File size: 60,884 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 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:12:58.124388Z"
},
"title": "Active Learning for Argument Strength Estimation",
"authors": [
{
"first": "Nataliia",
"middle": [],
"last": "Kees",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LMU Munich",
"location": {
"country": "Germany"
}
},
"email": "kees.nataliia@gmail.com"
},
{
"first": "Michael",
"middle": [],
"last": "Fromm",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LMU Munich",
"location": {
"country": "Germany"
}
},
"email": "fromm@dbs.ifi.lmu.de"
},
{
"first": "Evgeniy",
"middle": [],
"last": "Faerman",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LMU Munich",
"location": {
"country": "Germany"
}
},
"email": ""
},
{
"first": "Thomas",
"middle": [],
"last": "Seidl",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LMU Munich",
"location": {
"country": "Germany"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "High-quality arguments are an essential part of decision-making. Automatically predicting the quality of an argument is a complex task that recently got much attention in argument mining. However, the annotation effort for this task is exceptionally high. Therefore, we test uncertainty-based active learning (AL) methods on two popular argument-strength data sets to estimate whether sample-efficient learning can be enabled. Our extensive empirical evaluation shows that uncertainty-based acquisition functions can not surpass the accuracy reached with the random acquisition on these data sets.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "High-quality arguments are an essential part of decision-making. Automatically predicting the quality of an argument is a complex task that recently got much attention in argument mining. However, the annotation effort for this task is exceptionally high. Therefore, we test uncertainty-based active learning (AL) methods on two popular argument-strength data sets to estimate whether sample-efficient learning can be enabled. Our extensive empirical evaluation shows that uncertainty-based acquisition functions can not surpass the accuracy reached with the random acquisition on these data sets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Argumentative quality plays a significant role in different domains of social activity where information and idea exchange are essential, such as the public domain and the scientific world. Theoretical discussions about what constitutes a good argument can be traced back to the ancient Greeks (Smith, 2020) . Researchers nowadays continue exploring this topic, trying out approaches that employ empirical machine learning estimation techniques (Simpson and Gurevych, 2018) .",
"cite_spans": [
{
"start": 294,
"end": 307,
"text": "(Smith, 2020)",
"ref_id": null
},
{
"start": 445,
"end": 473,
"text": "(Simpson and Gurevych, 2018)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "One of the most expensive and time-consuming tasks for machine learning-driven argument strength prediction is data labeling. Here, the result is highly dependent on the quality of labels, while the annotation task demands cognitive and reasoning abilities. One way to guarantee good annotations is to perform labeling with schooled experts, raising project costs extensively. For this reason, a common approach involves employing crowd workers. As argument strength detection is a highly subjective task, crowd workers' labeling results are often identified by low reliability and prompt researchers to counter-check the results with more crowd workers and as specifically developed agreement-based techniques. Sometimes a threshold for agreement cannot be reached at all, which might lead to data loss (see e.g. (Habernal and Gurevych, 2016a; Toledo et al., 2019) .",
"cite_spans": [
{
"start": 814,
"end": 844,
"text": "(Habernal and Gurevych, 2016a;",
"ref_id": "BIBREF6"
},
{
"start": 845,
"end": 865,
"text": "Toledo et al., 2019)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This motivates us to investigate the applicability of some existing methods for reducing the amount of training data for automatic argument strength prediction. To this end, we look closely at the technique of active learning (AL). In this paper, we evaluate standard uncertainty-based acquisition functions for the argument strength prediction. We perform several experiments for the task of binary argument-pair classification (see Table 1 ) with several uncertainty-based data selection rounds. Our findings show that uncertainty-based AL techniques do not provide any advantages compared to random selection strategies. The coldstart problem and unreliable nature of annotations concerning argument strength might constitute the reasons for the failure of these techniques.",
"cite_spans": [],
"ref_spans": [
{
"start": 434,
"end": 442,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Argument 2 School uniforms are a BAD idea. I'm to lazy to explain it but trust me, I wore them 4 years.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Argument 1",
"sec_num": null
},
{
"text": "School uniform cant save person out of cold or heat like special clothes. It is not comfortable when you sit for an hours in a classroom. Table 1 : Example of an argument pair both arguing against school uniforms (Habernal and Gurevych, 2016b) 2 Related Work",
"cite_spans": [
{
"start": 213,
"end": 243,
"text": "(Habernal and Gurevych, 2016b)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 138,
"end": 145,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Argument 1",
"sec_num": null
},
{
"text": "In general, there is no agreement on how to operationalize argumentation quality (Toledo et al., 2019; Wachsmuth et al., 2017; Simpson and Gurevych, 2018; Persing and Ng, 2015; Lauscher et al., 2020) . In some studies, argument strength is regarded in its persuasiveness and quantified as the proportion of people persuaded by the given argument (Habernal and Gurevych, 2016b; Persing and Ng, 2015; Toledo et al., 2019) . Persuasiveness makes argument strength easy to operationalize and serves as a way of dealing with the unclear nature of the concept by approximating its meaning through relying on the majority's wisdom. This approach lies at the center of the crowd-sourcing data labeling efforts and is the most common approach undertaken in existing data sets. This limits the reliability of the labels attained in such a manner, though, due to the highly subjective nature of such labels.",
"cite_spans": [
{
"start": 81,
"end": 102,
"text": "(Toledo et al., 2019;",
"ref_id": null
},
{
"start": 103,
"end": 126,
"text": "Wachsmuth et al., 2017;",
"ref_id": null
},
{
"start": 127,
"end": 154,
"text": "Simpson and Gurevych, 2018;",
"ref_id": null
},
{
"start": 155,
"end": 176,
"text": "Persing and Ng, 2015;",
"ref_id": null
},
{
"start": 177,
"end": 199,
"text": "Lauscher et al., 2020)",
"ref_id": null
},
{
"start": 346,
"end": 376,
"text": "(Habernal and Gurevych, 2016b;",
"ref_id": null
},
{
"start": 377,
"end": 398,
"text": "Persing and Ng, 2015;",
"ref_id": null
},
{
"start": 399,
"end": 419,
"text": "Toledo et al., 2019)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Argument Quality Estimation",
"sec_num": "2.1"
},
{
"text": "Active learning is defined as a machine learning technique designed to assist in annotating unlabelled data sets by automatically selecting the most informative examples, which are subsequently labeled by human experts (the so-called oracles) (Hu, 2011; Cohn et al., 1996) . A popular approach to estimating the informativeness of single data points involves quantifying model uncertainty from a sample of stochastic forward passes for a given data point. Common techniques such as entropy, mutual information, or variation ratios (see Appendix A.1 for more details) reportedly help reach good results on a range of tasks on high-dimensional data, e.g., in Computer Vision or Natural Language Processing (Gal et al., 2017; Siddhant and Lipton, 2018; Hu, 2011) . The assumption behind this is that in this way, data points which are closest to the decision boundary can be selected, helping to fine-tune the line dividing the classes most efficiently.",
"cite_spans": [
{
"start": 243,
"end": 253,
"text": "(Hu, 2011;",
"ref_id": null
},
{
"start": 254,
"end": 272,
"text": "Cohn et al., 1996)",
"ref_id": "BIBREF0"
},
{
"start": 704,
"end": 722,
"text": "(Gal et al., 2017;",
"ref_id": "BIBREF5"
},
{
"start": 723,
"end": 749,
"text": "Siddhant and Lipton, 2018;",
"ref_id": null
},
{
"start": 750,
"end": 759,
"text": "Hu, 2011)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Active Learning",
"sec_num": "2.2"
},
{
"text": "So far, AL in argument mining has received little attention. In the work of (Ein-Dor et al., 2020), the authors propose an Iterative Retrospective Learning (IRL) variant for the argument mining task. Their approach, however, is focused on solving the class imbalance problem between arguments and nonarguments and is precisionrather than accuracyoriented as AL is. Another approach is suggested by (Simpson and Gurevych, 2018) . They apply the Gaussian process preference learning (GPPL) method for performing AL for estimating argument convincingness, which the authors expect to be helpful against the cold-start problem.",
"cite_spans": [
{
"start": 398,
"end": 426,
"text": "(Simpson and Gurevych, 2018)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Active Learning",
"sec_num": "2.2"
},
{
"text": "For our analysis, we use two publicly available data sets suitable for the task of pairwise argument strength prediction:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set",
"sec_num": "3"
},
{
"text": "\u2022 UKPConvArg1Strict, published by (Habernal and Gurevych, 2016b) , consists of 11,650 argument pairs distributed over 16 topics.",
"cite_spans": [
{
"start": 34,
"end": 64,
"text": "(Habernal and Gurevych, 2016b)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set",
"sec_num": "3"
},
{
"text": "\u2022 IBM-9.1kPairs, presented by (Toledo et al., 2019) consists of 9,125 argument pairs distributed over 11 topics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set",
"sec_num": "3"
},
{
"text": "Because supporting and opposing arguments often share the same vocabulary and semantics, we do not treat each stance within a given topic as a separate topic, contrary to the authors of the two data sets. Instead, we combine the \"for\" and \"against\" arguments within the same topic under the same topic index and, thus, avoid leakage of semantic information between train and test data split. This preprocessing makes the performance of our models not directly comparable with the performance from the original papers. However, reproducibility of the original papers' results is beyond the scope of this work, as our focus lies on testing AL acquisition functions instead of reaching higher performance with our models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set",
"sec_num": "3"
},
{
"text": "Due to the high computational costs of the AL process, we decide to select the three most representative topics from each data set. One way to reach high representativeness would be to select topics that are average in difficulty. Since we try to approximate a real-world setting where the labels are unknown, it is not clear at the beginning which topics are more challenging to learn than the others. For this reason, we decide to select our test topics according to their size. Thus, we cross-validate our models on each data set's smallest topic, the largest one, and the median-sized one. Thus, the topics we select according to this procedure are topics 10 (\"Is the school uniform a good or bad idea?\"), 13 (\"TV is better than books\") and 14 (\"Personal pursuit or advancing the common good?\") in UKPCon-vArg1Strict and topics 3 (\"Does social media bring more harm than good?\"), 4 (\"Should we adopt cryptocurrency?\") and 7 (\"Should we ban fossil fuels?\") in IBM-9.1kPairs data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set",
"sec_num": "3"
},
{
"text": "This study aims to test the hypothesis that uncertainty-based data acquisition strategies can help to achieve a better model performance than a mere random selection of the data for argument strength estimation. We test this by comparing different data selection strategies against random data selection, serving as a baseline.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Research Design",
"sec_num": "4.1"
},
{
"text": "We test our acquisition strategies on a task of pairwise (relative) argument strength comparison, constructed as a binary classification task, for which we use the UKPConvArg1Strict (Habernal and Gurevych, 2016b) and IBM-9.1kPairs (Toledo et al., 2019) data sets. The code to our experiments is publicly available. 1 In order to employ uncertainty-based acquisition functions, we need to measure model uncertainty at prediction time. This is possible either by using Bayesian methods or by approximating their effect via obtaining distributions for output predictions by some other means. Based on the ground work layed out by (Gal and Ghahramani, 2016) , who show that dropout training in deep neural networks help approximate Bayesian inference in deep Gaussian processes, we design our experiments as MC dropout. With this, we simulate several stochastic forward passes through the model at prediction time and sample repeatedly from softmax outputs to obtain prediction distributions.",
"cite_spans": [
{
"start": 182,
"end": 212,
"text": "(Habernal and Gurevych, 2016b)",
"ref_id": null
},
{
"start": 315,
"end": 316,
"text": "1",
"ref_id": null
},
{
"start": 627,
"end": 653,
"text": "(Gal and Ghahramani, 2016)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Research Design",
"sec_num": "4.1"
},
{
"text": "Similar to the procedure stipulated by (Toledo et al., 2019), we fine-tune the pre-trained BERT-Base Uncased English (Devlin et al., 2018) for the task of binary argument-pair classification by adding a single classification layer on top. The BERT architecture includes dropout layers with a probability of 0.1 (Devlin et al., 2018) . We keep it this way, which allows us to approximate model uncertainty as described above and test the uncertainty-based acquisition functions on the fine-tuned BERT-based. To do that, we enable dropout at inference time.",
"cite_spans": [
{
"start": 117,
"end": 138,
"text": "(Devlin et al., 2018)",
"ref_id": "BIBREF1"
},
{
"start": 311,
"end": 332,
"text": "(Devlin et al., 2018)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Method and Procedure",
"sec_num": "4.2"
},
{
"text": "In order to estimate topic difficulty and validate our topic selection procedure described above, we train and test the models on all available labels of both data sets separately with the method of k-fold cross-validation, where k stands for the respective number of topics in a given data set. We separate every topic and use it as test data, with model training performed on the rest of the data, which helps to isolate the topics and measure their respective difficulty.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method and Procedure",
"sec_num": "4.2"
},
{
"text": "Our active learning experiments are conducted in a setting of a 3-fold cross-validation, with 3 indi-cating the number of most representative topics selected by us from the given data sets, as mentioned in Section 3. Thus, in each fold in our experiments, we test on one of the three selected topics for each data set (holdout data) and train on the rest of the compete data set (train-dev).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method and Procedure",
"sec_num": "4.2"
},
{
"text": "The train-dev data in each fold consists of random splits into train (85%) and validation (15%) data, whereas the validation, or development, data are used for measuring the goodness of fit of the model trained on the training data. Having separated and fixed the validation data, a batch of 130 argument pairs is selected randomly from the train split. These data are used as initial training data on which bert-base-uncased is fine-tuned according to our classification task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method and Procedure",
"sec_num": "4.2"
},
{
"text": "Model evaluation is performed via accuracy measurement. Training on each of the three folds per data set is conducted ten times for improved reliability of the results. Thus, for each fold, we produce ten validation splits and ten initial training data batches to add some randomness into the experiments but in a controlled manner. They are kept fixed for every training fold to control for the effect of random initial data selection and enable a reliable comparison between the acquisition functions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method and Procedure",
"sec_num": "4.2"
},
{
"text": "We add another 130 argument pairs in each learning round and re-train the fine-tuned model. Within this setting, the whole data set would be selected within approx. 55 iterations for IBM-9.1kPairs data and approx. 72 iterations for UKPConvArg1Strict data (when calculated with the median-sized test split size). In an attempt to minimize the burden associated with heavy training, we decide to limit each active learning process to (less than) a half iterations, stopping at the 27 th iteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method and Procedure",
"sec_num": "4.2"
},
{
"text": "Further details on the hyperparameters and the computing and software infrastructure can be found in Appendix sections A.2 and A.4.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Method and Procedure",
"sec_num": "4.2"
},
{
"text": "We perform AL on three uncertainty-based acquisition functions one by one. In particular, we compare the performance of variation ratios, entropy, and BALD (Houlsby et al., 2011; Gal et al., 2017) against a random acquisition baseline. For each of the learning rounds, we acquire data based on the heuristics calculated over a sample of 20 stochastic forward pass outputs. Our expectation is that other measures will outperform the random acquisition.",
"cite_spans": [
{
"start": 156,
"end": 178,
"text": "(Houlsby et al., 2011;",
"ref_id": null
},
{
"start": 179,
"end": 196,
"text": "Gal et al., 2017)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Acquisition Functions",
"sec_num": "4.3"
},
{
"text": "For the estimation of the performance of the models trained on the whole data with k-fold crossvalidation, we reach a comparable performance of our BERT-based binary classification technique on both of the data sets (average accuracy on UKP-ConvArg1Strict: 0.76, on IBM-9.1kPairs: 0.77). This is a slightly worse performance than (Toledo et al., 2019) achieves with the same architecture; the reason could be attributed to a different topic attribution strategy, as well as to some differences in the used hardware or hyperparameters, such as batch size or the number of epochs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "We find that the topics selected by us from the UKPConvArg1Strict stand rather on the low end of difficulty, with model accuracy tending towards the upper end of the scale when validated on these topics: all of them are higher than the mean performance of 0.76 (see Appendix A.3 for more details). However, from the distribution point of view, two of the topics, namely 10 and 13, yield median model performance, making them, in our opinion, suitable representatives of the whole data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "As for the IBM-9.1kPairs data set, our selected topics produce on average comparable performance with the model performance on the whole topic set (accuracy of 0.776 vs. 0.77 respectively). They also represent the most difficult topic, the easiest topic, and one closely neighboring the median topic (accuracy of 0.78 being slightly higher than the median performance of 0.77). In this case, the selected topics provide a better representation of the whole data set and grant strong validity when it comes to generalizing the results of our experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "The series of experiments we conducted in order to test whether our proposed heuristics for AL data acquisition provide us with any significant improvement surprisingly do not reveal any heuristic which would perform better than in the case of a random acquisition. This is true both for UKP-ConvArg1Strict and IBM-9.1kPairs data; a detailed overview is presented in Tables 2 and 3. Statistical significance of the results has been tested with a Wilcoxon signed-rank test, which provides a nonparametric alternative to the paired T-test and is more suitable due to the non-Gaussian distribution of the differences in the results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "All heuristics result in performance that is lower than that of the random baseline. All of our results are statistically significant with p-values \u2264 0.0001.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Despite the fact that random acquisition turns out to be the best one in terms of performance, with our results being consistent through both data sets and the difference being statistically significant, it is still noticeable that the differences in each case are rather small (see Figures 1 and 2 for graphic visualization of the model performance during active learning rounds comparing the acquisition functions).",
"cite_spans": [],
"ref_spans": [
{
"start": 283,
"end": 298,
"text": "Figures 1 and 2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "The results of our experiments do not point to any acquisition functions which outperform random acquisition. This finding does not exclude the possible existence of some other suitable acquisition functions, even from the same class (such as uncertainty-based). This remains an open question and should be considered in further research on the topic. For the time being, the random acquisition should be considered the approach of choice when selecting data for labeling for the task of pairwise argument strength prediction. This is sensible both from an accuracy standpoint as well as due to the computational cheapness of a random process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "As the literature suggests, a possible reason why uncertainty-based methods perform so unimpressively is their proneness to picking outliers -a disadvantage that some other methods, such as diversity-based acquisition (e.g., (Sener and Savarese, 2018), do not have. This might be especially critical in the realm of argument strength prediction, as outliers might represent the arguments where relative argument strength difference is marginal, the data are noisy, or where the provided labeling is too subjective. Another critical factor is the cold-start problem, i.e., overfitting on the small initial data set of data, for which no initial informativeness estimation could be performed. This poses a drawback for the uncertainty-based methods, relying on the initial data sample for subsequent data acquisition.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "This paper evaluates the effect of uncertainty-based acquisition functions, such as variation ratios, entropy, and BALD, on the model performance in the realm of argument strength prediction. As no acquisition function tested helps improve model performance in comparison to the random acquisition, we have not found any justification for using uncertainty-based active learning for pairwise argument strength estimation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "This work has been funded by the German Federal Ministry of Education and Research (BMBF) under Grant No. 01IS18036A and by the Deutsche Forschungsgemeinschaft (DFG) within the project Relational Machine Learning for Argument Validation (ReMLAV), Grant Number SE 1039/10-1, as part of the Priority Program \"Robust Argumentation Machines (RATIO)\" (SPP-1999) . The authors of this work take full responsibility for its content. ",
"cite_spans": [
{
"start": 346,
"end": 356,
"text": "(SPP-1999)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": "8"
},
{
"text": "In our work, we refer in particular to the following uncertainty-based acquisition functions (Gal et al., 2017 ):",
"cite_spans": [
{
"start": 93,
"end": 110,
"text": "(Gal et al., 2017",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "\u2022 variation ratios: given a set of labels y T from T stochastic forward passes, variation ratio for a given input point is calculated as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "varrat(x) = 1 \u2212 f x T",
"eq_num": "(1)"
}
],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "with f x denoting the number of times the most commonly occurring category (mode of the distribution) has been sampled. This serves as an indication of how concentrated the predictions are, with 0.5 being the highest dispersion, i.e. uncertainty, and 0 being the highest concentration (certainty) in the case of binary classification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "\u2022 predictive entropy: stems from information theory and is calculated by averaging the softmax values for each class :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "predentr(x) = \u2212 c p(y = c|x, D train ) \u00d7 log 2 (p(y = c|x, D train )),",
"eq_num": "(2)"
}
],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "where p(y = c|x, D train ) stands for average probability of a data point adhering to a specific class given the outputs of the stochastic forward passes and the training data. c denotes the label class, i.e. we sum the values over all the classes to receive a measure of entropy for a given data point.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "\u2022 Bayesian Active Learning by Disagreement (BALD) (Houlsby et al., 2011), also called mutual information (Gal, 2016) , is a function of predictive entropy as described above and averaged predictive entropies that have been calculated separately for each output:",
"cite_spans": [
{
"start": 105,
"end": 116,
"text": "(Gal, 2016)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "bald ( (3) Should we limit autonomous cars? 1217 0.79 10 Should we support information privacy laws? 533 0.77 Total Size/Average Acc. 9 125 0.77 Table 5 : Topic sizes in IBM-9.1kPairs. Topics are provided with their corresponding numbers and size within the data set, as well as our model's performance at test time. The topics selected for testing the acquisition functions have been highlighted in italics.",
"cite_spans": [],
"ref_spans": [
{
"start": 145,
"end": 152,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "A.1 Uncertainty-based Acquisition Functions",
"sec_num": null
},
{
"text": "For the evaluation we initialized all methods for ten runs with different seeds and reported the mean accuracy score. We used early stopping with a patience of three on a pre-selected validation set for regularization. As loss function we used weighted binary-cross-entropy for the (relative) Argument Strength task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A.4 Hyperparameters",
"sec_num": null
},
{
"text": "We train our models on top of the pre-trained BERT-Base uncased with a dropout probability of 0.1. Learning rate is 2 \u22125 (same as in (Toledo et al., 2019) ). The batch size per GPU is 64 and the model is validated after every half epoch.",
"cite_spans": [
{
"start": 133,
"end": 154,
"text": "(Toledo et al., 2019)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "A.4 Hyperparameters",
"sec_num": null
},
{
"text": "https://github.com/nkees/ active-learning-argument-strength",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "The experiments were conducted on a Ubuntu 18.04 system with an AMD Ryzen Processor with 16 CPU-Cores, 126 GB memory, and a single NVIDIA RTX 2080 GPU with 11 GB memory. We further used Python 3.7, PyTorch 1.4 and the Huggingface-Transformer library (2.11.0). ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A.2 Computing & Software Infrastructure",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Active learning with statistical models",
"authors": [
{
"first": "David",
"middle": [
"A"
],
"last": "Cohn",
"suffix": ""
},
{
"first": "Zoubin",
"middle": [],
"last": "Ghahramani",
"suffix": ""
},
{
"first": "Michael",
"middle": [
"I"
],
"last": "Jordan",
"suffix": ""
}
],
"year": 1996,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David A. Cohn, Zoubin Ghahramani, and Michael I. Jordan. 1996. Active learning with statistical mod- els. CoRR, cs.AI/9603104.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "BERT: pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language under- standing. CoRR, abs/1810.04805.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Ranit Aharonov, and Noam Slonim. 2020. Corpus wide argument mining-a working solution",
"authors": [
{
"first": "Eyal",
"middle": [],
"last": "Liat Ein-Dor",
"suffix": ""
},
{
"first": "Lena",
"middle": [],
"last": "Shnarch",
"suffix": ""
},
{
"first": "Alon",
"middle": [],
"last": "Dankin",
"suffix": ""
},
{
"first": "Benjamin",
"middle": [],
"last": "Halfon",
"suffix": ""
},
{
"first": "Ariel",
"middle": [],
"last": "Sznajder",
"suffix": ""
},
{
"first": "Carlos",
"middle": [],
"last": "Gera",
"suffix": ""
},
{
"first": "Martin",
"middle": [],
"last": "Alzate",
"suffix": ""
},
{
"first": "Leshem",
"middle": [],
"last": "Gleize",
"suffix": ""
},
{
"first": "Yufang",
"middle": [],
"last": "Choshen",
"suffix": ""
},
{
"first": "Yonatan",
"middle": [],
"last": "Hou",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Bilu",
"suffix": ""
}
],
"year": null,
"venue": "Proceedings of the AAAI Conference on Artificial Intelligence",
"volume": "34",
"issue": "",
"pages": "7683--7691",
"other_ids": {
"DOI": [
"10.1609/aaai.v34i05.6270"
]
},
"num": null,
"urls": [],
"raw_text": "Liat Ein-Dor, Eyal Shnarch, Lena Dankin, Alon Hal- fon, Benjamin Sznajder, Ariel Gera, Carlos Alzate, Martin Gleize, Leshem Choshen, Yufang Hou, Yonatan Bilu, Ranit Aharonov, and Noam Slonim. 2020. Corpus wide argument mining-a working solution. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):7683-7691.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Uncertainty in Deep Learning",
"authors": [
{
"first": "Yarin",
"middle": [],
"last": "Gal",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yarin Gal. 2016. Uncertainty in Deep Learning. Ph.D. thesis, University of Cambridge.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Dropout as a bayesian approximation: Representing model uncertainty in deep learning",
"authors": [
{
"first": "Yarin",
"middle": [],
"last": "Gal",
"suffix": ""
},
{
"first": "Zoubin",
"middle": [],
"last": "Ghahramani",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of The 33rd International Conference on Machine Learning",
"volume": "48",
"issue": "",
"pages": "1050--1059",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model un- certainty in deep learning. In Proceedings of The 33rd International Conference on Machine Learn- ing, volume 48 of Proceedings of Machine Learning Research, pages 1050-1059, New York, New York, USA. PMLR.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Deep bayesian active learning with image data",
"authors": [
{
"first": "Yarin",
"middle": [],
"last": "Gal",
"suffix": ""
},
{
"first": "Riashat",
"middle": [],
"last": "Islam",
"suffix": ""
},
{
"first": "Zoubin",
"middle": [],
"last": "Ghahramani",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yarin Gal, Riashat Islam, and Zoubin Ghahramani. 2017. Deep bayesian active learning with image data. CoRR, abs/1703.02910.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "What makes a convincing argument? empirical analysis and detecting attributes of convincingness in web argumentation",
"authors": [
{
"first": "Ivan",
"middle": [],
"last": "Habernal",
"suffix": ""
},
{
"first": "Iryna",
"middle": [],
"last": "Gurevych",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1214--1223",
"other_ids": {
"DOI": [
"10.18653/v1/D16-1129"
]
},
"num": null,
"urls": [],
"raw_text": "Ivan Habernal and Iryna Gurevych. 2016a. What makes a convincing argument? empirical analysis and detecting attributes of convincingness in web argumentation. In Proceedings of the 2016 Con- ference on Empirical Methods in Natural Language Processing, pages 1214-1223, Austin, Texas. Asso- ciation for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Overview of the training results on the UKP-ConvArg1Strict dataset based on different uncertaintybased acquisation methods Figure 2: Overview of the training results on the IBM-9.1kPairs dataset based on different uncertainty-based acquisition methods",
"type_str": "figure",
"uris": null,
"num": null
},
"FIGREF1": {
"text": "x) = \u2212[ c p(y = c|x, D train ) \u00d7 log(p(y = c|x, D train ))] + E p(\u03c9|D train ) [ c p(y = c|w, \u03c9) \u00d7 log(p(y = c|x, \u03c9))].",
"type_str": "figure",
"uris": null,
"num": null
},
"TABREF0": {
"num": null,
"type_str": "table",
"html": null,
"text": "HeuristicMean Variat. Avg.Diff.",
"content": "<table><tr><td>random (b.)</td><td>0.747 0.0881</td><td>-</td></tr><tr><td>entropy</td><td>0.7388 0.0925</td><td>-0.0082</td></tr><tr><td colspan=\"2\">variation ratios 0.7368 0.0922</td><td>-0.0103</td></tr><tr><td>bald</td><td>0.7377 0.0928</td><td>-0.0093</td></tr></table>"
},
"TABREF1": {
"num": null,
"type_str": "table",
"html": null,
"text": "Results of active learning experiments on UKPConvArg1Strict. Abbreviations: b. stands for baseline, variat. stands for variation, avg.diff. stands for average difference. Negative average difference means that the challenger heuristic has not outperformed the baseline.",
"content": "<table><tr><td>Heuristic</td><td colspan=\"2\">Mean Variat. Avg.Diff.</td></tr><tr><td>random (b.)</td><td>0.7491 0.0855</td><td>-</td></tr><tr><td>entropy</td><td>0.7414 0.0878</td><td>-0.0077</td></tr><tr><td colspan=\"2\">variation ratios 0.7377 0.0923</td><td>-0.0114</td></tr><tr><td>bald</td><td>0.7412 0.0882</td><td>-0.0079</td></tr></table>"
},
"TABREF2": {
"num": null,
"type_str": "table",
"html": null,
"text": "Results of active learning experiments on IBM-9.1kPairs. Abbreviations: b. stands for baseline, variat. stands for variation, avg.diff. stands for average difference. Negative average difference means that the challenger heuristic has not outperformed the baseline.",
"content": "<table/>"
},
"TABREF3": {
"num": null,
"type_str": "table",
"html": null,
"text": "IvanHabernal and Iryna Gurevych. 2016b. Which argument is more convincing? analyzing and predicting convincingness of web arguments using bidirectional LSTM. In Proceedings of the 54th Annual Meeting of the Association for Computational",
"content": "<table><tr><td>Linguistics (Volume 1: Long Papers), pages 1589-</td></tr><tr><td>1599, Berlin, Germany. Association for Computa-</td></tr><tr><td>tional Linguistics.</td></tr><tr><td>Neil Houlsby, Ferenc Husz\u00e1r, Zoubin Ghahramani, and</td></tr><tr><td>M\u00e1t\u00e9 Lengyel. 2011. Bayesian active learning for</td></tr><tr><td>classification and preference learning.</td></tr><tr><td>Rong Hu. 2011. Active Learning for Text Classifi-</td></tr><tr><td>cation. Doctoral Thesis, Technological University</td></tr><tr><td>Dublin.</td></tr><tr><td>Anne Lauscher, Lily Ng, Courtney Napoles, and Joel</td></tr><tr><td>Tetreault. 2020. Rhetoric, logic, and dialectic: Ad-</td></tr><tr><td>vancing theory-based argument quality assessment</td></tr><tr><td>in natural language processing.</td></tr><tr><td>Isaac Persing and Vincent Ng. 2015. Modeling argu-</td></tr><tr><td>ment strength in student essays. In Proceedings</td></tr><tr><td>of the 53rd Annual Meeting of the Association for</td></tr><tr><td>Computational Linguistics and the 7th International</td></tr><tr><td>Joint Conference on Natural Language Processing</td></tr><tr><td>(Volume 1: Long Papers), pages 543-552.</td></tr><tr><td>Ozan Sener and Silvio Savarese. 2018. Active learn-</td></tr><tr><td>ing for convolutional neural networks: A core-set</td></tr><tr><td>approach. In International Conference on Learning</td></tr><tr><td>Representations.</td></tr><tr><td>Aditya Siddhant and Zachary C. Lipton. 2018. Deep</td></tr><tr><td>bayesian active learning for natural language pro-</td></tr><tr><td>cessing: Results of a large-scale empirical study.</td></tr><tr><td>CoRR, abs/1808.05697.</td></tr><tr><td>Edwin Simpson and Iryna Gurevych. 2018. Finding</td></tr><tr><td>convincing arguments using scalable Bayesian pref-</td></tr><tr><td>erence learning. Transactions of the Association for</td></tr><tr><td>Computational Linguistics, 6:357-371.</td></tr><tr><td>Robin Smith. 2020. Aristotle's Logic. In Edward N.</td></tr><tr><td>Zalta, editor, The Stanford Encyclopedia of Philos-</td></tr><tr><td>ophy, Fall 2020 edition. Metaphysics Research Lab,</td></tr><tr><td>Stanford University.</td></tr><tr><td>Assaf Toledo, Shai Gretz, Edo Cohen-Karlik, Roni</td></tr><tr><td>Friedman, Elad Venezian, Dan Lahav, Michal Ja-</td></tr><tr><td>covi, Ranit Aharonov, and Noam Slonim. 2019. Au-</td></tr><tr><td>tomatic argument quality assessment -new datasets</td></tr><tr><td>and methods.</td></tr><tr><td>Henning Wachsmuth, Nona Naderi, Yufang Hou,</td></tr><tr><td>Yonatan Bilu, Vinodkumar Prabhakaran, Tim Al-</td></tr><tr><td>berdingk Thijm, Graeme Hirst, and Benno Stein.</td></tr><tr><td>2017. Computational argumentation quality assess-</td></tr><tr><td>ment in natural language. In Proceedings of the 15th</td></tr><tr><td>Conference of the European Chapter of the Associa-</td></tr><tr><td>tion for Computational Linguistics: Volume 1, Long</td></tr><tr><td>Papers, pages 176-187, Valencia, Spain. Associa-</td></tr><tr><td>tion for Computational Linguistics.</td></tr></table>"
},
"TABREF4": {
"num": null,
"type_str": "table",
"html": null,
"text": "Topic sizes in UKPConvArg1Strict. Topics are provided with their corresponding numbers and size within the data set, as well as our model's performance at test time. The topics selected for testing the acquisition functions have been highlighted in italics.",
"content": "<table><tr><td colspan=\"2\">No. Topic</td><td colspan=\"2\">Size Acc.</td></tr><tr><td>0</td><td>Should flu vaccinations be mandatory?</td><td>731</td><td>0.75</td></tr><tr><td>1</td><td>Should gambling be banned?</td><td>503</td><td>0.8</td></tr><tr><td>2</td><td colspan=\"2\">Does online shopping bring more harm than good? 278</td><td>0.79</td></tr><tr><td>3</td><td>Does social media bring more harm than good?</td><td colspan=\"2\">2587 0.78</td></tr><tr><td>4</td><td>Should we adopt cryptocurrency?</td><td>719</td><td>0.82</td></tr><tr><td>5</td><td>Should we adopt vegetarianism?</td><td colspan=\"2\">1073 0.77</td></tr><tr><td>6</td><td>Should we sale violent video games to minors?</td><td>484</td><td>0.74</td></tr><tr><td>7</td><td>Should we ban fossil fuels?</td><td>263</td><td>0.73</td></tr><tr><td>8</td><td>Should we legalize doping in sport?</td><td>737</td><td>0.77</td></tr><tr><td>9</td><td/><td/><td/></tr></table>"
}
}
}
} |