Social hash for language models
Components of language processing engines, such as translation models and language models, can be customized for groups of users or based on user type values. Users can be organized into groups or assigned a value on a continuum based on factors such as interests, biographical characteristics, social media interactions, etc. In some implementations, translation engine components can be customized for groups of users by selecting the training data from content created by users in that group. In some implementations, the group identifier or continuum value can be part of the input into a general translation component allowing the translation component to take a language style of that user group into account when performing language processing tasks.
1. A computer implemented method of using a user-signal cluster translation engine, comprising:
receiving a content item comprising a source and is associated with a source author;
obtaining a user-signal cluster translation engine comprising a neural network, the user-signal translation engine having been trained using a plurality of training items and a training signal user cluster;
generating a signal user cluster by entering two or more the multiple characteristics into a sparse matrix or vector;
obtaining, via the user-signal cluster translation engine, a translation result of the source associated with the content item, wherein the user-signal cluster translation engine is modified based on computing a cosine difference value between vector representations of the translation result and a translation of the source; and
returning the translation result.
2. The method of claim 1 , wherein the source comprises acoustic speech data, and wherein the author is a speaker.
3. The method of claim 1 , wherein the signal user cluster is further formulated by embedding the sparse matrix or vector into a value with a lower dimension than the sparse matrix or vector.
4. The method of claim 1 , wherein the obtaining of the translation result comprises:
converting the source into a language input conversion by:
combining vectors corresponding to words or phrases of the source; and
combining the vectors corresponding to words or phrases of the source into a single vector representation of the source; and
providing the language input conversion as input to a model of the user-signal cluster translation engine.
5. The method of claim 1 , wherein the obtaining of the translation result comprises:
providing, to each function corresponding to input layer nodes of the neural network, the signal user cluster and the source representation included in the content item; and
providing, to functions corresponding to nodes at one or more hidden layers of the neural network, the signal user cluster.
6. The method of claim 1 , wherein the multiple characteristics comprise two or more of gender, age, country, region, city, level of education, life stage, or any combination thereof.
7. The method of claim 1 , wherein applying the user-signal cluster translation engine to the source representation included in the content item and the signal user cluster comprises using a translation model to determine a proposed translation of the source from multiple possible translations of the source based on a probability distribution showing how likely the possible translations are to appear in an output language.
8. A non-transitory computer-readable storage medium storing instructions that, when executed by a computing system, cause the computing system to perform operations for using a user-group cluster translation engine, the operations comprising:
receiving a content item comprising a source and is associated with a source author;
obtaining a user-signal cluster translation engine comprising a neural network, the user-signal translation engine having been trained using a plurality of training items and a training signal user cluster;
generating a signal user cluster by entering two or more multiple characteristics into a sparse matrix or vector;
obtaining, via the user-signal cluster translation engine, a translation result of the source associated with the content item, wherein the user-signal cluster translation engine is modified based on computing a cosine difference value between vector representations of the translation result and a translation of the source; and
returning the translation result.
9. The non-transitory computer-readable storage medium of claim 8 , wherein the source comprises acoustic speech data, and wherein the source author is a speaker.
10. The non-transitory computer-readable storage medium of claim 8 , wherein computing a language input conversion for the content item comprises:
combining vectors corresponding to words or phrases of the source; and
combining the vectors corresponding to words or phrases of the source into a single vector representation of the source.
11. The non-transitory computer-readable storage medium of claim 8 , wherein the obtaining of the translation result comprises providing the language input conversion as input to a model of the user-group cluster translation engine.
12. The non-transitory computer-readable storage medium of claim 8 , wherein the two or more user characteristics comprise one or more of gender, age, country, region, city, level of education, life stage, or any combination thereof.
13. The non-transitory computer-readable storage medium of claim 8 , wherein the obtaining of the translation result comprises using a translation model to determine a proposed translation of the source from multiple possible translations of the source based on a probability distribution showing how likely the possible translations of words or phrases are to appear in an output language.
14. A system for applying a user-signal cluster translation engine, comprising:
memory;
one or more processors operable to execute stored instructions that, when executed, cause the one or more processors to:
receive a content item comprising a source and is associated with a source author;
obtain a user-signal cluster translation engine comprising a neural network, the user-signal translation engine having been trained using a plurality of training items and a training signal user cluster;
generate a signal user cluster by entering two or more multiple characteristics into a sparse matrix or vector;
obtain, via the user-signal cluster translation engine, a translation result of the source associated with the content item, wherein the user-signal cluster translation engine is modified based on computing a cosine difference value between vector representations of the translation result and a translation of the source; and
return the translation result.
15. The system of claim 14 , wherein the one or more processors is further caused to compute the signal user cluster by embedding the sparse matrix or vector into a value with a lower dimension than the sparse matrix or vector.
16. The system of claim 14 , wherein the one or more processors is further caused to obtain the translation result by providing the language input conversion and the signal user cluster to each function corresponding to input layer nodes of the neural network.
17. The system of claim 14 , wherein the one or more processors is further caused to convert the language into a language input conversion by combining the vectors corresponding to words or phrases of the language into a single vector representation of the language.
18. The system of claim 14 , wherein the language comprises acoustic speech data, and wherein the author is a speaker.
19. The system of claim 14 , wherein the two or more user characteristics comprise one or more of gender, age, country, region, city, level of education, life stage, or any combination thereof.
20. The system of claim 14 , wherein the one or more processors is further caused to obtain the translation result using a translation model to determine a proposed translation of the language from multiple possible translations of the language based on a probability distribution showing how likely the possible translations are to appear in an output language.