File size: 97,825 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 |
{
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:27:57.116382Z"
},
"title": "OMEGA: A Probabilistic Approach to Referring Expression Generation in a Virtual Environment",
"authors": [
{
"first": "Maurice",
"middle": [],
"last": "Langner",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Ruhr-Universit\u00e4t Bochum",
"location": {}
},
"email": "maurice.langner@rub.de"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In recent years, referring expression generation algorithms were inspired by game theory and probability theory. In this paper, an algorithm is designed for the generation of referring expressions (REG) that base on both models by integrating maximization of utilities into the content determination process. It implements cognitive models for assessing visual salience of objects and additional features. In order to evaluate the algorithm properly and validate the applicability of existing models and evaluative information criteria, both, production and comprehension studies, are conducted using a complex domain of objects, providing new directions of approaching the evaluation of REG algorithms.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "In recent years, referring expression generation algorithms were inspired by game theory and probability theory. In this paper, an algorithm is designed for the generation of referring expressions (REG) that base on both models by integrating maximization of utilities into the content determination process. It implements cognitive models for assessing visual salience of objects and additional features. In order to evaluate the algorithm properly and validate the applicability of existing models and evaluative information criteria, both, production and comprehension studies, are conducted using a complex domain of objects, providing new directions of approaching the evaluation of REG algorithms.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Probabilistic and game-theoretic approaches to REG base on the maximization of utilities (Frank and Goodman, 2012; Goodman and Frank, 2016) or on corpus frequencies of attributes that are selected according to random float values (e.g. the PRO model, Gompel et al., 2019) . These models and their humanlikeness of production were tested on domains of limited size, mostly three objects, and a minimal set of 3 to 4 properties. The parameters of these models are learned using Maximum Likelihood Estimation. Due to the non-deterministic nature of probabilistic models, evaluation metrics used for deterministic models, especially the Dice Score, cannot be applied. Therefore, the likelihood of the predicted distribution, given the observed data, is computed using the Bayesian Information Criterion (BIC). With a limited set of possible referring expressions, the likelihood function, which bases on the multinomial distribution, produces reasonable results, but an increasing complexity of the domain may necessitate a different evaluation approach.",
"cite_spans": [
{
"start": 89,
"end": 114,
"text": "(Frank and Goodman, 2012;",
"ref_id": "BIBREF5"
},
{
"start": 115,
"end": 139,
"text": "Goodman and Frank, 2016)",
"ref_id": "BIBREF10"
},
{
"start": 251,
"end": 271,
"text": "Gompel et al., 2019)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Empirical data against which the reference games model (Frank and Goodman, 2012; Franke and J\u00e4ger, 2016) was tested has been collected in betting games in a forced-choice paradigm. Such data sets are not only non-reproducible for larger domains, but also circumvent the integration of underlying cognitive processes of content determination into the model. Hence, the aim of this paper is to develop a new probabilistic REG-algorithm that is able to handle the choice of complex attributes and spatial relations on the basis of a cognitively motivated tradeoff between salience, preference and discriminatory power (DP). An important aspect of this tradeoff must be the possibility of generating minimal and overspecified expressions in complex domains for overcoming the deficiency of models rooted in game theory, which typically produce unary expressions only. Overspecification is of essential importance for a humanlike modeling of reference, since overspecification has been proved to facilitate the identification task for the hearer if the redundant information is easily accessible (Paraboni et al., 2017) .",
"cite_spans": [
{
"start": 55,
"end": 80,
"text": "(Frank and Goodman, 2012;",
"ref_id": "BIBREF5"
},
{
"start": 81,
"end": 104,
"text": "Franke and J\u00e4ger, 2016)",
"ref_id": "BIBREF7"
},
{
"start": 1091,
"end": 1114,
"text": "(Paraboni et al., 2017)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In what follows, first the domain underlying the experimental studies will be presented. Then the production survey is described. Subsequently, cognitive models for the approximation of visual salience are introduced which are used in the OMEGA algorithm. Section 5 gives a detailed description of the developed model. Section 6 evaluates the algorithm performance in both humanlikeness and comprehension, giving insight into problems and advantages of deployed evaluation methods.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The domain for the empirical studies consists of 3D geometric primitives with different properties. Atomic properties are SHAPE and NAME TAG, while SIZE is a scalar property whose value is determined in the context, and COLOUR relies on the calculation of shades for RED, BLUE and GREEN. The light and dark shades are each prefixed with the according specifier light-or dark-. Two further relational properties are given as coordinates for x-axis and y-axis. The design of this corpus is different from other well-renowned corpora (e.g. TUNA) in several important aspects. Colors provide different degrees of visual difference. Size is varying by fixed ratios. The domain on which the production data from participants is collected must agree with the domain for the comprehension study. Last but not least, the type attribute is fixed as gift box, such that the generation of the SHAPE attribute does not vary. Each object is rendered with each possible attribute combination in a 3D setting with identical lighting and orientation, which results in a domain with 360 different objects. Figure 1 , six sample objects are given. Assuming that the top left picture is the target, the referring expression the cone would be insufficient for identification; a minimal expression would contain expressions for COORDINATES or COLOUR, SIZE and TYPE. Adding further attributes to these combinations results in overspecification, e.g. the small light-blue cone in the top row.",
"cite_spans": [],
"ref_spans": [
{
"start": 1088,
"end": 1096,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "The reference domain",
"sec_num": "2"
},
{
"text": "A web-based empirical survey has been conducted in order to collect production data from native speakers of English. This data reveals the attribute choice and reference habits of the participants. The production survey is split into two subexperiments. In the first part, speakers are rating the visual salience of attribute values for each attribute separately, e.g. in the set of COLOUR values, participants rate the salience of each value within the set of colours. Additionally, for SIZE and COLOUR, the correlation between the gradience and the salience ratio is assessed by letting the participants rate the visual difference between values. An example of this assessment of the visual difference is given in Figure 2 . The intention behind this survey design is to provide an attribute-specific rating of visual salience or prominence, respectively. These values serve as input to the REG models presented in the following sections. The downside of this is the dependence of models on corpora for each unseen domain. For this reason, section 4 proposes cognitive models that may replace corpus data and make REG models independent from corpora and collected production data. The second part of the study comprises production sessions. In each session, between 3 and 9 objects are depicted in a grid (see Figure 1 ). The speaker's task is to compose an expression that uniquely identifies the target object which is highlighted by a golden frame. Participants were able to select and deselect attribute values, candidate objects for spatial relations and in turn attributes for the description of the spatially related object. Attribute selection is restricted to the choice from domain attributes and their values in order to avoid the necessity of normalizing free text input after collecting the corpus data. Participants could also produce a surface realization from their selected attributes. An intriguing fact is that this restriction of attribute selection does not trigger a shift in preferential usage of attributes in comparison to e.g. the TUNA corpus and other psycholinguistic studies (e.g. Tarenskeen et al., 2015) . The production subdomains of 10 sessions were fixed, 15 subsequent sessions were randomized. The web application allowed for the integration of identical objects several times, such that identification may be possible only via using coordinates. Via Prolific, 100 native speakers of English from the United Kingdom were acquired. The produced corpus consists of about 2500 production sessions. The pie chart in figure 3 reveals the amount of overspecified phrases with 60% of all produced expressions in the data. The capability of producing overspecified expressions is consequently of high importance for humanlikeness in the REG task. The outer circle is divided into expressions with and without spatial relations, while the inner circle visualizes the sum of the respective generation quality. A well justified metric is needed in order to model the perceptual prominence of each object as a substitute for salience ratings. If this salience is a joint effect of the visual salience of the object's properties, a utility function for a REG model needs to integrate a computation of how much the object springs to the speaker's eye. For size, this is straightforward, since the domain objects were rendered with three different size factors. The prior production survey reveals a linear relation between the increase in size factor and the increase in perceptual size difference. Van Deemter and his coauthors (2016, p.141; van Gompel et al., 2014) prove that in REG, the ratio of size difference has essential importance on the production probability of this property. Therefore, for a REG model, this ratio models the salience for size in a set of objects. For COLOUR, comparing values is more complex. Cognitive experiments on colour perception led to the conception of the CIElab colour space (Mokrzycki and Tatol, 2011). Formulas have been construed that calculate the difference between two colours, the \u2206E value, that humans may still perceive. Multiple revisions of the CIElab \u2206E functions exist with different terms for compensating hue and saturation, but they all base on the Euclidean distance of colours which are represented as 3-dimensional vectors in the CIElab colour space. These \u2206E functions allow to compute the difference between two colours in a cognitively reasonable way and can be used in a utility function. A challenge to be faced is that in realistic settings corresponding metrics are required for other attributes as well, e.g. shapes. Although not implemented in the model presented in this paper, attempts are made to formalize the similarity measures between shapes by analysing their vertices (Mori et al., 2001 ) and points on the contour (Ling and Jacobs, 2007) . Latecki and Lakaemper 2000developed a model for shape comparison that is close to cognitive processes of noisy shape perception. Relational attributes may be modeled using an approach comparable to Kelleher and Genabith's (2004) measure of assessing the distance towards the center of the visual field, but this is subject to further research. The main advantage of approximating visual salience with these cognitive models is that values are calculated in situ on the present visual field, instead of optimizing parameters such that they fit a large data set best. In this way, referring expressions may be generated in a more individual way and be tailored to the present domain, while other methods rather capture the whole data on average. Finally, these cognitive models make REG models independent from experimental studies as given in section 3 under the premise that they perform equally well.",
"cite_spans": [
{
"start": 2112,
"end": 2136,
"text": "Tarenskeen et al., 2015)",
"ref_id": "BIBREF24"
},
{
"start": 3553,
"end": 3566,
"text": "(2016, p.141;",
"ref_id": null
},
{
"start": 3567,
"end": 3591,
"text": "van Gompel et al., 2014)",
"ref_id": "BIBREF9"
},
{
"start": 4770,
"end": 4788,
"text": "(Mori et al., 2001",
"ref_id": "BIBREF19"
},
{
"start": 4817,
"end": 4840,
"text": "(Ling and Jacobs, 2007)",
"ref_id": "BIBREF16"
},
{
"start": 5041,
"end": 5071,
"text": "Kelleher and Genabith's (2004)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 716,
"end": 724,
"text": "Figure 2",
"ref_id": "FIGREF1"
},
{
"start": 1312,
"end": 1320,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Empirical production survey",
"sec_num": "3"
},
{
"text": "The overspecifying, utility maximizing referring expression generation algorithm, short OMEGA, is an incremental, non-deterministic and probabilistic algorithm which uses a utility function in order to calculate which attributes are the most utile ones for production. This includes the choice of a candidate object for spatial relations and the attribute selection for the description of this candidate.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "At its core, a utility function as used in reference games (Frank and Goodman, 2012) computes production probabilities which guide content determination. The concept of this utility function is crucial for the performance of the whole model. At this point, an observation made on the empirical data from the prior study becomes a decisive factor.",
"cite_spans": [
{
"start": 59,
"end": 84,
"text": "(Frank and Goodman, 2012)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "When high DP coincides with high salience, the usage of the respective attribute increases drastically, while high DP for non-salient attributes does not influence attribute selection significantly. Additionally, van Deemter (2016) and van Gompel et al. (2014) show that larger proportional differences in gradable attributes, e.g. SIZE, trigger a higher usage in referring expressions than small differences that are cognitively less easily assessable. After all, this is the interference between DP and salience. The 'object-distractor contrast', as Mast (2016, p.140) terms it, in regard to a specific attribute is influenced both by the number of distractors that provide the same attribute value as well as by the visual distinctiveness of the attribute value in comparison with other values of the same attribute that apply to the distractors. Therefore, salience, preference and DP must be taken balanced when defining the utility function.",
"cite_spans": [
{
"start": 240,
"end": 260,
"text": "Gompel et al. (2014)",
"ref_id": "BIBREF9"
},
{
"start": 552,
"end": 570,
"text": "Mast (2016, p.140)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "The salience computation which is used in the utility definition of the OMEGA algorithm integrates the cognitive models of perception presented in the previous section. The formulae below describe how the salience values are computed. The attribute value of the target object is compared to the values of the distractors in regard to that attribute and the average value is calculated. Therefore, salience is represented as the average visual difference between the target and the set of distractors D. The more different, the higher is the salience. The factor \u03b1 projects the \u2206 E value onto the interval of numbers between 0 and 1, c is a colour vector.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "SAL colour (c, D) = i (\u03b1\u2206E(c, ci)) |D|",
"eq_num": "(1)"
}
],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "For size, the absolute value of the difference between s and s i is needed in order to prevent negative values when s < s i .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "SALsize(s, D) = i abs(s \u2212 si) |D|",
"eq_num": "(2)"
}
],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "The same method is used for discriminatory power which is a ratio of how many distractors an attribute value rules out. Considering the fact that attribute selection is performed across attributes (and not across attribute values), it is possible to calculate for each attribute, of which a value applies to the target, what the average DP is in comparison with values of the respective attribute that apply to the distractors.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "AV GDP (e, D) = e max(0, DP (e, D) \u2212 DP (ei, D)) |D|",
"eq_num": "(3)"
}
],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "In order to prevent negative utilities, the function only adds the difference of e to some e i \u2208 D to the sum if the result is > 0. In other words, if the DP of the target attribute is smaller than the DP of a competitive attribute, the negative distance is not added, but the average DP nevertheless is decreased since it is normalized against the number of all competitive attributes, i.e. including those with higher DP. For the OMEGA model, the utility function u is defined as the product of the average DP and the salience of the value v A of attribute A that is true of the target object in domain D. To this product, the preference degree estimated by the frequency of attributes in the production data can be added under the assumption that preference is not independent from the joint effect of salience and DP. The decision to integrate preference is also motivated by the findings of Ferreira and Paraboni (2014) , who prove that the integration of speaker preferences enhances the quality of their SVM classifiers for REG.",
"cite_spans": [
{
"start": 896,
"end": 924,
"text": "Ferreira and Paraboni (2014)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "U (vA, A, D) = AV GDP (vA, D) * SALA(vA, D) + P ref (A)",
"eq_num": "(4)"
}
],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "In RSA and other log-linear models a logarithm is used in order to map large utility values to a smaller interval of results (Qing and Franke, 2015; Franke and Degen, 2016) . This has the positive effect of assigning the lowest possible utility, namely negative infinity, to attributes with a literal interpretation of 0, or, as in OMEGA, irrelevant information. In combination with softmaximizing utilities, the model conforms with Grice's Maxim of Quantity. Softmaximization which is deployed in OMEGA means turning the utilities into weighted production probabilities, as shown in the formula (5). ",
"cite_spans": [
{
"start": 125,
"end": 148,
"text": "(Qing and Franke, 2015;",
"ref_id": "BIBREF21"
},
{
"start": 149,
"end": 172,
"text": "Franke and Degen, 2016)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "By means of the chain rule we are able to calculate the probabilities of all candidate referring expressions e which are sequences of attribute values a that are true of the referent. After attribute value a 1 is selected, the set of distractors D of domain M is updated, as well as the attribute a 1 belongs to is removed from the set of selectable properties. This probability is multiplied with the result of calculating the probability of a 2 given the updated data structures and the already selected attributes, and so on until a n has been achieved. This mechanism permits the OMEGA model to produce a probability distribution over all possible expressions that are true of the target.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "A closer analysis of the pseudo code in algorithm 1 will show how OMEGA implements the selection of reference objects and production of spatial relations, as well as the enrichment with redundant attributes for overspecification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "Line 4 initializes two empty data structures e and e rel for attributes that are selected during the production process. In line 5, rel c is a container for storing possible candidates from the set of distractors that may be used for a spatial relation. The subsequent while-loop (lines 6 to 14) calculates for each attribute a in the set of attributes A what production probability it provides in the given visual field. In line 9, the production probability is calculated, which integrates the utility function described earlier as well as the cognitive models for salience approximation. If the probability is higher than the one of the previous attribute, the maximally probable attribute is replaced with a. Then, data structures are updated. If some distractor is found for which the production probability for the present attribute a is larger then for the target, it is stored as a candidate for spatial relations (line 10). The while-loop terminates when either the set of distractors M or the set of attributes A is empty. In line 15, if some candidate for spatial relations has been found, OMEGA is recursively called on that object. Otherwise, the overspecification process starts (line 18). While overspecification is selected and A is not empty, maximal probable attributes are added to the expression using the same mechanism as in the first while loop.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "A possible point of criticism is that the usage of the formula for production probability may not be suitable for overspecification, since the utility tries to calculate the average salience and DP values on an empty set of distractors. In full awareness of this circumstance the functions for salience and DP are intended to integrate default values. For DP, the default value is 1, while for salience, the prior while A = \u2205 \u2227 M = \u2205 do 7:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "amax \u2190 NONE 8:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "for a \u2208 A do 9:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "p \u2190 P(a|r, D) 10:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "relc \u2190 relc \u222a {r \u2032 |r \u2032 \u2208 M \u2227 P(a|r \u2032 ,D) > p} 11:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "if p > amax then 12:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "amax \u2190 a 13:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "e.insert(a) 14:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "update M , A 15:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "if relc = \u2205 then 16:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "r rel \u2190 ARGMAX(relc) 17: e rel \u2190 OMEGA(r rel , D, M , A rel , \u2126) 18:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "while RANDOMBOOLEAN(\u2126) is T rue \u2227 A = \u2205 do 19:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "a \u2032 max \u2190 NONE 20:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "for a \u2032 \u2208 A \u2032 do 21:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "p \u2032 \u2190 P(a \u2032 |r, D) 22: if p \u2032 > a \u2032 max then 23: a \u2032 max \u2190 a \u2032 24:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "e.insert(a \u2032 ) 25:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "update A \u2032 26:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "return e, e rel ratings are used. In summary this means, that the overspecification process is guided exceptionally by salience and preference, which is comparable to PRO and IA (Dale and Reiter, 1995; Reiter and Dale, 1992) , while also being in perfect agreement with psycholinguistic studies on overspecification in REG (Tarenskeen et al., 2015) . A legitimate criticism is indeed, that OMEGA does not cover the cases correctly where both target description and relational description are underspecified, but producing a minimal expression in joint effect. A modification to OMEGA that allows to cover this small number of expressions found in the empirical data remains open for further research. An intriguing point is that both selection steps share the same set of attributes, meaning that for overspecification the residual attributes are used that have not been selected beforehand. As a result, attributes that were not produced in the first step, for reasons of an average DP value close or equal to zero, are reconsidered in the overspecification step. Nonetheless OMEGA makes a giant leap towards independence from corpus data by implementing the cognitive models of salience approximation as a substitute for corpus statistics. The evaluation in the next section gives insight in which way the cognitive models influence the performance of the model in comparison to its competitors. Table 2 lists the technical differences between the competitor models and OMEGA. On the domain displayed in figure 1, these algorithms may generate different expressions. IA generates deterministically according to the preference order until all distractors are ruled out. This may also trigger overspecification. For the top left object as target, IA will add attributes according to preference degree, namely first COLOUR and TYPE, then SIZE. Since this is not discriminating yet, it further adds NAME TAG, resulting in a minimal expression the small light-blue cone with a circular name tag'. Since IA is deterministic, this expression is generated at every instance. PRO and OMEGA produce a probability distribution over all possible expressions. PRO offers optimized parameters for each attribute plus the overspecification parameter, which are fine-tuned to match a full data set. In this case, the visual difference between the light-blue and light-green objects is rather unsalient, except for the red sphere the visual field is rather unichrome. While PRO uses the preference degree optimized for the full data set in order to probabilistically choose to (or not to) add colour, OMEGA recognizes the small visual salience on the domain in situ and assigns to colour a low production probability accordingly. Since DP influences the selection as well, highly discriminating attributes, e.g. coordinates, are considered. COLOUR and TYPE are being reconsidered only for the overspecification process. Additionally, the visual prominence of the NAME TAG attribute with the value circular for the target object is diminishingly small, although it has highest DP. PRO may consider this attribute as well, while OMEGA identifies the low perceptive recognizability, resulting in a tiny salience value, which decreases the production probability. Since OMEGA recognizes the comparably far more salient object, \"the red sphere\" with both unique and highly salient attribute values for COLOUR and TYPE, OMEGA will add a spatial description to the expression, resulting in, e.g., the leftmost light-blue cone to the left of the red sphere.",
"cite_spans": [
{
"start": 178,
"end": 201,
"text": "(Dale and Reiter, 1995;",
"ref_id": "BIBREF2"
},
{
"start": 202,
"end": 224,
"text": "Reiter and Dale, 1992)",
"ref_id": "BIBREF22"
},
{
"start": 323,
"end": 348,
"text": "(Tarenskeen et al., 2015)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [
{
"start": 1398,
"end": 1405,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "The OMEGA model",
"sec_num": "5"
},
{
"text": "For evaluation, the OMEGA model has been tested competitively against the PRO model and the standard IA algorithm. Both, humanlikeness in production and performance in a comprehension task, have been evaluated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "6"
},
{
"text": "For the production side, the Dice Score is not applicable due to the non-deterministic nature of both OMEGA and PRO. For this reason, often BIC is used to evaluate the performance of the models given the production corpus that is collected in the empirical production survey (van Deemter, 2016). For toy domains with only two or three attributes, the number of possible referring expressions is limited to 3 and 9, respectively. A probability distribution over those categories is therefore less noisy than in a domain with 6 attributes, which give 2 6 = 64 combinations of attributes. Due to the number of categories, of which many did not occur in the corpus while being generated by both algorithms and vice versa, the Maximum-Likelihood-Method (MLE) failed for both models. The reason might be that the optimizer mainly modeled noise due to data sparseness of the empirical data. Therefore, the likelihood of the predicted distributions, given the corpus data, could not be assessed. Instead of training model parameters, corpus frequencies were used as model parameters for PRO, but still the likelihood was 0. A fact that may explain this conversion towards 0 is that even the probabilities that are calculated from the data, given the same data, have a likelihood of only 19.245e \u221251 . Gompel et al. (2019) list their values for experiments in a scenario that equals the reference games by Franke and J\u00e4ger (2016) . In this simple domain, the smallest likelihood listed is about 4.9 * 10 \u221269 , which indicates that the likelihood on large domains with far more categories might be too small to be represented by a float value. Dice Score is inapplicable but may serve as a blueprint for an evaluation method that is at least capable of ap-proximating the performance of the models.",
"cite_spans": [
{
"start": 1293,
"end": 1313,
"text": "Gompel et al. (2019)",
"ref_id": "BIBREF8"
},
{
"start": 1397,
"end": 1420,
"text": "Franke and J\u00e4ger (2016)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Humanlikeness",
"sec_num": "6.1"
},
{
"text": "PRO and OMEGA are both models that predict a probability distribution over a discrete number of possible expressions. The empirical data provides about 2500 data points. A referring expression of participant q i in session s i is predicted by both models on session s i with a specific probability. For each data point, it is possible to determine with what chance it is produced by the models given the domain and the referent. Instead of calculating how likely the distribution predicted by a model is, given the empirical data, which also entails the interpretation of the empirical data as the 'truth' (Lewandowsky and Farrell, 2011, p.181) , one can at least calculate the average probability with which the models would generate the referring expressions produced by the participants.",
"cite_spans": [
{
"start": 606,
"end": 644,
"text": "(Lewandowsky and Farrell, 2011, p.181)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Humanlikeness",
"sec_num": "6.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "PDice = n i=1 PM (ei|ri, Di) N",
"eq_num": "(6)"
}
],
"section": "Humanlikeness",
"sec_num": "6.1"
},
{
"text": "Formula (6) captures the intuition behind this criterion, which could be circumscribed as a probabilistic Dice Score. The Model P M assigns a probability to the expression e i that is contained in data point i, given the corresponding referent r i and the domain D i . By dividing the sum of these probabilities by the number of data points N , the formula calculates the average probability model M assigns to the referring expressions that were produced by the participants of the study. The according graph (figure 4) not only reveals that OMEGA performs better regarding this data set, but also that some probabilities that PRO assigns to expressions are negative, which means that some calculations in the PRO model cause some of the probabilities to get a negative polarity.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Humanlikeness",
"sec_num": "6.1"
},
{
"text": "The reason for this is the optimization of parameters as probabilities for attributes, as well as the mathematical operations including overspecification eagerness. In figure (5), the probability of producing a unary expression COLOUR is (c \u2212 e), where c is the probability that is proportional to the preference degree of COLOUR, whereas e is the parameter of overspecification eagerness. After optimization, the parameter for a rather unsalient attribute may be optimized in such a way that c < e, which causes c \u2212 e to be negative. Effectively, the choice of subtraction or addition of a value e, that is not a probability by definition, from or to a probability, is not only theoretically questionable, but also leads to values that cannot be interpreted as probabilities. Indeed, data sparseness may have caused the optimizer to produce parameters that provided this configuration of c < e, but the problem is definitely rooted in the subtraction and addition of the probabilities that are assigned to the nodes of the decision tree. A possible remedy is using multiplication instead, such that the model is sound in the context of probability theory. Unfortunately, this has the downside of loosing both the elegant difference in mathematically combining overspecification eagerness with probabilities for continued overspecification and termination of the selection process. Nevertheless, the modification did not significantly change the performance of PRO for humanlikeness. As described above, the evaluation methods for REG algorithms on the basis of empirical production data cannot be straightforwardly applied to probabilistic, non-deterministic models and larger domains. For BIC, data sparseness is the main reason for failure, since MLE cannot produce reasonable parameter estimates if the likelihood function provides no gradient. Dice Score cannot be applied to non-deterministic algorithms, since the score varies as much as the output does. Nonetheless, instead of calculating the sequential identity of expressions (Dice), the average probability can be determined which the models assign to the referring expressions in the empirical data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Humanlikeness",
"sec_num": "6.1"
},
{
"text": "Humanlikeness alone does not help to fully evaluate the performance of REG models. Human speakers are not always optimal, e.g. due to errorprone speaking. Evaluating humanlikeness only would mean modeling humanlike reference without knowledge of how well the referring expressions are understood and serve the purpose of identifying the target.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "Another empirical study is necessary in order to assess how well human speakers comprehend the REs the models produce. Objective measures, e.g. identification time and error rate, give insight which model produces more optimal expressions in regard to comprehension.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "The setting of the comprehension study in which the OMEGA model and its competitors are evaluated is a sales dialogue in a store for custom gift packages. A sales agent, which is simulated by the REG models, guides the participant through the survey and instructs him or her to find storage units and gift packages contained in these stores. The listener's task is to find the storage and identify the target object. The participant can freely move around in the virtual environment and perceives the rooms from the first person view. After identification of ten storage units and ten objects, the survey ends. Throughout the sessions, the system chooses randomly from the set of REG models, but the ten object domains are identical to the 10 fixed sessions in the production session in section 3.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "For production of referring expressions, the sales agent (the generation system respectively) has access to the participant's position in the room and his direction of view. In this way, the model can give precise directional expressions that are tailored to the configuration of the virtual room and the participant's movements. Figure 7 depicts a screenshot from the participant's perspective on domain S1, for which OMEGA generates the following expressions: 1. \"Please got to the storage to the right of the antique column at the wall in front of you.\" 2. \"Yes, here it is. Now, please find the red cylindershaped gift box.\" The sales agent gives feedback for each identification. For storage units, the system recognizes the participant's proximity to the storage as well as the direction of view. Participants identify gift boxes in these storage units by clicking when the cursor hovers above the object of their choice. Then, the sales agent gives feedback and either directs to the next storage on success or repeats the expression on false identification.",
"cite_spans": [],
"ref_spans": [
{
"start": 330,
"end": 338,
"text": "Figure 7",
"ref_id": "FIGREF8"
}
],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "In this implementation, the referring expressions and instructions that are generated by the REG models are given as audio input. This reduces identification onset due to different reading speed and prevents distraction from the visual stimuli that may occur if the REs are presented as text. Participants can ask for repetition of the expression. The simulated sales agent then utters the expression again, but updates the RE according to the new player's position and direction of view.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "The virtual environment (see layout in Figure 6 ) is less complex than the setting used for the GIVE challenges in regard to the room configuration (Koller et al., 2009 (Koller et al., ,2010 Byron et al., 2007 Striegnitz et al., 2011) , since the number of rooms is smaller and there are no separate floors the participants need to be guided through. This was intentional, because the focus is shifted to referring expressions and not to instructions as used in navigation systems.",
"cite_spans": [
{
"start": 149,
"end": 169,
"text": "(Koller et al., 2009",
"ref_id": "BIBREF12"
},
{
"start": 170,
"end": 191,
"text": "(Koller et al., ,2010",
"ref_id": "BIBREF13"
},
{
"start": 192,
"end": 210,
"text": "Byron et al., 2007",
"ref_id": "BIBREF0"
},
{
"start": 211,
"end": 235,
"text": "Striegnitz et al., 2011)",
"ref_id": "BIBREF23"
}
],
"ref_spans": [
{
"start": 39,
"end": 48,
"text": "Figure 6",
"ref_id": "FIGREF7"
}
],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "The study was conducted on 36 participants with sufficient knowledge of English. The participants completed the study on a local computer with a mouse, headset or loudspeaker. Two participants were excluded. The resulting corpus consists of 680 completed identification tasks and the respective measures of identification time, error rate and repetition counts. Average identification times are evaluated across participants for object identification and storage locating separately. The differences between PRO and OMEGA as well as IA and OMEGA are significant, with OMEGA having the highest average in identification times. The reason for this may be, that OMEGA is the only algorithm that is capable of producing spatial relations. In case that participants await the speaker to finish, this may increase the onset of the identification process. When spatial relations are excluded, the identification times for OMEGA with an average of 8.32 seconds are significantly lower than for PRO with an average of 10.62 seconds (F=4.006, p=0.04), while the difference between IA (average of 10.01 seconds) and OMEGA is close to significance (F=3.39, p=0.06) . A straightforward conclusion to be drawn from this is that spatial relations prolong the identification process, while they may also increase proportionally with the length of the referring expression. For storage locating, there is no significant difference between PRO and OMEGA, but OMEGA is significantly faster than IA (F=3.89, p=0.049). An important observation to be made is that for IA on storage locating, the lowest times are more frequent than for PRO and OMEGA, but the distribution of times for IA is also much wider and outliers above 50 seconds are more frequent than for the competitors. A reason for this may be the generation quality, since IA tends far more often to produce minimal expressions, while PRO and OMEGA frequently overspecify.",
"cite_spans": [],
"ref_spans": [
{
"start": 1136,
"end": 1152,
"text": "(F=3.39, p=0.06)",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "In each comprehension session, during object identification, the application counted how often the participant selected a wrong object. This error rate is the second objective evaluation criterion to be used for comparing which of the three models performs better. Significance tests between data sets reveal that no difference between models is significant. Nonetheless, independent from the inclusion of spatial relations, OMEGA provides the smallest averages of errors with 0.87 and 0.85 respectively for expressions with and without spatial relations. IA is close to OMEGA with an average of 0.88 errors per session, while PRO is decently higher with 1.017 errors on average. Exclusion of spatial relations truncates the error rate at a maximum of three errors per session, but it also triggers a higher average. Since the difference is once more insignificant, no valid conclusion can be made, but the difference between the distributions indicates that spatial relations may prolong the identification process, but also show a tendency to reduce the error rate.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "As one of the most important observations, a comprehension study for REG models in a virtual environment is subject to variance between participants as much as other empirical studies that collect language data from human speakers and listeners. Nevertheless, the comprehension study elicited a significant difference between OMEGA and its competitors in relation to the time that participants need to identify the target object. The data set indicates that OMEGA produces referring expressions that are more optimized to listener comprehension than PRO and IA, with a lower average error when spatial relations are given.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comprehension",
"sec_num": "6.2"
},
{
"text": "The OMEGA model succeeds in integrating the tradeoff between the most important parameters of REG, namely discriminatory power, salience and preference. The maximization of their joint effect, which defines the utility, allows to incrementally and probabilistically build referring expressions that may also contain spatial relations. The OMEGA model outperforms both IA and PRO in a comprehension task in a virtual environment. Nonetheless, more research is needed to further optimize the cognitive models that compute the salience values for different attributes which are needed for the OMEGA model. These measures provide the advantage of calculating salience in situ on the preset visual field, resulting in more independence from empirical input and parameter tuning at an improved performance level. One of the most important findings is the inapplicability of information criteria like BIC and optimization methods like MLE to models, given complex domains with numerous, realistically infinite, possible referring expressions. This is a reasonable indicator for the shift to comprehension as the standard evaluative method for REG models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Generating instructions in virtual environments (GIVE): A challenge and an evaluation testbed for NLG",
"authors": [
{
"first": "Donna",
"middle": [],
"last": "Byron",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Koller",
"suffix": ""
},
{
"first": "Jon",
"middle": [],
"last": "Oberlander",
"suffix": ""
},
{
"first": "Laura",
"middle": [],
"last": "Stoia",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Striegnitz",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the Workshop on Shared Tasks and Comparative Evaluation in Natural Language Generation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Donna Byron, Alexander Koller, Jon Oberlander, Laura Stoia, and Kristina Striegnitz. 2007. Generating in- structions in virtual environments (GIVE): A chal- lenge and an evaluation testbed for NLG. Proceed- ings of the Workshop on Shared Tasks and Compar- ative Evaluation in Natural Language Generation.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Report on the first NLG challenge on generating instructions in virtual environments (GIVE)",
"authors": [
{
"first": "Donna",
"middle": [],
"last": "Byron",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Striegnitz",
"suffix": ""
},
{
"first": "Justine",
"middle": [],
"last": "Cassell",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Dale",
"suffix": ""
},
{
"first": "Johanna",
"middle": [],
"last": "Moore",
"suffix": ""
},
{
"first": "Jon",
"middle": [],
"last": "Oberlander",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the 12th European Workshop on Natural Language Generation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Donna Byron, Kristina Striegnitz, Justine Cassell, Robert Dale, Johanna Moore, and Jon Oberlander. 2009. Report on the first NLG challenge on gen- erating instructions in virtual environments (GIVE). Proceedings of the 12th European Workshop on Nat- ural Language Generation (ENLG 2009).",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Computational interpretations of the gricean maxims in the generation of referring expressions",
"authors": [
{
"first": "Robert",
"middle": [],
"last": "Dale",
"suffix": ""
},
{
"first": "Ehud",
"middle": [],
"last": "Reiter",
"suffix": ""
}
],
"year": 1995,
"venue": "Cognitive Science",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Robert Dale and Ehud Reiter. 1995. Computational in- terpretations of the gricean maxims in the generation of referring expressions. Cognitive Science, 19.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Computational Models of Referring: A Study in Cognitive Science",
"authors": [
{
"first": "",
"middle": [],
"last": "Kees Van Deemter",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kees van Deemter. 2016. Computational Models of Re- ferring: A Study in Cognitive Science.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Referring expression generation: Taking speakers' preferences into account",
"authors": [
{
"first": "Thiago",
"middle": [],
"last": "Ferreira",
"suffix": ""
},
{
"first": "Ivandr\u00e9",
"middle": [],
"last": "Paraboni",
"suffix": ""
}
],
"year": 2014,
"venue": "International Conference on Text, Speech and Dialogue",
"volume": "8655",
"issue": "",
"pages": "539--546",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Thiago Ferreira and Ivandr\u00e9 Paraboni. 2014. Refer- ring expression generation: Taking speakers' pref- erences into account. International Conference on Text, Speech and Dialogue, 8655:539-546.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Predicting pragmatic reasoning in language games",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Frank",
"suffix": ""
},
{
"first": "Noah",
"middle": [],
"last": "Goodman",
"suffix": ""
}
],
"year": 2012,
"venue": "Science",
"volume": "336",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Frank and Noah Goodman. 2012. Predicting pragmatic reasoning in language games. Science (New York, N.Y.), 336:998.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Reasoning in reference games: Individual-vs. population-level probabilistic modeling",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Franke",
"suffix": ""
},
{
"first": "Judith",
"middle": [],
"last": "Degen",
"suffix": ""
}
],
"year": 2016,
"venue": "PLOS ONE",
"volume": "11",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Franke and Judith Degen. 2016. Reasoning in reference games: Individual-vs. population-level probabilistic modeling. PLOS ONE, 11:e0154854.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Probabilistic pragmatics, or why bayes' rule is probably important for pragmatics",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Franke",
"suffix": ""
},
{
"first": "Gerhard",
"middle": [],
"last": "J\u00e4ger",
"suffix": ""
}
],
"year": 2016,
"venue": "Zeitschrift f\u00fcr Sprachwissenschaft",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Franke and Gerhard J\u00e4ger. 2016. Probabilistic pragmatics, or why bayes' rule is probably important for pragmatics. Zeitschrift f\u00fcr Sprachwissenschaft, 35.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Conceptualization in reference production: Probabilistic modeling and experimental testing",
"authors": [
{
"first": "Roger",
"middle": [],
"last": "Gompel",
"suffix": ""
},
{
"first": "Albert",
"middle": [],
"last": "Kees Van Deemter",
"suffix": ""
},
{
"first": "Rick",
"middle": [],
"last": "Gatt",
"suffix": ""
},
{
"first": "Emiel",
"middle": [],
"last": "Snoeren",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Krahmer",
"suffix": ""
}
],
"year": 2019,
"venue": "Psychological Review",
"volume": "126",
"issue": "",
"pages": "345--373",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Roger Gompel, Kees van Deemter, Albert Gatt, Rick Snoeren, and Emiel Krahmer. 2019. Conceptual- ization in reference production: Probabilistic model- ing and experimental testing. Psychological Review, 126:345-373.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Overspecification in reference: Modelling size contrast effects",
"authors": [
{
"first": "Albert",
"middle": [],
"last": "Roger Van Gompel",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Gatt",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 2014 conference on architectures and mechanisms in language processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Roger van Gompel, Albert Gatt, Emiel Krahmer, and Kees van Deemter. 2014. Overspecification in refer- ence: Modelling size contrast effects. Proceedings of the 2014 conference on architectures and mecha- nisms in language processing.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Pragmatic language interpretation as probabilistic inference",
"authors": [
{
"first": "Noah",
"middle": [],
"last": "Goodman",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Frank",
"suffix": ""
}
],
"year": 2016,
"venue": "Trends in Cognitive Sciences",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Noah Goodman and Michael Frank. 2016. Pragmatic language interpretation as probabilistic inference. Trends in Cognitive Sciences, 20.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Visual salience and reference resolution in simulated 3-d environments",
"authors": [
{
"first": "John",
"middle": [],
"last": "Kelleher",
"suffix": ""
},
{
"first": "Josef",
"middle": [],
"last": "Genabith",
"suffix": ""
}
],
"year": 2004,
"venue": "Artif. Intell. Rev",
"volume": "21",
"issue": "",
"pages": "253--267",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John Kelleher and Josef Genabith. 2004. Visual salience and reference resolution in simulated 3-d environments. Artif. Intell. Rev., 21:253-267.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "The software architecture for the first challenge on generating instructions in virtual environments",
"authors": [
{
"first": "Alexander",
"middle": [],
"last": "Koller",
"suffix": ""
},
{
"first": "Donna",
"middle": [],
"last": "Byron",
"suffix": ""
},
{
"first": "Justine",
"middle": [],
"last": "Cassell",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Dale",
"suffix": ""
},
{
"first": "Johanna",
"middle": [],
"last": "Moore",
"suffix": ""
},
{
"first": "Jon",
"middle": [],
"last": "Oberlander",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Striegnitz",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Demonstrations Session at EACL 2009",
"volume": "",
"issue": "",
"pages": "33--36",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexander Koller, Donna Byron, Justine Cassell, Robert Dale, Johanna Moore, Jon Oberlander, and Kristina Striegnitz. 2009. The software architecture for the first challenge on generating instructions in virtual environments. Proceedings of the Demon- strations Session at EACL 2009, pages 33-36.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Report on the second nlg challenge on generating instructions in virtual environments (GIVE-2)",
"authors": [
{
"first": "Alexander",
"middle": [],
"last": "Koller",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Striegnitz",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Gargett",
"suffix": ""
},
{
"first": "Donna",
"middle": [],
"last": "Byron",
"suffix": ""
},
{
"first": "Justine",
"middle": [],
"last": "Cassell",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Dale",
"suffix": ""
},
{
"first": "Johanna",
"middle": [],
"last": "Moore",
"suffix": ""
},
{
"first": "Jon",
"middle": [],
"last": "Oberlander",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the 6th International Natural Language Generation Conference",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexander Koller, Kristina Striegnitz, Andrew Gargett, Donna Byron, Justine Cassell, Robert Dale, Johanna Moore, and Jon Oberlander. 2010. Report on the sec- ond nlg challenge on generating instructions in vir- tual environments (GIVE-2). Proceedings of the 6th International Natural Language Generation Confer- ence.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Shape similarity measure based on correspondence of visual parts. Pattern Analysis and Machine Intelligence",
"authors": [
{
"first": "Jan",
"middle": [],
"last": "Longin",
"suffix": ""
},
{
"first": "Rolf",
"middle": [],
"last": "Latecki",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lakaemper",
"suffix": ""
}
],
"year": 2000,
"venue": "IEEE Transactions on",
"volume": "22",
"issue": "",
"pages": "1185--1190",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Longin Jan Latecki and Rolf Lakaemper. 2000. Shape similarity measure based on correspondence of vi- sual parts. Pattern Analysis and Machine Intelli- gence, IEEE Transactions on, 22:1185 -1190.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Computational Modeling in Cognition: Principles and Practice",
"authors": [
{
"first": "Stephan",
"middle": [],
"last": "Lewandowsky",
"suffix": ""
},
{
"first": "Simon",
"middle": [],
"last": "Farrell",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Stephan Lewandowsky and Simon Farrell. 2011. Com- putational Modeling in Cognition: Principles and Practice.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Shape classification using the inner-distance",
"authors": [
{
"first": "Haibin",
"middle": [],
"last": "Ling",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Jacobs",
"suffix": ""
}
],
"year": 2007,
"venue": "IEEE transactions on pattern analysis and machine intelligence",
"volume": "29",
"issue": "",
"pages": "286--99",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Haibin Ling and David Jacobs. 2007. Shape classifica- tion using the inner-distance. IEEE transactions on pattern analysis and machine intelligence, 29:286- 99.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Referring Expression Generation in Situated Interaction",
"authors": [
{
"first": "Vivien",
"middle": [],
"last": "Mast",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vivien Mast. 2016. Referring Expression Generation in Situated Interaction. Ph.D. thesis.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Color difference delta E -A survey. Machine Graphics and Vision",
"authors": [
{
"first": "Wojciech",
"middle": [],
"last": "Mokrzycki",
"suffix": ""
},
{
"first": "Maciej",
"middle": [],
"last": "Tatol",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "20",
"issue": "",
"pages": "383--411",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wojciech Mokrzycki and Maciej Tatol. 2011. Color difference delta E -A survey. Machine Graphics and Vision, 20:383-411.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Shape contexts enable efficient retrieval of similar shapes",
"authors": [
{
"first": "G",
"middle": [],
"last": "Mori",
"suffix": ""
},
{
"first": "Serge",
"middle": [],
"last": "Belongie",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Malik",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "G Mori, Serge Belongie, and J Malik. 2001. Shape contexts enable efficient retrieval of similar shapes. Proceedings of the IEEE Computer Society Confer- ence on Computer Vision and Pattern Recognition, 1:I-723.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Effects of cognitive effort on the resolution of overspecified descriptions",
"authors": [
{
"first": "Ivandr\u00e9",
"middle": [],
"last": "Paraboni",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Lan",
"suffix": ""
},
{
"first": "Matheus",
"middle": [],
"last": "Ana",
"suffix": ""
},
{
"first": "Fl\u00e1vio",
"middle": [],
"last": "Coutinho",
"suffix": ""
}
],
"year": 2017,
"venue": "Computational Linguistics",
"volume": "43",
"issue": "",
"pages": "1--14",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ivandr\u00e9 Paraboni, Alex Lan, Matheus Ana, and Fl\u00e1vio Coutinho. 2017. Effects of cognitive effort on the resolution of overspecified descriptions. Computa- tional Linguistics, 43:1-14.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Variations on a bayesian theme: Comparing bayesian models of referential reasoning",
"authors": [
{
"first": "Ciyang",
"middle": [],
"last": "Qing",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Franke",
"suffix": ""
}
],
"year": 2015,
"venue": "Bayesian Natural Language Semantics and Pragmatics",
"volume": "",
"issue": "",
"pages": "201--220",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ciyang Qing and Michael Franke. 2015. Variations on a bayesian theme: Comparing bayesian models of referential reasoning. In Henk Zeevat and Hans- Christian Schmitz, editors, Bayesian Natural Lan- guage Semantics and Pragmatics, pages 201-220.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "A fast algorithm for the generation of referring expressions",
"authors": [
{
"first": "Ehud",
"middle": [],
"last": "Reiter",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Dale",
"suffix": ""
}
],
"year": 1992,
"venue": "Proceedings of the 15th International Conference on Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ehud Reiter and Robert Dale. 1992. A fast algorithm for the generation of referring expressions. Proceed- ings of the 15th International Conference on Com- putational Linguistics, 1.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Report on the second second challenge on generating instructions in virtual environments (GIVE2.5)",
"authors": [
{
"first": "Kristina",
"middle": [],
"last": "Striegnitz",
"suffix": ""
},
{
"first": "Alexandre",
"middle": [],
"last": "Denis",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Gargett",
"suffix": ""
},
{
"first": "Konstantina",
"middle": [],
"last": "Garoufi",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Koller",
"suffix": ""
},
{
"first": "Mariet",
"middle": [],
"last": "Theune",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kristina Striegnitz, Alexandre Denis, Andrew Gargett, Konstantina Garoufi, Alexander Koller, and Mariet Theune. 2011. Report on the second second chal- lenge on generating instructions in virtual environ- ments (GIVE2.5). Applied Geomatics.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Overspecification of color, pattern, and size: Salience, absoluteness, and consistency",
"authors": [
{
"first": "Sammie",
"middle": [],
"last": "Tarenskeen",
"suffix": ""
},
{
"first": "Mirjam",
"middle": [],
"last": "Broersma",
"suffix": ""
},
{
"first": "Bart",
"middle": [],
"last": "Geurts",
"suffix": ""
}
],
"year": 2015,
"venue": "Frontiers in Psychology",
"volume": "6",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sammie Tarenskeen, Mirjam Broersma, and Bart Geurts. 2015. Overspecification of color, pattern, and size: Salience, absoluteness, and consistency. Frontiers in Psychology, 6.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"text": "Six sample objects from the domain In",
"uris": null,
"num": null
},
"FIGREF1": {
"type_str": "figure",
"text": "surface of the production survey, part I: rating the visual difference between colour gradients",
"uris": null,
"num": null
},
"FIGREF2": {
"type_str": "figure",
"text": "Generation qualities of the production corpus 4 Cognitive models for approximation of visual salience",
"uris": null,
"num": null
},
"FIGREF3": {
"type_str": "figure",
"text": "(e|r, D) = exp(\u03bbU tility(e; r, D)) e i \u2208D exp(\u03bbU tility(ei; r, D))",
"uris": null,
"num": null
},
"FIGREF4": {
"type_str": "figure",
"text": "Overspecifying Utility-Maximizing Referring Expression Generation Algorithm (OMEGA) 1: Input: reference domain D with target object r and a non-empty set of distractors M, a set of attributes A at least one of whose values is true of r. 2: Output: the referring expression with maximal probability 3: function OMEGA(r, D, M, A,",
"uris": null,
"num": null
},
"FIGREF5": {
"type_str": "figure",
"text": "Probabilistic Dice Score",
"uris": null,
"num": null
},
"FIGREF6": {
"type_str": "figure",
"text": "PRO paths for a small domain with colour (C), size (S) and border (B), where colour is discriminating",
"uris": null,
"num": null
},
"FIGREF7": {
"type_str": "figure",
"text": "Layout of the virtual environment for the comprehension study",
"uris": null,
"num": null
},
"FIGREF8": {
"type_str": "figure",
"text": "example domain from the virtual environment from the participant's perspective",
"uris": null,
"num": null
},
"TABREF1": {
"type_str": "table",
"content": "<table/>",
"html": null,
"num": null,
"text": ""
},
"TABREF3": {
"type_str": "table",
"content": "<table/>",
"html": null,
"num": null,
"text": ""
}
}
}
} |