Automatic post-editing model for neural machine translation
Techniques are disclosed for training and/or utilizing an automatic post-editing model in correcting translation error(s) introduced by a neural machine translation model. The automatic post-editing model can be trained using automatically generated training instances. A training instance is automatically generated by processing text in a first language using a neural machine translation model to generate text in a second language. The text in the second language is processed using a neural machine translation model to generate training text in the first language. A training instance can include the text in the first language as well as the training text in the first language.
1. A method implemented by one or more processors, the method comprising:
receiving an automatically generated training instance including training text in a target language, ground truth text in the target language, and preceding text in the target language,
wherein the ground truth text is for the training text,
wherein the preceding text is included in the training instance based on the preceding text preceding the ground truth text in an electronic resource from which the ground truth text is obtained, and
wherein the training text is generated by processing the ground truth text, and is generated independent of processing the preceding text;
processing the training text, and processing the preceding text, using an automatic post-editing model to generate predicted output, wherein the automatic post-editing model, when trained, is used in correcting one or more translation errors introduced by a neural machine translation model translating text from a source language into the target language;
determining a difference between the predicted output and the ground truth text for the training text; and
updating one or more weights in the automatic post-editing model based on the determined difference.
2. The method of claim 1 , further comprising automatically generating the training instance, wherein automatically generating the training instance comprises:
processing the ground truth text in the target language using the neural machine translation model to generate text in the source language; and
processing the generated text in the source language using the neural machine translation model to generate the training text in the target language.
3. The method of claim 2 , wherein the one or more translation errors introduced by the neural machine translation model are one or more words incorrectly translated from the source language to the target language.
4. The method of claim 2 , wherein the one or more translation errors introduced by the neural machine translation model are one or more words translated with an incorrect gender from the source language into the target language.
5. The method of claim 1 , wherein the automatic post-editing model is a transformer model that includes a transformer encoder and a transformer decoder, wherein processing the training text using the automatic post-editing model to generate the predicted output comprises:
processing the training text using the transformer encoder to generate an encoded representation of the training text; and
generating the predicted output, token-by-token, using the transformer decoder attended to the encoded representation of the training text.
6. The method of claim 5 , wherein generating the predicted output, token-by-token, using the transformer decoder attended to the encoded representation of the training text comprises:
generating, at each of a plurality of iterations of processing using the transformer decoder, a probability distribution over a vocabulary of tokens;
selecting, from the vocabulary of tokens and based on the probability distribution for the iteration, a corresponding token for the iteration; and
using the selected token as part of input to the transformer decoder in a subsequent iteration of the iterations of processing.
7. The method of claim 1 , wherein the automatic post-editing model is a sequence to sequence model.
8. The method of claim 1 , further comprising:
subsequent to updating one or more weights in the automatic post-editing model:
receiving input text generated using the neural machine translation model;
processing the input text using the trained automatic post-editing model to generate edited text; and
causing a client device to perform one or more actions based on the edited text.
9. The method of claim 8 , wherein causing the client device to perform one or more actions based on the edited text comprises:
processing the edited text using a text to speech engine to generate an audio waveform corresponding to the edited text; and
causing the client device to render the audio waveform via one or more speakers of the client device.
10. The method of claim 8 , wherein causing the client device to perform one or more actions based on the edited text comprises:
processing the edited text to determine one or more device actions of a device associated with the client device; and
causing the device to perform the one or more device actions.
11. The method of claim 10 , wherein the device associated with the client device is a light, a thermostat, or a camera.
12. A system comprising:
a memory for storing instructions; and
at least one processor, the at least one processor configured to execute the instructions to perform a method that includes:
receiving an automatically generated training instance including training text in a target language, ground truth text in the target language, and preceding text in the target language,
wherein the ground truth text is for the training text,
wherein the preceding text is included in the training instance based on the preceding text preceding the ground truth text in an electronic resource from which the ground truth text is obtained, and
wherein the training text is generated by processing the ground truth text, and is generated independent of processing the preceding text;
processing the training text, and processing the preceding text, using an automatic post-editing model to generate predicted output, wherein the automatic post-editing model, when trained, is used in correcting one or more translation errors introduced by a neural machine translation model translating text from a source language into the target language;
determining a difference between the predicted output and the ground truth text for the training text; and
updating one or more weights in the automatic post-editing model based on the determined difference.
13. The system of claim 12 , wherein the method further includes automatically generating the training instance, wherein automatically generating the training instance comprises:
processing the ground truth text in the target language using the neural machine translation model to generate text in the source language; and
processing the generated text in the source language using the neural machine translation model to generate the training text in the target language.
14. The system of claim 13 , wherein the one or more translation errors introduced by the neural machine translation model are one or more words incorrectly translated from the source language to the target language.
15. The system of claim 13 , wherein the one or more translation errors introduced by the neural machine translation model are one or more words translated with an incorrect gender from the source language into the target language.
16. The system of claim 12 , wherein the automatic post-editing model is a transformer model that includes a transformer encoder and a transformer decoder, wherein processing the training text using the automatic post-editing model to generate the predicted output comprises:
processing the training text using the transformer encoder to generate an encoded representation of the training text; and
generating the predicted output, token-by-token, using the transformer decoder attended to the encoded representation of the training text.
17. The system of claim 16 , wherein generating the predicted output, token-by-token, using the transformer decoder attended to the encoded representation of the training text comprises:
generating, at each of a plurality of iterations of processing using the transformer decoder, a probability distribution over a vocabulary of tokens;
selecting, from the vocabulary of tokens and based on the probability distribution for the iteration, a corresponding token for the iteration; and
using the selected token as part of input to the transformer decoder in a subsequent iteration of the iterations of processing.
18. The system of claim 12 , wherein the automatic post-editing model is a sequence to sequence model.
19. The system of claim 12 , wherein the method further include:
subsequent to updating one or more weights in the automatic post-editing model:
receiving input text generated using the neural machine translation model;
processing the input text using the trained automatic post-editing model to generate edited text; and
causing a client device to perform one or more actions based on the edited text.
20. The system of claim 19 , wherein causing the client device to perform one or more actions based on the edited text comprises:
processing the edited text using a text to speech engine to generate an audio waveform corresponding to the edited text; and
causing the client device to render the audio waveform via one or more speakers of the client device.