Place name and address multi-language translation system and method based on deep learning

Abstract

The invention relates to the field of language translation, and particularly discloses a geographical name address multi-language translation system and method based on deep learning, and the method comprises the steps: firstly obtaining a to-be-translated geographical name address input by a user, and carrying out the semantic analysis of the to-be-translated geographical name address through a language model based on a Transform architecture, so as to extract the semantic feature representation of the to-be-translated geographical name address; meanwhile, a proprietary term library is further introduced, a proprietary term set associated with the to-be-translated place name address is matched from the proprietary term library by using a fuzzy matching algorithm, and the semantic feature expression of the to-be-translated place name address is optimized by using proprietary terms, so that the translation efficiency of the to-be-translated place name address is improved based on the target translation language specified by the user. And translating by using the large language model to obtain a place name address translation result. By introducing the proprietary term library resource, the accuracy and semantic richness of place name address translation can be effectively improved, so that a more accurate and natural translation result is provided for a user.

Description

Technical Field

[0001] The present application relates to the field of language translation, and more specifically, to a multi-language translation system and method for place names and addresses based on deep learning. Background Art

[0002] As the globalization process accelerates, cross-border exchanges and cooperation become more frequent, and the demand for accurate and efficient multilingual translation of place names and addresses also increases. Whether it is international trade, travel or international logistics, it is inseparable from the accurate understanding and conversion of place names and addresses in different countries and regions.

[0003] Traditionally, the translation of place names and addresses mainly relies on pre-built bilingual or multilingual comparison tables or databases, which are realized by searching and matching. Although this method is simple and direct, it has poor flexibility and adaptability, especially when facing new place names or non-standard addresses, it is easy to make mistakes or fail to recognize them.

[0004] In addition, since place names and addresses may contain a large number of proper nouns, abbreviations, and specific cultural background information, these elements often do not have a direct one-to-one correspondence between different languages. Therefore, for place names and addresses with complex structures and deep cultural meanings, simply relying on literal translation cannot convey their true geographical location and semantic information, and it is difficult to provide high-quality translation results.

[0005] Therefore, there is a need for an optimized deep learning-based multilingual translation system and method for place names and addresses. Summary of the invention

[0006] In order to solve the above technical problems, this application is proposed.

[0007] According to one aspect of the present application, a method for translating place names and addresses into multiple languages ​​based on deep learning is provided, which includes: Acquire a place name and address to be translated input by a user, wherein the place name and address to be translated is described in a first language; Using a fuzzy matching algorithm to match a set of proprietary terms associated with the place name address to be translated from a proprietary term library; Semantically encoding the place name and address to be translated and each of the special terms in the set of special terms to obtain a set of semantic encoding vectors of the place name and address to be translated and a set of semantic encoding vectors of the special terms; Performing semantic query encoding based on fast matching alignment on the set of the semantic encoding vectors of the place name and address to be translated and the semantic encoding vectors of the proprietary terms to obtain a term-enhanced semantic encoding vector of the place name and address to be translated; Acquire a target translation language confirmed by a user, wherein the target translation language is a second language, and the second language is different from the first language; Based on the terminology-enhanced semantic encoding vector of the place name and address to be translated and the target translation language, a place name and address translation result is generated.

[0008] Specifically, the fuzzy matching algorithm is the Jaro-Winkler distance algorithm.

[0009] Specifically, semantic encoding is performed on the place name and address to be translated and each of the special terms in the set of special terms to obtain a set of semantic encoding vectors of the place name and address to be translated and a set of semantic encoding vectors of the special terms, including: The semantic encoding model based on the Transformer architecture is used to semantically encode the place name and address to be translated and each proper term in the set of proper terms to obtain a set of semantic encoding vectors of the place name and address to be translated and a set of semantic encoding vectors of the proper terms.

[0010] Specifically, the semantic query encoding based on fast matching alignment is performed on the set of the semantic encoding vector of the place name and address to be translated and the semantic encoding vector of the proprietary term to obtain the term enhanced semantic encoding vector of the place name and address to be translated, including: Determining a fine-grained matching search window for proprietary terms based on the semantic difference between the semantic encoding vector of the place name and address to be translated and each proprietary term semantic encoding vector in the set of proprietary term semantic encoding vectors; Fine-grained query response encoding based on the attention mechanism is performed on the semantic encoding vector of the place name and address to be translated and the semantic encoding vector of each proprietary term in the proprietary term fine-grained matching search window to obtain the term-enhanced semantic encoding vector of the place name and address to be translated.

[0011] Specifically, based on the semantic difference between the semantic encoding vector of the place name and address to be translated and each semantic encoding vector of the proprietary term in the set of semantic encoding vectors of the proprietary term, determining a proprietary term fine-grained matching search window includes: Calculating the cross entropy of the semantic encoding vector of the place name and address to be translated relative to each semantic encoding vector of the proprietary term in the set of semantic encoding vectors of the proprietary term to obtain a set of fast query positioning factors for the proprietary term; The semantic encoding vector of the proprietary term corresponding to the minimum value in the set of the proprietary term fast query positioning factors is used as the positioning matching proprietary term semantic encoding vector; Based on the located matching proper term semantic encoding vector, the proper term fine-grained matching search window is determined.

[0012] Specifically, the vector at the center of the proprietary term fine-grained matching search window is the positioning matching proprietary term semantic encoding vector, and each proprietary term semantic encoding vector in the proprietary term fine-grained matching search window is defined as a fine-grained query proprietary term semantic encoding vector.

[0013] Specifically, the semantic encoding vector of the place name and address to be translated and each semantic encoding vector of the proprietary term in the proprietary term fine-grained matching search window are subjected to fine-grained query response encoding based on an attention mechanism to obtain the term enhanced semantic encoding vector of the place name and address to be translated, including: The semantic encoding vector of the place name and address to be translated is linearly transformed to obtain a query vector and a value vector, and each fine-grained query proprietary term semantic encoding vector in the proprietary term fine-grained matching search window is used as a key vector, and the query vector, the value vector and the key vector are input into a fine-grained query encoding module based on a heterogeneous converter structure to obtain the term-enhanced semantic encoding vector of the place name and address to be translated.

[0014] Specifically, based on the terminology-enhanced semantic encoding vector of the place name and address to be translated and the target translation language, a place name and address translation result is generated, including: After the target translation language is added as a prompt word to the term-enhanced semantic encoding vector of the place name and address to be translated, it is input into a place name and address translator based on a large language model to obtain the place name and address translation result, and the place name and address translation result is described in a second language.

[0015] According to another aspect of the present application, a multilingual translation system for place names and addresses based on deep learning is provided, which includes: A place name and address acquisition module to be translated, used to acquire a place name and address to be translated input by a user, wherein the place name and address to be translated is described in a first language; A proprietary term matching module, used for matching a set of proprietary terms associated with the place name address to be translated from a proprietary term library using a fuzzy matching algorithm; A semantic encoding module, used for semantically encoding the place name and address to be translated and each of the special terms in the set of special terms to obtain a set of semantic encoding vectors of the place name and address to be translated and a set of semantic encoding vectors of the special terms; A proprietary term semantic query response encoding module, used for performing semantic query encoding based on fast matching alignment on the semantic encoding vector of the place name and address to be translated and the set of the proprietary term semantic encoding vector to obtain a term enhanced semantic encoding vector of the place name and address to be translated; a target translation language confirmation module, used to obtain a target translation language confirmed by a user, wherein the target translation language is a second language, and the second language is different from the first language; The place name and address translation result generating module is used to generate the place name and address translation result based on the terminology enhanced semantic encoding vector of the place name and address to be translated and the target translation language.

[0016] This application has at least the following technical effects: Compared with the prior art, the deep learning-based multilingual translation system and method for place name and address provided in the present application first obtains the place name and address to be translated input by the user, and uses a language model based on the Transformer architecture to perform semantic analysis on it to extract the semantic feature representation of the place name and address to be translated, and further introduces a proprietary terminology library, and uses a fuzzy matching algorithm to match the set of proprietary terms associated with the place name and address to be translated, so as to use the proprietary terms to optimize the semantic feature expression of the place name and address to be translated, and then on this basis, based on the target translation language specified by the user, uses a large language model for translation to obtain the place name and address translation result; by introducing proprietary terminology library resources, the present application can effectively improve the accuracy and semantic richness of place name and address translation, thereby providing users with more accurate and natural translation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0018] Figure 1 This is a flowchart of a method for multilingual translation of place names and addresses based on deep learning according to an embodiment of the present application.

[0019] Figure 2 This is a data flow diagram of a method for multi-language translation of place names and addresses based on deep learning according to an embodiment of the present application.

[0020] Figure 3 This is a flowchart of step S4 in the method for multilingual translation of place names and addresses based on deep learning according to an embodiment of the present application.

[0021] Figure 4 It is a block diagram of a multilingual translation system of place names and addresses based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0023] It should be noted that the acquisition and processing of all information or data in this application are carried out in compliance with the relevant national data protection laws and policies and with the authorization given by the owner of the corresponding device.

[0024] Figure 1 This is a flowchart of a method for multilingual translation of place names and addresses based on deep learning according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of a multi-language translation method of place names and addresses based on deep learning according to an embodiment of the present application. Figure 1 and Figure 2 As shown, the multilingual translation method of place names and addresses based on deep learning includes the following steps: S1, obtaining the place name and address to be translated input by the user, wherein the place name and address to be translated is described in a first language; S2, matching a set of proper terms associated with the place name and address to be translated from a proper term library using a fuzzy matching algorithm; S3, semantically encoding the place name and address to be translated and each proper term in the set of proper terms to obtain a semantic encoding vector of the place name and address to be translated and a set of proper term semantic encoding vectors; S4, performing semantic query encoding based on fast matching alignment on the semantic encoding vector of the place name and address to be translated and the set of proper term semantic encoding vectors to obtain a term-enhanced semantic encoding vector of the place name and address to be translated; S5, obtaining a target translation language confirmed by the user, wherein the target translation language is a second language, which is different from the first language; S6, generating a place name and address translation result based on the term-enhanced semantic encoding vector of the place name and address to be translated and the target translation language.

[0025] In the above-mentioned multilingual translation method of place names and addresses based on deep learning, the step S1 obtains the place name and address to be translated input by the user, and the place name and address to be translated is described in the first language. In an embodiment of the present application, the user inputs the place name and address to be translated in text form through an input device such as a keyboard, a touch screen or a voice recognition input.

[0026] Specifically, to ensure efficient and accurate reception of user input, first of all, the user interface (UI) is the first contact point of the entire process, which determines how users interact with the system and input the place name and address to be translated. In order to make this process as intuitive and convenient as possible, the system provides a variety of input methods for users to choose from: the text box allows users to directly type or paste the place name and address to be translated; the voice input function uses advanced voice recognition technology to support multiple mainstream languages ​​and convert the user's spoken language into accurate text information; the map selection method allows users to determine the place name and address to be translated by clicking or dragging a specific location on the electronic map. This design not only improves the user experience, but also provides flexibility for different types of users and meets the needs of users in different scenarios.

[0027] When the user completes the input, the system will immediately start the language detection algorithm to automatically identify the language type of the text provided by the user. This automated language detection can greatly simplify the user's operation steps, especially when the user is unsure or has not explicitly specified the input language. If the user has already specified the input language, the system's priority will be to follow the user's settings. In addition, for addresses that contain non-standard formats or special characters, the system will also attempt to standardize them, including removing extra spaces, correcting common spelling errors, etc., to ensure that the input data is neat and consistent, thereby improving the efficiency of subsequent semantic parsing and matching processes.

[0028] Considering that place names and addresses may contain a large number of proper nouns, abbreviations, and specific cultural background information, these elements often do not have a direct one-to-one correspondence between different languages. Therefore, the understanding of user input does not just stay at the literal meaning level, but also needs to be deeply analyzed in combination with context and other relevant information. For example, the address provided by the user may be incomplete or contain multiple possibilities. In this case, the system can enhance its understanding of the input content by asking the user to provide more information, or by using geographic information systems (GIS) and other external resources, such as historical search records, current geographic location, etc. At the same time, considering the need for cross-border place name and address translation, the system can also combine the user's historical interaction records or other related information sources, such as previously searched places, to help more accurately interpret the user's intentions and ensure that the translation results are not only accurate but also culturally relevant.

[0029] Throughout the entire process, security and privacy protection are always important links that cannot be ignored. When processing user input, the system must comply with strict security measures to protect user privacy. This includes but is not limited to encrypting and transmitting user input data, anonymizing personal sensitive information, etc. Especially in cases involving voice input, the system should ensure that the recording files will not be improperly stored or leaked to protect the user's personal information security. In addition, for user groups with special needs, such as the visually impaired, the system also provides corresponding auxiliary tools and technologies, such as screen reader support, to ensure that all users can enjoy the service equally.

[0030] When the user submits the place name and address to be translated, the system should promptly give a confirmation prompt and, if necessary, provide the option for the user to review or modify the input content to ensure that the user has the opportunity to correct any possible errors. At the same time, it also establishes a two-way communication bridge to improve the user experience. If the system detects any problems, such as unrecognizable characters or suspected incorrect input, it will provide the user with clear warning information and guidance on how to correct these problems. This instant feedback mechanism can not only improve user satisfaction, but also effectively reduce translation errors caused by misunderstandings.

[0031] Finally, in terms of compatibility and accessibility, the system design must take into account the characteristics of various devices and platforms to ensure that it can run smoothly on a variety of terminals such as desktop computers, tablets and smartphones. The system has been optimized and adjusted to meet the requirements of different screen sizes and operating systems to ensure the best display effect and user experience. At the same time, the system also pays special attention to user groups with special needs, such as the visually impaired, and provides them with additional support and services, such as screen reader adaptation, to ensure that they can use this function without obstacles.

[0032] In the above-mentioned multilingual translation method of place names and addresses based on deep learning, the step S2 uses a fuzzy matching algorithm to match a set of proper terms associated with the place name and address to be translated from a proper term library. Specifically, considering that the place name and address to be translated may contain some proper nouns and cultural background information, it is difficult for traditional translation methods to accurately understand their meanings. Therefore, the present application introduces a proper term library containing place names, address-related terms and their multiple language versions to improve the accuracy and cultural adaptability in the translation process. The proper term library not only contains the standard names of geographical entities such as countries, cities, streets, and buildings, but also covers the nicknames, abbreviations, and cultural-related terms of geographical entities in various cultural backgrounds, such as China's Hutongs and Japan's Machikai, etc., which provide rich contextual information for the translation of place names and addresses. On this basis, the present application uses a fuzzy matching algorithm to match a set of proper terms associated with the place name and address to be translated from the proper term library, thereby providing effective auxiliary information for subsequent translation of place names and addresses. In the embodiment of the present application, the Jaro-Winkler distance algorithm is used as the fuzzy matching algorithm to achieve fast matching of proprietary terms.

[0033] In the above-mentioned multilingual translation method of place names and addresses based on deep learning, the step S3 is to semantically encode the place names and addresses to be translated and each of the proper terms in the set of proper terms to obtain a set of semantic encoding vectors of the place names and addresses to be translated and a set of semantic encoding vectors of the proper terms. In a specific example of the present application, a semantic encoding model based on the Transformer architecture is used to semantically encode the place names and addresses to be translated and each of the proper terms in the set of proper terms to obtain a set of semantic encoding vectors of the place names and addresses to be translated and a set of semantic encoding vectors of the proper terms. Among them, the Transformer model has significant advantages in natural language tasks. It has accumulated extensive language knowledge by pre-training on large-scale text data, and can provide more accurate semantic representation for each word phrase in the place names and addresses to be translated, as well as each proper term in the set of proper terms. In addition, since the place name address to be translated consists of multiple words and phrases, it has a complex semantic association structure, and the Transformer model, based on the self-attention mechanism, can effectively capture the contextual semantic dependencies between each word or phrase in the place name address to be translated, and better understand the overall contextual meaning of the place name address to be translated, thereby providing a richer semantic representation for subsequent address translation.

[0034] In the above-mentioned multilingual translation method of place names and addresses based on deep learning, the step S4 performs semantic query encoding based on fast matching alignment on the set of semantic encoding vectors of the place names and addresses to be translated and the semantic encoding vectors of the proprietary terms to obtain the term-enhanced semantic encoding vectors of the place names and addresses to be translated. In the technical solution of the present application, in order to effectively utilize the semantic information of the proprietary terms to enhance the semantic expression of the place names and addresses to be translated, the present application further performs matching fusion processing on the semantic encoding vectors of the place names and addresses to be translated and the set of the semantic encoding vectors of the proprietary terms. In particular, in order to improve the efficiency of semantic matching, the present application proposes a semantic query encoding based on fast matching alignment, which performs semantic difference measurement and local context perception on the set of semantic encoding vectors of the place names and addresses to be translated and the semantic encoding vectors of the proprietary terms, thereby realizing the process from coarse positioning to fine-grained matching and then to deep semantic interactive encoding, so as to ensure that the semantic information of the proprietary terms is fully utilized in the subsequent translation process, and the original semantics and cultural background of the place names and addresses are better understood, thereby improving the translation quality.

[0035] Figure 3 FIG. 4 is a flowchart of step S4 in the method for multilingual translation of place names and addresses based on deep learning according to an embodiment of the present application. Figure 3 As shown, the step S4 includes: S41, determining a proper term fine-grained matching search window based on the semantic difference between the semantic coding vector of the place name and address to be translated and each proper term semantic coding vector in the set of proper term semantic coding vectors; S42, performing fine-grained query response encoding based on an attention mechanism on the semantic coding vector of the place name and address to be translated and each proper term semantic coding vector in the proper term fine-grained matching search window to obtain the term enhanced semantic coding vector of the place name and address to be translated.

[0036] Specifically, the step S41 includes: first, calculating the cross entropy of the semantic encoding vector of the place name address to be translated relative to each semantic encoding vector of the proprietary term in the set of semantic encoding vectors of the proprietary term to obtain a set of fast query positioning factors for the proprietary term, which is expressed as follows: ; ; ; in, represents the set of semantic encoding vectors of the proprietary terms, , , and Respectively represent the first, second, and third in the set of semantic encoding vectors of the proprietary term and semantic encoding vectors of specific terms, is the number of feature vectors in the set of semantic encoding vectors of the proprietary term, Indicates the The semantic encoding vector of the specific term The eigenvalues ​​at the positions, represents the semantic encoding vector of the place name to be translated, Indicates the first The eigenvalues ​​at the positions, represents the logarithmic function with base 2, , , and Respectively indicate the Relative to the , , and stated The cross entropy of The set of factors representing the fast query positioning of the proprietary term.

[0037] First, by calculating the cross entropy of the semantic encoding vector of the place name and address to be translated relative to the semantic encoding vector of each proprietary term, the semantic distance and mismatch between the two are quantified, and a set of proprietary term fast query positioning factors is generated to quickly narrow the search scope of the semantic query matching to the most likely relevant area. Among them, cross entropy, as a key indicator to measure the difference between two probability distributions, not only considers the direct semantic similarity between the semantic encoding vector of the place name and address to be translated and the semantic encoding vector of the proprietary term, but also implicitly captures the potential contextual relationship between the two, thereby providing a more accurate basis for subsequent fast query positioning.

[0038] Next, the semantic encoding vector of the proprietary term corresponding to the minimum value in the set of the proprietary term fast query positioning factors is used as the positioning matching proprietary term semantic encoding vector; then, based on the positioning matching proprietary term semantic encoding vector, the proprietary term fine-grained matching search window is determined. Among them, the vector at the center position of the proprietary term fine-grained matching search window is the positioning matching proprietary term semantic encoding vector, and each proprietary term semantic encoding vector in the proprietary term fine-grained matching search window is defined as the fine-grained query proprietary term semantic encoding vector, which is expressed by the formula: ; ; ; in, represents taking the semantic encoding vector of the proprietary term corresponding to the minimum value in the set of the proprietary term fast query positioning factors, Represents the semantic encoding vector of the positioning matching proprietary term, represents the position index of the semantic encoding vector of the positioned matching proprietary term in the set of the semantic encoding vectors of the proprietary term, represents the calculation of the norm of the vector, Indicates rounding down. represents the width of the fine-grained matching search window for the proprietary term, , , and querying a semantic encoding vector of a proprietary term for each fine-grained term in the proprietary term fine-grained matching search window, A set of semantic encoding vectors of the fine-grained query proprietary terms in the proprietary term fine-grained matching search window.

[0039] Specifically, the minimum value in the set of proprietary term quick query positioning factors calculated above is used to determine the best matching proprietary term semantic encoding vector, so as to ensure that the initially screened proprietary term semantic encoding vector has the highest semantic relevance to the place name address to be translated, and it is used as the positioning matching proprietary term semantic encoding vector, and a proprietary term fine-grained matching search window is delineated with it as the center, and each proprietary term semantic encoding vector in the window is designated as a fine-grained query proprietary term semantic encoding vector for the subsequent more refined local context perception and matching operations. In this way, resources can be concentrated on a smaller and more targeted data subset for in-depth analysis, which not only improves retrieval accuracy but also ensures computational efficiency, and enhances the model's ability to understand the local semantic patterns of proprietary terms.

[0040] Specifically, the step S42 includes: performing a linear transformation on the semantic encoding vector of the place name and address to be translated to obtain a query vector and a value vector, and using each fine-grained query proprietary term semantic encoding vector in the proprietary term fine-grained matching search window as a key vector, inputting the query vector, the value vector and the key vector into a fine-grained query encoding module based on a heterogeneous converter structure to obtain the term enhanced semantic encoding vector of the place name and address to be translated, which is expressed by the formula: ; ; ; in, and denote the query embedding matrix and the value embedding matrix respectively, and Represent different bias terms, and denote the query vector and value vector respectively, Indicates the first fine-grained query term semantic encoding vector, represents matrix multiplication, represents the soft maximization function, represents the transpose of a vector, Indicates the The feature scale value of the semantic encoding vector of the fine-grained query term, It represents the semantic encoding vector of the place name or address to be translated by the term enhancement.

[0041] Specifically, a linear transformation is performed on the semantic encoding vector of the place name and address to be translated to convert it into a query vector and a value vector suitable for the attention mechanism, and at the same time, the key vectors in the fine-grained matching search window of the proprietary terms are combined, and a fine-grained query encoding module based on a heterogeneous converter structure is used to simulate the attention interaction between the query, key and value. The semantic association between the place name and address to be translated and each proprietary term is processed in parallel through the multi-head self-attention layer in the fine-grained query encoding module to achieve deep semantic interaction fusion between the place name and address to be translated and each proprietary term, and generate a term-enhanced semantic encoding vector of the place name and address to be translated that integrates the semantic information of the place name and address to be translated and the proprietary terms, thereby providing a more accurate semantic context representation for subsequent translation.

[0042] In the above-mentioned multilingual translation method of place names and addresses based on deep learning, the step S5 obtains the target translation language confirmed by the user, and the target translation language is a second language, which is different from the first language. That is, by clarifying the target translation language, it is ensured that the translation process can correctly select and apply the corresponding translation rules and models, thereby accurately reflecting the original semantics and cultural background of the place name and address.

[0043] In order to help users easily confirm the target translation language, the system first provides an intuitive user interface (UI) on which users can select or enter the target language into which they want the place name and address to be translated. This interface usually includes a drop-down menu or scrolling list that lists all supported language options. Each language item is equipped with its corresponding flag icon or country / region name for users to quickly identify it. In addition, for users who are not familiar with foreign languages, the system can also use voice prompts to let users hear the pronunciation of the selected language, thereby increasing the accuracy of the selection.

[0044] In addition, to further enhance the user experience, the system uses machine learning algorithms to analyze the user's historical behavior patterns, predict and recommend personalized target language options. For example, if a user frequently translates Chinese addresses into English, the system will recommend English as the target language by default the next time the user performs a similar operation. Not only that, as users continue to use the service, the system will continue to optimize the recommendation logic, gradually adapt to the user's personal habits, and make the selection process more efficient.

[0045] Once the user selects the target translation language, the system will immediately display a brief confirmation message asking the user whether to confirm the current selection. If the user finds an error or changes his mind, he can go back and modify it at any time. This instant feedback mechanism not only helps to reduce the occurrence of misoperation. At the same time, in some cases, the system may ask the user to provide more information. For example, when two similar but different languages ​​​​(such as Simplified Chinese and Traditional Chinese) are detected, the system will prompt the user to clearly distinguish and ensure that the translation result meets expectations.

[0046] In the above-mentioned multilingual translation method of place names and addresses based on deep learning, the step S6 generates a place name and address translation result based on the term-enhanced semantic encoding vector of the place name and address to be translated and the target translation language. In a specific example of the present application, after the target translation language is added as a prompt word to the term-enhanced semantic encoding vector of the place name and address to be translated, it is input into a place name and address translator based on a large language model to obtain the place name and address translation result, and the place name and address translation result is described in a second language. That is, the term-enhanced semantic encoding vector of the place name and address to be translated is translated using a large language model to generate a place name and address translation result. In this process, in order to ensure that the large language model can clearly define the target translation language, the target translation language is first word-embedded and encoded, converted into a numerical form that the model can process, connected with the term-enhanced semantic encoding vector of the place name and address to be translated, and then input into the large language model for translation. In an embodiment of the present application, the large language model adopts an mBERT (Multilingual BERT) model, which is pre-trained in multiple languages ​​and can handle translation tasks between multiple languages, generating accurate and natural target language translation results. During the translation process, the model will take into account the specificity and cultural background of the proprietary terms in place names and addresses, ensuring that the translation results are not only linguistically accurate but also culturally appropriate. Ultimately, the translation results output by the translation system will be provided to the user, who can make further edits or use them directly based on the results. Through such a translation process, the accuracy and efficiency of place name and address translation can be effectively improved to meet the translation needs in different language environments.

[0047] In particular, when the set of the semantic encoding vector of the place name and address to be translated and the set of the semantic encoding vector of the proper terms respectively represent the semantic encoding features of the place name and address to be translated and the semantic encoding features of the proper terms associated with the place name and address to be translated, when encoding the semantic query based on fast matching alignment, the content differences in the expressions of the named entities in the proper terms and the place name and address to be translated will cause the semantic query encoding to obtain the term-enhanced semantic encoding vector of the place name and address to be translated to have a diversity imbalance in semantic reinforcement, resulting in differences in the spatial structure of the distribution of diversity semantic reinforcement, affecting the convergence of the place name and address translator input into the large language model, resulting in a decrease in the accuracy of the place name and address translation results.

[0048] Therefore, in a preferred example of the present application, the semantic encoding vector of the term-enhanced place name and address to be translated is optimized and then input into the large language model for place name and address translation; specifically, the optimization process includes: First, the absolute value sum and the square root of the sum of squares of all eigenvalues ​​of the term-enhanced semantic coding vector of the place name and address to be translated are calculated to obtain the first term-enhanced semantic coding spatial structure value of the place name and address to be translated and the second term-enhanced semantic coding spatial structure value of the place name and address to be translated, that is: ; ; in, Indicates that the term enhances the semantic encoding vector of the place name to be translated The eigenvalues ​​at the positions, Indicates the spatial structure value of the semantic encoding of the place name to be translated by the first term. Indicates that the second term enhances the semantic encoding spatial structure value of the place name address to be translated; Next, the number of eigenvalues ​​of all eigenvalues ​​of the semantic encoding vector of the term-enhanced place-name and address to be translated is determined, that is, the length of the semantic encoding vector of the term-enhanced place-name and address to be translated ; Then, for each eigenvalue of the semantic encoding vector of the term-enhanced place-name and address to be translated, the long-range dependency value of the semantic encoding of the first term-enhanced place-name and address to be translated obtained by subtracting the product of the eigenvalue and the number of eigenvalues ​​from the spatial structure value of the semantic encoding of the first term-enhanced place-name and address to be translated is calculated, that is: ; in, For the The corresponding first term enhances the semantic encoding long-range dependency value of the place name and address to be translated; At the same time, the second term enhanced semantic encoding long-range dependency value of the place name and address to be translated is obtained by calculating the product of the square root of the number of eigenvalues ​​multiplied by the eigenvalue minus the second term enhanced semantic encoding spatial structure value of the place name and address to be translated, that is: ; in, For the The corresponding second term enhances the semantic encoding long-range dependency value of the place name and address to be translated; Secondly, the exponential value calculated by taking the first term enhanced long-range dependency value of the semantic coding of the place name and address to be translated as the exponent of the natural constant and the inverse of the second term enhanced long-range dependency value of the semantic coding of the place name and address to be translated are weighted summed to obtain the optimized feature value corresponding to each feature value, that is: ; in, For the The corresponding optimized eigenvalue is and are different weighting parameters and can be set as needed; Finally, the optimized feature values ​​are combined into an optimized terminology-enhanced semantic encoding vector of the place name and address to be translated.

[0049] That is, in view of the possible spatial structure loss in the high-dimensional space of the feature set of the term-enhanced semantic coding vector of the place-name and address to be translated, which causes the weight of the translation decoder to implicitly infer the spatial structure information based on the features, resulting in inconsistent convergence. By establishing a long-distance feature dependency relationship based on the spatial structure representation of the term-enhanced semantic coding vector of the place-name and address to be translated relative to the overall feature scale of the term-enhanced semantic coding vector of the place-name and address to be translated, and then establishing the feature local connectivity of the term-enhanced semantic coding vector of the place-name and address to be translated, and capturing the spatial ambiguous information of the object feature value through the unstructured feature value point prediction of the term-enhanced semantic coding vector of the place-name and address to be translated, the semantic space inductive bias perception ability of the feature set of the term-enhanced semantic coding vector of the place-name and address to be translated is improved, thereby improving the convergence of the place-name and address translator input into the large language model and thus improving the accuracy of the place-name and address translation result.

[0050] In summary, based on the present application, the place name and address to be translated input by the user is first obtained, and a language model based on the Transformer architecture is used to perform semantic analysis on it to extract the semantic feature representation of the place name and address to be translated, and at the same time, a proprietary terminology library is further introduced, and a fuzzy matching algorithm is used to match a set of proprietary terms associated with the place name and address to be translated, so as to use the proprietary terms to optimize the semantic feature expression of the place name and address to be translated, and then on this basis, based on the target translation language specified by the user, a large language model is used for translation to obtain the place name and address translation result. By introducing proprietary terminology library resources, the accuracy and semantic richness of place name and address translation can be effectively improved, thereby providing users with more accurate and natural translation results.

[0051] Furthermore, the present application also provides a multilingual translation system for place names and addresses based on deep learning, which includes a client and a management end. Through microservice design, the scalability and high performance of the system are guaranteed. The system has functions such as multilingual fully automatic translation, batch processing and progress management, translation model management, knowledge base and function library support. In addition, the tool can also be embedded in third-party applications, supporting custom access control and permission management to ensure data security.

[0052] Specifically, the multilingual translation system for place names and addresses consists of a user layer, a service layer, a database layer, a general layer, and a network layer, and deeply integrates deep learning technology to improve translation quality and efficiency. The user layer provides an intuitive interface through the PC client and the management terminal, which is convenient for submitting requests and monitoring progress. The service layer is the core, using deep learning models based on architectures such as Transformer to process language conversion, and supports fine-tuning of data sets in specific fields to enhance accuracy. The database layer stores bilingual texts, special terminology libraries, and historical records, providing a basis for model training. The general layer uses the PyTorch framework to accelerate algorithm iteration, and combines Docker containerization with SpringCloud to achieve efficient resource utilization and service expansion. The overall architecture ensures high-quality instant translation, while having the ability to evolve and continuously optimize the user experience. In terms of security, the system adopts role-based access control and comprehensive log management mechanisms. This design takes into account both innovation and practicality, and is suitable for scenarios that require efficient cross-language communication.

[0053] In terms of translation model selection, the system uses a machine translation model based on the Transformer architecture and handles multilingual translation tasks through pre-trained models. In order to improve the professionalism of translation in specific fields, such as place name and address translation, the system supports fine-tuning of the model and integrates a professional terminology library.

[0054] The system also integrates knowledge base support, and supports the translation of domain-specific terms by integrating industry terminology libraries and custom vocabularies. Common place names, proper nouns, and vocabulary for specific fields can be defined in the knowledge base. Through close integration with the translation model, the system can ensure the consistency of terminology and improve translation accuracy. At the same time, administrators can edit, update and maintain the knowledge base content at any time. The knowledge base supports dynamic calls during translation to ensure that the latest terminology can be applied to translation tasks in a timely manner. At the same time, the system also provides a rich library of predefined functions for processing text preprocessing, post-processing, data formatting and other operations. Users can call these functions during the translation process to achieve flexible configuration, such as custom processing rules, text format requirements for specific fields, etc. The function library also allows developers to define new translation logic and rules, thereby achieving highly customized translation processes.

[0055] The system allows developers to embed translation functions into existing applications through open APIs and SDKs. The system provides a multi-language support interface, and developers can implement complex translation tasks by simply calling the API. Developers can integrate translation functions by embedding the system API. The API supports functions such as text translation, batch translation, and model selection. In addition, the system supports integrating translation functions into business systems such as CRM and ERP, and low-code development is achieved by calling the SDK provided by the system, which simplifies the integration process and reduces development costs.

[0056] The system provides a plug-in mechanism that allows developers to extend existing functions or develop specific plug-ins for specific needs (such as legal translation, medical document translation, etc.). Users can also expand model support for different fields through plug-ins.

[0057] Figure 4 FIG. 1 is a block diagram of a multilingual translation system for place names and addresses based on deep learning according to an embodiment of the present application. Figure 4As shown, according to the multilingual translation system 100 of place names and addresses based on deep learning of the embodiment of the present application, it includes: a place name and address acquisition module 110 to be translated, which is used to obtain the place name and address to be translated input by the user, and the place name and address to be translated is described in the first language; a special term matching module 120, which is used to match the set of special terms associated with the place name and address to be translated from the special term library by using a fuzzy matching algorithm; a semantic encoding module 130, which is used to semantically encode the place name and address to be translated and each special term in the set of special terms to obtain a semantic encoding vector of the place name and address to be translated and a semantic encoding vector of the special term. a collection of; a proprietary term semantic query response encoding module 140, used to perform semantic query encoding based on fast matching alignment on the semantic encoding vector of the place name and address to be translated and the collection of the proprietary term semantic encoding vector to obtain the term enhanced semantic encoding vector of the place name and address to be translated; a target translation language confirmation module 150, used to obtain the target translation language confirmed by the user, the target translation language is a second language, and the second language is different from the first language; a place name and address translation result generation module 160, used to generate a place name and address translation result based on the term enhanced semantic encoding vector of the place name and address to be translated and the target translation language.

[0058] The specific operations of each module in the above-mentioned deep learning-based multilingual translation system for place names and addresses have been referenced above. Figures 1 to 3 The description of the deep learning-based multilingual translation method of place name addresses has been introduced in detail, and therefore, its repeated description will be omitted.

[0059] The basic principle of the present invention is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the present invention to be implemented by adopting the above specific details.

Claims

1. A multi-language translation method for place names and addresses based on deep learning, characterized in that: include: Acquire a place name and address to be translated input by a user, wherein the place name and address to be translated is described in a first language; Using a fuzzy matching algorithm to match a set of proprietary terms associated with the place name address to be translated from a proprietary term library; Semantically encoding the place name and address to be translated and each of the special terms in the set of special terms to obtain a set of semantic encoding vectors of the place name and address to be translated and a set of semantic encoding vectors of the special terms; Performing semantic query encoding based on fast matching alignment on the set of the semantic encoding vectors of the place name and address to be translated and the semantic encoding vectors of the proprietary terms to obtain a term-enhanced semantic encoding vector of the place name and address to be translated; Acquire a target translation language confirmed by a user, wherein the target translation language is a second language, and the second language is different from the first language; Based on the terminology-enhanced semantic encoding vector of the place name and address to be translated and the target translation language, a place name and address translation result is generated.

2. The method for multilingual translation of place names and addresses based on deep learning according to claim 1 is characterized in that: The fuzzy matching algorithm is the Jaro-Winkler distance algorithm.

3. The method for multilingual translation of place names and addresses based on deep learning according to claim 2 is characterized in that: Semantically encoding the place name and address to be translated and each of the special terms in the set of special terms to obtain a set of semantic encoding vectors of the place name and address to be translated and a set of semantic encoding vectors of the special terms, including: The semantic encoding model based on the Transformer architecture is used to semantically encode the place name and address to be translated and each proper term in the set of proper terms to obtain a set of semantic encoding vectors of the place name and address to be translated and a set of semantic encoding vectors of the proper terms.

4. The method for multilingual translation of place names and addresses based on deep learning according to claim 3 is characterized in that: The method of performing semantic query encoding based on fast matching alignment on the set of the semantic encoding vector of the place name and address to be translated and the semantic encoding vector of the proprietary term to obtain the semantic encoding vector of the place name and address to be translated with term enhancement includes: Determining a fine-grained matching search window for proprietary terms based on the semantic difference between the semantic encoding vector of the place name and address to be translated and each proprietary term semantic encoding vector in the set of proprietary term semantic encoding vectors; Fine-grained query response encoding based on the attention mechanism is performed on the semantic encoding vector of the place name and address to be translated and the semantic encoding vector of each proprietary term in the proprietary term fine-grained matching search window to obtain the term-enhanced semantic encoding vector of the place name and address to be translated.

5. The method for multilingual translation of place names and addresses based on deep learning according to claim 4 is characterized in that: Determining a fine-grained matching search window for proprietary terms based on the semantic difference between the semantic encoding vector of the place name and address to be translated and each proprietary term semantic encoding vector in the set of proprietary term semantic encoding vectors includes: Calculating the cross entropy of the semantic encoding vector of the place name and address to be translated relative to each semantic encoding vector of the proprietary term in the set of semantic encoding vectors of the proprietary term to obtain a set of fast query positioning factors for the proprietary term; The semantic encoding vector of the proprietary term corresponding to the minimum value in the set of the proprietary term fast query positioning factors is used as the positioning matching proprietary term semantic encoding vector; Based on the located matching proper term semantic encoding vector, the proper term fine-grained matching search window is determined.

6. The method for multilingual translation of place names and addresses based on deep learning according to claim 5, characterized in that: The vector at the center of the proprietary term fine-grained matching search window is the positioning matching proprietary term semantic encoding vector, and each proprietary term semantic encoding vector in the proprietary term fine-grained matching search window is defined as a fine-grained query proprietary term semantic encoding vector.

7. The method for multilingual translation of place names and addresses based on deep learning according to claim 6, characterized in that: The semantic encoding vector of the place name and address to be translated and each semantic encoding vector of the proprietary term in the proprietary term fine-grained matching search window are subjected to fine-grained query response encoding based on an attention mechanism to obtain the term enhanced semantic encoding vector of the place name and address to be translated, including: The semantic encoding vector of the place name and address to be translated is linearly transformed to obtain a query vector and a value vector, and each fine-grained query proprietary term semantic encoding vector in the proprietary term fine-grained matching search window is used as a key vector, and the query vector, the value vector and the key vector are input into a fine-grained query encoding module based on a heterogeneous converter structure to obtain the term-enhanced semantic encoding vector of the place name and address to be translated.

8. The method for multilingual translation of place names and addresses based on deep learning according to claim 7, characterized in that: Generating a place name and address translation result based on the terminology-enhanced place name and address semantic encoding vector to be translated and the target translation language, including: After the target translation language is added as a prompt word to the term-enhanced semantic encoding vector of the place name and address to be translated, it is input into a place name and address translator based on a large language model to obtain the place name and address translation result, and the place name and address translation result is described in a second language.

9. A multilingual translation system for place names and addresses based on deep learning, characterized in that: include: A place name and address acquisition module to be translated, used to acquire a place name and address to be translated input by a user, wherein the place name and address to be translated is described in a first language; A proprietary term matching module, used for matching a set of proprietary terms associated with the place name address to be translated from a proprietary term library using a fuzzy matching algorithm; A semantic encoding module, used for semantically encoding the place name and address to be translated and each of the special terms in the set of special terms to obtain a set of semantic encoding vectors of the place name and address to be translated and a set of semantic encoding vectors of the special terms; A proprietary term semantic query response encoding module, used for performing semantic query encoding based on fast matching alignment on the semantic encoding vector of the place name and address to be translated and the set of the proprietary term semantic encoding vector to obtain a term enhanced semantic encoding vector of the place name and address to be translated; a target translation language confirmation module, used to obtain a target translation language confirmed by a user, wherein the target translation language is a second language, and the second language is different from the first language; The place name and address translation result generating module is used to generate the place name and address translation result based on the terminology enhanced semantic encoding vector of the place name and address to be translated and the target translation language.

Images