CN114676708B - Low-resource neural machine translation method based on multi-strategy prototype generation - Google Patents

Abstract

The invention relates to a low-resource neural machine translation method based on multi-strategy prototype generation, belonging to the technical field of natural language processing. The method comprises the following steps: first, a prototype sequence is searched by combining keyword matching and distributed representation matching, and if matching is not obtained, a usable pseudo prototype sequence is generated by using a pseudo prototype generation method. Second, to make efficient use of prototype sequences, conventional encoder-decoder frameworks have been improved. The encoding end receives the prototype sequence input by using an additional encoder; the decoding end uses the improved loss function to reduce the influence of the low-quality prototype sequence on the model while controlling the information flow by using a gating mechanism. The method provided by the invention can effectively improve the quantity and quality of the search results based on a small quantity of parallel corpus, and is suitable for neural machine translation in a low-resource environment and in a similarity language environment.

Description

Low-resource neural machine translation method based on multi-strategy prototype generation

Technical Field

The invention relates to a low-resource neural machine translation method based on multi-strategy prototype generation, and belongs to the technical field of natural language processing.

Background technique

In recent years, with the introduction of end-to-end translation models and attention mechanisms, neural machine translation (NMT) has made great progress. Its translation performance on mainstream language pairs has quickly surpassed statistical machine translation and has gradually developed into the current mainstream machine translation model. In order to improve the performance of neural machine translation, researchers have proposed various methods. Among them, the prototype method based on prototype sequence integration has received a lot of attention. In resource-rich scenarios, using similarity translation as the target-side prototype sequence can effectively improve the performance of neural machine translation. However, in low-resource scenarios, due to the lack of parallel corpus resources, the prototype sequence cannot be matched or the sequence quality is poor. Therefore, in low-resource scenarios, exploring how to effectively use prototype sequences to improve the performance of neural machine translation has very important research and application value.

The prototype sequence is a target-side sentence existing in the translation memory, which contains the semantic information of the target language. The prototype method utilizes the semantic information of the target side by introducing the prototype sequence in the translation process, so that it is implicitly used to guide processes such as word alignment and decoding constraints. At present, the research work in the field of prototype methods mainly focuses on the two stages of prototype retrieval and prototype utilization. The prototype sequence retrieval method has been well developed in resource-rich scenarios because there are large-scale translation memories in resource-rich scenarios. Therefore, the prototype method can obtain high-quality prototype sequences by retrieving the memory, thereby effectively improving the translation performance. However, in low-resource scenarios, limited by the scale and quality of parallel corpora, traditional prototype sequence retrieval methods often find it difficult to retrieve usable prototypes. The effect of the next translation task is limited. In addition, in terms of the utilization of prototype sequences, especially the way of integrating prototype sequences as encoding input into the translation model, researchers have proposed many improvement methods. For example, a dual encoder structure is used to encode the input sentence and the prototype sequence at the same time, and a gating mechanism is introduced at the decoding end to balance the information ratio between the source sentence and the prototype sequence. However, all of the above methods have brought improvements in translation performance, but they are still mainly aimed at resource-rich scenarios, and there are few specific improvements for low-resource scenarios. Therefore, the present invention proposes a low-resource neural machine translation method based on multi-strategy prototype generation, which better improves the performance of low-resource neural machine translation through an improved prototype acquisition method and a specific translation framework structure.

Summary of the invention

The present invention provides a low-resource neural machine translation method based on multi-strategy prototype generation. The efficiency and quality of prototype sequence acquisition are improved by combining the traditional retrieval method with the proposed pseudo-prototype generation method. At the same time, the retrieved prototypes are integrated into the codec framework by changing the neural network structure, thereby maximizing the use of the semantic information contained in the prototype sequence while weakening the impact of low-quality sequences. The performance of low-resource neural machine translation can be improved.

The technical solution of the present invention is: a low-resource neural machine translation method based on multi-strategy prototype generation, and the specific steps of the method are as follows:

Step 1, corpus preprocessing: preprocess parallel training corpora, verification corpora and test corpora of different sizes for model training, parameter tuning and effect testing; and build multi-language global replacement dictionaries and keyword dictionaries for pseudo-prototype generation;

Step 2, prototype generation: Prototype generation is performed using a prototype generation method based on a mixture of multiple strategies to ensure the availability of the prototype sequence; the specific idea of this step is: first, prototype retrieval is performed by combining fuzzy matching and distributed representation matching. If the prototype is not retrieved, the keyword in the input sentence is replaced by a word replacement operation to obtain a pseudo prototype sequence;

Step 3. Construct a translation model that incorporates the prototype sequence: Improve the encoder-decoder structure of the traditional attention-based neural machine translation model to better integrate the prototype sequence, use the corpus of Step 1 and Step 2 as the model input, and generate the final translation.

As a preferred embodiment of the present invention, the specific steps of Step 1 are:

Step 1.1. Use the general dataset IWSLT15 in the field of machine translation for model training. The translation tasks are English-Vietnamese, English-Chinese and English-German. For verification and testing, select tst2012 as the verification set for parameter optimization and model selection, and select tst2013 as the test set for test evaluation.

Step 1.2, use PanLex, Wikipedia, the laboratory's own English-Chinese-Southeast Asian dictionary and Google Translate interface to build an English-Vietnamese-Chinese-German global replacement dictionary;

Step 1.3: Based on Step 1.2, a key dictionary is obtained through tag screening, and all entities are retained during the screening process; to avoid excessive concentration of replacement on certain hot nouns, noun words are searched in the corpus and reversed according to the frequency of occurrence.

As a preferred embodiment of the present invention, the specific steps of Step 2 are:

Step 2.1, use fuzzy matching and distributed representation matching to perform prototype retrieval; the specific implementation is as follows: the translation memory is a set of L pairs of parallel sentences {(s _l ,t _l ):l＝1,…,L}, where s _l is the source sentence and t _l is the target sentence; for a given input sentence x, first use keyword matching to search in the translation memory; use fuzzy matching as the keyword matching method, which is defined as:

Where ED(x,s _i ) is the edit distance between x and s _i , |x| is the sentence length of x;

Different from the keyword-based matching method, distributed representation matching performs retrieval based on the distance between sentence vector representations. To some extent, it is a means of similarity retrieval using semantic information, and therefore provides a different retrieval perspective from keyword matching. Distributed representation matching based on cosine similarity is defined as:

where h _x and are the vector representations of x and _si respectively, and ||h _x || is the metric of vector h _x . To achieve fast calculation, we first use the multilingual pre-trained model mBERT to obtain the vector representations of sentences x and _si , and then use the faiss tool to perform similarity matching based on the representations.

When fuzzy matching can obtain the best matching source sentence s _best , the distributed representation matching is used to obtain the set of top-k matching results s′＝{s ₁ ,s ₂ ,…,s _k }. If s _best ∈s′, the target sentence t _best corresponding to s _best is selected as the prototype sequence. When fuzzy matching fails to retrieve a matching source sentence or , the best matching source sentence s _best is retrieved through distributed representation matching;

Step 2.2: If the prototype is not retrieved in Step 2.1, the input sentence is replaced with keywords to generate a pseudo-prototype, which is called pseudo-prototype generation based on word replacement. Specifically, it includes the following two replacement strategies:

Global replacement: When no match is found in the input sentence, the bilingual dictionary is used to replace the words in the input sentence as much as possible based on the maximization principle. The replaced sentence is called a pseudo-prototype sequence.

Keyword replacement: Extract important nouns and entities from the bilingual dictionary to build a keyword dictionary. When no match is found in the input sentence, use the dictionary to replace the keywords in the input sentence to generate a pseudo-prototype sequence. The upper limit of the number of replacements is less than the set threshold. It is expected that based on the shared vocabulary, the pseudo-prototype sequence that mixes source and important target vocabulary can provide guidance for the generation of translation.

As a preferred embodiment of the present invention, Step 3 includes:

The encoding end adopts a dual encoder structure, which can simultaneously receive sentence input and prototype sequence input, and then encode the input into the corresponding hidden state representation; the sentence encoder is a standard Transformer encoder, which is composed of multiple layers; each layer is composed of two sub-layers: a multi-head self-attention layer and a feedforward neural network layer, both of which use residual connections and layer regularization mechanisms; given an input sentence x = (x ₁ , x ₂ , …, x _m ), the sentence encoder encodes it into the corresponding hidden state sequence h _x = (h _x1 , h _x2 , …, h _xm ), where h _xi is the hidden state corresponding to _xi . The prototype encoder is consistent with the sentence encoder in terms of neural network structure. Given a prototype sequence t = (t ₁ , t ₂ , …, t _n ), the prototype encoder encodes it into the corresponding hidden state sequence h _t = (h _t1 , h _t2 , …, h _tn ), where h _ti is the hidden state corresponding to _ti .

As a preferred embodiment of the present invention, Step 3 includes:

The decoding end incorporates a gating mechanism and uses the self-learning ability of the neural network to optimize the ratio between sentence information and prototype information, thereby controlling the information flow during the decoding process. The improved decoder consists of three sublayers: (1) self-attention layer; (2) improved codec attention layer; (3) fully connected feedforward network layer. The improved codec attention layer consists of a sentence codec attention module and a prototype codec attention module. When receiving the output s _self of the multi-head self-attention layer at time i and the output h _x of the sentence encoder, the sentence codec attention module performs attention calculation.

As a preferred solution of the present invention, in Step 3, the sentence encoding and decoding attention module performs attention calculation including:

s _x =MultiHeadAtt(s _self ,h _x ,h _x )

Among them, MultiHeadAtt(·) is the attention calculation based on multiple heads. Similarly, the calculation of prototype encoding and decoding attention is:

s _t =MultiHeadAtt(s _self ,h _t ,h _t )

Subsequently, the sentence encoder-decoder attention output _sx and the prototype encoder-decoder attention output _st are concatenated to calculate the scale variable α:

α＝sigmoid(W _α [s _x ；s _t ]+b _α )

Where W _α and b _α are trainable parameters, and α is then used to calculate the final output of the encoder-decoder attention layer, calculated as:

s _{enc_dec} = α*s _x + (1-α)*s _t

Then s _{enc_dec} is fed into the fully connected feed-forward network as input:

s _ffn = f(s _{enc_dec} )

The definition of f(x) is: f(x) = max(0,xW ₁ +b ₁ )W ₂ +b ₂ , where W ₁ , W ₂ , b ₁ and b ₂ are all parameters. The final translation yi at time _i is calculated as follows:

P(y _i |y _＜i ；x；t,θ)＝softmax(σ(s _ffn ))

Where t is the prototype sequence and σ(·) is the linear transformation function.

The beneficial effects of the present invention are:

1. The present invention obtains prototype sequences by combining a prototype sequence retrieval method and a pseudo-prototype generation method based on word replacement, thereby maximizing the number of available prototype sequences in low-resource scenarios while ensuring sequence quality;

2. The present invention improves the encoder-decoder translation framework. The encoder uses a dual encoding structure. The sentence encoder and the prototype encoder encode the input sentence and the retrieved (generated) prototype respectively. The decoder uses a gating mechanism to control the proportion and flow of information.

3. Improve the loss calculation method of the neural machine translation model. While making the best use of the semantic information contained in high-quality prototype sequences, the model weakens the negative impact of low-quality prototype sequences on the translation model, ultimately improving the translation performance in low-resource scenarios and bringing better translation fluency.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the present invention;

FIG2 is a diagram showing the overall structure of the model proposed by the present invention;

FIG3 is a graph showing the effect of the number of keyword replacements on model performance in the present invention.

Detailed ways

Embodiment 1: As shown in FIG. 1 to FIG. 3 , a low-resource neural machine translation method based on multi-strategy prototype generation, the specific steps of the method are as follows:

Step 1, corpus preprocessing: preprocess parallel training corpora, verification corpora and test corpora of different sizes for model training, parameter tuning and effect testing; and build multi-language global replacement dictionaries and keyword dictionaries for pseudo-prototype generation;

Step 1.1. Use the general dataset IWSLT15 in the field of machine translation for model training. The translation tasks are English-Vietnamese, English-Chinese and English-German. For verification and testing, select tst2012 as the verification set for parameter optimization and model selection, and select tst2013 as the test set for test evaluation.

Step 1.2, use PanLex, Wikipedia, the laboratory's own English-Chinese-Southeast Asian dictionary and Google Translate interface to build an English-Vietnamese-Chinese-German global replacement dictionary;

Step 1.3: Based on Step 1.2, a key dictionary is obtained through tag screening, and all entities are retained during the screening process; to avoid excessive concentration of replacement on certain hot nouns, noun words are searched in the corpus and reversed according to the frequency of occurrence.

The processed parallel corpora are divided into two categories according to their scale: small-scale parallel corpora and large-scale parallel corpora. By applying the method of the present invention to parallel corpora of different scales, we can observe the impact of the increase in corpus scale on information utilization and verify that the proposed method is applicable to scenarios where parallel corpus resources are scarce. Table 1 shows the experimental data information.

Table 1 Experimental data

Step 2, prototype generation: Prototype generation is performed using a prototype generation method based on a mixture of multiple strategies to ensure the availability of the prototype sequence; the specific idea of this step is: first, prototype retrieval is performed by combining fuzzy matching and distributed representation matching. If the prototype is not retrieved, the keyword in the input sentence is replaced by a word replacement operation to obtain a pseudo prototype sequence;

Step 3. Construct a translation model that incorporates the prototype sequence: Improve the encoder-decoder structure of the traditional attention-based neural machine translation model to better integrate the prototype sequence, use the corpus of Step 1 and Step 2 as the model input, and generate the final translation.

As a preferred embodiment of the present invention, the specific steps of Step 2 are:

Step 2.1, use fuzzy matching and distributed representation matching to perform prototype retrieval; the specific implementation is as follows: the translation memory is a set of L pairs of parallel sentences {(s _l ,t _l ):l＝1,…,L}, where s _l is the source sentence and t _l is the target sentence; for a given input sentence x, first use keyword matching to search in the translation memory; in a low-resource environment, due to the small number of long sentences in the corpus, it is easy to cause the length of the matched fragment to be short, making it difficult to form an effective similarity measure. Therefore, this paper does not use N-gram matching, but uses fuzzy matching as a keyword matching method, which is defined as:

Where ED(x,s _i ) is the edit distance between x and s _i , |x| is the sentence length of x;

Different from the keyword-based matching method, distributed representation matching performs retrieval based on the distance between sentence vector representations. To some extent, it is a means of similarity retrieval using semantic information, and therefore provides a different retrieval perspective from keyword matching. Distributed representation matching based on cosine similarity is defined as:

where h _x and are the vector representations of x and _si respectively, and ||h _x || is the metric of vector h _x . To achieve fast calculation, we first use the multilingual pre-trained model mBERT to obtain the vector representations of sentences x and _si , and then use the faiss tool to perform similarity matching based on the representations.

When fuzzy matching can obtain the best matching source sentence s _best , the distributed representation matching is used to obtain the set of top-k matching results s′＝{s ₁ ,s ₂ ,…,s _k }. If s _best ∈s′, the target sentence t _best corresponding to s _best is selected as the prototype sequence. When fuzzy matching fails to retrieve a matching source sentence or , the best matching source sentence s _best is retrieved through distributed representation matching;

Step 2.2: If the prototype is not retrieved in Step 2.1, the input sentence is replaced with keywords to generate a pseudo-prototype, which is called pseudo-prototype generation based on word replacement. Specifically, it includes the following two replacement strategies:

Global replacement: When no match is found in the input sentence, the bilingual dictionary is used to replace the words in the input sentence as much as possible based on the maximization principle. The replaced sentence is called a pseudo-prototype sequence.

Keyword replacement: Extract important nouns and entities from the bilingual dictionary to build a keyword dictionary. When no match is found in the input sentence, use the dictionary to replace the keywords in the input sentence to generate a pseudo-prototype sequence. The upper limit of the number of replacements is less than the set threshold. It is expected that based on the shared vocabulary, the pseudo-prototype sequence that mixes source and important target vocabulary can provide guidance for the generation of translation.

In order to illustrate the prototype retrieval effect of the present invention in low-resource scenarios, the method proposed by the present invention is compared with the baseline prototype retrieval effect on different data scales. Table 2 shows the improvement in matching quality brought by the hybrid prototype retrieval model.

Table 2 Comparison of fuzzy matching and hybrid prototype retrieval effects

It can be seen from the results in Table 2 that in low-resource scenarios, the number of prototype sequences obtained by fuzzy matching retrieval is obviously insufficient. This problem can be alleviated to a certain extent by using a hybrid prototype matching that combines fuzzy matching and distributed representation matching. On the large-scale dataset WMT14, compared with using the fuzzy matching strategy alone, the combined use of distributed representation matching can obtain a more ideal matching result. As for the small-scale dataset IWSLT15, after combining fuzzy matching and distributed representation matching, it is possible to add a comprehensive consideration of the semantic level on the basis of keyword matching, further improving the quality of the prototype sequence. This proves that the hybrid prototype retrieval method proposed in the present invention is suitable for low-resource scenarios.

Figure 3 shows the impact of the number of keyword replacements on model performance in the Vietnamese-English translation task. In the process of pseudo-prototype generation using keyword replacement, we first set the replacement threshold based on experience, and then adjust the threshold based on the performance of the validation set. The evaluation results show that under the sequential dictionary traversal strategy, when the replacement threshold is set to a small value, the generated prototype sequence is limited in difference from the original text and it is difficult to provide effective guidance information for the translation process; when it is set to a large value, it tends to degenerate into a global replacement; when the number of keyword replacements is set to 3, the model performs best. The results on the test set show the same trend of change.

As a preferred embodiment of the present invention, Step 3 includes:

The encoding end adopts a dual encoder structure, which can simultaneously receive sentence input and prototype sequence input, and then encode the input into the corresponding hidden state representation; the sentence encoder is a standard Transformer encoder, which is composed of multiple layers; each layer is composed of two sub-layers: a multi-head self-attention layer and a feedforward neural network layer, both of which use residual connections and layer regularization mechanisms; given an input sentence x = (x ₁ , x ₂ , …, x _m ), the sentence encoder encodes it into the corresponding hidden state sequence h _x = (h _x1 , h _x2 , …, h _xm ), where h _xi is the hidden state corresponding to _xi . The prototype encoder is consistent with the sentence encoder in terms of neural network structure. Given a prototype sequence t = (t ₁ , t ₂ , …, t _n ), the prototype encoder encodes it into the corresponding hidden state sequence h _t = (h _t1 , h _t2 , …, h _tn ), where h _ti is the hidden state corresponding to _ti .

As a preferred embodiment of the present invention, Step 3 includes:

The decoding end incorporates a gating mechanism and uses the self-learning ability of the neural network to optimize the ratio between sentence information and prototype information, thereby controlling the information flow during the decoding process. The improved decoder consists of three sublayers: (1) self-attention layer; (2) improved codec attention layer; (3) fully connected feedforward network layer. The improved codec attention layer consists of a sentence codec attention module and a prototype codec attention module. When receiving the output s _self of the multi-head self-attention layer at time i and the output h _x of the sentence encoder, the sentence codec attention module performs attention calculation.

As a preferred solution of the present invention, in Step 3, the sentence encoding and decoding attention module performs attention calculation including:

s _x =MultiHeadAtt(s _self ,h _x ,h _x )

Among them, MultiHeadAtt(·) is the attention calculation based on multiple heads. Similarly, the calculation of prototype encoding and decoding attention is:

s _t =MultiHeadAtt(s _self ,h _t ,h _t )

Subsequently, the sentence encoder-decoder attention output _sx and the prototype encoder-decoder attention output _st are concatenated to calculate the scale variable α:

α＝sigmoid(W _α [s _x ；s _t ]+b _α )

Where W _α and b _α are trainable parameters, and α is then used to calculate the final output of the encoder-decoder attention layer, calculated as:

s _{enc_dec} = α*s _x + (1-α)*s _t

Then s _{enc_dec} is fed into the fully connected feed-forward network as input:

s _ffn = f(s _{enc_dec} )

The definition of f(x) is: f(x) = max(0,xW ₁ +b ₁ )W ₂ +b ₂ , where W ₁ , W ₂ , b ₁ and b ₂ are all parameters. The final translation yi at time _i is calculated as follows:

P(y _i |y _<i ；x；t,θ)＝softmax(σ(s _ffn ))

Where t is the prototype sequence and σ(·) is the linear transformation function.

In order to illustrate the translation effect of the present invention, the translations generated by the baseline system and the present invention are compared. Tables 3 and 4 show the improvement results on a small-scale comment corpus.

Table 3 BLEU value evaluation results (%)

Table 4 RIBES value evaluation results (%)

From the results in Table 3 and Table 4, it can be seen that the method proposed in the present invention can use an additional prototype encoder to learn the prototype sequence representation at the encoding end, and effectively utilize the semantic information contained in the high-quality prototype sequence at the decoding end to reduce the noise information introduced by the low-quality prototype sequence. It can effectively improve the quality and fluency of the translation in low-resource scenarios.

The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims (5)

1. A low-resource neural machine translation method based on multi-strategy prototype generation, characterized in that: the specific steps of the method are as follows: Step 1, corpus preprocessing: preprocess parallel training corpora, verification corpora and test corpora of different sizes for model training, parameter tuning and effect testing; and build multi-language global replacement dictionaries and keyword dictionaries for pseudo-prototype generation; Step 2, prototype generation: Prototype generation is performed using a prototype generation method based on a mixture of multiple strategies to ensure the availability of the prototype sequence; the specific idea of this step is: first, prototype retrieval is performed by combining fuzzy matching and distributed representation matching. If the prototype is not retrieved, the keyword in the input sentence is replaced by a word replacement operation to obtain a pseudo prototype sequence; Step 3: Construction of a translation model that incorporates prototype sequences: Improve the encoder-decoder structure of the traditional attention-based neural machine translation model to better incorporate prototype sequences, use the corpus from Step 1 and Step 2 as model input, and generate the final translation; The specific steps of Step 2 are: Step 2.1, use fuzzy matching and distributed representation matching to perform prototype retrieval; the specific implementation is as follows: the translation memory is a set of L pairs of parallel sentences {(s _l ,t _l ):l＝1,…,L}, where s _l is the source sentence and t _l is the target sentence; for a given input sentence x, first use keyword matching to search in the translation memory; use fuzzy matching as the keyword matching method, which is defined as: Where ED(x,s _i ) is the edit distance between x and s _i , |x| is the sentence length of x; Different from the keyword-based matching method, distributed representation matching performs retrieval based on the distance between sentence vector representations. To some extent, it is a means of similarity retrieval using semantic information, and therefore provides a different retrieval perspective from keyword matching. Distributed representation matching based on cosine similarity is defined as: where h _x and are the vector representations of x and _si respectively, and ||h _x || is the metric of vector h _x . To achieve fast calculation, we first use the multilingual pre-trained model mBERT to obtain the vector representations of sentences x and _si , and then use the faiss tool to perform similarity matching based on the representations. When fuzzy matching can obtain the best matching source sentence s _best , the distributed representation matching is used to obtain the set of top-k matching results s′＝{s ₁ ,s ₂ ,…,s _k }. If s _best ∈s′, the target sentence t _best corresponding to s _best is selected as the prototype sequence. When fuzzy matching fails to retrieve a matching source sentence or , the best matching source sentence s _best is retrieved through distributed representation matching; Step 2.2: If the prototype is not retrieved in Step 2.1, the input sentence is replaced with keywords to generate a pseudo-prototype, which is called pseudo-prototype generation based on word replacement. Specifically, it includes the following two replacement strategies: Global replacement: When no match is found in the input sentence, the bilingual dictionary is used to replace the words in the input sentence as much as possible based on the maximization principle. The replaced sentence is called a pseudo-prototype sequence. Keyword replacement: Extract important nouns and entities from the bilingual dictionary to build a keyword dictionary. When no match is found in the input sentence, use the dictionary to replace the keywords in the input sentence to generate a pseudo-prototype sequence. The upper limit of the number of replacements is less than the set threshold. It is expected that based on the shared vocabulary, the pseudo-prototype sequence that mixes source and important target vocabulary can provide guidance for the generation of translation. 2. According to the low-resource neural machine translation method based on multi-strategy prototype generation according to claim 1, it is characterized in that: the specific steps of Step 1 are: Step 1.1. Use the general dataset IWSLT15 in the field of machine translation for model training. The translation tasks are English-Vietnamese, English-Chinese and English-German. For verification and testing, select tst2012 as the verification set for parameter optimization and model selection, and select tst2013 as the test set for test evaluation. Step 1.2, use PanLex, Wikipedia, the laboratory's own English-Chinese-Southeast Asian dictionary and Google Translate interface to build an English-Vietnamese-Chinese-German global replacement dictionary; Step 1.3: Based on Step 1.2, a key dictionary is obtained through tag screening, and all entities are retained during the screening process; to avoid excessive concentration of replacement on certain hot nouns, noun words are searched in the corpus and reversed according to the frequency of occurrence. 3. According to the low-resource neural machine translation method based on multi-strategy prototype generation according to claim 1, it is characterized in that: the step 3 includes: Step 3.1. The encoding end adopts a dual encoder structure, which can simultaneously receive sentence input and prototype sequence input, and then encode the input into the corresponding hidden state representation; the sentence encoder is a standard Transformer encoder, which is composed of multiple layers; each layer is composed of two sub-layers: a multi-head self-attention layer and a feedforward neural network layer, both of which use residual connections and layer regularization mechanisms; given an input sentence x = (x ₁ , x ₂ , …, x _m ), the sentence encoder encodes it into the corresponding hidden state sequence h _x = (h _x1 , h _x2 , …, h _xm ), where h _xi is the hidden state corresponding to _xi , and the prototype encoder is consistent with the sentence encoder in terms of neural network structure. Given a prototype sequence t = (t ₁ , t ₂ , …, t _n ), the prototype encoder encodes it into the corresponding hidden state sequence h _t = (h _t1 , h _t2 , …, h _tn ), where h _ti is the hidden state corresponding to _ti . 4. The low-resource neural machine translation method based on multi-strategy prototype generation according to claim 1, characterized in that: Step 3 includes: The decoding end incorporates a gating mechanism and uses the self-learning ability of the neural network to optimize the ratio between sentence information and prototype information, thereby controlling the information flow during the decoding process. The improved decoder consists of three sublayers: (1) self-attention layer; (2) improved codec attention layer; (3) fully connected feedforward network layer. The improved codec attention layer consists of a sentence codec attention module and a prototype codec attention module. When receiving the output s _self of the multi-head self-attention layer at time i and the output h _x of the sentence encoder, the sentence codec attention module performs attention calculation. 5. According to the low-resource neural machine translation method based on multi-strategy prototype generation according to claim 4, it is characterized in that: in the step 3, the sentence encoding and decoding attention module performs attention calculation including: s _x =MultiHeadAtt(s _self ,h _x ,h _x ) Among them, MultiHeadAtt(·) is the attention calculation based on multiple heads. Similarly, the calculation of prototype encoding and decoding attention is: s _t =MultiHeadAtt(s _self ,h _t ,h _t ) Subsequently, the sentence encoder-decoder attention output _sx and the prototype encoder-decoder attention output _st are concatenated to calculate the scale variable α: α＝sigmoid(W _α [s _x ；s _t ]+b _α ) Where W _α and b _α are trainable parameters, and α is then used to calculate the final output of the encoder-decoder attention layer, calculated as: s _{enc_dec} = α*s _x + (1-α)*s _t Then s _{enc_dec} is fed into the fully connected feed-forward network as input: s _ffn = f(s _{enc_dec} ) The definition of f(x) is: f(x) = max(0,xW ₁ +b ₁ )W ₂ +b ₂ , where W ₁ , W ₂ , b ₁ and b ₂ are all parameters. The final translation yi at time _i is calculated as follows: P(y _i |y _＜i ；x；t,θ)＝softmax(σ(s _ffn )) Where t is the prototype sequence and σ(·) is the linear transformation function.