Neural machine translation method and neural machine translation device

Abstract

The present application discloses a neural machine translation method and a neural machine translation device. The neural machine translation method includes: obtaining a source sentence to be translated and a target sentence as a candidate translation of the source sentence to form an original training corpus, calculating the translation entropy of each word included in the source sentence; determining whether the translation entropy of each word is greater than a predetermined threshold; according to the determination result, determining the words with translation entropy greater than the predetermined threshold as the words prone to being missed in translation; replacing the candidate translations of the words prone to being missed in translation with preset characters to form a new target sentence; forming a new training corpus according to the source sentence and the new target sentence; performing parameter training on the neural machine translation model based on the original training corpus and the new training corpus, and performing machine translation by using the neural machine translation model with parameters trained.

Claims

1. A neural machine translation method, characterized in that, comprising: obtaining a source sentence to be translated and a target sentence as a candidate translation of the source sentence to form an original training corpus, calculating the translation entropy of each word included in the source sentence; judging whether the translation entropy of each word is greater than a predetermined threshold; determining, according to the judgment result, the words with translation entropy greater than the predetermined threshold as the words prone to being missed in translation; replacing the candidate translations of the words prone to being missed in translation with preset characters to form a new target sentence; forming a new training corpus according to the source sentence and the new target sentence; performing parameter training on a neural machine translation model based on the original training corpus and the new training corpus, and performing machine translation by using the neural machine translation model with parameters trained, wherein calculating the translation entropy of each word included in the source sentence includes: obtaining multiple candidate translations of each word and the translation probability of each candidate translation, where the translation probability represents the probability that the candidate translation is the target translation of the corresponding word, calculating the translation entropy of each word according to the number of candidate translations of each word and the translation probabilities of each candidate translation, wherein the calculation formula of the translation entropy is expressed as follows: where s denotes a word, K represents the number of candidate translations of the word s, p k represents the translation probability of each candidate translation, and E(s) represents the translation entropy of the word s.

2. The neural machine translation method according to claim 1, characterized in that, performing parameter training on a neural machine translation model based on the original training corpus and the new training corpus includes: performing parameter training on the neural machine translation model by using a maximum likelihood objective function and a gradient descent method based on the original training corpus and the new training corpus.

3. The neural machine translation method according to claim 2, characterized in that, The original training corpus is represented by The new training corpus is represented by The maximum likelihood objective function is: Among them, N represents the number of source sentences, n represents the nth sentence, and X {n} represents the source sentence, and Y {n} represents the target sentence, and Z {n} represents the new target sentence, λ is the correction coefficient, p() represents the translation probability obtained by the neural machine translation model, θ represents the parameters of the neural machine translation model, D represents all the corpora and D = D xy ∪D xz .

4. The neural machine translation method according to claim 3, characterized in that, the correction coefficient is 0.4 to 0.

6.

5. The neural machine translation method according to claim 1, characterized in that, further comprising: establishing a correspondence information library for each word prone to being missed in translation and the preset characters for each word prone to being missed in translation, and storing the correspondence information library in a memory.

6. The neural machine translation method according to claim 5, characterized in that, replacing the candidate translations of the words prone to being missed in translation with the preset characters includes: querying the preset characters for each word prone to being missed in translation from the correspondence information library, and replacing the candidate translations of the words prone to being missed in translation with the corresponding preset characters according to the alignment relationship.

7. The neural machine translation method according to claim 1, characterized in that, the neural machine translation model includes an end-to-end model, an encoder-decoder model, a coverage model, and a reconstruction model.

8. The neural machine translation method according to claim 1, characterized in that, the predetermined threshold is greater than or equal to 4.

9. A neural machine translation device, characterized in that, comprising: an obtaining unit configured to obtain a source sentence to be translated and a target sentence as a candidate translation of the source sentence to form an original training corpus, A calculation unit, configured to calculate the translation entropy of each word included in the source statement from the acquisition unit; A judgment unit, configured to receive the translation entropy of each word from the calculation unit and judge whether the translation entropy of each word is greater than a predetermined threshold; A determination unit, configured to determine, according to the judgment result of the judgment unit, the words with translation entropy greater than the predetermined threshold as the words prone to be missed in translation; A replacement unit, configured to receive the words prone to be missed in translation from the determination unit and replace the candidate translations of the words prone to be missed in translation with preset characters to form a new target statement, wherein a new training corpus is formed according to the source statement and the new target statement; A training unit, configured to receive the original training corpus from the acquisition unit and the new training corpus from the replacement unit, and perform parameter training on the neural machine translation model based on the original training corpus and the new training corpus; and A translation unit, configured to perform machine translation by using the neural machine translation model with parameters trained; wherein, the calculation unit is further configured to: Obtain multiple candidate translations of each word and the translation probability of each candidate translation, where the translation probability represents the probability that the candidate translation is the target translation of the corresponding word; Calculate the translation entropy of each word according to the number of candidate translations of each word and the translation probabilities of the respective candidate translations, wherein the calculation formula of the translation entropy is expressed as follows: Among them, s indicates a word, K represents the number of candidate translations of word s, p k represents the translation probability of each candidate translation, and E(s) represents the translation entropy of word s.

Images