Real-time Multilingual Translation System and Method for Geographical Names and Addresses Based on Neural Network Model

Abstract

The present application discloses a real-time multilingual translation system and method for place names and addresses based on a neural network model, which relates to the field of multilingual translation technology. The system first extracts the topological relationship of address components through structured analysis, and then uses a neural network model based on deep learning to deeply model the contextual dependencies across components, so as to mine the deep semantic representation of the address text through semantic association learning across components, and based on this, realizes place name disambiguation based on dynamic semantic reasoning; the present application solves the problem of semantic deviation in complex scenarios such as the same name in different places and the same place with different names, significantly improves the accuracy of distinguishing the polysemy of geographic entities, enables the translation results to dynamically adapt to the place name expression habits of different languages, effectively eliminates translation ambiguity caused by the limitations of static knowledge bases or imperfect rules, and achieves a more accurate multilingual address conversion effect.

Description

Technical Field

[0001] The present application relates to the technical field of multilingual translation, and more specifically, to a real-time multilingual translation system and method for geographical names and addresses based on a neural network model. Background Art

[0002] In the process of multilingual translation of geographical names and addresses, geographical name ambiguity is a very core and common problem. Existing technologies usually rely on rule matching, static geographical knowledge bases, or shallow semantic analysis to handle geographical name ambiguity problems. However, these methods have significant limitations in practical applications. Traditional rule matching methods are difficult to adapt to complex and changing context scenarios. Especially when geographical names have multiple meanings, relying solely on keyword matching easily leads to mistranslation, which directly results in inaccurate translation and affects users' ability to obtain correct information.

[0003] In addition, the existing technologies' utilization of address context information stays at the level of shallow feature matching and fails to establish a cross-component semantic reasoning mechanism. As a result, when facing complex situations such as the same name in different places or different names for the same place, the translation results are prone to semantic deviation. These problems together lead to bottlenecks in the accuracy of eliminating geographical name ambiguity and multilingual adaptability of existing technologies.

[0004] Therefore, an optimized real-time multilingual translation method for geographical names and addresses based on a neural network model is needed to solve the above technical problems. Summary of the Invention

[0005] According to one aspect of the present application, there is provided a real-time multilingual translation method for geographical names and addresses based on a neural network model, which includes:

[0006] Obtain the address text input by the user;

[0007] Use an address parsing model to structurally parse the address text to obtain a set of address components;

[0008] Select the address components to be strengthened from the set of address components and use the other address components in the set of address components as the context information of the address components to be strengthened;

[0009] Perform semantic embedding encoding on the address components to be strengthened to obtain the semantic embedding encoding features of the address components to be strengthened, and perform semantic enhancement on the semantic embedding encoding features of the address components to be strengthened based on the context information of the address components to be strengthened to obtain the semantic strengthened encoding features of the address components to be strengthened;

[0010] Based on the semantic strengthened encoding features of the address components to be strengthened, obtain the target language address text.

[0011] According to another aspect of the present application, there is provided a real-time multilingual translation system for geographical names and addresses based on a neural network model, which includes:

[0012] An address text acquisition module for acquiring the address text input by the user;

[0013] An address parsing module for structurally parsing the address text using an address parsing model to obtain a set of address components;

[0014] A semantic encoding module for respectively performing semantic encoding on the address components to be enhanced and the context information of the address components to be enhanced in the set of address components to obtain the semantic embedding encoding features of the address components to be enhanced and the semantic encoding features of the context information of the address components to be enhanced;

[0015] A semantic enhancement module for enhancing the semantic embedding encoding features of the address components to be enhanced based on the semantic encoding features of the context information of the address components to be enhanced to obtain the semantic enhanced encoding features of the address components to be enhanced;

[0016] A target language address text generation module for obtaining the target language address text based on the semantic enhanced encoding features of the address components to be enhanced.

[0017] Advantages of the present application: Compared with the prior art, the present application first extracts the topological relationship of address components through structural parsing, and then uses a neural network model based on deep learning to deeply model the cross-component context dependence, so as to mine the deep semantic representation of the address text through cross-component semantic association learning, and achieve geographical name disambiguation based on dynamic semantic reasoning. The present application solves the problem of semantic deviation in complex scenarios such as the same name for different places and different names for the same place, significantly improves the discrimination accuracy of the polysemy of geographical entities, enables the translation result to dynamically adapt to the naming expression habits of different languages, effectively eliminates translation ambiguities caused by the limitations of static knowledge bases or imperfect rules, and achieves a more accurate multilingual address conversion effect. Description of the Drawings

[0018] By describing the embodiments of the present application in more detail in conjunction with the drawings, the above and other objects, features and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 It is a flowchart of a real-time multilingual translation method for geographical names and addresses based on a neural network model according to an embodiment of the present application;

[0020] Figure 2Schematic diagram of data flow for the real-time multilingual translation method of geographical names and addresses based on a neural network model according to an embodiment of the present application;

[0021] Figure 3 Flowchart of sub-step S4 of the real-time multilingual translation method of geographical names and addresses based on a neural network model according to an embodiment of the present application;

[0022] Figure 4 Flowchart of sub-step S42 of the real-time multilingual translation method of geographical names and addresses based on a neural network model according to an embodiment of the present application;

[0023] Figure 5 Block diagram of the real-time multilingual translation system of geographical names and addresses based on a neural network model according to an embodiment of the present application. Detailed implementation manners

[0024] Next, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0025] As shown in the present application and the claims, unless the context clearly indicates an exception, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0026] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or the server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0027] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the operations before or below do not necessarily need to be precisely executed in sequence. On the contrary, as needed, various steps can be executed in reverse order or simultaneously. At the same time, other operations can also be added to these processes, or one or several steps can be removed from these processes.

[0028] Next, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0029] It should be noted in advance that the acquisition and processing of all information or data in this application are carried out on the premise of complying with the corresponding national data protection regulations and policies and obtaining the authorization from the authority manager.

[0030] In the technical solution of this application, a real-time multi-language translation method for geographical names and addresses based on a neural network model is proposed. Figure 1 It is a flowchart of the real-time multi-language translation method for geographical names and addresses based on a neural network model according to an embodiment of this application. Figure 2 It is a schematic diagram of data flow of the real-time multi-language translation method for geographical names and addresses based on a neural network model according to an embodiment of this application. As Figure 1 and Figure 2 shown, the real-time multi-language translation method for geographical names and addresses based on a neural network model according to an embodiment of this application includes the steps: S1, obtaining the address text input by the user; S2, using an address parsing model to perform structured parsing on the address text to obtain a set of address components; S3, selecting the address component to be enhanced from the set of address components and using the other address components in the set of address components as the context information of the address component to be enhanced; S4, performing semantic embedding encoding on the address component to be enhanced to obtain the semantic embedding encoding feature of the address component to be enhanced, and performing semantic enhancement on the semantic embedding encoding feature of the address component to be enhanced based on the context information of the address component to be enhanced to obtain the semantic enhanced encoding feature of the address component to be enhanced; S5, obtaining the target language address text based on the semantic enhanced encoding feature of the address component to be enhanced.

[0031] Specifically, for the S1 and S2, they obtain the address text input by the user and use an address parsing model to perform structured parsing on the address text to obtain a set of address components. Specifically, the original address text input by the user often exhibits non-linear arrangement characteristics, and may contain nested expressions, mixed language symbols, or regional abbreviation rules. These non-standard expression forms directly hinder the effectiveness of subsequent semantic analysis and ambiguity elimination. By performing structured parsing on the address text, continuous natural language strings can be segmented into address components with clear type labels (such as country, province, city, street, house number, etc.), so that the system can establish the spatial logical relationship and semantic dependence path between the components. It is worth mentioning that the address parsing model usually adopts a hybrid neural network architecture that integrates the multi-head attention mechanism and the conditional random field. In one example, the model first performs multi-granularity word segmentation on the address text input by the user. Among them, the system can adaptively adjust the word segmentation strategy according to different language characteristics (for example, Chinese uses BERT-based character-level processing, and English uses sub-word segmentation); then, the boundary features and type features of the address components are captured through a bidirectional Transformer encoder; in addition, the model integrates a geographical entity recognition module, which can distinguish ordinary words from geographical entities and dynamically link to a geographical knowledge graph to verify the components. Specifically, when there are format errors or component missing in the user input, the model will learn robust features through synthetic data generated by adversarial training, automatically complete the necessary components or trigger an interactive verification mechanism. The set of address components generated by structured parsing not only provides a discretized processing unit for subsequent semantic enhancement encoding, but also constructs a cross-component semantic inference channel.

[0032] Specifically, for the S3, it selects the address component to be enhanced from the set of address components and uses the other address components in the set of address components as the context information of the address component to be enhanced. Specifically, the semantics of each component in the address text does not exist in isolation, but jointly constitutes a complete spatial semantic network through the topological relationship with other components. The traditional method of treating address components as a parallel set makes it difficult to establish this cross-level semantic dependence relationship. In the technical solution of this application, the i-th address component is extracted from the set of address components as the address component to be enhanced, and the other address components in the set of address components are used as the context information of the address component to be enhanced. That is, by dynamically constructing an "object component - context" association framework, the neural network can flexibly capture the cross-level semantic dependence relationship for each address component, thereby providing a computable inference unit for subsequent semantic enhancement and ambiguity resolution.

[0033] Specifically, in step S4, semantic embedding encoding is performed on the address component to be enhanced to obtain the semantic embedding encoding feature of the address component to be enhanced, and semantic enhancement is performed on the semantic embedding encoding feature of the address component to be enhanced based on the context information of the address component to be enhanced, so as to obtain the semantic enhanced encoding feature of the address component to be enhanced. In a specific example of the present application, as Figure 3 shown, step S4 includes: S41, using the embedding matrix of the address component to be enhanced to perform semantic embedding encoding on the address component to be enhanced, so as to obtain the semantic embedding encoding vector of the address component to be enhanced and use it as the semantic embedding encoding feature of the address component to be enhanced; S42, performing semantic enhancement on the semantic embedding encoding feature of the address component to be enhanced based on the context information of the address component to be enhanced, so as to obtain the semantic enhanced encoding feature of the address component to be enhanced.

[0034] Specifically, in step S41, the embedding matrix of the address component to be enhanced is used to perform semantic embedding encoding on the address component to be enhanced, so as to obtain the semantic embedding encoding vector of the address component to be enhanced and use it as the semantic embedding encoding feature of the address component to be enhanced. Considering that when the traditional method directly inputs the address component in the form of a string into the translation model, it is unable to effectively capture its potential multi-level semantic features. Through semantic embedding encoding, such geographical entities can be mapped to a high-dimensional vector space, so that they form a distributed representation associated with administrative divisions and geographical levels in the semantic space. Therefore, in the technical solution of the present application, the embedding matrix of the address component to be enhanced is used to perform semantic embedding encoding on the address component to be enhanced, so as to obtain the semantic embedding encoding vector of the address component to be enhanced and use it as the semantic embedding encoding feature of the address component to be enhanced. Among them, by mapping the discrete address components to the continuous vector space defined by the pre-trained embedding matrix of the address component to be enhanced, the system can capture its implicit geographical attributes, functional classifications, and cultural association features. In an example, first, an embedding matrix of the address component to be enhanced is pre-trained based on a large amount of multilingual address corpora, and each row vector of the embedding matrix of the address component to be enhanced corresponds to the projection of a specific address component in the multi-level semantic space. When processing the address component to be enhanced (such as "Chaoyang District"), the system obtains its basic embedding vector through a matrix lookup operation, and then superimposes the attribute embedding based on the geographical knowledge graph (such as administrative division level, population density characteristics) and the context-aware position encoding (such as the strong association between "Chaoyang District" and business district characteristics in Beijing addresses). The semantic embedding encoding vector of the address component to be enhanced generated by semantic embedding encoding not only carries the lexical-level information of the address component, but also contains its spatial level, functional type, and cross-language alignment clues. This deep semantic encoding mechanism provides a feature basis with high information density for subsequent context semantic enhancement, significantly improving the system's ability to resolve complex place name ambiguities.

[0035] Specifically, in step S42, semantic enhancement is performed on the semantic embedding coding features of the address component to be enhanced based on the context information of the address component to be enhanced, so as to obtain the semantic enhanced coding features of the address component to be enhanced. In a specific example of the present application, as Figure 4 shown, step S42 includes: S421, performing semantic coding on the context information of the address component to be enhanced to obtain a semantic coding vector of the context information of the address component to be enhanced and using it as the semantic coding feature of the context information of the address component to be enhanced; S422, based on the semantic coding feature of the context information of the address component to be enhanced, performing semantic enhancement on the semantic embedding coding features of the address component to be enhanced to obtain the semantic enhanced coding features of the address component to be enhanced.

[0036] More specifically, in step S421, semantic coding is performed on the context information of the address component to be enhanced to obtain a semantic coding vector of the context information of the address component to be enhanced and using it as the semantic coding feature of the context information of the address component to be enhanced. In the embodiment of the present application, the specific steps of performing semantic coding on the context information of the address component to be enhanced are as follows: performing semantic embedding coding on each address component in the context information of the address component to be enhanced to obtain a context sequence of the semantic embedding coding vectors of the place name components; performing context semantic coding based on a bidirectional LSTM model on the context sequence of the semantic embedding coding vectors of the place name components to obtain a semantic coding vector of the context information of the address component to be enhanced. Specifically, the context information of the address text is not simply a set of adjacent components, but contains complex spatial topological logic and geographical subordination relationships. Traditional unidirectional recurrent neural networks can only capture dependencies in a single direction and cannot simultaneously perceive the jurisdiction and constraint of the superior administrative region over the subordinate place names and the reverse verification of the subordinate components on the superior semantics, resulting in a directional bias in semantic modeling. Therefore, in the technical solution of the present application, context semantic coding based on a bidirectional LSTM model is performed on the context information of the address component to be enhanced to obtain a semantic coding vector of the context information of the address component to be enhanced. That is, through bidirectional temporal modeling, the bidirectional influence of context components in the spatial hierarchy and syntactic structure is dynamically fused to construct a context semantic representation with global perception ability. Specifically, the bidirectional LSTM captures the progressive spatial hierarchy of address components (such as "province → city → district → street") through forward propagation, and at the same time extracts the reverse geographical subordination features (such as "street ← district ← city ← province") through backward propagation. This bidirectional information fusion enables the model to establish a geographical hierarchy attenuation relationship, achieve accurate geographical location disambiguation, and fundamentally break through the dependence limitation of traditional methods on local window information.

[0037] More specifically, the S422, based on the semantic coding features of the context information of the address component to be strengthened, performs semantic enhancement on the semantic embedding coding features of the address component to be strengthened, so as to obtain the semantic enhancement coding features of the address component to be strengthened. Specifically, when the traditional method simply concatenates or weighted sums the context features with the target component features, it is unable to effectively capture this implicit association across levels and dimensions, resulting in the loss of key context clues in the semantic enhancement process. Therefore, in the technical solution of the present application, based on the semantic coding features of the context information of the address component to be strengthened, the semantic embedding coding features of the address component to be strengthened are semantically enhanced to obtain the semantic enhancement coding features of the address component to be strengthened. That is, a graph structure topological space of the semantic kernel association between the address component to be strengthened and the context is established in the feature principal component dimension, so as to deeply explore the fine-grained semantic interaction rules of the address component to be strengthened, so as to obtain the semantic enhancement coding features of the address component to be strengthened. Specifically, first, based on the principal component kernel association coding network, the principal component vectors of the component to be strengthened (such as the administrative division feature axis of "Hongqiao Road") and the principal component vectors of the context component (such as the economic function axis of "Changning District, Shanghai") are mapped to a high-dimensional association space, generating a kernel association code that reflects the strength of cross-component interactions. This decoupling operation not only eliminates feature redundancy but also enables subsequent graph structure modeling to focus on the essential association patterns between address components. Then, through the principal component kernel association coding network, the system maps the decoupled principal component vectors of the semantic features of the address components to a high-dimensional association space, breaking through the traditional attention mechanism's reliance on explicit similarity calculation and capturing the abstract semantic associations between address components. Subsequently, a dynamic graph structure is constructed using the association topology matrix of the kernel association strength evaluation factors of the address components to be strengthened, in which nodes represent principal component feature dimensions (such as the postal code feature axis and the road grade axis), and edge weights are determined by the kernel association strength. Subsequently, through the multi-hop information propagation of the graph convolutional neural network, the system can capture the semantic resonance effect across levels and establish long-distance associations between address components. This semantic enhancement mechanism based on graph topology enables the differentiated features of homonymous address components in different contexts to be explicitly expressed, and generates distinct translation vectors through differential information aggregation through graph convolution, achieving accurate semantic mapping in cross-language scenarios.

[0038] Specifically, first, perform principal component analysis on the semantic encoding vector of the context information of the address component to be enhanced and the semantic embedding encoding vector of the address component to be enhanced, so as to obtain a set of principal component encoding vectors of the context semantics of the address component to be enhanced and a set of principal component encoding vectors of the semantics of the address component to be enhanced. Specifically, there is multicollinearity between the context semantic encoding of the address component (such as the temporal features generated by bidirectional LSTM) and the semantic embedding encoding (such as the distributed representation mapped by the embedding matrix) in the original feature space. When traditional methods directly splice these high-dimensional vectors, the collinearity between different semantic dimensions (such as the correlation between geographical orientation and administrative division) will make it difficult for the model to focus on key discriminative features, and this redundant association will interfere with the subsequent cross-component semantic interaction modeling. Therefore, in the technical solution of this application, in order to achieve semantic decoupling and information purification of the feature space, perform principal component analysis on the semantic encoding vector of the context information of the address component to be enhanced and the semantic embedding encoding vector of the address component to be enhanced, so as to obtain a set of principal component encoding vectors of the context semantics of the address component to be enhanced and a set of principal component encoding vectors of the semantics of the address component to be enhanced. Through the principal component analysis technology, an orthogonal transformation is performed on the semantic encoding vector of the context information of the address component to be enhanced and the semantic embedding encoding vector of the address component to be enhanced. The system can strip out independent principal component dimensions from the complex feature interweaving state, and these dimensions respectively carry the core semantic attributes of the address component (such as the geographical coordinate features of "Paris" as an urban entity) and the context constraint relationship (such as the orientation of "Texas" to the North American geographical reference system). In the specific implementation process, adopt a hierarchical orthogonalized feature reconstruction strategy, decompose the covariance matrix of the semantic encoding vector of the context information of the address component to be enhanced and the semantic embedding encoding vector of the address component to be enhanced, and screen out the principal component axes with the highest variance contribution rate through eigenvalue sorting. In this way, not only the redundant association between the address component features is eliminated, but also the feature dimensions with clear semantic pointers are reconstructed. In a specific example of this application, perform principal component analysis on the semantic encoding vector of the context information of the address component to be enhanced and the semantic embedding encoding vector of the address component to be enhanced with the following principal component analysis formula, so as to obtain a set of principal component encoding vectors of the context semantics of the address component to be enhanced and a set of principal component encoding vectors of the semantics of the address component to be enhanced; where, the principal component analysis formula is:

[0039] ;

[0040] ;

[0041] Wherein, is the semantic encoding vector of the context information of the address component to be enhanced, is the semantic embedding encoding vector of the address component to be enhanced, is the principal component analysis, is the set of principal component coding vectors of the context semantic features of the address components to be enhanced, is the set of principal component coding vectors of the semantic features of the address components to be enhanced, are respectively the 1st, 2nd, th, and the th principal component coding vectors of the context semantic features of the address components to be enhanced in the set of principal component coding vectors of the context semantic features of the address components to be enhanced, are respectively the 1st, 2nd, th, and the th principal component coding vectors of the semantic features of the address components to be enhanced in the set of principal component coding vectors of the semantic features of the address components to be enhanced, is the diagonal matrix of the semantic features related to the scheduling requirement time, and are respectively and the corresponding eigenvalues, is the diagonal matrix of the semantic features of the address components to be enhanced, and are respectively and the corresponding eigenvalues.

[0042] Next, for each pair of corresponding context semantic feature principal component encoding vectors and semantic feature principal component encoding vectors of the address components to be enhanced in the set of context semantic feature principal component encoding vectors of the address components to be enhanced and the set of semantic feature principal component encoding vectors of the address components to be enhanced, principal component kernel association encoding is performed respectively to obtain a set of kernel association encoding vectors between the semantic feature principal components of the address components to be enhanced. Since the linear combination of the set of context semantic feature principal component encoding vectors of the address components to be enhanced and the set of semantic feature principal component encoding vectors of the address components to be enhanced in the low-dimensional orthogonal space cannot capture complex semantic dependency relationships. Traditional linear associations cannot effectively model the non-linear coupling relationship between the two in the feature dimension, resulting in the inability to distinguish subtle differences between, for example, multiple cities with the same name. Therefore, in the technical solution of this application, to construct a hyperplane association topology between the address principal component features to learn the cross-dimensional semantic resonance pattern through a deep neural network, in the technical solution of this application, for each pair of corresponding context semantic feature principal component encoding vectors and semantic feature principal component encoding vectors of the address components to be enhanced in the set of context semantic feature principal component encoding vectors of the address components to be enhanced and the set of semantic feature principal component encoding vectors of the address components to be enhanced, principal component kernel association encoding is performed respectively to obtain a set of kernel association encoding vectors between the semantic feature principal components of the address components to be enhanced. Specifically, the system projects the decoupled principal component pairs (such as the cultural naming feature axis of "Chang'an Street" and the administrative division feature axis of "Xi'an City") into a high-dimensional association space through non-linear mapping, so that the semantic relationships that were originally linearly inseparable in the low-dimensional space (such as the road naming rules under the specific historical background of Beijing) can be explicitly expressed. In one example, a deep encoding architecture with heterogeneous feature fusion can be adopted. First, an independent fully connected neural network channel is constructed for each pair of principal components (such as the "landmark function" axis of the component to be enhanced and the "administrative division level" axis of the context component), and the cross-dimensional association pattern is learned through multi-layer non-linear transformation; subsequently, the system uses an attention gating mechanism to dynamically weight the association encoding, so that the differential association patterns of the same-name components in different contexts can be explicitly expressed. This deep association mechanism enables the system to handle extremely complex scenarios of the same name in different locations (such as 28 "San Jose" globally), and learn the unique association pattern between the context principal components of the country, state, and province where each "San Jose" is located and the city name feature axis through the kernel network, so as to dynamically select the naming specification that conforms to the cultural habits of the target language during translation.In a specific example of the present application, the corresponding context semantic feature principal component coding vectors and semantic feature principal component coding vectors of the address components to be enhanced in the set of context semantic feature principal component coding vectors of the address components to be enhanced and the set of semantic feature principal component coding vectors of the address components to be enhanced are respectively subjected to principal component kernel correlation coding according to the following correlation coding formula to obtain a set of kernel correlation coding vectors between the semantic feature principal components of the address components to be enhanced; wherein, the correlation coding formula is:.

[0043] ;

[0044] Wherein, represents the first norm of the vector, and respectively represent trainable weighted hyperparameters, is the th kernel correlation coding vector between the semantic feature principal components of the address components to be enhanced in the set of kernel correlation coding vectors between the semantic feature principal components of the address components to be enhanced.

[0045] Subsequently, calculate the kernel correlation strength evaluation factor between any two kernel correlation encoding vectors of the main components of the semantic features of the address components to be enhanced in the set of kernel correlation encoding vectors of the main components of the semantic features of the address components to be enhanced, so as to obtain the kernel correlation strength evaluation factor correlation topology matrix of the address components to be enhanced. Specifically, when dealing with address components with multi-level semantic nesting, for example, the main components of a certain component may simultaneously carry geographical level attributes, cultural metaphor features, and spatial orientation information, it is difficult for traditional shallow association metrics based on Euclidean distance or cosine similarity to depict the functional synergy relationship between these cross-dimensional features. The kernel correlation strength evaluation factor of the address components to be enhanced, as a task-oriented customized metric function, its core purpose is to transform the complex interaction patterns between high-dimensional kernel correlation encoding vectors into quantifiable topological connection weights, so as to construct an adjacency relationship matrix that can reflect the semantic synergy effect for the graph neural network. Therefore, in the technical solution of this application, calculate the kernel correlation strength evaluation factor between any two kernel correlation encoding vectors of the main components of the semantic features of the address components to be enhanced in the set of kernel correlation encoding vectors of the main components of the semantic features of the address components to be enhanced, so as to obtain the kernel correlation strength evaluation factor correlation topology matrix of the address components to be enhanced. That is, construct an interpretable semantic association strength map, and realize the directional strengthening and suppression of feature interaction through a customized metric function. Specifically, by calculating the kernel correlation strength evaluation factor of the address components to be enhanced, the complex interaction patterns between the kernel correlation encoding vectors of the main components of the semantic features of the address components to be enhanced can be transformed into quantifiable association weights. These weights not only reflect the collaborative contribution of the features to the translation task (such as the hierarchical constraint relationship between the "street type" and "administrative division" features), but also can dynamically adapt to the translation rules of different language pairs (such as strengthening the feature association of the Chinese character cultural circle when translating from Chinese to Japanese, and enhancing the feature association of the transliteration rule when translating from Chinese to English). The generated kernel correlation strength evaluation factor correlation topology matrix of the address components to be enhanced can accurately depict the influence intensity of different semantic dimensions on the translation decision. For example, when dealing with the translation of streets with the same name, the association weight between the cultural memorial attribute feature of "Zhongshan Road" and the historical background feature of the superior administrative region will be strengthened to ensure that its memorial meaning is retained when translated as "Zhongshan Road"; while when "Nanjing Road" is a commercial landmark, the association weight between its functional attribute feature and the urban economic feature dominates, driving the adoption of the internationally common "Nanjing Road" instead of a literal translation when translating. This dynamic quantization mechanism enables the system to adaptively process complex scenarios and achieve an accurate balance between retaining the original cultural features and conforming to the target language norms by adjusting the association strength of different semantic channels.In a specific example of the present application, the kernel correlation strength evaluation factor between any two kernel correlation coding vectors of the principal components of the semantic features of the address components to be strengthened is calculated by the following correlation strength calculation formula to obtain the kernel correlation strength evaluation factor correlation topology matrix of the address components to be strengthened; wherein, the correlation strength calculation formula is:.

[0046] ;

[0047] ;

[0048] Wherein, and are respectively the eigenvalues at the th and th positions in the kernel correlation coding vector between the principal components of the semantic features of the address components to be strengthened, is the number of eigenvalues in the kernel correlation coding vector between the principal components of the semantic features of the address components to be strengthened, is the kernel correlation strength evaluation factor between and and , is the kernel correlation strength evaluation factor correlation topology matrix of the address components to be strengthened.

[0049] Then, perform disorder optimization on the kernel space distribution of each to-be-strengthened kernel correlation coding vector among the principal components of the semantic features of the address components to be strengthened in the set of kernel correlation coding vectors among the principal components of the semantic features of the address components to be strengthened, so as to obtain a set of kernel correlation coding vectors among the principal components of the semantic features of the address components to be strengthened after optimization. In particular, since the disorder of the spatial distribution generated by the set of kernel correlation coding vectors among the principal components of the semantic features of the address components to be strengthened under the action of the random potential field will lead to information transmission distortion of the graph structure. For example, there may be a dimension break between the principal component features of the address components (such as the administrative level axis and the cultural attribute axis) and the context correlation topology, resulting in the obstruction of the semantic reasoning path across components. Traditional methods do not consider the compactness constraint of the feature principal components in the topological space, resulting in the attenuation or distortion of the high-order correlation signals during the graph convolution process, affecting the semantic consistency of the translation decision. In a preferred example of the present application, perform disorder optimization on the kernel space distribution of each to-be-strengthened kernel correlation coding vector among the principal components of the semantic features of the address components to be strengthened in the set of kernel correlation coding vectors among the principal components of the semantic features of the address components to be strengthened, so as to obtain a set of kernel correlation coding vectors among the principal components of the semantic features of the address components to be strengthened after optimization. That is, in the technical solution of the present application, the standardized transition of the topological space is realized by constructing a cross-section function matrix to eliminate the disorder of the spatial distribution caused by the random potential field. Specifically, the system projects the set of discrete kernel correlation coding vectors among the principal components of the semantic features of the address components to be strengthened into a normalized topological form space through the generation of compactified vectors and the iterative optimization of the Gaussian correlation coefficient, so that the graph convolutional network can perform multi-hop information propagation based on a stable spatial distribution. For example, in the cross-language address translation scenario, when dealing with address components with complex hierarchical relationships, the construction of the compactification matrix can ensure the spatial continuity of the correlation topology between the principal component axes of administrative levels such as "provincial - municipal - district-level", avoiding translation breaks caused by dimension misalignment.In this process, by adopting a dynamic space calibration mechanism based on the cross-sectional function matrix, first, the kernel correlation encoding vector between the principal components of the semantic features of each address component to be enhanced is multiplied by the initial cross-sectional function matrix of the address component features to be enhanced for dimensionality reduction and compactification processing of the high-dimensional feature space, eliminating the interference signals of redundant dimensions; subsequently, through two-dimensional splicing, a compact cross-sectional matrix of the address component features to be enhanced is formed, and the Gaussian correlation coefficient between it and the correlation topology matrix of the kernel correlation strength evaluation factor of the address component to be enhanced is calculated. Through an iterative optimization process that constrains the Gaussian coefficient to approach zero, the system gradually corrects the parameters of the initial cross-sectional function matrix of the address component features to be enhanced, making it a standardized transition bridge connecting the correlation encoding space of the principal components of the semantic features of the address component to be enhanced and the graph topology space; finally, the optimized initial cross-sectional function matrix of the address component features to be enhanced maps the kernel correlation encoding vector between the principal components of the semantic features of the address component to be enhanced to a compact topological subspace, ensuring that the message passing of the graph convolutional network follows the standardized spatial distribution law. By eliminating the disorder of the spatial distribution, the model can accurately capture cross-level and cross-dimensional semantic association patterns. For example, when processing multi-language mixed addresses, the implicit association between the principal component axis of the cultural attributes and the language rule topology can be completely transmitted. This standardized spatial distribution mechanism enables high-order semantic signals to be aggregated losslessly through multiple layers of graph convolution, ensuring that the translation results satisfy both local context constraints and global semantic consistency, fundamentally breaking through the topological distortion bottleneck of traditional methods in complex place name semantic reasoning.

[0050] That is, in the calculation process of the above graph convolutional neural network model, the correlation topology matrix of the kernel correlation strength evaluation factor of the address component to be enhanced As a topological form space, then the kernel correlation encoding vector between the principal components of the semantic features of each address component to be enhanced Will serve as the space master and slave to obey the spatial distribution form, that is , and, considering the dimensionality break between the kernel correlation encoding vector between the principal components of the semantic features of the address component to be enhanced And the correlation topology matrix of the kernel correlation strength evaluation factor of the address component to be enhanced A space standard transition gauge field needs to be constructed from the initial cross-sectional function matrix of the address component features to be enhanced To serve as the spatial standard transition gauge field.

[0051] On the other hand, it is also necessary to correct the problem of spatial distribution disorder caused by the random potential field in the initial cross-sectional function matrix of the address component features to be enhanced First, the kernel correlation encoding vector between the principal components of the semantic features of each address component to be enhanced And the corresponding initial cross-sectional function matrix of the address component features to be enhanced Perform matrix multiplication to obtain the compactified vector of the feature cross-section of the address component to be enhanced , and then the kernel correlation encoding vectors between the semantic feature principal components of each address component to be enhanced The corresponding compactified vector of the feature cross-section of the address component to be enhanced Are two-dimensionally stitched to obtain the compactified matrix of the feature cross-section of the address component to be enhanced :

[0052] ;

[0053] ;

[0054] Among them, Is the rd kernel correlation encoding vector between the semantic feature principal components of the address component to be enhanced in the set of kernel correlation encoding vectors between the semantic feature principal components of the address component to be enhanced, Is the initial function matrix of the feature cross-section of the address component to be enhanced, Is matrix multiplication, Is the th compactified vector of the feature cross-section of the address component to be enhanced in the sequence of compactified vectors of the feature cross-section of the address component to be enhanced, Is two-dimensional stitching processing, Is the compactified matrix of the feature cross-section of the address component to be enhanced.

[0055] In this way, the Gaussian correlation coefficient between the compactified matrix of the feature cross-section of the address component to be enhanced and the associated topological matrix of the kernel association strength evaluation factor of the address component to be enhanced can be calculated, and the initial function matrix of the feature cross-section of the address component to be enhanced can be iterated by making the Gaussian correlation coefficient tend to zero to obtain the optimized initial function matrix of the feature cross-section of the address component to be enhanced:

[0056] ;

[0057] Among them, Is subtraction by position, Is the norm of the matrix, Is And The variance of the set composed of all matrix values of Is the natural exponential function value with the natural constant as the base.

[0058] Thus, by optimizing the initial function matrix Optimize the kernel correlation coding vector between the principal components of the semantic features of the address components to be enhanced:

[0059] ;

[0060] Among them, is the function matrix of the feature cross-section for optimizing the initial address components to be enhanced, is the optimized kernel correlation coding vector between the principal components of the semantic features of the address components to be enhanced corresponding to the kernel correlation coding vector between the principal components of the semantic features of the th address components to be enhanced.

[0061] During the calculation process of the graph convolutional neural network model, the problem of the disordered spatial distribution brought by the kernel correlation coding vector between the principal components of each feature in the random potential field representation to the topological form space of the correlation topology matrix of the kernel correlation strength evaluation factor of the address components to be enhanced can be solved, thereby improving the calculation results of the graph convolutional neural network model.

[0062] Furthermore, the set of kernel correlation encoding vectors between the principal components of the semantic features of the address components to be enhanced after optimization and the correlation topology matrix of the kernel correlation strength evaluation factors of the address components to be enhanced are input into the graph convolutional neural network model to obtain the semantic enhancement encoding matrix of the address components to be enhanced. Specifically, although kernel correlation encoding can capture the direct correlations between principal components, it cannot explicitly represent the cross-level and cross-dimensional composite semantic influences. For example, the administrative division attribute of an address component may indirectly affect the road naming rules through multi-level nodes, while cultural features may form implicit associations with language habits through long-range connections in the graph structure. Such complex interaction patterns are difficult to capture by shallow networks and rely on the deep aggregation mechanism of graph convolution. Specifically, the graph convolutional neural network regards each kernel correlation encoding vector between the principal components of the semantic features of the address components to be enhanced as a graph node, and based on the connection weights defined by the correlation topology matrix of the kernel correlation strength evaluation factors of the address components to be enhanced, aggregates the semantic features of adjacent nodes layer by layer. For example, when processing an address containing multiple administrative levels, the features of the "provincial-level node" can affect the translation decision of the "street-level node" through multi-hop propagation of graph convolution, ensuring that the translation result not only conforms to the local context constraints but also inherits the characteristics of the global administrative system. At the same time, the attention gating mechanism in the graph structure can dynamically adjust the contribution degrees of different-level nodes. For example, when processing culturally sensitive place names, it enhances the propagation weight of historical evolution nodes and suppresses the interference of ordinary descriptive nodes. Taking the processing of address components with multiple administrative nesting relationships as an example, low-level nodes (such as road names) receive jurisdiction constraint signals from high-level nodes (such as provincial administrative regions) through multi-hop propagation, and at the same time, reversely transmit geographical specificity features (such as regional naming conventions) to form two-way semantic correction. Each layer of convolution operation is essentially a joint embedding process of the feature space and the topological space, gradually refining the implicit association rules across components through non-linear transformation. The generated semantic enhancement encoding matrix of the address components to be enhanced contains multi-granularity address component semantic associations from local to global. This global semantic reasoning mechanism based on the graph structure effectively solves the problem of over-reliance on local co-occurrence patterns in traditional methods and realizes the collaborative optimization of semantic fidelity and cross-language adaptability in place name translation. In a specific example of the present application, the set of kernel correlation encoding vectors between the principal components of the semantic features of the address components to be enhanced after optimization and the correlation topology matrix of the kernel correlation strength evaluation factors of the address components to be enhanced are input into the graph convolutional neural network model according to the following graph convolution formula to obtain the semantic enhancement encoding matrix of the address components to be enhanced; where the graph convolution formula is:

[0063] ;

[0064] where is the The kernel correlation coding vector between the principal components of the semantic features of the address components to be strengthened corresponds to the optimized kernel correlation coding vector between the principal components of the semantic features of the address components to be strengthened, represents the graph convolutional neural network model, is the semantic enhancement coding matrix of the address component to be enhanced.

[0065] In particular, the S5 obtains the target language address text based on the semantic reinforcement coding features of the address components to be strengthened. That is, in the technical solution of the present application, the semantic reinforcement coding matrix of the address components to be strengthened is input into the pre-trained neural machine translation model to obtain the target language address text. Specifically, when the traditional translation model directly processes the original address text, it is limited by the discreteness of the surface symbols and the heterogeneity of the language structure, and it is difficult to synchronously coordinate the complex mapping relationship between geographical hierarchical constraints, cultural metaphor features and multilingual grammatical rules. Therefore, in the technical solution of the present application, the semantic reinforcement coding matrix of the address components to be strengthened is input into the pre-trained neural machine translation model to achieve accurate mapping of structured semantic features to target language sequences. This deep semantic-driven generation mechanism realizes lossless conversion from structured semantic representation to multilingual sequence generation, ensuring that the translation results meet both geographical accuracy and language naturalness requirements.

[0066] In summary, the real-time multilingual translation method of place names and addresses based on a neural network model according to the embodiment of the present application is explained. It first extracts the topological relationship of the address components through structured analysis, and then uses a neural network model based on deep learning to deeply model the contextual dependencies across components, so as to mine the deep semantic representation of the address text through semantic association learning across components, and based on this, realize place name disambiguation based on dynamic semantic reasoning. In this way, the semantic deviation problem in complex scenarios such as the same name in different places and the same place with different names is solved, and the accuracy of the judgment of the ambiguity of geographic entities is significantly improved, so that the translation results can dynamically adapt to the place name expression habits of different languages, effectively eliminate the translation ambiguity caused by the limitations of the static knowledge base or imperfect rules, and achieve a more accurate multilingual address conversion effect.

[0067] Furthermore, a real-time place name and address multilingual translation system based on a neural network model is also provided.

[0068] Figure 5 FIG is a block diagram of a real-time multilingual translation system of place names and addresses based on a neural network model according to an embodiment of the present application. Figure 5As shown, the real-time multi-language translation system 300 for geographical names and addresses based on a neural network model according to an embodiment of the present application includes: an address text acquisition module 310 for acquiring address text input by a user; an address parsing module 320 for structurally parsing the address text using an address parsing model to obtain a set of address components; a semantic encoding module 330 for respectively performing semantic encoding on the address components to be enhanced and the context information of the address components to be enhanced in the set of address components to obtain semantic embedding encoding features of the address components to be enhanced and semantic encoding features of the context information of the address components to be enhanced; a semantic enhancement module 340 for enhancing the semantic embedding encoding features of the address components to be enhanced based on the semantic encoding features of the context information of the address components to be enhanced to obtain semantic enhanced encoding features of the address components to be enhanced; and a target language address text generation module 350 for obtaining a target language address text based on the semantic enhanced encoding features of the address components to be enhanced.

[0069] As described above, the real-time multi-language translation system 300 for geographical names and addresses based on a neural network model according to an embodiment of the present application can be implemented in various wireless terminals, such as a server with a real-time multi-language translation algorithm for geographical names and addresses based on a neural network model. In a possible implementation manner, the real-time multi-language translation system 300 for geographical names and addresses based on a neural network model according to an embodiment of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the real-time multi-language translation system 300 for geographical names and addresses based on a neural network model can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the real-time multi-language translation system 300 for geographical names and addresses based on a neural network model can also be one of the many hardware modules of the wireless terminal.

[0070] Alternatively, in another example, the real-time multi-language translation system 300 for geographical names and addresses based on a neural network model and the wireless terminal can also be separate devices, and the real-time multi-language translation system 300 for geographical names and addresses based on a neural network model can be connected to the wireless terminal through a wired and / or wireless network and transmit interaction information in accordance with a predefined data format.

[0071] The embodiments of the present disclosure have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of the technology in the market, or to enable other ordinary skilled persons in the technical field to understand the disclosed embodiments.

Claims

1. A real-time multi-language translation method for place names and addresses based on a neural network model, characterized in that, Including: Obtain the address text input by the user; Use the address parsing model to perform structured parsing on the address text to obtain a set of address components; Select the address component to be enhanced from the set of address components and use the other address components in the set of address components as the context information of the address component to be enhanced; Perform semantic embedding encoding on the address component to be enhanced to obtain the semantic embedding encoding feature of the address component to be enhanced, and perform semantic enhancement on the semantic embedding encoding feature of the address component to be enhanced based on the context information of the address component to be enhanced to obtain the semantic enhanced encoding feature of the address component to be enhanced, which includes: Use the embedding matrix of the address component to be enhanced to perform semantic embedding encoding on the address component to be enhanced, so as to obtain the semantic embedding encoding vector of the address component to be enhanced and use it as the semantic embedding encoding feature of the address component to be enhanced; Perform semantic embedding encoding on each address component in the context information of the address component to be enhanced to obtain the context sequence of the semantic embedding encoding vectors of the place name components; Perform context semantic encoding based on the bidirectional LSTM model on the context sequence of the semantic embedding encoding vectors of the place name components to obtain the context information semantic encoding vector of the address component to be enhanced and use it as the context information semantic encoding feature of the address component to be enhanced; Based on the context information semantic encoding feature of the address component to be enhanced, perform semantic enhancement on the semantic embedding encoding feature of the address component to be enhanced to obtain the semantic enhanced encoding feature of the address component to be enhanced; Based on the semantic enhanced encoding feature of the address component to be enhanced, obtain the target language address text.

2. The real-time toponym address multilingual translation method based on a neural network model according to claim 1, wherein Based on the context information semantic encoding feature of the address component to be enhanced, perform semantic enhancement on the semantic embedding encoding feature of the address component to be enhanced to obtain the semantic enhanced encoding feature of the address component to be enhanced, including: Perform kernel association information encoding based on principal component analysis on the context information semantic encoding vector of the address component to be enhanced and the semantic embedding encoding vector of the address component to be enhanced to obtain a set of kernel association encoding vectors between the principal components of the semantic features of the address component to be enhanced; Construct the kernel association strength evaluation factor between every two kernel association encoding vectors between the principal components of the semantic features of the address component to be enhanced in the set of kernel association encoding vectors between the principal components of the semantic features of the address component to be enhanced, and perform graph structure-based association encoding on the set of kernel association encoding vectors between the principal components of the semantic features of the address component to be enhanced based on the kernel association strength evaluation factor to obtain the semantic enhanced encoding matrix of the address component to be enhanced and use it as the semantic enhanced encoding feature of the address component to be enhanced.

3. The real-time multilingual translation method of geographical names and addresses based on a neural network model according to claim 2, characterized in that Perform kernel association information encoding based on principal component analysis on the context information semantic encoding vector of the address component to be enhanced and the semantic embedding encoding vector of the address component to be enhanced to obtain a set of kernel association encoding vectors between the principal components of the semantic features of the address component to be enhanced, including: Perform feature principal component analysis on the context information semantic encoding vector of the address component to be enhanced and the semantic embedding encoding vector of the address component to be enhanced to obtain a set of principal component encoding vectors of the context semantic features of the address component to be enhanced and a set of principal component encoding vectors of the semantic features of the address component to be enhanced; For each pair of corresponding context semantic feature principal component encoding vectors and semantic feature principal component encoding vectors of the address components to be enhanced in the set of context semantic feature principal component encoding vectors of the address components to be enhanced and the set of semantic feature principal component encoding vectors of the address components to be enhanced, perform principal component kernel correlation encoding to obtain a set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced.

4. The real-time toponym address multilingual translation method based on a neural network model according to claim 3, characterized in that Construct a kernel correlation strength evaluation factor between every two kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced in the set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced, and perform graph-structure-based correlation encoding on the set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced based on the kernel correlation strength evaluation factor to obtain a semantic enhancement encoding matrix for the address components to be enhanced, including: Calculate the kernel correlation strength evaluation factor of the address components to be enhanced between any two kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced in the set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced to obtain a correlation topology matrix of the kernel correlation strength evaluation factor of the address components to be enhanced; Based on the correlation topology matrix of the kernel correlation strength evaluation factor of the address components to be enhanced, perform graph-structure-based correlation encoding on the set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced to obtain a semantic enhancement encoding matrix for the address components to be enhanced.

5. The real-time toponym and address multilingual translation method based on a neural network model according to claim 4, characterized in that, Based on the correlation topology matrix of the kernel correlation strength evaluation factor of the address components to be enhanced, perform graph-structure-based correlation encoding on the set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced to obtain a semantic enhancement encoding matrix for the address components to be enhanced, including: Perform disorder optimization of the kernel space distribution of the address components to be enhanced on each kernel correlation encoding vector between the semantic feature principal components of the address components to be enhanced in the set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced to obtain an optimized set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced; Input the optimized set of kernel correlation encoding vectors between the semantic feature principal components of the address components to be enhanced and the correlation topology matrix of the kernel correlation strength evaluation factor of the address components to be enhanced into a graph convolutional neural network model to obtain a semantic enhancement encoding matrix for the address components to be enhanced.

6. The real-time multilingual translation method of geographical names and addresses based on a neural network model according to claim 5, characterized in that Based on the semantic enhancement encoding features of the address components to be enhanced, obtain the target language address text, including: Input the semantic enhancement encoding matrix of the address components to be enhanced into a pre-trained neural machine translation model to obtain the target language address text.

7. A real-time multilingual translation system for geographical names and addresses based on a neural network model, the system being used to execute the method as described in claim 6, characterized in that, Including: An address text acquisition module for acquiring the address text input by the user; An address parsing module for structurally parsing the address text using an address parsing model to obtain a set of address components; A semantic encoding module for respectively performing semantic encoding on the address components to be enhanced and the context information of the address components to be enhanced in the set of address components to obtain semantic embedding encoding features of the address components to be enhanced and semantic encoding features of the context information of the address components to be enhanced, which includes: Semantically embed and encode the address component to be enhanced using the address component embedding matrix to be enhanced, so as to obtain the semantic embedding encoding vector of the address component to be enhanced and use it as the semantic embedding encoding feature of the address component to be enhanced; Semantically embed and encode each address component in the context information of the address component to be enhanced to obtain the context sequence of the semantic embedding encoding vectors of the place name components; Perform context semantic encoding based on the bidirectional LSTM model on the context sequence of the semantic embedding encoding vectors of the place name components to obtain the context information semantic encoding vector of the address component to be enhanced and use it as the context information semantic encoding feature of the address component to be enhanced; A semantic enhancement module for semantically enhancing the semantic embedding encoding feature of the address component to be enhanced based on the context information semantic encoding feature of the address component to be enhanced to obtain the semantic enhancement encoding feature of the address component to be enhanced; A target language address text generation module for obtaining the target language address text based on the semantic enhancement encoding feature of the address component to be enhanced.

Images