Word2vec basic.py

Type:	Algolit extension
Datasets:	nearbySaussre
Technique:	word embeddings
Developed by:	a team of researchers led by Tomas Mikolov at Google, Claude Lévi-Strauss, Algolit

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

Each table is occupied with one of the multiple steps of the script word2vec_basic.py. Picture taken during the Algoliterary Encounter event in November 2017.

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

Graph generated by the word2vec_basic.py Tensorflow tutorial, based on the nearbySaussre dataset.

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

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.

wordlist.txt

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

[u'Introduction', u'saussure', u'today', u'Carol', u'sanders', u'Why', u'still', u'today', u'do', u'we', u'\ufb01nd', u'the', u'name', u'of', u'ferdinand', u'de', u'saussure', u'featuring', u'prominently', u'in', u'volumes', u'published', u'not', u'only', u'on', u'linguistics', u'but', u'on', u'a', u'multitude', u'of', u'topics',  ... ]

counted.txt

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

Counter({u'the': 22315, u'of': 16396, u'and': 8271, u'a': 8246, u'to': 7797, u'in': 7314, u'is': 5983, u'as': 4143, u'that': 3586, u'it': 2629, u'e': 2500, u'The': 2478, u's': 2332, u'language': 2281, u'saussure': 2201, u'which': 2101, u'by': 1962, u'this': 1944, u'on': 1937, u'be': 1808, u'or': 1751, u'r': 1713, u'not': 1689, u'an': 1680, ... })

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: u'the', 2: u'of', 3: u'and', 4: u'a', 5: u'to', 6: u'in', 7: u'is', 8: u'as', 9: u'that', 10: u'it', 11: u'e', 12: u'The', 13: u's', 14: u'language', 15: u'saussure', 16: u'which', 17: u'by', 18: u'this', 19: u'on', 20: u'be', 21: u'or', 22: u'r', 23: u'not', 24: u'an', ... }

data.txt

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

[1169, 15, 1289, 3020, 1427, 3697, 354, 1289, 269, 68, 1021, 1, 345, 2, 234, 34, 15, 4416, 0, 6, 3052, 293, 23, 64, 19, 31, 38, 19, 4, 0, 2, 3877, ... ]

disregarded.txt

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

[u'prominently', u'multitude', u'Volumes', u'titles', u'lee', u'poynton', u'intriguing', u'Plastic', u'glasses', u'fathers', u'kronenfeld', u'Afresh', u'Impact', u'titles', u'excite', u'premature', u'\u2018course', u'Sole', u'brilliant', u'precocious', u'centuries', u'examines', u'tracing', u'barely', u'praise', ... ]

reversed-input.txt

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

Introduction saussure today Carol sanders Why still today do we ﬁnd the name of ferdinand de saussure featuring UNK in volumes published not only on linguistics but on a UNK of topics UNK with UNK such as culture and text discourse and methodology in Social research and cultural studies UNK and UNK 2000 or the UNK UNK UNK and church UNK UNK 1996 ...

big-random-matrix.txt

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

 [[  7.91555882e-01   4.78600025e-01  -7.13676214e-01   2.30826855e-01
    6.61124229e-01   2.52689123e-01   6.37347698e-02   2.63915062e-01
    7.84061432e-01   6.69055700e-01   3.71650457e-01  -3.47790241e-01
   -4.34857845e-01  -9.00017262e-01   5.75044394e-01  -2.66819954e-01
    2.29521990e-01  -1.87541008e-01   7.47018099e-01  -8.54661465e-01]
 [  1.86723471e-01  -5.84969044e-01  -7.00650215e-01   7.50902653e-01
    2.52289057e-01  -9.68446016e-01  -1.12547159e-01  -9.01058912e-01
   -5.95885992e-01   3.08442831e-01   3.84899616e-01   7.09214926e-01
    9.58799362e-01  -8.78485441e-01  -3.27231169e-01   6.92137718e-01
    8.31190109e-01   1.67458773e-01   2.05923319e-01  -8.14627409e-01]
 [ -6.24799252e-01   9.01598454e-01   7.46447325e-01   5.45922041e-01
    4.28986549e-02  -2.75697231e-01   5.12938023e-01  -4.38443661e-01
    7.13398457e-01  -9.77021456e-01  -6.00349426e-01  -1.46302462e-01
   -9.75251198e-02  -1.80129766e-01   4.47291374e-01  -9.00330782e-01
    8.20701122e-02   9.37094688e-01  -8.20110321e-01  -7.58672953e-01] ... ]

training-words.txt

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

[ 323  323   52   52  107  107 2984 2984    3    3 1092 1092   48   48    4
    4    0    0 2898 2898   89   89   66   66   20   20   28   28    0    0
    4    4    0    0  142  142   28   28    0    0    0    0  173  173  697
  697 1054 1054  133  133    0    0    0    0   13   13 4364 4364 1146 1146
    2    2    1    1  201  201    2    2 1432 1432   26   26   12   12  201
  201    2    2  219  219    5    5  813  813  290  290    0    0 3071 3071
    5    5    1    1  280  280 2485 2485  705  705    6    6  144  144   28
   28    4    4 1125 1125    2    2  301  301    9    9    7    7 2851 2851
    6    6   16   16    0    0 3574 3574]

Or in words:

['One', 'One', 'can', 'can', 'then', 'then', 'enter', 'enter', 'and', 'and', 'remain', 'remain', 'In', 'In', 'a', 'a', 'UNK', 'UNK', 'synchronics', 'synchronics', 'This', 'This', 'would', 'would', 'be', 'be', 'for', 'for', 'UNK', 'UNK', 'a', 'a', 'UNK', 'UNK', 'distinction', 'distinction', 'for', 'for', 'UNK', 'UNK', 'UNK', 'UNK', 'historical', 'historical', 'questions', 'questions', 'somewhat', 'somewhat', 'like', 'like', 'UNK', 'UNK', 'UNK', 'UNK', 's', 's', 'separating', 'separating', 'off', 'off', 'of', 'of', 'the', 'the', 'book', 'book', 'of', 'of', 'god', 'god', 'from', 'from', 'The', 'The', 'book', 'book', 'of', 'of', 'nature', 'nature', 'to', 'to', 'give', 'give', 'himself', 'himself', 'UNK', 'UNK', 'Access', 'Access', 'to', 'to', 'the', 'the', 'latter', 'latter', 'eagleton', 'eagleton', 'argues', 'argues', 'in', 'in', 'fact', 'fact', 'for', 'for', 'a', 'a', 'Process', 'Process', 'of', 'of', 'reading', 'reading', 'that', 'that', 'is', 'is', 'dialectical', 'dialectical', 'in', 'in', 'which', 'which', 'UNK', 'UNK', 'undergo', 'undergo']

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.

[[0] [52] [107] [323] [2984] [52] [3] [107] [1092] [2984] [48] [3] [4] [1092] [48] [0] [2898] [4] [89] [0] [66] [2898] [20] [89] [66] [28] [20] [0] [28] [4] [0] [0] [4] [142] [28] [0] [142] [0] [28] [0] [173] [0] [0] [697] [1054] [173] [697] [133] [0] [1054] [133] [0] [0] [13] [4364] [0] [13] [1146] [4364] [2] [1146] [1] [201] [2] [1] [2] [1432] [201] [26] [2] [1432] [12] [26] [201] [12] [2] [219] [201] [5] [2] [813] [219] [290] [5] [0] [813] [290] [3071] [5] [0] [1] [3071] [5] [280] [2485] [1] [705] [280] [6] [2485] [144] [705] [28] [6] [4] [144] [1125] [28] [2] [4] [1125] [301] [9] [2] [7] [301] [9] [2851] [6] [7] [2851] [16] [0] [6] [3574] [16] [0] [4331]]