Intelligent customer service conversation method, apparatus, device, storage medium, and program product

Abstract

Embodiments of the present disclosure provide an intelligent customer service conversation method and device, equipment, storage medium and program product, which are related to the field of artificial intelligence, and the application scenarios include but are not limited to the intelligent customer service conversation scene. The method comprises: based on a target question, performing standard question prediction processing and standard intent prediction processing through a trained first large language model, obtaining a target standard question corresponding to the target question and a target standard intent corresponding to the target question based on the prediction result, and performing emotion level prediction processing through a trained second large language model to obtain an emotion type and an emotion level; obtaining a target reply strategy in a reply strategy set corresponding to the target standard question; determining a target reply rule matched with the target object information from the target reply strategy based on the target object information, the target standard intent, the emotion type and the emotion level; and replying to the target standard question based on the target reply rule to obtain a target standard reply of the target standard question.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and more specifically, to an intelligent customer service conversation method, apparatus, device, storage medium, and program product. Background Technology

[0002] In existing technologies, e-commerce intelligent customer service responds to user questions during intelligent customer service conversations. Existing technologies can retrieve and filter existing knowledge bases, inputting large amounts of pre-trained corpus data into a retrieval model for training, resulting in a trained retrieval model. The intelligent customer service then uses this trained model to classify, match, and answer new input questions. However, search matching may encounter situations where some words are not found or are incorrectly searched, leading to errors in the intelligent customer service's responses to user questions, resulting in low efficiency in answering user questions. Summary of the Invention

[0003] This disclosure addresses the shortcomings of existing methods by proposing an intelligent customer service conversation method, apparatus, device, computer-readable storage medium, and computer program product to solve the problem of how to improve the efficiency of intelligent customer service in e-commerce in responding to user questions.

[0004] Firstly, this disclosure provides an intelligent customer service conversation method, including: Get the target session of the session object; If the target conversation is a target question for the target object, then based on the target question, the first trained language model is used to perform standard question prediction and standard intent prediction. Based on the prediction results, the target standard question and the target standard intent corresponding to the target question are obtained. Based on the target question, the second trained language model is used to perform emotion level prediction to obtain the emotion type and the emotion level corresponding to the emotion type. The emotion level corresponding to the target question is used to represent the emotional degree of the conversation object, and the target standard intent corresponding to the target question is used to represent the intent of the conversation object. Obtain the target response strategy from the set of response strategies corresponding to the target standard question. The set of response strategies includes the response strategy corresponding to each standard question among multiple standard questions. The response strategy corresponding to each standard question includes at least one response rule. Each response rule is associated with at least one object information, a standard intention, an emotion type, and an emotion level corresponding to an emotion type. Obtain target object information, and based on the target object information, target standard intent, emotion type and emotion level, determine target response rules that match the target object information from the target response strategy; Based on the target response rules, the target standard questions are answered to obtain the target standard responses to the target standard questions, and the target standard responses are determined as the responses to the target questions.

[0005] In one embodiment, based on the target question, a pre-trained first language model is used to perform standard question prediction and standard intent prediction. Based on the prediction results, the target standard question and the target standard intent corresponding to the target question are obtained, including: Obtain prompt words. Prompt words are used to instruct the trained first language model to determine the candidate standard questions corresponding to the input question from the question set and to select the candidate standard intents corresponding to the input question from the intent set. The prompt words include the question set and the intent set. The question set includes multiple pre-set candidate standard questions, and the intent set includes multiple candidate standard intents. The prompt words and target question are input into the trained first language model. The trained first language model is used to perform standard intent prediction processing to obtain the candidate standard intent corresponding to the target question. The trained first language model is used to perform standard question prediction processing to obtain the candidate standard question corresponding to the target question. Based on the predicted candidate standard questions, the target standard question corresponding to the target question is obtained, and based on the predicted candidate standard intent, the target standard intent corresponding to the target question is obtained.

[0006] In one embodiment, obtaining the target standard problem corresponding to the target problem based on the predicted candidate standard problems includes: The predicted candidate standard problems are matched with each preset standard problem in the standard problem set of the database, and the preset standard problems in the standard problem set that match the predicted candidate standard problems are taken as the target standard problems.

[0007] In one embodiment, based on the target question, a pre-trained second-largest language model is used to predict the emotion level, resulting in the emotion type corresponding to the target question and the emotion level corresponding to the emotion type, including: The target question is input into the trained second language model, and word segmentation is performed to obtain the corresponding word segmentation for the target question. Based on the word segmentation corresponding to the target question, the word segmentation encoding representation corresponding to the target question is obtained through encoding processing; Based on the word segmentation encoding representation corresponding to the target question, the emotion type corresponding to the target question and the emotion level corresponding to the emotion type are obtained through classification processing.

[0008] In one embodiment, based on the target standard problem, the target rule judgment method corresponding to the target standard problem is determined from the set of rule judgment methods in the database; Based on target audience information, target standard intent, emotion type, and emotion level, target response rules that match the target audience information are determined from the target response strategies, including: Based on the target rule judgment method, target object information, target standard intent, emotion type and emotion level, target response rules that match the target object information are determined from the target response strategies.

[0009] In one embodiment, based on the target response rules, a target standard question is answered to obtain a target standard response to the target standard question, including: Based on the target response rules and the set of standard response templates in the database, the target standard response templates that match the target standard questions are determined from the set of standard response templates. Based on the target object's information and the target standard response template, the target standard response to the target standard question is obtained. The content dimension set, emotion type, and emotion level of the target standard response template are related. The content dimension set includes at least one of the following: regular fields, tone, and specific fields. The specific fields include relevant operations for the target object.

[0010] In one embodiment, the target object is an order or a product, and after determining the target standard response as a response to the target question, the method further includes: The target standard response is sent to the terminal so that the target standard response to the target question for the order or product is displayed in the terminal's intelligent customer service conversation interface.

[0011] In one embodiment, based on the target question, a dynamically adjusted instruction prediction process is performed using a trained first language model to obtain a dynamically adjusted instruction. The dynamically adjusted instruction is used to adjust the target standard response on the preset platform. Based on the dynamic adjustment instructions, emotion type, and the corresponding emotion level, the target decision function is matched from the set of decision functions of the preset platform. Based on the dynamic adjustment instructions, emotion type, and the corresponding emotion level, a decision result is generated through the target decision function and sent to the preset platform. The preset platform then adjusts the target standard response based on the decision result to obtain a new response, which is then sent to the conversation object.

[0012] In one embodiment, if there is an intercepted word in the intercepted word set in the database in the target session, then based on the intercepted word and the standard response template set in the database, a first standard response template matching the intercepted word is determined from the standard response template set, and the first standard response template is determined as the response to the intercepted word. The response to the intercepted word includes switching from the intelligent customer service session to the human customer service session. If there are no intercepted words in the target conversation and the target conversation consists of small talk, then based on the small talk, standard question prediction is performed using the trained first language model to determine the candidate standard question corresponding to the small talk. Based on the candidate standard question corresponding to the small talk, the set of small talk in the database, and the set of standard response templates, a second standard response template that matches the standard question corresponding to the small talk is determined from the set of standard response templates, and the second standard response template is determined as the response to the small talk.

[0013] In one embodiment, the first trained language model is obtained by training it in the following way: Retrieve a set of question samples and prompts. The question sample set includes object-related questions and greetings. The prompts include a set of questions, a set of greetings, a set of intentions, and a set of dynamic adjustment instructions; For each training round in the multi-round training, each question sample and prompt word are input into the first large language model to be trained. The first large language model to be trained performs standard intent prediction processing to obtain the candidate standard intent corresponding to each question sample. The first large language model to be trained performs standard question prediction processing to determine the candidate standard question corresponding to each question sample. The first large language model to be trained performs dynamic adjustment instruction prediction processing to obtain the dynamic adjustment instruction. Based on candidate standard questions, preset standard questions, candidate standard intentions, preset standard intentions, dynamic adjustment instructions, and preset standard dynamic adjustment instructions, the training loss of the first large language model to be trained in each round of training is determined. If the training termination condition is met, the training of the first language model to be trained is terminated, and the trained first language model is obtained. If the training termination condition is not met, the model parameters of the first language model are adjusted based on the training loss, and the adjusted first language model is trained in the next round.

[0014] In one embodiment, the content dimension set of the target standard response includes at least one of general fields, tone, and specific fields, where the specific fields include relevant operations for the target object. If the content dimension set of the target standard response includes specific fields, after determining that the target standard response is a response to the target question, it further includes: Retrieve the target response from the session object for the relevant operation; Based on the target response, the operation command is predicted and processed by the trained third language model to obtain the operation command. The operation instructions are sent to the preset platform so that the preset platform can perform corresponding operations on the target object based on the operation instructions.

[0015] Secondly, this disclosure provides an intelligent customer service conversation device, comprising: The first processing module is used to obtain the target session of the session object; The second processing module is used to perform standard question prediction and standard intent prediction based on the target question if the target conversation is a target question for the target object. Based on the target question, it performs standard question prediction and standard intent prediction based on the trained first language model. Based on the prediction results, it obtains the target standard question and the target standard intent corresponding to the target question. Based on the target question, it performs emotion level prediction based on the trained second language model to obtain the emotion level corresponding to the target question. The emotion level corresponding to the target question is used to represent the emotional degree of the conversation object, and the target standard intent corresponding to the target question is used to represent the intent of the conversation object. The third processing module is used to obtain the target response strategy from the set of response strategies corresponding to the target standard question. The set of response strategies includes the response strategy corresponding to each standard question among multiple standard questions. The response strategy corresponding to each standard question includes at least one response rule. Each response rule is associated with at least one object information, a standard intention, and an emotion level. The fourth processing module is used to obtain target object information, and based on the target object information, target standard intent and emotion level, determine the target response rules that match the target object information from the target response strategy; The fifth processing module is used to respond to the target standard question based on the target response rules, obtain the target standard response to the target standard question, and determine the target standard response as the response to the target question.

[0016] Thirdly, this disclosure provides an electronic device, including: a processor, a memory, and a bus; A bus is used to connect the processor and memory; Memory, used to store operation instructions; A processor is used to execute the intelligent customer service session method of the first aspect of this disclosure by invoking operation instructions.

[0017] Fourthly, this disclosure provides a computer-readable storage medium storing a computer program that is used to execute the intelligent customer service session method of the first aspect of this disclosure.

[0018] Fifthly, this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the intelligent customer service session method in the first aspect of this disclosure.

[0019] The technical solutions provided in this disclosure have at least the following beneficial effects: The process involves: 1) Obtaining the target session of the conversation object; if the target session is a target question for the target object, then using a pre-trained first language model to perform standard question prediction and standard intent prediction based on the target question. Based on the prediction results, the target standard question and target standard intent corresponding to the target question are obtained. Then, using a pre-trained second language model, emotion level prediction is performed based on the target question to obtain the emotion type and emotion level corresponding to the emotion type. The emotion level of the target question represents the emotional intensity of the conversation object, and the target standard intent represents the intent of the conversation object. 2) Obtaining target response strategies from the response strategy set corresponding to the target standard question. The response strategy set includes response strategies corresponding to each standard question in multiple standard questions. Each response strategy for a standard question includes at least one response rule, and each response rule is associated with at least one object information, one standard intent, one emotion type, and one emotion level corresponding to the emotion type. 3) Obtaining the target object information of the target object. Based on the target object information, target standard intent, emotion type, and emotion level, the target response rule matching the target object information is determined from the target response strategies. 4) Responding to the target standard question based on the target response rule to obtain the target standard response to the target standard question, and defining the target standard response as... The response addresses the target question. Leveraging the large language model's ability to classify questions, intentions, emotion types, and emotion levels, standard question prediction, standard intention prediction, and emotion level prediction are performed. The computational power of the large language model ensures the accuracy of the classification ability, guaranteeing the accuracy of the prediction results. This accurately identifies intentions (e.g., user intentions), emotion types, and emotion levels (e.g., the user's emotional level), providing data support for personalized responses (target standard responses). Based on target object information, target standard intentions, emotion types, and emotion levels, target response rules are determined. Based on these rules, the target standard question is answered, resulting in the target standard response. The system provides targeted standard answers to quasi-questions, resulting in different target standard answers based on different emotion types, emotion levels, and target audience information (personalized data). This provides both high recall performance for targeted standard answers and personalized, rule-based, accurate standard scripts (target standard answers), improving the accuracy of the scripts (responses) of e-commerce intelligent customer service. This, in turn, improves the efficiency of e-commerce intelligent customer service in answering user questions (target questions), avoiding risks and customer complaints caused by incorrect answers. The introduction of highly efficient intelligent customer service reduces the need for human customer service intervention, improves customer service response efficiency, and ultimately enhances the user's purchase Q&A experience. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments of this disclosure will be briefly introduced below.

[0021] Figure 1 This is a schematic diagram of the architecture of an intelligent customer service conversation system provided in an embodiment of this disclosure; Figure 2 A flowchart illustrating an intelligent customer service conversation method provided in an embodiment of this disclosure; Figure 3 A schematic diagram of an intelligent customer service session provided in an embodiment of this disclosure; Figure 4 A schematic diagram of an intelligent customer service session provided in an embodiment of this disclosure; Figure 5 A schematic diagram of an intelligent customer service session provided in an embodiment of this disclosure; Figure 6 A schematic diagram of an intelligent customer service session provided in an embodiment of this disclosure; Figure 7 A schematic diagram of an intelligent customer service session provided in an embodiment of this disclosure; Figure 8 A schematic diagram of an intelligent customer service session provided in an embodiment of this disclosure; Figure 9 A flowchart illustrating an intelligent customer service conversation method provided in an embodiment of this disclosure; Figure 10 A flowchart illustrating an intelligent customer service conversation method provided in an embodiment of this disclosure; Figure 11 This is a schematic diagram of the structure of an intelligent customer service conversation device provided in an embodiment of the present disclosure; Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0022] The embodiments of this disclosure are described below with reference to the accompanying drawings. It should be understood that the embodiments described below with reference to the accompanying drawings are exemplary descriptions for explaining the technical solutions of the embodiments of this disclosure, and do not constitute a limitation on the technical solutions of the embodiments of this disclosure.

[0023] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the terms “comprising” and “including” as used in embodiments of this disclosure mean that the corresponding feature can be implemented as the presented feature, information, data, step, operation, element, and / or component, but do not exclude implementation as other features, information, data, step, operation, element, component, and / or combinations thereof supported by the art. It should be understood that when we say that an element is “connected” or “coupled” to another element, the one element can be directly connected or coupled to the other element, or it can mean that the one element and the other element are connected through an intermediate element. Furthermore, “connected” or “coupled” as used herein can include wireless connection or wireless coupling. The term “and / or” as used herein indicates at least one of the items defined by the term; for example, “A and / or B” indicates implementation as “A,” or implementation as “B,” or implementation as “A and B.”

[0024] It is understood that in the specific embodiments of this disclosure, data related to intelligent customer service sessions is involved. When the above embodiments of this disclosure are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0025] To make the objectives, technical solutions, and advantages of this disclosure clearer, the embodiments of this disclosure will be described in further detail below with reference to the accompanying drawings.

[0026] This disclosure provides an intelligent customer service conversation method for an intelligent customer service conversation system, which relates to fields such as artificial intelligence.

[0027] To better understand and explain the solutions of the embodiments of this disclosure, some technical terms involved in the embodiments of this disclosure will be briefly explained below.

[0028] LLM (Large Language Model) is an artificial intelligence model designed to understand and generate human language. LLMs can be trained on massive amounts of text data and can perform a wide range of tasks, such as text summarization, translation, and sentiment analysis. LLMs have revolutionized learning paradigms across various research fields and show great potential in bridging the gap between classic recommenders and open-world knowledge. With their massive model and corpus size, LLMs demonstrate exceptional capabilities in areas such as problem-solving, logical reasoning, and creative writing. Learning from vast amounts of internet text, LLMs encode a wealth of open-world knowledge, ranging from basic factual information to complex social norms and logical structures. Therefore, LLMs can perform basic logical reasoning consistent with known facts and relationships.

[0029] Prompt: A prompt is a piece of text describing a task that is input by the user during the interaction between the user and the large language model.

[0030] BERT model: BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language representation model based on the Transformer architecture.

[0031] Large Language Model API: The business layer uses the large language model by calling the Large Language Model API (Application Programming Interface). The code in the Large Language Model API includes: model: Model category; wsid: Business ID; enable_stream: Whether to enable streaming response; headers: Request headers; json_data: The request body; query_id: Request ID, must be unique; message: a message; role: role (user / system); content: the content of the prompt; output_seq_len: Length of the output string.

[0032] The existing intelligent customer service conversation technologies have the following problems: (1) Retrieval-based approach: This approach retrieves and filters existing knowledge bases through search. Retrieval models typically utilize neural network technology, inputting large amounts of pre-trained corpus data for training. After training, the retrieval model can classify, match, and answer questions based on new inputs. The implementation of this approach relies heavily on large-scale pre-trained data and efficient retrieval algorithms. Problems with retrieval-based approaches include: While retrieval-based answers, derived from knowledge bases, can provide relatively high recall, simple search matching may result in some words not being found or being incorrectly searched, leading to incorrect responses. In e-commerce customer service scenarios, the accuracy of answering user questions is paramount; it is better to not answer than to answer incorrectly.

[0033] (2) Generative approach: This approach generates dialogues using the computational power of large language models, primarily employing deep learning techniques and large language models to generate text by learning from a large amount of data. This approach typically uses structures such as recurrent neural networks or transformers, capable of processing sequential data and generating new sequences. Unlike retrieval-based models, generative models directly generate target text during training, rather than through retrieval matching. Problems with generative approaches include: the generated answers are influenced by various factors, leading to uncontrollable results; because the technical capability of large language models is to generate relatively high-probability possible results through prediction, the final answer may be imprecise due to the scale of the original data and the training model.

[0034] In summary, both retrieval-based and generative solutions in existing technologies have certain drawbacks in terms of response accuracy and recall performance in intelligent customer service, and cannot be fully applied to real-world intelligent customer service scenarios.

[0035] Based on this, the present disclosure provides an intelligent customer service conversation method, apparatus, device, computer-readable storage medium, and program product, and the specific technical solutions will be described in detail below.

[0036] The solutions provided in this disclosure relate to artificial intelligence technology. The technical solutions of this disclosure will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this disclosure will now be described with reference to the accompanying drawings.

[0037] To better understand the solution provided in this disclosure, the solution will be described below in conjunction with a specific application scenario.

[0038] In one embodiment, Figure 1The diagram illustrates the architecture of an intelligent customer service conversation system applicable to embodiments of this disclosure. It is understood that the intelligent customer service conversation method provided in these embodiments can be applied to, but is not limited to, applications such as... Figure 1 In the application scenarios shown.

[0039] In this example, as Figure 1 As shown, the architecture of the intelligent customer service conversation system in this example may include, but is not limited to, server 10, terminal 20, and database 30. Server 10, terminal 20, and database 30 can interact with each other via network 40.

[0040] Server 10 obtains the target session of the session object; if the target session is a target question for the target object, Server 10 performs standard question prediction and standard intent prediction based on the target question using a pre-trained first language model. Based on the prediction results, it obtains the target standard question and the target standard intent corresponding to the target question. Then, based on the target question, it performs emotion level prediction using a pre-trained second language model to obtain the emotion type and emotion level corresponding to the emotion type. The emotion level of the target question is used to represent the emotional intensity of the session object, and the target standard intent is used to represent the intent of the session object. Server 10 obtains the set of response strategies corresponding to the target standard question. The target response strategy in the system includes a set of response strategies for each standard question among multiple standard questions. Each standard question's response strategy includes at least one response rule, and each response rule is associated with at least one type of object information, a standard intent, an emotion type, and an emotion level corresponding to that emotion type. Server 10 obtains the target object information and, based on this information, the target standard intent, emotion type, and emotion level, determines the target response rule that matches the target object information from the target response strategy. Server 10 responds to the target standard question based on the target response rule, obtaining the target standard response for the target standard question, and confirms the target standard response as the answer to the target question. Server 10 sends the target standard response to terminal 20, and the target standard response for the target question regarding the order or product is displayed in the terminal's intelligent customer service conversation interface. Server 10 then sends the target standard response to database 30 for storage.

[0041] It is understood that the above is only one example, and this embodiment is not limited here.

[0042] Terminals include, but are not limited to, smartphones (such as Android phones, iOS phones, etc.), mobile phone emulators, tablets, laptops, digital broadcast receivers, MIDs (Mobile Internet Devices), PDAs (Personal Digital Assistants), smart voice interaction devices, smart home appliances, and in-vehicle terminals.

[0043] A server can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server or server cluster that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0044] The aforementioned networks may include, but are not limited to, wired networks and wireless networks. Wired networks include local area networks (LANs), metropolitan area networks (MANs), and wide area networks (WANs). Wireless networks include Bluetooth, Wi-Fi, and other networks that enable wireless communication. Specific details can be determined based on actual application scenario requirements and are not limited here.

[0045] See Figure 2 , Figure 2 This illustration shows a flowchart of an intelligent customer service conversation method provided by an embodiment of the present disclosure. This method can be executed by any electronic device, such as a server. As an optional implementation, the method can be executed by a server. For ease of description, in the following description of some optional embodiments, execution by a server will be used as an example of the method's execution subject. Figure 2 As shown, the intelligent customer service conversation method provided in this embodiment includes the following steps: S201, Get the target session of the session object.

[0046] Specifically, e-commerce intelligent customer service includes a front-end and a back-end. The front-end is such as the e-commerce intelligent customer service chat interface, and the back-end is such as the server. The chat object is such as the user, and the target chat is such as the question entered by the user in the e-commerce intelligent customer service chat interface.

[0047] For example, such as Figure 3The architecture of the intelligent customer service chat system shown includes a front-end, an access layer, and a back-end. The front-end includes the front-end chat interface, while the back-end includes the intelligent customer service module, information retrieval, Wuji, CLS (CECloud LogService), and a large language model. The front-end chat interface handles user question-and-answer interactions; an example of a smart customer service chat interface in e-commerce is the front-end chat interface. The access layer serves as the common access and interaction layer between the front-end and back-end, providing HTTP... JSON data interaction; the intelligent customer service module provides intercept word filtering, question rewriting, sentiment recognition, response generation (question result recall), dynamic adjustment, and integrates the question rewriting process and sentiment recognition with a large language model, such as a pre-trained first language model, a pre-trained second language model, etc.; intercept word filtering is used to intercept non-safety questions input by users; question rewriting is used to integrate with the large language model for intent recognition and question rewriting into candidate standard questions; sentiment recognition is used to analyze user emotions and score them (emotion type and emotion level); response generation is used to generate responses (target standard answers) based on standard questions (target standard questions), sentiment scores (emotion type and emotion level), user intent (target standard intent), order status (personalized data, i.e., target object information), etc.; dynamic adjustment is used to generate decisions based on dynamic adjustment instructions and sentiment levels. The decision results enable the e-commerce platform (preset platform) to adjust its responses (target standard responses) and corresponding order processing based on the decision results; information retrieval is used to provide services for obtaining relevant data information on orders, products, and shopping carts. Information retrieval includes order modules, product modules, and shopping cart modules, which are used to provide services for obtaining relevant data information on orders, products, and shopping carts, respectively. The relevant data information on orders, products, and shopping carts is stored in DB (Database); Wuji is a database used to store standard question sets, rule judgment method sets, response strategy sets, standard response template sets, etc.; CLS is used for the persistence of conversation logs, for example, saving historical content in the front-end conversation interface to CLS; and the intelligent customer service conversation is converted into a human customer service conversation through the generation of enterprise WeChat customer service jump links.

[0048] S202, if the target conversation is a target question for the target object, then based on the target question, standard question prediction and standard intent prediction are performed using the first trained language model. Based on the prediction results, the target standard question and the target standard intent corresponding to the target question are obtained. Based on the target question, the emotion level prediction is performed using the second trained language model to obtain the emotion type and the emotion level corresponding to the emotion type. The emotion level corresponding to the target question is used to represent the emotional degree of the conversation object, and the target standard intent corresponding to the target question is used to represent the intent of the conversation object.

[0049] Specifically, target objects include orders and products; target conversations include target questions related to orders, such as "Haven't you shipped it yet?"; target conversations include target questions related to products, such as "What is the product model?"; target standard questions include "Why hasn't my order been shipped yet?"; target standard intents include order category - order delay - slow shipping; and emotion types include happy, peaceful, and angry; the emotion type "angry" corresponds to emotion levels such as emotion level 1 (slightly angry), emotion level 2 (somewhat angry), and emotion level 3 (very angry).

[0050] For example, if the target conversation is a target question for the target object, then based on the target question, the standard question prediction process is performed using the trained first language model to obtain the prediction result; based on the prediction result, the target standard question corresponding to the target question is obtained; where the prediction result is the candidate standard question.

[0051] For example, if the target conversation is a target question for the target object, then based on the target question, standard intent prediction is performed using the trained first language model to obtain the prediction result; based on the prediction result, the target standard intent corresponding to the target question is obtained; where the prediction result is the target standard intent.

[0052] For example, such as Figure 4 As shown, the prompt word and the target question (user question) are input into the trained first language model. The trained first language model performs standard question prediction processing to obtain the candidate standard question corresponding to the target question, and performs standard intent prediction processing to obtain the target standard intent (user intent) corresponding to the target question. The target question (user question) is input into the trained second language model. The trained second language model performs emotion level prediction processing to obtain the emotion type and emotion level corresponding to the target question.

[0053] S203, obtain the target response strategy from the response strategy set corresponding to the target standard question. The response strategy set includes the response strategy corresponding to each standard question among multiple standard questions. The response strategy corresponding to each standard question includes at least one response rule. Each response rule is associated with at least one object information, a standard intention, an emotion type, and an emotion level corresponding to an emotion type.

[0054] Specifically, for example, the database includes a set of response strategies, which includes multiple response strategies. Among the multiple response strategies, there is a target response strategy. Each of the multiple response strategies includes multiple response rules. Each of the multiple response rules is associated with at least one type of object information, a standard intent, an emotion type, and an emotion level corresponding to an emotion type. The object information includes, for example, the type of the target object and the state of the target object. Each response rule includes multiple condition fields.

[0055] For example, the target standard question is "Why hasn't my order been shipped yet?", and the target standard question "Why hasn't my order been shipped yet?" corresponds to a target response strategy in the response strategy set. This target response strategy has multiple response rules, and one of the response rules is, for example, "It is a Caochangdi order - It is a pre-sale item - Final payment pending - Angry - Emotion level 1 - Slow delivery". This response rule includes multiple condition fields, which are "It is a Caochangdi order", "It is a pre-sale item", "Final payment pending", "Angry", "Emotion level 1", and "Slow delivery". Among them, at least one object information is "It is a Caochangdi order", "It is a pre-sale item", and "Final payment pending", one standard intent is "Order type - Order delay - Slow delivery", one emotion type is "Angry", and the emotion level corresponding to the emotion type "Angry" is "Emotion level 1".

[0056] S204, Obtain target object information of the target object, and based on the target object information, the target standard intent, emotion type and the emotion level, determine the target response rule that matches the target object information from the target response strategy.

[0057] Specifically, for example, obtaining two types of target object information: one type, such as the target object's type, and the other type, such as the target object's status. The target object's type could be, for example, "is a grass field order," and its status could be, for example, "is a pre-sale item" or "pending final payment." The target's standard intent could be, for example, order type - order delay - slow delivery. The emotion type is "angry," and the corresponding emotion level for "angry" is "emotion level 1." The target response rule could be, for example, "is a grass field order - is a pre-sale item - pending final payment - angry - emotion level 1 - slow delivery."

[0058] It should be noted that, for example, "is a grass field order - is a pre-sale item - pending final payment" can correspond to different emotion types, different emotion levels, and different intentions; thus, the target response rules can be determined through multiple dimensions, namely "is a grass field order - is a pre-sale item - pending final payment", "emotion type", "emotion level", and "intention".

[0059] S205, based on the target response rules, respond to the target standard question, obtain the target standard response to the target standard question, and determine the target standard response as the response to the target question.

[0060] Specifically, for example, if the target standard question is "Why hasn't my order been shipped yet?", based on the target response rules, the target standard question "Why hasn't my order been shipped yet?" is answered to obtain the target standard response to the target standard question "Why hasn't my order been shipped yet?", which is "Dear, we found that your order has not yet been paid for. It can only be shipped after the final payment is made."

[0061] For example, the content dimension set, emotion type, and emotion level of the target standard response template are correlated. The content dimension set includes at least one of the following: general fields, tone, and specific fields. The specific fields include relevant operations for the target object.

[0062] For example, if the emotion type is "angry," and the corresponding emotion level 3 is "very angry," then the emotion type "angry" - emotion level 3 "very angry" corresponds to a target standard response template. This target standard response template's content dimensions include general fields, tone, and specific fields. Specific fields include relevant actions for the target object (e.g., the order). The general fields include "Your order is currently in {queue position} and is expected to ship at {shipping time}." The tone is mild and polite. The specific fields include relevant actions for the target object (e.g., the order): "To expedite delivery, you can choose from the following courier services: 1. Standard courier (free); 2. Express courier (additional fee of {express fee} yuan); Please reply with the number you choose, or tell us your needs."

[0063] For example, the emotion type is "peaceful", the emotion level 1 corresponding to "peaceful" is calm, and the emotion type "peaceful" - emotion level 1 "calm" corresponds to a target standard response template. The content dimension set of this target standard response template includes general fields and tone; among them, the general fields include "Your order is expected to be shipped at {shipping time}", and the tone is concise and direct; the general fields only need to include the shipping time.

[0064] It should be noted that the fields differ for different emotion types; and even for different emotion levels within the same emotion type, the fields differ. For example, the emotion type "anger" corresponds to emotion levels such as Emotion Level 1 (slight anger), Emotion Level 2 (somewhat angry), and Emotion Level 3 (very angry). The fields for Emotion Level 1 (slight anger) include the scheduled shipping time and the reason for the delay; these are standard fields. The fields for Emotion Level 2 (somewhat angry) include queue position, a more precise shipping time, and the reason for the delay; these are standard fields. The fields for Emotion Level 3 (very angry) include queue position, a more precise shipping time, the reason for the delay, and related actions for the order. Related actions for the order include providing alternative courier options and inquiring whether to switch to a faster courier. Among these, "queue position, more precise shipping time, and reason for the delay" are standard fields, while "related actions for the order" are specific fields.

[0065] For example, target response rule 1 is "It's a grass field order - it's a pre-sale item - final payment pending - angry - emotion level 1 - slow delivery", target response rule 2 is "It's a grass field order - it's a pre-sale item - final payment pending - angry - emotion level 2 - slow delivery", and target response rule 3 is "It's a grass field order - it's a pre-sale item - final payment pending - angry - emotion level 3 - slow delivery". Different target response rules (target response rule 1, target response rule 2, and target response rule 3) correspond to different target standard responses, that is, different emotion levels of the same emotion type correspond to different target standard responses.

[0066] In this embodiment, the following steps are taken: First, the target session of the conversation object is obtained. If the target session is a target question for the target object, then based on the target question, a first trained language model is used for standard question prediction and standard intent prediction. Based on the prediction results, the target standard question and the target standard intent corresponding to the target question are obtained. Then, based on the target question, a second trained language model is used for emotion level prediction to obtain the emotion type and the emotion level corresponding to the emotion type. The emotion level corresponding to the target question is used to characterize the emotional degree of the conversation object, and the target standard intent corresponding to the target question is used to characterize the intent of the conversation object. Next, the target response strategy in the response strategy set corresponding to the target standard question is obtained. The response strategy set includes response strategies corresponding to each standard question among multiple standard questions. Each response strategy corresponding to a standard question includes at least one response rule, and each response rule is associated with at least one object information, a standard intent, an emotion type, and an emotion level corresponding to an emotion type. Finally, the target object information of the target object is obtained. Based on the target object information, the target standard intent, the emotion type, and the emotion level, a target response rule matching the target object information is determined from the target response strategy. Based on the target response rule, the target standard question is answered to obtain the target standard answer to the target standard question, and the target standard answer is then... The response is then determined to be a response to the target question. Thus, leveraging the large language model's ability to classify questions, intentions, emotion types, and emotion levels, standard question prediction, standard intention prediction, and emotion level prediction are performed. The computational power of the large language model ensures the accuracy of the classification ability, guaranteeing the accuracy of the prediction results. This accurately identifies intentions (e.g., user intentions), emotion types, and emotion levels (e.g., the user's emotional level), providing data support for personalized responses (target standard responses). Based on the target object information, target standard intentions, emotion types, and emotion levels, target response rules are determined. Based on these rules, the target standard question is answered, resulting in the desired response. The system targets standard answers to standard questions, generating different standard answers based on different emotion types, emotion levels, and target audience information (personalized data). This provides high recall performance for target standard answers and yields personalized and rule-based accurate standard responses, improving the accuracy of responses from e-commerce intelligent customer service. This, in turn, increases the efficiency of e-commerce intelligent customer service in answering user questions (target questions), avoiding risks and customer complaints caused by incorrect answers. The introduction of highly efficient intelligent customer service reduces the need for human intervention, improves customer service response efficiency, and ultimately enhances the user's purchase and Q&A experience.

[0067] In one embodiment, based on the target question, standard question prediction and standard intent prediction are performed using a pre-trained first language model. Based on the prediction results, the target standard question and the target standard intent corresponding to the target question are obtained, including steps A1-A3: Step A1: Obtain prompt words. Prompt words are used to instruct the trained first language model to determine the candidate standard questions corresponding to the input question from the question set and to select the candidate standard intents corresponding to the input question from the intent set. The prompt words include the question set and the intent set. The question set includes multiple preset candidate standard questions, and the intent set includes multiple candidate standard intents.

[0068] Specifically, examples of prompt words are shown below: (Prompt): Imagine you are an e-commerce customer service representative tasked with rewriting code. Given [User Questions], [E-commerce Question Library], [Greeting Library], [User Intent Library], and [Dynamically Adjusting Instruction Library], each line represents a question. Please process the code as follows: 1. User Intent Recognition: First, analyze the core intent of the [user question] (e.g., order status, product inquiry, after-sales service, etc.); 2. Classification Judgment: Based on the identified user intent, classify the question into either "question" or "greeting"; 3. Standard Question Matching: (1) If it is classified as a question, then select the question with the closest intent from the [e-commerce question library] to replace the user's question. If there is no candidate standard question with similar intent, then output "No candidate standard question found"; (2) If the category is small talk, select the question with the closest intent from the [small talk library] to replace the user's question. If there is no candidate standard question with similar intent, output "No candidate standard question found"; Finally, generate a response according to the [requirements] and [output format]; [User Issues]=""" {query} """ [E-commerce FAQ]=""" Why hasn't my order been shipped yet? (Order) Does this product support a 7-day no-reason return policy? (Product or order) What is the material of this product? (Uncategorized) How do I turn off automatic membership renewal? (Uncategorized) Why can't members watch this show? (Uncategorized) What are the dimensions of this item? (Item or order) Are all the products genuine? (Uncategorized) When will this item be shipped after I purchase it? (Item or order) I want to return the item. (Order) How do I use this product? (Uncategorized) Can I exchange this item? (Product or order) How do I use the free order coupon? (Uncategorized) I haven't received my refund! Where can I view my orders? (Uncategorized) Which courier service should I use? (Uncategorized) Change the shipping address. (Uncategorized) When will this item be restocked? (Uncategorized) One item was missing from my shipment. (Uncategorized) The item I received was damaged. (Uncategorized) """ [Hush Greetings]=""" Hello. (Uncategorized) How's the weather today? (Uncategorized) I'm very unhappy. (Uncategorized) How can I be happy? (Uncategorized) I'm so angry! (Uncategorized) I'm so anxious! (Uncategorized) """ [User Intent Library] = { "Order Category": { "Order Monitoring": ["Slow logistics progress updates", "Shipping status missing"], "Order Delay": ["Slow Shipping", "Reason for Delay"], "Order Cancellation": ["Reason for Cancellation", "Order Cancellation"] }, "Product Category": { Product Inquiry: ["Size Parameters", "Material Specifications", "Features"], "Out of Stock Inquiry": ["Restock Time", "Availability", "Pre-order Time"] }, "After-sales service": { "Returns and Exchanges": ["Return", "Exchange for Size"], "Repair Service": ["Warranty Period", "Fault Reporting", "After-Sales Service Network"] } "Pricing Category": { "Pricing": ["Pricing is too high", "Pricing is unreasonable"], "Settlement": ["Price settlement error", "Price discount error"] } } [Dynamically adjust instruction library] = { "price": { "command": "execute_price_adjustment", "action": [ 1 :"High", 2 :"Low" ] }, "Order": { "command": "trigger_express_priority", "action": [ 1: "Extend shipping request", 2: "Delayed shipping request" ] } } [Requirement]=""" 1. You need to match the [User Question] with the [E-commerce Question Database] and the [Greeting Database] in sequence, and output candidate standard questions. If there are no candidate standard questions with similar intent, output "No candidate standard questions found"; 2. No need to rewrite standard questions, nor to answer user questions; 3. Please do not output questions other than those in the knowledge base; 4. First, perform user intent recognition, then complete classification judgment and standard question matching, and deduce dynamic adjustment instructions according to the [dynamic adjustment instruction library]; 5. User intent recognition needs to distinguish specific types (such as order / product / membership, etc.), which should be reflected in the `relate_entity` field; 6. Standard question matching must simultaneously satisfy intent matching and semantic similarity; """ [Output format]=""" {{ "category": User question classification (question / greeting); "standard_question": When categorized as a question, output the question most relevant to the user's question from the [e-commerce question library]. Do not rewrite it; delete the last parenthesis and its contents. If there are no candidate standard questions with similar intent, output "No candidate standard question found". When categorized as a greeting, output the question most relevant to the user's question from the [greeting library]. Do not rewrite it. If there are no candidate standard questions with similar intent, output "No candidate standard question found". Remove the parenthesis content after the question; for example, "The goods I received are damaged. (order)", only fill in "The goods I received are damaged". “relate_entity”: When categorized as an issue, fill in the order / product / no category in parentheses after the candidate standard issue in this field; the user intent category corresponding to the candidate standard issue (e.g., order / product / price / no category). "relate_content": Content and dynamic adjustment instructions based on user intent; }} """ It should be noted that the e-commerce question library is a set of questions, user questions are target questions, the user intent library is a set of intents, and the user intents in the user intent library are candidate standard intents.

[0069] Step A2: Input the prompt word and target question into the trained first language model, perform standard intent prediction processing through the trained first language model to obtain the candidate standard intent corresponding to the target question, and perform standard question prediction processing through the trained first language model.

[0070] Specifically, for example, the prompt word and the target question (What is the material?) are input into the trained first language model. The trained first language model performs standard question prediction processing to obtain the candidate standard question (What is the material of this product?) corresponding to the target question (What is the material of this product?). Based on the predicted candidate standard questions, the target standard question (What is the material of this product?) corresponding to the target question is obtained.

[0071] For example, after the large language model service is deployed, it provides an HTTP API, namely the large language model API, for the business layer to call, such as... Figure 3 The question rewriting process in the intelligent customer service module shown is integrated with the large language model. By calling the large language model API, the prompt words and target question are input into the trained large language model. The trained first large language model performs standard question prediction processing to obtain the candidate standard question corresponding to the target question.

[0072] For example, the prompt word and the target question (Why haven't you shipped the goods yet?) are input into the trained first language model. The trained first language model performs standard intent prediction processing to obtain the candidate standard intent (order category - order delay - slow delivery) corresponding to the target question (Why haven't you shipped the goods yet?).

[0073] Step A3: Based on the predicted candidate standard questions, obtain the target standard question corresponding to the target question, and based on the predicted candidate standard intent, obtain the target standard intent corresponding to the target question.

[0074] Specifically, for example, the predicted candidate standard problems are matched with each preset standard problem in the standard problem set of the database, and the preset standard problems in the standard problem set that match the predicted candidate standard problems are taken as target standard problems.

[0075] For example, the predicted candidate standard intents are determined as the target standard intents corresponding to the target question.

[0076] It should be noted that by leveraging the large language model's ability to classify questions and intents, standard question prediction and standard intent prediction are performed. The computational power of the large language model ensures the accuracy of the classification ability, thereby guaranteeing the accuracy of the prediction results and accurately identifying intents (such as user intents), providing data support for personalized responses (target standard answers).

[0077] In one embodiment, obtaining the target standard problem corresponding to the target problem based on the predicted candidate standard problems includes: The predicted candidate standard problems are matched with each preset standard problem in the standard problem set of the database, and the preset standard problems in the standard problem set that match the predicted candidate standard problems are taken as the target standard problems.

[0078] Specifically, for example, such as Figure 3 As shown, Wuji is a database containing a set of standard questions, a set of rule-based judgment methods, a set of response strategies, and a set of standard response templates. Preset standard questions in the standard question set include, for example, "Why hasn't my order been shipped yet?", "Does this product support a seven-day no-reason return policy?", and "What is this product made of?". For example, the predicted candidate standard question "Why hasn't my order been shipped yet?" is matched with each preset standard question in the standard question set of the database, and the preset standard question "Why hasn't my order been shipped yet?" that matches the predicted candidate standard question is taken as the target standard question.

[0079] It should be noted that the predicted candidate standard questions are matched with each preset standard question in the standard question set of the database. That is, the candidate standard questions output by the large language model are matched, and the most similar preset standard question is matched. The most similar preset standard question is then used as the target standard question, thereby improving the accuracy of determining the target standard question.

[0080] In one embodiment, based on the target question, a pre-trained second-largest language model is used to predict the emotion level, resulting in the emotion type corresponding to the target question and the emotion level corresponding to the emotion type, including: The target question is input into the trained second language model, and word segmentation is performed to obtain the corresponding word segmentation for the target question. Based on the word segmentation corresponding to the target question, the word segmentation encoding representation corresponding to the target question is obtained through encoding processing; Based on the word segmentation encoding representation corresponding to the target question, the emotion type corresponding to the target question and the emotion level corresponding to the emotion type are obtained through classification processing.

[0081] Specifically, the second largest language model is, for example, sentiment classification models based on the BERT model. For example, ... Figure 5 As shown, the target question (Why haven't you shipped the goods yet?) is input into the trained second language model. Through word segmentation, the word segments corresponding to the target question are obtained. Based on the word segments corresponding to the target question, the word segmentation encoding representation corresponding to the target question is obtained through encoding. The hidden part of the word segmentation encoding representation, such as the word segmentation encoding representation corresponding to the hidden punctuation marks, is also obtained. Based on the word segmentation encoding representation other than the partial word segmentation encoding representation, the emotion type corresponding to the target question and the emotion level corresponding to the emotion type are obtained through classification.

[0082] It should be noted that by leveraging the large language model's ability to classify emotion types and levels, and performing emotion level prediction processing, the computational power of the large language model ensures the accuracy of the classification ability, thereby guaranteeing the accuracy of the prediction results. This allows for the precise identification of emotion types and levels (such as the user's emotion level), providing data support for personalized responses (target standard replies).

[0083] In one embodiment, the trained second-largest language model is obtained by training in the following way: Obtain the training sample set: Obtain the labeled sentiment analysis dataset (IMDB movie review dataset, Twitter sentiment dataset, etc.); Based on the labeled sentiment analysis dataset, obtain the training sample set, and pair each training sample in the training sample set with the corresponding sentiment label (sentiment label, for example, 1 represents calm, 2 represents slightly angry, 3 represents very angry, etc.); To obtain the second language model to be trained: add a classifier (e.g., nn.Linear) to the output layer of the pre-trained BERT model to obtain the emotion classification model; determine the emotion classification model as the second language model to be trained; the emotion type (e.g., anger) and emotion level (e.g., emotion level 1, emotion level 2, emotion level 3, etc.) can be output through the classifier of the emotion classification model.

[0084] Based on the training sample set, the second language model to be trained is fine-tuned using the cross-entropy loss function to obtain a well-trained second language model.

[0085] It should be noted that the performance of the trained second language model is evaluated on the validation set, with performance metrics such as accuracy and F1 score.

[0086] In one embodiment, based on the target standard problem, the target rule judgment method corresponding to the target standard problem is determined from the set of rule judgment methods in the database; Based on target audience information, target standard intent, emotion type, and emotion level, target response rules that match the target audience information are determined from the target response strategies, including: Based on the target rule judgment method, target object information, target standard intent, emotion type and emotion level, target response rules that match the target object information are determined from the target response strategies.

[0087] Specifically, for example, if the target standard question is "Why hasn't my order been shipped yet?", the corresponding target rule judgment methods for this question include: Grassland order / Member card order, pre-sale / non-pre-sale (in stock), customized / non-customized, package status, estimated shipping time, dispute status, and whether the specific order has been sent / not sent in this session. Target object information includes, for example, order type and order status in order information; target object information also includes, for example, product type and product status in product information; order type, order status, product type, and product status are stored in... Figure 3 In CLS.

[0088] For example, such as Figure 3 As shown, Wuji is a database containing a set of standard questions, a set of rule judgment methods, a set of response strategies, and a set of standard response templates. The set of standard questions includes multiple preset standard questions. The set of judgment methods includes the rule judgment methods corresponding to each preset standard question. The set of response strategies includes multiple response strategies, including a target response strategy. Each response strategy includes multiple response rules, and each response rule includes multiple condition fields.

[0089] For example, as shown in Table 1, a preset standard question in the database is "Why hasn't my order been shipped yet?", which is the target standard question. The multiple target rule judgment methods corresponding to this preset standard question include: Caochangdi order / membership card order, pre-sale / non-pre-sale (in stock), customized / non-customized, package status, estimated shipping time, rights protection status, and whether the specific order has been sent / not sent in this round of conversation. This preset standard question corresponds to a response strategy, which includes multiple response rules. One such response rule is "It's a Caochangdi order - It's a pre-sale item - Final payment pending - Angry - Emotional level 1 - Slow shipping." This response rule includes multiple condition fields: "It's a Caochangdi order," "It's a pre-sale item," "Final payment pending," "Angry," "Emotional level 1," and "Slow shipping." The standard response template corresponding to this response rule is: "Dear customer, we found that your order has not yet been shipped. You can click the button below to transfer to a live customer service representative."

[0090] Table 1: Examples of Pre-defined Standard Questions in the Database

[0091] For example, based on the target standard question "Why hasn't my order been shipped yet?", the target rule judgment method corresponding to the target question is determined from the rule judgment method set in the database; based on the target rule judgment method, target object information, target standard intent, emotion type, and emotion level, the target response rule matching the target object information is determined from the target response strategy: "It's a Caochangdi order - It's a pre-sale item - Final payment pending - Angry - Emotion level 1 - Slow shipping"; where the target rule judgment method includes, for example, Caochangdi order / membership card order, pre-sale / non-pre-sale (in stock), customized / non-customized, package status, estimated shipping time, rights protection status, and whether the specific order has been sent / not sent in this round of conversation; target object information includes, for example, order type and order status in the order information, which is stored in... Figure 3 In CLS.

[0092] It should be noted that, for example, "is a grass field order - is a pre-sale item - pending final payment" can correspond to different emotion types, different emotion levels, and different intentions. Based on the target object information, target standard intentions, emotion types, and emotion levels, target response rules are determined. Based on the determined target response rules, the target standard questions are answered, resulting in target standard answers to the target standard questions. This provides high recall performance for target standard answers and obtains personalized and rule-accurate standard scripts (target standard answers), improving the accuracy of the scripts (responses) of e-commerce intelligent customer service, thereby improving the efficiency of e-commerce intelligent customer service in answering user questions (target questions).

[0093] In one embodiment, based on the target response rules, a target standard question is answered to obtain a target standard response to the target standard question, including: Based on the target response rules and the standard response template set in the database, a target standard response template that matches the target standard question is determined from the standard response template set. The content dimension set, emotion type and emotion level of the target standard response template are correlated. The content dimension set includes at least one of the following: regular fields, tone and specific fields. The specific fields include relevant operations for the target object. Based on the object information of the target object and the target standard response template, the target standard response to the target standard question is obtained.

[0094] Specifically, for example, the emotion type is "angry," and the corresponding emotion level 3 is "very angry." The emotion type "angry" - emotion level 3 "very angry" corresponds to a target standard response template. The content dimensions of this target standard response template include general fields, tone, and specific fields. Specific fields include relevant actions for the target object (e.g., the order). The general fields include "Your order is currently in {queue position} and is expected to ship at {shipping time}." The tone is mild and polite. Specific fields include relevant actions for the target object (e.g., the order): "To expedite delivery, you can choose from the following courier services: 1. Standard courier (free); 2. Express courier (additional fee of {express fee} yuan); Please reply with a number to select, or inform us of your needs."

[0095] For example, the emotion type is "peaceful", the emotion level 1 corresponding to "peaceful" is calm, and the emotion type "peaceful" - emotion level 1 "calm" corresponds to a target standard response template. The content dimension set of this target standard response template includes general fields and tone; among them, the general fields include "Your order is expected to be shipped at {shipping time}", and the tone is concise and direct; the general fields only need to include the shipping time.

[0096] It should be noted that the fields differ for different emotion types; and even for different emotion levels within the same emotion type, the fields differ. For example, the emotion type "anger" corresponds to emotion levels such as Emotion Level 1 (slight anger), Emotion Level 2 (somewhat angry), and Emotion Level 3 (very angry). The fields for Emotion Level 1 (slight anger) include the scheduled shipping time and the reason for the delay; these are standard fields. The fields for Emotion Level 2 (somewhat angry) include queue position, a more precise shipping time, and the reason for the delay; these are standard fields. The fields for Emotion Level 3 (very angry) include queue position, a more precise shipping time, the reason for the delay, and related actions for the order. Related actions for the order include providing alternative courier options and inquiring whether to switch to a faster courier. Among these, "queue position, more precise shipping time, and reason for the delay" are standard fields, while "related actions for the order" are specific fields. For example, as shown in Table 1, based on the target response rule "It is a grass field order - all goods - order shipped and awaiting receipt - angry - emotion level 1 - slow delivery", the target standard response template matching the target standard question "Why hasn't my order been shipped yet?" is determined from the standard response template set: "Dear, we found that your order was shipped on XX / XX / XX. You can check the logistics information through the order details." Based on the target object's information and the target standard response template, the target standard response for the target standard question is obtained: "Dear, we found that your order was shipped on February 15, 2021. You can check the logistics information through the order details." The target object's information includes the actual shipping time, i.e., February 15, 2021.

[0097] For example, such as Figure 6 As shown, based on the rewritten question (candidate standard question) and the standard question registration table (standard question set), standard question matching is performed to obtain the standard question ID (Identity document), which is the target standard question; based on the target standard question, the judgment method set, the target object information (e.g., order information / product information), the target standard intent, emotion type, emotion level, the rule registration table (response strategy set), and the standard response template set, rule calculation and matching are performed to obtain the response template ID; based on the response template ID, the response template registration table (standard response template set), and the target object information (e.g., order information / product information), the response is generated (target standard response).

[0098] It should be noted that the matching between standard questions (preset standard questions) and response templates (standard reply templates) is shown in Table 1 of the AI ​​(Artificial Intelligence) customer service knowledge base (database). Table 1 is stored in... Figure 3 The mapping relationship in Table 1 is implemented directly by hard-coding, which shortens the development cycle but reduces subsequent scalability. Since the standard question, response template, and the rule mapping relationship between them (response rules) will change frequently with business development, we consider using a configuration-based approach to allow flexible adjustments by product and operations teams. The relationship between the three is: standard question + (several) condition fields (a response rule, which includes multiple condition fields, and there is an AND relationship between the multiple condition fields) -> standard response template. Among them, the condition fields include standard intent, emotion type, emotion level, etc. The standard question and standard response template are relatively stable and can be managed through registry 1. The core is the modeling and maintenance of the rule mapping relationship. Based on the determined target response rule, the target standard question is answered to obtain the target standard response to the target standard question. This provides high recall performance for the target standard response and obtains accurate standard scripts (target standard responses) for the rule class, which improves the accuracy of the scripts (responses) of the e-commerce intelligent customer service and thus improves the efficiency of the e-commerce intelligent customer service in answering user questions (target questions).

[0099] In one embodiment, for example, such as Figure 7 As shown, business analysts write requirements documents, resulting in a requirements document; development engineers write rules (response rules) based on the requirements document (understanding the requirements), resulting in a DSL (domain-specific language); rules based on the DSL are published and stored in the rule base; the client will... Figure 3 Context information from the CLS is passed into the rule engine Drools ( Figure 3The generated response (shown) reads the corresponding rules from the rule base using the Drools rule engine. The Drools rule engine includes a pattern matcher and an executor. For example, contextual information (the order is a Caochangdi order, the goods in the order are pre-sale items, the balance is pending payment, the emotion type is angry, the emotion level is 2, and delivery is slow) is input into the Drools engine. The pattern matcher in the Drools engine matches the corresponding target response rule from the rule base (it's a Caochangdi order - it's a pre-sale item - the balance is pending payment - angry - emotion level 2 - slow delivery). Based on the target response rule, the pattern matcher in the Drools engine matches the corresponding target standard response template. Based on the emotion level of a certain emotion type, the tone and wording of the target standard response template are adjusted. The tone and wording of the target standard response template corresponding to emotion level 1 (calm) for the emotion type "peaceful" are: concise and direct. The tone and wording of the target standard response template corresponding to emotion level 1 (slightly angry) for the emotion type "angry" are: The tone and wording of the template are as follows: more polite tone, with added explanatory content; the tone and wording of the target standard response template corresponding to the emotion type "angry" and emotion level 3 (very angry) are as follows: gentler tone, providing more solutions, expressing apology, and generating a final reply; explanatory content can be added to the target standard response template. The implementation logic of the explanatory content includes: querying information such as the reason for the delay, queue position, and shipping time from the order system, and filling the target standard response template with the queried information; the target standard response template is as follows: "Your order is delayed due to {delay_reason} and is expected to be shipped at {estimated_ship_time}. We are working hard to speed up the processing. Thank you for your patience." Here, the variable fields delay_reason and estimated_ship_time are personalized data (object information of the target object); filling the target standard response template with personalized data generates the target standard response to the target standard question.

[0100] It should be noted that the standard response templates in the standard response template set may include variable fields, which are dynamically populated based on personalized data such as order and product status.

[0101] In one embodiment, for example, there are AND and OR relationships between rules (condition fields). Since the standard response (preset standard response) is filled in by product operations, the need to maintain the OR relationship is weak (i.e., there are fewer instances of A+B->E and C+D->E), and the relevant mapping can be achieved by configuring multiple rules, which is in line with the product operations and maintenance habits.

[0102] It should be noted that standard question A + (several) condition fields B (one answer rule) -> standard answer template E, and standard question C + (several) condition fields D (another answer rule) -> standard answer template E. In general, there is a one-to-one correspondence between standard questions, answer rules, and standard answer templates. Only in a few cases does A + B -> E and C + D -> E occur simultaneously.

[0103] For example, priority can be added between multiple mapping relationships (multiple response rules). This is mainly used for matching fallback rules (condition fields) such as "all products" in the AI ​​customer service knowledge base, and takes effect when other condition fields are not met.

[0104] For example, given a target question, the trained first language model outputs a candidate standard question A. However, none of the preset standard questions in the AI ​​customer service knowledge base match candidate standard question A. In this case, a pre-set response is used as the answer to candidate standard question A. The priority order of the responses is: Response A > Response B > Response C. Response A has the highest priority, so it is output first. Response A could be something like, "Sorry, please re-enter your question!" If the user enters another target question, and the resulting candidate standard question B does not match any of the preset standard questions, then response B is output. Response B could be something like, "Please enter the relevant auxiliary information for the order."

[0105] For example, the condition fields are as follows: {"key":"order.id","op":"EQAUL""value":"123456"}, / / Order id=123456; {"key":"order.trans_price_sum","op":"NOT_EQAUL","value":"0"}, / / The shipping cost for the order is not equal to 0.

[0106] It should be noted that the two condition fields are "order id=123456" and "the order's shipping fee is not equal to 0".

[0107] For example, the relationship between the condition fields is as follows: {"type":"combine","op":"AND", / / Connected by the "AND" operator, meaning both condition fields must be satisfied; "rules":[ {"type":"combine","op":"OR", / / Connected by the "OR" operator, meaning either condition field must satisfy the condition; "rules" {"key":"order.id","op":"EQAUL","value":"123456"}, / / Order id=123456; {"key":"order.id","op":"EQAUL","value":"456789"}, / / Order id=456789; ]}, {"key":"order.trans_price_sum","op":"NOT_EQAUL","value":"0"}, / / The shipping cost for the order is not equal to 0.

[0108] It should be noted that when connecting with the "OR" operator, either of the two condition fields must be satisfied, meaning either order id=123456 or order id=456789 must be satisfied; the relationship between order id=123456 and order id=456789 is OR. When connecting with the "AND" operator, both condition fields must be satisfied, meaning either order id=123456 or order id=456789, and the order's shipping fee must not be equal to 0.

[0109] In one embodiment, the target object is an order or a product, and after determining the target standard response as a response to the target question, the method further includes: The target standard response is sent to the terminal so that the target standard response to the target question for the order or product is displayed in the terminal's intelligent customer service conversation interface.

[0110] Specifically, for example, such as Figure 3 As shown, the front-end conversation interface is the terminal's intelligent customer service conversation interface; the target standard reply is "Dear customer, we found that your order was shipped on February 15, 2021. You can check the logistics information through the order details." This target standard reply is sent to the terminal, and the terminal's intelligent customer service conversation interface displays the target standard reply.

[0111] It should be noted that the introduction of highly efficient intelligent customer service reduces the need for human intervention, improves customer service response efficiency, and ultimately enhances the user's purchase and Q&A experience.

[0112] In one embodiment, based on the target question, a dynamically adjusted instruction prediction process is performed using a trained first language model to obtain a dynamically adjusted instruction. The dynamically adjusted instruction is used to adjust the target standard response on the preset platform. Based on the dynamic adjustment instructions, emotion type, and the corresponding emotion level, the target decision function is matched from the set of decision functions of the preset platform. Based on the dynamic adjustment instructions, emotion type, and the corresponding emotion level, a decision result is generated through the target decision function and sent to the preset platform. The preset platform then adjusts the target standard response based on the decision result to obtain a new response, which is then sent to the conversation object.

[0113] Specifically, dynamic adjustment instructions, such as urging shipment, and target decision functions, such as expedited logistics API functions.

[0114] For example, such as Figure 4 As shown, the prompt words and target question are input into a pre-trained first language model. The pre-trained first language model performs dynamic adjustment instruction prediction processing to obtain the dynamic adjustment instruction. The response generation forwards the dynamic adjustment instruction, emotion type, and corresponding emotion level to the dynamic adjustment. Based on the dynamic adjustment instruction, emotion type, and corresponding emotion level, the dynamic adjustment matches the target decision function from the decision function set of the preset platform (e-commerce platform), i.e., inverse feedback. The dynamic adjustment calls the target decision function from the decision function set of the preset platform. Based on the dynamic adjustment instruction, emotion type, and corresponding emotion level, the dynamic adjustment generates a decision result through the target decision function and sends the decision result to the preset platform. The preset platform adjusts the target standard response based on the decision result, obtains a new response, and sends the new response to the conversation partner.

[0115] For example, such as Figure 8 As shown, the dynamic adjustment obtains dynamic adjustment instructions (e.g., urging delivery), emotion type (e.g., anger), and the corresponding emotion level (e.g., emotion level 3). The dynamic adjustment then executes a decision, which includes: based on the dynamic adjustment instructions (e.g., urging delivery), emotion type (e.g., anger), and the corresponding emotion level (e.g., emotion level 3), the dynamic adjustment matches the target decision function (dynamic event, e.g., expedited logistics API function) from the preset platform's (e.g., e ... The system adjusts the response based on dynamic adjustment instructions (e.g., urging delivery), emotion type (e.g., anger), and the corresponding emotion level (e.g., emotion level 3). A decision result (e.g., expedited delivery) is generated through a target decision function. The dynamic adjustment sends the decision result to the preset platform. The preset platform adjusts the target standard response based on the decision result, obtaining a new response (e.g., "You have been upgraded to expedited delivery; the estimated delivery time is moved up to tomorrow. Thank you for your understanding!"). This new response is then sent to the conversation partner, and the preset platform adjusts the order processing accordingly. The event pool includes dynamic events of various types, such as logistics, orders, compensation, and discounted reissues.

[0116] It should be noted that the system generates decision results based on dynamically adjusted instructions, emotion types, and emotion levels; based on these decision results, the system dynamically adjusts responses and corresponding order processing in real time through the e-commerce platform, thus achieving intelligent service and improving the accuracy and efficiency of responses.

[0117] In one embodiment, if there is an intercepted word in the intercepted word set in the database in the target session, then based on the intercepted word and the standard response template set in the database, a first standard response template matching the intercepted word is determined from the standard response template set, and the first standard response template is determined as the response to the intercepted word. The response to the intercepted word includes switching from the intelligent customer service session to the human customer service session. If there are no intercepted words in the target conversation and the target conversation consists of small talk, then based on the small talk, standard question prediction is performed using the trained first language model to determine the candidate standard question corresponding to the small talk. Based on the candidate standard question corresponding to the small talk, the set of small talk in the database, and the set of standard response templates, a second standard response template that matches the candidate standard question corresponding to the small talk is determined from the set of standard response templates, and the second standard response template is determined as the response to the small talk.

[0118] Specifically, for example, such as Figure 3 The intercept word filtering function redirects responses to intercepted words from the intelligent customer service session to a human customer service session. It also checks the input (target session), for example, by identifying excessively long inputs (e.g., limiting the number of characters); invalid product IDs or order IDs; and by verifying login status (disallowing inquiries without being logged in), requiring users to verify their login status before accessing the online store. Response generation is handled by the intelligent customer service module, which outputs responses to intercepted words, such as "redirecting from intelligent customer service session to human customer service session."

[0119] For example, the intercepted terms in the database are shown in Table 2: Table 2: Examples of intercepted terms in the database

[0120] For example, if the target session contains the intercepted word "human" from the database's intercepted word set, then based on the intercepted word "human" and the standard response template set in the database, the first standard response template that matches the intercepted word is determined from the standard response template set: "Dear customer, for your inquiry about the product / order, you can click the button below to transfer to human customer service for processing.", and the first standard response template is determined as the response to the intercepted word.

[0121] For example, the greetings used in the database are shown in Table 3: Table 3: Examples of greetings used in the database

[0122] For example, if there are no intercepted words in the target conversation and the target conversation is a greeting like "How's the weather?", then based on this greeting, the first trained language model is used to predict the standard question, determining the candidate standard question corresponding to the greeting as "How's the weather today?". Based on this candidate standard question and the set of greetings in the database, a preset standard question matching the candidate standard question, "How's the weather today?", is determined from the set of greetings. Based on this preset standard question and the set of standard response templates, a second standard response template matching the preset standard question, "Sunny. In my eyes, if you are well, then it is a sunny day," is determined from the set of standard response templates, and this second standard response template is determined as the response to the greeting.

[0123] It should be noted that the introduction of highly efficient intelligent customer service reduces the need for human intervention, improves customer service response efficiency, and ultimately enhances the user's purchase and Q&A experience by addressing intercepted keywords or greetings.

[0124] In one embodiment, the first trained language model is obtained by training it in the following way: Retrieve a set of question samples and prompts. The question sample set includes object-related questions and greetings. The prompts include a set of questions, a set of greetings, a set of intentions, and a set of dynamic adjustment instructions; For each training round in the multi-round training, each question sample and prompt word are input into the first large language model to be trained. The first large language model to be trained performs standard intent prediction processing to obtain the candidate standard intent corresponding to each question sample. The first large language model to be trained performs standard question prediction processing to determine the candidate standard question corresponding to each question sample. The first large language model to be trained performs dynamic adjustment instruction prediction processing to obtain the dynamic adjustment instruction. Based on candidate standard questions, preset standard questions, candidate standard intentions, preset standard intentions, dynamic adjustment instructions, and preset standard dynamic adjustment instructions, the training loss of the first large language model to be trained in each round of training is determined. If the training termination condition is met, the training of the first language model to be trained is terminated, and the trained first language model is obtained. If the training termination condition is not met, the model parameters of the first language model are adjusted based on the training loss, and the adjusted first language model is trained in the next round.

[0125] Specifically, for example, if the training loss of the first language model to be trained is less than or equal to a preset loss threshold, then the training of the first language model to be trained ends, and a trained first language model is obtained; if the training loss of the first language model to be trained is greater than the preset loss threshold, then the model parameters of the first language model are adjusted based on the training loss, and the adjusted first language model is trained in the next round.

[0126] For example, obtain the existing customer service question dataset (question sample set), classify and label the question dataset as shown in Table 4; upload the question data in the question dataset to the first language model to be trained, train and fine-tune the first language model to obtain the trained first language model; publish and deploy the trained first language model.

[0127] Table 4: Classification and Labeling of Problem Dataset

[0128] It should be noted that the knowledge base is categorized into: question base, greeting base, and question-greeting base, etc.; user question categorization includes: similar questions only, greetings, and question-greetings, etc.; and whether it is product-related or order-related includes: product-related and order-related, etc. The intelligent customer service module connects the question rewriting process to the pre-trained first language model. Question rewriting is used to converge the question domain (classifying multiple questions into one question, for example, classifying 5 similar questions into the same standard question in Table 4). Since the question domain is too large, it is advisable to connect to the pre-trained first language model. By leveraging the pre-trained first language model's classification capabilities for questions, intents, and dynamic adjustment instructions, standard question prediction, standard intent prediction, and dynamic adjustment instruction prediction are performed. The computational power of the pre-trained first language model ensures the accuracy of the classification capability, thus ensuring the accuracy of the prediction results.

[0129] In one embodiment, the content dimension set of the target standard response includes at least one of general fields, tone, and specific fields, where the specific fields include relevant operations for the target object. If the content dimension set of the target standard response includes specific fields, after determining that the target standard response is a response to the target question, it further includes: Retrieve the target response from the session object for the relevant operation; Based on the target response, the operation command is predicted and processed by the trained third language model to obtain the operation command. The operation instructions are sent to the preset platform so that the preset platform can perform corresponding operations on the target object based on the operation instructions.

[0130] Specifically, for example, if the emotion type is "angry," and the corresponding emotion level 3 is "very angry," the content dimension set of the target standard response for emotion type "angry" - emotion level 3 "very angry" includes regular fields, tone, and specific fields. Specific fields include relevant actions for the target object (e.g., the order). The regular fields include "Your order is currently in {queue position} and is expected to ship at {shipping time}." The tone is mild and polite. The specific fields include relevant actions for the target object (e.g., the order): "To expedite delivery, you can choose the following courier services: 1. Standard courier (free); 2. Express courier (additional fee {express fee} yuan); Please reply with a number to select, or tell us your needs."

[0131] For example, a user responds to the message, "To expedite delivery, you can choose from the following courier services: 1. Standard courier (free); 2. Express courier (additional fee of {express fee} yuan); please reply with a number to choose from, or tell us your needs." The user's response (target response) is "Choose 2, it's best to use Company A for express delivery." This response is obtained. Based on this response, a pre-trained third language model is used to predict the operation instruction, resulting in the instruction, "Use Company A for express delivery." This instruction is then sent to a pre-set platform (e.g., an e-commerce platform). The pre-set platform, based on this instruction, performs the corresponding order processing, sending the order via Company A's express delivery.

[0132] It should be noted that the content dimension set of the target standard response corresponding to the emotion type "angry" - emotion level 3 "very angry" includes relevant operations for the target object (e.g., an order), and the conversation object (user) responds to these relevant operations. Based on the operation instructions, the target object is given corresponding order operations through a preset platform (e.g., an e-commerce platform) to more accurately meet the order needs of the conversation object, realize intelligent service, and improve the accuracy and efficiency of the conversation.

[0133] Applying the embodiments of this disclosure has at least the following beneficial effects: Leveraging the large language model's ability to classify questions, intentions, emotion types, emotion levels, and dynamic adjustment instructions, standard question prediction, standard intention prediction, emotion level prediction, and dynamic adjustment instruction prediction are performed. The computational power of the large language model ensures the accuracy of the classification capabilities, guaranteeing the accuracy of the prediction results. It precisely identifies intentions (e.g., user intentions), emotion types, and emotion levels (e.g., the user's emotional level), providing data support for personalized responses (target standard answers). Based on target object information, target standard intentions, and emotion levels, target answer rules are determined. Based on these rules, target standard questions are answered to obtain target standard answers, thus enabling personalized responses based on different emotion types, emotion levels, and target object information. (Personalized data) yields different target standard responses; it provides high recall performance for target standard responses and obtains personalized and rule-based accurate standard scripts (target standard responses), improving the accuracy of the scripts (responses) of e-commerce intelligent customer service, thereby improving the efficiency of e-commerce intelligent customer service in answering user questions (target questions), avoiding risks and customer complaints caused by incorrect responses from e-commerce intelligent customer service. The introduction of highly efficient intelligent customer service reduces the intervention of human customer service, improves customer service response efficiency, and ultimately enhances the user's purchase Q&A experience; based on dynamically adjusting instructions, emotion types, and emotion levels, decision results are generated, and based on the decision results, the target standard responses and corresponding order processing are dynamically adjusted in real time through the e-commerce platform, realizing intelligent service and improving the accuracy and efficiency of responses.

[0134] To better understand the methods provided in the embodiments of this disclosure, the solutions of the embodiments of this disclosure will be further explained below with reference to specific application scenarios.

[0135] In a specific application scenario, such as an intelligent customer service conversation scenario, see [link to example]. Figure 9 This illustrates the processing flow of an intelligent customer service conversation method, such as... Figure 9 As shown, the processing flow of the intelligent customer service conversation method provided in this embodiment includes the following steps: S701, the user enters the customer service conversation page (dialog box) and selects the question to consult.

[0136] S702, if the user selects a product, i.e., selects product consultation, then obtain the product list.

[0137] S703 retrieves product information from the product backend.

[0138] S704 If the user selects a product, the corresponding product information is saved in the session context, that is, the product is recorded in the context.

[0139] Specifically, the session context is as follows: Figure 3 CLS in the context.

[0140] S705: If you select "Inquiry about Shopping Cart Items" (i.e., select items in the shopping cart), then retrieve the shopping cart list.

[0141] S706 retrieves shopping cart product information from the shopping cart backend.

[0142] S707 If the user selects a product, the corresponding product information is saved in the session context, that is, the product is recorded in the context.

[0143] S708, if you select order consultation, i.e., order selection, you will get the order list.

[0144] S709 retrieves product information from the order backend.

[0145] S710: If the user selects an order, the corresponding order information is saved in the session context, that is, the order is recorded in the context.

[0146] S711, users enter their questions into the customer service chat page.

[0147] S712 submits the user's question and the corresponding conversation context (such as product inquiry, shopping cart product inquiry, order inquiry, etc.) selected by the user in S701-S710 to the customer service backend (the backend of the intelligent customer service).

[0148] S713: The customer service backend first checks for blocked words. If blocked words are found, the customer service will refuse to answer.

[0149] S714: The customer service backend obtains the prompt, calls the pre-trained first language model API to rewrite the question, determine the standard answer category (candidate standard question), target standard intent, and dynamic adjustment instructions; and calls the pre-trained second language model to determine the emotion type and emotion level.

[0150] S715: The first trained language model returns the standard answer classification (candidate standard question), target standard intent, and dynamic adjustment instructions to the customer service backend; the second trained language model returns the emotion type and emotion level to the customer service backend.

[0151] S716, the back-end customer service determines the target response rules based on the standard answer classification (candidate standard questions), target standard intent, emotion type, emotion level, and personalized data (target object information) through a rule engine.

[0152] S717, the customer service backend determines the target standard response template based on the target response rules.

[0153] S718, the customer service backend obtains product information from the product backend.

[0154] S719, the customer service backend retrieves order information from the order backend.

[0155] S720: The customer service backend fills in the standard response template based on product information or order information and generates a reply (target standard response).

[0156] S721, the customer service backend returns a reply (target standard reply) to the dialog box.

[0157] S722, the dialog box displays a response (target standard reply) to the user.

[0158] It should be noted that the steps for users to select a consultation partner include S701-S710, and the steps for users to ask questions (not to transfer to human customer service) include S711-S722. After S722, the customer service backend matches the target decision function from the e-commerce platform's decision function set based on the dynamic adjustment instructions, emotion type, and emotion level. The customer service backend generates a decision result through the target decision function based on the dynamic adjustment instructions, emotion type, and emotion level corresponding to the target question, and sends the decision result to the e-commerce platform. The e-commerce platform adjusts the response (target standard response) based on the decision result to obtain a new response, and sends the new response to the user.

[0159] In a specific application scenario, such as an intelligent customer service conversation scenario, see [link to example]. Figure 10 This illustrates the processing flow of an intelligent customer service conversation method, such as... Figure 10 As shown, the processing flow of the intelligent customer service conversation method provided in this embodiment includes the following steps: S801: The server trains the first language model to be trained, and obtains the trained first language model; and trains the second language model to be trained, and obtains the trained second language model.

[0160] S802, the server obtains the target session; if the target session is a target question for the target object, then execute S803; if the target session contains an intercept word from the database's intercept word set, then execute S809; if the target session does not contain an intercept word and the target session is a greeting, then execute S810.

[0161] Specifically, the target session is, for example, a question entered by a user in an e-commerce intelligent customer service session interface.

[0162] S803: Based on the acquired prompt words and target question, the server obtains the candidate standard intent, candidate standard question, and dynamic adjustment instructions corresponding to the target question through the trained first language model. Based on the target question, the server obtains the emotion type and emotion level corresponding to the emotion type through the trained second language model.

[0163] Specifically, for example, the acquired prompt words and target question are input into a pre-trained first language model. The pre-trained first language model is used for standard intent prediction to obtain candidate standard intents corresponding to the target question. The pre-trained first language model is used for standard question prediction to obtain candidate standard questions corresponding to the target question. The pre-trained first language model is used for dynamic adjustment instruction prediction to obtain dynamic adjustment instructions. Based on the target question, the pre-trained second language model is used for emotion level prediction to obtain the emotion type corresponding to the target question and the emotion level corresponding to the emotion type.

[0164] For example, the prompt and the target question (What is the material?) are input into the trained first language model. The trained first language model performs standard question prediction processing to obtain the candidate standard question (What is the material of this product?) corresponding to the target question (What is the material of this product?). Based on the predicted candidate standard questions, the target standard question (What is the material of this product?) corresponding to the target question is obtained.

[0165] S804, the server matches the predicted candidate standard problem with each preset standard problem in the standard problem set in the database, and takes the preset standard problem in the standard problem set that matches the predicted candidate standard problem as the target standard problem; and takes the candidate standard intent corresponding to the target problem as the target standard intent corresponding to the target problem.

[0166] Specifically, for example, the predicted candidate standard question "Why hasn't my order been shipped yet?" is matched with each preset standard question in the standard question set of the database, and the preset standard question "Why hasn't my order been shipped yet?" that matches the predicted candidate standard question is taken as the target standard question.

[0167] S805: Based on the target standard problem, the server determines the target rule judgment method corresponding to the target problem from the set of rule judgment methods in the database.

[0168] Specifically, for example, based on the target standard question "Why hasn't my order been shipped yet?", the target rule judgment method corresponding to the target question is determined from the set of rule judgment methods in the database; the target rule judgment method includes, for example, grass field order / member card order, pre-sale / non-pre-sale (in stock), customized / non-customized, package status, estimated delivery time, rights protection status, and whether the specific order has been sent / not sent in this round of conversation, etc.

[0169] S806, the server determines the target response rule that matches the target object information from the target response strategy based on the target rule judgment method, target object information, target standard intent, emotion type and emotion level.

[0170] S807, the server determines the target standard response template that matches the target standard question based on the target response rules and the set of standard response templates in the database.

[0171] S808, the server obtains the target standard answer to the target standard question based on the object information of the target object and the target standard answer template; then proceeds to S812 for execution.

[0172] S809, the server, based on the intercepted word and the set of standard response templates in the database, determines the first standard response template that matches the intercepted word from the set of standard response templates, and determines the first standard response template as the response to the intercepted word; proceed to S812 for execution.

[0173] Specifically, for example, if the target session contains the intercepted word "human" from the database's intercepted word set, then based on the intercepted word "human" and the standard response template set in the database, a first standard response template matching the intercepted word is determined from the standard response template set: "Dear customer, for your inquiry about the product / order, you can click the button below to transfer to human customer service for processing.", and the first standard response template is determined as the response to the intercepted word.

[0174] The S810 server, based on greetings, uses a pre-trained first language model to perform standard question prediction processing to determine the candidate standard questions corresponding to the greetings.

[0175] Specifically, for example, if there are no intercepted words in the target conversation and the target conversation is a greeting such as "How's the weather?", then based on this greeting, the standard question prediction process is performed using the trained first language model to determine the candidate standard question corresponding to the greeting as "How's the weather today?".

[0176] S811, the server, based on the candidate standard questions corresponding to the greetings, the set of greetings in the database, and the set of standard response templates, determines the second standard response template that matches the candidate standard questions corresponding to the greetings from the set of standard response templates, and determines the second standard response template as the response to the greetings.

[0177] Specifically, for example, if there are no intercepted words in the target conversation and the target conversation is a greeting such as "How's the weather?", then based on this greeting, a pre-trained large language model is used to perform standard question prediction processing to determine the candidate standard question corresponding to the greeting as "How's the weather today?"; based on this candidate standard question and the set of greetings in the database, a preset standard question matching the candidate standard question, "How's the weather today?", is determined from the set of greetings; based on this preset standard question and the set of standard response templates, a second standard response template matching the preset standard question, "Sunny. In my eyes, if you are well, then it is a sunny day," is determined from the set of standard response templates, and this second standard response template is determined as the response to the greeting.

[0178] S812, the server sends the target standard response, the response to the intercepted words, or the response to the greeting to the terminal.

[0179] S813, the terminal's intelligent customer service conversation interface displays the target standard response, the response to the intercepted words, or the response to the greeting.

[0180] Specifically, for example, such as Figure 3 As shown, the front-end conversation interface is the terminal's intelligent customer service conversation interface; the target standard reply is "Dear customer, we found that your order was shipped on February 15, 2021. You can check the logistics information through the order details." This target standard reply is sent to the terminal, and the terminal's intelligent customer service conversation interface displays the target standard reply.

[0181] The S814 server determines the decision outcome based on dynamic adjustment instructions, emotion type, and emotion level, and then sends the decision outcome to the e-commerce platform.

[0182] Specifically, for example, based on the dynamic adjustment instructions, emotion type, and emotion level, a target decision function is matched from the decision function set of the e-commerce platform; based on the dynamic adjustment instructions, emotion type, and emotion level corresponding to the target question, a decision result is generated through the target decision function, and the decision result is sent to the e-commerce platform.

[0183] S815, the e-commerce platform adjusts the target standard response based on the decision results, obtains a new response, and sends the new response to the terminal.

[0184] Applying the embodiments of this disclosure has at least the following beneficial effects: The combination of emotion recognition and large language models: Large language model technology accurately identifies user emotions, providing data support for personalized responses; Dynamic adjustment and automation: Based on dynamic adjustment instructions and emotion levels, the system dynamically adjusts target standard responses and corresponding order processing in real time through e-commerce platforms, achieving intelligent service and improving the accuracy and efficiency of responses; Multi-module collaboration: Emotion recognition, response generation, and dynamic adjustment work together to form a complete intelligent customer service solution; Leveraging the large language model's classification capabilities for questions, intentions, emotion types, emotion levels, and dynamic adjustment instructions, the system performs standard question prediction, standard intention prediction, emotion level prediction, and dynamic adjustment instruction prediction. The computational power of the large language model ensures the accuracy of classification capabilities, guaranteeing accurate prediction results and precise intent identification. For example, user intent, emotion type, and emotion level (e.g., the emotion level of a user's emotion) provide data support for personalized responses (target standard responses). Based on target object information, target standard intent, emotion type, and emotion level, target response rules are determined. Based on these rules, target standard questions are answered to obtain target standard responses. Thus, different target standard responses are obtained based on different emotion types, different emotion levels, and different target object information (personalized data). This provides both high recall performance for target standard responses and personalized and rule-based accurate standard scripts (target standard responses), improving the accuracy of the scripts (responses) of e-commerce intelligent customer service, thereby improving the efficiency of e-commerce intelligent customer service in answering user questions (target questions).

[0185] This disclosure also provides an intelligent customer service chat device, the structural schematic diagram of which is shown below. Figure 11 As shown, the intelligent customer service conversation device 90 includes a first processing module 901, a second processing module 902, a third processing module 903, a fourth processing module 904, and a fifth processing module 905.

[0186] The first processing module 901 is used to obtain the target session of the session object; The second processing module 902 is used to, if the target conversation is a target question for the target object, perform standard question prediction and standard intent prediction based on the target question using a trained first language model, obtain the target standard question and target standard intent corresponding to the target question based on the prediction results, and perform emotion level prediction based on the target question using a trained second language model to obtain the emotion level corresponding to the target question. The emotion level corresponding to the target question is used to represent the emotional degree of the conversation object, and the target standard intent corresponding to the target question is used to represent the intent of the conversation object. The third processing module 903 is used to obtain the target response strategy from the response strategy set corresponding to the target standard question. The response strategy set includes the response strategy corresponding to each standard question among multiple standard questions. The response strategy corresponding to each standard question includes at least one response rule. Each response rule is associated with at least one object information, a standard intention, and an emotion level. The fourth processing module 904 is used to obtain target object information, and based on the target object information, target standard intent and emotion level, determine the target response rules that match the target object information from the target response strategy; The fifth processing module 905 is used to respond to the target standard question based on the target response rules, obtain the target standard response to the target standard question, and determine the target standard response as the response to the target question.

[0187] In one embodiment, the second processing module 902 is specifically used for: Obtain prompt words. Prompt words are used to instruct the trained first language model to determine the candidate standard questions corresponding to the input question from the question set and to select the candidate standard intents corresponding to the input question from the intent set. The prompt words include the question set and the intent set. The question set includes multiple pre-set candidate standard questions, and the intent set includes multiple candidate standard intents. The prompt words and target question are input into the trained first language model. The trained first language model is used to perform standard intent prediction processing to obtain the candidate standard intent corresponding to the target question. The trained first language model is used to perform standard question prediction processing to obtain the candidate standard question corresponding to the target question. Based on the predicted candidate standard questions, the target standard question corresponding to the target question is obtained, and based on the predicted candidate standard intent, the target standard intent corresponding to the target question is obtained.

[0188] In one embodiment, the second processing module 902 is specifically used for: The predicted candidate standard problems are matched with each preset standard problem in the standard problem set of the database, and the preset standard problems in the standard problem set that match the predicted candidate standard problems are taken as the target standard problems.

[0189] In one embodiment, the second processing module 902 is specifically used for: The target question is input into the trained second language model, and word segmentation is performed to obtain the corresponding word segmentation for the target question. Based on the word segmentation corresponding to the target question, the word segmentation encoding representation corresponding to the target question is obtained through encoding processing; Based on the word segmentation encoding representation corresponding to the target question, the emotion level corresponding to the target question is obtained through classification processing.

[0190] In one embodiment, the fourth processing module 904 is further configured to: Based on the target standard problem, the target rule judgment method corresponding to the target standard problem is determined from the set of rule judgment methods in the database; The fourth processing module 904 is specifically used for: Based on the target rule judgment method, target object information, target standard intent and emotion level, target response rules that match the target object information are determined from the target response strategies.

[0191] In one embodiment, the fifth processing module 905 is specifically used for: Based on the target response rules and the standard response template set in the database, a target standard response template that matches the target standard question is determined from the standard response template set. The content dimension set, emotion type and emotion level of the target standard response template are correlated. The content dimension set includes at least one of the following: regular fields, tone and specific fields. The specific fields include relevant operations for the target object. Based on the object information of the target object and the target standard response template, the target standard response to the target standard question is obtained.

[0192] In one embodiment, where the target object is an order or a product, the fifth processing module 905 is further configured to: The target standard response is sent to the terminal so that the target standard response to the target question for the order or product is displayed in the terminal's intelligent customer service conversation interface.

[0193] In one embodiment, the second processing module 902 is further configured to: Based on the target question, the first trained language model is used to perform dynamic adjustment instruction prediction processing to obtain dynamic adjustment instructions. These dynamic adjustment instructions are used to adjust the target standard response on the preset platform. Based on the dynamic adjustment instructions, emotion type, and the corresponding emotion level, the target decision function is matched from the set of decision functions of the preset platform. Based on the dynamic adjustment instructions, emotion type, and the corresponding emotion level, a decision result is generated through the target decision function and sent to the preset platform. The preset platform then adjusts the target standard response based on the decision result to obtain a new response, which is then sent to the conversation object.

[0194] In one embodiment, the second processing module 902 is further configured to: If the target session contains a intercepted word from the database's intercepted word set, then based on the intercepted word and the standard response template set in the database, a first standard response template matching the intercepted word is determined from the standard response template set, and the first standard response template is determined as the response to the intercepted word. The response to the intercepted word includes transferring from the intelligent customer service session to a human customer service session. If there are no intercepted words in the target conversation and the target conversation consists of small talk, then based on the small talk, standard question prediction is performed using the trained first language model to determine the candidate standard question corresponding to the small talk. Based on the candidate standard question corresponding to the small talk, the set of small talk in the database, and the set of standard response templates, a second standard response template that matches the standard question corresponding to the small talk is determined from the set of standard response templates, and the second standard response template is determined as the response to the small talk.

[0195] In one embodiment, the first trained language model is obtained by training it in the following way: Retrieve a set of question samples and prompts. The question sample set includes object-related questions and greetings. The prompts include a set of questions, a set of greetings, a set of intentions, and a set of dynamic adjustment instructions; For each training round in the multi-round training, each question sample and prompt word are input into the first large language model to be trained. The first large language model to be trained performs standard intent prediction processing to obtain the candidate standard intent corresponding to each question sample. The first large language model to be trained performs standard question prediction processing to determine the candidate standard question corresponding to each question sample. The first large language model to be trained performs dynamic adjustment instruction prediction processing to obtain the dynamic adjustment instruction. Based on candidate standard questions, preset standard questions, candidate standard intentions, preset standard intentions, dynamic adjustment instructions, and preset standard dynamic adjustment instructions, the training loss of the first large language model to be trained in each round of training is determined. If the training termination condition is met, the training of the first language model to be trained is terminated, and the trained first language model is obtained. If the training termination condition is not met, the model parameters of the first language model are adjusted based on the training loss, and the adjusted first language model is trained in the next round.

[0196] In one embodiment, the content dimension set of the target standard response includes at least one of general fields, tone, and specific fields, where the specific fields include relevant operations for the target object. If the content dimension set of the target standard response includes specific fields, the fifth processing module 905 is further configured to: Retrieve the target response from the session object for the relevant operation; Based on the target response, the operation command is predicted and processed by the trained third language model to obtain the operation command. The operation instructions are sent to the preset platform so that the preset platform can perform corresponding operations on the target object based on the operation instructions.

[0197] Applying the embodiments of this disclosure has at least the following beneficial effects: The process involves: 1) Obtaining the target session of the conversation object; if the target session is a target question for the target object, then using a pre-trained first language model to perform standard question prediction and standard intent prediction based on the target question. Based on the prediction results, the target standard question and target standard intent corresponding to the target question are obtained. Then, using a pre-trained second language model, emotion level prediction is performed based on the target question to obtain the emotion type and emotion level corresponding to the emotion type. The emotion level of the target question represents the emotional intensity of the conversation object, and the target standard intent represents the intent of the conversation object. 2) Obtaining target response strategies from the response strategy set corresponding to the target standard question. The response strategy set includes response strategies corresponding to each standard question in multiple standard questions. Each response strategy for a standard question includes at least one response rule, and each response rule is associated with at least one object information, one standard intent, one emotion type, and one emotion level corresponding to the emotion type. 3) Obtaining the target object information of the target object. Based on the target object information, target standard intent, emotion type, and emotion level, the target response rule matching the target object information is determined from the target response strategies. 4) Responding to the target standard question based on the target response rule to obtain the target standard response to the target standard question, and defining the target standard response as... The response addresses the target question. Leveraging the large language model's ability to classify questions, intentions, emotion types, and emotion levels, standard question prediction, standard intention prediction, and emotion level prediction are performed. The computational power of the large language model ensures the accuracy of the classification ability, guaranteeing the accuracy of the prediction results. This accurately identifies intentions (e.g., user intentions), emotion types, and emotion levels (e.g., the user's emotional level), providing data support for personalized responses (target standard responses). Based on target object information, target standard intentions, emotion types, and emotion levels, target response rules are determined. Based on these rules, the target standard question is answered, resulting in the target standard response. The system provides targeted standard answers to quasi-questions, resulting in different target standard answers based on different emotion types, emotion levels, and target audience information (personalized data). This provides both high recall performance for targeted standard answers and personalized, rule-based, accurate standard scripts (target standard answers), improving the accuracy of the scripts (responses) of e-commerce intelligent customer service. This, in turn, improves the efficiency of e-commerce intelligent customer service in answering user questions (target questions), avoiding risks and customer complaints caused by incorrect answers. The introduction of highly efficient intelligent customer service reduces the need for human customer service intervention, improves customer service response efficiency, and ultimately enhances the user's purchase Q&A experience.

[0198] This disclosure also provides an electronic device, the structural schematic diagram of which is shown below. Figure 12 As shown, Figure 12The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of this disclosure.

[0199] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with this disclosure. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0200] Bus 4002 may include a pathway for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 12 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0201] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium capable of carrying or storing computer programs and capable of being read by a computer, without limitation herein.

[0202] The memory 4003 is used to store computer programs that execute embodiments of the present disclosure, and is controlled by the processor 4001 to execute them. The processor 4001 is used to execute the computer programs stored in the memory 4003 to implement the steps shown in the foregoing method embodiments.

[0203] Electronic devices include, but are not limited to, servers.

[0204] Applying the embodiments of this disclosure has at least the following beneficial effects: The process involves: 1) Obtaining the target session of the conversation object; if the target session is a target question for the target object, then using a pre-trained first language model to perform standard question prediction and standard intent prediction based on the target question. Based on the prediction results, the target standard question and target standard intent corresponding to the target question are obtained. Then, using a pre-trained second language model, emotion level prediction is performed based on the target question to obtain the emotion type and emotion level corresponding to the emotion type. The emotion level of the target question represents the emotional intensity of the conversation object, and the target standard intent represents the intent of the conversation object. 2) Obtaining target response strategies from the response strategy set corresponding to the target standard question. The response strategy set includes response strategies corresponding to each standard question in multiple standard questions. Each response strategy for a standard question includes at least one response rule, and each response rule is associated with at least one object information, one standard intent, one emotion type, and one emotion level corresponding to the emotion type. 3) Obtaining the target object information of the target object. Based on the target object information, target standard intent, emotion type, and emotion level, the target response rule matching the target object information is determined from the target response strategies. 4) Responding to the target standard question based on the target response rule to obtain the target standard response to the target standard question, and defining the target standard response as... The response addresses the target question. Leveraging the large language model's ability to classify questions, intentions, emotion types, and emotion levels, standard question prediction, standard intention prediction, and emotion level prediction are performed. The computational power of the large language model ensures the accuracy of the classification ability, guaranteeing the accuracy of the prediction results. This accurately identifies intentions (e.g., user intentions), emotion types, and emotion levels (e.g., the user's emotional level), providing data support for personalized responses (target standard responses). Based on target object information, target standard intentions, emotion types, and emotion levels, target response rules are determined. Based on these rules, the target standard question is answered, resulting in the target standard response. The system provides targeted standard answers to quasi-questions, resulting in different target standard answers based on different emotion types, emotion levels, and target audience information (personalized data). This provides both high recall performance for targeted standard answers and personalized, rule-based, accurate standard scripts (target standard answers), improving the accuracy of the scripts (responses) of e-commerce intelligent customer service. This, in turn, improves the efficiency of e-commerce intelligent customer service in answering user questions (target questions), avoiding risks and customer complaints caused by incorrect answers. The introduction of highly efficient intelligent customer service reduces the need for human customer service intervention, improves customer service response efficiency, and ultimately enhances the user's purchase Q&A experience.

[0205] This disclosure provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the steps and corresponding content of the aforementioned method embodiments.

[0206] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, can implement the steps and corresponding content of the aforementioned method embodiments.

[0207] It should be understood that although arrows indicate various operation steps in the flowcharts of the embodiments of this disclosure, the order in which these steps are implemented is not limited to the order indicated by the arrows. Unless explicitly stated herein, in some implementation scenarios of the embodiments of this disclosure, the implementation steps in each flowchart can be executed in other orders as required. Furthermore, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages based on the actual implementation scenario. Some or all of these sub-steps or stages can be executed at the same time, and each sub-step or stage can also be executed at different times. In scenarios where execution times differ, the execution order of these sub-steps or stages can be flexibly configured as required, and the embodiments of this disclosure do not limit this.

[0208] The above description is only an optional implementation method for some implementation scenarios of this disclosure. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this disclosure without departing from the technical concept of this disclosure also fall within the protection scope of the embodiments of this disclosure.

Claims

1. A method for intelligent customer service conversations, characterized in that, include: Get the target session of the session object; If the target conversation is a target question for a target object, then based on the target question, standard question prediction and standard intent prediction are performed using a trained first language model. Based on the prediction results, the target standard question and the target standard intent corresponding to the target question are obtained. Then, based on the target question, emotion level prediction is performed using a trained second language model to obtain the emotion type and the emotion level corresponding to the emotion type. The emotion level corresponding to the target question is used to characterize the conversational emotion level of the conversational object, and the target standard intent corresponding to the target question is used to characterize the conversational intent of the conversational object. Obtain the target response strategy from the set of response strategies corresponding to the target standard question. The set of response strategies includes the response strategy corresponding to each standard question among multiple standard questions. The response strategy corresponding to each standard question includes at least one response rule. Each response rule is associated with at least one object information, a standard intention, an emotion type, and the emotion level corresponding to the emotion type. Obtain target object information of the target object; based on the target object information, the target standard intent, the emotion type, and the emotion level, determine the target response rule that matches the target object information from the target response strategy. Based on the target response rules, the target standard question is answered to obtain the target standard response to the target standard question, and the target standard response is determined as the response to the target question.

2. The method according to claim 1, characterized in that, Based on the target question, the standard question prediction and standard intent prediction processes are performed using a trained first language model. Based on the prediction results, the target standard question and the target standard intent corresponding to the target question are obtained, including: Obtain prompt words, which are used to instruct the trained first language model to determine the candidate standard questions corresponding to the input question from the question set and to select the candidate standard intents corresponding to the input question from the intent set. The prompt words include the question set and the intent set. The question set includes multiple preset candidate standard questions, and the intent set includes multiple candidate standard intents. The prompt word and the target question are input into the trained first language model. The trained first language model is used to perform standard intent prediction processing to obtain the candidate standard intent corresponding to the target question. The trained first language model is used to perform standard question prediction processing to obtain the candidate standard question corresponding to the target question. Based on the predicted candidate standard problems, the target standard problem corresponding to the target problem is obtained, and based on the predicted candidate standard intent, the target standard intent corresponding to the target problem is obtained.

3. The method according to claim 2, characterized in that, The process of obtaining the target standard problem corresponding to the target problem based on the predicted candidate standard problems includes: The predicted candidate standard problem is matched with each preset standard problem in the standard problem set in the database, and the preset standard problem in the standard problem set that matches the predicted candidate standard problem is taken as the target standard problem.

4. The method according to claim 1, characterized in that, Based on the target question, the process involves using a trained second language model to predict the emotion level, thereby obtaining the emotion type corresponding to the target question and the emotion level corresponding to the emotion type, including: The target question is input into the trained second language model, and word segmentation is performed to obtain the word segmentation corresponding to the target question. Based on the word segmentation corresponding to the target question, the word segmentation encoding representation corresponding to the target question is obtained through encoding processing; Based on the word segmentation encoding representation corresponding to the target question, the emotion type corresponding to the target question and the emotion level corresponding to the emotion type are obtained through classification processing.

5. The method according to claim 1, characterized in that, The method further includes: Based on the target standard problem, the target rule judgment method corresponding to the target standard problem is determined from the set of rule judgment methods in the database; The step of determining target response rules that match the target object information from the target response strategy based on the target object information, the target standard intent, the emotion type, and the emotion level includes: Based on the target rule judgment method, the target object information, the target standard intent, the emotion type, and the emotion level, a target response rule matching the target object information is determined from the target response strategy.

6. The method according to claim 1, characterized in that, The step of responding to the target standard question based on the target response rule to obtain the target standard response to the target standard question includes: Based on the target response rules and the set of standard response templates in the database, a target standard response template that matches the target standard question is determined from the set of standard response templates. The content dimension set of the target standard response template, the emotion type, and the emotion level are correlated. The content dimension set includes at least one of regular fields, tone, and specific fields. The specific fields include relevant operations for the target object. Based on the object information of the target object and the target standard response template, the target standard response to the target standard question is obtained.

7. The method according to claim 1, characterized in that, The target object is an order or a product. After determining the target standard response as a response to the target question, the method further includes: The target standard response is sent to the terminal so that the target standard response to the target question regarding the order or the product is displayed in the terminal's intelligent customer service conversation interface.

8. The method according to claim 1, characterized in that, The method further includes: Based on the target question, a dynamic adjustment instruction prediction process is performed using a trained first language model to obtain a dynamic adjustment instruction. The dynamic adjustment instruction is used to adjust the target standard response on the preset platform. Based on the dynamic adjustment instruction, the emotion type, and the emotion level corresponding to the emotion type, a target decision function is matched from the decision function set of the preset platform; Based on the dynamic adjustment instruction, the emotion type, and the emotion level corresponding to the emotion type, a decision result is generated through the target decision function, and the decision result is sent to the preset platform so that the preset platform can adjust the target standard response based on the decision result to obtain a new response, and send the new response to the conversation object.

9. The method according to claim 1, characterized in that, Also includes: If the target session contains an intercepted word from the intercepted word set in the database, then based on the intercepted word and the standard response template set in the database, a first standard response template matching the intercepted word is determined from the standard response template set, and the first standard response template is determined as the response to the intercepted word. The response to the intercepted word includes transferring from the intelligent customer service session to the human customer service session. If the intercepted word does not exist in the target session and the target session is a greeting, then based on the greeting, the standard question prediction process is performed through the trained first language model to determine the candidate standard question corresponding to the greeting. Based on the candidate standard questions corresponding to the greetings, the set of greetings in the database, and the set of standard response templates, a second standard response template that matches the standard question corresponding to the greeting is determined from the set of standard response templates, and the second standard response template is determined as the response to the greeting.

10. The method according to claim 1, characterized in that, The first trained language model was obtained through the following method: Obtain a set of question samples and prompts. The question sample set includes object-related questions and greetings. The prompts include a set of questions, a set of greetings, a set of intentions, and a set of dynamic adjustment instructions. For each training round in the multi-round training, each question sample and the prompt word are input into the first large language model to be trained. The first large language model to be trained performs standard intent prediction processing to obtain the candidate standard intent corresponding to each question sample. The first large language model to be trained performs standard question prediction processing to determine the candidate standard question corresponding to each question sample. The first large language model to be trained performs dynamic adjustment instruction prediction processing to obtain the dynamic adjustment instruction. Based on the candidate standard problem, the preset standard problem, the candidate standard intent, the preset standard intent, the dynamic adjustment instruction, and the preset standard dynamic adjustment instruction, the training loss of the first large language model to be trained in each round of training is determined; If the training termination condition is met, the training of the first language model to be trained is terminated, and the trained first language model is obtained. If the training termination condition is not met, the model parameters of the first language model are adjusted based on the training loss, and the adjusted first language model is trained in the next round.

11. The method according to claim 1, characterized in that, The set of content dimensions for the target standard response includes at least one of general fields, tone, and specific fields. The specific fields include relevant operations for the target object. If the set of content dimensions for the target standard response includes specific fields, after determining the target standard response as a response to the target question, the method further includes: Obtain the target response from the session object for the relevant operation; Based on the target response, the operation command is predicted and processed by the trained third language model to obtain the operation command. The operation instructions are sent to a preset platform so that the preset platform performs corresponding operations on the target object based on the operation instructions.

12. An intelligent customer service conversation device, characterized in that, include: The first processing module is used to obtain the target session of the session object; The second processing module is configured to, if the target conversation is a target question for a target object, perform standard question prediction and standard intent prediction based on the target question using a trained first language model, obtain the target standard question and the target standard intent corresponding to the target question based on the prediction results, and perform emotion level prediction based on the target question using a trained second language model to obtain the emotion type and the emotion level corresponding to the emotion type. The emotion level corresponding to the target question is used to characterize the emotional degree of the conversation object, and the target standard intent corresponding to the target question is used to characterize the intent of the conversation object. The third processing module is used to obtain the target response strategy from the response strategy set corresponding to the target standard question. The response strategy set includes the response strategy corresponding to each standard question among multiple standard questions. The response strategy corresponding to each standard question includes at least one response rule. Each response rule is associated with at least one object information, a standard intention, an emotion type, and the emotion level corresponding to the emotion type. The fourth processing module is used to obtain target object information of the target object, and based on the target object information, the target standard intent, the emotion type and the emotion level, determine the target response rule that matches the target object information from the target response strategy; The fifth processing module is used to answer the target standard question based on the target answer rule, obtain the target standard answer to the target standard question, and determine the target standard answer as the answer to the target question.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-11.

Images