Credit exchange method and device, equipment, storage medium and program product
By generating structured and unstructured features and combining an intelligent question-answering model with an points knowledge graph and a consultation corpus, the problem of insufficient intelligence in existing points redemption systems is solved, achieving efficient and accurate points redemption services and improving user experience and the incentive effect of the points system.
Patent Information
- Application Number
- CN202511085758.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-04
- Publication Date
- 2025-11-21
AI Technical Summary
The existing points redemption system lacks sufficient intelligence. Users need to manually check their points balance and compare the redemption conditions of each product. The process is cumbersome and inefficient, and the recommendation accuracy is not high, resulting in a poor user experience, idle points, and weakening the incentive effect and commercial value of the points system.
By acquiring users' original consultation data, generating structured and unstructured features, and utilizing an points knowledge graph and consultation corpus, combined with an intelligent question-answering model, we can accurately identify consultation intent and keyword dependencies, thereby achieving efficient and accurate points redemption services.
It improves the efficiency and accuracy of points redemption, enhances the user experience, dynamically matches user conditions with product redemption conditions, and improves the incentive effect and commercial value of the points system.
Smart Images
Figure CN120996189A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to a points redemption method, apparatus, device, storage medium and program product. Background Technology
[0002] Current points redemption systems generally suffer from insufficient automation. Users must manually check their points balance and compare redemption conditions for each item, a cumbersome and inefficient process. While some systems offer recommendations based on consumer behavior, the accuracy of these recommendations is low. This traditional model not only results in a poor user experience and low participation, but also leaves a large number of points idle, severely weakening the incentive effect and commercial value of the points system. Summary of the Invention
[0003] This invention provides a points redemption method, apparatus, device, storage medium, and program product that can solve the problem that conventional methods cannot provide efficient and accurate points redemption services.
[0004] To solve the above-mentioned technical problems, the present invention is implemented as follows:
[0005] In a first aspect, embodiments of the present invention provide a points redemption method, comprising: acquiring a user's original consultation data; the original consultation data including consultation data related to redeeming goods using points; generating structured features of the user based on the original consultation data and the user's account information, and generating unstructured features of the original consultation data based on the original consultation data; wherein the user's account information includes at least the user's points data, the structured features are used to characterize structured data, and the unstructured features are used to characterize unstructured data; generating fusion features based on the structured features and the unstructured features; and obtaining consultation results of the original consultation data generated by the intelligent question answering model by inputting the fusion features into the intelligent question answering model; wherein the consultation results include intent classification results and / or dependency relationship results, the intent classification results are used to characterize the consultation intent of the entire original consultation data, and the dependency relationship results are used to characterize the dependency relationships between multiple keywords of the original consultation data.
[0006] Optionally, the points redemption method further includes: generating the user's structured features by inputting the original consultation data and the user's account information into a points knowledge graph; wherein the points knowledge graph is constructed based on the points of the product.
[0007] Optionally, the points redemption method further includes: obtaining the unstructured features of the original consultation data by inputting the original consultation data into a consultation corpus; wherein the consultation corpus is constructed based on text data related to points for the product.
[0008] Optionally, the points redemption method further includes: the intelligent question-answering model is trained based on an intelligent question-answering training set and intelligent question-answering tags; wherein, the intelligent question-answering training set includes the user's historical original consultation data, and the intelligent question-answering tags include the consultation intent corresponding to the user's historical original consultation data and the dependency relationship between multiple keywords in the user's historical original consultation data.
[0009] Optionally, the points redemption method further includes: performing feature fusion on the structured features, the unstructured features, and the original consultation data to obtain the fused features.
[0010] Optionally, the points redemption method further includes: obtaining the user's behavior analysis results based on the user's historical original consultation data; and recommending products to the user based on the user's behavior analysis results.
[0011] Secondly, embodiments of the present invention provide a points redemption device, comprising: generating fused features based on the structured features and the unstructured features; and an acquisition module for acquiring original consultation data of a user; wherein the original consultation data includes consultation data related to redeeming goods using points; a feature extraction module for generating structured features of the user based on the original consultation data and the user's account information, and generating unstructured features of the original consultation data based on the original consultation data; wherein the user's account information includes at least the user's points data, the structured features are used to characterize the structured data, and the unstructured features are used to characterize the unstructured data; and an intelligent question answering module for obtaining consultation results of the original consultation data generated by the intelligent question answering model by inputting the fused features into the intelligent question answering model; wherein the consultation results include intent classification results and / or dependency relationship results, the intent classification results are used to characterize the consultation intent of the entire original consultation data, and the dependency relationship results are used to characterize the dependency relationships between multiple keywords of the original consultation data.
[0012] Thirdly, embodiments of the present invention provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the points redemption method as described in the first aspect above.
[0013] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the points redemption method described in the first aspect above.
[0014] Fifthly, embodiments of the present invention provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the points redemption method as described in the first aspect above.
[0015] In this embodiment of the invention, by inputting the fusion features constructed based on the user's original consultation data into the intelligent question answering model, the model can accurately identify the semantic features of the consultation text and the dependency relationships between its keywords, thereby enhancing the model's ability to process semantic and numerical features and improving the accuracy of the consultation results produced by the intelligent question answering model. Attached Figure Description
[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:
[0017] Figure 1 This is a schematic diagram illustrating an application scenario of the points redemption method according to some embodiments of the present invention;
[0018] Figure 2 This is a flowchart illustrating a points redemption method according to some embodiments of the present invention;
[0019] Figure 3 This is a flowchart illustrating another points redemption method according to some embodiments of the present invention;
[0020] Figure 4 This is a schematic diagram of the structure of an integral knowledge graph according to some embodiments of the present invention;
[0021] Figure 5 This is a schematic diagram of the structure of a points redemption device according to some embodiments of the present invention; and
[0022] Figure 6 This is a schematic diagram of the structure of an electronic device according to some embodiments of the present invention. Detailed Implementation
[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] Figure 1 This is a schematic diagram illustrating an application scenario of the points redemption method according to some embodiments of the present invention. In some embodiments, such as Figure 1 As shown, application scenario 100 may include a user, user terminal 110, server 120, storage device 130, network 140, and / or processor.
[0025] User terminal 110 refers to the device that interacts with the user. The user (not in...) Figure 1 (As shown in the image) refers to individuals who need to redeem points. In some embodiments, user terminal 110 may include a PC, a mobile device, etc. In some embodiments, user terminal 110 can interact with other components in application scenario 100 (e.g., server 120, etc.) via network 140. For example, a user can upload the content they need to consult to server 120 via network 140 through user terminal 110, and server 120 can feed back the consultation results to user terminal 110 for the user to view at any time.
[0026] Server 120 can be used to process and store data involved in application scenario 100. For example, server 120 can receive raw inquiry data sent by users through user terminal 110. As another example, server 120 can send answers generated by intelligent question-answering models to user terminal 110 through network 140.
[0027] Storage device 130 can be used to store data involved in application scenario 100. For example, storage device 130 can store pre-trained intelligent question answering models, pre-built integral knowledge graphs, and / or pre-built corpora, etc. In some embodiments, the processor can store each user's original consultation data for each time in storage device 130, and store the consultation results generated by each intelligent question answering model in storage device 130.
[0028] Network 140 may include any suitable network 140 that facilitates the exchange of information and / or data. In some embodiments, one or more components of application scenario 100 may exchange information and / or data via network 140. For example, a processor may use network 140 to obtain raw consultation data uploaded by user terminal 110. In some embodiments, network 140 may include a local area network (LAN), a wide area network (WAN), a wired network 140, a wireless network 140, or any combination thereof.
[0029] Processor (not in) Figure 1 The processor (shown in the diagram) can be used to process data and / or information from various components and / or external data sources of application scenario 100. The processing device can execute program instructions based on this data, information, and / or processing results to perform one or more functions described in this invention. For example, the processor can execute the points redemption method described below. By way of example only, the processor may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), a special-purpose instruction processor (ASIP), a microprocessor, or any combination thereof. In some embodiments, the processor may be integrated into the components and / or external devices of application scenario 100.
[0030] It should be noted that application scenario 100 is provided for illustrative purposes only and is not intended to limit the scope of the invention. Those skilled in the art will recognize that various modifications or variations can be made based on the description provided. For example, application scenario 100 may also include a cloud computing platform. Furthermore, application scenario 100 may be implemented on other devices to achieve similar or different functions. However, these variations and modifications will not depart from the scope of the invention.
[0031] Figure 2 This is a schematic flowchart illustrating a points redemption method according to some embodiments of the present invention. For example... Figure 2 As shown, process 200 includes the following steps. In some embodiments, process 200 may be executed by a processor (e.g., the processor of a server).
[0032] In step 210, the user's original consultation data is obtained.
[0033] In step 220, based on the original consultation data and the user's account information, a structured feature of the user is generated, and based on the original consultation data, an unstructured feature of the original consultation data is generated.
[0034] In step 230, a fused feature is generated based on the structured feature and the unstructured feature.
[0035] In step 240, the fused features are input into the intelligent question-answering model to obtain the consultation result of the original consultation data generated by the intelligent question-answering model.
[0036] The following are examples related to step 210.
[0037] Original consultation data refers to data related to redeeming goods using points. In some embodiments, original consultation data may include multimedia data such as text and / or audio data, for example, "What goods can I redeem with my points?" or "How many points do I need to redeem?". In some embodiments, original consultation data may be input by a user through a user terminal. For example, a user can input original consultation data through a user terminal, and the user terminal uploads the original consultation data to a server via a network, where the server can receive the uploaded original consultation data.
[0038] The following are examples related to step 220.
[0039] In some embodiments, a user's account information may include the user's security data, such as the user's account name, biometrics, and / or password. In some embodiments, a user's account information may include the user's points data, such as the user's points balance, points validity period, and / or the user's points redemption history. In some embodiments, a user's account information may include membership data, such as the user's membership level and / or membership days. In some embodiments, a user's account information includes at least the user's points data.
[0040] Structured features refer to the characteristics used to characterize structured data. Structured data refers to data with a predefined fixed format and organization rules. In some embodiments, structured data may include fields such as a user's points balance, a user's membership level, and / or a product's ID. In some embodiments, structured data may be stored in a relational database (e.g., Oracle).
[0041] In some embodiments, the processor can generate the user's structured features by inputting the original consultation data and the user's account information into an points knowledge graph.
[0042] An integral knowledge graph is a knowledge graph that stores the relationships between items and points. A knowledge graph is a graph structure. In some embodiments, an integral knowledge graph may include multiple nodes and multiple edges. Nodes are used to store attributes (e.g., user attributes and / or item attributes), and edges are used to connect two nodes, storing dependencies between them. In some embodiments, edges may include undirected edges and directed edges.
[0043] Nodes in the points knowledge graph can be associated with users and / or products. For example, a product node might include the product name, the points required to redeem the product, the minimum membership level required to redeem the product, and / or the product category. Similarly, a user node might include the user's ID, the user's points balance, and / or the user's membership level. These are just examples. Figure 4 As shown, node 410 is the mapping of user ID 1001 in the points knowledge graph. Node 410 stores user UID 1001's points balance and membership level. In some embodiments, such as Figure 4 As shown, nodes can be stored in the knowledge graph as key-value pairs, such as {name: "headphones", required points: 3000, minimum level: "3"}, etc.
[0044] In some embodiments, the edges of the points knowledge graph can be stored in the relationships between nodes. For example, the edge between user A's user node and product B's product node can store the points relationship, membership level relationship, etc., between user A and product B. The points relationship refers to whether user A's points balance meets the requirements for purchasing product B, and the membership level relationship refers to whether user A's membership level meets the requirements for purchasing product B. For another example, such as... Figure 4 As shown, edge 420-1 stores the dependency relationship between the user corresponding to node 410 and the product corresponding to node 430-1. This is just an example. The value of the key "points balance" in node 410 is lower than the value of the key "required points" in node 430-1, and the value of the key "membership level" in node 410 is higher than the value of the key "minimum level" in node 430-1. Therefore, the dependency relationship stored in edge 420-1 is "points not satisfied" and "level satisfied". The processor can perform heat encoding on the dependency relationship stored in the edge. For example, the processor can set satisfied to 1 and unsatisfied to 0.
[0045] In some embodiments, the processor can input the user's original consultation data and the user's account information into the points knowledge graph to obtain the user's structured features. For example, the processor can parse multiple keywords from the original consultation data, and then input the parsed keywords and the user's account information into the points knowledge graph. The processor can then extract the paths between the keywords and the user from the knowledge graph, thereby generating structured features that can describe the original consultation data.
[0046] In some embodiments, the processor can map keywords in the original consultation data to an integral knowledge graph to obtain the node corresponding to the keyword. For example, the processor can use a natural language processing algorithm to extract the entity corresponding to the keyword and then map the entity to the integral knowledge graph.
[0047] Keywords refer to text in the original consultation data that carries important semantic meaning. Entities refer to standardized objects that need to be mapped to the points knowledge graph. For example, if the original consultation data is "Hello, what can I exchange my 5,000 points for?", then the keywords in this original consultation data include "5,000 points", "I", and "things". The processor can convert the keyword "5,000 points" into the entity 5,000 points, "I" into the user's ID, and "things" into the name of the product.
[0048] In some embodiments, the processor can parse multiple keywords from the original consultation data based on natural language processing algorithms. For example, the processor can first perform text cleaning on the original consultation data to obtain cleaned text, then perform word segmentation on the cleaned text to obtain multiple candidate keywords (e.g., person's name or synonyms of person and / or product name or synonyms of product, etc.), and then use a natural language processing library to identify the entities (e.g., numbers and / or names, etc.) of the multiple candidate keywords. In some embodiments, the processor can identify complex entities (e.g., product name and / or product points, etc.) based on a trained machine learning model (e.g., convolutional neural network model, etc.).
[0049] In some embodiments, the processor can perform text cleaning using regular expressions. In some embodiments, the processor can perform word segmentation using the NLTK (Natural Language Toolkit). In some embodiments, the processor can identify entities of candidate keywords using SpaCy. Both NLTK and SpaCy are open-source toolkits designed for natural language processing, integrating modules for handling common natural language processing tasks and standardized corpora.
[0050] In some embodiments, the processor can map the entities of candidate keywords to the integral knowledge graph based on the entities of the candidate keywords and the integral knowledge graph. For example, the processor can traverse the nodes of the integral knowledge graph, calculate the similarity between the entities of the candidate keywords and the nodes during each traversal, and set the node with the highest similarity as the mapping of the entity of the candidate keyword in the integral knowledge graph. In some embodiments, the processor can calculate the similarity between entities and nodes using string matching algorithms, fuzzy matching algorithms, and / or semantic similarity algorithms.
[0051] In some embodiments, the processor can map multiple keywords from the original query data to an points knowledge graph. For example, for the original query data "Can I redeem headphones with my points?", the processor can map "me", "points", and "headphones" from the original query data to user nodes and product nodes in the points knowledge graph, where the user node is the mapping between "me" and "points", and the headphone product node is the mapping between "headphones". For example, as... Figure 4 As shown, node 430-2 represents the mapping of the product "earphones" in the points knowledge graph, and node 430-1 represents the mapping of the product "watches" in the points knowledge graph.
[0052] In some embodiments, the processor can generate a structural feature vector that describes the original consultation data based on the mapping of the original consultation data in the integral knowledge graph. For example, the processor can extract paths and / or relationships related to the original consultation data from the integral knowledge graph, and then generate a structural feature vector based on the paths and / or relationships. In some embodiments, the processor can generate paths based on edges in the integral knowledge graph. For example, the processor can traverse the nodes connected to "User A" to search for edges that satisfy the integral value of the edge connecting "User A" and the keyword "5000 points" in the original consultation data, meaning that User A's points can purchase the product corresponding to the product node at the other end of the edge, and then determine other product nodes that meet the conditions, and set the set of edges connecting "User A" and other product nodes as the path between the user and the product.
[0053] In some embodiments, the processor can search for paths between multiple keywords mapped to nodes in an integral knowledge graph based on the original consultation data. For example, the processor can use a graph search algorithm (e.g., depth-first search or breadth-first search) to query the paths between the nodes corresponding to each keyword in the integral knowledge graph, such as the path between "User A" and "Product B" and "Product C".
[0054] In some embodiments, the processor can assign different relation weights to multiple attributes of an edge. For example, when the processor detects that the keyword "points" exists but "membership" does not, it can assign a higher relation weight to the points relation than to the membership level relation.
[0055] In some embodiments, the processor can construct a structured feature vector based on the searched path. For example, the processor can first perform hot encoding of the attributes of the edges and nodes in the path, then concatenate them, and finally generate a structured feature vector using a graph embedding model. As an example only, the processor can generate a structured feature vector based on the following formula.
[0056]
[0057] Among them, V Structured For the generated structured vectors, Embed() refers to the graph embedding model (e.g., the TransE model), P is the set of paths, and W... j e represents the weight of the edge. i For the starting entity (e.g., a user node, etc.), r jLet e be the attribute of the j-th edge in the path (e.g., points relationship, membership relationship, etc.). k is the target entity (e.g., product node, etc.), and ⊕ is the vector concatenation operator.
[0058] In some embodiments, the processor can construct an points knowledge graph based on the points of the products. For example, the processor can first generate an empty knowledge graph, then copy all products and their corresponding attributes (e.g., storing products and their attributes as key-value pairs) to various nodes of the empty knowledge graph to obtain multiple product nodes, and copy users and their corresponding attributes (e.g., storing users and their attributes as key-value pairs) to various nodes of the empty knowledge graph to obtain multiple user nodes. Then, based on the relationships between user nodes and product nodes (e.g., points relationships, membership level relationships, etc.), the processor constructs the attributes of the edges of the empty knowledge graph to obtain the points knowledge graph. This is just an example. Figure 4 As shown, the key-value pairs can be in the form of {"User ID":"UID1001","Points Balance":4000,"Member Level":5}, {"Product ID":"Headphones","Points Required for Redemption":3000,"Minimum Member Level Required for Redemption":3}, and the edges can be in the form of {"Points Balance"≥"Points Required for Redemption"}.
[0059] In some embodiments of the present invention, by associating user consultation data with user account information (e.g., user's points balance) through an points knowledge graph, the automated generation of structured features can be achieved, improving the efficiency and accuracy of structured feature extraction. Through the reasoning ability of the knowledge graph on the relationships between entities, the user's own conditions (e.g., user's membership level) and the redemption conditions of goods (e.g., the minimum membership level required to redeem a certain goods) can be dynamically matched. This not only solves the problem of excessively high costs in manually constructing structured features, but also enhances the intelligent question-answering model's ability to process complex query data.
[0060] Unstructured features refer to features used to characterize unstructured data. Unstructured data refers to data without a fixed format and / or without pre-defined logical rules. For example, text and / or audio corresponding to raw consultation data. In some embodiments, unstructured features may include unstructured feature vectors extracted from unstructured data.
[0061] In some embodiments, the processor can obtain the unstructured features of the original consultation data by inputting the original consultation data into a consultation corpus. For example, the processor can input the original consultation data into the consultation corpus and retrieve the text paragraphs and / or documents that best match the original consultation data from the consultation corpus. As an example only, the processor can parse the user's original consultation data to extract keywords and / or phrases (e.g., product names, user points, etc.); then, the processor can automatically construct a search query to retrieve the text paragraphs and / or documents in the consultation corpus that are most relevant to these keywords and / or phrases. In some embodiments, the processor can use a text matching algorithm (e.g., cosine similarity algorithm, etc.) to evaluate the relevance between the text paragraphs and / or documents in the consultation corpus and the keywords and / or phrases, thereby identifying the text paragraphs and / or documents with the highest relevance as the best-matching text paragraphs and / or documents.
[0062] In some embodiments, the processor can extract semantic features from the most matching text paragraphs and / or documents that help the processor understand the original query. For example, the processor can use a word frequency statistics algorithm (e.g., TF-IDF algorithm, etc.) to count the importance of each keyword and / or each phrase in the text paragraph and / or document, and then use a word embedding model (e.g., Word2Vec model, etc.) to convert the text paragraph and / or document into a numerical vector to capture the contextual relationships of keywords and / or phrases.
[0063] In some embodiments, the processor can generate unstructured feature vectors corresponding to the original consultation data based on the contextual relationships of keywords and / or phrases. For example, the processor can generate unstructured feature vectors based on the following formula.
[0064] V Unstructured =a·Norm(V TF-IDF )⊕(1-a)·V embed (2)
[0065] Among them, V Unstructured The generated unstructured feature vector, where a is the weight coefficient, and V TF-IDF The TF-IDF model calculates the word vectors corresponding to the importance of each keyword and / or phrase in the text paragraph and / or document, with Norm() as the normalization function, and V... embed is the average word vector, and ⊕ is the vector concatenation operator.
[0066] In some embodiments, the processor can set the generated unstructured feature vector as the unstructured feature corresponding to the original consultation data.
[0067] A consultation corpus refers to a corpus that includes information related to points-for-redeem products. In some embodiments, the consultation corpus may include multiple text paragraphs and / or documents related to points-for-redeem products.
[0068] In some embodiments, the processor can construct a consultation corpus based on multi-source heterogeneous data. For example, the processor can first clean the multi-source heterogeneous data, and then index the processed text to construct the consultation corpus. Multi-source heterogeneous data refers to data from multiple sources related to points redemption for goods. For example, the processor can obtain data related to points redemption, such as question and answer texts, product descriptions, user reviews, help documents, and / or policy statements, from storage devices (e.g., databases on storage devices).
[0069] In some embodiments, cleaning multi-source heterogeneous data may include: removing irrelevant information from the multi-source heterogeneous data, correcting typos in the multi-source heterogeneous data, and / or unifying terminology expressions in the multi-source heterogeneous data, so as to enrich the structure and query efficiency of the consultation corpus.
[0070] In some embodiments, indexing the processed text may include: the processor calling modern text indexing techniques (e.g., indexing tools such as Kiban or Logstash) to build an inverted index that supports Boolean queries and fuzzy matching on fields such as product category and product points to obtain a consultation corpus.
[0071] In some embodiments of the present invention, extracting unstructured features from the original consultation data through a consultation corpus can effectively parse the ambiguous semantics in user consultations (e.g., "things-products"), making up for the shortcomings of conventional unstructured feature extraction methods in natural language understanding; the consultation corpus constructed based on the points-related text of products can associate business keywords (e.g., products, points, etc.) with the user's original consultation data; the unstructured features generated based on the consultation corpus and the user's original consultation data can improve the accuracy of the intelligent question answering model in recognizing the user's true consultation intent.
[0072] The following are examples related to step 230.
[0073] In some embodiments, the processor may perform feature fusion on the structured features, the unstructured features, and the original consultation data to obtain the fused features.
[0074] In some embodiments, before feature fusion, the processor can align the dimensions of structured features, unstructured features, and the original consultation data. For example, the processor can map the structured feature vectors corresponding to the structured features, the unstructured feature vectors corresponding to the unstructured features, and the original feature vectors corresponding to the original consultation data to have the same dimensions (e.g., horizontal or vertical dimensions). More information about structured and unstructured feature vectors can be found in the preceding descriptions and will not be repeated here.
[0075] In some embodiments, the processor can transform the raw consultation data into raw feature vectors (i.e., vector form) based on a word embedding algorithm. For example, the processor can perform the vector form transformation based on the following formula.
[0076] V r (d)=[F(t1,d),F(t2,d),…,F(t n ,d)] (3)
[0077] Among them, V r (d) represents the original feature vector corresponding to the generated original consultation data, F() is the term frequency statistics function (e.g., TF-IDF term frequency statistics function), t n Let f be the nth word, and f be the dimension corresponding to the original consultation data.
[0078] In some embodiments, the processor can fuse the original feature vector, structured feature vector, and unstructured feature vector in various ways to obtain a fused feature vector. For example, the processor can concatenate the original feature vector, structured feature vector, and unstructured feature vector into a single vector to obtain the fused feature vector. As an example only, the processor can obtain the fused feature vector based on the following formula.
[0079] V fused =[V r (d)⊕FC(V Structured )⊕FC(V Unstructured (4)
[0080] Among them, V fused To fuse feature vectors, FC() is a fully connected layer, and ⊕ is the vector concatenation operator. In some embodiments, the processor can... r (d), V Structured 、 and V Unstructured The concatenation is performed along the row direction, and the resulting fused vector has a size of 1×(d+m+m), where m is V. Structured and V Unstructured The size, i.e., the number of entities or keywords in the original consultation data. In some embodiments, the processor can...r (d), V Structured 、 and V Unstructured The concatenation is performed along the column direction, and the resulting fused vector has a size of 1×d.
[0081] In some embodiments, the processor can set the fused feature vector as the fused feature.
[0082] In some embodiments described herein, by fusing structured features (e.g., integral relationships) with unstructured features (e.g., the semantics of the original consultation data), not only is the intelligent question-answering model's understanding of the user's consultation intent enhanced, but the problem of missing semantic associations under a single feature dimension is also solved, thereby improving the accuracy of points redemption for goods.
[0083] In some embodiments, the processor can reduce the dimensionality of the fused features using dimensionality reduction algorithms to alleviate the reasoning burden on the intelligent question-answering model. For example, the processor can use Principal Components Analysis (PCA) to remove non-critical information from the fused feature vectors corresponding to the fused features.
[0084] The following are examples related to step 240.
[0085] An intelligent question-answering model is a model used to answer user inquiries (e.g., raw inquiry data). In some embodiments, an intelligent question-answering model may include a trained neural network, such as a trained graph convolutional neural network (GNN).
[0086] In some embodiments, the processor can store the fused vectors as a heterogeneous graph. For example, the processor can construct the heterogeneous graph corresponding to the fused vectors using the following formula.
[0087] G = (V fused (5)
[0088] Among them, V fused E is the fusion vector, representing the attributes of entities in the heterogeneous graph. E is the set of edges, representing the attributes of the edges between entities in the heterogeneous graph. A is the adjacency matrix between entities, representing the connection relationship between entities in the heterogeneous graph (e.g., 0 represents non-connection, 1 represents connection).
[0089] In some embodiments, the intelligent question-answering model can answer user inquiries based on fused features. For example, the processor inputs a fused feature vector in the form of a heterogeneous graph to the input of the intelligent question-answering model. The intelligent question-answering model can extract convolutional features from the heterogeneous graph using the convolutional layers of its graph convolutional neural network, then assign different weights to each sub-feature of the convolutional features through a graph attention mechanism, and finally output the weighted convolutional features to an activation function (e.g., a SoftMax activation function) to generate a consultation result for the original consultation data.
[0090] In some embodiments, the consultation result includes intent classification results and / or dependency relationship results. Intent classification results characterize the consultation intent of the entire original consultation data. Dependency relationship results characterize the dependency relationships between multiple keywords in the original consultation data. For example, if a processor inputs the original consultation data "How many points do I have left? What products can I redeem?" into an intelligent question-answering model, the consultation result output by the intelligent question-answering model may include intent classification results "user-points" and "product-points," and the corresponding output dependency relationship results may include "user-acquisition-product."
[0091] In some embodiments, the intelligent question-answering model is trained based on an intelligent question-answering training set and intelligent question-answering labels. For example, a processor can train a graph convolutional neural network using the intelligent question-answering training set and intelligent question-answering labels to obtain the intelligent question-answering model.
[0092] In some embodiments, the intelligent question-answering training set includes the user's historical raw consultation data. For example, the processor can retrieve the user's raw consultation data from any past time from a storage device.
[0093] In some embodiments, the processor can convert historical raw consultation data into syntactic graphs, and then use these syntactic graphs as a training set for intelligent question answering. For example, the processor can set all words from all users in the history as nodes in the syntactic graph, with empty edges between these nodes. In some embodiments, the syntactic graph is heterogeneous with the aforementioned... Figure 1 A graph structure like this, consisting of nodes and edges.
[0094] In some embodiments, the processor can set the dependency relationship between the consultation intent corresponding to all historical raw consultation data that has been manually verified and the keywords of all historical raw consultation data that have been manually verified as the training label corresponding to the historical raw consultation data. For example only, the consultation intent corresponding to the raw consultation data "How many points do I have left? What products can I redeem?" is "user-points" and "product-points", and the corresponding dependency relationship result is "user-acquisition-product".
[0095] In some embodiments of the present invention, by training a graph convolutional neural network based on historical original consultation data and the consultation intent and dependency relationships corresponding to the historical original consultation data, the intelligent question answering model's ability to understand the user's true consultation intent is improved, thereby improving the accuracy of the answer.
[0096] In some embodiments, such as Figure 3 As shown, the processor can recommend products to the user based on the following steps.
[0097] In step 310, based on the user's historical raw consultation data, the user's behavior analysis results are obtained.
[0098] In some embodiments, the processor can obtain user behavior analysis results based on the user's historical raw consultation data. For example, the processor can statistically analyze the correlation between high-frequency words in the user's historical raw consultation data and subsequent redemption operations, and store closely related redemption operations and the remaining points at the time of redemption in the user's behavior analysis results, so as to make subsequent product recommendations based on the behavior analysis results. In some embodiments, the processor can generate a user behavior profile based on the user's historical raw consultation data and historical redemption operations, so that the processor can subsequently update the intelligent question-answering model based on this user behavior profile.
[0099] In step 320, products are recommended to the user based on the user's behavior analysis results.
[0100] In some embodiments, the processor can recommend products to users based on user behavior analysis results. For example, if the processor finds that the high-frequency words in a user's historical raw consultation data are "points balance," "membership level," and "product B," and the user subsequently redeems "product B" more frequently, then the processor can assume that the user prefers to redeem product B, and can increase the frequency of recommending product B in the future.
[0101] In some embodiments, the processor can collect the user's consultation text and the results generated by the intelligent question-answering model in real time, correct the generated results through manual verification, and then use the consultation text and the corrected results to update the parameters of the intelligent question-answering model to ensure that the intelligent question-answering model can update itself and form a complete closed-loop feedback mechanism from consultation to question answering.
[0102] In some embodiments of the present invention, by generating behavioral analysis results of users' points redemption of goods based on users' historical consultation data, not only can personalized product recommendations be realized, but also the recommendation strategy can be dynamically adjusted based on users' past consultation preferences, forming a closed-loop service mechanism of "consultation-behavioral analysis-precise recommendation", thereby improving the accuracy and user experience of points redemption services.
[0103] In some embodiments of the present invention, by inputting fusion features constructed based on users' original consultation data into the intelligent question-answering model, accurate identification of the semantic features of the consultation text and the dependency relationships between keywords in the consultation text is achieved. Furthermore, the fusion features simultaneously include structured and unstructured features, enabling the intelligent question-answering model to process both numerical and semantic features, significantly improving the efficiency and convenience of the points redemption service. Structured features generated based on the original consultation data and users' account information (e.g., points balance) provide high-quality data support for the intelligent question-answering model. By extracting unstructured features from the original consultation data, the semantic information of the consultation text can be fully preserved, ensuring that the fusion features input into the intelligent question-answering model possess the semantic features of the user's consultation text. Through this multi-feature fusion processing mechanism, the high accuracy of the consultation results output by the intelligent question-answering model can be ensured.
[0104] Please refer to Figure 5 This invention also provides a points redemption device 500, which includes an acquisition module 510, a feature extraction module 520, and an intelligent question-answering module 530.
[0105] In some embodiments, the acquisition module 510 may be used to acquire the user's original consultation data; wherein, the original consultation data includes consultation data related to redeeming goods with points.
[0106] In some embodiments, the feature extraction module 520 can be used to generate structured features of the user based on the original consultation data and the user's account information, and to generate unstructured features of the original consultation data based on the original consultation data; wherein, the user's account information includes at least the user's points data, the structured features are used to characterize the structured data, and the unstructured features are used to characterize the unstructured data.
[0107] In some embodiments, the feature extraction module 520 can be further used to generate structured features of the user by inputting the original consultation data and the user's account information into the points knowledge graph; wherein, the points knowledge graph is constructed based on the points of the product.
[0108] In some embodiments, the feature extraction module 520 can be further used to obtain unstructured features of the original consultation data by inputting the original consultation data into a consultation corpus; wherein the consultation corpus is constructed based on text data related to the points of the product.
[0109] In some embodiments, the feature extraction module 520 can be further used to perform feature fusion on structured features, unstructured features and original consultation data to obtain fused features.
[0110] In some embodiments, the intelligent question answering module 530 can be used to obtain the consultation results of the original consultation data generated by the intelligent question answering model by inputting the fused features into the intelligent question answering model; wherein, the consultation results include intent classification results and / or dependency relationship results, the intent classification results are used to characterize the consultation intent of the entire original consultation data, and the dependency relationship results are used to characterize the dependency relationship between multiple keywords in the original consultation data.
[0111] In some embodiments, the intelligent question-answering module 530 can be further used to obtain user behavior analysis results based on the user's historical raw consultation data; and to recommend products to the user based on the user behavior analysis results.
[0112] Please refer to Figure 6 The present invention also provides an electronic device 600, including a processor 610, a memory 620, and a computer program stored in the memory 620 and executable on the processor 610. When the computer program is executed by the processor 610, it implements the various processes of the above-described points redemption method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here.
[0113] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described points redemption method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.
[0114] This application also provides a computer program product, including computer instructions. When executed by a processor, the computer instructions implement the various processes of the above-described points redemption method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.
[0115] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0117] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.
Claims
1. A points redemption method, characterized in that, include: Obtain the user's original consultation data; wherein, the user's original consultation data includes consultation data related to redeeming goods with points; Based on the original consultation data and the user's account information, a structured feature of the user is generated, and an unstructured feature of the original consultation data is generated based on the original consultation data. Based on the structured features and the unstructured features, a fused feature is generated; and By inputting the fused features into the intelligent question-answering model, the consultation results of the original consultation data generated by the intelligent question-answering model are obtained; wherein, the consultation results include intent classification results and / or dependency relationship results, the intent classification results are used to characterize the consultation intent of the entire original consultation data, and the dependency relationship results are used to characterize the dependency relationship between multiple keywords of the original consultation data.
2. The points redemption method according to claim 1, characterized in that, The process of generating the user's structured characteristics based on the original consultation data and the user's account information includes: By inputting the original consultation data and the user's account information into the points knowledge graph, the user's structured features are generated; wherein, the points knowledge graph is constructed based on the points of goods.
3. The points redemption method according to claim 1, characterized in that, Based on the original consultation data, unstructured features of the original consultation data are generated, including: The unstructured features of the original consultation data are obtained by inputting the original consultation data into a consultation corpus; wherein the consultation corpus is constructed based on text data related to product points.
4. The points redemption method according to claim 1, characterized in that, The intelligent question-answering model is trained based on an intelligent question-answering training set and intelligent question-answering tags; wherein, the intelligent question-answering training set includes the user's historical original consultation data, and the intelligent question-answering tags include the consultation intent corresponding to the user's historical original consultation data and the dependency relationship between multiple keywords in the user's historical original consultation data.
5. The points redemption method according to claim 1, characterized in that, The generation of fused features based on the structured features and the unstructured features includes: The structured features, the unstructured features, and the original consultation data are fused to obtain the fused features.
6. The points redemption method according to claim 1, characterized in that, Also includes: Based on the user's historical original consultation data, obtain the user's behavioral analysis results; as well as Based on the user's behavior analysis results, products are recommended to the user.
7. A points redemption device, characterized in that, include: The acquisition module is used to acquire the user's original consultation data; wherein, the user's original consultation data includes consultation data related to redeeming goods with points; A feature extraction module is used to generate structured features of the user based on the original consultation data and the user's account information, and to generate unstructured features of the original consultation data based on the original consultation data; wherein the user's account information includes at least the user's points data, the structured features are used to characterize structured data, and the unstructured features are used to characterize unstructured data; and a fused feature is generated based on the structured features and the unstructured features; and The intelligent question answering module is used to obtain the consultation results of the original consultation data generated by the intelligent question answering model by inputting the fused features into the intelligent question answering model; wherein, the consultation results include intent classification results and / or dependency relationship results, the intent classification results are used to characterize the consultation intent of the entire original consultation data, and the dependency relationship results are used to characterize the dependency relationship between multiple keywords of the original consultation data.
8. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the points redemption method as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the points redemption method as described in any one of claims 1 to 6.
10. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the points redemption method as described in any one of claims 1 to 6.