AI-based intelligent recommendation method and device for frequently used shipping information

Abstract

This application relates to the field of computer application technology and discloses an AI-based intelligent recommendation method and apparatus for common shipping information. The method includes: acquiring a shipping information table input by a user; parsing the fields in the shipping information table to extract the filled-in structured data; collecting user behavior data and static knowledge base information formed from the user's historical order records to generate contextual information for the current recommendation scenario; processing the contextual information using a preset intelligent recommendation model, wherein the intelligent recommendation model dynamically calculates the weights of different information sources for the current recommendation scenario by introducing an attention mechanism, extracts information related to missing fields in the table based on the weights from each information source, and generates candidate recommendation content for the missing fields; displaying the candidate recommendation content to the user and receiving user feedback on the candidate recommendation content; updating the parameters of the intelligent recommendation model based on the feedback information to optimize the intelligent recommendation model.

Description

Technical Field

[0001] This application relates to the field of computer application technology, and in particular to an AI-based intelligent recommendation method and apparatus for frequently used shipping information. Background Technology

[0002] In e-commerce, logistics, and supply chain management scenarios, users frequently enter orders in batches using Excel spreadsheets. However, existing systems generally suffer from the following shortcomings:

[0003] 1. Passive and lagging validation methods: Traditional systems can only perform rule validation (such as field format checks) after the user uploads the form, and can only prompt basic errors such as "field missing" and "format error". They cannot predict missing content or provide intelligent recommendations in real time during the user's filling process, which leads to users having to modify the form repeatedly and upload it repeatedly, resulting in low operation efficiency.

[0004] 2. Limited Recommendation Capabilities: Existing recommendation functions often rely on single-dimensional data (such as only user history) and have fixed recommendation strategies, failing to dynamically adjust recommended content based on the real-time context of the current operation (such as order type or semantic relationships between existing fields in a table). For example, traditional systems cannot distinguish between "e-commerce returns and exchanges" and "B2B contract delivery" scenarios, resulting in mismatches between recommended address templates or goods information and the actual scenario, leading to insufficient accuracy.

[0005] 3. Lack of adaptive learning mechanism: Existing systems cannot continuously optimize the recommendation model based on user feedback on recommended content (such as selection, modification, ignoring, etc.), making it difficult for recommendation performance to improve with changes in user habits, and preventing the user experience from forming a virtuous cycle.

[0006] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0007] This application provides an AI-based intelligent recommendation method and apparatus for frequently used shipping information. It addresses the shortcomings of existing technologies, which only implement rule-based passive verification, fail to build an intelligent recommendation system integrating multi-source data (user behavior, static knowledge base, and real-time context), and lack dynamic weight calculation and adaptive learning capabilities. Current technical solutions remain at the stage of "reporting errors after discovering problems," failing to achieve the intelligent leap of "proactively predicting problems and providing solutions," and lacking a closed loop of "data collection - intelligent recommendation - feedback optimization." This results in cumbersome user operations, low system recommendation accuracy, and an inability to meet the needs of efficient batch order processing.

[0008] Firstly, this application provides an AI-based intelligent recommendation method for frequently used shipping information, the method comprising:

[0009] Obtain the shipping information form entered by the user, parse the fields in the shipping information form, and extract the entered structured data; collect user behavior data and static knowledge base information formed from the user's historical order records. The static knowledge base information includes at least commonly used addresses and product database data.

[0010] The parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation are fused together to generate context information for the current recommendation scenario.

[0011] The context information is processed using a preset intelligent recommendation model. The intelligent recommendation model dynamically calculates the weight of different information sources to the current recommendation scenario by introducing an attention mechanism. Based on the weight, information related to the missing fields in the table is extracted from each information source, and recommendation candidate content for the missing fields is generated.

[0012] Display recommended candidate content to users and receive user feedback on the recommended candidate content; update the parameters of the intelligent recommendation model based on the feedback information and optimize the intelligent recommendation model.

[0013] In some embodiments, parsing the fields in the shipping information table and extracting the filled-in structured data includes: identifying the column headers of the shipping information table according to preset field rules to determine the target fields corresponding to the shipping information and the goods information; for each target field, extracting the specific content of the filled cells in the corresponding column and marking the missing fields that are not filled in; performing format validation on the specific content, identifying data that conforms to the preset format specifications and converting it into a structured format, and marking data with incorrect format.

[0014] In some embodiments, the collection of user behavior data and static knowledge base information formed from users' historical order records includes: extracting complete data of historical orders from the user's historical order database, including the address entered, the selected product, the operation time, and the modification record, to form a user behavior dataset; and obtaining common address and product database data from a preset static knowledge base.

[0015] In some embodiments, the process of fusing the parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation to generate context information for the current recommendation scenario includes: associating the filled fields in the structured data with the historical filling habits in the user behavior data; filtering the address templates and product attributes in the static knowledge base in conjunction with the real-time context information of the current operation; and integrating the address templates, product attributes, and associated filled fields with the user behavior data by field dimension to form context information that includes the user's historical preferences, current scenario features, and existing filled content.

[0016] In some embodiments, the step of dynamically calculating the weights of different information sources for the current recommendation scenario by introducing an attention mechanism includes: inputting structured data, user behavior data, static knowledge base information, and real-time context information as multimodal inputs into the attention mechanism module of the intelligent recommendation model; for each missing field, calculating the contribution of each information source to the recommendation of the missing field according to the field type and current scenario characteristics; generating the weight value of the corresponding information source according to the contribution, so that the intelligent recommendation model prioritizes information with higher relevance to the current field.

[0017] In some embodiments, the step of extracting information related to the missing field in the table from each information source based on weight includes: classifying and filtering the multimodal information sources according to the calculated weights, and prioritizing the extraction of data that matches the missing field from high-weight information sources; wherein, for the recipient's phone number field, if the current scenario is a user's historical high-frequency delivery scenario, the historically associated phone number information of the corresponding recipient is first extracted from the user behavior data, and then the phone number information corresponding to the recipient with the same name is matched from the static address database.

[0018] In some embodiments, generating recommended candidate content for missing fields includes: deduplicating and prioritizing the extracted multi-source information, generating at least one recommended candidate content according to weight from high to low; performing correlation verification on the recommended content by combining semantic context reasoning, filtering out obviously mismatched information; and encapsulating the verified recommended candidate content in a structured form to form an interactive recommendation result.

[0019] In some embodiments, updating the parameters of the intelligent recommendation model based on feedback information and optimizing the intelligent recommendation model includes: receiving user feedback information on recommended candidate content, recording the field type, recommended content, and operation time corresponding to the feedback behavior in the feedback information; inputting the feedback information and the corresponding field type, recommended content, and operation time into the model training module, and adjusting the weight calculation logic of the attention mechanism through incremental learning; reducing the weight of the information source corresponding to the recommended content in the field type recommendation for recommended content rejected by the user or modified multiple times; and increasing the weight of the corresponding information source for recommended content that the user has not modified or has modified only once, thereby achieving adaptive optimization of the recommendation model.

[0020] In some embodiments, before processing the context information using a preset intelligent recommendation model, the method further includes: constructing an intelligent recommendation model with an attention mechanism, wherein the input layer of the intelligent recommendation model receives context information fused from multimodal data, and the output layer provides candidate recommendation content for each missing field; training the model using historical order data, user behavior logs, and labeled recommendation scenario data, and adjusting model parameters by comparing the matching degree between the recommendation results and the actual filled content; and generating sub-models for different business scenarios during the training process, enabling the intelligent recommendation model to automatically switch recommendation strategies based on real-time context.

[0021] Secondly, this application provides an AI-based intelligent recommendation device for frequently used shipping information, the device comprising:

[0022] The data extraction unit is used to obtain the shipping information table input by the user, parse the fields in the shipping information table, and extract the structured data that has been filled in; it also collects user behavior data and static knowledge base information formed by the user's historical order records. The static knowledge base information includes at least commonly used addresses and product database data.

[0023] The information generation unit is used to fuse the parsed structured data, user behavior data, static knowledge base information and real-time context information of the current operation to generate context information for the current recommendation scenario.

[0024] The information processing unit is used to process contextual information using a preset intelligent recommendation model. The intelligent recommendation model dynamically calculates the weights of different information sources to the current recommendation scenario by introducing an attention mechanism, and extracts information related to the missing fields in the table from each information source based on the weights to generate recommendation candidate content for the missing fields.

[0025] The intelligent recommendation unit is used to display recommended candidate content to users, receive user feedback on the recommended candidate content, and update the parameters of the intelligent recommendation model based on the feedback information to optimize the intelligent recommendation model.

[0026] This application discloses an AI-based intelligent recommendation method and device for frequently used shipping information. The method proactively recommends missing content by parsing table fields in real time, eliminating the need for users to manually fill in or repeatedly modify tables, significantly reducing the number of re-uploads and shortening order processing time. Based on multimodal data fusion and attention mechanisms, the system dynamically adjusts its recommendation strategy according to the current operational scenario (such as order type and existing content). For example, it intelligently associates frequently used addresses and phone numbers with recipient names, resulting in recommendations that better match actual needs and reduce human error rates. The method continuously optimizes the recommendation model based on user feedback, prioritizing frequently used addresses and products. Recommendation accuracy gradually improves with usage time, creating a personalized experience that becomes increasingly intelligent with use. The generated high-quality structured data lays the foundation for subsequent intelligent order allocation and supply chain optimization, enhancing the overall intelligence level of the business process.

[0027] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating the steps of an AI-based intelligent recommendation method for commonly used shipping information provided in an embodiment of this application.

[0030] Figure 2 This is a schematic block diagram of an AI-based intelligent recommendation device for frequently used shipping information provided in an embodiment of this application;

[0031] Figure 3 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0032] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0035] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0036] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0037] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0038] In e-commerce, logistics, and supply chain management scenarios, users frequently enter orders in batches using Excel spreadsheets. However, existing systems generally suffer from the following shortcomings:

[0039] 1. Passive and lagging validation methods: Traditional systems can only perform rule validation (such as field format checks) after the user uploads the form, and can only prompt basic errors such as "field missing" and "format error". They cannot predict missing content or provide intelligent recommendations in real time during the user's filling process, which leads to users having to modify the form repeatedly and upload it repeatedly, resulting in low operation efficiency.

[0040] 2. Limited Recommendation Capabilities: Existing recommendation functions often rely on single-dimensional data (such as only user history) and have fixed recommendation strategies, failing to dynamically adjust recommended content based on the real-time context of the current operation (such as order type or semantic relationships between existing fields in a table). For example, traditional systems cannot distinguish between "e-commerce returns and exchanges" and "B2B contract delivery" scenarios, resulting in mismatches between recommended address templates or goods information and the actual scenario, leading to insufficient accuracy.

[0041] 3. Lack of adaptive learning mechanism: Existing systems cannot continuously optimize the recommendation model based on user feedback on recommended content (such as selection, modification, ignoring, etc.), making it difficult for recommendation performance to improve with changes in user habits, and preventing the user experience from forming a virtuous cycle.

[0042] The core problem with existing technologies is that they only achieve rule-based passive verification, failing to build an intelligent recommendation system that integrates multi-source data (user behavior, static knowledge base, and real-time context), and lacking dynamic weight calculation and adaptive learning capabilities. Current technical solutions remain at the stage of "reporting errors after problems are discovered," failing to achieve the intelligent leap of "proactively predicting problems and providing solutions," and not forming a closed loop of "data collection - intelligent recommendation - feedback optimization." This results in cumbersome user operations, low system recommendation accuracy, and an inability to meet the needs of efficient batch order processing.

[0043] The ingenuity of this invention lies in breaking through the single-dimensional verification and fixed-strategy recommendation mode of existing technologies. Through multimodal data fusion (structured tabular data, user behavior data, static knowledge base, real-time context), attention mechanism to dynamically calculate information source weights, adaptive learning model optimization and other technical means, an intelligent recommendation system of "real-time parsing-missing prediction-active recommendation-feedback iteration" is constructed, realizing the technological leap from "passive error reporting" to "active suggestion" and from "single rule" to "context awareness".

[0044] Please see Figure 1 , Figure 1 This is a schematic flowchart illustrating an AI-based intelligent recommendation method for frequently used shipping information provided in an embodiment of this application. The method is applied to computer equipment, which can be deployed on a single server or a server cluster. It can also be deployed on handheld terminals, laptops, wearable devices, or robots, etc.

[0045] It should be noted that the acquisition of any information mentioned in the provided methods is in compliance with relevant regulations and is carried out with the user's consent, and will not infringe on the user's privacy or violate relevant laws and regulations.

[0046] like Figure 1 As shown, the specific steps of this AI-based intelligent recommendation method for commonly used shipping information include steps S101 to S104.

[0047] S101. Obtain the shipping information table entered by the user, parse the fields in the shipping information table, and extract the filled-in structured data; collect user behavior data and static knowledge base information formed by the user's historical order records. The static knowledge base information includes at least commonly used addresses and product database data.

[0048] Specifically, this step enables real-time parsing of user input forms, correlation and collection of historical behavioral data, and invocation of static knowledge bases to form a basic data input set.

[0049] Table parsing and structured data extraction include: Parsing engine: Using table parsing tools (such as pandas and openpyxl for Python, and Apache POI for Java) to read Excel tables in real time, supporting dynamic identification of header fields (through a pre-defined field mapping dictionary, such as "Shipping Address" corresponding to "address" and "Product Code" corresponding to "skull"). Field type validation: Performing basic format validation on filled fields (such as regular expression matching for phone numbers and date format conversion), generating a structured data dictionary filled_data (where the key is the field name and the value is the filled content and type). Missing field identification: Identifying a list of missing fields in the table (missing_fields) using a pre-defined set of order fields (such as required fields "Recipient Name" and "Contact Number", and optional field "Remarks").

[0050] User behavior data collection includes: Historical record extraction: querying the user's order records for the past 30 days from the user behavior database, including addresses, product combinations, and order types (such as "B2C retail," "B2B bulk," and "returns / exchanges"), forming a user behavior sequence user_history (structured data format, including timestamps, order scenario tags, and field preferences). High-frequency feature extraction: extracting frequently used addresses (such as home address and company address), commonly used product SKUs, and preferred logistics methods through statistical analysis (such as TF-IDF and sliding window).

[0051] The static knowledge base includes: a frequently used address library containing preset standard address templates (such as enterprise receiving address library and after-sales warehouse address library), with each address associated with scenario tags (such as "B2B shipping address" and "return and exchange address") and weights (calculated based on address usage frequency). A product library stores basic product information (SKU, product name, specifications, weight), supporting fuzzy search by product name or precise SKU matching, and associating products with their respective categories (such as "3C electronics" and "apparel and bags").

[0052] S102. The parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation are fused together to generate context information for the current recommendation scenario.

[0053] Specifically, by semantically fusing structured tabular data, user behavior, static knowledge base, and real-time operation scenarios, a context vector containing scenario semantics is generated, providing comprehensive input for intelligent recommendation.

[0054] Real-time context information extraction includes: Operation scenario labels: scene_labels are generated by inferring from the "order type" field already filled in the table (such as "B2B contract delivery" or "e-commerce return and exchange" manually marked by the user) or implicit features (such as inferring a B2B scenario if the "delivery address" contains a company name). Filling progress features: the proportion of filled fields and the completeness of key fields (such as address and product) are calculated to generate a filling progress vector progress_vector (such as [0.7, 0.9] representing the completion rate of the address field and the product field, respectively).

[0055] Multimodal data vectorization includes: Structured data embedding: Converting field values ​​in filled_data into vectors (e.g., address embedding using word segmentation + Word2Vec, product SKU embedding using One-Hot encoding) to generate table feature vectors (table_embedding). User behavior embedding: Encoding user_history into user preference vectors (user_embedding) using a time series model (e.g., LSTM) to capture historical operational habits (e.g., preference for SF Express logistics, frequently used recipient name patterns). Knowledge base embedding: Pre-training on frequently used address and product databases to generate address semantic vectors (address_embedding) and product semantic vectors (product_embedding), supporting fast retrieval of related information.

[0056] Context fusion includes: Feature concatenation: concatenating table_embedding, user_embedding, address_embedding, product_embedding, and scene_label (one-hot encoding) into the original feature vector raw_context. Scene enhancement encoding: performing dimensionality reduction and semantic enhancement on raw_context using a fully connected neural network (FCN) to generate the final context vector context_vector, which contains comprehensive information about table content, user habits, and scene semantics.

[0057] S103. The context information is processed using a preset intelligent recommendation model. The intelligent recommendation model dynamically calculates the weights of different information sources to the current recommendation scenario by introducing an attention mechanism. Based on the weights, information related to the missing fields in the table is extracted from each information source, and candidate recommendation content for the missing fields is generated.

[0058] Specifically, the intelligent recommendation model dynamically calculates the weights of multi-source information, focuses on the most relevant information source in the current scenario, and generates recommended candidate content for missing fields.

[0059] The intelligent recommendation model architecture includes: Model structure: It adopts an "encoder-attention layer-decoder" architecture. The encoder is a multi-layer Transformer or CNN used to extract deep features from the context vector. The attention layer supports multi-head attention, calculating the contribution weights of different information sources (tabular data, user history, knowledge base) to the missing field. Dynamic weight calculation includes: Attention mechanism formula: Attention(Q,K,V)=softmax(QK) T / dk^0.5)V; where the query vector Q is the semantic representation of the missing field (such as the pre-trained vector of "receiving address"), and the key K and value V are feature vectors from table data, user history, and knowledge base, respectively. The information source weight allocation dynamically adjusts the attention focus according to the scenario label (e.g., in the B2B scenario, the weight of the enterprise address database in the knowledge base is increased; in the return and exchange scenario, the weight of the after-sales address in the user history is prioritized).

[0060] Candidate content generation includes: Missing field classification: grouping missing_fields by type (e.g., address, product, logistics) and generating recommendations accordingly. Recommendation strategies include: Address field: retrieving highly matching addresses from a frequently used address database based on attention weights (e.g., filtering the address database based on scenario tags and then sorting by user's historical usage frequency). Product field: recommending products based on associations from the product database (e.g., recommending previously purchased models of the same brand if "phone" is already filled in, or recommending best-selling phone SKUs from the knowledge base). Logical reasoning: for fields such as "logistics method," combining order weight (if already filled in), user historical preferences, and address accessibility (logistics coverage area in the knowledge base) for comprehensive inference. Candidate set filtering: filtering low-confidence recommendations (e.g., candidate content with attention weight < 0.3) using thresholds, generating a Top-N recommendation list (e.g., providing 3 candidate values ​​for each missing field).

[0061] S104. Display recommended candidate content to the user and receive feedback from the user on the recommended candidate content; update the parameters of the intelligent recommendation model based on the feedback information and optimize the intelligent recommendation model.

[0062] Specifically, by collecting user feedback on recommended content, a closed loop of "recommendation-feedback-iteration" is constructed to continuously optimize the model's recommendation strategy.

[0063] Feedback information collection includes: Interaction behavior classification: recording user actions on recommended content (such as "accept", "modify", "ignore", "delete"), generating a feedback log (feedback_log) containing timestamps, recommended fields, candidate content, user action type, and modified value (if any). Quantifying feedback weights: assigning weights to different feedback behaviors (e.g., "accept" weight +1, "modify" weight +0.5, "ignore" weight -0.2) to adjust the loss function during model training.

[0064] Model parameter updates include: Supervised learning updates: The user's final input (regardless of whether the recommendation is adopted) is used as a label to construct training samples with the model's recommendation output. Gradient descent is used to optimize the weight parameters of the attention mechanism and the decoder network. Reinforcement learning optimization: A reward function is introduced (e.g., +1 for adopting a recommendation, -1 for frequent ignoring), and deep reinforcement learning (e.g., DQN, PPO) is used to adjust the recommendation strategy and improve long-term recommendation accuracy.

[0065] The incremental learning mechanism includes: Real-time incremental training: A feedback information buffer pool is set up. When a certain number of feedback samples are accumulated (e.g., 50 per user), a micro-update of the model is triggered to avoid frequent full training affecting performance. Cold start handling: For new users or new scenarios, a static knowledge base and general recommendation strategies are used first, and a personalized model is gradually built through feedback information.

[0066] In some embodiments, parsing the fields in the shipping information table and extracting the filled-in structured data includes: identifying the column headers of the shipping information table according to preset field rules to determine the target fields corresponding to the shipping information and the goods information; for each target field, extracting the specific content of the filled cells in the corresponding column and marking the missing fields that are not filled in; performing format validation on the specific content, identifying data that conforms to the preset format specifications and converting it into a structured format, and marking data with incorrect format.

[0067] By parsing the table column headers using preset field rules, identifying core business fields (shipping information, cargo information), extracting valid data, and marking formatting issues, structured input is provided for subsequent recommendations.

[0068] Field rules are preset and column headers are identified by establishing a field mapping dictionary (e.g., {"Recipient Name": "recipient_name", "Product Name": "product_name"}) to support fuzzy matching (e.g., both "Shipping Address" and "Packaging Address" are mapped to "address"). Regular expressions or string matching algorithms (e.g., Levenshtein distance) are used to validate table column headers, and columns that match successfully are marked as target fields (e.g., sender information includes "sender's phone number" and "sender's address", and goods information includes "product SKU" and "quantity").

[0069] Content extraction and missing cell marking are performed by traversing the columns corresponding to the target field, extracting the content of non-empty cells row by row, and storing it in a temporary dictionary {field name: [value1, value2, ...]}; empty cells are marked as <missing>, and a list of missing fields missing_fields is generated (e.g., if "recipient's phone number" is empty in a row, then this field in that row is marked as missing).

[0070] Format validation and structured conversion involve validating the extracted content (e.g., mobile phone numbers must match the regular expression ^1[3-9]\d{9}$, and addresses must include province, city, and district information). Data that conforms to the specifications is converted into a unified format (e.g., addresses are uniformly formatted as "province-city-district-street-house number"). Data with incorrect formats is marked as <format error> and the error type is recorded (e.g., "insufficient number of digits in mobile phone number"). Such missing or incorrect fields will be prioritized for supplementation during subsequent recommendations.

[0071] In some embodiments, the collection of user behavior data and static knowledge base information formed from users' historical order records includes: extracting complete data of historical orders from the user's historical order database, including the address entered, the selected product, the operation time, and the modification record, to form a user behavior dataset; and obtaining common address and product database data from a preset static knowledge base.

[0072] By collecting users' historical operation trajectories and standardized basic data, a personalized preference library and a general knowledge base are constructed to provide historical behavioral basis and standard data support for recommendations.

[0073] The user behavior dataset was constructed by querying users' historical orders from the order database over the past 12 months. Fields extracted included: information entered (shipping address, selected products (SKU, name, quantity), shipping method, and order type (tag); and action information (order time, number of modifications, and modified fields, such as changing "regular express" to "SF Express" in a particular modification). The data was arranged chronologically to form a time-series behavior sequence, user_action_sequence, used to analyze user input habits (such as whether they frequently modify their address or prefer fixed product combinations).

[0074] Static knowledge base acquisition includes: Commonly used address database: storing standard enterprise shipping addresses (including warehouse addresses and after-sales addresses), user-defined commonly used addresses (such as home and company addresses), each address is associated with scenario tags (such as "B2C retail address" and "B2B bulk shipping address") and usage frequency count; Product database data: containing basic product attributes (SKU, name, category, weight, size), associated attributes (such as "mobile phone" associated with "charger" and "screen protector" as package recommendations), supporting fuzzy search by SKU or name.

[0075] In some embodiments, the process of fusing the parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation to generate context information for the current recommendation scenario includes: associating the filled fields in the structured data with the historical filling habits in the user behavior data; filtering the address templates and product attributes in the static knowledge base in conjunction with the real-time context information of the current operation; and integrating the address templates, product attributes, and associated filled fields with the user behavior data by field dimension to form context information that includes the user's historical preferences, current scenario features, and existing filled content.

[0076] By associating the filled-in content of the form, user history preferences, static knowledge, and real-time scenarios, irrelevant information is filtered out to form contextual features that focus on the current operation, thus solving the problem of mismatched recommendation scenarios.

[0077] The system links filled fields with historical habits. By querying the user behavior data for filled fields in the structured data (such as "return and exchange" for "order type"), the system can generate a preference feature vector (such as whether the after-sales address is used first when returning or exchanging goods, and whether "quality problem" is always filled in as the reason).

[0078] The static knowledge base filtering uses real-time context (such as the scenario tag "B2B contract delivery") to filter out entries marked "company address" from the common address database and "high-frequency bulk purchase products" from the product database (such as filtering by category + purchase quantity threshold) to reduce interference from irrelevant data.

[0079] Field dimension integration establishes a feature matrix based on field type (address, product, logistics), with each row corresponding to one field and columns containing: the content already filled in the table (e.g., "receiver address" is filled with "XX province");

[0080] The user's history of high-frequency values ​​for this field (e.g., the high frequency of "recipient name" in the user's history is "Zhang San"); the static knowledge base recommendation values ​​(e.g., address templates filtered according to the scenario); finally forming context information_matrix containing 3 dimensions (current input, historical preferences, and knowledge base standards).

[0081] In some embodiments, the step of dynamically calculating the weights of different information sources for the current recommendation scenario by introducing an attention mechanism includes: inputting structured data, user behavior data, static knowledge base information, and real-time context information as multimodal inputs into the attention mechanism module of the intelligent recommendation model; for each missing field, calculating the contribution of each information source to the recommendation of the missing field according to the field type and current scenario characteristics; generating the weight value of the corresponding information source according to the contribution, so that the intelligent recommendation model prioritizes information with higher relevance to the current field.

[0082] By quantifying the importance of multi-source information to the current field through an attention mechanism, the model can focus on highly relevant information sources, thus solving the problem of insufficient accuracy in traditional fixed-strategy recommendations.

[0083] The multimodal input and module architecture includes: taking structured data (table features), user behavior data (time-series vectors), static knowledge base (address / product embedding), and real-time scene (one-hot tag encoding) as inputs to the multi-head attention module, and converting each information source into a vector with a unified dimension (e.g., 128 dimensions).

[0084] Field type and scene-aware calculation first determines the type of the missing field (e.g., "recipient's phone number" belongs to contact information, "product SKU" belongs to goods information), and combines it with scene tags (e.g., "B2C retail" relies more on the user's historical delivery address, "B2B" relies more on the knowledge base enterprise address), generating a query vector Q_field = [field_type, scene_label]; then, the similarity between Q_field and the key vectors K_i of each information source is calculated using the dot product: Similarity(Q,Ki)=Q*K i T / dk^0.5; where d_k is the dimension of the key vector, and the similarity is processed by Softmax to obtain the weight w_i of each information source (e.g., user behavior data weight 0.6, static knowledge base weight 0.3, table data weight 0.1).

[0085] The dynamic weight allocation automatically increases the weight of users' historical behavior in high-frequency scenarios (such as "regular restocking" scenarios where users place repeated orders every month), and increases the weight of static knowledge base in low-frequency scenarios (such as first-time B2B orders), balancing personalization and universality.

[0086] In some embodiments, the step of extracting information related to the missing field in the table from each information source based on weight includes: classifying and filtering the multimodal information sources according to the calculated weights, and prioritizing the extraction of data that matches the missing field from high-weight information sources; wherein, for the recipient's phone number field, if the current scenario is a user's historical high-frequency delivery scenario, the historically associated phone number information of the corresponding recipient is first extracted from the user behavior data, and then the phone number information corresponding to the recipient with the same name is matched from the static address database.

[0087] Based on attention weight, highly relevant information is extracted first, and a differentiated extraction strategy is designed in combination with field characteristics to achieve accurate information retrieval (such as prioritizing historical records when matching recipient's phone number).

[0088] The hierarchical filtering strategy sorts information sources by weight from high to low (e.g., user behavior data > static address database > data already filled in the table), and retrieves matching data in sequence: high weight source (weight ≥ 0.5): full retrieval and retention of matching results; medium weight source (0.3 ≤ weight < 0.5): retrieval of the top 5 high-frequency records; low weight source (weight < 0.3): only used as supplementary candidates.

[0089] Customized field extraction (taking recipient's phone number as an example): If the current scenario is a high-frequency shipping scenario in the user's history (e.g., the recipient's address appears ≥3 times in the past 30 days), firstly, extract the historical phone information corresponding to the recipient from the user behavior data (there may be multiple, such as home phone number and office phone number); if there is no matching phone number in the historical data, then search for records of recipients with the same name in the static address database, and prioritize the entries marked as "frequently used contact phone numbers"; if there are still no results, trigger the fallback strategy (e.g., prompt the user to fill in manually, or recommend the default contact phone number format).

[0090] In some embodiments, generating recommended candidate content for missing fields includes: deduplicating and prioritizing the extracted multi-source information, generating at least one recommended candidate content according to weight from high to low; performing correlation verification on the recommended content by combining semantic context reasoning, filtering out obviously mismatched information; and encapsulating the verified recommended candidate content in a structured form to form an interactive recommendation result.

[0091] By deduplicating, sorting, and semantically verifying information extracted from multiple sources, structured recommendation results are generated to ensure that the recommended content conforms to business logic and contextual relationships.

[0092] Deduplication and priority sorting use a hash table to deduplicat candidate values ​​extracted from multiple sources (e.g., different information sources extract the same address), sort by weight product (information source weight × frequency of candidate value within the source), and retain the top-3 candidates.

[0093] Semantic relevance verification is based on domain knowledge graphs (e.g., "recipient address" and "recipient phone number" must belong to the same region, and "product SKU" and "product name" must be semantically consistent). The logical consistency of candidate content is verified through rule engines or pre-trained language models (e.g., BERT). For example, the area code of the recommended "recipient phone number" must match the province of the "recipient address" (the area code for Beijing addresses is 010); the recommended "product SKU" must exist in the product database, and the category must match the filled "order type" (e.g., electronic product SKUs are not recommended for "fresh food" orders).

[0094] Structured encapsulation and interaction design encapsulate validated candidate content into weighted structured objects (e.g., {"field": "recipient_phone", "candidates": [{"value": "138XXXX1234", "score": 0.9}, ...]}); these are then displayed in the interactive interface according to priority, with high-scoring candidates placed first, and their recommendation sources are labeled (e.g., "historically used" or "company standard address"), making it easy for users to quickly identify them.

[0095] In some embodiments, updating the parameters of the intelligent recommendation model based on feedback information and optimizing the intelligent recommendation model includes: receiving user feedback information on recommended candidate content, recording the field type, recommended content, and operation time corresponding to the feedback behavior in the feedback information; inputting the feedback information and the corresponding field type, recommended content, and operation time into the model training module, and adjusting the weight calculation logic of the attention mechanism through incremental learning; reducing the weight of the information source corresponding to the recommended content in the field type recommendation for recommended content rejected by the user or modified multiple times; and increasing the weight of the corresponding information source for recommended content that the user has not modified or has modified only once, thereby achieving adaptive optimization of the recommendation model.

[0096] By quantifying the recommendation effect through user interaction feedback, the weight of information sources and model strategies are dynamically adjusted to form a closed loop of "recommendation-feedback-optimization" and improve the long-term recommendation accuracy.

[0097] Feedback information is collected and recorded by recording three types of feedback behaviors: Acceptance: The user directly selects a recommended candidate value (marked as action=accept); Modification: The user modifies the recommended value (recording the content before and after modification, marked as action=modify); Ignore: The user does not select any recommended value and manually enters new content (marked as action=ignore). Feedback log fields include: field type, recommended content, operation time, user ID, modified value (if necessary), operation IP, etc., for subsequent analysis.

[0098] Incremental learning and weight adjustment include: Model training module: converting feedback information into training samples (input is the context vector during recommendation, and labels are the user's final input values), and using stochastic gradient descent (SGD) to update the weight matrix in the attention mechanism; Dynamic weight adjustment rules: if a recommendation is "ignored" ≥3 times, the weight of the corresponding information source in that field type is reduced by 20%; if the "adopted" rate is ≥60%, the weight of the corresponding information source is increased by 15%; "modification" behavior is considered partial approval, and the weight is fine-tuned (+5%), while recording the modification pattern (such as the user's habit of adding a house number after the recommended address) to optimize the details of subsequent recommendations.

[0099] Cold start scenario adaptation is performed for new users or scenarios without historical behavior data. The initial weight is biased towards the static knowledge base (70% weight). As feedback information accumulates, the weight of user behavior data is gradually increased (the weight increases by 5% for every 10 new feedback items).

[0100] In some embodiments, before processing the context information using a preset intelligent recommendation model, the method further includes: constructing an intelligent recommendation model with an attention mechanism, wherein the input layer of the intelligent recommendation model receives context information fused from multimodal data, and the output layer provides candidate recommendation content for each missing field; training the model using historical order data, user behavior logs, and labeled recommendation scenario data, and adjusting model parameters by comparing the matching degree between the recommendation results and the actual filled content; and generating sub-models for different business scenarios during the training process, enabling the intelligent recommendation model to automatically switch recommendation strategies based on real-time context.

[0101] By building an attention mechanism model that supports multiple scenarios, improving generalization ability through training with historical data, generating sub-models to adapt to different business scenarios, and realizing dynamic switching of recommendation strategies.

[0102] The model architecture design includes: Input layer: receiving the context vector after multimodal fusion (dimension is configurable, such as 512 dimensions); Core layer: containing a multi-head attention module (such as 8 heads) and a feedforward neural network (FFN) for feature weighting and semantic enhancement; Output layer: for each missing field, outputting the probability distribution of candidate content (such as Softmax outputting the probability values ​​of Top-3 candidates).

[0103] The training data and optimization objectives use historical order data (annotated with the actual values ​​entered), user behavior logs (recording recommendation interactions), and scene label data (manually annotated or automatically generated) as the training set; the loss function combines cross-entropy loss (the degree of matching between the recommended value and the actual value) and reinforcement learning reward (the weight of the feedback behavior), and the formula is: Loss=α*CrossEntropy+(1-α)*ReinforcementReward; where α is the balancing coefficient (default 0.7).

[0104] Multi-scenario sub-model generation splits training data by order type (B2C, B2B, returns and exchanges), industry (fresh food, 3C, apparel), etc., and uses transfer learning to train sub-models (sharing the underlying attention module and fine-tuning the upper output layer); when the model is deployed, the corresponding sub-model is automatically loaded based on real-time scenario labels (inferred from the "order type" field in the table or address keywords), and the recommendation strategy is switched (e.g., cold chain logistics is recommended first in fresh food scenarios, and insurance services are recommended first in 3C scenarios).

[0105] In some embodiments, by modeling the recommendation of missing fields as a sequential decision problem, reinforcement learning (RL) is used to dynamically optimize the recommendation order and information source combination strategy to solve the recommendation efficiency problem in multi-field dependent scenarios (such as whether to prioritize recommending "recipient's phone number" to assist in inferring the address when "recipient address" is not filled in).

[0106] The state space definition includes: State S_t contains: the vector of currently filled fields F_t (e.g., [filled field 1: value 1, filled field 2: value 2]); the set of fields to be recommended M_t (e.g., {"receiver address", "receiver's phone number"}); the historical recommendation action sequence A_1^t (recording the recommended fields and content); and the real-time scene tag C (e.g., "cross-border e-commerce" "urgent order").

[0107] The action space design includes: Action A_t includes: selecting the next recommended field (based on field dependencies, such as prioritizing recommendations when "address" and "phone" are strongly correlated); selecting an information source combination strategy (such as joint retrieval of "user behavior data + static knowledge base"); and generating a recommendation candidate set size k (recommending 1 accurate value in high-frequency scenarios and 3 candidates in low-frequency scenarios).

[0108] The reward function design includes:

[0109] Immediate rewards: +10 points for user adoption of a recommendation, +5 points for modification, -5 points for ignoring; Long-term rewards: Reduced field completion time (+0.1 points for every second saved), improved accuracy of subsequent field recommendations (based on state transition probability); Constraint penalties: -20 points for recommending duplicate content or data with incorrect formatting.

[0110] Model training and policy execution utilize Deep Q-Network (DQN) or Policy Gradient (PPO) algorithms, with historical interaction logs as the initial experience pool, and balance exploration and utilization through an ε-greedy policy. For the "address-phone-zip code" strongly dependent field group, the training model prioritizes recommending "address" to reduce ambiguity in subsequent field inference. In cross-border scenarios, the "country / region" field is prioritized to lock the address format.

[0111] In some embodiments, by constructing a knowledge graph in the logistics domain, the entity relationships such as "address-recipient-goods-logistics rules" are structured, and the semantic reasoning ability is enhanced by graph neural networks (GNNs) to solve implicit association recommendations in complex business scenarios (such as recommending matching logistics channels based on the weight of goods).

[0112] The knowledge graph construction includes: entity types: address (province / city / district / street), recipient (enterprise / individual), product (SKU / category / weight / size), logistics rules (volume weight calculation, prohibited items list), and scenario tags (e-commerce / contract / sample); relationship definitions: address-recipient: "frequently used address" and "associated phone number"; product-logistics rules: "weight ≤ 1kg → recommend economy express delivery" and "contains lithium batteries → air freight prohibited"; scenario-address: "cross-border e-commerce → must include customs code".

[0113] The graph embedding and reasoning process uses TransE or GraphSAGE algorithms to embed entities and relations into vectors, generating a graph knowledge base (KG). When recommending the "Logistics Method" field: ① Extract the filled fields "Product Weight = 2.5kg" and "Recipient Address = **"; ② Query the logistics rules associated with "Weight 2.5kg + ** Address" in the KG, and filter out channels that "support international shipping and have a weight ≤ 3kg" (such as DHL and FedEx); ③ Combine user historical preferences (such as users choosing DHL for 90% of international shipments), prioritize recommending DHL and mark it with a prohibited shipping risk warning.

[0114] Anomaly detection and conflict resolution are achieved through graph logic verification of recommended content: for example, if there is an air transport restriction conflict between "recipient address = Xinjiang" and "product = lithium battery", the land transport channel is automatically replaced and the missing product attribute is marked ("transportation mode preference" needs to be added); potential associations are discovered through graph path search (e.g., "recipient company name = XX Technology" → associated with its historically frequently used warehouse addresses, improving the efficiency of enterprise customer recommendations).

[0115] In some embodiments, meta-learning is introduced to quickly adapt to new users / new scenarios (without historical behavior data), and personalized recommendation strategies are generated using limited sample interaction data, thus solving the problem of low recommendation accuracy in traditional models during cold starts.

[0116] The meta-learning framework is designed using a model-agnostic meta-learning (MAML) architecture, which includes: a base model with the same structure as the original intelligent recommendation model (including an attention mechanism); and a meta-trainer that updates the model parameters through a small number of tasks (such as the first 3 order records of different users) to enable it to quickly adapt to new tasks.

[0117] The meta-training data construction simulates a cold start scenario by sampling "user-scenario" pairs from historical data: for user A, the first two orders are retained as the "support set" and subsequent orders are retained as the "query set"; the training objective is to enable the model to predict the missing fields of the user's third order using the two order data.

[0118] The cold start recommendation process includes: When a new user places their first order: ① Collect some fields filled in during the first 3 times (such as "shipping address" and "product name" filled in for the first time) as a support set; ② The meta trainer quickly fine-tunes the basic model parameters (1-2 gradient updates) to generate a temporary personalized model; ③ The temporary model recommends the remaining fields based on the static knowledge base and a small amount of filled content (such as recommending "price protection service" and "cardboard packaging" based on "product name = laptop" and historical data of similar users); As the interaction data increases (≥5 times), gradually switch back to the main model, and at the same time add new user data to the meta training pool for continuous optimization.

[0119] In some embodiments, by aggregating shipping data from multiple enterprises / departments through federated learning while protecting user privacy and data compliance, a cross-domain recommendation model is constructed to solve the problem of data sparsity for a single enterprise (such as small and medium-sized merchants jointly optimizing recommendation strategies).

[0120] The federated learning architecture adopts a layered federated architecture: Client: Each enterprise deploys sub-models locally and stores anonymized user behavior data (such as regional codes with specific addresses hidden and product category statistics); Coordination server: Aggregates model parameters (gradients or weights) from each client and does not directly access the original data; Global model: Iteratively updated through the FedAvg algorithm, and the updated model is periodically distributed to the clients.

[0121] Data anonymization and security mechanisms include k-anonymization of address fields (e.g., "XX Province XX City" is retained at the city level, while districts / streets are hidden); homomorphic encryption technology is used for product data, sharing only category-level association rules (e.g., the co-occurrence probability of "fresh produce → insulated box packaging"); and differential privacy (DP) is introduced to add noise, ensuring that aggregation parameters cannot be used to deduce the original data.

[0122] When a business user fills in "Product Category = Fresh Produce": ① The local model combines cross-domain knowledge learned from the global model (in the fresh produce scenario, 58% of businesses recommend the "ice pack + insulated box" packaging combination); ② The model integrates the business's own packaging inventory data (prioritizing the recommendation of existing insulated box models); ③ The model generates recommended content that takes into account both industry commonalities and business characteristics, while avoiding the disclosure of other businesses' specific inventory information.

[0123] In some embodiments, by breaking away from the pure form-filling mode, natural language interaction (voice / text) and visual guidance are introduced, and a multimodal fusion model is used to understand user intent and dynamically generate guided recommendations to improve filling efficiency in complex scenarios (such as automatically suggesting historical addresses when the user says "send the address from last time").

[0124] Multimodal input processing includes: text / speech parsing: using BERT-NER to extract entities from user input (e.g., mapping "last address" to historical address ID, and "urgent" to "express type = SF Express"); visual interaction: adding a quick input panel next to the table, supporting operations such as clicking on historical address cards, dragging and dropping product lists to the table, and sliding to select logistics timeliness, generating interactive behavior vectors I_v.

[0125] Intent understanding and modality fusion are achieved by constructing a multimodal fusion model (such as the Transformer encoder). The inputs include: tabular structured data F_s; natural language intent vector I_t (generated through a text embedding model); visual interaction vector I_v (generated through operation sequence encoding); and attention mechanism is used to calculate the contribution of each modality to the missing field (e.g., when the voice input is "send to company", the language modality weight of the "receiver address" field is increased to 0.8).

[0126] Guided recommendation generation works as follows: For fuzzy input (e.g., a user enters "the usual"): ① Query the user's filling patterns from the last 3 similar orders and generate a "default filling template"; ② Highlight the recommended field in the table and label it "Recommended based on your historical habits," allowing one-click application of the template; For conflicting input (e.g., a user says "send to Beijing" but the table already has a "Shanghai address"): ① Trigger an ambiguity resolution mechanism, prompting a pop-up window "Address conflict detected, do you want to use the historical Beijing address?", and recommend the optimal solution based on historical modification records.

[0127] This application discloses an AI-based intelligent recommendation method and device for frequently used shipping information. The method proactively recommends missing content by parsing table fields in real time, eliminating the need for users to manually fill in or repeatedly modify tables, significantly reducing the number of re-uploads and shortening order processing time. Based on multimodal data fusion and attention mechanisms, the system dynamically adjusts its recommendation strategy according to the current operational scenario (such as order type and existing content). For example, it intelligently associates frequently used addresses and phone numbers with recipient names, resulting in recommendations that better match actual needs and reduce human error rates. The method continuously optimizes the recommendation model based on user feedback, prioritizing frequently used addresses and products. Recommendation accuracy gradually improves with usage time, creating a personalized experience that becomes increasingly intelligent with use. The generated high-quality structured data lays the foundation for subsequent intelligent order allocation and supply chain optimization, enhancing the overall intelligence level of the business process.

[0128] Please see Figure 2 , Figure 2 This application also provides a schematic block diagram of an AI-based intelligent recommendation device for frequently used shipping information. This AI-based intelligent recommendation device 200 is used to execute the aforementioned AI-based intelligent recommendation method for frequently used shipping information. The AI-based intelligent recommendation device for frequently used shipping information can be configured in a server or terminal.

[0129] The server can be a standalone server, a server cluster, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal can be an electronic device such as a mobile phone, tablet, laptop, desktop computer, user digital assistant, or wearable device.

[0130] like Figure 2 As shown, the AI-based intelligent recommendation device 200 for frequently used shipping information includes:

[0131] The data extraction unit 201 is used to obtain the shipping information table input by the user, parse the fields in the shipping information table, and extract the structured data that has been filled in; collect user behavior data and static knowledge base information formed by the user's historical order records, and the static knowledge base information includes at least common addresses and product database data;

[0132] The information generation unit 202 is used to fuse the parsed structured data, user behavior data, static knowledge base information and real-time context information of the current operation to generate context information for the current recommendation scenario.

[0133] The information processing unit 203 is used to process context information using a preset intelligent recommendation model. The intelligent recommendation model dynamically calculates the weights of different information sources to the current recommendation scenario by introducing an attention mechanism, and extracts information related to the missing fields in the table from each information source based on the weights to generate recommendation candidate content for the missing fields.

[0134] The intelligent recommendation unit 204 is used to display recommended candidate content to the user, receive user feedback on the recommended candidate content, and update the parameters of the intelligent recommendation model based on the feedback information to optimize the intelligent recommendation model.

[0135] In some embodiments, parsing the fields in the shipping information table and extracting the filled-in structured data includes: identifying the column headers of the shipping information table according to preset field rules to determine the target fields corresponding to the shipping information and the goods information; for each target field, extracting the specific content of the filled cells in the corresponding column and marking the missing fields that are not filled in; performing format validation on the specific content, identifying data that conforms to the preset format specifications and converting it into a structured format, and marking data with incorrect format.

[0136] In some embodiments, the collection of user behavior data and static knowledge base information formed from users' historical order records includes: extracting complete data of historical orders from the user's historical order database, including the address entered, the selected product, the operation time, and the modification record, to form a user behavior dataset; and obtaining common address and product database data from a preset static knowledge base.

[0137] In some embodiments, the process of fusing the parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation to generate context information for the current recommendation scenario includes: associating the filled fields in the structured data with the historical filling habits in the user behavior data; filtering the address templates and product attributes in the static knowledge base in conjunction with the real-time context information of the current operation; and integrating the address templates, product attributes, and associated filled fields with the user behavior data by field dimension to form context information that includes the user's historical preferences, current scenario features, and existing filled content.

[0138] In some embodiments, the step of dynamically calculating the weights of different information sources for the current recommendation scenario by introducing an attention mechanism includes: inputting structured data, user behavior data, static knowledge base information, and real-time context information as multimodal inputs into the attention mechanism module of the intelligent recommendation model; for each missing field, calculating the contribution of each information source to the recommendation of the missing field according to the field type and current scenario characteristics; generating the weight value of the corresponding information source according to the contribution, so that the intelligent recommendation model prioritizes information with higher relevance to the current field.

[0139] In some embodiments, the step of extracting information related to the missing field in the table from each information source based on weight includes: classifying and filtering the multimodal information sources according to the calculated weights, and prioritizing the extraction of data that matches the missing field from high-weight information sources; wherein, for the recipient's phone number field, if the current scenario is a user's historical high-frequency delivery scenario, the historically associated phone number information of the corresponding recipient is first extracted from the user behavior data, and then the phone number information corresponding to the recipient with the same name is matched from the static address database.

[0140] In some embodiments, generating recommended candidate content for missing fields includes: deduplicating and prioritizing the extracted multi-source information, generating at least one recommended candidate content according to weight from high to low; performing correlation verification on the recommended content by combining semantic context reasoning, filtering out obviously mismatched information; and encapsulating the verified recommended candidate content in a structured form to form an interactive recommendation result.

[0141] In some embodiments, updating the parameters of the intelligent recommendation model based on feedback information and optimizing the intelligent recommendation model includes: receiving user feedback information on recommended candidate content, recording the field type, recommended content, and operation time corresponding to the feedback behavior in the feedback information; inputting the feedback information and the corresponding field type, recommended content, and operation time into the model training module, and adjusting the weight calculation logic of the attention mechanism through incremental learning; reducing the weight of the information source corresponding to the recommended content in the field type recommendation for recommended content rejected by the user or modified multiple times; and increasing the weight of the corresponding information source for recommended content that the user has not modified or has modified only once, thereby achieving adaptive optimization of the recommendation model.

[0142] In some embodiments, before processing the context information using a preset intelligent recommendation model, the method further includes: constructing an intelligent recommendation model with an attention mechanism, wherein the input layer of the intelligent recommendation model receives context information fused from multimodal data, and the output layer provides candidate recommendation content for each missing field; training the model using historical order data, user behavior logs, and labeled recommendation scenario data, and adjusting model parameters by comparing the matching degree between the recommendation results and the actual filled content; and generating sub-models for different business scenarios during the training process, enabling the intelligent recommendation model to automatically switch recommendation strategies based on real-time context.

[0143] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the model training device and each module described above can be referred to the corresponding processes in the aforementioned AI-based intelligent recommendation method for common shipping information, and will not be repeated here.

[0144] The aforementioned AI-based intelligent recommendation device for frequently used shipping information can be implemented as a computer program, which can, for example... Figure 3 It runs on the computer device shown.

[0145] Please see Figure 3 , Figure 3 This is a schematic block diagram illustrating the structure of a computer device according to an embodiment of this application. The computer device may be a server or a terminal.

[0146] See Figure 3 The computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include storage media and internal memory.

[0147] The storage medium may store an operating system and a computer program. The computer program includes program instructions that, when executed, cause the processor to perform any of the AI-based intelligent recommendation methods for frequently used shipping information provided in the embodiments of this application.

[0148] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0149] Internal memory provides an environment for the execution of computer programs stored in the storage medium. When executed by a processor, this program allows the processor to perform any AI-based intelligent recommendation method for frequently used shipping information. The storage medium can be non-volatile or volatile.

[0150] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0151] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0152] For example, in one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps:

[0153] Obtain the shipping information form entered by the user, parse the fields in the shipping information form, and extract the entered structured data; collect user behavior data and static knowledge base information formed from the user's historical order records. The static knowledge base information includes at least commonly used addresses and product database data.

[0154] The parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation are fused together to generate context information for the current recommendation scenario.

[0155] The context information is processed using a preset intelligent recommendation model. The intelligent recommendation model dynamically calculates the weight of different information sources to the current recommendation scenario by introducing an attention mechanism. Based on the weight, information related to the missing fields in the table is extracted from each information source, and recommendation candidate content for the missing fields is generated.

[0156] Display recommended candidate content to users and receive user feedback on the recommended candidate content; update the parameters of the intelligent recommendation model based on the feedback information and optimize the intelligent recommendation model.

[0157] In some embodiments, parsing the fields in the shipping information table and extracting the filled-in structured data includes: identifying the column headers of the shipping information table according to preset field rules to determine the target fields corresponding to the shipping information and the goods information; for each target field, extracting the specific content of the filled cells in the corresponding column and marking the missing fields that are not filled in; performing format validation on the specific content, identifying data that conforms to the preset format specifications and converting it into a structured format, and marking data with incorrect format.

[0158] In some embodiments, the collection of user behavior data and static knowledge base information formed from users' historical order records includes: extracting complete data of historical orders from the user's historical order database, including the address entered, the selected product, the operation time, and the modification record, to form a user behavior dataset; and obtaining common address and product database data from a preset static knowledge base.

[0159] In some embodiments, the process of fusing the parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation to generate context information for the current recommendation scenario includes: associating the filled fields in the structured data with the historical filling habits in the user behavior data; filtering the address templates and product attributes in the static knowledge base in conjunction with the real-time context information of the current operation; and integrating the address templates, product attributes, and associated filled fields with the user behavior data by field dimension to form context information that includes the user's historical preferences, current scenario features, and existing filled content.

[0160] In some embodiments, the step of dynamically calculating the weights of different information sources for the current recommendation scenario by introducing an attention mechanism includes: inputting structured data, user behavior data, static knowledge base information, and real-time context information as multimodal inputs into the attention mechanism module of the intelligent recommendation model; for each missing field, calculating the contribution of each information source to the recommendation of the missing field according to the field type and current scenario characteristics; generating the weight value of the corresponding information source according to the contribution, so that the intelligent recommendation model prioritizes information with higher relevance to the current field.

[0161] In some embodiments, the step of extracting information related to the missing field in the table from each information source based on weight includes: classifying and filtering the multimodal information sources according to the calculated weights, and prioritizing the extraction of data that matches the missing field from high-weight information sources; wherein, for the recipient's phone number field, if the current scenario is a user's historical high-frequency delivery scenario, the historically associated phone number information of the corresponding recipient is first extracted from the user behavior data, and then the phone number information corresponding to the recipient with the same name is matched from the static address database.

[0162] In some embodiments, generating recommended candidate content for missing fields includes: deduplicating and prioritizing the extracted multi-source information, generating at least one recommended candidate content according to weight from high to low; performing correlation verification on the recommended content by combining semantic context reasoning, filtering out obviously mismatched information; and encapsulating the verified recommended candidate content in a structured form to form an interactive recommendation result.

[0163] In some embodiments, updating the parameters of the intelligent recommendation model based on feedback information and optimizing the intelligent recommendation model includes: receiving user feedback information on recommended candidate content, recording the field type, recommended content, and operation time corresponding to the feedback behavior in the feedback information; inputting the feedback information and the corresponding field type, recommended content, and operation time into the model training module, and adjusting the weight calculation logic of the attention mechanism through incremental learning; reducing the weight of the information source corresponding to the recommended content in the field type recommendation for recommended content rejected by the user or modified multiple times; and increasing the weight of the corresponding information source for recommended content that the user has not modified or has modified only once, thereby achieving adaptive optimization of the recommendation model.

[0164] In some embodiments, before processing the context information using a preset intelligent recommendation model, the method further includes: constructing an intelligent recommendation model with an attention mechanism, wherein the input layer of the intelligent recommendation model receives context information fused from multimodal data, and the output layer provides candidate recommendation content for each missing field; training the model using historical order data, user behavior logs, and labeled recommendation scenario data, and adjusting model parameters by comparing the matching degree between the recommendation results and the actual filled content; and generating sub-models for different business scenarios during the training process, enabling the intelligent recommendation model to automatically switch recommendation strategies based on real-time context.

[0165] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the steps of the risk warning method described in the first aspect above.

[0166] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0167] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based common shipping information intelligent recommendation method, characterized in that, The application comprises the following steps: acquiring a user input shipping information form, parsing the fields in the shipping information form, and extracting filled structured data; collecting user behavior data formed by historical order records and static knowledge base information, the static knowledge base information at least including commonly used address and commodity library data; fusing the parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation to generate context information of the current recommendation scenario; processing the context information by using a preset intelligent recommendation model, the intelligent recommendation model dynamically calculating the weights of different information sources for the current recommendation scenario by introducing an attention mechanism, extracting information related to the missing fields in the form from each information source based on the weights, and generating recommended candidate contents for the missing fields; showing the recommended candidate contents to the user and receiving feedback information of the user on the recommended candidate contents; updating the parameters of the intelligent recommendation model according to the feedback information and optimizing the intelligent recommendation model.

2. The method of claim 1, wherein, The parsing of the fields in the shipping information form and the extraction of the filled structured data comprise the following steps: identifying the column titles of the shipping information form according to preset field rules to determine target fields corresponding to the sender information and the cargo information; for each target field, extracting the specific content of the filled cells in the corresponding column, marking the missing fields that are not filled, performing format verification on the specific content, identifying the data that meets the preset format specification and converting it into a structured format, and marking the data with format errors.

3. The method of claim 1, wherein, The collection of user behavior data formed by historical order records and static knowledge base information comprises the following steps: extracting complete data of historical orders from a user historical order database, including filled addresses, selected commodities, operation time, and modification records, to form a user behavior data set; obtaining commonly used address and commodity library data from a preset static knowledge base.

4. The method of claim 1, wherein, The fusion of the parsed structured data, user behavior data, static knowledge base information, and real-time context information of the current operation to generate context information of the current recommendation scenario comprises the following steps: associating the filled fields in the structured data with the historical filling habits in the user behavior data; filtering the address templates and commodity attributes in the static knowledge base in combination with the real-time context information of the current operation; integrating the address templates, commodity attributes, and associated filled fields with the user behavior data by field dimension to form context information containing user historical preferences, current scenario characteristics, and existing filled contents.

5. The method of claim 1, wherein, The dynamic calculation of the weights of different information sources for the current recommendation scenario by introducing an attention mechanism comprises the following steps: inputting the structured data, user behavior data, static knowledge base information, and real-time context information as multi-modal inputs into the attention mechanism module of the intelligent recommendation model; for each missing field, calculating the contribution degree of each information source to the recommendation of the missing field according to the field type and the current scenario characteristics; generating weight values of the corresponding information sources according to the contribution degrees, so that the intelligent recommendation model pays more attention to the information with a higher association degree with the current field.

6. The method of claim 1, wherein, The extraction of information related to the missing fields in the form from each information source based on the weights comprises the following steps: According to the calculated weight, the multi-modal information source is screened and filtered, and the data matching the missing field in the high-weight information source is preferentially extracted; for the recipient phone field, if the current scene is a user historical high-frequency delivery scene, the historical associated phone information of the corresponding recipient is first extracted from the user behavior data, and then the phone information corresponding to the recipient with the same name is matched from the static address library.

7. The method of claim 1, wherein, The recommended candidate content of the missing field is generated, including: The extracted multi-source information is de-duplicated and prioritized, and at least one recommended candidate content is generated according to the weight from high to low; In combination with semantic context reasoning, the relevance of the recommended content is verified, and the obviously unmatched information is filtered; the recommended candidate content that passes the verification is encapsulated in a structured form to form an interactive recommended result.

8. The method of claim 1, wherein, The parameters of the intelligent recommendation model are updated according to the feedback information, and the intelligent recommendation model is optimized, including: Receiving feedback information of the user on the recommended candidate content, recording the field type corresponding to the feedback behavior, the recommended content and the operation time in the feedback information; The feedback information and the corresponding field type, the recommended content and the operation time are input into the model training module, and the weight calculation logic of the attention mechanism is adjusted through incremental learning; For the recommended content that the user rejects or modifies multiple times, the weight of the information source corresponding to the recommended content in the field type recommendation is reduced; For the recommended content that the user does not modify or modifies once, the weight of the corresponding information source is increased to realize the adaptive optimization of the recommendation model.

9. The method of claim 1, wherein, Before the context information is processed by using the preset intelligent recommendation model, further comprising: An intelligent recommendation model containing an attention mechanism is constructed, the input layer of the intelligent recommendation model receives the context information after multi-modal fusion, and the output layer corresponds to the recommended candidate content of each missing field; The model is trained by using historical order data, user behavior logs and labeled recommendation scene data, and the model parameters are adjusted by comparing the matching degree of the recommended result and the actual filled content; During the training process, sub-models are generated for different business scenarios, so that the intelligent recommendation model can automatically switch the recommendation strategy according to the real-time context. 10.An AI-based common shipment information intelligent recommendation device, characterized by Including: A data extraction unit is configured to obtain a delivery information table input by a user, parse the fields in the delivery information table, and extract the structured data that has been filled in; User behavior data formed by collecting user historical order records and static knowledge base information are collected, and the static knowledge base information at least includes common address and commodity library data; An information generation unit is configured to fuse the structured data parsed, the user behavior data, the static knowledge base information and the real-time context information of the current operation to generate context information of the current recommendation scene; An information processing unit is configured to process the context information by using a preset intelligent recommendation model, the intelligent recommendation model dynamically calculates the weight of different information sources for the current recommendation scene by introducing an attention mechanism, extracts information related to the missing fields in the table from each information source based on the weight, and generates recommended candidate content of the missing field; An intelligent recommendation unit is configured to display the recommended candidate content to the user and receive feedback information of the user on the recommended candidate content. The parameters of the intelligent recommendation model are updated according to the feedback information, and the intelligent recommendation model is optimized.

Images