Intelligent matching system for international supply and demand of commodities based on multilingual AI model

Abstract

The application provides a commodity international supply and demand intelligent matching system based on a multilingual AI model, the system comprising a knowledge graph construction module for collecting multilingual commodity data and constructing a cross-language commodity knowledge graph comprising attribute entities, scene nodes and various semantic relationships; an encoding model training module for training a semantic encoding model based on the knowledge graph through knowledge-enhanced representation learning to narrow the vector representation of the same attribute entity in different languages; a demand analysis module for attribute analysis, completion and normalization of user natural language queries to generate a structured attribute demand set; and a matching and explanation output module for matching and scoring candidate commodities through semantic encoding retrieval and knowledge graph reasoning, and generating interpretable matching information comprising attribute relationship paths. The application improves the accuracy and transparency of cross-language commodity matching.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-driven cross-border e-commerce information processing technology, specifically to an intelligent matching system for international supply and demand of goods based on a multilingual AI model. Background Technology

[0002] With the rapid development of global e-commerce, intelligent cross-language product matching has become a core technological challenge for cross-border e-commerce. Existing machine translation-based matching methods require translating multilingual queries into a single language before retrieval. The translation quality directly affects the matching results, and professional attribute terms such as product specifications and materials often lack precise translation correspondences between different languages, leading to the loss of attribute semantic information. Although multilingual pre-trained language models introduced in recent years can semantically represent different languages ​​in a shared vector space, they rely on surface co-occurrence information in the text and struggle to capture structured semantic relationships such as subordination, equivalence, and hierarchy among attributes within the product domain, making it difficult to simultaneously achieve high recall and precision in matching.

[0003] Meanwhile, existing systems suffer from severe deficiencies in the interpretability of matching decisions. The matching process is largely black-box; users cannot know which attribute dimensions of recommended products meet their needs, nor can they understand matching biases caused by differences in language expression. Systems generally lack the ability to quantify and standardize vague user needs, nor do they utilize business scenario information to complete the requirements. Matching results are limited to literal semantics, making it difficult to uncover implicit needs. Furthermore, the lack of transparent explanations for cross-language attribute value conversions further weakens user trust. Therefore, there is an urgent need for an intelligent system that deeply integrates structured product knowledge, possesses cross-language semantic alignment capabilities, and provides interpretable matching decisions to improve the accuracy, intelligence, and user trust in cross-border product matching. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes an intelligent matching system for international supply and demand of goods based on a multilingual AI model.

[0005] Therefore, the technical solution adopted by the present invention is as follows: A multilingual AI model-based intelligent matching system for international supply and demand of goods includes: Knowledge graph construction module: Collects multilingual product data and preprocesses it to obtain a multilingual product corpus; constructs a cross-language product knowledge graph based on the multilingual product corpus; Encoding model training module: Based on the cross-language product knowledge graph, a preset semantic encoding model is trained through knowledge-enhanced representation learning; Requirement parsing module: Obtain the natural language query from the user, and perform attribute parsing and attribute completion on the natural language based on the cross-language product knowledge graph to obtain a structured set of attribute requirements; Matching and Explanation Output Module: Based on the attribute requirement set, retrieve candidate products using the semantic encoding model; based on the cross-language product knowledge graph, connect the attribute relationship paths between the attribute requirement set and the candidate products, perform matching scoring on the candidate products, and generate interpretable matching information containing the attribute relationship paths.

[0006] Furthermore, the nodes of the cross-language product knowledge graph include attribute entities and scene nodes; the attribute entities include product entities, attribute name entities, and attribute value entities. The edges of the cross-language product knowledge graph include subordinate relationship edges between product entities and attribute name entities, descriptive relationship edges between attribute name entities and attribute value entities, equivalence relationship edges between attribute entities, hierarchical relationship edges between attribute entities, and scene and attribute association edges between scene nodes and attribute name entities; the equivalence relationship edges include cross-language equivalence edges and intra-language equivalence edges.

[0007] Furthermore, the construction process of the cross-language product knowledge graph is as follows: The first step is to construct a joint extraction model of attribute entities and relations. The multilingual product corpus is output to the joint extraction model, and the output is an initial triple containing product entities, attribute name entities, and attribute name value entities. The second step is to combine the cross-language equivalent entity pairs collected from existing multilingual knowledge resources into an equivalent entity pair set. The third step is to construct an initial knowledge graph based on all initial triples and sets of equivalent entity pairs. The initial triplet provides the product entity, attribute name entity, and attribute value entity nodes in the initial knowledge graph, as well as the subordinate relationship edges between the product entity and the attribute name entity, and the descriptive relationship edges between the attribute name entity and the attribute value entity; the equivalent entity pair set provides cross-language equivalent edges in the initial knowledge graph; The fourth step involves mining the hierarchical relationships between attribute entities through lexical rules, attribute value type analysis, and remote supervision from an external knowledge base, and establishing hierarchical relationship edges. The fifth step is to construct a binary classifier based on graph features to filter out incorrect hierarchical relationships; The sixth step involves domain experts sampling and reviewing the hierarchical relationships selected by the classifier. Hierarchical relationships that have been manually reviewed and confirmed are then formally inserted into the initial knowledge graph, ultimately forming a cross-language product knowledge graph. Step 7: Introduce scene nodes into the cross-language product knowledge graph. Each scene node is connected to several attribute name entities through scene and attribute association edges. The eighth step involves supplementing metadata into the numerical attribute name entities in the cross-language product knowledge graph. The metadata includes commonly used units of measurement and value ranges. Simultaneously, domain experts identify the corresponding quantitative constraints for common fuzzy descriptive terms in product categories and establish linguistic equivalent edges for each pair of fuzzy descriptive terms and quantitative constraints in the cross-language product knowledge graph.

[0008] Furthermore, the semantic encoding model is a dual-tower structure, containing a query encoding tower and a product encoding tower, both of which share the underlying text encoder; The semantic coding model introduces a knowledge injection mechanism into the input layer of the text encoder; The training tasks for the semantic coding model include cross-linguistic attribute alignment, fine-grained attribute differentiation, and translation prediction assistance.

[0009] Furthermore, the training samples of the semantic encoding model include positive sample pairs and negative sample pairs, and are constructed based on the cross-language product knowledge graph and the multilingual product corpus pool; The positive sample pairs consist of text pairs with the same semantics but different languages, including different language mentions of the same attribute entity connected from the cross-language product knowledge graph, and two parallel sentences in two languages ​​describing the same product extracted from the multilingual product corpus. Negative sample pairs include general negative samples and difficult negative samples. The difficult negative samples are cross-language text pairs that are literally similar but semantically different, and are obtained based on edge mining in the cross-language product knowledge graph.

[0010] Furthermore, the operation for obtaining the attribute requirement set is as follows: The system extracts demand items from the natural language of user queries using an adjusted language model and links them to attribute entities in the cross-language product knowledge graph. Implicit demands are completed based on the scene nodes and hierarchical relationships. Units and fuzzy descriptive words are standardized using metadata pre-stored in the attribute entities to form a structured set of attribute demands. Each demand belongs to one of three types of hard constraints: product type constraints, discrete attribute value constraints, or numerical attribute constraints.

[0011] Furthermore, the matching and interpretation output module is specifically used to: transform the attribute requirement set into a standard query description, and input it into the query encoding tower of the multilingual product semantic encoding model to obtain a query vector; The product coding tower is used to pre-encode the products in the merchant's product library into product vectors, and a vector index is built. Candidate products are retrieved based on the cosine similarity between the query vector and the product vectors in the index. The candidate products are validated using the hard constraints, and those that satisfy all hard constraints are retained. Recommended products are then selected by sorting the products by cosine similarity. Explainable matching information is obtained by collecting attribute entities involved in the verification and generating explanatory sentences, and additional explanations are added when there are differences in expression, scenario reasoning or unit conversion.

[0012] Compared with the prior art, the advantages of the present invention are as follows: 1. This invention effectively aligns the vector representations of the same attributes in different languages ​​through cross-language product knowledge graphs and knowledge-enhanced representation learning, avoiding information loss during the translation process and accurately capturing the subordinate, equivalence, and hierarchical relationships between attributes.

[0013] 2. This invention combines knowledge graph reasoning to generate attribute relationship paths and adds explanations to the matching criteria such as language differences, scenario reasoning, and unit conversion, so that users are clearly aware of how the recommended products meet their needs in specific dimensions, thereby enhancing trust.

[0014] 3. This invention extracts attributes from natural language queries, completes implicit requirements driven by scenarios, and quantitatively transforms fuzzy descriptive words, turning vague expressions into structured hard constraints, uncovering key conditions not explicitly stated by users, and making the matching results more in line with actual business intentions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the system execution flow of the present invention; Figure 2 This is a schematic diagram of the cross-language product knowledge graph acquisition process of the present invention; Figure 3 This is a schematic diagram of the cross-language product knowledge graph of the present invention. Detailed Implementation

[0017] To achieve the above objectives, this invention provides an intelligent international supply and demand matching system for commodities based on a multilingual AI model. Please refer to [link / reference]. Figure 1 ,include: Knowledge graph construction module: Collects multilingual product data and preprocesses it to obtain a multilingual product corpus; constructs a cross-language product knowledge graph based on the multilingual product corpus; the cross-language product knowledge graph contains attribute entities described in different languages, as well as cross-language equivalence relations and hierarchical relations between the attribute entities; Attribute entities include three types: product entities (e.g., dresses), attribute name entities (e.g., colors), and attribute value entities (e.g., red).

[0018] Product data, including text descriptions, specifications, and user reviews, is collected from multiple e-commerce platforms and standard product information databases, covering multiple languages. While collecting product titles, details, and specifications, the category path text for each product is also recorded.

[0019] The specific preprocessing process is as follows: The collected texts are preprocessed, including language identification, cleaning, and word segmentation, to obtain a multilingual product corpus consisting of word sequence. For segments containing multiple languages, the following regularization is applied: First, the segment is segmented into sub-words using a multilingual BPE word segmenter. Then, a pre-trained character-level transliteration model is used to convert sub-words identified as non-main language into transliterated strings based on the pronunciation of the main language. The corresponding original glyphs are retained after the transliterated strings as supplementary information.

[0020] The construction process of a cross-language product knowledge graph is as follows; please refer to [link / reference]. Figure 2 : The first step is to construct a joint extraction model of attribute entities and relations. The multilingual product corpus is output to the joint extraction model, and the output is an initial triple containing product entities, attribute name entities, and attribute name value entities; for example, <dress, color, red>, where "dress" is the product entity, "color" is the attribute name entity, and "red" is the attribute name entity.

[0021] The training of the joint extraction model employs a strategy that combines remote supervision and human verification.

[0022] The joint extraction model uses the pre-trained multilingual model XLMRoBERTa as the underlying text encoder; on top of the encoder, two task sub-networks are designed in parallel, namely the named attribute entity recognition sub-network and the relation classification sub-network. The encoder's role is to transform each input lexical unit into a semantic vector representation that contains contextual meaning; The named attribute entity recognition subnetwork is responsible for detecting word fragments in the text that belong to product entities, attribute name entities, and attribute name entities. It adopts the BIO annotation mode and connects a conditional random field (CRF) at the top layer. The relationship classification sub-network is responsible for determining whether there is a specific relationship between the attribute entities output by the named attribute entity recognition sub-network, which are respectively the subordinate relationship between the product entity and the attribute name entity, and the description relationship between the attribute name entity and the attribute value entity.

[0023] Finally, the relationship feature vector is fed into a two-layer fully connected network, and it is judged through binary classification whether the attribute entity pair has the subordinate relationship between the product and the attribute name or the description relationship between the attribute name and the attribute value.

[0024] In the second step, cross-lingual equivalent entity pairs with high confidence are collected from existing multilingual knowledge resources and combined into an equivalent entity pair set; The existing multilingual knowledge resources include existing cross-lingual knowledge graphs and the multilingual category attribute mapping tables disclosed on international e-commerce platforms, from which verified equivalent pairs such as (Chinese: 颜色, English: color), (Chinese: 制造商, English: manufacturer) can be directly extracted.

[0025] In the third step, an initial knowledge graph is constructed based on all initial triples and the equivalent entity pair set; the initial triples provide the product entity, attribute name entity and attribute value entity nodes in the graph, as well as the subordinate relationship edges between the product entity and the attribute name entity, and the description relationship edges between the attribute name entity and the attribute value entity; based on the preset graph attention network and the semantic vector representation corresponding to each attribute entity, the entity vector of each attribute entity in the initial knowledge graph is generated; and the cosine similarity between the two entity vectors in the entity pair is calculated. When the similarity is higher than a preset similarity threshold (for example, 0.9), a cross-lingual equivalent edge is added between the entity pair; The number of layers of the graph attention network is usually 1, 2, or 3, and the value is 2 in this embodiment; The training objectives of the graph attention network include two aspects: one is to reduce the distance of equivalent entity pairs in the vector space; the other is to make the equivalent entity pairs of different languages in the entire network language-indistinguishable.

[0026] For the first objective, the alignment loss function is defined as follows:

[0027] Wherein, is the set of equivalent entity pairs; and are respectively the entity and the entity vector obtained after being updated by the graph attention network; represents the L2 norm of the vector. By minimizing this loss, the vector distance of the seed pairs can be directly compressed.

[0028] For the second objective, a language discriminator (single-layer linear classifier) ​​is connected after the output layer of the graph attention network. The discriminator predicts the original language of an entity based on its entity vector. During training, a gradient reversal layer is used to update the parameters of the graph attention network in the direction where the discriminator has difficulty distinguishing languages, thereby eliminating language differences. The total training loss of the graph attention network is the weighted sum of the two.

[0029] The fourth step involves mining the hierarchical relationships between attribute entities through lexical rules, attribute value type analysis, and remote supervision from external knowledge bases (such as WordNet). In this embodiment, the hierarchical relationship is specifically defined as the hierarchical relationship. The fifth step is to construct a binary classifier based on graph features to filter out incorrect hierarchical relationships; The sixth step involves domain experts sampling and reviewing the hierarchical relationships selected by the classifiers. Hierarchical relationships that pass manual review are then formally inserted into the initial knowledge graph, ultimately forming a cross-language product knowledge graph. The seventh step is to introduce scenario nodes into the cross-language product knowledge graph. Each scenario node is connected to several attribute name entities through scenario and attribute association edges. The scenario and attribute association edges are manually defined by domain experts based on high-frequency business needs. The scenario and attribute association edges and their weights are pre-organized into scenario and attribute mapping templates. The scenario nodes represent typical business scenarios in which the product is used (such as beach vacation, outdoor sports, etc.). Step 8: Supplement metadata to the numerical attribute name entities in the cross-language product knowledge graph. The metadata includes commonly used units of measurement and value ranges.

[0030] Specifically, for each attribute name entity, the system iterates through all corresponding attribute value entities, automatically calculating its commonly used units of measurement (the top three most frequent units) and value range (i.e., typical value intervals, represented by minimum and maximum values). The statistical results are then used to construct a unit mapping table (containing the correspondence between common unit names and standard codes), which is stored as a node attribute on the attribute name entity. Furthermore, domain experts have identified the corresponding quantitative constraints for common fuzzy descriptive terms (such as "large capacity") in high-frequency product categories, and established equivalent edges in the same language for each pair of fuzzy descriptive terms and quantitative constraints in the cross-language product knowledge graph. For example, "large capacity" is equivalent to a volume ≥ 5 liters.

[0031] The final constructed cross-language product knowledge graph consists of nodes including product entities, attribute name entities, attribute value entities, and scene nodes; edges include: subordinate relationship edges between product entities and attribute name entities, descriptive relationship edges between attribute name entities and attribute value entities, equivalence relationship edges between attribute entities (including cross-language equivalence edges and intra-language equivalence edges), hierarchical relationship edges between attribute entities, and scene-attribute association edges between scene nodes and attribute name entities. Please refer to [link / reference needed]. Figure 3 .

[0032] Encoding model training module: based on the cross-lingual commodity knowledge graph, train a preset semantic encoding model through knowledge-enhanced representation learning, so as to reduce the representation difference of texts in different languages in the vector space when describing the same attribute entity; The semantic encoding model adopts a two-tower structure, including a query encoding tower and a commodity encoding tower, which share the underlying text encoder. The text encoder selects a pre-trained multilingual model (such as XLMRoBERTa), and its initial parameters are loaded from public pre-trained weights; The training samples of the semantic encoding model include positive sample pairs and negative sample pairs, which are constructed based on the cross-lingual commodity knowledge graph and the multilingual commodity corpus pool; Positive sample pairs consist of text pairs with the same semantics but different languages, and there are two specific sources: The first is to extract mentions in different languages connected by the same attribute entity from the cross-lingual commodity knowledge graph. For example, the Chinese mention "yanse" (color) and the English mention "color" form a positive sample pair; the mention refers to the specific word or phrase actually used by the attribute entity in the multilingual commodity data (for example, the mention of the attribute name entity 'Color' in English is 'color', and the mention in Chinese is 'yanse' (color)).

[0033] The second is to extract parallel sentences in two languages describing the same commodity from the multilingual commodity corpus pool by utilizing commodity identifiers and synonymous commodities aligned with the existing knowledge graph as positive sample pairs.

[0034] Negative sample pairs are divided into general negative samples and hard negative samples; General negative samples are constructed by randomly sampling mentions in different languages of two different attribute entities, for example, Chinese "yanse" (color) and English "material"; Hard negative samples specifically mine those cross-lingual text pairs that are literally similar but semantically different. For example, the English word "boot" corresponding to Chinese "xuezi" (boot) is a positive pair, but the Chinese word "qidong" (start / boot) should also form a hard negative pair with "boot"; or easily confused attribute pairs in the same language, such as "gonglü" (power) and "gonghao" (power consumption); The mining of hard negative samples is assisted by the semantic relationships in the cross-lingual commodity knowledge graph for judgment: if there is no equivalence relationship edge (including cross-lingual equivalence edge and same-language equivalence edge) between two attribute entities in the cross-lingual commodity knowledge graph, and the shortest path length between the two entities calculated according to other relationship edges in the graph (such as subordinate, description, hypernym-hyponym, scenario-attribute association, etc.) exceeds the preset hop count threshold, but the string similarity of their mentions is higher than the preset threshold, then the mentions corresponding to this pair of entities are used as hard negative samples.

[0035] To integrate structured knowledge from cross-language product knowledge graphs into text encoding, a knowledge injection mechanism is introduced at the input layer of the text encoder. The specific operation is as follows: For each lexical unit, if the lexical unit is identified as an attribute entity, the entity vector corresponding to the attribute entity is retrieved from the cross-language product knowledge graph. After adjusting the dimension of the entity vector through a linear mapping layer, it is added element-wise to the semantic vector representation output by XLMRoBERTa for the lexical unit to obtain the knowledge-enhanced semantic vector representation. Lexical units that are not identified as attribute entities directly use the semantic vector representation of the lexical unit.

[0036] The training tasks for semantic coding models include cross-linguistic attribute alignment, fine-grained attribute region tasks, and translation prediction assistance tasks.

[0037] The cross-language attribute alignment task aims to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. Its loss function adopts the InfoNCE contrastive loss form.

[0038] Fine-grained attribute differentiation task: Extract entity pairs belonging to the same product entity but with different attribute names from a cross-language product knowledge graph. Define the loss by maximizing the cosine distance between their corresponding entity vectors. This loss is directly applied to the entity representation layer inside the text encoder.

[0039] Translation prediction auxiliary task: A lightweight decoder is attached to the text encoder. This decoder receives the vector encoded from the source language text and predicts the corresponding sequence of attribute words in the target language. The translation prediction uses standard multi-class cross-entropy loss. This auxiliary task helps to further align cross-linguistic attribute semantics.

[0040] Once the semantic coding model converges, the query coding tower and the product coding tower can work independently, encoding product queries or product descriptions in any language into a unified semantic vector, and significantly reducing the distance between different language texts of entities with the same attributes in the vector space.

[0041] Requirement parsing module: Obtain the natural language query from the user, and perform attribute parsing and attribute completion on the natural language based on the cross-language product knowledge graph to obtain a structured set of attribute requirements; After obtaining the natural language of the user's query, the process of obtaining the attribute requirement set is as follows: A fine-tuned language model is used to extract requirement items from user queries and link them to attribute entities in a cross-language product knowledge graph. Subsequently, implicit requirements are completed using scene nodes and hierarchical relationships (e.g., inferring 'UPF value > 30' from 'beach vacation'), and units and fuzzy descriptive terms are standardized using metadata. This results in a structured set of attribute requirements, where each item falls under one of the following three hard constraints: If this item specifies the product category required by the user (e.g., 'dress'), it is called a product type constraint; If this item specifies that an attribute must take a specific discrete value (e.g., the color is 'red'), it is called a discrete attribute value constraint; If this item specifies that the value of an attribute must meet numerical comparison conditions (e.g., price less than $100, volume greater than 5 liters), it is called a numerical attribute constraint. The comparison conditions and units within the numerical attribute constraints are derived from quantity words and unit words in the user's original input, or from fuzzy descriptive words after being processed by cross-language product knowledge graph equivalence relations and unit normalization.

[0042] Matching and Explanation Output Module: Based on the attribute requirement set, retrieve candidate products using the semantic encoding model; based on the attribute relationship path connecting the attribute requirement set and candidate products in the cross-language product knowledge graph, match and score the candidate products, and generate interpretable matching information containing the attribute relationship path.

[0043] Since the query encoding tower of the semantic encoding model takes a text sequence as input, and the attribute requirement set is a set of structured entity identifiers and constraints, it is necessary to first transform the attribute requirement set into a compact and semantically complete text as a standard query description. The specific transformation process is as follows: Each item in the attribute requirement set is traversed, and each is transformed into a text phrase in the user's query language using a cross-language product knowledge graph. These phrases are then concatenated into a complete standard query description. In the cross-language product knowledge graph, each attribute entity is associated with multiple mentions in different languages. During the transformation, for each entity to be referenced, the text with the highest frequency of occurrence in the product corpus is selected from all its mentions in that language. The specific transformation rules are as follows: If the item is a product type constraint, then obtain the usage text of the corresponding product entity in the user's language and generate the phrase "Product Type: Usage Text"; If the item is a discrete attribute constraint, then obtain the usage text of the attribute name entity and the attribute value entity in the user language respectively, and generate the phrase "attribute name: attribute value"; If the item is a numeric attribute constraint (with comparison operators, numbers, and units), then the usage text of the attribute name entity is retrieved, and phrases are generated in the format "attribute name comparison operator number unit". All phrases are concatenated sequentially with spaces to obtain the standard query description.

[0044] The query encoding tower of the semantic encoding model of the standard query description input value is used to obtain a query vector; During the preparation phase, all products in the merchant's product library have been encoded into product vectors with the same dimensions as the query vector by using the product coding tower in the semantic coding model. A query index is then built for all product vectors using a vector index based on approximate nearest neighbor search (such as the HNSW index). The cosine similarity between the query vector and each product vector in the query index is calculated based on the query index, and the top M products are selected as the candidate product set. The cosine similarity is used as the matching score of the candidate products. M is used to balance the recall coverage and subsequent processing efficiency. In this embodiment, M is 200.

[0045] The candidate product set is validated using the hard constraints in the attribute requirement set, and candidate products that satisfy all hard constraints are retained.

[0046] In the numerical attribute constraints, some come directly from the numerical conditions in the user query, while others come from the quantitative constraints transformed by the requirement parsing module through fuzzy adjective equivalence relations. These hard constraints are used to verify each item in the candidate product set, eliminating those that do not meet the requirements. The verification process is as follows: For each candidate product in the candidate product set Locate the corresponding product entity in a cross-language product knowledge graph. Then, check each constraint in the attribute requirement set one by one: If the constraint is a product type, determine... Check whether it is the same as the product entity given in the requirement, or whether the two are directly connected by an equivalence relation edge (including cross-language equivalence edges); if true, then this item is considered satisfied. If the constraint is a discrete attribute value, first check... Is it connected to an entity with an attribute name through a dependency edge? ,and The entity must have the same attribute name as the entity in the requirement or be connected by an equivalence relation edge; then check... Does it connect to an attribute value entity via a description relation edge or an equivalence relation edge? ,and The item is satisfied when it is the same as the attribute value entity in the requirement or is connected by an equivalence relation edge; all of the above conditions must be met. If it is a numerical attribute constraint, first refer to the method of locating the attribute name entity in the discrete attribute value constraint to match the corresponding attribute name entity and attribute value entity. Then, from the mention text of the attribute value entity and the commonly used units of measurement and value range metadata stored in the attribute name entity node, parse out the normalized value and standard unit of the candidate product, and compare it with the value, standard unit and comparison operator in the requirement constraint; if the comparison result is true, then the condition is satisfied. If any hard constraint is not met, the candidate product is removed from the candidate product set, resulting in a simplified candidate set. The top D products with the highest cosine similarity are then selected as recommended products. If the simplified candidate set is empty, then the D products with the highest cosine similarity in the candidate product set are selected as recommended products. Cosine similarity is sorted from high to low, and D is usually taken as 10-20; The process of generating interpretable matching information is as follows: Recommended products for each output Collect the attribute name entity and attribute value entity in the hard constraint validation; By obtaining the usage text of these attribute entities in the user's query language through a cross-language product knowledge graph, and generating explanatory sentences according to the basic template "the [attribute name] of this product is [attribute value]", parentheses are added to explain when the following differences occur: If the actual matched attribute value entity differs from the expression in the user's original input (but the two are related by an equivalence relation edge), then add "(consistent with the meaning of the [user's original input] you require)"; If the requirement originates from scenario reasoning completion (e.g., when a user mentions "beach vacation", the system completes "UPF>30"), then add "(satisfies the requirement of [scenario name]"; If the numerical values ​​or units presented after unit standardization differ from the user's original input, then brief equivalent information is added. All explanatory phrases for the same product are then concatenated into a coherent text in a logical order of attributes, serving as the explanatory matching information for that product. This information is presented to the user along with the product identifier and title. This information allows users to confirm, item by item, how their needs are met by specific products, improving the transparency and credibility of cross-language product matching.

[0047] The proposed intelligent international supply and demand matching system for commodities based on a multilingual AI model constructs a cross-language commodity knowledge graph and trains a knowledge-enhanced semantic encoding model to align multilingual attribute semantics to a unified vector space, avoiding the loss of translation information and accurately capturing structured relationships. Simultaneously, it leverages knowledge graph reasoning to complete implicit needs, transform fuzzy descriptive terms, generate structured constraints, and output interpretable matching paths. This addresses the problems of insufficient semantic accuracy, opaque decision-making, and difficulty in parsing fuzzy needs, thereby improving matching accuracy, intelligence, and user trust.

[0048] In summary, this invention integrates cross-language product knowledge graphs with knowledge-enhanced semantic encoding to align multilingual attribute semantics to a unified vector space, avoiding the loss of translation information and capturing structured relationships. By combining graph reasoning to resolve fuzzy requirements, complete implicit conditions, and output interpretable matching results, it improves the accuracy, intelligence, and transparency of cross-language matching, overcoming the shortcomings of existing systems in semantic accuracy, interpretability, and requirement understanding.

[0049] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multilingual AI model-based intelligent matching system for international supply and demand of goods, characterized in that, include: Knowledge graph construction module: Collects multilingual product data and preprocesses it to obtain a multilingual product corpus pool; Construct a cross-language product knowledge graph based on the aforementioned multilingual product corpus; Encoding model training module: Based on the cross-language product knowledge graph, a preset semantic encoding model is trained through knowledge-enhanced representation learning; Requirement parsing module: Obtain the natural language query from the user, and perform attribute parsing and attribute completion on the natural language based on the cross-language product knowledge graph to obtain a structured set of attribute requirements; Matching and Explanation Output Module: Based on the attribute requirement set, retrieve candidate products using the semantic encoding model; based on the cross-language product knowledge graph, connect the attribute relationship paths between the attribute requirement set and the candidate products, perform matching scoring on the candidate products, and generate interpretable matching information containing the attribute relationship paths.

2. The system according to claim 1, characterized in that, The nodes of the cross-language product knowledge graph include attribute entities and scene nodes; the attribute entities include product entities, attribute name entities, and attribute value entities. The edges of the cross-language product knowledge graph include subordinate relationship edges between product entities and attribute name entities, descriptive relationship edges between attribute name entities and attribute value entities, equivalence relationship edges between attribute entities, hierarchical relationship edges between attribute entities, and scene and attribute association edges between scene nodes and attribute name entities; the equivalence relationship edges include cross-language equivalence edges and intra-language equivalence edges.

3. The system according to claim 2, characterized in that, The construction process of the cross-language product knowledge graph is as follows: The first step is to construct a joint extraction model of attribute entities and relations. The multilingual product corpus is output to the joint extraction model, and the output is an initial triple containing product entities, attribute name entities, and attribute name value entities. The second step is to combine the cross-language equivalent entity pairs collected from existing multilingual knowledge resources into an equivalent entity pair set. The third step is to construct an initial knowledge graph based on all initial triples and sets of equivalent entity pairs. The initial triplet provides the product entity, attribute name entity, and attribute value entity nodes in the initial knowledge graph, as well as the subordinate relationship edges between the product entity and the attribute name entity, and the descriptive relationship edges between the attribute name entity and the attribute value entity; the equivalent entity pair set provides cross-language equivalent edges in the initial knowledge graph; The fourth step involves mining the hierarchical relationships between attribute entities through lexical rules, attribute value type analysis, and remote supervision from an external knowledge base, and establishing hierarchical relationship edges. The fifth step is to construct a binary classifier based on graph features to filter out incorrect hierarchical relationships; The sixth step involves domain experts sampling and reviewing the hierarchical relationships selected by the classifier. Hierarchical relationships that have been manually reviewed and confirmed are then formally inserted into the initial knowledge graph, ultimately forming a cross-language product knowledge graph. Step 7: Introduce scene nodes into the cross-language product knowledge graph. Each scene node is connected to several attribute name entities through scene and attribute association edges. The eighth step involves supplementing metadata into the numerical attribute name entities in the cross-language product knowledge graph. The metadata includes commonly used units of measurement and value ranges. Simultaneously, domain experts identify the corresponding quantitative constraints for common fuzzy descriptive terms in product categories and establish linguistic equivalent edges for each pair of fuzzy descriptive terms and quantitative constraints in the cross-language product knowledge graph.

4. The system according to claim 3, characterized in that, The semantic coding model is a dual-tower structure, containing a query coding tower and a product coding tower, both of which share the underlying text encoder. The semantic coding model introduces a knowledge injection mechanism into the input layer of the text encoder; The training tasks for the semantic coding model include cross-linguistic attribute alignment, fine-grained attribute differentiation, and translation prediction assistance.

5. The system according to claim 4, characterized in that, The training samples of the semantic coding model include positive sample pairs and negative sample pairs, and are constructed based on the cross-language product knowledge graph and the multilingual product corpus pool. The positive sample pairs consist of text pairs with the same semantics but different languages, including different language mentions of the same attribute entity connected from the cross-language product knowledge graph, and two parallel sentences in two languages ​​describing the same product extracted from the multilingual product corpus. Negative sample pairs include general negative samples and difficult negative samples. The difficult negative samples are cross-language text pairs that are literally similar but semantically different, and are obtained based on edge mining in the cross-language product knowledge graph.

6. The system according to claim 5, characterized in that, The operation for obtaining the attribute requirement set is as follows: The system extracts demand items from the natural language of user queries using an adjusted language model and links them to attribute entities in the cross-language product knowledge graph. Implicit demands are completed based on the scene nodes and hierarchical relationships. Units and fuzzy descriptive words are standardized using metadata pre-stored in the attribute entities to form a structured set of attribute demands. Each demand belongs to one of three types of hard constraints: product type constraints, discrete attribute value constraints, or numerical attribute constraints.

7. The system according to claim 6, characterized in that, The matching and interpretation output module is specifically used to: convert the attribute requirement set into a standard query description, and input it into the query encoding tower of the multilingual product semantic encoding model to obtain a query vector; The product coding tower is used to pre-encode the products in the merchant's product library into product vectors, and a vector index is built. Candidate products are retrieved based on the cosine similarity between the query vector and the product vectors in the index. The candidate products are validated using the hard constraints, and those that satisfy all hard constraints are retained. Recommended products are then selected by sorting the products by cosine similarity. Explainable matching information is obtained by collecting attribute entities involved in the verification and generating explanatory sentences, and additional explanations are added when there are differences in expression, scenario reasoning or unit conversion.

Images