AI-based e-commerce intelligent search semantic understanding extension method and system

Abstract

The application discloses an AI-based e-commerce intelligent search semantic understanding extension method and system, and relates to the technical field of e-commerce platforms.The method comprises the following steps: obtaining a user input search keyword text, standardizing and preprocessing the search keyword text to obtain an effective word segmentation unit set, and constructing a user search basic data set; in the second step, based on a pre-trained e-commerce field natural language processing deep learning model, the semantic vectors corresponding to each effective word segmentation unit in the user search basic data set are extracted.The semantic vector extraction and integration structure based on the pre-trained e-commerce field natural language processing deep learning model can completely capture the deep semantics and potential search intention of the user input search keyword, solves the problem that the prior art can only match the literal meaning of the keyword and cannot identify the potential demand of the user, and realizes the AI-based accurate analysis effect of the search keyword semantics.

Description

Technical Field

[0001] This invention relates to the field of e-commerce platform technology, specifically to an AI-based method and system for extending the semantic understanding of e-commerce intelligent search. Background Technology

[0002] E-commerce is a business model that relies on internet technology to enable online display, trading, and distribution of goods. It is currently the mainstream channel for commodity consumption and distribution in China. Product search functionality is a core foundational function of e-commerce platforms. It is the primary means for users to quickly locate target products within the platform's vast product database, and its performance directly impacts the user experience and the platform's transaction conversion rate.

[0003] Most existing e-commerce platforms use keyword literal matching for their product search functions. This approach fails to deeply analyze the potential search intent of users' input and cannot provide users with relevant search content that closely matches their search needs. Consequently, the search results do not match the users' actual needs, resulting in low search efficiency and a poor user experience. Therefore, this paper proposes an AI-based method and system for semantic understanding extension in e-commerce intelligent search to address these issues. Summary of the Invention

[0004] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an AI-based method and system for extending semantic understanding in e-commerce intelligent search, which solves the problems mentioned in the background section.

[0005] Technical solution To achieve the above objectives, the present invention provides the following technical solution: an AI-based method for semantic understanding extension in e-commerce intelligent search, comprising the following steps: Step 1: Obtain the search keyword text input by the user, perform standardized preprocessing on the search keyword text to obtain a set of effective word segmentation units, and construct the basic dataset of user search. Step 2: Based on the pre-trained deep learning model for natural language processing in the e-commerce domain, extract the semantic vectors corresponding to each effective word segmentation unit in the user search basic dataset, calculate the semantic association evaluation value of the entire word segmentation, and integrate all semantic vectors according to the order of the effective word segmentation units in the search keyword text to generate the overall semantic evaluation value corresponding to the search keyword. Step 3: Based on the pre-built e-commerce domain knowledge graph, match entity nodes that are consistent with the overall semantic evaluation value dimension, filter out core related nodes, extract product category related features, popular search term related features, and brand related features corresponding to the core related nodes, and construct a search extended related dataset; Step 4: Calculate the search expansion range evaluation value based on the user search base dataset and the search expansion association dataset, generate a set of search association expansion words based on the search expansion range evaluation value, and push the set of search association expansion words to the user terminal; Step 5: Based on the search keywords and search association extended terms, retrieve the product database of the e-commerce platform and output the corresponding product search results.

[0006] Preferably, the specific steps for constructing the user search basic dataset are as follows: Remove meaningless symbols, special characters, and stop words from the search keyword text to obtain valid text content; The effective text content is segmented to obtain an initial segmentation set consisting of multiple segmentation units with independent semantics; Invalid word segments are removed from the initial word segmentation set to obtain the set of valid word segments, and the number of valid word segments is counted. By associating and integrating search keyword text, effective text content, effective word segmentation unit set, and effective word segmentation unit quantity, a structured user search basic dataset is formed.

[0007] Preferably, the specific steps for obtaining the semantic association evaluation value of all word segments are as follows: The set of effective word segmentation units is retrieved from the user search basic dataset. Based on the pre-trained e-commerce domain natural language processing deep learning model, the fixed-dimensional semantic vector corresponding to each effective word segmentation unit is extracted. Based on the order of effective word segmentation units in the search keyword text, calculate the difference between the semantic vectors corresponding to two adjacent effective word segmentation units, take the magnitude of the difference result, and obtain the semantic association evaluation value of a single group of adjacent word segmentation units. Traverse all adjacent valid word segments to obtain the semantic association evaluation value of all word segments.

[0008] Preferably, the specific steps for generating the overall semantic evaluation value corresponding to the search keywords are as follows: Retrieve all semantic association evaluation values ​​of the word segments, and sum them up to obtain the overall semantic coherence evaluation value. Based on the overall semantic coherence evaluation value, the semantic vectors corresponding to all effective word segments are sequentially concatenated and integrated according to the order of the effective word segments in the search keyword text to obtain the overall semantic evaluation value corresponding to the search keyword.

[0009] Preferably, the specific steps for obtaining the core associated nodes are as follows: Based on a pre-built knowledge graph of the e-commerce domain, entity nodes with the same dimension as the overall semantic evaluation value are matched, and the ratio of the vector magnitude of the entity node to the magnitude of the overall semantic evaluation value is calculated to obtain the matching evaluation value of the entity semantics. The system presets the effective range of entity semantic matching evaluation values ​​and filters entity nodes that fall within the effective range as core associated nodes.

[0010] Preferably, the specific steps for constructing the search extended related dataset are as follows: Traverse the first-order association path of the core association nodes in the e-commerce domain knowledge graph, extract the features of the corresponding product category dimension as product category association features, and count the number of corresponding valid product categories. Extract features from the dimensions of popular search terms within a fixed historical period of the corresponding platform as popular search term association features, and count the cumulative hits of the corresponding popular search terms; Extract the brand dimension features under the corresponding valid product categories as brand association features, and count the total number of corresponding associated brands; By integrating product category association features, the number of effective product categories, popular search term association features, cumulative hits of popular search terms, brand association features, and the total number of associated brands, a structured extended search association dataset is formed.

[0011] Preferably, the specific steps for obtaining the search extension range evaluation value are as follows: The number of effective word segments is retrieved from the user search basic dataset, and the number of effective product categories, the cumulative hit count of popular search terms, and the total number of associated brands are retrieved from the search extended related dataset. Calculate the magnitude of the difference between the overall semantic evaluation value and the vector corresponding to each core associated node, and sum all the magnitude results to obtain the cumulative semantic bias evaluation value; The cumulative evaluation value of semantic deviation, the number of effective product categories, the cumulative hit of popular search terms, and the total number of associated brands are all subjected to dimensionless normalization. All normalization results are summed to obtain the comprehensive evaluation value of the extended dimension. Divide the comprehensive evaluation value of the expanded dimensions by the number of effective word segments to obtain the final evaluation value of the search expansion range.

[0012] Preferably, the specific steps for generating the set of search association extended terms based on the search extension range evaluation value are as follows: The target number of search association expansion terms to be generated is determined based on the numerical range of the search expansion range evaluation value. Retrieve the text content corresponding to product category association features, popular search term association features, and brand association features from the search extended association dataset; The retrieved text content is semantically combined with the search keyword text in the user search base dataset. The combined text retains the complete semantics of product retrieval, resulting in an initial extended word set. The entities are sorted from high to low based on their semantic matching scores with the overall semantic evaluation value. Expansion words that meet the target number of generated words are selected, and duplicate content is removed to obtain the final set of search association expansion words.

[0013] Preferably, the specific steps for outputting the corresponding product search results are as follows: Retrieve the set of search keywords and search association extended terms, and use the search keywords and each search association extended term as search terms to traverse the product database of the e-commerce platform to obtain the initial set of product results corresponding to each search term; Merge all initial product result sets, remove duplicate product data, and obtain the final product result set; The final product result set is sorted according to its matching degree with the overall semantic evaluation value, and then pushed to the user terminal to complete the output.

[0014] An AI-based e-commerce intelligent search semantic understanding extension system includes: The data processing module is used to acquire the search keyword text input by the user, perform standardized preprocessing to obtain a set of effective word segmentation units, and build a basic dataset of user search. The semantic understanding module is used to extract the semantic vectors corresponding to each effective word segmentation unit based on a pre-trained deep learning model for natural language processing in the e-commerce domain, calculate the semantic association evaluation value of all word segments, and integrate them to generate the overall semantic evaluation value corresponding to the search keywords. The graph matching module is used to match entity nodes that are consistent with the overall semantic evaluation value dimension based on a pre-built e-commerce domain knowledge graph, filter out core related nodes, extract corresponding related features, and build a search extended related dataset. The extended calculation module is used to calculate the search expansion range evaluation value based on the user's basic search dataset and the search expansion association dataset, and to generate a set of search association expansion words based on the search expansion range evaluation value and push them to the user. The retrieval output module is used to retrieve product search results from the e-commerce platform's product database based on search keywords and search association extended terms.

[0015] Beneficial effects The present invention has the following beneficial effects: (1) The AI-based e-commerce intelligent search semantic understanding extension method and system, through the semantic vector extraction and integration structure based on the pre-trained e-commerce domain natural language processing deep learning model, can fully capture the deep semantics and potential search intent of the user's input search keywords, solve the problem that the existing technology can only match the literal meaning of keywords and cannot identify the user's potential needs, and achieve the AI-based accurate analysis effect of search keyword semantics.

[0016] (2) The AI-based e-commerce intelligent search semantic understanding extension method and system, through the knowledge graph entity matching structure based on AI-generated semantic evaluation value, can obtain multi-dimensional related features that are highly consistent with the user's search intent, solve the problem of insufficient matching degree between the search related content and user needs in the existing technology, and realize the AI-based accurate filtering effect of the search related content.

[0017] (3) The AI-based e-commerce intelligent search semantic understanding expansion method and system can dynamically adapt to the expansion scale of different search requests by setting a search expansion range evaluation value calculation structure based on multi-dimensional semantic features. This solves the problem of insufficient matching or over-expansion caused by the fixed expansion rules of the existing technology and realizes the AI-based adaptive adjustment effect of the search expansion range.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] Figure 1 This is a flowchart of the AI-based e-commerce intelligent search semantic understanding extension method of the present invention; Figure 2 This is a structural diagram of the AI-based e-commerce intelligent search semantic understanding extension system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a technical solution: an AI-based method and system for semantic understanding extension in e-commerce intelligent search, such as... Figure 1 As shown, it includes: Step 1: Obtain the search keyword text input by the user, perform standardized preprocessing on the search keyword text to obtain a set of effective word segmentation units, and construct the basic dataset of user search.

[0022] The specific steps for constructing the basic user search dataset are as follows: First, obtain the complete search keyword text entered by the user in the search bar of the e-commerce platform. Then, clean the content by using the platform's built-in text filtering rules to remove meaningless symbols, special characters, and stop words from the search keyword text, thus obtaining valid text content.

[0023] The text filtering rules are set based on common specifications in the e-commerce search field. Meaningless symbols and special characters include non-semantic punctuation, garbled text, and spaces entered by the user. Stop words adopt a set of words with no search meaning that are common in the e-commerce search field.

[0024] The effective text content is segmented using a word segmentation model adapted for the e-commerce domain, resulting in an initial word segmentation set composed of multiple word segmentation units with independent semantics.

[0025] The word segmentation model is trained on a massive corpus from the e-commerce domain and can accurately identify core semantic units such as product names, categories, and brands. Invalid word segments irrelevant to product retrieval are removed from the initial word segmentation set to obtain a set of valid word segments. The number of valid word segments within this set is simultaneously counted. The search keyword text, valid text content, the set of valid word segments, and the number of valid word segments are linked and integrated according to the unique session identifier of a single user search request. A key-value pair structure is used to construct the user search basic dataset, where the key is the unique session identifier corresponding to the current request, and the key value is all the corresponding data after integration. The dataset is written to the in-memory database of the e-commerce platform's retrieval server in real time and can be retrieved in real time as the retrieval process progresses. After the retrieval process is completed, it is stored according to the platform's data archiving rules.

[0026] It is worth noting that this step removes invalid and interfering information from the user input through standardized preprocessing, which ensures that the subsequent semantic extraction process only targets content with actual retrieval significance, thus avoiding interference from invalid information on the semantic understanding results.

[0027] Step 2: Based on the pre-trained deep learning model for natural language processing in the e-commerce domain, extract the semantic vectors corresponding to each effective word segmentation unit in the user search basic dataset, calculate the semantic association evaluation value of the entire word segmentation, and integrate all semantic vectors according to the order of the effective word segmentation units in the search keyword text to generate the overall semantic evaluation value corresponding to the search keyword.

[0028] The specific steps for obtaining the word segmentation semantic association evaluation value are as follows: Retrieve the user search base dataset corresponding to the current search request from the e-commerce platform's in-memory database, and extract the set of valid word segmentation units from the dataset.

[0029] We retrieve a pre-trained deep learning model for natural language processing in the e-commerce field. This model is pre-trained based on the full volume of product title text, user historical search corpus, product detail description text, and product category labeling data from the e-commerce platform over the past three years. The model adopts a general semantic encoding structure and outputs a fixed 128-dimensional semantic vector. After pre-training, the model parameters are fixed and deployed on the e-commerce platform's retrieval server, which can respond to semantic vector extraction requests in real time.

[0030] Each effective word segmentation unit in the set of effective word segmentation units is sequentially input into the model, and a fixed-dimensional semantic vector corresponding to each effective word segmentation unit is extracted. Based on the order of the effective word segmentation units in the original search keyword text, the semantic vectors corresponding to two adjacent effective word segmentation units are compared, and the magnitude of the difference is taken to obtain the semantic association evaluation value for a single set of adjacent word segmentation units.

[0031] Traverse all adjacent valid word segmentation unit combinations within the set of valid word segmentation units to complete the calculation of the semantic association evaluation value of all word segments. All calculation results are stored in association with the unique session identifier corresponding to the current search request.

[0032] The method for obtaining the semantic association evaluation value of adjacent word segments is as follows: In the formula, It represents the semantic association evaluation value of a single group of adjacent word segments, which is the magnitude of the difference between the semantic vectors of two adjacent effective word segments, and is a non-negative scalar. This represents a fixed 128-dimensional semantic vector output by a deep learning model from the preceding effective word segmentation units. The vector dimension is fixed by the pre-trained model, and the values ​​are taken from the real-time output of the model corresponding to the current search request. This represents a fixed 128-dimensional semantic vector output by a deep learning model from subsequent effective word segmentation units. The vector dimension is... Completely consistent, with values ​​derived from the real-time output of the model corresponding to this search request.

[0033] The specific steps for generating the overall semantic evaluation value corresponding to the search keywords are as follows: Retrieve all word segmentation semantic association evaluation values ​​corresponding to the current search request, and sum all word segmentation semantic association evaluation values ​​to obtain the overall semantic coherence evaluation value.

[0034] The overall semantic coherence evaluation value is obtained as follows: In the formula, This represents the overall semantic coherence evaluation value. Indicates the first Group adjacent word segmentation semantic association evaluation value, The value range is 1 to , This represents the total number of adjacent combinations of effective word segments, and its value is one less than the number of effective word segments. The value is derived from the number of effective word segments in the user's search dataset. Based on the obtained overall semantic coherence evaluation value, the semantic vectors corresponding to all effective word segments are sequentially concatenated and integrated according to the order of the effective word segments in the original search keyword text to obtain the overall semantic evaluation value corresponding to the search keyword.

[0035] The specific method of sequential splicing and integration is as follows: according to the order of the effective word segmentation units in the original search keyword text, the 128-dimensional semantic vectors corresponding to each effective word segmentation unit are spliced ​​together end to end to generate a 128-dimensional overall semantic evaluation value with the same dimension as the single word segmentation vector. During the splicing process, the overall semantic coherence evaluation value is used as the basis to ensure that the semantic relationship between adjacent words is completely preserved after integration.

[0036] The overall semantic evaluation value is a fixed 128-dimensional vector data, which is completely consistent with the semantic vector dimension output by the pre-trained model. It can be directly used for subsequent knowledge graph matching calculations. The calculation results are associated with and stored with the unique session identifier corresponding to the current search request.

[0037] It is worth noting that this step, by calculating the magnitude of the difference in semantic vectors between adjacent words, can accurately capture the semantic relationships between adjacent words in the search keyword text, restore the complete semantic logic of the user's input content, and avoid the semantic loss problem caused by the isolated processing of each word after traditional word segmentation. The integration of semantic vectors is completed based on the cumulative result of the semantic relationship evaluation value of word segmentation, which can ensure that the generated overall semantic evaluation value fully restores the complete semantic logic and potential search intent of the user's input search keywords, and avoids the semantic deviation problem caused by ignoring the relationship between words in the traditional semantic integration process.

[0038] Step 3: Based on the pre-built e-commerce domain knowledge graph, match entity nodes that are consistent with the overall semantic evaluation value dimension, filter out core related nodes, extract product category related features, popular search term related features, and brand related features corresponding to the core related nodes, and construct a search extended related dataset.

[0039] The specific steps for obtaining the core associated nodes are as follows: The pre-built e-commerce domain knowledge graph is retrieved. This knowledge graph is constructed based on the e-commerce platform's full product data, product classification system, brand registration information, and historical popular search term statistics. The graph has four core node types: product entities, product category entities, brand entities, and popular search term entities. The nodes are associated with three fixed relationships: subordinate, related, and same category. Each node generates a fixed 128-dimensional vector data, and the vector dimensions are completely consistent with the vector dimensions of the overall semantic evaluation value. The graph data is synchronized monthly according to the platform's product update rhythm and deployed in the e-commerce platform's graph database, which can respond to entity matching requests in real time.

[0040] The generated overall semantic evaluation value is input into the e-commerce domain knowledge graph. All entity nodes with the same dimension as the overall semantic evaluation value are matched, and the ratio of the vector magnitude of the entity node to the magnitude of the overall semantic evaluation value is calculated to obtain the matching evaluation value of the entity semantics.

[0041] The entity semantic matching evaluation value is obtained as follows: In the formula, The matching evaluation value of entity semantics is the ratio of the magnitude of the entity node vector to the magnitude of the overall semantic evaluation value, and is a non-negative scalar. This represents a fixed 128-dimensional vector corresponding to the entity node to be matched in the e-commerce domain knowledge graph. The vector dimension is completely consistent with the overall semantic evaluation value, and the value comes from the pre-built e-commerce domain knowledge graph. This represents the overall semantic evaluation value generated for this search request. It is a fixed 128-dimensional vector, and its value comes from the calculation results of the previous steps.

[0042] The effective range of the matching evaluation value of the preset entity semantics is 0.7 to 1.0. Entity nodes within this effective range are selected as core associated nodes. All core associated nodes are associated and stored with a unique session identifier corresponding to the current search request.

[0043] The specific steps for constructing the search-extended related dataset are as follows: Traverse all the first-order association paths of the core association nodes obtained by the filtering in the e-commerce domain knowledge graph. The first-order association path is the node connection path that has a direct relationship with the core association node. Extract the features of the corresponding product category dimension as product category association features, and simultaneously count the number of valid product categories corresponding to the core association nodes.

[0044] Features of popular search terms in the platform's historical data for the past 30 consecutive days corresponding to the core related nodes are extracted as popular search term association features, and the cumulative hit count of popular search terms corresponding to the core related nodes is counted simultaneously.

[0045] Extract the brand dimension features of the core related nodes under the effective product categories as brand association features, and simultaneously count the total number of related brands under the corresponding effective product categories.

[0046] The product category association features, the number of valid product categories, the popular search term association features, the cumulative number of popular search term hits, the brand association features, and the total number of associated brands are all associated and integrated according to the unique session identifier corresponding to the current search request. The search extended association dataset is constructed in a key-value pair structure, where the key name is the unique session identifier corresponding to the current request, and the key value is all the corresponding data after association and integration.

[0047] The dataset is written to the in-memory database of the e-commerce platform's retrieval server in real time, and can be retrieved in real time as the retrieval process progresses. After the retrieval process is completed, the data is stored according to the platform's data archiving rules.

[0048] It is worth noting that this step completes the matching and filtering of entity nodes by calculating the ratio of vector magnitudes. This can accurately filter out entity nodes that are highly matched with the user's search semantics and remove irrelevant nodes with insufficient matching evaluation values. This ensures that the subsequently extracted related features match the user's search needs and avoids irrelevant features interfering with the search expansion results. By extracting corresponding features through the first-order related paths of core related nodes, it can be ensured that all extracted related features have a direct semantic relationship with the user's search needs, avoiding the problem of feature overgeneralization caused by cross-level association.

[0049] Step 4: Calculate the search expansion range evaluation value based on the user search base dataset and the search expansion association dataset, generate a set of search association expansion words based on the search expansion range evaluation value, and push the set of search association expansion words to the user terminal.

[0050] The specific steps for obtaining the search extension range evaluation value are as follows: The system retrieves the number of valid word segments corresponding to the current search request from the user search base dataset, and the total number of valid product categories, cumulative hits of popular search terms, and total number of associated brands corresponding to the current search request from the search extended association dataset. It calculates the magnitude of the difference between the overall semantic evaluation value and the vector corresponding to each core association node, and sums all the magnitude results to obtain the cumulative semantic bias evaluation value.

[0051] The cumulative evaluation value of semantic bias is obtained as follows: In the formula, The cumulative semantic deviation evaluation value is the sum of the magnitudes of the differences between the overall semantic evaluation value and the vectors corresponding to all core associated nodes, and is a non-negative scalar. This represents the overall semantic evaluation value generated for this search request, and is a fixed 128-dimensional vector. Indicates the first A fixed 128-dimensional vector corresponding to each core associated node; The index of the core associated node, ranging from 1 to... ; This refers to the total number of core related nodes obtained from the current search request.

[0052] The cumulative evaluation value of semantic deviation, the number of effective product categories, the cumulative hits of popular search terms, and the total number of associated brands were all subjected to dimensionless normalization. The normalization method adopted was the maximum-minimum normalization method of the historical full data of the e-commerce platform. After normalization, all parameters were dimensionless scalars in the range of 0 to 1. All normalization results were summed to obtain the comprehensive evaluation value of the extended dimension.

[0053] The comprehensive evaluation value of the extended dimensions is obtained as follows: In the formula, This represents the comprehensive evaluation value of the extended dimensions, which is the summation result of multiple sets of related feature data. This represents the cumulative evaluation value of the normalized semantic bias. This represents the number of valid product categories after normalization. This represents the cumulative hit count of popular search terms after normalization. This represents the total number of related brands after normalization.

[0054] The final search expansion range evaluation value is obtained by dividing the comprehensive evaluation value of the expansion dimension by the number of effective word segments. The calculation result is associated with the unique session identifier corresponding to the current search request and stored in the in-memory database of the e-commerce platform's retrieval server.

[0055] The search extension range evaluation value is obtained as follows: In the formula, This represents the evaluation value of the search expansion range, which is the core basis for dynamically adjusting the expansion range for this search request, and is a non-negative scalar. This represents the comprehensive evaluation value of the extended dimensions, and the value is derived from the calculation results of the previous steps; This represents the number of valid word segmentation units, and is a positive integer whose value comes from the user search dataset.

[0056] The specific steps for generating a set of search association expanded terms based on the search expansion range evaluation value are as follows: Based on the calculated range of search expansion evaluation values, a positive correlation is established between the preset numerical intervals and the number of expanded terms generated. The higher the value of the search expansion evaluation value, the more target terms are generated. The correlation is set based on the historical search data statistics of the e-commerce platform over the past year, thus determining the target number of search association expanded terms generated.

[0057] Retrieve the text content corresponding to product category association features, popular search term association features, and brand association features from the search extended association dataset.

[0058] The text content corresponding to the three types of features retrieved is semantically combined with the original search keyword text in the user search dataset to obtain an initial expanded word set.

[0059] The specific rules for semantically compliant combinations are as follows: product category association features and brand association features are combined using a modifier-head structure of original search keywords plus feature text; popular search term association features directly retain the complete text content. All combined texts must possess complete product retrieval semantics, and content with no retrieval meaning and disordered word order must be removed. Sort the words by matching the corresponding features with the overall semantic evaluation value from high to low, filter out the extended words that meet the target number of generated words, remove duplicate content in the set, and obtain the final set of search association extended words.

[0060] Once the search suggestion set is generated, the sorted set of suggestions is pushed to the search interface of the e-commerce platform currently accessed by the user through the real-time interaction interface of the front end. The suggestions are then displayed in the drop-down suggestion area of ​​the search bar, and the user can directly click on the corresponding suggestion to complete the quick search.

[0061] It is worth noting that this step calculates the search expansion range evaluation value using objectively collected and monitorable data. This directly reflects the semantic clarity and scalability of the current search request, enabling dynamic adaptation of the search expansion scale. This avoids issues such as insufficient matching or over-expansion caused by a fixed expansion scale. Based on the search expansion range evaluation value, the number of generated expansion terms is dynamically adjusted to adapt to the semantic clarity of different search requests. This provides users with search association content that highly matches their search intent, while also covering expansion dimensions such as related products, popular search terms, and different brands of similar products, fully matching the user search needs of e-commerce search scenarios.

[0062] Step 5: Based on the search keywords and search association extended terms, retrieve the product database of the e-commerce platform and output the corresponding product search results.

[0063] The specific steps for outputting the corresponding product search results are as follows: The process involves retrieving the original search keywords and related search terms for the current search request. Each keyword and related term is then used as an independent search term. The inverted index of the e-commerce platform's full product database is traversed to obtain an initial product result set for each search term. The e-commerce platform's full product database is constructed using an inverted index structure, allowing for real-time response to product search requests for search terms. Finally, all initial product result sets corresponding to search terms are merged, and duplicate product data is removed to obtain the final product result set.

[0064] Calculate the matching evaluation value between the vector corresponding to each product in the final product result set and the overall semantic evaluation value. Sort the product results from high to low according to the matching evaluation value. After sorting, push the sorted product result set to the user's search results page through the e-commerce platform's search results rendering interface and complete the display output according to the platform's general pagination rules.

[0065] The matching evaluation value is obtained in the following way: In the formula, The matching evaluation value between the product and the overall semantic evaluation value is the ratio of the magnitude of the product's corresponding vector to the magnitude of the overall semantic evaluation value, and is a non-negative scalar. This represents a fixed 128-dimensional vector corresponding to the product to be sorted. The vector dimension is completely consistent with the overall semantic evaluation value, and the value comes from the pre-generated product vector data in the e-commerce platform's product library. This represents the overall semantic evaluation value generated for this search request, and its value is derived from the calculation results of previous steps.

[0066] It is worth noting that this step uses both the original search keywords and the search association terms as the retrieval basis, which can break through the limitations of traditional retrieval based solely on original keywords, cover the user's potential search needs, improve the relevance and comprehensiveness of search results, and sort the results based on the overall semantic evaluation value, ensuring that the top-ranked products have the highest matching evaluation value with the user's search intent, thereby improving the user's search experience and retrieval efficiency.

[0067] AI-based e-commerce intelligent search semantic understanding extension system, such as Figure 2 As shown, it includes: The data processing module is used to acquire the search keyword text input by the user, perform standardized preprocessing to obtain a set of effective word segmentation units, and build a basic dataset of user search. The semantic understanding module is used to extract the semantic vectors corresponding to each effective word segmentation unit based on a pre-trained deep learning model for natural language processing in the e-commerce domain, calculate the semantic association evaluation value of all word segments, and integrate them to generate the overall semantic evaluation value corresponding to the search keywords. The graph matching module is used to match entity nodes that are consistent with the overall semantic evaluation value dimension based on a pre-built e-commerce domain knowledge graph, filter out core related nodes, extract corresponding related features, and build a search extended related dataset. The extended calculation module is used to calculate the search expansion range evaluation value based on the user's basic search dataset and the search expansion association dataset, and to generate a set of search association expansion words based on the search expansion range evaluation value and push them to the user. The retrieval output module is used to retrieve product search results from the e-commerce platform's product database based on search keywords and search association extended terms.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0069] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An AI-based method for semantic understanding extension in e-commerce intelligent search, characterized in that: Includes the following steps: Step 1: Obtain the search keyword text input by the user, perform standardized preprocessing on the search keyword text to obtain a set of effective word segmentation units, and construct the basic dataset of user search. Step 2: Based on the pre-trained deep learning model for natural language processing in the e-commerce domain, extract the semantic vectors corresponding to each effective word segmentation unit in the user search basic dataset, calculate the semantic association evaluation value of the entire word segmentation, and integrate all semantic vectors according to the order of the effective word segmentation units in the search keyword text to generate the overall semantic evaluation value corresponding to the search keyword. Step 3: Based on the pre-built e-commerce domain knowledge graph, match entity nodes that are consistent with the overall semantic evaluation value dimension, filter out core related nodes, extract product category related features, popular search term related features, and brand related features corresponding to the core related nodes, and construct a search extended related dataset; Step 4: Calculate the search expansion range evaluation value based on the user search base dataset and the search expansion association dataset, generate a set of search association expansion words based on the search expansion range evaluation value, and push the set of search association expansion words to the user terminal; Step 5: Based on the search keywords and search association extended terms, retrieve the product database of the e-commerce platform and output the corresponding product search results.

2. The AI-based e-commerce intelligent search semantic understanding extension method according to claim 1, characterized in that, The specific steps for constructing the basic user search dataset are as follows: Remove meaningless symbols, special characters, and stop words from the search keyword text to obtain valid text content; The effective text content is segmented to obtain an initial segmentation set consisting of multiple segmentation units with independent semantics; Invalid word segments are removed from the initial word segmentation set to obtain the set of valid word segments, and the number of valid word segments is counted. By associating and integrating search keyword text, effective text content, effective word segmentation unit set, and effective word segmentation unit quantity, a structured user search basic dataset is formed.

3. The AI-based e-commerce intelligent search semantic understanding extension method according to claim 1, characterized in that, The specific steps for obtaining the semantic association evaluation value of all word segments are as follows: The set of effective word segmentation units is retrieved from the user search basic dataset. Based on the pre-trained e-commerce domain natural language processing deep learning model, the fixed-dimensional semantic vector corresponding to each effective word segmentation unit is extracted. Based on the order of effective word segmentation units in the search keyword text, calculate the difference between the semantic vectors corresponding to two adjacent effective word segmentation units, take the magnitude of the difference result, and obtain the semantic association evaluation value of a single group of adjacent word segmentation units. Traverse all adjacent valid word segments to obtain the semantic association evaluation value of all word segments.

4. The AI-based e-commerce intelligent search semantic understanding extension method according to claim 1, characterized in that, The specific steps for generating the overall semantic evaluation value corresponding to the search keywords are as follows: Retrieve all semantic association evaluation values ​​of the word segments, and sum them up to obtain the overall semantic coherence evaluation value. Based on the overall semantic coherence evaluation value, the semantic vectors corresponding to all effective word segments are sequentially concatenated and integrated according to the order of the effective word segments in the search keyword text to obtain the overall semantic evaluation value corresponding to the search keyword.

5. The AI-based e-commerce intelligent search semantic understanding extension method according to claim 1, characterized in that, The specific steps for obtaining the core associated nodes are as follows: Based on a pre-built knowledge graph of the e-commerce domain, entity nodes with the same dimension as the overall semantic evaluation value are matched, and the ratio of the vector magnitude of the entity node to the magnitude of the overall semantic evaluation value is calculated to obtain the matching evaluation value of the entity semantics. The system presets the effective range of the entity semantic matching evaluation value, and then filters the entity nodes that fall within the effective range as core associated nodes.

6. The AI-based e-commerce intelligent search semantic understanding extension method according to claim 1, characterized in that, The specific steps for constructing the search extended related dataset are as follows: Traverse the first-order association path of the core association nodes in the e-commerce domain knowledge graph, extract the features of the corresponding product category dimension as product category association features, and count the number of corresponding valid product categories. Extract features from the dimensions of popular search terms within a fixed historical period of the corresponding platform as popular search term association features, and count the cumulative hits of the corresponding popular search terms; Extract the brand dimension features under the corresponding valid product categories as brand association features, and count the total number of corresponding associated brands; By integrating product category association features, the number of effective product categories, popular search term association features, cumulative hits of popular search terms, brand association features, and the total number of associated brands, a structured extended search association dataset is formed.

7. The AI-based e-commerce intelligent search semantic understanding extension method according to claim 1, characterized in that, The specific steps for obtaining the search extension range evaluation value are as follows: The number of effective word segments is retrieved from the basic user search dataset, and the number of effective product categories, cumulative hits of popular search terms, and total number of associated brands are retrieved from the extended search dataset. Calculate the magnitude of the difference between the overall semantic evaluation value and the vector corresponding to each core associated node, and sum all the magnitude results to obtain the cumulative semantic bias evaluation value; The cumulative evaluation value of semantic deviation, the number of effective product categories, the cumulative hit of popular search terms, and the total number of associated brands are all subjected to dimensionless normalization. All normalization results are summed to obtain the comprehensive evaluation value of the extended dimension. Divide the comprehensive evaluation value of the expanded dimensions by the number of effective word segments to obtain the final evaluation value of the search expansion range.

8. The AI-based e-commerce intelligent search semantic understanding extension method according to claim 1, characterized in that, The specific steps for generating the set of search association expanded terms based on the search expansion range evaluation value are as follows: The target number of search association expansion terms to be generated is determined based on the numerical range of the search expansion range evaluation value. Retrieve the text content corresponding to product category association features, popular search term association features, and brand association features from the search extended association dataset; The retrieved text content is semantically combined with the search keyword text in the user search base dataset. The combined text retains the complete semantics of product retrieval, resulting in an initial extended word set. The entities are sorted from high to low based on their semantic matching scores with the overall semantic evaluation value. Expansion words that meet the target number of generated words are selected, and duplicate content is removed to obtain the final set of search association expansion words.

9. The AI-based e-commerce intelligent search semantic understanding extension method according to claim 1, characterized in that, The specific steps for outputting the corresponding product search results are as follows: Retrieve the set of search keywords and search association extended terms, and use the search keywords and each search association extended term as search terms to traverse the product database of the e-commerce platform to obtain the initial set of product results corresponding to each search term; Merge all initial product result sets, remove duplicate product data, and obtain the final product result set; The final product result set is sorted according to its matching degree with the overall semantic evaluation value, and then pushed to the user terminal to complete the output.

10. An AI-based e-commerce intelligent search semantic understanding extension system, applied to the AI-based e-commerce intelligent search semantic understanding extension method according to any one of claims 1 to 9, characterized in that, include: The data processing module is used to acquire the search keyword text input by the user, perform standardized preprocessing to obtain a set of effective word segmentation units, and build a basic dataset of user search. The semantic understanding module is used to extract the semantic vectors corresponding to each effective word segmentation unit based on a pre-trained deep learning model for natural language processing in the e-commerce domain, calculate the semantic association evaluation value of all word segments, and integrate them to generate the overall semantic evaluation value corresponding to the search keywords. The graph matching module is used to match entity nodes that are consistent with the overall semantic evaluation value dimension based on a pre-built e-commerce domain knowledge graph, filter out core related nodes, extract corresponding related features, and build a search extended related dataset. The extended calculation module is used to calculate the search expansion range evaluation value based on the user's basic search dataset and the search expansion association dataset, and to generate a set of search association expansion words based on the search expansion range evaluation value and push them to the user. The retrieval output module is used to retrieve product search results from the e-commerce platform's product database based on search keywords and search association extended terms.

Images