Actions

Word2vec basic.py: Difference between revisions

From Algolit
Line 41: Line 41:
 
* Step 6: Visualize the embeddings.
 
* Step 6: Visualize the embeddings.
 
** '''Algolit adaption''': select 3 words to be included in the graph
 
** '''Algolit adaption''': select 3 words to be included in the graph
−
<br>
 
  
 
===Source===
 
===Source===

Revision as of 22:12, 31 October 2017

Type: Algolit extension
Datasets: nearbySaussre
Technique: word embeddings
Developed by: a team of researchers led by Tomas Mikolov at Google, Claude Lévi-Strauss, Algolit
Graph generated by the word2vec_basic.py Tensorflow tutorial, based on the nearbySaussre dataset.

This is an annotated version of the basic word2vec script. The code is based on this Word2Vec tutorial provided by Tensorflow.

History

Word2vec consists of related models used to generate vectors from words (also called word embeddings). It is a two-layer neural network, produced by a team of researchers led by Tomas Mikolov at Google. The script that we use here is not the original version of word2vec. The original project is written in the programming language C, which made us look for a version of the script written in the programming language Python. Another python implementation of word2vec is provided by Gensim.

word2vec_basic_algolit.py

The structure of the annotated word2vec script is the following:

  • Step 1: Download data.
  • Algolit step 1: read data from plain text file
    • Algolit inspection: wordlist.txt
  • Step 2: Create a dictionary and replace rare words with UNK token.
    • Algolit inspection: counted.txt
    • Algolit inspection: dictionary.txt
    • Algolit inspection: data.txt
    • Algolit inspection: disregarded.txt
    • Algolit adaption: reversed-input.txt
  • Step 3: Function to generate a training batch for the skip-gram model
  • Step 4: Build and train a skip-gram model.
    • Algolit inspection: big-random-matrix.txt
    • Algolit adaption: select your own set of test words
  • Step 5: Begin training.
    • Algolit inspection: training-words.txt
    • Algolit inspection: training-window-words.txt
    • Algolit adaption: visualisation of the cosine similarity calculation updates
    • Algolit inspection: logfile.txt
  • Step 6: Visualize the embeddings.
    • Algolit adaption: select 3 words to be included in the graph

Source

The script word2vec_basic.py provides an option to download a dataset from Matt Mahoney's home page. It turns out to be a plain text document, without any punctuation or line breaks.

For the tests that we wanted to do with the script, we decided to work with an algoliterary dataset that circles around the structuralist linguistic theory of Ferdinand de Saussure: nearbySaussure.

Before we could use the three books that form this dataset as training material, we needed to remove all the punctuation from the file. To do this, we wrote a small python script text-punctuation-clean-up.py. The script saves a *stripped* version of the original book under another filename.

The book contains 424.811 words in total of which 24.651 words are unique.

wordlist.txt

From continuous text to list of words, exported as wordlist.txt. 

['xt', '1250', 'By', 'Claude', 'levistrauss', 'Translated', 'by', 'john', 'r', 'ussell', 'Illustrated', 'with', '48', 'pages', 'of', 'photographs', 'and', '48', 'line', 'drawings', 'Have', 'sought', 'a', 'human', 'society', 'reduced', 'To', 'its', 'most', 'basic', 'expression', 'His', 'search', 'has', 'taken', 'claude', 'levi', 'Strauss', 'eminent', 'french', 'anthropologist', 'And', 'one', 'of', 'the', 'founders', 'of', 'structural', 'Anthropology', 'to', 'the', 'far', 'corners', 'of', 'the', 'Earth', 'not', 'as', 'a', 'superficial', 'sightseer', 'but', 'As', 'a', 'close', 'student', 'of', 'man', 'and', 'the', 'varied', 'Cultures', 'he', 'has', 'erected', 'around', 'himself', 'While', 'a', 'professor', 'at', 'sao', 'paolo', 'univer', 'Sity', 'in', 'brazil' ... ]

counted.txt

From list of words to a list with the structure [(word, value)], exported as counted.txt. 

[['UNK', 18767], ('the', 10108), ('of', 5790), ('and', 4229), ('to', 3895), ('a', 3407), ('in', 3092), ('that', 1633), ('was', 1380), ('it', 1367), ('as', 1271), ('with', 1206), ('for', 1196), ('which', 1158), ('had', 1129), ('is', 1119), ('on', 1015), ('i', 1014), ('or', 945), ('they', 905), ('their', 886), ('by', 876), ('were', 868), ('one', 800), ('at', 794), ('from', 764), ('The', 762), ('be', 731), ('we', 726), ('he', 678), ('not', 668), ('his', 646), ('an', 596), ('this', 584), ('but', 576), ('have', 558), ('are', 555), ('all', 547), ('them', 509), ('its', 454), ('our', 452), ('would', 449), ('s', 445), ('so', 440), ('been', 396), ('my', 394), ('these', 386), ('who', 375), ('there', 361), ('And', 348), ('two', 346), ('no', 341), ('into', 336), ('up', 336), ('more', 335), ('when', 335), ('Of', 324), ('has', 296), ('if', 291), ('other', 289), ('out', 287), ('me', 282), ('only', 274), ('us', 272), ('could', 262), ('some', 250), ('To', 243), ('time', 232), ('can', 232), ('In', 229), ('made', 223), ('die', 222), ('what', 222), ('those', 221), ('than', 214), ('men', 209), ('where', 208), ('will', 202), ('first', 201), ('him', 198), ('A', 192), ('between', 191), ('each', 189), ('any', 185), ('own', 183), ('another', 182), ('way', 178) ... ]

dictionary.txt

Reversed dictionary, a list of the 5000 (=vocabulary size) most common words, accompanied by an index number, exported as dictionary.txt. 

{0: 'UNK', 1: 'the', 2: 'of', 3: 'and', 4: 'to', 5: 'a', 6: 'in', 7: 'that', 8: 'was', 9: 'it', 10: 'as', 11: 'with', 12: 'for', 13: 'which', 14: 'had', 15: 'is', 16: 'on', 17: 'i', 18: 'or', 19: 'they', 20: 'their', 21: 'by', 22: 'were', 23: 'one', 24: 'at', 25: 'from', 26: 'The', 27: 'be', 28: 'we', 29: 'he', 30: 'not', 31: 'his', 32: 'an', 33: 'this', 34: 'but', 35: 'have', 36: 'are', 37: 'all', 38: 'them', 39: 'its', 40: 'our', 41: 'would', 42: 's', 43: 'so', 44: 'been', 45: 'my', 46: 'these', 47: 'who', 48: 'there', 49: 'And', 50: 'two', 51: 'no', 52: 'into', 53: 'up', 54: 'more', 55: 'when', 56: 'Of', 57: 'has', 58: 'if', 59: 'other', 60: 'out', 61: 'me', 62: 'only', 63: 'us', 64: 'could', 65: 'some', 66: 'To', 67: 'time', 68: 'can', 69: 'In', 70: 'made', 71: 'die', 72: 'what', 73: 'those', 74: 'than', 75: 'men', 76: 'where', 77: 'will', 78: 'first', 79: 'him', 80: 'A', 81: 'between', 82: 'each', 83: 'any', 84: 'own', 85: 'another', 86: 'way' ... }

data.txt

The object data is created, the original texts where words are replaced with index numbers, exported as data.txt. 

[0, 0, 223, 0, 2465, 0, 21, 0, 1951, 0, 0, 11, 2574, 3339, 2, 3858, 3, 2574, 232, 1882, 427, 1493, 5, 189, 115, 1404, 66, 39, 116, 2493, 2328, 477, 1090, 57, 269, 0, 0, 0, 0, 382, 487, 49, 23, 2, 1, 0, 2, 0, 3917, 4, 1, 149, 1715, 2, 1, 0, 30, 10, 5, 4136, 0, 34, 192, 5, 1487, 1303, 2, 104, 3, 1, 2203, 0, 29, 57, 3905, 418, 144, 872, 5, 3282, 24, 248, 4672, 0, 0, 6, 227, 686, 2465, 1457, 0, 172, 1, 741, 1000, 49, 1, 4837, 0, 0, 2, 227, 66, 1, 0, 2639, 2, 31, 4563, 180, 8, 295, 105, 1, 116, 433, 56, 1, 0, 480, 7, 29, 131, 26, 2493, 0, 408, 29, 8, 0, 2480, 2639, 15, 1, 818, 2, 31, 2098, 105, 46, 480, 295, 589, 0, 0, 0, 2, 1, 3697, 3, 1, 2001, 516, 0, 429, 13, 19, 2578, 20, 2621, 1019, 1, 0, 0, 0, 115, 2, 1, 185, 1, 953, 47, 0, 5, 267, 2, 1468, 223, 1171, 504, 4, 20, 179, 1, 4349, 3, 0, 705, 3903, 147, 0, 2748, 2192, 1516, 190, 12, 166, 0, 16, 106, 0, 0, 2262, 2262, 0, 2480, 2639, 0, 0, 0, 2053, 0, 42, 2480, 2639, 0, 4004, 0, 339, 888, 3225, 0, 77, 27, 0, 62, 246, 0, 2, 3225, 2885, 0, 0, 373, 0, 3, 0, 2, 2173, 0, 0, 0, 36, 1036, 12, 310, 1214, 0, 0, 0, 297, 59, 3225, 3705, 0, 60, 16, 20, 0, 184, 0, 375, 2213, 1236, 3, 50, 627, 0, 2, 1, 196, 0, 1, 0, 36, 1412, 1737, 214, 0, 0, 3, 0, 4, 1, 185, 0, 6, 1, 1108, 19, 154, 36, 23, 56, 1, 2736, 480, 2, 481, 227 ... ]

disregarded.txt

List of disregarded words, that fall outside the vocabulary size, exported as disregarded.txt. 

['xt', '1250', 'Claude', 'Translated', 'john', 'ussell', 'Illustrated', 'claude', 'levi', 'Strauss', 'eminent', 'founders', 'structural', 'Earth', 'sightseer', 'Cultures', 'univer', 'Sity', 'Extensively', 'upland', 'jungles', 'tristes', 'amerindian', 'humain', 'seeking', 'intricate', 'detailed', 'accounts', 'Designs', 'rigid', 'hier', 'Archical', 'win', 'superstitionridden', 'weird', 'Continued', 'flap', 'Iv', 'cv', '981', 'l56t', 'Le', 'straus', '61157', 'Kansas', 'Books', 'issued', 'presentation', 'Please', 'report', 'cards', 'Change', 'promptly', 'Card', 'holders', 'records', 'films', 'pict', 'Checked', 'cards', 'Frontispiece', 'Carajiindians', 'araguaia', 'Caraji', 'geo', 'Graphically', 'culturally', 'Described', 'Date', 'duk', 'Auf2s', '67', 'Wl', 'Translated', 'John', 'russell', 'Criterion', 'hutchinson', 'publishers', 'ltd', 'london', '1961', 'Library', 'congress', 'catalog', '617203', 'Originally', 'tropiaues', 'librairie', 'plon', '1955', 'chapters', 'Xiv', 'xv', 'xvi', 'xxxix', 'Edition', 'omitted', 'Printed', 'britain', '15758', 'laurent', 'Minus', 'ergo', 'ante', 'haec', 'quam', 'tu', 'ceddere', 'cadentque', 'Lucretius', 'rerum', 'natura', '969', '15758', 'Contents', '65', 'iii', '133', '151', '160', '183', '198', 'vii', '286', 'crusoe', '323', '342', 'japim', '363', 'ix', '381', 'Bibliography', '399', '401', 'Illustrations', 'Frontispiece', 'carajaindians', '97', 'thepantanal', 'belle', 'regalia', 'preparations', 'mariddo', 'cigarette', 'Tucked', 'bracelet', 'wakletou', 'cf', 'plate', 'piercing', 'grading', 'threading', 'suckling', 'conjugal', 'felicity', 'affectionate', 'frolics', 'dozing', 'spinner', 'Plug', 'daydreamer', '46', 'smile', '47', 'amidst', 'mund6', 'dome', 'archer', 'medi', 'Terranean', 'cf', 'Plate', 'mothers', 'eyebrows', 'coated', 'Wax', '55', 'lucinda', '57', 'skinning' ... ]

reversed-input.txt

Reversed version of the initial dataset, where all the disregard words are replaced with UNK (unkown), exported as reversed-input.txt. 

UNK UNK By UNK levistrauss UNK by UNK r UNK UNK with 48 pages of photographs and 48 line drawings Have sought a human society reduced To its most basic expression His search has taken UNK UNK UNK UNK french anthropologist And one of the UNK of UNK Anthropology to the far corners of the UNK not as a superficial UNK but As a close student of man and the varied UNK he has erected around himself While a professor at sao paolo UNK UNK in brazil m levistrauss travelled UNK through the amazon basin And the dense UNK UNK of brazil To the UNK tropiques of his title It was here among the most primitive Of the UNK tribes that he found The basic UNK societies he was UNK Tristes tropiques is the story of his Experience among these tribes here Are UNK UNK UNK of the Caduveo and the elaborate painted UNK behind which they hide their Natural faces the UNK UNK UNK society of the bororo the Nambikwara who UNK a sort of security By giving wives to their chief the Disease and UNK tupi Kawahib whose UNK tribal dances Sometimes last for days UNK on back UNK UNK v v UNK Tristes tropiques UNK UNK UNK vi UNK s Tristes tropiques UNK L UNK city public library UNK will be UNK only On UNK of library card UNK UNK lost UNK and UNK of residence UNK UNK UNK are responsible for All books UNK UNK UNK Or other library materials UNK out on their UNK I UNK Two masked dancers and two girls UNK of the rio UNK the UNK are closely related both UNK UNK and UNK to the bororo UNK in the book they too are one Of the wandering tribes of central brazil ...

big-random-matrix.txt

A big random matrix is created, with a vector size of 5000x20, exported as big-random-matrix.txt. 

 [[  2.85661697e-01   9.69764948e-01  -7.59074926e-01  -6.15304947e-01
    6.77072048e-01  -3.78361940e-01  -6.71523094e-01   3.94770384e-01
    7.04541206e-02  -8.92262936e-01   5.87280035e-01   4.58304882e-02
    2.53162384e-01   1.90168381e-01  -6.61255836e-01  -3.75634432e-01
   -5.55147886e-01   4.49278116e-01   3.26536417e-01   8.64576340e-01]
 [ -6.70668364e-01  -5.53100824e-01  -3.71278524e-01   1.25042677e-01
   -1.46459818e-01  -6.10010624e-01   9.19621468e-01  -1.55832767e-01
   -7.70623922e-01  -1.44968033e-01  -6.36267662e-01  -1.87215090e-01
    7.09211111e-01  -6.57156706e-01   3.26824188e-02  -4.25864220e-01
   -5.86277485e-01   8.16827059e-01  -5.57327747e-01  -3.35038900e-01]
 [ -9.33161497e-01   8.45068693e-01  -8.14761639e-01  -5.67158937e-01
    5.23060560e-01   4.90430593e-01  -9.11595106e-01   4.36383963e-01
   -9.69607353e-01  -6.64181471e-01  -4.44166183e-01   7.78196335e-01
   -5.34924030e-01   6.49461985e-01   5.69838047e-01   2.50927448e-01
   -8.87476921e-01  -3.74064207e-01   4.24978733e-02   1.25571489e-01]
 [  9.89913464e-01   3.36525917e-01  -1.86083794e-01  -5.25027514e-01
   -8.87480021e-01   8.53247643e-02   4.10822868e-01   3.29172134e-01
    8.56166363e-01   5.12266636e-01   7.75470734e-01   7.89757490e-01
   -9.44452286e-02  -8.79762173e-01   1.57778263e-02  -8.59814644e-01
    4.55990076e-01   4.06166315e-01  -8.40348721e-01  -2.75753498e-01]
 [  5.79052448e-01  -3.62973213e-01  -8.79675150e-01  -9.98473167e-01
   -1.73240185e-01   7.07520723e-01   4.95352268e-01   4.99097586e-01
   -5.02996445e-02  -4.01979208e-01   5.94721079e-01   7.37986326e-01
   -6.61164761e-01   6.45744085e-01  -4.68054295e-01  -5.54257870e-01
    5.12778997e-01   7.89849758e-01   2.42011547e-02  -2.77193785e-01] ... ]

training-words.txt

Export a training batch of 64 words, with a vector size of 128x20, exported as training-words.txt. 

 [2831 2831 1906 1906   25   25    1    1  221  221   37   37    1    1 1840
 1840  655  655    3    3   22   22  971  971    4    4    1    1  481  481
 4235 4235  297  297    0    0    7    7 1343 1343   16   16   53   53  172
  172    1    1 1080 1080 1831 1831    0    0    2    2    0    0 1804 1804
    1    1  590  590  653  653    3    3   16   16  489  489    2    2    7
    7    8    8    5    5    0    0   56   56 1313 1313   13   13   14   14
   44   44 3432 3432    6    6    1    1   98   98  744  744   23   23   16
   16  489  489   56   56   85   85    4    4  224  224    5    5    0    0
 1080 1080    1    1    0    0  474  474]

<br>Or in words: 

<pre>
['thirteen', 'thirteen', 'Feet', 'Feet', 'from', 'from', 'the', 'the', 'ground', 'ground', 'all', 'all', 'the', 'the', 'poles', 'poles', 'met', 'met', 'and', 'and', 'were', 'were', 'tied', 'tied', 'to', 'to', 'the', 'the', 'central', 'central', 'pole', 'pole', 'Or', 'Or', 'UNK', 'UNK', 'that', 'that', 'pushed', 'pushed', 'on', 'on', 'up', 'up', 'through', 'through', 'the', 'the', 'roof', 'roof', 'horizontal', 'horizontal', 'UNK', 'UNK', 'of', 'of', 'UNK', 'UNK', 'completed', 'completed', 'the', 'the', 'main', 'main', 'structure', 'structure', 'and', 'and', 'on', 'on', 'top', 'top', 'of', 'of', 'that', 'that', 'was', 'was', 'a', 'a', 'UNK', 'UNK', 'Of', 'Of', 'palmleaves', 'palmleaves', 'which', 'which', 'had', 'had', 'been', 'been', 'folded', 'folded', 'in', 'in', 'the', 'the', 'same', 'same', 'direction', 'direction', 'one', 'one', 'on', 'on', 'top', 'top', 'Of', 'Of', 'another', 'another', 'to', 'to', 'form', 'form', 'a', 'a', 'UNK', 'UNK', 'roof', 'roof', 'the', 'the', 'UNK', 'UNK', 'hut', 'hut']

training-window-words.txt

Export a the 128 connected window words, one to the left, one to the right, with a vector size of 128x20, exported as training-window-words.txt. 

 [[1906] [18] [25] [2831] [1] [1906] [221] [25] [1] [37] [1] [221] [1840] [37] [655] [1] [1840] [3] [655] [22] [3] [971] [22] [4] [971] [1] [4] [481] [1] [4235] [297] [481] [0] [4235] [7] [297] [1343] [0] [16] [7] [1343] [53] [172] [16] [1] [53] [1080] [172] [1] [1831] [1080] [0] [2] [1831] [0] [0] [2] [1804] [0] [1] [590] [1804] [1] [653] [590] [3] [16] [653] [489] [3] [2] [16] [7] [489] [2] [8] [7] [5] [0] [8] [5] [56] [1313] [0] [13] [56] [1313] [14] [44] [13] [14] [3432] [6] [44] [3432] [1] [98] [6] [744] [1] [98] [23] [16] [744] [489] [23] [56] [16] [489] [85] [4] [56] [85] [224] [5] [4] [224] [0] [1080] [5] [0] [1] [1080] [0] [474] [1] [0] [8]]


Or in words: 

['Feet', 'or', 'from', 'thirteen', 'the', 'Feet', 'ground', 'from', 'the', 'all', 'the', 'ground', 'poles', 'all', 'met', 'the', 'poles', 'and', 'met', 'were', 'and', 'tied', 'were', 'to', 'tied', 'the', 'to', 'central', 'the', 'pole', 'Or', 'central', 'UNK', 'pole', 'that', 'Or', 'pushed', 'UNK', 'on', 'that', 'pushed', 'up', 'through', 'on', 'the', 'up', 'roof', 'through', 'the', 'horizontal', 'roof', 'UNK', 'of', 'horizontal', 'UNK', 'UNK', 'of', 'completed', 'UNK', 'the', 'main', 'completed', 'the', 'structure', 'main', 'and', 'on', 'structure', 'top', 'and', 'of', 'on', 'that', 'top', 'of', 'was', 'that', 'a', 'UNK', 'was', 'a', 'Of', 'palmleaves', 'UNK', 'which', 'Of', 'palmleaves', 'had', 'been', 'which', 'had', 'folded', 'in', 'been', 'folded', 'the', 'same', 'in', 'direction', 'the', 'same', 'one', 'on', 'direction', 'top', 'one', 'Of', 'on', 'top', 'another', 'to', 'Of', 'another', 'form', 'a', 'to', 'form', 'UNK', 'roof', 'a', 'UNK', 'the', 'roof', 'UNK', 'hut', 'the', 'UNK', 'was']

cosine similarity calculation updates

Visualisation of the cosine similarity calculation updates. 

...

logfile.txt

Save training log, exported as logfile.txt. 

Nearest to collective: Beyond, Although, luxury, confirmed, pointless, Born, colour, stick, scattered, somewhere,
Nearest to being: direcdy, appropriate, 8000, muito, disgusting, broad, southeast, Longer, completed, Before,
Nearest to social: photograph, Working, Hung, coasts, teacher, skins, cuts, extent, sheets, worth,

Nearest to collective: manioc, colour, work, grass, simply, adopted, it, particular, groups, concerned,
Nearest to being: jaguar, said, longer, sky, adopted, this, design, From, better, Longer,
Nearest to social: fall, make, photograph, yellow, given, than, took, men, worth, clouds,

Nearest to collective: manioc, colour, work, simply, grass, adopted, Beyond, horizons, particular, position,
Nearest to being: Longer, said, adopted, jaguar, longer, design, Before, sky, From, completed,
Nearest to social: photograph, fall, yellow, make, Hung, skins, given, worth, extent, teacher,

...

Nearest to collective: Beyond, Although, tubes, heightened, Born, line, horizons, tongue, occupied, unexpected,
Nearest to being: Difficulty, maintained, control, mass, Three, why, goiania, Behind, Children, negative,
Nearest to social: wooden, Tropical, leaf, finely, extent, considerations, northern, feeling, humanity, derisory,

Nearest to collective: Beyond, Although, tubes, heightened, Born, line, tongue, horizons, lower, unexpected,
Nearest to being: Difficulty, maintained, control, mass, Three, goiania, Behind, why, characteristics, Instead,
Nearest to social: wooden, Tropical, leaf, finely, extent, considerations, feeling, northern, humanity, derisory,

Nearest to collective: Beyond, Although, tubes, heightened, Born, line, tongue, lower, unexpected, horizons,
Nearest to being: Difficulty, maintained, mass, control, Three, goiania, Behind, why, characteristics, Instead,
Nearest to social: wooden, Tropical, leaf, finely, extent, considerations, northern, feeling, humanity, derisory,
Retrieved from "https://www.algolit.net/index.php?title=Word2vec_basic.py&oldid=10272"