Intelligent customer service dialogue management method and system fusing multi-layer memory

Abstract

The invention discloses an intelligent customer service dialogue management method and system fused with multi-layer memory, and belongs to the technical field of computers, and the method obtains a three-dimensional context through multi-mode context awareness, achieves the dynamic scheduling of a memory system through a memory coordinator, and combines a dynamic construction mechanism of a service situation portrait, thereby achieving the intelligent customer service dialogue management. The problem that the response of the intelligent customer service system is not matched with the complex and dynamic service demand of the user due to the single situation awareness dimension and the rigid memory scheduling mechanism is effectively solved, and the adaptive personalized service response can be generated according to the real-time change of the dialogue situation.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an intelligent customer service dialogue management method and system that integrates multi-layer memory. Background Technology

[0002] Current intelligent customer service systems suffer from significant deficiencies in dialogue management capabilities. Most systems rely solely on user-inputted text for intent recognition, failing to effectively integrate multi-dimensional information such as emotional features embedded in voice signals and visual scenes presented in image data. This results in a one-sided and inaccurate understanding of the user's state. In actual interactions, user emotional fluctuations and the visual environment often conceal key service needs, but existing technologies lack collaborative perception mechanisms for multimodal data, making it impossible to comprehensively capture the dynamic changes in the dialogue context. For example, they cannot analyze emotional intensity through the acoustic features of voice or extract scene entity relationships through multi-scale feature extraction from images, thus overlooking crucial contextual information that influences service decisions.

[0003] To further improve interaction quality, some technical solutions attempt to incorporate historical interaction records as a reference. However, this historical information is typically stored in a single static data pool, with a simple and inflexible retrieval mechanism. Systems often rely solely on keyword matching to retrieve past cases, failing to intelligently allocate different types of memory resources—such as user personalized preferences, operation process records, and historical solutions—based on the real-time evolution of the dialogue context. This rigid memory allocation method makes it difficult for the system to distinguish between service strategies in contextual memory, case solutions in experiential memory, and semantic fragments in vector memory, hindering cross-level memory collaborative optimization. Consequently, the generated responses are mechanical and rigid, lacking adaptability to complex user needs. For example, they cannot dynamically adjust service strategy priorities in emotionally charged scenarios, or integrate historical cases with real-time semantic fragments in semantically complex dialogues. Ultimately, the output responses are severely out of sync with users' dynamically changing service expectations, failing to meet the demands for deeply personalized services in real-world scenarios. Summary of the Invention

[0004] This application provides an intelligent customer service dialogue management method and system that integrates multi-layer memory, which can meet the needs of deeply personalized services in real-world scenarios. The technical solution is as follows: On the one hand, a method for intelligent customer service dialogue management that integrates multi-layered memory is provided, the method including: In response to dialogue requests from intelligent customer service, the system performs multimodal context perception on voice signals, image data, and text data from users to obtain emotional state representation, visual semantic understanding, and text intent parsing. Based on emotional state representation and text intent parsing, a memory coordinator is used to coordinate the multi-layer memory system to obtain a set of memory features that match the current dialogue context. The coordinated scheduling includes activating personalized service strategies in contextual memory based on emotional state representation and retrieving relevant historical solutions from experiential memory and vector memory based on text intent parsing. Based on memory feature sets and visual semantic understanding, a fusion decision-maker is used to integrate multi-source memories to obtain a fusion decision result that is adapted to the current dialogue context. Multi-source memory integration includes consistency verification of operation records in program memory and user preferences in context memory, as well as correlation fusion of case solutions in experience memory and semantic fragments in vector memory. Based on the fusion of decision results and service context profiles, a personalized service response is generated for each dialogue request through an intelligent decision engine. The service context profile is dynamically constructed based on emotional state representation, visual semantic understanding, and text intent parsing to reflect the user's service needs and preferences at the current stage of the dialogue. Attached Figure Description

[0005] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0006] Figure 1 This is a schematic diagram of the implementation environment of an intelligent customer service dialogue management method integrating multi-layer memory provided in an embodiment of this application; Figure 2 This is a flowchart of an intelligent customer service dialogue management method integrating multi-layer memory provided in an embodiment of this application; Figure 3 This is a partial flowchart of an intelligent customer service dialogue management method integrating multi-layer memory provided in an embodiment of this application; Figure 4 This is a partial flowchart of another intelligent customer service dialogue management method integrating multi-layer memory provided in the embodiments of this application. Detailed Implementation

[0007] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0008] In this application, the terms "first", "second", etc. are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first", "second", and "n", nor are there any restrictions on the quantity or execution order.

[0009] Artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain better results.

[0010] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge sub-models to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence.

[0011] Emotional state representation: This refers to the digital output used to quantitatively describe a user's current emotional state, obtained through the analysis of a user's non-textual modal information using a computational model. This representation is a structured data object, typically including emotion type (such as discrete labels like happy, frustrated, or anxious obtained through a classification model) and emotion intensity (such as continuous values ​​between 0.0 and 1.0, or low, medium, and high levels, obtained through a regression model). It serves as one of the key inputs for downstream modules to make personalized decisions.

[0012] Visual semantic understanding refers to the machine-readable structured description of a visual scene obtained through deep analysis of user-provided images or video frames. It includes not only scene entities identified by object detection models (such as a user holding a damaged product), but also relationships between entities inferred through relational models (such as hand-holding relationships), and scene classification of the overall environment (such as indoor home environment), providing crucial context for understanding the physical context in which the service occurs.

[0013] Text intent parsing refers to the formalized conclusions about the user's dialogue purpose formed after performing deep semantic analysis on the text sequence input by the user. This parsing result is typically manifested on two levels: first, intent type identifiers, which are one or more labels belonging to a predefined intent classification system (such as consultation, processing, complaint); second, semantic embedding vectors, vector points in a high-dimensional space used to capture complex, compound, or subtle intents that cannot be summarized by simple labels.

[0014] Memory Coordinator: This refers to the core control algorithm module responsible for the global management of access to multi-layered memory resources. It receives emotional state representations and text intent parsing as input, and dynamically determines, through built-in scheduling strategies (such as weight calculation based on attention mechanisms), what information to retrieve from episodic memory, experiential memory, procedural memory, and vector memory, as well as the priority of this information, thereby achieving on-demand and efficient collaborative activation of memory resources.

[0015] Cooperative scheduling: This refers to the specific operational process performed by the memory coordinator. It emphasizes not simple parallel retrieval, but rather intelligent resource allocation based on the input context. For example, when a high level of user emotion is detected, the scheduler will prioritize requesting soothing service strategies from the contextual memory. When complex semantic intent is identified, it will simultaneously configure a broad-range retrieval from the experience memory and a deep semantic matching from the vector memory, fusing the results of both in a dynamically calculated proportion.

[0016] Fusion Decision Maker: This refers to a dedicated algorithm module responsible for integrating, verifying, and optimizing information from multiple independent sources. It receives the memory feature set output by the memory coordinator and the visual semantic understanding provided by the perceptual layer. By performing specific operations such as logical consistency checks, semantic relevance fusion, and conflict resolution, it outputs a unified, internally consistent, and executable fusion decision result.

[0017] Multi-source memory integration: This specifically refers to information processing actions completed within the fusion decision-maker. Its core lies in resolving potential contradictions and achieving complementarity between information from different sources. For example, consistency verification compares the standard operating procedures of procedural memory with the user's historical preferences in contextual memory, identifying and correcting redundant steps or style inconsistencies. Relevance fusion involves splicing and completing cases from experiential memory with semantic fragments from vector memory to form a solution more suited to the current context.

[0018] Fusion decision results refer to structured data packets generated after a multi-source memory integration process, which can directly drive response generation. It is not simply a collection of information, but an optimized and validated decision object containing multiple fields, such as the final determined sequence of operations, details of the optimized solution, confidence scores for each solution, and recommended execution order.

[0019] Service context profile: refers to a user state model that is continuously and dynamically updated within a single conversation. It is a data structure maintained in memory, whose fields are refreshed in real time by fusing the latest emotional state representations, visual semantic understanding, and text intent parsing, thus forming a comprehensive context model that can fully reflect the user's current service needs, preference characteristics, emotional context, and conversation stage.

[0020] Intelligent decision engine: This refers to a module that receives fused decision results and service context profiles as input, and ultimately generates natural language responses or service instructions as output. It is typically a sequence generation model based on deep learning (such as the Transformer architecture). Its core responsibility is to transform structured decision data into response text that conforms to human language habits and is highly matched to the user's personalized context.

[0021] Personalized service response: This is the final output of the entire methodology. Its personalization is reflected in three aspects: content elements originate from decisions optimized through memory fusion, ensuring accuracy; tone elements (such as the emotional tone of word choice and sentence structure) are adapted to the user's real-time emotional state; and action suggestion elements (such as specific follow-up steps) comprehensively consider the user's historical preferences and current scenario constraints, ensuring feasibility.

[0022] Acoustic feature analysis refers to the process of extracting low-level features from the waveform of the original speech signal to characterize the physical properties of the sound. Specifically, this includes calculating features such as Mel-frequency cepstral coefficients (MFCCs), spectral centroid, and zero-crossing rate. These features constitute the initial input vector for subsequent sentiment analysis.

[0023] Prosodic pattern extraction refers to the process of analyzing high-level features related to language rhythm, stress, and intonation patterns from speech signals. The main features extracted include: fundamental frequency (F0) and its trajectory (used to represent intonation), energy (amplitude) and its variation (used to represent stress), and syllable duration (used to represent speech rate).

[0024] Multidimensional sentiment analysis refers to a comprehensive sentiment computing process that does not simply output a sentiment label, but simultaneously and jointly completes two tasks: sentiment type identification (classification task) and sentiment intensity estimation (regression or ranking classification task), thereby providing a richer sentiment description.

[0025] Multi-scale visual feature extraction refers to the technique of extracting visual features from different levels of abstraction of an input image using convolutional neural networks (CNNs). Low-level features (from shallow networks) capture details (such as edges and textures), mid-level features capture part information, and high-level features (from deep networks) capture global semantic information. Combining these features at different scales can more comprehensively describe the image content.

[0026] Scene semantic parsing refers to the process of understanding the overall content of an image and converting it into machine-readable semantics. It typically includes: scene classification (determining which type of scene the image belongs to, such as an office or a street), object detection and recognition (locating and identifying specific objects in the image), and attribute recognition (identifying the attributes of objects, such as color and state).

[0027] Multi-level semantic reasoning refers to combining low-level visual features extracted from images with high-level semantic labels, and using more complex models (such as Generative Neural Networks, GNNs) to perform logical reasoning in order to uncover semantic information in images that is not explicitly displayed but is implicit. For example, inferring that a person is working based on the fact that they are sitting in front of a computer.

[0028] Semantic structure analysis refers to the process of performing linguistic analysis on input text to obtain its basic semantic units and grammatical structure. It mainly includes word segmentation, part-of-speech tagging, named entity recognition, and syntactic dependency parsing, with the aim of understanding basic semantic relationships such as who did what to whom in a sentence.

[0029] Intent feature extraction refers to the process of further extracting the most discriminative feature vectors for intent classification tasks from the preliminary analysis of the text semantic representation (such as word vector sequences or sentence vectors). This can be achieved by using techniques such as attention mechanisms to automatically focus on keywords.

[0030] Intent reasoning refers to the process of deepening and refining the intent features extracted from the current single sentence by combining the contextual history of the current dialogue and information from multiple turns of interaction. It aims to capture the evolution of user intent during the dialogue process, thereby arriving at more accurate and forward-looking intent conclusions.

[0031] Time-frequency feature extraction refers to the process of analyzing the joint distribution characteristics of non-stationary signals (such as speech) in both time and frequency dimensions. The most typical method is the Short-Time Fourier Transform (STFT), which produces a spectrum that visually displays the changes in the signal's frequency components over time.

[0032] Energy distribution analysis refers to analyzing the pattern of changes in the amplitude (or energy) of a speech signal over time. This includes calculating the short-time average energy and analyzing the contour of the energy envelope, which is crucial for identifying emphasis, stress, and silence in speech.

[0033] Emotional state mapping refers to the process of taking various acoustic features (such as spectral features and prosodic features) extracted from the front end as input, and mapping them to a pre-trained machine learning model (such as DNN or SVM) through a pre-trained machine learning model. The output of this model is the emotional state representation.

[0034] Region feature extraction: This refers to extracting features from local regions within an image that may contain independent semantics, rather than being limited to the entire image. This is typically achieved using object detection models (such as YOLO and Faster R-CNN), which first generate candidate regions (bounding boxes) and then extract visual feature vectors from the image content within each region.

[0035] Spatial relationship analysis refers to the quantitative or qualitative analysis of the geometric relationships between multiple entities in an image after they have been identified. For example, by calculating the relative positions of bounding boxes (e.g., A is to the left of B) and overlapping areas (e.g., A intersects with B), a set of predicates describing spatial relationships is formed.

[0036] Scene semantic association reasoning: This refers to using graph models and other technologies to construct a scene graph by treating entities in an image as nodes and their spatial and semantic relationships as edges. Then, reasoning is performed on this graph to uncover deeper semantic relationships between entities that are not directly apparent (such as inferring that cooking is taking place based on people, knives, cutting boards, and vegetables).

[0037] Multi-level feature fusion refers to strategies for effectively integrating feature vectors from different levels (such as local region features and global scene features) or different modes (such as visual features and semantic labels). Common methods include vector concatenation, weighted summation, or feature selection and fusion based on attention mechanisms.

[0038] Context encoding refers to using sequence models such as recurrent neural networks (RNNs) or Transformer encoders to encode previously occurring statements in the current dialogue, generating a comprehensive vector representation that contains contextual information. This vector helps in understanding the referent and background of the current statement.

[0039] Sequence modeling refers to modeling data sequences arranged in chronological order (such as speech frame sequences or text word sequences) to capture their dynamic evolution patterns and long-term dependencies. RNN, LSTM, and Transformer are typical architectures for accomplishing this type of task.

[0040] Memory enhancement mechanism: refers to the introduction of information from external multi-layered memory (such as memory state features) as additional input or conditions during the reasoning process of neural network models (such as models used for intention reasoning), thereby enhancing the model's reasoning ability and context awareness range.

[0041] Intent evolution analysis refers to tracking and analyzing the dynamic changes of user intents throughout multiple rounds of dialogue. It focuses on how intents are generated, solidified, transformed, or dissolved as the dialogue progresses, aiming to understand the user's decision-making path and the process of need deepening.

[0042] Intent evolution trajectory: refers to the formal representation of the results of intent evolution analysis. It is usually a sequence arranged in chronological order. Each element in the sequence represents the user intent and its confidence level identified at a specific dialogue turn or moment, thus clearly showing the ins and outs of the intent.

[0043] Emotional intensity: This refers to a quantitative measure of the degree of emotional arousal in a user. It is a numerical value or level used to distinguish different levels of expression within the same emotional type. For example, mild annoyance and rage both belong to the category of anger, but have significantly different emotional intensities.

[0044] Emotion type: refers to the basic classification of a user's emotional experience. It uses discrete labels to identify the nature of the emotion, such as happy, sad, fearful, surprised, angry, neutral, etc., and is usually given by a classification model.

[0045] Service preference dimensions: These refer to the different aspects or axes used in contextual memory for structured storage of user preferences. Each dimension represents a separate type of preference, such as preference for information detail, preference for response speed, preference for communication style (formal / casual), and preference for channel (text / voice).

[0046] Dynamic weight allocation: This refers to an algorithmic process that automatically adjusts the weights of different factors (such as different service preference dimensions) in decision-making based on real-time changing input conditions (such as current emotional intensity). For example, when emotionally agitated, the weight of the response speed dimension will be significantly increased.

[0047] Cross-level related retrieval: refers to a joint retrieval strategy that simultaneously searches for relevant information from two different organizational levels of memory (such as structured experience memory and unstructured vector memory) and establishes semantic associations between the two types of retrieval results.

[0048] Priority fusion: When merging multiple alternative solutions or strategies, instead of simply listing them side by side, they are sorted according to their suitability for the current situation, importance, or urgency. The results of the sorting are then selectively integrated or output to ensure that the optimal solution is adopted first.

[0049] Synergy assessment: refers to a metric function used to evaluate whether two or more independently developed service strategies or solutions can cooperate and work together when applied to the same specific service scenario without causing conflicting objectives, resource competition, or logical contradictions.

[0050] Emotion intensity quantification refers to the process of transforming the abstract concept of emotion intensity into a specific numerical value or discrete level. For example, an acoustic feature can be mapped to a floating-point number between 0.0 (completely calm) and 1.0 (extreme excitement) using a regression model.

[0051] Emotion type classification refers to the process of inputting comprehensive acoustic features into a classification model (such as a Softmax classifier) ​​and assigning them to predefined emotion categories. The model's output is typically a probability distribution for each category, with the highest probability being taken as the classification result.

[0052] Preference weight distribution: This refers to a numerical vector where each element corresponds to a service preference dimension, and its value represents the importance of that preference dimension relative to other dimensions in a specific context. The sum of all weights is usually 1.

[0053] Historical service records refer to past interaction data related to a specific user, stored in contextual memory. Each record may include: service request content, the response provided, explicit user feedback (such as ratings), and implicit behavioral feedback (such as response speed, whether the question was asked repeatedly).

[0054] Service strategy generation: refers to the process of filtering, sorting, trimming or combining historical service records based on the currently active preference weight distribution, thereby generating one or more specific and executable service plans.

[0055] Intent type identification: This refers to the process of assigning the core intent in text intent parsing to a specific category within a predefined classification system. For example, identifying that the user's current intent belongs to the query-product-price type.

[0056] Semantic complexity assessment refers to the process of quantifying the semantic depth, breadth, or abstraction of a user query or request. Assessment factors may include: the number of entities in the query, the complexity of relationships, whether multi-hop reasoning is involved, and whether it contains fuzzy or abstract concepts. The output can be a rating (e.g., simple, complex) or a score.

[0057] Retrieval strategy configuration refers to the process of dynamically setting retrieval parameters based on the results of intent type identification and semantic complexity assessment. These parameters include, but are not limited to: the retrieval depth of experiential memory (how many layers of association are traversed), the retrieval range of vector memory (similarity threshold), and the fusion ratio of retrieval results between the two types of memory.

[0058] Service Scenario Synergy Assessment: This refers to a specific calculation function used to quantify the degree of matching and collaboration between a service strategy and a solution in a given service scenario. Its output is a synergy score; a higher score indicates a potentially better combined effect.

[0059] Synergy Distribution: This refers to a score matrix or list obtained by evaluating the synergy of each pair of service scenarios within the service strategy set and solution set. This distribution reflects the synergy potential of all possible strategy-solution combinations.

[0060] Strategy priority fusion: refers to an advanced fusion technique that, when making the final decision, not only considers the original scores of each option, but also introduces semantic complexity and emotional intensity as external weighting factors to recalibrate the priority of the options, thereby generating a more robust and context-sensitive final strategy sequence.

[0061] Scene entities and relationships: This is the core data structure extracted from visual semantic understanding. Scene entities are objects or people in an image that have independent semantics (such as a customer, a broken screen, or an invoice). Relationships are predicates that connect two entities and describe their interaction or spatial location (such as a customer pointing at the screen or an invoice placed on the table).

[0062] Dynamic conflict resolution refers to the process of detecting and resolving logical contradictions between information from different memory sources in real time during a dialogue flow. For example, if the program memory requires user authentication while the contextual memory indicates that the user strongly dislikes repeated authentication, a resolution rule is activated, which may generate a compromise solution that uses a more convenient authentication method.

[0063] Consistency enhancement refers to the process of optimizing the initial solution after dynamic conflict resolution. It aims to ensure that the final sequence of operations or service content maintains a higher degree of consistency with the user's long-term preferences, current emotional state, and visual scene constraints, thereby improving the user experience.

[0064] Semantic guidance: This refers to using the deep semantic context provided by text intent parsing to guide and constrain the direction of the multi-source memory integration process. For example, when intent parsing reveals that the user's core concern is privacy and security, the fusion process will tend to select those scheme fragments that are marked as highly secure in experiential memory and vector memory.

[0065] Complementary fusion: This refers to a fusion strategy that aims to combine the advantages of information with different characteristics. Specifically, it combines structurally complete but potentially rigid case solutions retrieved from experiential memory with flexible and context-rich semantic fragments retrieved from vector memory. The flexibility of the latter compensates for the shortcomings of the former, forming a reliable and novel solution.

[0066] Semantic gaps refer to gaps in information or semantic incoherence that occur when past experiences are directly applied to the current context of a conversation due to differences in situation. For example, an old refund policy example might lack instructions on how to handle complaints made through social media channels; this gap is a semantic gap.

[0067] Demand intensity score: This refers to a quantitative estimate of the urgency of a user's demand for a particular service, calculated by analyzing behavioral data in the current conversation (such as input frequency, sentence length, and number of repeated questions). The higher the score, the more urgent the demand, and the more direct and rapid the response strategy may be required.

[0068] Preference sensitivity refers to the degree to which a user cares about their personal preferences being adhered to, calculated by analyzing historical user interaction data (such as the intensity of feedback on service deviations and the consistency of choices across different options). Users with high sensitivity require stricter adherence to their historical preferences when making decisions.

[0069] Service content: This refers to the core entity information part of a personalized service response, namely the specific facts, answers, solutions, or confirmation information to be conveyed to the user. It originates directly from the integrated decision-making results and is the backbone of the response text.

[0070] Tone intensity: This refers to the strength of the emotional tone infused into the text when generating personalized service responses. It needs to match the user's emotional state. For example, a more assertive and empathetic tone might be used for users with high emotional intensity, while a more neutral and neutral tone would be used for users with calm emotions.

[0071] Action suggestion elements: These refer to the specific and clear instructions in a personalized service response that guide the user to the next step. For example, "Please click the link to complete the payment," "Please provide your order number so we can check," or "We suggest you restart your device and try again." These suggestions are derived from standard processes remembered by the program and successful cases learned from experience.

[0072] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0073] Figure 1 This is a schematic diagram illustrating the implementation environment of an intelligent customer service dialogue management method integrating multi-layer memory, as provided in an embodiment of this application. See also... Figure 1 This implementation environment may include node 110 and server 140.

[0074] Node 110 is connected to server 140 via a wireless or wired network. Optionally, node 110 can be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. Node 110 has an application installed and running that supports intelligent customer service dialogue management with integrated multi-layer memory.

[0075] Server 140 is a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. Server 140 can provide background services for applications running on node 110.

[0076] In traditional intelligent customer service systems, dialogue management is limited by a single-modal context awareness capability and a static memory retrieval mechanism. Specifically, intent recognition is based solely on the current text dialogue content, failing to effectively integrate acoustic features from speech signals to analyze user emotional states, nor utilize spatial relationships from image data for visual semantic understanding. This results in a lack of dimensionality in portraying users' true service needs. Furthermore, historical interaction records are fixed in a static data pool, relying on simple keyword matching rules for retrieval. They cannot intelligently allocate user preferences from contextual memory, operational procedures from program memory, or similar cases from experiential memory based on the dynamic evolution of the dialogue context. Consequently, the generated responses lack adaptability to complex user states, resulting in a disconnect between service strategies and the current stage of the dialogue, thus affecting the coherence of the interaction and user satisfaction.

[0077] For example, in scenarios where users initiate customer service requests via mobile devices, the voice signal exhibits high-frequency fluctuations and short pauses; image data captures the user's visual context of being in a confined space holding a damaged device; and text input contains vague statements like "this won't work." However, simply retrieving general repair guidelines from historical records based solely on the text keyword "damaged" fails to identify the user's anxiety state using acoustic features, nor does it assess the spatial constraints limiting the operational steps through visual semantic understanding. As a result, the mechanically pushed voice guidance instruction, "Please describe the specific problem," conflicts with the feasibility of voice interaction in a noisy environment. Furthermore, the recommended repair steps are not adapted to the physical location of the damaged device, preventing the user from performing the operation and further exacerbating their emotional state.

[0078] If the above problems are not addressed, the intelligent customer service system will continuously generate responses that do not match the user's dynamic service needs. The system will be unable to adjust service strategy priorities based on emotional state representation, nor can it perform spatial constraint verification of operation records based on visual semantic understanding, causing historical solution calls to become detached from the real-time dialogue context. Consequently, the risk of service interruption during the interaction process will significantly increase, user trust in the system will gradually decrease, and ultimately, the systemic failure of personalized service capabilities will result, failing to meet the basic requirements for response adaptability in real-world scenarios.

[0079] To address this, this application proposes an intelligent customer service dialogue management method that integrates multi-layered memory. Taking the server as the execution entity as an example, see [link to relevant documentation]. Figure 2 This includes the following steps.

[0080] 201. In response to dialogue requests from intelligent customer service, perform multimodal context perception on voice signals, image data and text data from users to obtain emotional state representation, visual semantic understanding and text intent parsing.

[0081] 202. Based on emotional state representation and text intent parsing, a memory coordinator is used to coordinate the multi-layer memory system to obtain a set of memory features that match the current dialogue context. The coordinated scheduling includes activating personalized service strategies in contextual memory based on emotional state representation and retrieving relevant historical solutions from experiential memory and vector memory based on text intent parsing.

[0082] 203. Based on memory feature sets and visual semantic understanding, a fusion decision-maker is used to integrate multi-source memories to obtain a fusion decision result that is adapted to the current dialogue context. Multi-source memory integration includes consistency verification of operation records in program memory and user preferences in context memory, as well as correlation fusion of case solutions in experience memory and semantic fragments in vector memory.

[0083] 204. Based on the fusion decision results and service context profile, a personalized service response corresponding to the dialogue request is generated through an intelligent decision engine. The service context profile is dynamically constructed based on emotional state representation, visual semantic understanding and text intent parsing, and is used to reflect the user's service needs and preference characteristics at the current dialogue stage.

[0084] This application provides an intelligent customer service dialogue management method that integrates multi-layered memory. Multimodal context awareness refers to the technical process of simultaneously processing speech signals, image data, and text data to obtain user state information. It can employ basic signal processing techniques to analyze energy fluctuations in speech signals to roughly distinguish emotional states, or use pre-trained simple image classification models to identify scene categories in image data, and utilize keyword matching methods to extract intent labels from text data. For example, the energy change trajectory of the speech signal can be used to determine the range of user emotional intensity, object detection results from image data can provide scene entity clues, and keyword frequency statistics from text data can assist in intent type identification. Its main purpose is to achieve a multi-dimensional understanding of the user context, avoiding the one-sidedness caused by single-modal input. Furthermore, the memory coordinator refers to the component that coordinates the data scheduling of a multi-layered memory system. Its implementation can adopt rule-based scheduling strategies, such as triggering different service strategy libraries according to preset emotion thresholds, or retrieving matching cases from historical records through simple similarity calculations. For example, when the value of the emotional state representation exceeds the preset threshold, the soothing service strategy is called first, and when the keywords of the text intent parsing match a specific pattern, the relevant solution set is retrieved. Its main purpose is to realize the dynamic allocation of memory resources to adapt to the changing needs of the current dialogue context.

[0085] Specifically, service context profiles can be understood as dynamic representations that reflect user needs and preferences. In practical applications, their construction can be based on statistical features of historical interactions, such as estimating the strength of current preferences by calculating the average of past user preferences, or generating profile vectors by using simple weighted combinations of real-time inputs. For example, numerical representations of emotional states and scene labels from visual semantic understanding can be linearly combined to form service demand indicators. The main purpose is to provide a real-time snapshot of the user's state to guide the response generation process.

[0086] Therefore, this application obtains a three-dimensional context through multimodal context awareness and uses a memory coordinator to realize the dynamic scheduling of the memory system. Combined with the dynamic construction mechanism of service context profile, it effectively solves the problem that the response of intelligent customer service system is not matched with the complex and dynamic service needs of users due to the single dimension of context awareness and the rigid memory scheduling mechanism. It can generate adaptive personalized service responses according to the real-time changes of the dialogue context.

[0087] In the intelligent customer service dialogue management process, upon receiving a user's dialogue request, the system first simultaneously analyzes voice signals, image data, and text data to achieve multimodal context awareness. Specifically, voice signals are used to extract acoustic features to identify the user's emotional state, image data is used to analyze the visual scene to understand the user's environment, and text data is used to analyze semantic content to determine the user's intent. This generates three key types of information: emotional state representation, visual semantic understanding, and text intent parsing. Furthermore, the memory coordinator activates personalized service strategies in contextual memory based on the emotional state representation; for example, when a high level of user emotion is detected, the priority of reassuring service strategies is automatically increased. Simultaneously, historical solutions are retrieved from experiential and vector memories based on text intent parsing. For example, the retrieval ratio of experiential and vector memories is dynamically adjusted for complex intent types, thus forming a memory feature set that matches the current dialogue context. The fusion decision-maker then integrates the memory feature set and visual semantic understanding to perform multi-source memory integration. This includes consistency verification between operation records in program memory and user preferences in contextual memory; for example, when visual semantic understanding identifies that the user is in a noisy environment, the recommended voice interaction steps in the operation record are automatically corrected. Furthermore, it fuses the case solutions in experiential memory with semantic fragments in vector memory, for example, using vector memory to supplement the semantically missing parts of the case solutions in the current dialogue context, thereby generating a fusion decision result adapted to the current dialogue situation. The intelligent decision engine ultimately generates personalized service responses based on the fusion decision result and service context profile. The service context profile continuously integrates the dynamic changes of emotional state representation, visual semantic understanding, and text intent parsing to reflect the evolution of user needs in real time. For example, it automatically adjusts the intensity of the response tone based on the decrease in emotional intensity during the dialogue progression.

[0088] This method overcomes the limitations of single text input through multimodal context awareness, comprehensively capturing users' implicit emotions and real-world scenarios, avoiding misjudgments caused by incomplete information. The collaborative scheduling mechanism of the memory coordinator dynamically adjusts memory retrieval strategies based on emotional state and textual intent, replacing the rigid retrieval based on fixed rules in traditional methods, thus improving the flexibility and contextual adaptability of historical information retrieval. The consistency verification and relevance fusion functions of the fusion decision-maker effectively eliminate logical contradictions between operation records and user preferences, and fill semantic gaps in case solutions, ensuring that the decision result closely matches the current dialogue context. The intelligent decision engine, combined with dynamically constructed service context profiles, generates responses that accurately respond to the phased evolution of user needs, completely changing the mechanical and rigid response problems caused by the single dimension of context awareness and the rigid memory scheduling mechanism in traditional intelligent customer service, achieving precise matching of complex and dynamic user service needs.

[0089] In some embodiments, to achieve highly adaptive cooperative scheduling, this application provides an endogenous two-layer memory architecture and a complete dynamic evolution mechanism. This design overcomes the limitations of "memory silos" and static retrieval commonly found in current intelligent agent systems, and by simulating the human memory processing process, enables the system to have the ability to continuously learn and evolve.

[0090] The memory module stores general memory and user memory. General memory is stored using an embedded semantic vector pool, encoding generalized information such as domain knowledge and standard solutions. This design supports efficient similarity-based feature retrieval and serves as the knowledge base for the system's analogies and reasoning. User memory is organized using a global slots structure, storing personalized user information (such as identity, historical preferences, and specific needs) in key-value pairs. This structured design supports millisecond-level location and retrieval via keywords, ensuring real-time access to personal context.

[0091] Furthermore, regarding the context-aware dynamic memory retrieval in the above embodiments, the memory coordinator does not access all memories equally. Therefore, an attention-based memory selector is introduced. This selector uses the real-time context in the user's memory as the query benchmark, combining current emotional intensity, dialogue stage, and other multi-dimensional contextual signals to dynamically calculate the association weight between the selected memory and various items in the general memory vector pool. This process simulates the human ability to selectively recall relevant knowledge based on emotion and topic focus during dialogue, effectively overcoming the industry challenge of diluting early key information in long-context dialogues, and achieving precise, on-demand activation of memory resources.

[0092] Furthermore, this application provides an incremental fusion and purification method for general memory, using a gated recurrent unit (GRU) network to fuse newly generated effective memory features with historical memories. The update and reset gate mechanisms of the GRU can adaptively determine the ratio of memory retention to update, achieving smooth iteration and accumulation of knowledge. To simulate the forgetting curve of human memory and avoid information redundancy, a memory decay mechanism based on fixed dialogue rounds is introduced. Simultaneously, a memory summary operation is periodically triggered: recent interactions are analyzed offline, refining high-frequency, high-value dialogue patterns into more concise rules of thumb or strategy patterns. This process is similar to memory consolidation; after verification, these essential elements are archived into the general memory pool, thus completing the sublimation from concrete interactions to abstract knowledge and driving the continuous autonomous evolution of system capabilities.

[0093] For updates to user memories, especially when writing critical personal information (such as changed phone numbers or confirmed long-term preferences) into the global slot, a lightweight verification and confirmation process is initiated. This process assesses the reasonableness and consistency of the information and, if necessary, initiates user confirmation through the interactive interface, completing the update after receiving positive feedback. This ensures the accuracy of user memories, reflects respect for user rights, and complies with data quality management standards.

[0094] This application further proposes the following technical solutions, taking a server as the execution subject as an example, see [link to relevant documentation]. Figure 3 This includes the following steps.

[0095] 301. Perform acoustic feature analysis and prosodic pattern extraction on the speech signal to obtain acoustic feature sequences and prosodic feature parameters.

[0096] 302. Perform multi-dimensional emotion analysis on acoustic feature sequences and prosodic feature parameters to obtain emotional state representations. Multi-dimensional emotion analysis includes joint identification of emotion intensity and emotion type.

[0097] 303. Perform multi-scale visual feature extraction and scene semantic parsing on image data to obtain visual feature maps and scene understanding information.

[0098] 304. Perform multi-level semantic reasoning on visual feature maps and scene understanding information to obtain visual semantic understanding. Multi-level semantic reasoning includes entity recognition and relationship understanding of visual scenes.

[0099] 305. Perform semantic structure analysis and intent feature extraction on the text data to obtain semantic representation vectors and intent feature representations.

[0100] 306. Perform context-aware intent reasoning on semantic representation vectors and intent feature representations to obtain text intent parsing. Context-aware intent reasoning includes intent evolution analysis in the current dialogue context.

[0101] Among these, acoustic feature analysis and prosodic pattern extraction refer to extracting physical acoustic characteristics and intonation variation patterns from speech signals. This can be achieved using spectral envelope analysis and fundamental frequency trajectory modeling based on short-time Fourier transform, aiming to capture dynamic emotional cues contained in speech. Multi-dimensional emotion analysis refers to simultaneously quantifying the intensity and specific category of emotions. This can be achieved using support vector machines combined with fuzzy logic classifiers to jointly discriminate acoustic features, aiming to avoid perceptual biases caused by single-dimensional emotion judgments. Multi-scale visual feature extraction and scene semantic parsing refer to the hierarchical capture and structured semantic transformation of visual information from macro-scenes to micro-details. This can be achieved using feature pyramid networks combined with semantic segmentation models, aiming to ensure a deep association between visual input and the service context. Multi-level semantic reasoning refers to the in-depth mining of object entities and their logical relationships in a visual scene. This can be achieved using a knowledge graph-based reasoning engine to model the topological relationships of scene entities, aiming to go beyond surface feature matching to understand implicit semantics. Semantic structure analysis and intent feature extraction refer to parsing the grammatical level and deep-seated needs of text. This can be achieved by combining dependency parsing with word vector clustering, aiming to avoid the shallow analysis limitations of keyword matching. Context-aware intent reasoning refers to dynamically tracking the evolution of user intent in the dialogue flow. This can be achieved by combining hidden Markov models with dialogue history state encoding, aiming to adapt to the continuity and dynamism of dialogue.

[0102] Through the above technical solutions, this application achieves deep perception of user state information. Multi-dimensional emotion analysis avoids the limitations of single-dimensional emotion judgment, enabling emotion state representation to simultaneously reflect both the intensity and specific category of emotion. Multi-level semantic reasoning delves into the logical connections between entities in the visual scene, allowing visual semantic understanding to transcend surface features and capture implicit semantics. Context-aware intent reasoning dynamically tracks the evolution of intent in the dialogue flow, adapting text intent parsing to the continuity and dynamism of the dialogue. Ultimately, the depth and accuracy of emotion state representation, visual semantic understanding, and text intent parsing are improved, effectively solving the problem of misjudgment of user needs caused by one-sided perception results.

[0103] This application further proposes the following steps: Time-frequency feature extraction and energy distribution analysis are performed on the acoustic feature sequence to obtain the spectral feature vector and energy change trajectory.

[0104] Based on prosodic feature parameters and energy change trajectories, a comprehensive acoustic feature representation is generated through a feature fusion network.

[0105] The spectral feature vector and acoustic integrated feature representation are input into the emotion recognition model. The emotion recognition model is used to map the emotion state to obtain an emotion state representation that includes emotion type and emotion intensity. The emotion state mapping includes dynamic weighted fusion of spectral features and acoustic integrated features based on an attention mechanism.

[0106] Among these, time-frequency feature extraction refers to the dynamic analysis of acoustic feature sequences in the time-frequency domain, which can be achieved using short-time Fourier transform or wavelet transform, aiming to capture the evolution of spectral patterns in non-stationary speech signals. Energy distribution analysis refers to quantifying the distribution characteristics of speech signal energy over time, which can be achieved using energy envelope extraction or power spectral density estimation, aiming to identify the instantaneous fluctuation trajectory of emotional intensity. Feature fusion network refers to a nonlinear modeling structure used to integrate multi-source acoustic features, which can be implemented using deep neural networks or graph neural networks, aiming to generate a comprehensive acoustic feature representation with context awareness. Emotion recognition model refers to a machine learning architecture for performing emotion classification tasks, which can be implemented using support vector machines or deep neural networks, aiming to map features to emotion type and intensity. Dynamic weighted fusion based on attention mechanism refers to a fusion strategy that adaptively adjusts weights according to feature contribution, which can be implemented using self-attention mechanism or gated attention mechanism, aiming to highlight key emotional cues to improve recognition robustness.

[0107] Specifically, the proposed solution first extracts time-frequency features and analyzes energy distribution in the acoustic feature sequence to obtain high-resolution spectral feature vectors and energy change trajectories, providing a temporal dynamic information foundation for emotion recognition. Subsequently, based on the complementary relationship between prosodic feature parameters and energy change trajectories, a feature fusion network generates a comprehensive acoustic feature representation. This process utilizes the nonlinear modeling capability of the feature fusion network to deeply correlate discrete prosodic features (such as intonation fluctuations and rhythm changes) with continuous energy dynamic information. Finally, the spectral feature vector and the comprehensive acoustic feature representation are input into the emotion recognition model. An attention-based dynamic weighted fusion is used to map the emotion state, adaptively evaluating the contribution of spectral features and comprehensive acoustic features to the current emotion recognition. Feature weights are dynamically adjusted to highlight key emotional cues (such as a surge in high-frequency energy during anger or a decay in low-frequency energy during sadness), thereby outputting an emotion state representation that includes emotion type and intensity. This hierarchical acoustic feature processing and intelligent fusion mechanism accurately captures subtle changes and intensity fluctuations in emotion, effectively overcoming the problems of insufficient feature mining and simplistic fusion mechanisms in traditional methods.

[0108] As a specific implementation method, the scheme of this application is implemented as follows: the acoustic feature sequence is obtained through the Mel frequency cepstral coefficient extraction module, the time-frequency feature extraction is implemented using the short-time Fourier transform module, and the energy distribution analysis is completed through the frame energy calculation module. The feature fusion network is specifically implemented as a deep neural network structure, whose input layer receives prosodic feature parameters and energy change trajectories, the hidden layer performs nonlinear feature association, and the output layer generates a comprehensive acoustic feature representation. The emotion recognition model adopts a recurrent neural network containing an attention mechanism, where the attention mechanism achieves dynamic weighted fusion by calculating feature weights, and finally outputs an emotion state representation.

[0109] Through the above-mentioned approach, this application can accurately capture subtle changes and intensity fluctuations in users' emotions, especially reducing recognition bias in scenarios with noise interference or emotional transitions, providing reliable multi-dimensional emotional basis for subsequent dialogue management, thereby supporting the generation of personalized service responses that are highly adapted to the user's current emotional state.

[0110] Specifically, in some embodiments of this application, multi-level semantic reasoning of visual feature maps and scene understanding information is proposed to obtain visual semantic understanding. However, in its implementation, the basic reasoning method fails to effectively model the spatial relationships and semantic associations between visual entities, resulting in a lack of depth and adaptability in visual semantic understanding in dynamic service scenarios. Specifically, relying solely on conventional visual feature extraction and simple semantic parsing cannot accurately capture the geometric constraints and contextual dependencies between entities in the scene. This makes it difficult for the system to accurately identify visual elements associated with user intent when processing complex image data, thereby affecting the generation of personalized responses in subsequent dialogue management.

[0111] In response, this application further proposes to extract regional features and analyze spatial relationships from visual feature maps to obtain a matrix of regional visual features and spatial relationships.

[0112] Based on scene understanding information and spatial relationship matrix, scene semantic association reasoning is performed through graph neural network to obtain enhanced scene semantic representation.

[0113] The region's visual features and enhanced scene semantic representations are input into the semantic fusion module. Multi-level feature fusion is performed through a cross-modal attention mechanism to obtain visual semantic understanding. The multi-level feature fusion includes selective enhancement of the region's visual features and scene semantic representations based on service context relevance.

[0114] Among these, region feature extraction refers to the local division of a visual feature map and the extraction of feature information. This can be achieved using region proposal networks or sliding window techniques, aiming to obtain detailed visual features of different regions in an image. A spatial relationship matrix is ​​a data structure representing the geometric constraints between visual entities. It can be constructed based on the relative positions and orientations of entities, aiming to explicitly model the spatial layout of objects in a scene. Graph neural networks are neural network models used to process graph-structured data. They can be implemented using graph convolutional networks or graph attention networks, aiming to infer deep semantic relationships between entities through message passing mechanisms between nodes. Cross-modal attention mechanisms are computational mechanisms that establish dynamic associations between features of different modalities. They can be implemented based on multi-head attention mechanisms, aiming to adaptively adjust feature fusion weights according to contextual importance. Service context relevance refers to the degree of relevance between the current dialogue context and visual elements. It can be dynamically calculated by analyzing user emotional states and dialogue history, aiming to achieve context-aware feature selective enhancement.

[0115] As a specific implementation method, when a user uploads a picture of a device malfunction to the intelligent customer service system, the system first extracts regional features from the visual feature map to identify the local features of each component of the device. Simultaneously, spatial relationship analysis is performed to determine the relative positional relationships between the components, forming a spatial relationship matrix. Next, based on scene understanding information (such as a cracked screen) and the spatial relationship matrix, a graph neural network is used to infer the semantic associations of the user's handheld device operations. Finally, in the semantic fusion stage, if user anxiety is detected, targeted visual semantic understanding is generated based on service context relevance and safety-related visual elements (such as "Do not touch the cracked screen"), providing accurate data for subsequent dialogue management.

[0116] The above solution effectively addresses the problem of insufficient spatial relationship and semantic association modeling in basic methods. It can accurately capture the geometric constraints and contextual dependencies between entities in a scene, and improve the depth and adaptability of visual semantic understanding in dynamic service scenarios. This ensures that intelligent customer service can accurately identify visual elements associated with user intent, laying a reliable foundation for generating personalized service responses.

[0117] This application further proposes the following steps for context-aware intent reasoning on semantic representation vectors and intent feature representations: By performing context encoding and sequence modeling on the semantic representation vector, a context-enhanced semantic representation is obtained.

[0118] Based on intent feature representation and memory state features, intent evolution analysis is performed through memory enhancement mechanism to obtain intent evolution trajectory, where memory state features are extracted from multi-layer memory system.

[0119] The context-enhanced semantic representation and intent evolution trajectory are input into the intent reasoning module. Cross-level feature fusion is performed through a multi-head attention mechanism to obtain text intent parsing. Cross-level feature fusion includes dynamic weight allocation of semantic representation and intent evolution trajectory based on the importance of dialogue stages.

[0120] Specifically, semantic representation vectors refer to the vector representation of text data, which can be implemented using embedding vectors generated by pre-trained language models. Their purpose is to preserve the semantic structural features of the original text. Intent feature representation refers to the set of intent-related features extracted from user input text. This can be implemented using convolutional neural networks to extract features from the semantic representation vectors, aiming to highlight key discriminative information for intent recognition. Context encoding and sequence modeling refer to the process of capturing temporal information from the semantic representation vectors. This can be implemented using long short-term memory networks or Transformer encoders, aiming to enhance the context-awareness of semantic representations by modeling the temporal dependencies of dialogue history. Memory state features refer to historical interaction state information extracted from a multi-layered memory system. This can be implemented using feature vectors retrieved from episodic memory, experiential memory, and vector memory, aiming to provide information on user historical interaction patterns and preferences. Memory enhancement mechanisms refer to mechanisms that combine historical memory information for intent analysis. This can be implemented using a neural Turing machine architecture, aiming to integrate historical state information into the current intent parsing process. Intent evolution trajectory refers to the path representation of how user intent changes as the dialogue progresses. It can be implemented using a sequence-to-sequence model to generate a sequence of hidden states, aiming to characterize the continuous evolution of intent. Multi-head attention mechanism refers to the mechanism of parallel computation of multiple attention heads to capture features from different subspaces. It can be implemented using the multi-head self-attention structure in the Transformer model, aiming to achieve efficient interaction between features. Cross-level feature fusion refers to the process of fusing feature information from different levels. It can be achieved by concatenating features and then processing them through a fully connected layer, aiming to integrate semantic representation and historical intent information. Dynamic weight allocation refers to the strategy of adjusting the contribution of features according to the importance of each dialogue stage. It can be implemented using a weight calculation function based on changes in dialogue rounds, aiming to adaptively match the required features at different dialogue stages.

[0121] As a specific implementation method, the solution of this application is implemented as follows: When the intelligent customer service system finds the previously recommended solution too complex when processing user input, a semantic representation vector is first generated using the BERT model. Then, a Transformer encoder is applied to perform context encoding and sequence modeling on the semantic representation vector to obtain a context-enhanced semantic representation that includes the dialogue history. Simultaneously, the user's historical interaction records are extracted from a multi-layered memory system as memory state features. Combined with the current intent feature representation, the evolutionary trajectory of the user's intent from seeking a simple solution to expressing dissatisfaction is analyzed through a memory enhancement mechanism. Finally, the context-enhanced semantic representation and the intent evolution trajectory are input into the intent reasoning module and fused using a multi-head attention mechanism. The weight of the intent evolution trajectory is dynamically increased based on the importance of the current dialogue's problem-solving stage, thereby accurately resolving the user's intent to obtain a simpler solution.

[0122] Through the above technical solution, this application effectively integrates historical state information in a multi-layer memory system, realizes dynamic adaptation to the importance of different dialogue stages, enables intent parsing to accurately capture the continuous evolution and contextual dependence of user intent, and improves the accuracy of subsequent dialogue management.

[0123] In some embodiments of this application, a memory coordinator is proposed to coordinate the scheduling of a multi-layer memory system. However, in its implementation, the scheduling mechanism relies only on basic emotional state representation and text intent parsing for simple invocation, lacking the ability to dynamically perceive changes in emotional intensity and semantic complexity of intent. This results in the inability to adjust the priority of service strategies according to subtle differences in user emotional fluctuations, and the inability to adaptively allocate historical information retrieval resources according to the complexity of intent. Consequently, the generated service strategies and solutions suffer from insufficient scenario coordination and low matching accuracy, ultimately resulting in mechanical and rigid response content that is difficult to meet the dynamic and personalized service needs of users.

[0124] In response, this application further proposes the following technical solution, see [link to technical solution]. Figure 4 Taking the server as the executing entity as an example, the following steps are included.

[0125] 401. Based on the emotional intensity and type of the emotional state representation, the service preference dimension in the contextual memory is dynamically weighted by a memory coordinator to obtain a set of emotionally adapted service strategies. The dynamic weighting includes adjusting the priority of service strategies according to the emotional intensity.

[0126] 402. Based on the intent type and semantic complexity of text intent parsing, a memory coordinator is used to perform cross-level association retrieval of experiential memory and vector memory to obtain a set of intent matching solutions. The cross-level association retrieval includes determining the retrieval ratio of experiential memory and vector memory based on semantic complexity.

[0127] 403. Perform policy consistency verification and priority fusion on the service policy set and solution set to obtain a memory feature set. Policy consistency verification includes evaluating the synergy between service policies and solutions based on service scenario matching degree.

[0128] In practical applications, dynamic weight allocation refers to a mechanism that adjusts the priority of service strategies based on emotional intensity. This can be implemented using threshold judgment based on rule engines or reinforcement learning models, aiming to enable service strategies to respond to user emotional fluctuations in real time. Cross-level related retrieval refers to a method that determines the ratio of experiential memory to vector memory retrieval based on semantic complexity. This can be implemented by dynamically adjusting the retrieval depth through a preset complexity mapping table, aiming to optimize the efficiency of historical information retrieval. Specifically, strategy consistency verification refers to the process of evaluating the synergy between service strategies and solutions based on service scenario matching. This can be implemented using scenario feature vector similarity calculation, aiming to eliminate logical conflicts between strategies and solutions.

[0129] Through the above technical solution, this application effectively solves the problem of rigidity in memory scheduling mechanism, enabling service strategy priority to be adjusted in real time according to user emotional fluctuations, and historical information retrieval resources to be adaptively allocated according to the complexity of intent, thereby improving the scenario synergy and matching accuracy of service strategy and solution, and generating more natural and personalized service responses.

[0130] This application further proposes a set of emotion-adaptive service strategies based on emotion state representation, using emotion intensity and emotion type, and dynamically assigning weights to the service preference dimension in episodic memory through a memory coordinator. This set includes: The emotional state representation is quantified in terms of emotional intensity and classified in terms of emotional type, resulting in emotional intensity values ​​and emotional category labels.

[0131] Based on emotion intensity values ​​and emotion category identifiers, a memory coordinator dynamically assigns weights to the service preference dimension in episodic memory to obtain the preference weight distribution for emotion adaptation. The dynamic weight assignment includes determining the activation level of the preference dimension based on the emotion intensity value through a dynamic weight assignment model.

[0132] Based on the preference weight distribution of emotion adaptation and historical service records in contextual memory, service strategies are generated through a memory coordinator to obtain a set of emotion-adapted service strategies. The service strategy generation includes prioritizing and filtering historical service records according to the preference weight distribution.

[0133] In practical applications, emotion intensity quantification refers to converting the intensity of a user's emotion into a calculable numerical indicator. This can be achieved using energy distribution analysis based on acoustic features or amplitude measurement based on facial expressions, aiming to provide a precise quantitative basis for subsequent weight allocation. Emotion type classification identifies the category label to which a user's emotion belongs. This can be achieved using machine learning-based emotion classifiers or rule-based emotion type determination tables, aiming to distinguish the different impacts of different emotion types on service strategies. Dynamic weight allocation refers to dynamically adjusting the importance weights of each service preference dimension based on real-time emotional states. This can be achieved using a weight calculation module based on fuzzy logic or a dynamic weight generator based on neural networks, aiming to enable the weights of service preference dimensions to adaptively adjust with changes in emotion. The dynamic weight allocation model refers to the algorithmic framework for implementing dynamic weight allocation. This can be achieved using an adaptive model based on reinforcement learning or a rule engine based on decision trees, aiming to accurately determine the activation level of each preference dimension based on the emotion intensity value. The activation level of a preference dimension can represent the intensity of activation of a specific service preference dimension. This can be achieved using continuous values ​​between 0 and 1 or discrete level indicators, aiming to quantify the relative importance of each preference dimension in the current emotional state. Service strategy generation refers to the process of generating specific service strategies based on weight distributions. This can be achieved using strategy matching algorithms based on historical data or strategy synthesis methods based on generative models, aiming to transform the emotion-adapted weight distribution into executable service strategies. Strategy priority ranking and filtering refers to sorting and selecting historical service records according to weight distributions. This can be achieved using weighted scoring-based ranking algorithms or threshold-based filtering mechanisms, aiming to select the service strategy that best matches the current emotional state from historical records.

[0134] As a preferred embodiment, the solution of this application is implemented as follows: When a user exhibits high anger, the intensity of the emotion is first determined to be at a high level of anger and the emotion category is identified as anger through acoustic feature analysis. The memory coordinator invokes a dynamic weight allocation model to automatically reduce the weight of the price sensitivity dimension and increase the weight of the response speed dimension based on the high anger intensity. Based on this preference weight distribution, the fastest service strategies from historical service records are selected from the contextual memory, excluding price-sensitive strategies, and finally generating a set of emotion-adaptive service strategies with rapid response as the core.

[0135] Through the above solution, this application achieves accurate matching between service strategies and user emotional states, avoiding the problem of mechanical call history caused by insufficient emotional adaptation, enabling the generated service response to adapt to the real-time dynamic changes in user emotions, and effectively improving the matching degree of dialogue context and user satisfaction.

[0136] This application further proposes intent types and semantic complexity based on text intent parsing, and obtains a set of intent matching solutions by performing cross-level association retrieval of experiential memory and vector memory through a memory coordinator, including: The intent type is identified and the semantic complexity is evaluated in the text intent parsing to obtain the intent type identifier and complexity level.

[0137] Based on intent type identifiers and complexity levels, a memory coordinator determines the retrieval strategy configuration for experiential memory and vector memory. The retrieval strategy configuration includes determining the retrieval depth and retrieval range for the two types of memory based on complexity levels.

[0138] Based on the retrieval strategy configuration, the structured case retrieval of experience memory and the semantic similarity retrieval of vector memory are performed in parallel through the memory coordinator, and the two types of retrieval results are fused together to obtain a set of solutions for intent matching.

[0139] Among these, intent type identification refers to the categorized abstract representation of user intent, which can be implemented using deep learning-based classification models or rule engines. Its purpose is to accurately distinguish the demand characteristics under different service scenarios such as queries, complaints, or inquiries. Complexity level can be understood as a hierarchical quantitative indicator of semantic complexity, specifically using semantic density analysis methods based on context windows. Its purpose is to provide a refined control basis for retrieval strategies. Retrieval depth refers to the hierarchical depth of retrieval in experience memory, which can be dynamically adjusted based on the branch paths of the case tree, aiming to adapt to the information mining needs under different complexity scenarios. Retrieval scope refers to the semantic coverage breadth of retrieval in vector memory, which can be flexibly defined based on similarity distribution curves, aiming to balance retrieval efficiency and result coverage. Relevance fusion refers to the complementary integration of structured cases and semantic fragments at the semantic level. This can be achieved using cross-modal attention mechanisms or graph structure association analysis methods, aiming to generate a logically coherent and context-adaptive set of solutions.

[0140] As a specific implementation method, the solution of this application is implemented as follows: When a user enters "Why hasn't my order been shipped yet?", the intent of the text is identified as a complaint intent with a semantic complexity assessment of medium. Based on this, the memory coordinator configures the retrieval depth of the experience memory to be medium, limiting the retrieval scope to recent high-frequency complaint cases, and the retrieval depth of the vector memory to be deeper, focusing the retrieval scope on highly semantically relevant segments. Subsequently, structured case retrieval (extracting the solution framework of historical complaints) and semantic similarity retrieval (obtaining semantic segments related to shipping delays) are performed in parallel in the experience memory. Finally, the solution steps in the case are weighted and fused with the semantic segments through semantic relevance calculation. For example, the semantic segments from the vector memory are used to supplement the missing timeliness explanations in the experience memory cases, generating a complete set of solutions for the complaint scenario.

[0141] Through the above solution, this application effectively solves the problem of rigid retrieval strategy configuration, realizes dynamic differentiated retrieval based on intent type and semantic complexity, improves the flexibility of cross-level related retrieval and the matching accuracy of solutions, and enables the intelligent customer service system to adaptively adjust the memory scheduling mechanism according to the diversity of user intent, thereby generating service responses that are more in line with complex semantic context.

[0142] This application further proposes to perform policy consistency verification and priority fusion on the service policy set and solution set to obtain a memory feature set, including: The service scenario synergy is evaluated between each service strategy in the service strategy set and each solution in the solution set, and the synergy distribution between strategies and solutions is obtained.

[0143] The suitability of strategies and schemes is verified by analyzing the distribution of synergy and the representation of emotional state, resulting in a verified combination of strategies and schemes. The suitability verification includes dynamically adjusting the synergy threshold based on the intensity of emotion.

[0144] The validated strategy combination and text intent parsing are fused by strategy priority to obtain a memory feature set. The strategy priority fusion includes a dual weight allocation mechanism based on semantic complexity and emotional intensity.

[0145] Service scenario synergy assessment refers to the technical feature that quantifies the matching degree of service strategies and solutions in specific service scenarios. It can be implemented using semantic correlation calculation based on graph matching or dynamic Bayesian network modeling. The aim is to precisely identify potential logical contradictions between service strategies and solutions, avoiding misjudgments caused by relying solely on keyword matching. Adaptability verification can be understood as a technical feature that dynamically calibrates the synergy distribution based on emotional state representation. Specifically, dynamically adjusting the synergy threshold based on emotional intensity can be achieved through nonlinear function mapping. For example, when the emotional intensity value exceeds a preset threshold, the synergy threshold is linearly reduced. The purpose is to allow the verification mechanism to automatically adjust its strictness according to user emotional fluctuations, preventing response mismatches in emotionally sensitive scenarios. Strategy priority fusion refers to the technical feature that dynamically assigns weights to the verified strategy solution combination. The dual weight allocation mechanism based on semantic complexity and emotional intensity can be specifically implemented using a weighted summation model combined with an adaptive decay factor. The purpose is to ensure that the generation of the memory feature set simultaneously responds to changes in intent complexity and emotional state evolution, ensuring that the fusion result is highly consistent with the current dialogue context.

[0146] As a preferred embodiment, the solution of this application is implemented as follows: When a user expresses a need for a replacement through voice complaint and the emotional intensity value is detected as high (e.g., tone frequency exceeds 250Hz and speech rate increases to over 200 words per minute), the synergy threshold is dynamically adjusted from the baseline value of 0.65 to 0.82, and the matching combination of service strategies and solutions is strictly screened. Simultaneously, considering the dual semantic complexity of product quality concerns and timeliness requirements implicit in the user's intent, the weight ratio of experienced memory cases is increased to 70%, and the soothing weight of tone elements is strengthened based on high emotional intensity characteristics. The final generated strategy solution combination prioritizes the use of historically successfully handled high-emotion complaint cases and integrates immediate emotional soothing language to form a highly context-appropriate service response.

[0147] Through the above solution, this application effectively solves the problems of fixed threshold and single fusion mechanism in policy consistency verification, enabling the generation of memory feature set to dynamically adapt to the dual changes of user emotional fluctuations and intent complexity, improving the contextual adaptability and response personalization of dialogue management, and ensuring that the service response generated by intelligent customer service in complex interaction scenarios is accurately matched with the user's real-time needs.

[0148] This application further proposes the following technical solution, taking the server as the execution subject as an example, including the following steps.

[0149] Based on scene entities and relationships in visual semantic understanding, a decision fusion engine is used to dynamically resolve conflicts and enhance consistency between procedural memory operation records and contextual memory user preferences in the memory feature set, resulting in a preference-enhanced operation sequence. Dynamic conflict resolution includes detecting and correcting logical contradictions between operation records and user preferences in the current service scenario.

[0150] Based on the semantic context of text intent parsing, a context-aware fusion scheme is obtained by semantically guiding complementary fusion of experiential memory case schemes and vector memory semantic fragments in the memory feature set through a fusion decision-maker. The semantically guided complementary fusion includes using the semantic fragments of vector memory to fill the semantic gaps of experiential memory cases in the current dialogue context.

[0151] Based on emotion state representation, a fusion decision-maker is used to perform multiple rounds of iterative optimization on the preference-enhanced operation sequence and the context-aware fusion scheme. The fusion decision result is output with emotion stability as a constraint. The multiple rounds of iterative optimization include dynamically adjusting the combination strategy of operation sequence and fusion scheme according to the changes in emotion intensity.

[0152] Dynamic conflict resolution refers to detecting and correcting logical contradictions between program memory operation records and contextual memory user preferences in the current service scenario. This can be achieved using a conflict detection engine based on a scenario constraint rule base or a graph structure matching algorithm. Its purpose is to ensure that the operation sequence conforms to both the user's historical preferences and the real-time visual scene, avoiding logical contradictions caused by mismatches in entity spatial relationships. Semantic-guided complementary fusion can be understood as using semantic fragments from vector memory to fill semantic gaps in the current dialogue context of experiential memory cases. This can be achieved using semantic vector space mapping technology or a context-aware fragment retrieval mechanism. Its purpose is to overcome semantic incoherence caused by missing context in experiential memory cases, enabling the fusion scheme to dynamically adapt to the semantic evolution of the dialogue. Multi-round iterative optimization specifically involves dynamically adjusting the combination strategy of the operation sequence and fusion scheme based on changes in emotional intensity. This can be achieved using a gradient optimizer based on emotional feedback or a dynamic weight allocation model. Its purpose is to select the optimal decision combination under the constraint of emotional stability, avoiding decision results that cause fluctuations in user emotions.

[0153] As a specific implementation method, the solution of this application is implemented as follows: When a user reports an electronic device malfunction and uploads a picture of the damaged device through the intelligent customer service system, the device model, location of damage, and severity in the picture are first identified based on visual semantic understanding. Then, a dynamic conflict resolution process is performed between the standard repair process in the program memory and the user's preferences in the contextual memory (such as a preference for quick replacement over repair). If the visual semantics indicate that the device screen is completely shattered, the operation sequence is modified to skip the detection step and directly trigger the replacement process. Simultaneously, based on the semantic context of logistics delays mentioned in the user's current dialogue, relevant semantic fragments are retrieved from the vector memory to fill the gaps in the service timeliness descriptions in the experiential memory cases, forming a complete replacement solution including expedited logistics options. Finally, considering that the user's emotional intensity is high, the combination strategy of the operation sequence and the fusion solution is dynamically adjusted, simplifying the replacement steps and prioritizing the insertion of emotional reassurance statements to ensure the emotional stability of the output decision.

[0154] Through the above solution, this application effectively solves the problems of scene logic conflict, semantic incoherence and emotional instability in the process of multi-source memory integration, so that the generated fusion decision results can accurately adapt to real-time visual scene constraints, dynamically fill semantic gaps in the dialogue context, and maintain stable output in the emotional dimension, thereby improving the intelligent customer service system's ability to respond to and adapt to users' dynamic needs in complex interactive scenarios.

[0155] This application further proposes a method based on scene entities and relationships in visual semantic understanding. By fusing a decision-maker, it dynamically resolves conflicts and enhances consistency between procedural memory operation records and episodic memory user preferences in the memory feature set, resulting in a preference-enhancing operation sequence, including: The program's memory operation records are analyzed step-by-step to obtain the sequence of operation steps and the execution dependencies between steps.

[0156] The sequence of operation steps is matched with the scene entity relationships in visual semantic understanding. When the execution order of operation steps in the sequence of operation steps is found to be inconsistent with the scene relationship constraints indicated by the scene entity relationships, it is marked as a scene logic conflict point. The consistency matching analysis includes evaluating the feasibility of operation steps based on entity spatial relationships.

[0157] Based on the emotional intensity represented by the logical conflict points of the scene and the emotional state, the user preferences of the context memory are prioritized and recalibrated by the fusion decision-maker to generate conflict resolution strategies.

[0158] Based on conflict resolution strategies and service context profiling, the step sequence in the program memory operation record is reconstructed and the parameters are optimized to output a preference-enhanced operation sequence.

[0159] Among them, program memory operation records refer to standardized service process records stored in program memory, which can be implemented using structured task trees or state machine models, aiming to ensure the standardization and traceability of service operations. Scene logic conflict points refer to specific locations where operational steps contradict visual scene constraints, which can be identified based on spatial geometric relationship calculations or semantic rule engines, aiming to accurately locate logical inconsistencies in the operation sequence. Emotional state representation refers to the numerical index quantifying the degree of user emotional fluctuations, which can be represented using acoustic feature energy values ​​or facial expression action unit intensity, aiming to provide a dynamic adjustment basis for conflict resolution. Service context profiles can be understood as dynamically constructed user demand representation models, which can be implemented based on real-time feature vectors from multimodal data fusion, aiming to capture changes in user preference sensitivity during the current dialogue phase.

[0160] As a specific implementation method, when a user uploads an image showing a kitchen scene and expresses anxiety via voice, the program's memory of cleaning the stove is analyzed to obtain a step-by-step sequence including taking cleaning agent → spraying → wiping. This sequence is matched with the entity relationship between the stove and cleaning agent, identified in visual semantic understanding, indicating they are located in different rooms. A conflict is found between the step of taking cleaning agent and spatial constraints, which is marked as a scene logic conflict point. Based on the user's high anxiety intensity, the core preference for quick problem-solving in the context memory is given higher weight, generating a conflict resolution strategy that prioritizes calling the backup cleaning agent. Combining the user's preference for simple operations in the service context profile, the original step sequence is reconstructed into wiping directly with the nearest disinfectant wipe, outputting an operation sequence that better matches the current scene and emotional state.

[0161] The above approach accurately identifies logical contradictions between operational steps and the visual scene, avoiding the generation of invalid steps detached from the actual environment. Simultaneously, it dynamically adjusts preference weights based on the user's emotional state, ensuring that the conflict resolution strategy addresses both logical inconsistencies and emotional needs. The final output sequence of operations conforms to both physical scene constraints and genuine user preferences, improving the scenario adaptability and user satisfaction of the intelligent customer service response.

[0162] This application further proposes a semantic context based on text intent parsing. Through a fusion decision-maker, it performs complementary fusion of experiential memory case schemes and vector memory semantic fragments in the memory feature set, resulting in a context-aware fusion scheme, including: The task structure of the experience memory case solution is analyzed and the solution features are extracted to obtain the case task framework and core solution elements.

[0163] Based on the semantic context of text intent parsing, the semantic integrity of the case task framework is evaluated, and the semantic gaps of the experience memory case solutions in the current dialogue context are identified.

[0164] Based on the identified semantic gaps, supplementary semantic segments that match the current semantic context are retrieved from the vector memory semantic segments, and the fusion weight of the supplementary semantic segments is determined by semantic relevance calculation.

[0165] Based on the fusion weights, supplementary semantic fragments are dynamically and weightedly fused with the core solution elements of experiential memory case schemes to generate context-aware fusion schemes.

[0166] In practical applications, task structure parsing refers to the logical hierarchical decomposition of experience memory case schemes. This can be achieved using a structured parser based on graph neural networks or a rule-template-based parser. The goal is to transform unstructured historical cases into a quantifiable task framework, avoiding the contextual disconnect problem caused by traditional keyword matching. The case task framework can be understood as the logical skeleton of the experience memory case scheme, specifically represented as a directed graph of task step sequences and step dependencies, providing a structured foundation for semantic integrity assessment. Semantic gaps refer to missing parts of the experience memory case scheme that do not match the current dialogue semantic context. These can be identified using difference measurement methods based on context embedding vectors or missing part detection algorithms based on semantic role labeling, aiming to accurately locate discontinuities in the scheme that are inconsistent with the user's real-time intent. In practical applications, semantic relevance calculation refers to quantifying the logical fit between supplementary semantic fragments and the current context. This can be achieved using cosine similarity calculation or a semantic matching model based on attention mechanisms, aiming to ensure the coherence of supplementary content with the dialogue context. Dynamic weighted fusion can be understood as a fusion mechanism that adaptively adjusts the contribution ratio of elements based on semantic relevance. Specifically, it can adopt a weighted average algorithm with adjustable weight coefficients or a fusion strategy optimizer based on reinforcement learning. The purpose is to balance the reliability of experience memory with the real-time adaptability of vector memory.

[0167] As a specific implementation method, the solution of this application is implemented as follows: When a user consults about troubleshooting smart devices, the experience memory case solution stores the standard fault handling process, but lacks specific operation steps for new model devices. The fusion decision-maker first performs task structure parsing on the case solution, extracting the core solution elements including device model identification, fault phenomenon diagnosis, and solution execution. Then, based on the semantic context of the user's current dialogue that "my new model cannot connect to Wi-Fi," the semantic integrity of the case task framework is evaluated, identifying semantic gaps in the compatibility verification step for new model devices. Next, semantic fragments related to Wi-Fi connection and new model are retrieved from the vector memory, and high-weight fragments are determined by calculating their semantic relevance to the current context. Finally, these fragments are dynamically weighted and fused with the core solution elements to generate a complete troubleshooting solution including new model verification steps.

[0168] Through the above technical solutions, this application effectively solves the problem of the integrity of experience memory case solutions in dynamic dialogues due to insufficient semantic coverage, avoids the semantic gap problem caused by relying solely on fixed retrieval mechanisms in traditional methods, and improves the logical coherence and context adaptability of the fusion solution in complex dialogue scenarios, thereby providing reliable technical support for generating highly personalized service responses.

[0169] This application further proposes a process based on emotional state representation, which involves iteratively optimizing the fusion scheme of preference-enhancing operation sequences and context-awareness through a fusion decision-maker, and outputting the fusion decision result with emotional stability as a constraint. The steps include: The weights for assessing emotional stability are determined based on the intensity and type of emotion as represented by the emotional state.

[0170] Based on the emotion stability assessment weight, the combination evaluation of the preference enhancement operation sequence and the context-aware fusion scheme is carried out to obtain multiple candidate decision combinations and the predicted emotion impact value of each candidate decision combination.

[0171] Based on the trend of emotion intensity change in emotion state representation, a dynamic threshold for emotion stability is determined.

[0172] Multiple rounds of emotion stability verification are performed on candidate decision combinations. In each iteration, based on the comparison between the rate of change of emotion intensity and the dynamic threshold of emotion stability, multiple reference decision combinations that have reached the preset stability are selected.

[0173] The weights of multiple reference decision combinations are optimized and adjusted by a fusion decision maker. Specifically, the weight ratio of the operation sequence and the fusion scheme in the combination is dynamically adjusted according to the emotion type characteristics of the emotion state representation, and the predicted value of emotion impact is recalculated.

[0174] When the rate of change of the predicted value of the emotion impact in multiple consecutive iterations is less than the preset convergence threshold, the reference decision combination with the best emotion stability is selected from multiple reference decision combinations as the fusion decision result.

[0175] The emotional stability assessment weight refers to the assessment parameters dynamically adjusted based on emotional state. It can be implemented using a weighted coefficient matrix based on emotional intensity and emotional type, specifically quantified through a pre-defined emotional sensitivity mapping table. Its purpose is to accurately reflect the sensitivity differences of different emotional states during the assessment process, avoiding optimization direction deviations caused by uniform assessment. The emotional impact prediction value can be understood as a quantitative indicator of the potential emotional changes in users triggered by candidate decision combinations. It can be calculated using an emotional prediction model based on a long short-term memory network, trained using historical interaction data. Its purpose is to provide a quantifiable dynamic reference benchmark for decision combination selection. The emotional stability dynamic threshold refers to a verification standard that adaptively adjusts with the emotional evolution trend. It can be dynamically generated based on an exponential decay function of the emotional intensity change rate. Its purpose is to enable the verification process to respond to rising or falling emotional trends, avoiding misjudgments in fluctuating scenarios using a fixed threshold. The emotional intensity change rate is specifically the rate of change of emotional intensity over time. It can be obtained by performing a first-order difference operation on continuous emotional state representations. Its purpose is to reflect the dynamics of emotional evolution in real time, providing a basis for threshold adjustment. Dynamic weighting adjustment refers to adjusting the contribution ratio of the operation sequence and the fusion scheme in real time based on emotion type characteristics. This can be achieved using a lookup mechanism based on emotion type-weight mapping rules, aiming to achieve refined emotion adaptation in decision combinations. The convergence threshold can be understood as a preset rate of change threshold for determining whether the iteration process terminates. It can be dynamically adjusted based on dialogue stage characteristics, aiming to ensure timely output of the optimal decision when emotion fluctuations are mild, avoiding response delays caused by excessive iteration.

[0176] Specifically, the proposed solution determines the weights for assessing emotional stability by using the emotional intensity value and emotional type as representations of emotional states. This ensures that the weight allocation precisely matches the sensitivity differences of different emotional states; for example, users with high levels of anger are given higher stability weights. Based on these dynamic weights, the operation sequences and fusion schemes are combined for evaluation, quantifying the predicted emotional impact of each candidate decision combination and establishing a dynamic reference benchmark closely related to real-time emotional characteristics. Simultaneously, a dynamic threshold for emotional stability is determined based on the trend of emotional intensity changes, allowing the verification standard to adaptively adjust with emotional evolution; for example, the threshold is tightened to increase verification rigor when emotions rapidly deteriorate. During multiple iterations, reference decision combinations are selected based on real-time comparisons of the rate of change of emotional intensity and the dynamic threshold, gradually filtering out the risk of emotional instability through an iterative screening mechanism. In particular, during the weight optimization and adjustment phase, the weight ratio of the operation sequences and fusion schemes is dynamically adjusted according to emotional type characteristics; for example, the weight of explanatory elements in anxiety-enhancing schemes and the weight of smoothness elements in pleasure-enhancing schemes. Finally, the convergence state is determined by the rate of change of the predicted value of emotion impact in multiple consecutive iterations. When the rate of change is lower than a preset threshold, the decision combination with the best emotional stability is output to ensure the long-term stability of the decision results in a dynamic emotional environment.

[0177] As a specific implementation method, the solution of this application is implemented as follows: When a user exhibits anxiety, the emotional state representation outputs a high anxiety intensity value and an anxiety type identifier. The fusion decision-maker first determines a higher emotional stability assessment weight based on this identifier and evaluates candidate decision combinations to obtain predicted emotional impact values. Simultaneously, based on the trend of rapidly increasing emotional intensity, the dynamic threshold is tightened to improve the verification rigor. During the iteration process, if the rate of change in emotional intensity exceeds the dynamic threshold, a reference decision combination that can effectively alleviate anxiety is selected. Subsequently, based on the anxiety type characteristics, the weight ratio of the explanatory elements of the solution is increased, and the predicted values ​​are recalculated. When the rate of change in predicted values ​​is lower than a preset threshold for multiple consecutive iterations, the decision combination with the optimal emotional stability is selected as the fusion decision result, and a service response containing detailed explanations is output.

[0178] Through the above technical solutions, this application effectively solves the problem of rigid emotional stability constraints in multi-round iterative optimization. Dynamically adjusted emotional stability evaluation weights ensure that decision evaluations accurately match user emotional sensitivity, avoiding optimization bias caused by uniform evaluations. Adaptive dynamic emotional stability thresholds enable the verification process to respond to emotional evolution trends, significantly reducing threshold rigidity. Dynamic adjustment of weight ratios based on emotional type characteristics achieves refined adaptation of decision combinations, overcoming the defects of fixed weight ratios. Furthermore, the convergence judgment mechanism based on the rate of change ensures timely output of the optimal decision when emotional fluctuations are mild, solving the problem of inaccurate convergence judgments. Ultimately, the generated service response can effectively maintain emotional stability in scenarios with fluctuating user emotions, improving the interactive experience of intelligent customer service.

[0179] Specifically, in some of the embodiments described above in this application, a personalized service response corresponding to a dialogue request is generated through an intelligent decision engine based on the fusion decision results and service context profile. However, in its implementation, the response generation lacks a refined evaluation mechanism for the dynamic matching degree between the decision results and the service context, as well as the collaborative generation and natural integration of multi-dimensional elements such as response content, tone, and action suggestions. This results in the generated response potentially deviating from the user's immediate needs due to insufficient adaptability, or appearing stiff and incoherent due to logical breaks between elements, failing to fully meet the user's expectations for a highly natural and contextualized service experience.

[0180] In response, this application further proposes the following technical solution, taking the server as the executing entity as an example, including the following steps.

[0181] The context fit of the integrated decision results and service context profiles is evaluated to obtain the context fit index. The context fit evaluation includes calculating the matching degree of the decision results based on service needs and preference characteristics.

[0182] Based on the adaptability index and dynamic demand characteristics in the service context profile, the intelligent decision engine generates elements to obtain multi-dimensional response elements, which include content elements, tone elements, and action suggestion elements.

[0183] By integrating and optimizing the naturalness of multi-dimensional response elements, personalized service responses are obtained. The integration of response elements includes the consistent integration of content, tone, and action suggestions based on the principle of dialogue coherence.

[0184] Context fit assessment refers to the process of quantitatively evaluating the degree of matching between decision results and service context. This can be achieved using rule-based matching algorithms or lightweight neural network models. The aim is to correlate abstract decision results with the dynamic characteristics of the user's current service context, avoiding matching biases caused by ignoring subtle changes in the context. The fit index can be understood as a quantitative representation of the matching degree between the decision and the context. It can be calculated based on the matching degree of service needs and preference characteristics, aiming to provide a dynamic adjustment basis for subsequent response generation. Multi-dimensional response elements specifically refer to the response components including content elements, tone elements, and action suggestion elements. Content elements can be textual information of service content, tone elements can be the intensity and type of emotional expression, and action suggestion elements can be suggestions of operational steps. The aim is to achieve multi-dimensional adaptive generation of the response. Response element fusion refers to the process of consistent integration of multi-dimensional response elements based on the principle of dialogue coherence. This can be achieved using semantic consistency verification and logical structure verification, aiming to eliminate logical breaks between elements. Naturalness optimization can be understood as improving the fluency of response elements. It can be based on the dialogue generation model to personalize the expression mode, with the aim of outputting a highly natural service response.

[0185] The specific implementation of this application's solution is as follows: When a user inquires about product returns through intelligent customer service, the system first calculates a contextual fit index based on the fusion decision-making results (including return policies and user historical preferences) and the service context profile (including the user's current emotions and intensity of needs). If the fit index is high, indicating a good match between the decision-making results and the user's context, detailed content elements (such as a return process description), gentle tone elements (such as using reassuring language), and clear action suggestion elements (such as providing a return link) are generated. Subsequently, these elements are fused to ensure that the tone is consistent with the content; for example, maintaining a patient tone when explaining complex processes and naturally embedding action suggestions into the dialogue. Finally, a personalized service response is output, such as, "We understand your urgent need for a return; we have prepared a fast track for you. Please click this link to complete the operation. If you have any questions during the process, please let us know."

[0186] In some of the embodiments described above in this application, a context fit assessment is proposed to quantify the degree of matching between the decision result and the service context. However, in its implementation, the assessment method adopts a fixed weight mechanism and fails to dynamically adjust the weight allocation of each matching degree according to the characteristics of the dialogue stage. It also fails to introduce the confidence of the service type identifier as a reliability consideration, which makes the fit index unable to accurately reflect the dynamic changes of the real-time dialogue context, thereby causing the generated personalized service response to deviate from the user's actual needs.

[0187] To address this, this application further proposes an assessment of the contextual adaptability of the fusion decision-making results and service context profiles, resulting in an adaptability index between the decision and the context, including: Structured feature extraction is performed on the fusion decision results to obtain the execution complexity including the number of operation steps, the service type identifier based on the service classification model, and the service timeliness level determined according to business rules.

[0188] A multi-dimensional quantitative analysis of the service context profile is conducted to obtain the demand intensity value based on the frequency of user input, the expected service type obtained through semantic analysis, and the preference sensitivity calculated based on historical interaction data.

[0189] Based on the matching degree between execution complexity and demand intensity value, the consistency between service type identifier and expected service type, and the coordination between service timeliness level and preference sensitivity, a dynamic weighted evaluation model is used to determine the context fit index. The dynamic weighted evaluation model adjusts the weight of each matching degree in real time based on the characteristics of the dialogue stage, and introduces the confidence level of service type identifier as a reliability factor in the evaluation process.

[0190] Structured feature extraction refers to extracting quantifiable structural attributes from the fused decision results. This can be achieved by counting the number of operation steps, calling pre-trained service classification models, and parsing business rule bases. Execution complexity can be a metric for the number of operation steps, implemented using step counters or process complexity algorithms, aiming to objectively reflect the ease or difficulty of decision execution. Service type identification refers to category labels generated based on the service classification model, implemented using neural network classifiers or rule engines, aiming to overcome the limitations of keyword matching and achieve a deep semantic understanding of service categories. Service timeliness level refers to the time sensitivity level determined according to business rules, implemented using rule matching engines or timeliness assessment matrices, aiming to ensure that timeliness assessments comply with industry standards. Multi-dimensional quantitative analysis refers to the numerical characterization of service context profiles, implemented using time series analysis, semantic parsing modules, and historical data mining algorithms. Demand intensity value refers to a quantitative indicator based on user input frequency, implemented using interaction frequency counters or behavioral pattern recognition models, aiming to dynamically capture users' urgent needs. Expected service type refers to the category of user expectations obtained through semantic analysis. This can be implemented using a context-aware intent classifier or a semantic similarity calculation module, aiming to deeply analyze implicit user expectations. Preference sensitivity refers to the personalized sensitivity calculated based on historical interaction data. This can be implemented using a historical trajectory analysis model or a sensitivity scoring algorithm, aiming to incorporate historical user behavior characteristics. The dynamic weighted evaluation model is an evaluation mechanism that dynamically adjusts weights based on dialogue stage characteristics. This can be implemented using a stage-aware weight allocator or an adaptive weighted network, aiming to dynamically align the evaluation mechanism with the dialogue evolution. The confidence level of the service type identifier is a quantitative value of the reliability of the service type identification result. This can be implemented using a confidence scoring module or a probability calibration algorithm, aiming to avoid low-reliability data interfering with the evaluation results.

[0191] As a specific implementation method, the solution of this application is implemented as follows: During the dialogue between the user and the intelligent customer service, when the user continuously and rapidly inputs multiple messages, the interaction frequency counter detects a significant increase in the user's input frequency, thereby calculating a higher demand intensity value. Simultaneously, the service classification model analyzes the fusion decision result and identifies the service type as an account query, but the confidence score module outputs a confidence level of only 0.65 for this identifier, lower than the preset threshold of 0.7. At this point, the dynamic weighted evaluation model, based on the characteristics of the current early stage of the dialogue, increases the weight of the demand intensity matching degree to 0.5, while automatically lowering its consistency weight to 0.2 due to the low confidence level of the service type identifier. When calculating the context fit index, the model prioritizes the matching relationship between demand intensity and execution complexity, weakening the influence of the service type identifier, thereby generating a fit index that better reflects the user's urgent needs.

[0192] Through the above technical solution, this application enables the context fit index to dynamically adjust the weight allocation according to the characteristics of the dialogue stage, and filters low-quality data through the confidence reliability factor of the service type identifier, thereby accurately reflecting the dynamic changes of the real-time dialogue context and effectively avoiding the problem of personalized service response deviating from the user's actual needs.

[0193] This application further proposes a process for generating multi-dimensional response elements based on adaptability indicators and dynamic demand characteristics using an intelligent decision engine, including the following steps: Based on the adaptability index and dynamic demand characteristics, an intelligent decision engine generates initial response elements that include service content, tone intensity, and action suggestions.

[0194] Multi-dimensional feature extraction is performed on the initial response elements to obtain semantic features, sentiment features, and logical structure features.

[0195] Based on semantic features, sentiment features, and logical structure features, a smart decision engine is used for collaborative optimization to obtain multi-dimensional response elements. The collaborative optimization process includes calibrating the semantic features for sentiment consistency based on sentiment features, and verifying the rationality of action suggestions based on logical structure features.

[0196] Specifically, the fit index refers to the quantitative indicator of the degree of matching between the decision and the context. It can be implemented using a dynamic weighted evaluation model, which adjusts the matching weights in real time based on the characteristics of the dialogue stage. Dynamic demand characteristics refer to the set of demand attributes that change in real time in the service context profile, which can be quantified in multiple dimensions based on user input frequency and semantic analysis. The intelligent decision engine refers to the core processing module for generating and optimizing execution elements, which can be implemented using a deep learning model or a rule reasoning engine. Initial response elements refer to the combination of initially generated service content, tone intensity, and action suggestions, which can be generated based on a context-aware generative model or a predefined rule base. Multi-dimensional feature extraction refers to the process of deconstructing semantic features, sentiment features, and logical structure features from the initial response elements, which can be achieved using feature extraction algorithms such as the Transformer architecture to achieve deep representation of semantic features. Collaborative optimization processing refers to the processing mechanism for consistency calibration and verification of extracted features, which can be achieved using optimization algorithms such as reinforcement learning to dynamically adjust weights to ensure the matching of semantics and sentiment and the logical feasibility of action suggestions.

[0197] As a specific implementation method, when a user expresses anxiety about a product malfunction via voice, the intelligent decision engine generates initial response elements based on adaptability indicators and dynamic demand characteristics. These elements include service content (troubleshooting steps), tone intensity (reassuring), and action suggestions (contacting customer service). Subsequently, multi-dimensional feature extraction is performed on the initial response elements to obtain semantic features (use of technical terminology), emotional features (anxiety level), and logical structure features (step order). Based on the emotional features, the semantic features are calibrated for emotional consistency, simplifying technical terms into more colloquial expressions. Simultaneously, based on the logical structure features, the rationality of the action suggestions is verified, ensuring that the troubleshooting steps conform to the spatial constraints of the product structure.

[0198] Through the above technical solution, this application effectively solves the problem of lack of in-depth correlation analysis of initial response elements, ensuring the overall consistency and rationality of service content, tone intensity and action suggestions in complex dialogue situations, thereby improving the accuracy of personalized service response and user experience.

[0199] This application further proposes to fuse and optimize the naturalness of multi-dimensional response elements to obtain personalized service responses, including: Semantic consistency is verified on multi-dimensional response elements to obtain semantic consistency scores and sentiment consistency scores.

[0200] Based on semantic consistency score and sentiment consistency score, the fusion degree of multi-dimensional response elements is optimized to obtain an optimized set of response elements.

[0201] Based on the optimized set of response elements and service context profile, a personalized service response is generated through a dialogue generation model. The dialogue generation model personalizes the expression of the response content based on user preference features.

[0202] Semantic consistency verification refers to the process of quantitatively analyzing the inherent logical connections between content elements, tone elements, and action suggestion elements. This can be achieved using rule-based logical reasoning engines or deep learning-based semantic matching models, aiming to accurately identify potential logical contradictions or emotional biases among elements. Semantic consistency scoring and emotional consistency scoring are quantitative indicators representing the logical coherence of content and the coordination of emotional expression, respectively. These can be achieved using normalized values ​​or discrete-level scoring, aiming to provide quantifiable decision-making basis for subsequent optimization. Fusion optimization refers to the process of dynamically adjusting the weight allocation and combination strategies of each response element based on the scoring results. This can be achieved using weighted average fusion algorithms or reinforcement learning-based optimization strategies, aiming to ensure the elimination of content conflicts at the semantic level and the achievement of coordination and unity at the emotional level. The dialogue generation model is a generative model used to transform optimized response elements into natural language expressions. It can be implemented using a pre-trained language model based on the Transformer architecture, aiming to generate fluent and natural dialogue content. User preference features refer to feature vectors extracted from service context profiles that reflect users' personalized expression needs. These features can be implemented using statistical features of historical interaction data or encoded representations of real-time emotional states, with the aim of dynamically matching response content with users' expression habits.

[0203] Specifically, this application first performs semantic consistency verification on multi-dimensional response elements. By analyzing the inherent relationships between content elements, tone elements, and action suggestion elements, it quantitatively assesses the degree of matching between semantic logic and emotional expression, obtaining semantic consistency scores and emotional consistency scores. Subsequently, based on these scores, the response elements are optimized for integration. The weight allocation and combination strategies of each element are dynamically adjusted according to the scoring results to eliminate potential logical contradictions and emotional biases, forming an optimized set of response elements. Finally, the optimized set of response elements is combined with user preference features from the service context profile, and a personalized service response is generated through a dialogue generation model. This model personalizes the expression style according to user preference features, ensuring that the response content is logically consistent, emotionally coordinated, and meets the user's personalized expression needs.

[0204] As a preferred embodiment, the solution of this application is implemented as follows: When the intelligent customer service system handles a user's complaint about order delays, it first performs semantic consistency verification on content elements (such as solution description), tone elements (such as apology intensity), and action suggestion elements (such as compensation measures). A pre-trained semantic matching model calculates the correlation between the content description and tone intensity, resulting in a semantic consistency score of 0.82. The sentiment analysis module evaluates the emotional consistency between tone and suggestion, resulting in a sentiment consistency score of 0.75. Based on these scores, the weights of tone elements are dynamically adjusted, optimizing the apology intensity from "very sorry" to "deeply sorry," and adjusting the compensation measure from providing coupons to prioritizing expedited delivery, thus raising the sentiment consistency score to 0.90. Subsequently, the optimized elements are input into a dialogue generation model based on the GPT architecture. This model combines user preference features from the service context profile (such as historical interactions showing a preference for concise expressions) to generate a response of "We deeply apologize and have prioritized expedited delivery for you," avoiding conflicts between lengthy explanations and emotional tone.

[0205] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0206] Through the above scheme, this application can deeply and quantitatively evaluate the intrinsic matching degree between semantic logic and emotional expression, effectively identify and eliminate the implicit conflict between content elements and emotional tone, and dynamically optimize the response expression method according to the user's real-time preference characteristics, thereby improving the natural fluency and personalized experience of the dialogue response, and making the response generated by the intelligent customer service more in line with the user's current psychological state and expression habits.

[0207] This application provides an intelligent customer service dialogue management system integrating multi-layer memory, comprising: The perception module is used to respond to dialogue requests with intelligent customer service by performing multimodal context perception on voice signals, image data and text data from users to obtain emotional state representation, visual semantic understanding and text intent parsing.

[0208] The scheduling module is used to coordinate the multi-layer memory system based on emotional state representation and text intent parsing through a memory coordinator to obtain a set of memory features that match the current dialogue context. Coordinated scheduling includes activating personalized service strategies in contextual memory based on emotional state representation and retrieving relevant historical solutions from experiential memory and vector memory based on text intent parsing.

[0209] The integration module is used to integrate multi-source memories based on memory feature sets and visual semantic understanding through a fusion decision-maker to obtain a fusion decision result adapted to the current dialogue context. Multi-source memory integration includes consistency verification of operation records in program memory and user preferences in contextual memory, as well as correlation fusion of case solutions in experience memory and semantic fragments in vector memory.

[0210] The generation module is used to generate personalized service responses corresponding to dialogue requests based on the fusion decision results and service context profiles through the intelligent decision engine. The service context profile is dynamically constructed based on emotional state representation, visual semantic understanding and text intent parsing, and is used to reflect the user's service needs and preference characteristics at the current stage of the dialogue.

[0211] It should be noted that the intelligent customer service dialogue management system integrating multi-layer memory provided in the above embodiments is only illustrated by the division of the above functional modules when performing dialogue management. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the intelligent customer service dialogue management system integrating multi-layer memory provided in the above embodiments and the intelligent customer service dialogue management method embodiments integrating multi-layer memory belong to the same concept. The specific implementation process is detailed in the method embodiments and will not be repeated here.

[0212] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent customer service dialogue management integrating multi-layer memory, characterized in that, The methods include: In response to dialogue requests from intelligent customer service, the system performs multimodal context perception on voice signals, image data, and text data from users to obtain emotional state representation, visual semantic understanding, and text intent parsing. Based on emotional state representation and text intent parsing, a memory coordinator is used to coordinate the multi-layer memory system to obtain a set of memory features that match the current dialogue context. The coordinated scheduling includes activating personalized service strategies in contextual memory based on emotional state representation and retrieving relevant historical solutions from experiential memory and vector memory based on text intent parsing. Based on memory feature sets and visual semantic understanding, a fusion decision-maker is used to integrate multi-source memories to obtain a fusion decision result that is adapted to the current dialogue context. Multi-source memory integration includes consistency verification of operation records in program memory and user preferences in context memory, as well as correlation fusion of case solutions in experience memory and semantic fragments in vector memory. Based on the fusion of decision results and service context profiles, a personalized service response is generated for each dialogue request through an intelligent decision engine. The service context profile is dynamically constructed based on emotional state representation, visual semantic understanding, and text intent parsing to reflect the user's service needs and preferences at the current stage of the dialogue.

2. The method according to claim 1, characterized in that, Multimodal context perception is performed on voice signals, image data, and text data from users to obtain emotional state representations, visual semantic understanding, and text intent parsing, including: Acoustic feature analysis and prosodic pattern extraction are performed on the speech signal to obtain acoustic feature sequences and prosodic feature parameters; Multidimensional emotion analysis is performed on acoustic feature sequences and prosodic feature parameters to obtain emotional state representations. Multidimensional emotion analysis includes joint identification of emotion intensity and emotion type. Multi-scale visual feature extraction and scene semantic parsing are performed on image data to obtain visual feature maps and scene understanding information; Multi-level semantic reasoning is performed on visual feature maps and scene understanding information to obtain visual semantic understanding. Multi-level semantic reasoning includes entity recognition and relationship understanding of visual scenes. Semantic structure analysis and intent feature extraction are performed on text data to obtain semantic representation vectors and intent feature representations; Context-aware intent reasoning is performed on semantic representation vectors and intent feature representations to obtain text intent parsing. Context-aware intent reasoning includes intent evolution analysis in the current dialogue context.

3. The method according to claim 1, characterized in that, Based on emotional state representation and text intent parsing, a memory coordinator is used to coordinate the multi-layered memory system, resulting in a set of memory features that match the current dialogue context, including: Based on the emotional intensity and type of emotional state representation, a memory coordinator dynamically assigns weights to the service preference dimension in episodic memory to obtain a set of emotionally adapted service strategies. The dynamic weight assignment includes adjusting the priority of service strategies according to the emotional intensity. Based on the intent type and semantic complexity of text intent parsing, a memory coordinator is used to perform cross-level association retrieval of experiential memory and vector memory to obtain a set of intent matching solutions. The cross-level association retrieval includes determining the retrieval ratio of experiential memory and vector memory based on semantic complexity. The service strategy set and solution set are subjected to strategy consistency verification and priority fusion to obtain a memory feature set. The strategy consistency verification includes a collaborative evaluation of service strategies and solutions based on service scenario matching degree.

4. The method according to claim 3, characterized in that, Based on the intensity and type of emotion represented by emotional state, a memory coordinator dynamically assigns weights to the service preference dimension in episodic memory, resulting in a set of emotion-adaptive service strategies, including: The emotional state representation is quantified in terms of emotional intensity and classified into emotional types to obtain emotional intensity values ​​and emotional category labels. Based on emotion intensity values ​​and emotion category identifiers, a memory coordinator dynamically assigns weights to the service preference dimension in episodic memory to obtain the preference weight distribution for emotion adaptation. The dynamic weight assignment includes determining the activation level of the preference dimension based on the emotion intensity value through a dynamic weight assignment model. Based on the preference weight distribution of emotion adaptation and historical service records in contextual memory, service strategies are generated through a memory coordinator to obtain a set of emotion-adapted service strategies. The service strategy generation includes prioritizing and filtering historical service records according to the preference weight distribution.

5. The method according to claim 3, characterized in that, The service policy set and solution set are subjected to policy consistency verification and priority fusion to obtain a memory feature set, including: The service scenario synergy is evaluated between each service strategy in the service strategy set and each solution in the solution set to obtain the synergy distribution between strategies and solutions; The suitability of strategies and schemes is verified by analyzing the distribution of synergy and the representation of emotional state, and the verified strategy and scheme combination is obtained. The suitability verification includes dynamically adjusting the synergy threshold according to the emotional intensity. The validated strategy combination and text intent parsing are fused by strategy priority to obtain a memory feature set. The strategy priority fusion includes a dual weight allocation mechanism based on semantic complexity and emotional intensity.

6. The method according to claim 1, characterized in that, Based on memory feature sets and visual semantic understanding, a fusion decision-maker integrates multi-source memories to obtain a fusion decision result adapted to the current dialogue context, including: Based on scene entities and relationships in visual semantic understanding, a decision fusion engine is used to dynamically resolve conflicts and enhance consistency between procedural memory operation records and contextual memory user preferences in the memory feature set, resulting in a preference-enhanced operation sequence. Dynamic conflict resolution includes detecting and correcting logical contradictions between operation records and user preferences in the current service scenario. Based on the semantic context of text intent parsing, a context-aware fusion scheme is obtained by semantically guiding the complementary fusion of experiential memory case schemes and vector memory semantic fragments in the memory feature set through a fusion decision-maker. The semantically guided complementary fusion includes using the semantic fragments of vector memory to fill the semantic gaps of experiential memory cases in the current dialogue context. Based on emotion state representation, a fusion decision-maker is used to perform multiple rounds of iterative optimization on the preference-enhanced operation sequence and the context-aware fusion scheme. The fusion decision result is output with emotion stability as a constraint. The multiple rounds of iterative optimization include dynamically adjusting the combination strategy of operation sequence and fusion scheme according to the changes in emotion intensity.

7. The method according to claim 6, characterized in that, Based on scene entities and relationships in visual semantic understanding, a decision fusion engine is used to dynamically resolve conflicts and enhance consistency between procedural memory operation records and episodic memory user preferences in the memory feature set, resulting in a preference-enhancing operation sequence, including: The program memory operation record is parsed step-by-step to obtain the operation step sequence and the execution dependencies between steps; The sequence of operation steps is matched with the scene entity relationships in visual semantic understanding. When the execution order of operation steps in the sequence of operation steps is found to be inconsistent with the scene relationship constraints indicated by the scene entity relationships, it is marked as a scene logic conflict point. The consistency matching analysis includes evaluating the feasibility of operation steps based on entity spatial relationships. Based on the emotional intensity of the scene's logical conflict points and emotional state representations, the user preferences of the scene memory are prioritized and recalibrated by a fusion decision-maker to generate conflict resolution strategies. Based on conflict resolution strategies and service context profiling, the step sequence in the program memory operation record is reconstructed and the parameters are optimized to output a preference-enhanced operation sequence.

8. The method according to claim 6, characterized in that, Based on the semantic context of text intent parsing, a context-aware fusion scheme is obtained by semantically guiding the complementary fusion of experiential memory case schemes and vector memory semantic fragments in the memory feature set through a fusion decision-maker, including: The task structure of the experience memory case solution is analyzed and the solution features are extracted to obtain the case task framework and core solution elements. Based on the semantic context of text intent parsing, the semantic integrity of the case task framework is evaluated, and the semantic gaps of the experience memory case solutions in the current dialogue context are identified. Based on the identified semantic gaps, supplementary semantic segments that match the current semantic context are retrieved from the vector memory semantic segments, and the fusion weight of the supplementary semantic segments is determined by semantic relevance calculation. Based on the fusion weights, supplementary semantic fragments are dynamically and weightedly fused with the core solution elements of experiential memory case schemes to generate context-aware fusion schemes.

9. The method according to claim 1, characterized in that, Based on the fusion of decision-making results and service context profiles, a personalized service response corresponding to the dialogue request is generated through an intelligent decision engine, including: The context fit of the integrated decision results and service context profiles is evaluated to obtain the context fit index. The context fit evaluation includes calculating the matching degree of the decision results based on service needs and preference characteristics. Based on the dynamic demand characteristics in the adaptation index and service context profile, the intelligent decision engine generates elements to obtain multi-dimensional response elements, which include content elements, tone elements and action suggestion elements. By integrating and optimizing the naturalness of multi-dimensional response elements, personalized service responses are obtained. The integration of response elements includes the consistent integration of content, tone, and action suggestions based on the principle of dialogue coherence.

10. An intelligent customer service dialogue management system integrating multi-layer memory, characterized in that, The system includes: The perception module is used to respond to dialogue requests with intelligent customer service by performing multimodal context perception on voice signals, image data and text data from users to obtain emotional state representation, visual semantic understanding and text intent parsing. The scheduling module is used to coordinate the multi-layer memory system based on emotional state representation and text intent parsing through a memory coordinator to obtain a set of memory features that match the current dialogue context. The coordinated scheduling includes activating personalized service strategies in contextual memory based on emotional state representation and retrieving relevant historical solutions from experiential memory and vector memory based on text intent parsing. The integration module is used to integrate multi-source memories based on memory feature sets and visual semantic understanding through a fusion decision-maker to obtain a fusion decision result adapted to the current dialogue context. Multi-source memory integration includes consistency verification of operation records in program memory and user preferences in context memory, as well as correlation fusion of case solutions in experience memory and semantic fragments in vector memory. The generation module is used to generate personalized service responses corresponding to dialogue requests based on the fusion decision results and service context profiles through the intelligent decision engine. The service context profile is dynamically constructed based on emotional state representation, visual semantic understanding and text intent parsing, and is used to reflect the user's service needs and preference characteristics at the current stage of the dialogue.

Images