File size: 61,498 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 | {
"paper_id": "R13-1012",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T14:55:12.553219Z"
},
"title": "A Unified Lexical Processing Framework Based on the Margin Infused Relaxed Algorithm. A Case Study on the Romanian Language",
"authors": [
{
"first": "Tiberiu",
"middle": [],
"last": "Boro\u0219",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Research Institute for Artificial Intelligence, \"Mihai Dr\u0103g\u0103nescu\"",
"location": {
"country": "Romanian Academy"
}
},
"email": "tibi@racai.ro"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "General natural language processing and text-to-speech applications require certain (lexical level) processing steps in order to solve some frequent tasks such as lemmatization, syllabification, lexical stress prediction and phonetic transcription. These steps usually require knowledge of the word's lexical composition (derivative morphology, inflectional affixes, etc.). For known words all applications use lexicons, but there are always out-of-vocabulary (OOV) words that impede the performance of NLP and speech synthesis applications. In such cases, either rule based or data-driven techniques are used to automatically process these OOV words and generate the desired results. In this paper we describe how the above mentioned tasks can be achieved using a Perceptron with the Margin Infused Relaxed Algorithm (MIRA) and sequence labeling.",
"pdf_parse": {
"paper_id": "R13-1012",
"_pdf_hash": "",
"abstract": [
{
"text": "General natural language processing and text-to-speech applications require certain (lexical level) processing steps in order to solve some frequent tasks such as lemmatization, syllabification, lexical stress prediction and phonetic transcription. These steps usually require knowledge of the word's lexical composition (derivative morphology, inflectional affixes, etc.). For known words all applications use lexicons, but there are always out-of-vocabulary (OOV) words that impede the performance of NLP and speech synthesis applications. In such cases, either rule based or data-driven techniques are used to automatically process these OOV words and generate the desired results. In this paper we describe how the above mentioned tasks can be achieved using a Perceptron with the Margin Infused Relaxed Algorithm (MIRA) and sequence labeling.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Natural Language Processing (NLP) applications and Text-to-Speech (TTS) synthesis systems require a set of pre-processing steps that include tasks such as lemmatization, syllabification, lexical stress prediction and phonetic transcription. Because these all these tasks require knowledge of the word composition (derivative morphology, inflectional affixes, part of speech, etc.) we will refer to them as lexical processing steps.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This paper presents a unified lexical processing framework based on the Margin Infused Relaxed Algorithm (MIRA) (Crammer and Singer, 2003) designed to solve the basic text-preprocessing tasks involved in both text-to-speech (TTS) synthesis and general NLP applications. Assuming that all existing systems use lexicons for known words, we focused our research in handling the difficult problems generated by presence of out-of vocabulary (OOV) or previously unseen in the training data words that negatively impact the performance of the above mentioned tasks. Our current research is focused on the Romanian language, but the methods presented here are data-driven and with proper lexicons and feature templates, they can be used for other (Latin based) languages as well. We show how we achieved state-of-the-art results on Romanian by using the MIRA framework.",
"cite_spans": [
{
"start": 112,
"end": 138,
"text": "(Crammer and Singer, 2003)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "There are various methods proposed in the literature for each of the previously mentioned lexical subtasks. For each of them, we will offer a short literature review of available methods and we will compare our results with the current state-of-art systems.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lexical processing with MIRA",
"sec_num": "2"
},
{
"text": "The previously proposed methods vary from rule-based to data-driven and different authors employ different classifiers (in data-driven approaches), such as Maximum Entropy Classifiers, Classification and Regression Trees, Support Vector Machines (SVM), Structured SVMs, Conditional Random Fields, etc. While these are all powerful methodologies, we chose the Perceptron classifier with the MIRA update learning as our sequence labeling classifier because of its robustness and its ability to obtain highly accurate results that compare to the ones obtained using CRFs. All the lexical processing methods that we propose, share the following similarities:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lexical processing with MIRA",
"sec_num": "2"
},
{
"text": "-All of them are reformulated as sequence labeling tasks; -We use the same classifier for all our tasks (MIRA); -The classification context is based on different and mostly lexical (except for lemmatization and lexical stress prediction, which use the morphosyntactic) feature sets; -The performance is measured in terms of word accuracy rates (WAR); -All the tests are reported on OOV words, as we assume that all systems use lookup lexicons for known words; -All our tests are performed on Romanian and we report the feature sets that yielded the best results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lexical processing with MIRA",
"sec_num": "2"
},
{
"text": "Syllabification is the process of decomposing words into their phonological units, which is an important requirement in modern approaches to TTS synthesis and speech recognition. All languages have phonetic rules that govern the syllabification process, but it is often the case that these rules are contradicted by etymological principles, a fact which complicates the task of automatic syllabification. Phonetic transcription (letter to sound -L2S) or the position of the lexical stress both provide useful information for syllabification, but more often than not, L2S and lexical stress are not accurate enough on OOV words to help the syllabification process. Also, syllabification lexicons are usually larger than L2S lexicons, thus providing more training data, which helps the syllabification system obtain better results than L2S. Because of the above mentioned reasons, we strictly based our method on purely lexical features (i.e. the word's letters).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification",
"sec_num": "3"
},
{
"text": "Several algorithms have been proposed for the syllabification task divided between rule-based and data-driven. While, rule-based methods are centered on theoretical aspects of the syllabification problem, data-driven methods are usually preferable, since they are language independent and they only require the construction of syllabified words lexicons.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification",
"sec_num": "3"
},
{
"text": "In the following description, we use the term juncture point to denote the places where hyphen marks (syllable breaks) are placed within a word.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification",
"sec_num": "3"
},
{
"text": "The look-up procedure was introduced by Weijters (1991) . It constructs a table of n-grams from the training corpus and uses this table to predict juncture points. Each n-gram contains the focus character (the character that is being analyzed to determine if a juncture point should or should not occur after) with left and right context, including hyphen marks. When syllabification is performed on a new word, the algorithm determines if a focus character should be followed by a hyphen, using the majority of similar n-grams.",
"cite_spans": [
{
"start": 40,
"end": 55,
"text": "Weijters (1991)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification",
"sec_num": "3"
},
{
"text": "The IB1 (Daelemans et al., 1997 ) algorithm creates n-grams (of predetermined size) from word juncture points and stores them into a database. When a new word has to be split into syllables, every n-gram around the word's possible junctures is matched against the n-grams already available from the training step. N-grams are compared using a distance measure to determine how similar two n-grams are to one another. Marchand and Damper (2007) introduced Syllabification by Analogy (SbA) which follows the principles of the Pronunciation by Analogy (PbA) algorithm. It works by applying a \"full pattern match\" on the input string using entries in a dictionary compiled from the training corpora. Marchand and Damper also investigate the possibility of using syllabification to improve grapheme to phoneme performance on English words. Barlett et al. (2008) use structured SVMs to predict tags for letters in a given word and compare results obtained using different tagging strategies. Their method outperforms the results of the SbA method.",
"cite_spans": [
{
"start": 8,
"end": 31,
"text": "(Daelemans et al., 1997",
"ref_id": "BIBREF4"
},
{
"start": 417,
"end": 443,
"text": "Marchand and Damper (2007)",
"ref_id": null
},
{
"start": 835,
"end": 856,
"text": "Barlett et al. (2008)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification",
"sec_num": "3"
},
{
"text": "Our sequence labeling approach is inspired after Barlett et al. (2008) . In their paper they experimented with different tagging strategies and according to their results, the numbered ONC (onset-nucleus-coda) achieved the highest performance. This is why we employed the same tagging strategy for our system. The main difference between our approach and theirs, is the features set we designed and the classifier we used (MIRA).",
"cite_spans": [
{
"start": 49,
"end": 70,
"text": "Barlett et al. (2008)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "A widely accepted fact is that a syllable is composed of a nucleus vowel with or without surrounding consonants which are divided into the onset (the consonants preceding the vowel) and the coda (the consonants succeeding the vowel). The ONC tagging strategy assigns a tag to every letter of a word based on its role inside the parent syllable. There are three types of tags: O-onset, N-nucleus and C-coda. The numbered ONC makes every tag unique, inside a syllable, by adding an index to the tag. To exemplify, we will use the syllabification of the Romanian word \"avertisment\" (English \"warning\"). The correct tag sequence for this word is:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "N 1 O 1 N 1 C 1 O 1 N 1 C 1 O 1 N 1 C 1 C 2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "Determining where the junctures appear inside the word is easily attained by looking for tag sequences that are unacceptable inside the same syllable such as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "C i - O j , N i -N 1 , C i -N j , N i -O j etc.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "(for whatever indexes i and j). By doing so, we obtain the break sequence:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "N 1 -O 1 N 1 C 1 -O 1 N 1 C 1 -O 1 N 1 C 1 C 2 ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "and with a 1-1 correspondence between tags and letters, we get the sequence \"a-ver-tis-ment\", which is the correct syllabification of the word. After iterating through several feature sets we selected the one that yielded the highest results:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "(l -2 ,l -1 ,l), (l -3 ,l -2 ,l -1 ,l), (l -4 ,l -3 ,l -2 ,l -1 ,l), (l,l 1 ,l2), (l,l 1 ,l 2 ,l 3 ), (l,l 1 ,l 2 ,l 3 ,l 4 ), (l -1 ,l,l 1 ), (l -2 ,l -1 ,l,l 1 ,l 2 ),",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "where l is used to mark the current letter and l i is used to denote the letter at relative distance i from the current one.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syllabification with MIRA",
"sec_num": "3.1"
},
{
"text": "To test this approach we used a training corpus consisting of 600K syllabified words, compiled from the Romanian Academy Explanatory Dictionary. Using 10-fold validation we obtained and accuracy of 99.01% on OOV words. To our knowledge, the best performing system for Romanian syllabification is presented in Ungurean et al. (2011) . In their approach, they use Katz-Backoff for determining the most probable n-gram letter split sequence using the output of a stochastic search algorithm. Their method obtained a maximum accuracy of 97.04% using a window of 5 letter n-grams.",
"cite_spans": [
{
"start": 309,
"end": 331,
"text": "Ungurean et al. (2011)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments and results",
"sec_num": "3.2"
},
{
"text": "Lemmatization is the process of determining a word's canonical form from its inflectional form. It is a technique useful in various natural language processing applications such as datamining and document classification. Lemmatization is related to the technique called stemming, which is the process of extracting the longest common subsequence between word forms.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lemmatization",
"sec_num": "4"
},
{
"text": "In the case of English, the lemmatization process is fairly simple, but for highly inflectional languages, such as Romanian, this process poses a series of challenges. There are several approaches to this task, with a trend toward rule-based transformations applied to the sequence of characters. The best-performing Romanian lemmatizer 1 (to the best of our knowledge) is implemented after the 1 http://ws.racai.ro:9191 methodology proposed in Ion (2007) . The method builds a lookup table storing for each POS tag (named CTAG), the transformations required for word form to canonical form conversion. When the method has to predict the lemma for a previously unseen word with an associated CTAG (supplied by the POS tagging process), it searches the lookup table for the transformation rules of the CTAG and applies all of them to the unseen word, thus obtaining a set of candidate lemmas from which it probabilistically chooses the most likely one.",
"cite_spans": [
{
"start": 445,
"end": 455,
"text": "Ion (2007)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Lemmatization",
"sec_num": "4"
},
{
"text": "In order to use the MIRA framework, we had to reformulate lemmatization as a sequence labeling task. Our labels are designed to encode the following transformations:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lemmatization with MIRA",
"sec_num": "4.1"
},
{
"text": "-'*'means leave current letter unchanged -'_nil_'means that the current letter must be removed from the word's lemma -'_r(<character sequence>) -means that the current letter has to be replaced with the character sequence in brackets (<character sequence>). To exemplify, we will use the 2 nd person, plural verb \"\u00eembr\u0103ca\u021bi\" (English \"dressed\"), which has the canonical form \"\u00eembr\u0103ca\" (\"to dress\"). The letter tag sequence is shown in Lemmatization has to take into account the information provided by the word's morphosyntactic-description (MSD) tag (Ion, 2007) . This means that we either have to train different models for different MSDs or we have to incorporate the MSD information inside the features we use. The Romanian MSDs inventory is very large (more than 600 MSDs) and consequently, the MIRA model obtained by training with MSDs is extremely large, difficult to train and use. Tufi\u015f (1999) presents a strategy for coping with the large Romanian MSD inventory, in which he eliminates lexiconrecoverable morpho-syntactic attributes from the MSDs. The resulting tagset is much smaller and the resulting POS tags are called CTAGs (from Corpus POS tags).",
"cite_spans": [
{
"start": 551,
"end": 562,
"text": "(Ion, 2007)",
"ref_id": "BIBREF7"
},
{
"start": 890,
"end": 902,
"text": "Tufi\u015f (1999)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Lemmatization with MIRA",
"sec_num": "4.1"
},
{
"text": "In order to reduce our lemmatization model size, we converted every word's MSD from our training set into a CTAG, based on the above mentioned methodology. This reduced our model size about 5 times.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lemmatization with MIRA",
"sec_num": "4.1"
},
{
"text": "The context used by the labeler is composed of both lexical and morpho-syntactic features (CTAGs): (l -2 ,l -1 ,l,C), (l -3 ,l -2 ,l -1 ,l,C), (l -4 ,l -3 ,l -2 ,l -1 ,l,C), (l,l 1 ,l 2 ,C), (l,l 1 ,l 2 ,l 3 ,C), (l,l 1 ,l 2 ,l 3 ,l 4 ,C), (l -1 ,l,l 1 ,C), (l -2 ,l -1 ,l,l 1 ,l 2 ,C) , where l is used to mark the current letter, l i is used to denote the letter at relative distance i from the current one and C is used to denote the word form's CTAG.",
"cite_spans": [],
"ref_spans": [
{
"start": 99,
"end": 285,
"text": "(l -2 ,l -1 ,l,C), (l -3 ,l -2 ,l -1 ,l,C), (l -4 ,l -3 ,l -2 ,l -1 ,l,C), (l,l 1 ,l 2 ,C), (l,l 1 ,l 2 ,l 3 ,C), (l,l 1 ,l 2 ,l 3 ,l 4 ,C), (l -1 ,l,l 1 ,C), (l -2 ,l -1 ,l,l 1 ,l 2 ,C)",
"ref_id": null
}
],
"eq_spans": [],
"section": "Lemmatization with MIRA",
"sec_num": "4.1"
},
{
"text": "Using a training corpus composed of 1M words we withheld 10% for each individual CTAG as the test set. The results of our experiments are shown in Table 1 . The overall accuracy of 94% which is 12% higher than the results presented in Ion (2007) .",
"cite_spans": [
{
"start": 235,
"end": 245,
"text": "Ion (2007)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [
{
"start": 147,
"end": 154,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "4.2"
},
{
"text": "In Table 1 , all CTAGS beginning with an \"N\" are nouns, \"A\" are adjectives and \"V\" are verbs. The best result (100%) is for invariant adjectives (\"A\") for which the lemma is the word form. This behavior is preserved for all CTAGs for which lemma is equal to the word form: NSRN (noun, singular, nominative/accusative, nondefinite form) with 99.5%, ASN (adjective, singular, non-definite form) with 98.95%, etc. At the opposite pole we find words with CTAGs that are harder to lemmatize: NPN (noun, plural, nondefinite form) with 81.51% or NPOY (noun, plural, dative/genitive, definite form) with 83.01% due to their root alternation when going from singular (the number of the lemma) to plural, e.g. for \"stadioanelor\" (NPOY, English \"to the stadiums\") lemma is \"stadion\" (English \"stadium\") where in bold we have the inflectional ending corresponding to the CTAG NPOY and in italic we have the root of the word. ",
"cite_spans": [],
"ref_spans": [
{
"start": 3,
"end": 10,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Experimental results",
"sec_num": "4.2"
},
{
"text": "Phonetic transcription (PT; also referred to as grapheme-to-phoneme (G2P) or letter-to-sound (L2S)) can be formalized as finding a relation between letters and corresponding phonemes, which is not a straightforward task and may pose some challenges for languages such as English. For Romanian, phonetic transcription rules are relatively simple compared to English or French (Burileanu, 1999) , but there are several exceptions that need to be managed. For the purpose of language independence, data-driven methods are preferable as they only require words and their phonetic transcription equivalents for training, which are easier to obtain than wide coverage set of phonetic transcription rules. Several Machine Learning (ML) methods have been proposed for the PT task: Black et al. (1998 ), Jiampojamarn et al. (2008 ), Pagel et al. (1998 , Bisani and Ney (2002) , Marchand and Damper (2000) and Demberg (2007) . Jiampojamarn et al. (2008) presented a MIRA based method for L2S conversion of words. Their best result on the English CMU lexicon was 71%. However, the feature template provided in their paper did not turn out to be suitable in our tests. Instead we came up with a different one, which turned out to be the most discriminative for Romanian L2S: (l -2 ,l -1 ,l), (l -3 ,l -2 ,l -1 ,l), (l -4 ,l -3 ,l -2 ,l -1 ,l), (l,l 1 ,l 2 ), (l,l 1 ,l 2 ,l 3 ), (l,l 1 ,l 2 ,l 3 ,l 4 ), (l -1 ,l,l 1 ), (l -2 ,l -1 ,l,l 1 ,l 2 ), (l -2 ,l -1 ,l,l 1 ), (l -1 ,l,l 1 ,l 2 ) , where l is used to mark the current letter, l i is used to denote the letter at relative distance i from the current one.",
"cite_spans": [
{
"start": 375,
"end": 392,
"text": "(Burileanu, 1999)",
"ref_id": "BIBREF2"
},
{
"start": 773,
"end": 791,
"text": "Black et al. (1998",
"ref_id": "BIBREF1"
},
{
"start": 792,
"end": 820,
"text": "), Jiampojamarn et al. (2008",
"ref_id": "BIBREF8"
},
{
"start": 821,
"end": 842,
"text": "), Pagel et al. (1998",
"ref_id": "BIBREF1"
},
{
"start": 845,
"end": 866,
"text": "Bisani and Ney (2002)",
"ref_id": null
},
{
"start": 869,
"end": 895,
"text": "Marchand and Damper (2000)",
"ref_id": null
},
{
"start": 900,
"end": 914,
"text": "Demberg (2007)",
"ref_id": "BIBREF5"
},
{
"start": 917,
"end": 943,
"text": "Jiampojamarn et al. (2008)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 1263,
"end": 1476,
"text": "(l -2 ,l -1 ,l), (l -3 ,l -2 ,l -1 ,l), (l -4 ,l -3 ,l -2 ,l -1 ,l), (l,l 1 ,l 2 ), (l,l 1 ,l 2 ,l 3 ), (l,l 1 ,l 2 ,l 3 ,l 4 ), (l -1 ,l,l 1 ), (l -2 ,l -1 ,l,l 1 ,l 2 ), (l -2 ,l -1 ,l,l 1 ), (l -1 ,l,l 1 ,l 2 )",
"ref_id": null
}
],
"eq_spans": [],
"section": "Phonetic transcription",
"sec_num": "5"
},
{
"text": "All the data-driven methods for phonetic transcription require alignments between letters and phonemes. For so-called phonetic (or pseudo-phonetic) languages (e.g. Romanian), the task of grapheme to phoneme conversion is significantly easier and more accurate than for many other languages (such as English). However, there are several issues, common to several languages. The simplest example is that not all words have the same number of phonemes and letters and even if this condition is satisfied, it still does not imply a one-to-one alignment (e.g. experience -IH K S P IH R IY AH N S, where the letter x spawns two phonemes \"K\" + \"S\" and the ending \"e\" is silent; a similar phenomenon happens when we phonetically transcribe the word Romanian \"experien\u0163\u0103\" (experience) into e k s p e r i e n ts @, where again x spawns \"k\"+\"s\"). Expectation-Maximization (EM) can be used to find one-toone or many-to-many alignments between letters and phonemes (Black et al., 1998; Jiampojamarn et al., 2008; Pagel et al. 1998) . Although it is arguable that in the case of Romanian such alignments can be easily attained using simple heuristics, we preferred to use EM on our training data, to keep our system portable to other languages.",
"cite_spans": [
{
"start": 952,
"end": 972,
"text": "(Black et al., 1998;",
"ref_id": "BIBREF1"
},
{
"start": 973,
"end": 999,
"text": "Jiampojamarn et al., 2008;",
"ref_id": "BIBREF8"
},
{
"start": 1000,
"end": 1018,
"text": "Pagel et al. 1998)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Phonetic transcription",
"sec_num": "5"
},
{
"text": "Our training data was extracted from the Romanian Speech Synthesis Corpus (RSS) (Stan et al., 2011) and it is comprised of a small number of words (8K). However, due to the preponderantly phonetical nature of Romanian, this number seems to be sufficient for training a highly accurate L2S data-driven method. Using 10-fold validation we obtained an accuracy of 96.29% on OOV words, which is comparable to the state-of-the art results (96.99%) of a rulebased system reported in Ungurean et al. (2011) .",
"cite_spans": [
{
"start": 80,
"end": 99,
"text": "(Stan et al., 2011)",
"ref_id": "BIBREF12"
},
{
"start": 477,
"end": 499,
"text": "Ungurean et al. (2011)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments and results",
"sec_num": "5.1"
},
{
"text": "In natural speech certain syllables inside a word have a higher prominence compared to the neighboring syllables of the same word. When this phenomenon occurs, it is said that the syllable is carrying lexical stress. Lexical stress prediction is critical in prosody generation for TTS systems as it governs the correct pronunciation of diverse words and it is used to discriminate between homographs. Oancea and B\u0103dulescu (2003) introduced their rule-based method for lexical stress prediction on Romanian. They trained and tested their method on the same lexicon (4500 words) achieving a 94% accuracy. Ungurean et al. (2009) used Katz back-off smoothing, for lexical stress assignment based on letter n-grams. Their algorithm works by calculating the probability of every possible combination of stress pattern on an input string. According to their evaluation, this method achieves an accuracy of over 99% for OOV words.",
"cite_spans": [
{
"start": 401,
"end": 428,
"text": "Oancea and B\u0103dulescu (2003)",
"ref_id": null
},
{
"start": 603,
"end": 625,
"text": "Ungurean et al. (2009)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Lexical stress prediction",
"sec_num": "6"
},
{
"text": "Our tagging strategy is inspired after the numbered ONC style encoding used for syllabification. In this case we designed a numbered tagging strategy, in which the \"BPS\" tag used to label letters which appear before the primary lexical stress; \"APS\" was used on letters that appear after the primary lexical stress and \"PS\" to label the letter which carries the primary lexical stress. To exemplify, we will show the labels for the word \"\u00eemr\u0103ca\u0163i\" (bolded and underlined a, receives the primary lexical stress). This type of encoding is available for Romanian, which only uses primary lexical stress. For other languages, which support multiple degrees of lexical stress, the encoding requires adaptations. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lexical stress prediction with MIRA",
"sec_num": "6.2"
},
{
"text": "Franz\u00e9n and Horne (1997) conducted a study on stress patterns in Romanian. They showed that stress is rather influenced by derivational affixes than by inflectional ones, especially for nouns and verbs. Since the vast majority of derivational affixes change the grammatical category of a word, we were motivated to split our training data into 5 categories: nouns (N), verbs (V), adjectives (A), adverbs (R) and mixed (M). This is where the main difference between our approach and other methods can be seen: splitting the training data based on the part-ofspeech increases the overall accuracy by 3.9% (see Table 3 ). When predicting the primary lexical stress position for a given word, a model is chosen based on the POS tag of the given word. If the POS is different from the first four categories or if it is unknown (if there is no context available), the system uses the mixed model, which is a model created by training on the entire lexicon regardless of the POS.",
"cite_spans": [],
"ref_spans": [
{
"start": 608,
"end": 615,
"text": "Table 3",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Experiments and results",
"sec_num": "6.3"
},
{
"text": "V",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "POS # tokens # errors Accuracy",
"sec_num": null
},
{
"text": "The lexical feature templates we used for lexical stress prediction are identical to the ones we used for lemmatization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "POS # tokens # errors Accuracy",
"sec_num": null
},
{
"text": "In this paper we addressed the task of lexical processing for OOV words, which are one of the main sources of errors in both speech synthesis and natural language processing applications. We presented a unified data-driven framework that is designed to accurately handle the lemmatization, syllabification, phonetic transcription and lexical stress prediction of OOV words. Although, our main focus was on Romanian, the advantage of using data-driven methods is that with proper training lexicons and, in some cases, with minor adjustments, they can be applied to any other language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "Our results are better than state-of-the-art results cited for Romanian in the case of syllabification (99% vs. 97%) and lemmatization (94% vs. 82%), and only slightly worse for phonetic transcription (96.3% vs. 97%) and lexical stress prediction (98.8% vs. 99%), which can be explained by the fact that we did not incorporate any explicit knowledge of Romanian into our algorithms. In this context, we should emphasize that we successfully employed the MIRA framework described in this paper (without any modifications) to do phonetic transcription for English, French, German and Dutch and lemmatization for Serbian with very good results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "The methods we presented are already implemented in a natural language preprocessing tool written entirely in JAVA for portability and available as an open-source package.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
}
],
"back_matter": [
{
"text": "The work reported here was funded by the project METANET4U by the European Commission under the Grant Agreement No 270893",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Automatic syllabification with structured SVMs for letter-to-phoneme conversion",
"authors": [
{
"first": "S",
"middle": [],
"last": "Bartlett",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Kondrak",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Cherry",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "568--576",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bartlett, S., Kondrak, G., & Cherry, C. 2008. Automatic syllabification with structured SVMs for letter-to-phoneme conversion. Proceedings of ACL- 08: HLT, 568-576.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Issues in building general letter to sound rules",
"authors": [
{
"first": "A",
"middle": [
"W"
],
"last": "Black",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Lenzo",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Pagel",
"suffix": ""
}
],
"year": 1998,
"venue": "The Third ESCA/COCOSDA Workshop (ETRW) on Speech Synthesis",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Black, A. W., Lenzo, K., & Pagel, V. 1998. Issues in building general letter to sound rules. In The Third ESCA/COCOSDA Workshop (ETRW) on Speech Synthesis.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "A phonetic converter for speech synthesis in Romanian",
"authors": [
{
"first": "D",
"middle": [],
"last": "Burileanu",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Sima",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Neagu",
"suffix": ""
}
],
"year": 1999,
"venue": "Proc. of the XIVth International Congress on Phonetic Sciences ICPhS'99",
"volume": "",
"issue": "",
"pages": "503--506",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Burileanu, D., Sima, M., & Neagu, A. 1999. A phonetic converter for speech synthesis in Romanian. In Proc. of the XIVth International Congress on Phonetic Sciences ICPhS'99 (pp. 503- 506).",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Ultraconservative online algorithms for multiclass problems",
"authors": [
{
"first": "K",
"middle": [],
"last": "Crammer",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Singer",
"suffix": ""
}
],
"year": 2003,
"venue": "The Journal of Machine Learning Research",
"volume": "3",
"issue": "",
"pages": "951--991",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Crammer, K. and Singer, Y. 2003. Ultraconservative online algorithms for multiclass problems. The Journal of Machine Learning Research, 3:951-991.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "IGTree: Using trees for compression and classification in lazy learning algorithms",
"authors": [
{
"first": "W",
"middle": [],
"last": "Daelemans",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Van Den",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bosch",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Weijters",
"suffix": ""
}
],
"year": 1997,
"venue": "Artificial Intelligence Review",
"volume": "11",
"issue": "1",
"pages": "407--423",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daelemans, W., Van Den Bosch, A., & Weijters, T. 1997. IGTree: Using trees for compression and classification in lazy learning algorithms. Artificial Intelligence Review, 11(1), 407-423.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Phonological constraints and morphological preprocessing for grapheme-to-phoneme conversion",
"authors": [
{
"first": "V",
"middle": [],
"last": "Demberg",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Schmid",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Mohler",
"suffix": ""
}
],
"year": 2007,
"venue": "Annual Meeting-Association for Computational Linguistics",
"volume": "45",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Demberg, V., Schmid, H., & Mohler, G. 2007. Phonological constraints and morphological preprocessing for grapheme-to-phoneme conversion. In Annual Meeting-Association for Computational Linguistics (Vol. 45, No. 1, p. 96).",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Word stress in Romanian. Lund Working Papers in Linguistics",
"authors": [
{
"first": "V",
"middle": [],
"last": "Franz\u00e9n",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Horne",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "46",
"issue": "",
"pages": "75--91",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz\u00e9n, V., & Horne, M. 2009. Word stress in Romanian. Lund Working Papers in Linguistics, 46, 75-91.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Word Sense Disambiguation Methods Applied to English and Romanian",
"authors": [
{
"first": "R",
"middle": [],
"last": "Ion",
"suffix": ""
}
],
"year": 2007,
"venue": "Romanian Academy",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ion, R. 2007. Word Sense Disambiguation Methods Applied to English and Romanian. PhD thesis (in Romanian). Romanian Academy, Bucharest.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Joint processing and discriminative training for letter-to-phoneme conversion",
"authors": [
{
"first": "S",
"middle": [],
"last": "Jiampojamarn",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Cherry",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Kondrak",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "905--913",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jiampojamarn, S., Cherry, C., & Kondrak, G. 2008. Joint processing and discriminative training for letter-to-phoneme conversion. Proceedings of ACL- 08: HLT, 905-913.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Syllable-based generalizations in English phonology",
"authors": [
{
"first": "D",
"middle": [],
"last": "Kahn",
"suffix": ""
}
],
"year": 1976,
"venue": "",
"volume": "156",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kahn, D. 1976. Syllable-based generalizations in English phonology (Vol. 156). Bloomington: Indiana University Linguistics Club.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data",
"authors": [
{
"first": "J",
"middle": [],
"last": "Lafferty",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Mccallum",
"suffix": ""
},
{
"first": "F",
"middle": [
"C"
],
"last": "Pereira",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lafferty, J., McCallum, A., & Pereira, F. C. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Stressed syllable determination for Romanian words within speech synthesis applications",
"authors": [
{
"first": "E",
"middle": [],
"last": "Oancea",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Badulescu",
"suffix": ""
}
],
"year": 2002,
"venue": "International Journal of Speech Technology",
"volume": "5",
"issue": "3",
"pages": "237--246",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Oancea, E., & Badulescu, A. 2002. Stressed syllable determination for Romanian words within speech synthesis applications. International Journal of Speech Technology, 5(3), 237-246.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "The Romanian Speech Synthesis (RSS) corpus: building a high quality HMM-based speech synthesis system using a high sampling rate",
"authors": [
{
"first": "A",
"middle": [],
"last": "Stan",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Yamagishi",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "King",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Aylett",
"suffix": ""
}
],
"year": 2011,
"venue": "Speech Communication",
"volume": "53",
"issue": "3",
"pages": "442--450",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Stan, A., Yamagishi, J., King, S., & Aylett, M. 2011. The Romanian Speech Synthesis (RSS) corpus: building a high quality HMM-based speech synthesis system using a high sampling rate. Speech Communication, 53(3), 442-450.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Tiered tagging and combined language models classifiers. Text, Speech and Dialogue",
"authors": [
{
"first": "D",
"middle": [],
"last": "Tufi\u015f",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "843--843",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tufi\u015f, D. 1999. Tiered tagging and combined language models classifiers. Text, Speech and Dialogue (pp. 843-843). Springer Berlin/Heidelberg.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Hybrid Syllabification and Letter-To-Phone Conversion For TTS Synthesis",
"authors": [
{
"first": "C",
"middle": [],
"last": "Ungurean",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Burileanu",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Popescu",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Dervi\u015f",
"suffix": ""
}
],
"year": 2011,
"venue": "U.P.B. Sci. Bull., Series C",
"volume": "73",
"issue": "3",
"pages": "1454--234",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ungurean, C., Burileanu, D., Popescu, V. and Dervi\u015f, A. 2011. Hybrid Syllabification and Letter-To- Phone Conversion For TTS Synthesis. In U.P.B. Sci. Bull., Series C, Vol. 73, Iss. 3, 2011, ISSN 1454-234x",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "A simple look-up procedure superior to NETtalk",
"authors": [
{
"first": "A",
"middle": [],
"last": "Weijters",
"suffix": ""
}
],
"year": 1991,
"venue": "Proceedings of the International Conference on Artificial Neural Networks -ICANN-91",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Weijters, A. 1991. A simple look-up procedure superior to NETtalk?. In Proceedings of the International Conference on Artificial Neural Networks -ICANN-91, Espoo, Finland",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"content": "<table><tr><td>.</td><td/><td/><td/><td/></tr><tr><td>\u00ee m b r \u0103</td><td>c</td><td>a</td><td>\u021b</td><td>i</td></tr><tr><td>* * * * *</td><td>*</td><td>*</td><td colspan=\"2\">_nil_ _nil_</td></tr><tr><td colspan=\"4\">Table 1 -Lemmatization example for word</td><td/></tr><tr><td/><td>\"\u00eembr\u0103ca\u021bi\"</td><td/><td/><td/></tr></table>",
"text": "",
"num": null,
"html": null,
"type_str": "table"
},
"TABREF3": {
"content": "<table/>",
"text": "Lexical stress tagging for the word \"\u00eembr\u0103ca\u0163i\"",
"num": null,
"html": null,
"type_str": "table"
},
"TABREF5": {
"content": "<table/>",
"text": "Lexical stress accuracy",
"num": null,
"html": null,
"type_str": "table"
}
}
}
} |