Method for enhancing the translation accuracy of place names and addresses by using context awareness and geographical libraries

Abstract

The present application discloses a method for enhancing the translation accuracy of place names and addresses by using context awareness and geographical knowledge base, which relates to the field of intelligent translation technology. By integrating the context awareness mechanism and the multi-dimensional attributes of the geographical knowledge base, a dynamic collaborative disambiguation framework is constructed to break through the semantic limitations of traditional place name and address translation. Specifically, firstly, the context semantic encoding of non-place name text segments is used to capture the dynamic association information hidden in the address text, and the multi-dimensional attributes of the candidate entities are extracted by combining the geographical knowledge base with the structured feature embedding of the candidate entities, so as to achieve cross-modal disambiguation decision-making in the semantic aggregation stage. Through the geographical feature cross-validation mechanism guided by context semantics, the translation ambiguity problem caused by static rule dependence and geographical attribute fragmentation in the prior art is effectively solved, thus achieving a more accurate multi-language address conversion effect.

Description

Technical Field

[0001] This application relates to the field of intelligent translation technology, and more specifically, to a method for enhancing the translation accuracy of place names and addresses by using context awareness and a geographical knowledge base. Background Art

[0002] In the context of globalization today, address translation, as an important part of cross-language communication, its accuracy is crucial for ensuring the correct transmission of information. In the field of traditional place name and address translation, existing technologies generally face the problem of insufficient translation accuracy caused by language polysemy and geographical entity complexity. Conventional methods usually rely on static dictionary matching or rule-based translation strategies, and it is difficult to cope with the ambiguity challenges of the same place name in different contexts. For example, when there are administrative regions, natural geographical entities or cultural landmarks with the same name in the source language, existing systems often lack the ability to dynamically identify context associations, resulting in deviations between the translation results and the target geographical locations.

[0003] In addition, the application of geographical knowledge bases in existing technologies is not deep enough. Existing translation models do not fully integrate the spatial hierarchical relationships, administrative region subordination degrees, and historical naming features in structured geographical knowledge bases, making the disambiguation process limited to local text segment analysis and unable to perform cross-verification through multi-dimensional geographical attributes, thus affecting the accuracy of the final translation results.

[0004] Therefore, a method for enhancing the translation accuracy of place names and addresses by using context awareness and a geographical knowledge base is needed to solve the above technical problems. Summary of the Invention

[0005] In order to solve the above technical problems, this application is proposed.

[0006] According to one aspect of this application, a method for enhancing the translation accuracy of place names and addresses by using context awareness and a geographical knowledge base is provided, which includes:

[0007] S1: Receive the source language address text to be translated;

[0008] S2: Perform standardized preprocessing and structured semantic parsing on the source language address text to obtain a list of candidate place name entities and non-place name text segments;

[0009] S3: Extract the first disambiguation place name entity from the list of candidate place name entities;

[0010] S4: Extract multiple candidate place name entity entries that match the first disambiguation place name entity from the geographical knowledge base;

[0011] S5: Input the multiple candidate toponym entity entries and non-toponym text fragments that match the first toponym entity to be disambiguated into the disambiguation model to obtain the disambiguated toponym entity of the first toponym entity to be disambiguated;

[0012] Step 6: Loop through S3 to S5 to obtain a list of disambiguated toponym entities;

[0013] Step 7: Obtain the target language address text based on the list of disambiguated toponym entities.

[0014] This application has at least the following technical effects: Compared with the prior art, a method for enhancing the translation accuracy of toponym addresses by using context awareness and a geographical library provided by this application constructs a dynamic collaborative disambiguation framework by integrating the context awareness mechanism and the multi-dimensional attributes of the geographical knowledge base to break through the semantic limitations of traditional toponym address translation. Specifically, first, the context semantic encoding of non-toponym text fragments is used to capture the dynamic association information hidden in the address text, and the candidate entities are embedded with structured features in combination with the geographical knowledge base to extract the multi-dimensional attributes of the candidate entities, so as to achieve cross-modal disambiguation decision-making in the semantic aggregation stage. This application can effectively solve the translation ambiguity problem caused by static rule dependence and geographical attribute fragmentation in the prior art through the geographical feature cross-validation mechanism guided by context semantics, thereby achieving a more accurate multi-language address conversion effect. Description of the Drawings

[0015] 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. Together with the embodiments of the present application, they are used to explain 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.

[0016] Figure 1 It is a flowchart of a method for enhancing the translation accuracy of toponym addresses by using context awareness and a geographical library according to an embodiment of the present application;

[0017] Figure 2 It is a schematic diagram of data flow of a method for enhancing the translation accuracy of toponym addresses by using context awareness and a geographical library according to an embodiment of the present application;

[0018] Figure 3 It is a flowchart of sub-step S5 of a method for enhancing the translation accuracy of toponym addresses by using context awareness and a geographical library according to an embodiment of the present application;

[0019] Figure 4 It is a flowchart of sub-step S51 of a method for enhancing the translation accuracy of toponym addresses by using context awareness and a geographical library according to an embodiment of the present application. Detailed Implementation Modes

[0020] 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.

[0021] As shown in the present application and the claims, unless the context clearly indicates otherwise, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0022] 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.

[0023] 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 executed precisely in sequence. On the contrary, various steps can be processed in reverse order or simultaneously as needed. At the same time, other operations can also be added to these processes, or one or several operations can be removed from these processes.

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

[0026] In the technical solution of the present application, a method for enhancing the translation accuracy of place names and addresses by using context awareness and a geographical library is proposed. Figure 1 FIG. [ID] is a flowchart of a method for enhancing the translation accuracy of place names and addresses by using context awareness and a geographical library according to an embodiment of the present application. Figure 2 FIG. [ID] is a schematic diagram of data flow of a method for enhancing the translation accuracy of place names and addresses by using context awareness and a geographical library according to an embodiment of the present application. As Figure 1 and Figure 2As shown, a method for enhancing the translation accuracy of geographical names and addresses by using context awareness and geographical libraries according to an embodiment of the present application includes the steps of: S1, receiving a source language address text to be translated; S2, performing standardized preprocessing and structured semantic parsing on the source language address text to obtain a list of candidate geographical name entities and non-geographical name text segments; S3, extracting a first disambiguation geographical name entity from the list of candidate geographical name entities; S4, extracting a plurality of candidate geographical name entity entries matching the first disambiguation geographical name entity from a geographical knowledge base; S5, inputting the plurality of candidate geographical name entity entries matching the first disambiguation geographical name entity and the non-geographical name text segments into a disambiguation model to obtain a disambiguated geographical name entity of the first disambiguation geographical name entity; S6, repeatedly executing S3 to S5 to obtain a list of disambiguated geographical name entities; S7, obtaining a target language address text based on the list of disambiguated geographical name entities.

[0027] Specifically, in S1, a source language address text to be translated is received. In practical applications, the source language address text may come from cross-border logistics documents, multilingual map service interfaces, or fragmented information manually input by users, and its manifestation often includes spelling variants, abbreviations, non-standard separators, and even mixed punctuation marks.

[0028] Specifically, in S2, the source language address text is subjected to standardized preprocessing and structured semantic parsing to obtain a list of candidate place name entities and non-place name text segments. In the technical solution of this application, first, the source language address text is cleaned to obtain a purified source language address text. It should be understood that the original source language address text is interfered by multimodal noise, such as unnecessary punctuation marks, spelling mistakes, or irrelevant characters, etc. These factors will interfere with the subsequent processing steps. Therefore, in the technical solution of this application, the source language address text is cleaned to obtain a purified source language address text; specifically, a structured preprocessing channel is established, and the interference information outside the address elements is eliminated through a patterned noise filtering mechanism to ensure that the input of the subsequent named entity recognition module conforms to the geographical entity parsing specification. By removing irrelevant characters and correcting possible errors, the named entity recognition can be made more accurate, which helps to more accurately locate and parse the geographical entity names and their context information in the text to be translated. In one example, the source language address text can be cleaned through the following steps: a series of predefined rules are used to remove or correct unnecessary elements in the source language address text. For example, all non-alphanumeric characters (except for specific reserved characters, such as spaces) are filtered out, common spelling mistakes are corrected, and the format is unified (such as dates, phone numbers, etc.). In addition, natural language processing techniques are also used to identify and correct potential language errors or non-standard expressions. Specifically, by combining machine learning models, the system can automatically learn how to effectively clean the input text according to a large number of labeled data sets to ensure that the output is a purified source language address text with reasonable structure and clear content. In this way, the system can more efficiently extract useful information from the source language address text and provide high-quality input data for the next steps.

[0029] Next, the purified source language address text is tokenized to obtain a set of source language address text words. It should be understood that although the original address text has been cleaned and denoised, it still needs to be further decomposed into smaller information units so that the subsequent named entity recognition and place name disambiguation and other steps can be executed more accurately. Therefore, in the technical solution of this application, the purified source language address text is tokenized to obtain a set of source language address text words. The generated set of source language address text words can better support the system to identify which parts belong to place name entities (such as cities, street names), and which are non-place name information (such as house numbers, building names, etc.). This is crucial for correctly parsing the address structure and determining the relationship between each component in the subsequent steps.

[0030] Furthermore, named entity recognition and classification are performed on each source language address text word in the set of source language address text words to obtain a list of candidate place name entities and non-place name text segments. It should be understood that address text usually contains rich geographical information and other non-place name information (such as house numbers, unit names, etc.), and this information is crucial for accurately translating the entire address. Therefore, in the technical solution of this application, named entity recognition and classification are performed on each source language address text word in the set of source language address text words to obtain a list of candidate place name entities and non-place name text segments. That is, by performing named entity recognition and classification on each source language address text word in the set of source language address text words, the geographical entities and non-geographical entities in the address text are accurately distinguished, and the geographical entities are further classified and recognized to facilitate subsequent disambiguation and translation work. Specifically, through effective named entity recognition, the system can accurately extract the place name entities involved in the address, providing basic data support for subsequent fuzzy query and ambiguity elimination using the geographical knowledge base. At the same time, the separated non-place name text segments can also be used as part of the context information to help better understand and resolve the polysemy of place name entities. In one example, by adopting advanced natural language processing techniques to deeply analyze the cleaned and segmented address text, for example, using a pre-trained language model, the system can identify which words represent geographical entities such as countries, cities, and streets, and which words belong to non-place name entities (such as house numbers composed of numbers). During this process, the model not only relies on the information of the word itself but also considers its surrounding context to improve the recognition accuracy. In addition, this step may also combine with external knowledge bases (such as geographical databases) to enhance the credibility of the recognition results and ensure that the identified place name entities conform to the actual situation. In this way, the system can efficiently extract all relevant place name entities from the address text to form a candidate list and regard the remaining part as non-place name text segments, improving the overall accuracy and reliability of address translation.

[0031] Specifically, for S3 and S4, a first to-be-disambiguated place name entity is extracted from the list of candidate place name entities, and using the first to-be-disambiguated place name entity as a keyword, a fuzzy query is performed in the geographical knowledge base to obtain multiple candidate place name entity entries that match the first to-be-disambiguated place name entity. That is, by leveraging the rich geographical information provided by the geographical knowledge base, the ability to understand and parse place names is enhanced. Among them, the geographical knowledge base contains detailed geographical information, including but not limited to the spatial hierarchical relationship of place names, historical changes, and cultural backgrounds, etc. These information are crucial for correctly understanding and parsing place names. In the technical solution of this application, by matching and querying the to-be-disambiguated place name entity with the data in the geographical knowledge base, more background information about this place name can be obtained, such as its administrative division and surrounding geographical features. This not only helps to eliminate ambiguity but also provides a more comprehensive and accurate translation result. In addition, the fuzzy query allows the system to find possible matches even in the face of spelling mistakes or approximate names, further improving the robustness and practicality of the system. In one example, the following steps can be used to perform a fuzzy query in the geographical knowledge base with the first to-be-disambiguated place name entity as a keyword: First, a to-be-disambiguated place name entity is selected from the list of candidate place name entities obtained through word segmentation and named entity recognition. The selection criterion can be determined based on factors such as the importance of the place name in the text or its frequency of occurrence, etc.; then, using this to-be-disambiguated place name entity as a keyword, a fuzzy query operation is performed in the pre-constructed geographical knowledge base. The query results usually are a series of possible matching place name entity entries, and each entry carries detailed geographical attribute information. These information will subsequently be used in the subsequent disambiguation process. By combining context information and other non-place name text fragments, the system can more accurately determine the place name entity most suitable for the current context, thereby improving the accuracy and reliability of the entire address translation process.

[0032] Specifically, for S5, multiple candidate place name entity entries that match the first to-be-disambiguated place name entity and non-place name text fragments are input into the disambiguation model to obtain the disambiguated place name entity of the first to-be-disambiguated place name entity. Specifically, in a specific example of this application, as Figure 3 shown, S5 includes: S51, using the non-place name text fragments as context information, performing semantic-level disambiguation aggregation analysis on multiple candidate place name entity entries to obtain the semantic aggregation coding feature of the disambiguated place name entity query response; S52, performing semantic decoding on the semantic aggregation coding feature of the disambiguated place name entity query response to obtain the disambiguated place name entity of the first to-be-disambiguated place name entity.

[0033] Specifically, in S51, using non-place-name text fragments as context information, semantic-level disambiguation aggregation analysis is performed on multiple candidate place-name entity entries to obtain disambiguated place-name entity query response semantic aggregation coding features. In particular, in a specific example of the present application, as Figure 4 shown, S51 includes: S511, using non-place-name text fragments as context information, performing semantic understanding on the context information to obtain a semantic encoding vector of the place-name entity context information to be disambiguated; S512, using the candidate place-name entity entry embedding matrix to perform structured encoding on each candidate place-name entity entry in the multiple candidate place-name entity entries to obtain multiple candidate place-name entity entry semantic embedding encoding vectors; S513, performing place-name entity semantic disambiguation aggregation coding on the semantic encoding vector of the place-name entity context information to be disambiguated and the multiple candidate place-name entity entry semantic embedding encoding vectors to obtain a disambiguated place-name entity query response semantic aggregation coding vector as the disambiguated place-name entity query response semantic aggregation coding feature.

[0034] More specifically, in S511, using non-place-name text fragments as context information, semantic understanding is performed on the context information to obtain a semantic encoding vector of the place-name entity context information to be disambiguated. That is, in the embodiment of the present application, first, the place-name entity embedding matrix to be disambiguated is used to perform semantic embedding encoding on the place-name entity context information to be disambiguated to obtain a context sequence of the semantic embedding encoding vector of the place-name entity context information to be disambiguated. Since traditional methods often ignore the rich background information provided by non-place-name text fragments, it is difficult to make the correct choice when faced with place names with the same name but different meanings. By converting non-place-name text fragments into semantic embedding vectors, the system can better capture the association between non-place-name text fragments and the place-name entity to be disambiguated, thereby improving the accuracy of the disambiguation process. This not only helps to identify the place name that best fits the current context, but also enhances the robustness and adaptability of the entire translation process. In one example, the semantic embedding encoding of the context information can be performed through the following steps: First, construct a non-place-name text fragment embedding matrix, and then, for each address text containing the place-name entity to be disambiguated, extract the non-place-name part and convert it into the corresponding semantic embedding encoding vector of the place-name entity context information to be disambiguated through the above embedding matrix. During this process, the language model will consider the relationships within and between text fragments to ensure that the finally obtained semantic embedding encoding vector of the place-name entity context information to be disambiguated can accurately reflect the meaning of the non-place-name text fragment and its relevance to the place-name entity to be disambiguated.

[0035] Next, perform context semantic encoding on the context sequence of the semantic embedding encoding vector of the disambiguation place name entity context information based on a bidirectional LSTM model to obtain the semantic encoding vector of the disambiguation place name entity context information. Considering that it is difficult to capture the complex semantic relationships and context dynamic changes in non-place name text segments relying solely on static word vectors or simple embedding methods. Especially when dealing with polysemous place names, relying only on the information of the place name itself cannot effectively distinguish place name entities with different meanings, while context information provides additional clues to help identify the correct place name. In addition, in complex address structures, the key place name may be far from the words that provide important context information. Since the bidirectional LSTM model combines the forward and backward information flows, it can overcome the vanishing gradient problem encountered by traditional RNNs or LSTMs to some extent when dealing with long-term dependencies, thus making more effective use of the useful information in the entire sentence to assist in the disambiguation of place name entities. Therefore, in the technical solution of this application, perform context semantic encoding on the context sequence of the semantic embedding encoding vector of the disambiguation place name entity context information based on a bidirectional LSTM model to obtain the semantic encoding vector of the disambiguation place name entity context information. Among them, the bidirectional LSTM model can process the context sequence of the semantic embedding encoding vector of the disambiguation place name entity context information from both the forward and backward directions at the same time, and even those words that are far from the disambiguation place name entity but still provide important clues can be taken into account, greatly enhancing the model's ability to understand complex contexts. By capturing the context semantic connections in non-place name text segments through the bidirectional LSTM model, it is possible to provide a rich semantic background for each disambiguation place name entity more accurately.

[0036] More specifically, in S512, the candidate geographical name entity entry embedding matrix is used to perform structured encoding on each candidate geographical name entity entry among multiple candidate geographical name entity entries, so as to obtain multiple candidate geographical name entity entry semantic embedding coding vectors. That is, by converting each candidate geographical name entity into a vector representation in a high-dimensional space, the system can better understand and compare the differences between these candidate entities, thereby improving the accuracy of the disambiguation process. This structured encoding not only takes into account the characteristics of the geographical name itself, but also includes other geographical attributes related to it (such as the administrative division to which it belongs, adjacent landmarks, etc.), which helps to evaluate the fitness of candidate geographical name entities from multiple dimensions and finally select the translation result that best matches the current context. In one example, the structured encoding of each candidate geographical name entity entry among multiple candidate geographical name entity entries can be performed through the following steps: First, construct a candidate geographical name entity entry embedding matrix, which can map the input geographical name entity into a high-dimensional vector space, where each dimension represents the characteristics of the geographical name in a specific aspect. For each candidate geographical name entity entry from the geographical knowledge base, the system will calculate a corresponding candidate geographical name entity entry semantic embedding coding vector according to the specific information it contains (such as geographical location coordinates, administrative region to which it belongs, historical background, etc.). The candidate geographical name entity entry semantic embedding coding vector can accurately reflect the core characteristics of the original geographical name entity and its relationship with other entities. In this way, the system can effectively convert complex geographical information into a numerical form that is easy to process, facilitating semantic analysis and disambiguation aggregation coding in subsequent steps.

[0037] More specifically, in S513, semantic disambiguation aggregation encoding is performed on the semantic encoding vector of the to-be-disambiguated place name entity context information and the semantic embedding encoding vectors of multiple candidate place name entity entries to obtain a disambiguated place name entity query response semantic aggregation encoding vector as the disambiguated place name entity query response semantic aggregation encoding feature. It should be understood that in the prior art, when dealing with geographically named entities with the same name, due to insufficient context association and lack of integration of multi-dimensional attributes in the geographical knowledge base, it is impossible to effectively distinguish the deep semantic differences of fields with the same name. For example, when there is a lack of explicit geographical hierarchy identification in the source text, traditional methods relying on surface feature matching are prone to cross-domain mapping errors. Therefore, in order to achieve the deep coupling of text context features and geographical knowledge features, in the technical solution of this application, semantic disambiguation aggregation encoding is performed on the semantic encoding vector of the to-be-disambiguated place name entity context information and the semantic embedding encoding vectors of multiple candidate place name entity entries to obtain a disambiguated place name entity query response semantic aggregation encoding vector. That is, through a data-driven feature enhancement and dynamic weight allocation mechanism, the dimensional deviation of the heterogeneous feature space is eliminated, and the semantic weights of candidate entities are optimized based on global context awareness. Specifically, first, deconvolution encoding is used to enhance the features of the to-be-disambiguated context semantic encoding vector, so as to improve its representation ability through adaptive learning and ensure its alignment with the candidate entry semantic embedding encoding vector in terms of feature scale to eliminate the dimensional deviation in the cross-domain comparison process; subsequently, the enhanced context encoding vector and each candidate entry encoding vector are input into a single semantic query unit, and a set of to-be-disambiguated-candidate place name entity single semantic query score encoding vectors containing multi-dimensional semantic association information is generated through a non-linear transformation layer to capture the expressions of semantic relevance at different abstraction levels, such as implicit features such as administrative division subordination relationship and spatial proximity; further, the competition and cooperation relationship between candidate entries is modeled through a relational gating proxy module, which uses a gating mechanism to dynamically suppress noisy candidates and strengthen highly relevant entries; finally, weighted aggregation is performed on the set of to-be-disambiguated-candidate place name entity single semantic query score encoding vectors based on weights. In this process, dynamic information screening is realized through the self-attention mechanism to ensure that the aggregation result focuses on the geographical entity attributes strongly associated with the context. The generated disambiguated place name entity query response semantic aggregation encoding vector can capture the interaction between the implicit context constraints in non-place name text segments (such as the hierarchical implications of administrative division modifiers "province" and "city") and the geographical attributes of candidate entries (such as spatial topological relationships and administrative subordination degrees). In this way, the semantic sensitivity and geographical knowledge fusion ability of the disambiguation model are improved.

[0038] Specifically, first, feature enhancement based on deconvolution coding is performed on the semantic coding vector of the context information of the to-be-disambiguated place name entity to obtain the semantic enhanced coding vector of the context information of the to-be-disambiguated place name entity. Among them, the semantic enhanced coding vector of the context information of the to-be-disambiguated place name entity has the same feature scale as each candidate place name entity semantic embedding coding vector in the multiple candidate place name entity entries. It should be understood that the semantic coding vector of the context information of the to-be-disambiguated place name entity generated by the bidirectional LSTM may focus on local language patterns (such as the position features of address entities in the text sequence), while the semantic vectors generated by the candidate entity entry embedding matrix are based on the coding of the structured attributes of the geographical knowledge base (such as administrative division levels, spatial topological relationships, historical evolution information). The feature distributions and dimensional differences between the two directly restrict the accuracy of semantic alignment. Traditional methods suffer from semantic gaps during cross-modal matching due to the heterogeneity of the feature space. Through the deconvolution coding operation, the system uses a learnable deconvolution kernel network to perform non-linear feature expansion on the context coding, adaptively improving its representation ability, and at the same time forcing the output feature dimension to align with the scale of the candidate entry embedding coding to ensure comparability between the two in a unified semantic space. During this process, not only the limitations of manually designed interpolation rules are avoided, but also multi-channel convolution operations are used to capture multi-granularity associations of context semantics, such as implicit administrative division level relationships or functional scenario features. While retaining the original semantic topology structure, the semantic enhanced coding vector of the context information of the to-be-disambiguated place name entity after feature enhancement can establish deep associations with the candidate entry coding in terms of attributes such as spatial proximity and membership. In a specific example, the following feature enhancement formula is used to perform feature enhancement based on deconvolution coding on the semantic coding vector of the context information of the to-be-disambiguated place name entity to obtain the semantic enhanced coding vector of the context information of the to-be-disambiguated place name entity; where the feature enhancement formula is:

[0039] ;

[0040] Among them, is the semantic coding vector of the context information of the to-be-disambiguated place name entity, represents deconvolution coding, represents the deconvolution weight matrix, represents the one-norm of the vector, is the semantic enhanced coding vector of the context information of the to-be-disambiguated place name entity.

[0041] Next, the semantic enhanced encoding vector of the to-be-disambiguated place name entity context information and each of the semantic embedding encoding vectors of multiple candidate place name entity entries are respectively input into the place name entity single semantic query unit to obtain a set of initial to-be-disambiguated - candidate place name entity single semantic query score encoding vectors. Considering that traditional methods relying on surface feature similarity metrics (such as cosine similarity) cannot effectively capture the deep semantic associations between the context and candidate entries. For example, the context of "Paris" in the source text may imply educational attributes (such as "University"), while the homonymous candidate entries in the geographical knowledge base (such as Paris in France and Paris in Texas) have essential differences in structured attributes such as administrative level and functional scenario. Through the single semantic query unit, the system constructs an implicit semantic representer based on a deep neural network architecture to measure the relevance between the context semantic enhanced encoding and the candidate entry embedding encoding from multiple dimensions, breaking through the semantic gap of traditional methods. Specifically, the single semantic query unit adopts a non-linear transformation layer and a cross-attention mechanism to perform cross-modal interaction between the semantic enhanced encoding vector of the to-be-disambiguated place name entity context information (representing functional scenarios such as "University") and the semantic embedding encoding vector of the candidate place name entity entry (including attributes such as administrative affiliation and historical evolution), generating a multi-dimensional initial to-be-disambiguated - candidate place name entity single semantic query score encoding vector. Among them, different dimensions respectively represent the spatial proximity matching degree, functional scenario consistency, or historical name association strength, rather than a single scalar score. By capturing deep semantic associations, it provides highly discriminative inputs for the relational gating agent module, ultimately achieving the topological consistency and geographical logical integrity of cross-language address translation. In a specific example, the semantic enhanced encoding vector of the to-be-disambiguated place name entity context information and each of the semantic embedding encoding vectors of multiple candidate place name entity entries are respectively input into the place name entity single semantic query unit according to the following semantic query formula to obtain a set of initial to-be-disambiguated - candidate place name entity single semantic query score encoding vectors; where the semantic query formula is:

[0042] ;

[0043] ;

[0044] Wherein, is the semantic embedding encoding vectors of the multiple candidate place name entity entries, are respectively the 1st, 2nd, th, and th semantic embedding encoding vectors among the semantic embedding encoding vectors of multiple candidate place name entity entries, represents function, denotes the semantic weight matrix, denotes the bias term, is the corresponding initial disambiguation-candidate toponym entity monomer semantic query score encoding vector.

[0045] In particular, it should be understood that when the set of the initial disambiguation-candidate toponym entity monomer semantic query score encoding vectors is obtained through feature concatenation and simple interaction calculation, its distribution in the semantic query encoding space may not accurately reflect the complex association between the context and the geographical knowledge base due to insufficient dynamic characteristics of the features or spatial alignment deviation. For example, when there are entities with the same name but different geographical locations in the address text (such as "Paris" may correspond to Paris, France or Paris, Texas, USA), traditional static encoding methods are difficult to capture the dynamic influence of non-toponym context (such as adjacent words "Texas" or "France") on entity disambiguation, resulting in a deviation in the weight assignment of candidate entity matching. Therefore, preferably, the set of the initial disambiguation-candidate toponym entity monomer semantic query score encoding vectors is subjected to feature dynamic optimization mapping to obtain the set of the disambiguation-candidate toponym entity monomer semantic query score encoding vectors. That is, a microdifferential manifold alignment mechanism of cross-modal features is established through feature dynamic optimization mapping to transform the set of the initial disambiguation-candidate toponym entity monomer semantic query score encoding vectors from the static feature space to the dynamically optimized semantic query space, thereby enhancing the semantic alignment ability of the feature distribution and enabling it to more flexibly reflect the association between the context and the geographical knowledge base. Specifically, the disambiguation-candidate toponym entity monomer semantic dynamic eigenvector representation is constructed by cascading features ontology, and the disambiguation-candidate toponym entity monomer semantic interaction potential vector is calculated, introducing the dynamic characteristics in the mapping process, so that the feature representation can adapt to context changes through the combination of linear transformation and residual connection while retaining the original semantic information. Further, the inherent alignment mechanism adjusts the feature projection path under the condition of gauge symmetry through the coupling constant and the covariant matrix, ensuring that the alignment of the feature space and the semantic query space not only meets the requirements of dynamic optimization but also avoids instability caused by noise amplification. In addition, the introduction of the covariant matrix enables the feature projection process to follow the spatial hierarchical relationship (such as the "province-city-district" structure) and historical naming features (such as the old name "New Towne") in the geographical knowledge base, realizing cross-verification of multi-dimensional geographical attributes. In this way, the optimized set of the initial disambiguation-candidate toponym entity monomer semantic query score encoding vectors not only reduces the translation deviation caused by language polysemy but also enables the model to maintain high robustness through the synergistic effect of dynamic characteristics and inherent alignment, ensuring the strict consistency between the target language address text and the real geographical location.

[0046] Specifically, when the single-entity semantic query unit of the place name entity performs single-entity semantic query score encoding through the direct feature splicing of the semantic enhanced encoding vector of the place name entity context information to be disambiguated and the semantic embedding encoding vector of the corresponding candidate place name entity entry it is expected to improve the representation accuracy of the single-entity semantic query score by optimizing the dynamic characteristic representation from the feature space to the semantic query encoding space and enhancing the inherent alignment characteristic between the feature space and the semantic query encoding space.

[0047] Based on this, if the concatenated feature of the semantic enhanced encoding vector of the place name entity context information to be disambiguated and the semantic embedding encoding vector of the corresponding candidate place name entity entry is denoted as i.e., then the single-entity semantic interaction potential vector between the place name entity to be disambiguated and the candidate place name entity under the mapping representation is calculated first:

[0048] ;

[0049] That is, the concatenated feature ontology is used as the single-entity semantic dynamic eigenvector representation of the place name entity to be disambiguated - candidate place name entity to obtain the projection representation to the eigenstate under the spatial mapping process.

[0050] Then, the inherent alignment between the feature space and the semantic query encoding space is performed:

[0051] ;

[0052] where is the coupling constant, calculated in the same way as in the case of deconvolution enhancement to maintain symmetry, i.e.:

[0053] ;

[0054] and is the covariant matrix, i.e., let to obtain.

[0055] Thus, with the single-entity semantic interaction potential vector as the generator of the inherent alignment, the representation accuracy of the single-entity semantic query score is improved by using the covariant path under the gauge symmetry condition as the mapping gauge fixing guarantee.

[0056] Furthermore, based on the self-distribution characteristics of the feature set of the set of disambiguation-candidate place name entity single-semantic query score encoding vectors, the single-semantic matching degree of each disambiguation-candidate place name entity single-semantic query score encoding vector in the set of disambiguation-candidate place name entity single-semantic query score encoding vectors is determined to obtain a set of disambiguation-candidate place name entity single-semantic matching degrees. It should be understood that when the traditional method isolates and evaluates the candidate entity matching degree, it ignores the dynamic influence of the global context on the individual semantic association. For example, when the distributions of the disambiguation-candidate place name entity single-semantic query score encoding vectors of multiple candidate entities show significant differences, relying only on the absolute score is vulnerable to noise interference and cannot distinguish the relative importance of highly relevant entries in the whole. In the technical solution of this application, the adaptive optimization of the matching degree is realized through the context awareness mechanism to avoid semantic deviation caused by local evaluation. Specifically, the system uses probability density estimation or distribution fitting methods to analyze the distribution form (such as long-tailed distribution, normal distribution) of the set of disambiguation-candidate place name entity single-semantic query score encoding vectors, and generates a matching degree weight according to the degree of deviation of each disambiguation-candidate place name entity single-semantic query score encoding vector from the distribution center (such as Z-score standardization). For example, if the score vector of a certain candidate entity is significantly higher than the distribution mean in the multi-dimensional semantic space, its matching degree will be strengthened; conversely, vectors close to or lower than the mean will be suppressed. In this way, the misjudgment problem caused by the static threshold or isolated matching in the traditional method is solved, and the context consistency of the disambiguation decision is improved through dynamic weight assignment, supporting the reliable application of high-precision address translation in multi-language service scenarios. In a specific example, the single-semantic matching degree of each disambiguation-candidate place name entity single-semantic query score encoding vector in the set of disambiguation-candidate place name entity single-semantic query score encoding vectors is determined by the following semantic matching formula to obtain a set of disambiguation-candidate place name entity single-semantic matching degrees; where the semantic matching formula is:

[0057] ;

[0058] wherein, denotes function, is the number of candidate place name entity semantic embedding encoding vectors in the semantic embedding encoding vectors of the multiple candidate place name entity entries, denotes exponential operation, denotes vector multiplication, denotes the square of the one-norm of the vector, is the transpose of the k-th disambiguation-candidate place name entity single-semantic query score encoding vector, is corresponding disambiguation-candidate place name entity single-semantic matching degree.

[0059] Then, the set of single - entity semantic matching degrees of the disambiguation - candidate geographical name entities is input into the relational gating proxy module to obtain the set of self - attention weights of the single - entity query semantics of the disambiguation - candidate geographical name entities. It should be understood that when the system preliminarily filters out multiple candidate geographical name entities through semantic matching degrees (for example, multiple "Paris" with the same name but different geographical locations), directly relying on the original matching degrees for attention weight assignment may face two core problems: First, the matching degree value only reflects the surface semantic association between the text segment and the candidate entity, and cannot capture the implicit geographical hierarchy relationship in the address text (such as the state - belonging relationship in "Paris, Texas") or historical naming features (such as the geographical scope difference of "New York" in different historical stages); Second, the traditional self - attention mechanism has limited ability to integrate multi - source heterogeneous information (such as street descriptions and administrative division modifiers in the address text), and is prone to weight assignment deviation due to local noise interference. In the technical solution of this application, the original matching degree is non - linearly transformed through a gating function. While retaining the basic semantic association strength, predefined spatial relationship constraints in the geographical knowledge base (such as the subordination relationship between the administrative region to which the candidate entity belongs and the superior administrative division mentioned in the context) and semantic association rules (such as the co - existence probability of natural geographical entities and cultural landmarks) are injected. In this way, the robustness of the model to noise interference is improved, the semantic consistency of the disambiguation decision is enhanced, the semantic deviation problem caused by static weight assignment in the traditional method is solved, and fine - grained context - aware control is achieved through the gating mechanism, providing high - discrimination weight input for subsequent self - attention aggregation, and supporting the accurate mapping of the geographical topology integrity and administrative hierarchy logic of cross - language address translation. In a specific example, the set of single - entity semantic matching degrees of the disambiguation - candidate geographical name entities is input into the relational gating proxy module according to the following gating formula to obtain the set of self - attention weights of the single - entity query semantics of the disambiguation - candidate geographical name entities; where the gating formula is:

[0060] ;

[0061] Wherein, represents the gating function, is a preset threshold, represents the self - attention weight of the single - entity query semantics of the corresponding disambiguation - candidate geographical name entity.

[0062] Further, based on the set of semantic self-attention weights for the disambiguation-candidate place name entity monomers, the set of semantic query score encoding vectors for the disambiguation-candidate place name entity monomers is aggregated to obtain a disambiguated place name entity query response semantic aggregation encoding vector. That is, through a dynamic weight assignment mechanism, namely the self-attention mechanism, the importance of each candidate place name entity is flexibly adjusted, so as to achieve attention to key semantic information and ignore irrelevant or noisy information. This method of weighted aggregation not only improves the model's ability to understand complex contexts, but also makes the final translation result more accurate and reliable. In a specific example, the following aggregation formula is used to aggregate the set of semantic query score encoding vectors for the disambiguation-candidate place name entity monomers to obtain a disambiguated place name entity query response semantic aggregation encoding vector; where, the aggregation formula is:

[0063] ;

[0064] Wherein, is the disambiguated place name entity query response semantic aggregation encoding vector.

[0065] Specifically, the S52 performs semantic decoding on the disambiguated place name entity query response semantic aggregation encoding features to obtain the disambiguated place name entity of the first disambiguation-candidate place name entity. That is, in the technical solution of the present application, semantic decoding based on the RNN model is performed on the disambiguated place name entity query response semantic aggregation encoding vector to obtain the disambiguated place name entity of the first disambiguation-candidate place name entity. It should be understood that although the disambiguated place name entity query response semantic aggregation encoding vector integrates context information and the structured features of the geographical knowledge base, it is essentially still a complex sequence containing multi-dimensional semantic associations. If the traditional method directly selects candidate entities through simple classification or rule matching, it is easy to cause misjudgment due to ignoring the implicit dynamic dependence relationships in the sequence. The RNN model, through its recurrent network structure, can parse the semantic segments in the aggregation encoding step by step in time, capture the strongly associated parts of the disambiguated place name entity query response semantic aggregation encoding vector, suppress the redundant information in the vector, and thus dynamically construct a disambiguation decision path.

[0066] Specifically, in step S6, steps S3 to S5 are repeatedly executed to obtain a list of disambiguated place name entities. It should be understood that the source language address text may contain multiple place name entities, each with its specific context and may have complex relationships with other place name entities. Therefore, processing only a single place name entity cannot meet the overall translation requirements. In the technical solution of this application, steps S3 to S5 are repeatedly executed to obtain a list of disambiguated place name entities. By gradually parsing and resolving the ambiguities of each place name entity to be translated, a complete and unambiguous list of place name entities is constructed. Among them, the list of disambiguated place name entities ensures that each place name has been analyzed and verified in detail. By disambiguating each place name entity in the source language address text one by one, the exact meaning of each place name in a specific context can be effectively captured, thereby reducing the problem of geographical information distortion caused by incorrect identification or translation.

[0067] Specifically, in step S7, based on the list of disambiguated place name entities, the target language address text is obtained. In the technical solution of this application, the list of disambiguated place name entities is input into a pre-trained neural machine translation model to obtain the target language address text. Among them, the neural machine translation model will translate each place name entity according to the learned language rules, vocabulary mapping relationships, and knowledge in a specific domain. It should be noted that during the translation process, the model not only focuses on the conversion of a single place name entity, but also comprehensively considers the characteristics of the entire address structure to ensure the logical coherence and consistency among various parts. For example, the neural machine translation model will pay attention to details such as whether the hierarchical relationship of administrative divisions is correct and whether the combination of street names and house numbers is reasonable. Finally, after this series of complex calculations and adjustments, the system will output the target language address text. In this way, the comprehensive and accurate conversion from the source language address text to the target language address text is completed. Through this process, it can be ensured that users in different languages can understand and use the converted address information without obstacles.

[0068] In summary, the method for enhancing the translation accuracy of place name addresses by using context awareness and geographical library according to the embodiments of this application is elucidated. By integrating the context awareness mechanism and the multi-dimensional attributes of the geographical knowledge base, a dynamic collaborative disambiguation framework is constructed to break through the semantic limitations of traditional place name address translation. Specifically, first, the context semantics of non-place name text fragments is encoded to capture the dynamic association information hidden in the address text, and the candidate entities are embedded with structured features in combination with the geographical knowledge base to extract the multi-dimensional attributes of the candidate entities, so as to achieve cross-modal disambiguation decision-making in the semantic aggregation stage. In this way, through the geographical feature cross-verification mechanism guided by context semantics, the translation ambiguity problem caused by static rule dependence and geographical attribute fragmentation in the prior art can be effectively solved, thereby achieving a more accurate multi-language address conversion effect.

[0069] The embodiments of the present disclosure have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to technologies in the market, or to enable other ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for enhancing the accuracy of place name and address translation using context awareness and geographic database, characterized in that: include: S1: receiving a source language address text to be translated; S2: Perform standardized preprocessing and structured semantic parsing on the source language address text to obtain a candidate list of place name entities and non-place name text fragments; S3: extracting a first place-name entity to be disambiguated from the place-name entity candidate list; S4: extracting multiple candidate place name entity entries matching the first place name entity to be disambiguated from the geographic knowledge base; S5: Inputting a plurality of candidate place name entity entries and non-place name text segments matching the first place name entity to be disambiguated into an ambiguity resolution model to obtain a disambiguated place name entity of the first place name entity to be disambiguated, including: Using non-place name text fragments as context information, semantic understanding of the context information is performed to obtain a semantic encoding vector of the context information of the place name entity to be disambiguated; Using the candidate place name entity entry embedding matrix, structurally encoding each of the multiple candidate place name entity entries is performed to obtain semantic embedding encoding vectors of the multiple candidate place name entity entries; Performing feature enhancement based on deconvolution coding on the semantic encoding vector of the context information of the place name entity to be disambiguated, so as to obtain the semantic enhanced encoding vector of the context information of the place name entity to be disambiguated; Performing a single semantic query on the place name entity by using the semantic enhancement encoding vector of the context information of the place name entity to be disambiguated and the semantic embedding encoding vectors of each candidate place name entity item in the semantic embedding encoding vectors of multiple candidate place name entity items, so as to obtain a set of single semantic query score encoding vectors of the candidate place name entity to be disambiguated; Based on the self-distribution characteristics of the feature set of the set of monomer semantic query score encoding vectors of the to-be-disambiguated-candidate place name entities, determine the monomer semantic matching degree of each to-be-disambiguated-candidate place name entity monomer semantic query score encoding vector in the set of monomer semantic query score encoding vectors of the to-be-disambiguated-candidate place name entities, so as to obtain a set of monomer semantic matching degrees of the to-be-disambiguated-candidate place name entities; The set of semantic matching degrees of entity monomers of to-be-disambiguated and candidate place names is input into the relation gating agent module to obtain the set of semantic self-attention weights of entity monomers of to-be-disambiguated and candidate place names; Based on the set of semantic self-attention weights of the to-be-disambiguated-candidate place name entity monomer query, the set of to-be-disambiguated-candidate place name entity monomer semantic query score encoding vectors is aggregated to obtain the disambiguated place name entity query response semantic aggregate encoding vector and use it as the disambiguated place name entity query response semantic aggregate encoding feature; Semantically decoding the semantically aggregated encoding features of the disambiguated place name entity query response to obtain a disambiguated place name entity of the first place name entity to be disambiguated; Step 6: Loop through S3 to S5 to obtain a list of disambiguated place name entities; Step 7: Based on the disambiguated place name entity list, obtain the target language address text.

2. The method for enhancing the accuracy of place name and address translation by using context awareness and geographic library according to claim 1, characterized in that S2 include: Perform text cleaning on the source language address text to obtain the purified source language address text: Performing word segmentation processing on the purified source language address text to obtain a set of source language address text words; Named entity recognition and classification are performed on each source language address text word in the set of source language address text words to obtain a place name entity candidate list and a non-place name text segment.

3. The method for enhancing the accuracy of place name and address translation by using context awareness and geographic library according to claim 2, characterized in that: S4 includes: using the first place-name entity to be disambiguated as a keyword, performing a fuzzy query in a geographic knowledge base to obtain a plurality of candidate place-name entity entries that match the first place-name entity to be disambiguated.

4. The method for enhancing the accuracy of place name and address translation by using context awareness and geographic library according to claim 3 is characterized in that: Using non-place name text fragments as context information, the context information is semantically understood to obtain the semantic encoding vector of the context information of the place name entity to be disambiguated, including: Using the embedding matrix of the place name entity to be disambiguated, semantic embedding encoding is performed on the context information of the place name entity to be disambiguated, so as to obtain a context sequence of the semantic embedding encoding vector of the context information of the place name entity to be disambiguated; The context sequence of the semantic embedding encoding vector of the context information of the place name entity to be disambiguated is subjected to context semantic encoding based on the bidirectional LSTM model to obtain the semantic encoding vector of the context information of the place name entity to be disambiguated.

5. The method for enhancing the accuracy of place name and address translation by using context awareness and geographic library according to claim 4, characterized in that: The semantic enhancement encoding vector of the context information of the place name entity to be disambiguated and the semantic embedding encoding vectors of each candidate place name entity item in the semantic embedding encoding vectors of multiple candidate place name entity items are used for place name entity monomer semantic query to obtain a set of to-be-disambiguated-candidate place name entity monomer semantic query score encoding vectors, including: The semantic enhancement coding vector of the context information of the place name entity to be disambiguated and the semantic embedding coding vectors of each candidate place name entity item in the multiple candidate place name entity item semantic embedding coding vectors are respectively input into the place name entity monomer semantic query unit to obtain a set of initial to-be-disambiguated-candidate place name entity monomer semantic query score coding vectors; The set of initial to-be-disambiguated-candidate place name entity monomer semantic query score encoding vectors is subjected to dynamic feature optimization mapping to obtain a set of to-be-disambiguated-candidate place name entity monomer semantic query score encoding vectors.

6. The method for enhancing the accuracy of place name and address translation using context awareness and geographic library according to claim 5, characterized in that: Semantically decoding the semantic aggregation encoding features of the disambiguation place name entity query response to obtain the disambiguation place name entity of the first place name entity to be disambiguated, including: The semantically aggregated encoding vector of the disambiguated place name entity query response is semantically decoded based on the RNN model to obtain the disambiguated place name entity of the first place name entity to be disambiguated.

7. The method for enhancing the accuracy of place name and address translation by using context awareness and geographic library according to claim 6, characterized in that: Based on the disambiguated place name entity list, the target language address text is obtained, including: The disambiguated place name entity list is fed into a pre-trained neural machine translation model to obtain the target language address text.

Images