File size: 67,455 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 | {
"paper_id": "R09-1013",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T15:00:09.916695Z"
},
"title": "Prototype-based Active Learning for Lemmatization",
"authors": [
{
"first": "Walter",
"middle": [],
"last": "Daelemans",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Antwerp Antwerp",
"location": {
"country": "Belgium"
}
},
"email": "walter.daelemans@ua.ac.be"
},
{
"first": "Hendrik",
"middle": [
"J"
],
"last": "Groenewald",
"suffix": "",
"affiliation": {},
"email": "handre.groenewald@nwu.ac.za"
},
{
"first": "Gerhard",
"middle": [
"B"
],
"last": "Van Huyssteen",
"suffix": "",
"affiliation": {},
"email": "gvhuyssteen@csir.co.za"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Annotation of training data for machine learning is often a laborious and costly process. In Active Learning (AL), criteria are investigated that allow ordering the unannotated data in such a way that those instances potentially contributing most to the speed of learning can be annotated first. Within this context we explore a new approach that focuses on prototypicality as a criterion for the selection of instances to act as training data in order to optimize prediction accuracy. In parallel with the prototype-based active classification (PBAC) approach of Cebron & Berthold (2009), we investigate whether the basic PBAC assumption rings true for linguistic data. The NLP task we address is lemmatization, the reduction of inflected word forms to their base-form. We operationalize prototypicality as features (i.e. word frequency and word length) of the already available training data items, and combine this with a measure of uncertainty (entropy). We show that the selection of less prototypical instances first, provides performance that is better than when data is randomly selected or when state of the art AL methods are used. We argue that this improvement is possible due to the fact that language processing tasks have highly disjunctive instance spaces, as there are often few regularities and many irregularities.",
"pdf_parse": {
"paper_id": "R09-1013",
"_pdf_hash": "",
"abstract": [
{
"text": "Annotation of training data for machine learning is often a laborious and costly process. In Active Learning (AL), criteria are investigated that allow ordering the unannotated data in such a way that those instances potentially contributing most to the speed of learning can be annotated first. Within this context we explore a new approach that focuses on prototypicality as a criterion for the selection of instances to act as training data in order to optimize prediction accuracy. In parallel with the prototype-based active classification (PBAC) approach of Cebron & Berthold (2009), we investigate whether the basic PBAC assumption rings true for linguistic data. The NLP task we address is lemmatization, the reduction of inflected word forms to their base-form. We operationalize prototypicality as features (i.e. word frequency and word length) of the already available training data items, and combine this with a measure of uncertainty (entropy). We show that the selection of less prototypical instances first, provides performance that is better than when data is randomly selected or when state of the art AL methods are used. We argue that this improvement is possible due to the fact that language processing tasks have highly disjunctive instance spaces, as there are often few regularities and many irregularities.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Supervised machine learning techniques are still superior to unsupervised machine learning techniques for many NLP tasks. However, annotation of training data is often a laborious and costly process. In Active Learning (AL) [1, 2] criteria are investigated that allow ordering the unannotated data in such a way that those instances potentially contributing most to the speed of learning can be annotated first. Rather than relying on random samples to act as instances in training data, AL entails the selection of instances to act as training data in order to optimize prediction accuracy. Ideally, AL leads to the creation of a supervised learning classifier at a fraction of the annotation effort needed when selecting new training items to be annotated randomly, and without loss in accuracy. Also in domain adaptation, AL has been proposed [3] as a feasible approach. Research on AL centers around the development and comparison of different approaches that could be used to order the available unannotated examples in such a way that those selected first are the ones most informative for learning. Another research area is the design of suitable stopping criteria.",
"cite_spans": [
{
"start": 224,
"end": 227,
"text": "[1,",
"ref_id": "BIBREF0"
},
{
"start": 228,
"end": 230,
"text": "2]",
"ref_id": "BIBREF1"
},
{
"start": 846,
"end": 849,
"text": "[3]",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "In this paper, we explore a new approach that focuses on prototypicality (i.e. the degree to which some examples are better, more representative examples of a category than others) as a criterion for ordering the data. In parallel with the prototype-based active classification (PBAC) approach of Cebron & Berthold [4] , we investigate whether the basic PBAC assumption rings true for linguistic data. We operationalize prototypicality on the basis of different features of the already available training data items and show that the selection of less prototypical instances first provides performance that is better than when data is randomly selected or when state of the art AL methods are used (i.e. a committee-based entropy method). The NLP task we address is lemmatization, the reduction of inflected word forms to their base-form.",
"cite_spans": [
{
"start": 315,
"end": 318,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "In Section 2 work related to Prototype Theory, Active Learning, and lemmatization is discussed. Section 3 outlines our approach, Section 4 describes our experiments on Afrikaans lemmatization and the remainder of the paper discusses these results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "In studies on human cognition, prototypes have been studied for many years [5] . Prototype effects are especially prevalent in language structure and language usage, and have been a central topic in the Cognitive Linguistics paradigm [6, 7] . It is widely accepted that language structures (including lexical items) show prototype effects: some instances are better examples than others. For example, with regard to plural formation the {s} morpheme (like in tables) could be considered more prototypical than the {en} morpheme (as in oxen); likewise, the lexical item chair would probably be considered to be a more proto-typical English lexical item than, say, cache. In both these cases frequency plays a central role in determining which example is more prototypical than the other (e.g. chair is more frequently used than cache). When working with linguistic data for natural language processing (NLP) purposes, we can therefore safely assume that some examples in a data set (of unknown examples) are better examples, or more representative, than others.",
"cite_spans": [
{
"start": 75,
"end": 78,
"text": "[5]",
"ref_id": null
},
{
"start": 234,
"end": 237,
"text": "[6,",
"ref_id": "BIBREF5"
},
{
"start": 238,
"end": 240,
"text": "7]",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Prototype Theory",
"sec_num": "2.1"
},
{
"text": "Various interdependent physiological, referential, statistical and/or psychological factors determine prototypicality [7] ; in our current research we focus on frequency (i.e. most commonly or productively used; cf. the statistical hypothesis in Prototype Theory), and size/length (i.e. prototypical concepts are often represented by shorter words; cf. insights on basic-level effects in Prototype Theory [6] ). Other factors include membership function (i.e. centrality and salience within a family resemblance model), activation time (i.e. time to process/classify/identify), association/chaining (i.e. the link between form, function, and meaning), conventionalization (i.e. how well-known a word is), acquisition (i.e. more prototypical items are learned first), etc. [6] ). These factors are not considered in our current research, but could also be operationalized in future work.",
"cite_spans": [
{
"start": 118,
"end": 121,
"text": "[7]",
"ref_id": "BIBREF6"
},
{
"start": 405,
"end": 408,
"text": "[6]",
"ref_id": "BIBREF5"
},
{
"start": 772,
"end": 775,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Prototype Theory",
"sec_num": "2.1"
},
{
"text": "Predominant approaches to AL include uncertainty-based sampling [8] , Support Vector Machine methods [9] , and query-by-committee [2] . The latter is a popular method in Active Learning, where entropy computed on the basis of a committee of classifiers is used as a criterion for sample selection.",
"cite_spans": [
{
"start": 64,
"end": 67,
"text": "[8]",
"ref_id": "BIBREF7"
},
{
"start": 101,
"end": 104,
"text": "[9]",
"ref_id": "BIBREF8"
},
{
"start": 130,
"end": 133,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Active Learning",
"sec_num": "2.2"
},
{
"text": "Recently, Cebron & Berthold [4] presented a novel approach to AL, which they call prototype-based active classification (PBAC). In their algorithm a new, labeled prototype is added in each learning iteration to fine-tune the classification of the datasets; prototypical (i.e. representative) examples are selected first, and examples at the classification boundary (i.e. less prototypical examples) are only selected/focused on automatically when it becomes necessary. In all their experiments, they only use non-linguistic data.",
"cite_spans": [
{
"start": 28,
"end": 31,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Active Learning",
"sec_num": "2.2"
},
{
"text": "In the PBAC algorithm the relative importance of each data point is calculated as a combination of (a) its representativeness of the data set as a whole (i.e. based on density estimates on the unannotated data), and (b) the uncertainty of a classifier to assign a class to it (i.e. measured as entropybased on the annotated datathat is inversely related to the voting confidence of the classifier). These two measures are then combined as a new data selection criterion (i.e. the uncertainty distribution), which is calculated as the weighted sum of the representativeness (called potentials) and the classification uncertainty, to the extent that \"the remaining potential on the data point still prevents unrepresentative samples from being chosen\", which \"helps to prevent selection of rare or borderline cases\"; see Mazzoni et al. [10] for the detrimental effects of choosing irrelevant data points in AL). Thus, the uncertainty distribution is being used to choose prototypical examples for classification, an approach which promises, with regard to non-linguistic data, to outperform AL with random initialization and closest-to-boundary selection; the algorithm also proved to be stable, and reaches levels of accuracy close to the final one after only a few iterations.",
"cite_spans": [
{
"start": 834,
"end": 838,
"text": "[10]",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Active Learning",
"sec_num": "2.2"
},
{
"text": "One central aspect of Prototype Theory that is important for the PBAC algorithm is the radial model of categorization. Lakoff [6] makes it clear that \"the center, or prototype, of the category is predictable. And while the non-central members are not predictable from the central members, they are \"motivated\" by it, in the sense that they bear family resemblances to it.\" Hence, in the PBAC algorithm, the value of the radius of the neighborhood (i.e. a positive constant defining a neighborhood) is one of the parameters that determine the performance of the algorithm. Cebron & Berthold [4] found that a \"...larger radius seems to be beneficial in the first iterations, whereas a small radius leads to more regions with high potential. This causes more exploration and leads to a more detailed (but slower) exploration of the datasets, which proves beneficial in later iterations\".",
"cite_spans": [
{
"start": 126,
"end": 129,
"text": "[6]",
"ref_id": "BIBREF5"
},
{
"start": 590,
"end": 593,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Active Learning",
"sec_num": "2.2"
},
{
"text": "In their conclusion, Cebron & Berthold [4] indicate that future work could include tuning the parameters of the PBAC algorithm to a specific problem; in this research, we look at a specific linguistic problem, viz. lemmatization.",
"cite_spans": [
{
"start": 39,
"end": 42,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Active Learning",
"sec_num": "2.2"
},
{
"text": "AL has been applied to a large range of natural language processing (NLP) tasks, including document classification ( [9] , Part-of-Speech tagging [11] , and parse selection [12] . To our knowledge, no literature has been published on using AL in the development of lemmatizers.",
"cite_spans": [
{
"start": 117,
"end": 120,
"text": "[9]",
"ref_id": "BIBREF8"
},
{
"start": 146,
"end": 150,
"text": "[11]",
"ref_id": "BIBREF10"
},
{
"start": 173,
"end": 177,
"text": "[12]",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Lemmatization",
"sec_num": "2.3"
},
{
"text": "Lemmatization is a common NLP task for most languages, and can simply be defined as \"a normalisation step on textual data, where all inflected forms of a lexical word are reduced to its common headword-form, i.e. lemma\" [13] . For example, the grouping of the inflected forms swim, swimming and swam under the base-form swim is seen as an instance of lemmatization. The last part of this definition applies to this project, as the emphasis is on recovering the base-form from the inflected form of the word. The base-form or lemma is the simplest form of a word as it would appear as headword in a dictionary.",
"cite_spans": [
{
"start": 220,
"end": 224,
"text": "[13]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Lemmatization",
"sec_num": "2.3"
},
{
"text": "Our experiments in this paper deal specifically with lemmatization for Afrikaans. Inflection is a productive, but rather simple (in comparison to languages like Spanish or Finnish) morphological process in Afrikaans, with nine basic categories of inflection, viz. plural, diminutive, comparative, superlative, partitive genitive, infinitive, past tense, participle, and attributive. The -e suffix is by far the most frequent affix, occurring across many of these inflec-tional categories [14] . We could therefore predict that words ending on -e would be prototypical examples of inflected words in Afrikaans.",
"cite_spans": [
{
"start": 488,
"end": 492,
"text": "[14]",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Lemmatization",
"sec_num": "2.3"
},
{
"text": "The broad aim of our research is to investigate whether the basic PBAC assumption rings true for linguistic data. Note that we don\"t implement the PBAC algorithm directly; our implementation was developed in parallel with the research of Cebron & Berthold [4] , and is merely related to and compatible with the broad approach taken in the PBAC algorithm. As such, our research can be seen as a contribution towards a better understanding of the representativeness parameter in the PBAC approach.",
"cite_spans": [
{
"start": 256,
"end": 259,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Assumptions",
"sec_num": "3.1"
},
{
"text": "A central assumption of the PBAC approach is that less prototypical cases (i.e. peripheral cases or exceptions) are not important for machine classification, since they do not contribute much information to the construction of a global model. This contrasts directly with our view that, with regard to linguistic data, less prototypical instances are in actual fact important for learning. Our view is based on two grounds: firstly, it is generally accepted in Cognitive Linguistics [6] that outliers contribute as much to the construction of cognitive models as central members (see Section 2.2); hence the interest that Cognitive Linguistics takes in studying not only prototypical instances, but also those peripheral, less prototypical instances of language usage [15] . Secondly, in memory-based language processing [16] it has been argued, on the basis of comparative machine learning experiments on natural language processing data, that exceptions are crucial for obtaining high generalization accuracy. It therefore seems as if the assumption of the PBAC approach is at odds with what is widely believed about natural language data.",
"cite_spans": [
{
"start": 483,
"end": 486,
"text": "[6]",
"ref_id": "BIBREF5"
},
{
"start": 768,
"end": 772,
"text": "[15]",
"ref_id": "BIBREF15"
},
{
"start": 821,
"end": 825,
"text": "[16]",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Assumptions",
"sec_num": "3.1"
},
{
"text": "For purposes of this paper, our assumption is that long words (e.g. manifestations) are less prototypical than short words (e.g. chair); likewise, we assume that low frequency words (e.g. cache) are less prototypical than high frequency words (e.g. chair). (See 2.1 above for a motivation of these assumptions.)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Assumptions",
"sec_num": "3.1"
},
{
"text": "Based on our above stated point of view, we hypothesize that less prototypical linguistic examples should provide better results quicker in AL; this is in contrast with the PBAC approach that would predict that more prototypical instances should provide better results quicker.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Hypothesis",
"sec_num": "3.2"
},
{
"text": "Our basic hypothesis is that, contrary to what is expected from the PBAC approach, adding less prototypical instances to a baseline classifier (seeded with randomly selected data) at the start of the learning process has a bigger impact than adding prototypical instances (specifically with regard to linguistic data). The reason for this is that less prototypical instances (in our case long, low frequency words, such as manifestations) contribute new information to the classifier, and are therefore more informative for learning than prototypical instances (i.e. short, high frequency words, such as chair).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Hypothesis",
"sec_num": "3.2"
},
{
"text": "Similarly, words with high entropy should provide more information to the construction of a classification model than words with low entropy. Hence, using long, low frequency words with high entropy should provide better results in AL.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Hypothesis",
"sec_num": "3.2"
},
{
"text": "Recall that in the PBAC approach the criterion for data selection is calculated as the weighted sum of (1) the representativeness, and (2) the classification uncertainty.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Approach",
"sec_num": "3.3"
},
{
"text": "With regard to (1), we must determine for a certain dataset and/or a certain task what factors determine the representativeness (i.e. prototypicality) of category members. These factors must then be operationalized as density estimates on the unannotated data so that it could be used to select the data point with the highest estimate as prototype. Such operationalizations could be implemented as features of the training instances, or otherwise as organizational principles of the data set. In our current research, our experiments are geared towards the exploration of word frequency and word length as density estimates in the task of lemmatization (see Section 3).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Approach",
"sec_num": "3.3"
},
{
"text": "With regard to (2), we follow the PBAC approach by using entropy as an indicator of the degree of uncertainty or disagreement among different classifiers to assign a class to it, where entropy is inversely related to the voting confidence of the classifier. High entropy therefore indicates higher levels of uncertainty. Words with high entropy are believed to be less prototypical and should therefore be beneficial at the start of the learning process. Entropy correlates with exploitation in the PBAC approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Approach",
"sec_num": "3.3"
},
{
"text": "Our criterion for data selection, called the combination distribution (CD), is calculated as a weighted combination of word frequency, word length and entropy. The formula for the combination distribution is as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Approach",
"sec_num": "3.3"
},
{
"text": ") ( ) ( ) ( 3 2 1 WF w WL w Entropy w CD \uf02b \uf02b \uf03d (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Approach",
"sec_num": "3.3"
},
{
"text": "where w 1 , w 2 and w 3 indicate the different weights.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Approach",
"sec_num": "3.3"
},
{
"text": "For purposes of our experiments, we want to construct a model of the data where we can (a) distinguish between more or less prototypical instances; and (b) select different subsets of the data for AL purposes. In this way, we want to explore, for the task of lemmatization, word frequency and word length as parameters of representativeness, and entropy as an indication of classifier uncertainty. In our experiments, we operationalize these three factors in the following way.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Setup",
"sec_num": "4.1"
},
{
"text": "Concerning word frequency, the data set (see 4.2) is ordered based on frequency counts for the words in the data set, which were calculated on the basis of frequency counts (on types) obtained from an Afrikaans corpus containing more than 160 million tokens [17] . It should be noted that, if frequency is viewed as a density estimate of the distribution of instances in memory, it means that such a distribution will have a high density with regard to low frequency words, which could be viewed as a large group of similar training instances (i.e. in the core of a radial representation). High frequency words (which are less frequent in the data set) will appear, on the other hand, outside the boundaries of this core. Since words appearing inside the boundaries are deemed to be more prototypical than words that fall outside of the boundaries, this representation of the word frequency is so to speak the inverse of a commonsense representation; if viewed as a density estimate, low frequency words are therefore actually more prototypical than high frequency words.",
"cite_spans": [
{
"start": 258,
"end": 262,
"text": "[17]",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Setup",
"sec_num": "4.1"
},
{
"text": "The same argument is also valid when considering word length as a feature. Longer words are less prototypical than shorter words, since the data set contains larger numbers of short words grouped together, than longer words. Short words will appear inside the boundaries and are therefore viewed as more prototypical than longer words. Word length was calculated by counting the number of characters comprising each word in the data set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Setup",
"sec_num": "4.1"
},
{
"text": "Entropy is calculated on the basis of a class distribution obtained from a committee of three different classifiers, each using a different machine learning algorithm. The algorithms that were used are the default TiMBL implementations of IB1, IGTree and IB2 [18] , where k=1. k refers to the k-nearest distances, rather than the k-nearest neighbors. This means the class distribution may contain several instances, despite k having a value of 1. The class distribution therefore consists of all the classes of the instances contained within a distance of 1 from the classified instance, as indicated by the committee of classifiers. The formula used for the calculation of the entropy of a word is shown in Equation 1:",
"cite_spans": [
{
"start": 259,
"end": 263,
"text": "[18]",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Setup",
"sec_num": "4.1"
},
{
"text": ") ( log ) ( ) ( 1 i n i i w p w p w Entropy \uf0e5 \uf03d \uf02d \uf03d (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Setup",
"sec_num": "4.1"
},
{
"text": "where n is the number of classes in the distribution and p(w i ) is the proportional number of a particular class relative to the total number of classes in the class distribution output by the committee.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Setup",
"sec_num": "4.1"
},
{
"text": "Data for the Afrikaans lemmatiser was constructed by extracting word-forms that contain substrings that correspond to inflectional affixes (at the surface level) from an Afrikaans lexicon, together with an equal number of instances where lemma and word-form are equal. The extraction yielded 72,226 instances, which were manually lemmatized as training data. Each instance consisted of 20 features (letters of the word-form as separate features). 271 classes were automatically derived by means of a comparison based on the longest common substring of the extracted word-forms and their manually provided lemmas. The classes indicate the transformation that a wordform must undergo in order to obtain its linguistically correct lemma, specifying the character string to be removed, the relative position of the operation (i.e. L (left), R (right) and M (middle)), and the replacement string. If a word-form and its lemma are identical, the class awarded will be \"0\", denoting the word should be left in the same form. This annotation scheme yields classes like those in the third column of Table 1 . The classifiers are not prohibited from predicting impossible classes (e.g. \"Lge>\" is not a valid class for the word bote, since the word does not containing the string \"ge\").",
"cite_spans": [],
"ref_spans": [
{
"start": 1090,
"end": 1097,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data",
"sec_num": "4.2"
},
{
"text": "Word-form Lemma Class \"geel\" 'yellow'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Table 1. Inflected words with their lemmas and classes as found in the Afrikaans training data",
"sec_num": null
},
{
"text": "\"geel\" 'yellow' 0 \"geslaap\" 'slept'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Table 1. Inflected words with their lemmas and classes as found in the Afrikaans training data",
"sec_num": null
},
{
"text": "\"slaap\" 'sleep\" Lge> \"hondjie\" 'puppy'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Table 1. Inflected words with their lemmas and classes as found in the Afrikaans training data",
"sec_num": null
},
{
"text": "\"hond\" 'dog' Rjie> \"bote\" 'ships'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Table 1. Inflected words with their lemmas and classes as found in the Afrikaans training data",
"sec_num": null
},
{
"text": "\"boot\" 'ship' Rte>ot",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Table 1. Inflected words with their lemmas and classes as found in the Afrikaans training data",
"sec_num": null
},
{
"text": "In all our experiments, we use a k-Nearest Neighbor (k-NN) approach as learner (i.e. memory-based learning). In this approach, classification of a new instance is based on local extrapolation from memorized similar instances. We employed the standard k-NN algorithm, IB1, with default algorithmic parameter settings as implemented in the TiMBL software package [18] .This package also contains implementations of IGTree (a decision tree based approximation of k-NN, and IB2 (a variation of IB1 in which only instances misclassified with the current contents of memory are added to that memory). These variations were used in computing the entropy measure (see 4.1 above).",
"cite_spans": [
{
"start": 361,
"end": 365,
"text": "[18]",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Implementation",
"sec_num": "4.3"
},
{
"text": "Experiments were performed by using the entire data set, consisting of 72,226 words, where every word is a single instance in the training data. 10-fold cross validation was used throughout the evaluation process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Implementation",
"sec_num": "4.3"
},
{
"text": "We started by training the system with a seed memory (10% of the data set) containing randomly-selected instances. We then arranged the remaining instances in the training data set according to the parameters to be evaluated (i.e. word frequency (WF), word length (WL), and entropy, as well as using the combination distribution (CD) described in 3.3 above). In each case the instances were added both in a high-to-low and in a low-to-high order to the learner in sets of 6,500 instances.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Implementation",
"sec_num": "4.3"
},
{
"text": "We are interested in obtaining a learning curve with a steeper gradient than that of the baseline experiment (indicated as \"Random\" in Figure 1 ) in order to show that our data selection method performs better than a random selection. We also want to compare our method with a standard state of the art approach in AL, which we consider to be the committee-based entropy method (indicated as \"Entropy (HL)\" in Figure 1 ).",
"cite_spans": [],
"ref_spans": [
{
"start": 135,
"end": 143,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 410,
"end": 418,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Implementation",
"sec_num": "4.3"
},
{
"text": "The learning curves for the various parameters, the combination distribution, as well as a random distribution (the base-line experiment) are indicated in Figure 1 . Using the same set of randomly selected instances for computing the accuracy obtained in the first fold of every 10-fold crossvalidation experiment results in the learning curves of all experiments starting at the same point on the graph. For the calculation of the combination distribution (see Equation 1), we also experimented with different weight values, but found that the best combination distribution curve was obtained with equal weight values. Figure 1 shows that adding unprototypical words to the seed memory at the start of the learning process clearly outperforms the experiments where prototypical words were added first. This is true for both the evaluated parameters and can be observed by comparing the unprototypical learning curves (e.g. Figure 1 is that the combination distribution (CD) with equal weights yields a steeper learning curve than any of the other individual parameters, including that of the committee-based entropy method.",
"cite_spans": [],
"ref_spans": [
{
"start": 155,
"end": 163,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 620,
"end": 628,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 924,
"end": 932,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "Even though the gains of this approach seem small at first, the significance of our results is appreciated when considering the difference in the number of training instances required by each of the distributions to reach a certain accuracy figure. The combination distribution, for example, requires 19,864 instances to achieve an accuracy of 0.89, compared to the 24,656 instances required by the random distribution to achieve the same accuracy. In this case it means that 4,792 less instances are needed when using the combination distribution, representing a significant saving in terms of the annotation effort.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "Entropy computed on the basis of a committee of classifiers is a popular method for selecting instances in AL (see 2.2 above). Our results indicate that the performance of this method can be improved by combining entropy with other parameters of representativeness, selected on the basis of Prototype Theory. This approach also improves notably upon the random baseline. However, contrary to intuition and to results for AL in other areas than language processing, it is the selection of less prototypical instances first that provides the best improvement, both for word frequency and word length. A possible explanation for this is that language processing tasks have highly disjunc-tive instance spaces, as there are often few regularities and many irregularities, and pockets of exceptions [16] . Starting from a random seeding may already provide sufficient structure (as there is little structure of the instance space), and in such a case, finding the boundary cases is as least as important as finding the central cases of classes. A meta-learning analysis in which the prototype-based selection approach is investigated for a larger range of language processing tasks with class systems of different complexities could shed more light on this issue.",
"cite_spans": [
{
"start": 794,
"end": 798,
"text": "[16]",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
},
{
"text": "Our research shows in any case that a prototypebased selection approach indeed improves upon a committee-based and random baseline approach, but not necessarily the way expected in the PBAC approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
},
{
"text": "In this paper we have shown that a prototype-based selection strategy for AL improves upon both random baseline and entropy-based committee approaches. Interestingly, for our language processing problem, prototypicality works not in the way expected and documented in other research (more prototypical instances first is better than less prototypical first), but exactly the other way round. A possible explanation for this is the lack of structure found in instance spaces of language processing problems, which typically show large disjunctivity.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7."
},
{
"text": "Future work includes the investigation of more natural language processing tasks with more operationalizations of prototypicality to investigate whether our findings indeed point to a different superior selection strategy for language processing tasks than for other types of problems. Another aspect to be investigated is the interaction of this approach with possible stopping criteria for AL.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7."
}
],
"back_matter": [
{
"text": "Van Huyssteen is jointly affiliated with the Human Language Technologies Research Group, Meraka Institute, CSIR, Pretoria, South Africa. Support by the CSIR is hereby acknowledged.We would like to extend our gratitude to JA Pienaar, who was involved in the initial conceptualization of this project.Part of this research was made possible through a research grant by the South African National Research Foundation (GUN: 65462).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": "8."
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Active Learning with Statistical Models",
"authors": [
{
"first": "D",
"middle": [
"A"
],
"last": "Cohn",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Ghahramani",
"suffix": ""
},
{
"first": "M",
"middle": [
"I"
],
"last": "Jordan",
"suffix": ""
}
],
"year": 1996,
"venue": "Journal of Artificial Intelligence Research",
"volume": "4",
"issue": "",
"pages": "129--145",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cohn, D.A., Ghahramani, Z. & Jordan, M.I. 1996. Active Learning with Statistical Models. Journal of Artificial Intelligence Research. 4: 129-145.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Selective sampling using the query by committee algorithm",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Freund",
"suffix": ""
},
{
"first": "H",
"middle": [
"S"
],
"last": "Seung",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Shamir",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Tishby",
"suffix": ""
}
],
"year": 1997,
"venue": "Machine Learning",
"volume": "28",
"issue": "",
"pages": "133--168",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Freund, Y., Seung, H.S., Shamir, E. & Tishby, N. 1997. Selective sampling using the query by commit- tee algorithm. Machine Learning. 28: 133-168.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Domain adaptation with active learning for word sense disambiguation",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Chan",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ng",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 45th Conference of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "49--56",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chan, Y. & Ng, H. 2007. Domain adaptation with active learning for word sense disambiguation. In Proceedings of the 45th Conference of the Associa- tion for Computational Linguistics. pp. 49-56.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Active learning for object classification: from exploration to exploitation",
"authors": [
{
"first": "N",
"middle": [],
"last": "Cebron",
"suffix": ""
},
{
"first": "M",
"middle": [
"R"
],
"last": "Berthold",
"suffix": ""
}
],
"year": 2009,
"venue": "Data Mining and Knowledge Discovery",
"volume": "18",
"issue": "",
"pages": "283--299",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cebron, N. & Berthold, M.R. 2009. Active learning for object classification: from exploration to exploita- tion. Data Mining and Knowledge Discovery. 18: 283-299.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Women, Fire, and Dangerous Things",
"authors": [
{
"first": "G",
"middle": [],
"last": "Lakoff",
"suffix": ""
}
],
"year": 1987,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lakoff, G. 1987. Women, Fire, and Dangerous Things. Chicago: University of Chicago Press.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Words and other wonders: papers on lexical and semantic topics",
"authors": [
{
"first": "D",
"middle": [],
"last": "Geeraerts",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geeraerts, D. 2006. Words and other wonders: pa- pers on lexical and semantic topics. Berlin: Walter de Gruyter.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Sample selection for statistical grammar induction",
"authors": [
{
"first": "R",
"middle": [],
"last": "Hwa",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the Joint SIGDAT Conference on Empirical Methods in NLP and Very Large Corpora",
"volume": "",
"issue": "",
"pages": "45--52",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hwa, R. 2000. Sample selection for statistical gram- mar induction. In Proceedings of the Joint SIGDAT Conference on Empirical Methods in NLP and Very Large Corpora. pp. 45-52.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Less is more: Active learning with support vector machines",
"authors": [
{
"first": "G",
"middle": [],
"last": "Schohn",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Cohn",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 17th International Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "839--846",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Schohn, G. & Cohn, D. 2000. Less is more: Active learning with support vector machines. In Proceed- ings of the 17th International Conference on Machine Learning. San Francisco: Morgan Kaufmann. pp. 839-846.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Active learning with irrelevant examples",
"authors": [
{
"first": "D",
"middle": [],
"last": "Mazzoni",
"suffix": ""
},
{
"first": "K",
"middle": [
"L"
],
"last": "Wagstaff",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Burl",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 17th European Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "695--702",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mazzoni, D., Wagstaff, K.L. & Burl, M. 2006. Active learning with irrelevant examples. In Proceedings of the 17th European Conference on Machine Learning. pp. 695-702.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Minimizing manual annotation cost in supervised training from corpora",
"authors": [
{
"first": "S",
"middle": [],
"last": "Engelson",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Dagan",
"suffix": ""
}
],
"year": 1996,
"venue": "Proceedings of the 34th Meeting of the Association for Computational Natural Language Learning",
"volume": "",
"issue": "",
"pages": "319--326",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Engelson, S. & Dagan, I. 1996. Minimizing manual annotation cost in supervised training from corpora. In Proceedings of the 34th Meeting of the Association for Computational Natural Language Learning. San Fransico: Morgan Kaufmann. pp. 319-326.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Active learning and the total cost of annotation",
"authors": [
{
"first": "J",
"middle": [],
"last": "Baldridge",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Osborne",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "9--16",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Baldridge, J. & Osborne, M. 2004. Active learning and the total cost of annotation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing. Barcelona, Spain. pp. 9-16.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Outomatiese Lemma-identifisering vir Afrikaans",
"authors": [
{
"first": "H",
"middle": [
"J"
],
"last": "Groenewald",
"suffix": ""
},
{
"first": "G",
"middle": [
"B"
],
"last": "Van Huyssteen",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Groenewald, H.J. & Van Huyssteen, G.B. 2008. Ou- tomatiese Lemma-identifisering vir Afrikaans [Automatic Lemmatisation for Afrikaans].",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Cognitive Grammar: A Basic Introduction",
"authors": [
{
"first": "R",
"middle": [
"W"
],
"last": "Langacker",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Langacker, R.W. 2008. Cognitive Grammar: A Basic Introduction. Oxford: Oxford University Press.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Memorybased language processing",
"authors": [
{
"first": "W",
"middle": [],
"last": "Daelemans",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Van Den",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bosch",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daelemans, W. & Van den Bosch, A. 2005. Memory- based language processing. Cambridge: Cambridge University Press.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Media24 Corpus. Me-dia24: Cape Town",
"authors": [
{
"first": "Pharos",
"middle": [],
"last": "Dictionaries",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pharos Dictionaries. 2007. Media24 Corpus. Me- dia24: Cape Town.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "TiMBL: Tilburg Memory Based Learner, version 5.1, Reference Guide",
"authors": [
{
"first": "W",
"middle": [],
"last": "Daelemans",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Zavrel",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Van Der Sloot",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Van Den",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bosch",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daelemans, W., Zavrel, J., Van der Sloot, K. & Van den Bosch, A. 2004. TiMBL: Tilburg Memory Based Learner, version 5.1, Reference Guide. ILK Technic- al Report 04-02. Tilburg: University of Tilburg.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Learning curves",
"uris": null,
"num": null,
"type_str": "figure"
},
"FIGREF1": {
"text": "Word Frequency [Low to High] and Word Length [Long to Short] with the prototypical curves (e.g. Word Frequency [High to Low] and [Word Length Short to Long]). (With regard to the learning curves representing word frequency, refer to 4.1 for an explanation of why [High to Low] is indicated as better than [Low to High] in Figure 1.) Another finding from",
"uris": null,
"num": null,
"type_str": "figure"
}
}
}
} |