Computing numeric representations of words in a high-dimensional space

• US 10,241,997 B1
• Filed: 08/21/2017
• Issued: 03/26/2019
• Est. Priority Date: 01/15/2013
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google LLC (Alphabet Inc.)
• Inventors: Mikolov, Tomas, Chen, Kai, Corrado, Gregory S., Dean, Jeffrey A.
• Primary Examiner(s): Abebe, Daniel
• Application Number:
  US15/682,374
• Time in Patent Office: 582 Days
• Field of Search: 704243
• CPC Class Codes: G06F 40/279 Recognition of textual entities
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G10L 15/06 Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice G10L15/14 takes precedence

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for computing numeric representations of words. One of the methods includes obtaining a set of training data, wherein the set of training data comprises sequences of words; training a classifier and an embedding function on the set of training data, wherein training the embedding function comprises obtained trained values of the embedding function parameters; processing each word in the vocabulary using the embedding function in accordance with the trained values of the embedding function parameters to generate a respective numerical representation of each word in the vocabulary in the high-dimensional space; and associating each word in the vocabulary with the respective numeric representation of the word in the high-dimensional space.

Claims

1. A method performed by one or more computers, the method comprising:

receiving a request for a numeric representation for a particular word in a high-dimensional space;

determining, by accessing a data set that is stored on one or more non-transitory computer readable media and that associates each word in a vocabulary of words with a respective numeric representation of the word in the high-dimensional space, the numeric representation of the particular word; and

providing the numeric representation of the particular word in response to the request, wherein the data set has been generated by;

obtaining a set of training data, wherein the set of training data comprises sequences of words;

training a plurality of classifiers and an embedding function on the set of training data, wherein the embedding function receives an input word and maps the input word to a numeric representation in the high-dimensional space in accordance with a set of embedding function parameters, wherein each of the classifiers corresponds to a respective position surrounding the input word in a sequence of words, and wherein each of the classifiers processes the numeric representation of the input word to generate a respective word score for each word in a pre-determined set of words, wherein each of the respective word scores represents a predicted likelihood that the corresponding word will be found in the corresponding position relative to the input word, and wherein training the embedding function comprises determining trained values of the embedding function parameters;

processing each word in the vocabulary using the embedding function in accordance with the trained values of the embedding function parameters to generate a respective numeric representation of each word in the vocabulary;

associating each word in the vocabulary with the respective numeric representation of the word in the high-dimensional space; and

storing the data set on the one or more non-transitory computer storage media.

2. The method of claim 1, wherein the numeric representations are continuous representations represented using floating-point numbers.

3. The method of claim 1, wherein positions of numeric representations in the high-dimensional space reflect semantic similarities between words represented by the numeric representations.

4. The method of claim 1, wherein positions of numeric representations in the high-dimensional space reflect syntactic similarities between words represented by the numeric representations.

5. The method of claim 1, wherein the embedding function maps the input word to a floating point vector.

6. The method of claim 1, wherein training the plurality of classifiers and the embedding function comprises preforming a backpropagation training technique to determine the trained values of the embedding function parameters.

7. The method of claim 1, wherein a dimensionality of the high-dimensional space is on the order of one thousand.

8. The method of claim 1, wherein each of the plurality of classifiers comprises a respective set of classifier parameters, and wherein training the plurality of classifiers and the embedding function on the set of training data comprises determining trained values of each of the sets of classifier parameters.

9. A method performed by one or more computers, the method comprising:

receiving a request for a numeric representation for a particular word in a high-dimensional space;

determining, by accessing a data set that is stored on one or more non-transitory computer readable media and that associates each word in a vocabulary of words with a respective numeric representation of the word in the high-dimensional space, the numeric representation of the particular word; and

providing the numeric representation of the particular word in response to the request, wherein the data set has been generated by;

obtaining a set of training data, wherein the set of training data comprises sequences of words;

training a classifier and an embedding function on the set of training data, wherein the embedding function receives a plurality of words surrounding an unknown word in a sequence of words and maps the plurality of words into a numeric representation in accordance with a set of embedding function parameters, wherein the classifier processes the numeric representation of the sequence of words to generate a respective word score for each word in a pre-determined set of words, wherein each of the respective word scores measure a predicted likelihood that the corresponding word is the unknown word, and wherein training the embedding function comprises determining trained values of the embedding function parameters;

processing each word in the vocabulary using the embedding function in accordance with the trained values of the embedding function parameters to generate a respective numeric representation of each word in the vocabulary in the high-dimensional space;

associating each word in the vocabulary with the respective numeric representation of the word in the high-dimensional space; and

storing the data set on the one or more non-transitory computer storage media.

10. The method of claim 9, wherein the numeric representations are continuous representations represented using floating-point numbers.

11. The method of claim 9, wherein positions of numeric representations in the high-dimensional space reflect semantic similarities between words represented by the numeric representations.

12. The method of claim 9, wherein positions of numeric representations in the high-dimensional space reflect syntactic similarities between words represented by the numeric representations.

13. The method of claim 9, wherein the embedding function maps each of the plurality of words to a respective floating point vector and outputs a single merged vector that is a combination of the respective floating point vectors.

14. The method of claim 9, wherein training the classifier and the embedding function comprises preforming a backpropagation training technique to determine the trained values of the embedding function parameters.

15. The method of claim 9, wherein a dimensionality of the high-dimensional space is on the order of one thousand.

16. The method of claim 9, wherein the classifier has a set of classifier parameters, and wherein training the plurality of classifiers and the embedding function on the set of training data comprises determining trained values of the set of classifier parameters.

17. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising:

receiving a request for a numeric representation for a particular word in a high-dimensional space;

determining, by accessing a data set that is stored on one or more non-transitory computer readable media and that associates each word in a vocabulary of words with a respective numeric representation of the word in the high-dimensional space, the numeric representation of the particular word; and

providing the numeric representation of the particular word in response to the request, wherein the data set has been generated by;

obtaining a set of training data, wherein the set of training data comprises sequences of words;

training a plurality of classifiers and an embedding function on the set of training data, wherein the embedding function receives an input word and maps the input word to a numeric representation in the high-dimensional space in accordance with a set of embedding function parameters, wherein each of the classifiers corresponds to a respective position surrounding the input word in a sequence of words, and wherein each of the classifiers processes the numeric representation of the input word to generate a respective word score for each word in a pre-determined set of words, wherein each of the respective word scores represents a predicted likelihood that the corresponding word will be found in the corresponding position relative to the input word, and wherein training the embedding function comprises determining trained values of the embedding function parameters;

processing each word in the vocabulary using the embedding function in accordance with the trained values of the embedding function parameters to generate a respective numeric representation of each word in the vocabulary;

associating each word in the vocabulary with the respective numeric representation of the word in the high-dimensional space; and

storing the data set on the one or more non-transitory computer storage media.

18. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising:

receiving a request for a numeric representation for a particular word in a high-dimensional space;

determining, by accessing a data set that is stored on one or more non-transitory computer readable media and that associates each word in a vocabulary of words with a respective numeric representation of the word in the high-dimensional space, the numeric representation of the particular word; and

providing the numeric representation of the particular word in response to the request, wherein the data set has been generated by;

obtaining a set of training data, wherein the set of training data comprises sequences of words;

training a classifier and an embedding function on the set of training data, wherein the embedding function receives a plurality of words surrounding an unknown word in a sequence of words and maps the plurality of words into a numeric representation in accordance with a set of embedding function parameters, wherein the classifier processes the numeric representation of the sequence of words to generate a respective word score for each word in a pre-determined set of words, wherein each of the respective word scores measure a predicted likelihood that the corresponding word is the unknown word, and wherein training the embedding function comprises determining trained values of the embedding function parameters;

processing each word in the vocabulary using the embedding function in accordance with the trained values of the embedding function parameters to generate a respective numeric representation of each word in the vocabulary in the high-dimensional space;

associating each word in the vocabulary with the respective numeric representation of the word in the high-dimensional space; and

storing the data set on the one or more non-transitory computer storage media.