File size: 56,151 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 |
{
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T12:13:57.773056Z"
},
"title": "Applying Graph Neural Networks for Vietnamese Dependency Parsing",
"authors": [
{
"first": "Nguyen",
"middle": [],
"last": "Duc",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Thien",
"middle": [],
"last": "Nguyen",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Thi",
"middle": [
"Thu"
],
"last": "Trang",
"suffix": "",
"affiliation": {},
"email": "trangntt@soict.hust.edu.vn"
},
{
"first": "Truong",
"middle": [],
"last": "Dang Quan\u01f5",
"suffix": "",
"affiliation": {},
"email": "dangquangtruong98@gmail.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper presents a state-of-the-art model to solve the Vietnamese dependency parsing task (HA My Linh, 2020) in VLSP 2020 1 Evaluation Campaign. In this model, the Bidirectional Long Short-Term Memory (BiLSTM) network is used to extract the contextual information, while the graph neural network captures high-order information. Some preprocessing for Vietnamese raw texts are included for the training, such as word segmentation, part-of-speech (POS) tagging for the model. We modified the network with suitable word embedding mechanisms, i.e., fastText, to represent the semantic information of words more accurately. Therefore, Vietnamese words that are marked as unknown tokens now can have the right embedding; thus, they will be well modeled in dependency parsing. Experiments on the raw text dataset show that the model achieved an average of 72.85% of unlabeled attachment score (UAS) and 64.35% of labeled attachment score (LAS). With the Segmentation and POS tagging dataset, we achieved a higher average of 81.71% (UAS) and 73.19% (LAS).",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper presents a state-of-the-art model to solve the Vietnamese dependency parsing task (HA My Linh, 2020) in VLSP 2020 1 Evaluation Campaign. In this model, the Bidirectional Long Short-Term Memory (BiLSTM) network is used to extract the contextual information, while the graph neural network captures high-order information. Some preprocessing for Vietnamese raw texts are included for the training, such as word segmentation, part-of-speech (POS) tagging for the model. We modified the network with suitable word embedding mechanisms, i.e., fastText, to represent the semantic information of words more accurately. Therefore, Vietnamese words that are marked as unknown tokens now can have the right embedding; thus, they will be well modeled in dependency parsing. Experiments on the raw text dataset show that the model achieved an average of 72.85% of unlabeled attachment score (UAS) and 64.35% of labeled attachment score (LAS). With the Segmentation and POS tagging dataset, we achieved a higher average of 81.71% (UAS) and 73.19% (LAS).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "In recent years, dependency parsing is a fascinating research topic and has a large number of applications in natural language processing. This task is to automatically identify the relationship between words in a sentence and label the relationship between the head and the dependency word, and thus, establish the grammatical structure of the sentence. Traditional graph-based dependency parsing only extracts the parent-child relationship and ignores deeper relationships. Hence, we decided to experiment with the idea of extracting deeper relation-ships of the neighbor nodes, which is extensively covered in the paper Ji et al. (2019) .",
"cite_spans": [
{
"start": 623,
"end": 639,
"text": "Ji et al. (2019)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This state-of-the-art model achieved good performance due to its ability to represent incorrect Out-Of-Vocabulary (OOV) words in the input layer for Vietnamese. Normally, words that are not found in the vocabulary will be marked as unknown tokens before feeding to the embedding layer. This caused the model to embed OOV words incorrectly; therefore, it created the loss of information in calculating attention distribution. In this paper, we modified the pre-trained layer of word embedding for the graph neural networks with a more suitable embedding mechanism for Vietnamese, which solved the issue well.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The rest of this paper is organized as follows: Section 2 presents the architecture and its components of graph neural networks. The experiments are shown in Section 3. Finally, Section 4 concludes the paper and gives some perspectives for the work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Normally, Graph-based dependency parser search through the space of possible trees for a given sentence encoded as directed graphs and use methods from graph theory (Maximum Spanning Tree or greedy algorithm) for the optimal solutions. However, in the Graph Neural Network (GNN) model, the dependency parser utilizes the neural network to assign a weight to each edge, then construct a MST from the edge weight (Dozat et al., 2017) . For maximum accuracy, we need to analyze the surface form and the deep structure of the graph. There are three main components in the model: Encoder extracts the surface form and the contextual information and turns them into the nodes (words) representations for the next components; The graph attention network (a subset of GNN, using the structure from Veli\u010dkovi\u0107 et al. (2017) ) layers then extract the deep structure and high-order information to illustrate the head-dependent relationships of the nodes; the final component is the decoder, used to create the dependency tree from the output of the GNN. We will discuss the details in the following sections.",
"cite_spans": [
{
"start": 411,
"end": 431,
"text": "(Dozat et al., 2017)",
"ref_id": "BIBREF2"
},
{
"start": 790,
"end": 814,
"text": "Veli\u010dkovi\u0107 et al. (2017)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "2"
},
{
"text": "First, we used the VNCoreNLP -suggested by Vu et al. (2018) -to segment and perform the POS tagging on the raw text. VNCoreNLP used a transformation rule-based learning model for the segmentation of the Vietnamese document, thus, obtained faster and better accuracy than all previous segmentation tools, as the model accounted for the fact that Vietnamese words are created from syllables including the space character (Nguyen et al., 2017) . The VNCoreNLP performed the task of labeling words with POS tag Vu et al. (2018) via MarMot (a CRF framework), state of the art POS and morphological tagger (M\u00fcller et al., 2013) Word embedding is the most popular representation method for words in a document because it captures the context of words, semantic and syntactic similarity, relation with other words, etc. Using word embedding makes it easier to represent words with less memory than using a one-hot vector while also showing the relationship between words.",
"cite_spans": [
{
"start": 43,
"end": 59,
"text": "Vu et al. (2018)",
"ref_id": "BIBREF11"
},
{
"start": 419,
"end": 440,
"text": "(Nguyen et al., 2017)",
"ref_id": "BIBREF9"
},
{
"start": 507,
"end": 523,
"text": "Vu et al. (2018)",
"ref_id": "BIBREF11"
},
{
"start": 600,
"end": 621,
"text": "(M\u00fcller et al., 2013)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Pre-processing",
"sec_num": "2.1"
},
{
"text": "With a huge training corpus (e.g., a total of 100 billion words with a 3-million-word vocab in Google News), the pre-trained model can cover much more context for word embedding than the auto-updating mechanism of the word embedding in the end-to-end abstractive summarization model with its training corpus (e.g., a total of 240 million words with a 50k-word vocab in Daily Mail/CNN) (Anh and Trang, 2019).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pre-processing",
"sec_num": "2.1"
},
{
"text": "In this paper, we adopted a suitable pre-trained model for Vietnamese with 300-dimensional word embeddings, i.e., fastText from Facebook , for the word embedding layer. The fastText trained on the Wikipedia dataset with character n-grams of length 5 by CBOW 2 method. fast-Text is more suitable in our case as when the GNN model meets unknown vocab, the fastText generates an embedding of the vocab with value 0, resulting in error reductions; meanwhile, the Word2Vec and the GloVe does not do that. This method enables fastText to handle OOV 3 words by constructing the vector for OOV words from its characters. Both GloVe and Word2Vec are unable to do so.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pre-processing",
"sec_num": "2.1"
},
{
"text": "According to Kiperwasser and Goldberg (2016) , we can apply BiLSTM model to create the dependency tree as illustrated in Figure [ 1] Firstly, each word is embedded using a vector combined from three different vectors: randomly initialized word embedding, pre-trained word embedding, and partof-speech embedding.",
"cite_spans": [
{
"start": 13,
"end": 44,
"text": "Kiperwasser and Goldberg (2016)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [
{
"start": 121,
"end": 129,
"text": "Figure [",
"ref_id": null
}
],
"eq_spans": [],
"section": "Encoder",
"sec_num": "2.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "x i = e(w i ) \u2295 e (w i ) \u2295 e(pos i )",
"eq_num": "(1)"
}
],
"section": "Encoder",
"sec_num": "2.2"
},
{
"text": "As a result, the x i illustrated the sentence of the word i in [2] . Given the position i of the word, the BiLSTM model can compute state vectors \u2212 \u2192 c i and \u2190 \u2212 c i where the \u2212 \u2192 c i is draw from the start of the sentence to the position i and \u2190 \u2212 c i is from the end of the sentence to i.",
"cite_spans": [
{
"start": 63,
"end": 66,
"text": "[2]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Encoder",
"sec_num": "2.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u2212 \u2192 c i = \u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2192 LST M (x i ) \u2295 \u2190\u2212\u2212\u2212\u2212\u2212\u2212\u2212 LST M (x i )",
"eq_num": "(2)"
}
],
"section": "Encoder",
"sec_num": "2.2"
},
{
"text": "The two vectors \u2212 \u2192 c i and \u2190 \u2212 c i then concatenate to become the context-dependent representation of the word i. Thus we can use multilayers perceptron (MLP) to define two-node representations of the word i the probability of being the head role vector and probability of being the dependent role vector (Dozat et al., 2017) :",
"cite_spans": [
{
"start": 306,
"end": 326,
"text": "(Dozat et al., 2017)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Encoder",
"sec_num": "2.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "h i = M LP h (c i ), d i = M LP d (c i )",
"eq_num": "(3)"
}
],
"section": "Encoder",
"sec_num": "2.2"
},
{
"text": "The score function is a SoftMax function, where the representations of the word i and j is the input, therefore complementing the analysis of the surface form of the segmented sentence. As a result, the output of the BiLSTM component is a complete weight graph model. (Dozat et al., 2017) ",
"cite_spans": [
{
"start": 268,
"end": 288,
"text": "(Dozat et al., 2017)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Encoder",
"sec_num": "2.2"
},
{
"text": "\u03c3(i, j) = Softmax i (h T j Ad j + b T 1 h j + b T 2 h j ) (4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Encoder",
"sec_num": "2.2"
},
{
"text": "In the implementation, the GNN component can utilize at most three layers, each layer consists of 4 graph neural network units as illustrated in Figure [1] -where the representation of the vectors is calculated from the same representation in the previous layer using this formula where g is the LeakyReLU function, t is the layer, v i is the vector representation of i, and a ij is the edge weight of ",
"cite_spans": [],
"ref_spans": [
{
"start": 145,
"end": 155,
"text": "Figure [1]",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "v t i = g \uf8eb \uf8ed W j\u2208N (i) \u03b1 t ij v t\u22121 j + Bv t\u22121 i \uf8f6 \uf8f8",
"eq_num": "(5)"
}
],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "We can apply the formula [5] to analyze the high order information of the nodes which is represented in three ways: grandparents, grandchildren, and siblings ( Figure [2] ) (Eisner, 1997) .",
"cite_spans": [
{
"start": 173,
"end": 187,
"text": "(Eisner, 1997)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [
{
"start": 160,
"end": 170,
"text": "Figure [2]",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "Specifically, the head representation of node i should attend to the neighbors' representation as they are the parents of the i. Therefore the model can calculate h i from the h j of the previous layer t \u2212 1 using the formula [5]:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\uf8f1 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f2 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f3 h t i = g W 1 j\u2208N (i) \u03b1 t ji h t\u22121 j + B 1 h t\u22121 i d t i = g W 2 j\u2208N (i) \u03b1 t ij d t\u22121 j + B 2 d t\u22121 i",
"eq_num": "("
}
],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "6) The dependent node d i 's computation operation is the same as the head node's one i. Thus the equation [6] can assist to analyse the order of the relationship of grandparents and grandchild.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "To examine the sibling relationships, the head representation of the node i check the neighborhood where they are dependent on node i. Thus the formula will update the h i in the following way: and siblings.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\uf8f1 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f2 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f3 h t i = g W 1 j\u2208N (i) \u03b1 t ji d t\u22121 j + B 1 h t\u22121 i d t i = g W 2 j\u2208N (i) \u03b1 t ij h t\u22121 j + B 2 d t\u22121 i",
"eq_num": "(7)"
}
],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\uf8f1 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f2 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f4 \uf8f3 h t i = g(W 1 j\u2208N (i) (\u03b1 t ji h t\u22121 j + \u03b1 t ji d t\u22121 j ) +B 1 h t\u22121 i ) d t i = g(W 2 j\u2208N (i) (\u03b1 t ij h t\u22121 i + \u03b1 t ji d t\u22121 j ) +B 2 d t\u22121 i )",
"eq_num": "(8)"
}
],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "As the equations [8] illustrated, the edge weight a ij is the decisive element responsible for the update of the relationship information. The edge weight is figured with the following formula:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03b1 t ij = \uf8f1 \uf8f2 \uf8f3 Softmax i (h T i Ad j + b T 1 h i + b T 2 d j ) i \u2208 N t k (j) 0, otherwise",
"eq_num": "(9)"
}
],
"section": "GNN Layers",
"sec_num": "2.3"
},
{
"text": "After the high-order information is extracted from the GNN and enhanced the nodes representations, the node representation will be used to build the dependency tree via Biaffine parser (the setting is identical to Dozat et al. (2017)) 3 Experiments",
"cite_spans": [
{
"start": 214,
"end": 234,
"text": "Dozat et al. (2017))",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Decoder",
"sec_num": "2.4"
},
{
"text": "The VLSP provided the datasets and separated them into training datasets and raw text datasets. The data for training was further divided into two packages: the first package consists of 5070 sentences, with a large domain from the social media comments on restaurants and hotels (100 sentences), to the story of the Little Prince (1570 sentences) and the VietTreeBank -VTB (3400 sentences); the second package includes 3000 sentences with diverse origins.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "3.1"
},
{
"text": "The raw text data for prediction includes the two packages above, and 20 raw text files crawled from VnExpress news articles. The VTB files and the files with index 1,3,4,7,8,10,14 were accurately tokenized and labeled.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "3.1"
},
{
"text": "The graph-based dependency parsing neural network model has one important characteristic: the raw text dataset's sentences have to be tokenized for the training to be carried out successfully. Therefore the VNCoreNLP -an NLP pipeline used for POS tagging, named entity recognition and dependency parsing is useful here in this case [4] . This tool is capable of providing highly accurate annotation for the input sentences, therefore improving the score of the training model.",
"cite_spans": [
{
"start": 332,
"end": 335,
"text": "[4]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "3.1"
},
{
"text": "The training operation consists of two methods: First, we have to decode the output of the final layer of the GNN component (denoted by)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03b1 t ij = \u03c3 t (i, j) = P t (i|j)",
"eq_num": "(10)"
}
],
"section": "Training",
"sec_num": "3.2"
},
{
"text": "which are the tree structures (computed by P (i|j)) and the dependency edge labels (measured by P (r|i, j), which indicated the probability a tree (i, j) holds a dependency relation r, using another MLP from biaffine parser (Dozat et al., 2017) , the loss function of the classifier is computed with the equation:",
"cite_spans": [
{
"start": 224,
"end": 244,
"text": "(Dozat et al., 2017)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L 0 = \u2212 1 n (i,j,r)\u2208T (logP \u03c4 (i|j) + logP (r|i, j))",
"eq_num": "(11)"
}
],
"section": "Training",
"sec_num": "3.2"
},
{
"text": "Second, the model can supervise on P t (i|j) from each layer of the GNN component, therefore the layer-wise loss will be computed with the equation:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L = \u03c4 t=1 L t = \u03c4 t=1 \u2212 1 n (i,j,r)\u2208T logP (r|i, j)",
"eq_num": "(12)"
}
],
"section": "Training",
"sec_num": "3.2"
},
{
"text": "The main objective is to minimize the loss of combination of them:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L = \u03bb 1 L 0 + \u03bb 2 L",
"eq_num": "(13)"
}
],
"section": "Training",
"sec_num": "3.2"
},
{
"text": "We have implemented and operated the model on the AWS Server (AWS Deep Learning AMI (Ubuntu 18.04) Version 34.0 installed in the EC2 Instance p3.2xlarge -GPU NVIDIA Tesla v100 16 GB, Memory 61 GB, SSD 100 GB, CPU 8 Virtual Cores) successfully. The hyperparameters configuration in Table [ 1] has slight modifications. For the word embedding, we used fastText with Vietnamese data as the primary pre-trained model, which has 300 dimensions instead of 100 dimensions of GloVe that Ji et al. (2019) used. Then, we concatenate the pre-trained word embedding with 200-dimension randomly initialize word embedding and 100-dimension part-ofspeech embedding. Randomly embedding vectors obtained from binomial distribution. The training operation took approximately one hour.",
"cite_spans": [
{
"start": 479,
"end": 495,
"text": "Ji et al. (2019)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [
{
"start": 281,
"end": 288,
"text": "Table [",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "3.3"
},
{
"text": "The main evaluators for the dependency parsing problem are LAS and UAS. The results are coming from the script evaluator 2018. For the labeled data, the highest UAS is 81.89% from the VTB package, meanwhile the package Test VNExpress 14 achieved the highest LAS 73.57%. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "3.3"
},
{
"text": "To conclude, our experiment on using the graph neural network for graph-based dependency parsing suggests that understanding the deep structure of the representations of words via nodes' message passing improved a slightly better accuracy and efficiency than other traditional graph-based dependency parsers. In future works, we are planning to improve the performance of the model by applying Conditional Random Fields in the labeling process for the nodes before extracting the high-order information via graph neural network.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "4"
},
{
"text": "Continuous Bag of Words 3 Out-of-vocabulary",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "The authors wish to thank VLSP organizers for their reviews and encouragement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Abstractive text summarization using pointergenerator networks with pre-trained word embedding",
"authors": [],
"year": 2019,
"venue": "Proceedings of the Tenth International Symposium on Information and Communication Technology",
"volume": "",
"issue": "",
"pages": "473--478",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dang Trung Anh and Nguyen Thi Thu Trang. 2019. Abstractive text summarization using pointer- generator networks with pre-trained word embed- ding. In Proceedings of the Tenth International Sym- posium on Information and Communication Technol- ogy, pages 473-478.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Enriching word vectors with subword information",
"authors": [
{
"first": "Piotr",
"middle": [],
"last": "Bojanowski",
"suffix": ""
},
{
"first": "Edouard",
"middle": [],
"last": "Grave",
"suffix": ""
},
{
"first": "Armand",
"middle": [],
"last": "Joulin",
"suffix": ""
},
{
"first": "Tomas",
"middle": [],
"last": "Mikolov",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1607.04606"
]
},
"num": null,
"urls": [],
"raw_text": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2016. Enriching word vec- tors with subword information. arXiv preprint arXiv:1607.04606.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Stanford's graph-based neural dependency parser at the CoNLL 2017 shared task",
"authors": [
{
"first": "Timothy",
"middle": [],
"last": "Dozat",
"suffix": ""
},
{
"first": "Peng",
"middle": [],
"last": "Qi",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies",
"volume": "",
"issue": "",
"pages": "20--30",
"other_ids": {
"DOI": [
"10.18653/v1/K17-3002"
]
},
"num": null,
"urls": [],
"raw_text": "Timothy Dozat, Peng Qi, and Christopher D. Manning. 2017. Stanford's graph-based neural dependency parser at the CoNLL 2017 shared task. In Proceed- ings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, pages 20-30, Vancouver, Canada. Association for Computational Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Three new probabilistic models for dependency parsing: An exploration",
"authors": [
{
"first": "Jason",
"middle": [],
"last": "Eisner",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jason Eisner. 1997. Three new probabilistic models for dependency parsing: An exploration. arXiv preprint cmp-lg/9706003.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Vlsp 2020 shared task: Universal dependency parsing for vietnamese",
"authors": [
{
"first": "Luong Nguyen Thi Luong Phan Thi Hue Le",
"middle": [],
"last": "Vu Xuan",
"suffix": ""
},
{
"first": "Ha My",
"middle": [],
"last": "Van Cuong",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Linh",
"suffix": ""
},
{
"first": "Minh",
"middle": [],
"last": "Nguyen Thi",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Huyen",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of The seventh international workshop on Vietnamese Language and Speech Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "VU Xuan Luong NGUYEN Thi Luong PHAN Thi Hue LE Van Cuong HA My Linh, NGUYEN Thi Minh Huyen. 2020. Vlsp 2020 shared task: Uni- versal dependency parsing for vietnamese. In Pro- ceedings of The seventh international workshop on Vietnamese Language and Speech Processing (VLSP 2020), Hanoi, Vietnam.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Graph-based dependency parsing with graph neural networks",
"authors": [
{
"first": "Tao",
"middle": [],
"last": "Ji",
"suffix": ""
},
{
"first": "Yuanbin",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Man",
"middle": [],
"last": "Lan",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "2475--2485",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tao Ji, Yuanbin Wu, and Man Lan. 2019. Graph-based dependency parsing with graph neural networks. In Proceedings of the 57th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 2475- 2485.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Bag of tricks for efficient text classification",
"authors": [
{
"first": "Armand",
"middle": [],
"last": "Joulin",
"suffix": ""
},
{
"first": "Edouard",
"middle": [],
"last": "Grave",
"suffix": ""
},
{
"first": "Piotr",
"middle": [],
"last": "Bojanowski",
"suffix": ""
},
{
"first": "Tomas",
"middle": [],
"last": "Mikolov",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1607.01759"
]
},
"num": null,
"urls": [],
"raw_text": "Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2016. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Simple and accurate dependency parsing using bidirectional lstm feature representations",
"authors": [
{
"first": "Eliyahu",
"middle": [],
"last": "Kiperwasser",
"suffix": ""
},
{
"first": "Yoav",
"middle": [],
"last": "Goldberg",
"suffix": ""
}
],
"year": 2016,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "4",
"issue": "",
"pages": "313--327",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eliyahu Kiperwasser and Yoav Goldberg. 2016. Sim- ple and accurate dependency parsing using bidirec- tional lstm feature representations. Transactions of the Association for Computational Linguistics, 4:313-327.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Efficient higher-order crfs for morphological tagging",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "M\u00fcller",
"suffix": ""
},
{
"first": "Helmut",
"middle": [],
"last": "Schmid",
"suffix": ""
},
{
"first": "Hinrich",
"middle": [],
"last": "Sch\u00fctze",
"suffix": ""
}
],
"year": 2013,
"venue": "Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "322--332",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Thomas M\u00fcller, Helmut Schmid, and Hinrich Sch\u00fctze. 2013. Efficient higher-order crfs for morphological tagging. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 322-332.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A fast and accurate vietnamese word segmenter",
"authors": [
{
"first": "",
"middle": [],
"last": "Dat Quoc Nguyen",
"suffix": ""
},
{
"first": "Thanh",
"middle": [],
"last": "Dai Quoc Nguyen",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Vu",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Dras",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Johnson",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dat Quoc Nguyen, Dai Quoc Nguyen, Thanh Vu, Mark Dras, and Mark Johnson. 2017. A fast and accurate vietnamese word segmenter. CoRR, abs/1709.06307.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Graph attention networks",
"authors": [
{
"first": "Petar",
"middle": [],
"last": "Veli\u010dkovi\u0107",
"suffix": ""
},
{
"first": "Guillem",
"middle": [],
"last": "Cucurull",
"suffix": ""
},
{
"first": "Arantxa",
"middle": [],
"last": "Casanova",
"suffix": ""
},
{
"first": "Adriana",
"middle": [],
"last": "Romero",
"suffix": ""
},
{
"first": "Pietro",
"middle": [],
"last": "Lio",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1710.10903"
]
},
"num": null,
"urls": [],
"raw_text": "Petar Veli\u010dkovi\u0107, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "VnCoreNLP: A Vietnamese natural language processing toolkit",
"authors": [
{
"first": "Thanh",
"middle": [],
"last": "Vu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Dat Quoc Nguyen",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Dai Quoc Nguyen",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Dras",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Johnson",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations",
"volume": "",
"issue": "",
"pages": "56--60",
"other_ids": {
"DOI": [
"10.18653/v1/N18-5012"
]
},
"num": null,
"urls": [],
"raw_text": "Thanh Vu, Dat Quoc Nguyen, Dai Quoc Nguyen, Mark Dras, and Mark Johnson. 2018. VnCoreNLP: A Vietnamese natural language processing toolkit. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Demonstrations, pages 56-60, New Orleans, Louisiana. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Graph-based dependency parsing with bidirectional LSTM",
"authors": [
{
"first": "Wenhui",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Baobao",
"middle": [],
"last": "Chang",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "2306--2315",
"other_ids": {
"DOI": [
"10.18653/v1/P16-1218"
]
},
"num": null,
"urls": [],
"raw_text": "Wenhui Wang and Baobao Chang. 2016. Graph-based dependency parsing with bidirectional LSTM. In Proceedings of the 54th Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 2306-2315, Berlin, Germany. Association for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"text": "The architecture of Graph Neural Networks",
"uris": null
},
"FIGREF1": {
"num": null,
"type_str": "figure",
"text": "Relations between nodes v i and v j (i and j are forming the neighborhood)(Wang and Chang, 2016):",
"uris": null
},
"TABREF1": {
"type_str": "table",
"content": "<table/>",
"html": null,
"text": "Hyper-parameter.",
"num": null
},
"TABREF3": {
"type_str": "table",
"content": "<table><tr><td>Dataset</td><td>UAS LAS</td></tr><tr><td colspan=\"2\">Test from VTB 73.18 64.66</td></tr><tr><td>VNExpress 1</td><td>68.77 58.75</td></tr><tr><td>VNExpress 3</td><td>74.10 65.81</td></tr><tr><td>VNExpress 7</td><td>61.67 55.56</td></tr><tr><td>VNExpress 8</td><td>68.96 61.43</td></tr><tr><td colspan=\"2\">VNExpress 10 73.19 64.13</td></tr><tr><td>VNExpress 14</td><td>68.4 60.72</td></tr><tr><td>Total</td><td>72.85 64.35</td></tr></table>",
"html": null,
"text": "Test on labeled datasets.",
"num": null
},
"TABREF4": {
"type_str": "table",
"content": "<table/>",
"html": null,
"text": "Test on raw-text datasets.",
"num": null
},
"TABREF5": {
"type_str": "table",
"content": "<table/>",
"html": null,
"text": "] shows results from VLSP 2020 private tests for dependency parsing on labeled datasets, meanwhile raw-text datasets' results are shown onTable [3].",
"num": null
}
}
}
} |