A multilingual place name translation method combining syllable segmentation and adaptive learning

Abstract

The present application relates to the field of place name translation technology and discloses a place name translation method that combines multilingual syllable segmentation with adaptive learning. The method performs word segmentation and common name translation annotation on the source language place name address based on a general dictionary, converts the unannotated proper noun portion into an IPA phoneme sequence, and then performs syllable segmentation and translation matching on the IPA phoneme sequence of each word segment. Multiple possible candidate transliteration results are matched for each word segment. The candidate transliteration results of the proper noun segment are then combined with the common name translation result to construct a candidate set of place name and address target language translation results. The candidate set is then adaptively learned and aggregated using a deep learning algorithm to comprehensively consider the semantic information and contextual coherence of multiple transliteration versions to generate the final place name and address target language translation. The present application can solve the problems of inaccurate, stiff, and ambiguous transliterations in traditional place name and address translation methods, thereby improving translation efficiency and quality.

Description

Technical Field

[0001] The present application relates to the technical field of place name translation, and more specifically, to a place name translation method that combines multilingual syllable segmentation with adaptive learning. Background Art

[0002] In today's increasingly globalized world, it has become crucial to accurately and efficiently translate place names and addresses between different languages. Whether it is international trade, logistics and transportation, cross-border tourism, or emergency rescue and geographic information system applications, the clarity and accuracy of place names and addresses are directly related to the smooth progress of activities and the effective transmission of information. However, due to language differences, cultural backgrounds, naming habits, and the diversity of writing systems, the translation of place names and addresses is far from a simple word replacement. Place names and addresses usually contain a large number of proper nouns (proper nouns), which often have no direct corresponding target language vocabulary and need to be converted through transliteration. Existing place name and address translation solutions mainly rely on pre-established comparison dictionaries or rule-based transliteration systems, or attempt to use general statistical machine translation. Although these methods can handle the translation problems of some place names and addresses to a certain extent, they generally have limitations.

[0003] Specifically, due to the vast differences in phonological systems across languages, including inconsistent phonemic inventories, varying syllable structure rules, and restrictions on legal phoneme combinations (phonotactics), a single source syllable may be approximated by multiple target transliteration units. This makes it difficult for direct, simple one-to-one or fixed-rule mappings to capture all transliteration possibilities and select the optimal solution, often resulting in inaccurate, stilted, and even ambiguous transliterations. General machine translation models are also prone to producing errors or non-compliant transliteration results when processing place names and addresses containing a large number of proper nouns, making existing methods unable to cope with the complex and diverse translation needs of place names and addresses with significant cross-linguistic differences.

[0004] Therefore, it is necessary to provide an optimized place name translation method that combines multilingual syllable segmentation with adaptive learning 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 the present application, a method for translating place names by combining multilingual syllable segmentation with adaptive learning is provided, which includes:

[0007] Obtain the source language address and target translation language input by the user;

[0008] Based on a universal dictionary and the target translation language, splitting, matching and translation annotation are performed on the source language place name and address to be translated to obtain a sequence of source language place name and address word segments containing common name translation annotations;

[0009] Performing transliteration matching based on syllable segmentation on each unlabeled source language place name and address word segment to be translated in the sequence of source language place name and address word segments containing common name translation labels to obtain a target language transliteration result candidate set for each unlabeled source language place name and address word segment to be translated;

[0010] Assembling the candidate sets of target language transliteration results of the unlabeled source language place name and address word segments to be translated with the sequence of source language place name and address word segments containing common name translation labels to obtain a candidate set of place name and address target language translation results;

[0011] Adaptive learning aggregation optimization is performed on the candidate set of the place name and address target language translation results to obtain the place name and address target language translation.

[0012] Beneficial effects: Compared with the existing technology, the place name translation method provided by this application combines multilingual syllable segmentation with adaptive learning. It performs word segmentation and common name translation annotation on the source language place name address based on a general dictionary, and converts the unannotated proper noun part into an IPA phoneme sequence. Then, based on the multi-source place name historical transliteration knowledge base, it performs syllable division and translation matching on the IPA phoneme sequence of each word segment, matches multiple possible candidate transliteration results for each word segment, and then combines the various candidate transliteration results of the proper noun segment with the common name translation result to construct a candidate set of target language translation results of multiple transliteration versions of the place name address. The candidate set is then adaptively learned and aggregated using a deep learning algorithm to comprehensively consider the semantic information and contextual coherence of the multiple transliteration versions to generate the final target language translation of the place name address. This application can solve the problems of inaccurate, stiff, and ambiguous transliteration in traditional place name and address translation methods, and improve translation efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0014] Figure 1 The flowchart of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application.

[0015] Figure 2 Schematic diagram of data flow of a place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application.

[0016] Figure 3 This is a flowchart of sub-step S2 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application.

[0017] Figure 4 This is a flowchart of sub-step S3 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application.

[0018] Figure 5 This is a flowchart of sub-step S5 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application.

[0019] Figure 6 This is a flowchart of sub-step S52 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application.

[0020] Figure 7 This is a flowchart of sub-step S523 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] As used in this application and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not intended to refer to the singular but may include the plural. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements.

[0022] Although the present application makes various references to certain modules in the system according to embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are illustrative only, and different aspects of the system and method can use different modules.

[0023] Flowcharts are used in this application to illustrate the operations performed by the systems according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the various steps may be processed in reverse order or simultaneously, as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

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

[0025] It should be noted in advance that all data processing in this application is carried out in compliance with the relevant local data protection laws and policies and with the authorization of the relevant authority administrator.

[0026] Figure 1 The flowchart of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the place name translation method combining multilingual syllable segmentation and adaptive learning according to the embodiment of the present application. Figure 1 and Figure 2 As shown, the multilingual place name translation method combining syllable segmentation with adaptive learning includes the following steps: S1, obtaining a source language place name address to be translated and a target translation language input by a user; S2, splitting, matching, and translation annotation of the source language place name address to be translated based on a universal dictionary and the target translation language to obtain a sequence of source language place name and address word segments containing common name translation annotations; S3, performing transliteration matching based on syllable segmentation on each unannotated source language place name and address word segment to be translated in the sequence of source language place name and address word segments containing common name translation annotations to obtain a target language transliteration result candidate set of each unannotated source language place name and address word segment to be translated; S4, assembling the target language transliteration result candidate set of each unannotated source language place name and address word segment to be translated with the sequence of source language place name and address word segments containing common name translation annotations to obtain a candidate set of place name and address target language translation results; S5, performing adaptive learning aggregation optimization on the candidate set of place name and address target language translation results to obtain a place name and address target language translation.

[0027] In the place name translation method that combines multilingual syllable segmentation with adaptive learning, the step S1 obtains the source language place name address to be translated and the target translation language input by the user. Specifically, in order to establish the initial parameters of the translation task and ensure that the system is adapted to the multilingual environment, the present application is based on user interaction and multilingual input interface design, and dynamically configures the language rule library and transliteration knowledge base of the subsequent processing flow by receiving the source language place name address and the specified target translation language input by the user. In the specific implementation process, the original place name and target language identifier (such as "English") input by the user are captured through the front-end interface or API interface, and the legality of the input format is verified (such as filtering non-text characters). In this way, the translation direction can be accurately locked, and language type constraints can be provided for subsequent word segmentation, transliteration and semantic optimization to avoid rule conflicts caused by unclear target language.

[0028] During implementation, users may directly type in a complete address string, such as "Jianguomenwai Street, Chaoyang District, Beijing," or they may enter a specific place name, such as "Xiangshan." Furthermore, users must explicitly specify the target translation language, such as English, Japanese, or French. This target language identifier not only determines the selection of subsequent transliteration rules but also influences the behavior logic of the multilingual dictionary matching module. To improve user convenience, the system typically provides drop-down menus and language code selectors to guide users in correctly entering the target language type, preventing the entire translation process from failing due to language recognition errors.

[0029] After receiving the original input, the system immediately initiates a format validation mechanism. Since place names and addresses often contain non-standard characters such as spaces, numbers, and special symbols, the input content needs to be cleaned and normalized. For example, multiple consecutive spaces need to be removed, illegal control characters need to be filtered, and uppercase and lowercase formats need to be unified. For Chinese input, it is also necessary to determine whether pinyin or other Latin alphabet spellings are mixed; and for non-Latin writing systems such as Arabic and Russian, character encoding conversion is required to ensure that subsequent processing modules can parse correctly. In addition, the system also needs to identify and separate the common name part (such as "district", "road", "avenue") and the proper name part (such as "Xiangshan") in the address so that different translation strategies can be used for subsequent processing.

[0030] Taking into account the complexity of the multilingual environment, the system must have flexible language detection capabilities. When the user does not explicitly specify the target language, the system can try to automatically infer the target language type based on contextual clues, but for accuracy reasons, it is still recommended that the user actively select the target language. The language detection module is trained based on large-scale multilingual texts. It can quickly identify the language system to which the input place name belongs, and call the corresponding phoneme mapping rule library in combination with the target language settings. For example, when a Chinese place name is translated into English, the system will give priority to calling the mapping table from Chinese pinyin to English pronunciation habits; when a Chinese place name is translated into Japanese, a transliteration rule set based on Katakana will be enabled.

[0031] In order to improve the user experience and enhance the intelligence level of the system, the system has also introduced input suggestions and error correction mechanisms. During the user input process, the system can dynamically prompt possible place name options based on some of the characters that have been entered, reducing the interference caused by typing errors. In addition, after submitting the complete address, the system will perform a grammatical structure analysis on the input content to identify whether there are spelling errors, duplications or illogical combinations. For example, if the user enters "Beijing District, Beijing", the system may prompt that "Beijing District" may have spelling deviations and recommends correcting it to "Dongcheng District" or "Xicheng District" and other standardized place name expressions.

[0032] Throughout the entire input processing process, the system adheres to the design philosophy of "dynamic configuration." This means that based on the source language place name and address and target language identifier provided by the user, the corresponding processing rule library and transliteration knowledge base are loaded in real time. This on-demand loading method not only improves the system's response speed but also enhances its adaptability to different language combinations. For example, when translating from Chinese to English, the system uses an IPA (International Phonetic Alphabet)-based syllable segmentation model and an English pronunciation rule database. When translating from Chinese to German, the system switches to the German phonetic rule library to ensure that the generated transliteration conforms to the pronunciation habits of the target language.

[0033] In the above-mentioned multilingual syllable segmentation and adaptive learning combined place name translation method, step S2 is to split, match and translate the source language place name address to be translated based on a universal dictionary and the target translation language to obtain a sequence of source language place name address word segments containing common name translation annotations. Figure 3 Flowchart of sub-step S2 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application. Figure 3As shown, step S2 includes the steps of: S21, performing word segmentation processing on the source language place name and address to be translated based on universal dictionary matching to obtain a sequence of source language place name and address word fragments to be translated; S22, based on the universal dictionary and the target translation language, performing translation annotation on each source language place name and address word fragment to be translated in the sequence of source language place name and address word fragments to obtain the sequence of source language place name and address word fragments containing common name translation annotations.

[0034] Specifically, the step S21 performs word segmentation processing on the source language place name address to be translated based on universal dictionary matching to obtain a sequence of word fragments of the source language place name address to be translated. It should be understood that since the place name address is usually composed of multiple words or phrases, and the address contains multiple components such as street names, district names, building names, directions, etc., different address components require different translation strategies and processing methods. Therefore, in order to separate the common name and the proper name parts in the source language place name address to be translated and give priority to the common part that can be translated, the present application is based on the rule matching mechanism of the multilingual geographic term dictionary, and performs word boundary recognition and segmentation on the source language place name address to be translated through a predefined universal dictionary (covering common common names and their multilingual translation comparison table). During the specific implementation process, a bidirectional maximum matching algorithm (BMM) is used to scan the source language place name address to be translated from left to right and from right to left to identify the common name segments in the source language place name address to be translated, and the place name address is divided into common name segments and unmatched proper name segments according to the matching results, so as to preliminarily distinguish the common parts of the place name address that can be directly translated (such as "street", "district") and the proper names that require further transliteration (such as specific place names, building names), laying the foundation for subsequent processing.

[0035] Specifically, in step S22, based on the general dictionary and the target translation language, each source language place name and address word segment in the sequence of source language place name and address word segments to be translated is translationally annotated to obtain the sequence of source language place name and address word segments containing the generic name translation annotation. It should be understood that since the translation of generic names in place names and addresses needs to follow the standardized translation rules of the target language (for example, the Chinese "路" corresponds to the English "Road" rather than the transliteration "Lu"), and no complex transliteration processing is required. Therefore, based on the cross-language generic name mapping table stored in the general dictionary, this application performs semantic annotation on the segmented generic name segments through dictionary query. In the specific implementation process, by traversing the sequence of source language place name and address word segments to be translated, for each identified generic name segment, the pre-stored target language translation in the general dictionary is queried, and the translation result is used as the translation annotation for this generic name segment. For example, if a segment matches the generic name entry in the general dictionary (such as "山" being marked as a generic name in Chinese), it is directly replaced with the standard translation in the target language (such as "Mountain"), and its position marker is retained; while the segments that do not match are marked as proper names to be transliterated. In this way, the standardized translation in the target language can be quickly and accurately assigned to the generic name part of the place name and address, providing a basis for subsequent transliteration and semantic optimization steps.

[0036] In the above place name translation method combining multilingual syllable segmentation and adaptive learning, in step S3, for each unannotated source language place name and address word segment in the sequence of source language place name and address word segments containing the generic name translation annotation, syllable segmentation-based transliteration matching is performed to obtain the candidate set of target language transliteration results for each unannotated source language place name and address word segment. Specifically, since proper name transliteration needs to solve the problem of cross-language phonological mismatch, therefore, in order to capture the cross-language phoneme correspondence and generate transliteration results, based on the phonological neutrality characteristic of the International Phonetic Alphabet (IPA), this application constructs a phoneme-transliteration unit probability mapping table through a multi-source place name historical transliteration knowledge base, and uses the historical transliteration data in the transliteration knowledge base to query the possible transliteration results of each unannotated source language place name and address word segment. Among them, Figure 4 is a flowchart of sub-step S3 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application. As Figure 4As shown, step S3 includes steps: S31, performing IPA phoneme mapping on the unannotated source language place name and address word segment to be translated to obtain an IPA phoneme sequence of the place name and address word segment to be translated; S32, constructing a multi-source place name historical transliteration knowledge base, and performing syllable segmentation and target language transliteration unit matching on the IPA phoneme sequence of the place name and address word segment to be translated based on the multi-source place name historical transliteration knowledge base to obtain multiple initial candidate units of target language transliteration results; S33, calculating the phoneme sequence edit distance between each initial candidate unit of the target language transliteration result in the multiple initial candidate units of the target language transliteration result and the IPA phoneme sequence of the place name and address word segment to be translated, and selecting the initial candidate units of the target language transliteration result with the phoneme sequence edit distance less than a preset threshold to form a candidate set of the target language transliteration result of the unannotated source language place name and address word segment.

[0037] Specifically, in step S31, IPA phoneme mapping is performed on the unannotated source language place name and address word segment to be translated to obtain an IPA phoneme sequence of the place name and address word segment to be translated. Specifically, since there are significant differences in phonemes between the writing systems of different languages (such as Chinese characters, Japanese kana, Cyrillic letters, etc.), directly performing transliteration based on the orthography of the source language may lead to distorted phoneme representation. For example, the actual pronunciation of the Japanese Chinese character "京都" is , and the pronunciation of "J" in the Spanish place name "Jerez" is / x / (similar to "ch" in German). If directly mapped according to English spelling habits, it will be wrongly converted to . Therefore, in order to eliminate orthographic interference and achieve cross-language phoneme alignment, based on the phonetic neutrality characteristic of the International Phonetic Alphabet (IPA), this application performs phoneme-level standardization conversion by calling multi-language IPA mapping tables (such as Chinese pinyin to IPA, Japanese kana to IPA, Spanish orthography to IPA, etc.) and combining phonetic change rules (such as Chinese erhua sounds, French liaison), converting the source language place name and address word segment into a unified IPA phoneme sequence independent of the writing form. In this way, it is possible to avoid phoneme misjudgment caused by differences in writing systems (such as avoiding pronouncing "Х" in Russian as "X" according to the Latin letter), providing a high-fidelity phonetic input for subsequent syllable segmentation.

[0038] Specifically, the step S32 constructs a multi-source place name historical transliteration knowledge base, and based on the multi-source place name historical transliteration knowledge base, performs syllable segmentation on the IPA phoneme sequence of the place name and address word fragment to be translated and matches it with the target language transliteration unit to obtain multiple target language transliteration result initial candidate units. It should be understood that, taking into account the possible ambiguity and cultural sensitivity of the transliteration results, for example, when an English place name is transliterated into Chinese, a syllable can be transliterated into multiple similar Chinese characters, and different Chinese character combinations may contain different cultural meanings or produce ambiguity. Therefore, in order to generate a transliteration result that is both accurate and in line with the cultural background, the present application constructs a multi-source place name historical transliteration knowledge base to utilize historical transliteration experience to perform syllable segmentation on the IPA phoneme sequence of the place name and address word fragment to be translated, and performs query matching at the syllable level to find the target language transliteration unit from real historical data that is similar to the pronunciation of the source language. During the specific implementation process, multiple historical transliteration cases similar to the IPA phoneme sequence of the place name and address word fragment to be translated are first retrieved from the knowledge base. Then, according to the target language syllable template of each historical transliteration case, the IPA phoneme sequence of the place name and address word fragment to be translated is divided into corresponding syllable units, and the syllable units are matched with the target language transliteration units in the historical transliteration cases to generate multiple initial candidate units of the target language transliteration results.

[0039] Specifically, step S33 calculates the phoneme sequence edit distance between each of the multiple target language transliteration result initial candidate units and the IPA phoneme sequence of the place name and address word segment to be translated, and selects the target language transliteration result initial candidate units whose phoneme sequence edit distance is less than a preset threshold to form the target language transliteration result candidate set of the unlabeled source language place name and address word segment to be translated. Specifically, since the transliteration process is essentially a phoneme approximation mapping rather than a completely equivalent conversion, and the degree of phonological deviation of different candidate units directly affects the acceptability of the translated name, in order to quantify the similarity between each transliteration result initial candidate unit and the source language pronunciation and eliminate significantly distorted solutions to ensure the quality of the candidate set, this application uses a phoneme sequence edit distance algorithm to calculate the minimum editing operation cost, including insertion, deletion, replacement, etc., from the source language place name and address word segment IPA phoneme sequence to the target language transliteration result initial candidate unit phoneme sequence through dynamic programming. The smaller the edit distance, the higher the similarity between the target language transliteration result initial candidate unit and the source language pronunciation. By setting a reasonable edit distance threshold (e.g., 0.3 × sequence length), only initial transliteration candidate units with an edit distance below this threshold are retained to form the final candidate set of target language transliteration results. This approach effectively suppresses excessive phonological distortion while retaining reasonable transliteration variants, ensuring that the candidate set is both phonologically reasonable and historically correct.

[0040] In the place name translation method combining multilingual syllable segmentation with adaptive learning, the step S4 assembles the target language transliteration result candidate set of each unlabeled source language place name and address word segment to be translated with the sequence of the source language place name and address word segment containing the common name translation annotation to obtain a candidate set of place name and address target language translation results. It should be understood that in the sequence of the source language place name and address word segment to be translated, the common name part has been standardized translated by the standard translation name, while the proper name part has generated multiple transliteration result candidates. In order to further evaluate the impact of different transliteration combinations on the overall translation effect of the place name and address, this application assembles the standard translation name of the common name part and the transliteration result candidate set of the proper name part in the source language place name and address to be translated to generate all possible complete place name and address target language translation strings, and normalizes them according to the grammatical rules and expression habits of the target language to construct a candidate set of place name and address target language translation results containing multiple transliteration versions. In this way, a candidate space containing multimodal transliteration possibilities can be constructed, providing rich material for the final place name and address translation.

[0041] In the place name translation method combining multilingual syllable segmentation with adaptive learning, the step S5 performs adaptive learning aggregation optimization on the candidate set of the place name and address target language translation results to obtain the place name and address target language translation. It should be understood that different transliteration versions in the candidate set of the place name and address target language translation results may have problems such as expression differences and insufficient cultural adaptability, and the traditional scoring mechanism only relies on phoneme similarity and cannot capture the semantic rationality and contextual adaptability of the translation in the target language. Therefore, in order to further improve the accuracy and cultural adaptability of place name and address translation, this application introduces an adaptive learning aggregation optimization mechanism based on a deep learning algorithm, which performs contextual semantic learning and aggregation optimization on the candidate set of the place name and address target language translation results to comprehensively consider the semantic information and contextual coherence of multiple transliteration versions to generate the final place name and address target language translation. Among them, Figure 5 Flowchart of sub-step S5 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application. Figure 5 As shown, the step S5 includes the steps of: S51, extracting the contextual semantic features of each place name and address target language translation result in the candidate set of the place name and address target language translation results to obtain a candidate set of place name and address target language translation result semantic coding vectors; S52, performing aggregation analysis on the candidate set of place name and address target language translation result semantic aggregation analysis coding vectors to obtain place name and address target language translation result semantic aggregation analysis coding vectors; S53, performing feature decoding on the place name and address target language translation result semantic aggregation analysis coding vectors to obtain the place name and address target language translation.

[0042] Specifically, in a specific example of the present application, the step S51 includes: performing contextual semantic encoding based on the Transformer architecture on each place name and address target language translation result in the candidate set of the place name and address target language translation results to obtain a candidate set of semantic encoding vectors of the place name and address target language translation results. It should be understood that in order to evaluate the naturalness and standardization of each place name and address target language translation result in the target language semantic space, the present application is based on the deep context perception capability of the Transformer architecture, and performs contextual semantic encoding on each place name and address target language translation result to utilize the self-attention mechanism of the Transformer architecture to model the global semantic dependency in the place name and address target language translation results, capture the place name structure characteristics and cultural implicit meanings of the place name and address target language translation results, and convert them into a quantifiable semantic space distribution to obtain a candidate set of semantic encoding vectors of the place name and address target language translation results, thereby providing basic feature support for the semantic aggregation optimization of subsequent multi-transliteration versions.

[0043] Specifically, the step S52 performs aggregation analysis on the candidate set of semantic coding vectors of the place name and address target language translation results to obtain semantic aggregation analysis coding vectors of the place name and address target language translation results. Specifically, the present application takes into account that simply performing mean aggregation on the candidate set of semantic coding vectors of the place name and address target language translation results or selecting a certain best translation result semantic coding feature as the final representation may not be sufficient to capture the complementary information or potential better combinations contained in multiple candidate items. Therefore, in order to further capture the semantic complementarity in multiple transliteration result candidates and dig out potential better combinations, the present application proposes a semantic aggregation optimization method, which first performs deep aggregation analysis on the candidate set of semantic coding vectors of the place name and address target language translation results to extract the main semantic direction of the set, and further embeds the individual semantic offsets of the semantic coding vectors of each place name and address target language translation result relative to the main semantic direction in the main semantic direction of the set to simulate the interaction and potential information fusion between multiple candidate translation results, and generate a semantic aggregation analysis coding vector of the place name and address target language translation results. In this way, we can achieve deep integration and optimization of multiple candidate translation results at the semantic level, avoid the information loss of the direct voting mechanism, and help improve the accuracy and robustness of the target language translation of place names and addresses. Figure 6 Flowchart of sub-step S52 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application. Figure 6As shown, the step S52 includes the steps of: S521, performing base state feature aggregation coding on the candidate set of semantic coding vectors of the place name and address target language translation result to obtain the base state feature aggregation coding vector of the place name and address target language translation result; S522, extracting the semantic compensation excitation state features of each place name and address target language translation result semantic coding vector in the candidate set of semantic coding vectors of the place name and address target language translation result relative to the base state feature aggregation coding vector of the place name and address target language translation result to obtain a set of semantic compensation excitation state feature coding vectors of the place name and address target language translation result; S523, based on the set of semantic compensation excitation state feature coding vectors of the place name and address target language translation result, performing gain modulation on the base state feature aggregation coding vector of the place name and address target language translation result to obtain the semantic aggregation analysis coding vector of the place name and address target language translation result.

[0044] More specifically, step S521 includes: first, inputting the candidate set of semantic encoding vectors of the place name and address target language translation results into a K-Means clustering network to obtain K semantic cluster center encoding vectors of the place name and address target language translation results, which can be expressed as follows:

[0045] ;

[0046] ;

[0047] in, The candidate set of semantic encoding vectors representing the target language translation results of place names and addresses, 、 、 and Respectively The first, second, and and The semantic encoding vector of the target language translation result of the place name and address, represents the set of real numbers, The dimension of the semantic encoding vector representing the target language translation result of the place name and address, Indicates the number of semantic encoding vectors of the target language translation results of place names and addresses, Represents the K-Means clustering network, Represents the set of semantic cluster center encoding vectors of the target language translation results of place names and addresses, 、 ,and Respectively The first, second and The semantic cluster center encoding vector of the target language translation result of place names and addresses, Indicates the number of semantic cluster center encoding vectors of the target language translation results of place names and addresses.

[0048] That is, the K-Means clustering network is used to perform unsupervised learning processing on the candidate set of semantic encoding vectors of the target language translation results of place names and addresses, to achieve data concentration and structured detection, to identify the centroid of the area with high data density from the high-dimensional semantic feature space, and to obtain K semantic clustering center encoding vectors of the target language translation results of place names and addresses that can represent the main semantic directions, so as to provide structured pattern primitives for the subsequent ground-state feature aggregation based on the self-attention mechanism, reduce the complexity of the feature space, and lay the foundation for integrating the semantic information of multiple transliteration versions, conducting adaptive feature learning and aggregation optimization.

[0049] Then, the semantic cluster center encoding vectors of the K place name and address target language translation results are aggregated based on the self-attention mechanism to obtain the aggregated encoding vector of the place name and address target language translation results, which can be expressed as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] ;

[0055] in, 、 and Represent the query weight matrix, key weight matrix, and value weight matrix respectively. 、 and represent the query matrix, key matrix and value matrix respectively, represents the self-attention layer, represents the normalized exponential function, represents the transpose of a vector, Representation layer normalization function, Represents the ground state feature aggregation encoding vector of the target language translation result of the place name and address.

[0056] That is, the self-attention mechanism is used to explore the intrinsic semantic associations and global structural information between the semantic cluster center encoding vectors of K place name and address target language translation results, breaking through the limitation of K-Means clustering that only provides local center representation, so that the semantic cluster center encoding vector of each place name and address target language translation result can fuse the contextual semantic information of other centers, and generate the ground state feature aggregation encoding vector of the place name and address target language translation result, thereby effectively capturing the deep connection between different transliteration candidate versions at the semantic level, and providing a core feature representation that is both locally representative and globally coherent for subsequent adaptive learning aggregation optimization.

[0057] More specifically, step S522 is expressed as follows:

[0058] ;

[0059] ;

[0060] in, represents feature cascade, represents the sigmoid activation function, represents the compensation weight matrix, represents the compensation bias term, represents dot product, The set of semantically compensated excited state feature encoding vectors representing the target language translation results of place names and addresses, 、 、 and Respectively The first, second, and and The semantically compensated excited state feature encoding vector of the target language translation result of each place name and address.

[0061] That is, by calculating the deviation component of the semantic encoding vector of each place name and address target language translation result relative to the base state, the individual-specific semantic information in each candidate translation result that is not explained by the base state features is separated, and it is made explicit as the semantic compensation excitation state feature encoding vector of the place name and address target language translation result, so that the network can quantify and process the unique semantic differences of different transliteration versions, which not only retains the global semantic consensus but also highlights the semantic characteristics of individual candidates, providing richer feature representation for subsequent adaptive learning.

[0062] Figure 7 Flowchart of sub-step S523 of the place name translation method combining multilingual syllable segmentation and adaptive learning according to an embodiment of the present application. Figure 7As shown, the step S523 includes the steps of: S5231, calculating the excited-state significance modulation weight factor of each semantic compensation excited-state feature coding vector of the place-name and address target language translation result in the set of semantic compensation excited-state feature coding vectors to obtain a set of semantic compensation significant modulation weight factors of the place-name and address target language translation result; S5232, based on the set of semantic compensation significant modulation weight factors of the place-name and address target language translation result, weighted modulating the set of semantic compensation excited-state feature coding vectors of the place-name and address target language translation result to obtain a set of semantic compensation excited-state feature significant coding vectors of the place-name and address target language translation result; S5233, inputting the set of semantic compensation excited-state feature significant coding vectors of the place-name and address target language translation result and the ground-state feature aggregation coding vector of the place-name and address target language translation result into a feature gain superposition network to obtain the semantic aggregation analysis coding vector of the place-name and address target language translation result.

[0063] In particular, considering that when calculating the semantic compensation excited state representation of the semantic encoding vector of each place name and address target language translation result relative to the ground state feature aggregation encoding vector of the place name and address target language translation result, Essentially, it defines the derived metric of the metric space of each semantic encoding vector of the target language translation result of each place name and address relative to the ground state space of the ground state feature aggregation encoding vector of the target language translation result of the place name and address, that is, the semantic encoding vector of each target language translation result of the place name and address is relatively high-dimensional. ’s feature space as a metric to construct a low-dimensional quantity in the ground state space of the ground state feature aggregation coding vector of the place name and address target language translation result through the derived mapping. In this way, it is necessary to consider the interface measurement constraints under the derived metric, that is, the metric of the excitation state is limited to the geometric boundary of the space transition under the induced mapping, rather than a simple high-dimensional-low-dimensional space transformation. Based on this, in a preferred example of the present application, the step S5231 includes: first, based on the key difference boundary definition between each place name and address target language translation result semantic coding vector in the candidate set of the place name and address target language translation result semantic coding vector and the ground state feature aggregation coding vector of the place name and address target language translation result, the semantic compensation excitation state feature coding vector of each place name and address target language translation result is feature corrected to obtain a set of optimized place name and address target language translation result semantic compensation excitation state feature coding vectors.

[0064] Specifically, first calculate Relative to The reduced-order incentive factor:

[0065] ;

[0066] ;

[0067] in, represents the L1 norm, represents the L2 norm, and They represent different reduced-order excitation factors.

[0068] Then, the dimensionality reduction tangent space structure is introduced to decompose the induced mapping in the form of tangent space so that it can be projected onto the orthogonal basis in the basis space:

[0069] ;

[0070] ;

[0071] in, and Represent different dimensionality reduction tangent space vectors respectively.

[0072] Next, the reduced-dimensional tangent space structure is used as a metric representation under the local submanifold to perform boundary-space geometry reconstruction, so as to transform the metric complexity of the excited state from the high-dimensional set representation of the semantic encoding vector of the target language translation result of the place name address to the interface geometry constraint measured by the tangent space:

[0073] ;

[0074] in, It is an adjustment factor used to compensate for excessive spatial curvature parameters The impact brought represents an exponential function with a natural constant as the base, express The corresponding semantic encoding vector of the target language translation result of the corrected place name and address.

[0075] Then Correction for ,in, The first in the set of semantic compensation excited state feature encoding vectors representing the optimized translation result of place name and address into target language The semantically compensated excited-state feature encoding vectors of the target language translation results of place names and addresses are optimized. In this way, when performing the derived metric from the submanifold metric to the basis space, the fixed points in the tangential metric flow can be renormalized by making the excited-state metric depend on the boundary rather than the spatial transformation to achieve effective measurement of the boundary and avoid excessive expansion of the representation of the semantically compensated excited-state feature encoding vectors of the target language translation results of each place name and address.

[0076] Then, each optimized place name and address target language translation result semantic compensation excited state feature coding vector in the set of the semantic compensation excited state feature coding vectors is input into a feature transformation compression module based on the ReLU function and then globally normalized to obtain a set of semantic compensation significance modulation weight factors of the place name and address target language translation result, which can be expressed as follows:

[0077] ;

[0078] ;

[0079] in, Indicates The exponential function operation with base , represents the ReLU activation function, represents the attention score transformation vector, represents the transpose of the feature, represents the learnable weight parameter matrix, The set of semantic compensation saliency modulation weight factors representing the target language translation results of place names and addresses, 、 、 and Respectively The first, second, and and The semantic compensation saliency modulation weight factor of the target language translation results of place names and addresses is used.

[0080] That is, through the feature transformation compression module based on the ReLU function, the semantic compensation excitation state feature coding vector of the optimized place name and address target language translation result is nonlinearly mapped and dimensionally screened to highlight the effective semantic difference features and suppress redundant information. After global normalization processing, the weights of each feature are made comparable across dimensions, thereby giving a reasonable importance measurement to the individual semantic deviations of different transliteration candidate versions, and constructing a set of semantic compensation significance modulation weight factors of the place name and address target language translation results that can be used for subsequent weighted modulation, providing a scientific importance assessment basis for the subsequent weighted modulation process, helping to focus on key semantic differences and weaken noise interference in aggregation optimization, and improving the semantic accuracy and context adaptability of the place name and address target language translation.

[0081] In a specific example of the present application, step S5232 is expressed as follows:

[0082] ;

[0083] ;

[0084] in, The set of semantically compensated excited state feature saliency encoding vectors representing the target language translation results of place names and addresses, 、 、 and Represents the first, second, and and The semantically compensated excited state feature saliency encoding vector of the target language translation results of place names and addresses.

[0085] That is, through weighted modulation, the semantically compensated excited-state feature encoding vectors of each place-name and address target language translation result can be adjusted according to its corresponding semantically compensated salience modulation weight factor, thereby enhancing important semantic features and weakening relatively unimportant or potentially noisy features. In this way, the generated set of semantically compensated excited-state feature salience encoding vectors of place-name and address target language translation results can more accurately reflect the key semantic differences between different translation candidate results, thereby improving the quality and accuracy of place-name and address translation and making the generated translation more capable of meeting the needs of practical applications.

[0086] In a specific example of the present application, step S5233 is expressed as follows:

[0087] ;

[0088] in, represents a multilayer perceptron, Represents the semantic aggregation analysis encoding vector of the target language translation result of the place name and address.

[0089] That is, through the processing of the feature gain superposition network, the excited state features reflecting individual differences and the ground state features reflecting global commonalities are effectively combined. This fusion not only enriches the dimension of semantic representation, but also enhances the ability to capture complex semantic relationships, which helps to better weigh various semantic factors in the subsequent translation decision-making process, thereby generating more accurate, more contextual, and more practical place name and address target language translation results semantic aggregation analysis encoding vectors, thereby improving the performance and practicality of the entire place name translation method.

[0090] Specifically, step S53 performs feature decoding on the semantic aggregation analysis encoding vector of the target language translation result of the place name and address to obtain the target language translation of the place name and address. Specifically, to reconstruct the abstract semantic representation into a natural language symbol sequence that conforms to the morphological rules of the target language, this application, based on an adversarial training-enhanced Transformer decoder, guides the generation process by introducing linguistic constraints (such as the target language's orthographic rules and place name capitalization conventions). In the specific implementation process, a two-stage decoding strategy is adopted: first, a standard Transformer decoder is used to decode the semantic aggregation analysis encoding vector of the target language translation result of the place name and address into a coarse-grained word fragment sequence in the target language. Subsequently, a conditional random field (CRF) layer is used to impose structural constraints and correct formatting errors (such as correcting it to "Lhasa City"). Simultaneously, an adversarial training mechanism is introduced, and the generated result is input into a discriminator (such as a pre-trained geographic entity recognition model) to evaluate its distribution consistency with real place names. The decoder parameters are optimized through gradient backpropagation, prompting the decoder to generate more natural and accurate place name and address translations. For example, when translating the Spanish place name "Ciudad de México" into English, the discriminator, by learning the distribution of English place names, can optimize the decoder to generate "Mexico City" rather than the literal translation "Ciudad de Mexico." This translation approach, combining adversarial training with linguistic constraints, effectively improves the quality of place name and address translation, making it more consistent with the target language's expression habits and cultural context.

[0091] In summary, a place name translation method combining multilingual syllable segmentation and adaptive learning based on an embodiment of the present application is illustrated, which performs word segmentation and common name translation annotation on the source language place name address based on a general dictionary, and converts the unannotated proper noun part into an IPA phoneme sequence. Then, based on a multi-source place name historical transliteration knowledge base, syllable division and translation matching are performed on the IPA phoneme sequence of each word segment, and multiple possible candidate transliteration results are matched for each word segment. Then, each candidate transliteration result of the proper noun segment is combined with the common name translation result to construct a candidate set of target language translation results of multiple transliterated versions of the place name address, and a deep learning algorithm is used to perform adaptive feature learning and aggregation optimization on the candidate set to comprehensively consider the semantic information and contextual coherence of the multiple transliterated versions to generate the final target language translation of the place name address. This method can solve the problems of inaccurate, stiff, and ambiguous transliterations in traditional place name and address translation methods, and improve translation efficiency and quality.

[0092] The basic principles of the present invention have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, and effects mentioned in the present invention are merely illustrative and non-limiting, and should not be construed as necessarily possessed by each embodiment of the present invention. Furthermore, the specific details of the above embodiments are provided for illustrative purposes and to facilitate understanding, and are not intended to be limiting. These details do not necessarily limit the present invention to being implemented using these specific details.

[0093] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiment described above is only schematic. For example, the unit division is only a logical function division, and there may be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0094] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be encompassed therein. Any reference to a figure in a claim should not be construed as limiting the claim to which it relates.

[0095] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units stated in the system claims can also be implemented by one unit through software or hardware.

[0096] Finally, it should be noted that the above description has been provided for purposes of illustration and description. Furthermore, the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to be limiting. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art will appreciate that the technical solutions of the present invention may be modified or replaced with equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multilingual place name translation method combining syllable segmentation and adaptive learning, characterized in that: include: Obtain the source language address and target translation language input by the user; Based on a universal dictionary and the target translation language, splitting, matching and translation annotation are performed on the source language place name and address to be translated to obtain a sequence of source language place name and address word segments containing common name translation annotations; Performing transliteration matching based on syllable segmentation on each unlabeled source language place name and address word segment to be translated in the sequence of source language place name and address word segments containing common name translation labels to obtain a target language transliteration result candidate set of each unlabeled source language place name and address word segment to be translated, including: performing IPA phoneme mapping on the unlabeled source language place name and address word segment to be translated to obtain an IPA phoneme sequence of the place name and address word segment to be translated; Constructing a multi-source place name historical transliteration knowledge base, and performing syllable segmentation on the IPA phoneme sequence of the place name address fragment to be translated based on the multi-source place name historical transliteration knowledge base and matching it with the target language transliteration unit to obtain a plurality of target language transliteration result initial candidate units; calculating a phoneme sequence edit distance between each of the plurality of target language transliteration result initial candidate units and the IPA phoneme sequence of the place name and address word segment to be translated, and selecting the target language transliteration result initial candidate units whose phoneme sequence edit distance is less than a preset threshold to form a target language transliteration result candidate set for the unlabeled source language place name and address word segment to be translated; Assembling the candidate sets of target language transliteration results of the unlabeled source language place name and address word segments to be translated with the sequence of source language place name and address word segments containing common name translation labels to obtain a candidate set of place name and address target language translation results; Performing adaptive learning aggregation optimization on the candidate set of place name and address target language translation results to obtain place name and address target language translations, including: extracting contextual semantic features of each place name and address target language translation result in the candidate set of place name and address target language translation results to obtain a candidate set of semantic encoding vectors of the place name and address target language translation results; Performing aggregation analysis on the candidate set of semantic coding vectors of the place name and address target language translation results to obtain a semantic aggregation analysis coding vector of the place name and address target language translation results; The semantic aggregation analysis encoding vector of the place name and address target language translation result is feature decoded to obtain the place name and address target language translation.

2. The multilingual place name translation method combining syllable segmentation and adaptive learning according to claim 1 is characterized in that: Based on a universal dictionary and the target translation language, the source language place name and address to be translated are split, matched, and translated and annotated to obtain a sequence of source language place name and address word segments containing common name translation annotations, including: Performing word segmentation processing on the source language place name and address to be translated based on universal dictionary matching to obtain a sequence of word segments of the source language place name and address to be translated; Based on the universal dictionary and the target translation language, translation annotation is performed on each source language place name and address word segment to be translated in the sequence of source language place name and address word segments to be translated to obtain the sequence of source language place name and address word segments containing common name translation annotations.

3. The method for translating place names by combining multilingual syllable segmentation with adaptive learning according to claim 2, characterized in that: Extracting contextual semantic features of each place name and address target language translation result in the candidate set of place name and address target language translation results to obtain a candidate set of semantic encoding vectors of the place name and address target language translation results, including: Contextual semantic encoding based on the Transformer architecture is performed on each place name and address target language translation result in the candidate set of place name and address target language translation results to obtain a candidate set of semantic encoding vectors of the place name and address target language translation results.

4. The method for translating place names by combining multilingual syllable segmentation with adaptive learning according to claim 3, characterized in that: Performing aggregation analysis on the candidate set of semantic coding vectors of the place name and address target language translation results to obtain a semantic aggregation analysis coding vector of the place name and address target language translation results, including: Performing ground state feature aggregation coding on the candidate set of semantic coding vectors of the place name and address target language translation results to obtain a ground state feature aggregation coding vector of the place name and address target language translation results; Extracting the semantic compensation excited state features of each semantic coding vector of the place name and address target language translation result relative to the place name and address target language translation result base state feature aggregation coding vector in the candidate set of semantic coding vectors of the place name and address target language translation result to obtain a set of semantic compensation excited state feature coding vectors of the place name and address target language translation result, which is specifically expressed by the formula: ； ； in, Indicates the The semantic encoding vector of the target language translation result of the place name and address, Represents the ground state feature aggregation encoding vector of the target language translation result of the place name address, represents feature cascade, represents the sigmoid activation function, represents the compensation weight matrix, represents the compensation bias term, represents dot product, The set of semantically compensated excited state feature encoding vectors representing the target language translation results of place names and addresses, 、 、 and Respectively The first, second, and and The semantic compensation excited state feature encoding vector of the target language translation result of the place name address; Based on the set of semantically compensated excited-state feature coding vectors of the place-name and address target language translation results, gain modulation is performed on the ground-state feature aggregation coding vector of the place-name and address target language translation results to obtain the semantic aggregation analysis coding vector of the place-name and address target language translation results.

5. The method for translating place names by combining multilingual syllable segmentation with adaptive learning according to claim 4, characterized in that: Performing base state feature aggregation coding on the candidate set of semantic coding vectors of the place name and address target language translation results to obtain a base state feature aggregation coding vector of the place name and address target language translation results, including: Inputting the candidate set of semantic encoding vectors of the place name and address target language translation results into a K-Means clustering network to obtain K semantic clustering center encoding vectors of the place name and address target language translation results; The semantic cluster center encoding vectors of the K place name and address target language translation results are aggregated based on the self-attention mechanism to obtain the ground state feature aggregation encoding vector of the place name and address target language translation result.

6. The method for translating place names by combining multilingual syllable segmentation with adaptive learning according to claim 5, characterized in that: Based on the set of semantically compensated excited-state feature coding vectors of the place-name and address target language translation result, gain modulation is performed on the ground-state feature aggregation coding vector of the place-name and address target language translation result to obtain the semantic aggregation analysis coding vector of the place-name and address target language translation result, including: Calculating the excited-state significance modulation weight factor of each place-name-address target language translation result semantically compensated excited-state feature coding vector in the set of place-name-address target language translation result semantically compensated excited-state feature coding vectors to obtain a set of place-name-address target language translation result semantically compensated significance modulation weight factors; Based on the set of semantic compensation saliency modulation weight factors of the place name and address target language translation results, weighted modulation is performed on the set of semantic compensation excited state feature coding vectors of the place name and address target language translation results to obtain a set of semantic compensation excited state feature saliency coding vectors of the place name and address target language translation results; The set of semantically compensated excited-state feature saliency coding vectors of the place name and address target language translation result and the ground-state feature aggregation coding vector of the place name and address target language translation result are input into a feature gain superposition network to obtain the semantic aggregation analysis coding vector of the place name and address target language translation result, wherein the feature gain superposition network is a multi-layer perceptron.

7. The method for translating place names by combining multilingual syllable segmentation with adaptive learning according to claim 6, characterized in that: Calculating the excited-state saliency modulation weight factor of each place-name-address target language translation result semantically compensated excited-state feature coding vector in the set of the place-name-address target language translation result semantically compensated excited-state feature coding vectors to obtain a set of place-name-address target language translation result semantically compensated saliency modulation weight factors, including: Based on the key difference boundary definition between each place name and address target language translation result semantic coding vector in the candidate set of the place name and address target language translation result semantic coding vector and the place name and address target language translation result ground state feature aggregation coding vector, feature correction is performed on each place name and address target language translation result semantic compensation excitation state feature coding vector to obtain a set of optimized place name and address target language translation result semantic compensation excitation state feature coding vectors, specifically including: calculating Relative to The reduced-order incentive factor: ； ； in, represents the L1 norm, represents the L2 norm, and Represent different reduced-order excitation factors respectively; Then, the dimensionality reduction tangent space structure is introduced to decompose the induced mapping in the form of tangent space so that it can be projected onto the orthogonal basis in the basis space: ； ； in, and Represent different dimensionality reduction tangent space vectors respectively; Next, the reduced-dimensional tangent space structure is used as a metric representation under the local submanifold to perform boundary-space geometry reconstruction, so as to transform the metric complexity of the excited state from the high-dimensional set representation of the semantic encoding vector of the target language translation result of the place name address to the interface geometry constraint measured by the tangent space: ； in, It is an adjustment factor used to compensate for excessive spatial curvature parameters The impact brought represents an exponential function with a natural constant as the base, express The corresponding semantic encoding vector of the target language translation result of the corrected place name and address; Then Correction , specifically ,in, The first in the set of semantic compensation excited state feature encoding vectors representing the optimized translation result of place name and address into target language An optimized semantically compensated excited state feature encoding vector for the target language translation result of the place name and address; Each optimized place name and address target language translation result semantic compensation excited state feature coding vector in the set of the optimized place name and address target language translation result semantic compensation excited state feature coding vector is input into a feature transformation compression module based on the ReLU function and then globally normalized to obtain a set of semantic compensation significance modulation weight factors of the place name and address target language translation result.

Images