Robot customer service problem intelligent processing method and system

Abstract

The invention provides a robot customer service question intelligent processing method and system. The method comprises the steps that user questions are acquired, and feature extraction and integration are conducted on the user questions; constructing a user dynamic causal reasoning model, and generating a multi-level problem solving strategy according to user problems; constructing a self-adaptive response strategy generation model according to the generated multi-level problem solving strategy and the integrated feature vector to generate a response strategy most adaptive to the current corresponding context; executing the response strategy and judging whether to perform learning triggering according to the feedback of the user; performing real-time adaptive learning and optimization on the adaptive response strategy generation model through user feedback and historical user feedback; and analyzing the long-term trend of the historical user feedback and the model parameters of the adaptive response strategy generation model after multiple times of optimization, performing global optimization in combination with the long-term trend, and generating a final response strategy. The method and the device can be excellently expressed in a complex user service scene.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and in particular relates to a method and system for intelligently processing robot customer service problems. Background Art

[0002] Robotic customer service systems have become a key tool in the interaction between modern enterprises and users. These systems handle a large number of user requests in an automated way, provide real-time services, and reduce operating costs. Current robotic customer service systems mainly rely on predefined rules, knowledge bases, or machine learning models trained with large amounts of labeled data. These systems perform well when processing simple, structured queries, such as checking account balances, order status, or operating instructions. However, as user needs become more diverse and complex, existing robotic customer service systems face many challenges and limitations.

[0003] First, existing systems lack the ability to deeply understand complex problems. Many user questions are not just simple information queries, but comprehensive problems involving multiple factors. For example, a problem a user may encounter is caused by the combined effect of multiple independent factors. Existing systems find it difficult to parse the root causes of these complex problems and can often only provide superficial and one-sided answers, which leads to a decline in user experience and satisfaction. Second, existing systems are not adaptable enough to new problems. Traditional robot customer service systems rely on static rules or models, which need to be frequently updated and maintained to cope with changing user needs and problem types. This manual update and maintenance method is time-consuming and laborious, and cannot respond to the emergence of new problems in a timely manner, causing the system to perform poorly when faced with unseen problems. In addition, many robot customer service systems also have obvious deficiencies in handling user emotions and contextual understanding. Emotion recognition and contextual understanding are important factors in improving user experience, but most current systems lack accurate recognition of user emotions and effective use of conversation context, resulting in stiff responses and lack of human touch.

[0004] Problems with existing technologies include insufficient use of user feedback and interaction data. Although many systems can collect a large amount of interaction data, how to effectively use this data for system optimization and self-improvement remains a difficult problem. Most systems cannot achieve real-time learning and optimization based on user feedback, and therefore cannot continuously improve their problem handling capabilities and service levels. Especially in the face of complex and changing user needs and market environments, such systems that lack dynamic adaptability find it difficult to provide high-quality service experiences. Summary of the invention

[0005] The purpose of the present invention is to design a method and system for intelligently processing robot customer service problems, achieve in-depth understanding of complex problems through causal reasoning and multi-level analysis, use biological neural network simulation and recursive neural network to improve context understanding and decision-making capabilities, improve the rapid adaptability of the system through few-sample learning, and achieve self-optimization capabilities in combination with reinforcement learning. The above innovations are organically combined to overcome the technical defects of existing robot customer service systems, allowing the present invention to perform well in complex user service scenarios.

[0006] In order to achieve the above object, a first aspect of the present invention provides a method for intelligently processing robot customer service problems, the method comprising the following steps:

[0007] S1. Obtain user questions and extract and integrate features of user questions, including:

[0008] S101, obtaining the user's complete input dialogue, and preprocessing the dialogue, converting the preprocessed dialogue into text representation, embedding each word in the text as a semantic feature, and designing a sentiment embedding model to extract sentiment features from the input dialogue, and then designing a context feature extractor to capture the contextual relationship of the user in multiple rounds of dialogue to obtain context features;

[0009] S102, generating a complexity score according to the semantic features, the emotional features and the context features, and generating a comprehensive feature vector by combining the semantic features, the emotional features, the context features and the complexity score;

[0010] S2. Build a user dynamic causal reasoning model based on the integrated feature vectors and generate a multi-level problem-solving strategy based on user questions, including:

[0011] S201, build a dynamic causal reasoning model, take the comprehensive feature vector as input, infer the causal dependency behind the user's question, and for each feature, establish a causal function by analyzing the dependency between contextual features and semantic features;

[0012] S202, classifying user problems into three levels according to the complexity of the user problems, including: surface feature level, potential cause level and root cause level;

[0013] S203, generating three problem-solving strategies for the three levels respectively, including: a surface level strategy for problems at the surface feature level, a potential level strategy for problems at the potential cause level, and a root level strategy for problems at the root cause level;

[0014] S204. If a strategy fails to effectively solve the problem, the causal relationship between features is re-evaluated through real-time user feedback signals, and the causal dependency weights in the causal function are dynamically adjusted;

[0015] S3. Based on the generated multi-level problem-solving strategies and the integrated feature vectors, an adaptive response strategy generation model is constructed to generate the response strategy that best suits the current corresponding context, including:

[0016] S301, constructing an adaptive answer strategy generation model to generate answer strategies at different levels according to the level of user questions and their corresponding comprehensive feature vectors;

[0017] S302, adding regular terms based on sentiment characteristics and question complexity to constrain the generation of the answering strategy;

[0018] S303, generating an initial response strategy, and dynamically adjusting the response strategy according to the user's real-time feedback to obtain an adjusted response strategy;

[0019] S304, merging the multi-level response strategies generated by the user to generate a final response strategy, and updating the model parameters of the adaptive response strategy generation model in real time according to the feedback from the interaction with the user;

[0020] S4. Execute the acceptance strategy and determine whether to trigger learning based on user feedback, including:

[0021] Execute the final response strategy and listen to the user's feedback, extract the feedback signal from the user's feedback, and determine whether to trigger learning based on the feedback signal:

[0022] If the feedback signal shows that the response effect is poor, the learning module is triggered to update the response model;

[0023] If the feedback signal indicates that the response is good, learning is not triggered;

[0024] S5. Real-time adaptive learning and optimization of the adaptive response strategy generation model through user feedback and historical user feedback;

[0025] S6. Analyze the long-term trends of the model parameters of the historical user feedback and the adaptive response strategy generation model after multiple optimizations, perform global optimization based on the long-term trends, optimize the response strategy that best suits the current corresponding context, and generate the final response strategy.

[0026] Furthermore, each word in the text is embedded as a vector to introduce context-related feature representation, and the semantic information of each word is captured by using the context information of the entire conversation; the emotion embedding model maps the conversation content to the emotion feature space, captures the user's emotional expression in a specific sentence, and combines the conversation context to obtain the dynamic process of emotion changing over time; the context feature extractor is used to capture the contextual relationship of users in multiple rounds of conversations.

[0027] Furthermore, the dynamic causal reasoning model is expressed as follows:

[0028] Define each feature in the causal graph as a node v i , the edges between all nodes have dependency weight w i Indicates causal relationships between features;

[0029] Through the analysis of historical data, the causal relationship model is derived, which is expressed as follows:

[0030]

[0031] in, represents the loss function of the causal relationship model, Pa(v i ) represents the feature v i The set of causal parent nodes, f i represents the causal function of the estimated feature, i represents the number of features;

[0032] The surface symptom layer is used to process semantic features and emotional features, and directly reflects the problem currently expressed by the user;

[0033] The potential cause layer is used to handle complex problems, identifying the potential causes behind the surface problems based on contextual features and problem complexity and through cause-effect diagrams;

[0034] The root cause layer is used to combine sentiment features, context features, and historical data of multiple rounds of dialogue to trace the root cause of the problem;

[0035] The surface level strategy directly generates standardized responses to questions at the surface feature layer;

[0036] The potential hierarchy strategy generates a dialogue strategy for guiding the user to describe the problem in the next step for the problem at the potential cause layer;

[0037] The root-level strategy generates complex response plans for root-cause level questions based on sentiment fluctuations and contextual information in the long-term conversation history.

[0038] Further, the S301 specifically includes: for the surface symptom layer, the response strategy is based on semantic features and emotional features; for the potential cause layer and the root cause layer, a detailed response is generated based on semantic features and emotional features and combined with context features and complexity scores;

[0039] At the same time, hierarchical adaptive coefficients are introduced into the surface symptom layer, potential cause layer and root cause layer to control the complexity of the response strategy at different problem levels. The largest hierarchical adaptive coefficient indicates that the problem is the most complex.

[0040] Furthermore, the user's real-time feedback signal includes emotional fluctuations and context changes; the multi-level response strategies generated by the user are merged by weighted summation.

[0041] Furthermore, the feedback signal includes emotional changes and semantic matching, wherein the emotional changes represent emotional fluctuations from the current dialogue turn to the previous dialogue turn, and the semantic matching represents the semantic matching between the robot's answer and the user's question.

[0042] Furthermore, when the learning trigger is activated, the adaptive response strategy generation model is updated according to the feedback signal, and the adaptive response strategy generation model is updated using a gradient descent method to optimize the adaptive response strategy generation model by minimizing the loss function.

[0043] Furthermore, the S5 specifically includes:

[0044] S501. Design semantic and sentiment-based regularization terms and construct loss functions based on semantic and sentiment features (F,R final ), which is expressed as follows:

[0045]

[0046] Among them, F represents the real-time feedback signal, R final represents the final response strategy, α and β represent the weights of balancing the feedback signal and the response strategy deviation; F optimal Represents the preset ideal feedback signal; R optimal represents the ideal response strategy; γ represents the weight coefficient of the regularization term based on semantics and sentiment; Represents a regularization term based on semantics and sentiment, which is used to ensure that the adaptive answer strategy generation model balances semantic and sentiment factors during adjustment, as shown below:

[0047]

[0048] Among them, e represents the emotional feature, X represents the semantic feature, and e optimal represents the ideal emotional characteristics, X optimal represents the ideal semantic feature, λ1 and λ2 represent the weights for balancing sentiment and semantic bias;

[0049] S502, introducing an adaptive learning rate to dynamically adjust according to the strength and complexity of the feedback signal, to ensure that when the feedback signal is the largest, the adaptive response strategy generation model can be adjusted as quickly as possible; when the feedback signal is the weakest, the adaptive response strategy generation model can be updated; the adaptive learning rate η is expressed as follows:

[0050]

[0051] Among them, η0 represents the initial learning rate; κ represents the adjustment coefficient; q represents the complexity of the problem;

[0052] S503, constructing a historical feedback accumulation item, which is used to combine feedback signals from multiple rounds of dialogues to guide the adaptive response strategy generation model to perform long-term optimization;

[0053] S504: after the adaptive response strategy generation model is updated, a dynamic verification mechanism is designed to verify whether the adaptive response strategy generation model is updated, as shown below:

[0054]

[0055] if If it is negative, it means that the model optimization is successful; otherwise, adjustments will be made to ensure that the model gradually converges to the ideal response effect.

[0056] Furthermore, the S6 specifically includes:

[0057] S601, constructing a long-term feedback trend function according to the problem classification features, analyzing the long-term change trend of the feedback signal, and finding the long-term deviation point and optimization space; wherein the problem classification features are extracted from the long-term feedback signal trend of the answering strategy of the multi-level problem;

[0058] S602, globally optimizing the model parameter θ of the adaptive response strategy generation model according to the long-term change trend of the feedback signal;

[0059] S603, introducing a strategy evolution factor in the long-term optimization process of the adaptive response strategy generation model, and dynamically adjusting the rules of the adaptive response strategy generation model according to the long-term feedback signal;

[0060] S604. An adaptive architecture expansion mechanism is designed to automatically expand different model modules or update the functions of existing modules according to the volatility of feedback signals, the complexity of conversations, and the diversity of user needs;

[0061] S605. After the model is updated and the strategy is optimized, the effect of long-term optimization is verified through testing.

[0062] In a second aspect of the present invention, a robot customer service problem intelligent processing system is provided, the system comprising:

[0063] The user question acquisition subsystem is used to obtain user questions and extract and integrate the features of user questions, including:

[0064] S101, obtaining the user's complete input dialogue, and preprocessing the dialogue, converting the preprocessed dialogue into text representation, embedding each word in the text as a semantic feature, and designing a sentiment embedding model to extract sentiment features from the input dialogue, and then designing a context feature extractor to capture the contextual relationship of the user in multiple rounds of dialogue to obtain context features;

[0065] S102, generating a complexity score according to the semantic features, the emotional features and the context features, and generating a comprehensive feature vector by combining the semantic features, the emotional features, the context features and the complexity score;

[0066] The user problem analysis subsystem is used to build a user dynamic causal reasoning model based on the integrated feature vectors and generate a multi-level problem-solving strategy based on the user problems, including:

[0067] S201, build a dynamic causal reasoning model, take the comprehensive feature vector as input, infer the causal dependency behind the user's question, and for each feature, establish a causal function by analyzing the dependency between contextual features and semantic features;

[0068] S202, classifying user problems into three levels according to the complexity of the user problems, including: surface feature level, potential cause level and root cause level;

[0069] S203, generating three problem-solving strategies for the three levels respectively, including: a surface level strategy for problems at the surface feature level, a potential level strategy for problems at the potential cause level, and a root level strategy for problems at the root cause level;

[0070] S204. If a strategy fails to effectively solve the problem, the causal relationship between features is re-evaluated through real-time user feedback signals, and the causal dependency weights in the causal function are dynamically adjusted;

[0071] The user question answering subsystem is used to construct an adaptive answering strategy generation model based on the generated multi-level problem solving strategies and the integrated feature vectors to generate the answering strategy that best suits the current corresponding context, including:

[0072] S301, constructing an adaptive answer strategy generation model to generate answer strategies at different levels according to the level of user questions and their corresponding comprehensive feature vectors;

[0073] S302, adding regular terms based on sentiment characteristics and question complexity to constrain the generation of the answering strategy;

[0074] S303, generating an initial response strategy, and dynamically adjusting the response strategy according to the user's real-time feedback to obtain an adjusted response strategy;

[0075] S304, merging the multi-level response strategies generated by the user to generate a final response strategy, and updating the model parameters of the adaptive response strategy generation model in real time according to the feedback from the interaction with the user;

[0076] The response learning subsystem is used to execute the response strategy and determine whether to trigger learning based on user feedback, including:

[0077] Execute the final response strategy and listen to the user's feedback, extract the feedback signal from the user's feedback, and determine whether to trigger learning based on the feedback signal:

[0078] If the feedback signal shows that the response effect is poor, the learning module is triggered to update the response model;

[0079] If the feedback signal indicates that the response is good, learning is not triggered;

[0080] The optimization subsystem is used to perform real-time adaptive learning and optimization of the adaptive response strategy generation model through user feedback and historical user feedback, analyze the long-term trends of the model parameters of the adaptive response strategy generation model after historical user feedback and multiple optimizations, perform global optimization based on the long-term trends, optimize the response strategy that best suits the current corresponding context, and generate the final response strategy.

[0081] The beneficial technical effects of the present invention are at least as follows:

[0082] The present invention proposes an adaptive robot customer service system based on multi-level intelligent analysis, aiming to solve the above-mentioned problems in the prior art and improve the performance of the robot customer service system in complex scenarios. First of all, the present invention adopts the technology of causal reasoning and multi-level problem analysis, breaking through the limitation that the existing system cannot deeply understand complex problems. Through the causal reasoning model, the system can analyze the root cause of user problems, rather than just dealing with surface symptoms. This innovation enables the system to provide more accurate and effective solutions, thereby improving user satisfaction. Through the multi-level problem analysis framework, the system can distinguish the different levels of the problem, including surface symptoms, potential causes and root causes, making the solution more targeted.

[0083] Secondly, the present invention introduces biological neural network simulation and recursive neural network technology to enhance the system's decision-making ability and context understanding ability. By imitating the information processing method of biological neural networks, the system can better understand the user's intentions and problem background when facing complex decision-making scenarios. The introduction of recursive neural networks allows the system to capture the dynamic evolution of problems and changes in user intentions when dealing with long-term conversations, thereby providing more coherent and more humane responses during the conversation.

[0084] In addition, the present invention improves the system's adaptability to new problems through few-sample learning technology. Traditional robot customer service systems require a large amount of labeled data for training, while the present invention uses few-sample learning to enable the system to quickly learn new types of problems and solutions with very little data. This rapid adaptability not only greatly reduces the frequency of manual intervention and system updates, but also enables the system to respond to users' new needs and problems more promptly.

[0085] Finally, the adaptive learning module of the present invention combines user feedback and reinforcement learning mechanisms, and has the ability to continuously optimize itself. The system can dynamically adjust its strategies and model parameters based on real-time feedback from users, and perform self-calibration and optimization. Through this ability to continuously learn and improve, the system can continuously improve its service level and maintain its competitiveness in a complex and changing market environment.

[0086] In summary, the main innovations of the present invention include: achieving in-depth understanding of complex problems through causal reasoning and multi-level analysis, using biological neural network simulation and recursive neural network to improve context understanding and decision-making capabilities, improving the rapid adaptability of the system through few-sample learning, and combining reinforcement learning to achieve self-optimization capabilities. The above innovations are organically combined to overcome the technical defects of existing robot customer service systems, allowing the present invention to perform well in complex user service scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0088] Figure 1 The present invention is a flowchart of a method for intelligently processing robot customer service problems according to an embodiment of the present invention.

[0089] Figure 2 This is a framework diagram of a robot customer service problem intelligent processing system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0090] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0091] In one or more embodiments, Figure 1 As shown, a method for intelligently processing robot customer service problems of the present invention is disclosed, and the method includes steps 1 to 6, including:

[0092] S1. Obtain user questions and extract and integrate features of user questions.

[0093] In intelligent customer service systems, the diversity and complexity of user input data requires targeted preprocessing steps. In particular, in order to accurately understand user needs, the system must not only process the input text, but also capture the sentiment, contextual information, and complexity of the question. The key to this step is to extract multi-dimensional features from the natural language data input by the user, laying the foundation for subsequent causal reasoning and problem-solving strategy generation.

[0094] Specifically, first, when a user interacts with the customer service system through text or voice, the input data T = {t1, t2, ..., t n} represents the user's complete input dialog.

[0095] For speech input, it will first be converted into text through the speech recognition model, and then the text needs to be processed to ensure that it can be understood and used by subsequent models. Text preprocessing includes common steps such as removing stop words and standardizing vocabulary morphology, but more importantly, generating contextual embedding representations of vocabulary, that is, capturing the semantic information of each word through a context-aware embedding model. Unlike traditional methods, the present invention introduces context-related feature representations here, which can make full use of the context of the entire conversation. Each word t i is embedded as a vector x i , which is generated by relying on the contextual information in the conversation rather than embedding words in isolation. The embedding method is defined by the following formula:

[0096] x i =ContextEmbedding(t i ,T)

[0097] ContextEmbedding here is not a simple word vector generation, but a semantic representation of words based on the entire conversation history. This context-aware embedding method ensures that the system not only understands the meaning of a single sentence, but also accurately grasps the core information in a longer conversation, thereby providing sufficient semantic basis for subsequent analysis.

[0098] Furthermore, in addition to semantic understanding, the user's emotional state plays an important role in the customer service system. Traditional systems only rely on the semantic structure of sentences and tend to ignore the user's emotions, which often affects the system's response strategy. Therefore, in the text preprocessing stage, the present invention designs an emotional embedding model to extract emotional features from the input dialogue.

[0099] The present invention uses a specially trained emotion recognition module to map the conversation content to the emotion vector space. Emotion embedding not only captures the user's emotional expression in a specific sentence, but also combines the conversation context and considers the dynamic process of emotion changing over time. The emotion feature is represented by e, which is expressed by the following formula:

[0100] e=EmotionEmbedding(T)

[0101] Unlike general emotion classification models, this invention incorporates subtle changes in user emotions, especially emotional fluctuations in multiple rounds of conversations, into emotion recognition. This allows the system to not only know whether the user is expressing confusion or anger, but also capture the trend of emotional changes, helping the system predict the user's next emotional response. This has a direct impact on the subsequent generation of response strategies, because the system can adjust its own strategies more flexibly to ensure an appropriate response to the user's emotional state.

[0102] Furthermore, in customer service scenarios, a single conversation sentence often cannot independently express the user's problem. Context understanding is a key step to ensure that the system can give accurate responses. Therefore, based on the recurrent neural network (RNN) or Transformer architecture, the present invention designs a context feature extractor specifically for capturing the contextual relationship of users in multi-round conversations.

[0103] The context feature extractor takes the entire conversation history as input and outputs a vector c representing the current conversation state. This vector not only takes into account the current sentence, but also the previous conversation history, ensuring that the system can make comprehensive use of all conversation information to make decisions.

[0104] c=ContextModel(T)

[0105] In this way, the system can understand the continuity of the user's questions. Even if the user's expression is not clear, the system can still infer the complete intention through the context. Traditional methods usually only process single sentence input and ignore the contextual relationship of multiple rounds of dialogue. The context feature extraction of the present invention ensures that the system can handle complex dialogues more intelligently.

[0106] Furthermore, in customer service systems, different questions have different complexities. Some simple questions can be solved with standardized responses, while complex questions require more detailed explanations and processing. Therefore, the complexity assessment of the question becomes a key part of the preprocessing stage.

[0107] The present invention designs a question complexity assessment module, which generates a complexity score q by analyzing multiple dimensions such as dialogue length, syntactic complexity, keyword frequency, etc. This complexity score not only considers the surface features of the dialogue, but also combines the context vector c to ensure its accuracy.

[0108] q = ComplexityScore(T,X,c)

[0109] This evaluation method allows the system to distinguish questions of different complexity and provide key references for generating response strategies in subsequent steps. If the question is more complex, the system may need to introduce multi-step dialogues to gradually guide users to find solutions; while simple questions can be quickly solved through direct responses.

[0110] Furthermore, finally, the present invention integrates all the extracted features to generate a comprehensive feature vector f. The feature vector combines the semantic feature X, the emotional feature e, the context feature c and the problem complexity q, ensuring that the subsequent causal reasoning and strategy generation can obtain sufficient information input.

[0111] f=[X,e,c,q]

[0112] This feature vector is not just a simple accumulation of semantic information, it also takes into account multiple dimensions of information such as sentiment, context, and complexity. It incorporates every aspect of user input and provides a complete perspective for the system's subsequent decision-making.

[0113] Through the preprocessing of the above steps, the present invention not only performs standard text processing on the user input, but also innovatively combines sentiment analysis, context understanding and question complexity assessment to provide a comprehensive multi-dimensional feature vector for subsequent steps. This feature vector can fully reflect the user's intention and question difficulty, ensuring that the system can perform more accurate causal reasoning and response strategy generation in subsequent steps.

[0114] S2. Build a user dynamic causal reasoning model based on the integrated feature vector to generate a multi-level problem-solving strategy based on user questions.

[0115] In the customer service system, handling complex user problems requires going beyond the traditional simple question-and-answer model, especially when the problem has multi-level causes, the system must be able to deeply understand the logic behind the user input. This process not only relies on surface semantics, but also requires causal reasoning to identify hidden causal relationships and gradually solve problems through multi-level problem analysis. The present invention will use the multidimensional features f = [X, e, c, q] (semantic feature X, emotional feature e, contextual feature c, and problem complexity q) extracted from step 1 to perform causal reasoning, identify the relationship between features, and analyze the problem in layers to generate a solution strategy.

[0116] Specifically, the feature vector f obtained from step 1 is used as input to construct a causal relationship graph The present invention can infer the causal dependencies behind user questions. For each user question, there is a potential causal relationship between features. For example, the question complexity q may be jointly affected by the context information c and the semantic feature X. To this end, the present invention establishes a dynamic causal reasoning model, which determines the dependency between features by minimizing the residual.

[0117] The present invention defines each feature in the causal graph as a node v i , the edges w between these nodes i Represents the causal relationship between features. Through the analysis of historical data, the following causal relationship model is derived:

[0118]

[0119] Among them, Pa(v i ) represents the feature v i The set of causal parent nodes, and f i To estimate the causal function of the feature. For each feature, such as the problem complexity q, the present invention establishes a causal function f by analyzing the dependency of the context feature c and the semantic feature X. q :

[0120] q≈f q (X,c)

[0121] This causal reasoning not only reveals the dynamic dependencies between features, but also helps the system identify the potential causes of complex user problems by interpreting the features. In this way, the system can update the causal relationships between features in real time and adjust the response strategies for different user problems.

[0122] Furthermore, after constructing the cause-effect relationship diagram, the system needs to analyze the user's problem in a hierarchical manner. This is a key innovation in the patent solution, because user problems are usually not isolated, but have different levels of causes. The present invention divides the problem into three levels: surface symptoms, potential causes, and root causes.

[0123] Surface symptom layer: This layer mainly processes semantic features X and sentiment features e, which can directly reflect the problem currently expressed by the user. For example, when a user asks "I can't log in to my account", this is a surface symptom, and the system can generate a direct response based on semantic features and existing knowledge base.

[0124] Potential cause layer: For complex problems, the system identifies the potential causes behind the surface problems through the cause-effect diagram. Contextual features c and problem complexity q play a key role in this layer. For example, if the user has mentioned other account-related problems in previous conversations, this may be a potential cause, and the system can further confirm these potential problems by asking the user.

[0125] Root cause layer: The last layer is to trace the root cause of the problem. This requires the integration of emotional features e, context c, and historical data from multiple rounds of conversations. For long-standing unresolved problems, the user's emotional changes are often one of the root causes. The system can infer the deep cause of the problem by gradually analyzing the emotional changes in the conversation.

[0126] Through this hierarchical analysis, the system can deal with problems layer by layer from the surface to the root, avoiding the limitations of simple problem-solving systems that only deal with surface symptoms and ignore the underlying causes.

[0127] Furthermore, after identifying the different levels of the problem, the system needs to generate corresponding response strategies for each level. This part is closely related to the causal reasoning model, because the system's response strategy must match the identified causal dependencies and problem levels.

[0128] Surface level strategy: For surface symptoms, the system can directly generate standardized responses. At this time, the generation of responses mainly depends on semantic features X, for example:

[0129] Response1 = f r (X)

[0130] The f in this formula r Represents a standard semantic reply function, which is mainly used to generate quick responses based on semantic features.

[0131] Latent hierarchical strategy: For potential reasons, the system generates a dialogue strategy to guide the user to further describe the problem. At this time, context features c and problem complexity q are used as key parameters for strategy generation. For example:

[0132] Response2 = f r (X,c,q)

[0133] This response strategy guides the user through multiple rounds of dialogue, allowing the user to gradually describe more details of the potential problem.

[0134] Root-level strategy: For the root cause, the system needs to generate more complex responses based on the emotional fluctuations and contextual information in the long-term conversation history. At this time, the system can combine the emotional changes of multiple rounds of conversations to generate responses tailored to the user's emotions to ensure that the user feels understood and supported.

[0135] Furthermore, after generating a response strategy, the system will dynamically adjust the weights in the causal relationship graph based on user feedback. Specifically, if a strategy fails to effectively solve the problem, the system will re-evaluate the causal relationship between features through user feedback signals and dynamically adjust the causal dependency weights w i The update formula is:

[0136]

[0137] Through this feedback-driven adjustment, the system is able to continuously optimize the causal reasoning model, thereby making more accurate inferences and responses when dealing with similar problems in the future.

[0138] S3. Based on the generated multi-level problem-solving strategy and the integrated feature vector, an adaptive response strategy generation model is constructed to generate the response strategy that best suits the current corresponding context.

[0139] In step 2, the system passes through the cause-effect diagram The user questions are analyzed hierarchically to generate questions at various levels, including surface symptoms, potential causes, and root causes. These outputs (hierarchical questions L, semantic features X, emotional features e, contextual features c, and question complexity q) will be used as input to generate adaptive response strategies. The system needs to generate flexible response strategies based on the level, emotional state, and contextual features of the questions, and dynamically adjust the response strategies during the actual interaction process.

[0140] Specifically, first, according to the hierarchical question L and its characteristics (semantics X, context c, problem complexity q) output in step 2, the present invention defines an adaptive answer strategy generation function f s This function combines these inputs to generate different levels of response strategies R i .

[0141] The generation of response strategies depends on the different levels of the problem. For the surface symptom layer, the response strategy is mainly based on semantic features X and emotional features e; for the potential cause layer and the root cause layer, the complexity of the strategy will increase, and it is necessary to combine the context c and the problem complexity q to generate a more detailed response.

[0142] The present invention introduces the hierarchical adaptive coefficient λ i , which controls the complexity of the answering strategy at different question levels. The generated answering strategy is expressed as:

[0143] R i =f s (λ i X,c,q)

[0144] Among them, R i represents the response strategy generated at level i, λ i is the adaptive complexity coefficient of the i-th level, which depends on the problem level L i . i The larger it is, the more complex the problem is and the more context and historical information needs to be considered in strategy generation.

[0145] For example, for the root cause layer, λi It may take a larger value to ensure that the answer can cover multiple aspects of the user's problem; for the surface symptom layer, λ i The smaller the value, the simpler the response strategy will be.

[0146] Furthermore, in order to avoid the answer strategy being too complex or too simple, the present invention introduces a regularization term Ω(R) when generating the answer. The regularization term constrains the generation of the answer strategy according to the user's emotional characteristics e and the complexity of the question q, ensuring that the strategy can adapt to the user's current emotional state and the complexity of the question. The improved answer strategy generation formula is:

[0147]

[0148] Among them, Ω(R) is the regularization term, defined as:

[0149] Ω(R)=α·∥Re∥ 2 +β·∥Rq∥ 2

[0150] Among them, α and β control the matching degree of the response strategy with the emotional feature e and the complexity of the question q. Through this regularization term, the system can make dynamic adjustments according to the user's current emotional state and the complexity of the question when generating a response, avoiding the generation of complex responses for simple questions or the generation of brief responses for complex questions.

[0151] Furthermore, after the system generates the initial response strategy, it will dynamically adjust the response strategy according to the user's real-time feedback F. The feedback signal F includes the emotional fluctuation Δe and the context change Δc, which are used to modify the initial response strategy.

[0152] The adjusted response strategy formula is:

[0153]

[0154] Among them, f a (F) is the feedback adjustment function, which includes the immediate emotion change Δe and the context change Δc. Driven by feedback, the system can adjust the response strategy according to the user's emotional fluctuations to avoid generating blunt or inappropriate responses.

[0155] Furthermore, user questions usually involve multiple levels (surface symptoms, potential causes, and root causes), so the system needs to generate a response strategy R for each level. i The present invention combines the response strategies at each level into the final response by weighted summation:

[0156]

[0157] Among them, α iIt is a weight determined by the complexity of the question q. In this way, the system can comprehensively consider problems at all levels in the final answer to ensure that the answer strategy is comprehensive and accurate.

[0158] Furthermore, in order to improve the long-term response capability of the system, the system will record user feedback data after each user interaction and update the response strategy generation model based on these historical feedbacks. s and the adjustment function f a The optimization continuously improves the response strategy through the gradient descent method. The optimization formula is:

[0159]

[0160] Among them, γ is the learning rate, Represents the loss gradient based on historical feedback. Through this mechanism, the system can gradually accumulate experience in dealing with complex problems and optimize the quality of future responses.

[0161] S4. Execute the acceptance strategy and determine whether to trigger learning based on user feedback.

[0162] Specifically, in step 3, the final response strategy R has been generated fi na l , this strategy is generated through multi-level strategy merging and optimization, combining the user's semantic features X, emotional features e, context features c and question complexity q. At this stage, the system will execute the generated answer strategy R final and interact with users through actual conversations.

[0163] The system will respond according to the response strategy R final Send the corresponding reply and monitor the user's subsequent reaction to ensure that each step of the response matches the user's question and emotional state. The system monitors the response effect by processing the user's reply and reaction in real time. This execution process is based on the previous adaptive response strategy generation model (see step 3) to ensure that each response is context-aware.

[0164] For example, user feedback after execution f feedback The feedback will be monitored, including the user's response time, emotional fluctuations and problem solving. By systematically analyzing the feedback, the effectiveness of the current response strategy can be judged.

[0165] Furthermore, the system needs to extract feedback signals from the user's real-time response to provide input for subsequent learning and optimization. The core variables of feedback include emotional change Δe and semantic matching degree ΔX. The present invention defines a feedback signal function f feedback , whose output is the feedback signal F. The specific definition of the feedback signal F is as follows:

[0166] F=f feedback (Δe,ΔX)

[0167] Among them, Δe is the change of user emotional characteristics, indicating the emotional fluctuation from the current dialogue round to the previous dialogue round; ΔX is the change of semantic matching, indicating the semantic matching between the system response and the user question. The system quantifies these signals through natural language processing technology and provides them to the learning trigger module for further processing.

[0168] For example, Δe can be quantified by analyzing the user's tone, vocabulary, and voice features; ΔX can be quantified by using a text similarity algorithm (such as cosine similarity) to determine the semantic deviation between the system response and the user's question.

[0169] Furthermore, the system needs to decide whether to trigger the learning module to further optimize the model according to the real-time feedback signal F. The present invention defines a learning trigger function T(F), the output of which determines whether to perform learning.

[0170] The core mechanism of learning triggering depends on the strength of the feedback signal. If the feedback signal shows that the response effect is not good (for example, the emotional fluctuation is significant or the semantic matching is low), the learning module is triggered to update the response model; if the feedback signal shows that the response effect is good, learning is not triggered. The trigger condition formula is:

[0171]

[0172] Among them, θ is the threshold for triggering learning, which depends on the design of the system. When F exceeds the threshold, it means that the response is biased and learning needs to be triggered; when it is lower than the threshold, it means that the response is relatively successful and no further learning needs to be triggered.

[0173] Furthermore, when the learning trigger is activated, the system will generate a model f for the response strategy based on the feedback signal F s The model is updated using the gradient descent method, which optimizes the response strategy generation model by minimizing the loss function. Defined as:

[0174]

[0175] Among them, F optimal is the ideal feedback signal set by the system, R optimal is the ideal response strategy, α and β are weight parameters that control the relative importance of the feedback signal and the response strategy in the optimization. By minimizing the loss, the system can continuously adjust the generated response strategy R, thereby improving the response quality of subsequent conversations.

[0176] Through learning triggers and model updates, the system can gradually improve its response capabilities in different conversation rounds. This learning mechanism enables the system to adapt to the needs of different users and diverse conversation scenarios, and provide more intelligent services.

[0177] S5. Perform real-time adaptive learning and optimization on the adaptive response strategy generation model through user feedback and historical user feedback.

[0178] In step 4, the system executes the generated response strategy R final And collect real-time feedback signal F. This feedback signal quantifies the user's emotional change Δe and semantic matching degree ΔX, representing the degree of adaptation of the current response strategy to user needs.

[0179] This step will use F as the core input to trigger the adaptive learning mechanism by adjusting the response model f s parameters, so that it can provide more accurate and personalized responses in the next interaction. At the same time, the system will also integrate historical feedback from multiple rounds of dialogue to optimize the long-term performance of the model. Based on the feedback signal, the core task of this step is to minimize the loss function To update the parameters θ of the answer strategy generation model. The loss function not only considers the current feedback, but also combines the balance between emotional and semantic feedback.

[0180] Specifically, first, the present invention defines an innovative loss function A regularization term based on semantics and sentiment is added to enable the model to consider the balance between semantic and sentiment features when updating. The loss function is as follows:

[0181]

[0182] Among them, α and β are the weights of balancing feedback signals and response strategy deviations; F optimal It is the ideal feedback signal preset by the system; Ro p tim a l is the ideal response strategy; the regularization term It is an innovative design targeting user sentiment and semantic features, ensuring that the model balances semantic and sentiment factors when adjusting. This regularization term is defined as:

[0183]

[0184] The innovation of this regularization term is that it forces the system to not only pursue the matching of the response strategy and the feedback signal during the optimization process, but also maintain the consistency of the user's emotions (e) and semantics (X). This design is very suitable for the customer service field, because the change of user emotions often has a significant impact on the success of the response strategy.

[0185] In order to ensure that the model update process is more efficient and dynamic, the present invention introduces an adaptive learning rate η. This rate is dynamically adjusted according to the strength and complexity of the feedback signal F, so as to ensure that when the feedback signal is large (indicating that the user is dissatisfied with the current response strategy), the model can be adjusted quickly; when the feedback signal is weak (indicating that the response strategy is basically successful), the model is updated in a more stable manner. The innovative formula of the adaptive learning rate η is designed as:

[0186]

[0187] Among them, η0 is the initial learning rate; κ is the adjustment coefficient; q is the problem complexity. The purpose of adding ∥q∥ is to further slow down the learning rate and avoid over-adjustment when the problem complexity is high.

[0188] When the feedback signal F is strong (indicating that the response strategy is not ideal), the learning rate will increase; when the problem complexity q is high, the system tends to make more cautious adjustments to avoid over-reliance on single feedback for adjustments in complex scenarios.

[0189] In the customer service system, a single feedback cannot fully represent the long-term needs of users. Therefore, the present invention designs a historical feedback accumulation term H, which is used to combine the feedback signals of multiple rounds of dialogue to guide the model to perform long-term optimization. By accumulating historical feedback, the model can better cope with emotional fluctuations and semantic changes in complex scenarios. The calculation formula of the historical feedback accumulation term is:

[0190] H new =ρH old +(1-ρ)F

[0191] Among them, H new is the new historical feedback accumulation value; H old is the accumulated historical feedback; ρ is the feedback decay factor, which controls the weight of old feedback and new feedback. Through this design, the system can combine the new feedback signal with the past historical feedback after each interaction, gradually accumulate the understanding of different user emotions and semantic needs, and optimize the long-term response strategy generation model f s .

[0192] After the model is updated, the system will re-verify the model's performance through feedback signals. If the updated model performs well, the feedback signal F will be significantly close to the ideal feedback F. optimal Otherwise, the system will readjust the learning rate η and the regularization term The weight parameters λ1 and λ2.

[0193] This dynamic verification mechanism makes the model update process more flexible and avoids the situation where too many parameters are adjusted at one time, which leads to unstable model performance. The present invention measures the update effect through the following formula:

[0194]

[0195] if If it is negative, it means that the model optimization is successful; otherwise, the system will make further adjustments to ensure that the model gradually converges to the ideal response effect.

[0196] S6. Analyze the long-term trends of the model parameters of the historical user feedback and the adaptive response strategy generation model after multiple optimizations, perform global optimization based on the long-term trends, optimize the response strategy that best suits the current corresponding context, and generate the final response strategy.

[0197] In step 5, the system has optimized the response strategy generation model f through real-time adaptive learning and historical feedback accumulation mechanism. s At this point, the system has accumulated a large amount of real-time feedback F and historical feedback H, and the model has been updated many times. However, long-term optimization not only needs to rely on the feedback of a single conversation, but also requires system-level upgrades and long-term strategy optimization to adapt to future complex conversation scenarios and diverse user needs.

[0198] The task of this step is to further optimize the response strategy and gradually update the entire system based on the data accumulated in the long-term interaction of the system to ensure efficient and intelligent services in the long run.

[0199] Specifically, the main input of this step is the accumulated historical feedback H and the model parameters θ after multiple optimizations. new , these parameters come from the optimization results in step 5. The system needs to analyze the long-term trend of these historical data to find structural problems or potential optimization space in the response strategy. Through the long-term feedback trend function T(H), the long-term change trend of the feedback signal is analyzed to find the long-term deviation points and optimization space. The trend function is defined as:

[0200]

[0201] Among them, H i Represents the historical feedback in the i-th conversation, and n is the number of cumulative conversations. By calculating the average feedback signal, the system can identify potential problems in long-term responses, such as insufficient responses to certain question types or deviations in emotional processing.

[0202] In this process, the present invention introduces the problem classification feature P. The system analyzes the long-term feedback according to the feedback signals of different problem categories P to ensure the targeted optimization. For example, for technical support problems P tech , the system can focus on optimizing strategies related to the problem complexity q; while for emotional support problems P emotional , the system pays more attention to the matching degree of emotional feature e.

[0203] Furthermore, the core of long-term optimization is to continuously improve the overall response capability of the system through the accumulation of a large amount of dialogue data. To achieve this goal, the system needs to gradually update the core parameters of the model and further improve the system performance by generating new strategies. In combination with the long-term feedback trend, the present invention performs global optimization on the model parameter θ. This optimization minimizes the long-term loss function accomplish:

[0204]

[0205] in, It is the ideal historical feedback signal in each round of interaction; is the ideal strategy in each round of response; α and β represent the importance weights of historical feedback and response strategy respectively.

[0206] Through this optimization, the system can continuously converge to the ideal response strategy based on long-term trends, ensuring that even in complex and changing scenarios, the system still maintains high response accuracy and flexibility.

[0207] Furthermore, long-term optimization is not only about tuning the existing model parameters, but the system also needs to "evolve" the response strategy itself according to the long-term trend. This evolution is reflected in the structural update of the response strategy and the improvement of the generation rules. To this end, the present invention introduces a strategy evolution factor μ, which dynamically adjusts the rules of the strategy generation model based on long-term feedback signals. The strategy evolution formula is:

[0208] R evolved =R new +μ·T(H)

[0209] Among them, R new is the response strategy updated through long-term feedback, T(H) is the result of the feedback trend function, and μ is the strategy evolution factor, which is used to control the rate of strategy evolution. Through this evolutionary process, the system can gradually adapt to different types of questions and emotional needs, and make more creative and adaptive improvements in the response strategy.

[0210] Furthermore, in the long-term operation, as the amount of data increases and the scenarios become more complex, the system also needs to be optimized and expanded at the architecture level. This is not just about adjusting the parameters of the model, but also about optimizing the architecture of the entire response system to adapt to higher complexity. An adaptive architecture expansion mechanism is designed. According to the volatility of feedback signals, the complexity of conversations, and the diversity of user needs, the system will automatically expand different model modules or update the functions of existing modules. The trigger conditions for adaptive architecture expansion are:

[0211]

[0212] Among them, τ q and τ F are the thresholds for problem complexity and feedback bias, respectively. When the complexity or feedback bias exceeds the threshold, the system will trigger an architecture upgrade. For example, the system may add more deep learning model modules for emotion recognition, or introduce more complex causal reasoning modules to handle highly complex problems.

[0213] Furthermore, after the model is updated and the strategy evolves, the system needs to pass a series of tests to verify the effect of long-term optimization. System-level verification simulates multiple rounds of dialogue scenarios and conducts automated testing in combination with historical data. The main indicators of verification include:

[0214] Response accuracy: how well the system's response matches the ideal answer;

[0215] Emotion matching: the processing effect of the emotional feature e in the system response;

[0216] Ability to handle complex conversations: The system's ability to cope with high-complexity scenarios, mainly measured by ∥q∥.

[0217] The system evaluates the optimization effect of the model through dynamic testing of these indicators, and further adjusts the strategy and model parameters based on the test results to ensure that the system maintains good performance in the long run.

[0218] Step S6 aims to comprehensively optimize and upgrade the response strategy and system architecture through the accumulation and trend analysis of long-term feedback signals. Combined with the strategy evolution factor μ and the adaptive architecture expansion mechanism, the system can continuously improve its intelligence level and response capabilities while coping with complex scenarios. At the same time, the system ensures the effectiveness of continuous optimization through rigorous verification tests after each optimization, providing an efficient and stable intelligent customer service solution for long-term use in patent scenarios.

[0219] In one or more embodiments, Figure 2As shown, the present invention discloses a robot customer service problem intelligent processing system, the system comprising:

[0220] The user question acquisition subsystem 101 is used to acquire user questions and extract and integrate features of the user questions, specifically including:

[0221] S101, obtaining the user's complete input dialogue, and preprocessing the dialogue, converting the preprocessed dialogue into text representation, embedding each word in the text as a semantic feature, and designing a sentiment embedding model to extract sentiment features from the input dialogue, and then designing a context feature extractor to capture the contextual relationship of the user in multiple rounds of dialogue to obtain context features;

[0222] S102, generating a complexity score according to the semantic features, the emotional features and the context features, and generating a comprehensive feature vector by combining the semantic features, the emotional features, the context features and the complexity score;

[0223] The user problem analysis subsystem 102 is used to construct a user dynamic causal reasoning model based on the integrated feature vectors and generate a multi-level problem solving strategy based on the user problem, specifically including:

[0224] S201, build a dynamic causal reasoning model, take the comprehensive feature vector as input, infer the causal dependency behind the user's question, and for each feature, establish a causal function by analyzing the dependency between contextual features and semantic features;

[0225] S202, classifying user problems into three levels according to the complexity of the user problems, including: surface feature level, potential cause level and root cause level;

[0226] S203, generating three problem-solving strategies for the three levels respectively, including: a surface level strategy for problems at the surface feature level, a potential level strategy for problems at the potential cause level, and a root level strategy for problems at the root cause level;

[0227] S204. If a strategy fails to effectively solve the problem, the causal relationship between features is re-evaluated through real-time user feedback signals, and the causal dependency weights in the causal function are dynamically adjusted;

[0228] The user question answering subsystem 103 is used to construct an adaptive answering strategy generation model based on the generated multi-level problem solving strategies and the integrated feature vectors to generate the answering strategy that best suits the current corresponding context, including:

[0229] S301, constructing an adaptive answer strategy generation model to generate answer strategies at different levels according to the level of user questions and their corresponding comprehensive feature vectors;

[0230] S302, adding regular terms based on sentiment characteristics and question complexity to constrain the generation of the answering strategy;

[0231] S303, generating an initial response strategy, and dynamically adjusting the response strategy according to the user's real-time feedback to obtain an adjusted response strategy;

[0232] S304, merging the multi-level response strategies generated by the user to generate a final response strategy, and updating the model parameters of the adaptive response strategy generation model in real time according to the feedback from the interaction with the user;

[0233] The response learning subsystem 104 is used to execute the response strategy and determine whether to trigger learning based on user feedback, including:

[0234] Execute the final response strategy and listen to the user's feedback, extract the feedback signal from the user's feedback, and determine whether to trigger learning based on the feedback signal:

[0235] If the feedback signal shows that the response effect is poor, the learning module is triggered to update the response model;

[0236] If the feedback signal indicates that the response is good, learning is not triggered;

[0237] The optimization subsystem 105 is used to perform real-time adaptive learning and optimization of the adaptive response strategy generation model through user feedback and historical user feedback, analyze the long-term trends of the model parameters of the adaptive response strategy generation model after historical user feedback and multiple optimizations, perform global optimization based on the long-term trends, optimize the response strategy that best suits the current corresponding context, and generate the final response strategy.

[0238] These are just some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and making equivalent changes according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A method for intelligently processing robot customer service problems, characterized in that: The method comprises the following steps: S1. Obtain user questions and extract and integrate features of user questions, including: S101, obtaining the user's complete input dialogue, and preprocessing the dialogue, converting the preprocessed dialogue into text representation, embedding each word in the text as a semantic feature, and designing a sentiment embedding model to extract sentiment features from the input dialogue, and then designing a context feature extractor to capture the contextual relationship of the user in multiple rounds of dialogue to obtain context features; S102, generating a complexity score according to the semantic features, the emotional features and the context features, and generating a comprehensive feature vector by combining the semantic features, the emotional features, the context features and the complexity score; S2. Build a user dynamic causal reasoning model based on the integrated feature vectors and generate a multi-level problem-solving strategy based on user questions, including: S201, build a dynamic causal reasoning model, take the comprehensive feature vector as input, infer the causal dependency behind the user's question, and for each feature, establish a causal function by analyzing the dependency between contextual features and semantic features; S202, classifying user problems into three levels according to the complexity of the user problems, including: surface feature level, potential cause level and root cause level; S203, generating three problem-solving strategies for the three levels respectively, including: a surface level strategy for problems at the surface feature level, a potential level strategy for problems at the potential cause level, and a root level strategy for problems at the root cause level; S204. If a strategy fails to effectively solve the problem, the causal relationship between features is re-evaluated through real-time user feedback signals, and the causal dependency weights in the causal function are dynamically adjusted; S3. Based on the generated multi-level problem-solving strategies and the integrated feature vectors, an adaptive response strategy generation model is constructed to generate the response strategy that best suits the current corresponding context, including: S301, constructing an adaptive answer strategy generation model to generate answer strategies at different levels according to the level of user questions and their corresponding comprehensive feature vectors; S302, adding regular terms based on sentiment characteristics and question complexity to constrain the generation of the answering strategy; S303, generating an initial response strategy, and dynamically adjusting the response strategy according to the user's real-time feedback to obtain an adjusted response strategy; S304, merging the multi-level response strategies generated by the user to generate a final response strategy, and updating the model parameters of the adaptive response strategy generation model in real time according to the feedback from the interaction with the user; S4. Execute the acceptance strategy and determine whether to trigger learning based on user feedback, including: Execute the final response strategy and listen to the user's feedback, extract the feedback signal from the user's feedback, and determine whether to trigger learning based on the feedback signal: If the feedback signal shows that the response effect is poor, the learning module is triggered to update the response model; If the feedback signal indicates that the response is good, learning is not triggered; S5. Real-time adaptive learning and optimization of the adaptive response strategy generation model through user feedback and historical user feedback; S6. Analyze the long-term trends of the model parameters of the historical user feedback and the adaptive response strategy generation model after multiple optimizations, perform global optimization based on the long-term trends, optimize the response strategy that best suits the current corresponding context, and generate the final response strategy.

2. According to claim 1, a method for intelligently processing robot customer service problems is characterized in that: Each word in the text is embedded as a vector to introduce context-related feature representation, and the semantic information of each word is captured by using the context information of the entire conversation; the emotion embedding model maps the conversation content to the emotion feature space, captures the user's emotional expression in a specific sentence, and combines the conversation context to obtain the dynamic process of emotion changing over time; the context feature extractor is used to capture the contextual relationship of users in multiple rounds of conversations.

3. The method for intelligently processing robot customer service problems according to claim 1, characterized in that: The dynamic causal inference model is expressed as follows: Define each feature in the causal graph as a node v i , the edges between all nodes have dependency weight w i Indicates causal relationships between features; Through the analysis of historical data, the causal relationship model is derived, which is expressed as follows: in, represents the loss function of the causal relationship model, Pa(v i ) represents the feature v i The set of causal parent nodes, f i represents the causal function of the estimated feature, i represents the number of features; The surface symptom layer is used to process semantic features and emotional features, and directly reflects the problem currently expressed by the user; The potential cause layer is used to handle complex problems, identifying the potential causes behind the surface problems based on contextual features and problem complexity and through cause-effect diagrams; The root cause layer is used to combine sentiment features, context features, and historical data of multiple rounds of dialogue to trace the root cause of the problem; The surface level strategy directly generates standardized responses to questions at the surface feature layer; The potential hierarchy strategy generates a dialogue strategy for guiding the user to describe the problem in the next step for the problem at the potential cause layer; The root-level strategy generates complex responses to root-cause level questions based on sentiment fluctuations and contextual information in the long-term conversation history.

4. The method for intelligently processing robot customer service problems according to claim 1, characterized in that: The S301 specifically includes: for the surface symptom layer, the response strategy is based on semantic features and emotional features; for the potential cause layer and the root cause layer, a detailed response is generated based on semantic features and emotional features and combined with context features and complexity scores; At the same time, hierarchical adaptive coefficients are introduced into the surface symptom layer, potential cause layer and root cause layer to control the complexity of the response strategy at different problem levels. The largest hierarchical adaptive coefficient indicates that the problem is the most complex.

5. The method for intelligently processing robot customer service problems according to claim 4, characterized in that: The user's real-time feedback signal includes emotional fluctuations and context changes; the multi-level response strategies generated by the user are merged by weighted summation.

6. The method for intelligently processing robot customer service problems according to claim 1, characterized in that: The feedback signal includes emotional changes and semantic matching, wherein the emotional changes represent emotional fluctuations from the current dialogue turn to the previous dialogue turn, and the semantic matching represents the semantic matching between the robot's answer and the user's question.

7. The method for intelligently processing robot customer service problems according to claim 6, characterized in that: When the learning trigger is activated, the adaptive response strategy generation model is updated according to the feedback signal. The adaptive response strategy generation model is updated using a gradient descent method to optimize the adaptive response strategy generation model by minimizing the loss function.

8. The method for intelligently processing robot customer service problems according to claim 6, characterized in that: The S5 specifically includes: S501. Design semantic and sentiment-based regularization terms and construct loss functions based on semantic and sentiment features (F,R final ), which is expressed as follows: Among them, F represents the real-time feedback signal, R final represents the final response strategy, α and β represent the weights of balancing the feedback signal and the response strategy deviation; F optimal Represents the preset ideal feedback signal; R optimal represents the ideal response strategy; γ represents the weight coefficient of the regularization term based on semantics and sentiment; Represents a regularization term based on semantics and sentiment, which is used to ensure that the adaptive answer strategy generation model balances semantic and sentiment factors during adjustment, as shown below: Among them, e represents the emotional feature, X represents the semantic feature, and e optimal represents the ideal emotional characteristics, X optimal represents the ideal semantic feature, λ1 and λ2 represent the weights for balancing sentiment and semantic bias; S502, introducing an adaptive learning rate to dynamically adjust according to the strength and complexity of the feedback signal, to ensure that when the feedback signal is the largest, the adaptive response strategy generation model can be adjusted as quickly as possible; when the feedback signal is the weakest, the adaptive response strategy generation model can be updated; the adaptive learning rate η is expressed as follows: Among them, η0 represents the initial learning rate; κ represents the adjustment coefficient; q represents the complexity of the problem; S503, constructing a historical feedback accumulation item, which is used to combine feedback signals from multiple rounds of dialogues to guide the adaptive response strategy generation model to perform long-term optimization; S504: after the adaptive response strategy generation model is updated, a dynamic verification mechanism is designed to verify whether the adaptive response strategy generation model is updated, as shown below: if If it is negative, it means that the model optimization is successful; otherwise, adjustments will be made to ensure that the model gradually converges to the ideal response effect.

9. The method for intelligently processing robot customer service problems according to claim 6, characterized in that: The S6 specifically includes: S601, constructing a long-term feedback trend function according to the problem classification features, analyzing the long-term change trend of the feedback signal, and finding the long-term deviation point and optimization space; wherein the problem classification features are extracted from the long-term feedback signal trend of the answering strategy of the multi-level problem; S602, globally optimizing the model parameter θ of the adaptive response strategy generation model according to the long-term change trend of the feedback signal; S603, introducing a strategy evolution factor in the long-term optimization process of the adaptive response strategy generation model, and dynamically adjusting the rules of the adaptive response strategy generation model according to the long-term feedback signal; S604. An adaptive architecture expansion mechanism is designed to automatically expand different model modules or update the functions of existing modules according to the volatility of feedback signals, the complexity of conversations, and the diversity of user needs; S605. After the model is updated and the strategy is optimized, the effect of long-term optimization is verified through testing.

10. A robot customer service problem intelligent processing system, characterized in that: The system comprises: The user question acquisition subsystem is used to obtain user questions and extract and integrate the features of user questions, including: S101, obtaining the user's complete input dialogue, and preprocessing the dialogue, converting the preprocessed dialogue into text representation, embedding each word in the text as a semantic feature, and designing a sentiment embedding model to extract sentiment features from the input dialogue, and then designing a context feature extractor to capture the contextual relationship of the user in multiple rounds of dialogue to obtain context features; S102, generating a complexity score according to the semantic features, the emotional features and the context features, and generating a comprehensive feature vector by combining the semantic features, the emotional features, the context features and the complexity score; The user problem analysis subsystem is used to build a user dynamic causal reasoning model based on the integrated feature vectors and generate a multi-level problem-solving strategy based on the user problems, including: S201, build a dynamic causal reasoning model, take the comprehensive feature vector as input, infer the causal dependency behind the user's question, and for each feature, establish a causal function by analyzing the dependency between contextual features and semantic features; S202, classifying user problems into three levels according to the complexity of the user problems, including: surface feature level, potential cause level and root cause level; S203, generating three problem-solving strategies for the three levels respectively, including: a surface level strategy for problems at the surface feature level, a potential level strategy for problems at the potential cause level, and a root level strategy for problems at the root cause level; S204. If a strategy fails to effectively solve the problem, the causal relationship between features is re-evaluated through real-time user feedback signals, and the causal dependency weights in the causal function are dynamically adjusted; The user question answering subsystem is used to construct an adaptive answering strategy generation model based on the generated multi-level problem solving strategies and the integrated feature vectors to generate the answering strategy that best suits the current corresponding context, including: S301, constructing an adaptive answer strategy generation model to generate answer strategies at different levels according to the level of user questions and their corresponding comprehensive feature vectors; S302, adding regular terms based on sentiment characteristics and question complexity to constrain the generation of the answering strategy; S303, generating an initial response strategy, and dynamically adjusting the response strategy according to the user's real-time feedback to obtain an adjusted response strategy; S304, merging the multi-level response strategies generated by the user to generate a final response strategy, and updating the model parameters of the adaptive response strategy generation model in real time according to the feedback from the interaction with the user; The response learning subsystem is used to execute the response strategy and determine whether to trigger learning based on user feedback, including: Execute the final response strategy and listen to the user's feedback, extract the feedback signal from the user's feedback, and determine whether to trigger learning based on the feedback signal: If the feedback signal shows that the response effect is poor, the learning module is triggered to update the response model; If the feedback signal indicates that the response is good, learning is not triggered; The optimization subsystem is used to perform real-time adaptive learning and optimization of the adaptive response strategy generation model through user feedback and historical user feedback, analyze the long-term trends of the model parameters of the adaptive response strategy generation model after historical user feedback and multiple optimizations, perform global optimization based on the long-term trends, optimize the response strategy that best suits the current corresponding context, and generate the final response strategy.

Images