File size: 64,420 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 |
{
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:28:21.366442Z"
},
"title": "Studying the Impact of Filling Information Gaps on the Output Quality of Neural Data-to-Text",
"authors": [
{
"first": "Craig",
"middle": [],
"last": "Thomson",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Aberdeen",
"location": {
"country": "UK"
}
},
"email": "c.thomson@abdn.ac.uk"
},
{
"first": "Zhijie",
"middle": [],
"last": "Zhao",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Aberdeen",
"location": {
"country": "UK"
}
},
"email": ""
},
{
"first": "Somayajulu",
"middle": [],
"last": "Gowri",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Aberdeen",
"location": {
"country": "UK"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "It is unfair to expect neural data-to-text to produce high quality output when there are gaps between system input data and information contained in the training text. Thomson et al. (2020) identify and narrow information gaps in Rotowire, a popular data-to-text dataset. In this paper, we describe a study which finds that a state-of-the-art neural data-to-text system produces higher quality output, according to the information extraction (IE) based metrics, when additional input data is carefully selected from this newly available source. It remains to be shown, however, whether IE metrics used in this study correlate well with humans in judging text quality.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "It is unfair to expect neural data-to-text to produce high quality output when there are gaps between system input data and information contained in the training text. Thomson et al. (2020) identify and narrow information gaps in Rotowire, a popular data-to-text dataset. In this paper, we describe a study which finds that a state-of-the-art neural data-to-text system produces higher quality output, according to the information extraction (IE) based metrics, when additional input data is carefully selected from this newly available source. It remains to be shown, however, whether IE metrics used in this study correlate well with humans in judging text quality.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The ecological validity (de Vries et al., 2020) of data-to-text tasks requires that tasks resemble, as closely as possible, real-world problems. Only if this is the case can neural data-to-text solutions be operationally deployed with confidence. In the context of data-to-text, one of the issues with ecological validity is that most real-world tasks involve sizeable input data, with longer and more complex texts than are found in 'toy-sized' datasets. We must be able to see a path to a machine learning task which closely resembles a real-world scenario and allows us to investigate important research questions. We should aim to improve both the dataset and the task continuously. Generating summaries of basketball games from tabled data with the Rotowire dataset, as introduced by Wiseman et al. (2017a) for the English language, moves us closer to an ecologically valid data-to-text task.",
"cite_spans": [
{
"start": 789,
"end": 811,
"text": "Wiseman et al. (2017a)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The original Rotowire dataset has been found to contain gaps between the information in the input data, and the information content of the training text. This makes the task unfair for evaluating neural data-to-text systems that are required to generate high quality output text, with hardly any factual errors. The SportSett:Basketball dataset (Thomson et al., 2020) addresses these data issues by fixing information gaps in the input data, whilst maintaining the original human-authored texts. Given that there is now at least an order of magnitude more data per game, we should consider which subset of data to train the system on, and what if any pre-processing should be preformed.",
"cite_spans": [
{
"start": 345,
"end": 367,
"text": "(Thomson et al., 2020)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We added some of this newly available data to an existing state-of-the-art neural data-to-text system (Rebuffel et al., 2020) and found improvement across a range of metrics. We used this system since, at the time of writing, it is one of the most recent, best performing and easiest to configure. We also discuss here which types of data could be added in the future, as well as some difficulties that may be encountered in doing so.",
"cite_spans": [
{
"start": 102,
"end": 125,
"text": "(Rebuffel et al., 2020)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Many systems have been designed and evaluated using the Rotowire dataset (Wiseman et al., 2017a; Puduppully et al., 2019a,b; Wang, 2019; Gong et al., 2019; Iso et al., 2019; Rebuffel et al., 2020) . Most of these works focus on adjusting the architecture of the system whilst using similar input data. Gong et al. (2019) is a notable exception, as their architecture change allows box score and other data from previous games in their input. This was, however, still data from the original Rotowire dataset.",
"cite_spans": [
{
"start": 73,
"end": 96,
"text": "(Wiseman et al., 2017a;",
"ref_id": "BIBREF15"
},
{
"start": 97,
"end": 124,
"text": "Puduppully et al., 2019a,b;",
"ref_id": null
},
{
"start": 125,
"end": 136,
"text": "Wang, 2019;",
"ref_id": "BIBREF14"
},
{
"start": 137,
"end": 155,
"text": "Gong et al., 2019;",
"ref_id": "BIBREF2"
},
{
"start": 156,
"end": 173,
"text": "Iso et al., 2019;",
"ref_id": "BIBREF4"
},
{
"start": 174,
"end": 196,
"text": "Rebuffel et al., 2020)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Some works have attempted to better align data to text in other datasets, with techniques such as semantic control (Du\u0161ek et al., 2019) . For Rotowire, this has been investigated by Wang (2019) , which aims to prevent generation of sentences not grounded in the data. With this approach, some of the most common sentence types from the human narrative, such as the subsequent opponents of each team, are not generated. This difference is crucial when determining ecological validity of the task. The aim is to replicate the human author in the act of writing a summary of the basketball game, including its full narrative structure.",
"cite_spans": [
{
"start": 115,
"end": 135,
"text": "(Du\u0161ek et al., 2019)",
"ref_id": "BIBREF1"
},
{
"start": 182,
"end": 193,
"text": "Wang (2019)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "3 Information Gaps in Data-to-Text Shown in Figure 1 is an example textual summary from the Rotowire dataset. An example partial box score is shown in Table 1 . There are numerous cases where information conveyed by the text is not present in the same form in the box score or other game data. These information gaps should be investigated in order to improve the machine learning task.",
"cite_spans": [],
"ref_spans": [
{
"start": 44,
"end": 52,
"text": "Figure 1",
"ref_id": null
},
{
"start": 151,
"end": 158,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "We performed a machine-assisted corpus analysis, using the spaCy syntactic parser (Honnibal, 2015) to group sentences which only differ by entity. We do this using an abstraction process where we replace named entities with special tokens comprised of their part-of-speech and entity label. Some manual rules are added to the parser to handle domain specific syntax. By this process the sentence (S01 from Figure 1 ) 'The Atlanta Hawks (41-9) beat the Washington Wizards (31-19) 105-96 on Wednesday.', is transformed to 'PROPN-ORG (X-Y) beat the PROPN-ORG (X-Y) X-Y on NOUN-DATE.'.",
"cite_spans": [
{
"start": 82,
"end": 98,
"text": "(Honnibal, 2015)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 406,
"end": 414,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "We then count and read these abstract sentence types to find statements common to the narrative, but with attribute types not present in the data. For example, sentences with the same abstract form as S1 occur 26 times in the training corpus, with more than 800 further sentences of a similar form (using defeat instead of beat, or also including the location/stadium). It is the most common type of thing to say in the opening sentence of these summaries (which teams played, when, and where). There are, however, important attributes in these sentences which are not provided by the original Rotowire data. When generating game summaries, systems will often hallucinate these attributes as they deem it probable that such language is included in the summary, but the attribute is not available to the copy attention mechanism. In the case of our above example, the day of the week is not available in the data. The stadium in which the game was played, as well as the city and/or state within which the stadium stands, are also not available despite being common in variants of this opening sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "In S02 and S11 from Figure 1 we notice that the games being discussed are not the game being summarised, they are previous or subsequent games for these teams. This is common in the training corpus as well. Handling data for previous games is complex (see subsection 3.2). However, data for the subsequent game can be easily obtained provided that a yearly partition scheme like that proposed by Thomson et al. (2020) is used. If such a partition scheme is not used, we cannot guarantee that a previous or subsequent game was not used to condition the system during training.",
"cite_spans": [
{
"start": 396,
"end": 417,
"text": "Thomson et al. (2020)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [
{
"start": 20,
"end": 28,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Missing Game Information",
"sec_num": "3.1"
},
{
"text": "We also see in S09, a mention of the conference/division structure of the league. These are known sets, which can change over time but are fixed within a season. In the NBA there are 2 conferences, each with 3 divisions of 5 teams.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Missing Game Information",
"sec_num": "3.1"
},
{
"text": "The hierarchical encoder of Rebuffel et al. (2020) takes as input a set of entities, where each entity is a set of 24 tuples 1 , and each tuple describes an attribute of that entity. An example entity would be a PLAYER, which might have attributes such as 'NAME-Kyrie Irving', 'POINTS-30', 'TEAM-Celtics', and 'REB-8'. If there are not 24 attributes of the entity then it is padded with 'NULL-NULL' tuples.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Missing Game Information",
"sec_num": "3.1"
},
{
"text": "To model attributes of the current and subsequent games, we include in our input data an additional entity of type GAME, as well as two additional entities of type NEXT-GAME. These entities were chosen because our machine-assisted corpus analysis highlighted that sentences about the game date and location, as well as those for upcoming games for each team, were common in the human-authored texts, but not supported by the original Rotowire data. The newly available data from SportSett allowed us to fill these gaps. In the two NEXT-GAME entities (one each for the two teams which are the focus of the current game summary), we include attributes for season, month, day of the week, stadium, capacity, and finally both team names plus their respective division and conference names. For the GAME entity, we include the same attributes as for NEXT-GAME, plus the attendance for the game (which was obviously not available for NEXT-GAME as those events have yet to take place).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Missing Game Information",
"sec_num": "3.1"
},
{
"text": "The Atlanta Hawks (41-9) beat the Washington Wizards (31-19) 105-96 on Wednesday.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S01:",
"sec_num": null
},
{
"text": "The Hawks bounced back after losing their first game of 2015, a 115-100 loss at the hands of the New Orleans Pelicans on Monday.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S02:",
"sec_num": null
},
{
"text": "Jeff Teague was Atlanta's top scorer against the Wizards, recording 26 points on 9-of-13 shooting from the field. Kyle Korver was kept in check with just six points in a team-high 37 minutes. S04:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S03:",
"sec_num": null
},
{
"text": "He helped get his teammates involved as he dished out six assists.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S03:",
"sec_num": null
},
{
"text": "Al Horford has put up at least 20 points and 10 rebounds in three of his last five games.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S05:",
"sec_num": null
},
{
"text": "The Wizards have now lost four straight, which is their longest losing streak of the season. S07:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S06:",
"sec_num": null
},
{
"text": "They have lost by single-digits in all four games, and are now 0-3 against Atlanta this season. S08:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S06:",
"sec_num": null
},
{
"text": "John Wall led Washington with 24 points and nine assists. Bradley Beal was coming off an 18-point, 11-rebound effort against Charlotte on Monday. S09:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S06:",
"sec_num": null
},
{
"text": "He proceeded to post 23 points in 39 minutes in this matchup of two of the Eastern Conference's top three teams. S10:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S06:",
"sec_num": null
},
{
"text": "Washington did a great job slowing down Korver, but it wasn't enough to get them the win. S11:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S06:",
"sec_num": null
},
{
"text": "Washington will take their losing streak to Charlotte on Thursday, while the Hawks will welcome the Golden State Warriors to town Friday. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "S06:",
"sec_num": null
},
{
"text": "In subsection 3.1 we described adding data which easily fit our chosen encoder. It is worth noting, however, that there is much more data available in the SportSett database, and it can be presented to our neural system in different formats. One argument is that we should take all the atomic data, along with, perhaps, its structure, then create an encoder which accepts data in that form. The atomic entities are players, which are grouped into teams, divisions, conferences and leagues. The atomic events are plays, the act of one or more players acquiring countable statistics. See Thomson et al. (2020) Section 2.1 for a more detailed analysis of entities and dimensions in this dataset. The NLG system would then be tasked with learning both the language, as well as the underlying mathematics for the statistics. Whilst we could include all atomic events as training data, this would greatly increase the size of each input sample. We could alternatively in-clude carefully selected aggregated forms of data, although creating rules to determine what should be included may be time consuming, and likely domain specific. There would also be many combinations of these derived attributes. The key question is should all possible attributes which are to be realised be available to copy attention, or, should all attributes be transformable from the atomic data? An approach combining these different types of input data could also be used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data of Varying Forms",
"sec_num": "3.2"
},
{
"text": "Examples of aggregated forms of data could be anything from the percentage of shots which were successful for a player/team, to the average points a player has scored over an arbitrary span of games. One common inclusion in summaries is the aggregated statistic 'double-double' 2 , where players are said to have recorded double-digits in exactly two of points, rebounds, assists, blocks and steals. Mentions of previous games in the summary frequently use phrases in the form 'X of his/their last Y'. This can be seen in S05 of Figure 1 where we learn that Al Holford has scored greater than 20 points, and recorded more than 10 rebounds in exactly 3 of his last 5 games. There would be an impractical quantity of combinations for 'X out of Y' based statements, even if Y had a maximum of 5-10. Other similar aggregations, such as 'scored a combined 60 points over his last 3 games.' compound this problem.",
"cite_spans": [],
"ref_spans": [
{
"start": 529,
"end": 537,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data of Varying Forms",
"sec_num": "3.2"
},
{
"text": "It is unclear whether models could learn such mathematical operations since even very large and powerful models, such as GPT-3 (Brown et al., 2020) , currently only demonstrate simple addition and subtraction. This important and difficult aspect of defining the problem requires further research, but is essential if we are to create a machine learning task which is ecologically valid. The original Rotowire dataset contains a mix of countable statistics (at the game level only) and derived statistics such as percentages.",
"cite_spans": [
{
"start": 127,
"end": 147,
"text": "(Brown et al., 2020)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data of Varying Forms",
"sec_num": "3.2"
},
{
"text": "The information extraction (IE) metrics of Relation Generation (RG), Content Selection (CS), and Content Ordering (CO), have been used extensively to compare systems operating on the Rotowire based datasets. These metrics are also key to the system design philosophy of Wiseman et al. (2017a) which aims to \"[exploit] the fact that post-hoc information extraction is significantly easier than generation itself.\" All of these metrics are based on IE models which learn to link statements in the text to tuples in the data. These models are trained on a corpus which has been automatically annotated using rules. Names or numbers in the text are linked to the possible data tuples which could represent them. During evaluation, tuples are predicted for the test text summaries, with a name, value and a type, e.g. 'Atlanta-96-TEAM-PTS'. These predicted tuples can be compared with the tuples in the data in order to determine whether facts predicted from the text match those in the data. For example, if the fact extracted from the text is 'Atlanta-96-TEAM-PTS', but in the data we see 'Atlanta-105-TEAM-PTS', then based on the match of both name and type, we can determine that the number is wrong.",
"cite_spans": [
{
"start": 270,
"end": 292,
"text": "Wiseman et al. (2017a)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation by Information Extraction",
"sec_num": "4"
},
{
"text": "The CS metrics use these tuples to measure how many of the tuples from the predicted text exist in the gold standard text. The CO metric measures the order of these tuples. For more details see (Wiseman et al., 2017b) (we have purposefully cited the arXiv version of this paper as it includes an additional appendix detailing the procedure).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation by Information Extraction",
"sec_num": "4"
},
{
"text": "We extend the IE based metrics using the data now available in Thomson et al. (2020) . Details can be found on GitHub 3 . We make two modifications:",
"cite_spans": [
{
"start": 63,
"end": 84,
"text": "Thomson et al. (2020)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Extended IE Metrics",
"sec_num": "4.1"
},
{
"text": "\u2022 Extend the annotation logic such that it can detect the additional entities and attributes we added in subsection 3.1. For example, days of the week for both the current game and the subsequent game for each team.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extended IE Metrics",
"sec_num": "4.1"
},
{
"text": "\u2022 Use a season-based partition scheme so that the IE model is not being used to evaluate data upon which it was previously conditioned. We use the 2014, 2015 and 2016 seasons to train, 2017 to validate, and 2018 to test. This is the same problem in the partition scheme which was identified for the text generation system by Thomson et al. (2020) .",
"cite_spans": [
{
"start": 325,
"end": 346,
"text": "Thomson et al. (2020)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Extended IE Metrics",
"sec_num": "4.1"
},
{
"text": "5 Experimental Setup",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extended IE Metrics",
"sec_num": "4.1"
},
{
"text": "We created two datasets:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NLG System Setup",
"sec_num": "5.1"
},
{
"text": "D1: Where we emulated as closely as possible the data format and content used by Rebuffel et al. (2020) except using season-based partitions.",
"cite_spans": [
{
"start": 81,
"end": 103,
"text": "Rebuffel et al. (2020)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "NLG System Setup",
"sec_num": "5.1"
},
{
"text": "D2: Keeping all data from D1, but adding a new entity for the GAME, and two for the NEXT-GAME as detailed in subsection 3.1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NLG System Setup",
"sec_num": "5.1"
},
{
"text": "We then trained models using the system of Rebuffel et al. (2020) on each dataset, with 10 different random seeds, to determine whether adding the additional information improved the results. We also tested whether changing the early-stop strategy impacted the results, taking a snapshot from training using each of the best BLEU (Papineni et al., 2002) , RG, CS-PREC (precision), CS-REC (recall), and CO. To summarise, we used two datasets, with 10 random seed each, all with 4 different early-stop strategies for 80 models total (2*10*4). We then calculated BLEU, RG, CS-PREC, CS-REC, and CO for each model.",
"cite_spans": [
{
"start": 330,
"end": 353,
"text": "(Papineni et al., 2002)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "NLG System Setup",
"sec_num": "5.1"
},
{
"text": "We trained IE models following the general procedure proposed by (Wiseman et al., 2017b) . We trained with different random seeds and learning rates, then chose the best 3 LSTM and the best 3 Convolutional models to ensemble. We then used the model ensemble, as well as BLEU-4 (Papineni et al., 2002) , to evaluate the NLG system itself. Table 2 shows the results of our evaluation. We find a statistically significant difference (p < 0.005) for all information extraction based metrics (RG, CS-PREC, CS-REC, and CO) when we add the additional information as described in subsection 3.1. Information extraction based metrics increased in all cases when adding data, regardless of earlystopping method. Whilst BLEU scores also appeared to increase, we did not find the changes in them to be statistically significant. This is not surprising given that BLEU is known to not correlate for NLG (Reiter and Belz, 2009; Reiter, 2018) , and even in machine translation it only correlates when differences are large (Mathur et al., 2020) .",
"cite_spans": [
{
"start": 65,
"end": 88,
"text": "(Wiseman et al., 2017b)",
"ref_id": "BIBREF16"
},
{
"start": 277,
"end": 300,
"text": "(Papineni et al., 2002)",
"ref_id": "BIBREF6"
},
{
"start": 890,
"end": 913,
"text": "(Reiter and Belz, 2009;",
"ref_id": "BIBREF11"
},
{
"start": 914,
"end": 927,
"text": "Reiter, 2018)",
"ref_id": "BIBREF10"
},
{
"start": 1008,
"end": 1029,
"text": "(Mathur et al., 2020)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [
{
"start": 338,
"end": 345,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Automated Metric Setup",
"sec_num": "5.2"
},
{
"text": "Our results show that identifying data which should be included, then modelling it within the system architecture, increased all information extraction based metrics. Existing metrics have only been evaluated in limited ways for this domain. Improved metrics could help us evaluate systems, as well as find and categorise information gaps.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Future Work",
"sec_num": "7"
},
{
"text": "The subset of data selected for input, the form it takes (atomic versus aggregated), as well as the inclusion of system components/techniques (copy attention mechanism, hierarchical encoder, separate document plan, fact grounding, etc.), are all variables which could affect system performance. We plan in future work to perform ablation studies to determine which such variables, and in which combination, produce the best results. As part of this, we aim to create a unified code-base which will allow for components to be selected and configured in combination, for as many different data forms as possible.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Future Work",
"sec_num": "7"
},
{
"text": "Beyond this, we hope to move away from end-toend system designs. This is similar in spirit to the idea proposed in Puduppully et al. (2019a) , where a single model is not attempting to learn everything, the document plan is learned separately. We would extend such ideas to the data itself, if we can use both the known ontology from the structured data, as well as relationships and other information extracted with NLU or other tools, then this additional information could be input to systems which realize the language, meaning they are not left to solve both data analytic and language problems with a single model. If we can define our data operations in terms of standard data models, such as relational models, then we will be closer to a general approach for filling the information gap in data-to-text.",
"cite_spans": [
{
"start": 115,
"end": 140,
"text": "Puduppully et al. (2019a)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Future Work",
"sec_num": "7"
},
{
"text": "This is configurable in the encoder of Rebuffel et al. (2020), although we did not change it.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://en.wikipedia.org/wiki/ Double-double",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/nlgcat/adding_data",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We would like to thank our reviewers for their insightful feedback and questions.The work presented here is partially funded by the Engineering and Physical Sciences Research Council (EPSRC), which funds Craig Thomson under a National Productivity Investment Fund Doctoral Studentship (EP/R512412/1).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF1": {
"ref_id": "b1",
"title": "Semantic noise matters for neural natural language generation",
"authors": [
{
"first": "Ond\u0159ej",
"middle": [],
"last": "Du\u0161ek",
"suffix": ""
},
{
"first": "David",
"middle": [
"M"
],
"last": "Howcroft",
"suffix": ""
},
{
"first": "Verena",
"middle": [],
"last": "Rieser",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 12th International Conference on Natural Language Generation",
"volume": "",
"issue": "",
"pages": "421--426",
"other_ids": {
"DOI": [
"10.18653/v1/W19-8652"
]
},
"num": null,
"urls": [],
"raw_text": "Ond\u0159ej Du\u0161ek, David M. Howcroft, and Verena Rieser. 2019. Semantic noise matters for neural natural lan- guage generation. In Proceedings of the 12th Inter- national Conference on Natural Language Genera- tion, pages 421-426, Tokyo, Japan. Association for Computational Linguistics.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Table-to-text generation with effective hierarchical encoder on three dimensions (row, column and time)",
"authors": [
{
"first": "Xiaocheng",
"middle": [],
"last": "Heng Gong",
"suffix": ""
},
{
"first": "Bing",
"middle": [],
"last": "Feng",
"suffix": ""
},
{
"first": "Ting",
"middle": [],
"last": "Qin",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "3143--3152",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1310"
]
},
"num": null,
"urls": [],
"raw_text": "Heng Gong, Xiaocheng Feng, Bing Qin, and Ting Liu. 2019. Table-to-text generation with effective hier- archical encoder on three dimensions (row, column and time). In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Process- ing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3143-3152, Hong Kong, China. Association for Computational Linguistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Learning to select, track, and generate for data-to-text",
"authors": [
{
"first": "Hayate",
"middle": [],
"last": "Iso",
"suffix": ""
},
{
"first": "Yui",
"middle": [],
"last": "Uehara",
"suffix": ""
},
{
"first": "Tatsuya",
"middle": [],
"last": "Ishigaki",
"suffix": ""
},
{
"first": "Hiroshi",
"middle": [],
"last": "Noji",
"suffix": ""
},
{
"first": "Eiji",
"middle": [],
"last": "Aramaki",
"suffix": ""
},
{
"first": "Ichiro",
"middle": [],
"last": "Kobayashi",
"suffix": ""
},
{
"first": "Yusuke",
"middle": [],
"last": "Miyao",
"suffix": ""
},
{
"first": "Naoaki",
"middle": [],
"last": "Okazaki",
"suffix": ""
},
{
"first": "Hiroya",
"middle": [],
"last": "Takamura",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "2102--2113",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1202"
]
},
"num": null,
"urls": [],
"raw_text": "Hayate Iso, Yui Uehara, Tatsuya Ishigaki, Hiroshi Noji, Eiji Aramaki, Ichiro Kobayashi, Yusuke Miyao, Naoaki Okazaki, and Hiroya Takamura. 2019. Learning to select, track, and generate for data-to-text. In Proceedings of the 57th Annual Meeting of the Association for Computational Lin- guistics, pages 2102-2113, Florence, Italy. Associa- tion for Computational Linguistics.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Tangled up in BLEU: Reevaluating the evaluation of automatic machine translation evaluation metrics",
"authors": [
{
"first": "Nitika",
"middle": [],
"last": "Mathur",
"suffix": ""
},
{
"first": "Timothy",
"middle": [],
"last": "Baldwin",
"suffix": ""
},
{
"first": "Trevor",
"middle": [],
"last": "Cohn",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "4984--4997",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.448"
]
},
"num": null,
"urls": [],
"raw_text": "Nitika Mathur, Timothy Baldwin, and Trevor Cohn. 2020. Tangled up in BLEU: Reevaluating the eval- uation of automatic machine translation evaluation metrics. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 4984-4997, Online. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Bleu: a method for automatic evaluation of machine translation",
"authors": [
{
"first": "Kishore",
"middle": [],
"last": "Papineni",
"suffix": ""
},
{
"first": "Salim",
"middle": [],
"last": "Roukos",
"suffix": ""
},
{
"first": "Todd",
"middle": [],
"last": "Ward",
"suffix": ""
},
{
"first": "Wei-Jing",
"middle": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the 40th annual meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "311--318",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic eval- uation of machine translation. In Proceedings of the 40th annual meeting of the Association for Compu- tational Linguistics, pages 311-318.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Data-to-text generation with content selection and planning",
"authors": [
{
"first": "Ratish",
"middle": [],
"last": "Puduppully",
"suffix": ""
},
{
"first": "Li",
"middle": [],
"last": "Dong",
"suffix": ""
},
{
"first": "Mirella",
"middle": [],
"last": "Lapata",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 33rd AAAI Conference on Artificial Intelligence",
"volume": "33",
"issue": "",
"pages": "6908--6915",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ratish Puduppully, Li Dong, and Mirella Lapata. 2019a. Data-to-text generation with content selec- tion and planning. In Proceedings of the 33rd AAAI Conference on Artificial Intelligence, volume 33, pages 6908-6915, Honolulu, Hawaii.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Data-to-text generation with entity modeling",
"authors": [
{
"first": "Ratish",
"middle": [],
"last": "Puduppully",
"suffix": ""
},
{
"first": "Li",
"middle": [],
"last": "Dong",
"suffix": ""
},
{
"first": "Mirella",
"middle": [],
"last": "Lapata",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "2023--2035",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1195"
]
},
"num": null,
"urls": [],
"raw_text": "Ratish Puduppully, Li Dong, and Mirella Lapata. 2019b. Data-to-text generation with entity model- ing. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2023-2035, Florence, Italy. Association for Computational Linguistics.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A hierarchical model for data-to-text generation",
"authors": [
{
"first": "Cl\u00e9ment",
"middle": [],
"last": "Rebuffel",
"suffix": ""
},
{
"first": "Laure",
"middle": [],
"last": "Soulier",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [],
"last": "Scoutheeten",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Gallinari",
"suffix": ""
}
],
"year": 2020,
"venue": "Advances in Information Retrieval",
"volume": "",
"issue": "",
"pages": "65--80",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cl\u00e9ment Rebuffel, Laure Soulier, Geoffrey Scoutheeten, and Patrick Gallinari. 2020. A hierarchical model for data-to-text generation. In Advances in Information Retrieval, pages 65-80, Cham. Springer International Publishing.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A structured review of the validity of BLEU",
"authors": [
{
"first": "Ehud",
"middle": [],
"last": "Reiter",
"suffix": ""
}
],
"year": 2018,
"venue": "Computational Linguistics",
"volume": "44",
"issue": "3",
"pages": "393--401",
"other_ids": {
"DOI": [
"10.1162/coli_a_00322"
]
},
"num": null,
"urls": [],
"raw_text": "Ehud Reiter. 2018. A structured review of the validity of BLEU. Computational Linguistics, 44(3):393- 401.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "An investigation into the validity of some metrics for automatically evaluating natural language generation systems",
"authors": [
{
"first": "Ehud",
"middle": [],
"last": "Reiter",
"suffix": ""
},
{
"first": "Anja",
"middle": [],
"last": "Belz",
"suffix": ""
}
],
"year": 2009,
"venue": "Computational Linguistics",
"volume": "35",
"issue": "4",
"pages": "529--558",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ehud Reiter and Anja Belz. 2009. An investigation into the validity of some metrics for automatically evalu- ating natural language generation systems. Compu- tational Linguistics, 35(4):529-558.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Sportsett:basketball -a robust and maintainable dataset for natural language generation",
"authors": [
{
"first": "Craig",
"middle": [],
"last": "Thomson",
"suffix": ""
},
{
"first": "Ehud",
"middle": [],
"last": "Reiter",
"suffix": ""
},
{
"first": "Somayajulu Gowri",
"middle": [],
"last": "Sripada",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of IntelLanG",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Craig Thomson, Ehud Reiter, and Somayajulu Gowri Sripada. 2020. Sportsett:basketball -a robust and maintainable dataset for natural language generation. In Proceedings of IntelLanG 2020.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Towards ecologically valid research on language user interfaces",
"authors": [
{
"first": "Dzmitry",
"middle": [],
"last": "Harm De Vries",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Bahdanau",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Harm de Vries, Dzmitry Bahdanau, and Christopher Manning. 2020. Towards ecologically valid re- search on language user interfaces.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Revisiting challenges in datato-text generation with fact grounding",
"authors": [
{
"first": "Hongmin",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 12th International Conference on Natural Language Generation",
"volume": "",
"issue": "",
"pages": "311--322",
"other_ids": {
"DOI": [
"10.18653/v1/W19-8639"
]
},
"num": null,
"urls": [],
"raw_text": "Hongmin Wang. 2019. Revisiting challenges in data- to-text generation with fact grounding. In Proceed- ings of the 12th International Conference on Nat- ural Language Generation, pages 311-322, Tokyo, Japan. Association for Computational Linguistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Challenges in data-to-document generation",
"authors": [
{
"first": "Sam",
"middle": [],
"last": "Wiseman",
"suffix": ""
},
{
"first": "Stuart",
"middle": [],
"last": "Shieber",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Rush",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "2253--2263",
"other_ids": {
"DOI": [
"10.18653/v1/D17-1239"
]
},
"num": null,
"urls": [],
"raw_text": "Sam Wiseman, Stuart Shieber, and Alexander Rush. 2017a. Challenges in data-to-document generation. In Proceedings of the 2017 Conference on Empiri- cal Methods in Natural Language Processing, pages 2253-2263, Copenhagen, Denmark. Association for Computational Linguistics.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Challenges in data-to-document generation",
"authors": [
{
"first": "Sam",
"middle": [],
"last": "Wiseman",
"suffix": ""
},
{
"first": "Stuart",
"middle": [
"M"
],
"last": "Shieber",
"suffix": ""
},
{
"first": "Alexander",
"middle": [
"M"
],
"last": "Rush",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sam Wiseman, Stuart M. Shieber, and Alexander M. Rush. 2017b. Challenges in data-to-document gen- eration.",
"links": null
}
},
"ref_entries": {
"TABREF1": {
"type_str": "table",
"html": null,
"text": "",
"content": "<table/>",
"num": null
},
"TABREF2": {
"type_str": "table",
"html": null,
"text": "\u00b1 0.386 0.70 \u00b1 0.021 0.39 \u00b1 0.015 0.38 \u00b1 0.009 0.19 \u00b1 0.006 D2 BLEU 17.39 \u00b1 1.189 0.75 \u00b1 0.034 0.43 \u00b1 0.033 0.40 \u00b1 0.019 0.21 \u00b1 0.015 D1 RG 16.97 \u00b1 0.435 0.71 \u00b1 0.016 0.38 \u00b1 0.015 0.38 \u00b1 0.009 0.18 \u00b1 0.008 D2 RG 17.00 \u00b1 1.207 0.77 \u00b1 0.029 0.42 \u00b1 0.028 0.40 \u00b1 0.013 0.21 \u00b1 0.009 D1 CS-PREC 17.08 \u00b1 0.358 0.71 \u00b1 0.018 0.39 \u00b1 0.012 0.38 \u00b1 0.009 0.19 \u00b1 0.007",
"content": "<table><tr><td colspan=\"2\">Dataset Stopping Metric</td><td>BLEU</td><td>RG</td><td>CS-PREC</td><td>CS-REC</td><td>CO</td></tr><tr><td colspan=\"7\">D1 17.18 D2 BLEU CS-PREC 17.30 \u00b1 1.301 0.76 \u00b1 0.026 0.44 \u00b1 0.034 0.41 \u00b1 0.015 0.21 \u00b1 0.015</td></tr><tr><td>D1</td><td>CS-REC</td><td colspan=\"5\">17.12 \u00b1 0.314 0.71 \u00b1 0.017 0.39 \u00b1 0.011 0.38 \u00b1 0.007 0.19 \u00b1 0.005</td></tr><tr><td>D2</td><td>CS-REC</td><td colspan=\"5\">17.27 \u00b1 1.191 0.77 \u00b1 0.026 0.43 \u00b1 0.029 0.41 \u00b1 0.015 0.21 \u00b1 0.013</td></tr><tr><td>D1</td><td>CO</td><td colspan=\"5\">17.09 \u00b1 0.540 0.70 \u00b1 0.022 0.39 \u00b1 0.012 0.38 \u00b1 0.010 0.19 \u00b1 0.003</td></tr><tr><td>D2</td><td>CO</td><td colspan=\"5\">17.34 \u00b1 1.348 0.76 \u00b1 0.025 0.44 \u00b1 0.034 0.41 \u00b1 0.014 0.22 \u00b1 0.011</td></tr><tr><td>Gold</td><td>N/A</td><td>-</td><td>0.92</td><td>-</td><td>-</td><td>-</td></tr></table>",
"num": null
},
"TABREF3": {
"type_str": "table",
"html": null,
"text": "Experiment results; Comparing D1 to D2 within every cell pair is statistically significant (p < 0.005) with the exception of entries in the BLEU column. Note that BLEU, CS, and CO all inherently achieve 100% on gold standard texts.",
"content": "<table/>",
"num": null
}
}
}
} |