Research:Language-Agnostic Topic Classification/Wikidata model performance

From Meta, a Wikimedia project coordination wiki
Jump to navigation Jump to search

This page provides details about the current (as of May 2020) model for Wikidata topic classification based on statements:

  • Code: https://github.com/geohci/wikidata-topic-model
  • Model architecture: multi-label fastText supervised model
  • Training data: 4,911,836 Wikidata items that have English sitelinks (and corresponding topic labels)
  • Epochs: 25
  • Learning rate: 0.1
  • Window size: 20
  • Min count (under which QID is not retained in vocab): 3
  • No pre-trained embeddings used
  • Embeddings dimension: 50
  • Total number of model params: 3200 (50 x 64)
  • Vocab size: 429,206
  • Total number of embeddings params: 21,460,300 (vocab size * embeddings dimension)
  • Model size on disk: 89.1 MB

Test Results[edit]

Regarding the test results below, there is a strong limitation in that there is no labeled data for Wikidata items that do not have English Wikipedia sitelinks. It is very possible that the performance degrades for these items (assuming they also have different types of properties and values).

Overall performance:

  • Precision: 0.886 micro; 0.837 macro
  • Recall: 0.764 micro; 0.629 macro
  • F1: 0.817 micro; 0.712 macro
  • Avg pre.: 0.872 micro; 0.746 macro

Label-specific:

topic n TP FP TN FN precision recall f1 avg_pre
Culture.Biography.Biography* 85659 83162 2455 175965 2497 0.971 0.971 0.971 0.990
Culture.Biography.Women 10669 7693 2531 250879 2976 0.752 0.721 0.736 0.801
Culture.Food and drink 1738 823 166 262175 915 0.832 0.474 0.604 0.586
Culture.Internet culture 2326 1646 154 261599 680 0.914 0.708 0.798 0.816
Culture.Linguistics 4705 3454 246 259128 1251 0.934 0.734 0.822 0.863
Culture.Literature 8879 6314 840 254360 2565 0.883 0.711 0.788 0.840
Culture.Media.Books 2924 2448 251 260904 476 0.907 0.837 0.871 0.892
Culture.Media.Entertainment 2500 1207 279 261300 1293 0.812 0.483 0.606 0.632
Culture.Media.Films 8160 7409 398 255521 751 0.949 0.908 0.928 0.948
Culture.Media.Media* 35326 29419 3174 225579 5907 0.903 0.833 0.866 0.935
Culture.Media.Music 13236 11143 1384 249459 2093 0.890 0.842 0.865 0.932
Culture.Media.Radio 1434 1108 47 262598 326 0.959 0.773 0.856 0.841
Culture.Media.Software 754 190 117 263208 564 0.619 0.252 0.358 0.390
Culture.Media.Television 5156 3765 493 258430 1391 0.884 0.730 0.800 0.839
Culture.Media.Video games 1800 1505 43 262236 295 0.972 0.836 0.899 0.896
Culture.Performing arts 1923 1024 232 261924 899 0.815 0.533 0.644 0.660
Culture.Philosophy and religion 7148 3128 1164 255767 4020 0.729 0.438 0.547 0.576
Culture.Sports 42560 39322 2485 219034 3238 0.941 0.924 0.932 0.970
Culture.Visual arts.Architecture 7527 5418 888 255664 2109 0.859 0.720 0.783 0.858
Culture.Visual arts.Comics and Anime 1418 790 127 262534 628 0.862 0.557 0.677 0.724
Culture.Visual arts.Fashion 609 278 108 263362 331 0.720 0.456 0.559 0.576
Culture.Visual arts.Visual arts* 12362 8219 1754 249963 4143 0.824 0.665 0.736 0.821
Geography.Geographical 14848 11192 1561 247670 3656 0.878 0.754 0.811 0.877
Geography.Regions.Africa.Africa* 5943 3570 815 257321 2373 0.814 0.601 0.691 0.740
Geography.Regions.Africa.Central Africa 0 0 2 264077 0 0.000 0 0 nan
Geography.Regions.Africa.Eastern Africa 352 179 38 263689 173 0.825 0.509 0.629 0.597
Geography.Regions.Africa.Northern Africa 950 514 165 262964 436 0.757 0.541 0.631 0.626
Geography.Regions.Africa.Southern Africa 915 532 151 263013 383 0.779 0.581 0.666 0.677
Geography.Regions.Africa.Western Africa 555 233 106 263418 322 0.687 0.420 0.521 0.565
Geography.Regions.Americas.Central America 2382 1286 242 261455 1096 0.842 0.540 0.658 0.674
Geography.Regions.Americas.North America 41312 29991 5616 217151 11321 0.842 0.726 0.780 0.876
Geography.Regions.Americas.South America 4847 3518 730 258502 1329 0.828 0.726 0.774 0.831
Geography.Regions.Asia.Asia* 33925 25211 3444 226710 8714 0.880 0.743 0.806 0.885
Geography.Regions.Asia.Central Asia 546 287 70 263463 259 0.804 0.526 0.636 0.613
Geography.Regions.Asia.East Asia 8549 6137 1008 254522 2412 0.859 0.718 0.782 0.821
Geography.Regions.Asia.North Asia 761 245 168 263150 516 0.593 0.322 0.417 0.441
Geography.Regions.Asia.South Asia 11637 8775 853 251589 2862 0.911 0.754 0.825 0.877
Geography.Regions.Asia.Southeast Asia 4326 2714 441 259312 1612 0.860 0.627 0.726 0.738
Geography.Regions.Asia.West Asia 8414 6248 628 255037 2166 0.909 0.743 0.817 0.869
Geography.Regions.Europe.Eastern Europe 9697 7692 844 253538 2005 0.901 0.793 0.844 0.903
Geography.Regions.Europe.Europe* 53802 40392 7230 203047 13410 0.848 0.751 0.796 0.891
Geography.Regions.Europe.Northern Europe 20509 13631 2302 241268 6878 0.856 0.665 0.748 0.833
Geography.Regions.Europe.Southern Europe 9731 6624 1340 253008 3107 0.832 0.681 0.749 0.823
Geography.Regions.Europe.Western Europe 14305 10560 2201 247573 3745 0.828 0.738 0.780 0.858
Geography.Regions.Oceania 11602 9044 614 251863 2558 0.936 0.780 0.851 0.885
History and Society.Business and economics 6855 2908 1427 255797 3947 0.671 0.424 0.520 0.561
History and Society.Education 5418 2947 855 257806 2471 0.775 0.544 0.639 0.663
History and Society.History 7601 3247 1162 255316 4354 0.736 0.427 0.541 0.593
History and Society.Military and warfare 10000 6068 1008 253071 3932 0.858 0.607 0.711 0.760
History and Society.Politics and government 11151 6158 1136 251792 4993 0.844 0.552 0.668 0.738
History and Society.Society 9341 4801 960 253778 4540 0.833 0.514 0.636 0.680
History and Society.Transportation 10462 8078 648 252969 2384 0.926 0.772 0.842 0.881
STEM.Biology 23603 21686 332 240144 1917 0.985 0.919 0.951 0.974
STEM.Chemistry 978 498 151 262950 480 0.767 0.509 0.612 0.618
STEM.Computing 1893 858 255 261931 1035 0.771 0.453 0.571 0.552
STEM.Earth and environment 3239 2141 247 260593 1098 0.897 0.661 0.761 0.765
STEM.Engineering 3870 2470 323 259886 1400 0.884 0.638 0.741 0.745
STEM.Libraries & Information 496 225 56 263527 271 0.801 0.454 0.579 0.588
STEM.Mathematics 314 74 32 263733 240 0.698 0.236 0.352 0.337
STEM.Medicine & Health 3591 1745 337 260151 1846 0.838 0.486 0.615 0.650
STEM.Physics 697 172 114 263268 525 0.601 0.247 0.350 0.357
STEM.STEM* 41446 33119 1809 220824 8327 0.948 0.799 0.867 0.943
STEM.Space 1625 1208 70 262384 417 0.945 0.743 0.832 0.835
STEM.Technology 3399 1127 342 260338 2272 0.767 0.332 0.463 0.498