neural-network - 预训练的 GloVe 矢量文件(例如 glove.6B.50d.txt)中的 "unk"是什么？

Question

我在下载的手套矢量文件 glove.6B.50d.txt 中找到“unk”标记 from https://nlp.stanford.edu/projects/glove/ .它的值如下:

unk -0.79149 0.86617 0.11998 0.00092287 0.2776 -0.49185 0.50195 0.00060792 -0.25845 0.17865 0.2535 0.76572 0.50664 0.4025 -0.0021388 -0.28397 -0.50324 0.30449 0.51779 0.01509 -0.35031 -1.1278 0.33253 -0.3525 0.041326 1.0863 0.03391 0.33564 0.49745 -0.070131 -1.2192 -0.48512 -0.038512 -0.13554 -0.1638 0.52321 -0.31318 -0.1655 0.11909 -0.15115 -0.15621 -0.62655 -0.62336 -0.4215 0.41873 -0.92472 1.1049 -0.29996 -0.0063003 0.3954

它是用于未知单词的标记还是某种缩写？

Answer 1

预训练的 GloVe 文件中的 unk 标记不是未知标记!

请参阅 Jeffrey Pennington(GloVe 作者)写道的 google groups thread:

The pre-trained vectors do not have an unknown token, and currently the code just ignores out-of-vocabulary words when producing the co-occurrence counts.

这是一种像任何其他语言一样在语料库中出现“unk”时学习的嵌入(这似乎偶尔会发生!)

相反，Pennington 建议(在同一篇文章中):

...I've found that just taking an average of all or a subset of the word vectors produces a good unknown vector.

您可以使用以下代码执行此操作(应该适用于任何预训练的 GloVe 文件):

import numpy as np

GLOVE_FILE = 'glove.6B.50d.txt'

# Get number of vectors and hidden dim
with open(GLOVE_FILE, 'r') as f:
    for i, line in enumerate(f):
        pass
n_vec = i + 1
hidden_dim = len(line.split(' ')) - 1

vecs = np.zeros((n_vec, hidden_dim), dtype=np.float32)

with open(GLOVE_FILE, 'r') as f:
    for i, line in enumerate(f):
        vecs[i] = np.array([float(n) for n in line.split(' ')[1:]], dtype=np.float32)

average_vec = np.mean(vecs, axis=0)
print(average_vec)

对于 glove.6B.50d.txt 这给出:

[-0.12920076 -0.28866628 -0.01224866 -0.05676644 -0.20210965 -0.08389011
  0.33359843  0.16045167  0.03867431  0.17833012  0.04696583 -0.00285802
  0.29099807  0.04613704 -0.20923874 -0.06613114 -0.06822549  0.07665912
  0.3134014   0.17848536 -0.1225775  -0.09916984 -0.07495987  0.06413227
  0.14441176  0.60894334  0.17463093  0.05335403 -0.01273871  0.03474107
 -0.8123879  -0.04688699  0.20193407  0.2031118  -0.03935686  0.06967544
 -0.01553638 -0.03405238 -0.06528071  0.12250231  0.13991883 -0.17446303
 -0.08011883  0.0849521  -0.01041659 -0.13705009  0.20127155  0.10069408
  0.00653003  0.01685157]

而且由于使用较大的 glove 文件执行此操作需要大量计算，因此我继续为您计算了 glove.840B.300d.txt 的矢量:

0.22418134 -0.28881392 0.13854356 0.00365387 -0.12870757 0.10243822 0.061626635 0.07318011 -0.061350107 -1.3477012 0.42037755 -0.063593924 -0.09683349 0.18086134 0.23704372 0.014126852 0.170096 -1.1491593 0.31497982 0.06622181 0.024687296 0.076693475 0.13851812 0.021302193 -0.06640582 -0.010336159 0.13523154 -0.042144544 -0.11938788 0.006948221 0.13333307 -0.18276379 0.052385733 0.008943111 -0.23957317 0.08500333 -0.006894406 0.0015864656 0.063391194 0.19177166 -0.13113557 -0.11295479 -0.14276934 0.03413971 -0.034278486 -0.051366422 0.18891625 -0.16673574 -0.057783455 0.036823478 0.08078679 0.022949161 0.033298038 0.011784158 0.05643189 -0.042776518 0.011959623 0.011552498 -0.0007971594 0.11300405 -0.031369694 -0.0061559738 -0.009043574 -0.415336 -0.18870236 0.13708843 0.005911723 -0.113035575 -0.030096142 -0.23908928 -0.05354085 -0.044904727 -0.20228513 0.0065645403 -0.09578946 -0.07391877 -0.06487607 0.111740574 -0.048649278 -0.16565254 -0.052037314 -0.078968436 0.13684988 0.0757494 -0.006275573 0.28693774 0.52017444 -0.0877165 -0.33010918 -0.1359622 0.114895485 -0.09744406 0.06269521 0.12118575 -0.08026362 0.35256687 -0.060017522 -0.04889904 -0.06828978 0.088740796 0.003964443 -0.0766291 0.1263925 0.07809314 -0.023164088 -0.5680669 -0.037892066 -0.1350967 -0.11351585 -0.111434504 -0.0905027 0.25174105 -0.14841858 0.034635577 -0.07334565 0.06320108 -0.038343467 -0.05413284 0.042197507 -0.090380974 -0.070528865 -0.009174437 0.009069661 0.1405178 0.02958134 -0.036431845 -0.08625681 0.042951006 0.08230793 0.0903314 -0.12279937 -0.013899368 0.048119213 0.08678239 -0.14450377 -0.04424887 0.018319942 0.015026873 -0.100526 0.06021201 0.74059093 -0.0016333034 -0.24960588 -0.023739101 0.016396184 0.11928964 0.13950661 -0.031624354 -0.01645025 0.14079992 -0.0002824564 -0.08052984 -0.0021310581 -0.025350995 0.086938225 0.14308536 0.17146006 -0.13943303 0.048792403 0.09274929 -0.053167373 0.031103406 0.012354865 0.21057427 0.32618305 0.18015954 -0.15881181 0.15322933 -0.22558987 -0.04200665 0.0084689725 0.038156632 0.15188617 0.13274793 0.113756925 -0.095273495 -0.049490947 -0.10265804 -0.27064866 -0.034567792 -0.018810693 -0.0010360252 0.10340131 0.13883452 0.21131058 -0.01981019 0.1833468 -0.10751636 -0.03128868 0.02518242 0.23232952 0.042052146 0.11731903 -0.15506615 0.0063580726 -0.15429358 0.1511722 0.12745973 0.2576985 -0.25486213 -0.0709463 0.17983761 0.054027 -0.09884228 -0.24595179 -0.093028545 -0.028203879 0.094398156 0.09233813 0.029291354 0.13110267 0.15682974 -0.016919162 0.23927948 -0.1343307 -0.22422817 0.14634751 -0.064993896 0.4703685 -0.027190214 0.06224946 -0.091360025 0.21490277 -0.19562101 -0.10032754 -0.09056772 -0.06203493 -0.18876675 -0.10963594 -0.27734384 0.12616494 -0.02217992 -0.16058226 -0.080475815 0.026953284 0.110732645 0.014894041 0.09416802 0.14299914 -0.1594008 -0.066080004 -0.007995227 -0.11668856 -0.13081996 -0.09237365 0.14741232 0.09180138 0.081735 0.3211204 -0.0036552632 -0.047030564 -0.02311798 0.048961394 0.08669574 -0.06766279 -0.50028914 -0.048515294 0.14144728 -0.032994404 -0.11954345 -0.14929578 -0.2388355 -0.019883996 -0.15917352 -0.052084364 0.2801028 -0.0029121689 -0.054581646 -0.47385484 0.17112483 -0.12066923 -0.042173345 0.1395337 0.26115036 0.012869649 0.009291686 -0.0026459037 -0.075331464 0.017840583 -0.26869613 -0.21820338 -0.17084768 -0.1022808 -0.055290595 0.13513643 0.12362477 -0.10980586 0.13980341 -0.20233242 0.08813751 0.3849736 -0.10653763 -0.06199595 0.028849555 0.03230154 0.023856193 0.069950655 0.19310954 -0.077677034 -0.144811