File size: 55,899 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 |
{
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T06:44:22.963759Z"
},
"title": "Term-Recency for TF-IDF, BM25 and USE Term Weighting",
"authors": [
{
"first": "Divyanshu",
"middle": [],
"last": "Marwah",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Statistics Trinity College Dublin Dublin",
"location": {
"country": "Ireland"
}
},
"email": "marwahd@tcd.ie"
},
{
"first": "Joeran",
"middle": [],
"last": "Beel",
"suffix": "",
"affiliation": {},
"email": "joeran.beel@scss.tcd.ie"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Effectiveness of a recommendation in an Information Retrieval (IR) system is determined by relevancy scores of retrieved results. Term weighting is responsible for computing the relevance scores and consequently differentiating between the terms in a document. However, current term weighting formula like TF-IDF weigh terms only based on term frequency and inverse document frequency irrespective of other important factors. This results in uncertainty in cases when both TF and IDF values are same for more than one document, hence resulting in same term weight values. In this paper, we propose a modification of TF-IDF and other term-weighting schemes that weights terms additionally based on the recency of a term, i.e. the metric based on the year the term occurred for the first time and the document frequency. We modified the term weighting schemes TF-IDF, BM25 and Universal Sentence Encoder (USE) to additionally consider the recency of a term and evaluated them on three datasets. Our modified TF-IDF outperformed the standard TF-IDF on all three datasets; the modified USE outperformed the standard USE on two of the three datasets; the modified BM25 did not outperform the standard BM25 term-weighting scheme.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "Effectiveness of a recommendation in an Information Retrieval (IR) system is determined by relevancy scores of retrieved results. Term weighting is responsible for computing the relevance scores and consequently differentiating between the terms in a document. However, current term weighting formula like TF-IDF weigh terms only based on term frequency and inverse document frequency irrespective of other important factors. This results in uncertainty in cases when both TF and IDF values are same for more than one document, hence resulting in same term weight values. In this paper, we propose a modification of TF-IDF and other term-weighting schemes that weights terms additionally based on the recency of a term, i.e. the metric based on the year the term occurred for the first time and the document frequency. We modified the term weighting schemes TF-IDF, BM25 and Universal Sentence Encoder (USE) to additionally consider the recency of a term and evaluated them on three datasets. Our modified TF-IDF outperformed the standard TF-IDF on all three datasets; the modified USE outperformed the standard USE on two of the three datasets; the modified BM25 did not outperform the standard BM25 term-weighting scheme.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Term Weighting is one of the most crucial tasks in information retrieval and recommender systems. It is method of quantifying terms in a document to determine the importance of the words in the document and the corpus (El-Khair, 2009) . Apart from recommendation engine and information retrieval, term weighting is effective in many scenarios such as text mining, text classification, duplicate image detection (Chum et al., 2008) , document clustering, and even in medical science research. In text categorization and data mining, efficient. term weighting brings a considerable boost in effectiveness (Domeniconi et al., 2015) . Several term weighting approaches are used in different applications basically derived from the frequency and distribution of words in documents (Domeniconi et al., 2015) .",
"cite_spans": [
{
"start": 218,
"end": 234,
"text": "(El-Khair, 2009)",
"ref_id": "BIBREF5"
},
{
"start": 411,
"end": 430,
"text": "(Chum et al., 2008)",
"ref_id": "BIBREF3"
},
{
"start": 603,
"end": 628,
"text": "(Domeniconi et al., 2015)",
"ref_id": "BIBREF4"
},
{
"start": 776,
"end": 801,
"text": "(Domeniconi et al., 2015)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "TF-IDF is one of the classic term weighting approaches, that is most frequently used and was found to be used, for instance, by 83% of textbased research paper recommender systems (Beel et al., 2017) . TF-IDF as the name suggests, is made up of two parts, term frequency (TF) and inverse document frequency (IDF). TF gives the number of times a term occurs in a document. The basis is that the more frequently a term occurs, the more it is important for the context of the document (Beel et al., 2017) . IDF is computed as the inverse frequency of documents containing the searched term. The idea behind this is that a rare term should be given higher importance as compared to frequently occurring terms such as articles, pronouns, etc. There have been numerous researches on TF-IDF, and many extensions and alternatives are suggested. Some other term weighting models used are BM25, LM Dirichlet, Divergence from independence, etc. Text and sentence embedding models such as Universal Sentence Encoder (USE) (Cer et al., 2018) , Google's BERT, InferSent, etc are also used in text classification tasks. These different approaches depend on the type, size of corpus, types of queries, and they use different term metrics to determine the effectiveness of term in a document and corpus.",
"cite_spans": [
{
"start": 180,
"end": 199,
"text": "(Beel et al., 2017)",
"ref_id": "BIBREF0"
},
{
"start": 482,
"end": 501,
"text": "(Beel et al., 2017)",
"ref_id": "BIBREF0"
},
{
"start": 1010,
"end": 1028,
"text": "(Cer et al., 2018)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In case of information retrieval task, there are certain limitations in standard term weighting approaches. Analyzing the simple approach of TF-IDF, that weights term based on the frequency distribution in the corpus. The real issue in this method is the assumption that frequency distribution remains constant with time, without contemplating the diverse contexts for different terms. In short periods, this holds, however, over longer time this assumption fails. For example, consider two terms, \"COVID19\" and \"neural networks\", that have different origin years. Now, there are probably fewer documents containing the term \"COVID19\" than documents containing the term \"neural networks\", simply because \"COVID19\" is a relatively new term, while \"neural networks\" is a term being used since decades. However, they would be weighted similarly without considering the difference in the origins. The issue that terms have temporal distributions of frequency, not just space distribution is unaccounted when using the standard term weighting methodologies.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Considering this uncertainty in term weighting, we suggest a time-normalized term weighting approach, which reflects the age of a term. As the vocabulary changes over time, our intuition is to identify a term's age based on its first usage and current year and distinguish between the documents based on the age of the terms used. Hence, we propose to weigh terms not only on their frequency distributions but also temporal distributions. Furthermore, we demonstrate the significance of adding a time-based feature by comparing our method with state-of-the-art baseline models, that is, TF-IDF, BM25 and USE embedding. Experimental results show substantial improvements over the baseline models for similar recommendations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "TF-IDF is a relatively old approach and there have been many studies comparing the results of TF-IDF with other states of the art term weighting schemes. Also, different researches have suggested novel variants and enhancing algorithms solving various issues. For instance, (Beel et al., 2017) points out the lack of personalization in classic TF-IDF. The authors have highlighted the issue of access to the document corpus for calculating IDF and another issue of ignoring the information from the user's document collection for recommendations and user modelling. Thus, a novel term weighting is suggested, that does not require the document corpus and uses the user's document collection for user modelling.",
"cite_spans": [
{
"start": 274,
"end": 293,
"text": "(Beel et al., 2017)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "In another paper, (Domeniconi et al., 2015 ) points out the problem of using IDF in text classification. The basic idea behind IDF is that a term occurring frequently has negligible distinguishing power, however, in the case of text classification, this might not be true, because, highly frequent terms in different documents of the same category can be helpful in text classification. Hence, the authors suggested a supervised learning approach to calculate IDF excluding the category under consideration.",
"cite_spans": [
{
"start": 18,
"end": 42,
"text": "(Domeniconi et al., 2015",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "(Park et al., 2005) suggests a novel approach to term weighting based on the term positions along with the TF and IDF terms. The authors studied the term patterns that occur in the documents using the wavelet transform method. The paper also suggests that the documents are ranked more relevant if the query terms are close to each other.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Utilizing temporal feature has also proved to be an efficient way for recommenders and time normalized recommendations are certainly receiving growing application in recent times (Campos et al., 2014) . One of the researches (Kacem et al., 2014) , suggests usage of time-normalized term weighting for user modelling. The authors have used the time of social/web search of terms to form the short and long-term contexts and further creating a user profile based on the same. The comparative study of this algorithm with the standard TF-IDF suggests a significant improvement in results centered on the time normalized user models. Considering this research of temporal context's effect on term weights, we propose a Time Normalized TF-IDF algorithm for information retrieval and recommender system, discussed and implemented in this paper.",
"cite_spans": [
{
"start": 179,
"end": 200,
"text": "(Campos et al., 2014)",
"ref_id": "BIBREF1"
},
{
"start": 225,
"end": 245,
"text": "(Kacem et al., 2014)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "In a classic term weighting approach, the terms are weighted irrespective of the different contexts or usage or recency. In this paper, we try to emphasize on the importance of term recency in relevant results retrieval. The premise for this algorithm is that, if a term is devised newly then there are probably a lesser number of documents containing the term compared to the term which is being used for a longer duration of time. In this algorithm, we introduce a time factor along with the regular TF-IDF values. This time-based factor is formulated from the origin year of the word and the document frequency of the term giving the metric as documents per year. For a given term w in document d D, where D is the document corpus D with size N , term-age is calculated as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "t w,D = log(df w,D /(y dif f + 1)) 1 ,",
"eq_num": "(1)"
}
],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "where y dif f is calculated as :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "y dif f = y current \u2212 y origin (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "where y origin is the year of first usage of the word and Y current is the present year. This current year remains constant in the calculations, giving us the sort of age for the word. We take the logarithm of the terms to normalize the value, since this can go to a large number based on the size of the corpus. Also, we take up the absolute value of log, so that we don't have negative weight values. The y origin can be traced from multiple places depending on the problem statement. For example, if a research paper recommender is being developed, the origin year can be retrieved as the year of first occurrence of the term in the recommendation corpus. Or in case of web search, time of first search of the term can be used. Likewise, for some instances the terms can be traced to their etymology and the year of first occurrence can be fetched. Now the updated formula for term weight calculation for tTF-IDF is given as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "wt w,d = t w,D * tf w,d * log( N df w,d )",
"eq_num": "(3)"
}
],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "where t ( w, D) is the time-factor calculated value in equation (1) tf ( w, d) is the number of times term w occurs in a document d, and df ( w, D) is the number of documents in which w appears in D. Likewise, in case of time normalized BM25 (tBM25) model, the term age, t ( w, D) is multiplied to the classic BM25 formula for the time normalized model. For USE embedding approach, cosine similarity is used to calculate the term weights. In the time normalized model, we multiply the term age factor, t ( w, D) with the cosine similarity function to get the updated time normalized USE (tUSE) model. Now, assume that a term is new and occurs in reasonable number of documents, then the value of t ( w, D) will be large and hence the term weight will be large. Similarly, if the term is being used for many years and is occurring in many documents, it will relatively reduce the value of the time-factor, thus giving it low importance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "A caution which needs to be taken while implementing this algorithm is to check for more commonly occurring non relevant terms which are normalized by using IDF should not get boosted. This can be taken care of while calculating the value of y origin , and such terms can be ignored so they don't boost up the term weights based on non-relevant terms.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "4 Implementation 4.1 Data 4.1.1 TREC Washington Post Corpus (Post, 2018) This collection contains 608,180 news articles and blog posts, along with 50 queries from TREC -2018 news background linking task (Soboroff et al., 2018) , and expected set of results. For the purpose of testing our hypothesis, we use a sample of 20909 documents with approximately 2400 relevant documents. However, this has been done only for time-based index due to scalability and resource constraints. And the term age is still calculated considering the entire corpus and does not affect the algorithmic logic. The relevant fields in this dataset are id, URL, title, author, and article text.",
"cite_spans": [
{
"start": 60,
"end": 72,
"text": "(Post, 2018)",
"ref_id": null
},
{
"start": 203,
"end": 226,
"text": "(Soboroff et al., 2018)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "4.1.2 Web Answer Passage(WebAP) Dataset (Keikha et al., 2014 (Keikha et al., , 2015 This collection contains 8027 articles from the web, which are answers to 82 TREC queries. The dataset contains the following fields: unique document id, target question id, and passage. The results contain 50 relevant documents, given as question id, document number and relevance as ranked from 1 to 50. (Wang et al., 2013) This dataset is collected from CiteULike and Google Scholar and contains 17013 documents with the following fields: document id, title of the research paper, and abstract. We are given another file in this dataset, that contains the referenced articles for every document. We have randomly selected 116 test topics having exactly 10 citations to be used as our ground truth.",
"cite_spans": [
{
"start": 40,
"end": 60,
"text": "(Keikha et al., 2014",
"ref_id": "BIBREF7"
},
{
"start": 61,
"end": 83,
"text": "(Keikha et al., , 2015",
"ref_id": "BIBREF8"
},
{
"start": 390,
"end": 409,
"text": "(Wang et al., 2013)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Time Normalized Term Weighting",
"sec_num": "3"
},
{
"text": "We implemented text-based recommendation systems using the data mentioned in the last section. This is implemented using TF-IDF, BM25 and USE embedding models. Further, we devised an algorithm to calculate t ( w, D) as described earlier. And scoring is done using customized plugins. Finally, we compare the results of different algorithms using the evaluation metrics described later. The first index is created with the same mapping structure as given in the input dataset files. For the second index, we add a time normalized term weight parameter as a payload to the terms while indexing the documents. A third index is created using the time normalized index for USE embedding model. In this index, we calculate the term vectors using the pre-trained TensorFlow model (Cer et al., 2018) and store it in a 512-length vector field. This methodology remains the same for all the datasets. Following steps are used for calculations of term age:",
"cite_spans": [
{
"start": 773,
"end": 791,
"text": "(Cer et al., 2018)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "\u2022 Consider the article text of the document and fetch the origin year for every word from etymonline.com",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "\u2022 Calculate the difference in number of years from the year of first occurrence, to the current year. We have assumed the base year for our corpus to be 2017(TREC News), 2015(Web AP) and 2019(CiteULike) since that is the year of latest publications. This has been done for uniform term weighting across the corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "\u2022 Now the term weight is calculated using the formula given in section 3.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "An important part to note is, that every term is not given a term weight, this happens if the origin year of the term is not traceable, or the terms are most frequently used such as articles, or prepositions. We have used the following evaluation metrics to evaluate the significance of retrieved results:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "\u2022 Precision @10: We have calculated the precision value for top 10 fetched results on the given set of input queries and take an average of the results for comparison.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "\u2022 Recall: For calculating the recall, we have considered the queries having less than 100 results, in case of TREC news. And then recall is calculated as the number of relevant retrieved document divided by the number of relevant documents present in the index. For other datasets, since the number of relevant results is fixed, so the value of precision and recall remains the same.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "\u2022 F1 Score: Since F1 score uses both the precision and recall values, so for calculating the precision scores, we have used the same results from the recall measure and used a fixed denominator as the number of retrieved results. Formula used for F1 score is given as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "F 1 = 2 * P recision * Recal P recision + Recall (4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "\u2022 Normalized Discounted Cumulative Gain: DCG is calculated at specific rank position p, given by",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "DCG p = p i=1 rel i / log 2 (i + 1)",
"eq_num": "(5)"
}
],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "where, rel i is the relevance score of a document at position i. And NDCG is calculated by considering the DCG of ideal order along with the DCG values and is given by",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "nDCG p = DCG p /IDCG p",
"eq_num": "(6)"
}
],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "5 Results and Discussion",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "In 2 out of 3 algorithms, our term-recency modification improved the performance notably. When measured by p@10, tTF-IDF outperformed TF-IDF by an average 47% and tUSE outperformed USE in 2 of the 3 datasets by 14.3% but performed 50% worse in the other dataset ( Figure 1) . The time normalized BM25 version, however, performed 32% worse than BM25. NDCG@10 leads to similar results ( Figure 2 ). For CiteULike dataset, NDCG cannot be calculated, since there is no ranking specified for the citations. On closer analysis of the BM25 model, we see 27 out of 50 queries gave better or almost similar results in case of time normalized model when compared to the classic approach. These results are also promising and need to be worked upon for better results in the future work.",
"cite_spans": [],
"ref_spans": [
{
"start": 264,
"end": 273,
"text": "Figure 1)",
"ref_id": null
},
{
"start": 385,
"end": 393,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "For calculating the recall and F1 scores, we fetch the top 100 results for the given query sets. For Figure 2 : NDCG@10 comparison Figure 3 : Precision, Recall, F1@100 scores for TREC News uniformity in the metric calculation, we compute the precision scores as well. The result metrics are shown in Figure 3 . We see a 150% improvement in the tTF-IDF model over TF-IDF and a 31% improvement in tUSE model over the USE model. Recall and F1 scores for WebAP and CiteULike would be same as precision scores since the number of relevant results in dataset remains the same, so they are not shown.",
"cite_spans": [],
"ref_spans": [
{
"start": 101,
"end": 109,
"text": "Figure 2",
"ref_id": null
},
{
"start": 131,
"end": 139,
"text": "Figure 3",
"ref_id": null
},
{
"start": 300,
"end": 308,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "Analyzing the tUSE model in WebAP dataset, we see it does not perform well against the USE model. One of the possible reasons for this might be the size of the corpus used, that is, TREC news corpus has approximately 600k documents while Web AP dataset has just 6k documents, which is 100 times less than the former dataset. However, this is an inference based on the results retrieved and has not been verified. There might be other possible reasons, such as the size of documents, size of queries used, number of proper nouns in the queries, etc. Or probably term age might not be a relevant metric for this dataset. These possible reasons still need to be analyzed before affirming out a conclusion on these contrasting results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Architecture/Methodology",
"sec_num": "4.2"
},
{
"text": "In this paper, we proposed a novel algorithm for term weighting. The presented approach shows the significance of temporal distribution along with existing space distribution of terms. We suggest the scheming of a term recency parameter based on the origin of the word and the usage in the document corpus. This factor is used along with the standard weighting values (such as TF and IDF) for relevance scoring in the information retrieval system. The algorithm is tested on a news dataset, with queries trying to find links with the documents, Web answer retrieval dataset and research papers citations dataset. We have also extended the algorithm to other text embedding models that are BM25 and USE.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "Experiments conducted on the IR system show that term-recency based TF-IDF and tUSE model outperforms the classic TF-IDF and classic USE algorithms with a significant margin when measured in terms of average precision, recall, F1 and NDCG. It has set up a strong premise for our ongoing research on ways to improve recommendation effectiveness. Future works for this can be to find ways to improve the time-based BM25 model and testing the algorithm's performance in other tasks such as text classification, and user modelling. Furthermore, we also plan to test different normalization factors for calculating the term age and then using it in the scoring algorithm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "This is an updated formula with an added 1 in the denominator, for our experiments, we used the older version of formula",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We would like to thank Sneha Srivastava, Rachit Rastogi and Shivam Khanduri for their comments, suggestions and feedback.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "TF-IDuF: A novel term-weighting scheme for user modeling based on users' personal document collections",
"authors": [
{
"first": "Joeran",
"middle": [],
"last": "Beel",
"suffix": ""
},
{
"first": "Stefan",
"middle": [],
"last": "Langer",
"suffix": ""
},
{
"first": "Bela",
"middle": [],
"last": "Gipp",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Joeran Beel, Stefan Langer, and Bela Gipp. 2017. TF- IDuF: A novel term-weighting scheme for user mod- eling based on users' personal document collections. University of Illinois.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Time-aware recommender systems: a comprehensive survey and analysis of existing evaluation protocols. User Modeling and User-Adapted Interaction",
"authors": [
{
"first": "Pedro",
"middle": [
"G"
],
"last": "Campos",
"suffix": ""
},
{
"first": "Fernando",
"middle": [],
"last": "D\u00edez",
"suffix": ""
},
{
"first": "Iv\u00e1n",
"middle": [],
"last": "Cantador",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "24",
"issue": "",
"pages": "67--119",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pedro G. Campos, Fernando D\u00edez, and Iv\u00e1n Cantador. 2014. Time-aware recommender systems: a compre- hensive survey and analysis of existing evaluation protocols. User Modeling and User-Adapted Inter- action, 24(1-2):67-119.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Universal sentence encoder",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Cer",
"suffix": ""
},
{
"first": "Yinfei",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Sheng-Yi",
"middle": [],
"last": "Kong",
"suffix": ""
},
{
"first": "Nan",
"middle": [],
"last": "Hua",
"suffix": ""
},
{
"first": "Nicole",
"middle": [],
"last": "Limtiaco",
"suffix": ""
},
{
"first": "Rhomni",
"middle": [],
"last": "St John",
"suffix": ""
},
{
"first": "Noah",
"middle": [],
"last": "Constant",
"suffix": ""
},
{
"first": "Mario",
"middle": [],
"last": "Guajardo-Cespedes",
"suffix": ""
},
{
"first": "Steve",
"middle": [],
"last": "Yuan",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1803.11175"
]
},
"num": null,
"urls": [],
"raw_text": "Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, and Chris Tar. 2018. Universal sentence encoder. arXiv preprint arXiv:1803.11175.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Near duplicate image detection: min-hash and tf-idf weighting",
"authors": [
{
"first": "Ondrej",
"middle": [],
"last": "Chum",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Philbin",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Zisserman",
"suffix": ""
}
],
"year": 2008,
"venue": "BMVC",
"volume": "810",
"issue": "",
"pages": "812--815",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ondrej Chum, James Philbin, and Andrew Zisserman. 2008. Near duplicate image detection: min-hash and tf-idf weighting. In BMVC, volume 810, pages 812-815.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "A study on term weighting for text categorization: A novel supervised variant of tf. idf",
"authors": [
{
"first": "Giacomo",
"middle": [],
"last": "Domeniconi",
"suffix": ""
},
{
"first": "Gianluca",
"middle": [],
"last": "Moro",
"suffix": ""
},
{
"first": "Roberto",
"middle": [],
"last": "Pasolini",
"suffix": ""
},
{
"first": "Claudio",
"middle": [],
"last": "Sartori",
"suffix": ""
}
],
"year": 2015,
"venue": "DATA",
"volume": "",
"issue": "",
"pages": "26--37",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Giacomo Domeniconi, Gianluca Moro, Roberto Pa- solini, and Claudio Sartori. 2015. A study on term weighting for text categorization: A novel super- vised variant of tf. idf. In DATA, pages 26-37.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Term Weighting",
"authors": [
{
"first": "Ibrahim Abu",
"middle": [],
"last": "El-Khair",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "3037--3040",
"other_ids": {
"DOI": [
"10.1007/978-0-387-39940-9_943"
]
},
"num": null,
"urls": [],
"raw_text": "Ibrahim Abu El-Khair. 2009. Term Weighting, pages 3037-3040. Springer US, Boston, MA.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Time-sensitive user profile for optimizing search personlization",
"authors": [
{
"first": "Ameni",
"middle": [],
"last": "Kacem",
"suffix": ""
},
{
"first": "Mohand",
"middle": [],
"last": "Boughanem",
"suffix": ""
},
{
"first": "Rim",
"middle": [],
"last": "Faiz",
"suffix": ""
}
],
"year": 2014,
"venue": "International conference on user modeling, adaptation, and personalization",
"volume": "",
"issue": "",
"pages": "111--121",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ameni Kacem, Mohand Boughanem, and Rim Faiz. 2014. Time-sensitive user profile for optimizing search personlization. In International conference on user modeling, adaptation, and personalization, pages 111-121. Springer.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Retrieving passages and finding answers",
"authors": [
{
"first": "Mostafa",
"middle": [],
"last": "Keikha",
"suffix": ""
},
{
"first": "Jae",
"middle": [
"Hyun"
],
"last": "Park",
"suffix": ""
},
{
"first": "W",
"middle": [
"Bruce"
],
"last": "Croft",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Sanderson",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 2014 Australasian Document Computing Symposium",
"volume": "",
"issue": "",
"pages": "81--84",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mostafa Keikha, Jae Hyun Park, W. Bruce Croft, and Mark Sanderson. 2014. Retrieving passages and finding answers. In Proceedings of the 2014 Aus- tralasian Document Computing Symposium, pages 81-84.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Web answer passages (webap) dataset",
"authors": [
{
"first": "Mostafa",
"middle": [],
"last": "Keikha",
"suffix": ""
},
{
"first": "Jae",
"middle": [
"Hyun"
],
"last": "Park",
"suffix": ""
},
{
"first": "W",
"middle": [
"Bruce"
],
"last": "Croft",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Sanderson",
"suffix": ""
}
],
"year": 2015,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mostafa Keikha, Jae Hyun Park, W. Bruce Croft, and Mark Sanderson. 2015. Web answer passages (we- bap) dataset.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A novel document retrieval method using the discrete wavelet transform",
"authors": [
{
"first": "A",
"middle": [
"F"
],
"last": "Laurence",
"suffix": ""
},
{
"first": "Kotagiri",
"middle": [],
"last": "Park",
"suffix": ""
},
{
"first": "Marimuthu",
"middle": [],
"last": "Ramamohanarao",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Palaniswami",
"suffix": ""
}
],
"year": 2005,
"venue": "ACM Transactions on Information Systems (TOIS)",
"volume": "23",
"issue": "3",
"pages": "267--298",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Laurence A. F. Park, Kotagiri Ramamohanarao, and Marimuthu Palaniswami. 2005. A novel document retrieval method using the discrete wavelet trans- form. ACM Transactions on Information Systems (TOIS), 23(3):267-298.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Trec washington post corpus",
"authors": [],
"year": 2018,
"venue": "Washington Post",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Washington Post. 2018. Trec washington post corpus.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Trec 2018 news track overview",
"authors": [
{
"first": "Ian",
"middle": [],
"last": "Soboroff",
"suffix": ""
},
{
"first": "Shudong",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "Donna",
"middle": [],
"last": "Harman",
"suffix": ""
}
],
"year": 2018,
"venue": "The Twenty-Seventh Text RE-trieval Conference (TREC 2018) Proceedings",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ian Soboroff, Shudong Huang, and Donna Harman. 2018. Trec 2018 news track overview. In The Twenty-Seventh Text RE-trieval Conference (TREC 2018) Proceedings.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Collaborative topic regression with social regularization for tag recommendation",
"authors": [
{
"first": "Hao",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Binyi",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Wu-Jun",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2013,
"venue": "Twenty-Third International Joint Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hao Wang, Binyi Chen, and Wu-Jun Li. 2013. Collabo- rative topic regression with social regularization for tag recommendation. In Twenty-Third International Joint Conference on Artificial Intelligence.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "Figure 1: P@10 comparison",
"type_str": "figure",
"num": null
}
}
}
} |