Intelligent question and answer implementation method and system based on large model and semantic map

Abstract

The invention discloses an intelligent question and answer implementation method and system based on a large model and a semantic atlas, and relates to the technical field of supply chain intellectualization, and the method comprises the steps: obtaining a structured distribution data table and an unstructured text data stream, carrying out the data cleaning and normalization processing, and obtaining a structured distribution data table and an unstructured text data stream; constructing a unified knowledge base to perform semantic understanding and knowledge extraction on the unified knowledge base, extracting key entities and semantic relationships through named entity recognition and relationship extraction tasks, and fusing structured distribution data features to construct a preliminary semantic map; and integrating the preliminary semantic map with the spatial dimension features acquired by the data interface and the real-time task execution log of the shop patrol assistant, endowing nodes with space-time attributes through graph neural network training, and outputting a dynamic semantic map integrated with space-time dynamic mode features. According to the method, geofence coordinates and crowd density characteristics are injected by utilizing an Atde API, a store patrol assistant real-time replenishment state log is superposed, and the causal association strength modeling precision is improved through ST-GNN training.

Description

Technical Field

[0001] The present invention relates to the field of intelligent supply chain technology, and in particular to a method and system for implementing intelligent question answering based on a large model and semantic graph. Background Art

[0002] The current field of intelligent retail supply chain relies on the evolution of multi-source data fusion technology framework and has formed a standardized technical route: the mainstream general solution uses a rule engine to achieve structured cleaning and pipeline transmission of cross-platform API data, supporting inventory and order data integration of O2O platforms; combined with pre-trained language models to achieve entity recognition and relationship extraction in product description texts, and build static network topology based on knowledge graphs for out-of-stock problem attribution analysis.

[0003] However, existing solutions have significant bottlenecks in spatiotemporal dynamic modeling: First, the multi-platform store entity alignment method relies on a single feature and does not integrate the inverse weight of geographic coordinates and cross-platform name frequency features, resulting in a high error rate in platform ID mapping; second, the static graph is not connected to the geo-fence data of the AutoNavi API and the minute-level status updates of store inspection logs, resulting in insufficient accuracy in spatiotemporal correlation strength modeling and the inability to capture real-time business fluctuations. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an intelligent question-answering implementation method based on a large model and semantic graph to solve the problem of high attribution error rate of the root cause of morning peak out-of-stock caused by insufficient spatiotemporal fusion of multi-source retail data.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a method for implementing intelligent question answering based on a large model and semantic graph, which includes obtaining a structured distribution data table and an unstructured text data stream, performing data cleaning and normalization processing, and constructing a unified knowledge base; Perform semantic understanding and knowledge extraction on the unified knowledge base, extract key entities and semantic relationships through named entity recognition and relationship extraction tasks, and integrate structured distribution data features to build a preliminary semantic graph; Integrate the preliminary semantic graph with the spatial dimension features obtained from the data interface and the real-time task execution logs of the store inspection assistant. Through graph neural network training, assign spatiotemporal attributes to nodes and output a dynamic semantic graph that integrates spatiotemporal dynamic pattern features. Based on the natural language questions input by users, a large language model is used to analyze intent and entities. Furthermore, a dynamic semantic graph that integrates spatiotemporal dynamic pattern features is used to retrieve multi-dimensional store sales data, task execution records, and user feedback aggregate data to generate a structured query vector. The structured query vector is input into a dynamic semantic graph that integrates spatiotemporal dynamic pattern features for multi-hop reasoning. The store nodes are traversed to analyze the status and edge attributes, the spatiotemporal patterns are identified, and user feedback data nodes are associated to generate structured answers.

[0007] As a preferred solution of the intelligent question-answering implementation method based on a large model and semantic graph described in the present invention, the structured distribution data table includes a brand product field, a real-time shelf status field, an accurate inventory field, a historical sales trend field, and a forecast data field; The unstructured text data stream includes the main content of user evaluations, product question-and-answer interaction texts, and text store inspection records.

[0008] As a preferred solution of the intelligent question answering method based on a large model and semantic graph described in the present invention, the construction of a unified knowledge base includes the following steps: Clean the structured distribution data table and the unstructured text data stream, and output the cleaned structured distribution data table and the cleaned unstructured text data stream; Calculate the similarity of the cleaned structured distribution data and perform weighted summation to generate comprehensive similarity; Execute the K-medoids clustering algorithm on store entities whose comprehensive similarity exceeds the similarity threshold, and output cross-platform entity clustering groups; Assign a globally unique standardized store ID code to each cluster group, replace the original store code field in the cleaned structured distribution data table, and generate a normalized structured distribution data table; The normalized structured distribution data table is associated with the cleaned unstructured text data stream to build a unified knowledge base.

[0009] As a preferred solution of the intelligent question answering method based on a large model and a semantic graph according to the present invention, the construction of a preliminary semantic graph includes the following steps: Load the pre-trained large language model parameters and initialize the entity recognition and relationship extraction tasks. At the same time, extract user reviews and store inspection text records from the unified knowledge base, and generate a standardized text parsing input stream through segmentation and encoding. The standardized text input stream is processed through the embedding layer and Transformer encoding layer of the large language model, and a structured entity set is generated through entity recognition function parsing and label mapping; The relation extraction function based on the large language model generates a set of semantic relations. After integrating the characteristics of product distribution data, the association strength is calculated through the self-attention mechanism to construct a preliminary semantic graph.

[0010] As a preferred solution of the intelligent question answering method based on a large model and a semantic graph according to the present invention, the output of the dynamic semantic graph integrating spatiotemporal dynamic pattern features comprises the following steps: Based on the store's geographic location in the preliminary semantic graph, we obtain business district labels, geo-fence coordinates, and crowd density data. After data parsing and binding with the global store ID field primary key, we output a spatially enhanced semantic graph. Using the spatially enhanced semantic graph as input, the global store ID field of the store node is extracted to generate a data request list. The store inspection assistant interface is called to obtain real-time task logs. The node status attributes are updated through field mapping and a timestamp field is added to output a spatiotemporally enhanced dynamic semantic graph. The store node attribute fields are converted into a node feature matrix, and the adjacency matrix is ​​generated by combining the problematic relationship edges. Through iterative training of the graph convolutional neural network model, a dynamic semantic graph integrating spatiotemporal dynamic pattern features is output.

[0011] As a preferred solution of the intelligent question answering method based on a large model and semantic graph of the present invention, the generating of a structured query vector includes the following steps: After normalizing the natural language input by the user, the large language model performs word segmentation, intent classification, and named entity recognition to generate parsing result tuples. Using the regional entity field value in the parsed result tuple as the index key, the dynamic semantic graph that integrates spatiotemporal dynamic pattern features is used to match the store nodes in the corresponding business district. All question type nodes are traversed to generate structured search results. Based on the set of associated store nodes in the structured search results, GMV trends, replenishment efficiency and user feedback data are extracted, a three-dimensional feature vector is constructed, and then assembled into a structured query vector.

[0012] As a preferred solution of the intelligent question-answering method based on a large model and semantic graph of the present invention, the generating of structured answers includes the following steps: Extract the corresponding field values ​​according to the structured query vector field structure, match the store nodes in the corresponding business district in the dynamic semantic graph that integrates spatiotemporal dynamic pattern features, filter the node types, and output a set of high-potential stores; Based on the high-potential store set, we traverse the three-level links of products, tasks, and questions, integrate relationship weights and status data, and build a complete traversal path tree; Calculate task completion rate and complaint burst coefficient based on traversal path tree, and generate spatiotemporal pattern matching report; Based on the task completion rate and complaint outbreak coefficient, the root cause is output, the estimated GMV loss value is calculated, and a structured answer is generated.

[0013] In a second aspect, the present invention provides an intelligent question-answering implementation system based on a large model and semantic graph, including a knowledge base construction module that obtains structured distribution data tables and unstructured text data streams, performs data cleaning and normalization processing, and constructs a unified knowledge base; The graph construction module performs semantic understanding and knowledge extraction on the unified knowledge base. It extracts key entities and semantic relationships through named entity recognition and relationship extraction tasks, and integrates the characteristics of structured distribution data to build a preliminary semantic graph. The spatiotemporal enhancement module integrates the preliminary semantic graph with the spatial dimension features obtained from the data interface and the real-time task execution logs of the store inspection assistant. It assigns spatiotemporal attributes to nodes through graph neural network training and outputs a dynamic semantic graph that integrates spatiotemporal dynamic pattern features. The vector generation module uses a large language model to parse intent and entities based on natural language questions input by users. It then uses a dynamic semantic graph that integrates spatiotemporal dynamic pattern features to retrieve multi-dimensional store sales data, task execution records, and user feedback aggregate data to generate structured query vectors. The answer generation module inputs the structured query vector into a dynamic semantic graph that integrates spatiotemporal dynamic pattern features for multi-hop reasoning, traverses store nodes to analyze status and edge attributes, identifies spatiotemporal patterns, associates user feedback data nodes, and generates structured answers.

[0014] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the method for implementing intelligent question answering based on a large model and a semantic graph as described in the first aspect of the present invention.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for implementing intelligent question answering based on a large model and a semantic graph as described in the first aspect of the present invention.

[0016] The beneficial effects of the present invention are as follows: by comprehensively analyzing similarities and generating global store IDs through K-medoids clustering, the error rate of store mapping across multiple platforms is reduced; by using the AutoNavi API to inject geo-fence coordinates and crowd density features, and superimposing the real-time replenishment status log of the store patrol assistant, the causal relationship strength modeling accuracy of "breakfast milk out of stock → delivery delay" is improved through ST-GNN training. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 This is a flowchart of the implementation method of intelligent question answering based on large models and semantic graphs.

[0019] Figure 2 Schematic diagram of building a unified knowledge base.

[0020] Figure 3 Construct a schematic for a preliminary semantic graph.

[0021] Figure 4 Schematic diagram for generating structured query vectors and answers. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0025] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a method for implementing intelligent question answering based on a large model and a semantic graph, comprising the following steps: S1. Obtain structured distribution data tables and unstructured text data streams, perform data cleaning and normalization, and build a unified knowledge base.

[0026] S1.1. Clean the structured distribution data table and the unstructured text data stream, and output the cleaned structured distribution data table and the cleaned unstructured text data stream.

[0027] Specifically, we call API interfaces to initiate programmatic requests for multiple O2O platforms, including Meituan, Ele.me, and JD Daojia, and output the original API response data streams from each O2O platform. We extract the brand product field, real-time shelf status field, accurate inventory field, historical sales trend field, and forecast data field from the original API response data streams of each O2O platform, map them to the store code field, and convert them into a unified store coding rule to generate a data set with standardized store codes. We then construct a data table structure and reorganize the data set with standardized store codes (such as brand product columns, real-time shelf status columns, accurate inventory columns, historical sales trend columns, and forecast data columns) to generate a structured distribution data table. Configure a list of URLs for user review pages and product Q&A detail pages from multiple O2O platforms, such as Meituan, Ele.me, and JD Daojia, as crawler entry points. Initialize the web crawler framework's request queue and schedule crawler threads to access the corresponding user review pages and product Q&A detail pages at a preset frequency. Apply web page parsing technology to extract the user review content and product Q&A interaction text. Simultaneously, retrieve text store inspection records submitted by salespersons through the store inspection assistant mini-program storage area interface. Organize these into continuous user review content, product Q&A interaction text, and text store inspection records based on the dimensions of multiple O2O platforms, such as Meituan, Ele.me, and JD Daojia, forming a complete unstructured text data stream. Load the structured distribution data table into the rules engine of the B1 data cleansing product. Enable regular expression matching to remove invalid characters from the brand product field and the real-time listing status field. Standardize the field formats and use outlier filtering rules to remove negative values ​​in the precise inventory field, as well as abnormal fluctuations in the historical sales trend field and forecast data field (such as month-over-month growth exceeding 500%). Output the cleaned structured distribution data table. Remove HTML tags, advertising codes, and special control characters from the body of user reviews, product Q&A interaction text, and store inspection records in the unstructured text data stream. Then, enable outlier filtering rules to remove reviews with more than 20 consecutive repeated characters and Q&A / store inspection records with more than 30% garbled characters. Output the cleaned unstructured text data stream. S1.2. Calculate the similarity of the cleaned structured distribution data separately and then perform weighted summation to generate a comprehensive similarity.

[0028] Specifically, the core module of the One-ID algorithm is executed on the store name field, address field, and latitude and longitude fields of the cleaned structured distribution data table. Similarity functions are used to calculate the name text similarity, address edit distance score, and latitude and longitude spherical distance difference between stores on different O2O platforms, such as Meituan, Ele.me, and JD Daojia. To further explain, calculating the name text similarity between stores on different platforms means performing Chinese word segmentation and stop word filtering on each store name string to generate a standardized store name word vector; inputting the standardized store name word vector into the cosine similarity algorithm, calculating the word frequency-inverse document frequency weight value between each pair of store names on different O2O platforms such as Meituan, Ele.me and JD Daojia, and outputting a name text similarity score in the range of 0-1.

[0029] Calculating the address edit distance score involves extracting the store address field string value, performing standardized mapping of provincial, municipal, and district administrative units and road name normalization on each address string to generate a formatted address string. The formatted address strings from different O2O platforms are then combined into an edit distance algorithm, and the minimum number of single-character operations between the address strings on platforms such as Meituan, Ele.me, and JD Daojia is calculated, converting the result into an address edit distance score in the range of 0-1.

[0030] Calculating the spherical distance of longitude and latitude involves extracting the values ​​of the longitude and latitude fields, converting the decimal coordinates into radians in the spherical coordinate system, and inputting the longitude and latitude radians of different O2O platforms into the Haversine formula to calculate the shortest path distance on the surface.

[0031] Apply an inverse function to the longitude and latitude spherical distance to convert it into a 0-1 similarity score (1 / (1+longitude and latitude spherical distance)), and output the standardized longitude and latitude similarity; The name text similarity score (weight 40%), address edit distance score (weight 30%) and normalized latitude and longitude similarity (weight 30%) are weighted and summed to obtain the comprehensive similarity. S1.3. Execute the K-medoids clustering algorithm on store entities whose comprehensive similarity exceeds the similarity threshold, and output cross-platform entity clustering groups for stores on multiple O2O platforms such as Meituan, Ele.me, and JD Daojia.

[0032] To further explain, executing the K-medoids clustering algorithm means scanning the comprehensive similarity scores of all store entity pairs, retaining store entity pairs whose comprehensive similarity exceeds the similarity threshold, and generating a set of high-confidence matching entity pairs; randomly selecting K store entities from the set of high-confidence matching entity pairs as initial cluster centers (the K value is adaptively determined based on the total number of entities), where K is not less than the integer value of the total number of stores on platforms such as Meituan, Ele.me, and JD Daojia divided by 5; assigning each store entity to the nearest cluster center (the distance is defined as 1-comprehensive similarity score); reselecting the entity with the highest average similarity to other entities in each cluster as the new cluster center within each cluster; repeating the assignment and center reselection until the cluster center no longer changes after three consecutive iterations, and outputting the final stable cluster group structure, where each cluster group contains store entities from different O2O platforms such as Meituan, Ele.me, and JD Daojia, and the entities within the cluster group share the same real-world store identifier.

[0033] It should be noted that the similarity threshold is set based on the accuracy requirement of strict cross-platform entity matching, aiming to screen entity pairs with extremely high confidence to ensure the accuracy of multi-platform store clustering.

[0034] S1.4. Assign a globally unique standardized store ID code to each cluster group, replace the original store code field in the cleaned structured distribution data table, and generate a normalized structured distribution data table containing all fields, including the brand product field of the global store ID, the real-time shelf status field, and the accurate inventory field.

[0035] S1.5. Use the global store ID field in the normalized structured distribution data table as the primary key, associate the store attribute label (the business district type label derived from the address field + latitude and longitude fields) and the user review text content and product question and answer interaction text in the cleaned unstructured text data stream to build a unified knowledge base.

[0036] To further explain, building a unified knowledge base means mapping brand product fields, real-time shelf status fields, precise inventory fields, historical sales trend fields, and forecast data fields into columns to build a product distribution status dimension table; mapping address fields into STRING type columns, longitude and latitude fields into GEO_POINT type columns, and business district type label fields into STRING type columns to build a store basic information dimension table; mapping user evaluation text content fields into TEXT type columns, product Q&A interaction text fields into TEXT type columns, and text store tour record fields into TEXT type columns to build a user voice dimension table; creating a product distribution status dimension table, a store basic information dimension table, and a user voice dimension table in the distributed storage engine to complete the construction of the unified knowledge base.

[0037] S2. Perform semantic understanding and knowledge extraction on the unified knowledge base, extract key entities and semantic relationships through named entity recognition and relationship extraction tasks, and integrate structured distribution data features to build a preliminary semantic graph.

[0038] S2.1. Load the pre-trained large language model parameters and initialize the entity recognition task and relationship extraction task. At the same time, extract the user evaluation content and store inspection text records from the unified knowledge base, and generate a standardized text parsing input stream through segmentation and encoding processing.

[0039] Specifically, obtain the storage path of the model parameter file of a pre-trained large language model (such as BERT or similar variants); load the model parameter file into memory, instantiate the inference calculation graph of the basic semantic understanding engine; initialize the entity category label set of the named entity recognition task in the basic semantic understanding engine; To further explain, the entity category label set includes four categories: specific product SKU name, store ID, problem type and time point.

[0040] It should be noted that the training of the pre-trained large language model is as follows: first, basic training is performed on a large amount of general corpus using masked language modeling and next sentence prediction tasks, and the cross-entropy loss function is optimized through multiple rounds of iterations using the Transformer with a BERT-like architecture; secondly, specialized training is performed in the retail field, and vertical corpora such as cross-platform product descriptions, store inspection reports, and user complaint texts in the unified knowledge base are input. On the basis of the MLM task, entity boundary detection and relationship type classification tasks are added, and the entity position offset loss and relationship triplet confidence loss are jointly optimized; finally, spatiotemporal dynamic pattern features are incorporated for graph adaptive training, and product distribution data (distribution rate, GMV curve) and user feedback time series are converted into feature vectors and spliced ​​into the text embedding layer. The modeling ability of the semantic topology of retail scenarios such as "store-product-problem" is enhanced through graph topology regularization constraints and relationship edge weight modeling, and finally a trained large language model that supports complex spatiotemporal pattern parsing is output.

[0041] Synchronously initialize the relationship type template library of the relationship extraction task (including predefined relationship patterns such as "there is a problem" and "happened in"); extract all text records of the user evaluation body content field from the user voice dimension table of the unified knowledge base; extract all text content of the text store inspection record from the store inspection assistant applet log storage area associated with the unified knowledge base; merge the text records of the user evaluation body content field and the text content of the text store inspection record data stream to form an unstructured text data input stream; split the unstructured text data input stream into text fragment batches according to a fixed length of 512 characters; input the text fragment batches into the text parsing interface of the basic semantic understanding engine one by one, perform UTF-8 encoding conversion and special character escape processing on each text fragment, and generate a standardized text parsing input stream.

[0042] S2.2. The standardized text input stream is processed through the embedding layer and Transformer encoding layer of the large language model, and a structured entity set is generated through entity recognition function parsing and label mapping.

[0043] Specifically, each text segment of the standardized text parsing input stream is input into the embedding layer of the pre-trained large language model, and the text segment is converted into a 768-dimensional dense vector representation to generate a text vector sequence. The text vector sequence is input into the Transformer encoding layer of the pre-trained large language model, and the context-aware feature vector is calculated through the multi-head self-attention mechanism, outputting an enhanced text feature vector sequence. The enhanced text feature vector sequence is input into the named entity recognition function of the pre-trained large language model, and a position label sequence is generated through character-level position label prediction (B / I / O label set) based on the Softmax classifier. Perform entity classification mapping on a sequence of position labels: The continuous character sequence of the tag category "BRAND_SKU" is mapped to the specific product SKU name entity field and associated with the brand product field in the unified knowledge base; The tag category "STORE_ID" is mapped to the store ID entity field and associated with the global store ID field; The label category "ISSUE_TYPE" is mapped to the issue type entity field (such as "damaged product"); The tag category "TIMESTAMP" is parsed into the minute-level time point entity field using the regular expression "\d{4}-\d{2}-\d{2} \d{2}:\d{2}"; The output is a structured entity collection containing the specific product SKU name entity field, store ID entity field, problem type entity field, and time point entity field.

[0044] S2.3. Generate a set of semantic relationships based on the relation extraction function of the large language model. After integrating the characteristics of product distribution data, calculate the association strength through the self-attention mechanism and construct a preliminary semantic graph.

[0045] Specifically, the structured entity set is input into the relation extraction function of the pre-trained large language model. The relation extraction function matches the "specific product SKU name - store ID - question type" format, generates triples, and verifies their matching with the unified knowledge base fields. At the same time, it identifies the "question type - time point" relationship to generate relationship pairs and outputs a semantic relationship set. The semantic relationship set is combined with the store-average SKU numerical segment values ​​and the distribution completion rate field values ​​extracted from the unified knowledge base's product distribution status dimension table to form a composite feature vector. This is then input into the feature fusion layer of the pre-trained large language model for weighted integration, outputting an enhanced feature vector that combines text semantic relationships with structured data features. The enhanced feature vector is input into the self-attention mechanism layer of the pre-trained large language model. The semantic relevance scores between entity nodes (e.g., the relevance strength between the "breakfast milk" node and the "Chaoyang District store" node) and the confidence scores of the relationship edges (0-1 floating point values) are calculated. The output is a set of semantic relevance scores containing entity relevance strengths and relationship confidences. Based on the specific product SKU name entity field, store ID entity field, problem type entity field and time point entity field of the semantic relationship set, four types of nodes are created: product node, store node, problem type node and solution node. Among them, the product node attribute maps the specific product SKU name entity field and the unified knowledge base brand product field, the store node attribute maps the store ID entity field and the unified knowledge base business district type label field, the problem type node attribute maps the problem type entity field, and the solution node attribute is automatically generated based on the predefined relationship model. At the same time, the time point entity field is mapped to the timestamp attribute, and the relationship confidence score of the semantic association score set is mapped to the confidence attribute attached to the relationship edge, and a preliminary graph structure containing node attributes and relationship edge metadata is output; The preliminary graph structure connects relationship edges through the graph layout function of the pre-trained large language model: product nodes are connected to problem type nodes through the "existing problem" relationship edge, problem type nodes are connected to time point attributes through the "occurred in" relationship edge, and store nodes are connected to business district attributes through the "attributed location" relationship edge. The entity association strength in the semantic association score set is used to adjust the edge weights, and finally an initial semantic graph of multidimensional metadata with timestamp attributes, confidence attributes, and spatial relationships is output.

[0046] S3. Integrate the preliminary semantic graph of multidimensional metadata with the spatial dimension features obtained from the data interface and the real-time task execution log of the store inspection assistant. Through graph neural network training, assign spatiotemporal attributes to nodes and output a spatiotemporally enhanced dynamic semantic graph.

[0047] S3.1. Based on the store location in the preliminary semantic graph, obtain business district labels, geo-fence coordinates, and crowd density data. After data parsing and binding with the global store ID field primary key, output a spatially enhanced semantic graph.

[0048] Specifically, based on the latitude and longitude field values ​​and store name field values ​​of the store nodes in the preliminary semantic graph, the store location request parameter list is extracted and combined. The store location request parameter list is used to request the offline sales tag field through the Geocoding API of the AutoNavi Map API, the geofencing boundary coordinate field through the GeoFencing API, and the real-time heat map data field of the business district through the Heatmap API, thus outputting the original geographic data set. Perform JSON format parsing on the original geographic dataset: extract the "business_area" field value from the geocoding service response as the offline sales tag field, extract the "fence_vertices" field value from the geofence service response as the geofence boundary coordinate field, and extract the "density_index" field value from the heatmap service response as the business district population density field to generate a structured spatial feature data table; Using the global store ID field of the store node in the preliminary semantic graph as the primary key association mapping, the offline sales label field of the structured spatial feature data table is bound to the "business district" attribute field of the store node, the geofence boundary coordinate field is bound to the "geofence coordinate" attribute field, and the business district crowd density field is bound to the "surrounding crowd density" attribute field. The output is a spatially enhanced semantic graph.

[0049] S3.2. Using the spatially enhanced semantic graph as input, extract the global store ID field of the store node to generate a data request list. Call the store inspection assistant interface to obtain the real-time task log. Update the node status attributes through field mapping and add a timestamp field to output a spatiotemporally enhanced dynamic semantic graph.

[0050] Specifically, the global store ID field values ​​of all store nodes in the spatially enhanced semantic graph are extracted to generate a real-time data request ID list. At the same time, the current timestamp field value is obtained as the data query deadline field. The real-time data request ID list and the current timestamp field value are used to request the latest data through the real-time task execution log interface of the store inspection assistant applet: the OCR task result interface is called to obtain the recognition result text field of the shelf completion photo field (including the product name field and the shelf time field), and the inventory check interface is called to obtain the digital inventory status update data field (including the accurate inventory quantity field and the inspection time field). The original real-time task log dataset is output. Perform JSON parsing on the original real-time task log dataset: extract the "product_status" field value from the OCR task result interface response and map it to the shelf completion status field. Extract the "current_stock" field value from the inventory check interface response and map it to the latest inventory status field. Generate a structured real-time status data table containing the global store ID field, shelf completion status field, and latest inventory status field. Using the global store ID field of the store node in the spatially enhanced semantic graph as the primary key, the latest inventory status field value of the structured real-time status data table is matched with the inventory status attribute field of the updated store node, and the shelf completion status field value is matched with the task completion status attribute field of the updated store node. Unmatched nodes retain the original status attribute field value. After the update is completed, a "last updated time" attribute field (taken from the current timestamp field value) is added to all store nodes, and the business district field, geographic fence coordinate field, and surrounding population density field in the original spatially enhanced semantic map remain unchanged, outputting a spatiotemporally enhanced dynamic semantic map.

[0051] S3.3. Convert the store node attribute fields into a node feature matrix, combine them with the problem relationship edges to generate an adjacency matrix, and iteratively train the graph convolutional neural network model to output a dynamic semantic graph that integrates spatiotemporal dynamic pattern features.

[0052] Specifically, we extract the store node attribute fields (business district, surrounding population density, and latest inventory status) from the spatiotemporal-enhanced dynamic semantic graph and convert them into numerical vectors to form row vectors of the node feature matrix. These vectors are then connected to the "problem" edge in the spatiotemporal-enhanced semantic graph to generate an adjacency matrix (e.g., the matrix element value corresponding to connected nodes is 1, and the matrix element value corresponding to disconnected nodes is 0). Load the standard parameter initialization file for the graph convolutional layer of the graph convolutional neural network model, use the node feature matrix and adjacency matrix as input, and perform iterative training: The graph convolution layer performs a neighbor node feature aggregation operation, performing a weighted summation of each store node's business district field, surrounding crowd density field, and directly connected problem type node features, outputting a new node representation matrix containing spatial association features. This new node representation matrix is ​​input into the next round of training iterations, and the training is repeated. Each iteration uses the ReLU activation function to process the convolution result to update the node feature matrix. At the same time, the connection strength of the "existing problem" relationship edge is updated based on the feature propagation result. When the loss function change value converges after 10 consecutive iterations, the training is terminated and the trained dynamic semantic map is output. In the trained dynamic semantic graph, the store nodes retain the spatial location attribute field (geofence boundary coordinate field), the problem type nodes add the time dynamic attribute field (latest inventory status field + task completion status field), and the edge relationships add the spatiotemporal association weight field (for example, the correlation strength between "breakfast milk shortage problem" and "office building lunchtime order peak" is 0.92). The output is a dynamic semantic graph that integrates spatiotemporal dynamic pattern features.

[0053] S4. Based on the natural language input by the user, a large language model is called to parse the intent and entities. The dynamic semantic graph that integrates spatiotemporal dynamic pattern features is then used to retrieve multi-dimensional store sales data, task execution records, and user feedback aggregate data to generate a structured query vector.

[0054] S4.1. After normalizing the natural language input by the user, generate a parsing result tuple through word segmentation and intent classification and named entity recognition using a large language model.

[0055] Specifically, the natural language question string entered by the user on the front-end interface (for example, "Why is breakfast milk often out of stock in Chaoyang District?") is normalized by removing leading and trailing spaces and control characters. This normalized question string is then input into the tokenizer function of the pre-trained large language model for subword segmentation. A sequence of token ID vectors with special tags is generated, and a sequence of token embedding vectors is output. Intent classification is then performed. Pooled feature vectors are extracted at the special tag positions, and the probability distribution of intent categories is calculated using the Softmax classifier. The "Inventory Attribution Analysis" intent classification label field, which is in the Probability Distribution > Intent Classification Probability Threshold, is selected as the valid output. It should be noted that the setting of the intent classification probability threshold is based on the decision boundary determined by the ROC curve analysis of the large language model on the retail intent classification validation set.

[0056] At the same time, named entity recognition is performed on the word embedding vector sequence. The BIO tagging scheme is used to optimize the label sequence output through the CRF layer. The continuous entity tag fields are merged to generate the product category entity field "breakfast milk" and the region entity field "Chaoyang District". The intent classification label field is mapped to a fixed-format string, and the entity field is converted into a key-value pair dictionary structure, which is combined to generate the parsing result tuple. S4.2. Use the regional entity field value in the parsed result tuple as the index key, match the store nodes in the corresponding business district in the dynamic semantic graph that integrates spatiotemporal dynamic pattern features, traverse all question type nodes, and generate structured retrieval results.

[0057] Specifically, the regional entity field value (e.g., "Chaoyang District") in the parsed result tuple is used as the index key input. The "business district attribute field" of all store nodes in the dynamic semantic graph that integrates spatiotemporal dynamic pattern features is traversed to perform an exact match of the attribute values. Store nodes with attribute field values ​​equal to "Chaoyang District" are selected to generate the "Chaoyang District associated store node set." Based on the global store ID field of all store nodes in the Chaoyang District associated store node set, a topological traversal is performed along the "existing problem" relationship edge in the dynamic semantic graph that integrates spatiotemporal dynamic pattern features. All directly associated problem type nodes are found. Nodes whose product name attribute field value contains "breakfast milk" in the problem type node are filtered out to generate the "breakfast milk out-of-stock problem node set." A newly created association index table records the mapping relationship between each store node ID field and the corresponding problem node ID field, maintaining independent storage structures for the "Chaoyang District associated store node set" and the "breakfast milk out-of-stock problem node set." This outputs structured search results for the associated store node set, which includes global store ID attribute fields and geofence coordinate attribute fields, and the associated problem node set, which includes problem type attribute fields and occurrence timestamp attribute fields.

[0058] S4.3. Based on the set of associated store nodes in the structured search results, extract GMV trends, replenishment efficiency, and user feedback data, construct a three-dimensional feature vector, and assemble it into a structured query vector.

[0059] Specifically, the associated store node set in the structured search results is used as input, and all global store ID attribute field values ​​are extracted to form a store ID index list. The store ID index list is used to query the product distribution status dimension table, retrieve the GMV contribution rate change trend curve field for breakfast products, and filter the records where the product name field matches "breakfast milk", outputting a multi-dimensional dynamic sales data field block. The store inspection assistant task record interface is called through the store ID index list to request the average completion delay rate field of historical replenishment tasks and limit the time range, outputting a task execution record field block. Using the store ID index list, we search the user feedback dashboard system's aggregate database, extracting complaint period statistics fields such as "breakfast milk delivery anomaly," and calculating the peak percentage of complaints during the morning hours (e.g., 07:00-09:00). This data is then output as a user feedback statistics field block. We also merge the multi-dimensional dynamic sales data field block, the task execution record field block, and the user feedback statistics field block, grouping and aligning them by store ID field to form a multimodal data block. Extract the peak percentage of complaints in the morning period from the user feedback statistics field block of the multimodal data block as the time dimension feature; extract the geo-fence coordinate attribute field value of the first store node from the associated store node set as the spatial dimension feature; extract the average completion delay rate field value of historical replenishment tasks from the task execution record field block of the multimodal data block as the execution link feature; and encapsulate the time dimension feature, spatial dimension feature, and execution link feature into a three-dimensional feature vector. Rename the time dimension field in the three-dimensional feature vector to, for example, "Morning Time," the spatial dimension field to, for example, "Chaoyang District Business District," and the execution step dimension field to, for example, "Replenishment Task Efficiency." Then, assemble a structured query vector based on the preset field structure, including the intent label field (for example, "Inventory Attribution Analysis"), key entity fields (for example, product category: "Breakfast Milk," region: "Chaoyang District"), the time dimension field, the spatial dimension field, and the execution step field.

[0060] It should be noted that user feedback refers to unstructured text information, which is used to capture users' real-time feedback, questions, and records on products or services. User feedback mainly includes three types of unstructured text: User review content: Product reviews or service feedback posted by users on O2O platforms (such as Meituan and Ele.me), such as comments on the quality of "breakfast milk" or delivery time.

[0061] Product Q&A interaction text: Q&A records between users and merchants on the product details page, such as inquiries about inventory status or complaints about out-of-stock issues.

[0062] Text store inspection records: On-site inspection reports submitted by sales staff through the store inspection assistant mini-program, including product listing status or problem records.

[0063] S5. Input the structured query vector into the dynamic semantic graph that integrates spatiotemporal dynamic pattern features for multi-hop reasoning. Traverse the store nodes to analyze the status and edge attributes, identify spatiotemporal patterns, and associate user feedback data nodes to generate structured answers.

[0064] S5.1. Extract the corresponding field values ​​according to the structured query vector field structure, match the store nodes in the corresponding business district in the dynamic semantic graph that integrates spatiotemporal dynamic pattern features, filter the node types, and output a set of high-potential stores.

[0065] Specifically, according to the field structure of the structured query vector, the region entity field value is extracted as the spatial index key, the time dimension field value is used as the time window identifier, and the product category entity field value is used as the target product identifier. The business district attribute field values ​​of all store nodes are scanned in the dynamic semantic graph that integrates spatiotemporal dynamic pattern features. A string exact match is performed to filter out store nodes whose business district attribute field values ​​are equal to the spatial index key, generating a set of region-matching store nodes. Attribute filtering is performed on the set of region-matching store nodes, checking the node type attribute field value of each store node, and outputting a set of high-potential store nodes. S5.2. Based on the set of high-potential stores, traverse the three-level links of products, tasks, and questions, integrate relationship weights and status data, and build a complete traversal path tree.

[0066] Specifically, starting with the high-potential store node set as the input, traverse the outbound edges of each store node, filter the edges with the relationship type of "product-store", connect the target product nodes along the edges, and output the product node traversal set; then, starting with the high-potential store node set, traverse the outbound edges, filter the edges with the relationship type of "store-task", connect the store inspection task nodes along the edges, and output the task node traversal set; then, starting with the task node traversal set, traverse the outbound edges, filter the edges with the relationship type of "task-problem", connect the delivery problem nodes along the edges, and output the problem node traversal set; During the traversal process, the spatiotemporal weight field values ​​of the "product-store," "store-task," and "task-problem" relationship edges are synchronously recorded. The inventory status attribute field values ​​of the product nodes and the task completion status attribute field values ​​of the task nodes on the connection path are copied to generate a path record containing a list of relationship edge weights, inventory status, and task status. Finally, all path records are grouped by the ID field of the high-potential store node. A tree structure is constructed with the high-potential store node as the root node, the target product node as the first-level child node (connected by the "product-store" edge), the store inspection task node as the second-level child node (connected by the "store-task" edge), and the delivery problem node as the third-level child node (connected by the "task-problem" edge). The complete traversal path tree containing the path record attribute data is output. S5.3. Calculate the task completion rate and complaint burst coefficient based on the traversal path tree and generate a spatiotemporal pattern matching report.

[0067] Specifically, extract the task completion status attribute field values ​​and occurrence timestamp attribute field values ​​of all secondary child nodes in the complete traversal path tree, filter the task nodes whose occurrence timestamp attribute field values ​​are in the "07:00-09:00" interval, and generate a morning task node set; count the number of nodes in the morning task node set with the task completion status attribute field value of "True" as the completion number, count the total number of nodes in the set as the total number, and calculate the task completion rate field value; Among them, the task completion rate field value = (number of completed tasks / total number × 100%); Mark the corresponding first-level child nodes whose task completion rate field values ​​are lower than the task completion rate threshold as "promotional products not put on shelves in time" gap nodes, and output a list of gap nodes and completion rate values. Extract the problem type attribute field values ​​and occurrence timestamp attribute field values ​​of all third-level child nodes in the complete traversal path tree, filter out problem nodes with timestamp field values ​​in the "07:00-09:00" range, and generate a set of morning problem nodes. Count the number of nodes in the morning problem node set whose problem type attribute field value equals "breakfast delivery delay" as the number of complaint nodes, count the total number of nodes as the total number of problem nodes, and calculate the complaint burst coefficient. Among them, complaint outbreak coefficient = (number of complaint nodes / total number of problem nodes); It should be noted that the task completion rate threshold is set based on the business standard value of the retail industry, such as 85%, which is a rigid baseline to ensure that the promotional products are put on the shelves in time to meet the peak demand for breakfast consumption.

[0068] Integrate the "promotional products not put on shelves in time" gap node field, the morning task completion rate field, and the "breakfast delivery delay" complaint burst coefficient field to generate a spatiotemporal pattern matching report. S5.4. Based on the task completion rate and complaint outbreak coefficient, output the root cause, calculate the estimated GMV loss value, and generate a structured answer.

[0069] Specifically, based on the spatiotemporal pattern matching report, the morning task completion rate field values ​​and complaint burst coefficient field values ​​are extracted. When the morning task completion rate is less than the task completion rate threshold and the complaint burst coefficient is greater than the complaint burst coefficient threshold, the root cause field string "Insufficient salesperson morning inspection coverage" is output. The GMV contribution rate change curve field for "breakfast milk" is retrieved from the product distribution status dimension table using the global store ID field corresponding to the gap node, the estimated GMV loss value is calculated, and the quantitative impact field string "Estimated GMV Loss" is output. Among them, estimated GMV loss value = (average value of morning period in the same period in history - current value) / average value of morning period in the same period in history × 100%; It should be noted that the complaint outbreak coefficient threshold is a critical point of strong causal correlation determined through attribution analysis of historical events, such as 0.7. When the proportion of complaints about delivery delays during the breakfast period is >0.7, it is judged as a systemic service failure.

[0070] Get the global store ID field of all gap nodes and generate the solution field string = "Store patrol schedule is forcibly adjusted to start at 7 am"; encapsulate the root cause field, quantitative impact field and solution field combination into a structured answer framework output; use the global store ID field in the gap node list as input to call the store list database interface to extract the relevant store name list field output, call the store patrol task dashboard interface to obtain the task execution efficiency dashboard real-time screenshot field output, and at the same time use the morning period and the global store ID field of the gap node as input to obtain the morning period complaint distribution heat map field output from the user feedback dashboard system; combine the store name list field, task efficiency screenshot field and complaint heat map field into the supporting attachment field dictionary output; merge the structured answer framework and the supporting attachment field dictionary into a structured answer.

[0071] This embodiment also provides an intelligent question-answering system based on a large model and a semantic graph, including: The knowledge base construction module obtains structured distribution data tables and unstructured text data streams, performs data cleaning and normalization, and builds a unified knowledge base; The graph construction module performs semantic understanding and knowledge extraction on the unified knowledge base. It extracts key entities and semantic relationships through named entity recognition and relationship extraction tasks, and integrates the characteristics of structured distribution data to build a preliminary semantic graph. The spatiotemporal enhancement module integrates the preliminary semantic graph with the spatial dimension features obtained from the data interface and the real-time task execution logs of the store inspection assistant. It assigns spatiotemporal attributes to nodes through graph neural network training and outputs a dynamic semantic graph that integrates spatiotemporal dynamic pattern features. The vector generation module uses a large language model to parse intent and entities based on natural language questions input by users. It then uses a dynamic semantic graph that integrates spatiotemporal dynamic pattern features to retrieve multi-dimensional store sales data, task execution records, and user feedback aggregate data to generate structured query vectors. The answer generation module inputs the structured query vector into a dynamic semantic graph that integrates spatiotemporal dynamic pattern features for multi-hop reasoning, traverses store nodes to analyze status and edge attributes, identifies spatiotemporal patterns, associates user feedback data nodes, and generates structured answers.

[0072] This embodiment also provides a computer device, which is suitable for the case of an intelligent question-answering implementation method based on a large model and a semantic graph, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the intelligent question-answering implementation method based on a large model and a semantic graph as proposed in the above embodiment.

[0073] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0074] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for implementing intelligent question answering based on a large model and a semantic graph as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0075] In summary, the present invention reduces the error rate of store mapping across multiple platforms by generating global store IDs through comprehensive similarity and K-medoids clustering. It also utilizes the AutoNavi API to inject geofence coordinates and crowd density features, overlays the real-time replenishment status log of the store patrol assistant, and improves the accuracy of causal relationship strength modeling for the "breakfast milk out-of-stock → delivery delay" problem through ST-GNN training.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for implementing intelligent question answering based on a large model and semantic graph, characterized by: include, Obtain structured distribution data tables and unstructured text data streams, perform data cleaning and normalization, and build a unified knowledge base; Perform semantic understanding and knowledge extraction on the unified knowledge base, extract key entities and semantic relationships through named entity recognition and relationship extraction tasks, and integrate structured distribution data features to build a preliminary semantic graph; Integrate the preliminary semantic graph with the spatial dimension features obtained from the data interface and the real-time task execution logs of the store inspection assistant. Through graph neural network training, assign spatiotemporal attributes to nodes and output a dynamic semantic graph that integrates spatiotemporal dynamic pattern features. Based on the natural language questions input by users, a large language model is used to analyze intent and entities. Combined with a dynamic semantic graph that integrates spatiotemporal dynamic pattern features, it retrieves multi-dimensional store sales data, task execution records, and user feedback aggregate data to generate a structured query vector. The structured query vector is input into a dynamic semantic graph that integrates spatiotemporal dynamic pattern features for multi-hop reasoning. The store nodes are traversed to analyze the status and edge attributes, identify spatiotemporal patterns, and associate user feedback data nodes to generate structured answers.

2. The method for implementing intelligent question answering based on a large model and semantic graph according to claim 1, characterized in that: The structured distribution data table includes a brand product field, a real-time shelf status field, an accurate inventory field, a historical sales trend field, and a forecast data field; The unstructured text data stream includes the main content of user evaluations, product question-and-answer interaction texts, and text store inspection records.

3. The method for implementing intelligent question answering based on a large model and semantic graph according to claim 1, characterized in that: The construction of the unified knowledge base includes the following steps: Clean the structured distribution data table and the unstructured text data stream, and output the cleaned structured distribution data table and the cleaned unstructured text data stream; Calculate the similarity of the cleaned structured distribution data and perform weighted summation to generate comprehensive similarity; Execute the K-medoids clustering algorithm on store entities whose comprehensive similarity exceeds the similarity threshold, and output cross-platform entity clustering groups; Assign a globally unique standardized store ID code to each cluster group, replace the original store code field in the cleaned structured distribution data table, and generate a normalized structured distribution data table; The normalized structured distribution data table is associated with the cleaned unstructured text data stream to build a unified knowledge base.

4. The method for implementing intelligent question answering based on a large model and semantic graph according to claim 1, characterized in that: The construction of the preliminary semantic graph includes the following steps: Load the pre-trained large language model parameters and initialize the entity recognition and relationship extraction tasks. At the same time, extract user reviews and store inspection text records from the unified knowledge base, and generate a standardized text parsing input stream through segmentation and encoding. The standardized text input stream is processed through the embedding layer and Transformer encoding layer of the large language model, and a structured entity set is generated through entity recognition function parsing and label mapping; The relation extraction function based on the large language model generates a set of semantic relations. After integrating the characteristics of product distribution data, the association strength is calculated through the self-attention mechanism to construct a preliminary semantic graph.

5. The method for implementing intelligent question answering based on a large model and semantic graph according to claim 1, characterized in that: The output of the dynamic semantic graph integrating spatiotemporal dynamic pattern features comprises the following steps: Based on the store's geographic location in the preliminary semantic graph, we obtain business district labels, geo-fence coordinates, and crowd density data. After data parsing and binding with the global store ID field primary key, we output a spatially enhanced semantic graph. Using the spatially enhanced semantic graph as input, the global store ID field of the store node is extracted to generate a data request list. The store inspection assistant interface is called to obtain real-time task logs. The node status attributes are updated through field mapping and a timestamp field is added to output a spatiotemporally enhanced dynamic semantic graph. The store node attribute fields are converted into a node feature matrix, and the adjacency matrix is ​​generated by combining the problematic relationship edges. Through iterative training of the graph convolutional neural network model, a dynamic semantic graph integrating spatiotemporal dynamic pattern features is output.

6. The method for implementing intelligent question answering based on a large model and semantic graph according to claim 1, characterized in that: Generating a structured query vector includes the following steps: After normalizing the natural language input by the user, the large language model performs word segmentation, intent classification, and named entity recognition to generate parsing result tuples. Using the regional entity field value in the parsed result tuple as the index key, the dynamic semantic graph that integrates spatiotemporal dynamic pattern features is used to match the store nodes in the corresponding business district. All question type nodes are traversed to generate structured search results. Based on the set of associated store nodes in the structured search results, GMV trends, replenishment efficiency and user feedback data are extracted, a three-dimensional feature vector is constructed, and then assembled into a structured query vector.

7. The method for implementing intelligent question answering based on a large model and semantic graph according to claim 1, characterized in that: Generating structured answers includes the following steps: Extract the corresponding field values ​​according to the structured query vector field structure, match the store nodes in the corresponding business district in the dynamic semantic graph that integrates spatiotemporal dynamic pattern features, filter the node types, and output a set of high-potential stores; Based on the high-potential store set, we traverse the three-level links of products, tasks, and questions, integrate relationship weights and status data, and build a complete traversal path tree; Calculate task completion rate and complaint burst coefficient based on traversal path tree, and generate spatiotemporal pattern matching report; Based on the task completion rate and complaint outbreak coefficient, the root cause is output, the estimated GMV loss value is calculated, and a structured answer is generated.

8. An intelligent question-answering system based on a large model and a semantic graph, based on the intelligent question-answering method based on a large model and a semantic graph according to any one of claims 1 to 7, characterized in that: include, The knowledge base construction module obtains structured distribution data tables and unstructured text data streams, performs data cleaning and normalization, and builds a unified knowledge base; The graph construction module performs semantic understanding and knowledge extraction on the unified knowledge base. It extracts key entities and semantic relationships through named entity recognition and relationship extraction tasks, and integrates the characteristics of structured distribution data to build a preliminary semantic graph. The spatiotemporal enhancement module integrates the preliminary semantic graph with the spatial dimension features obtained from the data interface and the real-time task execution logs of the store inspection assistant. It assigns spatiotemporal attributes to nodes through graph neural network training and outputs a dynamic semantic graph that integrates spatiotemporal dynamic pattern features. The vector generation module uses a large language model to parse intent and entities based on natural language questions input by users. It then uses a dynamic semantic graph that integrates spatiotemporal dynamic pattern features to retrieve multi-dimensional store sales data, task execution records, and user feedback aggregate data to generate structured query vectors. The answer generation module inputs the structured query vector into a dynamic semantic graph that integrates spatiotemporal dynamic pattern features for multi-hop reasoning, traverses store nodes to analyze status and edge attributes, identifies spatiotemporal patterns, associates user feedback data nodes, and generates structured answers.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the intelligent question answering method based on a large model and a semantic graph are implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the intelligent question answering method based on a large model and a semantic graph are implemented as described in any one of claims 1 to 7.

Images