A cross-border e-commerce multilingual content generation system based on a knowledge graph

Abstract

The application discloses a kind of based on knowledge graph's cross-border e-commerce multilingual content generation system, it is related to cross-border electronic commerce information processing technical field, including: analyzing commodity information constructs cross-border commodity knowledge graph;According to target language, sales market and release platform form commodity generation subgraph;Commodity generation subgraph is input into improved CompGCN, and restricted message is formed by bidirectional combination branch;With field capacity gate restricted message, encode field state and form content generation control sequence;Decoding control sequence forms target language text segment, and forms return message along the return path back;According to message difference update rule relationship door value, redirection propagation path, local decoding text segment;According to field node index arrangement and splicing update text segment, form cross-border e-commerce multilingual content.The application can reduce fact bias, rule conflict and field overrun rate.

Description

Technical Field

[0001] This invention relates to the field of cross-border e-commerce information processing technology, and in particular to a cross-border e-commerce multilingual content generation system based on knowledge graphs. Background Technology

[0002] In the field of cross-border e-commerce information processing technology, before goods can enter sales platforms in different countries, the original data needs to be converted into text that conforms to local language habits and publishing requirements. Existing platforms typically extract the content to be expressed from the product data, then use automated generation tools to complete the language conversion, subsequently adjust the text length according to platform requirements, and finally hand it over to humans for verification of accuracy and compliance. Some solutions utilize knowledge graphs to store the relationships between product information, providing content sources for automated generation tools.

[0003] Existing processing methods require repeatedly transferring product information between different processing stages. After language conversion, the information is easily detached from its original attribute context, causing a shift in the product's meaning. When the sales region changes, the generation process struggles to adapt to local expression requirements in a timely manner; when the publishing platform changes, the already generated text needs to be readjusted.

[0004] Existing solutions typically compress the length of the text after it is generated, which can easily lead to the deletion of key content that affects product recognition. When errors occur in the generated results, it is difficult to locate the original information causing the error in the text. It is necessary to rely on manual verification sentence by sentence and regenerate the entire content, which can easily affect the already correct expression. As a result, it is difficult to balance the efficiency and accuracy of multilingual product content production.

[0005] Therefore, how to provide a knowledge graph-based multilingual content generation system for cross-border e-commerce is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a knowledge graph-based multilingual content generation system for cross-border e-commerce. This invention uses an improved CompGCN to associate product facts with language rules, generates multilingual content according to field capacity, and corrects generation deviations through text feedback. It has the advantages of accurate content, strong market adaptability, and high correction efficiency.

[0007] A knowledge graph-based multilingual content generation system for cross-border e-commerce according to an embodiment of the present invention includes the following modules: The product graph module is used to parse product information into fact nodes, align multilingual terms and market rules with fact nodes, instantiate attribute relationships, terminology relationships and rule relationships, and construct a cross-border product knowledge graph. The subgraph generation module is used to trim the cross-border commodity knowledge graph according to the target language, sales market and publishing platform and connect it to field nodes. It measures the field capacity based on the field length and fact importance, marks the propagation path from fact nodes to field nodes, and forms a commodity generation subgraph. The bidirectional combination module is used to input the product-generated subgraph into the improved CompGCN, bind positive and negative relation vectors between the bidirectional combination branches, and combine the fact node vector and attribute relation vector in the positive combination branch to form a fact message, which is then formed into a restricted message through term relation mapping and rule gating. The capacity gating module is used to control the limited message by field capacity, write the actual propagation volume into the field status, pass the remaining message volume to the next field node along the propagation path, and encode the field status into content to generate a control sequence. The text return module is used to generate control sequences from sequence decoding content, form target language text fragments, combine text fragment vectors and inverse relation vectors in reverse combination branches, return them along the propagation path, and aggregate them into return messages at fact nodes; The differential feedback module is used to perform message differential on fact messages and feedback messages, update the relationship threshold according to the message difference update rule and redirect the propagation path, generate a control sequence for the content corresponding to the message difference and re-execute local decoding to form an updated text fragment; The content assembly module is used to rearrange the sequence and concatenate the fields of the updated text fragments according to the field node index, forming multilingual content for cross-border e-commerce.

[0008] Optionally, the product map module specifically includes: Perform field parsing on product information, bind attribute identifiers and attribute values ​​to product identifiers, and encode them as fact nodes; Compare the attribute identifiers corresponding to fact nodes and multilingual terms. If the attribute identifiers are the same, instantiate the term relation and encode the language identifiers into the term relation. Compare the attribute identifiers of fact nodes with the applicable attributes defined by market rules. If the comparison results are consistent, instantiate the rule relationship, encode the market identifier and platform identifier into the rule relationship, and instantiate the attribute relationship between fact nodes based on the attribute association record. Multi-relation graph encoding is performed on fact nodes and instantiated relations according to relation type and relation direction to construct a cross-border commodity knowledge graph.

[0009] Optionally, the subgraph generation module specifically includes: Filter terminology relationships based on the target language, filter rule relationships based on the sales market and publishing platform, and retain the filtered terminology relationships, the rule relationships connected to the filtered terminology nodes, and the corresponding nodes. Configure the encoded field node according to the content field, compare the attribute identifier of the fact node with the attribute identifier that the field node is allowed to accept, and connect the rule node and field node corresponding to the fact node that matches the comparison. Convert the allowed text length of the field into the total capacity, read the fact importance corresponding to the fact node, allocate the total capacity among the fact nodes connected to the same field node according to the fact importance, and use the allocation result as the field capacity. The relationship type and direction are encoded according to the order of calls from fact nodes to field nodes, the propagation path is marked, and the nodes and relationships traversed by the propagation path are preserved to form a product generation subgraph.

[0010] Optionally, the bidirectional combination module specifically includes: The improved CompGCN relation composition layer is split into forward composition branches and reverse composition branches, and the forward relation vector and the corresponding reverse relation vector are bound between the two composition branches; The forward combination branch performs cyclic correlation operations on the fact node vector and attribute relationship vector to form fact messages; Cyclic correlation operations are performed on factual messages and terminology relation vectors. The regular relation vectors are transformed into regular relation gates through linear projection and sigmoid activation. The cyclic correlation results are then modulated element by element through the regular relation gates to form restricted messages.

[0011] Optionally, the binding of positive and inverse relation vectors in the bidirectional combination module includes the following steps: The forward relation vector is transformed into the inverse relation vector using the inverse transformation matrix; The forward combination branch performs a cyclic correlation operation on the input vector and the forward relation vector, while the reverse combination branch performs a cyclic convolution on the cyclic correlation result and the inverse relation vector to form the input recovery vector; Element-wise difference is performed on the input recovered vector and the input vector of the forward combined branch. The difference result is then propagated backward along the bidirectional combined branch to update the forward relation vector and the inverse transformation matrix.

[0012] Optionally, the capacity gating module specifically includes: Restricted messages are mapped to a sequence of message units, with each message unit occupying one field capacity unit; Set the receiving position according to the sum of the field capacities corresponding to the same field node, extract message units from each message unit sequence according to the corresponding field capacity, and complete the allocation according to the attention weight of the extracted message units and the receiving position; The number of message units that have been allocated to the receiving position is counted to form the actual propagation volume. The propagation volume component in the field status is updated according to the actual propagation volume, and the corresponding receiving position component is updated according to the vector of the allocated message units. Extract message units that have not been assigned to the receiving position, count the number of unassigned units to form the remaining message volume, and pass the unassigned message units carrying the remaining message volume along the propagation path to the next field node; The updated field states are arranged according to the field node index, and position encoding is performed on the field states. The encoded results are concatenated to form the content generation control sequence.

[0013] Optionally, the text feedback module specifically includes: A local attention mask is generated based on the range of field states in the content generation control sequence, blocking attention connections across field states, and decoding is performed based on the preserved attention connections to form a target language text fragment. Encode the target language text fragments into text fragment vectors, trace the propagation path back from the corresponding field node to the fact node, and call the bound inverse relation vector according to the reverse order of the relations in the propagation path; The reverse combination branch performs continuous circular convolutions in the order of calling the inverse relation vector, using the result of the previous circular convolution as the input of the next circular convolution, until the result of the circular convolution reaches the fact node; Calculate the attention weights between the fact node vector and the circular convolution results that reach the corresponding fact node along the same propagation path, and perform a weighted summation on the circular convolution results according to the attention weights to form the feedback message.

[0014] Optionally, the differential feedback module specifically includes: Perform element-wise difference analysis on the fact message and the feedback message, and take the absolute value of the difference to form the message difference; The message difference is concatenated with the rule relation vector, and the updated rule relation gate value is output after linear mapping and S-shaped activation. Based on the updated rule relationship threshold, the propagation branch of the corresponding fork node in the original propagation path is switched, and the original propagation path is continued along the switched propagation branch to form the replaced propagation path; The updated rule relation threshold modulates the term mapping result to form an updated restricted message. The updated restricted message is then transmitted along the replaced propagation path, and the corresponding field status is updated based on the transmission result. The updated field state is encoded, and the corresponding part in the control sequence is generated by replacing the content with the encoded result. The replaced part is then partially decoded, and the local content in the corresponding target language text fragment is replaced with the decoded result to form the updated text fragment.

[0015] Optionally, replacing the propagation path in the differential feedback module includes the following steps: Compare the propagation branches before and after the rule relation threshold update, and identify the rule relations in which the propagation branch switch occurs as abnormal relations; Tracing from the bifurcation node corresponding to the abnormal relationship along the switched propagation branch to the first merging node intersecting with the original propagation path, the propagation branch between the bifurcation node and the first merging node is extracted to form an alternative path segment; Replace the original propagation path segment between the fork node and the first merging node with an alternative path segment.

[0016] Optionally, the content assembly module specifically includes: The control sequence part is generated by locating and updating the content corresponding to the text fragment, determining the field node based on the state of the field, and taking the first sequence position as the fragment start position. Updated text fragments are arranged according to the field node index, and updated text fragments belonging to the same field node are arranged according to the fragment's starting position; Updated text fragments with different field node indexes are sorted in ascending order of field node index; updated text fragments with the same field node index are sorted in ascending order of fragment start position. The updated text fragments belonging to the same field node after being arranged are concatenated continuously, and the concatenation results are combined according to the field node index to form multilingual content for cross-border e-commerce.

[0017] The beneficial effects of this invention are: (1) This invention associates product information with target language expressions through cross-border product knowledge graphs, and uses an improved CompGCN to combine product content and market restrictions along relational paths, thereby reducing attribute mistranslation, terminology confusion and illegal expressions, and improving the factual accuracy and market adaptability of multilingual content.

[0018] (2) The present invention allocates field capacity based on field length and fact importance, and controls the entry of messages into the content field by accepting the position, and transfers unaccepted content to the next field, thereby reducing the problem of key product information being truncated and improving field space utilization and content integrity.

[0019] (3) The present invention uses the reverse combination branch to send the generated text back to the fact node, corrects the rule relationship threshold and propagation path through message difference, performs local decoding only on the content corresponding to the deviation, retains the correct text fragments, reduces the amount of repeated generation, and improves the content correction efficiency. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a knowledge graph-based cross-border e-commerce multilingual content generation system proposed in this invention. Figure 2This is a schematic diagram of the improved CompGCN bidirectional combined branch in this invention; Figure 3 This is a schematic diagram illustrating the process of text return, message differentiation, and propagation path correction in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A knowledge graph-based multilingual content generation system for cross-border e-commerce includes the following modules: The product graph module is used to parse product information into fact nodes, align multilingual terms and market rules with fact nodes, instantiate attribute relationships, terminology relationships and rule relationships, and construct a cross-border product knowledge graph. The subgraph generation module is used to trim the cross-border commodity knowledge graph according to the target language, sales market and publishing platform and connect it to field nodes. It measures the field capacity based on the field length and fact importance, marks the propagation path from fact nodes to field nodes, and forms a commodity generation subgraph. The bidirectional combination module is used to input the product-generated subgraph into the improved CompGCN, bind positive and negative relation vectors between the bidirectional combination branches, and combine the fact node vector and attribute relation vector in the positive combination branch to form a fact message, which is then formed into a restricted message through term relation mapping and rule gating. The capacity gating module is used to control the limited message by field capacity, write the actual propagation volume into the field status, pass the remaining message volume to the next field node along the propagation path, and encode the field status into content to generate a control sequence. The text return module is used to generate control sequences from sequence decoding content, form target language text fragments, combine text fragment vectors and inverse relation vectors in reverse combination branches, return them along the propagation path, and aggregate them into return messages at fact nodes; The differential feedback module is used to perform message differential on fact messages and feedback messages, update the relationship threshold according to the message difference update rule and redirect the propagation path, generate a control sequence for the content corresponding to the message difference and re-execute local decoding to form an updated text fragment; The content assembly module is used to rearrange the sequence and concatenate the fields of the updated text fragments according to the field node index, forming multilingual content for cross-border e-commerce.

[0023] In this embodiment, the product map module specifically includes: Perform field parsing on product information, bind attribute identifiers and attribute values ​​to product identifiers, and encode them as fact nodes; Product information is categorized by product identifier, and attributes and their corresponding values ​​are read item by item. Product identifiers are transformed using an embedded table, and then the corresponding attribute vectors are read based on the attribute identifiers. Numerical attribute values ​​are first converted to the unit of measurement used for the corresponding attribute, and then normalized according to the attribute value range; textual attribute values ​​are first segmented, and then average pooling is performed on the segmented vectors. The product identifier vector and attribute vector are combined, and the processed attribute values ​​are concatenated with the combined result, forming a fact node vector through linear mapping. When the same attribute has more than one attribute value, each is used to form a separate fact node.

[0024] Compare the attribute identifiers corresponding to fact nodes and multilingual terms. If the attribute identifiers are the same, instantiate the term relation and encode the language identifiers into the term relation. Multilingual terms are retrieved based on the attribute identifiers corresponding to fact nodes, retaining standard terms with the same attribute identifiers. Standard terms are segmented and average pooled; the processing results are combined with the corresponding attribute vectors and linearly mapped to form term node vectors. Language vectors are read based on the language identifiers corresponding to the standard terms; these language vectors are combined with the term relationship base vectors to form term relationship vectors; term relationships are established according to the direction from fact nodes to term nodes.

[0025] Compare the attribute identifiers of fact nodes with the applicable attributes defined by market rules. If the comparison results are consistent, instantiate the rule relationship, encode the market identifier and platform identifier into the rule relationship, and instantiate the attribute relationship between fact nodes based on the attribute association record. Market rules are retrieved based on the attribute identifiers corresponding to fact nodes, retaining those with consistent applicable attributes. The rule text is segmented and averaged pooled; the processing result is combined with the attribute vectors corresponding to the applicable attributes, and then linearly mapped to form rule node vectors. The market vector corresponding to the market identifier is combined with the action vector corresponding to the rule action, and the combination result is modulated using the platform vector corresponding to the platform identifier to form a rule relationship vector. Terminology nodes connected to the same fact node are located, and rule relationships are established according to the direction from the terminology node to the rule node.

[0026] Starting from the attribute corresponding to the current fact node, read the attribute association record, and locate the target fact node under the same product based on the target attribute pointed to by the record. Combine the type vector corresponding to the association type with the direction vector corresponding to the relationship direction to form the attribute relationship vector; connect the two fact nodes according to the direction specified by the attribute association record.

[0027] Multi-relation graph encoding is performed on fact nodes and instantiated relations according to relation type and relation direction to construct a cross-border commodity knowledge graph.

[0028] Node numbers are assigned according to the order of node generation, and each relationship is organized into a relationship record consisting of a source node, a relationship type, and a target node. A reverse transformation is performed on the forward relationship vector to form a reverse relationship in the opposite direction; self-connection relationships are configured for nodes without adjacency message input. Node vectors and relationship vectors are linearly mapped to the same dimension, and then relationship records are aggregated according to the source node and relationship type to form multi-relational adjacency data for use by the improved CompGCN.

[0029] In this embodiment, the subgraph generation module specifically includes: Filter terminology relationships based on the target language, filter rule relationships based on the sales market and publishing platform, and retain the filtered terminology relationships, the rule relationships connected to the filtered terminology nodes, and the corresponding nodes. Perform a consistency comparison between the language identifier carried by the terminology relationship and the target language identifier, and retain terminology relationships that match the comparison. Perform a consistency comparison between the market identifier and platform identifier carried by the rule relationship and the sales market identifier and the publishing platform identifier, respectively, and retain rule relationships that match in both comparisons and whose source terminology node belongs to the retained node, and retain the nodes connected by the filter relationship.

[0030] Configure the encoded field node according to the content field, compare the attribute identifier of the fact node with the attribute identifier that the field node is allowed to accept, and connect the rule node and field node corresponding to the fact node that matches the comparison. The content field configuration is parsed, and the field types are converted into field node vectors through the embedded table. Each fact node is compared against its corresponding attribute identifier based on the allowed attribute identifier. If a match is found, the rule node corresponding to the fact node is located, and a field connection is established from the rule node to the field node, using the field type vector as the relation vector for the field connection. Adjacent field nodes are connected according to their index order, establishing a successor relationship from the current field node to the next field node, and this successor relationship is then incorporated into the propagation path.

[0031] Convert the allowed text length of the field into the total capacity, read the fact importance corresponding to the fact node, allocate the total capacity among the fact nodes connected to the same field node according to the fact importance, and use the allocation result as the field capacity. The content field configuration limits the allowed text length based on the number of target language tokens, using the maximum allowed number of tokens as the total field capacity. The product attribute configuration assigns fact importance to each attribute identifier, allocating the total field capacity to fact nodes connected to the same field node according to the proportion of fact importance; any remaining capacity that cannot be evenly divided is allocated sequentially from highest to lowest fact importance. The capacity allocated to each fact node becomes the field capacity for the corresponding restricted message.

[0032] The relationship type and direction are encoded according to the order of calls from fact nodes to field nodes, the propagation path is marked, and the nodes and relationships traversed by the propagation path are preserved to form a product generation subgraph.

[0033] Starting with the fact node corresponding to the origin of the attribute relationship, the traversal proceeds along the attribute relationship direction to the fact node corresponding to the destination. At the destination fact node, the terminology relationship matching the target language is invoked, and applicable rule relationships are invoked based on the sales market and publishing platform. The path then connects to the field node that is allowed to accept the corresponding product attribute. The relationship type code and direction code are saved according to the actual invocation order; when propagation branches occur, the invocation order of each branch reaching the field node is saved separately. The nodes and relationships traversed along the propagation path are preserved to form a product generation subgraph.

[0034] In this embodiment, the bidirectional combination module specifically includes: The improved CompGCN relation composition layer is split into forward composition branches and reverse composition branches, and the forward relation vector and the corresponding reverse relation vector are bound between the two composition branches; The original data path of the CompGCN relation combination layer is retained as the forward combination branch, and a reverse combination branch is added in the same graph convolutional layer. The input of the forward combination branch is connected to the fact node, and the output is connected to the term relation processing position; the reverse combination branch has a backpropagation input, receives the forward recurrent correlation results during relation binding training, and receives the target language text fragment vector during text verification, and the output is connected to the fact node along the propagation path. The two branches call the forward relation vector and the reverse relation vector respectively according to the same relation type code.

[0035] The forward combination branch performs cyclic correlation operations on the fact node vector and attribute relationship vector to form fact messages; The fact node vector is used as the shifted vector, and the attribute relation vector is used as the relation kernel vector. The shifted vector is shifted cyclically according to the output position of the fact message. Each shifted component is multiplied by the corresponding component of the relation kernel vector, and the products are accumulated to form the message component at the current position. All output positions are traversed to form a fact message with the same dimension as the fact node vector, and the fact message is passed to the corresponding propagation path.

[0036] Cyclic correlation operations are performed on factual messages and terminology relation vectors. The regular relation vectors are transformed into regular relation gates through linear projection and sigmoid activation. The cyclic correlation results are then modulated element by element through the regular relation gates to form restricted messages.

[0037] The factual message is used as the shifted vector, and the term relation vector is used as the relation kernel vector. A cyclic correlation operation is performed on both to form the term mapping result. The rule relation vector is linearly mapped to a scalar, and a sigmoid activation function is used to form the rule relation gate. Each component of the term mapping result is modulated with the rule relation gate to form the restricted message.

[0038] In this embodiment, the binding of positive and negative relation vectors in the bidirectional combination module includes the following steps: The forward relation vector is transformed into the inverse relation vector using the inverse transformation matrix; Configure an independent inverse transformation matrix for each relation type, with the input and output dimensions of the matrix consistent with the relation vector dimensions. Select the forward relation vector and corresponding inverse transformation matrix based on the relation type code, transform the forward relation vector into an inverse relation vector through matrix transformation, and associate the forward relation vector and inverse relation vector with the same relation type code.

[0039] The forward combination branch performs a cyclic correlation operation on the input vector and the forward relation vector, while the reverse combination branch performs a cyclic convolution on the cyclic correlation result and the inverse relation vector to form the input recovery vector; The forward combination branch performs a cyclic correlation operation on the input vector and the forward relation vector, and feeds the cyclic correlation result into the reverse combination branch. The reverse combination branch performs a cyclic convolution on the cyclic correlation result and the inverse relation vector in the opposite vector movement direction, forming an input recovery vector with the same dimension as the input vector of the forward combination branch.

[0040] Element-wise difference is performed on the input recovered vector and the input vector of the forward combined branch. The difference result is then propagated backward along the bidirectional combined branch to update the forward relation vector and the inverse transformation matrix.

[0041] The difference between the input recovery vector and the input vector of the forward combination branch is calculated element by element. Each difference is squared and summed to form the recovery loss. The recovery loss is used as the training objective for binding the forward and inverse relation vectors. The gradient is propagated along the inverse combination branch and the forward combination branch to update the forward relation vector and the corresponding inverse transformation matrix.

[0042] In this embodiment, the capacity gating module specifically includes: Restricted messages are mapped to a sequence of message units, with each message unit occupying one field capacity unit; The target language terms pointed to by the term relationships are located along the propagation path corresponding to the restricted message. Word segmentation is performed on the target language terms, and word units are transformed using a word vector table. The word unit vectors are concatenated with the restricted message, and a message unit is formed through linear mapping. The message units are then arranged according to their word unit positions.

[0043] Set the receiving position according to the sum of the field capacities corresponding to the same field node, extract message units from each message unit sequence according to the corresponding field capacity, and complete the allocation according to the attention weight of the extracted message units and the receiving position; Summarize the field capacities of fact nodes connected to the same field node, and establish the receiving positions based on the summary results. Select message units from the current message unit sequence that do not exceed the corresponding field capacity according to the word order; unselected message units remain unassigned. Convert the receiving position index to the receiving position code, combine the receiving position code with the field node vector, and form the receiving position vector through linear mapping. Convert the selected message unit vector to the query vector, convert the empty receiving position vector to the key vector, calculate the dot product, and then perform Softmax normalization. Validate message units and receiving positions in descending order of attention weight. Establish an assignment relationship when the message unit is in an unassigned state and the receiving position is in an empty state; when the weights are the same, determine the validation order according to the word position and the receiving position index.

[0044] The number of message units that have been allocated to the receiving position is counted to form the actual propagation volume. The propagation volume component in the field status is updated according to the actual propagation volume, and the corresponding receiving position component is updated according to the vector of the allocated message units. When initializing the field state, the propagation quantity component is set to the zero vector, and the receiving position component is set to the corresponding receiving position vector. After allocating message units corresponding to the same field node, the number of message units in all occupied receiving positions is counted to form the actual propagation quantity. The ratio of the actual propagation quantity to the sum of the field capacities is calculated, and the obtained ratio is linearly mapped to form a propagation quantity vector, which replaces the propagation quantity component. Each allocated message unit vector is element-wise added to the corresponding receiving position vector, and the sum replaces the corresponding receiving position component.

[0045] Extract message units that have not been assigned to the receiving position, count the number of unassigned units to form the remaining message volume, and pass the unassigned message units carrying the remaining message volume along the propagation path to the next field node; Preserving the lexical order of unassigned message units, the remaining message volume is converted into a residual header vector via an embedding table. Unassigned message units are then sequentially appended to the residual header vector to form a residual sequence. The next field node is located based on the successor relationship of the current field node in the propagation path. The next field node parses the residual header vector and extracts message units from the residual sequence according to the parsed remaining message volume, participating in the allocation of the next field node's successor position.

[0046] The updated field states are arranged according to the field node index, and position encoding is performed on the field states. The encoded results are concatenated to form the content generation control sequence.

[0047] The field node index is converted into field position codes via an embedded table. The vector dimension of the field position code is the same as that of the field state component. The field position codes are then layered element by element onto the corresponding field state component. First, the propagation components are arranged, and then the receiving position components are arranged in ascending order according to their receiving position indices. The field states are then concatenated according to their field node index order to form the content generation control sequence for sequence decoding.

[0048] In this embodiment, the text return module specifically includes: A local attention mask is generated based on the range of field states in the content generation control sequence, blocking attention connections across field states, and decoding is performed based on the preserved attention connections to form a target language text fragment. The corresponding permutation range is determined based on the concatenation position of the field status in the content generation control sequence. When calculating the attention score, the attention score between sequence positions within the same permutation range is retained, while the attention score across permutation ranges is replaced with a negative infinity value. The retained attention score is then normalized using the Softmax function. A weighted sum is performed on the vectors of the corresponding sequence positions according to the normalized result. The weighted sum is then linearly mapped to the target language vocabulary. Target language lexical units are determined sequentially based on the vocabulary probability. Lexical unit generation for the current permutation range is stopped when a segment termination marker is detected.

[0049] Encode the target language text fragments into text fragment vectors, trace the propagation path back from the corresponding field node to the fact node, and call the bound inverse relation vector according to the reverse order of the relations in the propagation path; The target language text fragments are converted from words into word vectors using a word vector table. These word vectors are then arranged in the order of word generation, and average pooling is performed on them to form the text fragment vectors. The field node to which the field state used to generate the target language text fragment is located is identified as the corresponding field node, and the propagation path is determined by this corresponding field node. The propagation path is traversed backward from the field node to the fact node. For each relation encountered, the inverse relation vector bound to the forward relation vector is called according to the relation type.

[0050] The reverse combination branch performs continuous circular convolutions in the order of calling the inverse relation vector, using the result of the previous circular convolution as the input of the next circular convolution, until the result of the circular convolution reaches the fact node; The text fragment vector is used as the input for the first circular convolution. Circular convolutions are then performed with the currently invoked inverse relation vector in reverse traversal order. The current input is replaced with the result of the previous circular convolution, and circular convolutions are performed with the next inverse relation vector. After completing the reverse processing of all relations in the propagation path, the final circular convolution result is passed to the corresponding fact node.

[0051] Calculate the attention weights between the fact node vector and the circular convolution results that reach the corresponding fact node along the same propagation path, and perform a weighted summation on the circular convolution results according to the attention weights to form the feedback message.

[0052] The fact node vector is transformed into a query vector through linear projection. The circular convolution results reaching the corresponding fact node along the same propagation path are transformed into key and value vectors through linear projection. The dot product of the query vector and the key vector is calculated. The dot product result is normalized within the range of circular convolution results corresponding to the same fact node and the same propagation path using the Softmax function. The value vector is then weighted and summed according to the normalized result to form the return message, preserving the association between the return message and the propagation path.

[0053] In this embodiment, the differential feedback module specifically includes: Perform element-wise difference analysis on the fact message and the feedback message, and take the absolute value of the difference to form the message difference; After the returned message reaches the fact node, the system searches for the fact message formed by the same fact node during this forward propagation, following the propagation path taken during the returned message. If the propagation paths are consistent, message pairing is established, and the paired messages are aligned according to their vector positions.

[0054] The message difference is concatenated with the rule relation vector, and the updated rule relation gate value is output after linear mapping and S-shaped activation. The rule relationships that participate in forming restricted messages are located along the propagation path after pairing is completed. The message difference is input into the corresponding rule relationship's threshold update process. The updated rule relationship threshold overwrites the original rule relationship threshold, maintaining the binding relationship between the rule relationship and the propagation branch.

[0055] Based on the updated rule relationship threshold, the propagation branch of the corresponding fork node in the original propagation path is switched, and the original propagation path is continued along the switched propagation branch to form the replaced propagation path; The successor path connected to the branch node through each rule relationship is determined as the corresponding propagation branch. The corresponding rule relationship threshold is compared among the propagation branches that can reach the original field node, and the propagation branch with the highest threshold is activated; if the thresholds are the same, the original propagation branch is maintained. If the activated propagation branches before and after the threshold update are different, the original propagation branch is closed, and the corresponding propagation branch after the update is activated.

[0056] The updated rule relation threshold modulates the term mapping result to form an updated restricted message. The updated restricted message is then transmitted along the replaced propagation path, and the corresponding field status is updated based on the transmission result. The updated rule relation threshold is applied to each component of the term mapping result to form an updated restricted message. The updated restricted message is converted into a sequence of message units, preserving the association between message units and source fact nodes, and message units are selected according to lexical order, not exceeding the capacity of the corresponding field. After a message unit reaches a field node, message units associated with the same fact node in the original field state are removed, and the corresponding receiving position is restored to an empty state. The attention weight between the selected message unit and the empty receiving position is calculated, and allocation is completed according to the attention weight. The propagation quantity component and the receiving position component are updated based on the allocation result. Message units that are not selected remain in an unallocated state.

[0057] The updated field state is encoded, and the corresponding part in the control sequence is generated by replacing the content with the encoded result. The replaced part is then partially decoded, and the local content in the corresponding target language text fragment is replaced with the decoded result to form the updated text fragment.

[0058] When the target language text fragment is initially generated, each target language word is associated with the sequence position with the highest attention weight. After the field state is updated, the replacement range is determined based on the concatenation order and number of components of the field state in the content generation control sequence, and the encoding result of the updated field state covers the corresponding range. Attention scores that cross the boundary of the replacement range are replaced with negative infinity values, and local decoding is performed on the replacement range. Based on the sequence position associated with the locally decoded words, local content in the target language text fragment is located, and the located content is replaced with the local decoding result, while retaining the remaining content in the target language text fragment to form a complete updated text fragment.

[0059] In this embodiment, the replacement of the propagation path in the differential feedback module includes the following steps: Compare the propagation branches before and after the rule relation threshold update, and identify the rule relations in which the propagation branch switch occurs as abnormal relations; Before updating the rule relationship threshold, the currently enabled propagation branch of the forked node is retained. After the threshold update is completed, the corresponding scalar rule relationship thresholds are compared among the propagation branches that can reach the original field node. The propagation branch with the largest threshold is determined as the updated enabled propagation branch; if the thresholds are the same, the original propagation branch is retained. If the enabled propagation branches before and after the update are different, the rule relationship controlling the original propagation branch is determined as an abnormal relationship.

[0060] Tracing from the bifurcation node corresponding to the abnormal relationship along the switched propagation branch to the first merging node intersecting with the original propagation path, the propagation branch between the bifurcation node and the first merging node is extracted to form an alternative path segment; Starting from the branch node corresponding to the abnormal relationship, the tracing proceeds node by node along the enabled propagation branch in the direction of the relationship. Upon reaching a subsequent branch node, the tracing continues along the corresponding enabled propagation branch. At each node reached, it is compared with nodes in the original propagation path that are after the tracing starting point; the first matching node is designated as the first merging node. Re-entering already visited nodes is prohibited during the tracing process.

[0061] Replace the original propagation path segment between the fork node and the first merging node with an alternative path segment.

[0062] Close the path segment from the fork node to the first merge node in the original propagation path, and enable the alternative path segment. The path segments before the fork node and after the first merge node remain enabled.

[0063] In this embodiment, the content assembly module specifically includes: The control sequence part is generated by locating and updating the content corresponding to the text fragment, determining the field node based on the state of the field, and taking the first sequence position as the fragment start position. When concatenating field states, maintain the correspondence between the sequence position of the field state and the field node index. After completing the partial content replacement, locate the corresponding content to generate a control sequence part based on the field state of the updated text fragment, obtain the field node index, and use the first sequence position of the content to generate the control sequence part as the fragment's starting position.

[0064] Updated text fragments are arranged according to the field node index, and updated text fragments belonging to the same field node are arranged according to the fragment's starting position; Updated text fragments with different field node indexes are sorted in ascending order of field node index; updated text fragments with the same field node index are sorted in ascending order of fragment start position. The updated text fragments belonging to the same field node after being arranged are concatenated continuously, and the concatenation results are combined according to the field node index to form multilingual content for cross-border e-commerce.

[0065] Before concatenation, the start and end markers of the updated text fragments are removed, and the word sequences corresponding to the same field nodes are connected end-to-end according to the order of arrangement. The concatenation results of different field nodes are used as corresponding content fields, and the content fields are combined according to the field node index to form multilingual content for cross-border e-commerce.

[0066] Example 1: To verify the feasibility of this invention in practice, it was applied to a multilingual content publishing scenario for cross-border e-commerce products. This scenario generates product titles, selling points, and details in the target language based on product information, and must meet the content requirements of the sales market and the publishing platform. Existing fixed template translations are prone to altering product facts, mixing terminology, and the generated content may violate market rules or exceed field lengths; when local errors are discovered, it is usually necessary to regenerate the entire field.

[0067] In implementation, product information is categorized according to product identifiers, and fact nodes are formed by binding attribute identifiers with attribute values. Multilingual terminology and market rules are matched based on attribute identifiers to establish terminological relationships between fact nodes and terminology nodes, and rule relationships between terminology nodes and rule nodes. Attribute relationships are established between corresponding fact nodes based on the association records between product attributes. All nodes and relationships are organized according to relationship type and direction, forming a cross-border product knowledge graph.

[0068] After receiving the target language, sales market, and publishing platform, the terminology relationship between the language identifier and the target language is preserved, as are the rule relationships between the market identifier and the platform identifier, respectively, consistent with the sales market and publishing platform. Field nodes are formed based on the field configurations of the title, selling points, and details. Field nodes that can bear corresponding product attributes are connected to rule nodes. The allowed text length of a field is converted into its total capacity and allocated to fact nodes connected to the same field node according to fact importance, forming the corresponding field capacity. The propagation path is marked according to the order of calls from fact nodes to field nodes, preserving the nodes and relationships traversed by the propagation path to form a product generation subgraph.

[0069] The product generation subgraph inputs to the improved CompGCN. The forward combination branch performs cyclic correlation operations on the fact node vector and attribute relation vector to form a fact message; the fact message and term relation vector then undergo cyclic correlation operations to form a term mapping result. The regular relation vector is linearly projected and subjected to sigmoid activation to form a regular relation gate; the regular relation gate modulates the term mapping result to form a restricted message. The reverse combination branch calls the inverse relation vector bound to the forward relation vector according to the relation type; the forward relation vector is converted into an inverse relation vector through the corresponding inverse transformation matrix.

[0070] The target language terms are located along the propagation path corresponding to the restricted message. These terms are then divided into lexical units, and their vectors are combined with the restricted message to form a message unit sequence. Receiving positions are set according to the sum of the field capacities corresponding to the same field node. Message units are extracted from each message unit sequence according to their corresponding field capacities, and the attention weight between the extracted message units and their receiving positions is calculated. The allocation is then completed according to these attention weights. The number of allocated message units constitutes the actual propagation volume, and the vectors of the allocated message units are used to update the receiving position components in the field states.

[0071] For example, the title field has 40 continuation positions, and a total of 43 message units arrive at the title field. Based on the field capacity corresponding to each fact node, the 40 message units are allocated to the continuation positions, resulting in an actual propagation volume of 40. The remaining 3 message units, maintaining their token order, are passed along the propagation path to the next field node as the remaining message volume. The field states are arranged according to the field node index and positional encoding is performed; the encoding results are concatenated to form the content generation control sequence.

[0072] A local attention mask is generated based on the arrangement range of field states in the content generation control sequence to block attention connections across field states. The remaining attention connections are decoded to form target language text fragments. These text fragments are encoded into text fragment vectors and propagated back along the path from the corresponding field node to the fact node, calling the inverse relation vector in reverse order. Reverse combination branches continuously perform circular convolutions, using the result of the previous circular convolution as input for the next. Once the circular convolution results reach the fact node, they are aggregated according to attention weights to form a feedback message.

[0073] Fact messages and return messages are matched according to fact nodes and propagation paths. Element-by-element differencing is performed on the matching results, and the absolute value of the difference is taken to form the message difference. The message difference is concatenated with the corresponding rule relation vector, and then linearly mapped and activated to form the updated rule relation threshold. When the threshold update causes a propagation branch switch, the propagation path is traced from the fork node along the switched propagation branch to the first merging node that intersects with the original propagation path. The original path segment is replaced with the alternative path segment between the two nodes, forming the replaced propagation path.

[0074] In the decoding result of the title field, the target language text fragment contains 40 tokens, among which the content formed by two consecutive tokens does not conform to the publishing platform rules. After the text fragment is returned, the corresponding fact message and the returned message form a message difference. The updated rule relation threshold remodulates the term mapping result, and the restricted message updates the field state along the replaced propagation path. The system only performs local decoding on the control sequence portion of the content corresponding to the two tokens, replacing the original content with the newly generated two tokens, while the remaining 38 tokens remain unchanged.

[0075] The field node is determined based on the field status of the updated text fragment, and the first sequence position of the corresponding content's control sequence is used as the fragment's starting position. Updated text fragments from different field nodes are arranged according to their field node indices, while those from the same field node are arranged according to their fragment starting positions. After arrangement, the updated text fragments corresponding to the same field node are concatenated, and then the content fields are combined according to their field node indices to form multilingual content for cross-border e-commerce.

[0076] Using the same product information, target language, sales market, publishing platform, and field restrictions, the final content generated by different methods was verified, and the results are as follows.

[0077] Table 1 Comparison of Multilingual Content Generation Effects in Cross-Border E-commerce

[0078] As shown in Table 1, this invention demonstrates improvements across all five indicators. The product factual error rate decreased from 6.8% in the manual template translation method to 3.7%, and the terminology inconsistency rate decreased from 8.1% to 4.3%. This indicates that this invention does not simply replace the language of product content, but rather improves the accuracy of the correspondence between the target language text and product facts. This invention parses product attributes into fact nodes, limits the target language expression through terminology relationships, and then transmits the generated text back to the fact nodes along the propagation path. Based on the differences between the factual message and the transmitted message, the corresponding text is corrected, reducing factual bias and terminology misuse.

[0079] The market rule violation rate of this invention is reduced from 5.6% using manual template translation methods to 3.1%, and also by 1.0 percentage point compared to knowledge graph-assisted generation methods. While knowledge graph-assisted generation methods can link product information and market rules, they lack feedback processing when rule deviations occur in the generated content. This invention modulates the term mapping results using rule relationship thresholds, and then updates the rule relationship thresholds based on message differences. When a propagation branch switches, only the path segment between the fork node and the first merging node is replaced, and local decoding is performed on the affected text fragments, thereby reducing the proportion of content that does not meet the requirements of the sales market and publishing platform.

[0080] The field length exceeding rate decreased from 7.4% to 3.9%, and the content field missing rate decreased from 6.3% to 3.6%. Traditional length control typically meets field limits by truncating text, which can easily lead to missing product content. This invention allocates message units to the receiving position according to the field capacity. Message units not received by the current field continue to be passed to the next field node, and field length control no longer relies on directly deleting content that exceeds the limit. The simultaneous reduction of both indicators indicates that the generated content can maintain the integrity of the content fields while meeting the field length limits.

[0081] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based multilingual content generation system for cross-border e-commerce, characterized in that, include: The product graph module is used to parse product information into fact nodes, align multilingual terms and market rules with fact nodes, instantiate attribute relationships, terminology relationships and rule relationships, and construct a cross-border product knowledge graph. The subgraph generation module is used to trim the cross-border commodity knowledge graph according to the target language, sales market and publishing platform and connect it to field nodes. It measures the field capacity based on the field length and fact importance, marks the propagation path from fact nodes to field nodes, and forms a commodity generation subgraph. The bidirectional combination module is used to input the product-generated subgraph into the improved CompGCN, bind positive and negative relation vectors between the bidirectional combination branches, and combine the fact node vector and attribute relation vector in the positive combination branch to form a fact message, which is then formed into a restricted message through term relation mapping and rule gating. The capacity gating module is used to control the limited message by field capacity, write the actual propagation volume into the field status, pass the remaining message volume to the next field node along the propagation path, and encode the field status into content to generate a control sequence. The text return module is used to generate control sequences from sequence decoding content, form target language text fragments, combine text fragment vectors and inverse relation vectors in reverse combination branches, return them along the propagation path, and aggregate them into return messages at fact nodes; The differential feedback module is used to perform message differential on fact messages and feedback messages, update the relationship threshold according to the message difference update rule and redirect the propagation path, generate a control sequence for the content corresponding to the message difference and re-execute local decoding to form an updated text fragment; The content assembly module is used to rearrange the sequence and concatenate the fields of the updated text fragments according to the field node index, forming multilingual content for cross-border e-commerce.

2. The knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 1, characterized in that, The product map module specifically includes: Perform field parsing on product information, bind attribute identifiers and attribute values ​​to product identifiers, and encode them as fact nodes; Compare the attribute identifiers corresponding to fact nodes and multilingual terms. If the attribute identifiers are the same, instantiate the term relation and encode the language identifiers into the term relation. Compare the attribute identifiers of fact nodes with the applicable attributes defined by market rules. If the comparison results are consistent, instantiate the rule relationship, encode the market identifier and platform identifier into the rule relationship, and instantiate the attribute relationship between fact nodes based on the attribute association record. Multi-relation graph encoding is performed on fact nodes and instantiated relations according to relation type and relation direction to construct a cross-border commodity knowledge graph.

3. The knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 1, characterized in that, The subgraph generation module specifically includes: Filter terminology relationships based on the target language, filter rule relationships based on the sales market and publishing platform, and retain the filtered terminology relationships, the rule relationships connected to the filtered terminology nodes, and the corresponding nodes. Configure the encoded field node according to the content field, compare the attribute identifier of the fact node with the attribute identifier that the field node is allowed to accept, and connect the rule node and field node corresponding to the fact node that matches the comparison. Convert the allowed text length of the field into the total capacity, read the fact importance corresponding to the fact node, allocate the total capacity among the fact nodes connected to the same field node according to the fact importance, and use the allocation result as the field capacity. The relationship type and direction are encoded according to the order of calls from fact nodes to field nodes, the propagation path is marked, and the nodes and relationships traversed by the propagation path are preserved to form a product generation subgraph.

4. The knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 1, characterized in that, The bidirectional combination module specifically includes: The improved CompGCN relation composition layer is split into forward composition branches and reverse composition branches, and the forward relation vector and the corresponding reverse relation vector are bound between the two composition branches; The forward combination branch performs cyclic correlation operations on the fact node vector and attribute relationship vector to form fact messages; Cyclic correlation operations are performed on factual messages and terminology relation vectors. The regular relation vectors are transformed into regular relation gates through linear projection and sigmoid activation. The cyclic correlation results are then modulated element by element through the regular relation gates to form restricted messages.

5. The knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 4, characterized in that, The binding of positive and inverse relation vectors in the bidirectional combination module includes the following steps: The forward relation vector is transformed into the inverse relation vector using the inverse transformation matrix; The forward combination branch performs a cyclic correlation operation on the input vector and the forward relation vector, while the reverse combination branch performs a cyclic convolution on the cyclic correlation result and the inverse relation vector to form the input recovery vector; Element-wise difference is performed on the input recovered vector and the input vector of the forward combined branch. The difference result is then propagated backward along the bidirectional combined branch to update the forward relation vector and the inverse transformation matrix.

6. The knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 1, characterized in that, The capacity gating module specifically includes: Restricted messages are mapped to a sequence of message units, with each message unit occupying one field capacity unit; Set the receiving position according to the sum of the field capacities corresponding to the same field node, extract message units from each message unit sequence according to the corresponding field capacity, and complete the allocation according to the attention weight of the extracted message units and the receiving position; The number of message units that have been allocated to the receiving position is counted to form the actual propagation volume. The propagation volume component in the field status is updated according to the actual propagation volume, and the corresponding receiving position component is updated according to the vector of the allocated message units. Extract message units that have not been assigned to the receiving position, count the number of unassigned units to form the remaining message volume, and pass the unassigned message units carrying the remaining message volume along the propagation path to the next field node; The updated field states are arranged according to the field node index, and position encoding is performed on the field states. The encoded results are concatenated to form the content generation control sequence.

7. The knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 1, characterized in that, The text return module specifically includes: A local attention mask is generated based on the range of field states in the content generation control sequence, blocking attention connections across field states, and decoding is performed based on the preserved attention connections to form a target language text fragment. Encode the target language text fragments into text fragment vectors, trace the propagation path back from the corresponding field node to the fact node, and call the bound inverse relation vector according to the reverse order of the relations in the propagation path; The reverse combination branch performs continuous circular convolutions in the order of calling the inverse relation vector, using the result of the previous circular convolution as the input of the next circular convolution, until the result of the circular convolution reaches the fact node; Calculate the attention weights between the fact node vector and the circular convolution results that reach the corresponding fact node along the same propagation path, and perform a weighted summation on the circular convolution results according to the attention weights to form the feedback message.

8. The knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 1, characterized in that, The differential feedback module specifically includes: Perform element-wise difference analysis on the fact message and the feedback message, and take the absolute value of the difference to form the message difference; The message difference is concatenated with the rule relation vector, and the updated rule relation gate value is output after linear mapping and S-shaped activation. Based on the updated rule relationship threshold, the propagation branch of the corresponding fork node in the original propagation path is switched, and the original propagation path is continued along the switched propagation branch to form the replaced propagation path; The updated rule relation threshold modulates the term mapping result to form an updated restricted message. The updated restricted message is then transmitted along the replaced propagation path, and the corresponding field status is updated based on the transmission result. The updated field state is encoded, and the corresponding part in the control sequence is generated by replacing the content with the encoded result. The replaced part is then partially decoded, and the local content in the corresponding target language text fragment is replaced with the decoded result to form the updated text fragment.

9. A knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 8, characterized in that, The replacement of the propagation path in the differential feedback module includes the following steps: Compare the propagation branches before and after the rule relation threshold update, and identify the rule relations in which the propagation branch switch occurs as abnormal relations; Tracing from the bifurcation node corresponding to the abnormal relationship along the switched propagation branch to the first merging node intersecting with the original propagation path, the propagation branch between the bifurcation node and the first merging node is extracted to form an alternative path segment; Replace the original propagation path segment between the fork node and the first merging node with an alternative path segment.

10. A knowledge graph-based multilingual content generation system for cross-border e-commerce according to claim 1, characterized in that, The content assembly module specifically includes: The control sequence part is generated by locating and updating the content corresponding to the text fragment, determining the field node based on the state of the field, and taking the first sequence position as the fragment start position. Updated text fragments are arranged according to the field node index, and updated text fragments belonging to the same field node are arranged according to the fragment's starting position; Updated text fragments with different field node indexes are sorted in ascending order of field node index; updated text fragments with the same field node index are sorted in ascending order of fragment start position. The updated text fragments belonging to the same field node after being arranged are concatenated continuously, and the concatenation results are combined according to the field node index to form multilingual content for cross-border e-commerce.

Images