Actions

Difference between revisions of "A One Hot Vector"

From Algolit

 
(14 intermediate revisions by 2 users not shown)
Line 8: Line 8:
 
|}
 
|}
  
=one-hot-vectors=
+
A '''one-hot-vector''' is a word-representation technique that uses ''distributional similarity'' to find patterns in the phrases that are used in company of a word. One-hot-vectors are basically a big matrix of 0's, with as many rows and columns as there are unique words. A text with 500 unique words, will be represented by a matrix of 500x500. With this matrix as its central tool, a script will go through the sentences of a dataset and count how often a word appears next to another word.
  
"''Meaning is this illusive thing that were trying to capture''" (Richard Socher in [https://www.youtube.com/watch?v=xhHOL3TNyJs&index=2&list=PLcGUo322oqu9n4i0X3cRJgKyVy7OkDdoi CS224D Lecture 2 - 31st Mar 2016 (Youtube)])
+
==Recipe for a one hot vector==
 +
If this is our example sentence ...
  
[0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0]
+
"The algoliterary explorers discovered a multidimensional landscape made of words disguised as numbers."
  
with one 0 for each word in a vocabulary
+
... these are the 14 words we work with ...
where the 1 is representing the place of a word in the vector
 
> In this kind of vector representation: none of the words are similar, they are all a 1.
 
  
=Note that=
+
a
"''Words are represented once in a vector. So words with multiple meanings, like "bank", are more difficult to represent. There is research to multivectors for one word, so that it does not end up in the middle.''" (Richard Socher, idem.)]
+
algoliterary
 +
as
 +
discovered
 +
disguised
 +
explores
 +
landscape
 +
made
 +
multidimensional
 +
numbers
 +
of
 +
the
 +
words
 +
.
  
For more notes on this lecture visit http://pad.constantvzw.org/public_pad/neural_networks_3
+
... a single vector in a one-hot-vector looks like this ...
 +
 
 +
[0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 +
 
 +
... and a full fourteen-dimensional matrix like this ...
 +
 
 +
[[0 0 0 0 0 0 0 0 0 0 0 0 0 0]  a
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  algoliterary
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  as
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  discovered
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  disguised
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  explores
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  landscape
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  made
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  multidimensional
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  numbers
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  of
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  the
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  words
 +
  [0 0 0 0 0 0 0 0 0 0 0 0 0 0]] .
 +
 
 +
... with one 0 for each unique word in a vocabulary, and a row for each unique word.
 +
 
 +
The following step is to count how often a word appears next to another ...
 +
 
 +
"The algoliterary explorers discovered a multidimensional landscape made of words disguised as numbers."
 +
 
 +
[[0 0 0 1 0 0 0 0 1 0 0 0 0 0]  a
 +
  [0 0 0 0 0 1 0 0 0 0 0 1 0 0]  algoliterary
 +
  [0 0 0 0 1 0 0 0 0 1 0 0 0 0]  as
 +
  [1 0 0 0 0 1 0 0 0 0 0 0 0 0]  discovered
 +
  [0 0 1 0 0 0 0 0 0 0 0 0 1 0]  disguised
 +
  [0 1 0 1 0 0 0 0 0 0 0 0 0 0]  explores
 +
  [0 0 0 0 0 0 0 1 1 0 0 0 0 0]  landscape
 +
  [0 0 0 0 0 0 1 0 0 0 1 0 0 0]  made
 +
  [1 0 0 0 0 0 1 0 0 0 0 0 0 0]  multidimensional
 +
  [0 0 1 0 0 0 0 0 0 0 0 0 0 1]  numbers
 +
  [0 0 0 0 0 0 0 1 0 0 0 0 1 0]  of
 +
  [0 1 0 0 0 0 0 0 0 0 0 0 0 0]  the
 +
  [0 0 0 0 1 0 0 0 0 0 1 0 0 0]  words
 +
  [0 0 0 0 0 0 0 0 0 1 0 0 0 0]] .
 +
 
 +
==Algolit's one-hot-vector scripts==
 +
Two one-hot-vector scripts were created during one of the Algolit sessions, both creating the same matrix but in a different way. To download and run them, use the following links: [https://gitlab.constantvzw.org/algolit/algolit/blob/master/algoliterary_encounter/one-hot-vector/one-hot-vector_gijs.py one-hot-vector_gijs.py] & [https://gitlab.constantvzw.org/algolit/algolit/blob/master/algoliterary_encounter/one-hot-vector/one-hot-vector_hans.py one-hot-vector_hans.py]
  
 +
==Note that==
 +
"''Words are represented once in a vector. So words with multiple meanings, like "bank", are more difficult to represent. There is research to multivectors for one word, so that it does not end up in the middle.''" (Richard Socher, [https://cs224d.stanford.edu/lectures/CS224d-Lecture2.pdf#6 CS224d, Deep Learning for Natural Language Processing at Stanford University (2016)].)
  
 +
For more notes on this lecture visit http://pad.constantvzw.org/public_pad/neural_networks_3
  
 
[[Category:Algoliterary-Encounters]]
 
[[Category:Algoliterary-Encounters]]

Latest revision as of 22:05, 31 October 2017

Type: Algoliterary exploration
Technique: word-embeddings
Developed by: Algolit

A one-hot-vector is a word-representation technique that uses distributional similarity to find patterns in the phrases that are used in company of a word. One-hot-vectors are basically a big matrix of 0's, with as many rows and columns as there are unique words. A text with 500 unique words, will be represented by a matrix of 500x500. With this matrix as its central tool, a script will go through the sentences of a dataset and count how often a word appears next to another word.

Recipe for a one hot vector

If this is our example sentence ...

"The algoliterary explorers discovered a multidimensional landscape made of words disguised as numbers."

... these are the 14 words we work with ...

a
algoliterary
as
discovered
disguised
explores
landscape
made
multidimensional
numbers
of
the
words
.

... a single vector in a one-hot-vector looks like this ...

[0 0 0 0 0 0 0 0 0 0 0 0 0 0] 

... and a full fourteen-dimensional matrix like this ...

[[0 0 0 0 0 0 0 0 0 0 0 0 0 0]  a
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  algoliterary
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  as
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  discovered
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  disguised
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  explores
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  landscape
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  made
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  multidimensional
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  numbers
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  of
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  the
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]  words
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0]] .

... with one 0 for each unique word in a vocabulary, and a row for each unique word.

The following step is to count how often a word appears next to another ...

"The algoliterary explorers discovered a multidimensional landscape made of words disguised as numbers."
[[0 0 0 1 0 0 0 0 1 0 0 0 0 0]  a
 [0 0 0 0 0 1 0 0 0 0 0 1 0 0]  algoliterary
 [0 0 0 0 1 0 0 0 0 1 0 0 0 0]  as
 [1 0 0 0 0 1 0 0 0 0 0 0 0 0]  discovered
 [0 0 1 0 0 0 0 0 0 0 0 0 1 0]  disguised
 [0 1 0 1 0 0 0 0 0 0 0 0 0 0]  explores
 [0 0 0 0 0 0 0 1 1 0 0 0 0 0]  landscape
 [0 0 0 0 0 0 1 0 0 0 1 0 0 0]  made
 [1 0 0 0 0 0 1 0 0 0 0 0 0 0]  multidimensional
 [0 0 1 0 0 0 0 0 0 0 0 0 0 1]  numbers
 [0 0 0 0 0 0 0 1 0 0 0 0 1 0]  of
 [0 1 0 0 0 0 0 0 0 0 0 0 0 0]  the
 [0 0 0 0 1 0 0 0 0 0 1 0 0 0]  words
 [0 0 0 0 0 0 0 0 0 1 0 0 0 0]] .

Algolit's one-hot-vector scripts

Two one-hot-vector scripts were created during one of the Algolit sessions, both creating the same matrix but in a different way. To download and run them, use the following links: one-hot-vector_gijs.py & one-hot-vector_hans.py

Note that

"Words are represented once in a vector. So words with multiple meanings, like "bank", are more difficult to represent. There is research to multivectors for one word, so that it does not end up in the middle." (Richard Socher, CS224d, Deep Learning for Natural Language Processing at Stanford University (2016).)

For more notes on this lecture visit http://pad.constantvzw.org/public_pad/neural_networks_3