Common delivery information intelligent recommendation method and device based on AI

Abstract

The invention relates to the technical field of computer application, and discloses an AI-based common delivery information intelligent recommendation method and device, and the method comprises the steps: obtaining a delivery information table inputted by a user, analyzing fields in the delivery information table, and extracting filled structured data; collecting user behavior data and static knowledge base information formed by historical ordering records of a user, and generating context information of a current recommendation scene; a preset intelligent recommendation model is used for processing the context information, the intelligent recommendation model dynamically calculates the weights of different information sources for the current recommendation scene by introducing an attention mechanism, information related to missing fields in the table is extracted from all the information sources based on the weights, and recommendation candidate content of the missing fields is generated; displaying the recommended candidate content to the user, and receiving feedback information of the user on the recommended candidate content; and updating parameters of the intelligent recommendation model according to the feedback information, and optimizing the intelligent recommendation model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application, and in particular to an AI-based common delivery information intelligent recommendation method and device. BACKGROUND

[0002] In the e-commerce, logistics and supply chain management scenarios, users enter batch orders through Excel tables, which is a high-frequency operation. However, the existing systems generally have the following defects: 1. Passive lagging verification mode: Traditional systems can only perform rule verification (such as field format checking) after the user uploads the table, and can only prompt basic errors such as "missing fields" and "format errors". It cannot predict missing content or provide intelligent recommendations in real time during user filling, resulting in the need for users to repeatedly modify tables and repeatedly upload, which is low in operation efficiency.

[0003] 2. Single and limited recommendation capability: Existing recommendation functions rely on single-dimensional data (such as only user history records), and the recommendation strategy is fixed, which cannot dynamically adjust the recommended content in combination with the real-time context of the current operation (such as order type and semantic association of fields in the table). For example, traditional systems cannot distinguish between "e-commerce return and exchange" and "B2B contract delivery" scenarios, resulting in mismatched address templates or goods information and insufficient accuracy.

[0004] 3. Lack of adaptive learning mechanism: Existing systems cannot continuously optimize the recommendation model based on user feedback (such as selection, modification, and ignoring behavior) on recommended content, and the recommendation effect is difficult to improve with changes in user usage habits, and the user experience cannot form a virtuous circle.

[0005] Therefore, there is an urgent need for a method to solve at least one of the above problems. SUMMARY

[0006] The present application provides an AI-based common delivery information intelligent recommendation method and device, which is used to solve the problem that the prior art only implements passive verification based on rules, does not construct an intelligent recommendation system combining multi-source data (user behavior, static knowledge base, and real-time context), and lacks dynamic weight calculation and adaptive learning capability. The current technical solution is limited to the stage of "reporting errors after discovering problems", and has not achieved the intelligent leap of "actively predicting problems and providing solutions". It also does not form a closed loop of "data collection-intelligent recommendation-feedback optimization", resulting in cumbersome user operations, low system recommendation accuracy, and inability to meet the needs of efficient batch order processing.

[0007] In a first aspect, the present application provides an AI-based common delivery information intelligent recommendation method, which comprises: obtain a delivery information form filled by a user, parse fields in the delivery information form, and extract structured data filled by the user; collect user behavior data formed by historical order records of the user and static knowledge base information, the static knowledge base information at least including commonly used address data and commodity library data; fuse the parsed structured data, the user behavior data, the static knowledge base information, and real-time context information of a current operation, and generate context information of a current recommendation scenario; process the context information by using a preset intelligent recommendation model, the intelligent recommendation model dynamically calculates weights of different information sources on the current recommendation scenario by introducing an attention mechanism, extracts information related to a missing field in the form from the information sources based on the weights, and generates recommended candidate content of the missing field; show the recommended candidate content to the user, receive feedback information of the user on the recommended candidate content, update parameters of the intelligent recommendation model according to the feedback information, and optimize the intelligent recommendation model.

[0008] In some embodiments, the parsing of the fields in the delivery information form and the extraction of the filled structured data include: identifying column titles of the delivery information form according to a preset field rule, determining target fields corresponding to sending information and commodity information; for each target field, extracting specific content of a filled cell in a corresponding column, and marking a missing field that is not filled; performing format verification on the specific content, identifying data that meets a preset format specification and converting the data into a structured format, and marking data that has a format error.

[0009] In some embodiments, the collection of the user behavior data formed by the historical order records of the user and the static knowledge base information includes: extracting complete data of historical orders from a user historical order database, including filled addresses, selected commodities, operation times, and modification records, to form a user behavior data set; and obtaining commonly used address data and commodity library data from a preset static knowledge base.

[0010] In some embodiments, the fusion of the parsed structured data, the user behavior data, the static knowledge base information, and the real-time context information of the current operation to generate the context information of the current recommendation scenario includes: associating filled fields in the structured data with historical filling habits in the user behavior data; filtering address templates and commodity attributes in the static knowledge base in combination with real-time context information of the current operation; integrating the address templates, the commodity attributes, and the associated filled fields with the user behavior data according to a field dimension, to form context information including historical preferences of the user, characteristics of the current scenario, and filled content.

[0011] In some embodiments, the dynamically calculating the weight of different information sources to the current recommendation scene by introducing an attention mechanism comprises: inputting structured data, user behavior data, static knowledge base information and real-time context information as multi-modal input into an 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 the current scene characteristics; and generating a weight value of the corresponding information source according to the contribution, so that the intelligent recommendation model pays more attention to the information with higher relevance to the current field.

[0012] In some embodiments, the extracting information related to the missing field in the table from the information sources based on the weight comprises: performing hierarchical screening on the multi-modal information sources according to the calculated weight, and preferentially extracting data matching the missing field in the high-weight information source; and 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.

[0013] In some embodiments, the generating the recommendation candidate content of the missing field comprises: performing deduplication and priority sorting on the extracted multi-source information, and generating at least one recommendation candidate content in descending order of weight; performing relevance verification on the recommendation content in combination with semantic context reasoning, and filtering obviously unmatched information; encapsulating the recommendation candidate content that passes the verification in a structured form to form an interactive recommendation result.

[0014] In some embodiments, the updating the parameters of the intelligent recommendation model according to the feedback information to optimize the intelligent recommendation model comprises: receiving feedback information of the user on the recommendation candidate content, recording the field type corresponding to the feedback behavior, the recommendation content and the operation time in the feedback information; inputting the feedback information and the corresponding field type, the recommendation content and the operation time into the model training module, and adjusting the weight calculation logic of the attention mechanism through incremental learning; for the recommendation content that is rejected or modified multiple times by the user, reducing the weight of the corresponding information source in the field type recommendation; and for the recommendation content that is not modified or modified once by the user, increasing the weight of the corresponding information source, so as to realize the adaptive optimization of the recommendation model.

[0015] In some embodiments, before processing the context information by using the preset intelligent recommendation model, the method further comprises: constructing an intelligent recommendation model comprising an attention mechanism, wherein an input layer of the intelligent recommendation model receives the context information after multi-modal fusion, and an output layer corresponds to recommended candidate contents of each missing field; training the model by using historical order data, user behavior logs and labeled recommendation scene data, adjusting model parameters by comparing matching degrees of recommendation results and actual filled contents; and generating sub-models for different business scenes in the training process, so that the intelligent recommendation model can automatically switch recommendation strategies according to real-time context.

[0016] In a second aspect, the application provides an AI-based commonly used shipping information intelligent recommendation device, which comprises: A data extraction unit is configured to obtain a shipping information form input by a user, analyze fields in the shipping information form, and extract filled structured data; collect user behavior data formed by historical order records of the user and static knowledge base information, wherein the static knowledge base information at least comprises commonly used address and commodity library data; An information generation unit is configured to fuse the analyzed structured data, the user behavior data, the static knowledge base information and real-time context information of a current operation, and generate context information of a current recommendation scene; An information processing unit is configured to process the context information by using a preset intelligent recommendation model, wherein the intelligent recommendation model dynamically calculates weights of different information sources for the current recommendation scene by introducing an attention mechanism, extracts information related to missing fields in the form from the information sources based on the weights, and generates recommended candidate contents of the missing fields; An intelligent recommendation unit is configured to display the recommended candidate contents to the user, receive feedback information of the recommended candidate contents from the user, update parameters of the intelligent recommendation model according to the feedback information, and optimize the intelligent recommendation model.

[0017] The application discloses an AI-based commonly used shipping information intelligent recommendation method and device. The provided method actively recommends missing contents by real-time analyzing form fields, and the user does not need to manually fill in or repeatedly modify the form, thereby significantly reducing the number of repeated uploads and shortening the order processing time. The method is based on multi-modal data fusion and an attention mechanism, and the system can dynamically adjust the recommendation strategy according to the current operation scene (such as the order type and the existing filled content). For example, the system can intelligently associate the recipient name with the commonly used address and telephone number, and the recommendation result is more in line with the actual demand, thereby reducing the human error rate. The method continuously optimizes the recommendation model through user feedback, and the frequently used address, commodity and other information are preferentially recommended, the recommendation accuracy is gradually improved with the use time, and a personalized experience of “the more you use, the smarter you become” is formed. The generated high-quality structured data lays a foundation for subsequent intelligent order distribution, supply chain optimization and other links, and improves the intelligent level of the overall business process.

[0018] It should be understood that the above general description and the following detailed description are only exemplary and explanatory and are not restrictive of the application. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 is a step schematic flow chart of a commonly used delivery information intelligent recommendation method based on AI provided by the embodiments of the present application; Figure 2 is a schematic block diagram of a commonly used delivery information intelligent recommendation device based on AI provided by the embodiments of the present application; Figure 3 is a structural schematic block diagram of a computer device provided by the embodiments of the present application.

[0021] It should be understood that the above general description and the following detailed description are only exemplary and explanatory and are not restrictive of the application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0023] The flow chart shown in the drawings is only an example description, and is not necessarily to include all the contents and operations / steps, and is not necessarily to be executed in the described order. For example, some operations / steps can be decomposed, combined or partially combined, so that the actual execution order can be changed according to the actual situation.

[0024] It should be understood that the terms used in this specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification and the appended claims of the present application, unless otherwise clearly indicated by the context, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0025] It should also be understood that the term "and / or" as used herein refers to any or all possible combinations of one or more of the associated listed items, and includes all possible combinations.

[0026] Some embodiments of the present application will be described in detail with reference to the drawings. The following examples and features in the examples can be combined with each other in the case of no conflict.

[0027] In the e-commerce, logistics and supply chain management scenarios, users enter batch orders through Excel tables, which is a high-frequency operation. However, the existing systems have the following defects: 1. Passive and lagging verification method: Traditional systems can only perform rule verification (such as field format checking) after the user uploads the table, and can only prompt basic errors such as "missing field" and "format error". It cannot predict missing content or provide intelligent recommendations in real time during user filling, resulting in the need for users to repeatedly modify the table and repeatedly upload, which is low in operation efficiency.

[0028] 2. Single and limited recommendation capability: Existing recommendation functions mostly rely on single-dimensional data (such as only user historical records), and the recommendation strategy is fixed, which cannot dynamically adjust the recommended content in combination with the real-time context of the current operation (such as order type, semantic association of fields in the table). For example, traditional systems cannot distinguish between "e-commerce return and exchange" and "B2B contract delivery" scenarios, resulting in mismatched address templates or goods information recommended by the actual scenario, which is insufficient in accuracy.

[0029] 3. Lack of adaptive learning mechanism: Existing systems cannot continuously optimize the recommendation model according to user feedback (such as selection, modification, and ignoring behavior) on recommended content, and the recommendation effect is difficult to improve with the change of user usage habits, and the user experience cannot form a virtuous circle.

[0030] The core problem of the prior art is that only passive verification based on rules is implemented, and an intelligent recommendation system combining multi-source data (user behavior, static knowledge base, real-time context) is not constructed, and there is a lack of dynamic weight calculation and adaptive learning ability. The current technical solution stays at the stage of "reporting errors after discovering problems", and has not realized the intelligent leap of "actively predicting problems and providing solutions", and has not formed a closed loop of "data collection-intelligent recommendation-feedback optimization", resulting in cumbersome user operation, low recommendation accuracy of the system, and inability to meet the needs of efficient batch order processing.

[0031] The creativity of the present application is to break through the single dimension verification and fixed strategy recommendation mode of the prior art, through the technical means of multi-modal data fusion (structured table data, user behavior data, static knowledge base, real-time context), attention mechanism dynamic calculation of information source weight, adaptive learning model optimization, etc., to build an intelligent recommendation system of "real-time analysis - missing prediction - active recommendation - feedback iteration", realize the technical leap from "passive error reporting" to "active suggestion", from "single rule" to "context awareness".

[0032] Please refer to Figure 1 , Figure 1 is a schematic flow chart of an AI-based common shipping information intelligent recommendation method provided by the present application. The method is applied to a computer device, which can be deployed on a single server or a server cluster. It can also be deployed on a handheld terminal, a notebook computer, a wearable device, or a robot, etc.

[0033] It should be noted that the acquisition of any information mentioned in the provided method is in accordance with relevant regulations and with the consent of the user, and does not infringe on the user's privacy or violate relevant laws and regulations.

[0034] As Figure 1 shown, the specific steps of the AI-based common shipping information intelligent recommendation method include steps S101 to S104.

[0035] S101, acquire the shipping information table input by the user, parse the fields in the shipping information table, and extract the filled structured data; collect user behavior data formed by user historical order records and static knowledge base information, and the static knowledge base information at least includes common address and commodity library data.

[0036] Specifically, this step realizes real-time analysis of the user input table, associated collection of historical behavior data, and calling of the static knowledge base, forming a basic data input set.

[0037] The table analysis and structured data extraction includes: an analysis engine: using a table analysis tool (such as pandas of Python, openpyxl, Apache POI of Java) to read an Excel table in real time, supporting dynamic identification of table header fields (through a preset field mapping dictionary, such as “delivery address” corresponding to address, “product code” corresponding to sku). Field type verification: performing basic format verification (such as mobile phone regular matching, date format conversion) on the filled fields to generate a structured data dictionary filled_data (key is field name, value is filled content and type). Missing field identification: identifying the list of fields missing in the table through a preset complete set of order fields (such as the mandatory fields “recipient name” “contact number”, and the optional field “remark”).

[0038] The user behavior data collection includes: historical record extraction: querying the user's order records in the past 30 days from the user behavior database, including the address, product combination, and order type (such as “B2C retail” “B2B bulk” “return and exchange”) in the historical order, forming a user behavior sequence user_history (structured data format, containing timestamp, order scenario label, and filled field preference). High-frequency feature extraction: extracting high-frequency features such as frequently used addresses (such as home address, company address), commonly used product SKUs, and preferred logistics methods through statistical analysis (such as TF-IDF, sliding window).

[0039] The static knowledge base calling includes: a commonly used address library: containing preset standard address templates (such as enterprise delivery address library, after-sales warehouse address library), each address being associated with a scenario label (such as “B2B delivery address” “return and exchange address”) and a weight (calculated according to the address usage frequency). Product library data: storing product basic information (SKU, product name, specification, weight), supporting fuzzy search by product name or accurate matching by SKU, and associating the product category (such as “3C electronics” “clothing and luggage”).

[0040] S102, 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 of the current recommendation scenario.

[0041] Specifically, by semantically fusing the structured table data, user behavior, static knowledge base, and real-time operation scenario, a context vector containing scenario semantics is generated to provide comprehensive input for intelligent recommendation.

[0042] Real-time context information extraction includes: operation scene label: generate scene label scene_label by filling the "order type" field in the table (such as user manually annotating "B2B contract delivery" "e-commerce return and exchange") or implicit feature inference (such as "delivery address" contains enterprise name, which is inferred as B2B scene), generate scene label scene_label. Fill in the progress feature: calculate the proportion of filled fields, the integrity of key fields (such as address, goods), and generate the progress vector progress_vector (such as [0.7, 0.9] respectively indicating the completion degree of address field and goods field filling).

[0043] Multi-modal data vectorization includes: structured data embedding: convert field values in filled_data into vectors (such as address using word segmentation + Word2Vec embedding, and product SKU using One-Hot encoding), generate table feature vector table_embedding. User behavior embedding: encode user_history through time series model (such as LSTM) into user preference vector user_embedding, capture historical operation habits (such as preference for speed freight, commonly used recipient name pattern). Knowledge base embedding: pre-train common address library and product library to generate address semantic vector address_embedding and product semantic vector product_embedding, support fast retrieval of associated information.

[0044] Context fusion includes: feature splicing: splice table_embedding, user_embedding, address_embedding, product_embedding and scene_label (one-hot encoding) into raw_context. Scene enhanced coding: reduce and enhance the semantics of raw_context through fully connected neural network (FCN) to generate final context vector context_vector, which contains the comprehensive information of table content, user habit and scene semantics.

[0045] S103, 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 for the missing fields.

[0046] Specifically, the weight of multi-source information is dynamically calculated by the intelligent recommendation model, the most relevant information source of the current scene is focused, and the recommended candidate content for the missing fields is generated.

[0047] The intelligent recommendation model architecture comprises: a model structure: an "encoder-attention layer-decoder" architecture is adopted, the encoder is a multi-layer Transformer or CNN, and is used for extracting deep features of a context vector; the attention layer supports multi-head self-attention (Multi-Head Attention), and calculates contribution weights of different information sources (table data, user history, knowledge base) to a missing field. Dynamic weight calculation comprises: an attention mechanism formula: Attention(Q, K, V) = softmax(QK T / dk^0.5) V; wherein a query vector Q is a semantic representation (such as a pre-training vector of "mail address") of the missing field, and key K and value V are respectively obtained from feature vectors of table data, user history and knowledge base. Information source weight distribution is dynamically adjusted according to a scene label to adjust an attention focus (for example, in a B2B scene, the weight of an enterprise address library in the knowledge base is increased; in a return and exchange scene, the weight of an after-sales address in the user history is given priority).

[0048] Candidate content generation comprises: missing field classification: missing_fields are grouped according to types (such as address type, commodity type and logistics type), and recommendations are respectively generated. The recommendation strategy comprises: address field: high-matching-degree addresses are retrieved from a commonly used address library according to the attention weight (such as filtering the address library according to a scene label, and then sorting according to the user historical use frequency). Commodity field: commodity library association recommendation (such as having filled in "mobile phone", recommending a historical purchase of the same brand model or a hot-selling mobile phone SKU in the knowledge base). Logical reasoning: for fields such as "logistics mode", the order weight (if filled in), user historical preference and address accessibility (knowledge base logistics coverage area) are comprehensively inferred. The candidate set screening is filtered by a threshold to filter low-confidence recommendations (such as candidate contents with an attention weight <0.3), and a Top-N recommendation list (such as providing three candidate values for each missing field) is generated.

[0049] S104, the user is shown the recommended candidate content, and feedback information of the user on the recommended candidate content is received; parameters of the intelligent recommendation model are updated according to the feedback information, and the intelligent recommendation model is optimized.

[0050] Specifically, by collecting the interactive feedback of the user on the recommended content, a "recommendation-feedback-iteration" closed loop is constructed, and the model recommendation strategy is continuously optimized.

[0051] The feedback information collection includes: interaction behavior classification: record the user's operation on the recommended content (such as "adopt", "modify", "ignore", "delete"), generate a feedback log feedback_log, which contains a timestamp, a recommended field, candidate content, a user operation type, and a modified value (if any). Quantitative feedback weight: assign weights to different feedback behaviors (such as "adopt" weight +1, "modify" weight +0.5, "ignore" weight -0.2), which is used to adjust the loss function of model training.

[0052] The model parameter update includes: supervised learning update: the user's final filled content (whether to adopt the recommendation or not) is used as a label to construct a training sample with the model recommendation output, and the weight parameters and decoder network of the attention mechanism are optimized through gradient descent. Reinforcement learning optimization: introduce a reward function (such as +1 for adopting a recommendation, and -1 for frequent ignoring), use deep reinforcement learning (such as DQN, PPO) to adjust the recommendation strategy, and improve the long-term recommendation accuracy.

[0053] The incremental learning mechanism includes: real-time incremental training: set up a feedback information buffer pool, when a certain number of feedback samples (such as 50 per user) are accumulated, trigger model micro-update, avoid frequent full training affecting performance. Cold start processing: for new users or new scenarios, prefer to use static knowledge base and general recommendation strategy, gradually build personalized model through feedback information.

[0054] In some embodiments, the parsing of the fields in the shipping information form, extracting the filled structured data, includes: identifying the column titles of the shipping information form according to the preset field rules, determining the 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 checking on the specific content, identifying data that meets the preset format specification and converting it to a structured format, and marking data with format errors.

[0055] By parsing the table column title through the preset field rule, the core business fields (sender information, cargo information) are identified, the effective data is extracted and the format problem is marked, providing structured input for subsequent recommendation.

[0056] Field rule preset and column title identification: By establishing a field mapping dictionary (e.g., {"recipient name": "recipient_name", "product name": "product_name"}), fuzzy matching is supported (e.g., "shipping address" and "recipient address" are both mapped to "address"). Regular expressions or string matching algorithms (e.g., Levenshtein distance) are used to verify table column titles, and columns that match are marked as target fields (e.g., "sender phone number" and "sender address" are included in the sender information, and "product SKU" and "quantity" are included in the goods information).

[0057] Content extraction and missing field marking: By traversing the columns corresponding to the target fields, non-empty cell contents are extracted row by row and stored in a temporary dictionary {field name: [value1, value2,...]}; empty cells are marked as <missing>, and a missing field list missing_fields is generated (e.g., if the "recipient phone number" is empty in a certain row, the field is marked as missing for that row).

[0058] Format verification and structured conversion: By performing format verification on the extracted content (e.g., the mobile phone number must match the regular expression ^1[3-9]\d{9}$, and the address must contain province, city, and district level information), data that meets the specifications is converted to a unified format (e.g., addresses are unified as "province-city-district-street-door number"); data with format errors is marked as <format exception> and the error type (e.g., "mobile phone number is not enough digits") is recorded, and in the future, such missing or erroneous fields will be prioritized for supplementation.

[0059] In some embodiments, the user behavior data formed by collecting user historical order records and static knowledge base information includes: extracting complete data of historical orders from the user historical order database, including filled addresses, selected goods, operation times, and modification records, to form a user behavior data set; obtaining commonly used address and goods library data from a pre-set static knowledge base.

[0060] By collecting user historical operation trajectories and standardized basic data, a personalized preference library and a general knowledge base are constructed to provide historical behavior basis and standard data support for recommendation.

[0061] User behavior data set construction: By querying the user's historical orders in the past 12 months from the order database, the following fields are extracted: filled information: recipient address, selected goods (including SKU, name, quantity), logistics method, order type (label); operation information: order time, modification times, modified fields (e.g., changing "ordinary express" to "special express" in a certain modification); the data is arranged in chronological order to form a time series behavior sequence user_action_sequence, which is used to analyze user filling habits (e.g., whether to frequently modify addresses, preference for fixed goods combinations).

[0062] Static knowledge base acquisition includes: common address library: store enterprise standard shipping address (including warehouse address, after-sales address), user-defined common address (such as home, company address), each address is associated with scene label (such as "B2C retail address" "B2B bulk delivery address") and usage frequency count; commodity library data: contains commodity basic attributes (SKU, name, category, weight, size), associated attributes (such as "mobile phone" associated "charger" "protective film" as package recommendation item), supports fuzzy retrieval by SKU or name.

[0063] In some embodiments, the parsed structured data, user behavior data, static knowledge base information and real-time context information of the current operation are fused to generate context information of the current recommendation scene, including: associating the filled fields in the structured data with the historical filling habits in the user behavior data; combining the real-time context information of the current operation, filtering the address templates and commodity attributes in the static knowledge base; integrating the address templates, commodity attributes and associated filled fields with the user behavior data according to the field dimension, forming context information containing user historical preferences, current scene characteristics and existing filled content.

[0064] By associating the filled content of the table, user historical preferences, static knowledge and real-time scene, filtering irrelevant information, forming context features focusing on the current operation, solving the problem of mismatching recommendation scenes.

[0065] The association of filled fields and historical habits queries the historical filling preferences (such as whether to use after-sales address first when returning or exchanging, whether to fill "quality problem" as the reason) in the user behavior data under this scene according to the filled fields (such as "order type" filled with "return or exchange") in the structured data, and generates a preference feature vector (such as [after-sales address usage rate: 0.8, quality problem reason proportion: 0.6]).

[0066] Static knowledge base filtering filters out entries marked as "enterprise address" from the common address library and "bulk purchase high-frequency commodities" (such as through category + purchase quantity threshold screening) from the commodity library according to real-time context (such as scene label "B2B contract delivery"), reducing irrelevant data interference.

[0067] Field dimension integration establishes a feature matrix according to field type (address type, commodity type, logistics type), each row corresponds to a field, and the columns contain: table filled content (such as "recipient address" filled with "XX province"). User history: the high-frequency value of the user history field (e.g., the high-frequency of the user history "recipient name" is "Zhang San"); recommended value of the static knowledge base (e.g., address template filtered according to the scene); and finally, the context information context_matrix containing three dimensions (current filling, historical preference, and knowledge base standard).

[0068] In some embodiments, the dynamic calculation of the weight of different information sources for the current recommendation scene by introducing the attention mechanism includes: 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 scene characteristics; and generating a weight value of the corresponding information source according to the contribution degree, so that the intelligent recommendation model pays more attention to the information source with a higher association degree with the current field.

[0069] The importance of multi-source information to the current field is quantified by the attention mechanism, so that the model focuses on high-correlation information sources, solving the problem of insufficient recommendation accuracy of traditional fixed strategies.

[0070] The multi-modal input and module architecture includes: inputting the structured data (table features), user behavior data (time series vector), static knowledge base (address / commodity embedding), and real-time scene (label one-hot encoding) as inputs of the multi-head attention module, and converting each information source into a vector with unified dimensions (such as 128 dimensions).

[0071] The field type and scene perception calculation first judges the type of the missing field (such as "recipient phone" belonging to the contact information category and "product SKU" belonging to the goods information category), combines the scene label (such as "B2C retail" relying more on user historical delivery address and "B2B" relying more on knowledge base enterprise address), and generates a query vector Q_field = [field_type, scene_label]; the similarity between Q_field and each information source key vector K_i is calculated by dot product: Similarity(Q, K_i) = Q * K_i T / dk^0.5; where d_k is the key vector dimension, and the similarity is obtained after Softmax to get the weight w_i of each information source (such as user behavior data weight 0.6, static knowledge base weight 0.3, and table data weight 0.1).

[0072] The dynamic weight allocation automatically increases the weight of user historical behavior for high-frequency scenes (such as the "regular replenishment" scene where the user orders every month); and increases the weight of the static knowledge base for low-frequency scenes (such as the first B2B order), balancing personalization and universality.

[0073] In some embodiments, the extracting information related to the missing field in the table from each information source based on the weight comprises: performing hierarchical screening on the multi-modal information source according to the calculated weight, and preferentially extracting data matching the missing field in a high-weight information source; and wherein, for the recipient phone field, if the current scenario is a user historical high-frequency delivery scenario, 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.

[0074] According to the attention weight, high-correlation information is preferentially extracted, and a differentiated extraction strategy is designed in combination with field characteristics, so as to realize accurate information retrieval (such as preferentially matching the historical record of the recipient phone).

[0075] The hierarchical screening strategy orders the information sources according to the weight from high to low (such as user behavior data > static address library > table filled data), and sequentially retrieves matching data: 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; and low-weight source (weight < 0.3): only as a supplement candidate.

[0076] The field customization extraction (taking the recipient phone as an example) first extracts the historical phone information (there can be multiple, such as home phone and office phone) of the recipient from the user behavior data if the current scenario is a user historical high-frequency delivery scenario (such as the address of the recipient appearing ≥ 3 times in the last 30 days); if there is no matching phone in the historical data, the same-name recipient record is retrieved from the static address library, and the item marked as “common contact phone” is preferentially selected; if there is still no result, a bottom-up strategy (such as prompting the user to manually fill in, or recommending a default contact phone format) is triggered.

[0077] In some embodiments, the generating of the recommended candidate content of the missing field comprises: deduplicating and prioritizing the extracted multi-source information, generating at least one recommended candidate content according to the weight from high to low; in combination with semantic context reasoning, verifying the relevance of the recommended content, and filtering obviously unmatched information; encapsulating the recommended candidate content that passes the verification in a structured form to form an interactive recommended result.

[0078] By deduplicating, sorting, and semantically verifying the information extracted from multiple sources, a structured recommended result is generated to ensure that the recommended content conforms to the business logic and context association.

[0079] Deduplication and priority sorting use a hash table to deduplicate the candidate values extracted from multiple sources (such as different information sources extracting the same address), and sort them according to the weight product (information source weight × frequency of candidate value in source), and retain the top-3 candidates.

[0080] Semantic relevance verification is based on domain knowledge graph (e.g., "recipient address" and "recipient phone" should belong to the same region, and "product SKU" and "product name" should be semantically consistent), and the logical consistency of candidate content is verified through a rule engine or a pre-trained language model (e.g., BERT): for example, the recommended "recipient phone" area code should match the "recipient address" province (010 for Beijing address); the recommended "product SKU" should exist in the product library, and the category should match the "order type" that has been filled (e.g., "fresh food" orders do not recommend electronic product SKUs).

[0081] Structured packaging and interaction design encapsulate the verified candidate content as a structured object with weighted labels (e.g., {"field": "recipient_phone", "candidates": [{"value": "138XXXX1234", "score": 0.9},...]}); in the interactive interface, high-score candidates are displayed in priority, and the recommended source (e.g., "historical frequently used" and "enterprise standard address") is marked to facilitate quick identification by the user.

[0082] In some embodiments, the updating of the parameters of the intelligent recommendation model according to the feedback information and the optimization of the intelligent recommendation model include: 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; inputting the feedback information and the corresponding field type, the recommended content, and the operation time into a model training module, and adjusting the weight calculation logic of the attention mechanism through incremental learning; for the recommended content that is rejected or modified multiple times by the user, the weight of the information source corresponding to the recommended content in the field type recommendation is reduced; for the recommended content that is not modified or modified once by the user, the weight of the corresponding information source is increased, realizing the adaptive optimization of the recommendation model.

[0083] The recommendation effect is quantified through user interaction feedback, the information source weight and the model strategy are dynamically adjusted, a "recommendation-feedback-optimization" closed loop is formed, and the long-term recommendation accuracy is improved.

[0084] Feedback information collection and recording record three types of feedback behaviors: adoption: the user directly selects the recommended candidate value (marked as action=accept); modification: the user modifies the recommended value (records the content before and after modification, and is marked as action=modify); ignore: the user does not select any recommended value and manually fills in 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., which are used for subsequent analysis.

[0085] Incremental learning and weight adjustment includes: model training module: convert feedback information into training samples (input is context vector when recommending, label is user's final fill-in value), use stochastic gradient descent (SGD) to update weight matrix in attention mechanism; Dynamic weight adjustment rule: if a recommended content is "ignored" for > 3 times, the weight of the corresponding information source in this field type is reduced by 20%; If the "adopt" times account for > 60%, the weight of the corresponding information source is increased by 15%; The "modify" behavior is considered as partial recognition, and the weight is fine-tuned (+5%), while recording the modification mode (such as the user habit of adding a house number after the recommended address), and optimizing the details of subsequent recommendations.

[0086] Cold start scenario adaptation adjusts the initial weight to default to the static knowledge base (weight ratio 70%) for new users or scenarios without historical behavior data, and gradually increases the weight of user behavior data (weight increases by 5% for every 10 new feedbacks) as feedback information accumulates.

[0087] In some embodiments, before processing the context information using the preset intelligent recommendation model, it further includes: constructing an intelligent recommendation model containing an attention mechanism, the input layer of the intelligent recommendation model receiving the multi-modal fused context information, and the output layer corresponding to the recommended candidate content of each missing field; Train the model using historical order data, user behavior logs and labeled recommendation scenario data, adjust the model parameters by comparing the matching degree of the recommended results and the actual fill-in content; In the training process, generate sub-models for different business scenarios, so that the intelligent recommendation model can automatically switch the recommendation strategy according to the real-time context.

[0088] By constructing an attention mechanism model supporting multiple scenarios, the generalization ability is improved through historical data training, and sub-models are generated to adapt to different business scenarios, realizing dynamic switching of recommendation strategies.

[0089] Model architecture design includes: input layer: receives multi-modal fused context vector (dimension configurable, such as 512 dimensions); Core layer: contains multi-head attention module (such as 8 heads) and feedforward neural network (FFN), used for feature weighting and semantic enhancement; Output layer: for each missing field, output candidate content probability distribution (such as Softmax output Top-3 candidate probability value).

[0090] The training data and optimization target are obtained by using historical order data (annotated real fill-in values), user behavior logs (recorded recommendation interactions), and scene label data (artificially annotated or automatically generated) as the training set; the loss function combines cross-entropy loss (matching degree of recommended value and real value) and reinforcement learning reward (feedback behavior weight), and the formula is: Loss = a * CrossEntropy + (1-a) * ReinforcementReward; wherein a is a balance coefficient (default 0.7).

[0091] The multi-scene sub-model is generated by splitting the training data according to order types (B2C, B2B, return and exchange), industries (fresh food, 3C, and clothing), and other dimensions, and using transfer learning to train the sub-model (sharing the bottom attention module and fine-tuning the upper output layer); during model deployment, the corresponding sub-model is automatically loaded according to the real-time scene label (inferred through the "order type" field or address keywords), and the recommendation strategy is switched (for example, in the fresh food scene, cold chain logistics is preferentially recommended, and in the 3C scene, price protection service is preferentially recommended).

[0092] In some embodiments, by modeling the missing field recommendation as a sequence decision problem, the recommendation order and information source combination strategy are dynamically optimized through reinforcement learning (RL), and the recommendation efficiency problem in the multi-field dependent scene is solved (for example, when the "delivery address" is not filled in, whether to preferentially recommend the "recipient's phone number" to assist in inferring the address).

[0093] The state space definition includes: the state S_t includes: the current table filled field vector F_t (such as [filled field 1: value 1, filled field 2: value 2]); the set of fields to be recommended M_t (such as {"delivery address", "recipient's phone number"}); the historical recommendation action sequence A_1^t (records the recommended fields and contents); and the real-time scene label C (such as "cross-border e-commerce" and "urgent order").

[0094] The action space design includes: the action A_t includes: selecting the next recommended field field (based on field dependency, such as preferentially recommending when "address" and "phone number" are strongly associated); selecting the information source combination strategy policy (such as joint retrieval of "user behavior data + static knowledge base"); and generating the recommended candidate set size k (1 precise value is recommended in high-frequency scenarios, and 3 candidates are recommended in low-frequency scenarios).

[0095] The reward function design includes: Instant reward: +10 points for user adoption of recommended value, +5 points for modification, and -5 points for ignoring; long-term reward: field fill-in time is shortened (each saved 1 second +0.1 points), and subsequent field recommendation accuracy is improved (based on state transition probability); constraint penalty: -20 points for recommending repeated content or format error data.

[0096] Model training and strategy execution use deep Q network (DQN) or policy gradient (PPO) algorithm, with historical interaction logs as the initial experience pool, and balance exploration and utilization through the epsilon-greedy strategy; for the "address-phone-postcode" strong dependent field group, the trained model preferentially recommends "address" to reduce subsequent field reasoning ambiguity, and in the cross-border scenario, preferentially recommends the "country / region" field to lock the address format.

[0097] In some embodiments, by constructing a logistics field knowledge graph, structuring entity relationships such as "address-recipient-goods-logistics rules", and enhancing semantic reasoning capability through graph neural network (GNN), the implicit association recommendation in complex business scenarios (such as recommending matching logistics channels according to the weight of the goods) is solved.

[0098] Knowledge graph construction includes: entity types: address (province / city / district / street), recipient (enterprise / personal), goods (SKU / category / weight / size), logistics rules (volume weight calculation, prohibited goods list), scenario label (e-commerce / contract / sample); relationship definition: address-recipient: "common address" "associated phone"; goods-logistics rules: "weight≤1kg→recommend economy express" "contains lithium battery→prohibit air transport"; scenario-address: "cross-border e-commerce→need to contain customs code".

[0099] Graph embedding and reasoning process uses TransE or GraphSAGE algorithm to embed entities and relationships into vectors to generate graph knowledge base KG; when recommending "logistics method" field: ① extract filled fields "goods weight = 2.5kg" "recipient address = **"; ② query "weight 2.5kg+** address" associated logistics rules in KG, and filter out channels that support international transportation and weight≤3kg (such as DHL, FedEx); ③ combine user historical preferences (such as user 90% international pieces choose DHL), preferentially recommend DHL and mark the prohibited transport risk prompt.

[0100] Abnormality detection and conflict resolution through graph logic verification of recommended content: such as "recipient address = Xinjiang" and "goods = lithium battery" exist air transport prohibited conflict, automatically replace the land transport channel, and mark the missing attribute of the goods (need to supplement "transportation preference").

[0101] In some embodiments, by introducing meta-learning (Meta-Learning) for new users / new scenarios (no historical behavior data), generating personalized recommendation strategies using few-shot interaction data, the problem of low recommendation accuracy of traditional model cold start is solved.

[0102] The meta-learning framework design adopts a model-independent meta-learning (MAML) architecture, including: a basic model consistent with the original intelligent recommendation model structure (including attention mechanism); a meta-trainer that updates model parameters through a small number of tasks (such as the first three orders of different users) to quickly adapt to new tasks.

[0103] The meta-training data construction simulates the cold start scenario by sampling "user-scenario" pairs from historical data: for user A, the first two orders are kept as the "support set", and the subsequent orders are kept as the "query set"; the training target is to make the model learn to predict the missing fields of the third order of the user through two order data.

[0104] The cold start recommendation process includes: when a new user places an order for the first time: ① collect the first three filled fields (such as the first filled "delivery address" and "product name") as the 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 "value protection service" and "paper box packaging" based on "product name = notebook computer" and historical data of similar users); as the interaction data increases (≥5 times), gradually switch back to the main model, and add new user data to the meta-training pool for continuous optimization.

[0105] In some embodiments, by aggregating delivery data from multiple enterprises / departments through federated learning (Federated Learning) under the premise of protecting user privacy and data compliance, a cross-domain recommendation model is constructed to solve the problem of sparse data in a single enterprise (such as joint optimization of recommendation strategies by small and medium-sized businesses).

[0106] The federated learning architecture design adopts a hierarchical federated architecture: client: each enterprise locally deploys a sub-model and stores desensitized user behavior data (such as region code without specific address and commodity category statistics); coordination server: aggregates model parameters (gradient or weight) of each client without directly accessing raw data; global model: iteratively updated through FedAvg algorithm, and regularly issues updated model to the client.

[0107] Data desensitization and security mechanism: k-anonymity processing on address fields (such as keeping to the city level and hiding the district / street); using homomorphic encryption technology for commodity data, only sharing category-level association rules (such as the co-occurrence probability of "fresh food → insulation box packaging"); introducing differential privacy (DP) to add noise to ensure that aggregated parameters cannot be used to infer the original data.

[0108] When a certain enterprise user fills in "commodity category = fresh", ① the local model combines the cross-domain knowledge learned by the global model (in the fresh scenario, 58% of enterprises recommend "ice bag + insulation box" packaging combination); ② fusion of the enterprise's own packaging inventory data (preferably recommend existing inventory insulation box model); ③ generate recommendations that take into account industry commonality and enterprise characteristics, while avoiding revealing specific inventory information of other enterprises.

[0109] In some embodiments, by breaking the pure table filling mode, introducing natural language interaction (voice / text) and visual guidance, understanding user intent through multi-modal fusion model, dynamically generating guided recommendations, improving filling efficiency in complex scenarios (such as the user speaking "send the last address" automatically associating historical address).

[0110] Multi-modal input processing includes: text / voice analysis: using BERT-NER to extract entities in user input (such as "last address" mapped to historical address ID, "express" mapped to "express type = express express"); visual interaction: add a quick input panel next to the table, support clicking historical address card, dragging the commodity list to the table, sliding to select logistics time, etc. Operation generates interaction behavior vector I_v.

[0111] Intention understanding and modal fusion through the construction of a multi-modal fusion model (such as a Transformer encoder), input includes: table structured data F_s; natural language intent vector I_t (generated by text embedding model); visual interaction vector I_v (generated by operation sequence encoding); through attention mechanism to calculate the contribution of each modality to the missing field (such as voice input "send company", "recipient address" field language modality weight increased to 0.8).

[0112] Guided recommendation generation through fuzzy input (such as user input "old"): ① query the filling mode of the user's last 3 similar orders, generate "default filling template"; ② highlight the recommended fields in the table and label "recommended based on your historical habits", allow one-key application template; Conflict input (such as the user says "send Beijing" but the table has filled in "Shanghai address"): ① trigger ambiguity resolution mechanism, pop-up prompt "detect address conflict, do you want to use Beijing historical address?", Combined with historical modification records to recommend the optimal solution.

[0113] The application discloses an AI-based common delivery information intelligent recommendation method and device, the provided method actively recommends missing content by real-time analysis of table fields, and users do not need to manually fill in or repeatedly modify the table, thereby significantly reducing the number of repeated uploads and shortening the order processing time. The method is based on multi-modal data fusion and attention mechanism, and the system can dynamically adjust the recommendation strategy according to the current operation scene (such as order type and existing filled content). For example, the recipient's name is intelligently associated with the common address and telephone number, the recommended result is more in line with the actual demand, and the human error rate is reduced. The method continuously optimizes the recommendation model through user feedback, and the frequently used address, commodity and other information will be preferentially recommended, the recommendation accuracy is gradually improved with the use time, and a personalized experience of 'the more you use, the more intelligent' is formed. The high-quality structured data generated lays a foundation for subsequent intelligent order distribution, supply chain optimization and other links, and improves the intelligent level of the overall business process.

[0114] Please refer to Figure 2 , Figure 2 Embodiments of the application also provide a schematic block diagram of an AI-based common delivery information intelligent recommendation device, which is used for executing the AI-based common delivery information intelligent recommendation method described above. The AI-based common delivery information intelligent recommendation device 200 can be configured in a server or a terminal.

[0115] The server can be a stand-alone server, a server cluster, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDN) and big data and artificial intelligence platforms, and the like. The terminal can be a mobile phone, a tablet computer, a notebook computer, a desktop computer, a user digital assistant and a wearable device, and the like.

[0116] As shown in Figure 2 , the AI-based common delivery information intelligent recommendation device 200 comprises: A data extraction unit 201 is configured to acquire a delivery information table input by a user, analyze fields in the delivery information table, and extract filled structured data; collect user behavior data formed by historical order records of the user and static knowledge base information, and the static knowledge base information at least includes common address and commodity library data; An information generation unit 202 is configured to fuse the analyzed structured data, the user behavior data, the static knowledge base information and real-time context information of the current operation, and generate context information of the current recommendation scene; The information processing unit 203 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 a current recommendation scene by introducing an attention mechanism, extracts information related to a missing field in a table from each information source based on the weight, and generates a recommendation candidate content for the missing field. The intelligent recommendation unit 204 is configured to display the recommendation candidate content to a user, receive feedback information of the user on the recommendation candidate content, and update parameters of the intelligent recommendation model according to the feedback information to optimize the intelligent recommendation model.

[0117] In some embodiments, the parsing of the field in the shipping information table and the extraction of the filled structured data include: identifying the column title of the shipping information table according to a preset field rule, and determining the target field corresponding to the sender information and the cargo information; for each target field, extracting the specific content of the filled cell in the corresponding column, and marking the missing field that is not filled; and performing format checking on the specific content, identifying the data that meets the preset format specification and converting the data into a structured format, and marking the data that has format errors.

[0118] In some embodiments, the user behavior data formed by collecting user historical order records and the static knowledge base information include: extracting complete data of historical orders from a user historical order database, including filled addresses, selected goods, operation time and modification records, to form a user behavior data set; and obtaining common address and goods library data from a preset static knowledge base.

[0119] In some embodiments, the fusion of the parsed structured data, the user behavior data, the static knowledge base information and the real-time context information of the current operation to generate the context information of the current recommendation scene includes: associating the filled field in the structured data with the historical filling habit in the user behavior data; filtering the address template and the goods attribute in the static knowledge base in combination with the real-time context information of the current operation; and integrating the address template, the goods attribute and the associated filled field with the user behavior data according to the field dimension to form the context information containing the user historical preference, the current scene characteristics and the existing filled content.

[0120] In some embodiments, the dynamic calculation of the weight of different information sources for the current recommendation scene by introducing the attention mechanism includes: inputting the structured data, the user behavior data, the static knowledge base information and the real-time context information as multi-modal input into an 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 scene characteristics; and generating a weight value of the corresponding information source according to the contribution degree, so that the intelligent recommendation model pays more attention to the information with a higher association degree with the current field.

[0121] In some embodiments, the extracting information related to the missing field in the form from each information source based on the weight comprises: performing hierarchical screening on the multi-modal information source according to the calculated weight, and preferentially extracting data matching the missing field in the high-weight information source; and for the recipient phone field, if the current scenario is a user historical high-frequency delivery scenario, 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.

[0122] In some embodiments, the generating the recommended candidate content of the missing field comprises: performing deduplication and priority sorting on the extracted multi-source information, and generating at least one recommended candidate content in descending order of weight; performing relevance verification on the recommended content in combination with semantic context reasoning, and filtering obviously unmatched information; and encapsulating the recommended candidate content that passes the verification in a structured form to form an interactive recommended result.

[0123] In some embodiments, the updating the parameters of the intelligent recommendation model according to the feedback information and optimizing the intelligent recommendation model comprises: 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; inputting the feedback information and the corresponding field type, the recommended content, and the operation time into the model training module, and adjusting the weight calculation logic of the attention mechanism through incremental learning; for the recommended content that is rejected or modified multiple times by the user, reducing the weight of the information source corresponding to the recommended content in the field type recommendation; and for the recommended content that is not modified or modified once by the user, increasing the weight of the corresponding information source, to realize self-adaptive optimization of the recommendation model.

[0124] In some embodiments, before the processing the context information by using the preset intelligent recommendation model, further comprising: constructing an intelligent recommendation model containing an attention mechanism, wherein an input layer of the intelligent recommendation model receives the context information after multi-modal fusion, and an output layer corresponds to recommended candidate content of each missing field; training the model by using historical order data, user behavior logs, and labeled recommendation scenario data, adjusting the model parameters by comparing the matching degree of the recommended result and the actual filled content; and in the training process, generating a sub-model for different business scenarios, so that the intelligent recommendation model can automatically switch the recommendation strategy according to the real-time context.

[0125] It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the model training device and each module described above can refer to the corresponding process in the foregoing AI-based common delivery information intelligent recommendation method embodiments, which will not be described herein.

[0126] The AI-based common shipping information intelligent recommendation device described above can be implemented in the form of a computer program, which can run on a computer device as shown in Figure 3 .

[0127] Please refer to Figure 3 , Figure 3 is a structural schematic block diagram of a computer device provided by an embodiment of the present application. The computer device can be a server or a terminal.

[0128] Please refer to Figure 3 , the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a storage medium and an internal memory.

[0129] The storage medium can store an operating system and a computer program. The computer program includes program instructions which, when executed, can cause the processor to perform any AI-based common shipping information intelligent recommendation method provided by an embodiment of the present application.

[0130] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0131] The internal memory provides an environment for the execution of the computer program in the storage medium, which, when executed by the processor, can cause the processor to perform any AI-based common shipping information intelligent recommendation method. The storage medium can be non-volatile or volatile.

[0132] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 3 the structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0133] It should be appreciated that the processor can be a central processing unit (CPU), the processor can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, and the like. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor and the like.

[0134] For example, in one embodiment, the processor is configured to run a computer program stored in the memory to implement the following steps: Obtaining a user-inputted shipping information form, parsing fields in the shipping information form to extract filled structured data; collecting user behavior data formed by user 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 calculates the weight of different information sources on the current recommendation scenario by introducing an attention mechanism, extracts information related to the missing fields in the form from each information source based on the weight, and generates recommended candidate content for the missing fields; Displaying the recommended candidate content to the user, receiving feedback information of the user on the recommended candidate content; updating parameters of the intelligent recommendation model according to the feedback information, and optimizing the intelligent recommendation model.

[0135] In some embodiments, the parsing of the fields in the shipping information form to extract the filled structured data comprises: identifying column titles of the shipping information form according to a preset field rule, determining target fields corresponding to the sender information and the goods information; for each target field, extracting specific contents of filled cells in the corresponding column, and marking missing fields that are not filled; performing format checking on the specific contents, identifying data conforming to the preset format specification and converting the data into a structured format, and marking data with format errors.

[0136] In some embodiments, the user behavior data formed by collecting user historical order records and static knowledge base information comprises: extracting complete data of historical orders from a user historical order database, including filled addresses, selected goods, operation time and modification records, to form a user behavior data set; and obtaining commonly used address and goods library data from a preset static knowledge base.

[0137] In some embodiments, 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: associating filled fields in the structured data with historical filling habits in the user behavior data; filtering address templates and goods attributes in the static knowledge base in combination with real-time context information of the current operation; and integrating the address templates, goods attributes and associated filled fields with the user behavior data according to field dimensions to form context information containing user historical preferences, current scenario characteristics and existing filled content.

[0138] In some embodiments, the dynamic calculation of weights of different information sources for the current recommendation scenario by introducing an attention mechanism comprises: inputting the structured data, user behavior data, static knowledge base information and real-time context information as multi-modal inputs into an attention mechanism module of an intelligent recommendation model; calculating the contribution of each information source to the recommendation of a missing field according to the field type and the current scenario characteristics for each missing field; and generating a weight value of the corresponding information source according to the contribution, so that the intelligent recommendation model pays more attention to information with a higher degree of association with the current field.

[0139] In some embodiments, the extraction of information related to the missing fields in the table from the information sources based on the weights comprises: hierarchical screening of the multi-modal information sources according to the calculated weights, and preferentially extracting data matching the missing fields from high-weight information sources; and for the recipient phone field, if the current scenario is a user historical high-frequency delivery scenario, 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.

[0140] In some embodiments, the generation of recommended candidate content for the missing fields comprises: deduplication and priority sorting of the extracted multi-source information, generation of at least one recommended candidate content from high to low according to the weights; association verification of the recommended content in combination with semantic context reasoning, filtering of obviously unmatched information; and encapsulation of the recommended candidate content that passes the verification in a structured form to form an interactive recommendation result.

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

[0142] In some embodiments, before the processing of the context information by using the preset intelligent recommendation model, the method further comprises: constructing an intelligent recommendation model comprising an attention mechanism, wherein an input layer of the intelligent recommendation model receives the context information fused in multiple modes, and an output layer corresponds to recommended candidate content of each missing field; training the model by using historical order data, user behavior logs and labeled recommended scene data, adjusting model parameters by comparing the matching degree of the recommended result and the actually filled content; and in the training process, generating a sub-model for different business scenarios, so that the intelligent recommendation model can automatically switch the recommendation strategy according to real-time context.

[0143] The application further provides a computer readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor implements the steps of the risk early warning method according to the first aspect.

[0144] The computer readable storage medium can be an internal storage unit of the computer device, for example, a hard disk or a memory of the computer device. The computer readable storage medium can also be an external storage device of the computer device, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card and the like.

[0145] The above merely describes specific embodiments of the application, but the protection scope of the application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed in the application, and these modifications or replacements should be covered in the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection 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