite_spans": [
{
"start": 217,
"end": 242,
"text": "Bastianelli et al. (2014)",
"ref_id": "BIBREF1"
},
{
"start": 943,
"end": 961,
"text": "(Liu et al., 2015)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "7"
},
{
"text": "This paper evaluates and improves upon a two-step NLU pipeline that gradually tames the variation of language so that it can be understood and acted upon by a robot with a limited repertoire of domain concepts and behaviors. After enumerating the extensions needed for the annotation schema itself and contributing a dataset of Dialogue-AMR for the new Minecraft collaborative dialogue domain, we achieve promising results with roughly 200 instances of training data. We have integrated our updated pipeline into a software stack for a physical robot and are now performing a series of experiments where we use the same dialogue-management system, but vary the NLU component in order to compare task success with the two-step NLU pipeline against a baseline NLU system with a simple syntactic parser. We hypothesize that the NLU pipeline described here, and the deeper semantics of Dialogue-AMR specifically, will be especially advantageous for tracking and grounding user utterances involving coreference (e.g., Go to the sign and send a picture of it.), light verb constructions, which AMR represents identically to parallel synthetic verbs (e.g., make a left turn; turn left), negation (e.g., no, not the door on the right, the left!), and complex, nested prepositions (e.g., move through the doorway in front of you on the left)-all utterances where a simple syntactic parse has been found to lack information needed for interpretation of the intent and grounding. The extrinsic evaluation will also provide an opportunity to explore whether or not the conversion system variant with the best overall Smatch scores corresponds to the best real-world performance, or if we should consider other metrics, such as S 2 match (Opitz et al., 2020) and SemBleu (Song and Gildea, 2019) . As our results did not demonstrate a clear \"best\" rule-based, keyword or classifier approach to domain extension, we will continue to experiment with all three variants and consider which is the most time-efficient to extend, either by adding to the keyword dictionary or adding annotations. Overall, we are optimistic that the semantic representation of Dialogue-AMR, which provides a deeper understanding of both what a person said and what they really meant in the conversational context, will enhance human-robot collaboration.",
"cite_spans": [
{
"start": 1725,
"end": 1745,
"text": "(Opitz et al., 2020)",
"ref_id": "BIBREF16"
},
{
"start": 1758,
"end": 1781,
"text": "(Song and Gildea, 2019)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions & Future Work",
"sec_num": "8"
},
{
"text": "Smatch is an evaluation algorithm for scoring AMR graphs(Cai and Knight, 2013).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The Minecraft AMR corpus includes AMRs for the locations of blocks (expressed as Cartesian coordinates) as each movement takes place; because our focus is natural language dialogue, we removed these instances from our test set.6 Bonn et al.(2020)report an F-score of .66 on a Minecraft test set after retraining the Zhang et al. (2019) parser on Minecraft data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Graph-to-graph meaning representation transformations for human-robot dialogue",
"authors": [
{
"first": "Mitchell",
"middle": [],
"last": "Abrams",
"suffix": ""
},
{
"first": "Claire",
"middle": [],
"last": "Bonial",
"suffix": ""
},
{
"first": "Lucia",
"middle": [],
"last": "Donatelli",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the Society for Computation in Linguistics 2020",
"volume": "",
"issue": "",
"pages": "250--253",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mitchell Abrams, Claire Bonial, and Lucia Donatelli. 2020. Graph-to-graph meaning representation trans- formations for human-robot dialogue. In Proceed- ings of the Society for Computation in Linguistics 2020, pages 250-253, New York, New York. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "HuRIC: a human robot interaction corpus",
"authors": [
{
"first": "Emanuele",
"middle": [],
"last": "Bastianelli",
"suffix": ""
},
{
"first": "Giuseppe",
"middle": [],
"last": "Castellucci",
"suffix": ""
},
{
"first": "Danilo",
"middle": [],
"last": "Croce",
"suffix": ""
},
{
"first": "Luca",
"middle": [],
"last": "Iocchi",
"suffix": ""
},
{
"first": "Roberto",
"middle": [],
"last": "Basili",
"suffix": ""
},
{
"first": "Daniele",
"middle": [],
"last": "Nardi",
"suffix": ""
}
],
"year": 2014,
"venue": "LREC",
"volume": "",
"issue": "",
"pages": "4519--4526",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Emanuele Bastianelli, Giuseppe Castellucci, Danilo Croce, Luca Iocchi, Roberto Basili, and Daniele Nardi. 2014. HuRIC: a human robot interaction cor- pus. In LREC, pages 4519-4526.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Dialogue-AMR: Abstract Meaning Representation for dialogue",
"authors": [
{
"first": "Claire",
"middle": [],
"last": "Bonial",
"suffix": ""
},
{
"first": "Lucia",
"middle": [],
"last": "Donatelli",
"suffix": ""
},
{
"first": "Mitchell",
"middle": [],
"last": "Abrams",
"suffix": ""
},
{
"first": "Stephanie",
"middle": [
"M"
],
"last": "Lukin",
"suffix": ""
},
{
"first": "Stephen",
"middle": [],
"last": "Tratz",
"suffix": ""
},
{
"first": "Matthew",
"middle": [],
"last": "Marge",
"suffix": ""
},
{
"first": "Ron",
"middle": [],
"last": "Artstein",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Traum",
"suffix": ""
},
{
"first": "Clare",
"middle": [],
"last": "Voss",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 12th Language Resources and Evaluation Conference",
"volume": "",
"issue": "",
"pages": "684--695",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Claire Bonial, Lucia Donatelli, Mitchell Abrams, Stephanie M. Lukin, Stephen Tratz, Matthew Marge, Ron Artstein, David Traum, and Clare Voss. 2020. Dialogue-AMR: Abstract Meaning Representation for dialogue. In Proceedings of the 12th Lan- guage Resources and Evaluation Conference, pages 684-695, Marseille, France. European Language Re- sources Association.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Augmenting Abstract Meaning Representation for human-robot dialogue",
"authors": [
{
"first": "Claire",
"middle": [],
"last": "Bonial",
"suffix": ""
},
{
"first": "Lucia",
"middle": [],
"last": "Donatelli",
"suffix": ""
},
{
"first": "Stephanie",
"middle": [
"M"
],
"last": "Lukin",
"suffix": ""
},
{
"first": "Stephen",
"middle": [],
"last": "Tratz",
"suffix": ""
},
{
"first": "Ron",
"middle": [],
"last": "Artstein",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Traum",
"suffix": ""
},
{
"first": "Clare",
"middle": [],
"last": "Voss",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the First International Workshop on Designing Meaning Representations",
"volume": "",
"issue": "",
"pages": "199--210",
"other_ids": {
"DOI": [
"10.18653/v1/W19-3322"
]
},
"num": null,
"urls": [],
"raw_text": "Claire Bonial, Lucia Donatelli, Stephanie M. Lukin, Stephen Tratz, Ron Artstein, David Traum, and Clare Voss. 2019. Augmenting Abstract Meaning Representation for human-robot dialogue. In Pro- ceedings of the First International Workshop on De- signing Meaning Representations, pages 199-210, Florence, Italy. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Spatial AMR: Expanded spatial annotation in the context of a grounded Minecraft corpus",
"authors": [
{
"first": "Julia",
"middle": [],
"last": "Bonn",
"suffix": ""
},
{
"first": "Martha",
"middle": [],
"last": "Palmer",
"suffix": ""
},
{
"first": "Zheng",
"middle": [],
"last": "Cai",
"suffix": ""
},
{
"first": "Kristin",
"middle": [],
"last": "Wright-Bettner",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 12th Language Resources and Evaluation Conference",
"volume": "",
"issue": "",
"pages": "4883--4892",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Julia Bonn, Martha Palmer, Zheng Cai, and Kristin Wright-Bettner. 2020. Spatial AMR: Expanded spatial annotation in the context of a grounded Minecraft corpus. In Proceedings of the 12th Lan- guage Resources and Evaluation Conference, pages 4883-4892, Marseille, France. European Language Resources Association.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "ISO 24617-2: A semantically-based standard for dialogue annotation",
"authors": [
{
"first": "Harry",
"middle": [],
"last": "Bunt",
"suffix": ""
},
{
"first": "Jan",
"middle": [],
"last": "Alexandersson",
"suffix": ""
},
{
"first": "Jae-Woong",
"middle": [],
"last": "Choe",
"suffix": ""
},
{
"first": "Alex",
"middle": [
"Chengyu"
],
"last": "Fang",
"suffix": ""
},
{
"first": "Koiti",
"middle": [],
"last": "Hasida",
"suffix": ""
},
{
"first": "Volha",
"middle": [],
"last": "Petukhova",
"suffix": ""
},
{
"first": "Andrei",
"middle": [],
"last": "Popescu-Belis",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Traum",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12)",
"volume": "",
"issue": "",
"pages": "430--437",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Harry Bunt, Jan Alexandersson, Jae-Woong Choe, Alex Chengyu Fang, Koiti Hasida, Volha Petukhova, Andrei Popescu-Belis, and David Traum. 2012. ISO 24617-2: A semantically-based standard for dia- logue annotation. In Proceedings of the Eighth In- ternational Conference on Language Resources and Evaluation (LREC'12), pages 430-437.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Smatch: an evaluation metric for semantic feature structures",
"authors": [
{
"first": "Shu",
"middle": [],
"last": "Cai",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "748--752",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shu Cai and Kevin Knight. 2013. Smatch: an evalua- tion metric for semantic feature structures. In Pro- ceedings of the 51st Annual Meeting of the Associa- tion for Computational Linguistics (Volume 2: Short Papers), pages 748-752, Sofia, Bulgaria. Associa- tion for Computational Linguistics.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"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
},
"BIBREF8": {
"ref_id": "b8",
"title": "Annotation of tense and aspect semantics for sentential AMR",
"authors": [
{
"first": "Lucia",
"middle": [],
"last": "Donatelli",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Regan",
"suffix": ""
},
{
"first": "William",
"middle": [],
"last": "Croft",
"suffix": ""
},
{
"first": "Nathan",
"middle": [],
"last": "Schneider",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Joint Workshop on Linguistic Annotation, Multiword Expressions and Constructions",
"volume": "",
"issue": "",
"pages": "96--108",
"other_ids": {
"DOI": [
"10.1207/s15516709cog0303_1"
]
},
"num": null,
"urls": [],
"raw_text": "Lucia Donatelli, Michael Regan, William Croft, and Nathan Schneider. 2018. Annotation of tense and as- pect semantics for sentential AMR. In Proceedings of the Joint Workshop on Linguistic Annotation, Mul- tiword Expressions and Constructions (LAW-MWE- CxG-2018), pages 96-108.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The FrameNet Constructicon. Signbased construction grammar",
"authors": [
{
"first": "J",
"middle": [],
"last": "Charles",
"suffix": ""
},
{
"first": "Russell",
"middle": [],
"last": "Fillmore",
"suffix": ""
},
{
"first": "Russell",
"middle": [],
"last": "Lee-Goldman",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Rhodes",
"suffix": ""
}
],
"year": 2012,
"venue": "",
"volume": "",
"issue": "",
"pages": "309--372",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Charles J Fillmore, Russell Lee-Goldman, and Russell Rhodes. 2012. The FrameNet Constructicon. Sign- based construction grammar, pages 309-372.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Two constraints on speech act ambiguity",
"authors": [
{
"first": "Elizabeth",
"middle": [],
"last": "Hinkelman",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Allen",
"suffix": ""
}
],
"year": 1989,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Elizabeth Hinkelman and James Allen. 1989. Two con- straints on speech act ambiguity.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Compositional semantic parsing across graphbanks",
"authors": [
{
"first": "Matthias",
"middle": [],
"last": "Lindemann",
"suffix": ""
},
{
"first": "Jonas",
"middle": [],
"last": "Groschwitz",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Koller",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "4576--4585",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1450"
]
},
"num": null,
"urls": [],
"raw_text": "Matthias Lindemann, Jonas Groschwitz, and Alexan- der Koller. 2019. Compositional semantic parsing across graphbanks. In Proceedings of the 57th An- nual Meeting of the Association for Computational Linguistics, pages 4576-4585, Florence, Italy. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Toward abstractive summarization using semantic representations",
"authors": [
{
"first": "Fei",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Flanigan",
"suffix": ""
},
{
"first": "Sam",
"middle": [],
"last": "Thomson",
"suffix": ""
},
{
"first": "Norman",
"middle": [],
"last": "Sadeh",
"suffix": ""
},
{
"first": "Noah A",
"middle": [],
"last": "Smith",
"suffix": ""
}
],
"year": 2015,
"venue": "Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fei Liu, Jeffrey Flanigan, Sam Thomson, Norman Sadeh, and Noah A Smith. 2015. Toward abstrac- tive summarization using semantic representations. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Applying the Wizard-of-Oz technique to multimodal human-robot dialogue",
"authors": [
{
"first": "Matthew",
"middle": [],
"last": "Marge",
"suffix": ""
},
{
"first": "Claire",
"middle": [],
"last": "Bonial",
"suffix": ""
},
{
"first": "Brendan",
"middle": [],
"last": "Byrne",
"suffix": ""
},
{
"first": "Taylor",
"middle": [],
"last": "Cassidy",
"suffix": ""
},
{
"first": "A",
"middle": [
"William"
],
"last": "Evans",
"suffix": ""
},
{
"first": "Susan",
"middle": [
"G"
],
"last": "Hill",
"suffix": ""
},
{
"first": "Clare",
"middle": [],
"last": "Voss",
"suffix": ""
}
],
"year": 2016,
"venue": "RO-MAN 2016: IEEE International Symposium on Robot and Human Interactive Communication",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matthew Marge, Claire Bonial, Brendan Byrne, Taylor Cassidy, A. William Evans, Susan G. Hill, and Clare Voss. 2016. Applying the Wizard-of-Oz technique to multimodal human-robot dialogue. In RO-MAN 2016: IEEE International Symposium on Robot and Human Interactive Communication.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Exploring variation of natural human commands to a robot in a collaborative navigation task",
"authors": [
{
"first": "Matthew",
"middle": [],
"last": "Marge",
"suffix": ""
},
{
"first": "Claire",
"middle": [],
"last": "Bonial",
"suffix": ""
},
{
"first": "Ashley",
"middle": [],
"last": "Foots",
"suffix": ""
},
{
"first": "Cory",
"middle": [],
"last": "Hayes",
"suffix": ""
},
{
"first": "Cassidy",
"middle": [],
"last": "Henry",
"suffix": ""
},
{
"first": "Kimberly",
"middle": [],
"last": "Pollard",
"suffix": ""
},
{
"first": "Ron",
"middle": [],
"last": "Artstein",
"suffix": ""
},
{
"first": "Clare",
"middle": [],
"last": "Voss",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Traum",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the First Workshop on Language Grounding for Robotics",
"volume": "",
"issue": "",
"pages": "58--66",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matthew Marge, Claire Bonial, Ashley Foots, Cory Hayes, Cassidy Henry, Kimberly Pollard, Ron Art- stein, Clare Voss, and David Traum. 2017. Explor- ing variation of natural human commands to a robot in a collaborative navigation task. In Proceedings of the First Workshop on Language Grounding for Robotics, pages 58-66.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Collaborative dialogue in Minecraft",
"authors": [
{
"first": "Anjali",
"middle": [],
"last": "Narayan-Chen",
"suffix": ""
},
{
"first": "Prashant",
"middle": [],
"last": "Jayannavar",
"suffix": ""
},
{
"first": "Julia",
"middle": [],
"last": "Hockenmaier",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "5405--5415",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1537"
]
},
"num": null,
"urls": [],
"raw_text": "Anjali Narayan-Chen, Prashant Jayannavar, and Ju- lia Hockenmaier. 2019. Collaborative dialogue in Minecraft. In Proceedings of the 57th Annual Meet- ing of the Association for Computational Linguistics, pages 5405-5415, Florence, Italy. Association for Computational Linguistics.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "AMR similarity metrics from principles",
"authors": [
{
"first": "Juri",
"middle": [],
"last": "Opitz",
"suffix": ""
},
{
"first": "Letitia",
"middle": [],
"last": "Parcalabescu",
"suffix": ""
},
{
"first": "Anette",
"middle": [],
"last": "Frank",
"suffix": ""
}
],
"year": 2020,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "8",
"issue": "",
"pages": "522--538",
"other_ids": {
"DOI": [
"10.1162/tacl_a_00329"
]
},
"num": null,
"urls": [],
"raw_text": "Juri Opitz, Letitia Parcalabescu, and Anette Frank. 2020. AMR similarity metrics from principles. Transactions of the Association for Computational Linguistics, 8:522-538.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Penman Natural Language Group",
"authors": [],
"year": 1989,
"venue": "Information Sciences Institute",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Penman Natural Language Group. 1989. The Penman user guide. Technical report, Information Sciences Institute.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Wizard of Oz Studies in HRI: A Systematic Review and New Reporting Guidelines",
"authors": [
{
"first": "Laurel",
"middle": [],
"last": "Riek",
"suffix": ""
}
],
"year": 2012,
"venue": "Journal of Human-Robot Interaction",
"volume": "",
"issue": "1",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Laurel Riek. 2012. Wizard of Oz Studies in HRI: A Systematic Review and New Reporting Guidelines. Journal of Human-Robot Interaction, 1(1).",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Semantic Parsing in Spoken Language Understanding using Abstract Meaning Representation",
"authors": [
{
"first": "Hongyuan",
"middle": [],
"last": "Shen",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hongyuan Shen. 2018. Semantic Parsing in Spoken Language Understanding using Abstract Meaning Representation. Ph.D. thesis, Brandeis University.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Sembleu: A robust metric for amr parsing evaluation",
"authors": [
{
"first": "Linfeng",
"middle": [],
"last": "Song",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Gildea",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1905.10726"
]
},
"num": null,
"urls": [],
"raw_text": "Linfeng Song and Daniel Gildea. 2019. Sembleu: A robust metric for amr parsing evaluation. arXiv preprint arXiv:1905.10726.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Semantics and pragmatics of questions and answers for dialogue agents",
"authors": [
{
"first": "David",
"middle": [],
"last": "Traum",
"suffix": ""
}
],
"year": 2003,
"venue": "proceedings of the International Workshop on Computational Semantics",
"volume": "",
"issue": "",
"pages": "380--394",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Traum. 2003. Semantics and pragmatics of questions and answers for dialogue agents. In pro- ceedings of the International Workshop on Compu- tational Semantics, pages 380-394.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "AMR Parsing as Sequence-to-Graph Transduction",
"authors": [
{
"first": "Sheng",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Xutai",
"middle": [],
"last": "Ma",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Duh",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Van Durme",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sheng Zhang, Xutai Ma, Kevin Duh, and Benjamin Van Durme. 2019. AMR Parsing as Sequence-to- Graph Transduction. In Proceedings of the 57th An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Florence, Italy. Association for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"type_str": "figure",
"num": null,
"text": "c2 / commander) :direction (b / back)) :ARG1 (g / go-02 :completable -:ARG0 r :direction (b / back) :time (a / after :op1 (n / now))) :ARG2 (r / robot)) Move back in (a) Standard-AMR (parser output), (b) Dialogue-AMR (conversion system output)."
},
"FIGREF1": {
"uris": null,
"type_str": "figure",
"num": null,
"text": "Standard and Dialogue-AMR comparison for Commander instructing robot Move forward three feet."
},
"TABREF0": {
"html": null,
"type_str": "table",
"text": "Retrained AMR parser Smatch results on SCOUT Continuous-trial test set.",
"num": null,
"content": "<table><tr><td>Parser</td><td>Training</td><td>P</td><td>R</td><td>F</td></tr><tr><td/><td>AMR 2.0</td><td colspan=\"3\">.47 .77 .58</td></tr><tr><td>Zhang et al.</td><td colspan=\"4\">2.0 + DialAMR .73 .77 .75 AMR 3.0 .52 .80 .63</td></tr><tr><td/><td colspan=\"4\">3.0 + DialAMR .88 .89 .89</td></tr><tr><td/><td>AMR 2.0</td><td colspan=\"3\">.53 .77 .63</td></tr><tr><td>Lindemann</td><td colspan=\"4\">2.0 + DialAMR .92 .94 .93 AMR 3.0 .55 .81 .65</td></tr><tr><td/><td colspan=\"4\">3.0 + DialAMR .91 .95 .93</td></tr></table>"
},
"TABREF2": {
"html": null,
"type_str": "table",
"text": "Summary of Smatch scores & Robot-Concept Relation classification accuracy for each variant conversion system, including our G2G system before and after Minecraft domain extension, tested on SCOUT and Minecraft.",
"num": null,
"content": "<table/>"
},
"TABREF4": {
"html": null,
"type_str": "table",
"text": "Smatch scores for best-performing domainextended (ext.) G2G variants using automatically obtained Standard-AMR input from retrained parser.",
"num": null,
"content": "<table/>"
}
}
}
} |