User preference analysis method based on memory enhanced multi-modal large model

Abstract

The invention discloses a user preference analysis method based on a memory enhanced multi-modal large model. Comprising the steps of multi-modal interaction data acquisition and preprocessing, time sequence preference level identification and classification, cross-modal preference consistency analysis, implicit preference extraction and verification, cross-domain preference boundary identification and migration rule generation, multi-level preference conflict coordination and decision, and memory enhancement preference model updating and application. According to the method, accurate time sequence classification of different types of preferences is realized through the adaptive time decay model; coordinating and fusing multi-channel preference signals by using a cross-modal alignment mapping function; and efficient capture of implicit preferences is realized by adopting a low-interference verification strategy.

Description

Technical Field

[0001] The present invention belongs to the field of multimodal data processing, and in particular to a user preference parsing method based on a memory-enhanced multimodal large model. Background Art

[0002] With the rapid development of artificial intelligence technology, the interaction between intelligent systems and users is becoming increasingly diverse, expanding from single text input to multimodal interactions such as images, voice, and gestures. In this complex interactive environment, accurately understanding and responding to user preferences has become the key to improving user experience. As the core technology of personalized intelligent services, user preference parsing directly affects the adaptability of the system and user satisfaction. This is especially important for products such as intelligent assistants, recommendation systems, and creative design tools. Therefore, building a parsing system that can accurately capture and understand users' multimodal preferences has great theoretical value and application significance.

[0003] At present, the research on user preference parsing mainly focuses on capturing preferences in a single mode and at a specific time point. Traditional methods such as collaborative filtering, content-based filtering, and rule-based reasoning systems usually only consider explicit preference signals from a single source. Recent studies have begun to try to integrate multimodal information, such as the CLIP model that uses contrastive learning to achieve joint representation of text and images, and the MultiSense system that combines voice and facial expressions to understand user emotions. In terms of timing, some studies have introduced time decay models, such as BPR-MF, which uses an exponential decay function to weight historical behaviors. For preference transfer, methods such as TransRec have proposed knowledge transfer techniques based on domain mapping. In terms of conflict handling, existing systems mostly use simple priority rules or strategies in which recent preferences cover old preferences.

[0004] Despite some progress, current preference parsing technology still has problems: First, in terms of multi-level temporal preference conflict coordination, existing technologies treat all preferences as the same time level and cannot effectively distinguish short-term temporary preferences, medium-term task preferences, and long-term stable preferences, resulting in an inability to judge whether "high saturation" temporary demands in specific scenarios should cover "natural saturation" long-term preferences, and when temporary preferences should expire. Second, in terms of implicit preference extraction and verification, existing systems mainly rely on explicit user feedback, and are unable to capture implicit preference signals contained in user behaviors such as slight hesitations and repeated adjustments, or affect user experience through frequent inquiries and verifications. These technical gaps seriously restrict the intelligent system's deep understanding and accurate response to user preferences. Summary of the invention

[0005] The purpose of the invention is to provide a user preference analysis method based on a memory-enhanced multimodal large model to solve the above-mentioned problems existing in the prior art.

[0006] Technical solution: A user preference parsing method based on a memory-enhanced multimodal large model, comprising the following steps: Collect and preprocess multimodal interaction data to generate standardized feature vector sets and cleaned multimodal interaction sequences; Based on the cleaned multimodal interaction sequence, the temporal preference hierarchy is identified and classified to generate a multi-level temporal preference classification result and a temporal preference conversion rule set; Based on the multi-level temporal preference classification results and temporal preference conversion rule set, multi-level preference conflict coordination and decision-making are carried out to generate a unified preference decision result; Based on the unified preference decision results, the memory-enhanced preference model is updated and applied to generate an updated user preference model, which is then applied to the interactive interface and content generation to obtain personalized recommendation results.

[0007] Beneficial effects: The present invention can capture explicit and implicit preferences, process cross-modal and cross-domain preferences, coordinate multi-level preference conflicts, and possess long-term memory and progressive learning capabilities; the present invention breaks through the limitations of traditional preference parsing, realizes a preference understanding mechanism that is closer to human cognition, improves the personalized service quality of intelligent systems, solves the core challenges faced by current preference parsing systems, and provides a technical foundation for building a truly user-centric artificial intelligence system. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A flowchart of the steps of a user preference parsing method based on a memory-enhanced multimodal large model provided in an embodiment of the present application.

[0009] Figure 2 A flowchart of the steps for constructing an adaptive time decay model provided in an embodiment of the present application.

[0010] Figure 3 A flowchart of the steps for constructing a semantic-visual alignment space provided in an embodiment of the present application.

[0011] Figure 4 A flowchart of the steps for constructing a low-interference verification strategy provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0013] It should be noted that in order to clearly show the steps of this application, serial numbers are marked for each step in the specification. These serial numbers are only used for the convenience of explanation and do not limit the order of execution of the steps. In actual operation, according to the technical requirements of the specific implementation scenario, the steps can be executed in a different order than that shown in the specification, and in some cases, parallel processing between steps can be achieved.

[0014] like Figure 1 As shown, a user preference parsing method based on a memory-enhanced multimodal large model includes the following steps: S1, collect and preprocess multimodal interaction data to generate standardized feature vector sets and cleaned multimodal interaction sequences; S2. Based on the cleaned multimodal interaction sequence, the temporal preference hierarchy is identified and classified to generate a multi-level temporal preference classification result and a temporal preference conversion rule set; S3, performing cross-modal preference consistency analysis based on the standardized feature vector set and generating a cross-modal preference consistency report; S4, extracting and verifying implicit preferences based on the cleaned multimodal interaction sequence, and generating an implicit preference mapping table; S5. Based on the multi-level temporal preference classification results, cross-domain preference boundary identification and migration rule generation are performed to generate a cross-domain preference migration rule set; S6. Based on the multi-level temporal preference classification results and the temporal preference conversion rule set, multi-level preference conflict coordination and decision-making are performed to generate a unified preference decision result; S7. Update and apply the memory-enhanced preference model based on the unified preference decision results, generate an updated user preference model, and apply it to the interactive interface and content generation to obtain personalized recommendation results.

[0015] The multimodal interaction data in this embodiment refers to the information including text, pictures, videos, audio, expressions and other modes transmitted between the user and the system during the human-computer dialogue process. Multimodal interaction sequence: historical multimodal interaction data between the user and the system arranged in chronological order, reflecting the contextual history of user behavior and system response. For example, the user first sends a picture, then a description text, and finally sends a voice opinion, which constitutes a complete multimodal interaction sequence. Multi-level temporal preference classification results: based on the understanding of the multimodal interaction sequence and combined with the contextual relationship, the user's implicit preferences are classified in a hierarchical and staged manner. For example, the user inputs "I think this picture is too bright", and the corresponding multi-level temporal preference classification result is "Picture adjustment → brightness → reduce brightness". Temporal preference conversion rule set: converts the preferences expressed by the user in different historical contexts into a set of rules for clear image processing operation instructions based on the memory bank and the set rules. For example, for text that expresses the problem of image brightness, a rule of "involving image brightness → automatic matching memory → image processing instructions" is formulated, which is eventually converted into an "image brightness reduction" instruction that can directly call relevant image processing tools. Unified preference decision results: Through comprehensive analysis of context information and preference memory, the user preference decision results are uniformly output to guide system behavior. For example, for a user's expression "I think this image is too bright", the unified preference decision result is finally generated by combining the image in the context and the similar text processing flow mapping stored in the memory library: "Automatically reduce the brightness of the current image."

[0016] According to one aspect of the present application, step S1 further comprises: S11. Capture multi-source interaction data: read user device interface data, synchronously collect text instructions, image operations, interaction behaviors and time information through a multi-channel monitoring mechanism, construct the original multi-source interaction data packet, and ensure the time synchronization and source integrity of the data.

[0017] S12. Signal separation and modality recognition: read multi-source interactive data packets, classify the data using modality recognition algorithms, separate mixed signals into text modality data streams, image modality data streams, and behavioral modality data streams, and retain the temporal correspondence between modalities.

[0018] S13, noise filtering and anomaly detection: read each modal data stream, identify and remove noise signals caused by user misoperation or system interference through statistical outlier detection and fluctuation threshold filtering, and output denoised modal data stream.

[0019] S14, context association and conversation segmentation: read the denoised modal data stream, divide the continuous interaction data into meaningful conversation units based on time interval and semantic coherence analysis, establish context association within and between modalities, and output segmented conversation sequence and context association graph.

[0020] S15, feature extraction and standardization: read the segmented conversation sequence, extract semantic features, visual features and behavioral features for different modal data, make the features of each modality comparable through normalization, and output a standardized feature vector set and a cleaned multimodal interaction sequence.

[0021] This embodiment achieves the acquisition and integration of high-quality multimodal data, and establishes a solid foundation for subsequent preference analysis. The multi-channel monitoring mechanism is used for synchronous acquisition, and the signal is separated through the modal recognition algorithm, which effectively retains the temporal correspondence between modalities; combined with statistical outlier detection and fluctuation threshold filtering, it can improve data quality and reduce the misjudgment rate; the conversation segmentation method based on time interval and semantic coherence analysis solves the problem of continuous interaction data being difficult to divide in traditional methods; feature extraction and standardized processing for different modalities ensure the comparability and consistency of multimodal data, laying a data foundation for subsequent preference recognition.

[0022] According to one aspect of the present application, step S2 is further: S21. Analyze interaction frequency and duration: Read the cleaned multimodal interaction sequence, calculate the frequency, duration, and repetition pattern of each preference expression, establish the time distribution characteristics of the preference, and output the preference time distribution map.

[0023] S22. Construct a time decay function: Based on the preference time distribution mapping, construct a nonlinear time decay function, which can dynamically adjust the decay rate according to the historical distribution characteristics of the preference expression, so that different types of preferences have different "memory retention periods" and output an adaptive time decay model.

[0024] S23. Analyze multi-scale time windows: read the cleaned multimodal interaction sequence, use multi-scale sliding window technology to analyze the stability and change trend of preferences in short-term windows (hours), medium-term windows (days) and long-term windows (months), and output a multi-scale preference stability matrix.

[0025] S24, hierarchical clustering and preference classification: read the multi-scale preference stability matrix and the data processed by the adaptive time decay model, use the hierarchical clustering algorithm to divide the preference expression into short-term temporary preference (high volatility, low stability), medium-term task preference (medium stability) and long-term stable preference (low volatility, high stability), and output the multi-level temporal preference classification results.

[0026] S25, timing preference conversion boundary detection: read the multi-level timing preference classification results, analyze the conditions and triggers for the conversion of preferences from one level to another, establish conversion rules and boundary conditions, and output the timing preference conversion rule set for subsequent preference conflict coordination.

[0027] This embodiment solves the key problem that traditional preference models ignore the differences in time dimensions. Through the preference time distribution mapping technology, the time-varying characteristics of user preferences can be captured; the constructed nonlinear time decay function dynamically adjusts the decay rate according to the historical distribution of preferences, so that different types of preferences have corresponding "memory retention cycles"; the multi-scale sliding window technology analyzes the stability of preferences from three dimensions: short-term, medium-term, and long-term, providing unprecedented time granularity; the hierarchical clustering algorithm accurately classifies preferences into short-term temporary preferences, medium-term task preferences, and long-term stable preferences, improving the system's ability to distinguish different temporal preferences; the temporal preference conversion boundary detection technology effectively identifies the conditions and triggers for preference level conversion, providing an important basis for subsequent conflict coordination.

[0028] Because in the memory enhancement system, the user's preferences usually change at multiple time scales (short-term temporary preferences, medium-term task preferences, and long-term stable preferences), and the existing solutions do not solve the conflict problem between preferences at different time scales. For example, the user's long-term preference when processing landscape photos may be "natural saturation", but "high saturation" may be temporarily required in a specific scene. The system cannot automatically determine which preference should take precedence, nor can it determine when the temporary preference should expire. There is a lack of a mechanism that can automatically identify, classify, and coordinate multi-level temporal preference conflicts, and it is impossible to reasonably handle short-term preference changes while maintaining the stability of long-term preferences. Existing memory enhancement models usually regard all preferences as the same level, without considering the hierarchical structure of the time dimension. Therefore, according to one aspect of the present application, step S22 is further: S221. Extract the time characteristics of preference expression; read the preference time distribution map, and extract four key time characteristics for each preference expression: frequency periodicity characteristic (the periodic law of preference occurrence), persistence stability characteristic (the duration of a single preference expression), repetition interval characteristic (the time interval between multiple occurrences of the same preference), and recent activity characteristic (the difference between the timestamp of the most recent preference expression and the current time). These four characteristics are combined to form a preference time feature vector.

[0029] S222, clustering of preference categories by time patterns; read the preference time feature vectors of all preferences, use the K-means++ clustering algorithm to divide the preferences into multiple categories according to their time behavior patterns, evaluate the optimal number of clusters by the silhouette coefficient and Davies-Bouldin index, and obtain preference time behavior categories, each of which represents a set of preferences with similar time performance characteristics.

[0030] S223. Calculate category-specific attenuation parameters; read the preferred time behavior category, analyze the time characteristic distribution of each category, use the maximum likelihood estimation method to determine the attenuation function form (exponential, power law or hyperbolic) and parameter range that best suits the category, and output the category attenuation parameter set, which includes the function type and parameter value range for each category.

[0031] S224, construct an adaptive attenuation function; read the category attenuation parameter set, set the weight coefficient and attenuation rate parameter, and construct a mixed attenuation function template containing adjustable parameters: D(t, c, p) = w1(c)*exp(-α(p)*t)+w2(c)*(1 / (1+β(p)*t))+w3(c)*(t (-γ(p)) ); where t represents the time variable, c represents the preference category, p represents the personalized parameter, w1, w2, w3 are weight coefficients (satisfying w1+w2+w3=1), and α, β, γ are decay rate parameters. This function combines three modes: exponential decay, hyperbolic decay, and power law decay. It can adapt to the memory characteristics of different types of preferences and output a decay function template for subsequent personalized parameter learning and tuning.

[0032] S225, personalized parameter learning and tuning; read the preference time feature vector, preference time behavior category and attenuation function template, use the Bayesian optimization algorithm to learn the personalized attenuation parameters of each user on historical data, minimize the error between the predicted preference strength and the actual user feedback, and obtain the user-specific personalized attenuation parameters.

[0033] S226. Implement a dynamic adjustment mechanism; build a responsive adjustment mechanism so that the attenuation function can dynamically update parameters based on the user's latest interaction data. Read personalized attenuation parameters and real-time user feedback data, use online learning methods to continuously fine-tune the attenuation parameters, so that the function can adapt to changes in user preference stability, and output an adaptive time attenuation model with self-adjustment capabilities.

[0034] S227, attenuation model verification and calibration; read the adaptive time decay model, verify the model performance on the retained test set of historical data, calculate the prediction accuracy, recall rate and F1 score, and compare and analyze with the fixed decay rate model, perform final calibration of the model parameters according to the verification results, and output the verified and calibrated optimized adaptive time decay model, which will be passed to step S23 for multi-scale time window analysis.

[0035] like Figure 2 As shown, according to one aspect of the present application, the temporal preference level identification and classification includes the step of constructing an adaptive time decay model, and the step of constructing an adaptive time decay model includes: Based on the cleaned multimodal interaction sequence, a preference time distribution map is constructed to extract the time features of preference expression and generate a preference time feature vector. Cluster analysis is performed on the preference time feature vector to obtain the preference time behavior category. Calculate category-specific attenuation parameters for preferred time behavior categories and generate category attenuation parameter sets; construct a mixed attenuation function based on the category attenuation parameter sets and generate an attenuation function template; Based on the preferred time feature vector, preferred time behavior category and attenuation function template, personalized parameter learning and tuning are performed to generate personalized attenuation parameters. Based on the personalized attenuation parameters, a dynamic adjustment mechanism is implemented to generate an adaptive time attenuation model for multi-scale time window analysis.

[0036] According to one aspect of the present application, the steps of extracting the temporal characteristics of preference expression and cluster analysis include: Frequency periodicity features, continuous stability features, repetition interval features and recent activity features are extracted from the preference time distribution map; the frequency periodicity features, continuous stability features, repetition interval features and recent activity features are combined to form a preference time feature vector; the K-means++ clustering algorithm is applied to the preference time feature vector to generate the initial clustering result; the optimal number of clusters is evaluated by the silhouette coefficient and Davies-Bouldin index to generate the preference time behavior category.

[0037] This embodiment solves the problem that the traditional fixed decay rate model cannot adapt to the memory characteristics of different preference types by constructing an adaptive time decay function. It can accurately distinguish the time attributes of different preferences and assign the most appropriate "memory retention period" to each preference, thereby improving the accuracy of temporal preference classification.

[0038] According to one aspect of the present application, step S3 is further: S31. Construct a semantic-visual alignment space; read the standardized feature vector set, and construct a shared semantic space of the text modality and the image modality through a contrastive learning method, so that the preference expressions of different modalities can be compared in the same space, and output a cross-modal alignment mapping function.

[0039] S32. Conduct behavior-semantic association modeling; read the standardized feature vector set and context association graph, analyze the correspondence between user behavior modality and semantic expression, establish a mapping model from behavior sequence to semantic intent, and output the behavior-semantic mapping matrix.

[0040] S33, calculating the intra-modal consistency; reading the standardized feature vector set of each modality respectively, calculating the consistency degree of preference expression at different time points within the same modality, identifying preference conflicts and changes within the modality, and outputting the internal consistency score of each modality.

[0041] S34, measure inter-modal consistency; read the cross-modal alignment mapping function, behavior-semantic mapping matrix and the standardized feature vector set of each modality, calculate the semantic consistency of preference expressions between different modalities, identify the expression differences between modalities, and output the inter-modal consistency matrix.

[0042] S35. Locate and quantify consistency conflicts; read the internal consistency score and inter-modal consistency matrix, accurately locate the conflict points of preference expressions between different modalities, quantify the degree of conflict and the scope of influence, and output a cross-modal preference consistency report, including the conflict intensity, location and possible explanations.

[0043] This embodiment breaks through the limitations of traditional single-modal preference analysis through cross-modal preference consistency analysis and realizes the deep integration of multi-modal preference signals. By constructing a semantic-visual alignment space, preference expressions of text and image modalities can be compared in the same semantic space; behavioral-semantic association modeling effectively solves the problem of understanding user behavior intentions; intra-modal consistency calculation can accurately identify preference conflicts and changes within the same modality; inter-modal consistency measurement realizes the semantic consistency calculation of preference expressions between different modalities; consistency conflict location and quantification technology can accurately locate and quantify the conflicts of preference expressions between different modalities, improve the system's ability to understand complex multi-modal signals, and reduce the preference misunderstanding rate caused by modal inconsistency.

[0044] like Figure 3 As shown, according to one aspect of the present application, step S31 is further: S311, extracting text semantic features and image semantic features from the standardized feature vector set; S312, based on the cleaned multimodal interaction sequences, screen out time- and semantically related text-image pairs and construct a selected parallel corpus; S313. Use text semantic features, image semantic features and selected parallel corpora to build a double-tower neural network model for shared latent space learning and generate a cross-modal alignment mapping network; S314, verifying and fine-tuning the semantic consistency of the cross-modal alignment mapping network to generate an improved cross-modal alignment mapping network; S315. Constructing a preference expression-specific adaptation layer based on the improved cross-modal alignment mapping network to generate a preference-enhanced cross-modal mapping network, i.e., a semantic-visual alignment space; S316. Evaluate and optimize the preference-enhanced cross-modal mapping network to generate a cross-modal alignment mapping function for inter-modal consistency measurement.

[0045] In one embodiment of the present application, modal feature extraction and initial representation are performed; a standardized feature vector set is read, and feature extraction is performed on text modality and image modality data respectively. For text data, a pre-trained bidirectional encoder representation (BERT) variant model is used to extract context-sensitive word vectors to obtain text semantic features; for image data, a pre-trained visual transformer (Transformer) model is used to extract visual features to obtain image semantic features. Both features are mapped to an initial feature space of 512 dimensions, retaining the original semantic information.

[0046] Screen and construct parallel data pairs. Read the cleaned multimodal interaction sequence, identify the text-image pairs that are closely related in time, and construct the initial parallel corpus. Then design a dual filtering mechanism based on temporal correlation and semantic correlation to remove noise samples, retain high-quality text-image pairs, and output a selected parallel corpus as training data for cross-modal alignment.

[0047] Shared latent space learning. Read text semantic features, image semantic features and selected parallel corpora to build a dual-tower neural network model. The text tower and image tower each contain multiple layers of transformations, and finally map to the same dimension of latent space. Use cross-modal contrast loss function to train the model, optimize the training process through random batch processing and hard negative sample mining strategy, and output the initial cross-modal alignment mapping network.

[0048] Verify and fine-tune semantic consistency. Read the cross-modal alignment mapping network and validation set data, perform cross-modal retrieval tasks (text-image and image-text bidirectional retrieval) to evaluate the mapping quality, and calculate indicators such as R@1, R@5, and average ranking. Based on the verification results, use the gradient subsampling strategy to focus on the samples that are difficult to align, fine-tune the mapping network parameters, and output the improved cross-modal alignment mapping network.

[0049] Construct a dedicated adaptation layer for preference expression. Read the improved cross-modal alignment mapping network and build a dedicated adaptation layer for the special nature of preference expression. This layer can strengthen the semantic features related to preference and weaken irrelevant features. Fine-tune with a small amount of labeled preference expression data to achieve migration from the general semantic space to the preference-specific semantic space, and output a preference-enhanced cross-modal mapping network.

[0050] Perform alignment space evaluation and optimization. Read the preference-enhanced cross-modal mapping network, construct three evaluation indicators: inter-modal alignment, semantic fidelity, and preference sensitivity, and conduct a comprehensive evaluation on real user data. Based on the evaluation results, use the Bayesian hyperparameter optimization method to perform final tuning on the network structure and training strategy, and output the final cross-modal alignment mapping function, which will be used for subsequent inter-modal consistency measurement.

[0051] According to one aspect of the present application, shared latent space learning includes the steps of constructing a cross-modal contrast loss function, specifically: Construct a global alignment loss component to ensure that the overall representation distribution of different modalities is similar; construct a local semantic consistency loss component to ensure that the distance between semantically similar content in the cross-modal space is within a preset range; construct a structure preservation loss component to ensure that the semantic structure within the modality is preserved after mapping; The global alignment loss component, the local semantic consistency loss component and the structure preservation loss component are combined to form a cross-modal contrastive loss function. The cross-modal contrastive loss function, text semantic features, image semantic features and a selected parallel corpus are used to train a dual-tower neural network model to generate a cross-modal alignment mapping network.

[0052] According to one aspect of the present application, the step of constructing a preference expression dedicated adaptation layer includes: Based on the improved cross-modal alignment mapping network, a dedicated adaptation layer is constructed for the special properties of preference expression to form a preference-specific adaptation layer. The preference-specific adaptation layer is used to strengthen the semantic features related to preference, weaken irrelevant features, and generate enhanced preference expression features. The preference-specific adaptation layer is fine-tuned through a small amount of labeled preference expression data to obtain the fine-tuning parameters; based on the improved cross-modal alignment mapping network, preference-specific adaptation layer and fine-tuning parameters, the migration from the general semantic space to the preference-specific semantic space is achieved to generate a preference-enhanced cross-modal mapping network.

[0053] This embodiment solves the technical problem that text and image modal preference expressions cannot be directly compared. It can compare and associate preference expressions of different modalities in a unified semantic space, improve the accuracy of cross-modal preference understanding, solve the problem of difficulty in integrating preference information of different modalities in traditional methods, and lay a solid foundation for multimodal consistency analysis.

[0054] According to one aspect of the present application, step S4 is further: S41. Identify micro-behavior patterns; read the behavioral modal data stream in the cleaned multimodal interaction sequence, use a fine-grained behavioral analysis algorithm to identify the micro-behavior characteristics of the user's implicit intentions such as slight hesitation, repeated operations, and changes in dwell time, and output a micro-behavior feature set.

[0055] S42. Mining contextual condition associations; reading micro-behavior feature sets and contextual association graphs, analyzing the changing patterns of user behavior patterns under specific contextual conditions, identifying the associations between environmental factors and preference expressions, and outputting context-preference condition probability tables.

[0056] S43, generate implicit preference hypotheses; read the micro-behavior feature set and the context-preference conditional probability table, generate candidate hypotheses about the user's implicit preferences based on the Bayesian reasoning framework, assign an initial confidence level to each hypothesis, and output the implicit preference hypothesis set.

[0057] S44. Construct a low-interference verification strategy; read the implicit preference hypothesis set, and design a verification strategy that minimizes user interference for each implicit preference hypothesis to be verified, including indirect inquiries that are naturally integrated into the conversation, comparison of alternative solutions, and fine-tuning parameters to observe reactions, etc., and output a low-interference verification solution.

[0058] S45, verification execution and hypothesis update; execute low-interference verification scheme, collect verification feedback data, update the confidence of implicit preference hypothesis according to user response, confirm high-confidence hypothesis as valid preference, and output implicit preference mapping table, which includes semantic description, trigger conditions and confidence score of implicit preference.

[0059] This embodiment solves the major problem that traditional systems can only capture explicit preferences but ignore implicit preferences through implicit preference extraction and verification. Micro-behavior pattern recognition technology can capture subtle behavioral characteristics of users, such as signals of implicit intentions such as hesitation and repeated operations; contextual condition association mining analyzes the relationship between environmental factors and preference expression, improving the context sensitivity of preference understanding; the implicit preference hypothesis generation method based on Bayesian reasoning enables the system to actively infer the user's potential preferences; the low-interference verification strategy design realizes the verification of implicit preference hypotheses without disturbing the user; the verification execution and hypothesis update mechanism forms a closed-loop implicit preference learning system, which enables the model to continuously optimize the understanding of the user's implicit preferences and improves the system's ability to capture preference signals hidden in interactive behaviors.

[0060] During the interaction process, users usually only express some preferences explicitly, while a large number of preferences are implicit and not clearly expressed. The preference capture module in the existing scheme mainly relies on the user's explicit feedback and cannot effectively capture implicit preferences, such as the user's slight hesitation, repeated adjustment of the same parameters, preference patterns in specific scenarios, etc. There is currently a lack of a mechanism that can extract implicit preference signals from user interactions and confirm these implicit preferences through low-interference verification methods. Existing technologies either ignore implicit preferences or verify through frequent inquiries, resulting in a degraded user experience. This is a unique problem because it involves how to accurately capture subtle preference signals hidden in interactive behaviors without interrupting the user experience. Therefore, Figure 4 As shown, according to one aspect of the present application, step S44 is further: S441. Based on the cleaned multimodal interaction sequence, generate an implicit preference hypothesis set and perform verification demand analysis and classification to generate classified preference hypotheses; S442. Based on the classified preference hypothesis, construct natural dialogue fusion verification templates, parameter fine-tuning reaction observation schemes, and alternative scheme comparison strategies, respectively, and generate dialogue fusion verification template libraries, parameter fine-tuning verification scheme sets, and alternative scheme comparison verification strategies; S443, construct a multi-round progressive verification plan based on the dialogue fusion verification template library, the parameter fine-tuning verification solution set and the alternative solution comparison verification strategy to generate a timing verification plan; S444. Construct an adaptive control mechanism for the verification process based on the timing verification plan and generate verification control parameters; S445. Integrate the dialogue fusion verification template library, parameter fine-tuning verification scheme set, alternative scheme comparison verification strategy, timing verification plan and verification control parameters to form a low-interference verification strategy for verification execution and hypothesis update.

[0061] In one embodiment of the present application, verification requirements analysis and classification are performed. The implicit preference hypothesis set is read, and the difficulty of each hypothesis is evaluated. Considering the three dimensions of interaction complexity required for verification, user attention cost, and certainty of verification results, the hypotheses are divided into three categories: high-certainty hypothesis, medium-certainty hypothesis, and low-certainty hypothesis, and differentiated verification strategies are designed for different categories.

[0062] Build natural dialogue fusion verification templates. Design a series of natural dialogue templates for high-certainty hypotheses so that verification questions can be seamlessly integrated into the normal dialogue process. Read the user's dialogue history data, extract common language patterns and communication habits, customize personalized dialogue templates based on these features, ensure that verification questions do not interrupt the rhythm of the dialogue or cause user awareness, and output the dialogue fusion verification template library.

[0063] Construct parameter fine-tuning reaction observation scheme. For medium-certainty hypotheses, design an indirect verification method based on small parameter adjustments. Read the relevant hypotheses in the implicit preference hypothesis set, design 3-5 small parameter change schemes for each hypothesis, and control the change range near the user's perception threshold. Verify the hypothesis by observing the user's reaction to these fine-tuning, and output the parameter fine-tuning verification scheme set.

[0064] Construct an alternative solution comparison strategy. For low-certainty hypotheses, design a verification strategy based on limited alternative solutions. Read the hypothesis to be verified in the implicit preference hypothesis set, build 2-3 opposing alternative solutions for each hypothesis, design a natural presentation method, let users make choices without realizing that they are being tested, and output the alternative solution comparison verification strategy.

[0065] Generate a multi-round progressive verification plan. Read the dialogue fusion verification template library, parameter fine-tuning verification solution set, and alternative solution comparison verification strategy, design a multi-round progressive verification plan, ensure that the verification process is dispersed in multiple interactions, and avoid user fatigue caused by centralized verification. According to the user's interaction frequency and habits, assign the best verification time to each hypothesis and output a timing verification plan.

[0066] Control the adaptability of the verification process. Build a dynamic adaptation mechanism for the verification process, read real-time user feedback signals (such as response time, participation, etc.), and automatically adjust the verification intensity and frequency according to the user status. When user fatigue or inattention is detected, the verification frequency is automatically reduced or suspended, and verification control parameters with adaptive capabilities are output.

[0067] Integrate low-interference verification scheme. Integrate the dialogue fusion verification template library, parameter fine-tuning verification scheme set, alternative scheme comparison verification strategy, timing verification plan and verification control parameters to form a complete verification scheme. Perform conflict check and optimization on the overall scheme to ensure the coordination and consistency between various verification strategies, and output the final low-interference verification scheme, which will be used for verification execution in step S45.

[0068] According to one aspect of the present application, the steps of constructing a natural dialogue fusion verification template include: Filter high-certainty hypotheses from the classified preference hypotheses to form high-certainty hypotheses to be verified; analyze user conversation history data, extract users' common language patterns and communication habits, and form users' personalized language features; Based on the high-certainty hypothesis to be verified and the user's personalized language features, a series of natural dialogue templates are designed so that verification questions can be seamlessly integrated into the normal dialogue process to form an initial dialogue template set; the initial dialogue template set is customized according to the user's personalized language features to ensure that verification questions will not interrupt the rhythm of the dialogue or cause user awareness, and a dialogue fusion verification template library is generated.

[0069] According to one aspect of the present application, the step of constructing a parameter fine-tuning reaction observation scheme comprises: Select medium-certainty hypotheses from the classified preference hypotheses to form medium-certainty hypotheses to be verified; design 3-5 small parameter change plans for the medium-certainty hypotheses to be verified to form a parameter change plan set; Optimize the parameter change plan set, control the change range near the user's perception threshold, and form an optimized parameter change plan; construct indicators for observing user reactions to parameter changes, and form a reaction observation indicator set; The optimized parameter change scheme and the reaction observation indicator set are combined to form a parameter fine-tuning verification scheme set, which is used to verify the hypothesis by observing the user's reaction to the fine-tuning.

[0070] This embodiment solves the problem of interfering with the user experience when verifying implicit preference assumptions. It realizes the verification of implicit preference assumptions without interrupting the normal user experience, reduces the verification cost and user disturbance, and maintains the reliability of the verification results, solving the problem of the traditional explicit inquiry verification method causing a decline in user experience.

[0071] According to one aspect of the present application, step S45 is further: S451. Scheduling and executing verification tasks based on the low-interference verification scheme and implicit preference hypothesis set, and collecting original verification feedback data; S452, extracting and cleaning feedback signals from the original verification feedback data to generate structured feedback data; S453. Establish a Bayesian evidence accumulation model based on structured feedback data and implicit preference hypothesis set to generate a hypothesis posterior probability table; S454, performing multi-source evidence consistency evaluation on the hypothesis posterior probability table and the structured feedback data to generate an evidence consistency score; S455, making hypothesis confirmation and elimination decisions based on the hypothesis posterior probability table and the evidence consistency score, and generating hypothesis status update results; S456, performing hypothesis generation for incremental learning based on the hypothesis state update result and the user interaction data to generate a new implicit preference hypothesis; S457. Construct and update an implicit preference mapping table based on the assumed state update result and the newly added implicit preference assumption to generate an implicit preference mapping table.

[0072] In one embodiment of the present application, verification tasks are scheduled and executed. The low-interference verification scheme and implicit preference hypothesis set are read, and the execution order and timing of the verification tasks are arranged according to the timing verification plan. The corresponding verification strategy is triggered at the appropriate interaction node, and user response data, including selection results, response time, and interaction mode, is collected, and the original verification feedback data is output.

[0073] Extract and clean feedback signals. Read the original verification feedback data, apply signal processing technology to extract effective feedback signals, and filter out noise and outliers. For text feedback, extract sentiment polarity and semantic tendency; for behavioral feedback, extract operation mode and time characteristics; for selection feedback, extract preference strength and certainty, and output cleaned structured feedback data.

[0074] Construct a Bayesian evidence accumulation model. Construct a Bayesian evidence accumulation model, read the structured feedback data and the prior probability in the implicit preference hypothesis set, and calculate the posterior probability of each feedback data corresponding to the hypothesis. Adopt an incremental update mechanism to combine new feedback with historical evidence accumulation to form a continuously updated evidence chain and output a hypothesis posterior probability table.

[0075] Conduct consistency assessment of multi-source evidence. Read the hypothesis posterior probability table and structured feedback data to evaluate the consistency of evidence from different verification strategies. When there is a conflict between evidence from different sources, weight the evidence according to its reliability and timeliness, resolve the conflict of evidence, and output the evidence consistency score.

[0076] Make hypothesis confirmation and elimination decisions. Read the hypothesis posterior probability table and evidence consistency score, and set a dynamic threshold to determine the status of the hypothesis. When the posterior probability exceeds the high threshold and the consistency score is good, the hypothesis is confirmed as a valid preference; when the posterior probability is lower than the low threshold or the consistency score is extremely poor, the hypothesis is eliminated; the remaining hypotheses remain in a pending verification state, and the hypothesis status update results are output.

[0077] Generate incremental learning hypotheses. Based on the hypothesis status update results and the latest user interaction data, apply incremental learning technology to generate new hypotheses to fill the gaps of eliminated hypotheses. The new hypothesis generation process uses the characteristic patterns of confirmed hypotheses to improve the quality and relevance of hypotheses and output new implicit preference hypotheses.

[0078] Construct and update the implicit preference mapping table. Integrate the confirmed assumptions and newly added implicit preference assumptions in the assumption status update results to construct a complete implicit preference mapping table. The mapping table contains the semantic description, trigger conditions, confidence score and related context of each confirmed preference, sorted by confidence, for use by subsequent modules. At the same time, the assumptions that need further verification are looped back to step S44 to enter the next round of verification process.

[0079] According to one aspect of the present application, the steps of establishing a Bayesian evidence accumulation model and hypothesis confirmation and elimination decision-making include: Extract the prior probability of each hypothesis from the implicit preference hypothesis set to form the hypothesis prior probability; calculate the posterior probability of the hypothesis corresponding to each feedback data based on the structured feedback data and the hypothesis prior probability to generate the single feedback posterior probability; An incremental update mechanism is used to combine the single feedback posterior probability with the historical evidence accumulation to form a continuously updated evidence chain and generate a hypothesis posterior probability table. Dynamic thresholds are set based on the hypothesis posterior probability table and evidence consistency scores to generate confirmation thresholds and elimination thresholds. The confirmation threshold and elimination threshold are applied to judge the hypothesis. When the posterior probability exceeds the confirmation threshold and the consistency score is good, the hypothesis is confirmed as a valid preference. When the posterior probability is lower than the elimination threshold or the consistency score is extremely poor, the hypothesis is eliminated, and the remaining hypotheses remain in a state of pending verification to generate hypothesis status update results.

[0080] This embodiment forms a closed-loop system for implicit preference learning. It can continuously learn and verify implicit preferences from users' daily interactions, forming a self-improving closed-loop learning system, improving the system's ability to capture and confirm implicit preferences, and solving the problem that traditional systems do not understand implicit preferences well enough or the verification method is too intrusive.

[0081] Some preferences established by users in one field (such as landscape image processing) may be applicable to other fields (such as portrait processing), while some may not be applicable. Existing solutions cannot automatically identify which preferences can be transferred across domains and which preferences need to be isolated within the domain, resulting in the system either over-generalizing preferences (incorrectly applying specific domain preferences to unrelated fields) or being overly conservative (unable to utilize transferable preference knowledge). There is currently a lack of a mechanism that can automatically identify the boundaries of preference migration, and it is impossible to accurately determine the scope of application and migration conditions of specific preferences. This involves how to extract implicit domain boundary information from historical interaction data and how to verify migration assumptions in new scenarios. This involves the representation and reasoning of domain knowledge, rather than simple memory storage and retrieval. Therefore, according to one aspect of the present application, step S5 is further: S51. Extract application domain features; read the cleaned multimodal interaction sequence and multi-level temporal preference classification results, extract domain features of different application scenarios, establish domain feature vectors, and output application domain feature space.

[0082] S52, calculating the similarity between domains; reading the domain feature vector in the application domain feature space, calculating the similarity matrix between different application domains, identifying the correlation strength between domains, and outputting the domain similarity matrix.

[0083] S53. Analyze the preference-domain correlation; read the multi-level temporal preference classification results and the application domain feature space, analyze the applicability and performance differences of specific preferences in different domains, establish the correlation mapping between preferences and domain attributes, and output the preference-domain correlation table.

[0084] S54, automatically identify migration boundaries; read the domain similarity matrix and preference-domain association table, use the decision tree algorithm to identify under what conditions preferences can be migrated across domains and under what conditions intra-domain isolation is required, and output the preference migration boundary definition.

[0085] S55, generate and verify migration rules; based on the preference migration boundary definition, generate specific cross-domain preference migration rules, and verify the effectiveness of these rules through historical data simulation, adjust the rule parameters, and output the final cross-domain preference migration rule set.

[0086] This embodiment solves the problem of transferring preference knowledge between different application fields, improves the system's ability to generalize user preferences, avoids the dilemma of over-generalization and over-conservatism, and enables users to enjoy the convenience of established preferences in new scenarios.

[0087] In one embodiment of the present application, automatic identification of migration boundaries specifically includes: analyzing domain feature differences; reading the domain similarity matrix and the application domain feature space, calculating feature distances between different domains, using the Mahalanobis distance to measure differences in the multidimensional feature space, identifying key distinguishing features, and outputting a domain distinguishing feature set and a feature difference weight table.

[0088] Evaluate preference sensitivity. Read the preference-domain association table and the multi-level temporal preference classification results, and calculate the sensitivity of each preference to performance changes in different domains. Construct a sensitivity scoring function: S(p, d1, d2) = |P(p|d1) - P(p|d2)| / max(P(p|d1), P(p|d2)); where p represents preference, d1 and d2 represent two different domains, and P(p|d) represents the probability of preference p being applicable in domain d. Calculate the sensitivity of all preferences between all domain pairs and output the preference sensitivity matrix.

[0089] Select features for the migration decision tree. Read the domain-differentiated feature set, feature difference weight table, and preference sensitivity matrix, build a feature selection algorithm based on information gain, and identify the most critical feature subset for preference migration decisions. Calculate the information gain of each feature: IG(f) = H(T) - Σ(v∈V(f)) |T_v| / |T| * H(T_v); where f is the feature, T is the training set, V(f) is the value set of feature f, H represents information entropy, and T_v is the data subset corresponding to the value of feature f in the training data when v is taken. Select the feature subset with the highest information gain and output the decision-critical feature set.

[0090] Learn the migration condition rules. Read the decision key feature set, preference sensitivity matrix and historical preference migration success cases, build a decision tree model to learn the migration condition rules. Build the initial decision tree, each internal node represents a test of a certain feature, each leaf node represents a migration decision (can be migrated / cannot be migrated), and output the initial migration condition decision tree.

[0091] Build a migration risk assessment model. Read the migration condition decision tree and historical user feedback data, build a migration risk assessment model, and quantify the potential negative impact of incorrect migration. Design a risk scoring function: R(p, d1, d2) = I(p) * D(d1, d2) * (1-C(p, d1, d2)); where I(p) is the preference importance, D(d1, d2) is the domain distance, and C(p, d1, d2) is the confidence of the migration. Calculate the risk score for each potential migration decision and output the migration risk score table.

[0092] Perform decision boundary optimization and pruning. Read the migration condition decision tree and migration risk score table, and apply pruning and boundary optimization techniques to improve the generalization ability of the decision tree. Use cost-complexity pruning to reduce overfitting, adjust the decision threshold according to the risk score, set more conservative migration conditions for high-risk nodes, and output the optimized migration decision rule tree.

[0093] Visualize and verify the migration boundary. Read the migration decision rule tree and domain similarity matrix, and visualize the migration boundary in the feature space for easy understanding and verification. Evaluate the boundary accuracy through randomly generated test cases, calculate the precision, recall and F1 score, make final adjustments based on the verification results, and output the verified preference migration boundary definition, which will be used for migration rule generation in step S55.

[0094] This embodiment solves the key problem of accurately migrating preference knowledge between different domains by automatically identifying migration boundaries. This breakthrough enables the system to accurately determine under what conditions preferences can be migrated across domains and under what conditions intra-domain isolation is required, avoiding the dilemma of over-generalization and over-conservatism, and improving the system's ability to reuse existing preference knowledge in new scenarios.

[0095] In multimodal interactions, users express their preferences through different channels such as text, images, and interactive behaviors, and the signals of these channels are sometimes inconsistent or even contradictory. For example, a user expresses in text that he wants the image to be "more natural", but always chooses a high-saturation effect in the interaction. Existing solutions cannot effectively deal with this problem of multimodal signal inconsistency, resulting in a deviation in the system's understanding of the user's true preferences. There is currently a lack of a mechanism that can detect, quantify, and coordinate the inconsistency of multimodal preference signals, and it is impossible to extract the user's true intentions from contradictory signals. This involves how to design a cross-modal preference consistency metric and how to establish a reliable mapping relationship between different modalities. This involves deep semantic understanding and fusion of multimodal information, rather than preference processing within a single modality. Therefore, according to one aspect of the present application, step S6 is further: S61. Detect preference conflicts; read multi-level temporal preference classification results, cross-modal preference consistency reports and implicit preference mapping tables, comprehensively detect conflict points between preferences at different levels, different modalities and different types (explicit / implicit), and output a preference conflict map.

[0096] S62. Construct a preference priority matrix; based on the temporal preference conversion rule set, the cross-modal preference consistency report, and the confidence in the implicit preference mapping table, construct a multi-dimensional preference priority scoring system, assign dynamic weights to each preference, and output the preference priority matrix.

[0097] S63. Build a context-aware conflict resolution strategy; read the preference conflict map, preference priority matrix and current application context data, design a context-based conflict resolution strategy set, adopt different coordination methods for different types of conflicts, and output a conflict resolution strategy set.

[0098] S64. Perform conflict coordination and verification; apply the conflict resolution strategy set to handle conflicts in the preference conflict map, generate a coordinated preference set, verify the consistency of the coordination results and user satisfaction through historical data simulation, and output a conflict coordination report.

[0099] S65. Generate a unified preference decision; read the conflict coordination report and all non-conflicting preferences, integrate them to form the final unified preference decision, ensure the consistency and coherence of the decision results in all dimensions, and output the unified preference decision results.

[0100] This embodiment solves the problem of multiple preference conflicts in complex scenarios. It can make decisions that meet the user's true intentions in complex multi-preference conflict scenarios, improve user experience and satisfaction, and solve the problem that traditional systems make confusing decisions or simply take the latest preference when facing preference conflicts.

[0101] In one embodiment of the present application, the context-aware conflict resolution strategy is specifically constructed as follows: classifying conflict types and extracting features; reading the preference conflict map, analyzing the structural features of the conflict, and classifying the conflict into four basic types: temporal hierarchical conflict, modal expression conflict, explicit and implicit conflict, and domain applicability conflict. For each conflict type, a specific feature vector is extracted, including the conflict intensity, the number of preferences involved, the historical resolution mode, etc., and a conflict type feature table is output.

[0102] Build a historical conflict resolution case library. Read historical user interaction data and preference update records, identify conflict cases that have been successfully resolved in the past, extract conflict features and solutions, and build a historical conflict resolution case library. Structure the case representation, including fields such as conflict description, resolution strategy, and user satisfaction score, to form the basis for case retrieval.

[0103] Extract and represent contextual features. Read the current application context data and extract key contextual features, including task type, time pressure, user status, and environmental conditions. Design a contextual representation model to map multi-dimensional contextual information into a vector representation for subsequent strategy matching and output the current contextual feature vector.

[0104] Perform case similarity calculation and retrieval. Read the conflict type feature table, historical conflict resolution case library and current situation feature vector, and construct a multi-factor similarity calculation formula: Sim(c1, c2) = w1*SimType(c1, c2) + w2*SimContext(c1, c2) + w3*SimUser(c1, c2); where SimType represents conflict type similarity, SimContext represents situation similarity, SimUser represents user feature similarity, w1, w2, w3 are weight coefficients (satisfying w1+w2+w3=1), and c1 and c2 are two different conflict cases. Retrieve the most matching historical case based on similarity and output a set of similar cases.

[0105] Generate context-adaptive strategies. Read similar case sets and current context feature vectors, and generate solution strategies suitable for the current context through case adaptation technology. Perform context-adaptive adjustments on the retrieved historical strategies, taking into account factors such as the current user status, task importance, and time constraints, and output a set of candidate solution strategies.

[0106] Conduct strategy utility evaluation and optimization. Construct a multi-dimensional strategy utility evaluation framework and calculate the expected utility of each strategy in the candidate solution strategy set: U(s) = Σ(i=1 to n) w_i * u_i(s); where s is the strategy, u_i is the scoring function of the i-th utility dimension, and w_i is the weight. The evaluation dimensions include user satisfaction, decision consistency, execution efficiency, and cognitive load. Sort and filter the strategies according to the utility scores, and output the optimization strategy ranking table.

[0107] Perform adaptive strategy combination and integration. Read the optimization strategy ranking table and preference conflict map, analyze the dependencies between conflicts, and design an overall coordination solution. For interrelated conflicts, ensure that consistent resolution strategies are adopted; for independent conflicts, select the best resolution strategy for each. Integrate all strategies into a unified solution and output a complete conflict resolution strategy set, which will be used for conflict coordination execution in step S64.

[0108] This embodiment solves the complex problem of using different coordination methods for preference conflicts in different scenarios through a context-aware conflict resolution strategy. It can dynamically select the most suitable conflict resolution strategy according to the specific situation, rather than using fixed conflict resolution rules, which improves the flexibility and accuracy of conflict coordination and solves the problem of rigid conflict handling in traditional systems in complex situations.

[0109] In one embodiment of the present application, the process of generating a unified preference decision is specifically as follows: integrating the post-conflict preference set; reading the conflict coordination report and all non-conflicting preferences, merging the coordinated conflicting preferences with the original non-conflicting preferences to form a preliminary unified preference set; checking the logical consistency within the set to ensure that there are no residual contradictions, and outputting the preliminary unified preference set.

[0110] Recalculate preference priorities. Read the preliminary unified preference set and preference priority matrix, and recalculate the preference priorities based on the adjustments made during the conflict coordination process. Consider the changes in weights during the conflict resolution process and the new contextual conditions, update the priority score of each preference, and output the updated priority preference set.

[0111] Identify decision constraints. Read the update priority preference set and the current application context data to identify the constraints that the decision must meet, including system capacity limitations, resource constraints, and user-defined bottom-line requirements. Formalize these constraints as boundary conditions in the decision space and output the decision constraint set.

[0112] Generate a preference implementation plan. Read and update the priority preference set and decision constraint set, and design a preference implementation plan that satisfies the constraints. For multiple preferences that cannot be satisfied at the same time, determine the execution order based on the priority; for preferences that can be partially satisfied, determine the degree of satisfaction and parameter values, and output the preference implementation plan.

[0113] Conduct cross-dimensional consistency checks. Read the preference implementation plan and build a multi-dimensional consistency check mechanism to ensure the consistency of decisions in multiple dimensions. Check dimensions including temporal consistency (continuity of decisions at different time points), semantic consistency (compatibility of different expressions), and functional consistency (coordination of different functional domains). Identify and fix potential inconsistencies and output a consistency check report.

[0114] Fine-tune and optimize parameters. Read the preference implementation plan and consistency check report, and fine-tune the parameters for the detected inconsistencies. Use the simulated annealing algorithm to search for the optimal solution in the parameter space, fine-tune the specific parameter values ​​while maintaining the preference intent, make the overall decision more harmonious and consistent, and output the optimized parameter solution.

[0115] Generate and encapsulate the final decision. Integrate the optimization parameter scheme and update the high priority preferences in the priority preference set to generate the final unified decision. Encapsulate the decision into a standard format, including the core decision content, implementation parameters, priority information and explanatory text, to facilitate system execution and user understanding, and output the unified preference decision result, which will be used for the memory enhancement preference model update in step S7.

[0116] This embodiment achieves a final decision output that is consistent in multiple dimensions and complies with constraints by generating a unified preference decision. It is able to generate consistent, coherent and constrained decisions in a complex multi-preference environment, avoiding the problems of fragmented, contradictory or constraint-violating decisions in traditional systems, and providing users with a holistic solution that meets their true intentions.

[0117] According to one aspect of the present application, step S7 is further: S71. Extract difference increments; read the unified preference decision results and the existing user preference model, calculate the difference between the new and old models, extract the meaningful changes, and output the preference increment data.

[0118] S72. Build a memory-optimized storage strategy; read the incremental preference data and multi-level temporal preference classification results, design differentiated storage strategies based on the time level attributes of the preferences, assign different storage priorities and update frequencies to different types of preferences, and output a memory-optimized storage solution.

[0119] S73. Update the incremental preference model; integrate the incremental preference data into the existing preference model according to the memory optimization storage solution, adopt a progressive update mechanism to avoid catastrophic forgetting, maintain the stability and consistency of the model, and output the updated user preference model.

[0120] S74, constructing a personalized parameter mapping; reading the updated user preference model, establishing a mapping relationship between the user semantic expression and the specific operation parameters, forming a personalized parameter converter, and outputting a personalized parameter mapping table.

[0121] S75. Generate and apply recommendation strategies; based on the updated user preference model and personalized parameter mapping table, generate personalized recommendation strategies that meet the user's current needs and long-term preferences, apply them to interactive interface and content generation, output personalized recommendation results, and collect user feedback for the next round of preference model updates.

[0122] This embodiment achieves the long-term accumulation and optimal utilization of preference knowledge. It can continuously learn and improve the understanding of user preferences like humans, maintain the long-term stability of memory, and allow moderate updates and adjustments, solving the problem of untimely update of preference models or loss of important information during the update process in traditional systems.

[0123] In one embodiment of the present application, the process of updating the preference model increment is specifically as follows: constructing an incremental data structure; reading the incremental preference data, and constructing a dedicated incremental data structure, which can efficiently represent the preference change, including fields such as old value, new value, change type (new, modified, deleted), timestamp, and confidence. Organizing the incremental data into a tree structure to reflect the hierarchical relationship between preferences, and outputting a structured incremental preference tree.

[0124] Analyze the impact of updates. Read the incremental preference tree and the existing user preference model, and analyze the impact of each incremental change on the model. Identify directly and indirectly affected model components through dependency graph analysis, predict possible chain reactions, and output the update impact map.

[0125] Predict and prevent catastrophic forgetting. Read the update impact map and apply the feature preservation detection algorithm to predict areas where catastrophic forgetting may occur. Design important feature protection mechanism: L_protect = λ * ||M_new(x_i) - M_old(x_i)|| 2 ; Where M_new and M_old are the models before and after the update, x_i is the important sample, and λ is the protection strength parameter. The protection mechanism is integrated into the update process and the forgetting prevention strategy is output.

[0126] Perform incremental updates of knowledge distillation. Read the existing user preference model, incremental preference tree, and forgetting prevention strategy, and design an incremental update method based on knowledge distillation. Use the original model as the teacher model and the new model as the student model, and transfer stable knowledge through the distillation loss function: L_distill = α*L_new + (1-α)*KL(P_teacher||P_student); where L_new is the learning loss of new data, KL is the KL divergence, α is the balance parameter, P_teacher is the probability distribution of the teacher model, and P_student is the probability distribution of the student model. New knowledge is gradually integrated through this mechanism, and a distillation update plan is output.

[0127] Balance stability and plasticity. Build an adaptive learning rate adjustment mechanism to balance the stability and plasticity of the model. Read the distillation update scheme and the user's historical interaction stability analysis, and dynamically adjust the learning rate according to the historical stability of the user's preference: η_t = η_base * exp(-β*S(t)); where η_t is the learning rate at time t, S(t) is the preference stability index, β is the decay parameter, and η_base is the initial learning rate. Use lower learning rates in areas with high stability to ensure slow updates; use higher learning rates in areas with high plasticity to allow fast adaptation and output adaptive learning parameters.

[0128] Perform incremental update execution and monitoring. Perform model update operations according to the distillation update scheme and adaptive learning parameters. Adopt a phased update strategy, first update the model components that are less affected, then gradually update the core components, monitor the changes in performance indicators during the update process, roll back immediately if an abnormality is found, and output the update execution log and the preliminary updated model.

[0129] Perform consistency verification and repair after the update. Read the updated model, design a multi-dimensional consistency verification scheme, and test the model performance on the retained data set. Calculate the decision consistency index before and after the update: C = |{x | M_new(x) ==M_old(x), x ∈ X_stable}| / |X_stable|; where X_stable is the sample set that should remain stable. For areas with low consistency, apply the repair strategy for fine-tuning to ensure the continuity of key functions, and finally output the updated user preference model that has been verified and repaired, which will be used for the personalized parameter mapping construction in step S74.

[0130] This embodiment solves the problem of catastrophic forgetting that may occur during the model update process by updating the preference model increment. It can maintain the stability of existing preference knowledge while continuously absorbing new preference knowledge, solving the dilemma of traditional systems that either fail to fully absorb new knowledge or lose important old knowledge during the update process, and provides technical support for long-term stable preference learning.

[0131] In another embodiment of the present application, a user preference parsing system based on a memory-enhanced multimodal large model includes three main modules: a memory module, a content generation module, and a preference capture module. The preference capture module captures the user's conversation habits and image enhancement preferences and stores them in the memory module. The content generation module performs conversation generation and image enhancement based on the user's conversation and the memory in the memory module.

[0132] The memory module includes two parts: preference memory and memory retriever. The preference memory stores the mapping between the current user dialogue and its corresponding image enhancement process. It is automatically saved to the memory through the preference capture module during the dialogue with the user and is continuously updated with the interaction with the user. The memory retriever includes a text classifier (TextClassifier) ​​and a dense retrieval encoder (Dense Retrieval Encoder). The classifier is used to determine whether there is an intention of image enhancement in each user's dialogue input. The dense retriever is used to match the saved memory to extract the user preferences in the memory.

[0133] The content generation module includes the multimodal large language model backbone (MLLM) and the image enhancement toolbox. The multimodal large language model backbone is used to achieve continuous and natural multimodal dialogue with users. The image enhancement toolbox contains a variety of mainstream image enhancement tools, which can be called in sequence according to instructions to achieve more complex and flexible image enhancement effects.

[0134] The preference capture module is connected to the memory module and reuses the multimodal large language model in the content generation module. The preference capture module uses the text classifier in the memory module to determine whether the user's input dialogue contains image enhancement intentions. Then, the multimodal large language model is combined with the image enhancement tool in the toolbox to generate the image enhancement process to be selected, and interact with the user to confirm the user's real needs, and finally update the memory library to complete the capture of user preferences. At the same time, after each enhancement is completed, user feedback is obtained through dialogue to adjust the enhancement process.

[0135] During the entire operation process of the algorithm, the preference capture module captures the user's personal preferences in continuous conversations and saves them in the memory module. The memory module is responsible for extracting the correct memory in the conversation and mapping the user input to its personalized enhancement method. The content generation module is responsible for receiving the enhancement method extracted by the memory module while conducting the conversation and enhancing the image.

[0136] According to one aspect of the present application, a user preference parsing method based on a memory-enhanced multimodal large model comprises the following steps: Step 1: First, determine whether the user's input contains image enhancement intent. The classifier of the retriever in the memory bank is a text classification model, which is built on the basis of the transformer (Roberta) model represented by the robust optimized bidirectional encoder. The classification (CLS) vector of its encoded text is fed into a feedforward neural network (FFN) model to perform a binary classification task to determine whether the dialogue contains image enhancement intent.

[0137] Step 2: If not, use the multimodal language model backbone in the content generation module for regular conversation; if included, check whether the corresponding memory is empty, which is completed through the dense retrieval model in the retriever. This is a text encoder that encodes the text into a vector in the retrieval space. Similar texts have higher similarity. In the present invention, when the dot product value of two vectors is greater than 0.5, they are matched. If multiple texts in the memory library meet the requirements, the maximum value is taken. Otherwise, it does not match and the memory is determined to be empty.

[0138] Step 3: If the memory is empty, the system uses the multimodal language model backbone in the content generation module combined with the tools in the image enhancement toolbox to infer the user's possible enhancement intentions, generate several candidate enhancement processes, and ask the user for confirmation to obtain the accurate user intentions and establish a mapping relationship to update the memory library.

[0139] Step 4: If the memory is not empty, directly match the most similar process memory and perform enhancement processing.

[0140] Step 5. Finally, the system will ask the user about his / her satisfaction. If the user is satisfied, the process ends. Otherwise, the language model will adjust the direction based on the user's input, adjust the image enhancement process, and update the memory bank.

[0141] Step 6: Perform image enhancement again according to the adjusted process and execute step 5.

[0142] In an embodiment of image enhancement of the present application, user: <Picture> I think this picture is too dark; determine whether there is an intention: yes; obtain the toolbox process; generate query text. System: Do you want to do the following? 1. Increase brightness; 2. Perform low-light enhancement; 3. No modification intention. User: I choose 1. Analyze input; get the process [{brightness: +10}]; store memory: [I think this picture is too dark: [{brightness: +10}]]; process the picture. System: Are you satisfied? <Processed picture [{brightness: +10}]>. User: "It is dark overall, especially the face of the person, and there is a small amount of noise." Analyze how to adjust: get the new process [{brightness: +25}, {noise reduction: 30}]; process the picture; update memory: [I think this picture is too dark: [{brightness: +25}, {noise reduction: 30}]]]. System: Are you satisfied? <Processed picture [{brightness: +25}, {noise reduction: 30}]>. User: Satisfied. Subsequent dialogue includes: User: <Image> This image is too dark. Matching memory: [I think this image is too dark: [{brightness: +25}, {noise reduction: 30}]], processing the image. System: Are you satisfied? <Processed image [{brightness: +25}, {noise reduction: 30}]>. User: Satisfied.

[0143] In this embodiment, when the memory bank is empty, 1. The user inputs: I think this picture is too dark. is a placeholder token for the image). 2. The text classifier in the "I think this picture is too dark" retriever determines that the text may contain image processing intent. 3. Determine whether the memory bank is empty, yes. 4. Call the preference capture module, the module obtains the saved fixed prompt words and dialogues with the large model trunk to obtain the inquiry text: Prompt word example: Based on the user's description of the image problem and the existing image enhancement tool list, generate a concise text to inquire about the user's intention. Follow the steps below: Semantic analysis: identify keywords in the user's description; tool matching: filter all relevant tools from the tool library (1-3 related items are allowed); option generation: list matching items with numbers, keeping the original tool name; exclusion rules: ignore tools that are not related to the user's problem; output format: use a friendly question template: "Do you want to do the following?" + line break column item, and finally add a non-intention option; the tool list is [brightness adjustment, low light enhancement, denoising, super resolution]; the user input is I think this picture is too dark. 5. System output: Do you want to perform the following operations? (1) Increase brightness; (2) Perform low-light enhancement; (3) No modification intention. 6. User input: I choose 1. 7. Get the user's answer text, call the large language model again to judge, and get the accurate answer. Prompt word example: Judge the user's needs based on the user input, and strictly output a tool name; the options include (1) Increase brightness; (2) Perform low-light enhancement; (3) No modification intention; the user input is I choose 1. 8. The large language model judges: (1) Increase brightness. 9. According to the output tool name, match the enhancement process in the image enhancement toolbox (some processes are preset for the first startup), increase the brightness, corresponding to [{brightness: +10}]. 10. Save the user input and enhancement process mapping in the memory library: "I think this picture is too dark" [{brightness: +10}], and process the picture. If the output has no intention, it means that the user has no intention to enhance it, and the memory is stored. I think this picture is too dark - None. 11. Encode the user input in the memory bank using a dense retrieval encoder. 12. After processing, output the picture and output text to ask about the effect. The system outputs: Are you satisfied? <Processed picture>. 13. User input: "It is dark overall, especially the face of the person, and there is a small amount of noise." Obtain user feedback text and call the large model for judgment. The example is as follows: Prompt words: Based on the user's feedback on the image enhancement effect, perform the following operations: (1) Satisfaction judgment: If the feedback contains clear affirmation (such as "satisfied" / "good" / "ok") or no modification suggestions, directly output: "satisfied". (2) Parameter adjustment rules (only when unsatisfied): a) Adjust based on the current process: Current process: [{brightness: +10}]; b) Adjustment rules: 1) Modify the parameters in the process according to user feedback; 2) Add the following process according to user feedback: {{process to be selected in the toolbox}}; (3) Output format: When unsatisfied, output dictionary list: [{"step": value}, ...], get the adjusted process: [{brightness: +25}, {noise reduction: 30}]. 14. The new process updates the memory library, [{brightness: +25}, {noise reduction: 30}] replaces [{brightness: +10}]. 15. Ask again, loop until the output is satisfactory.

[0144] When the memory bank is not empty, 1. When there are memories in the memory bank, the dense retriever encodes the input and performs a dot product with the data in the memory bank. If the dot product value is <0.5, the retrieval fails. If it is >0.5, the maximum dot product value is taken if there are multiple results. 2. If the retrieval fails, repeat 1; if the retrieval is successful, the mapping process is enhanced. If the mapping is unintended, no processing is performed.

[0145] The text classifier uses a large number of regular conversations and conversations with intent for classification training. The dense retriever can use the retriever in the existing RAG, which can match the text similarity, or it can be retrained for comparative learning. The large language model can use the COT model of R1 and o1 to enhance the effect. The large language model can fine-tune the instructions to obtain better process optimization capabilities, and the remaining judgment tasks basically do not require training. The image enhancement toolbox can be a multi-layer structure, such as requirements-tools-processes, which are gradually refined and accurate. The process format of this embodiment is for illustration, and it can be more complex. In addition to processing images, the present application can also process audio, video, text, etc.

[0146] This embodiment uses the multimodal large language model backbone and memory module to achieve seamless connection between spoken dialogue and professional image processing process. The memory bank in the memory module stores a large amount of memory data mapped between natural language and image enhancement process. These data are collected and updated in the dialogue with the help of the preference capture module, which can accurately match each round of fuzzy and popular input of the user to the corresponding complex image enhancement scheme, so that users can obtain high-quality image processing effects only through daily dialogue without mastering professional terms. This embodiment has the ability to understand the image enhancement intentions behind different expressions of different users. The large language model built into the system can understand various language styles and expressions. The preference capture module captures and analyzes the personalized preferences of users by analyzing the continuous interaction with users, and accurately records the expression characteristics and demand changes of users when parsing various complex dialogues. The retriever in the memory module can quickly extract the historical records that best match the current input from the memory bank during the interaction, so as to achieve accurate recognition and response to user habits. The system's memory bank can store user preference information for a long time. During each interaction, the memory bank automatically records the user's language habits and expression characteristics corresponding to the specific needs behind them. During continuous interaction, the memory bank continuously receives information from the preference capture module, updates the memory bank, and gradually builds a complete user portrait. As the usage time increases, the system's understanding of the user continues to deepen, thereby providing more personalized and customized image enhancement services in subsequent interactions, which not only improves the accuracy of the response, but also optimizes the user experience.

[0147] The present invention effectively solves the coordination problem of multi-level temporal preference conflicts through a temporal preference hierarchy identification and classification module and a multi-level preference conflict coordination and decision module. Specifically: the system uses interaction frequency and duration analysis to identify the temporal distribution characteristics of preferences and establishes a preference time distribution map. Then, the decay rate is dynamically adjusted for different types of preferences through an adaptive time decay function to achieve a differentiated "memory retention cycle". Multi-scale time window analysis analyzes preference stability at three scales: short time window, medium time window, and long time window, while hierarchical clustering accurately classifies preferences into short-term temporary preferences, medium-term task preferences, and long-term stable preferences. Temporal preference conversion boundary detection further analyzes the conditions and triggers for preference conversion between different levels. When faced with preference conflicts, the preference conflict detection in the module can fully identify conflicts between preferences at different levels; the preference priority matrix assigns dynamic weights to each preference based on temporal characteristics; the context-aware conflict resolution strategy dynamically selects the best solution according to the specific scenario; and finally, unified preference decision generation ensures that the decision results remain consistent in all dimensions. This enables the system to intelligently prioritize based on specific situations when temporary demands for “high saturation” conflict with long-term preferences for “natural saturation,” and to accurately determine when temporary preferences should expire.

[0148] The implicit preference extraction and verification module solves the problem of capturing implicit preferences and verifying them with low interference: micro-behavior pattern recognition can capture the behavioral characteristics of users' implicit intentions such as slight hesitation and repeated operations; contextual condition association mining analyzes the changing rules of user behavior patterns under specific contexts; implicit preference hypothesis generation generates and preliminarily evaluates implicit preference hypotheses based on the Bayesian reasoning framework. The key lies in the design of low-interference verification strategies, verification execution, and hypothesis updating. After classifying the verification requirements, three low-interference verification methods are designed: natural dialogue fusion verification seamlessly integrates verification problems into normal dialogues; parameter fine-tuning reaction observation indirectly verifies through small parameter changes; alternative scheme comparison strategy allows users to make choices without being aware of being tested. Multi-round progressive verification plans and adaptive control of the verification process ensure that the verification process is decentralized and the intensity is dynamically adjusted according to the user status. The Bayesian evidence accumulation model and multi-source evidence consistency evaluation continuously update the credibility of the hypothesis to form a closed-loop learning system, which improves the system's ability to capture and verify implicit preferences while avoiding interference with user experience.

[0149] The problem of accurate boundary identification of preference cross-domain migration is solved through the cross-domain preference boundary identification and migration rule generation module: application domain feature extraction and inter-domain similarity calculation provide the basis for identifying domain relationships; preference-domain correlation analysis deeply explores the relationship between preferences and domain attributes. The core lies in the automatic identification of migration boundaries, which calculates the feature distance between different domains through domain feature difference analysis; preference sensitivity evaluation quantifies the performance changes of preferences in different domains; migration decision tree feature selection and migration condition rule learning establish an accurate migration judgment mechanism; migration risk assessment model quantifies the potential negative impact of incorrect migration; decision boundary optimization and pruning improve the generalization ability of decision trees; migration boundary visualization and verification ensure the accuracy of boundary definition. Finally, migration rule generation and verification generate specific rules based on the migration boundary definition and verify the effectiveness. This enables the system to accurately determine when the preference for landscape photo processing can be transferred to portrait processing, avoiding the dilemma of over-generalization and over-conservatism.

[0150] The problem of inconsistent multimodal preference signals is solved through the cross-modal preference consistency analysis module: the semantic-visual alignment space construction realizes the expression of text and image modalities in a unified semantic space through modal feature extraction, parallel data pair screening, contrast loss function design and shared latent space learning. Behavior-semantic association modeling establishes a mapping between user behavior and semantic intention. Through intra-modal consistency calculation and inter-modal consistency measurement, the system can accurately identify the differences in preference expression within the same modality and between different modalities. Consistency conflict location and quantification further accurately locate the conflict point and quantify the degree of conflict. When the user expresses in the text that he wants the image to be "more natural" but the behavior prefers high saturation, the system can deeply understand the contradiction between different modal signals through these technologies, and combine the conflict coordination mechanism to comprehensively analyze the user's true intention. This multimodal consistency analysis improves the system's ability to understand complex preference signals and reduces the preference misunderstanding rate caused by modal inconsistency.

[0151] The present invention constructs a complete memory-enhanced multimodal preference parsing closed-loop system. From multimodal data acquisition and processing to temporal preference classification, from cross-modal consistency analysis to implicit preference mining, from cross-domain boundary identification to conflict coordination decision-making, and then to memory-enhanced model update, a complete technical chain of data → model → application → optimization is formed. It can capture explicit and implicit preferences, handle cross-modal and cross-domain preferences, coordinate multi-level preference conflicts, and have long-term memory and progressive learning capabilities. The present invention breaks through the limitations of traditional preference parsing, realizes a preference understanding mechanism that is closer to human cognition, improves the personalized service quality of intelligent systems, solves the core challenges faced by current preference parsing systems, and provides a technical foundation for building a truly user-centric artificial intelligence system.

[0152] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A user preference parsing method based on a memory-enhanced multimodal large model, characterized in that: The following steps are involved: Collect and preprocess multimodal interaction data to generate standardized feature vector sets and cleaned multimodal interaction sequences; Based on the cleaned multimodal interaction sequence, the temporal preference hierarchy is identified and classified to generate a multi-level temporal preference classification result and a temporal preference conversion rule set; Based on the multi-level temporal preference classification results and temporal preference conversion rule set, multi-level preference conflict coordination and decision-making are carried out to generate a unified preference decision result; Based on the unified preference decision results, the memory-enhanced preference model is updated and applied to generate an updated user preference model, which is then applied to the interactive interface and content generation to obtain personalized recommendation results.

2. The method according to claim 1, characterized in that The temporal preference hierarchy identification and classification includes the steps of constructing an adaptive time decay model. The steps of constructing an adaptive time decay model include: Based on the cleaned multimodal interaction sequence, a preference time distribution map is constructed, the preference expression time features are extracted, and the preference time feature vector is generated; Perform cluster analysis on the preferred time feature vectors to obtain the preferred time behavior categories; calculating category-specific decay parameters for the preferred time behavior categories and generating a category decay parameter set; Constructing a mixed attenuation function based on the category attenuation parameter set and generating an attenuation function template; Based on the preferred time feature vector, preferred time behavior category and attenuation function template, personalized parameter learning and tuning are performed to generate personalized attenuation parameters; A dynamic adjustment mechanism is implemented based on personalized decay parameters to generate an adaptive time decay model.

3. The method according to claim 2, characterized in that The steps of extracting the time characteristics of preference expression and performing cluster analysis to obtain the preference time behavior categories include: Extract frequency periodicity features, continuous stability features, repetition interval features and recent activity features from the preference time distribution map, and combine them to form a preference time feature vector; Apply K-means++ clustering algorithm to the preference time feature vector to generate initial clustering results; Based on the initial clustering results, the optimal number of clusters was evaluated by the silhouette coefficient and Davies-Bouldin index to generate the preferred time behavior categories.

4. The method according to claim 2, characterized in that: The steps of constructing a mixed attenuation function and generating an attenuation function template include: The structure of the hybrid function is determined based on the category decay parameter set, including three modes: exponential decay, hyperbolic decay and power law decay; Set the weight coefficient and attenuation rate parameters to construct a parameterized mixed attenuation function: D(t, c, p) = w1(c)*exp(-α(p)*t)+w2(c)*(1 / (1+β(p)*t))+w3(c)*(t (-γ(p)) ), where t represents the time variable, c represents the preference category, p represents the personalization parameter, w1, w2, w3 are weight coefficients and satisfy w1+w2+w3=1, α, β, γ are decay rate parameters; Construct a decay function template based on a parameterized mixed decay function.

5. The method according to claim 2, characterized in that: The steps of performing personalized parameter learning and tuning and implementing a dynamic adjustment mechanism to generate an adaptive time decay model include: Based on the preference time feature vector, preference time behavior category and decay function template, a Bayesian optimization algorithm is used to learn personalized decay parameters on historical data and minimize the error between the predicted preference strength and the real user feedback; A responsive adjustment mechanism is built based on personalized decay parameters. The personalized decay parameters are continuously fine-tuned through online learning methods according to real-time user feedback data to form an adaptive time decay model with self-adjustment capabilities. The performance of the adaptive time decay model is verified on the reserved test set of historical data, and the prediction accuracy, recall rate and F1 score are calculated to generate the verification results. The parameters of the adaptive time decay model are finally calibrated according to the verification results to generate the optimized adaptive time decay model.

6. The method according to claim 1, characterized in that After generating the multi-level temporal preference classification results and the temporal preference conversion rule set, it also includes: Conduct cross-modal preference consistency analysis based on standardized feature vector sets and generate a cross-modal preference consistency report; Based on the cleaned multimodal interaction sequence, implicit preference extraction and verification are performed to generate an implicit preference mapping table; Based on the results of multi-level temporal preference classification, cross-domain preference boundary identification and migration rule generation are performed to obtain a cross-domain preference migration rule set.

7. The method according to claim 6, characterized in that The cross-modal preference consistency analysis includes the steps of constructing a semantic-visual alignment space. The steps of constructing a semantic-visual alignment space include: Extract text semantic features and image semantic features from the standardized feature vector set; Based on the cleaned multimodal interaction sequences, temporally and semantically related text-image pairs are selected to construct a curated parallel corpus. A dual-tower neural network model is constructed using text semantic features, image semantic features, and a selected parallel corpus to learn shared latent spaces and generate a cross-modal alignment mapping network. Verify and fine-tune the semantic consistency of the cross-modal alignment mapping network to generate an improved cross-modal alignment mapping network; Based on the improved cross-modal alignment mapping network, a preference expression-specific adaptation layer is constructed to generate a preference-enhanced cross-modal mapping network, i.e., a semantic-visual alignment space. The preference-enhanced cross-modal mapping network is evaluated and optimized to generate a cross-modal alignment mapping function.

8. The method according to claim 7, characterized in that Shared latent space learning includes the steps of constructing a cross-modal contrastive loss function. The steps of constructing a cross-modal contrastive loss function include: Construct a global alignment loss component to ensure that the overall representation distributions of different modalities are similar; Construct a local semantic consistency loss component to ensure that the distance between semantically similar content in the cross-modal space is within a preset range; Constructing a structure-preserving loss component to ensure that the semantic structure within the modality is preserved after mapping; The global alignment loss component, the local semantic consistency loss component and the structure preservation loss component are combined to form a cross-modal contrast loss function; A dual-tower neural network model is trained using a cross-modal contrast loss function, text semantic features, image semantic features, and a selected parallel corpus to generate a cross-modal alignment mapping network.

9. The method according to claim 7, characterized in that: The steps of constructing a preference expression-specific adaptation layer and generating a preference-enhanced cross-modal mapping network include: Based on the improved cross-modal alignment mapping network, a dedicated adaptation layer is constructed for the special properties of preference expression to form a preference-specific adaptation layer; Use preference-specific adaptation layers to strengthen preference-related semantic features and weaken irrelevant features to generate enhanced preference expression features; Based on the enhanced preference expression features, the preference expression data is extracted and the preference specific adaptation layer is fine-tuned to obtain the fine-tuning parameters; Based on the improved cross-modal alignment mapping network, preference-specific adaptation layer and fine-tuning parameters, the migration from the general semantic space to the preference-specific semantic space is achieved to generate a preference-enhanced cross-modal mapping network.

10. The method according to claim 6, characterized in that Implicit preference extraction and verification includes the steps of constructing a low-intrusion verification strategy. The steps of constructing a low-intrusion verification strategy include: Based on the cleaned multimodal interaction sequence, generate implicit preference hypothesis set and conduct verification demand analysis and classification to generate classified preference hypothesis; Based on the classified preference hypothesis, natural dialogue fusion verification templates, parameter fine-tuning reaction observation schemes, and alternative scheme comparison strategies are constructed to generate dialogue fusion verification template libraries, parameter fine-tuning verification scheme sets, and alternative scheme comparison verification strategies. Based on the dialogue fusion verification template library, parameter fine-tuning verification solution set and alternative solution comparison verification strategy, a multi-round progressive verification plan is constructed to generate a timing verification plan; Construct an adaptive control mechanism for the verification process based on the timing verification plan and generate verification control parameters; Integrate the dialogue fusion verification template library, parameter fine-tuning verification solution set, alternative solution comparison verification strategy, timing verification plan and verification control parameters to form a low-interference verification strategy.

Images