TWI814216B - Method and device for establishing translation model based on triple self-learning - Google Patents

Abstract

A device for establishing a translation model based on triple self-learning includes a storage module and a processing module. The processing module stores a default translation model for translating an original text into a target language and a translation model that has been trained and used for translating an original text into the target language. The auxiliary translation model performs the first self-learning on the default translation model, and based on the default translation model stored in the storage module, the auxiliary translation model, and an auxiliary translation model that has been trained and used to convert an original language A word translation model that translates a word into a translation word in the target language performs second-level self-learning and third-level self-learning on the default translation model to improve the translation accuracy of the default translation model.

Description

Translation model establishment method and device based on triple self-learning

The present invention relates to a translation model establishment method, and in particular to a translation model establishment method and device based on triple self-learning.

Machine Translation (English: Machine Translation) belongs to the field of computational linguistics, which studies the use of computer programs to translate text or speech from one natural language to another natural language. Current translation machines can sometimes obtain understandable translation results, but in order to obtain more meaningful translation results, it is often necessary to make appropriate edits and adjustments when entering sentences to facilitate computer program analysis. Therefore, it is necessary to improve the results of machine translation. Human intervention is required.

In the article "mT5: A massively multilingual pre-trained text-to-text transformer", Linting Xue et al. proposed a translation model based on pre-training of the data set based on machine learning, which greatly reduced human intervention.

However, existing translation models will cause very obvious translation errors in the translation of professional terms or idioms, so domain adaptive training is required. However, domain adaptive training requires a large amount of labeled data. If If there is insufficient labeled data, domain adaptation training cannot be performed.

Therefore, the purpose of the present invention is to provide a translation model establishment method based on triple self-learning that can improve the translation accuracy of the translation model.

Therefore, the present invention's translation model establishment method based on triple self-learning is implemented by a model establishment device, which stores a default translation model for translating an original text into a translation text in a target language, a trained translation model an auxiliary translation model that is completed and used to translate a source text into a translation text in the target language, and a vocabulary translation model that has been trained and is used to translate a vocabulary in a source language into a translation vocabulary in the target language, The method includes one step (A), one step (B), one step (C), one step (D), one step (E), one step (F), one step (G), and one step (H). , and step (I).

In the step (A), the model building device inputs a plurality of original texts into the default translation model, so that the default translation model outputs a plurality of target language translations respectively corresponding to the original texts.

In step (B), for each target language translation, the model building device inputs the target language translation into the auxiliary translation model, so that the auxiliary translation model outputs a first value corresponding to the target language translation. Original language translation, and obtain a corresponding relationship between the target language translation and the first original language translation

In step (C), the model building device adjusts the default translation model according to the original texts and the first original language translations corresponding to the original texts.

In this step (D), for each original text, the model building device obtains a plurality of residuals in the first original language translation that are different from the original text based on the original text and the first original language translation corresponding to the original text. Different original language translation words, and a plurality of residual original words in the original language that are different from the first original language translation.

In this step (E), for each original text, the model building device inputs the residual original text vocabulary into the vocabulary translation model, so that the vocabulary translation model outputs a plurality of target languages corresponding to the residual original text vocabulary. Translate vocabulary correctly.

In this step (F), for each target language translation, the model building device generates a target based on the target language translation, the correspondence relationship, the residual original vocabulary, and the correct translation vocabulary of the target language. Language patch translation.

In this step (G), for each target language patch translation, the model building device inputs the target language patch translation into the auxiliary translation model, so that the auxiliary translation model outputs a pair of the target language patch translation. Second original language translation.

In this step (H), the model building device adjusts the default translation model according to the original texts and the second original language translations corresponding to the original texts.

In step (I), the model building device adjusts the default translation model according to the residual original vocabulary corresponding to the original text and the correct translation vocabulary of the target language.

Another object of the present invention is to provide a translation model establishment device based on triple self-learning that can improve the translation accuracy of the translation model.

Therefore, the translation model establishment device based on triple self-learning of the present invention includes a storage module and a processing module.

The storage module stores a default translation model used to translate a source text into a translation text in a target language, an auxiliary translation model that has been trained and used to translate a source text into a translation text in the target language, and A word translation model that has been trained and used to translate a word in a source language into a translated word in the target language.

The processing module is electrically connected to the storage module, and the processing module inputs a plurality of original texts into the default translation model, so that the default translation model outputs a plurality of target language translations respectively corresponding to the original texts, and for each A target language translation is input into the auxiliary translation model, so that the auxiliary translation model outputs a first original language translation corresponding to the target language translation, and obtains a translation related to the target language The corresponding relationship between the text and the first original language translation, and then adjust the default translation model according to the original text and the first original language translation corresponding to the original text, and for each original text, according to the corresponding relationship between the original text and the original language of the first original language translation, obtain a plurality of residual original language translation words that are different from the original language translation in the first original language translation, and a plurality of residual original language translation words that are different from the first original language translation in the original language of the residual original vocabulary, and then for each original text, input the residual original vocabulary into the vocabulary translation model, so that the vocabulary translation model outputs a plurality of correctly translated vocabulary in the target language corresponding to the residual original vocabulary, and For each target language translation, a target language patch translation is generated based on the target language translation, the corresponding relationship, the residual original vocabulary, and the target language correct translation vocabulary, and for each target language patch Translate text, input the target language patch translation text into the auxiliary translation model, so that the auxiliary translation model outputs a second source language translation text corresponding to the target language patch translation text, and then correspond to the original text according to the original text The default translation model is adjusted based on the second original language translation, and finally the default translation model is adjusted based on the residual original vocabulary and the correct translation vocabulary in the target language corresponding to the original text.

The effect of the present invention is to: use the translation model building device to adjust the default translation model according to the original texts and the first original language translations corresponding to the original texts, and then adjust the default translation model according to the original texts and the second original language corresponding to the original texts. The language translation text adjusts the default translation model, and adjusts the default translation model according to the residual original vocabulary corresponding to the original text and the correct translation vocabulary of the target language, so as to improve the translation accuracy of the default translation model.

Before the present invention is described in detail, it should be noted that in the following description, similar elements are designated with the same numbering.

Referring to FIG. 1 , an embodiment of the translation model building device 1 based on triple self-learning of the present invention includes a storage module 11 and a processing module 12 electrically connected to the storage module 11 . It is worth noting that in this embodiment, the translation model establishment device 1 based on triple self-learning is, for example, a computer host, but is not limited to this.

The storage module 11 stores a default translation model used to translate an original text into a translation text in a target language, an auxiliary translation model that has been trained and used to translate an original text into a translation text in the target language, A vocabulary translation model that has been trained and used to translate a vocabulary in a source language into a translated vocabulary in the target language, multiple word-type vocabulary training materials, multiple compound word vocabulary training materials, and multiple complex mixed vocabulary training materials. Each word class vocabulary training material has a word class vocabulary in the original language and a word class vocabulary in the target language corresponding to the word class vocabulary in the original language. Each compound word vocabulary training material has a compound word vocabulary in the original language and a compound word vocabulary in the original language. Each compound word training material of the target language compound word has a source language compound word and a target language compound word corresponding to the source language compound word.

It is worth noting that in this embodiment, the preset translation model is, for example, the original model of the mT5 model proposed by Linting Xue et al., and the auxiliary translation model is, for example, the Google mT5 model obtained after training the mT5 model by Google. The original The language is, for example, Chinese, and the target language is, for example, English, but is not limited thereto.

It should be noted again that word-type vocabulary refers to the smallest unit of vocabulary that can be used independently and contains complete semantics. Such as nouns (for example: China Trust), verbs (for example: loss), unit quantifiers (for example: ten thousand yuan), prepositions (for example: Da), etc. Word formation requires at least one free morpheme, such as the preposition "da".

Compound words consist of at least two or more morphemes, and can be formed by collocation of free morphemes and bound morphemes. For example, there are mainly the following types of compound words in Chinese: ① Positive type: the front morpheme modifies the back morpheme, such as: pig iron, "生" is the front morpheme, and "铁" is the back morpheme; ② Parallel type: two morphemes The semantics are similar, such as: purchasing, the morphemes "cai" and "purchase" both mean "buy"; ③Subject-predicate type: one morpheme is a verb, and the other morpheme is the subject of this morpheme, such as: Maoshuan, "hair" As a verb, the semantic reference is "Maoli", "loss" is the main word, and the semantic reference is "loss"; and ④ verb-object type: one morpheme is a verb, and another morpheme is the object of this morpheme, such as: pending order , "shan" is the object of the verb "hang".

Complex mixed vocabulary refers to vocabulary that is formed by using grammatical abbreviation (abbreviation) from an originally complete sentence or multiple phrases. There are two main types of complex mixed forms: 1. Contraction of complete sentence: abbreviation of a complete sentence to form a short word with unique semantic meaning. For example: wire transfer, the deep structure of "wire transfer" is a complete sentence: "remit money by telegram", which is a specialized term used in the financial field after grammatical abbreviation; and ②contraction of multiple phrases: combining two The complete phrase abbreviation is a short word with unique semantic meaning. For example: short-term lending, the deep structure of "short-term lending" is a complete sentence: "short-term lending" ([short-term ADVP] + [loan VP]), which is a specialized term used in the financial field after grammatical abbreviation.

Referring to Figure 1, Figure 2, and Figure 3, it is explained how the translation model establishment device 1 based on triple self-learning of the present invention performs one embodiment of the translation model establishment method based on triple self-learning of the present invention. This embodiment includes a vocabulary translation model. A creation program and a translation model creation program.

The vocabulary translation model establishment program includes steps 21 to 23.

In step 21 , the processing module 12 trains a first preset deep learning model based on the word-type vocabulary training data to establish a word-class translation model for translating word-type vocabulary.

In step 22, the processing module 12 trains a second default deep learning model based on the compound word vocabulary training data to establish a compound word translation model for translating compound word vocabulary.

In step 23, the processing module 12 trains a third default deep learning model based on the complex mixed vocabulary training data to establish a complex hybrid model for translating complex mixed vocabulary.

The word class translation model, the compound word translation model, and the complex hybrid model together constitute the vocabulary translation model.

It is worth noting that in this embodiment, the first default deep learning model is, for example, a rule-based model, and the second default deep learning model is, for example, a noise channel type model. model), the third default deep learning model is, for example, a Sequence-to-Sequence Attention Model. In other implementations, it can also be a model used to translate professional terms or/and idioms. A translation model based on domain adaptive phrases replaces the vocabulary translation model, such as a Domain Adaptive Faster RNN Model for Object Detection, but is not limited to this.

It should be noted that in this embodiment, step 22 follows step 21, and step 23 follows step 22. In other embodiments, steps 21 to 23 are not sequential and can be performed in any order.

The translation model establishment program includes steps 31 to 39.

In step 31 , the processing module 12 inputs a plurality of original texts into the default translation model, so that the default translation model outputs a plurality of target language translations respectively corresponding to the original texts.

In step 32, for each target language translation, the processing module 12 inputs the target language translation into the auxiliary translation model, so that the auxiliary translation model outputs a first original language corresponding to the target language translation. Translate the text, and obtain a corresponding relationship between the target language translation and the first source language translation.

It should be noted that during the translation process of the auxiliary translation model, the processing module 12 obtains the corresponding relationship based on the attention alignment mechanism (transformer attention alignment mechanism), but is not limited to this.

In step 33, the processing module 12 adjusts the default translation model according to the original texts and the first original language translations corresponding to the original texts.

Referring to Figure 4, step 33 includes sub-steps 331~332.

In sub-step 331, the processing module 12 generates a first loss value based on the original texts and the first original language translations corresponding to the original texts.

In sub-step 332, the processing module 12 adjusts the preset translation model according to the first loss value.

In step 34, for each original text, the processing module 12 obtains a plurality of residual originals that are different from the original text in the first original language translation based on the original text and the first original language translation corresponding to the original text. The vocabulary of the language translation, and a plurality of residual original words in the original text that are different from the first original language translation.

For example, if the original text is "CITIC Group shut down its boilers in 2020 to comply with the government's environmental protection policy, resulting in a decline in nickel pig iron production and revenue, with a gross loss of NT$16,200,000", and the corresponding translation of the original text into the first original language It is "Citigroup shut down furnaces in 2020 to comply with the government's environmental protection policy, resulting in a decline in pig iron production and revenue, resulting in a loss of 16.2 million yuan." The original language translation of these residuals is "Citigroup, pig iron, loss, 16.2 million yuan". These residuals The original vocabulary of the difference is "CITIC, nickel pig iron, gross loss, 16,200 thousand yuan", in which "Citigroup, CITIC" is a translation error of the company name, "pig iron, nickel pig iron" is a translation error of the proper nouns in the field, "loss, gross loss" It is a translation error of professional terminology, and "16.2 million, 16,200 thousand yuan" is an error of data translation.

In step 35, for each original text, the processing module 12 inputs the residual original text vocabulary into the vocabulary translation model, so that the vocabulary translation model outputs a plurality of correct translations in the target language corresponding to the residual original text vocabulary. Vocabulary.

It is worth noting that in this embodiment, the processing module 12 uses word stemming to classify the residual original words to determine whether each residual original word belongs to a word category, a compound word, or a complex mixed type. Each residual original word is input into the corresponding model in the word translation model.

In step 36, for each target language translation, the processing module 12 generates a target language patch based on the target language translation, the correspondence, the residual original vocabulary, and the correct translation vocabulary of the target language. Translation.

Referring to Figure 5, step 36 includes sub-steps 361~362.

In sub-step 361, for each target language translation, the processing module 12 obtains a plurality of correspondences related to the residual original words in the target language translation based on the corresponding relationship and the residual original words. Location.

In sub-step 362, for each target language translation, the processing module 12 replaces the words at the corresponding positions in the target language translation with the correctly translated words in the target language corresponding to the residual original words, to generate patch translations for that target language.

In step 37, for each target language patch translation, the processing module 12 inputs the target language patch translation into the auxiliary translation model, so that the auxiliary translation model outputs a second translation corresponding to the target language patch translation. Original language translation.

In step 38 , the processing module 12 adjusts the default translation model according to the original texts and the second original language translations corresponding to the original texts.

Referring to Figure 6, step 38 includes sub-steps 381~382.

In step 381, the processing module 12 generates a second loss value based on the original texts and the second original language translations corresponding to the original texts.

In step 382, the processing module 12 adjusts the preset translation model according to the second loss value.

In step 39, the processing module 12 adjusts the default translation model according to the residual original vocabulary corresponding to the original text and the correct translation vocabulary in the target language.

Referring to Figure 7, step 39 includes sub-steps 391~393.

In sub-step 391, for each original text, the processing module 12 inputs the residual original words corresponding to the original text into the default translation model, so that the default translation model outputs a plurality of residual words corresponding to the original text. The target language prediction of the original vocabulary is the translation vocabulary.

In sub-step 392, the processing module 12 generates a third loss value based on the correct translation vocabulary of the target language corresponding to the original text and the predicted translation vocabulary of the target language.

In sub-step 393, the processing module 12 adjusts the preset translation model according to the third loss value.

It is worth noting that in this embodiment, the processing module 12 uses soft word alignment in the paper "Multi-Reference Training with Pseudo-References for Neural Translation and Text Generation" written by Renjie Zheng et al. alignment algorithm, in sub-steps 331 and 34, compare the original text with the first original language translation corresponding to the original text to obtain the first loss value, the residual original language translation vocabulary and The residual original text vocabulary; in sub-step 381, compare the original text with the second original language translation corresponding to the original text to obtain the second loss value; and in sub-step 392, compare the original text with the second original language translation corresponding to the original text. Compare the correctly translated words in the target language and the predicted translated words in the target language corresponding to the original text to obtain the third loss value, but it is not limited to this.

It should be noted again that in this embodiment, the first loss value, the second loss value, and the third loss value are calculated using the cross-entropy (Cross-Entropy) loss function based on the comparison results. But it is not limited to this.

To sum up, the present invention is a translation model establishment method and device based on triple self-learning. The translation model establishment device 1 adjusts the default translation model according to the original texts and the first original language translation corresponding to the original texts ( That is, perform the first level of self-learning), and then adjust the default translation model according to the original texts and the second original language translations corresponding to the original texts (that is, perform the second level of self-learning), and adjust the default translation model according to the residual text corresponding to the original texts. The original vocabulary and the correctly translated vocabulary in the target language are adjusted to adjust the default translation model (that is, perform the third level of self-learning) to improve the translation accuracy of the default translation model, so the purpose of the present invention can indeed be achieved.

However, the above are only examples of the present invention. They cannot be used to limit the scope of the present invention. All simple equivalent changes and modifications made based on the patent scope of the present invention and the contents of the patent specification are still within the scope of the present invention. within the scope covered by the patent of this invention.

1········· Translation model establishment device 11······· Storage module 12······· Processing module 21~23·· Vocabulary translation model establishment program 31~39·· The translation model establishment procedure includes steps 331~332 sub-steps 361~362 sub-steps 381~382 sub-steps 391~393 sub-steps

Other features and effects of the present invention will be clearly presented in the embodiments with reference to the drawings, in which: Figure 1 is a block diagram illustrating an embodiment of a translation model building device based on triple self-learning of the present invention; Figure 2 is a flow chart illustrating a vocabulary translation model establishment procedure of an embodiment of the translation model establishment method based on triple self-learning of the present invention; Figure 3 is a flow chart illustrating a translation model establishment procedure of this embodiment of the translation model establishment method based on triple self-learning of the present invention; Figure 4 is a flow chart to assist in explaining the sub-steps of step 33 of Figure 3; Figure 5 is a flow chart to assist in explaining the sub-steps of step 36 in Figure 3; Figure 6 is a flow chart to assist in explaining the sub-steps of step 38 of Figure 3; and FIG. 7 is a flow chart to assist in explaining the sub-steps of step 39 of FIG. 3 .

31~39·· Translation model establishment program

Claims (8)

A translation model establishment method based on triple self-learning is implemented by a model establishment device. The model establishment device stores a default translation model for translating an original text into a translation text in a target language, a trained and used With an auxiliary translation model that translates an original text into a translation text in the target language, and a vocabulary translation model that has been trained and is used to translate a vocabulary in the original language into a translation vocabulary in the target language, the method includes The following steps: (A) Input multiple original texts into the default translation model, so that the default translation model outputs multiple target language translations corresponding to the corresponding original texts; (B) For each target language translation, The target language translation is input to the auxiliary translation model, so that the auxiliary translation model outputs a first original language translation corresponding to the target language translation, and obtains a correlation between the target language translation and the first original language The corresponding relationship between the translated texts; (C) Adjust the default translation model according to the original texts and the first original language translations corresponding to the original texts; (D) For each original text, adjust the default translation model according to the original texts and the first original language translations corresponding to the original texts. The original language translation text obtains a plurality of residual original language translation words that are different from the original language translation text in the first original language translation text, and a plurality of residual original language translation text words that are different from the first original language translation text in the original language text. Original vocabulary; (E) For each original text, input the residual original vocabulary into the vocabulary translation model, so that the vocabulary translation model outputs a plurality of correctly translated vocabulary in the target language corresponding to the residual original vocabulary; (F ) for each target language translation, according to the corresponding relationship and the and other residual original words, and obtain a plurality of corresponding positions of the residual original words in the target language translation; (G) for each target language translation, obtain the corresponding positions in the target language translation. The vocabulary is replaced with the correctly translated vocabulary of the target language corresponding to the residual original vocabulary to generate the target language patch translation; (H) for each target language patch translation, input the target language patch translation into the An auxiliary translation model, so that the auxiliary translation model outputs a second original language translation corresponding to the target language patch translation; (I) adjusting the default translation according to the original text and the second original language translation corresponding to the original text model; and (J) adjust the default translation model based on the residual original vocabulary corresponding to the original text and the correct translation vocabulary in the target language. The translation model establishment method based on triple self-learning as described in claim 1, wherein in step (D), a soft word alignment algorithm is used to compare the original text and the first original language translation to obtain the The vocabulary of the residual original language translation and the vocabulary of the residual original language. As for the translation model establishment method based on triple self-learning described in claim 1, the model establishment device also stores multiple word-type vocabulary training data, multiple compound word vocabulary training data, and multiple complex mixed vocabulary training data. Each One word class vocabulary training material has a word class vocabulary in the original language and a target language word class vocabulary corresponding to the word class vocabulary in the original language. Each compound word vocabulary training material has a compound word vocabulary in the original language and a compound word vocabulary corresponding to the original language. Target language compound word vocabulary, each complex mixture The type vocabulary training material has a complex mixed vocabulary in the original language and a pair of complex mixed vocabulary in the original language and a pair of complex mixed words in the target language. Before step (A), it also includes the following steps: (L) According to the word type vocabulary training data to train a first preset deep learning model to establish a word class translation model for translating word class vocabulary; (M) based on the compound word vocabulary training data, a second preset deep learning model Carry out training to establish a compound word translation model for translating compound word vocabulary; and (N) train a third default deep learning model based on the complex mixed vocabulary training data to establish a compound word translation model for translating complex mixed words. A complex hybrid model of type vocabulary, the word class translation model, the compound word translation model, and the complex hybrid model together constitute the vocabulary translation model. The method for establishing a translation model based on triple self-learning as described in claim 1, wherein step (C) includes the following sub-steps: (C-1) Generating a translation text based on the original text and the first original language corresponding to the original text a first loss value; and (C-2) adjusting the default translation model according to the first loss value. Step (I) includes the following sub-steps: (I-1) based on the original texts and the corresponding first text of the original texts. The two original language translations generate a second loss value; and (I-2) adjust the default translation model according to the second loss value. Step (J) includes the following sub-steps: (J-1) for each original text, The residual original text corresponding to the original text Vocabulary is input into the preset translation model, so that the preset translation model outputs a plurality of target language predicted translation vocabularies respectively corresponding to the residual original text vocabulary; (J-2) Correctly translate the vocabulary according to the target language corresponding to the original text and The target language predicts the translation vocabulary to generate a third loss value; and (J-3) adjusts the default translation model according to the third loss value. A device for establishing a translation model based on triple self-learning, including: a storage module that stores a default translation model used to translate an original text into a target language, a trained model used to translate an original text an auxiliary translation model into a translation in the target language, and a vocabulary translation model that has been trained and used to translate a vocabulary in a source language into a translation vocabulary in the target language; and a processing module, electrically connected The storage module inputs a plurality of original texts into the default translation model, so that the default translation model outputs a plurality of target language translations respectively corresponding to the original texts, and for each target language translation, the target language The translated text is input to the auxiliary translation model, so that the auxiliary translation model outputs a first original language translation corresponding to the target language translation, and obtains a relationship between the target language translation and the first original language translation. corresponding relationship, and then adjust the default translation model according to the original text and the first original language translation corresponding to the original text, and for each original text, obtain multiple a residual original language translation vocabulary that is different from the original text in the first original language translation, and a plurality of residual original language vocabulary that is different from the first original language translation in the original text, and then for each Original text, input the residual original text vocabulary into the vocabulary translation model, so that the vocabulary translation model outputs A plurality of correctly translated words in the target language respectively corresponding to the residual original words, and for each target language translation, the processing module obtains a plurality of words related to the residual words based on the corresponding relationship and the residual original words. The corresponding positions of the original words in the target language translation are different, and for each target language translation, the processing module replaces the words in the corresponding positions in the target language translation with the words corresponding to the residual original words. Wait for the target language to correctly translate the vocabulary to generate the target language patch translation, and for each target language patch translation, input the target language patch translation into the auxiliary translation model, so that the auxiliary translation model outputs a corresponding target language A second source language translation of the language patch translation, and then adjust the default translation model according to the original text and the corresponding second source language translation of the original text, and finally adjust the residual original vocabulary and target according to the corresponding original text Language-correct translation vocabulary adjusts this default translation model. The triple self-learning translation model building device of claim 5, wherein the processing module uses a soft word alignment algorithm to compare the original text and the first original language translation to obtain the residual original language The vocabulary of the translated text and the residual vocabulary of the original text. The triple self-learning translation model building device as described in claim 5, wherein the storage module also stores multiple word-type vocabulary training data, multiple compound word vocabulary training data, and multiple complex mixed vocabulary training data, Each word class vocabulary training material has a word class vocabulary in the original language and a word class vocabulary in the target language corresponding to the word class vocabulary in the original language. Each compound word vocabulary training material has a compound word vocabulary in the original language and a compound word vocabulary in the original language. target language compound word vocabulary, each complex mixture The type vocabulary training data has a source language complex mixed vocabulary and a pair of target language complex mixed vocabulary of the original language complex hybrid vocabulary. The processing module converts a first preset depth based on the word type vocabulary training data. The learning model is trained to establish a word class translation model for translating word class vocabulary, and a second default deep learning model is trained based on the compound word vocabulary training data to establish a word class translation model for translating compound word vocabulary. Compound word translation model, and based on the complex mixed vocabulary training data, a third default deep learning model is trained to establish a complex hybrid model for translating complex mixed vocabulary, the word class translation model, the The compound word translation model and the complex hybrid model together constitute the vocabulary translation model. The device for establishing a triple self-learning translation model as claimed in claim 5, wherein the processing module generates a first loss value based on the original texts and the first original language translations corresponding to the original texts, and generates a first loss value based on the first The loss value adjusts the default translation model, and the processing module generates a second loss value based on the original texts and the second original language translations corresponding to the original texts, and adjusts the default translation model according to the second loss value , and for each original text, the processing module inputs the residual original words corresponding to the original text into the default translation model, so that the default translation model outputs a plurality of target languages corresponding to the residual original words respectively. Predict the translation vocabulary, generate a third loss value based on the correct translation vocabulary in the target language corresponding to the original text and predict the translation vocabulary in the target language, and adjust the default translation model based on the third loss value.

Images