File size: 68,208 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 |
{
"paper_id": "W07-0308",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T04:41:26.101266Z"
},
"title": "Experiences of an In-Service Wizard-of-Oz Data Collection for the Deployment of a Call-Routing Application",
"authors": [
{
"first": "Mats",
"middle": [],
"last": "Wir\u00e9n",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "TeliaSonera TeliaSonera SE",
"location": {
"postCode": "123 86, SE-751 42",
"settlement": "Farsta, Uppsala",
"country": "Sweden, Sweden"
}
},
"email": ""
},
{
"first": "Robert",
"middle": [],
"last": "Eklund",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "TeliaSonera TeliaSonera SE",
"location": {
"postCode": "123 86, SE-751 42",
"settlement": "Farsta, Uppsala",
"country": "Sweden, Sweden"
}
},
"email": ""
},
{
"first": "Fredrik",
"middle": [],
"last": "Engberg\u00b2",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "TeliaSonera TeliaSonera SE",
"location": {
"postCode": "123 86, SE-751 42",
"settlement": "Farsta, Uppsala",
"country": "Sweden, Sweden"
}
},
"email": ""
},
{
"first": "Johan",
"middle": [],
"last": "Westermark\u00b2",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "TeliaSonera TeliaSonera SE",
"location": {
"postCode": "123 86, SE-751 42",
"settlement": "Farsta, Uppsala",
"country": "Sweden, Sweden"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper describes our experiences of collecting a corpus of 42,000 dialogues for a call-routing application using a Wizard-of-Oz approach. Contrary to common practice in the industry, we did not use the kind of automated application that elicits some speech from the customers and then sends all of them to the same destination, such as the existing touch-tone menu, without paying attention to what they have said. Contrary to the traditional Wizard-of-Oz paradigm, our data-collection application was fully integrated within an existing service, replacing the existing touch-tone navigation system with a simulated callrouting system. Thus, the subjects were real customers calling about real tasks, and the wizards were service agents from our customer care. We provide a detailed exposition of the data collection as such and the application used, and compare our approach to methods previously used.",
"pdf_parse": {
"paper_id": "W07-0308",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper describes our experiences of collecting a corpus of 42,000 dialogues for a call-routing application using a Wizard-of-Oz approach. Contrary to common practice in the industry, we did not use the kind of automated application that elicits some speech from the customers and then sends all of them to the same destination, such as the existing touch-tone menu, without paying attention to what they have said. Contrary to the traditional Wizard-of-Oz paradigm, our data-collection application was fully integrated within an existing service, replacing the existing touch-tone navigation system with a simulated callrouting system. Thus, the subjects were real customers calling about real tasks, and the wizards were service agents from our customer care. We provide a detailed exposition of the data collection as such and the application used, and compare our approach to methods previously used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Spoken-dialogue systems for applications such as customer care increasingly use statistical language models (SLMs) and statistically-based semantic classification for recognition and analysis of utterances. A critical step in designing and deploying such a system is the initial data collection, which must provide a corpus that is both representative of the intended service and sufficiently large for development, training and evaluation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background and introduction",
"sec_num": "1"
},
{
"text": "For at least 20 years, Wizard-of-Oz methodology has been regarded as a superior (though not unproblematic) method of collecting high-quality, machine-directed speech data in the absence of a runnable application. 1 Normally, these data will be useful for several purposes such as guiding dialogue design and training speech recognizers. Still, the Wizard-of-Oz option is often dismissed in favour of simpler methods on the ground that it does not scale well in terms of cost and time (for example, Di Fabbrizio et al. 2005) . Consequently, Wizard-of-Oz has typically been used for data collections that are more limited in the number of subjects involved or utterances collected. One exception from this is the data collection for the original AT&T \"How May I Help You\" system (Gorin et al. 1997; Ammicht et al. 1999) , which comprised three batches of transactions with live customers, each involving up to 12,000 utterances. Other well-known instances are \"Voyager\" (Zue et al. 1989) and the individual ATIS collections (Hirschman et al. 1993) which involved up to a hundred subjects or (again) up to 12,000 utterances.",
"cite_spans": [
{
"start": 501,
"end": 523,
"text": "Fabbrizio et al. 2005)",
"ref_id": "BIBREF4"
},
{
"start": 777,
"end": 796,
"text": "(Gorin et al. 1997;",
"ref_id": "BIBREF6"
},
{
"start": 797,
"end": 817,
"text": "Ammicht et al. 1999)",
"ref_id": "BIBREF1"
},
{
"start": 968,
"end": 985,
"text": "(Zue et al. 1989)",
"ref_id": null
},
{
"start": 1022,
"end": 1045,
"text": "(Hirschman et al. 1993)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background and introduction",
"sec_num": "1"
},
{
"text": "While it is true that Wizard-of-Oz is a labourintensive method, the effort can often be motivated on the ground that it enables significant design and evaluation to be carried out before implementation, thereby reducing the amount of re-design necessary for the actual system. However, one should also bear in mind the crucial advantage brought about by the possibility in a production environment of running the Wizard-of-Oz collection in-service rather than in a closed lab setting. As we shall discuss, the fact that real customers with real problems are involved instead of role-playing subjects with artificial tasks circumvents the key methodological problem that has been raised as an argument against Wizard-of-Oz, namely, lack of realism.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background and introduction",
"sec_num": "1"
},
{
"text": "The aim of this paper is to describe our experiences of running a Wizard-of-Oz collection in a production environment with real customers, with the double purpose of guiding dialogue design and collecting a sufficient amount of data for the first training of a speech recognizer. We also review what other options there are for the initial data collection and compare our Wizard-of-Oz approach with those.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background and introduction",
"sec_num": "1"
},
{
"text": "The rest of this paper is organized as follows: Section 2 describes the call-routing problem and our particular domain. Section 3 gives an overview of the options for the initial data collection and the major trade-offs involved in selecting a method. Section 4 describes the application that was developed for our Wizard-of-Oz data collection, whereas Section 5 describes the actual data collection, summary statistics for the collected data and some experimental results obtained. Section 6 contains a discussion of our overall experiences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background and introduction",
"sec_num": "1"
},
{
"text": "Call routing is the task of directing a caller to a service agent or a self-serve application based on their description of the issue. Increasingly, speechenabled routing is replacing traditional touch-tone menues whereby callers have to navigate to the appropriate destinations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The call-routing task and domain",
"sec_num": "2"
},
{
"text": "The domain of interest in this paper is (the entrance to) the TeliaSonera 2 residential customer care in Sweden, comprising the entire range of services offered: fixed and mobile telephony, broadband and modem-based Internet, IP telephony, digital television, triple play, etc. Around 14 million calls are handled annually, and before the speech-enabled call-routing system was launched in 2006, touch-tone navigation was used throughout. The speech-enabled system involves an SLM-based speech recognizer and a statistically-based classifier. 3 The task of the classifier is to map a spoken utterance to an application category which corresponds to a selfserve application, (a queue to) a human agent, a disambiguation category or a discourse category. Whereas self-serve applications and service agents are the desired goals to reach, disambiguation and discourse categories correspond to intermediate states in the routing dialogue. More specifically, disambiguation categories correspond to cases where the classifier has picked up some information about the destination, but needs to know more in order to route the call. Discourse categories correspond to domain-independent utterances such as greetings (\"Hi, my name is John Doe\"), channel checks (\"Hello?\") and meta questions (\"Who am I talking to?\"). Altogether, there are 124 application categories used by the current classifier.",
"cite_spans": [
{
"start": 543,
"end": 544,
"text": "3",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The call-routing task and domain",
"sec_num": "2"
},
{
"text": "Basically, there are three options for making the initial data collection for a call-routing application: to collect human-human dialogues in a call center, to use an automated data-collection application, or to use a Wizard-of-Oz approach. We shall now describe each of these.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Options for initial data collection",
"sec_num": "3"
},
{
"text": "The simplest possible approach to the initial data collection is to record conversations between service agents and customers in a call center. This is an inexpensive method since it does not require any data-collection application to be built. Also, there is no customer impact. However, the data obtained tend not to be sufficiently representative, for two reasons: First, typically only a subset of the services of a call center is carried out by human agents, and hence many services will not be covered. Second, the characteristics of humanhuman conversations differ from those of humanmachine interaction. Still, this option has sometimes been preferred on the grounds of simplicity and lack of negative customer impact.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Human-human dialogues",
"sec_num": "3.1"
},
{
"text": "Due to the nature of the task, it is easy to put out a fully automated mock-up system in a live service that engages in the initial part of a call-routing dialogue. Typically, such a system will play an open prompt, record the customers' speech, play another prompt saying that the system did not understand, again record the speech, and finally direct all calls to a single destination, such as a general-skills service agent or the entry to the existing touch-tone menu. We estimate that a system of this kind could be implemented and integrated into a call center in about a person week. An example of this approach is the AT&T \"Ghost Wizard\" (referred to in Di Fabbrizio et al. 2005) .",
"cite_spans": [
{
"start": 665,
"end": 687,
"text": "Fabbrizio et al. 2005)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Automated applications",
"sec_num": "3.2"
},
{
"text": "This basic approach can be improved upon by detecting silences and touch-tone events, and in these cases playing designated prompts that try to get the caller on track. Furthermore, if data from previous call-routing applications are available, it is possible to use these to handle domainindependent utterances. Such utterances correspond to discourse categories as mentioned in Section 2, and the idea then is to play prompts that encourage the caller to describe the issue. A description of such an approach is provided by Di Fabbrizio et al. (2005) . A problem with the automated approach is that customer impact can be quite negative, since the application does not actually do anything except for recording their speech (possibly through several turns), and routing them to a \"dummy\" destination where they will have to start over. Of course, one way of avoiding this is to include a human in the loop who listens to the customer's speech and then routes the call to the right destination. Apparently, this is the approach of Di Fabbrizio et al. (2005) , which consequently is not fully automated.",
"cite_spans": [
{
"start": 529,
"end": 552,
"text": "Fabbrizio et al. (2005)",
"ref_id": "BIBREF4"
},
{
"start": 1035,
"end": 1058,
"text": "Fabbrizio et al. (2005)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Automated applications",
"sec_num": "3.2"
},
{
"text": "Apart from customer impact, the problem with an automated system is that we do not learn the full story about caller behaviour. In particular, since typically only a minority of callers will state their issue in an unambiguous way within the given few turns, less information about the callers' actual issues will be obtained. In particular, for callers who completely fail to speak or who give no details about their issue, we will have no possibility of finding out what they wanted and why they failed. Furthermore, since the system lacks the ability to respond intelligently to indomain utterances, no follow-up dialogue such as disambiguation can be collected.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Automated applications",
"sec_num": "3.2"
},
{
"text": "Although Wizard-of-Oz is arguably the best method for collecting machine-directed data in the absence of a running application, it is not without methodological problems. The basic critique has always been aimed at the lack of realism (for example, von Hahn 1986). In a thorough analysis, Allwood & Haglund (1992) point out that in a Wizard-of-Oz simulation, both the subjects and the wizard(s) are playing roles, occupied and assigned. The researcher acting as the wizard is occupying the role of a researcher interested in obtaining \"as natural as possible\" language and speech data, while playing the role of the system. The subject, on the other hand, is occupying the role of a subject in a scientific study, and playing the role of a client (or similar), communicating with a system while carrying out tasks that are not genuine to the subject, but given to them by the experiment leader (who might be identical with the wizard).",
"cite_spans": [
{
"start": 289,
"end": 313,
"text": "Allwood & Haglund (1992)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Wizard-of-Oz",
"sec_num": "3.3"
},
{
"text": "It turns out, however, that a traditional Wizardof-Oz approach with made-up tasks according to a scenario is anyway not an option when collecting data for deploying a call-routing system. The reason for this is that we want to learn not just how callers express themselves, but also what kind of tasks they have, which obviously rules out prewritten scenarios. If the existing system uses touch-tone navigation, usually not too much can be ascertained about this, and trying to design a set of tasks just by looking at the existing destinations would miss the point.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Wizard-of-Oz",
"sec_num": "3.3"
},
{
"text": "By instead integrating a Wizard-of-Oz application in an existing, live service, we can circumvent the key methodological problems, while addressing all the problems of the previously described approaches and even obtaining some independent advantages: 1. Since the callers' experience will be like that of the intended application, albeit with human speech understanding, the customer impact will be at least as good. In fact, it is even possible to issue a kind of guarantee against maltreatment of customers by instructing the wizards to take over calls that become problematic (this is further discussed in Section 4). 2. Since real customers are involved, no roleplaying from the point of view of the subjects takes place, and hence the data become highly realistic. 3. The fact that scenarios are superfluous-or even run counter to the goal of the data collection-means that the main source of methodological problems disappears, and that the data collection as such is considerably simplified compared to traditional Wizard-of-Oz. 4. By letting service agents be wizards, we move away even further from role-playing, given that the interaction metaphor in speech-enabled call routing is natural-language dialogue with a (general-skills) service agent.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Wizard-of-Oz",
"sec_num": "3.3"
},
{
"text": "5. Service agents possess the expertise necessary for a call-routing wizard: they know when additional information is required from the caller, when a call is ready for routing, and where to actually route the call. Hence, wizard guidelines and training become less complex than in traditional Wizard-of-Oz. 4 6. Service agents have excellent skills in dealing with customers. Hence, during the data collection they will be able to provide valuable feedback on dialogue and prompt design that can be carried over to the intended application. In spite of these advantages, Wizard-of-Oz appears to have been used only very rarely for collecting call-routing data. The sole such data collection that we are aware of was made for the original AT&T \"How May I Help you\" system (Gorin et al. 1997; Ammicht et al. 1999) . The one disadvantage of the Wizard-of-Oz approach is that it is more laborious than automated solutions, mainly because several person months of wizard work is required. On the other hand, as we have seen, it is still less laborious than a traditional Wizard-of-Oz, since there are no scenarios and since wizard guidelines can be kept simple.",
"cite_spans": [
{
"start": 772,
"end": 791,
"text": "(Gorin et al. 1997;",
"ref_id": "BIBREF6"
},
{
"start": 792,
"end": 812,
"text": "Ammicht et al. 1999)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Wizard-of-Oz",
"sec_num": "3.3"
},
{
"text": "Our data-collection application consists of two parts: The first part is the Prompt Piano Client (PPC), which is running on the service agent's PC. This is essentially a GUI with \"keys\" corresponding to prerecorded prompts by which the wizard interacts with the caller, thereby simulating the intended system. The PPC interface is shown in PLATE 1. The second part is the Prompt Piano Server (PPS), which is an IVR (interactive voice response) server with a Dialogic telephony board, running Envox, Nuance and Dialogic software. This handles playing of prompts as well as recording of calls. Two kinds of recordings are made: call logs (that is, the callers' speech events as detected by the Nuance speech recognizer) and complete dialogues (\"open mic\").",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data-collection application",
"sec_num": "4"
},
{
"text": "To set up a data collection, the contact center solution is modified so that a percentage of the incoming calls to the customer care is diverted to the PPS. The PPS in turn transfers each call to a wizard (that is, to a PPC) using tromboning.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data-collection application",
"sec_num": "4"
},
{
"text": "Allocation of the wizards is performed by the Telia CallGuide contact center platform using skillbased routing. Whenever a wizard answers a call, two audio streams are established, one from the customer to the wizard so that she can hear the customer's speech, and one from an audio source in the PPS to the customer. An initial open prompt is played automatically by the PPS, and the wizard is then free to start playback of prompts. This is realized by sending control messages from the PPC to the audio source on the PPS via TCP/IP, while listening to the customer throughout.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data-collection application",
"sec_num": "4"
},
{
"text": "Depending on the caller's response, different things will happen: If the caller provides an unambiguous description of the issue, the wizard will transfer the call to the correct queue and end the recording by pressing the \"end / route customer\" button. This signals to the PPS that the call should be released using the Explicit Call Transfer (ECT) supplementary service, freeing the two channels used for the tromboned call in the PPS.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data-collection application",
"sec_num": "4"
},
{
"text": "If, on the other hand, the caller does not provide an unambiguous description of the issue, the wizard will play a follow-up prompt aimed at getting more information from the caller by choosing from the buttons/prompts situated to the right (fields II and III of the GUI; see Plate 1). These parts of the GUI are fully configurable; the number and layout of buttons as well as the names of sound files for the corresponding prompts are declared separately. (Declarations include specifying whether the prompt associated with a particular button allows barge-in or not.) Thus, it is possible not just to vary individual prompts, but also to simulate call-routing dialogues to various depths by varying the number of buttons/prompts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data-collection application",
"sec_num": "4"
},
{
"text": "Apart from routing the call, a possible action of the wizard is to enter into the call. This is realized by establishing a two-way direct audio stream with the customer, enabling the parties to talk to each other. As pointed out in Section 3.3, one purpose of this is to let wizards take over calls that are problematic, thereby making sure that callers do not get maltreated during the data collection and reducing the risk that they hang up. A similar functionality was available in the data-collection application for AT&Ts \"How May I Help You\" system (Walker et al. 2000) .",
"cite_spans": [
{
"start": 555,
"end": 575,
"text": "(Walker et al. 2000)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data-collection application",
"sec_num": "4"
},
{
"text": "The Prompt Piano Client interface as configured towards the end of the data collection. The interface is divided into three fields with buttons. I: The leftmost field provides caller information, like A-nr (the phone number the customer is calling from) and Cid (the phone number the customer provides as reason for the call). The wizard has two option buttons, Mina \u00e5tg\u00e4rder ('my actions'), at hand: the button Bryt in / Prata med kund ('barge-in/talk to client') which is used for entering into the call, and the button Avsluta / Koppla kund ('end/route customer') which is used to terminate the recording prior to routing the call to the appropriate destination. (Both of these options are associated with prompts being played.) II: The second field, Kunden\u2026 ('the customer\u2026'), contains buttons corresponding to renewed open prompts for the purpose of error-handling, \u2026\u00e4r tyst ('\u2026 is silent'), \u2026trycker p\u00e5 knappar ('uses the touch-tone keypad'), \u2026ber om hj\u00e4lp ('asks for help'), \u2026avbryter ('interrupts'), \u2026pratar f\u00f6r l\u00e4nge ('talks for too long'), \u2026s\u00e4ger inget om \u00e4rendet ('doesn't say anything about the reason for the call'), \u2026\u00e4r sv\u00e5r att uppfatta ('is hard to understand'). III: The third field, Jag undrar om det g\u00e4ller\u2026 ('I would like to know if it is about\u2026'), contains buttons corresponding to disambiguation prompts asking for additional information, e.g. whether the customer's reason for the call is about fixed ('fast') or mobile ('mobilt') telephony, broadband ('bredband') or something else. All buttons also have hot-key possibilities for agents who prefer this over point-and-click.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PLATE 1:",
"sec_num": null
},
{
"text": "With the exception of the initial open prompt, the wizards have full control over when and in what order prompts are played and actions are executed. Thus, whereas an automated system will start playing the next prompt after an end-of-speech timeout typically within the range of 0.75-1.5 seconds, a wizard may decide to impose longer delays if she considers that the caller has not yet yielded the turn. On the other hand, the wizard may also respond more rapidly. Thus, the problem of response delays, which has sometimes had distorting impact in Wizard-of-Oz simulations, does not appear in our application (cf. Oviatt et al. 1992) .",
"cite_spans": [
{
"start": 615,
"end": 634,
"text": "Oviatt et al. 1992)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "PLATE 1:",
"sec_num": null
},
{
"text": "The PPS application was developed in the Envox graphical scripting language, which makes it possible to write event-driven applications controlled from an external source such as the PPC, and also supports Nuance call logging (for recording customer utterances) and Dialogic transaction recording (for recording entire conversations between two parties, in this case the customer and the PPS, or the customer and the wizard). 5 Design, implementation and testing of the Prompt Piano (PPC and PPS) took four person weeks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PLATE 1:",
"sec_num": null
},
{
"text": "The agents/wizards were involved in the development from the very start to ensure that the application (and in particular the GUI) was 5 VXML was not used since it appeared that real-time control of an IVR from an external source would then have been more difficult to implement. Furthermore, VXML browsers generally have no support for features such as transaction recording during tromboned transfer and delayed invocation of the ECT supplementary service in conjunction with call transfer. Hence, in a VXML framework, additional components would be required to solve these tasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PLATE 1:",
"sec_num": null
},
{
"text": "optimized according to their needs and wishes. The Prompt Piano GUI was reconfigured several times during the course of the data collection, both for the purpose of carrying out prompt-design experiments and in response to (individual or group) requests for changes by the agents/wizards.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PLATE 1:",
"sec_num": null
},
{
"text": "The purpose of the data collection was twofold: to obtain speech data that could be used for initial training of the speech recognizer, and to obtain data that could be used to guide dialogue design of the intended application. Thus, whereas the former only involved caller responses to open prompts, the latter required access to complete call-routing dialogues, including error-handling and disambiguation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview",
"sec_num": "5.1"
},
{
"text": "Organization. Ten wizards were used for the data collection. Initially, one week was used for training of the wizards and basic tuning of the prompts. This process required four person weeks (not all wizards were present all the time). After a break of three weeks, the data collection then went on for five weeks in a row, with the ten wizards acquiring around 42,000 call-routing dialogues. (This figure includes around 2,000 useable dialogues that were collected during the initial week.) This was more than had been anticipated, and much more than the 25,000 that had been projected as a minimum for training, tuning and evaluation of the speech recognizer. Thus, although 50 person weeks were used by the wizards for the actual collection, 32 person weeks would actually have been sufficient to reach the minimum of 25,000 dialogues. On average, 195 dialogues were collected per working day per wizard (mean values ranging from 117 dialogues per day to 317 dialogues per day; record for a wizard on a single day was 477).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview",
"sec_num": "5.1"
},
{
"text": "Barge-in. Initially, barge-in was allowed for all prompts. However, it turned out to be useful to have one very short prompt with barge-in disabled, just asking the caller to state the reason for the call. The main usage of this was in cases where callers were repeatedly barging in on the system to the extent that the system could not get its message through.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview",
"sec_num": "5.1"
},
{
"text": "Utterance fragments. As a consequence of, on the one hand, wizards having full control over when and whether to start playing a prompt and, on the other hand, the speech recognizer having a fixed end-of-speech timeout, it would sometimes happen that more than one sound file would be recorded between two prompts in the Nuance call logs. An example of this would be: \"Eeh, I... I'm wondering whether... can you tell me the pricing of broadband subscriptions?\", where both of the two silent pauses would trigger the end-of-speech timeout. Although this constitutes a mismatch between data collection and final system, in practice this caused no problem: on the contrary, the sound files were simply treated as separate utterances for the purpose of training the speech recognizer, which means that the most informative fragment, typically at the end, was not lost. In addition, these data are potentially very valuable for research on turn-taking (in effect, intelligent end-of-speech detection).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview",
"sec_num": "5.1"
},
{
"text": "Wizards entering into calls. The event of wizards taking over calls in order to sort out problematic dialogues occurred on average in 5% of the calls. The figure was initially a bit higher, presumably because the wizards were less skillful in using the prompts available, and because the prompts were less well-developed. As a side-effect of this, we have obtained potentially very valuable data for error-handling, with both human-machine and human-human data for the same callers and issues (compare Walker et al., 2000) .",
"cite_spans": [
{
"start": 502,
"end": 522,
"text": "Walker et al., 2000)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Overview",
"sec_num": "5.1"
},
{
"text": "Post-experimental interviews. We also used the facility of letting wizards take over calls as a way of conducting post-experimental interviews. This was achieved by having wizards route the calls to themselves and then handle the issue, whereupon the wizard would ask the caller if they would accept being interviewed. In this way, we were able to assess customer satisfaction on the fly with respect to the intended system and even getting user feedback on specific design features already during the data collection.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview",
"sec_num": "5.1"
},
{
"text": "Several design experiments were run during the data collection. Here, we shall only very briefly describe one of them, in which we compared two styles of disambiguation prompts, one completely open and one more directed. As can be seen in ?') , and the open prompt ('Could you please tell me a little bit more about the reason for you call?') prompts. Totals and ratios are given for utterances/words, disfluencies and number of concepts acquired before the disambiguation prompt was played (\"In\") and after the customer had replied to the disambiguation prompt (\"Out\"). Also, ratios are given for number of concepts compared to number of utterances and words, as well as totals and ratios for the differences (DIFFs) between concepts in and concepts out, i.e., how many concepts you \"win\" by asking the disambiguation prompt.",
"cite_spans": [],
"ref_spans": [
{
"start": 239,
"end": 242,
"text": "?')",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5.2"
},
{
"text": "Furthermore, in order to see to what extent these prompts also made callers provide more information, we manually tagged the transcribed utterances with semantic categories. Following the evaluation methodology suggested by Boye & Wir\u00e9n (2007, Section 5) , we then computed the difference with respect to \"concepts\" for utterances immediately following and preceding the two kinds of prompts.",
"cite_spans": [
{
"start": 224,
"end": 254,
"text": "Boye & Wir\u00e9n (2007, Section 5)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5.2"
},
{
"text": "Although the number of concepts gained is only slightly higher 6 for the open prompt (as a function of concepts per utterance), there are some palpable differences between the directed and the open prompt. One, shown in TABLE 1, is that there are no instances where an already instantiated concept (e.g. fixedTelephony) is changed to something else (e.g. broadband), while this happens 18 times following the open prompt. The other, not shown in TABLE 1, is that, following the directed prompt, one never \"gains\" more than one new concept, while there are 26 instances following the open prompt where the gain is two concepts, and even two instances where the gain is three concepts (which also means that one concept is changed).",
"cite_spans": [
{
"start": 63,
"end": 64,
"text": "6",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5.2"
},
{
"text": "Finally, when one analyses the syntactic characteristics following the two different types of prompts, there is an obvious shift from the telegraphic \"noun-only\" responses that amount to more than 70% of the directed prompt responses, to the responses following the open prompt, where 40% are complete sentences and 21% are noun phrases. Also, the syntax is more varied following the open prompt. 7",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5.2"
},
{
"text": "6 However, the difference is not statistically significant, either using a t test (two-sampled, two-tailed: p=0.16 with equal variances assumed; p=0.158 equal variances not assumed) or Mann-Whitney U test (two-tailed: p=0.288). 7 The distributions are, in descending order, for the directed prompt:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5.2"
},
{
"text": "Noun=85, Sentence=11, Yes/No=8, Noun Phrase=8, no response=3, Yes/No+Noun=2, Adverbial Phrase=1, Adjective Phrase=1; for the open prompt: Sentence=49, Noun Phrase=26, Noun=24, Verb",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5.2"
},
{
"text": "We claimed in Section 3.3 that by using an inservice Wizard-of-Oz data collection, we have been able to effectively overcome all problems of the alternative methods discussed there. A relevant question is then if there are any remaining, independent problems of the approach described here. On the methodological side, there is clearly a certain amount of role playing left in the sense that service agents are acting as the system (albeit a system whose interaction metaphor is a service agent!). Interestingly, we noticed early on that the agents sometimes failed in properly simulating the intended system in one respect: Since they would often grasp what the caller wanted before he or she had finished speaking, they would start playing the next prompt so early that they were barging in on the caller. Thus, in their willingness to provide quick service, they were stepping outside of their assigned role. However, they soon learnt to avoid this, and it was never a problem except for the first few days.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "Apart from this, the main disadvantage of Wizard-of-Oz collections clearly is the amount of work involved compared to the other methods. As we have seen, the Prompt Piano design and implementation took four person weeks, training of the wizards took another four person weeks, and collection of 25,000 dialogues required 32 person weeks-hence altogether 40 person weeks (although we actually used 50 person weeks, since we went on collecting more data). This could be compared with possibly a single person week required for the fully automated approach. The more Phrase=11, Adjective Phrase=5, Adverbial Phrase=2, no response=2, Yes/No=1, Interjection=1. elaborate automated methods would come somewhere in between, also depending on whether a human agent is used for routing callers or not.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "In the TeliaSonera case, the main desiderata favouring Wizard-of-Oz were highly representative data, no negative customer impact and need for early evaluation and design, particularly because this was the first deployment of natural-language call routing in Scandinavia. In other words, it was decided to accept a higher initial cost in return for reduced costs downstream, due to higher quality and less re-design of the implemented system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "It is impossible to quantify the downstream savings made by choosing Wizard-of-Oz since we have no baseline. However, one indication of the quality of the data is the initial performance of the classifier of the deployed system. (By \"initial\", we mean the period during which no data from the live system had yet been used for training or updating of the system.) In our case, the initial accuracy was 75%, using 113 application categories. We regard this as a high figure, also considering that it was achieved in spite of several new products having been introduced in the meantime that were not covered by the speech recognizer. The initial training of the speech recognizer and classifier used 25,000 utterances. As a comparison, when an additional 33,000 utterances (mostly from the live system) had been used for training, the accuracy increased to 85%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "For backgrounds on Wizard-of-Oz methodology, seeDahlb\u00e4ck et al. (1993) andFraser & Gilbert (1991).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "TeliaSonera (www.teliasonera.com) is the largest telco in the Scandinavian -Baltic region.3 The speech recognizer and classifier are delivered by Nuance (www.nuance.com).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Furthermore, as a side effect, it is possible to facilitate the subsequent process of manually tagging the data by keeping track of where each call is routed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Many colleagues have provided invaluable help and support throughout this project. Here we can only mention some of them: Johan Boye, Joakim Gustafson, Linda Bell, Fredrik Bystr\u00f6m, Robert Sandberg, Erik N\u00e4slund, Erik Demmelmaier, Viktoria Ahlbom, Inger Thall and Marco Petroni. Last but not least we thank our skilled wizards: Christina Carlson, Marie Hagdorn, Gunilla Johannisson, Ana-Maria Loriente, Maria Mellgren, Linda Norberg, Anne T\u00e4rk, Mikael Wikner, Eva Wintse and Jeanette \u00d6berg.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Communicative Activity Analysis of a Wizard of Oz Experiment",
"authors": [
{
"first": "Jens & Bj\u00f6rn",
"middle": [],
"last": "Allwood",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Haglund",
"suffix": ""
}
],
"year": 1992,
"venue": "PLUS ESPRIT project P5254",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Allwood, Jens & Bj\u00f6rn Haglund. 1992. Communicative Activity Analysis of a Wizard of Oz Experiment. Internal Report, PLUS ESPRIT project P5254.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Knowledge Collection For Natural Language Spoken Dialog Systems",
"authors": [
{
"first": "Egbert",
"middle": [],
"last": "Ammicht",
"suffix": ""
},
{
"first": "Allen",
"middle": [],
"last": "Gorin & Tirso",
"suffix": ""
},
{
"first": "Alonso",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 1999,
"venue": "Proc. Eurospeech",
"volume": "3",
"issue": "",
"pages": "1375--1378",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ammicht, Egbert, Allen Gorin & Tirso Alonso. 1999. Knowledge Collection For Natural Language Spoken Dialog Systems. Proc. Eurospeech, Budapest, Hungary, Volume 3, pp. 1375-1378.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Multi-slot semantics for natural-language call routing systems",
"authors": [
{
"first": "Johan & Mats",
"middle": [],
"last": "Boye",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wir\u00e9n",
"suffix": ""
}
],
"year": 2007,
"venue": "Proc. Bridging the Gap: Academic and Industrial Research in Dialog Technology. NAACL Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Boye, Johan & Mats Wir\u00e9n. 2007. Multi-slot semantics for natural-language call routing systems. Proc. Bridging the Gap: Academic and Industrial Research in Dialog Technology. NAACL Workshop, Rochester, New York, USA.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Wizard of Oz Studies -Why and How",
"authors": [
{
"first": "Nils",
"middle": [],
"last": "Dahlb\u00e4ck",
"suffix": ""
},
{
"first": "Arne",
"middle": [],
"last": "J\u00f6nsson & Lars Ahrenberg",
"suffix": ""
}
],
"year": 1993,
"venue": "Readings in Intelligent User Interfaces",
"volume": "6",
"issue": "",
"pages": "258--266",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dahlb\u00e4ck, Nils, Arne J\u00f6nsson & Lars Ahrenberg, Wizard of Oz Studies -Why and How. 1993. Knowledge-Based Systems, vol. 6, no. 4, pp. 258- 266. Also in: Mark Maybury & Wolfgang Wahlster (eds.). 1998. Readings in Intelligent User Interfaces, Morgan Kaufmann.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Automated Wizard-of-Oz for Spoken Dialogue Systems",
"authors": [
{
"first": "Di",
"middle": [],
"last": "Fabbrizio",
"suffix": ""
},
{
"first": "Gokhan",
"middle": [],
"last": "Giuseppe",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Tur & Dilek Hakkani-T\u00fcr",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. Interspeech",
"volume": "",
"issue": "",
"pages": "1857--1860",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Di Fabbrizio, Giuseppe, Gokhan Tur & Dilek Hakkani-T\u00fcr. 2005. Automated Wizard-of-Oz for Spoken Dialogue Systems. Proc. Interspeech, Lisbon, Portugal, pp. 1857-1860.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Simulating speech systems",
"authors": [
{
"first": "Norman",
"middle": [
"M & G"
],
"last": "Fraser",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Gilbert",
"suffix": ""
}
],
"year": 1991,
"venue": "Computer Speech and Language",
"volume": "5",
"issue": "",
"pages": "81--99",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fraser, Norman M. & G. Nigel Gilbert. Simulating speech systems. 1991. Computer Speech and Language, vol. 5, pp. 81-99.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "How may I help you",
"authors": [
{
"first": "A",
"middle": [
"L"
],
"last": "Gorin",
"suffix": ""
},
{
"first": "G",
"middle": [
"& J H"
],
"last": "Riccardi",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wright",
"suffix": ""
}
],
"year": 1997,
"venue": "Speech Communication",
"volume": "23",
"issue": "",
"pages": "113--127",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gorin, A. L., G. Riccardi & J. H. Wright. 1997. How may I help you? Speech Communication, vol. 23, pp. 113-127.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Pragmatic considerations in man-machine discourse",
"authors": [
{
"first": "Walther",
"middle": [],
"last": "Von Hahn",
"suffix": ""
}
],
"year": 1986,
"venue": "Proc. COLING",
"volume": "",
"issue": "",
"pages": "520--526",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "von Hahn, Walther. 1986. Pragmatic considerations in man-machine discourse. Proc. COLING, Bonn, Germany, pp. 520-526.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Multi-Site Data Collection and Evaluation in Spoken Language Understanding",
"authors": [
{
"first": "L",
"middle": [],
"last": "Hirschman",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Bates",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Dahl",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Fisher",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Garofolo",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Pallett",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Hunicke-Smith",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Price",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Rudnicky",
"suffix": ""
},
{
"first": "&",
"middle": [
"E"
],
"last": "Tzoukermann",
"suffix": ""
}
],
"year": 1993,
"venue": "Proc. ARPA Human Language Technology",
"volume": "",
"issue": "",
"pages": "19--24",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hirschman, L., M. Bates, D. Dahl, W. Fisher, J. Garofolo, D. Pallett, K. Hunicke-Smith, P. Price, A. Rudnicky & E. Tzoukermann. 1993. Multi-Site Data Collection and Evaluation in Spoken Language Understanding. Proc. ARPA Human Language Technology, Princeton, New Jersey, USA, pp. 19-24 .",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Martin Fong & Michael Frank. 1992. A rapid semi-automatic simulation technique for investigating interactive speech and handwriting",
"authors": [
{
"first": "Sharon",
"middle": [],
"last": "Oviatt",
"suffix": ""
},
{
"first": "Philip",
"middle": [],
"last": "Cohen",
"suffix": ""
}
],
"year": null,
"venue": "Proc. ICSLP",
"volume": "",
"issue": "",
"pages": "1351--1354",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Oviatt, Sharon, Philip Cohen, Martin Fong & Michael Frank. 1992. A rapid semi-automatic simulation technique for investigating interactive speech and handwriting. Proc. ICSLP, Banff, Alberta, Canada, pp. 1351-1354.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Learning to Predict Problematic Situations in a Spoken Dialogue System: Experiments with How May I Help You?",
"authors": [
{
"first": "Marilyn",
"middle": [],
"last": "Walker",
"suffix": ""
},
{
"first": "Irene",
"middle": [],
"last": "Langkilde",
"suffix": ""
},
{
"first": "Jerry",
"middle": [],
"last": "Wright",
"suffix": ""
},
{
"first": "Allen",
"middle": [],
"last": "Gorin & Diane Litman",
"suffix": ""
}
],
"year": 2000,
"venue": "Proc. North American Meeting of the Association for Computational Linguistics (NAACL)",
"volume": "",
"issue": "",
"pages": "210--217",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Walker, Marilyn, Irene Langkilde, Jerry Wright, Allen Gorin & Diane Litman. 2000. Learning to Predict Problematic Situations in a Spoken Dialogue System: Experiments with How May I Help You? Proc. North American Meeting of the Association for Computational Linguistics (NAACL), pp. 210-217.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Stephanie Seneff & Michael Soclof. 1989. The Collection and Preliminary Analysis of a Spontaneous Speech Database",
"authors": [
{
"first": "Victor",
"middle": [],
"last": "Zue",
"suffix": ""
},
{
"first": "Nancy",
"middle": [],
"last": "Daly",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Glass",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Goodine",
"suffix": ""
},
{
"first": "Hong",
"middle": [],
"last": "Leung",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Phillips",
"suffix": ""
},
{
"first": "Joseph",
"middle": [],
"last": "Polifroni",
"suffix": ""
}
],
"year": null,
"venue": "Proc. DARPA Speech and Natural Language Workshop",
"volume": "",
"issue": "",
"pages": "126--134",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zue, Victor, Nancy Daly, James Glass, David Goodine, Hong Leung, Michael Phillips, Joseph Polifroni, Stephanie Seneff & Michael Soclof. 1989. The Collection and Preliminary Analysis of a Spontaneous Speech Database. Proc. DARPA Speech and Natural Language Workshop, pp. 126-134.",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"html": null,
"type_str": "table",
"text": "Summary statistics for the directed prompt ('I need some additional information about the reason for your call. Is it for example about an order, price information or support",
"num": null,
"content": "<table><tr><td>, utterances following the open</td></tr><tr><td>disambiguation prompt are on average 3.6 times</td></tr><tr><td>longer than utterances following the directed</td></tr><tr><td>prompt.</td></tr></table>"
}
}
}
} |