E-commerce intelligent customer service optimization method, related device, storage medium and program product

Abstract

The invention relates to the field of large models, and provides an e-commerce intelligent customer service optimization method, a related device, a storage medium and a program product, which are used for an e-commerce customer service system and comprise a federal query engine and an intelligent customer service. The intelligent customer service integrates an event knowledge graph and a plurality of static knowledge graphs which are federated nodes; the method comprises the steps that if it is determined that a commodity query intention conforms to a multi-hop query mode based on a federated query engine, cross-knowledge-graph joint query is conducted, a first class of results based on all static knowledge graphs and a second class of results based on event knowledge graphs are obtained, and the first class of results and the second class of results are fused and then transmitted into an intelligent customer service; and the intelligent customer service generates a multi-hop reasoning chain based on the fusion result, generates an initial answer according to the combination of the multiple reasoning chains and the fusion result, adjusts the temperature parameter of each hop to enhance the initial answer according to the knowledge graph coverage rate and confidence of the fusion result and the number of reasoning hops, and obtains and outputs a target answer. According to the scheme, the expansibility and accuracy of intelligent customer service and the questioning and answering ability can be balanced.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of large model technology, and in particular to an e-commerce intelligent customer service optimization method, related devices, storage media and program products. Background Art

[0002] In the field of e-commerce customer service, customer service robots have become an indispensable tool. Customer service robots will automatically answer users' questions, improve customer service efficiency, and respond to users in a timely manner to improve user experience.

[0003] During research and practice in the existing technologies, the inventors of the present embodiments discovered that, in vertical domains, the quality of responses from intelligent customer service systems based on large models is currently mediocre. To enhance their professionalism, domain knowledge graphs (KGs) are often embedded in intelligent customer service systems for training. While this significantly improves the accuracy of intent recognition and answer generation, the responses provided by intelligent customer service systems embedded with domain KGs are limited by the domain KG itself. Summary of the Invention

[0004] The embodiments of the present application provide an e-commerce intelligent customer service optimization method, related devices, storage media and program products, which can balance the scalability and accuracy of intelligent customer service, as well as the commission-based question-answering capabilities, and solve the long-tail problem.

[0005] In a first aspect, embodiments of the present application provide an e-commerce intelligent customer service optimization method from the perspective of a business server. The method is applied to an e-commerce customer service system of an e-commerce platform. The e-commerce customer service system includes a federated query engine and intelligent customer service. The intelligent customer service integrates an event knowledge graph and static knowledge graphs of multiple business types. The event knowledge graph and each static knowledge graph are federated nodes. The method includes:

[0006] Receiving a first input from the user on the conversation interface;

[0007] In response to the user's first input, if the federated query engine pre-calculates the product query intent corresponding to the first input and determines that it meets the multi-hop query mode, a cross-knowledge graph joint query is performed on the product query intent to obtain a first type of result based on each static knowledge graph and a second type of result based on the event knowledge graph;

[0008] The first and second types of results are integrated through multi-party secure computation, and the integrated results are transmitted to the intelligent customer service;

[0009] Based on the fusion results, the intelligent customer service generates a multi-hop reasoning chain, and generates an initial answer according to the multiple reasoning chains combined with the fusion results. According to the knowledge graph coverage and confidence of the fusion results, the temperature parameter of each hop is adjusted according to the number of reasoning hops, the initial answer is enhanced, and the target answer is obtained and output.

[0010] In some embodiments, the method further comprises:

[0011] Build a similarity index based on the product attribute graph;

[0012] Align new product entities with similar products in terms of attributes;

[0013] The association relationships of similar products are migrated to the new product entity, and the weights of the association relationships of the new product entity are dynamically adjusted.

[0014] In a second aspect, embodiments of the present application provide an e-commerce intelligent customer service optimization device that implements the functions corresponding to the e-commerce intelligent customer service optimization method provided in the first aspect. These functions can be implemented in hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the aforementioned functions, and the modules can be software and / or hardware.

[0015] In some implementations, the e-commerce intelligent customer service optimization device is applied to the e-commerce customer service system of an e-commerce platform, which includes a federated query engine and intelligent customer service; the intelligent customer service integrates an event knowledge graph and static knowledge graphs of multiple business types, where both the event knowledge graph and each static knowledge graph are federated nodes. The e-commerce intelligent customer service optimization device includes:

[0016] An input / output module, configured to receive a first input from a user on the conversation interface;

[0017] a federated query module for responding to a first input received by the input / output module, and if the federated query engine pre-calculates the product query intent corresponding to the first input and determines that the multi-hop query mode is met, then performing a cross-knowledge graph joint query on the product query intent to obtain a first type of result based on each static knowledge graph and a second type of result based on the event knowledge graph;

[0018] Based on multi-party secure computing, the first and second types of results are integrated and the fused results are passed into the inference module.

[0019] The reasoning module is used to generate a multi-hop reasoning chain based on the fusion results obtained by the federated query module, and generate an initial answer according to the multiple reasoning chains combined with the fusion results. According to the knowledge graph coverage and confidence of the fusion results, the temperature parameter of each hop is adjusted according to the number of reasoning hops, the initial answer is enhanced, and the target answer is obtained and output through the input and output module.

[0020] In some implementations, the reasoning module is specifically configured to:

[0021] If the knowledge graph coverage of the first-hop reasoning is lower than the preset coverage and the confidence is lower than the preset confidence, then the temperature parameter of the first-hop is increased to supplement the content not covered by the static knowledge graph;

[0022] Alternatively, if the knowledge graph coverage of the first-hop reasoning is higher than the preset coverage and the confidence is higher than the preset confidence, the temperature parameter of the second hop is lowered, and the second hop is the next hop of the first hop.

[0023] In some implementations, the reasoning module is further configured to:

[0024] If the confidence of the first type of results returned by any knowledge graph in the multi-hop path is lower than the preset confidence, the temperature parameters of the subsequent hops are reduced.

[0025] In some implementations, the reasoning module is further configured to:

[0026] Based on a dynamic update strategy, detecting update events of the event knowledge graph, and synchronizing the update events to the event knowledge graph through a streaming graph;

[0027] Alternatively, the static knowledge graph or the event knowledge graph is automatically corrected based on the real-time feedback of the user on the target answer.

[0028] In some implementations, the reasoning module is further configured to:

[0029] Migrate the association relationships of similar products with similar attributes to the new product to the new product, and adjust the relationship weights based on the characteristics of the new product;

[0030] Generate multiple query samples based on the new product attributes and the migrated association relationship;

[0031] Inject the association relationship of similar products into each query sample;

[0032] Perform a federated query on the federated knowledge graph for each query sample to obtain multiple candidate answers, and mark conflicting answers among the multiple candidate answers that are inconsistent with the federated knowledge graph;

[0033] Calculate the cumulative probability of all conflicting answers for each query sample;

[0034] Aggregate the total cumulative probability of each query sample according to the knowledge consistency penalty parameter value;

[0035] Calculating target loss based on the total cumulative probability;

[0036] If the knowledge graph coverage is greater than the preset coverage and the average conflict probability of the conflicting answers is greater than the first threshold, the knowledge consistency penalty parameter value is added to the conflicting answer to update the target loss until the proportion of the knowledge consistency penalty parameter value to the total loss is less than the preset proportion, and the operation is stopped, wherein the knowledge consistency penalty parameter represents the matching rate between the conflicting answer and the said federated knowledge graph.

[0037] In some implementations, the federated query module is further configured to: obtain an initial e-commerce domain knowledge graph from multiple channels, as well as dynamic data of the e-commerce platform;

[0038] Splitting the initial e-commerce domain knowledge graph into multiple static knowledge graphs according to business types, adding logical labels to multi-hop paths in the sub-knowledge graphs; and constructing an event knowledge graph based on the dynamic data;

[0039] Establish entity correspondence across knowledge graphs;

[0040] The event knowledge graph and the multiple static knowledge graphs are respectively set as federation nodes.

[0041] In some implementations, the reasoning module is specifically configured to:

[0042] Determining multiple entities in the product query intent;

[0043] Linking the first entity to the user data node to query product purchase data, linking the second entity to the product data node to verify product parameters, and linking the third entity to the service data node to query after-sales outlets;

[0044] Connect product purchase data with after-sales outlets within a preset range of the user's current location in time and space;

[0045] Determine target outlets that meet product after-sales conditions based on time and space connections.

[0046] In a third aspect, an embodiment of the present application provides an e-commerce intelligent customer service optimization device for implementing the above-mentioned e-commerce intelligent customer service optimization method in real time, and the e-commerce intelligent customer service optimization device includes: at least one processor and a memory; wherein the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the steps in the above-mentioned first aspect, any one of the implementation methods of the first aspect, or any one of the e-commerce intelligent customer service optimization methods provided in the second aspect.

[0047] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having the function of implementing the e-commerce intelligent customer service optimization method provided in the first aspect above. The function can be implemented by hardware, or the corresponding software can be implemented by hardware. The hardware or software includes one or more modules corresponding to the above functions, and the modules can be software and / or hardware. Specifically, the computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for the processor to load to execute the steps in any one of the implementation methods of the first aspect and the first aspect in the embodiment of the present application.

[0048] Compared with the prior art, the solution provided by the embodiment of the present application sets the event knowledge graph and multiple static knowledge graphs of different business types as federated nodes respectively. In this way, when a user inquires, when it is determined that the user's product query intention meets multiple query patterns, a cross-domain KG combination can be achieved through federated query. On the one hand, each knowledge graph is deployed independently, especially the event knowledge graph is isolated from the static knowledge graph, the cross-domain knowledge acquisition of the federated query engine is decoupled from the prompt engineering of the intelligent customer service, and cross-domain knowledge of multi-hop reasoning is provided, so that each knowledge graph performs its own function, which can reduce the privacy risk during public query. On the other hand, federated query can improve the response speed and answer scalability. Under the premise of ensuring the autonomous query of each knowledge graph, the accurate question-answering capability of the global intelligent customer service is realized, and the multi-hop query covers complex problems, which can break through the information limitations of a single knowledge graph. On the other hand, by hierarchically controlling the temperature parameters according to the number of inference hops and structuring the inference chain of each hop path, it can ensure that the inference chain meets both flexibility and reliability, that is, the controllable generation of intelligent customer service answers is achieved.

[0049] It can be seen that by implementing multi-hop reasoning across federation domains, dynamically generating multiple reasoning chains, and evaluating the results of the covered knowledge graph in each hop to dynamically adjust the temperature parameters, through the collaboration of these three, in a balance between scalability (federation multi-hop) and accuracy (temperature control), it is possible to achieve the upgrade of intelligent customer service from "single point response" to "chain reasoning". BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of an application environment provided in an embodiment of the present application;

[0051] Figure 2a A schematic diagram of a process for pre-processing between training intelligent customer service in an embodiment of the present application;

[0052] Figure 2b A flowchart of the intelligent customer service training method in an embodiment of the present application;

[0053] Figure 3 A flow chart of the e-commerce intelligent customer service optimization method in an embodiment of the present application;

[0054] Figure 4 This is a schematic diagram of a reasoning example for optimizing the answers of e-commerce intelligent customer service in an embodiment of this application;

[0055] Figure 5 This is a structural diagram of an e-commerce intelligent customer service optimization device in an embodiment of the present application;

[0056] Figure 6 A schematic diagram of the structure of a physical device for implementing the e-commerce intelligent customer service optimization method in an embodiment of the present application;

[0057] Figure 7 A structural diagram of a server for implementing an e-commerce intelligent customer service optimization method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0058] The terms "first," "second," and the like in the description and claims of the embodiments of this application and the accompanying drawings are used to distinguish similar objects (e.g., the first and second categories of results in the embodiments of this application represent candidate answers to user queries based on different knowledge graphs, respectively), and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those steps or modules that are clearly listed, but may include other steps or modules that are not clearly listed or that are inherent to these processes, methods, products, or devices. The division of modules that appears in the embodiments of the present application is merely a logical division. In actual applications, there may be other division methods. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between each other shown or discussed can be through some interfaces, and the indirect coupling or communication connection between modules can be electrical or other similar forms, which are not limited in the embodiments of the present application. Moreover, the modules or submodules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed in multiple circuit modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment of the present application.

[0059] The embodiments of the present application provide an e-commerce intelligent customer service optimization method, related devices and storage media, which can be used in servers or terminal devices. Specifically, they can be used in scenarios such as cross-platform after-sales and supply chain consulting.

[0060] In some embodiments, this solution is applied to Figure 1 The communication system shown in FIG. 1 may include at least one service server, at least one user terminal ( Figure 1 Only one is marked in the figure. The business server includes an e-commerce customer service system, which deploys an e-commerce intelligent customer service optimization device. This device includes a federated query engine and multiple intelligent customer service agents. Each intelligent customer service agent has an embedded e-commerce domain knowledge graph for different business types (including static knowledge graphs and event knowledge graphs for multiple business types). The intelligent customer service agent includes a large model generation layer and a temperature parameter regulator.

[0061] Business Server: This server can be used to provide various data searches to the e-commerce customer service system, as well as provide users with product search and consultation, after-sales service, supply chain consultation, and other business services. This server can be used to deploy an e-commerce intelligent customer service optimization device.

[0062] When based on Figure 1 When the communication system shown implements the above-mentioned e-commerce intelligent customer service optimization method, after receiving a user inquiry from a user terminal, a federated query is performed on the static knowledge graphs and event knowledge graphs participating in the federated query through the federated query engine, the queried KG knowledge is fused, and the fusion result is input into the intelligent customer service. The intelligent customer service generates candidate answers based on the fusion result, and then dynamically adjusts the candidate answers based on the knowledge graph coverage and knowledge confidence, and finally obtains the target answer and outputs it to the user terminal.

[0063] The solutions provided in the embodiments of this application involve technologies such as artificial intelligence (AI), natural language processing (NLP), and machine learning (ML).

[0064] Among them, it should be specially noted that the server involved in the embodiment of the present application (such as a business server, a search engine) can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, as well as basic cloud computing services such as big data and artificial intelligence platforms. The e-commerce intelligent customer service optimization device involved in the embodiment of the present application can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, a personal digital assistant, etc., but is not limited to this. The e-commerce intelligent customer service optimization device and the server can be directly or indirectly connected through wired or wireless communication, and the embodiment of the present application does not limit this.

[0065] In response to multiple technical issues caused by knowledge graph-based intelligent customer service conversations, the embodiments of this application mainly adopt the following technical solutions:

[0066] 1. To address the issue of multiple relationship reasoning failures caused by complex knowledge graph traversal, we split the knowledge graph into sub-KGs, separated dynamic rules from static sub-KGs, and iteratively aggregated multi-hop neighbor features through GNNs to capture implicit relationship weights and conduct joint queries across sub-KGs. This improves response speed and multi-hop reasoning depth.

[0067] 2. Mixed queries of real-time data and static KGs can cause slow response times. By modeling real-time data as a temporary dynamic subgraph and logically connecting it with the static KG, static KG queries and real-time data requests can be initiated asynchronously and simultaneously.

[0068] 3. Due to the reduced scalability of answer generation caused by KG-based accuracy reasoning, this can be compensated by dynamically adjusting the temperature parameters of intelligent customer service based on knowledge graph coverage and knowledge confidence, thereby improving answer scalability and retaining answer accuracy.

[0069] 4. To address issues such as incorrect answers due to delayed KG updates, dynamic rules can be separated from the static KG and injected into the dialogue system in real time. Alternatively, a dynamic update mechanism can be used to monitor product library changes. This can reduce answer delays and improve scalability.

[0070] The following combination Figure 2a-Figure 7 The technical solutions of the embodiments of the present application are exemplified.

[0071] Since the e-commerce intelligent customer service optimization method of the embodiment of the present application is based on the federated query engine and intelligent customer service to complete the user's product inquiries, the pre-deployment process is introduced before introducing the e-commerce intelligent customer service optimization method. The method is applied to the e-commerce customer service system of the e-commerce platform, which includes a federated query engine and intelligent customer service; the intelligent customer service integrates the event knowledge graph and the static knowledge graphs of multiple business types, and the event knowledge graph and each static knowledge graph are all federated nodes. Figure 2a As shown, Figure 2a A schematic diagram of the preprocessing process for training intelligent customer service. The preprocessing process includes:

[0072] 101. Obtain the initial e-commerce knowledge graph from multiple channels, as well as dynamic data of e-commerce platforms.

[0073] The initial e-commerce knowledge graph can be derived from user behavior data and product-related knowledge graphs across different e-commerce retail platforms. For example, when cross-platform user behavior data is included, cross-platform after-sales service can be performed through federated knowledge graph queries. Another example is when cross-platform product-related knowledge graphs are included, cross-supply chain consulting can be performed through federated knowledge graph queries.

[0074] Dynamic data includes promotional rules, inventory changes, price adjustments, logistics dynamics, user reviews, and product popularity data with strong timeliness.

[0075] 102. Split the initial e-commerce domain knowledge graph into multiple static knowledge graphs according to business types, add logical labels to multi-hop paths in the sub-knowledge graphs; and construct an event knowledge graph based on the dynamic data.

[0076] In e-commerce retail platforms, sub-KGs are split according to business types, for example, into product KGs, logistics KGs, after-sales rules, user portraits, etc., to enable subsequent independent updates and independent queries by the federated query engine.

[0077] Specifically: when constructing each sub-KG, add logical labels to high-frequency multi-hop paths. For example, [Product]-[Applicable Scenario]->[Outdoor Sports]-[Climate Requirements]->

[0078] Alternatively, create a scenario bridge node for cross-domain association. For example, set up a virtual node "(Waterproof Grade ≥ IPX7) ∧ (Altitude Adaptability > 4000m)" for association logic between "Hiking Shoes" and "Tibet Travel."

[0079] 103. Establish entity correspondence across knowledge graphs.

[0080] 104. Set the event knowledge graph and the multiple static knowledge graphs as federation nodes respectively.

[0081] For example, by setting these sub-KGs as federated nodes respectively, it is possible to perform asynchronous execution of cross-domain static knowledge queries and dynamic data queries for static query requirements or dynamic query requirements, and then fuse the results of the two after the query.

[0082] For example, a user asked, "My Huawei phone, purchased in Chaoyang District, Beijing, has a screen problem. Where is the nearest authorized repair center?"

[0083] Federal Query Process:

[0084] S1. Location Entity: User → Order → Product → Brand → Service Point

[0085] S2. Sub-KG query decomposition:

[0086] KG1 (user data): Get purchase records

[0087] KG2 (product data): Verify mobile phone model

[0088] KG3 (Service Data): Query maintenance points in Chaoyang District, Beijing

[0089] S3. Time-space connection: Purchase time is within the warranty period AND current location is within the service radius

[0090] S4. Result fusion: Return the nearest repair point that supports the model

[0091] In the embodiments of this application, on the one hand, cross-domain KG combination is achieved through federated query, so that under the premise of data privacy and data isolation, intelligent customer service can access cross-domain knowledge graphs, improve response speed, answer scalability, and realize the precise question-and-answer capabilities of global intelligent customer service while ensuring the data autonomy of each business unit. On the other hand, considering that the combination of user historical behavior data and KG may leak privacy (such as inferring addresses through purchase records), the embodiments of this application, under the premise that user data does not leave the domain, jointly train the KG-based intelligent customer service through federated learning, desensitize sensitive fields in user behavior data (such as mobile phone numbers), and use differential privacy technology to perturb query results. This can reduce privacy risks.

[0092] In some implementations, to solve the long-tail problem (e.g., new products are put on the shelves, some KG data is missing, or KG is not updated in time), it is also possible to transfer the knowledge of the KG of similar products to the new product, and then generate diversified query samples based on the transferred KG and the parameters of the new product, and introduce a KG consistency penalty term in the loss function. Specifically, Figure 2b As shown, the embodiment of the present application also includes:

[0093] 201. Migrate the association relationships of similar products with similar attributes to the new product to the new product, and adjust the relationship weights based on the characteristics of the new product.

[0094] For example, the association relationship of similar mobile phones can be reused to migrate "Mobile Phone Y—Compatible Accessories→Charger A" to "Mobile Phone X".

[0095] 202. Generate multiple query samples based on the new product attributes and the migrated association relationship.

[0096] For example, based on product information: {KG fragment of product X}, five query samples that users may ask are generated. The query samples cover parameters, usage scenarios, compatibility, etc.

[0097] 203. Inject the association relationship of similar products into each inquiry sample.

[0098] Through injection, the large model of intelligent customer service can be guided to focus on the transferred knowledge.

[0099] For example, the question: Does mobile phone X support fast charging? After injecting the KG of similar mobile phones, we get:

[0100] [Knowledge Context]

[0101] Product: Mobile Phone X

[0102] Brand: Brand A

[0103] Compatible accessories: Charger B (Type: Fast Charge)

[0104] Waterproof rating: IP68

[0105] Question: Does the iPhone X support fast charging?

[0106] 204. Perform a federated query on the federated knowledge graph for each query sample to obtain multiple candidate answers, and mark conflicting answers among the multiple candidate answers that are inconsistent with the federated knowledge graph.

[0107] For example, model input: the current batch of questions {𝑄 𝑖}, its candidate answer set {𝐴 𝑖,𝑗}.

[0108] Conflict Detection:

[0109] For each problem 𝑄 𝑖 , retrieve relevant knowledge (such as product attributes, relationships) from KG, and mark all candidate answers 𝐴 that contradict KG 𝑖,𝑗 conflict For example, if KG states "Mobile phone X screen size = 6.7 inches", then "6.1 inches" and "7.0 inches" are both conflicting answers.

[0110] During the training process, for each query sample, the KG is queried in real time (rather than pre-generating a list of conflicting answers) to dynamically detect whether there are conflicting answers, ensuring the timeliness of knowledge updates.

[0111] For example, take the intelligent customer service’s response to a new e-commerce product:

[0112] Question: "What is the battery life of Smartwatch Y?"

[0113] KG Knowledge: The battery life is 7 days.

[0114] Conflicting answers: {"3 days", "5 days", "10 days"}.

[0115] It should be noted that if there is erroneous KG knowledge in the federated knowledge graph (such as "Mobile Phone X Waterproof Level = None"), then in order to cause the penalty term to reinforce the error, that is, to reduce the sensitivity of the intelligent customer service to erroneous KG, it is also possible to construct negative samples that conflict with the erroneous knowledge in the federated knowledge graph with a confidence level below the first threshold, and then conduct adversarial training on the intelligent customer service based on the negative samples, so that the intelligent customer service can identify the KG entries of true knowledge and erroneous knowledge. The model's sensitivity to erroneous KG can also be reduced by improving the above loss function. For example, for low-confidence conflicts, the value of the knowledge consistency penalty term can be weakened, and vice versa. For example, the improved loss function is:

[0116] Loss=CE+λ⋅∑(c j ⋅p j ), c j The confidence level of the conflicting answers corresponding to the KG knowledge

[0117] 205. Calculate the cumulative probability of all conflicting answers for each query sample.

[0118] For example, for each problem 𝑄 𝑖 , summing up the probabilities of all conflicting answers:

[0119]

[0120] Among them, 𝑝 𝑖,𝑗 For candidate answer 𝐴 𝑖,𝑗 The predicted probability of .

[0121] 206. Aggregate the total cumulative probability of each query sample according to the knowledge consistency penalty parameter value.

[0122] Sum up the penalty term 𝜆 for all problems:

[0123]

[0124] The purpose of summing the conflict probabilities is to suppress all conflicting answers that conflict with the facts, ensuring that domain knowledge (such as product parameters, policies and regulations) is strictly followed during model training, thereby significantly improving the accuracy and reliability of intelligent customer service.

[0125] 207. Calculate target loss according to the total cumulative probability.

[0126] The loss function is designed as follows:

[0127] Loss=Cross Entropy+λ⋅∑(Model output probability 冲突答案 )

[0128] 208. If the knowledge graph coverage is greater than the preset coverage and the average conflict probability of the conflicting answers is greater than the first threshold, the knowledge consistency penalty parameter value is increased for the conflicting answers to update the target loss until the proportion of the knowledge consistency penalty parameter value to the total loss is less than the preset proportion, and the operation is stopped.

[0129] The knowledge consistency penalty parameter represents the matching rate between the conflicting answers and the federated knowledge graph.

[0130] In this embodiment, only the probability exceeds the threshold (such as 𝑝 𝑖,𝑗 >0.1) to avoid noise interference. During backpropagation, the gradient is calculated only for the probability of the conflicting answer to prevent the penalty term λ from affecting the learning of other normal answers.

[0131] Based on the previous question: "How long is the battery life of smartwatch Y?", let's illustrate the training process:

[0132] S1. The model may initially assign a high probability (such as 0.4) to "5 days".

[0133] S2. By applying a summation penalty (0.4 + 0.2 + 0.1 = 0.7), the loss function forces the model to reduce the total probability of all conflicting answers.

[0134] S3. Final model output: “7 days” (probability 0.95), and the probability of conflicting answers is less than 0.02.

[0135] Result: Response accuracy improved, and users no longer received error messages such as "5 days".

[0136] For example, let's take cold-start product Q&A and long-tail questions caused by dynamic knowledge base updates as examples:

[0137] 1. Cold start product Q&A

[0138] Question: "What is the battery life of the new earphones Y?"

[0139] KG Knowledge: The battery life of a similar product "Headphones X" is 20 hours, and the technical documentation indicates that the difference between products in the same series is less than 10%.

[0140] Conflicting answer: "5 hours" (obviously contradicts products in the same series).

[0141] Effect: The model reduces the probability of predicting "5 hours", even though it is not directly labeled in the training data.

[0142] 2. Dynamic knowledge update

[0143] Scenario: The KG is updated with "Mobile Z supports IP68 waterproofing", but the training data is not yet covered.

[0144] Conflicting answers: "Not waterproof" is automatically detected as a conflicting answer.

[0145] Effect: The model adjusts its output instantly, without the need for re-fine-tuning.

[0146] It can be seen that in this embodiment, the knowledge consistency penalty item is dynamically adjusted according to the knowledge graph coverage and the average conflict probability to suppress the probability of conflicting answers, thereby improving the security and accuracy of answers to new products.

[0147] As described in the above embodiment, after the above intelligent customer service is obtained through model training, the user's static inquiry needs and dynamic inquiry needs can be responded to based on the intelligent customer service. Figure 3 A method for optimizing e-commerce intelligent customer service is provided. The method can be executed by a business server, which is an e-commerce retail platform. The method is applied to the e-commerce platform's customer service system, which includes a federated query engine and intelligent customer service. The intelligent customer service integrates an event knowledge graph and static knowledge graphs of multiple business types, where both the event knowledge graph and each static knowledge graph are federated nodes. Embodiments of this application include:

[0148] 301. Receive a first input from a user on a conversation interface.

[0149] 302. In response to the user's first input, the federated query engine pre-calculates the product query intention corresponding to the first input and determines that it conforms to the multi-hop query mode, then performs a cross-knowledge graph joint query on the product query intention to obtain a first type of result based on each static knowledge graph and a second type of result based on the event knowledge graph.

[0150] For example, a user asks: "Why can't phone X use 5G in a certain country?"

[0151] The federated query engine performs the following S1-S3 operations:

[0152] S1: parse the key entities in the question ("mobile phone X", "a country", "5G").

[0153] S2: Obtain relevant data fragments through multi-hop queries across sub-KGs (e.g., product KG → policy KG → operator KG). Specifically:

[0154] For example, based on the user question above: "Why can't phone X use 5G in a certain country?"

[0155] First hop (product KG): Get the frequency band list supported by mobile phone X.

[0156] Second hop (Policy KG): Query the 5G frequency band standard of the target country.

[0157] Third hop (operator KG): Verify local base station coverage.

[0158] For example, a user asked, "My Huawei phone, purchased in Chaoyang District, Beijing, has a screen problem. Where is the nearest authorized repair center?"

[0159] First hop KG1 (user data): obtain purchase records

[0160] Second hop KG2 (product data): Verify mobile phone model

[0161] The third jump KG3 (service data): query the maintenance point in Chaoyang District, Beijing

[0162] Then, we connect "the purchase time of mobile phone X is within the warranty period" with "the current location is within the service radius" in time and space to obtain the integrated result: the nearest repair point that supports this model is returned.

[0163] S3: Return structured knowledge fragments (i.e., KG data) to the intelligent customer service.

[0164] 303. The federated query engine fuses the first and second category results based on multi-party secure computing, and obtains the fused results and transmits them to the intelligent customer service.

[0165] It can be understood that the fusion result includes knowledge graph coverage and knowledge confidence, which provides a basis for intelligent customer service to adjust temperature parameter values.

[0166] For example Figure 4 In the inference example diagram shown, the fusion result is in JSON format:

[0167] {

[0168] "Mobile phone X band": ["A", "B", "C"],

[0169] "A country's 5G standard": ["D", "E"],

[0170] "Carrier Coverage": {"Carrier F": ["D"], "Carrier G": ["E"]}

[0171] }

[0172] Among them, A, B…E are the covered knowledge graphs.

[0173] In some implementations, the federated query engine can also return multi-hop neighbor features. For example, a hierarchical graph neural network can be used to identify potential paths between multi-hop entities, aggregate multi-hop neighbor features, and mine implicit relationship weights. Dynamic weights are assigned to each hop neighbor feature based on the query intent of the first input and an attention mechanism. Multi-hop neighbor features refer to the path nodes, relationships, and their attributes obtained through a multi-hop traversal in the KG (e.g., entity A → relationship R1 → entity B → relationship R2 → entity C).

[0174] In an embodiment of the present application, the multi-hop depth can be controlled. For example, the maximum number of hops of the current multi-hop query (for example, 3) can be dynamically specified based on the user's query intent and preset query parameters. Then, when the federated query engine returns the fusion result to the large model generation layer of the intelligent customer service, it will automatically stage the query path that exceeds the hop number, thereby achieving dynamic and flexible adaptation to the user's query intent and adapting to the needs of different scenarios.

[0175] Using federated query results as input for prompt engineering ensures that generated content is based on trusted knowledge, forming a closed data flow loop.

[0176] The cross-domain knowledge acquisition of the federated query engine is decoupled from the prompt engineering of intelligent customer service, and cross-domain knowledge of multi-hop reasoning is provided, each performing its own function.

[0177] 304. The intelligent customer service generates a multi-hop reasoning chain based on the fusion result.

[0178] Intelligent customer service can embed the fusion results returned from the federated query engine, i.e., KG data, into the inference instruction template. Figure 4 In the inference example diagram shown, the inference instructions obtained are as follows:

[0179] Please reason step by step based on the following knowledge:

[0180] - Mobile X supports frequency bands: {A / B / C}

[0181] - 5G standard frequency band in a certain country: {D / E}

[0182] - Frequency bands covered by local operators: {Operator F: D, Operator G: E}

[0183] User question: Why can't my phone X use 5G in a certain country?

[0184] After embedding the fusion results into the above inference instructions, the large model generation layer can generate the above multi-hop inference chain, for example:

[0185] {

[0186] "Inference Chain": [

[0187] {"Step 1": "Get the frequency band of mobile phone X from product KG"},

[0188] {"Step 2": "Query the frequency bands supported by the target country from the policy KG"},

[0189] {"Step 3": "Compare the differences"}

[0190] ],

[0191] "Conclusion": "Mobile X does not support 5G in this country"

[0192] In other implementations, if a multi-hop path is returned, the reasoning hint template may require the model to reference the path nodes step by step.

[0193] For example, based on the following path reasoning:

[0194] 1. [Mobile X] Supported frequency bands → {A, B}

[0195] 2. [Country] Requests frequency band → {C, D}

[0196] → Conclusion: Incompatible.

[0197] In other implementations, the federated query engine can return not only the explicit path described above, but also aggregated implicit neighbor features. If multi-hop neighbor features are returned, inference prompt templates within the large model generation layer of intelligent customer service are needed to guide the large model in interpreting the statistical results. For example, based on the frequency band matching degree (20%), phone X does not meet the country's 5G standards.

[0198] In the embodiments of the present application, whether the federated query engine returns multi-hop neighbor features depends on the business scenario's demand for reasoning depth and the trade-off between system performance. If deep reasoning is required, the original multi-hop path is returned, preserving the complete logical chain. If a fast response is required, aggregate features or terminal entities are returned, sacrificing some interpretability for improved efficiency. This embodiment of the present application is not limited to this.

[0199] 305. The intelligent customer service generates an initial answer based on the multiple reasoning chains and the fusion results.

[0200] For example, Figure 4 In the inference example diagram, the user asks: "Why can't mobile phone X use 5G in a certain country?"

[0201] The reasoning process of generating an initial answer by combining the fusion results of the multiple reasoning chains is as follows:

[0202] 1. Mobile X supports frequency bands: A, B, C (Source: Product Database)

[0203] 2. 5G frequency bands required by target countries: D, E (Source: Communications Policy Library)

[0204] 3. Local operator F has not deployed D / E band base stations (Source: Operator Data)

[0205] → Conclusion: Phone X cannot use 5G in this country.

[0206] 306. Intelligent customer service adjusts the temperature parameters of each hop according to the knowledge graph coverage and confidence of the fusion result and the number of reasoning hops, enhances the initial answer, obtains the target answer and outputs it.

[0207] In some implementations, the target answer can be obtained specifically by the following methods:

[0208] If the knowledge graph coverage of the first-hop reasoning is lower than the preset coverage and the confidence is lower than the preset confidence, the temperature parameter of the first-hop is increased to supplement the content not covered by the static knowledge graph.

[0209] Alternatively, if the knowledge graph coverage of the first-hop reasoning is higher than the preset coverage and the confidence is higher than the preset confidence, the temperature parameter of the second hop is lowered, and the second hop is the next hop of the first hop.

[0210] It can be seen that by dynamically adjusting the temperature, the accuracy and diversity of e-commerce intelligent customer service's answers to product inquiries can be balanced.

[0211] For example, the dynamic temperature tiered control mechanism based on the number of inference hops is shown in Table 1:

[0212] Inference hop count Temperature (T) Purpose Jump 1 T=0.3 Ensure single-hop facts are accurate Jump 2-3 T=0.5 Balancing Accuracy and Relevance Divergence >3 hops T=0.7 Allows creative completion of long chains of reasoning

[0213] Table 1

[0214] Under normal circumstances, the temperature increases with the number of hops (T: 0.3 → 0.5 → 0.7). After outputting the structured reasoning chain, the conclusion is generated: "Mobile phone X's frequency band (A / B / C) does not meet the country's 5G standard (D / E), and operator F has not deployed a compatible base station."

[0215] In some implementations, if the confidence level of the first-category result returned by any knowledge graph in a multi-hop path is lower than a preset confidence level, the temperature parameter for subsequent hops is reduced (e.g., T = 0.2). This indicates that conflict detection based on the confidence level of the first-category result reduces the weight of neighbor features in subsequent hops, thereby limiting subsequent error propagation.

[0216] In an embodiment of the present application, by setting the event knowledge graph and multiple static knowledge graphs of different business types as federated nodes respectively, when a user inquires, when it is determined that the user's product query intention meets multiple query patterns, a cross-domain KG combination can be achieved through a federated query. On the one hand, each knowledge graph is deployed independently, especially the event knowledge graph is isolated from the static knowledge graph, the cross-domain knowledge acquisition of the federated query engine is decoupled from the prompt engineering of the intelligent customer service, and cross-domain knowledge of multi-hop reasoning is provided, so that each knowledge graph performs its own function and can reduce the privacy risk during public queries; on the other hand, the response speed and answer scalability can be improved through federated query. Under the premise of ensuring the autonomous query of each knowledge graph, the accurate question-answering capability of the global intelligent customer service is realized, and complex questions are covered by multi-hop query, which can break through the information limitations of a single knowledge graph. On the other hand, by hierarchical control of temperature parameters according to the number of inference hops and structuring the inference chain of each hop path, it can be ensured that the inference chain meets both flexibility and reliability, that is, the controllable generation of intelligent customer service answers is achieved.

[0217] It can be seen that by implementing multi-hop reasoning across federation domains, dynamically generating multiple reasoning chains, and evaluating the results of the covered knowledge graph in each hop to dynamically adjust the temperature parameters, through the collaboration of these three, in a balance between scalability (federation multi-hop) and accuracy (temperature control), it is possible to achieve the upgrade of intelligent customer service from "single point response" to "chain reasoning".

[0218] Optionally, in some other embodiments of the present application, the static knowledge graph includes user data nodes and service data nodes, and the event knowledge graph is a product data node; the cross-knowledge graph joint query of the product query intent is performed to obtain the first type of results based on each static knowledge graph, and the second type of results based on the event knowledge graph, including:

[0219] Determining multiple entities in the product query intent;

[0220] Linking the first entity to the user data node to query product purchase data, linking the second entity to the product data node to verify product parameters, and linking the third entity to the service data node to query after-sales outlets;

[0221] Connect product purchase data with after-sales outlets within a preset range of the user's current location in time and space;

[0222] Determine target outlets that meet product after-sales conditions based on time and space connections.

[0223] It can be seen that in the after-sales consultation scenario of a product, through the federated query cross-domain knowledge graph, it is possible to quickly and accurately calculate whether the product meets the repair conditions and locate the target outlets that meet the repair conditions for the product.

[0224] Optionally, in other embodiments of the present application, each e-commerce platform joins a federated node, taking the federated query engine of the target e-commerce platform as an example. Each e-commerce platform aggregates and fine-tunes parameters through federated learning to avoid sharing their own original domain knowledge graphs. The method also includes:

[0225] Each e-commerce platform generates question-answer pairs based on the local domain knowledge graph;

[0226] The federated query engine obtains question-answer pairs from various e-commerce platforms;

[0227] The question-answer pair is passed to the generation layer of the intelligent customer service. The generation layer combines the question-answer pair to generate candidate answers. The candidate answers include domain knowledge graph coverage and knowledge confidence.

[0228] The temperature parameter regulator of intelligent customer service automatically adjusts the temperature parameters according to the coverage and confidence of the candidate answers to adjust the candidate answers, and finally obtains and outputs diverse target answers.

[0229] For example, take e-commerce return and exchange consultation as an example:

[0230] User question: "The phone I bought on Platform A is overheating. Does the same model on Platform B also have this problem? Can I exchange it on Platform C?"

[0231] Processing Flow

[0232] S1: Federated cross-domain query:

[0233] Obtain the fault record of the mobile phone from platform A sub-KG.

[0234] Query the evaluation keywords of the same mobile phone from platform B's sub-KG (such as the frequency of "fever").

[0235] Read the exchange policy from the platform C sub-KG.

[0236] S2: Temperature parameter calculation:

[0237] The fault report of platform A is clear → Confidence ↑ → T=0.4 (deterministic mode).

[0238] The occurrence rate of the keyword "fever" on platform B is 30% → medium risk → maintain T=0.6.

[0239] S3: Generate answer:

[0240] "According to Platform A's test report, the phone you purchased may have a heat dissipation issue (95% confidence level). Users of the same product on Platform B report a 30% probability of overheating. We recommend contacting Platform A for a return or exchange (policy link). If cross-platform processing is required, Platform C currently supports... [leave room for moderate suggestions]."

[0241] As can be seen, federated queries aggregate multi-source KG data, avoiding information bias within a single platform and making them more suitable for cross-platform after-sales and cross-supply chain consulting scenarios. A single platform also eliminates the need for a large, global domain knowledge graph, reducing local storage load. A single platform can analyze missing or incorrect domain knowledge graphs locally without syncing with other platforms participating in federated cross-domain queries. Knowledge can be integrated into the federated query engine through differential privacy protection.

[0242] Figures 1 to 4 Any technical feature mentioned in any embodiment corresponding to any one of the items is also applicable to the embodiments of this application. Figures 5 to 7 The corresponding embodiments will not be described in detail later.

[0243] The above describes an e-commerce intelligent customer service optimization method in an embodiment of the present application. The following introduces an e-commerce intelligent customer service optimization device that executes the above e-commerce intelligent customer service optimization method.

[0244] See Figure 5 ,like Figure 5 The schematic diagram of the structure of an e-commerce intelligent customer service optimization device 40 is shown, which can be applied to the e-commerce customer service system of the e-commerce platform, and balances the accuracy and diversity of answers through federated query cross-domain knowledge graph. The e-commerce intelligent customer service optimization device 40 in the embodiment of the present application can achieve the corresponding Figure 1-Figure 4 The steps in the e-commerce intelligent customer service optimization method executed by the e-commerce intelligent customer service optimization device 40 in any corresponding embodiment. The functions implemented by the e-commerce intelligent customer service optimization device 40 can be implemented by hardware, or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above functions, and the modules can be software and / or hardware. The e-commerce intelligent customer service optimization device 40 may include an input and output module 401, a federated query module 402, and an inference module 403. The functional implementation of the input and output module 401, the federated query module 402, and the inference module 403 can refer to Figure 1-Figure 4 The operations performed in any corresponding embodiment are not described in detail here.

[0245] In some implementations, the input / output module 401 may be configured to receive a first input from a user on a conversation interface;

[0246] The processing module 402 may be configured to respond to a first input received by the input / output module, and if the federated query engine pre-calculates the product query intent corresponding to the first input and determines that the query intent meets a multi-hop query mode, perform a cross-knowledge graph joint query on the product query intent to obtain a first type of result based on each static knowledge graph and a second type of result based on the event knowledge graph;

[0247] The first type of results and the second type of results are fused based on multi-party secure computing, and the fused results are input into the reasoning module 403.

[0248] The reasoning module 403 can be used to generate a multi-hop reasoning chain based on the fusion results obtained by the federated query module, and generate an initial answer according to the multiple reasoning chains combined with the fusion results. According to the knowledge graph coverage and confidence of the fusion results, the temperature parameter of each hop is adjusted according to the number of reasoning hops, the initial answer is enhanced, and the target answer is obtained and output through the input and output module.

[0249] In some implementations, the reasoning module 403 is specifically configured to:

[0250] If the knowledge graph coverage of the first-hop reasoning is lower than the preset coverage and the confidence is lower than the preset confidence, then the temperature parameter of the first-hop is increased to supplement the content not covered by the static knowledge graph;

[0251] Alternatively, if the knowledge graph coverage of the first-hop reasoning is higher than the preset coverage and the confidence is higher than the preset confidence, the temperature parameter of the second hop is lowered, and the second hop is the next hop of the first hop.

[0252] In some implementations, the reasoning module 403 is further configured to:

[0253] If the confidence of the first type of results returned by any knowledge graph in the multi-hop path is lower than the preset confidence, the temperature parameters of the subsequent hops are reduced.

[0254] In some implementations, the reasoning module 403 is further configured to:

[0255] Based on a dynamic update strategy, detecting update events of the event knowledge graph, and synchronizing the update events to the event knowledge graph through a streaming graph;

[0256] Alternatively, the static knowledge graph or the event knowledge graph is automatically corrected based on the real-time feedback of the user on the target answer.

[0257] In some implementations, the reasoning module 403 is further configured to:

[0258] Migrate the association relationships of similar products with similar attributes to the new product to the new product, and adjust the relationship weights based on the characteristics of the new product;

[0259] Generate multiple query samples based on the new product attributes and the migrated association relationship;

[0260] Inject the association relationship of similar products into each query sample;

[0261] Perform a federated query on the federated knowledge graph for each query sample to obtain multiple candidate answers, and mark conflicting answers among the multiple candidate answers that are inconsistent with the federated knowledge graph;

[0262] Calculate the cumulative probability of all conflicting answers for each query sample;

[0263] Aggregate the total cumulative probability of each query sample according to the knowledge consistency penalty parameter value;

[0264] calculating a target loss based on the total cumulative probability;

[0265] If the knowledge graph coverage is greater than the preset coverage and the average conflict probability of the conflicting answers is greater than the first threshold, the knowledge consistency penalty parameter value is added to the conflicting answer to update the target loss until the proportion of the knowledge consistency penalty parameter value to the total loss is less than the preset proportion, and the operation is stopped, wherein the knowledge consistency penalty parameter represents the matching rate between the conflicting answer and the said federated knowledge graph.

[0266] In some implementations, the federated query module 402 is further configured to: obtain initial e-commerce domain knowledge graphs from multiple channels, as well as dynamic data of e-commerce platforms;

[0267] Splitting the initial e-commerce domain knowledge graph into multiple static knowledge graphs according to business types, adding logical labels to multi-hop paths in the sub-knowledge graphs; and constructing an event knowledge graph based on the dynamic data;

[0268] Establish entity correspondence across knowledge graphs;

[0269] The event knowledge graph and the multiple static knowledge graphs are respectively set as federation nodes.

[0270] In some implementations, the reasoning module 403 is specifically configured to:

[0271] Determining multiple entities in the product query intent;

[0272] Linking the first entity to the user data node to query product purchase data, linking the second entity to the product data node to verify product parameters, and linking the third entity to the service data node to query after-sales outlets;

[0273] Connect product purchase data with after-sales outlets within a preset range of the user's current location in time and space;

[0274] Determine target outlets that meet product after-sales conditions based on time and space connections.

[0275] This solution, on the one hand, uses a federated query engine to conduct federated queries across cross-domain knowledge graphs (multiple static knowledge graphs + event knowledge graphs). The intelligent customer service system uses the fusion results of the federated query engine as input to the prompting process, ensuring that the target answer is based on trustworthy knowledge. Furthermore, during the generation of the target answer, the initial answer is dynamically adjusted by dynamically adjusting the temperature parameter value based on factors such as knowledge graph coverage. This ensures that the final target answer balances the scalability and accuracy of the intelligent customer service system, thereby improving the intelligent customer service system's question-answering capabilities.

[0276] The above describes the e-commerce intelligent customer service optimization device 40 and the e-commerce intelligent customer service optimization device 50 for executing the e-commerce intelligent customer service optimization method in the embodiment of the present application from the perspective of modular functional entities. The following describes the e-commerce intelligent customer service optimization device 40 and the e-commerce intelligent customer service optimization device 50 for executing the e-commerce intelligent customer service optimization method in the embodiment of the present application from the perspective of hardware processing. It should be noted that in the embodiment of the present application Figure 5 The physical device corresponding to the input / output module 401 in the illustrated embodiment may be an input / output unit, a transceiver, a radio frequency circuit, a communication module, an output interface, etc., and the physical device corresponding to the processing module 402 may be a processor. Figure 5 The e-commerce intelligent customer service optimization device 40 shown can have the following features: Figure 6 The structure shown, when Figure 5 The e-commerce intelligent customer service optimization device 40 shown has the following features: Figure 6 When the structure shown is Figure 6 The processor and transceiver in the embodiment can realize the same or similar functions as the input and output module 401 and the processing module 402 provided in the aforementioned device embodiment corresponding to the e-commerce intelligent customer service optimization device 40. Figure 6 The memory in the storage processor is used to execute the computer program that needs to be called when the e-commerce intelligent customer service optimization method is executed.

[0277] The present application also provides another e-commerce intelligent customer service optimization device. For ease of explanation, only the parts related to the present application are shown. For specific technical details not disclosed, please refer to the method section of the present application. The e-commerce intelligent customer service optimization device can be any e-commerce intelligent customer service optimization device, including a mobile phone, tablet computer, personal digital assistant (PDA), point of sales (POS), in-vehicle computer, etc.

[0278] The embodiment of the present application also provides another e-commerce intelligent customer service optimization device for implementing the above-mentioned e-commerce intelligent customer service optimization method, such as Figure 7 As shown, Figure 7 This is a schematic diagram of a server structure provided by an embodiment of the present application. The server 720 may vary significantly due to different configurations or performance, and may include one or more central processing units (CPUs) 722 (e.g., one or more processors), memory 732, and one or more storage media 730 (e.g., one or more mass storage devices) for storing application programs 742 or data 744. The memory 732 and storage media 730 may be either transient or persistent storage. The program stored in the storage medium 730 may include one or more modules (not shown), each of which may include a series of instruction operations on the server. Furthermore, the CPU 722 may be configured to communicate with the storage medium 730 to execute the series of instruction operations in the storage medium 730 on the server 720.

[0279] The server 720 may also include one or more power supplies 726, one or more wired or wireless network interfaces 750, one or more input and output interfaces 757, and / or one or more operating systems 741, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc.

[0280] In the above embodiment, the service server (for example Figure 5 The steps performed by the e-commerce intelligent customer service optimization device 40 shown in the figure can be based on the Figure 7 The structure of the server 720 shown in FIG. Figure 5 The steps performed by the e-commerce intelligent customer service optimization device 40 shown can be based on the Figure 7 For example, the processor 722 calls the instructions in the memory 732 to perform the following operations:

[0281] receiving a first input from the user on the conversation interface via the input / output interface 757;

[0282] In response to the first input, if the federated query engine pre-calculates the product query intent corresponding to the first input and determines that it meets the multi-hop query mode, a cross-knowledge graph joint query is performed on the product query intent to obtain a first type of result based on each static knowledge graph and a second type of result based on the event knowledge graph;

[0283] Based on multi-party secure computation, the first type of results and the second type of results are integrated to obtain a fusion result;

[0284] A multi-hop reasoning chain is generated based on the fusion result, and an initial answer is generated according to the multiple reasoning chains combined with the fusion result. According to the knowledge graph coverage and confidence of the fusion result, the temperature parameter of each hop is adjusted according to the number of reasoning hops, the initial answer is enhanced, and the target answer is obtained and output through the input and output interface 757.

[0285] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0286] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0287] In the several embodiments provided in the embodiments of the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0288] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of the present embodiment according to actual needs.

[0289] In addition, the functional modules in each embodiment of the present application can be integrated into a processing module, or each module can exist physically separately, or two or more modules can be integrated into a module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0290] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0291] The computer program product includes one or more computer instructions. When the computer program is loaded and executed on a computer, the process or function described in accordance with the embodiments of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium may be any available medium that can be stored on a computer or a data storage device such as a server or data center that includes one or more available media. The available medium may be a magnetic medium (e.g., a floppy disk, hard disk, or magnetic tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0292] The above is a detailed introduction to the technical solutions provided in the embodiments of the present application. Specific examples are used in the embodiments of the present application to illustrate the principles and implementation methods of the embodiments of the present application. The description of the above embodiments is only used to help understand the methods and core ideas of the embodiments of the present application. At the same time, for those skilled in the art, according to the ideas of the embodiments of the present application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the embodiments of the present application.

Claims

1. An e-commerce intelligent customer service optimization method, characterized in that: The method is applied to an e-commerce customer service system of an e-commerce platform, which includes a federated query engine and intelligent customer service; the intelligent customer service integrates an event knowledge graph and static knowledge graphs of multiple business types, where the event knowledge graph and each static knowledge graph are federated nodes; the method includes: Receiving a first input from the user on the conversation interface; In response to the first input, if the federated query engine pre-calculates the product query intent corresponding to the first input and determines that it meets the multi-hop query mode, a cross-knowledge graph joint query is performed on the product query intent to obtain a first type of result based on each static knowledge graph and a second type of result based on the event knowledge graph; The first and second types of results are integrated through multi-party secure computation, and the integrated results are transmitted to the intelligent customer service; Based on the fusion results, the intelligent customer service generates a multi-hop reasoning chain, and generates an initial answer according to the multiple reasoning chains combined with the fusion results. According to the knowledge graph coverage and confidence of the fusion results, the temperature parameter of each hop is adjusted according to the number of reasoning hops, the initial answer is enhanced, and the target answer is obtained and output.

2. The method according to claim 1, characterized in that According to the knowledge graph coverage and confidence of the fusion result, the temperature parameter of each hop is adjusted according to the number of reasoning hops, and the initial answer is enhanced to obtain the target answer, including: If the knowledge graph coverage of the first-hop reasoning is lower than the preset coverage and the confidence is lower than the preset confidence, then the temperature parameter of the first-hop is increased to supplement the content not covered by the static knowledge graph; Alternatively, if the knowledge graph coverage of the first-hop reasoning is higher than the preset coverage and the confidence is higher than the preset confidence, the temperature parameter of the second hop is lowered, and the second hop is the next hop of the first hop.

3. The method according to claim 2, characterized in that The method further comprises: If the confidence of the first type of results returned by any knowledge graph in the multi-hop path is lower than the preset confidence, the temperature parameters of the subsequent hops are reduced.

4. The method according to any one of claims 1 to 3, characterized in that The method further comprises: Based on a dynamic update strategy, detecting update events of the event knowledge graph, and synchronizing the update events to the event knowledge graph through a streaming graph; Alternatively, the static knowledge graph or the event knowledge graph is automatically corrected based on the real-time feedback of the user on the target answer.

5. The method according to claim 4, characterized in that The method further comprises: Migrate the association relationships of similar products with similar attributes to the new product to the new product, and adjust the relationship weights based on the characteristics of the new product; Generate multiple query samples based on the new product attributes and the migrated association relationship; Inject the association relationship of similar products into each query sample; Perform a federated query on the federated knowledge graph for each query sample to obtain multiple candidate answers, and mark conflicting answers among the multiple candidate answers that are inconsistent with the federated knowledge graph; Calculate the cumulative probability of all conflicting answers for each query sample; Aggregate the total cumulative probability of each query sample according to the knowledge consistency penalty parameter value; calculating a target loss based on the total cumulative probability; If the knowledge graph coverage is greater than the preset coverage and the average conflict probability of the conflicting answers is greater than the first threshold, the knowledge consistency penalty parameter value is added to the conflicting answer to update the target loss until the proportion of the knowledge consistency penalty parameter value to the total loss is less than the preset proportion, and the operation is stopped, wherein the knowledge consistency penalty parameter represents the matching rate between the conflicting answer and the said federated knowledge graph.

6. The method according to claim 5, characterized in that The method further comprises: Obtain initial e-commerce domain knowledge graphs from multiple channels, as well as dynamic data from e-commerce platforms; Splitting the initial e-commerce domain knowledge graph into multiple static knowledge graphs according to business types, adding logical labels to multi-hop paths in the sub-knowledge graphs; and constructing an event knowledge graph based on the dynamic data; Establish entity correspondence across knowledge graphs; The event knowledge graph and the multiple static knowledge graphs are respectively set as federation nodes.

7. The method according to any one of claims 1 to 3, characterized in that The static knowledge graph includes user data nodes and service data nodes, and the event knowledge graph is a product data node; the cross-knowledge graph joint query of the product query intention is performed to obtain the first type of results based on each static knowledge graph and the second type of results based on the event knowledge graph, including: Determining multiple entities in the product query intent; Linking the first entity to the user data node to query product purchase data, linking the second entity to the product data node to verify product parameters, and linking the third entity to the service data node to query after-sales outlets; Connect product purchase data with after-sales outlets within a preset range of the user's current location in time and space; Determine target outlets that meet product after-sales conditions based on time and space connections.

8. An e-commerce intelligent customer service optimization device, characterized in that: The e-commerce intelligent customer service optimization device includes: at least one processor and memory; The memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that The method comprises instructions which, when executed on a computer, cause the computer to perform the method according to any one of claims 1 to 7.

10. A computer program product comprising instructions, the computer program product comprising program instructions, which, when executed on a computer or a processor, cause the computer or the processor to perform the method according to any one of claims 1 to 7.

Images