CN110472251A - Translation model training method, sentence translation method, equipment and storage medium - Google Patents

Abstract

The present application discloses a method for training a translation model, including: obtaining a training sample set, which includes a plurality of training samples; determining a disturbance sample set corresponding to each training sample in the training sample set, and the disturbance sample set includes at least A perturbation sample, the semantic similarity between the perturbation sample and the corresponding training sample is higher than the first preset value; the initial translation model is trained by using a plurality of training samples and a set of perturbation samples corresponding to each training sample to obtain a target translation model. The solution provided by the embodiment of the present application introduces disturbance samples during model training, so the robustness and translation quality of machine translation can be improved.

Description

Translation model training method, sentence translation method, equipment and storage medium

technical field

The present application relates to the field of computer technology, and in particular to a translation model training method, a sentence translation method, computer equipment, terminal equipment, and a computer-readable storage medium.

Background technique

With the development of artificial intelligence, machine translation has been widely used, such as simultaneous interpretation and chat content translation, which are based on machine translation to convert one input language into another language output.

Neural machine translation is a machine translation model based entirely on neural networks. It has achieved a good level of translation in many language pairs and has been widely used in various machine translation products. However, since the neural machine translation model is based on a complete neural network, the global nature of its modeling causes each output of the target end to depend on each word input by the source end, making it overly sensitive to small disturbances in the input. For example, in the translation from Chinese to English, the user inputs "They are not afraid of difficulties to make Go AI", the English translation given by the machine translation model is "They are not afraid of difficulties to make Go AI", however, when the user enters a similar The sentence "They are not afraid of difficulties to make Go AI", the output of machine translation has changed drastically, and the result is "They are not afraid to make Go AI", although the user just replaced one of the words with a synonym, the translation result But drastic changes have taken place.

It can be seen that the stability of the current neural machine translation, that is, the robustness is relatively poor.

Contents of the invention

The embodiment of the present application provides a translation model training method, which can improve the robustness and translation quality of machine translation. Embodiments of the present application also provide corresponding sentence translation methods, computer equipment, terminal equipment, and computer-readable storage media.

The first aspect of the present application provides a method for training a translation model, including:

Obtain a training sample set, the training sample set includes a plurality of training samples;

Determine a disturbance sample set corresponding to each training sample in the training sample set, the disturbance sample set includes at least one disturbance sample, and the semantic similarity between the disturbance sample and the corresponding training sample is higher than a first preset value;

An initial translation model is trained by using the plurality of training samples and the perturbation sample set corresponding to each training sample to obtain a target translation model.

The second aspect of the present application provides a method for sentence translation, including:

receiving a first sentence to be translated expressed in a first language;

Translate the first sentence to be translated using a target translation model to obtain a translation result sentence expressed in a second language, wherein the target translation model is trained using a plurality of training samples and each of the plurality of training samples The samples are obtained by training the disturbance sample set corresponding to each sample, the disturbance sample set includes at least one disturbance sample, and the semantic similarity between the disturbance sample and the corresponding training sample is higher than the first preset value;

outputting the translation result sentence expressed in the second language.

The third aspect of the present application provides a translation model training device, including:

an acquisition unit, configured to acquire a training sample set, the training sample set including a plurality of training samples;

A determination unit, configured to determine a disturbance sample set corresponding to each training sample in the training sample set acquired by the acquisition unit, the disturbance sample set includes at least one disturbance sample, and the semantics of the disturbance sample and the corresponding training sample The similarity is higher than the first preset value;

A model training unit, configured to train an initial translation model by using the plurality of training samples obtained by the acquiring unit and the perturbation sample sets corresponding to each training sample determined by the determining unit, so as to obtain a target translation model.

The fourth aspect of the present application provides a sentence translation device, including:

a receiving unit, configured to receive a first sentence to be translated expressed in a first language;

a translation unit, configured to use a target translation model to translate the first sentence to be translated received by the receiving unit to obtain a translation result sentence expressed in a second language, wherein the target translation model uses a plurality of training samples and each of the plurality of training samples corresponding to the disturbance sample set training, the disturbance sample set includes at least one disturbance sample, and the semantic similarity between the disturbance sample and the corresponding training sample is higher than that of the first preset set value;

An output unit, configured to output the translation result sentence expressed in the second language translated by the translation unit.

A fifth aspect of the present application provides a computer device, the computer device comprising: an input/output (I/O) interface, a processor, and a memory, where program instructions are stored in the memory;

The processor is configured to execute program instructions stored in the memory, and execute the method as described in the first aspect.

A sixth aspect of the present application provides a terminal device, where the terminal device includes: an input/output (I/O) interface, a processor, and a memory, where program instructions are stored in the memory;

The processor is configured to execute program instructions stored in the memory, and execute the method as described in the second aspect.

The seventh aspect of the present application provides a computer-readable storage medium, including instructions, which, when run on a computer device, cause the computer device to execute the method described in the first aspect or the method described in the second aspect .

Another aspect of the present application provides a computer program product containing instructions, which when run on a computer, causes the computer to execute the method described in the first aspect or the second aspect above.

In the embodiment of the present application, disturbance samples are used in the translation model training. The semantic similarity between the disturbance samples and the training samples is higher than the first preset value, that is, the semantic similarity between the disturbance samples and the training samples is very similar, so that the trained target translation The model can also translate correctly when it receives sentences with noise. Thereby improving the robustness of machine translation, as well as the translation quality.

Description of drawings

Fig. 1 is a schematic diagram of an embodiment of a system for translation model training in the embodiment of the present application;

Fig. 2 is a schematic diagram of an embodiment of the method for translation model training in the embodiment of the present application;

Fig. 3 is a schematic diagram of the structure of the initial translation model in the embodiment of the present application;

Fig. 4 is a schematic diagram of an embodiment of the method for sentence translation in the embodiment of the present application;

Fig. 5 is a schematic diagram of an application scenario of sentence translation in the embodiment of the present application;

FIG. 6 is a schematic diagram of another application scenario of sentence translation in the embodiment of the present application;

FIG. 7 is a schematic diagram of another application scenario of sentence translation in the embodiment of the present application;

Fig. 8 is a schematic diagram of another application scenario of sentence translation in the embodiment of the present application;

FIG. 9 is a schematic diagram of an embodiment of the device for training the translation model in the embodiment of the present application;

Fig. 10 is a schematic diagram of an embodiment of the device for sentence translation in the embodiment of the present application;

Fig. 11 is a schematic diagram of an embodiment of computer equipment in the embodiment of the present application;

Fig. 12 is a schematic diagram of an embodiment of a terminal device in an embodiment of the present application.

Detailed ways

Embodiments of the present application are described below in conjunction with the accompanying drawings. Apparently, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Those of ordinary skill in the art know that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

The embodiment of the present application provides a translation model training method, which can improve the robustness and translation quality of machine translation. Embodiments of the present application also provide corresponding sentence translation methods, computer equipment, terminal equipment, and computer-readable storage media. Each will be described in detail below.

With the development of artificial intelligence, the accuracy of machine translation is getting higher and higher, which greatly facilitates users. For example, machine translation is used in scenarios such as simultaneous interpretation and text translation. Machine translation is usually model-based translation, that is, by pre-training a translation model, the trained translation model can receive a sentence in one language and then convert the sentence into another language for output. At present, neural machine translation is a machine translation model based entirely on neural networks. The accuracy of translation is high, but the anti-noise ability of this model is not good. Once there is a slight disturbance in the input sentence, the output sentence will be inaccurate. Therefore, the embodiment of the present application provides a translation model training method. During the translation model training, various disturbance samples are introduced into the training samples, thereby ensuring that when the trained translation model receives a sentence with disturbance, It can also be translated correctly.

It should be noted that, in the embodiment of the present application, the disturbance includes noise.

The following describes the translation model training process in the embodiment of the present application with reference to the accompanying drawings.

FIG. 1 is a schematic diagram of an embodiment of a translation model training system in an embodiment of the present application.

As shown in FIG. 1 , an embodiment of the translation model training system in the embodiment of the present application includes a computer device 10 and a database 20 , and training samples are stored in the database 20 .

The computer device 10 obtains a training sample set from the database 20, and then uses the training sample set to perform translation model training to obtain a target translation model.

The process of the model training can be understood by referring to an embodiment of the method for training the translation model in FIG. 2 .

As shown in Figure 2, an embodiment of the translation model training method provided by the embodiment of the present application includes:

101. Acquire a training sample set, where the training sample set includes multiple training samples.

In the embodiment of the present application, the training samples in the training sample set refer to samples without disturbance.

102. Determine a disturbance sample set corresponding to each training sample in the training sample set, the disturbance sample set includes at least one disturbance sample, and the semantic similarity between the disturbance sample and the corresponding training sample is higher than a first preset value .

In the embodiment of the present application, a disturbance sample refers to a sample that contains disturbance information or noise, but the similarity between the semantics and the training sample is basically the same. The disturbance information can be a word with the same meaning but different words, or it can be a word that can be used in other situations. A word that does not change the semantics of the sentence significantly.

The first preset value in the embodiment of the present application can be a specific value, such as: 90% or 95%, etc., this is only an example, and does not limit the value of the first preset value, the first preset value Can be set according to needs.

The relationship between training samples and disturbance samples can be understood by referring to the following examples:

Training sample: "They are not afraid of difficulties to make Go AI".

Disturbance sample: "They are not afraid of difficulties to make Go AI".

It can be seen from the above examples that the semantics of the training samples and the disturbance samples are very close, but a different word, such as "not afraid" is used to replace the original word "not afraid".

103. Train an initial translation model by using the plurality of training samples and the perturbation sample set corresponding to each training sample to obtain a target translation model.

During model training, the training samples and the corresponding perturbed samples are used to train together.

In the embodiment of the present application, disturbance samples are used in the translation model training. The semantic similarity between the disturbance samples and the training samples is higher than the first preset value, that is, the semantic similarity between the disturbance samples and the training samples is very similar, so that the trained target translation The model can also translate correctly when it receives sentences with noise. Thereby improving the robustness of machine translation, as well as the translation quality.

Optionally, in another embodiment of the translation model training method provided in the embodiment of this application,

Each of the training samples is a training sample pair, and the training sample pair includes a training input sample and a training output sample;

Correspondingly, the determination of the perturbation sample set corresponding to each training sample may include:

determining a disturbance input sample set corresponding to each training input sample, and a disturbance output sample corresponding to the disturbance output sample set, the disturbance input sample set includes at least one disturbance input sample, and the disturbance output sample is the same as the training output same sample;

Correspondingly, using the plurality of training samples and the perturbation sample sets corresponding to each training sample to train the initial translation model to obtain the target translation model may include:

The initial translation model is trained by using a plurality of training sample pairs, a disturbance input sample set corresponding to each training input sample, and a disturbance output sample corresponding to the disturbance output sample set, so as to obtain a target translation model.

In this embodiment of the present application, the training input samples are in the first language, and the training output samples are in the second language. First language and second language are different. In the embodiment of the present application, Chinese is used as an example for the first language, and English is used as an example for the second language. However, Chinese and English should not be interpreted as limitations on the translation model in the embodiment of the present application. The translation model in the embodiment of the present application may be applicable to mutual translation between any two different languages. As long as the training samples of the corresponding two languages are used in the training, the translation between the two languages can be realized.

In the embodiment of the present application, each training input sample may have multiple disturbance input samples, but the disturbance output samples corresponding to each disturbance input sample are the same as the training output samples.

The correspondence between the above training input samples, training output samples, disturbance input samples and disturbance output samples can be understood by referring to Table 1.

Table 1

It can be seen from Table 1 above that when the training input sample is x, the training output sample is y, and there are multiple disturbance input samples corresponding to x, namely x′1, x′2, and x′3, etc., and each disturbance input sample corresponds to The perturbed output samples of are all y. In this way, it can be ensured that the trained target translation model will output a translation result of y no matter whether the input is x or x′1, x′2 and x′3. Thus, the robustness and translation quality of the target translation model translation are further guaranteed.

Of course, Table 1 is only an example, and the disturbance input samples corresponding to the training input samples may be less than or more than those listed in Table 1.

The perturbation input sample is introduced above, and the generation of the perturbation input sample is described below.

A perturbed input sample can be generated by:

The determining the disturbance input sample set corresponding to each training input sample in the training sample set may include:

determining a first word in each of the training input samples, the first word being a word to be replaced;

The first words are respectively replaced with at least one second word to obtain the disturbed input sample set, and the semantic similarity between the second words and the first words is higher than a second preset value.

In the embodiment of the present application, a perturbed sentence is generated from the vocabulary level, an input sentence is given, and then the first word to be modified is sampled, the position of the first word is determined, and the first word at these positions Replace with the second word in the list.

The word list will contain many words, and the choice of the second word can be understood by referring to the following formula.

Wherein, E[ _xi ] is the word vector of the first word x _i , and cos(E[ _xi ],E[x]) measures the similarity between the first word x _i and the second word x. Since word vectors can capture the semantic information of words, through this replacement method, the first word x _i in the current sentence can be better replaced with a second word x that has similar semantic information.

Another way to generate perturbed input samples could be:

The determining the disturbance input sample set corresponding to each training input sample in the training sample set may include:

determining a word vector for each term in each of the training input samples;

Each time a different Gaussian noise vector is superimposed on the word vector of each word to obtain the disturbance sample set.

In the embodiment of the present application, the perturbed sentences are generated from the feature level. Given a sentence, you can get the vector of each word in the sentence, and then add Gaussian noise to the word vector of each word to simulate possible disturbance types. You can refer to the following formula for understanding:

E[x′ _i ]＝E[xi _i ]+ε,ε～N(0,δ ² I)

The above formula shows that E[ _xi ] identifies the word vector of the word x _i , E[x′ _i ] is the word vector of the word after adding Gaussian noise, the vector ε is sampled from the Gaussian noise with a variance of δ ² , δ is a hyperparameter.

This technical solution is a general solution, which can freely define any strategy of adding disturbance input.

The generation process of the disturbance input samples is described above, and the structure of the translation model in the embodiment of the present application is introduced below.

As shown in FIG. 3 , the initial translation model provided by the embodiment of the present application includes an encoder, a classifier, and a decoder.

The encoder is used to receive training input samples and corresponding disturbance input samples, and output the first intermediate representation result and the second intermediate representation result, the first intermediate representation result is the intermediate representation result of the training input sample, and the second intermediate representation result is the corresponding The intermediate representation result of perturbing the input samples.

A classifier is used to distinguish between the first intermediate representation result and the second intermediate representation result.

The decoder is configured to output training output samples according to the first intermediate representation result, and output training output samples according to the second intermediate representation result.

The model objective function of the initial translation model includes a classification objective function associated with the classifier and the encoder, a training objective function and a perturbation objective function associated with the encoder and the decoder;

Wherein, the classification objective function includes the training input samples, the corresponding disturbance input samples, parameters of the encoder and parameters of the classifier;

said training objective function includes said training input samples, said training output samples, parameters of said encoder and parameters of said decoder;

The perturbation objective function includes the perturbation input samples, the training output samples, parameters of the encoder and parameters of the decoder.

In the embodiment of the present application, the training input sample can be represented by x, the corresponding disturbance input sample can be represented by x′, the training output sample and the disturbance output sample are both represented by y, the first intermediate representation result can be represented by H _x , and the second The intermediate representation result can be represented by H _x′ , the classification objective function can be represented by L _inv (x,x′), the training objective function can be represented by L _true (x,y), and the disturbance objective function can be represented by L _noisy (x′, y) said.

The initial translation model in this embodiment of the present application may be a neural machine translation model.

The training goal of the initial translation model is to make the translation behavior of the initial translation model for x and x′ basically consistent. The encoder is responsible for converting the sentence x of the first language into H _x , and the decoder outputs the sentence y of the target language with H _x as input. The training goal of the embodiment of the present application is to train a disturbance-invariant encoder and decoder.

Since x' is a small change of x, there will be similar semantic information. Given an input pair (x, x′), the training objectives during translation model training are: (1) the encoded representation H _x should be as close as possible to H _x′ ; (2) given H _x′ , the decoder should output the same y. Therefore, two training objectives are introduced in the embodiment of this application to enhance the robustness of the encoder and decoder:

Introducing _Linv (x,x′) encourages the encoder to output similar representations for x and x′, thus achieving a perturbation-invariant encoder, which is achieved by adversarial learning.

The introduction of L _noisy (x',y) guides the decoder to generate target language sentence y for input x' containing disturbance.

Two newly introduced training objectives enable neural machine translation models to be robust against drastic changes in the output space due to small perturbations in the input. At the same time, the training target L _true (x,y) on the original data x and y will be introduced to ensure that the translation quality is enhanced while improving the robustness of the neural machine translation model.

Therefore, the model objective function of the initial translation model is:

where θ _enc is the parameter of the encoder, θ _dec is the parameter of the decoder, and θ _dis is the parameter of the classifier. α and β are used to control the importance between the original translation task and the stability of the machine translation model.

The goal of a perturbation-invariant encoder is that when the encoder inputs a correct sentence x and its corresponding perturbed sentence x′, the representations generated by the encoder for the two sentences are indistinguishable, which can directly benefit the decoder to generate robust Output. In the embodiment of the present application, the encoder can be used as the generator G, which defines the process of generating the hidden representation H _x sequence. At the same time, a classifier D is introduced to distinguish the original input representation H _x from the perturbed input H _x′ . The role of the generator G is to generate similar representations for x and x', so that the classifier D cannot distinguish them, while the role of the classifier D is to try to distinguish them.

Formally, the adversarial learning objective is defined as:

L _inv (x,x′; θ _enc ,θ _dis )＝E _x～S [-log D(G(x))]

+E _x′～N(x) [-log(1-D(G(x′)))]

Given an input, a classifier outputs a classification value whose goal is to maximize the classification value of the correct sentence x while minimizing the classification value of the perturbed sentence x′.

Stochastic gradient descent is used to optimize the model objective function J(θ). In the forward propagation, in addition to a batch of data containing x and y, it also contains a batch of data x' and y. Through these two batches of data, the value of J(θ) can be calculated, and then the gradient of J(θ) corresponding to the model parameters can be calculated, and these gradients will be used to update the model parameters. Because the goal of _Linv is to maximize the classification value of the correct sentence x while minimizing the classification value of the perturbed sentence x′, so the gradient of _Linv for the parameter set θ _enc is multiplied by -1, and other gradients propagate normally. Thus, the values of θ _enc , θ _dec and θ _dis in the initial training model can be calculated, so as to train a target translation model with anti-noise capability.

That is to say, in this embodiment of the present application, the initial translation model is trained using multiple training sample pairs and the disturbance input sample sets corresponding to each of the training input samples, and the disturbance output samples corresponding to the disturbance output sample sets , to get the target translation model, including:

inputting each training input sample, corresponding perturbed input sample, and corresponding training output sample into said model objective function;

Optimize the model objective function in a gradient descent manner to determine the values of the parameters of the encoder, the values of the parameters of the decoder and the values of the parameters of the classifier, wherein, for the classification target The gradient of the parameters of the encoder in the function is multiplied by -1.

The above describes the training process of the target translation model, and the following describes the process of using the target translation model for sentence translation.

As shown in Figure 4, an embodiment of the method for sentence translation provided by the embodiment of the present application includes:

201. Receive a first sentence to be translated expressed in a first language.

In this embodiment of the present application, the first language may be any type of language supported by the target translation model.

202. Translate the first sentence to be translated by using a target translation model to obtain a translation result sentence expressed in a second language, wherein the target translation model uses multiple training samples and each of the multiple training samples Each of the training samples corresponds to a set of perturbed samples, the set of perturbed samples includes at least one perturbed sample, and the semantic similarity between the perturbed sample and the corresponding training sample is higher than the first preset value.

The target translation model can be understood by referring to the above-mentioned embodiment of the model training process, and details are not repeated here.

203. Output the translation result sentence expressed in the second language.

The second language is a language different from the first language, for example: the first language is Chinese and the second language is English.

In the embodiment of the present application, because the target translation model has anti-noise capability, it can also translate correctly when receiving a sentence with noise. Thereby improving the robustness of machine translation, as well as the translation quality.

Optionally, in another embodiment of the sentence translation method provided in the embodiment of the present application, it may also include:

receiving a second sentence to be translated expressed in the first language, the second sentence to be translated is a disturbance sentence of the first sentence to be translated, the second sentence to be translated is the same as the first sentence to be translated The similarity is higher than the first preset value;

Translating the second sentence to be translated by using a target translation model to obtain the translation result sentence corresponding to the first sentence to be translated;

Output the sentence using the translation result.

In this embodiment of the present application, the first sentence to be translated is not limited to the training input sample in the above example, and may be one of the above disturbance input samples.

The sentence translation solution in the embodiment of the present application can be understood by referring to the following two scenario examples.

(A)-(C) in FIG. 5 is an example diagram of a scene of text translation in a social application according to an embodiment of the present application.

As shown in (A) in Figure 5, to translate "They are not afraid of difficulties to make Go AI" in social applications into English, long press the text part, and the page shown in (B) in Figure 5 will appear , Function boxes such as "copy", "forward", "delete" and "translation" have appeared in the page shown in (B) in Figure 5, of course (B) in Figure 5 is just an example, "translate English" can also be changed to "translation", and then a drop-down box will appear to select the corresponding translated text. After the user clicks "Translate English" on the page shown in (B) in Figure 5, the translation result "They are not afraid of difficulties to make Go AI" shown in (C) in Figure 5 will appear.

(A)-(C) in FIG. 5 are diagrams illustrating another scene example of text translation in a social application according to the embodiment of the present application.

As shown in A in Figure 6, if you want to translate "they are not afraid of difficulties to make Go AI" in social applications into English, press and hold the text part, and the page shown in (B) in Figure 6 will appear. Function boxes such as "copy", "forward", "delete" and "translate" have appeared in the page shown in (B) in Figure 5, and the user

After clicking "Translate English" on the page shown in (B) in Figure 6, the translation result "They are not afraid of difficulties to make Go AI" shown in (C) in Figure 6 will appear.

By (A)-(C) among Fig. 5, and the process of (A)-(C) among Fig. 6 and result are compared in visible, although the sentence to be translated of (A) among Fig. 5 is " they are not afraid of Difficult to make Go AI", the sentence to be translated in (A) of Figure 6 is "They made Go AI without fear of difficulty", for these two semantically similar sentences, Figure 5 (C) and Figure 6 ( In C), the same translation result "They are not afraid of difficulties to make Go AI" was obtained respectively. It can be seen that the sentence translation solution provided by the embodiment of the present application has better robustness and better translation quality.

FIG. 7 is a schematic diagram of the application of sentence translation in the simultaneous interpretation scene according to the embodiment of the present application.

As shown in Figure 7, in the simultaneous interpretation scene, the speaker said "They are not afraid of difficulties to make Go AI" in Chinese, and the sentence heard by the audience using the English channel was "They are not afraid of difficulties to make Go AI".

FIG. 8 is a schematic diagram of another application of sentence translation in the simultaneous interpretation scene according to the embodiment of the present application.

As shown in Figure 8, in the simultaneous interpretation scene, the speaker said in Chinese "They are not afraid of difficulties to make Go AI", and the sentence heard by the audience using the English channel was "They are not afraid of difficulties tomake Go AI".

From the comparison of the examples in Fig. 7 and Fig. 8, it can be seen that for inputs with similar semantics, the translation results are the same. It can be seen that the sentence translation solution provided by the embodiment of the present application has better robustness and better translation quality.

It should be noted that the above two application scenarios are just examples, and the solutions of the embodiments of the present application can be used in various translation scenarios, and the forms of the involved terminal devices are not limited to those shown in Figures 5 to 8 form.

The above embodiments have introduced the training process of the target translation model in the embodiment of the present application and the process of using the target translation model for sentence translation. The following describes the translation model training device, sentence translation device, and calculation in the embodiment of the present application in conjunction with the accompanying drawings. Festival equipment and terminal equipment.

As shown in FIG. 9, an embodiment of the translation model training device 30 provided in the embodiment of the present application includes:

An acquisition unit 301, configured to acquire a training sample set, which includes a plurality of training samples;

A determination unit 302, configured to determine a disturbance sample set corresponding to each training sample in the training sample set acquired by the acquisition unit 301, the disturbance sample set includes at least one disturbance sample, and the disturbance sample is the same as the corresponding training sample The semantic similarity of is higher than the first preset value;

A model training unit 303, configured to train an initial translation model using the plurality of training samples obtained by the acquiring unit 301 and the perturbation sample sets corresponding to each training sample determined by the determining unit 302, so as to obtain the target translation Model.

In the embodiment of the present application, disturbance samples are used in the translation model training. The semantic similarity between the disturbance samples and the training samples is higher than the first preset value, that is, the semantic similarity between the disturbance samples and the training samples is very similar, so that the trained target translation The model can also translate correctly when it receives sentences with noise. Thereby improving the robustness of machine translation, as well as the translation quality.

Optionally, the determining unit 302 is configured to determine the disturbance corresponding to each training input sample when each training sample is a training sample pair, and the training sample pair includes a training input sample and a training output sample. an input sample set, and a disturbance output sample corresponding to the disturbance output sample set, the disturbance input sample set includes at least one disturbance input sample, and the disturbance output sample is the same as the training output sample;

A model training unit 303, configured to train an initial translation model using a plurality of training sample pairs, a set of perturbed input samples corresponding to each of the training input samples, and a set of perturbed output samples corresponding to the set of perturbed output samples, so as to obtain the target translation Model.

Optionally, the determining unit 302 is configured to:

determining a first word in each of the training input samples, the first word being a word to be replaced;

The first words are respectively replaced with at least one second word to obtain the disturbed input sample set, and the semantic similarity between the second words and the first words is higher than a second preset value.

Optionally, the determining unit 302 is configured to:

determining a word vector for each term in each of the training input samples;

Each time a different Gaussian noise vector is superimposed on the word vector of each word to obtain the disturbance sample set.

Optionally, the initial translation model includes an encoder, a classifier and a decoder;

The encoder is configured to receive the training input sample and the corresponding disturbance input sample, and output a first intermediate representation result and a second intermediate representation result, the first intermediate representation result being an intermediate representation result of the training input sample, The second intermediate representation result is an intermediate representation result of the corresponding perturbed input sample;

said classifier for distinguishing said first intermediate representation result from said second intermediate representation result;

The decoder is configured to output the training output samples according to the first intermediate representation result, and output the training output samples according to the second intermediate representation result.

Optionally, the model objective function of the initial translation model includes a classification objective function related to the classifier and the encoder, a training objective function and a perturbation objective function related to the encoder and the decoder;

Wherein, the classification objective function includes the training input samples, the corresponding disturbance input samples, parameters of the encoder and parameters of the classifier;

said training objective function includes said training input samples, said training output samples, parameters of said encoder and parameters of said decoder;

The perturbation objective function includes the perturbation input samples, the training output samples, parameters of the encoder and parameters of the decoder.

Optionally, the model training unit 303 is used for:

inputting each training input sample, corresponding perturbed input sample, and corresponding training output sample into said model objective function;

Optimize the model objective function in a gradient descent manner to determine the values of the parameters of the encoder, the values of the parameters of the decoder and the values of the parameters of the classifier, wherein, for the classification target The gradient of the parameters of the encoder in the function is multiplied by -1.

The translation model training device 30 provided in the embodiment of the present application can be understood by referring to the corresponding content in the above method embodiment, and will not be repeated here.

As shown in Figure 10, an embodiment of the sentence translation device provided by the embodiment of the present application includes:

a receiving unit 401, configured to receive a first sentence to be translated expressed in a first language;

The translation unit 402 is configured to use a target translation model to translate the first sentence to be translated received by the receiving unit 401 to obtain a translation result sentence expressed in a second language, wherein the target translation model uses multiple The training sample and the disturbance sample set corresponding to each training sample in the plurality of training samples are trained, the disturbance sample set includes at least one disturbance sample, and the semantic similarity between the disturbance sample and the corresponding training sample is higher than the first a preset value;

The output unit 403 is configured to output the translation result sentence expressed in the second language translated by the translation unit 402 .

In the embodiment of the present application, because the target translation model has anti-noise capability, it can also translate correctly when receiving a sentence with noise. Thereby improving the robustness of machine translation, as well as the translation quality.

Optionally, the receiving unit 401 is further configured to receive a second sentence to be translated expressed in the first language, the second sentence to be translated is a disturbance sentence of the first sentence to be translated, and the second sentence to be translated is The similarity between the second sentence to be translated and the first sentence to be translated is higher than a first preset value;

The translation unit 402 is further configured to use a target translation model to translate the second sentence to be translated to obtain the translation result sentence corresponding to the first sentence to be translated;

The output unit 403 is further configured to output the translation result sentence.

The device 40 for sentence translation above can be understood by referring to the corresponding content in the method embodiment, and will not be repeated here.

FIG. 11 is a schematic structural diagram of a computer device 50 provided by an embodiment of the present application. The computer device 50 includes a processor 510 , a memory 540 and an input/output (I/O) interface 530 . The memory 540 may include a read-only memory and a random access memory, and provides operation instructions and data to the processor 510 . A portion of memory 540 may also include non-volatile random access memory (NVRAM).

In some embodiments, the memory 540 stores the following elements, executable modules or data structures, or their subsets, or their extensions:

In the embodiment of the present application, during the process of determining the ground marker, by calling the operation instruction stored in the memory 540 (the operation instruction can be stored in the operating system),

Obtain a training sample set, the training sample set includes a plurality of training samples;

Determine a disturbance sample set corresponding to each training sample in the training sample set, the disturbance sample set includes at least one disturbance sample, and the semantic similarity between the disturbance sample and the corresponding training sample is higher than a first preset value;

An initial translation model is trained by using the plurality of training samples and the perturbation sample set corresponding to each training sample to obtain a target translation model.

In the embodiment of the present application, disturbance samples are used in the translation model training. The semantic similarity between the disturbance samples and the training samples is higher than the first preset value, that is, the semantic similarity between the disturbance samples and the training samples is very similar, so that the trained target translation The model can also translate correctly when it receives sentences with noise. Thereby improving the robustness of machine translation, as well as the translation quality.

The processor 510 controls operations of the computer device 50, and the processor 510 may also be referred to as a CPU (Central Processing Unit, central processing unit). Memory 540 may include read-only memory and random-access memory, and provides instructions and data to processor 510 . A portion of memory 540 may also include non-volatile random access memory (NVRAM). In a specific application, various components of the computer device 50 are coupled together through the bus system 520, wherein the bus system 520 may include a power bus, a control bus, and a status signal bus, etc. in addition to a data bus. However, for clarity of illustration, the various buses are labeled as bus system 520 in the figure.

The methods disclosed in the foregoing embodiments of the present application may be applied to the processor 510 or implemented by the processor 510 . The processor 510 may be an integrated circuit chip and has a signal processing capability. In the implementation process, each step of the above method may be implemented by an integrated logic circuit of hardware in the processor 510 or instructions in the form of software. The above-mentioned processor 510 may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Various methods, steps, and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, and the like. The steps of the method disclosed in connection with the embodiments of the present application may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the field such as random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, register. The storage medium is located in the memory 540, and the processor 510 reads the information in the memory 540, and completes the steps of the above method in combination with its hardware.

Optionally, the processor 510 is used for:

When each of the training samples is a pair of training samples, and the pair of training samples includes training input samples and training output samples, determine the set of disturbance input samples corresponding to each training input sample, and the corresponding set of disturbance output samples The perturbed output samples of the perturbed input samples set include at least one perturbed input sample, the perturbed output samples are the same as the training output samples;

The initial translation model is trained by using a plurality of training sample pairs, a disturbance input sample set corresponding to each training input sample, and a disturbance output sample corresponding to the disturbance output sample set, so as to obtain a target translation model.

Optionally, the processor 510 is used for:

determining a first word in each of the training input samples, the first word being a word to be replaced;

The first words are respectively replaced with at least one second word to obtain the disturbed input sample set, and the semantic similarity between the second words and the first words is higher than a second preset value.

Optionally, the processor 510 is used for:

determining a word vector for each term in each of the training input samples;

Each time a different Gaussian noise vector is superimposed on the word vector of each word to obtain the disturbance sample set.

Optionally, the initial translation model includes an encoder, a classifier and a decoder;

The encoder is configured to receive the training input sample and the corresponding disturbance input sample, and output a first intermediate representation result and a second intermediate representation result, the first intermediate representation result being an intermediate representation result of the training input sample, The second intermediate representation result is an intermediate representation result of the corresponding perturbed input sample;

said classifier for distinguishing said first intermediate representation result from said second intermediate representation result;

The decoder is configured to output the training output samples according to the first intermediate representation result, and output the training output samples according to the second intermediate representation result.

Optionally, the model objective function of the initial translation model includes a classification objective function related to the classifier and the encoder, a training objective function and a perturbation objective function related to the encoder and the decoder;

Wherein, the classification objective function includes the training input samples, the corresponding disturbance input samples, parameters of the encoder and parameters of the classifier;

said training objective function includes said training input samples, said training output samples, parameters of said encoder and parameters of said decoder;

The perturbation objective function includes the perturbation input samples, the training output samples, parameters of the encoder and parameters of the decoder.

Optionally, the processor 510 is used for:

inputting each training input sample, corresponding perturbed input sample, and corresponding training output sample into said model objective function;

Optimize the model objective function in a gradient descent manner to determine the values of the parameters of the encoder, the values of the parameters of the decoder and the values of the parameters of the classifier, wherein, for the classification target The gradient of the parameters of the encoder in the function is multiplied by -1.

The above description of the computer device 50 can be understood by referring to the descriptions in FIG. 1 to FIG. 3 , and will not be repeated here.

When the above sentence translation process is performed by a terminal device, such as any terminal device such as a mobile phone, a tablet computer, a PDA (Personal Digital Assistant), a POS (Point of Sales, sales terminal), a vehicle-mounted computer, etc., the terminal is a mobile phone. example:

Fig. 12 shows a block diagram of a partial structure of a mobile phone related to the terminal device provided by the embodiment of the present invention. Referring to FIG. 12 , the mobile phone includes: a radio frequency (Radio Frequency, RF) circuit 1110, a memory 1120, an input unit 1130, a display unit 1140, a sensor 1150, an audio circuit 1160, a wireless fidelity (wireless fidelity, WiFi) module 1170, and a processor 1180 , and camera 1190 and other components. Those skilled in the art can understand that the structure of the mobile phone shown in FIG. 12 does not constitute a limitation to the mobile phone, and may include more or less components than shown in the figure, or combine some components, or arrange different components.

The following is a specific introduction to each component of the mobile phone in conjunction with Figure 12:

The RF circuit 1110 can be used for receiving and sending signals during sending and receiving information or talking, and the RF circuit 1110 is also a transceiver. Specifically, after receiving the downlink information of the base station, it is processed by the processor 1180; in addition, the data for designing uplink is sent to the base station. Generally, the RF circuit 1110 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low noise amplifier (Low Noise Amplifier, LNA), a duplexer, and the like. In addition, RF circuitry 1110 may also communicate with networks and other devices via wireless communications. The above-mentioned wireless communication can use any communication standard or protocol, including but not limited to Global System of Mobile Communication (Global System of Mobile communication, GSM), General Packet Radio Service (General Packet Radio Service, GPRS), Code Division Multiple Access (Code Division Multiple Access, CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (Long Term Evolution, LTE), email, Short Messaging Service (Short Messaging Service, SMS), etc.

The memory 1120 can be used to store software programs and modules, and the processor 1180 executes various functional applications and data processing of the mobile phone by running the software programs and modules stored in the memory 1120 . The memory 1120 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playback function, an image playback function, etc.) and the like; Data created by the use of mobile phones (such as audio data, phonebook, etc.), etc. In addition, the memory 1120 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage devices.

The input unit 1130 can be used to receive the sentence to be translated and the translation indicator light input by the user. Specifically, the input unit 1130 may include a touch panel 1131 and other input devices 1132 . The touch panel 1131, also referred to as a touch screen, can collect touch operations of the user on or near it (for example, the user uses any suitable object or accessory such as a finger or a stylus on the touch panel 1131 or near the touch panel 1131). operation), and drive the corresponding connection device according to the preset program. Optionally, the touch panel 1131 may include two parts, a touch detection device and a touch controller. Among them, the touch detection device detects the user's touch orientation, and detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into contact coordinates, and sends it to the to the processor 1180, and can receive and execute commands sent by the processor 1180. In addition, the touch panel 1131 can be implemented in various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the touch panel 1131 , the input unit 1130 may also include other input devices 1132 . Specifically, other input devices 1132 may include, but are not limited to, one or more of physical keyboards, function keys (such as volume control keys, switch keys, etc.), trackballs, mice, joysticks, and the like.

The display unit 1140 can be used to display the translated results. The display unit 1140 may include a display panel 1141. Optionally, the display panel 1141 may be configured in the form of a liquid crystal display (Liquid Crystal Display, LCD) or an organic light-emitting diode (Organic Light-Emitting Diode, OLED). Further, the touch panel 1131 can cover the display panel 1141, and when the touch panel 1131 detects a touch operation on or near it, it sends it to the processor 1180 to determine the type of the touch event, and then the processor 1180 determines the type of the touch event according to the The type provides a corresponding visual output on the display panel 1141 . Although in FIG. 12, the touch panel 1131 and the display panel 1141 are used as two independent components to realize the input and input functions of the mobile phone, in some embodiments, the touch panel 1131 and the display panel 1141 can be integrated to form a mobile phone. Realize the input and output functions of the mobile phone.

The handset may also include at least one sensor 1150, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor, wherein the ambient light sensor may adjust the brightness of the display panel 1141 according to the brightness of the ambient light, and the proximity sensor may turn off the display panel 1141 and/or when the mobile phone is moved to the ear. or backlight. As a kind of motion sensor, the accelerometer sensor can detect the magnitude of acceleration in various directions (generally three axes), and can detect the magnitude and direction of gravity when it is stationary, and can be used to identify the application of mobile phone posture (such as horizontal and vertical screen switching, related Games, magnetometer attitude calibration), vibration recognition related functions (such as pedometer, tap), etc.; as for other sensors such as gyroscope, barometer, hygrometer, thermometer, infrared sensor, etc. repeat.

The audio circuit 1160, the speaker 1161, and the microphone 1162 can provide an audio interface between the user and the mobile phone. The audio circuit 1160 can transmit the electrical signal converted from the received audio data to the speaker 1161, and the speaker 1161 converts it into an audio signal for output; After being received, it is converted into audio data, and then the audio data is processed by the output processor 1180, and then sent to another mobile phone through the RF circuit 1110, or the audio data is output to the memory 1120 for further processing.

WiFi is a short-distance wireless transmission technology. The mobile phone can help users send and receive emails, browse web pages, and access streaming media through the WiFi module 1170, which provides users with wireless broadband Internet access. Although Fig. 12 shows a WiFi module 1170, it can be understood that it is not an essential component of the mobile phone, and can be completely omitted as required without changing the essence of the invention.

The processor 1180 is the control center of the mobile phone. It uses various interfaces and lines to connect various parts of the entire mobile phone. By running or executing software programs and/or modules stored in the memory 1120, and calling data stored in the memory 1120, execution Various functions and processing data of the mobile phone, so as to monitor the mobile phone as a whole. Optionally, the processor 1180 may include one or more processing units; preferably, the processor 1180 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface and application programs, etc. , the modem processor mainly handles wireless communications. It can be understood that the foregoing modem processor may not be integrated into the processor 1180 .

The camera 1190 is used to collect images.

The mobile phone also includes a power supply (such as a battery) for supplying power to each component. Preferably, the power supply can be logically connected to the processor 1180 through the power management system, so that functions such as charging, discharging, and power consumption management can be realized through the power management system.

Although not shown, the mobile phone may also include a camera, a Bluetooth module, etc., which will not be repeated here.

In the embodiment of the present invention, the processor 1180 included in the terminal also has the following control functions:

receiving a first sentence to be translated expressed in a first language;

Translate the first sentence to be translated using a target translation model to obtain a translation result sentence expressed in a second language, wherein the target translation model is trained using a plurality of training samples and each of the plurality of training samples The samples are obtained by training the disturbance sample set corresponding to each sample, the disturbance sample set includes at least one disturbance sample, and the semantic similarity between the disturbance sample and the corresponding training sample is higher than the first preset value;

outputting the translation result sentence expressed in the second language.

Optionally, you can also:

receiving a second sentence to be translated expressed in the first language, the second sentence to be translated is a disturbance sentence of the first sentence to be translated, the second sentence to be translated is the same as the first sentence to be translated the similarity is higher than the first preset value;

Translating the second sentence to be translated by using a target translation model to obtain the translation result sentence corresponding to the first sentence to be translated;

Output the translation result sentence.

In the above embodiments, all or part of them may be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product.

The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application will be generated in whole or in part. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website, computer, server, or data center Transmission to another website site, computer, server, or data center by wired (eg, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be stored by a computer, or a data storage device such as a server or a data center integrated with one or more available media. The available medium may be a magnetic medium (for example, a floppy disk, a hard disk, or a magnetic tape), an optical medium (for example, DVD), or a semiconductor medium (for example, a Solid State Disk (SSD)).

Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above-mentioned embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium can include: ROM, RAM, disk or CD, etc.

The above has introduced in detail the translation model training method, sentence translation method, device and equipment provided by the embodiment of the present application. In this paper, specific examples have been used to illustrate the principle and implementation of the present application. The description of the above embodiments It is only used to help understand the method of the present application and its core idea; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation and application scope. In summary, The contents of this specification should not be understood as limiting the application.

Claims (15)

1. a kind of method of translation model training characterized by comprising

Training sample set is obtained, includes multiple training samples in the training sample set；

Determine each corresponding disturbance sample set of training sample, the disturbance sample set in the training sample set Sample is disturbed including at least one, the disturbance sample is higher than the first preset value with the semantic similarity of corresponding training sample；

Use the multiple training sample and the corresponding disturbance sample set training initial translation of each training sample Model, to obtain target translation model.

2. the method according to claim 1, wherein each training sample is a training sample pair, institute Training sample is stated to including training input sample and training output sample；

Accordingly, the corresponding disturbance sample set of each training sample of the determination, comprising:

Determine the corresponding disturbance input sample set of each trained input sample and disturbance output sample set pair The disturbance output sample answered, the disturbance input sample set includes at least one disturbance input sample, and the disturbance exports sample This is identical as the training output sample；

Accordingly, just using the multiple training sample and the corresponding disturbance sample set training of each training sample Beginning translation model, to obtain target translation model, comprising:

Using multiple training samples to each trained corresponding disturbance input sample set of input sample, Yi Jisuo The corresponding disturbance output sample training initial translation model of disturbance output sample set is stated, to obtain target translation model.

3. according to the method described in claim 2, it is characterized in that, in the determination training sample set each training it is defeated Enter the corresponding disturbance input sample set of sample, comprising:

Determine that the first word in each trained input sample, first word are word to be replaced；

First word is replaced respectively at least one second word, to obtain the disturbance input sample set, described The semantic similarity of two words and first word is higher than the second preset value.

4. according to the method described in claim 2, it is characterized in that, in the determination training sample set each training it is defeated Enter the corresponding disturbance input sample set of sample, comprising:

Determine the term vector of each word in each trained input sample；

It is superimposed a different Gaussian noise vector, on the term vector of each word every time to obtain the disturbance sample Set.

5. according to any method of claim 2-4, which is characterized in that the initial translation model includes encoder, divides Class device and decoder；

The encoder exports the first intermediate representation for receiving the trained input sample and corresponding disturbance input sample As a result with the second intermediate representation as a result, the first intermediate representation result be the trained input sample intermediate representation as a result, The second intermediate representation result is the intermediate representation result of the corresponding disturbance input sample；

The classifier is for distinguishing the first intermediate representation result and the second intermediate representation result；

The decoder is used for according to the first intermediate representation result output training output sample, according to second middle table Show the result output training output sample.

6. according to the method described in claim 5, it is characterized in that, the model objective function of the initial translation model include with The classifier and the relevant class object function of the encoder, trained mesh relevant to the encoder and the decoder Scalar functions and disturbance objective function；

It wherein, include the trained input sample, the corresponding disturbance input sample, the volume in the class object function The code parameter of device and the parameter of the classifier；

The training objective function include the trained input sample, it is described training output sample, the encoder parameter and The parameter of the decoder；

It is described disturbance objective function include the disturbance input sample, it is described training output sample, the encoder parameter and The parameter of the decoder.

7. according to the method described in claim 6, it is characterized in that, it is described using multiple training samples to each training The corresponding disturbance input sample set of input sample and the corresponding disturbance of disturbance output sample set export sample Training initial translation model, to obtain target translation model, comprising:

Each trained input sample, corresponding disturbance input sample and corresponding training output sample are inputted into the model mesh Scalar functions；

Model objective function described in method optimizing according to gradient decline, to determine value, the institute of the parameter of the encoder State the value of the value of the parameter of decoder and the parameter of the classifier, wherein for encoder in class object function The gradient of parameter is multiplied by -1.

8. a kind of method of statement translation characterized by comprising

Receive the first sentence to be translated expressed with first language；

The described first sentence to be translated is translated using target translation model, to obtain the translation knot expressed with second language Fruit sentence, wherein the target translation model is using each training sample in multiple training samples and the multiple training sample What corresponding disturbance sample set training obtained, the disturbance sample set includes at least one disturbance sample, described to disturb Dynamic sample is higher than the first preset value with the semantic similarity of corresponding training sample；

Export the translation result sentence expressed with second language.

9. according to the method described in claim 8, it is characterized in that, the method also includes:

The second sentence to be translated expressed with the first language is received, second sentence to be translated is described first to be translated The similarity of the disturbance sentence of sentence, second sentence to be translated and the described first sentence to be translated is higher than the first preset value；

The described second sentence to be translated is translated using target translation model, to obtain and the described first sentence pair to be translated The translation result sentence answered；

Export the translation result sentence.

10. a kind of device of translation model training characterized by comprising

Acquiring unit includes multiple training samples in the training sample set for obtaining training sample set；

Determination unit, for determining, each training sample is respectively corresponded in the training sample set that the acquiring unit obtains Disturbance sample set, the disturbance sample set includes at least one disturbance sample, the disturbance sample and corresponding trained sample This semantic similarity is higher than the first preset value；

Model training unit, the multiple training sample and the determination unit for being obtained using the acquiring unit are determined The corresponding disturbance sample set training initial translation model of each training sample, to obtain target translation model.

11. device according to claim 10, which is characterized in that

The determination unit, for being a training sample pair in each training sample, the training sample is to including instruction When practicing input sample and training output sample, the corresponding disturbance input sample set of each trained input sample is determined, with And the disturbance exports the corresponding disturbance of sample set and exports sample, the disturbance input sample set includes at least one disturbance Input sample, the disturbance output sample are identical as the training output sample；

The model training unit, for using multiple training samples to each trained input sample is corresponding disturbs Dynamic input sample set and the corresponding disturbance of disturbance output sample set export sample training initial translation model, with Obtain target translation model.

12. a kind of device of statement translation characterized by comprising

Receiving unit, for receiving the first sentence to be translated expressed with first language；

Translation unit, for being turned over using target translation model first sentence to be translated received to the receiving unit It translates, to obtain the translation result sentence expressed with second language, wherein the target translation model is to use multiple training samples It is obtained with the corresponding disturbance sample set training of training sample each in the multiple training sample, the disturbance sample Set includes at least one disturbance sample, and the disturbance sample, which is higher than first with the semantic similarity of corresponding training sample, to be preset Value；

Output unit, the translation result sentence expressed with second language translated for exporting the translation unit.

13. a kind of computer equipment, which is characterized in that the computer equipment includes: input/output (I/O) interface, processor And memory, program instruction is stored in the memory；

The processor executes method as claimed in claim 1 for executing the program instruction stored in memory.

14. a kind of terminal device, which is characterized in that the terminal device includes: input/output (I/O) interface, processor and deposits Reservoir is stored with program instruction in the memory；

The processor executes method as claimed in claim 8 or 9 for executing the program instruction stored in memory.

15. a kind of computer readable storage medium, including instruction, which is characterized in that when described instruction is transported on a computing device When row, so that the computer equipment is executed as described in method of any of claims 1-7 or claim 8 or 9 Method.