A cross-border e-commerce English scenario semantic accurate recognition and adaptation system
By constructing a semantic tree and a structured adaptation module for English scenarios in cross-border e-commerce, the problems of semantic recognition accuracy and data format matching of English text data in cross-border e-commerce were solved, achieving efficient semantic recognition and automated processing.
Patent Information
- Application Number
- CN202611360047.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-17
AI Technical Summary
Existing semantic recognition methods for English text data in cross-border e-commerce struggle to achieve high accuracy and consistency in cross-border e-commerce scenarios, and the recognition results do not match the data format of downstream business systems, resulting in low processing efficiency.
A semantic tree for English scenarios in cross-border e-commerce is constructed. Through semantic tree construction modules, scenario matching modules, conflict disambiguation modules, and structure generation modules, accurate identification and structured adaptation of English text data are achieved. Semantic matching and conflict disambiguation are performed using the scenario semantic tree, and the data is encapsulated according to the adaptation interface format of downstream business systems.
It significantly improves the accuracy and consistency of semantic recognition in cross-border e-commerce scenarios, shortens processing latency, improves the level of business automation, and eliminates the data format gap between recognition results and downstream systems.
Smart Images

Figure CN123117118A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semantic recognition technology, and in particular to a system for accurate semantic recognition and adaptation in English scenarios for cross-border e-commerce. Background Technology
[0002] Currently, with the rapid development of cross-border e-commerce, English, as the most widely used common language in cross-border transactions, generates massive amounts of English product text data and user query data daily on these platforms. How to accurately identify the core attribute information of products from this English text data and effectively adapt it to downstream business systems has become a key technical challenge restricting the data processing efficiency and user experience of cross-border e-commerce platforms. However, due to the significant structural diversity, domain specificity, and regional cultural differences in cross-border e-commerce English text data, existing general semantic recognition methods struggle to achieve high accuracy and consistency in semantic recognition when faced with this data, and also find it difficult to efficiently connect the recognition results to downstream business processes. This results in problems such as low product search accuracy and insufficient efficiency in product information entry into the database.
[0003] To address the aforementioned technical challenges, existing technologies have proposed several semantic recognition solutions. For example, some existing solutions directly employ general-domain natural language processing models to perform semantic encoding and classification of e-commerce English text. However, the training corpora for these general models are mostly derived from general-domain texts such as news and encyclopedias. Their semantic representation focuses on the overall sentiment or topic classification of sentences, lacking knowledge accumulation and specialized modeling of the three-level strong association structure of "product category - product attribute - attribute value" in the vertical field of cross-border e-commerce. As a result, when processing product titles and detailed descriptions, the models tend to treat core specifications such as "capacity," "size," and "color" as ordinary modifiers, making it difficult to accurately identify them as business entities with clear hierarchical relationships, thus resulting in a significantly low accuracy rate for attribute recognition. For example, some existing solutions use rule-based matching methods based on general thesaurus to extract product attribute words from English text. However, such solutions rely solely on the thesaurus mapping relationships in general dictionaries and fail to establish a dedicated semantic ambiguity discrimination and resolution mechanism for the phenomenon of homonyms caused by regional cultural differences in English in cross-border e-commerce scenarios. When different keywords in the same English text data point to leaf concepts of different product categories and do not have a common direct parent category, existing solutions cannot actively trace back the context information of the English text to determine the actual semantic reference of the above keywords in the current text, which can easily lead to incorrect semantic matching results and seriously affect the accuracy of search ranking and product recommendation. For example, after completing semantic recognition, the output of existing solutions is mostly plain text labels or unstructured vector representations. They do not establish a mapping relationship and adaptation and encapsulation mechanism from the recognition results to structured business fields according to the specific data format requirements of downstream business systems. This results in a significant data format gap between the recognition results and downstream business systems such as product information management and search ranking. Often, manual secondary editing and format conversion are required before it can be put into use. The overall processing efficiency is low and it is difficult to meet the needs of cross-border e-commerce businesses for data real-time performance and automation. Summary of the Invention
[0004] To achieve the above objectives, this invention provides a cross-border e-commerce English scene semantic accurate recognition and adaptation system, characterized in that the system includes a semantic tree construction module, a scene matching module, a conflict disambiguation module, a structure generation module, and an adaptation encapsulation module, wherein: The semantic tree construction module is used to set root nodes and intermediate nodes based on domain entity words, attribute words, and hierarchical relationships of entities in the original corpus of cross-border e-commerce, set leaf nodes based on attribute words and entity subordination relationships, and configure a set of synonyms and near-synonyms for each leaf node to construct a semantic tree for English cross-border e-commerce scenarios. The scene matching module is used to input cross-border e-commerce English text data as structured text data into the scene semantic tree, and match keywords with leaf nodes, synonyms, and near-synonyms according to the matching strategy corresponding to the text type label to obtain the initial semantic node matching results. The conflict disambiguation module is used to determine cross-node semantic conflict when the leaf nodes matched by two sets of keywords do not have a common direct parent node. It extracts the auxiliary qualifiers of the corresponding context, uses the auxiliary qualifiers to correct the semantic orientation of the leaf nodes, and obtains the disambiguated target semantic node matching result. The structured generation module is used to call the structured output template based on the text type label, and fill the template field with the product attribute keywords and attribute values corresponding to the leaf nodes in the target semantic node matching results to generate structured semantic recognition results. The adaptation and encapsulation module is used to encapsulate the structured semantic recognition results according to the adaptation interface format of downstream businesses and generate adaptation output data.
[0005] Compared with the prior art, the present invention has the following beneficial effects: 1. The cross-border e-commerce English scene semantic accurate recognition and adaptation method provided by this invention overcomes the technical defect of general semantic models that cannot accurately identify the core attributes of products due to domain bias in training corpora by constructing a scene semantic tree specifically for cross-border e-commerce English scenes and using it as the structured knowledge foundation for semantic recognition. This scene semantic tree uses a fixed hierarchical structure of "product category - product attribute - attribute value," explicitly encoding the hierarchical relationships between domain entity words and the subordinate relationships between attribute words and entities into a tree topology, ensuring that the attribute value represented by each leaf node has a clear category affiliation path. When keywords in structured text data are input into the scene semantic tree for matching, the matching result naturally carries the complete semantic hierarchy information of the keyword in the cross-border e-commerce domain, rather than just outputting isolated concept labels. Based on this hierarchical information, the subsequent cross-node semantic conflict determination can accurately identify whether there is a logical contradiction in the category affiliation between different keyword-hit nodes, and correct the actual semantic pointing of the conflicting nodes to the correct leaf node under the same category branch by extracting contextual auxiliary qualifiers. Thus, the category constraint information provided by the tree structure and the local contextual information provided by the context are fully utilized in the entire process of semantic disambiguation, which significantly improves the accuracy and consistency of semantic recognition results in cross-border e-commerce scenarios.
[0006] 2. This invention further utilizes a text-type tag-driven structured output template invocation mechanism and a downstream business adaptation interface encapsulation mechanism. During the semantic recognition result generation stage, the corresponding product attribute field combinations and data organization structure are determined based on the data source type. This ensures that each product attribute keyword and attribute value in the recognition result is output in a structured field format, rather than as unstructured plain text tags. Based on this, the structured semantic recognition results are mapped to field names and reassembled and encapsulated according to the adaptation interface format specified by the downstream business system. This eliminates the data format gap between the recognition results and the downstream product information management system and search ranking system, enabling the recognition results to be directly and automatically invoked and processed by the downstream system. This avoids the intervention of manual secondary editing and format conversion, thereby significantly shortening the end-to-end processing latency from English text data acquisition to business system data entry or search ranking application while improving semantic recognition accuracy. This comprehensively improves the processing efficiency and business automation level of English semantic recognition tasks in cross-border e-commerce scenarios. Attached Figure Description
[0007] Figure 1 This is a system architecture diagram of a cross-border e-commerce English scene semantic accurate recognition and adaptation system provided in an embodiment of the present invention.
[0008] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0009] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments belong to some, but not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0010] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “said” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.
[0011] Furthermore, the timing of the steps in the following method embodiments is merely an example and not a strict limitation.
[0012] In practice, the server-side equipment deployed in a cross-border e-commerce English scene semantic accurate recognition and adaptation system may consist of one or more devices. This system can be implemented as a business instance, a virtual machine, or hardware devices. For example, it can be implemented as a business instance deployed on one or more devices in a cloud node. Simply put, it can be understood as software deployed on a cloud node to provide a cross-border e-commerce English scene semantic accurate recognition and adaptation system to various user terminals. Alternatively, it can be implemented as a virtual machine deployed on one or more devices in a cloud node, with application software installed to manage various user terminals. Or, it can also be implemented as a server composed of numerous identical or different types of hardware devices, with one or more hardware devices configured to provide a cross-border e-commerce English scene semantic accurate recognition and adaptation system to various user terminals.
[0013] In terms of implementation, a cross-border e-commerce English scene semantic accurate recognition and adaptation system and the user terminal are mutually adapted. That is, if the cross-border e-commerce English scene semantic accurate recognition and adaptation system is implemented as an application installed on a cloud service platform, then the user terminal is a client that establishes a communication connection with the application; or if the cross-border e-commerce English scene semantic accurate recognition and adaptation system is implemented as a website, then the user terminal is implemented as a webpage; or if the cross-border e-commerce English scene semantic accurate recognition and adaptation system is implemented as a cloud service platform, then the user terminal is implemented as a mini-program in an instant messaging application.
[0014] like Figure 1 The diagram shown is a system architecture diagram of a cross-border e-commerce English scene semantic accurate recognition and adaptation system provided by an embodiment of the present invention.
[0015] The cross-border e-commerce English scene semantic accurate recognition and adaptation system 100 described in this invention can be set up in a cloud server. In terms of implementation, it can be used as one or more service devices, or as an application installed in the cloud (e.g., a mobile service operator's server, server cluster, etc.), or it can be developed into a website. Depending on the functions implemented, the cross-border e-commerce English scene semantic accurate recognition and adaptation system 100 may include a semantic tree construction module 101, a scene matching module 102, a conflict disambiguation module 103, a structure generation module 104, and an adaptation encapsulation module 105. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.
[0016] In this embodiment of the invention, in a cross-border e-commerce English scene semantic accurate recognition and adaptation system, each of the above modules can be implemented independently and can call other modules. Here, "calling" can be understood as a module connecting to multiple modules of another type and providing corresponding services to those connected modules. In the cross-border e-commerce English scene semantic accurate recognition and adaptation system provided by this embodiment of the invention, without modifying the program code, the applicable scope of the system architecture can be adjusted by adding modules and directly calling them, achieving cluster-based horizontal expansion to quickly and flexibly expand the cross-border e-commerce English scene semantic accurate recognition and adaptation system. In practical applications, the above modules can be set in the same device or different devices, or they can be set in virtual devices, such as service instances in a cloud server.
[0017] The following describes, with reference to specific embodiments, the various components and specific workflows of a cross-border e-commerce English scenario semantic accurate recognition and adaptation system: The semantic tree construction module 101 is used to set root nodes and intermediate nodes based on domain entity words, attribute words, and hierarchical relationships of entities in the original corpus of cross-border e-commerce, set leaf nodes based on attribute words and entity subordination relationships, and configure a set of synonyms and near-synonyms for each leaf node to construct a semantic tree for English scenarios in cross-border e-commerce. In this embodiment of the invention, the semantic tree construction module, in performing the following steps, sets root nodes and intermediate nodes based on domain entity words, attribute words, and hierarchical relationships of entities from the original corpus of cross-border e-commerce, and sets leaf nodes based on attribute words and entity subordination relationships: Obtain raw corpus data from cross-border e-commerce platforms. The raw corpus data includes product title text, product detail description text, product specification parameter list, user search query logs, and user review interaction text. The word segmentation and part-of-speech tagging of the original corpus data are processed, and the processed noun phrases and noun compound structures are combined into a candidate domain entity word set. Extract the attribute field names from the product specification parameter list to obtain a set of candidate attribute words. The attribute field names include product physical attribute names, product functional attribute names, and product service attribute names. Domain entities are selected based on the co-occurrence frequency of each candidate domain entity in the candidate domain entity set in the original corpus data, and attribute words are selected based on the frequency of association between each candidate attribute word and the domain entity in the original corpus data.
[0018] The construction of the semantic tree for English scenarios in cross-border e-commerce includes: Based on the hierarchical relationship between domain entity words, the topmost domain entity word is set as the root node of the cross-border e-commerce scenario semantic tree; Set the domain entity words located below the root node and having a direct hierarchical relationship with the root node as the intermediate nodes of the scene semantic tree; continue to set the next level intermediate nodes according to the hierarchy between domain entity words, until all words with hierarchical relationships in the domain entity words are traversed; Based on the subordinate relationship between attribute words and domain entity words, the attribute words are set as leaf nodes connected to the nodes represented by the corresponding domain entity words, and a set of synonym expressions and a set of near-synonyms are configured for each leaf node; The set of synonyms contains words that are semantically identical to the attribute words corresponding to the leaf nodes, and the set of near-synonyms contains words that are semantically similar to the attribute words corresponding to the leaf nodes. Construct a semantic tree for English scenarios in cross-border e-commerce based on the root node, intermediate nodes, leaf nodes, and hierarchical relationships.
[0019] The semantic tree construction module retrieves product title text, product detail description text, product specification parameter list, user search query logs, and user review interaction text stored in the cross-border e-commerce platform's database. It reads all five types of text into local memory to complete the full loading, without missing any type of raw corpus data. After all the data is read, a complete raw corpus dataset is generated as the basic input carrier for subsequent processing.
[0020] The English space segmentation rule is used to perform character-by-character segmentation of each English text in the original corpus dataset. The English space is used as the unique segmentation identifier to segment independent word units. The English part-of-speech tagging rule is called simultaneously to match fixed part-of-speech tags to each segmented word unit. The part-of-speech tags are limited to five basic tags: noun, compound noun, adjective, verb and preposition. Only word units with tagging results of noun and compound noun are retained. All the selected and retained word units are uniformly collected and stored in the candidate domain entity word set. The candidate domain entity word set is a fixed one-dimensional text storage array. Each storage unit in the array stores only one independent noun phrase or noun compound structure text.
[0021] The entire list of product specifications in the original corpus dataset is traversed individually. The text of the field name to the left of the colon separator in the specification list is read line by line. All the read field name texts are divided into three fixed category containers: physical attribute name, functional attribute name, and service attribute name. The physical attribute name container stores texts such as size, weight, and material; the functional attribute name container stores texts such as battery life, transmission, and power; and the service attribute name container stores texts such as warranty, delivery, and after-sales service. All texts in the three categories are merged and summarized to generate a candidate attribute word set. The candidate attribute word set is stored in the same one-dimensional text storage array format as the candidate domain entity word set.
[0022] The semantic tree construction module sets a fixed co-occurrence frequency threshold of 20 times. It iterates through each noun phrase text in the candidate domain entity word set one by one, and counts the total number of complete and continuous occurrences of the noun phrase text in a single text segment of the entire original corpus dataset. If the total number of occurrences is greater than or equal to 20, the noun phrase text is transferred to the domain entity word dedicated storage array for retention. If the total number of occurrences is less than 20, it is directly discarded and does not participate in subsequent hierarchical construction operations. After all candidate noun phrases are traversed, a complete domain entity word array is generated.
[0023] The semantic tree construction module sets a fixed association frequency threshold of 12 times. For each attribute field name text in the candidate attribute word set, it iterates through each domain entity word text in the domain entity word array and counts the total number of times the two appear synchronously in the same original corpus text segment. If the total number of synchronous occurrences is greater than or equal to 12, the attribute field name text is transferred to the attribute word's dedicated storage array for retention. If the total number of synchronous occurrences is less than 12, it is directly discarded and does not participate in the subsequent semantic tree leaf node mounting operation. After all candidate attribute words are traversed, a complete attribute word array is generated.
[0024] Read all domain entity texts in the domain entity word array, retrieve the platform's built-in cross-border e-commerce category hierarchy lookup table. The lookup table pre-records the superior and inferior entity identifier codes corresponding to each domain entity word. Domain entity words that do not have superior entity identifier codes in the lookup table are determined as the top-level domain entity words. Generate a semantic tree root node for this type of entity word. Each top-level domain entity word independently corresponds to a unique root node. The internal storage fields of the root node only contain the corresponding domain entity word text and the unique node number.
[0025] Read the lower-level entity identifier code corresponding to each root node, match the domain entity words in the domain entity word array that have the code, generate the first-level intermediate node directly belonging to the corresponding root node for the matched domain entity words, and generate an intermediate node independently for each successfully matched domain entity word. Write the corresponding domain entity word text, its own unique node number, and the node number of the parent root node into the storage fields of the intermediate node to complete all the generation operations of the first-level intermediate nodes.
[0026] Read the lower-level entity identifier code corresponding to each first-level intermediate node, match the domain entity words in the domain entity word array that have the code, generate a second-level intermediate node belonging to the first-level intermediate node for the matched domain entity words, write the corresponding domain entity word text, its own unique node number, and the node number of the parent first-level intermediate node into the internal storage fields of the second-level intermediate node, and repeat the hierarchical matching and generation operation until all domain entity words with upper and lower level identifier codes in the domain entity word array have generated corresponding intermediate nodes, and there are no domain entity words left with unmatched generated nodes.
[0027] The platform's built-in entity-attribute hierarchy mapping table is retrieved. The mapping table pre-records the domain entity text that uniquely corresponds to each attribute text. Each attribute text in the attribute text array is traversed one by one. The corresponding domain entity text is obtained by matching it with the mapping table. The root node or intermediate node corresponding to the domain entity text in the semantic tree is located. An independent leaf node is generated and attached to the root node or intermediate node. The basic storage fields inside the leaf node are written with the corresponding attribute text, its own unique node number, and the node number of the parent node.
[0028] The system retrieves the pre-generated sets of synonyms and near-synonyms based on semantic similarity calculations. The sets of synonyms are written entirely into the dedicated synonym storage field of the corresponding leaf node, and the sets of near-synonyms are written entirely into the dedicated near-synonyms storage field of the corresponding leaf node. The sets of synonyms contain only English words whose semantic similarity to the attribute words bound to the leaf node is greater than or equal to 0.85. The sets of near-synonyms contain only English words whose semantic similarity is greater than 0.6 and less than 0.85. Both sets are stored in a one-dimensional text array format, and there is no situation where the word texts of the two sets are stored interchangeably.
[0029] Integrate all generated root nodes, multi-level intermediate nodes, and mounted leaf nodes, and synchronously bind the pre-determined hierarchical relationships between each node. Use a tree-structured storage file to uniformly store all node data and node association mapping relationships. This tree-structured storage file is the fully constructed semantic tree for cross-border e-commerce English scenarios. Each node data carries a unique node number and a parent node number. You can retrieve all child nodes in a forward direction by the node number, or retrieve the parent node in a reverse direction.
[0030] The scene matching module 102 is used to input cross-border e-commerce English text data into the scene semantic tree as structured text data, and match keywords with leaf nodes, synonyms, and near-synonyms according to the matching strategy corresponding to the text type tags to obtain the initial semantic node matching result. In this embodiment of the invention, configuring a corresponding set of synonym expressions and a set of near-synonyms for each leaf node includes: For each leaf node, extract related words that appear together with the attribute word in the same product description text or the same search query text from the original domain corpus. Calculate the semantic similarity between related words and attribute words, and classify related words with semantic similarity greater than a first preset threshold into a synonym expression set, and classify related words with semantic similarity greater than a second preset threshold and less than or equal to the first preset threshold into a near-synonym expression set, wherein the first preset threshold is greater than the second preset threshold; Semantic similarity is calculated based on the co-occurrence frequency of related words and attribute words in the original domain corpus, the overlap of the context word sets of related words and attribute words, and the edit distance between related words and attribute words.
[0031] Cross-border e-commerce English text data is input into a semantic tree as structured text data. Following the matching strategy corresponding to the text type tags, keywords are matched against leaf nodes, synonyms, and near-synonyms to obtain initial semantic node matching results. Specifically, this is used for: Based on the source identifier and data interface field name carried in the cross-border e-commerce English text data, determine the data source type of the cross-border e-commerce English text data. The data source types include product title source type, product detail description source type, product specification parameter source type, and user search query source type. Based on the data source type, cross-border e-commerce English text data is assigned corresponding text type labels to filter the target matching strategy corresponding to the text type label from multiple matching strategies. The target matching strategy specifies the matching priority order and matching tolerance parameter. The matching priority order and matching tolerance parameter corresponding to different text type labels are different from each other. According to the matching priority order in the target matching strategy, the keywords in the structured text data are compared one by one with the attribute words corresponding to each leaf node of the scene semantic tree, the words in the synonym expression set of each leaf node, and the words in the near-synonym expression set of each leaf node. Record the leaf nodes that are successfully matched as the hit leaf nodes, and use the hit leaf nodes and their hierarchical paths in the scene semantic tree as the initial semantic node matching results corresponding to the cross-border e-commerce English text data.
[0032] The step of assigning corresponding text type tags to cross-border e-commerce English text data based on the data source type, in order to filter target matching strategies corresponding to the text type tags from multiple matching strategies, includes: When the data source type is the product title source type, assign a product title text type label to the cross-border e-commerce English text data, and select the first matching strategy corresponding to the product title text type label as the target matching strategy. The first matching strategy stipulates that the matching priority order is the product brand words, product category words, and product specification attribute words, and stipulates that the matching error tolerance of the product title source type is zero. When the data source type is the product details description source type, the English text data of cross-border e-commerce is assigned a product description text type label, and the second matching strategy corresponding to the product description text type label is selected as the target matching strategy. The second matching strategy stipulates that the matching priority order is the product specification attribute words, product function description words, and product brand words, and the matching fault tolerance of the product details description source type is the first fault tolerance. When the data source type is the product specification parameter source type, assign a product specification text type label to the cross-border e-commerce English text data, and select the third matching strategy corresponding to the product specification text type label as the target matching strategy. The third matching strategy stipulates that the matching priority order is based on the order of product attribute name words and product attribute value words, and stipulates that the matching fault tolerance of the product specification parameter source type is zero. When the data source type is the user search query source type, the cross-border e-commerce English text data is assigned a user query text type label, and the fourth matching strategy corresponding to the user query text type label is selected as the target matching strategy. The fourth matching strategy stipulates that the matching priority order is the core words of the user query intent, the user query limiting words, and the user query supplementary words. It also stipulates that the matching fault tolerance of the user search query source type is the second fault tolerance, which is higher than the first fault tolerance.
[0033] Locate each leaf node that has completed the mounting operation, read the attribute word text bound to the leaf node, traverse all product description text fragments and user search query text fragments in the entire domain original corpus dataset, and search for text lines in each text fragment that contain the complete attribute word. Extract and collect all independent English words that appear in the same text line as the attribute word and are separated by English commas, pauses, and spaces. Store all the collected words in a dedicated one-dimensional array of related words. This one-dimensional array is the original related word storage carrier for the corresponding leaf node.
[0034] The process involves retrieving the text of each independent related word from the one-dimensional array of related words, simultaneously retrieving the text of the attribute words bound to the current leaf node, and performing a comprehensive semantic similarity calculation. The first layer of semantic similarity calculation involves counting the total number of times related words and attribute words appear synchronously and completely in all text fragments of the original domain corpus. This total number is then divided by the total number of text fragments in the original domain corpus to obtain a normalized co-occurrence score. The second layer of semantic similarity calculation involves extracting the context words of each related word (four consecutive words before and after it) and the context words of each attribute word (four consecutive words before and after it). The number of identical words in both sets is then counted, and this number is divided by the number of identical words in each set. The total number of words after merging and deduplication yields the context overlap score. The third layer of semantic similarity calculation involves comparing all English letters and characters of the related words and attribute words character by character, counting the total number of characters that differ between the two texts, dividing the total number of differing characters by the maximum total length of the two texts to obtain the normalized edit distance difference. Subtracting the normalized edit distance difference from the value of 1 yields the edit distance matching score. The final semantic similarity value is obtained by multiplying the normalized co-occurrence score by a fixed weight of 0.4, the context overlap score by a fixed weight of 0.35, and the edit distance matching score by a fixed weight of 0.25, and then summing all the values to obtain a unique result value. This value is the semantic similarity between the related words and attribute words.
[0035] The first preset threshold is fixed at 0.85 and the second preset threshold is fixed at 0.6. The calculated semantic similarity value is compared with 0.85. If the semantic similarity value is greater than 0.85, the associated word is transferred from the one-dimensional array of associated words to the one-dimensional storage array of the exclusive synonym expression set of the corresponding leaf node. The one-dimensional storage array of synonym expression set only retains the word text that meets the comparison condition, and there is no situation where other types of words are mixed in the storage.
[0036] The calculated semantic similarity values are compared with two fixed thresholds of 0.6 and 0.85. If the semantic similarity value is greater than 0.6 and less than or equal to 0.85, the associated word is transferred from the one-dimensional array of associated words to the one-dimensional storage array of the corresponding leaf node's exclusive synonym expression set. The one-dimensional storage array of synonym expression set only retains the word text that meets the interval comparison conditions. The one-dimensional storage array of synonym expression set and the one-dimensional storage array of near-synonym expression set are stored independently of each other, and there are no duplicate word texts in the two arrays.
[0037] Read the cross-border e-commerce English text data processed upstream, call the structured text data output by the preprocessing process, and internally split and store independent keyword text, source identifier field, and data interface field name field. The complete structured text data is then batch transmitted to the constructed cross-border e-commerce English scene semantic tree for reading and matching. The cross-border e-commerce English scene semantic tree fully retains all root nodes, intermediate nodes, leaf nodes, and one-dimensional storage arrays of synonym expression sets and near-synonym expression sets bound to the leaf nodes, which serve as the fixed retrieval carrier for the matching operation.
[0038] The system reads the source identifier field text and data interface field name text stored internally in the structured text data line by line. It then retrieves the platform's built-in source type lookup table, pre-enters a unique data source type code corresponding to each type of source identifier and interface field name in the lookup table, and performs a complete character match between the read field text and the content entered in the lookup table. If a match is successful, it retrieves the data source type bound to the corresponding code. The data source types are strictly limited to four categories: product title source type, product details description source type, product specification parameter source type, and user search query source type. No other additional data source types participate in the judgment process.
[0039] The built-in text tag mapping table is retrieved. The text tag mapping table is pre-entered with unique text type tag characters bound to four types of data sources. The corresponding text type tag character is read according to the determined data source type. The character is written into the exclusive tag storage field in the header of the current structured text data to complete the assignment operation. The structured text data with the assigned value carries a unique text type tag for subsequent matching strategy filtering operations.
[0040] The built-in matching strategy repository is retrieved. The matching strategy repository independently stores four complete matching strategy files. Each matching strategy file records a set of matching priority order and a set of matching fault tolerance parameters. Each of the four matching strategies is bound to a unique text type label character. A complete character comparison is performed between the text type label characters carried by the structured text data and the label characters bound to the header of the matching strategy file. If the comparison is completely consistent, the matching strategy file is retrieved as the target matching strategy for the current structured text data. The other three matching strategy files do not participate in this matching execution process.
[0041] The matching priority order recorded in the target matching strategy file is retrieved. The keyword text stored in the structured text data is extracted in the order of the word categories recorded in the priority order. For each extracted keyword text, all leaf nodes in the semantic tree of the cross-border e-commerce English scenario are traversed. The attribute word text bound to the current leaf node is read and the keyword text is compared with the full character. If the characters are completely consistent, the comparison is considered successful. If the characters are different, all word texts in the one-dimensional storage array of the synonym expression set of the current leaf node are read and compared with the keyword text one by one. When all word texts in the synonym expression set have been compared and there is no matching result, all word texts in the one-dimensional storage array of the near-synonym expression set of the current leaf node are read and compared with the keyword text one by one. After a single keyword has completed the comparison of all leaf nodes, synonyms, and near-synonyms, the next priority keyword is extracted and the complete comparison process is repeated.
[0042] Establish an independent storage array for matched leaf nodes. Whenever a keyword matches the leaf node's attribute words, synonym set, and near-synonym set, store the unique node number of the matched leaf node and the attribute word text bound to the leaf node in the matched leaf node storage array. Simultaneously, retrieve all parent node numbers associated with this leaf node up to the root node number. Concatenate the root node number, multi-level intermediate node numbers, and the current leaf node number in order from top to bottom to form a hierarchical path text. Bind the hierarchical path text to the corresponding leaf node data and store it in the matched leaf node storage array.
[0043] After completing a full comparison of all keywords according to the priority order of the target matching strategy, read all the stored contents in the matching leaf node storage array, and encapsulate all the matching leaf node information recorded in the array and the hierarchical path text corresponding to each matching leaf node into an independent structured data file. This structured data file is the initial semantic node matching result corresponding to the current cross-border e-commerce English text data.
[0044] When the data source type corresponding to the current structured text data is determined to be the product title source type, the product title text type tag characters bound in the text tag mapping table are read and written into the structured text data tag storage field. The first matching strategy file bound to the tag characters in the matching strategy repository is retrieved as the target matching strategy. The first matching strategy file records the matching priority order as follows: product brand words, product category words, and product specification attribute words. The first matching strategy file records the matching fault tolerance parameter as zero fault tolerance. Zero fault tolerance means that the matching process only allows the precise character comparison of leaf node attribute words and synonym expression set words, and does not retrieve the one-dimensional storage array of synonym expression set to participate in the comparison operation.
[0045] When the data source type corresponding to the current structured text data is determined to be the product details description source type, the product description text type tag character bound in the text tag mapping table is read and written into the structured text data tag storage field. The second matching strategy file bound to the tag character in the matching strategy repository is retrieved as the target matching strategy. The matching priority order recorded in the second matching strategy file is as follows: product specification attribute words, product function description words, and product brand words. The matching fault tolerance parameter recorded in the second matching strategy file is the first fault tolerance. The first fault tolerance is fixed to allow the retrieval of a one-dimensional storage array of synonym expressions to participate in the matching, and a single keyword matching returns a maximum of 3 successfully matched leaf node records.
[0046] When the data source type corresponding to the current structured text data is determined to be the product specification parameter source type, the product specification text type label character bound in the text label mapping table is read and written into the structured text data label storage field. The third matching strategy file bound to the label character in the matching strategy repository is retrieved as the target matching strategy. The matching priority order recorded in the third matching strategy file is product attribute name words and product attribute value words. The matching fault tolerance parameter recorded in the third matching strategy file is zero fault tolerance. Zero fault tolerance means that the matching process only allows the precise character comparison of leaf node attribute words and synonym expression set words, and does not retrieve the one-dimensional storage array of synonym expression set to participate in the comparison operation.
[0047] When the data source type corresponding to the current structured text data is determined to be the user search query source type, the user query text type tag characters bound in the text tag mapping table are read and written into the structured text data tag storage field. The fourth matching strategy file bound to the tag characters in the matching strategy repository is retrieved as the target matching strategy. The fourth matching strategy file records the matching priority order as follows: user query intent core words, user query limiting words, and user query supplementary words. The fourth matching strategy file records the matching fault tolerance parameter as the second fault tolerance. The second fault tolerance is fixed to allow the retrieval of the one-dimensional storage array of synonym expression sets to participate in the matching, and a maximum of 8 successfully matched leaf node records are returned in a single keyword matching. The upper limit of 8 records returned in a single matching of the second fault tolerance is greater than the upper limit of 3 records returned in a single matching of the first fault tolerance.
[0048] The conflict disambiguation module 103 is used to determine that there is a cross-node semantic conflict when the leaf nodes hit by the two sets of keywords do not have a common direct parent node, extract the auxiliary qualifiers of the corresponding context, use the auxiliary qualifiers to correct the semantic orientation of the leaf nodes, and obtain the disambiguated target semantic node matching result. In this embodiment of the invention, the conflict disambiguation module determines a cross-node semantic conflict when the leaf nodes of two sets of keywords do not have a common direct parent node. Specifically, it is used to: Obtain the first leaf node matched by the first keyword and the second leaf node matched by the second keyword in the initial semantic node matching results; Get the direct parent node of the first leaf node in the scene semantic tree, and get the direct parent node of the second leaf node in the scene semantic tree; When the direct parent node of the first leaf node and the direct parent node of the second leaf node are different nodes, it is determined that the first leaf node and the second leaf node do not have a common direct parent node, and it is determined that there is a cross-node semantic conflict in the cross-border e-commerce English text data.
[0049] When extracting auxiliary qualifiers from the corresponding context, and using these qualifiers to correct the semantic orientation of leaf nodes to obtain the disambiguated target semantic node matching result, this is specifically used for: After determining that there is a cross-node semantic conflict, extract the context text fragments adjacent to the first keyword and the second keyword from the cross-border e-commerce English text data; Extract the first auxiliary qualifier from the context text fragment corresponding to the first keyword, and extract the second auxiliary qualifier from the context text fragment corresponding to the second keyword; Match the first auxiliary qualifier with each leaf node under the branch where the first leaf node is located in the scene semantic tree, and take the successfully matched leaf node as the first target leaf node after the first keyword is corrected. The second auxiliary qualifier is matched with each leaf node under the branch where the second leaf node is located in the scene semantic tree, and the leaf node that is successfully matched is taken as the second target leaf node after the second keyword is corrected. Replace the first leaf node with the first target leaf node, and replace the second leaf node with the second target leaf node to obtain the disambiguated target semantic node matching result.
[0050] Read the structured data file of the initial semantic node matching results output by the scene matching module, traverse all the matching leaf node records stored in the file, select any two matching records corresponding to independent keywords and mark them as the first keyword matching record and the second keyword matching record respectively, extract the complete node identification information from the first keyword matching record to generate the complete data carrier of the first leaf node, extract the complete node identification information from the second keyword matching record to generate the complete data carrier of the second leaf node, and both the first leaf node and the second leaf node fully carry three types of fixed storage fields: exclusive node number, hierarchical path text, and bound attribute word text.
[0051] Retrieve the hierarchical path text stored inside the first leaf node, and split the hierarchical path text into characters according to a preset uniform delimiter. The split character fragments are, from top to bottom, the root node number, the multi-level intermediate node number, and the first leaf node's own number. The second to last character fragment in the split result is the direct parent node number of the first leaf node. Based on this number, retrieve the complete node data of the direct parent node of the first leaf node from the semantic tree node storage file of the cross-border e-commerce English scenario.
[0052] The hierarchical path text stored inside the second leaf node is retrieved. The hierarchical path text is segmented using the same character splitting rules as the first leaf node. The segmented character fragments are, from top to bottom, the root node number, the multi-level intermediate node number, and the second leaf node's own number. The second to last character fragment in the segmentation result is the direct parent node number of the second leaf node. Based on this number, the complete node data of the direct parent node of the second leaf node is retrieved from the semantic tree node storage file of the cross-border e-commerce English scenario.
[0053] Extract the unique ID of the direct parent node of the first leaf node and the unique ID of the direct parent node of the second leaf node. Perform a full character-by-character comparison on the two sets of IDs. If all characters of the two sets of IDs are completely identical, it is determined that the two leaf nodes have a common direct parent node. If any character of the two sets of IDs is different, it is determined that the first leaf node and the second leaf node do not have a common direct parent node. Simultaneously, write a fixed conflict identifier character into the conflict marker field of the cross-border e-commerce English text data to complete the cross-node semantic conflict determination operation.
[0054] After detecting that a fixed conflict identifier character has been written into the conflict marker field of the cross-border e-commerce English text data, the original complete cross-border e-commerce English text data is retrieved. The fixed context capture window length is set to 4 consecutive English words before and after the keyword. Taking the complete text of the first keyword as the center, 4 words are read forward and 4 words are read backward. All the words read are concatenated according to the original text order to generate the context text fragment corresponding to the first keyword. Taking the complete text of the second keyword as the center, 4 words are read forward and 4 words are read backward. All the words read are concatenated according to the original text order to generate the context text fragment corresponding to the second keyword. The two types of context text fragments are stored independently and there is no word cross-use.
[0055] English part-of-speech tagging is performed on the context text fragments corresponding to the first keyword. Only the words whose part-of-speech tagging results are limited to adjectives or nouns are retained. All the retained words are collected and stored in a one-dimensional text array as the first auxiliary limiting word set. The word text with the first and first position in the first auxiliary limiting word set is extracted as the unique first auxiliary limiting word. The remaining words do not participate in the subsequent matching and correction process.
[0056] The context text fragments corresponding to the second keyword are filtered using the same part-of-speech filtering rules as the context of the first keyword. Only the word texts whose part-of-speech tags are limited to adjectives or nouns are retained. All the retained words are collected and stored in a one-dimensional text array as the second auxiliary limiting word set. The word text with the first and second-ranked words in the second auxiliary limiting word set is extracted as the unique second auxiliary limiting word. The remaining words do not participate in the subsequent matching and correction process.
[0057] Read the complete hierarchical path text of the first leaf node, locate the direct parent node number corresponding to the first leaf node, search all child nodes under the direct parent node in the semantic tree node storage file of the cross-border e-commerce English scenario, filter all nodes of type leaf node and uniformly collect them into an array of all candidate leaf nodes under the branch where the first leaf node is located. Each storage unit in the array carries complete data including complete node number, bound attribute words, synonym expression set, and near-synonym expression set.
[0058] The first auxiliary qualifier text is compared with the attribute words bound to each leaf node in the candidate leaf node array, all words in the synonym expression set, and all words in the near-synonym expression set. If any of the compared characters are completely matched, the leaf node is determined to be a successful match. The complete node data of the successfully matched leaf node is extracted separately to generate the first target leaf node. If there is no successfully matched node in the candidate leaf node array, the original first leaf node is retained without modification.
[0059] Read the complete hierarchical path text of the second leaf node, locate the direct parent node number corresponding to the second leaf node, search all child nodes under the direct parent node in the semantic tree node storage file of the cross-border e-commerce English scenario, filter all nodes of type leaf node and uniformly collect them into an array of all candidate leaf nodes under the branch where the second leaf node is located. Each storage unit in the array carries complete data including complete node number, bound attribute words, synonym expression set, and near-synonym expression set.
[0060] The second auxiliary qualifier text is compared with the attribute words bound to each leaf node in the candidate leaf node array, all words in the synonym expression set, and all words in the near-synonym expression set. If any of the compared characters are completely matched, the leaf node is determined to be a successful match. The complete node data of the successfully matched leaf node is extracted separately to generate the second target leaf node. If there is no successfully matched node in the candidate leaf node array, the original second leaf node is retained without modification.
[0061] Create a new independent node to replace the storage array, delete the complete data of the first leaf node originally stored in the array, write all node information of the first target leaf node into the array, delete the complete data of the second leaf node originally stored in the array, write all node information of the second target leaf node into the array, and keep the data of the other non-conflicting matching leaf nodes in the array in their original storage state without modification.
[0062] The system reads all corrected leaf node records in the storage array and replaces each leaf node with its corresponding complete hierarchical path text. All data is then packaged into an independent structured data file, which is the target semantic node matching result after the ambiguity correction operation.
[0063] The structured generation module 104 is used to call the structured output template according to the text type label, fill the product attribute keywords and attribute values corresponding to the leaf nodes in the target semantic node matching result into the template field, and generate the structured semantic recognition result. In this embodiment of the invention, the structured generation module executes the function of calling a structured output template based on text type tags, filling the template fields with the product attribute keywords and attribute values corresponding to the leaf nodes matched in the target semantic node matching results, and generating structured semantic recognition results. Specifically, it is used for: The corresponding template identifier is determined based on the text type label, and the structured output template corresponding to the text type label is retrieved from the preset structured template library according to the template identifier; Traverse each matched leaf node in the target semantic node matching results, read the product attribute keywords and attribute values corresponding to each matched leaf node, and determine the product attribute category to which each matched leaf node belongs in the scene semantic tree. According to the product attribute category to which each hit leaf node belongs, the product attribute keywords and attribute values corresponding to each hit leaf node are filled into the corresponding product attribute fields in the structured output template, and the filled structured output template is used as the structured semantic recognition result.
[0064] After reading the structured data file of the target semantic node matching results after dedisambiguation, extract the text type label characters stored in the fixed header of the file, and retrieve the built-in label template mapping table of the system. Each set of text type label characters in the mapping table is uniquely bound to a fixed template identifier code. Perform a character-by-character complete match between the extracted text type label characters and the label characters in the mapping table. After a successful match, read the corresponding bound template identifier code as the sole retrieval basis for retrieving the template.
[0065] Retrieve the structured template library from the local persistent storage. The structured template library has independent storage partitions created according to the template identifier code. Each storage partition contains a complete structured output template text file. Use the template identifier code obtained in the previous step as the search keyword to accurately locate the corresponding storage partition. Read the complete structured output template text file in the partition into the memory cache area. Only the single structured output template that is currently matched is retained in the cache. Files in other template partitions are not loaded into memory.
[0066] Read all matching leaf node records in the structured data file of the target semantic node matching results line by line. Each matching leaf node record stores three types of data: product attribute keyword text, product attribute value text, and complete hierarchical path text. When reading a matching leaf node record, the product attribute keyword and product attribute value text in that record are cached separately, and the hierarchical path text carried by that record is parsed simultaneously to determine the product attribute category.
[0067] The hierarchical path text carried by a single hit leaf node record is split using a fixed delimiter. After splitting, the root node number, multi-level intermediate node numbers, and leaf node numbers are obtained from top to bottom. The class identification text bound to the first intermediate node above the leaf node in the hierarchical path is extracted. This class identification text is pre-divided into three fixed product attribute categories: physical attributes, functional attributes, and service attributes. The product attribute category name to which the current hit leaf node belongs is determined based on the matching and attribution relationship of the class identification text.
[0068] Retrieve all preset fields within the structured output template in the memory cache. Each product attribute field in the template is marked with a unique category identifier text at the front. Perform a full character comparison between the product attribute category name to which the currently hit leaf node belongs and the category identifier text marked at the front of the template field. If the characters are completely consistent, the position of the target product attribute field to be filled is located.
[0069] Write the currently cached product attribute keyword text into the attribute name subfield of the target product attribute field, and write the currently cached product attribute value text into the attribute value subfield of the target product attribute field. For records where no valid product attribute value text is read, write the fixed null flag character into the attribute value subfield. After a single matching leaf node data is filled, switch to the next matching leaf node record and repeat the complete process of category determination, field positioning, and content writing until all matching leaf node records in the target semantic node matching result have been filled.
[0070] After all the product attribute keywords and attribute values corresponding to the leaf nodes are matched and the corresponding fields are filled, the complete text of the structured output template that has completed all write operations in the memory cache is exported. The filled structured output template is then packaged separately to generate a standardized structured data file, which is the structured semantic recognition result.
[0071] The adaptation and encapsulation module 105 is used to encapsulate the structured semantic recognition results according to the adaptation interface format of the downstream business and generate adaptation output data.
[0072] In this embodiment of the invention, when the adaptation and encapsulation module performs the encapsulation process of the structured semantic recognition result according to the adaptation interface format of the downstream business to generate adaptation output data, it is specifically used for: Obtain the interface format specifications of downstream business systems. The interface format specifications contain the data organization structure and field naming rules specified by the downstream business systems. According to the data organization structure in the interface format specification, the product attribute fields and the attribute recognition values corresponding to each product attribute field in the structured semantic recognition result are reorganized into a target data structure that conforms to the data organization structure. According to the field naming rules in the interface format specification, map the field names of each data item in the target data structure to target field names that can be recognized by the downstream business system; After the field name mapping is completed, the target data structure is encapsulated into adapted output data that can be directly called by the downstream business system according to the data transmission protocol of the downstream business system.
[0073] The system reads the standardized structured data file of the structured semantic recognition result output by the structured generation module, sends a fixed identifier read request to the dedicated configuration storage partition of the downstream business system, and pre-stores the corresponding interface format specification text file of the downstream business system in the configuration storage partition. After receiving the request, the system completely returns the interface format specification text file to the local memory cache area. The interface format specification text file completely retains three types of indispensable content: fixed data organization structure text, field naming mapping table, and data transmission protocol identifier.
[0074] The interface format specification text file within the memory cache is split to extract the data organization structure definition content. The data organization structure uniformly adopts a multi-level nested key-value pair hierarchy architecture. The architecture is fixedly divided into three levels: the first-level overall product object, the second-level attribute category sub-objects, and the third-level attribute key-value units. The text of all product attribute fields and the attribute identification value text bound to each product attribute field are read layer by layer from the standardized structured data file of the structured semantic recognition results. The single group of product attribute fields and corresponding attribute identification values are paired and stored in the third-level attribute key-value units. According to the product attribute belonging to the category, they are collected to the corresponding second-level attribute category sub-objects. All second-level attribute category sub-objects are uniformly collected into the internal storage of the first-level overall product object. The multi-level nested key-value pair carrier after collection is the target data structure.
[0075] Extract the field naming mapping lookup table attached to the interface format specification text file. The lookup table uses the original field names inside the system as the primary key for retrieval and the field names that can be recognized by downstream businesses as the mapping subkeys. Traverse all third-level attribute key-value units inside the target data structure, read the original product attribute field names stored inside the unit, and perform full character exact matching between the original field names and the retrieval primary key in the lookup table. If a match is successful, use the corresponding mapping subkey in the lookup table to overwrite and replace the original field names in the third-level attribute key-value units. After all third-level attribute key-value units have completed the field name replacement operation, the target data structure with completed field mapping is generated.
[0076] Extract the data transmission protocol identifier characters recorded in the interface format specification text file. The pre-defined compatible transmission protocols are limited to two types: JSON and Protobuf. The protocol identifier characters are uniquely bound to the two types of protocols. Based on the extracted protocol identifier characters, the unique data transmission protocol to be used is determined. After the target data structure with the field names mapped, the serialization encoding operation specified by the corresponding protocol is performed. The JSON transmission protocol uses a fixed character encoding rule of enclosing field names in double quotes, separating fields and values with colons, and distinguishing levels with curly braces. The Protobuf transmission protocol uses a fixed binary encoding rule of concatenating the pre-defined binary field number with the corresponding data length. After serialization encoding, a binary data stream with a complete call identifier and a verification length field is generated. This binary data stream is the adapted output data that can be directly called by downstream business systems.
[0077] 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 present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.
[0078] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.
[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.
Claims
1. A cross-border e-commerce English scene semantic accurate identification adaptation system, characterized in that, The system includes a semantic tree construction module, a scene matching module, a conflict disambiguation module, a structure generation module, and an adaptation and encapsulation module, wherein: The semantic tree construction module is used to set root nodes and intermediate nodes based on domain entity words, attribute words, and hierarchical relationships of entities in the original corpus of cross-border e-commerce, set leaf nodes based on attribute words and entity subordination relationships, and configure a set of synonyms and near-synonyms for each leaf node to construct a semantic tree for English cross-border e-commerce scenarios. The scene matching module is used to input cross-border e-commerce English text data as structured text data into the scene semantic tree, and match keywords with leaf nodes, synonyms, and near-synonyms according to the matching strategy corresponding to the text type label to obtain the initial semantic node matching results. The conflict disambiguation module is used to determine cross-node semantic conflict when the leaf nodes matched by two sets of keywords do not have a common direct parent node. It extracts the auxiliary qualifiers of the corresponding context, uses the auxiliary qualifiers to correct the semantic orientation of the leaf nodes, and obtains the disambiguated target semantic node matching result. The structured generation module is used to call the structured output template based on the text type label, and fill the template field with the product attribute keywords and attribute values corresponding to the leaf nodes in the target semantic node matching results to generate structured semantic recognition results. The adaptation and encapsulation module is used to encapsulate the structured semantic recognition results according to the adaptation interface format of downstream businesses and generate adaptation output data. 2.The cross-border e-commerce English scene semantic accurate recognition and adaptation system of claim 1, wherein, The semantic tree construction module, in the process of setting root nodes and intermediate nodes based on domain entity words, attribute words, and hierarchical relationships of entities from the original corpus of cross-border e-commerce, and setting leaf nodes based on attribute words and entity subordination relationships, includes: Obtain raw corpus data from cross-border e-commerce platforms. The raw corpus data includes product title text, product detail description text, product specification parameter list, user search query logs, and user review interaction text. The word segmentation and part-of-speech tagging of the original corpus data are processed, and the processed noun phrases and noun compound structures are combined into a candidate domain entity word set. Extract the attribute field names from the product specification parameter list to obtain a set of candidate attribute words. The attribute field names include product physical attribute names, product functional attribute names, and product service attribute names. Domain entities are selected based on the co-occurrence frequency of each candidate domain entity in the candidate domain entity set in the original corpus data, and attribute words are selected based on the frequency of association between each candidate attribute word and the domain entity in the original corpus data.
3. The cross-border e-commerce English scene semantic accurate recognition and adaptation system as described in claim 2, characterized in that, The construction of the semantic tree for English scenarios in cross-border e-commerce includes: Based on the hierarchical relationship between domain entity words, the topmost domain entity word is set as the root node of the cross-border e-commerce scenario semantic tree; Set the domain entity words located below the root node and having a direct hierarchical relationship with the root node as the intermediate nodes of the scene semantic tree; continue to set the next level intermediate nodes according to the hierarchy between domain entity words, until all words with hierarchical relationships in the domain entity words are traversed; Based on the subordinate relationship between attribute words and domain entity words, the attribute words are set as leaf nodes connected to the nodes represented by the corresponding domain entity words, and a set of synonym expressions and a set of near-synonyms are configured for each leaf node; The set of synonyms contains words that are semantically identical to the attribute words corresponding to the leaf nodes, and the set of near-synonyms contains words that are semantically similar to the attribute words corresponding to the leaf nodes. Construct a semantic tree for English scenarios in cross-border e-commerce based on the root node, intermediate nodes, leaf nodes, and hierarchical relationships.
4. The cross-border e-commerce English scene semantic accurate recognition and adaptation system as described in claim 1, characterized in that, The configuration of a corresponding set of synonym expressions and a set of near-synonyms for each leaf node includes: For each leaf node, extract related words that appear together with the attribute word in the same product description text or the same search query text from the original domain corpus. Calculate the semantic similarity between related words and attribute words, and classify related words with semantic similarity greater than the first preset threshold into the synonym expression set, and classify related words with semantic similarity greater than the second preset threshold and less than or equal to the first preset threshold into the near-synonym expression set, wherein the first preset threshold is greater than the second preset threshold; Semantic similarity is calculated based on the co-occurrence frequency of related words and attribute words in the original domain corpus, the overlap of the context word sets of related words and attribute words, and the edit distance between related words and attribute words.
5. The cross-border e-commerce English scene semantic accurate recognition and adaptation system as described in claim 4, characterized in that, The scenario matching module inputs cross-border e-commerce English text data as structured text data into the scenario semantic tree, and matches keywords with leaf nodes, synonyms, and near-synonyms according to the matching strategy corresponding to the text type tags to obtain initial semantic node matching results. Specifically, it is used for: Based on the source identifier and data interface field name carried in the cross-border e-commerce English text data, determine the data source type of the cross-border e-commerce English text data. The data source types include product title source type, product detail description source type, product specification parameter source type, and user search query source type. Based on the data source type, cross-border e-commerce English text data is assigned corresponding text type labels to filter the target matching strategy corresponding to the text type label from multiple matching strategies. The target matching strategy specifies the matching priority order and matching tolerance parameter. The matching priority order and matching tolerance parameter corresponding to different text type labels are different from each other. According to the matching priority order in the target matching strategy, the keywords in the structured text data are compared one by one with the attribute words corresponding to each leaf node of the scene semantic tree, the words in the synonym expression set of each leaf node, and the words in the near-synonym expression set of each leaf node. Record the leaf nodes that are successfully matched as the hit leaf nodes, and use the hit leaf nodes and their hierarchical paths in the scene semantic tree as the initial semantic node matching results corresponding to the cross-border e-commerce English text data.
6. The cross-border e-commerce English scene semantic accurate recognition and adaptation system as described in claim 5, characterized in that, The step of assigning corresponding text type tags to cross-border e-commerce English text data based on the data source type, in order to filter target matching strategies corresponding to the text type tags from multiple matching strategies, includes: When the data source type is the product title source type, assign a product title text type label to the cross-border e-commerce English text data, and select the first matching strategy corresponding to the product title text type label as the target matching strategy. The first matching strategy stipulates that the matching priority order is the product brand words, product category words, and product specification attribute words, and stipulates that the matching error tolerance of the product title source type is zero. When the data source type is the product details description source type, the English text data of cross-border e-commerce is assigned a product description text type label, and the second matching strategy corresponding to the product description text type label is selected as the target matching strategy. The second matching strategy stipulates that the matching priority order is the product specification attribute words, product function description words, and product brand words, and the matching fault tolerance of the product details description source type is the first fault tolerance. When the data source type is the product specification parameter source type, assign a product specification text type label to the cross-border e-commerce English text data, and select the third matching strategy corresponding to the product specification text type label as the target matching strategy. The third matching strategy stipulates that the matching priority order is based on the order of product attribute name words and product attribute value words, and stipulates that the matching fault tolerance of the product specification parameter source type is zero. When the data source type is the user search query source type, the cross-border e-commerce English text data is assigned a user query text type label, and the fourth matching strategy corresponding to the user query text type label is selected as the target matching strategy. The fourth matching strategy stipulates that the matching priority order is the core words of the user query intent, the user query limiting words, and the user query supplementary words. It also stipulates that the matching fault tolerance of the user search query source type is the second fault tolerance, which is higher than the first fault tolerance.
7. The cross-border e-commerce English scene semantic accurate recognition and adaptation system as described in claim 1, characterized in that, The conflict disambiguation module determines a cross-node semantic conflict when the leaf nodes of two sets of keywords do not share a common direct parent node. Specifically, it is used for: Obtain the first leaf node matched by the first keyword and the second leaf node matched by the second keyword in the initial semantic node matching results; Get the direct parent node of the first leaf node in the scene semantic tree, and get the direct parent node of the second leaf node in the scene semantic tree; When the direct parent node of the first leaf node and the direct parent node of the second leaf node are different nodes, it is determined that the first leaf node and the second leaf node do not have a common direct parent node, and it is determined that there is a cross-node semantic conflict in the cross-border e-commerce English text data.
8. The cross-border e-commerce English scene semantic accurate recognition and adaptation system as described in claim 7, characterized in that, The conflict disambiguation module, when extracting auxiliary qualifiers from the corresponding context, using these qualifiers to correct the semantic orientation of leaf nodes, and obtaining the disambiguated target semantic node matching result, is specifically used for: After determining that there is a cross-node semantic conflict, extract the context text fragments adjacent to the first keyword and the second keyword from the cross-border e-commerce English text data; Extract the first auxiliary qualifier from the context text fragment corresponding to the first keyword, and extract the second auxiliary qualifier from the context text fragment corresponding to the second keyword; Match the first auxiliary qualifier with each leaf node under the branch where the first leaf node is located in the scene semantic tree, and take the successfully matched leaf node as the first target leaf node after the first keyword is corrected. The second auxiliary qualifier is matched with each leaf node under the branch where the second leaf node is located in the scene semantic tree, and the leaf node that is successfully matched is taken as the second target leaf node after the second keyword is corrected. Replace the first leaf node with the first target leaf node, and replace the second leaf node with the second target leaf node to obtain the disambiguated target semantic node matching result.
9. The cross-border e-commerce English scene semantic accurate recognition and adaptation system as described in claim 1, characterized in that, The structured generation module, when executing the call of a structured output template based on text type tags, fills the template fields with the product attribute keywords and attribute values corresponding to the matched leaf nodes from the target semantic node matching results, generating structured semantic recognition results. Specifically, it is used for: The corresponding template identifier is determined based on the text type label, and the structured output template corresponding to the text type label is retrieved from the preset structured template library according to the template identifier; Traverse each matched leaf node in the target semantic node matching results, read the product attribute keywords and attribute values corresponding to each matched leaf node, and determine the product attribute category to which each matched leaf node belongs in the scene semantic tree. According to the product attribute category to which each hit leaf node belongs, the product attribute keywords and attribute values corresponding to each hit leaf node are filled into the corresponding product attribute fields in the structured output template, and the filled structured output template is used as the structured semantic recognition result.
10. The cross-border e-commerce English scene semantic accurate recognition and adaptation system as described in claim 1, characterized in that, When the adaptation and encapsulation module encapsulates the structured semantic recognition results according to the adaptation interface format of downstream services to generate adaptation output data, it is specifically used for: Obtain the interface format specifications of downstream business systems. The interface format specifications contain the data organization structure and field naming rules specified by the downstream business systems. According to the data organization structure in the interface format specification, the product attribute fields and the attribute recognition values corresponding to each product attribute field in the structured semantic recognition result are reorganized into a target data structure that conforms to the data organization structure. According to the field naming rules in the interface format specification, map the field names of each data item in the target data structure to target field names that can be recognized by the downstream business system; After the field name mapping is completed, the target data structure is encapsulated into adapted output data that can be directly called by the downstream business system according to the data transmission protocol of the downstream business system.
