File size: 54,739 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 |
{
"paper_id": "2022",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:13:25.584826Z"
},
"title": "The Document Vectors Using Cosine Similarity Revisited",
"authors": [
{
"first": "Zhang",
"middle": [],
"last": "Bingyu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Research University Higher School of Economics / Moscow",
"location": {
"country": "Russia"
}
},
"email": ""
},
{
"first": "Nikolay",
"middle": [],
"last": "Arefyev",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Research University Higher School of Economics / Moscow",
"location": {
"country": "Russia"
}
},
"email": "nick.arefyev@gmail.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The current state-of-the-art test accuracy (97.42%) on the IMDB movie reviews dataset was reported by Thongtan and Phienthrakul (2019) and achieved by the logistic regression classifier trained on the Document Vectors using Cosine Similarity (DV-ngrams-cosine) proposed in their paper and the Bag-of-Ngrams (BON) vectors scaled by Naive Bayesian weights. While large pre-trained Transformerbased models have shown SOTA results across many datasets and tasks, the aforementioned model has not been surpassed by them, despite being much simpler and pre-trained on the IMDB dataset only. In this paper, we describe an error in the evaluation procedure of this model, which was found when we were trying to analyze its excellent performance on the IMDB dataset. We further show that the previously reported test accuracy of 97.42% is invalid and should be corrected to 93.68%. We also analyze the model performance with different amounts of training data (subsets of the IMDB dataset) and compare it to the Transformer-based RoBERTa model. The results show that while RoBERTa has a clear advantage for larger training sets, the DV-ngramscosine performs better than RoBERTa when the labelled training set is very small (10 or 20 documents). Finally, we introduce a sub-sampling scheme based on Naive Bayesian weights for the training process of the DV-ngrams-cosine, which leads to faster training and better quality.",
"pdf_parse": {
"paper_id": "2022",
"_pdf_hash": "",
"abstract": [
{
"text": "The current state-of-the-art test accuracy (97.42%) on the IMDB movie reviews dataset was reported by Thongtan and Phienthrakul (2019) and achieved by the logistic regression classifier trained on the Document Vectors using Cosine Similarity (DV-ngrams-cosine) proposed in their paper and the Bag-of-Ngrams (BON) vectors scaled by Naive Bayesian weights. While large pre-trained Transformerbased models have shown SOTA results across many datasets and tasks, the aforementioned model has not been surpassed by them, despite being much simpler and pre-trained on the IMDB dataset only. In this paper, we describe an error in the evaluation procedure of this model, which was found when we were trying to analyze its excellent performance on the IMDB dataset. We further show that the previously reported test accuracy of 97.42% is invalid and should be corrected to 93.68%. We also analyze the model performance with different amounts of training data (subsets of the IMDB dataset) and compare it to the Transformer-based RoBERTa model. The results show that while RoBERTa has a clear advantage for larger training sets, the DV-ngramscosine performs better than RoBERTa when the labelled training set is very small (10 or 20 documents). Finally, we introduce a sub-sampling scheme based on Naive Bayesian weights for the training process of the DV-ngrams-cosine, which leads to faster training and better quality.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The word2vec algorithm originally published by Mikolov et al. (2013) is among the most famous methods to train vector representations of words. Soon after the emergence of word2vec, a similar method to build vector representations of documents was originally proposed by Le and Mikolov (2014) and further studied by Mesnil et al. (2015) . It is known under different names, including Paragraph Vectors, Sentence Vectors, doc2vec, etc. This method jointly learns word embeddings and document embeddings such that a binary classifier can predict if a given word occurs in a particular document given only the corresponding embeddings. More formally, the following objective is minimized:",
"cite_spans": [
{
"start": 47,
"end": 68,
"text": "Mikolov et al. (2013)",
"ref_id": "BIBREF7"
},
{
"start": 271,
"end": 292,
"text": "Le and Mikolov (2014)",
"ref_id": "BIBREF2"
},
{
"start": 316,
"end": 336,
"text": "Mesnil et al. (2015)",
"ref_id": "BIBREF6"
},
{
"start": 384,
"end": 402,
"text": "Paragraph Vectors,",
"ref_id": null
},
{
"start": 403,
"end": 420,
"text": "Sentence Vectors,",
"ref_id": null
},
{
"start": 421,
"end": 429,
"text": "doc2vec,",
"ref_id": null
},
{
"start": 430,
"end": 434,
"text": "etc.",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "d\u2208D w\u2208W d [\u2212 log \u03c3(v T d v w ) \u2212 w \u2032 \u223cV log \u03c3(\u2212v T d v w \u2032 )]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(1) Here D denotes the set of documents, W d is the list of words that make up the document d, w \u2032 is a word randomly sampled from the full vocabulary V , also known as a negative sample (Goldberg and Levy, 2014) . Finally, v d and v w are the learnt embeddings of d and w. Intuitively, for each document, an embedding is learnt that has high similarity to the embeddings of those words that occur in this document and low similarity to the embeddings of some random words.",
"cite_spans": [
{
"start": 187,
"end": 212,
"text": "(Goldberg and Levy, 2014)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Later Li et al. (2015) switched from single words to n-grams and observed significant improvements. Building on that, Thongtan and Phienthrakul (2019) studied different objective functions. They have found that the cosine similarity outperforms the dot product, which led to a modified model called the Document Vectors using Cosine Similarity (we will call it DV-ngrams-cosine for short). The new objective is:",
"cite_spans": [
{
"start": 6,
"end": 22,
"text": "Li et al. (2015)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "d\u2208D u\u2208U d [\u2212 log \u03c3(\u03b1cos(v d , v u )) \u2212 u \u2032 \u223cV log \u03c3(\u2212\u03b1cos(v d , v u \u2032 ))],",
"eq_num": "(2)"
}
],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "where U d denotes the set of all n-grams in d, v u is the embedding of the n-gram u from d, v u \u2032 is the embedding of a randomly sampled n-gram, and \u03b1 is a hyperparameter.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In the same paper, the authors proposed an ensemble consisting of the document embeddings from DV-ngrams-cosine and the Bag-of-N-grams vectors scaled by Naive Bayesian weights (NBweighted BON for short). They concatenated these two representations and trained the logistic regression classifier on top. The ensemble was reported to have very high test accuracy (97.42%) on the IMDB movie reviews dataset (Maas et al. (2011) ). To the best of our knowledge, this accuracy remains the SOTA result on IMDB. Even large Transformerbased models pre-trained on a huge amount of texts, both in-domain and out-of-domain, have shown lower accuracy on this dataset (Yang et al., 2019; Suchin et al., 2020; Arefyev et al., 2021) .",
"cite_spans": [
{
"start": 404,
"end": 423,
"text": "(Maas et al. (2011)",
"ref_id": "BIBREF5"
},
{
"start": 654,
"end": 673,
"text": "(Yang et al., 2019;",
"ref_id": "BIBREF11"
},
{
"start": 674,
"end": 694,
"text": "Suchin et al., 2020;",
"ref_id": "BIBREF8"
},
{
"start": 695,
"end": 716,
"text": "Arefyev et al., 2021)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This extraordinary performance of such a simple model motivated us to thoroughly study the model and its implementation trying to understand the reasons behind its success. Unfortunately, during this study, we found a bug in the implementation of the evaluation procedure of the ensemble, which had made the estimation of the accuracy incorrect.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In our paper, we re-evaluate the ensemble as well as its individual components. We show that the originally reported test accuracy of the ensemble (97.42%) is incorrect and shall be corrected to 93.68%, which is only 0.55% higher than the accuracy on pure DV-ngrams-cosine embeddings.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Additionally, we analyze how the amount of training data affects the performance of the ensemble, as well as its individual components, and also the Transformer-based RoBERTa model (Liu et al., 2020) , which has recently shown SOTA or near-SOTA results over a variety of tasks and datasets. Surprisingly, we have observed that DV-ngramscosine outperforms RoBERTa when the number of labelled training examples is small (10 or 20). We also ensemble RoBERTa with DV-ngrams-cosine, but only have achieved a marginal improvement. Finally, we propose a modification for the training process of DV-ngrams-cosine that results in faster training and better accuracy. The code reproducing our experiments is publicly available 1 .",
"cite_spans": [
{
"start": 181,
"end": 199,
"text": "(Liu et al., 2020)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In the aforementioned ensemble proposed by Thongtan and Phienthrakul (2019) , the NBweighted BON and the DV-ngrams-cosine are concatenated and fed into the logistic regression classifier. However, we have found that in the original implementation the two vectors concatenated to obtain a single training or test example usually correspond to two different documents of the same 1 https://github.com/Bgzh/dv_cosine_revisited class (see details in Appendix A). Specifically, the DV-ngrams-cosine vectors and the BON vectors are built from two different files having different orders of examples. As a result, after the concatenation, each input to the logistic regression corresponds to a combination of two examples. Due to the special structure of the files, those examples are guaranteed to belong to the same class and the same subset. For instance, a positive example from the test set is concatenated with another positive example from the test set.",
"cite_spans": [
{
"start": 43,
"end": 75,
"text": "Thongtan and Phienthrakul (2019)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Re-evaluation of the ensemble",
"sec_num": "2"
},
{
"text": "In Appendix B.3 we provide an analysis that shows the reasons of high performance of this concatenation of two representations. From this analysis it follows that most examples from IMDB are correctly classified with high confidence (a large logit) using any of two representations, i.e. they are easy examples. Less than 10% of examples are classified incorrectly by each representation (hard examples), but they often obtain low confidence (a logit near zero). Hard examples are more often combined with easy examples just because of their dominance. In these cases, the logit from the easy example often outweigh the logit from the hard one resulting in the correct final prediction.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-evaluation of the ensemble",
"sec_num": "2"
},
{
"text": "Thus, in both the training and the test sets, hard examples are often combined with simpler examples, making the classification task easier. In this process, the knowledge of the true labels is implicitly exploited to combine the examples this way, in both training and testing. This leads to an incorrect estimation of the classification accuracy for future examples.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-evaluation of the ensemble",
"sec_num": "2"
},
{
"text": "After fixing this issue, we have observed that the combination of different representations of the same document leads to the test accuracy of 93.68% instead of 97.42% originally reported. Compared to the pure DV-ngrams-cosine embeddings, the ensemble improves the test accuracy by 0.55%, not 4.29% reported previously. This improvement also better agrees with the improvements of less than 1% observed by Li et al. (2015) for similar ensembles with the predecessor model DV-ngram. As a sanity check, Appendix B additionally reports the accuracy for different schemes of combining the two representations, showing that higher accuracy can be achieved only by those schemes that exploit the knowledge of the test labels.",
"cite_spans": [
{
"start": 406,
"end": 422,
"text": "Li et al. (2015)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Re-evaluation of the ensemble",
"sec_num": "2"
},
{
"text": "In his section we further analyze the performance of the ensemble described above, comparing it to its individual components as well as to the recently introduced Transformer-based RoBERTa model (Liu et al., 2020) . We study the performance of these models depending on the number of labelled examples in the training set. For a more fair comparison, the most important hyperparameters of each model were tuned on the validation set, employing the train/validation/test split of the IMDB dataset provided by (Suchin et al., 2020) . Subsets of different sizes from 10 to 20000 examples were randomly sampled from the training set. The logistic regression classifier was trained on these subsets using the DV-ngram-cosine embeddings, the NB-weighted BON vectors, or their concatenation as its input representation.",
"cite_spans": [
{
"start": 195,
"end": 213,
"text": "(Liu et al., 2020)",
"ref_id": "BIBREF4"
},
{
"start": 508,
"end": 529,
"text": "(Suchin et al., 2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Further analysis of performance",
"sec_num": "3"
},
{
"text": "We tuned the L2-regularization strength C of the classifier individually for each subset of the training set. Additionally, we multiplied the DV-ngramcosine embeddings before concatenating them to the BON vectors in order to balance the magnitudes of the two representations, which may help the classifier to benefit from both representations. The scaling factor was also selected on the validation set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Further analysis of performance",
"sec_num": "3"
},
{
"text": "The pre-trained RoBERTa base model 2 was finetuned on a part (10 out of 30) of the same subsets of the training set, using the validation set for early stopping. We used a batch size of 32, with a maximum learning rate of 1e-5, recommended by fairseq 3 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Further analysis of performance",
"sec_num": "3"
},
{
"text": "As shown in Fig. 1 , the fine-tuned RoBERTa model usually achieves higher test accuracy. But when the number of labelled training examples is very small (10 or 20), the logistic regression on the DV-ngrams-cosine embeddings shows higher mean test accuracy and lower standard deviation. This result corroborated the notion that small models can be a better choice when the data are scarce.",
"cite_spans": [],
"ref_spans": [
{
"start": 12,
"end": 18,
"text": "Fig. 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Further analysis of performance",
"sec_num": "3"
},
{
"text": "On the other hand, logistic regression on the BON vectors performs significantly worse than all other models across all training set sizes. Finally, we don't observe any significant improvements from the ensembling when the training set size is less than 20k, as the difference is within one standard deviation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Further analysis of performance",
"sec_num": "3"
},
{
"text": "It is important to notice that the DV-ngramscosine embeddings were pre-trained on the indomain examples from the whole IMDB dataset, while RoBERTa was pre-trained on a huge but general-domain corpus. It is likely that the domain adaptation techniques (Suchin et al., 2020) will help RoBERTa when the number of labelled examples is small. However, for our study, we decided to compare the most standard approaches to training the corresponding models.",
"cite_spans": [
{
"start": 251,
"end": 272,
"text": "(Suchin et al., 2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Further analysis of performance",
"sec_num": "3"
},
{
"text": "In this section, we improve the training procedure of DV-ngrams-cosine by applying a sub-sampling procedure based on the Naive Bayesian weights of ngrams (NB Sub-Sampling) in order to make the model focus more on sentiment-related ngrams while building the document embeddings.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NB Sub-Sampling",
"sec_num": "4"
},
{
"text": "Inspired by the previous works (Wang and Manning (2012), Arefyev et al. (2021)), we trained a multinomial Naive Bayesian Classifier and exploited its weights to calculate the importance of each ngram f i for the final classification task:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NB Sub-Sampling",
"sec_num": "4"
},
{
"text": "h i = | log p(f i |y = 1) \u2212 log p(f i |y = 0)| (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NB Sub-Sampling",
"sec_num": "4"
},
{
"text": "In each epoch we put an ngram into training with the probability",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NB Sub-Sampling",
"sec_num": "4"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "p(f i ) = min(exp(h i /n a )/n b , 1),",
"eq_num": "(4)"
}
],
"section": "NB Sub-Sampling",
"sec_num": "4"
},
{
"text": "Test Accuracy % Models trained on the original training set of IMDB (25K) NB-weighted BON 91.29 DV-ngrams-cosine 93.13 DV-ngrams-cosine + NB-weighted BON (Thongtan and Phienthrakul, 2019) #97.42 DV-ngrams-cosine + NB-weighted BON (re-evaluated) 93.68 Models trained using the train/dev split from (Suchin et al., 2020) where n a and n b are the hyperparameters. The choices are purely empirical. We tried different combinations of n a and n b and found 2 and 3 (respectively) to be the best in them. The comparison of the training process with and without NB sub-sampling is shown in Fig. 2 (refer to Appendix C for details of the experiments and the accuracy on the validation set).",
"cite_spans": [
{
"start": 154,
"end": 187,
"text": "(Thongtan and Phienthrakul, 2019)",
"ref_id": "BIBREF9"
},
{
"start": 226,
"end": 244,
"text": "BON (re-evaluated)",
"ref_id": null
},
{
"start": 297,
"end": 318,
"text": "(Suchin et al., 2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 584,
"end": 590,
"text": "Fig. 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "The runs with NB sub-sampling progress faster and show a distinct advantage after 2500 steps. After 30k steps, the runs with NB sub-sampling stagnated and kept fluctuating in a small region; the vanilla runs stagnated after 50k steps, in a lower area. It is also worth noticing that although the labels of the training set are used during pre-training for sub-sampling, we did not observe any significant overfitting due to that. Neither the validation score nor the test score showed a tendency to decay long after reaching the plateau, indicating that this sub-sampling scheme can be used as an add-on to the original model, boosting its performance while not creating additional overfitting trouble.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": null
},
{
"text": "The ensemble proposed in (Thongtan and Phienthrakul (2019)) and described in Section 2 combines two different representations of documents, which are the DV-ngrams-cosine embeddings and the NB-weighted BON vectors. However, we have observed in Section 3 that the BON vectors are quite weak on their own, while RoBERTa outperforms all other models unless the number of examples is very small. Thus, it is interesting if DVngram-cosine can help RoBERTa. In this section, we combine the DV-ngrams-cosine (with or without NB sub-sampling) with the output of the last hidden layer of RoBERTa, and test on the IMDB dataset. Again, the train/validation/test splits by Suchin et al. (2020) were used. A scaling factor on the DV-ngrams-cosine and the hyperparameter C in the logistic regression were tuned on the validation set.",
"cite_spans": [
{
"start": 661,
"end": 681,
"text": "Suchin et al. (2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Ensemble DV-ngrams-cosine and RoBERTa",
"sec_num": "5"
},
{
"text": "The results are shown in Table 1 . Although RoBERTa is a much stronger model than DVngram-cosine, combining them has shown a small improvement of 0.13-0.15%.",
"cite_spans": [],
"ref_spans": [
{
"start": 25,
"end": 32,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Ensemble DV-ngrams-cosine and RoBERTa",
"sec_num": "5"
},
{
"text": "The ensemble featuring the DV-ngrams-cosine reported by Thongtan and Phienthrakul (2019) was re-evaluated. The test accuracy of this ensemble on the IMDB dataset was corrected from 97.42% to 93.68%. The DV-ngrams-cosine embeddings with the logistic regression on top were compared with RoBERTa using different amounts of training data.",
"cite_spans": [
{
"start": 56,
"end": 88,
"text": "Thongtan and Phienthrakul (2019)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "In this comparison, the DV-ngrams-cosine has surprisingly outperformed RoBERTa for a small number of training examples (10 or 20 documents). A sub-sampling scheme based on the Naive Bayesian weights was introduced to the training process of the DV-ngrams-cosine, resulting in faster training and better quality.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "132",
"sec_num": null
},
{
"text": "https://pytorch.org/hub/huggingface_ pytorch-transformers/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/pytorch/fairseq/ blob/main/examples/roberta/README. custom_classification.md",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We are grateful to our anonymous reviewers. This research was partially supported by the Basic Research Program at the HSE University.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Nb-mlm -efficient domain adaptation of masked language models for sentiment analysis",
"authors": [
{
"first": "Nikolay",
"middle": [],
"last": "Arefyev",
"suffix": ""
},
{
"first": "Dmitry",
"middle": [],
"last": "Kharchev",
"suffix": ""
},
{
"first": "Artem",
"middle": [],
"last": "Shelmanov",
"suffix": ""
}
],
"year": 2021,
"venue": "EMNLP",
"volume": "",
"issue": "",
"pages": "9114--9124",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nikolay Arefyev, Dmitry Kharchev, and Artem Shel- manov. 2021. Nb-mlm -efficient domain adaptation of masked language models for sentiment analysis. EMNLP, pages 9114-9124.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "word2vec explained: deriving mikolov et al.'s negative-sampling word-embedding method",
"authors": [
{
"first": "Yoav",
"middle": [],
"last": "Goldberg",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoav Goldberg and Omer Levy. 2014. word2vec ex- plained: deriving mikolov et al.'s negative-sampling word-embedding method. CoRR.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Distributed representations of sentences and documents. ICML",
"authors": [
{
"first": "V",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "Quoc",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Tomas",
"middle": [],
"last": "Mikolov",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "1188--1196",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "V. Quoc Le and Tomas Mikolov. 2014. Distributed representations of sentences and documents. ICML, pages 1188-1196.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Learning document embeddings by predicting n-grams for sentiment classification of long movie reviews",
"authors": [
{
"first": "Bofang",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Tao",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Xiaoyong",
"middle": [],
"last": "Du",
"suffix": ""
},
{
"first": "Deyuan",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Zhe",
"middle": [],
"last": "Zhao",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bofang Li, Tao Liu, Xiaoyong Du, Deyuan Zhang, and Zhe Zhao. 2015. Learning document embeddings by predicting n-grams for sentiment classification of long movie reviews. CoRR.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Ro{bert}a: A robustly optimized {bert} pretraining approach",
"authors": [
{
"first": "Yinhan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Myle",
"middle": [],
"last": "Ott",
"suffix": ""
},
{
"first": "Naman",
"middle": [],
"last": "Goyal",
"suffix": ""
},
{
"first": "Jingfei",
"middle": [],
"last": "Du",
"suffix": ""
},
{
"first": "Mandar",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Danqi",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Mike",
"middle": [],
"last": "Lewis",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "Veselin",
"middle": [],
"last": "Stoyanov",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Ro{bert}a: A robustly optimized {bert} pretraining approach.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Learning word vectors for sentiment analysis",
"authors": [
{
"first": "Andrew",
"middle": [
"L"
],
"last": "Maas",
"suffix": ""
},
{
"first": "Raymond",
"middle": [
"E"
],
"last": "Daly",
"suffix": ""
},
{
"first": "Peter",
"middle": [
"T"
],
"last": "Pham",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "Andrew",
"middle": [
"Y"
],
"last": "Ng",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Potts",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "142--150",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Lin- guistics.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Ensemble of generative and discriminative techniques for sentiment analysis of movie reviews",
"authors": [
{
"first": "Gr\u00e9goire",
"middle": [],
"last": "Mesnil",
"suffix": ""
},
{
"first": "Tomas",
"middle": [],
"last": "Mikolov",
"suffix": ""
},
{
"first": "Marc'aurelio",
"middle": [],
"last": "Ranzato",
"suffix": ""
},
{
"first": "Yoshua",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gr\u00e9goire Mesnil, Tomas Mikolov, Marc'Aurelio Ran- zato, and Yoshua Bengio. 2015. Ensemble of gen- erative and discriminative techniques for sentiment analysis of movie reviews. international conference on learning representations.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Efficient estimation of word representations in vector space",
"authors": [
{
"first": "Tomas",
"middle": [],
"last": "Mikolov",
"suffix": ""
},
{
"first": "Kai",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Greg",
"middle": [],
"last": "Corrado",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Dean",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representa- tions in vector space. CoRR.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Don't stop pretraining: Adapt language models to domains and tasks",
"authors": [
{
"first": "Gururangan",
"middle": [],
"last": "Suchin",
"suffix": ""
},
{
"first": "Marasovi\u0107",
"middle": [],
"last": "Ana",
"suffix": ""
},
{
"first": "Swayamdipta",
"middle": [],
"last": "Swabha",
"suffix": ""
},
{
"first": "Lo",
"middle": [],
"last": "Kyle",
"suffix": ""
},
{
"first": "Beltagy",
"middle": [],
"last": "Iz",
"suffix": ""
},
{
"first": "Downey",
"middle": [],
"last": "Doug",
"suffix": ""
},
{
"first": "Noah",
"middle": [],
"last": "Smith",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2020,
"venue": "ACL",
"volume": "",
"issue": "",
"pages": "8342--8360",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gururangan Suchin, Marasovi\u0107 Ana, Swayamdipta Swabha, Lo Kyle, Beltagy Iz, Downey Doug, and Noah Smith A. 2020. Don't stop pretraining: Adapt language models to domains and tasks. ACL, pages 8342-8360.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Sentiment classification using document embeddings trained with cosine similarity",
"authors": [
{
"first": "Tan",
"middle": [],
"last": "Thongtan",
"suffix": ""
},
{
"first": "Tanasanee",
"middle": [],
"last": "Phienthrakul",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop",
"volume": "",
"issue": "",
"pages": "407--414",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tan Thongtan and Tanasanee Phienthrakul. 2019. Sen- timent classification using document embeddings trained with cosine similarity. In Proceedings of the 57th Annual Meeting of the Association for Com- putational Linguistics: Student Research Workshop, pages 407-414, Florence, Italy. Association for Com- putational Linguistics.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Baselines and bigrams: Simple, good sentiment and topic classification",
"authors": [
{
"first": "Sida",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "90--94",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sida Wang and Christopher Manning. 2012. Baselines and bigrams: Simple, good sentiment and topic clas- sification. In Proceedings of the 50th Annual Meet- ing of the Association for Computational Linguistics (Volume 2: Short Papers), pages 90-94, Jeju Island, Korea. Association for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Xlnet: Generalized autoregressive pretraining for language understanding",
"authors": [
{
"first": "Zhilin",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Zihang",
"middle": [],
"last": "Dai",
"suffix": ""
},
{
"first": "Yiming",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "G",
"middle": [
"Jaime"
],
"last": "Carbonell",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Quoc Le",
"suffix": ""
}
],
"year": 2019,
"venue": "ADVANCES IN NEURAL IN-FORMATION PROCESSING SYSTEMS 32 (NIPS 2019)",
"volume": "",
"issue": "",
"pages": "5754--5764",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhilin Yang, Zihang Dai, Yiming Yang, G. Jaime Car- bonell, Ruslan Salakhutdinov, and V. Quoc Le. 2019. Xlnet: Generalized autoregressive pretraining for lan- guage understanding. ADVANCES IN NEURAL IN- FORMATION PROCESSING SYSTEMS 32 (NIPS 2019), pages 5754-5764.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "The performance of different models on training sets of different sizes. The mean values and standard deviations were calculated over 10 random subsets for RoBERTa and 30 random subsets for other models for each training set size. BON in the legend implies NBweighted BON.",
"type_str": "figure",
"num": null
},
"FIGREF1": {
"uris": null,
"text": "Training process with and without NB subsampling. The test accuracy of the logistic regression built on top of the document vectors is plotted. The mean values and standard deviations were calculated over 3 runs for each type.",
"type_str": "figure",
"num": null
},
"TABREF0": {
"text": "Test results on the IMDB dataset. # indicates incorrect previously reported results.",
"type_str": "table",
"content": "<table><tr><td>(20K/5K)</td><td/></tr><tr><td>DV-ngrams-cosine with NB sub-sampling RoBERTa DV-ngrams-cosine + RoBERTa DV-ngrams-cosine with NB sub-sampling + RoBERTa</td><td>93.36 95.79 95.92 95.94</td></tr><tr><td>Table 1:</td><td/></tr></table>",
"num": null,
"html": null
}
}
}
} |