Using meta-information in neural machine translation
Systems and methods for neural machine translation are provided. In one example, a neural machine translation system translates text and comprises processors and a memory storing instructions that, when executed by at least one processor among the processors, cause the system to perform operations comprising, at least, obtaining a text as an input to a neural network system, supplementing the input text with meta information as an extra input to the neural network system, and delivering an output of the neural network system to a user as a translation of the input text, leveraging the meta information for translation.
1. A translation system for translating text, the translation system comprising:
processors; and
a memory storing instructions that, when executed by at least one processor among the processors, cause the translation system to perform operations comprising, at least:
training a neural network system using a training algorithm that utilizes first meta information associated with one of a training pair of a source language sentence and an associated target language human translation, or a document or corpus of training pairs of source language sentences and associated target language human translations;
inputting text to the neural network system for a translation of the input text, the input text comprising at least one source word in a source language which is translatable into two or more target words in a target language different from the source language;
inputting second meta information corresponding to the input text to the neural network system, the second meta information indicating a context for the translation and comprising a product category of an electronic commerce listing, a topic of the electronic commerce listing, or both;
selecting a target word for the translation using the neural network system, the target word selected from the two or more target words of the target language based at least in part on the second meta information; and
delivering an output of the neural network system comprising the translation of the input text comprising the selected target word from the two or more target words, wherein the translation is generated based at least in part on the second meta information.
2. The translation system of claim 1 , further comprising:
an encoder-decoder architecture in which an input sentence is first encoded into a fixed-length representation and from which a decoder generates target words.
3. The translation system of claim 1 , wherein the neural network system is trained with an attention mechanism that provides a soft alignment to source sentence words at each decoder step.
4. The translation system of claim 1 , wherein the operations further comprise:
merging information from a last recurrent neural network decoder state, a last decoded word embedding, a current context or attention vector, and a vector representation of the second meta information to generate at least part of the output.
5. The translation system of claim 1 , wherein the input text and the second meta information are received from a client device, and wherein the output of the neural network system is delivered to the client device.
6. A method, at a translation system, comprising:
training a neural network system using a training algorithm that utilizes first meta information associated with one of a training pair of a source language sentence and an associated target language human translation, or a document or corpus of training pairs of source language sentences and associated target language human translations;
inputting text to the neural network system for a translation of the input text, the input text comprising at least one source word in a source language which is translatable into two or more target words in a target language different from the source language;
inputting second meta information corresponding to the input text to the neural network system, the second meta information indicating a context for the translation and comprising a product category of an electronic commerce listing, a topic of the electronic commerce listing, or both;
selecting a target word for the translation using the neural network system, the target word selected from the two or more target words of the target language based at least in part on the second meta information; and
delivering an output of the neural network system comprising the translation of the input text comprising the selected target word from the two or more target words, wherein the translation is generated based at least in part on the second meta information.
7. The method of claim 6 , further comprising:
encoding an input sentence into a fixed-length representation and generating target words.
8. The method of claim 6 , further comprising:
training the neural network system with an attention mechanism to provide a soft alignment to source sentence words at one or more decoder operations.
9. The method of claim 6 , further comprising:
merging information from a last recurrent neural network decoder state, a last decoded word embedding, a current context or attention vector, and a vector representation of the second meta information to generate at least part of the output.
10. A non-transitory machine-readable medium including instructions, which when read by a machine, cause the machine to perform operations comprising, at least:
training a neural network system using a training algorithm that utilizes first meta information associated with one of a training pair of a source language sentence and an associated target language human translation, or a document or corpus of training pairs of source language sentences and associated target language human translations;
inputting text to the neural network system for a translation of the input text, the input text comprising at least one source word in a source language which is translatable into two or more target words in a target language different from the source language;
inputting second meta information corresponding to the input text to the neural network system, the second meta information indicating a context for the translation and comprising a product category of an electronic commerce listing, a topic of the electronic commerce listing, or both;
selecting a target word for the translation using the neural network system, the target word selected from the two or more target words of the target language based at least in part on the second meta information; and
delivering an output of the neural network system comprising the translation of the input text comprising the selected target word from the two or more target words, wherein the translation is generated based at least in part on the second meta information.
11. The medium of claim 10 , wherein the operations further comprise:
encoding an input sentence into a fixed-length representation and generating target words.
12. The medium of claim 10 , wherein the operations further comprise:
training the neural network system with an attention mechanism to provide a soft alignment to source sentence words at one or more decoder operations.
13. The medium of claim 10 , wherein the operations further comprise:
merging information from a last recurrent neural network decoder state, a last decoded word embedding, a current context or attention vector, and a vector representation of the second meta information to generate at least part of the output.