File size: 58,163 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 | {
"paper_id": "R09-1012",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T15:00:37.275335Z"
},
"title": "Combining Finite State and Corpus-based Techniques for Unknown Word Prediction",
"authors": [
{
"first": "Kostadin",
"middle": [],
"last": "Cholakov",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Groningen",
"location": {
"postBox": "PO Box 716",
"postCode": "9700 AS",
"settlement": "Groningen",
"country": "The Netherlands"
}
},
"email": "k.cholakov@rug.nl"
},
{
"first": "Gertjan",
"middle": [],
"last": "Van Noord",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Groningen",
"location": {
"postBox": "PO Box 716",
"postCode": "9700 AS",
"settlement": "Groningen",
"country": "The Netherlands"
}
},
"email": "g.j.m.van.noord@rug.nl"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Many NLP systems make use of various lexicons and dictionaries. However, unknown words are a major problem for such resources when applied to real-life data. We propose a method that combines nite state techniques and web queries to deliver possible analyses for a given unknown word and to generate its paradigm. We ensure the general applicability of our approach by applying it to a test set of Dutch words.",
"pdf_parse": {
"paper_id": "R09-1012",
"_pdf_hash": "",
"abstract": [
{
"text": "Many NLP systems make use of various lexicons and dictionaries. However, unknown words are a major problem for such resources when applied to real-life data. We propose a method that combines nite state techniques and web queries to deliver possible analyses for a given unknown word and to generate its paradigm. We ensure the general applicability of our approach by applying it to a test set of Dutch words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Unknown words are a major hindrance for the performance of NLP tools that make use of lexicons and dictionaries. To overcome this problem, most applications try to extract (partially) the necessary morphological knowledge by implementing various heuristics and unknown word guessers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we present a two-phase method which delivers an accurate morphological analysis for a given unknown word and generates its paradigm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "It deals with open-class words nouns, adjectives and verbs based on the assumption that the other word classes are already covered by most lexicons. We test its eciency and accuracy by applying it to real-life Dutch data. Dutch is a language with a productive inectional morphology that exhibits quite a few interesting phenomena and thus poses a challenge for morphological processing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In the rst phase we use nite state techniques which are one of the most common approaches for morphological processing (Beesley and Karttunen, 2003; Petitpierre and Russel, 1995) since they can conveniently be used for both analysis and generation. We employ a small set of non-deterministic`unweighted'nite state transducers (FSTs) whose manually encoded rules cover regular morphological phenomena.",
"cite_spans": [
{
"start": 119,
"end": 148,
"text": "(Beesley and Karttunen, 2003;",
"ref_id": "BIBREF1"
},
{
"start": 149,
"end": 178,
"text": "Petitpierre and Russel, 1995)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Since our method deals with unknown words, it does not have access to any additional information apart from the limited knowledge provided by the word structure. Restricted only by that limited knowledge the FSTs, in analysis mode, identify all possible forms and roots allowed by the word structure. For example, the word schnabbel (a gig) is analysed as a singular noun, a base adjective and a rst person singular present verb. As a consequence, three possible roots are produced. Since there is no way to know which of them is the correct one, in generation mode, all possible paradigms for each of these roots are generated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The problem of disambiguating the output of the FSTs is dealt with in the second phase. We use Yahoo to search the web for each root and generated paradigm form and, based on the number of occurrences found, we try to identify the correct root and paradigm of the unknown word. Commercial search engines have already been successfully used for various NLP tasks (Keller and Lapata, 2003) and it is our claim that they are sucient for ours as well. Since the whole paradigm of the unknown word is generated, it would be very dicult to nd a large number of occurrences for each form in a wrong paradigm. For example, the generated adjective and verb forms for schnabbel have no or very few occurrences on the web and they can be safely rejected.",
"cite_spans": [
{
"start": 362,
"end": 387,
"text": "(Keller and Lapata, 2003)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A similar approach, described in (Adolphs, 2008) , applies nite state techniques to generate possible inectional classes for unknown German words. However, disambiguation is done by using metrics based on frequency counts obtained from a corpus. Thus disambiguation depends heavily on the size and the gender of the corpus which is a drawback in comparison with the virtually unlimited data in the web our method has access to. If a word is, for instance, both a noun and a verb, it is possible that it would occur only as a noun in a given corpus and the method would fail to deal with the morphological ambiguity. (Nakov et al., 2003) use a rule-based approach to guess the morphological classes of unknown German nouns where each induced rule is ranked in the manner of (Mikheev, 1997) . However, it is not clear if the method can scale to other word classes. (van den Bosch and Daelemans, 1999) apply memory-based learning to provide a detailed morphological analysis of Dutch. The method is tested on frequent dictionary words and only an estimate is provided about its expected performance on real-world data.",
"cite_spans": [
{
"start": 33,
"end": 48,
"text": "(Adolphs, 2008)",
"ref_id": "BIBREF0"
},
{
"start": 616,
"end": 636,
"text": "(Nakov et al., 2003)",
"ref_id": "BIBREF7"
},
{
"start": 773,
"end": 788,
"text": "(Mikheev, 1997)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We should mention that the work described here is part of an algorithm for the automated acquisition of lexical types for words unknown to the Alpino grammar and parser (van Noord, 2006) . The information provided by the generated paradigms is used as features in a statistical classier which predicts lexical types for each unknown word. We also take into account occurrences of the unknown word in dierent contexts to extract additional features, including the type(s) assigned by the Alpino POS tagger (Prins and van Noord, 2001 ) and types which Alpino allows as plausible in the particular context. Therefore, both the morphology of the unknown word and its context are considered in the prediction process. For more details, see (Cholakov, 2009) .",
"cite_spans": [
{
"start": 169,
"end": 186,
"text": "(van Noord, 2006)",
"ref_id": "BIBREF10"
},
{
"start": 505,
"end": 531,
"text": "(Prins and van Noord, 2001",
"ref_id": "BIBREF9"
},
{
"start": 735,
"end": 751,
"text": "(Cholakov, 2009)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The remainder of the paper is organised as follows. Section 2 presents the morphological phenomena which are relevant for our experiments. Section 3 describes the FSTs and investigates their coverage and degree of non-determinism. Section 4 describes the web heuristics used to disambiguate the output of the FSTs and presents the experiments with the test data. Section 5 concludes the paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this section we present the morphological phenomena which the FSTs account for. We begin by presenting some rules that have eect on all the three POS classes we consider.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "In Dutch, the vowels a, e, o, u and i are either long or short. A vowel is long if it is doubled (e.g., maan moon), if it is in a vowel combination (lief sweet, dear) or when it is at the end of a syllable (maken to make). The general rule is that the type of a vowel, short or long, is preserved in all word forms. In particular contexts, a vowel is kept long by doubling it:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "(1) ma-ken maak (to make, INF1st PER.SG.PRES)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "After removing the -en sux from the innitive, we get the form *mak and the vowel turns into a short one. To prevent this, a is doubled. When en is added to form plural in (2), u would become a long vowel *stu-ken. To prevent this, the following consonant is doubled stuk-ken.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "(2) stuk stukken (piecepieces) However, there are some exceptions to these rules. Depending on whether the syllable containing the vowel in question is stressed or not, the type of the vowel can change for words with stems ending in -el, -er and -ig. If the syllable is stressed, then the vowel preserves its type in all word forms as shown in (3-a). If the respective syllable is not stressed, then the vowel is not doubled and it turns into a short one as in (3-b).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "(3) a.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "de-len deel (partspart) b.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "re-ge-len re-gel (rulesrule)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "The same also applies to the doubling of consonants.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "In (4-a) r is doubled to keep the vowel short. However, this is not the case in (4-b) because the stress falls on another syllable and thus, e turns into a long vowel.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "a. sper sper-ren (to bar, 1st PER.SG.PRESINF) b.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "coun-ter coun-te-ren (to strike back, 1st PER.SG.PRESINF)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "One last important rule is that a morpheme cannot end in -v or -z and they are replace by f and s, respectively: rei-zenreis (to travel, INF1st PER.SG.PRES) and le-venleef (to live, INF1st PER.SG.PRES).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "Noun Inection. Most Dutch nouns form plural by adding -en to the singular form, as shown in (2). However, some nouns take -s to form their plural: jongen-jongens (boy-boys), tram-trams.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "Adjective Inection. Most adjectives in Dutch have base, comparative and superlative forms. Comparative is normally formed by adding -er to the base form: snel -sneller (fast-faster). Superlative is formed by adding the sux -st to the base form: snel -snelst. However, base forms that end in -s take only -t to form superlative:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "geeps -geepst (palethe most pale)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "Additionally, when adjectives are used attributively, they get an -e sux. The only exception is when they precede a neutral noun which is not used with a denite article or a pronoun. Some adjectives, for example the ones ending in -en which are mostly adjectives denoting material, do not exhibit that kind of inection: de gouden ring (the gold ring).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "Verb Inection. The starting point for verb inection is the verb stem. The only thing which needs to be explained in connection with the results of our experiments is the formation of the past participle (psp). It is formed by adding the prex ge-to the stem and, depending on the nal consonant of the stem, either a t or d sux is attached. However, if the stem already ends in -t or -d, no sux is added. An exception to these rules are verbs with separable particles which form psp by inserting ge between the separable particle and the verb stemopgeruimd (to clean). Verbs starting with be-, er-, ge-, her-, mis-, ont-and ver-form psp without ge-: vertel -verteld (to tell). These particles are also known as inseparable. For a detailed discussion of Dutch morphology, see (de Haas and Trommelen, 1993) .",
"cite_spans": [
{
"start": 773,
"end": 802,
"text": "(de Haas and Trommelen, 1993)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Morphological Phenomena",
"sec_num": "2"
},
{
"text": "We employ a set of FSTs to cover the morphological phenomena presented in the previous section. For example, if the input word is stukken from (2), the transducer for plural nouns should produce stuk as a possible root form in analysis mode, and then, given this root form, it should output stukken in generation mode. In our experiments, the root of a given noun is its singular form, the root of an adjective its base form, and the one of a verb is its stem.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "We have used the Stuttgart Finite State Transducer (SFST) tools to implement and run a number of separate transducers which are shown in Table 1 We use words from the CELEX morphological database (CELEX, 1995) as a development set in order to: i) investigate if all target phenomena are covered by the FSTs and ii) to have some notion of their degree of non-determinism. CELEX contains about 380K word forms corresponding to nearly 125K headwords. It is a very suitable resource for our purposes since it covers a large number of dierent morphological phenomena.",
"cite_spans": [
{
"start": 196,
"end": 209,
"text": "(CELEX, 1995)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 137,
"end": 144,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "Three word sets are randomly selected: 2000 plural nouns, 2000 superlative adjectives and 2000 rst person singular verbs and they are processed with the FSTs. We take superlative adjective forms and plural nouns in order to ensure that we deal with comparable adjectives and countable nouns. The paradigms of the selected words are extracted from CELEX, so we can check the paradigms generated by the FSTs against them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "The results for the analysis of the three extracted word sets are given in Table 2 . First, the words of each set are analysed with the respective transducer and the output is a set of possible root forms for each word. The number of analyses is divided by the number of analysed words to get the analyses per word ratio. This indicates how deterministic the applied transducer is in analysis mode.",
"cite_spans": [],
"ref_spans": [
{
"start": 75,
"end": 82,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "Next, we use the candidate roots to generate paradigms for each word and we check them against the paradigm we extracted from CELEX for the respective word. If the correct paradigm was found, the root which it was generated from is saved and thus, a list of correct roots for each of the three word sets is produced. The words which failed to be analysed are very irregular forms which makes it impossible for an analysis to be produced, e.g. vaklui -vakman (experts-expert), since our method deals only with regular inection.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "The FSTs do not have access to information about the word stress and therefore, cases like (3-a) and (3-b) are ambiguous because it is not clear whether the vowel should be doubled or not. For example, the possible roots for regelen would be *regeel and regel.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "The number of analyses for adjectives is so high because it is not clear whether the word is an adjective that ends in s and takes only -t to form superlative as shown in (5). If so, there is also no way to know if the root form ends in -s or if s is a replacement for a -z. Thus, for almost each adjective three analyses will be delivered boost (angry): *boo, *boos and booz. This is also the reason for the non-determinism of the rst person singular verb transducer the output of the analysis for leef is leev and *leef.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "There are also words for which no correct paradigms are generated. This is due to the fact that those words have at least one irregular form in their paradigmsschip-schepen (ship-ships). However, we do not see these irregularities as an obstacle for the performance of our method since they form a closed class and are supposed to be already included in most lexicons and dictionaries.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "Next, each of the three lists with correct roots is processed by the transducers for the respective POS in generation mode to investigate the non-determinism of the generation. Except for the psp transducer, the nondeterminism is caused by cases like (4-a) and (4-b) where, depending on which syllable is stressed, the nal l, r and g can be doubled or not. The nondeterminism of the psp generator is due to the fact that some of the verb particles can be both separable and inseparable. For example, we have omklemd (to clasp) but there is also omgekanteld (to tip over). Therefore, both *omkanteld and omgekanteld are generated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Finite State Morphology",
"sec_num": "3"
},
{
"text": "Since we showed that the FSTs cover all target morphological phenomena, the only remaining issue is to disambiguate their non-deterministic output. We resolve this problem by using Yahoo to obtain the search hits for a given root and the word forms it generates. If a given form is found more times than a certain threshold (500 in our experiments), it is very likely that the form is a valid member of the paradigm. A paradigm is considered valid if all its forms have passed the threshold. Further, we limit the search only to pages considered by Yahoo to be in Dutch and which are from the Netherlands. In order to evaluate the performance of our method, a test set that consists of nouns, adjectives and verbs which have between 40 and 100 occurrences in large Dutch newspaper corpora (\u223c530M words) has been created. This selection is based on the assumption that unknown words are typically less frequent. The corpora have already been parsed with the Alpino parser and many words have been added to the lexicon of the Alpino grammar. The lexical entries of the chosen words provide our gold standard.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "Verbs are the easiest POS for processing with our method because in most cases, all 6 forms sg1, sg2/3, pl/inf, past-sg, past-pl and psp are dierent and the chance of nding enough counts for all forms in a wrong paradigm is minimal. We start by checking if there are enough search hits for the root, i.e. the sg1 form in this case. In order to be sure that the occurrences found are actually the ones of the root, the query sent to Yahoo consists of the rst person singular personal pronoun and the root ik + root.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "However, it is very dicult to capture cases like opruimen with that query since the separable particle occurs at the end of the sentence. To deal with this, we automatically check whether the given candidate root starts with a separable particle and if so, we ignore that particle and search only for the verb stem ik ruim.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "Since the combination of ik and the root form can be very rare for some verbs, we set a threshold of only 50 occurrences. Even if a wrong root is able to get through, it would be always discarded, if there are not enough search hits for one of the forms it generates. The same threshold is also used for the past singular and the past plural forms since those are not that frequently used in Dutch.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "For adjectives, we want to nd the base, comparative and superlative forms and their inected counterparts. However, not every adjective has all six forms. Non-comparable adjectives like amorf (amorphous) have only the base and the base inected form. There are also some adjectives like romantisch (romantic) which form superlative by using the word meest (most) in front of the base form.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "We start again by looking if there are enough search hits for the proposed root which is the base adjective form in this case. If all six forms are found, then the generated paradigm is taken to be the nal result. However, in order to be able to deal with cases like romantisch, we also allow for paradigms where only the two base and the two comparative forms are found.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "The adjectives for which no paradigm has been generated are processed again but this time we assume that these adjectives are non-comparable and thus, we search only for the two base forms. If their counts are above the threshold, the paradigm is considered to be valid. We do this in a separate round in order to avoid situations where a wrong root is able to produce two valid base forms due to chance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "Nouns are the most complicated POS to process because there are only 2 forms, singular and plural, which makes it more dicult to obtain reliable results. One example is schel -schellen (bell-bells). If the input word is schellen, there are two possible roots for it schel and *schellen. In the latter case, there are enough hits for the generated plural form *schellens because schellen happens to be a family name and schellens is its genitive form Schellens methode (Schellen's method). Thus a wrong paradigm will be generated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "To prevent this, we search for the combination of the indenite article een and the root. However, not all nouns combine with een (e.g. mass nouns) and that is why, if there are not enough search hits, we also search for the combination of the root and is (to be, 3rd PER.SG.PRES). The queries for een schellen and schellen is return less than 100 hits and this root is correctly discarded.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "Once a root form has passed the threshold, we can also determine the denite article of the noun. Depending on their gender and number, Dutch nouns are used either with the de denite article (for masculine and feminine, and also plural) or the het article (for neuter). We search for occurrences of de + root and het + root and return the article with the higher number of search hits.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "The query for the generated noun plural form is de+plural. Since the generation is not deterministic, there might be more than one plural form which passes the threshold. In this case, the most frequent one is taken to be the nal output. If there is more than one root that was able to generate a paradigm, the paradigm of the one with the most search hits is taken to be the nal outcome.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "However, there are many compounds whose root forms, combined with the indenite article have very low number of occurrences: Marshall-plan (the Marshall plan), zaak-Bouterse (the Bouterse case), zendingswerk (missionary work). Normally, the compound head is a common word and the paradigm could be generated by processing the head instead of the whole compound. In most of the cases, the head is the rightmost part of the compound plan in Marshall-plan and werk (work) in zendingswerk.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "For words which are joined by a hyphen, splitting the compound is straightforward the word on the right side is considered to be the head of the compound. However, if this word starts with a capital letter, like in zaak-Bouterse, we assume that the head of the compound is on the left side since the right part is most probably a name. For compounds without a hyphen, we take the chunk from the third letter to the end of the word. If there are at least 10,000 search hits for this chunk, it is considered to be the head of the compound. Otherwise, the chunk from the fourth letter is taken and so on, until we nd a chunk with enough occurrences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "Let us put it all together. The query for the root is sent but the threshold set for it is 100 because nouns tend to occur less frequently together with the indenite article. After a valid root is found, the denite article is determined and then the search hits for the generated plural form(s) are obtained. If no paradigm was generated for a given word, we assume that it might be a low-frequent compound and we try to split it. If there is a valid head, the same procedure is applied to it in an attempt to generate its paradigm which is also the paradigm of the whole compound.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Web Search Heuristics",
"sec_num": "4.1"
},
{
"text": "The test set, described in Section 4.1, consists of 2593 unique words but the gold standard contains a total of 2781 entries 1368 nouns, 729 adjectives and 684 verbs. This is due to the fact that some morphologically ambiguous words have more than one valid paradigm. For instance, many psp can also act as adjectives: gebruind is the psp of the verb bruinen (to turn/make brown) but it is also an adjective`tanned, sunburnt'. That is why this word is listed both as an adjective and a verb. There are 188 cases of morphological ambiguity. However, our method is able do deal with them. When an input word comes in, the FSTs for each POS try to process it and if some of them manage to generate a paradigm, the web heuristics for the respective POS are applied. After the disambiguation phase, only the verb and the adjective paradigms for gebruind survived' and they are correctly taken to be the nal outcome. However, in very rare cases there is morphological ambiguity within the same POS. For instance, the word kussen is the plural form of kus (a kiss) but it also means`a cushion' whose plural form is kussens. In that case, only the paradigm whose root form has more search hits is returned. No paradigm has been generated for the uncountable noun smelt (smelt). All other nouns without paradigms are compounds which form plural in an irregular way. The same is also valid for the verbs which have not been covered all of them have irregular past or psp forms which are not handled by our technique. Most of the adjectives not covered are non-comparable adjectives ending in -en and designating material, e.g. satijnen (satin). Such adjectives have only a base form and thus, no valid paradigm is found for them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Error Analysis",
"sec_num": "4.2"
},
{
"text": "Next, we see in Table 4 that some of the generated noun and verb paradigms are wrong. Most of the wrong noun ones are compounds whose head is the word kind (child). For example, the compound pleegkind (foster child) is correctly split and we try to generate a paradigm for its head kind. However, it has an irregular plural form kinderen. Nevertheless, the`regular' plural form *kinden has more than 1000 occurrences in the web and it is taken to be a valid one. A manual examination of the results of the web search showed that all the occurrences were actually a typo the actual meaning was kind en (child and) but in many web forums, blogs, etc. they were written as one word.",
"cite_spans": [],
"ref_spans": [
{
"start": 16,
"end": 23,
"text": "Table 4",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Results and Error Analysis",
"sec_num": "4.2"
},
{
"text": "Another source of error for the noun paradigms are words from French origin, e.g. secretaris-generaal (secretary-general). In this case the head of the compound is on the left of the hyphen and thus it is the part that is inected. The right plural form is secretarissen-generaal. However, our compound splitting heuristics takes the right part to be the head which, in this case, happens to be the noun generaal (general). Its paradigm is generated but this is not the correct paradigm of the compound.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Error Analysis",
"sec_num": "4.2"
},
{
"text": "The only verb with a wrong paradigm generated is the irregular verb schijten (to shit). The incorrect psp generated by the psp FST, geschijt, has enough search hits because it happens to be a less frequent case of nominalisation namely, ge + root. This error, however, can be tolerated, since it is potentially possible in the rare case of irregular verbs which allow for such nominalisation and have roots that end in -t or -d.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Error Analysis",
"sec_num": "4.2"
},
{
"text": "The achieved results show clearly that simple but carefully designed web queries can be successfully used to disambiguate the output of a non-deterministicnite state morphology for Dutch. Our method is able to deal with all major and common morphological phenomena except for few cases that include irregular forms or very rare combinations of factors.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Error Analysis",
"sec_num": "4.2"
},
{
"text": "We proposed a combination of nite state techniques and specially designed web queries and heuristics to deliver morphological analyses for a given unknown word and to generate its paradigm. Our approach does not require access to sophisticated linguistic information but instead employs the web as a linguistic resource. The successful application of the method to real-life Dutch data proved its eciency and high accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
},
{
"text": "Naturally, languages with a large number of inectional variants would be more problematic for our approach due to the much larger morphological ambiguity they exhibit. However, we also expect that the high number of forms in the paradigms would facilitate the disambiguation of the FSTs output. As we showed, it is much easier to validate a paradigm that consists of 6 forms since the web results are more reliable for it. We will investigate the scalability of our method in future research.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Acquiring a poor man's inectional lexicon for German",
"authors": [
{
"first": "P",
"middle": [],
"last": "Adolphs",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 6th International Conference on Language Resources and Evaluation (LREC)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Adolphs, P. (2008). Acquiring a poor man's inec- tional lexicon for German. In Proceedings of the 6th International Conference on Language Resources and Evaluation (LREC), Marrakech, Morocco.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Finite State Morphology",
"authors": [
{
"first": "K",
"middle": [
"R"
],
"last": "Beesley",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Karttunen",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Beesley, K. R. and Karttunen, L. (2003). Finite State Morphology. CSLI Publications, Palo Alto, CA, USA.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "The CELEX lexical database-Dutch",
"authors": [],
"year": 1995,
"venue": "CELEX",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "CELEX (1995). The CELEX lexical database-Dutch, English, German. CD-ROM.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Towards morphologically enhanced automated lexical acquisition",
"authors": [
{
"first": "K",
"middle": [],
"last": "Cholakov",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the 14th ESSLLI Student Session",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cholakov, K. (2009). Towards morphologically en- hanced automated lexical acquisition. In Proceed- ings of the 14th ESSLLI Student Session, Bordeaux, France.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Morfologisch Handboek van het Nederlands: Een overzicht van de woordvorming. SDU, 's Gravenhage, The Netherlands",
"authors": [
{
"first": "W",
"middle": [],
"last": "De Haas",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Trommelen",
"suffix": ""
}
],
"year": 1993,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "de Haas, W. and Trommelen, M. (1993). Morfologisch Handboek van het Nederlands: Een overzicht van de woordvorming. SDU, 's Gravenhage, The Nether- lands.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Using the web the obtain frequencies for unseen bigrams",
"authors": [
{
"first": "F",
"middle": [],
"last": "Keller",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Lapata",
"suffix": ""
}
],
"year": 2003,
"venue": "Computational Linguistics",
"volume": "29",
"issue": "3",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Keller, F. and Lapata, M. (2003). Using the web the obtain frequencies for unseen bigrams. Computa- tional Linguistics, 29(3):459484.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Automatic rule induction for unknown-word guessing",
"authors": [
{
"first": "A",
"middle": [],
"last": "Mikheev",
"suffix": ""
}
],
"year": 1997,
"venue": "Computational Linguistics",
"volume": "23",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mikheev, A. (1997). Automatic rule induction for unknown-word guessing. Computational Linguist- ics, 23.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Guessing morphological classes of unknown German nouns",
"authors": [
{
"first": "P",
"middle": [],
"last": "Nakov",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Bonev",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Angelova",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Gius",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Hahn",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of RANLP 2003",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nakov, P., Bonev, Y., Angelova, G., Gius, E., and von Hahn, W. (2003). Guessing morphological classes of unknown German nouns. In Proceedings of RANLP 2003, Borovets, Bulgaria.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Mmorph the multext morphology program",
"authors": [
{
"first": "D",
"middle": [],
"last": "Petitpierre",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Russel",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Petitpierre, D. and Russel, G. (1995). Mmorph the multext morphology program. Technical report, Ca- rouge, Switzerland.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Unsupervised POS-tagging improves parcing accuracy and parsing eciency",
"authors": [
{
"first": "R",
"middle": [],
"last": "Prins",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Van Noord",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Van Den",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bosch",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Daelemans",
"suffix": ""
}
],
"year": 1999,
"venue": "Proceedings of the 37th Annual Meeting of the ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Prins, R. and van Noord, G. (2001). Unsupervised POS-tagging improves parcing accuracy and parsing eciency. In Proceedings of IWPT, Beijing, China. van den Bosch, A. and Daelemans, W. (1999). Memory-based morphological analysis. In Proceed- ings of the 37th Annual Meeting of the ACL, pages 285292, San Francisco, CA.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "At last parsing is now operational",
"authors": [
{
"first": "G",
"middle": [],
"last": "Van Noord",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of TALN",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "van Noord, G. (2006). At last parsing is now opera- tional. In Proceedings of TALN, Leuven, Belgium.",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"html": null,
"content": "<table><tr><td>POS</td><td>transducers</td></tr><tr><td colspan=\"2\">nouns singular, plural</td></tr><tr><td>adj</td><td>base, comparative, superlative</td></tr><tr><td>verbs</td><td>sg1, sg2/3, pl/inf, past-sg, past-pl, psp</td></tr></table>",
"type_str": "table",
"num": null,
"text": "."
},
"TABREF1": {
"html": null,
"content": "<table/>",
"type_str": "table",
"num": null,
"text": "Transducers used"
},
"TABREF3": {
"html": null,
"content": "<table/>",
"type_str": "table",
"num": null,
"text": "Analysis mode results"
},
"TABREF4": {
"html": null,
"content": "<table><tr><td>transducer</td><td>forms/root</td></tr><tr><td>plural</td><td>1.21</td></tr><tr><td>base</td><td>1.21</td></tr><tr><td>comparative</td><td>1.38</td></tr><tr><td>pl/inf</td><td>1.36</td></tr><tr><td>psp</td><td>1.11</td></tr></table>",
"type_str": "table",
"num": null,
"text": "shows the average number of generated forms per root for the transducers with nondeterministic generation."
},
"TABREF5": {
"html": null,
"content": "<table/>",
"type_str": "table",
"num": null,
"text": "Generation: non-deterministic transducers"
},
"TABREF6": {
"html": null,
"content": "<table><tr><td/><td>overall</td><td>nouns</td><td>adjectives</td><td>verbs</td></tr><tr><td>total</td><td>2781</td><td>1368</td><td>729</td><td>684</td></tr><tr><td>coverage</td><td>96.55%</td><td>98%</td><td>98.91%</td><td>90.94%</td></tr><tr><td colspan=\"3\">accuracy 99.63% 99.33%</td><td>100%</td><td>99.84%</td></tr></table>",
"type_str": "table",
"num": null,
"text": "presents the overall results and the results for each POS. Coverage indicates the number of words with a paradigm generated and accuracy is the number of the generated correct paradigms."
},
"TABREF7": {
"html": null,
"content": "<table/>",
"type_str": "table",
"num": null,
"text": "Web experiment results"
}
}
}
} |