Voice real-time translation and proofreading system for cross-border e-commerce

Abstract

The invention relates to the technical field of voice real-time translation, in particular to a voice real-time translation and proofreading system oriented to cross-border e-commerce, which comprises the following modules: a voice semantic processing module used for receiving voice input, converting the voice input into text data, performing semantic analysis on the text data, extracting commodity names, attribute information and logic relationships, and sending the extracted commodity names, attribute information and logic relationships to a server; semantic structure data is generated and stored in a context cache; and the anaphora analysis module is used for reading the semantic structure data from the context cache, detecting anaphora words in the semantic structure data, determining anaphora objects and updating the semantic data subjected to anaphora resolution to the context cache. According to the method, word segmentation, part-of-speech tagging, syntactic analysis and dependency analysis are carried out on a text, commodity names, the number, colors, sizes and attribute relations are extracted to generate structured semantic data, the structured semantic data are stored in a context cache, and when anaphora words are detected, anaphora objects are determined through semantic similarity and context logic analysis, and the semantic data are updated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of real-time speech translation, and particularly relates to a real-time speech translation and proofreading system for cross-border e-commerce. BACKGROUND

[0002] The real-time speech translation system for cross-border e-commerce is mainly used for converting the information of goods, attribute description and transaction-related content input by the seller through voice into target language text in real time, so as to realize the communication and information transmission between users of different languages. The system usually includes a voice recognition module, a text generation module and a basic semantic processing module, which are used to extract the name of goods, quantity, color, size and other attribute information in the voice and convert them into understandable written expression, so as to assist the buyer to quickly obtain the information of goods and complete cross-border transaction.

[0003] The traditional sentence-by-sentence translation system usually only processes the semantics of a single sentence, and lacks effective use of the context and contextual information of the previous text. When the seller uses a pronoun to point to the goods or attributes mentioned in the previous text in continuous voice, the system cannot correctly identify the referent, resulting in confusion of the referent relationship, loss of information or logical errors in the translation, and the formation of coherent and accurate product descriptions. SUMMARY

[0004] In order to make up for the above shortcomings, the present application provides a real-time speech translation and proofreading system for cross-border e-commerce, aiming to improve the problem of inaccurate pronoun recognition in the traditional sentence-by-sentence translation system, which leads to logical confusion in the translation.

[0005] In the first aspect, the present application provides the following technical scheme, a real-time speech translation and proofreading system for cross-border e-commerce, comprising the following modules:

[0006] A voice semantic processing module is used to receive voice input and convert it into text data, perform semantic analysis on the text data, extract the name of goods, attribute information and logical relationship, generate semantic structure data and store it in the context cache;

[0007] A reference resolution module is used to read the semantic structure data from the context cache, detect the pronouns therein and determine the referents, and update the semantic data after reference resolution to the context cache;

[0008] A size correction module is used to receive the semantic data after reference resolution, identify the size and size information therein, perform standardized conversion according to the source market size database and target market mapping rules, and output the corrected semantic data;

[0009] A preliminary translation generation module is used to generate a preliminary translation based on the corrected semantic data, and perform logical verification, attribute combination checking and conditional mapping processing on the preliminary translation, and output the corrected translation;

[0010] an error evaluation and backtracking module configured to evaluate the credibility of the corrected translation, identify potential errors according to attribute consistency, logical relationships and historical transaction data, and trigger a local backtracking mechanism to correct the previous translation when a high-risk error is identified;

[0011] a final translation output module configured to output the corrected translation as a final translation and update the context cache for reference by subsequent voice input.

[0012] By adopting the above technical solutions, the referential words and corresponding commodities or attributes in continuous voice are accurately identified, the logical coherence and information integrity of the translation are maintained, and thus the problems of missing referential information and logical confusion in the traditional sentence-by-sentence translation system are solved.

[0013] Preferably, the semantic analysis of the text data comprises:

[0014] performing word segmentation and part-of-speech tagging on the text data to identify keywords related to commodity names, quantities, colors, sizes and other attributes;

[0015] establishing logical associations between attributes and commodity names in the text through syntactic parsing and dependency analysis to generate a preliminary semantic structure;

[0016] mapping commodity names and their related attributes into structured data units and labeling their types, values and position indices using entity recognition and attribute extraction algorithms;

[0017] constructing a relationship network between commodity attributes to determine attribute combinations and conditional dependency relationships and form a complete semantic structure data.

[0018] Preferably, the referential resolution of the semantic structure data comprises:

[0019] identifying referential words in the semantic structure data, including personal pronouns, demonstrative pronouns and quantitative pronouns;

[0020] matching the previously stored semantic structure data in the context cache to determine the referential objects according to commodity names, attributes and logical relationships;

[0021] confirming the correspondence between the referential words and specific commodities or attributes through semantic similarity calculation and context logical analysis;

[0022] updating the referentially resolved semantic data to the context cache, including updating the mapping information of the referential words and the related attribute dependency relationships.

[0023] Preferably, the standardized conversion of the referentially resolved semantic data comprises:

[0024] Identify size and dimension information in semantic data, including specific numerical values, letter codes and descriptive quantifiers;

[0025] Match the size information against a source market size database to determine its corresponding standardized reference;

[0026] Invoke target market mapping rules to convert the source market size to a target market standard expression, calculating the corresponding numerical value or standard terminology;

[0027] Fine-tune the conversion result in combination with product category and historical transaction data to conform to target market conventions;

[0028] Output the corrected semantic data, including standardized size, dimension and related attribute information.

[0029] Preferably, the generating a preliminary translation includes:

[0030] Receiving the corrected semantic data, including product name, attributes, size and dimension information;

[0031] Mapping the data to a text expression in the target language according to semantic structure and contextual relationships;

[0032] Processing conditional dependencies and attribute combination relationships to ensure accurate matching of attributes in the translation;

[0033] Performing preliminary logical checks to verify the reasonableness of relationships between attributes;

[0034] Outputting a preliminary translation, forming translation data containing product description, attributes, size and condition information.

[0035] Preferably, the logical verification, attribute combination check and condition mapping processing includes:

[0036] Performing consistency checks on each attribute in the preliminary translation and product information to ensure that the attribute combination conforms to the actual product characteristics and semantic logic;

[0037] Resolving the interaction between attributes according to conditional dependency relationships, identifying condition constraints and combination rules, and mapping them to the translation;

[0038] Checking the logical correctness of quantity, size and unit attributes to prevent information conflicts or unreasonable expressions;

[0039] Outputting a corrected translation that has undergone logical verification and condition mapping processing.

[0040] Preferably, the credibility assessment includes:

[0041] Generate confidence indicators for each attribute information, quantity, unit and logical relationship in the corrected translation, and compare them with historical transaction data and semantic structure data in the context cache;

[0042] Analyze attribute consistency and logical relationship integrity to identify potential inconsistencies or abnormal expressions;

[0043] Integrate all confidence indicators to form a comprehensive translation credibility score, and mark low credibility content.

[0044] Preferably, the marking of potential errors includes:

[0045] According to the credibility score, identify parts of the corrected translation that may have numerical, unit, attribute or logical relationship abnormalities;

[0046] Generate potential error markers for the identified low credibility content to indicate the translation fragments that need to be corrected later;

[0047] Associate the potential error markers with the semantic structure data in the context cache to record the relevant attributes and logical dependencies.

[0048] Preferably, the local backtracking mechanism includes:

[0049] According to the credibility assessment and potential error marking, identify the previous translation fragments that need to be corrected;

[0050] Extract relevant semantic structure data and attribute logical dependencies from the context cache to determine the correction range and correction strategy;

[0051] Update the marked low credibility translation fragments, including adjusting numerical values, units, attribute combinations or logical expressions, and replacing the original translation with the corrected translation content;

[0052] Re-store the corrected translation in the context cache.

[0053] Preferably, updating the context cache includes:

[0054] Receive the corrected translation and its corresponding semantic structure data;

[0055] Identify and index the semantic structure data, compare the product name, attributes, quantities and logical relationships with previously cached data;

[0056] Write new or corrected information to the context cache, and delete or overwrite data that is no longer applicable;

[0057] After completing the cache update, the system reads the required semantic structure data from the context cache when processing subsequent voice input.

[0058] The present application has the following beneficial effects:

[0059] 1、In the present application, by performing word segmentation, part-of-speech tagging, syntax analysis and dependency analysis on the text, the product name, quantity, color, size and attribute relationship are extracted to generate structured semantic data and stored in the context cache. When the pronoun is detected, the semantic similarity and context logic analysis are used to determine the referent and update the semantic data, realizing the accurate correspondence between the previous product or attribute and the pronoun, and solving the problem of loss of reference information or logical confusion in traditional sentence-by-sentence translation.

[0060] 2、In the present application, by identifying the size, letter code and descriptive quantifier after receiving the corrected semantic data, matching the source market size reference and calling the target market mapping rule to calculate the corresponding numerical value or standard term, and combining the product category and historical transaction data for fine tuning, and performing unit and attribute logical verification, the size information in the translation is accurately expressed, solving the understanding deviation and potential dispute problem caused by size difference in cross-border transactions.

[0061] 3、In the present application, by generating a preliminary translation in real time, simultaneously processing context logic, attribute combination and condition mapping, marking potential errors through previous translation credibility evaluation and triggering a local backtracking mechanism to correct the translation, low-latency generation is realized while ensuring high accuracy and logical consistency, solving the problem of information conflict and attribute expression error caused by error accumulation in continuous speech translation. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 A method flowchart of a speech real-time translation and proofreading system for cross-border e-commerce is provided. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0064] In the first embodiment of the present application, the present application provides a speech real-time translation and proofreading system for cross-border e-commerce, as shown in Figure 1 The system includes the following modules:

[0065] A speech semantic processing module is used to receive speech input and convert it into text data, perform semantic analysis on the text data, extract product name, attribute information and logical relationship, generate semantic structure data and store it in the context cache;

[0066] Further, the semantic analysis of the text data includes:

[0067] The text data is segmented and tagged with parts of speech to identify keywords related to product name, quantity, color, size, and other attributes;

[0068] Through syntactic analysis and dependency relation analysis, the logical association between each attribute and the product name in the text is established, and a preliminary semantic structure is generated;

[0069] Using entity recognition and attribute extraction algorithms, the product name and its related attributes are mapped to structured data units, and their types, values, and position indexes are labeled;

[0070] A relationship network between product attributes is constructed to determine attribute combinations and conditional dependency relationships, forming a complete semantic structure data.

[0071] Specifically, the voice semantic processing module first receives the user's voice input signal, and converts the voice waveform into text data through an optional speech recognition algorithm; this conversion can use an end-to-end speech recognition model based on a deep neural network to map a sequence of acoustic features to a corresponding text sequence , where represents the acoustic feature vector at time , and represents the th word or character in the text; this model can be trained using a connectionist temporal classification (CTC) loss function to ensure the accuracy of the speech-to-text mapping; after the text is generated, the system inputs the text data into the text semantic analysis module for subsequent processing;

[0072] In the text semantic analysis process, first, the text data is segmented and tagged with parts of speech to identify keywords related to product name, quantity, color, size, and other attributes in the text; the segmentation process can use a combination of dictionary matching and statistical language models to split the text into several word units and tag their parts of speech , where represents the th word, and the part-of-speech information can be used to assist in subsequent attribute recognition and logical analysis; then, through syntactic analysis and dependency relation analysis, the logical association between each attribute and the product name in the text is established, and a preliminary semantic structure is generated; dependency relation analysis can be achieved by generating a dependency tree , where is a set of nodes representing words, is a set of edges representing semantic dependency relationships, and the modification and constraint relationships between attributes and product names are determined by analyzing the dependency tree structure;

[0073] After establishing the preliminary semantic structure, the system calls entity recognition and attribute extraction algorithms to map the product name and its related attributes into structured data units wherein each contains field type, value and position index, which are used to accurately represent the product and its attribute information; attribute types include but are not limited to quantity, color, size, material, etc.; position index is used to identify the start and end positions of the corresponding words in the original text, so as to backtrack or modify later; on this basis, a relationship network between product attributes is constructed wherein is a set of attribute nodes, is a set of conditional dependency relationships between nodes; by analyzing the network, the attribute combination and conditional constraint relationship are determined, for example, if the product is a piece of clothing, the color and size combination relationship needs to be logically consistent, so as to form a complete semantic structure data;

[0074] The generated semantic structure data is stored in the context cache as the basis for subsequent module processing; the context cache not only stores the product name and attributes, but also records the logical relationship and conditional dependency, which is used to support subsequent operations such as anaphora resolution, size correction and translation generation; during storage, a unique identifier can be generated for each structured unit for quick indexing and retrieval; this semantic data structure can be further cross-verified with historical transaction data and product standard library to improve the support capability of the subsequent processing module for attribute accuracy.

[0075] An anaphora resolution module is used to read the semantic structure data from the context cache, detect the anaphora words therein and determine the anaphora object, and update the anaphora resolved semantic data to the context cache;

[0076] Further, anaphora resolution of the semantic structure data includes:

[0077] Identify the anaphora words in the semantic structure data, including personal pronouns, demonstrative pronouns and quantity pronouns;

[0078] Match the previously stored semantic structure data in the context cache, determine the anaphora object according to the product name, attribute and logical relationship;

[0079] Confirm the corresponding relationship between the anaphora words and the specific product or attribute through semantic similarity calculation and context logical analysis;

[0080] Update the anaphora resolved semantic data to the context cache, including updating the mapping information of the anaphora words and the related attribute dependency relationship.

[0081] ​Specifically, the reference resolution module first reads the stored semantic structure data from the context cache, which includes the commodity name, attribute, quantity and logical relationship information; the module scans the semantic structure data through the pronoun recognition algorithm to identify the personal pronouns, demonstrative pronouns and quantity pronouns in it, forming a candidate pronoun set , wherein each represents a pronoun node in the text or semantic unit; during the identification process, the subject reference, object reference and quantity reference can be distinguished in combination with the part-of-speech tagging information and the dependency relationship, to ensure that the candidate set accurately covers the potential reference target;

[0082] The module matches the candidate pronoun with the previously stored context semantic structure data in the context cache; the context semantic structure data can be represented as a set , wherein each contains commodity name, attribute and logical relationship information; the reference matching can be achieved by constructing a semantic similarity matrix , wherein the matrix element represents the similarity score of the pronoun and the semantic unit , which can be calculated based on the cosine similarity of the word embedding vectors ,

[0083] ;

[0084] , wherein and represent the vector representation of the pronoun and the semantic unit respectively; during the calculation process, the context logical constraints such as the previous attribute matching, quantity consistency and conditional dependency relationship can be combined to improve the accuracy of the reference matching;

[0085] After determining the candidate reference object, the system combines the context logical analysis to perform the final resolution of the pronoun; the logical analysis includes verifying the rationality of the reference object and the pronoun in terms of attribute type, quantity and conditional relationship, and a confidence function can be used to represent the credibility of the reference relationship, and the formula is as follows:

[0086] ;

[0087] , wherein represents the context logical matching score, and are optional weight coefficients, satisfying ; through this confidence evaluation, the semantic unit with the highest score can be selected as the mapping object of the pronoun;

[0088] ​updating the semantic data after the anaphora resolution to the context cache, including updating the mapping information of the anaphoric words and the related attribute dependency relationship; in the updating process, the mapping object of each anaphoric word is recorded, and the anaphoric relationship is linked to the semantic structure data, so that the subsequent modules can directly refer to the parsed attribute and commodity relationship; the updated context cache ensures the completeness and logical consistency of the semantic structure data, provides an accurate reference basis for subsequent voice input, and supports potential backtracking correction and translation generation. The mapping object of each anaphoric word is recorded, and the anaphoric relationship is linked to the semantic structure data, so that the subsequent modules can directly refer to the parsed attribute and commodity relationship; the updated context cache ensures the completeness and logical consistency of the semantic structure data, provides an accurate reference basis for subsequent voice input, and supports potential backtracking correction and translation generation.

[0089] The size correction module is used for receiving the semantic data after the anaphora resolution, identifying the size and size information therein, performing standardized conversion according to the source market size database and the target market mapping rule, and outputting the corrected semantic data;

[0090] Further, the standardized conversion of the semantic data after the anaphora resolution includes:

[0091] Identifying the size and size information in the semantic data, including specific numerical values, letter codes and descriptive quantifiers;

[0092] Matching the size information according to the source market size database to determine the corresponding standardized reference;

[0093] Calling the target market mapping rule to convert the source market size into the target market standard expression, and calculating the corresponding numerical value or standard term of the target market;

[0094] Combining the commodity category and the historical transaction data to fine-tune the conversion result to meet the target market convention;

[0095] Outputting the corrected semantic data, including the standardized size, size and related attribute information.

[0096] Specifically, the size correction module first receives the semantic data after the anaphora resolution, which contains the commodity name, attribute information and size and size information after the anaphora resolution; the module scans the semantic data through the size identification algorithm, extracts specific numerical values, letter codes and descriptive quantifiers to form a size set , wherein each corresponds to the size representation of a certain attribute unit in the semantic structure; in the size identification process, the unit identification and attribute type matching algorithm can be combined to ensure that numerical size, letter size and fuzzy descriptive size are accurately identified;

[0097] The module matches the size information according to the source market size database to determine the standardized reference; the source market size database can be represented as a mapping set , wherein is the original market size, ​To correspond to the standardized benchmark value; the correction module matches each With Comparison, select the closest standardized benchmark , expressed as:

[0098] ;

[0099] Where Size difference function, can be calculated according to the numerical difference or category mapping distance; In the matching process, historical transaction data and commodity category information can be combined to assist in confirming the rationality of the benchmark matching;

[0100] Module calls target market mapping rules to convert source market size to target market standard expression; Target market mapping rules can be expressed as mapping function , Where Target market standard size or standard term; The conversion formula can be expressed as:

[0101] ;

[0102] Where Indicates the corrected target market size; Fine tuning coefficient Can be introduced in the mapping process, combined with commodity category and historical transaction preferences to make subtle adjustments to the conversion result, ensuring that the converted size meets the target market conventions and attribute logical relationships;

[0103] The module outputs the corrected semantic data, including the standardized size, size and related attribute information; The corrected semantic data is written into the context cache, updating the standardized size field of the corresponding commodity name and attribute; This output not only serves as input for subsequent translation generation and logical verification, but also provides accurate reference for potential backtracking correction and multi-market size consistency processing, ensuring the standardization and logical consistency of size and size information in the entire semantic processing chain.

[0104] The preliminary translation generation module is used to generate a preliminary translation based on the corrected semantic data, and to perform logical verification, attribute combination checking and condition mapping processing on the preliminary translation, and to output a corrected translation.

[0105] Further, generating a preliminary translation includes:

[0106] Receiving corrected semantic data, including commodity name, attribute, size and size information;

[0107] According to the semantic structure and context relationship, map the data to the text expression of the target language;

[0108] Process condition dependency and attribute combination relationship to ensure accurate matching of each attribute in the translation;

[0109] Perform preliminary logical checks to verify the reasonableness of the relationships between attributes;

[0110] Output the preliminary translation, forming translation data containing product descriptions, attributes, sizes, and condition information.

[0111] Further, the logical verification, attribute combination check, and condition mapping process includes:

[0112] Conduct consistency checks on each attribute in the preliminary translation and the product information to ensure that the attribute combination conforms to the actual product characteristics and semantic logic;

[0113] According to the condition dependency relationship, analyze the interaction between attributes, identify condition constraints and combination rules, and map them to the translation;

[0114] Check the logical correctness of quantity, size, and unit attributes to prevent information conflicts or unreasonable expressions;

[0115] Output the corrected translation after logical verification and condition mapping processing.

[0116] Specifically, the preliminary translation generation module first receives corrected semantic data, which contains product names, attributes, sizes, and size information; the module analyzes the logical dependencies between each attribute and the product name based on the semantic structure and contextual relationships, and generates a mapping table where is the attribute name, is the attribute value, is the target position index of the attribute in the translation; the mapping table is used for subsequent text generation and condition dependency processing;

[0117] The module converts the semantic data into target language text expressions according to the mapping table; the text generation process combines template-driven and dynamic splicing mechanisms to arrange product names, sizes, sizes, and related attributes according to the target language order, grammar rules; condition dependencies and attribute combination relationships are reflected in the mapping table, and by analyzing the dependency functions between attributes in the table determine the attribute combination method, where:

[0118]

[0119] According to the condition dependency function, the module ensures that the relevant attribute combination is correctly presented in the target language when generating the translation, while labeling possible condition restrictions and contextual constraint information;

[0120] The preliminary logical check is performed after the translation is generated, and the module checks the quantity, size, and unit attributes in the text to verify whether they are consistent with the semantic data in logic; the logical verification can use the attribute consistency function Indicates:

[0121]

[0122] For all attributes, calculate the overall logical consistency:

[0123]

[0124] If , mark the potentially conflicting attributes and trigger subsequent corrections;

[0125] The module parses the interactions between attributes according to the conditional mapping rules, maps the conditional constraints and combination rules to the translation, and ensures the reasonableness of the quantity, size, unit, and relationship between attributes; for example, if a clothing attribute depends on size information, the corresponding description automatically includes the size information when generating the translation, and matches the target market habit of speech;

[0126] The module outputs the corrected translation after logical verification and conditional mapping processing, forming a complete product description text containing product name, attributes, size, and condition information; the output translation updates the context cache, providing reliable basis for subsequent translation generation and potential backtracking correction, ensuring the consistency and integrity of semantics, logic, and conditional dependencies in the entire translation processing chain.

[0127] The error evaluation and backtracking module is used for credibility evaluation of the corrected translation, and potential errors are marked according to attribute consistency, logical relationship, and historical transaction data. When high-risk errors are identified, a local backtracking mechanism is triggered to correct the previous translation;

[0128] Further, the credibility evaluation includes:

[0129] Generate confidence indicators for each attribute information, quantity, unit, and logical relationship in the corrected translation, and compare them with historical transaction data and semantic structure data in the context cache;

[0130] Analyze attribute consistency and logical relationship integrity to identify parts with potential inconsistencies or abnormal expressions;

[0131] Combine all confidence indicators to form an overall translation credibility score to mark low-confidence content.

[0132] Further, marking potential errors includes:

[0133] According to the credibility score, identify parts of the corrected translation that may have numerical, unit, attribute, or logical relationship abnormalities;

[0134] Generate potential error markers for the identified low-confidence content to indicate translation segments that need subsequent backtracking correction; ​

[0135] The potential error identification is associated with the semantic structure data in the context cache, and the relevant attributes and logical dependencies are recorded.

[0136] Further, the local backtracking mechanism includes:

[0137] According to the credibility evaluation and the potential error annotation, a previous translation fragment that needs to be corrected is identified;

[0138] The relevant semantic structure data and attribute logical dependencies are extracted from the context cache to determine the correction range and correction strategy;

[0139] The annotated low-credibility translation fragment is updated, including adjusting numerical values, units, attribute combinations, or logical expressions, and replacing the original translation with the corrected translation content;

[0140] The corrected translation is stored back into the context cache.

[0141] Specifically, the error evaluation and backtracking module first receives the corrected translation and its corresponding semantic structure data, and performs preliminary analysis on each attribute information, quantity, unit, and logical relationship in the corrected translation. The module generates a confidence index for each attribute in combination with the semantic structure data in the context cache and the historical transaction data , where represents the i-th attribute, and the confidence index can be represented as:

[0142] ;

[0143] where is the distribution of the corresponding attribute in the historical transaction data, is the same type of semantic structure data in the context cache, is a logical consistency index for quantifying the rationality of the attribute value and the context logic matching degree;

[0144] The module then analyzes attribute consistency and logical relationship integrity to identify possible inconsistent or abnormal expressions in the corrected translation. The overall translation credibility score is calculated by synthesizing the confidence indexes:

[0145] ;

[0146] where is the total number of attributes in the corrected translation, and if is lower than the preset threshold, the module annotates the low-credibility content and generates a potential error identification , each identification contains the corresponding attribute, position, and related logical dependency information. These potential error identifications are also associated with the semantic structure data in the context cache to record the attributes and logical dependencies, facilitating subsequent correction.​

[0147] In the local backtracking mechanism, the module identifies the previous translation fragment that needs to be corrected according to the low confidence identifier, extracts the relevant semantic structure data and attribute logical dependency from the context cache, determines the correction range and correction strategy; the correction strategy can include adjusting numerical value, unit, attribute combination or logical expression, and the updated translation fragment replaces the original low confidence content; to ensure logical consistency, the module can calculate the confidence of the corrected attribute again during the backtracking correction process:

[0148] ;

[0149] Wherein is the logical consistency index of the corrected attribute; the corrected translation and its updated semantic structure data are stored back into the context cache to provide reference data for subsequent speech input processing and translation generation; the whole process ensures that the translation can be locally backtracked and corrected when potential errors are found, while maintaining the consistency and logical integrity of the translation and semantic structure data.

[0150] The final translation output module is configured to output the corrected translation as the final translation and update the context cache for subsequent speech input.

[0151] Further, updating the context cache comprises:

[0152] receiving the corrected translation and its corresponding semantic structure data;

[0153] identifying and indexing the semantic structure data, comparing the product name, attribute, quantity and logical relationship with the previously cached data;

[0154] writing the new or corrected information into the context cache, and deleting or covering the data that is no longer applicable;

[0155] After completing the cache update, the system reads the required semantic structure data from the context cache when processing subsequent speech input.

[0156] Specifically, the final translation output module first receives the translation and its corresponding semantic structure data corrected by the error evaluation and backtracking module; the module identifies and indexes the semantic structure data, compares the product name, attribute, quantity and logical relationship with the semantic structure data already stored in the context cache; during the comparison process, the semantic similarity identifies repeated or updated information, wherein is the newly received semantic unit, is the semantic unit in the cache, and the formula can be expressed as:

[0157]

[0158] wherein similarity of product name, similarity of attribute and value, consistency score of logical relationship, is an adjustable weight coefficient, satisfying ;

[0159] The module determines the cache update strategy according to the similarity comparison result, writes the newly added or corrected semantic units into the context cache, and deletes or covers the data that has become invalid or is no longer applicable, thereby ensuring the accuracy and integrity of the cache data. During the cache update process, the module can generate an index identifier for each semantic unit, including the product identifier, attribute type, value range, and logical dependency, so as to facilitate fast retrieval and subsequent speech input processing.

[0160] After the cache update is completed, when processing subsequent speech input, the system can read the semantic structure data in the context cache to obtain product information, attribute combinations, and logical dependencies, thereby supporting the translation generation module, the anaphora resolution module, and the error evaluation and backtracking module to accurately analyze and process the new input content. The entire process ensures the consistency and logical integrity of the final translation output and the cache data, thereby providing a repeatable reference data basis for continuous speech input and real-time translation.

[0161] Finally, it should be noted that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features, as long as they are within the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. made within the scope of the present application should be included in the protection scope of the present application.

Claims

1. A real-time voice translation and proofreading system for cross-border e-commerce, characterized in that, Includes the following modules: The speech semantic processing module is used to receive speech input and convert it into text data, perform semantic analysis on the text data, extract product names, attribute information and logical relationships, generate semantic structure data and store it in the context cache. The pronoun parsing module is used to read semantic structure data from the context cache, detect pronouns and determine the referents, and update the semantic data after pronoun resolution to the context cache. The size correction module is used to receive semantic data after referential resolution, identify the size and size information in it, perform standardization conversion according to the source market size database and target market mapping rules, and output corrected semantic data. The preliminary translation generation module is used to generate a preliminary translation based on the corrected semantic data, and to perform logical verification, attribute combination check and condition mapping processing on the preliminary translation, and output the corrected translation. The error assessment and backtracking module is used to assess the credibility of the corrected translation, mark potential errors based on attribute consistency, logical relationships and historical transaction data, and trigger a local backtracking mechanism to correct the previous translation when a high-risk error is identified. The final translation output module is used to output the corrected translation as the final translation and update the context cache to provide a reference for subsequent voice input.

2. The real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, The semantic analysis of the text data includes: Text data is segmented and part-of-speech tagged to identify keywords related to product name, quantity, color, size, and other attributes; By using syntactic parsing and dependency relation analysis, the logical relationship between each attribute in the text and the product name is established, and a preliminary semantic structure is generated. Using entity recognition and attribute extraction algorithms, product names and their related attributes are mapped into structured data units, and their type, value and location index are labeled. Construct a network of relationships between product attributes, determine attribute combinations and conditional dependencies, and form complete semantic structure data.

3. The real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, Dereference resolution of semantically structured data includes: Identify pronouns in semantic structure data, including personal pronouns, demonstrative pronouns, and quantifier pronouns; Match the semantic structure data of the preceding context cache, and determine the referent based on the product name, attributes and logical relationships; By calculating semantic similarity and analyzing contextual logic, the correspondence between pronouns and specific products or attributes is confirmed; The semantic data after referential resolution is updated to the context cache, including updating the mapping information of referents and related attribute dependencies.

4. A real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, Standardization transformation of semantic data after referential resolution includes: Identify size and sizing information in semantic data, including specific numerical values, letter codes, and descriptive quantifiers; The size information is matched with the size database from the source market to determine the corresponding standardized benchmark; Invoke the target market mapping rules to convert the source market size into the target market standard expression, and calculate the corresponding numerical value or standard terminology for the target market; By combining product categories and historical transaction data, the conversion results are fine-tuned to conform to the practices of the target market; Output corrected semantic data, including standardized dimensions, sizes, and related attribute information.

5. A real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, The generation of the preliminary translation includes: Receive corrected semantic data, including product name, attributes, size and sizing information; Based on semantic structure and contextual relationships, the data is mapped to textual expressions in the target language; Handle conditional dependencies and attribute combinations to ensure accurate matching of attributes in the translation; Perform a preliminary logic check to verify whether the relationships between attributes are reasonable; Output the initial translation, forming translated data that includes product description, attributes, size, and condition information.

6. A real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, The logical verification, attribute combination check, and condition mapping processing include: The consistency of each attribute in the preliminary translation with the product information is checked to ensure that the attribute combination conforms to the actual product characteristics and semantic logic. Based on the conditional dependencies, the interactions between the attributes are analyzed, the conditional constraints and combination rules are identified, and they are mapped to the translation. Verify the logical correctness of quantity, size, and unit attributes to prevent information conflicts or unreasonable expressions; Output the corrected translation after logical verification and conditional mapping.

7. A real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, The credibility assessment includes: Confidence indices are generated for each attribute, quantity, unit, and logical relationship in the corrected translation, and then compared with historical transaction data and semantic structure data in the context cache. Analyze attribute consistency and logical relationship integrity to identify potentially inconsistent or anomalous expressions. An overall credibility score for the translation is generated by combining various confidence indicators, and content with low credibility is marked.

8. A real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, The potential errors in the annotation include: Based on the credibility score, identify parts of the corrected translation that may contain abnormalities in numerical values, units, attributes, or logical relationships; Potential error markers are generated for identified low-confidence content to indicate translated segments that require subsequent backtracking and correction; Associate potential error identifiers with semantic structure data in the context cache, and record relevant attributes and logical dependencies.

9. A real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, The local backtracking mechanism includes: Based on credibility assessment and potential error annotation, identify the preceding translated text segments that need correction; Extract relevant semantic structure data and attribute logical dependencies from the context cache to determine the scope and strategy for correction. Update the low-confidence translation fragments marked with annotations, including adjusting the values, units, attribute combinations or logical expressions, and replace the original translation with the corrected translation content; The revised translation will be stored back in the context cache.

10. A real-time voice translation and proofreading system for cross-border e-commerce according to claim 1, characterized in that, The update context cache includes: Receive the revised translation and its corresponding semantic structure data; The semantic structure data is identified and indexed, and the product name, attributes, quantity, and logical relationships are compared with the previously cached data; Write new or corrected information to the context cache, and delete or overwrite data that is no longer applicable; After the cache update is completed, the system reads the required semantic structure data from the context cache when processing subsequent voice input.

Images