Ioas-based multi-agent collaborative cross-language system translation method and system

Abstract

The application provides a multi-agent collaborative cross-language system translation method and system based on IoAs, and relates to the technical field of intelligent translation. The method comprises the following steps: obtaining to-be-translated text information and conference theme information, and obtaining first fused semantic information; obtaining first target semantic information through a first translation agent; obtaining second target semantic information through a second translation agent; obtaining second fused semantic information through the first target semantic information and the conference theme information, and determining second target language translation text; and obtaining third fused semantic information through the second target semantic information and the conference theme information, and determining first target language translation text. According to the application, the translation accuracy of key text can be improved by referring to the conference theme information, and the semantic consistency of multi-language translation can be improved by cross-attention mechanism to mutually correct the semantic information of multiple languages, so that the overall translation accuracy is improved and the communication cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent translation, and in particular to a multi-agent collaborative cross-language system translation method and system based on IoAs. BACKGROUND

[0002] In the related art, machine translation usually uses a single language model for translation, making it difficult to achieve simultaneous translation of multiple languages. Moreover, after translating a text based on one language into multiple languages, the multiple languages often have inconsistent semantics. In international conferences, international business, and other scenarios, ambiguities are easy to arise, which can lead to additional communication costs or even losses.

[0003] The information disclosed in the background section of this application is only intended to deepen the understanding of the general background of the application and should not be considered as acknowledging or implying in any form that this information constitutes prior art known to those skilled in the art. SUMMARY

[0004] The present application provides a multi-agent collaborative cross-language system translation method and system based on IoAs, which can solve the technical problem of inconsistent semantics of multiple languages after translating a text based on one language into multiple languages in the related art.

[0005] According to a first aspect of the present application, a multi-agent collaborative cross-language system translation method based on IoAs is provided, comprising:

[0006] Obtaining text information to be translated and conference theme information;

[0007] Determining first fusion semantic information of the text information to be translated and the conference theme information through a first cross-attention mechanism of an intelligent translation model;

[0008] Processing the first fusion semantic information through a first translation agent of the intelligent translation model to obtain first target semantic information;

[0009] Processing the first fusion semantic information through a second translation agent of the intelligent translation model to obtain second target semantic information;

[0010] Obtaining second fusion semantic information according to the first target semantic information, the conference theme information, and a second cross-attention mechanism of the intelligent translation model;

[0011] Determining a second target language translation text according to the second fusion semantic information and the second target semantic information;

[0012] Obtaining third fusion semantic information according to the second target semantic information, the conference theme information, and a third cross-attention mechanism of the intelligent translation model;

[0013] According to the third fusion semantic information and the first target semantic information, the first target language translation text is determined.

[0014] According to the first target language translation text and the second target language translation text, the translation text is output.

[0015] According to the present application, by the first cross attention mechanism of the intelligent translation model, the first fusion semantic information of the to-be-translated text information and the conference theme information is determined, comprising:

[0016] Obtain the first semantic vector of the plurality of texts in the to-be-translated text information, and the second semantic vector of the plurality of texts in the conference theme information;

[0017] According to the first cross attention mechanism, the first query matrix, the first key-value matrix and the first weight matrix are obtained;

[0018] Through the first semantic vector, the first semantic matrix of the to-be-translated text information is obtained;

[0019] By multiplying the first semantic matrix and the first query matrix, the first query information matrix is obtained;

[0020] Through the second semantic vector, the second semantic matrix of the conference theme information is obtained;

[0021] By multiplying the second semantic matrix and the first key-value matrix, the first key-value information matrix is obtained;

[0022] By multiplying the second semantic matrix and the first weight matrix, the first weight information matrix is obtained;

[0023] Through the first query information matrix, the first key-value information matrix and the first weight information matrix, the first fusion semantic matrix is obtained;

[0024] According to the first fusion semantic matrix, the first fusion semantic information is obtained.

[0025] According to the present application, according to the second fusion semantic information and the second target semantic information, the second target language translation text is determined, comprising:

[0026] Through the third translation intelligent agent of the intelligent translation model, the second fusion semantic information is processed to obtain the third target semantic information;

[0027] According to the second target semantic information and the third target semantic information, the third semantic matrix is obtained;

[0028] The third semantic matrix is processed through a first self-attention mechanism of the intelligent translation model to obtain a second target semantic feature matrix, wherein the number of vectors in the second target semantic feature matrix is the same as the number of second target semantic information;

[0029] The second target language translation text is determined according to the second target semantic feature matrix.

[0030] According to the present application, the first target language translation text is determined according to the third fused semantic information and the first target semantic information, comprising:

[0031] The third fused semantic information is processed through a fourth translation agent of the intelligent translation model to obtain fourth target semantic information;

[0032] The fourth semantic matrix is obtained according to the fourth target semantic information and the first target semantic information;

[0033] The fourth semantic matrix is processed through a second self-attention mechanism of the intelligent translation model to obtain a first target semantic feature matrix, wherein the number of vectors in the first target semantic feature matrix is the same as the number of first target semantic information;

[0034] The first target language translation text is determined according to the first target semantic feature matrix.

[0035] According to the present application, the training method of the intelligent translation model comprises:

[0036] The training text and the training topic information are obtained, and the first training fused semantic information of the training text and the training topic information is determined, and the first correlation weight of each text in the training text and each text in the training topic information is determined;

[0037] The first training fused semantic information is processed through a first translation agent of the intelligent translation model to obtain first training semantic information, and the first training fused semantic information is processed through a second translation agent of the intelligent translation model to obtain second training semantic information;

[0038] The second training fused semantic information of the first training semantic information and the training topic information is determined through a second cross-attention mechanism, and the second correlation weight of each text in the first training semantic information and the training topic information is determined;

[0039] The third training fused semantic information of the second training semantic information and the training topic information is determined through a third cross-attention mechanism, and the third correlation weight of each text in the second training semantic information and the training topic information is determined;

[0040] According to the second training semantic information, the second training fused semantic information, the labeled translation text of the second target language, the third training fused semantic information, the first training semantic information and the labeled translation text of the first target language, a translation loss function is determined;

[0041] According to the first correlation weight, the second correlation weight and the third correlation weight, an attention loss function is determined;

[0042] According to the translation loss function and the attention loss function, a loss function of the intelligent translation model is determined;

[0043] According to the loss function of the intelligent translation model, the intelligent translation model is trained to obtain a trained intelligent translation model.

[0044] According to the present application, according to the second training semantic information, the second training fused semantic information, the labeled translation text of the second target language, the third training fused semantic information, the first training semantic information and the labeled translation text of the first target language, a translation loss function is determined, comprising:

[0045] The second semantic prediction information is obtained by processing the second training semantic information through the second semantic recognition model of the second target language;

[0046] The second semantic fusion information is obtained by processing the second training fused semantic information through the second semantic recognition model;

[0047] The second semantic annotation information is obtained by processing the semantic vector of the labeled translation text of the second target language through the second semantic recognition model;

[0048] The first semantic prediction information is obtained by processing the first training semantic information through the first semantic recognition model of the first target language;

[0049] The first semantic fusion information is obtained by processing the third training fused semantic information through the first semantic recognition model;

[0050] The first semantic annotation information is obtained by processing the semantic vector of the labeled translation text of the first target language through the first semantic recognition model;

[0051] According to the second semantic prediction information, the second semantic fusion information, the second semantic annotation information, the first semantic prediction information, the first semantic fusion information and the first semantic annotation information, a translation loss function is determined.

[0052] According to the present application, according to the second semantic prediction information, the second semantic fusion information, the second semantic annotation information, the first semantic prediction information, the first semantic fusion information and the first semantic annotation information, a translation loss function is determined, comprising:

[0053] According to the formula

[0054] ,

[0055] determining a translation loss function wherein, is the first semantic fusion information, is the first semantic annotation information, is the first semantic prediction information, is the second semantic fusion information, is the second semantic annotation information, is the second semantic prediction information, and sim is a similarity function, and is a preset weight.

[0056] According to the present application, the attention loss function is determined according to the first correlation weight, the second correlation weight and the third correlation weight, comprising:

[0057] According to the formula

[0058] ,

[0059] determining an attention loss function wherein, is the first correlation weight between the first text in the training topic information and the jth text in the training text, is the first correlation weight between the second text in the training topic information and the jth text in the training text, is the first correlation weight between the ith text in the training topic information and the jth text in the training text, is the first correlation weight between the mth text in the training topic information and the jth text in the training text, and m is the number of texts in the training topic information, is the number of texts in the training text, and j≤ , i≤m, is the annotation weight of the first text in the training topic information, is the annotation weight of the second text in the training topic information, is the annotation weight of the ith text in the training topic information, is the annotation weight of the mth text in the training topic information, is the second correlation weight between the first text in the training topic information and the kth first training semantic information, is the second correlation weight between the second text in the training topic information and the kth first training semantic information, is the second correlation weight between the ith text in the training topic information and the kth first training semantic information, a third association weight between the first text in the training theme information and the s-th second training semantic information, a number of the first training semantic information, k≤ , a third association weight between the first text in the training theme information and the s-th second training semantic information, a third association weight between the second text in the training theme information and the s-th second training semantic information, a third association weight between the i-th text in the training theme information and the s-th second training semantic information, a third association weight between the m-th text in the training theme information and the s-th second training semantic information, a number of the second training semantic information, s≤ , and j、 , i, m, k、 , s、 are positive integers, 、 、 a preset weight value, and sim is a similarity function.

[0060] According to a second aspect of the present application, a multi-agent collaborative cross-language system translation system based on IoAs is provided, comprising:

[0061] An acquisition module is configured to acquire to-be-translated text information and conference theme information.

[0062] A first fusion module is configured to determine first fusion semantic information of the to-be-translated text information and the conference theme information through a first cross-attention mechanism of an intelligent translation model.

[0063] A first target semantic information module is configured to obtain first target semantic information by processing the first fusion semantic information through a first translation agent of the intelligent translation model.

[0064] A second target semantic information module is configured to obtain second target semantic information by processing the first fusion semantic information through a second translation agent of the intelligent translation model.

[0065] A second fusion module is configured to obtain second fusion semantic information according to the first target semantic information, the conference theme information, and a second cross-attention mechanism of the intelligent translation model.

[0066] A second target language module is configured to determine a second target language translation text according to the second fusion semantic information and the second target semantic information.

[0067] The third fusion module is configured to obtain third fused semantic information according to the second target semantic information, the conference theme information, and a third cross attention mechanism of the intelligent translation model;

[0068] The first target language module is configured to determine a first target language translation text according to the third fused semantic information and the first target semantic information.

[0069] The translation text module is configured to output a translation text according to the first target language translation text and the second target language translation text.

[0070] By adopting the above technical solutions, the following technical effects can be achieved:

[0071] According to the present application, the to-be-translated text information can be translated by multiple translation intelligent agents, and in the translation process, the conference theme information is referred to to improve the translation accuracy of key texts. Moreover, the semantic information of multiple languages can be mutually corrected by the cross attention mechanism to improve the semantic consistency of multi-language translation, thereby improving the overall translation accuracy and reducing the communication cost. Furthermore, the first cross attention mechanism can be used to make the first semantic vector correspond to the first fused semantic information that has fused the feature information of the conference theme information, so that the features of the text related to the conference theme information are enhanced, the translation accuracy is improved, and the content of the translation is more consistent with the conference theme, thereby reducing the probability of ambiguity. In addition, the cross attention mechanism and the self-attention mechanism can be used to fuse the features of the semantic vectors corresponding to different target languages, so that the fused semantic vectors not only carry the feature information related to the conference theme information, but also carry the feature information of the semantic vectors of different languages. As a result, the output translation texts of different languages are not only consistent with the conference theme, but also highly consistent in terms of semantics, thereby reducing the probability of ambiguity and reducing the translation errors of different languages, and reducing the understanding differences of participants of different languages. In the training process, the segmentation function can be set as the translation loss function, and when the cross attention mechanism does not play a positive role, the training intensity is improved, thereby improving the accuracy and semantic recognition precision of the cross attention mechanism, and improving the training efficiency and translation accuracy. Moreover, the attention loss function can be set by using the association weight and the annotation weight, so that the association weight can be made more similar to the distribution of the relevance between the text and the main theme in the training theme in the training, so as to fuse more features related to the main theme in the fusion process, reduce the introduction of noise features irrelevant to the main theme, improve the fusion accuracy, and provide more accurate semantic information for the translation process.

[0072] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, but not limiting the present application. Other features and aspects of the present application will become more apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0073] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other embodiments can be obtained from these drawings without creative labor.

[0074] Figure 1 An exemplary flowchart of the IoAs-based multi-agent collaborative cross-language system translation method according to an embodiment of the present application is shown.

[0075] Figure 2 An exemplary schematic diagram of the intelligent translation model according to an embodiment of the present application is shown.

[0076] Figure 3 An exemplary block diagram of the IoAs-based multi-agent collaborative cross-language system translation system according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0077] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments only constitute some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0078] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and some embodiments may not be described again for the same or similar concepts or processes.

[0079] Figure 1 An exemplary flowchart of the IoAs-based multi-agent collaborative cross-language system translation method according to an embodiment of the present application is shown, which comprises:

[0080] Step S1, obtaining the to-be-translated text information and the conference theme information;

[0081] Step S2, determining the first fusion semantic information of the to-be-translated text information and the conference theme information through the first cross-attention mechanism of the intelligent translation model;

[0082] Step S3, processing the first fusion semantic information through the first translation agent of the intelligent translation model to obtain the first target semantic information;

[0083] In step S4, the first fusion semantic information is processed by a second translation intelligent agent of the intelligent translation model to obtain second target semantic information.

[0084] In step S5, the second fusion semantic information is obtained according to the first target semantic information, the conference theme information, and a second cross attention mechanism of the intelligent translation model.

[0085] In step S6, the second target language translation text is determined according to the second fusion semantic information and the second target semantic information.

[0086] In step S7, third fusion semantic information is obtained according to the second target semantic information, the conference theme information, and a third cross attention mechanism of the intelligent translation model.

[0087] In step S8, the first target language translation text is determined according to the third fusion semantic information and the first target semantic information.

[0088] In step S9, the translation text is output according to the first target language translation text and the second target language translation text.

[0089] The IoAs-based multi-intelligent agent cooperative cross-language system translation method according to the embodiments of the present application can translate the text information to be translated by multiple translation intelligent agents, and in the translation process, the conference theme information is referred to to improve the translation accuracy of key texts, and the cross attention mechanism can be used to mutually correct the semantic information of multiple languages to improve the semantic consistency of multi-language translation, thereby improving the overall translation accuracy and reducing the communication cost.

[0090] According to an embodiment of the present application, the IoAs (Intelligence Office Automatic System) can be used to deploy the intelligent translation model, which can be used for translation in international conferences and international business scenarios, and can also be used for multi-language translation.

[0091] Figure 2 An exemplary schematic diagram of the intelligent translation model according to an embodiment of the present application is shown.

[0092] According to an embodiment of the present application, in step S1, the conference theme information can be obtained, which is a brief introduction to the main topic of the conference. For example, the conference is an international business conference, which mainly discusses the roles, division of labor, price of goods, freight mode, delivery date, payment method, etc. of the parties in a business cooperation, and the conference theme information can be used as a basis for checking key texts (such as words, sentences, etc.) in subsequent translation process to reduce the probability of ambiguity.

[0093] According to an embodiment of the present application, the text information to be translated can be obtained. The participant can directly input the text information, or the participant's voice information can be processed based on a speech recognition model to convert into text information, and the present application does not limit this. Moreover, the text information input by the participant or processed from the speech by the speech recognition model can be preprocessed, for example, to remove stop words and the like, to obtain the text information to be translated that can be input into the intelligent translation model.

[0094] According to an embodiment of the present application, in step S2, the conference theme information can carry the keynote information of the current conference, and has a guiding effect on the speech or input text of the participant in the conference. Therefore, the text content related to the conference theme information is usually present in the text information to be translated, and this part of text content is usually the key content in the text information to be translated. The translation of this part of text content can be supervised and checked by the conference theme information, which can improve the translation accuracy of the key content and reduce the probability of ambiguity in the translation of the key content. For example, the first cross-attention mechanism can be used to enhance the features of the key content, thereby improving the accuracy in the translation process.

[0095] According to an embodiment of the present application, the first fusion semantic information of the text information to be translated and the conference theme information is determined by the first cross-attention mechanism of the intelligent translation model, including: obtaining a first semantic vector of a plurality of texts in the text information to be translated, and a second semantic vector of a plurality of texts in the conference theme information; obtaining a first query matrix, a first key-value matrix and a first weight matrix according to the first cross-attention mechanism; obtaining a first semantic matrix of the text information to be translated by the first semantic vector; obtaining a first query information matrix by multiplying the first semantic matrix and the first query matrix; obtaining a second semantic matrix of the conference theme information by the second semantic vector; obtaining a first key-value information matrix by multiplying the second semantic matrix and the first key-value matrix; obtaining a first weight information matrix by multiplying the second semantic matrix and the first weight matrix; obtaining a first fusion semantic matrix by the first query information matrix, the first key-value information matrix and the first weight information matrix; and obtaining the first fusion semantic information according to the first fusion semantic matrix.

[0096] According to an embodiment of the present application, the text information to be translated can include a plurality of texts (for example, a plurality of words or phrases), and the natural language processing model (for example, a natural language processing model based on transformer) can be used to process the plurality of texts in the text information to be translated to obtain a first semantic vector of each text, and the same method can be used to process the plurality of texts in the conference theme information to obtain a second semantic vector of the plurality of texts.

[0097] According to one embodiment of the present application, the first semantic vector and the second semantic vector can be fused through the first cross-attention mechanism, so that the feature information of the first semantic vector is fused with the features of the conference theme information, thereby strengthening the feature information of the text related to the conference theme, improving the translation accuracy, and reducing the probability of translation ambiguity. For example, a certain word has different translations in different fields, such as the English word “Volume” which has the meaning of “volume” in the publishing field and the meaning of “volume” in the acoustics field. Through the first cross-attention mechanism, the relevant text in the text to be translated is enhanced in features, so that the final translation content is more consistent with the conference theme information. In the presence of polysemy, the meaning related to the conference theme information can be more accurately translated, and the probability of ambiguity can be reduced.

[0098] According to one embodiment of the present application, the first cross-attention mechanism can provide a first query matrix, a first key-value matrix, and a first weight matrix. The first query information matrix can be obtained by multiplying the first semantic matrix composed of the first semantic vector with the first query matrix. The first key-value information matrix can be obtained by multiplying the second semantic matrix composed of the second semantic vector with the first key-value matrix. The first weight information matrix can be obtained by multiplying the second semantic matrix with the first weight matrix. The first query information matrix, the first key-value information matrix, and the first weight information matrix are operated by the operation method of the cross-attention mechanism to obtain the first fused semantic matrix. The number of vectors contained in the first fused semantic matrix is consistent with the number of first semantic vectors, that is, the plurality of first fused semantic information in the first fused semantic matrix correspond to the first semantic vectors respectively, so that the first semantic vectors are fused with the features of the conference theme information, thereby having consistent features with the conference theme information in translation, and reducing the probability of ambiguity in translation.

[0099] In this way, the first semantic vector can be fused with the features of the conference theme information in the first fused semantic information corresponding to the first semantic vector through the first cross-attention mechanism, so that the features of the text related to the conference theme information are enhanced, the translation accuracy is improved, and the translation is more consistent with the content of the conference theme, thereby reducing the probability of ambiguity.

[0100] According to one embodiment of the present application, in step S3, the first translation intelligent agent is mainly used to translate the text information to be translated into the language of the first target language, in the example, the text information to be translated is Chinese text information, the first target language is English, the first translation intelligent agent can process the first fused semantic information to obtain the first target semantic information, and the first target semantic information can be directly input into a multi-layer perception network level (for example, a network level composed of a multi-layer fully connected layer and a softmax activation layer) to obtain a plurality of words or phrases of the first target language, and the plurality of words or phrases are spliced to obtain the text of the first target language. In step S4, the second translation intelligent agent is mainly used to translate the text information to be translated into the language of the second target language, in the example, the text information to be translated is Chinese text information, the second target language is French, the second translation intelligent agent can process the first fused semantic information to obtain the second target semantic information, and the second target semantic information can be directly input into a multi-layer perception network level (for example, a network level composed of a multi-layer fully connected layer and a softmax activation layer) to obtain a plurality of words or phrases of the second target language, and the plurality of words or phrases are spliced to obtain the text of the second target language. The first translation intelligent agent and the second translation intelligent agent can both be transformer-based decoder models, and can process the first fused semantic information to obtain a plurality of semantic vectors, i.e., the first target semantic information and the second target semantic information. Moreover, the semantic features of the conference theme information are fused in the first target semantic information and the second target semantic information, so the translated text based on the first target semantic information and the second target semantic information is more consistent with the main points described by the conference theme information and has less ambiguity. However, when translating the same language into two different languages, there can be semantic differences between the different languages, resulting in differences in the semantic understanding of participants speaking different languages. Therefore, after obtaining the first target semantic information and the second target semantic information, the text can not be directly output through a multi-layer perception network level, but the first target semantic information and the second target semantic information are used to mutually correct each other, so that the meanings of the semantic information of different languages are as consistent as possible, the translation error of different languages is reduced, and the understanding difference of participants speaking different languages is reduced.

[0101] According to one embodiment of the present application, in step S5, the second semantic vector of the first target semantic information and the conference theme information can be fused by a second cross-attention mechanism to obtain second fused semantic information, and the fusion manner is similar to the fusion manner when the first fused semantic information is obtained, which will not be described here. After fusion, the features related to the conference theme information in the first target semantic information processed by the first translation agent are fused again, that is, the features related to the conference theme information are strengthened again, and the semantic consistency in the proofreading process is improved. That is, the second fused semantic information fused with the features of the conference theme information is used to proofread the second target semantic information fused with the conference theme information, so that the features of semantic information in different languages are close to the common theme (i.e. conference theme information), the consistency of semantic information in different languages is improved, and the proofreading effect and the translation accuracy of the second target language after proofreading are improved.

[0102] According to one embodiment of the present application, in step S6, the second target semantic information can be proofread by the second fused semantic information, and the proofread second target language translation text is output.

[0103] According to one embodiment of the present application, the second target language translation text is determined according to the second fused semantic information and the second target semantic information, comprising: processing the second fused semantic information by a third translation agent of the intelligent translation model to obtain third target semantic information; obtaining a third semantic matrix according to the second target semantic information and the third target semantic information; processing the third semantic matrix by a first self-attention mechanism of the intelligent translation model to obtain a second target semantic feature matrix, wherein the number of vectors in the second target semantic feature matrix is the same as the number of the second target semantic information; and determining the second target language translation text according to the second target semantic feature matrix.

[0104] According to one embodiment of the present application, the third translation agent can be used to translate the first target language into the second target language, the third translation agent can process the second fused semantic information (semantic information of the first target language) after the features of the conference theme information are fused, to obtain the semantic vector corresponding to the text of the second target language, that is, the third target semantic information, and the third target semantic information can also be processed through multiple levels of perception network levels to obtain the text of the second target language, but it is obtained by converting the text information to be translated into the semantic vector of the first target language and then converting it into the semantic vector of the second target language. Therefore, in this case, the third target semantic information is not directly used to obtain the translated text of the second target language, but the feature information contained in the third target semantic information can be used to correct the second target semantic information, that is, only as a correction basis, and the consistency of the semantic vector of the first target language and the semantic vector of the second target language in the semantic level can be improved, so that the meaning of the translated text of the second target language and the translated text of the third target language is closer, and the probability of translation error of different languages is reduced.

[0105] According to one embodiment of the present application, the second target semantic information (including a plurality of vectors) and the third target semantic information (including a plurality of vectors) can be combined to form a third semantic matrix, and the first self-attention mechanism is used to process the third semantic matrix, so that each semantic vector in the third semantic matrix fuses the relevant features of other semantic vectors, and the second target semantic information can also fuse the relevant features of the third target semantic information, so that the second target semantic information carries part of the features of the third target semantic information after fusing the features, so that the fused second target semantic information has higher consistency with the second fused semantic information corresponding to the third target semantic information in terms of semantic features, thereby improving the consistency of the semantic vector of the first target language and the semantic vector of the second target language (that is, the fused second target semantic information) in terms of semantics.

[0106] According to one embodiment of the present application, the fused second target semantic information can form a second target semantic feature matrix, and each fused second target semantic information in the second target semantic feature matrix can be input into a multi-layer perception network level to obtain the word or phrase of the second target language corresponding to each fused second target semantic information. A plurality of words or phrases can form a translated text of the second target language.

[0107] According to one embodiment of the present application, similarly, in step S7, according to the second target semantic information, the conference theme information and the third cross-attention mechanism of the intelligent translation model, the third fused semantic information, i.e., the semantic information of the second target language fused with the features of the second semantic vector of the conference theme information, is obtained and used to correct the first target semantic information.

[0108] According to one embodiment of the present application, in step S8, according to the third fused semantic information and the first target semantic information, the first target language translation text is determined, including: processing the third fused semantic information through the fourth translation intelligent agent of the intelligent translation model to obtain the fourth target semantic information; obtaining the fourth semantic matrix according to the fourth target semantic information and the first target semantic information; processing the fourth semantic matrix through the second self-attention mechanism of the intelligent translation model to obtain the first target semantic feature matrix, wherein the number of vectors in the first target semantic feature matrix is the same as the number of the first target semantic information; and determining the first target language translation text according to the first target semantic feature matrix. That is, the correction method is similar to the above-mentioned correction method of the second target semantic information, which will not be described here.

[0109] According to one embodiment of the present application, in step S9, the first target language translation text and the second target language translation text are the translation text. Based on the above correction process, the first target language translation text and the second target language translation text are highly related to the conference theme information, and the translation text is fused with the feature information of the semantic vectors of different languages, so that the translation texts of different languages are highly consistent in semantics, the semantic consistency of different languages is improved, the translation error of different languages is reduced, and the understanding difference of participants of different languages is reduced.

[0110] In this way, the semantic vectors corresponding to different target languages can be fused through cross-attention mechanism and self-attention mechanism, so that the fused semantic vectors not only carry feature information related to the conference theme information, but also carry feature information of semantic vectors of different languages, so that the output translation texts of different languages not only conform to the conference theme, but also are highly consistent in semantics, reducing the probability of ambiguity, reducing the translation error of different languages, and reducing the understanding difference of participants of different languages.

[0111] According to one embodiment of the present application, the intelligent translation model can be trained before use, and the training method of the intelligent translation model comprises: obtaining training text and training topic information, determining first training fusion semantic information of the training text and the training topic information, and determining a first correlation weight of each text in the training text and each text in the training topic information; processing the first training fusion semantic information through a first translation agent of the intelligent translation model to obtain first training semantic information, and processing the first training fusion semantic information through a second translation agent of the intelligent translation model to obtain second training semantic information; determining second training fusion semantic information of the first training semantic information and the training topic information through a second cross-attention mechanism, and determining a second correlation weight of the first training semantic information and each text in the training topic information; determining third training fusion semantic information of the second training semantic information and the training topic information through a third cross-attention mechanism, and determining a third correlation weight of the second training semantic information and each text in the training topic information; determining a translation loss function according to the second training semantic information, the second training fusion semantic information, a labeled translation text in the second target language, the third training fusion semantic information, the first training semantic information, and a labeled translation text in the first target language; determining an attention loss function according to the first correlation weight, the second correlation weight, and the third correlation weight; determining a loss function of the intelligent translation model according to the translation loss function and the attention loss function; and training the intelligent translation model according to the loss function of the intelligent translation model to obtain a trained intelligent translation model.

[0112] According to one embodiment of the present application, the first training fusion semantic information is obtained in a similar manner to the above-mentioned first fusion semantic information, and will not be described again here. Moreover, the first correlation weight is obtained when the first cross-attention mechanism processes the semantic vector of the training text and the semantic vector of the training topic information, that is, the semantic vector of the training text is multiplied by the first query matrix to obtain a query vector, the semantic vector of the training topic information is multiplied by the first key-value matrix to obtain a key-value vector, and the inner product of the query vector and the key-value vector is the first correlation weight of the semantic vector of the training text and the semantic vector of the training topic information. The first correlation weight between each semantic vector of the training text and each semantic vector of the training topic information can be determined in the same manner.

[0113] According to one embodiment of the present application, the first training semantic information, the second training semantic information, the second training fusion semantic information, and the third training fusion semantic information are obtained in a similar manner to the above-mentioned first target semantic information, second target semantic information, second fusion semantic information, and third fusion semantic information, respectively, and will not be described again here. The second correlation weight and the third correlation weight are obtained in a similar manner to the above-mentioned first correlation weight, and will not be described again here.

[0114] According to one embodiment of the present application, the translation loss function is determined according to the second training semantic information, the second training fused semantic information, the labeled translation text in the second target language, the third training fused semantic information, the first training semantic information and the labeled translation text in the first target language, comprising: processing the second training semantic information through the second semantic recognition model of the second target language to obtain second semantic prediction information; processing the second training fused semantic information through the second semantic recognition model to obtain second semantic fusion information; processing the semantic vector of the labeled translation text in the second target language through the second semantic recognition model to obtain second semantic label information; processing the first training semantic information through the first semantic recognition model of the first target language to obtain first semantic prediction information; processing the third training fused semantic information through the first semantic recognition model to obtain first semantic fusion information; processing the semantic vector of the labeled translation text in the first target language through the first semantic recognition model to obtain first semantic label information; and determining the translation loss function according to the second semantic prediction information, the second semantic fusion information, the second semantic label information, the first semantic prediction information, the first semantic fusion information and the first semantic label information.

[0115] According to one embodiment of the present application, the second semantic recognition model can be a decoder model of transformer, which can process the second training semantic information to obtain a plurality of semantic recognition vectors, and the plurality of semantic recognition vectors can be mapped to the characters in the second target language through a plurality of perception network levels, where the character output can not be performed, but the semantic vectors are fused, for example, fused to the last semantic vector (each semantic vector is weighted and summed with the last semantic vector) to obtain the second semantic prediction information. Similarly, the second semantic fusion information can be obtained by processing the second training fused semantic information through the second semantic recognition model, and the second semantic label information can be obtained by processing the semantic vector of the labeled translation text in the second target language (for example, the semantic vector obtained by processing the labeled translation text through the natural language processing model based on transformer) through the second semantic recognition model.

[0116] According to one embodiment of the present application, similarly, the first semantic prediction information, the first semantic fusion information and the first semantic label information can be obtained through the first semantic recognition model, and then the translation loss function is determined.

[0117] According to one embodiment of the present application, the translation loss function is determined according to the second semantic prediction information, the second semantic fusion information, the second semantic label information, the first semantic prediction information, the first semantic fusion information and the first semantic label information, comprising: determining the translation loss function according to formula (1) ,

[0118] (1),

[0119] in, For the first semantic fusion information, This is the first semantic annotation information. This is the first semantic prediction information. For second semantic fusion information, This is the second semantic annotation information. The second semantic prediction information is represented by sim, where sim is the similarity function. and Preset weights.

[0120] According to an embodiment of the present invention, in the first piecewise function of formula (1), for and Similarity, for example, cosine similarity. for and The similarity, if This indicates that the first semantic fusion information is closer to the first semantic annotation information, meaning that the fusion process of the cross-attention mechanism has a positive impact on semantic accuracy, that is, it improves semantic accuracy. In this case, it can be... As the value of the piecewise function, this value can be reduced during training, thereby improving the relationship between 1 and... The error between them, that is, further improvement and The similarity improves semantic accuracy. Otherwise, it will... As the value of the piecewise function, it is less than 1. By amplifying the piecewise function values, the training intensity is enhanced, and simultaneously during training... Decrease, make Increase, and simultaneously improve and Similarity and and The similarity is further improved to enhance semantic accuracy. The meaning of the second piecewise function is similar to that of the first piecewise function, and will not be repeated here.

[0121] According to one embodiment of the present invention, a translation loss function can be obtained by weighted summation of two piecewise functions. During the training process, the translation loss function is continuously reduced, thereby continuously improving the semantic recognition accuracy and translation accuracy, and improving the accuracy of the cross-attention mechanism.

[0122] In this way, the segmented function can be set as the translation loss function, and when the cross attention mechanism does not play a positive role, the training intensity is improved, so as to improve the accuracy and semantic recognition accuracy of the cross attention mechanism, thereby improving the training efficiency and translation accuracy.

[0123] According to one embodiment of the application, the attention loss function is determined according to the first correlation weight, the second correlation weight and the third correlation weight, comprising: determining the attention loss function according to formula (2) ,

[0124] (2),

[0125] wherein, is the first correlation weight between the first text in the training theme information and the jth text in the training text, is the first correlation weight between the second text in the training theme information and the jth text in the training text, is the first correlation weight between the ith text in the training theme information and the jth text in the training text, is the first correlation weight between the mth text in the training theme information and the jth text in the training text, and m is the number of texts in the training theme information, is the number of texts in the training text, and j≤ , i≤m, is the annotation weight of the first text in the training theme information, is the annotation weight of the second text in the training theme information, is the annotation weight of the ith text in the training theme information, is the annotation weight of the mth text in the training theme information, is the second correlation weight between the first text in the training theme information and the kth first training semantic information, is the second correlation weight between the second text in the training theme information and the kth first training semantic information, is the second correlation weight between the ith text in the training theme information and the kth first training semantic information, is the second correlation weight between the mth text in the training theme information and the kth first training semantic information, is the number of first training semantic information, and k≤ , is the third correlation weight between the first text in the training theme information and the s th second training semantic information, is the third correlation weight between the second text in the training theme information and the s th second training semantic information, To train the third association weights between the i-th text and the s-th second training semantic information in the topic information, To train the third association weight between the m-th text and the s-th second training semantic information in the topic information, For the second training semantic information, s≤ , and j i, m, k ,s, All are positive integers. , , The preset weights are sim, which is the similarity function.

[0126] According to an embodiment of the present invention, during the fusion process, the fusion objects of the first cross-attention mechanism, the second cross-attention mechanism, and the third cross-attention mechanism are all training topic information. During the fusion process, the association weights (first association weight, second association weight, and third association weight) between each semantic vector and the semantic vector of each text in the training topic information are determined. The distribution of the association weights should be the same as or similar to the distribution of the relevance between the texts and the main idea in the training topic. That is, if a word in the training topic has a high relevance to the main idea (for example, it describes important information in the main idea), then the association weight should also be high. This allows important information to be fused during the fusion process and reduces the introduction of noise.

[0127] According to an embodiment of the present invention, in formula (2), It can be used as a distribution vector of the relevance between text and theme in the training topic. The higher the value, the higher the relevance of the i-th text to the main theme of the training topic information. This is the sum of the first association weights between the i-th text in the training topic information and each text in the training text. In other words, it's the fusion weight of the semantic vector of each text in the training text with respect to the semantic vector of the i-th text in the training topic information when fusing features from the training topic information. A higher fusion weight means a higher degree of feature fusion for the semantic vector of the i-th text during fusion; in other words, more features of the semantic vector of the i-th text are obtained. As mentioned above, the more relevant the i-th text in the training topic information is to the main theme of the training topic information, the higher the degree of feature fusion for the semantic vector of the i-th text should be. Therefore... and The higher the similarity between them, the better. In formula (2), it can be made As the first term of the attention loss function, thus during training, making Closer to 1, that is, make and The similarity between the two is improved, so that more features related to the theme are obtained in the fusion process, and the introduction of noise features unrelated to the theme is reduced. The meanings of the second and third terms of formula (2) are similar to the first term, which will not be described here. The three weighted sums are obtained, and the attention loss function is obtained, so that the attention loss function is reduced in the training process, so that more features related to the theme are obtained in the fusion process, and the introduction of noise features unrelated to the theme is reduced.

[0128] In this way, the attention loss function can be set by associating the weight and the annotation weight, so that the distribution of the association weight and the relevance of the text in the training theme to the theme can be closer in the training, so that more features related to the theme can be fused in the fusion process, and the introduction of noise features unrelated to the theme is reduced, the fusion accuracy is improved, and more accurate semantic information is provided for the translation process.

[0129] According to an embodiment of the present application, the translation loss function and the attention loss function are weighted and summed to obtain the loss function of the intelligent translation model, and the parameters of the intelligent translation model are adjusted by gradient descent method through back propagation, so that the loss function is reduced, and the training process is completed after multiple training, and the trained intelligent translation model is obtained. It can be used in the translation process of multiple national languages.

[0130] The IoA-based multi-agent cooperative cross-language system translation method according to the embodiment of the present application can translate the text information to be translated by multiple translation agents, and in the translation process, the translation accuracy of key text is improved by referring to the conference theme information. Moreover, the semantic consistency of multi-language translation is improved by cross-attention mechanism to mutually correct the semantic information of multiple languages, thereby improving the overall translation accuracy and reducing the communication cost. The first semantic vector corresponding to the first fusion semantic information can be fused with the feature information of the conference theme information by the first cross-attention mechanism, thereby enhancing the features of the text related to the conference theme information, improving the translation accuracy, and making the translation content more consistent with the conference theme, thereby reducing the probability of ambiguity. The semantic vectors corresponding to different target languages can be fused by the cross-attention mechanism and the self-attention mechanism, thereby making the fused semantic vectors not only carry the feature information related to the conference theme information, but also carry the feature information of the semantic vectors of different languages, so that the output translation text of different languages is not only consistent with the conference theme, but also highly consistent in semantics, thereby reducing the probability of ambiguity and reducing the translation error of different languages, and reducing the understanding difference of participants of different languages. In the training process, the segmentation function can be set as the translation loss function, and the training intensity can be improved when the cross-attention mechanism does not play a positive role, thereby improving the accuracy and semantic recognition precision of the cross-attention mechanism, thereby improving the training efficiency and translation accuracy. The attention loss function can be set by the correlation weight and the labeling weight, so that the correlation weight is more similar to the distribution of the relevance between the text and the main theme in the training theme, so that more features related to the main theme can be fused in the fusion process, and the introduction of noise features unrelated to the main theme is reduced, thereby improving the fusion accuracy and providing more accurate semantic information for the translation process.

[0131] Figure 3 An exemplary block diagram of an IoA-based multi-agent cooperative cross-language system translation system according to an embodiment of the present application is shown, which includes:

[0132] The acquisition module is configured to acquire text information to be translated and conference theme information.

[0133] The first fusion module is configured to determine first fusion semantic information of the text information to be translated and the conference theme information by the first cross-attention mechanism of the intelligent translation model.

[0134] The first target semantic information module is configured to process the first fusion semantic information by the first translation agent of the intelligent translation model to obtain first target semantic information.

[0135] The second target semantic information module is configured to process the first fused semantic information through a second translation intelligent agent of the intelligent translation model to obtain second target semantic information.

[0136] The second fusion module is configured to obtain second fused semantic information according to the first target semantic information, the conference theme information, and a second cross attention mechanism of the intelligent translation model.

[0137] The second target language module is configured to determine second target language translation text according to the second fused semantic information and the second target semantic information.

[0138] The third fusion module is configured to obtain third fused semantic information according to the second target semantic information, the conference theme information, and a third cross attention mechanism of the intelligent translation model.

[0139] The first target language module is configured to determine first target language translation text according to the third fused semantic information and the first target semantic information.

[0140] The translation text module is configured to output translation text according to the first target language translation text and the second target language translation text.

[0141] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein for executing various aspects of the present application.

[0142] Those skilled in the art should understand that the above-described embodiments of the present application shown in the description and drawings are only used as examples and do not limit the present application. The purpose of the present application has been fully and effectively achieved. The functional and structural principles of the present application have been shown and described in the embodiments, and the embodiments of the present application can be modified or changed without departing from the principles.

[0143] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-agent collaborative cross-language translation method based on IoAs, characterized in that, include: Obtain information about the text to be translated and the conference topic; The first cross-attention mechanism of the intelligent translation model is used to determine the first fused semantic information between the text information to be translated and the conference topic information; The first translation agent of the intelligent translation model processes the first fused semantic information to obtain the first target semantic information; The second translation agent of the intelligent translation model processes the first fused semantic information to obtain the second target semantic information. Based on the first target semantic information, the conference topic information, and the second cross-attention mechanism of the intelligent translation model, the second fused semantic information is obtained; Based on the second fused semantic information and the second target semantic information, the translated text in the second target language is determined; Based on the second target semantic information, the conference topic information, and the third cross-attention mechanism of the intelligent translation model, the third fused semantic information is obtained; Based on the third fused semantic information and the first target semantic information, the translated text in the first target language is determined; Output the translated text based on the translated text in the first target language and the translated text in the second target language.

2. The multi-agent collaborative cross-language translation method based on IoAs according to claim 1, characterized in that, Through the first cross-attention mechanism of the intelligent translation model, the first fused semantic information of the text information to be translated and the conference topic information is determined, including: Obtain the first semantic vector of multiple texts in the text information to be translated, and the second semantic vector of multiple texts in the conference topic information; Based on the first cross-attention mechanism, the first query matrix, the first key-value matrix, and the first weight matrix are obtained; The first semantic matrix of the text information to be translated is obtained through the first semantic vector; The first query information matrix is ​​obtained by multiplying the first semantic matrix with the first query matrix. The second semantic vector is used to obtain the second semantic matrix of the conference topic information; The first key-value information matrix is ​​obtained by multiplying the second semantic matrix with the first key-value matrix; The first weight information matrix is ​​obtained by multiplying the second semantic matrix by the first weight matrix; The first fused semantic matrix is ​​obtained by using the first query information matrix, the first key-value information matrix, and the first weight information matrix; Based on the first fusion semantic matrix, the first fusion semantic information is obtained.

3. The multi-agent collaborative cross-language translation method based on IoAs according to claim 1, characterized in that, Based on the second fused semantic information and the second target semantic information, the translated text in the second target language is determined, including: The third translation agent of the intelligent translation model processes the second fused semantic information to obtain the third target semantic information. Based on the second and third target semantic information, the third semantic matrix is ​​obtained; The third semantic matrix is ​​processed through the first self-attention mechanism of the intelligent translation model to obtain the second target semantic feature matrix, wherein the number of vectors in the second target semantic feature matrix is ​​the same as the number of second target semantic information. Based on the semantic feature matrix of the second target, the translated text in the second target language is determined.

4. The multi-agent collaborative cross-language translation method based on IoAs according to claim 1, characterized in that, Based on the third fused semantic information and the first target semantic information, the translated text for the first target language is determined, including: The fourth translation agent of the intelligent translation model processes the third fused semantic information to obtain the fourth target semantic information. Based on the semantic information of the fourth target and the semantic information of the first target, the fourth semantic matrix is ​​obtained; The fourth semantic matrix is ​​processed through the second self-attention mechanism of the intelligent translation model to obtain the first target semantic feature matrix, wherein the number of vectors in the first target semantic feature matrix is ​​the same as the number of first target semantic information. Based on the semantic feature matrix of the first target, the translated text in the first target language is determined.

5. The multi-agent collaborative cross-language translation method based on IoAs according to claim 1, characterized in that, The training method for the intelligent translation model includes: Acquire training text and training topic information, determine the first training fusion semantic information of training text and training topic information, and determine the first association weight between each text in the training text and each text in the training topic information; The first translation agent of the intelligent translation model processes the first training fused semantic information to obtain the first training semantic information, and the second translation agent of the intelligent translation model processes the first training fused semantic information to obtain the second training semantic information. The second cross-attention mechanism is used to determine the second training fusion semantic information of the first training semantic information and the training topic information, and to determine the second association weight of each text in the first training semantic information and the training topic information. The third cross-attention mechanism is used to determine the third training fusion semantic information of the second training semantic information and the training topic information, and to determine the third association weight of each text in the second training semantic information and the training topic information. The translation loss function is determined based on the second training semantic information, the second training fused semantic information, the annotated translated text of the second target language, the third training fused semantic information, the first training semantic information and the annotated translated text of the first target language; The attention loss function is determined based on the first association weight, the second association weight, and the third association weight; Based on the translation loss function and the attention loss function, determine the loss function of the intelligent translation model; The intelligent translation model is trained based on its loss function to obtain the trained intelligent translation model.

6. The multi-agent collaborative cross-language translation method based on IoAs according to claim 5, characterized in that, Based on the second training semantic information, the second training fused semantic information, the annotated translated text in the second target language, the third training fused semantic information, and the first training semantic information and the annotated translated text in the first target language, the translation loss function is determined, including: The second training semantic information is processed by the second semantic recognition model of the second target language to obtain the second semantic prediction information; The second semantic fusion information is obtained by processing the second training fused semantic information through the second semantic recognition model; The semantic vector of the annotated translated text in the second target language is processed by the second semantic recognition model to obtain the second semantic annotation information; The first training semantic information is processed by the first semantic recognition model of the first target language to obtain the first semantic prediction information; The first semantic fusion information is obtained by processing the third training fused semantic information through the first semantic recognition model; The semantic vector of the annotated and translated text in the first target language is processed by the first semantic recognition model to obtain the first semantic annotation information; The translation loss function is determined based on the second semantic prediction information, the second semantic fusion information, the second semantic annotation information, the first semantic prediction information, the first semantic fusion information, and the first semantic annotation information.

7. The multi-agent collaborative cross-language translation method based on IoAs according to claim 6, characterized in that, Based on the second semantic prediction information, the second semantic fusion information, the second semantic annotation information, the first semantic prediction information, the first semantic fusion information, and the first semantic annotation information, the translation loss function is determined, including: According to the formula ， Determine the translation loss function ,in, For the first semantic fusion information, This is the first semantic annotation information. This is the first semantic prediction information. For second semantic fusion information, This is the second semantic annotation information. The second semantic prediction information is represented by sim, where sim is the similarity function. and Preset weights.

8. The multi-agent collaborative cross-language translation method based on IoAs according to claim 5, characterized in that, Based on the first association weight, the second association weight, and the third association weight, the attention loss function is determined, including: According to the formula ， Determine the attention loss function ,in, To train the first association weight between the first text in the topic information and the j-th text in the training text, To train the first association weight between the second text in the topic information and the j-th text in the training text, To train the first association weight between the i-th text in the topic information and the j-th text in the training text, This is used to define the first association weight between the m-th text in the training topic information and the j-th text in the training text, where m is the number of texts in the training topic information. The number of texts in the training text, j≤ , i≤m, To train the labeled weights of the first text in the topic information, To train the annotation weights of the second text in the topic information, To train the annotation weights of the i-th text in the topic information, To train the annotation weights of the m-th text in the topic information, To train the second association weights between the first text and the kth first training semantic information in the topic information, To train the second association weight between the second text and the kth first training semantic information in the topic information, To train the second association weights between the i-th text and the k-th first training semantic information in the topic information, To train the second association weights between the m-th text and the k-th first training semantic information in the topic information, Let k ≤ 1. The number of semantic information units in the first training iteration. , To train the third association weight between the first text and the s-th second training semantic information in the topic information, To train the third association weight between the second text and the s-th second training semantic information in the topic information, To train the third association weights between the i-th text and the s-th second training semantic information in the topic information, To train the third association weight between the m-th text and the s-th second training semantic information in the topic information, For the second training semantic information, s≤ , and j i, m, k ,s, All are positive integers. , , The preset weights are sim, which is the similarity function.

9. A multi-agent collaborative cross-language translation system based on IoAs, characterized in that, include: The acquisition module is used to acquire the text to be translated and the meeting topic information; The first fusion module is used to determine the first fused semantic information of the text information to be translated and the conference topic information through the first cross-attention mechanism of the intelligent translation model; The first target semantic information module is used to process the first fused semantic information through the first translation agent of the intelligent translation model to obtain the first target semantic information; The second target semantic information module is used to process the first fused semantic information through the second translation agent of the intelligent translation model to obtain the second target semantic information; The second fusion module is used to obtain the second fused semantic information based on the first target semantic information, the conference topic information, and the second cross-attention mechanism of the intelligent translation model; The second target language module is used to determine the translated text in the second target language based on the second fused semantic information and the second target semantic information. The third fusion module is used to obtain the third fused semantic information based on the second target semantic information, the conference topic information, and the third cross-attention mechanism of the intelligent translation model; The first target language module is used to determine the translated text in the first target language based on the third fused semantic information and the first target semantic information; The Translated Text module is used to output translated text based on the translated text in the first target language and the translated text in the second target language.

Images