A digital human interaction method based on time-frequency-domain frequency analysis model reasoning

Abstract

The application discloses a kind of digital human interaction methods based on time-frequency-domain frequency analysis model inference, it is related to digital human interaction technical field.A kind of digital human interaction methods based on time-frequency-domain frequency analysis model inference, including have: real-time data acquisition, interactive constraint analysis and digital human interaction optimization.The application is by constructing interactive behavior state sequence, the time-frequency state unit and domain frequency state unit in continuous interactive round are carried out association analysis, reflect the current interactive stage of user;On the basis of identifying current user interaction state, further extract the user feedback behavior after digital human reply, form user interaction habit mode, and the interaction habit mode is as important basis of interaction inference, so that digital human can generate corresponding reply strategy under the same user interaction state for different interaction habits, improve the matching degree of reply mode and user interaction preference.

Description

Technical Field

[0001] This invention relates to the field of digital human interaction technology, and in particular to a digital human interaction method based on time-frequency-domain frequency analysis to construct a model for reasoning. Background Technology

[0002] With the development of digital human technology, digital humans have been widely used in scenarios such as intelligent customer service, online consultation and e-commerce shopping guide. Most existing digital humans generate responses based on the user's current inquiry content, historical dialogue records or user intent. Although they can meet basic interaction needs, they generally focus on semantic understanding of the current round and lack analysis of the user's continuous interactive behavior. It is difficult to comprehensively characterize the evolution of user behavior by combining the time-frequency and domain-frequency characteristics of user inquiries.

[0003] Meanwhile, existing technologies typically employ the same response strategy for users with similar behavioral characteristics but different interaction habits, resulting in insufficient personalization capabilities and a lack of mechanisms to build response constraints based on overall online user interaction feedback. Consequently, the continuity and relevance of response strategies need to be improved.

[0004] Therefore, it is necessary to propose a digital human interaction method based on time-frequency-domain frequency analysis to construct model reasoning, so as to improve the accuracy, continuity and adaptability of digital human interaction. Summary of the Invention

[0005] This invention aims to provide a digital human interaction method based on time-frequency-domain frequency analysis to construct a model for reasoning, thereby achieving dynamic optimization of digital human response strategies and improving the relevance, continuity, and stability of digital human interaction.

[0006] A digital human interaction method based on time-frequency-domain frequency analysis for model inference includes the following steps: Step S1: Real-time data acquisition; real-time acquisition of online user interaction data; the user interaction data package contains user interaction number, user query data and user interaction timestamp; For online users: Obtain corresponding historical interaction data based on user interaction ID; Generate corresponding interaction behavior state sequences based on historical interaction data and user interaction data; Output the current user interaction state and interaction habit pattern based on the interaction behavior state sequence; Predict the target time-frequency state and target domain-frequency state of the online user in the next interaction round based on the current user interaction state; Combine the current user interaction state, interaction habit pattern, target time-frequency state, and target domain-frequency state to obtain user interaction feedback features; Step S2: Interaction constraint analysis; For all current online users, obtain the online user set; Extract user interaction feedback features from the online user set to obtain the user interaction feedback feature set; Construct single-round joint response constraints for the digital human based on the user interaction feedback feature set; Step S3: Digital Human Interaction Optimization; Construct an interaction reasoning model based on single-round joint response constraints, and generate corresponding digital human interaction content based on the interaction reasoning model; Obtain new user interaction data generated by online users in response to the digital human interaction content, provide feedback correction to the single-round joint response constraints and interaction reasoning model, and complete the real-time update of digital human interaction.

[0007] As a preferred embodiment of the present invention, the specific steps for generating a corresponding sequence of interactive behavior states based on historical interaction data and user interaction data include: Extract user interaction timestamps from historical interaction data and user interaction data to sort all user query data; divide the sorted user query data according to the preset interaction round division rules to obtain several consecutive interaction rounds; For any consecutive interaction rounds, extract all user query data within the consecutive interaction rounds to form the corresponding target query set; for each target query set, calculate the query time interval and query time density based on the user interaction timestamp; perform feature extraction based on the query time interval and query time density to obtain the corresponding time-frequency state unit; identify the attribute domain of the target query set to obtain the corresponding product attribute domain; calculate the coverage ratio of the product attribute domain to the preset product attribute domain space to obtain the corresponding domain-frequency state unit; According to the time sequence of consecutive interaction rounds, the time-frequency state unit and the domain-frequency state unit are associated and combined to obtain the joint behavior state node; all joint behavior state nodes are combined to obtain the interaction behavior state sequence.

[0008] As a preferred embodiment of the present invention, the specific steps for outputting the current user interaction state and interaction habit mode based on the interaction behavior state sequence include: Evolutionary features are extracted based on the sequence of interactive behavior states to obtain behavioral evolution features; feature clustering is performed on the behavioral evolution features to obtain different interactive evolution stages; According to the arrangement order of the joint behavior state nodes in the interaction behavior state sequence, a preset number of joint behavior state nodes are extracted as the current behavior analysis window; feature matching is performed on the interaction state relationship of the joint behavior state nodes in the current behavior analysis window to obtain the current interaction evolution stage; the current interaction evolution stage is determined as the current user interaction state. In continuous interaction rounds, user inquiry data after receiving a digital human response is set as user feedback data, and the corresponding digital human response data is extracted. The user feedback data and the corresponding digital human response data are then combined to obtain user feedback data pairs. Feature extraction is performed on the user feedback data pairs to obtain the interaction habit patterns.

[0009] As a preferred embodiment of the present invention, the specific steps for predicting the target time-frequency state and target domain-frequency state of an online user in the next interaction round based on the current user interaction state include: Match the historical behavior evolution path corresponding to the current user interaction state in the interaction behavior state sequence to obtain the target behavior evolution path; combine the interaction habit pattern to make user-personalized corrections to the target behavior evolution path to obtain the corresponding predicted behavior evolution path. Extract the joint behavior state node corresponding to the next interaction round in the predicted behavior evolution path to obtain the predicted behavior state node; parse the predicted time-frequency state unit and the predicted domain-frequency state unit corresponding to the predicted behavior state node; determine the predicted time-frequency state unit as the target time-frequency state of the next interaction round, and determine the predicted domain-frequency state unit as the target domain-frequency state of the next interaction round.

[0010] As a preferred embodiment of the present invention, the specific steps for constructing a single-round joint response constraint for a digital human based on a set of user interaction feedback features include: Based on the user interaction feedback feature set, output the corresponding state feedback set, habit feedback set, time-frequency feedback set, and domain-frequency feedback set; perform correlation analysis on each state feedback set, habit feedback set, time-frequency feedback set, and domain-frequency feedback set, calculate the correlation influence relationship between different feedback sets, and generate joint feedback relationship nodes for each online user; construct a user feedback correlation network according to the joint feedback relationship nodes between each online user. Based on the user feedback association network, the current online user set is clustered to obtain several target user groups; the joint feedback features of the target user groups are output; and the corresponding response guidance targets are output based on the joint feedback features of the groups. Based on the response guidance objectives corresponding to each target user group, query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints are constructed. The constraint association priority is calculated based on the group joint feedback characteristics corresponding to each target user group, and a corresponding joint response constraint set is generated. Conflict detection is performed based on each joint response constraint set to obtain the single-round joint response constraints of the digital human.

[0011] As a preferred embodiment of the present invention, the specific steps for constructing an interactive reasoning model based on single-round joint response constraints and generating corresponding digital human interactive content based on the interactive reasoning model include: The single-round joint response constraints of the digital human are analyzed to obtain response constraint parameters. The response constraint parameters are organized according to query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints. The correlation between the response constraint parameters is established to construct an interactive reasoning model. Input the user query data from all user interaction data in the current interaction round into the interaction inference model, perform inference analysis on the user query data according to the response constraint parameters, generate the corresponding digital human interaction content, and output the digital human interaction content to online users to complete the digital human response in the current interaction round.

[0012] As a preferred embodiment of the present invention, the specific steps for conflict detection based on each joint response constraint set include: When there are constraint conflicts between different joint response constraint sets, constraint coordination rules are constructed based on the degree of change in user interaction state, the degree of change in target time-frequency state, and the degree of change in target domain-frequency state. Based on the constraint coordination rules, the joint response constraint sets with conflicts are coordinated and corrected to obtain the single-round joint response constraints of the digital human.

[0013] The present invention has the following advantages: 1. This invention constructs an interactive behavior state sequence and performs correlation analysis on time-frequency state units and domain-frequency state units in consecutive interaction rounds. It no longer relies solely on the content of a single round of queries, but instead identifies the current user interaction state by combining the continuous evolution process of user interaction behavior, accurately reflecting the user's current interaction stage. Based on the identification of the current user interaction state, it further extracts user feedback behavior after the digital human replies, forming a user interaction habit pattern. This interaction habit pattern is used as an important basis for interaction reasoning, enabling the digital human to generate corresponding response strategies for different interaction habits under the same user interaction state, thereby improving the matching degree between the response method and user interaction preferences.

[0014] 2. This invention is based on the historical behavior evolution path matching of interactive behavior state sequence, and combines interactive habit patterns to make personalized corrections to the behavior evolution path, predicting the target time-frequency state and target domain-frequency state of the next interaction round, so that the subsequent digital human response is not only generated based on the current user state, but can also be adjusted according to the development trend of user behavior, thereby improving the continuity of the response strategy. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a digital human interaction method based on time-frequency-domain frequency analysis for model construction and reasoning, as used in an embodiment of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this invention.

[0017] An example is a digital human interaction method based on time-frequency-domain frequency analysis to construct a model for reasoning, such as... Figure 1As shown, it includes the following steps: Step S1: Real-time data acquisition; real-time acquisition of online user interaction data; the user interaction data package contains user interaction number, user query data and user interaction timestamp; The system continuously monitors interaction requests initiated by each user terminal through a preset data collection interface. When a new inquiry is detected from a user, the system automatically records the corresponding interaction information and organizes it according to the user dimension to form user interaction data.

[0018] User interaction data includes user interaction IDs, user query data, and user interaction timestamps. The user interaction ID is a unique identifier assigned to each online user, used to distinguish the interaction process of different users and serving as an index for linking historical interaction data with current interaction data. User query data refers to the text, speech-to-text, or other parsable query content entered by the user during the current interaction process. It reflects the user's current consultation needs and serves as an important data source for subsequent analysis of user interaction behavior. The user interaction timestamp records the precise time information corresponding to each query sent by the user, and can be recorded using a unified time format such as year, month, day, hour, minute, second, and millisecond. It describes the chronological order and time interval of each user query.

[0019] To ensure data consistency in subsequent analysis, when user query data originates from different terminals or data formats, all user query data undergoes unified preprocessing. This preprocessing includes character encoding standardization, time format standardization, and anomaly cleaning. Character encoding standardization converts different encoding formats to a unified one; time format standardization converts different time representations to a unified time standard; and anomaly cleaning removes empty, duplicate, and obviously erroneous data records. After preprocessing, using the user interaction number as an index, the corresponding user query data and user interaction timestamps are stored in chronological order into the corresponding interaction dataset. This forms user interaction data that continuously reflects the user interaction process, providing a complete, continuous, and unified data foundation for subsequent historical interaction data association, interaction behavior state sequence construction, and interaction state prediction.

[0020] For online users: Obtain corresponding historical interaction data based on user interaction ID; Generate corresponding interaction behavior state sequences based on historical interaction data and user interaction data; Output the current user interaction state and interaction habit pattern based on the interaction behavior state sequence; Predict the target time-frequency state and target domain-frequency state of the online user in the next interaction round based on the current user interaction state; Combine the current user interaction state, interaction habit pattern, target time-frequency state, and target domain-frequency state to obtain user interaction feedback features; The specific steps for generating corresponding interaction behavior state sequences based on historical interaction data and user interaction data include: Extract user interaction timestamps from historical interaction data and user interaction data to sort all user query data; divide the sorted user query data according to the preset interaction round division rules to obtain several consecutive interaction rounds; Historical interaction data refers to all interaction records between the current online user and the digital human before this interaction, while user interaction data refers to the interaction records collected in real time during the current interaction. Together, they reflect the complete user interaction process. First, using the user interaction number as an index, the historical interaction data and user interaction data are merged, and the user interaction timestamp corresponding to each user query data is extracted.

[0021] All user query data are sorted in order of user interaction timestamps from earliest to latest to form a continuous time series. After sorting, the continuous time series is divided according to the preset interaction round division rules. An interaction round refers to a set of continuous interaction processes completed by a user with a digital human around the same consultation goal within a continuous time period.

[0022] In the specific division, the time interval between two adjacent user query data is used as the criterion. When the time interval between two adjacent user query data is less than the preset time interval threshold, they are divided into the same interaction round. When the time interval is greater than or equal to the preset time interval threshold, it is considered that the previous interaction round has ended, and the current user query data is used as the starting data of the new interaction round, thereby obtaining several consecutive interaction rounds arranged continuously in time, providing a unified time window for subsequent time-frequency state analysis and domain-frequency state analysis.

[0023] For any consecutive interaction rounds, extract all user query data within the consecutive interaction rounds to form the corresponding target query set; for each target query set, calculate the query time interval and query time density based on the user interaction timestamp; perform feature extraction based on the query time interval and query time density to obtain the corresponding time-frequency state unit; identify the attribute domain of the target query set to obtain the corresponding product attribute domain; calculate the coverage ratio of the product attribute domain to the preset product attribute domain space to obtain the corresponding domain-frequency state unit; The target query set refers to the data set consisting of all user query data within the same consecutive interaction round. For each consecutive interaction round, all user query data contained therein are extracted to form the corresponding target query set. The time interval between adjacent user queries is calculated based on the user interaction timestamps. The time interval is equal to the timestamp of the next user query minus the timestamp of the previous user query. At the same time, the duration of the current consecutive interaction round and the total number of user queries are counted. The query time density is calculated by dividing the total number of user queries by the duration to reflect the frequency of user queries in the current interaction round.

[0024] Feature extraction is performed based on the query time interval and query time density. Specifically, a normalization process can be used to convert the query time interval and query time density to a uniform numerical range, and then fuse them according to preset weights to obtain the corresponding time-frequency state unit. The time-frequency state unit is used to characterize the query rhythm and time continuity of the user in the current interaction round. The shorter the query time interval and the higher the query time density, the faster the user's interaction rhythm.

[0025] At the same time, attribute domain identification is performed on each user query data in the target query set. Specifically, natural language processing technology is used to perform semantic parsing on the user query content, extract the product attribute keywords involved, and map each keyword to the corresponding product attribute domain based on a pre-established product attribute domain dictionary. For example, different attribute categories such as price, specifications, color, delivery method, after-sales service, and functional parameters are obtained, thereby obtaining all product attribute domains involved in the current target query set.

[0026] The ratio of the number of product attribute domains involved in the current target query set to the total number of all attribute domains in the preset product attribute domain space is calculated to obtain the product attribute domain coverage ratio. The preset product attribute domain space is a standard attribute set pre-established based on the target product or business scenario, used to uniformly describe all possible product attribute types.

[0027] Based on the coverage ratio of product attribute domains, corresponding domain frequency state units are generated. The domain frequency state units are used to represent the breadth of attribute domains involved in the content that the user is interested in during the current interaction round. The higher the coverage ratio, the more dispersed the content that the user is interested in, and the lower the coverage ratio, the more concentrated the content that the user is interested in.

[0028] Time-frequency analysis is mainly used to characterize the rhythm and temporal continuity of user inquiries during continuous interaction. Its technical effect is not to judge the user's emotions or psychological state, but to more accurately reflect the user's current level of interactive activity and the pace of interaction, thereby providing a basis for digital humans to adjust their response methods.

[0029] This embodiment can identify changes in the user's interaction rhythm by combining the inquiry time interval and inquiry time density in consecutive interaction rounds. For example, in an e-commerce shopping guide scenario, when a user asks questions such as "Do you have stock?", "Can you ship today?", and "When will it arrive?" in a short period of time, the short time interval between each inquiry and the high inquiry time density indicate that the user wants to obtain decision-making information quickly. The digital human can prioritize integrating core information such as inventory, shipping, and logistics in a single response, reducing multiple rounds of confirmation. On the other hand, when the interval between each user's inquiry is longer, such as asking about product functions, usage methods, and after-sales policies every few minutes, it indicates that the user prefers to gradually understand the product information. The digital human can gradually unfold the explanation according to the user's inquiry rhythm, thereby improving the matching degree between the response method and the user's actual interaction rhythm.

[0030] Domain frequency analysis is mainly used to characterize the attribute domain coverage of content that users pay attention to during continuous interaction. Its technical effect is to identify whether the content that users are currently paying attention to is concentrated on a few attributes or scattered across multiple attribute domains, so that digital humans can organize their response content more reasonably.

[0031] This embodiment analyzes the scope of user attention by statistically analyzing the coverage ratio of product attribute areas involved in consecutive interaction rounds. For example, in the scenario of buying a mobile phone, if the user continuously inquires about "processor performance," "RAM," and "screen refresh rate," the product attribute area is mainly focused on performance, with a relatively small domain coverage. The digital human can then provide a more in-depth introduction around performance parameters. If the user subsequently inquires about "color," "price," "installment payment," "after-sales service," and "delivery," it indicates that the user's scope of attention has expanded to multiple product attribute areas, increasing the domain coverage. In this case, the digital human can comprehensively organize multiple attribute information in the response, rather than just answering the current question, thereby reducing repeated questions from the user and improving interaction efficiency.

[0032] According to the time sequence of consecutive interaction rounds, the time-frequency state unit and the domain-frequency state unit are associated and combined to obtain the joint behavior state node; all joint behavior state nodes are combined to obtain the interaction behavior state sequence.

[0033] Furthermore, this embodiment does not use time-frequency analysis and domain-frequency analysis independently, but combines the two to construct an interactive behavior state sequence. Time-frequency analysis reflects when the user asks questions and how quickly, while domain-frequency analysis reflects what aspects the user asks and how broad their focus is. These two aspects respectively describe the time and content dimensions of user interaction behavior. For example, in a home appliance sales scenario, if a user continuously inquires about price, discounts, delivery, and installation within one minute, it indicates that their interaction pace is not only fast but also covers a wide range of areas of interest, usually meaning the user has entered the purchase decision-making stage. Another user, while also inquiring about price, delivery, and installation, does so at longer intervals, suggesting they are more likely in the information gathering stage. Although the two types of users ultimately focus on similar attribute areas, due to their different time-frequency characteristics, this invention can identify differences in their interactive behaviors and provide a more comprehensive behavioral basis for subsequent interaction reasoning. This allows the digital human to organize its response content while considering both the product attributes the user is currently interested in and maintaining consistency with the user's actual interaction pace, thereby improving the matching degree between the digital human's response content and the user's actual interaction needs.

[0034] In this context, a joint behavioral state node refers to a behavioral state description unit composed of time-frequency state units and domain-frequency state units corresponding to the same consecutive interaction round. It is used to fully represent the user's interaction rhythm and focus within a given interaction round. Specifically, according to the chronological order of the consecutive interaction rounds, a one-to-one correspondence is established between the time-frequency state units and the corresponding domain-frequency state units for each consecutive interaction round, and these are combined to form joint behavioral state nodes. Each joint behavioral state node retains the temporal position of the corresponding interaction round, reflecting the evolution of user behavior throughout the interaction process. Based on the chronological order of the consecutive interaction rounds, all joint behavioral state nodes are sequentially connected to form an interaction behavior state sequence. This interaction behavior state sequence is not a simple data arrangement, but rather a trajectory of user behavior changes organized according to a continuous temporal relationship. It continuously reflects the evolutionary patterns of the user's inquiry rhythm and focus throughout the interaction process, providing a unified and continuous behavioral analysis foundation for subsequent user interaction state identification, interaction habit pattern analysis, and prediction of target time-frequency and target domain-frequency states.

[0035] The specific steps for outputting the current user interaction state and interaction habit pattern based on the interaction behavior state sequence include: Evolutionary features are extracted based on the sequence of interactive behavior states to obtain behavioral evolution features; feature clustering is performed on the behavioral evolution features to obtain different interactive evolution stages; The interactive behavior state sequence consists of multiple joint behavior state nodes arranged in chronological order. Each joint behavior state node contains a time-frequency state unit and a domain-frequency state unit for the corresponding interaction round, which can reflect the user's interaction rhythm and focus content in that interaction round.

[0036] The state changes between adjacent joint behavior state nodes are extracted according to the time sequence of the interaction behavior state sequence. The change amplitude of the time-frequency state unit and the change amplitude of the domain-frequency state unit between adjacent nodes are calculated respectively, and the direction of continuous change and the number of continuous rounds are counted to form the behavior evolution feature. The behavior evolution feature is used to describe the change trend of user interaction behavior in the continuous interaction process, such as the interaction rhythm gradually accelerating, the area of ​​attention gradually expanding, or remaining stable.

[0037] Feature clustering is performed on all behavioral evolution characteristics. In practice, cluster analysis methods can be used. First, each behavioral evolution characteristic is converted into a feature vector of a uniform dimension. Then, the similarity between each feature vector is calculated, and similar feature vectors are continuously merged based on the similarity until a preset number of clusters or a clustering termination condition is reached. Finally, several representative interaction evolution stages are obtained. Interaction evolution stages refer to a set of states in which user interaction behaviors have similar evolutionary patterns, such as the interaction exploration stage, the demand confirmation stage, and the decision consultation stage. Different interaction evolution stages correspond to different behavioral change patterns, providing a foundation for subsequent user interaction state identification.

[0038] According to the arrangement order of the joint behavior state nodes in the interaction behavior state sequence, a preset number of joint behavior state nodes are extracted as the current behavior analysis window; feature matching is performed on the interaction state relationship of the joint behavior state nodes in the current behavior analysis window to obtain the current interaction evolution stage; the current interaction evolution stage is determined as the current user interaction state. The current behavior analysis window refers to a data window composed of several joint behavior state nodes continuously extracted from the end of the interaction behavior state sequence. It is used to reflect the changes in the user's interaction behavior over a recent period. In specific implementation, according to a preset number of nodes, several recent joint behavior state nodes are extracted from the interaction behavior state sequence in chronological order as the current behavior analysis window, and the time-frequency state unit and domain-frequency state unit corresponding to each joint behavior state node are extracted respectively. The change relationship between the joint behavior state nodes in the current behavior analysis window is compared one by one with the behavior evolution characteristics corresponding to the aforementioned interaction evolution stages. Specifically, this includes comparing whether the time-frequency state change trends are consistent, whether the domain-frequency state change trends are consistent, and whether the joint change order is consistent, and calculating the matching degree between each interaction evolution stage and the current behavior analysis window. When the matching degree corresponding to a certain interaction evolution stage reaches a preset matching threshold and is the highest value among all interaction evolution stages, the interaction evolution stage is determined as the current interaction evolution stage corresponding to the current behavior analysis window, and the current interaction evolution stage is further regarded as the current user interaction state. The current user interaction state is used to reflect the development stage of the user's current continuous interaction behavior, providing a basis for predicting the user's behavior changes in the next interaction round.

[0039] In continuous interaction rounds, user inquiry data after receiving a digital human response is set as user feedback data, and the corresponding digital human response data is extracted. The user feedback data and the corresponding digital human response data are then combined to obtain user feedback data pairs. Feature extraction is performed on the user feedback data pairs to obtain the interaction habit patterns.

[0040] User feedback data refers to the data from a new round of inquiries initiated by the user after the digital human has completed a response; it reflects the user's feedback on the previous round of responses. Digital human response data refers to the response content output by the digital human in the corresponding round of interaction.

[0041] In practice, in each round of continuous interaction, the completion of the digital human's reply is used as the dividing point. The first and subsequent user inquiry data after the reply are extracted as user feedback data. At the same time, the corresponding digital human reply data before the generation of the user feedback data is extracted. Each set of user feedback data is paired with the corresponding digital human reply data according to the time correspondence to form user feedback data pairs.

[0042] Feature extraction is performed on each pair of user feedback data. Specifically, this includes extracting the product attribute domains involved in the digital human's reply content, the coverage of the reply information, the length of the reply content, as well as the changes in the inquiry rhythm, attribute focus, and repeated inquiries in the corresponding user feedback data. These features are then numerically represented to form feedback behavior characteristics.

[0043] Statistical analysis is performed on all feedback behavior characteristics to identify recurring feedback patterns in long-term user interactions. For example, users tend to ask for more details after receiving a concise reply, or they tend to end the consultation after receiving a comprehensive reply. When a certain type of feedback pattern reaches a preset frequency or proportion in historical interactions, the feedback pattern is identified as the corresponding user's interaction habit pattern. The interaction habit pattern describes the stable interaction preferences formed by users when faced with different reply methods, providing a basis for subsequent behavior prediction and adjustment of digital human reply strategies.

[0044] The specific steps for predicting the target time-frequency state and target domain-frequency state of online users in the next interaction round based on the current user interaction state include: Match the historical behavior evolution path corresponding to the current user interaction state in the interaction behavior state sequence to obtain the target behavior evolution path; combine the interaction habit pattern to make user-personalized corrections to the target behavior evolution path to obtain the corresponding predicted behavior evolution path. Historical behavior evolution path refers to the continuous state change trajectory composed of multiple joint behavior state nodes in chronological order during historical interaction, which can reflect the entire process of user interaction state gradually evolving from the initial stage to the subsequent stage; target behavior evolution path refers to the historical behavior change trajectory that is closest to the current user interaction state; predictive behavior evolution path is a predictive path formed after combining individual user interaction habits, which is more in line with the actual behavior characteristics of the current user.

[0045] In practice, the current user interaction state is used as the retrieval condition. All historical behavior evolution paths starting with the same interaction state are searched within the interaction behavior state sequence. The similarity between each historical behavior evolution path and the current interaction behavior state sequence is calculated. The similarity is calculated comprehensively based on the arrangement order of the joint behavior state nodes, the time-frequency state change trend, and the domain-frequency state change trend. The more consistent these three trends are, the closer the corresponding historical behavior evolution path is to the current user behavior. The historical behavior evolution path with the highest similarity is selected as the target behavior evolution path. Subsequently, interaction habit patterns are introduced to personalize the target behavior evolution path.

[0046] Specifically, based on interaction habit patterns, stable interaction preferences formed by users over a long period of time are identified. For example, whether users tend to focus on the same product attribute area, whether they are accustomed to asking questions quickly and continuously, and whether they tend to gradually expand the content of their inquiries. The change trend of subsequent joint behavior state nodes in the target behavior evolution path is adjusted according to these interaction preferences. When the interaction habit pattern shows that the user has the characteristic of continuously focusing on a single attribute area, the domain frequency state change amplitude in the target behavior evolution path is reduced. When the interaction habit pattern shows that the user has the characteristic of asking questions quickly and continuously, the continuity of subsequent time frequency state change trends is increased.

[0047] After the correction is completed, the corresponding predicted behavior evolution path is obtained, so that the prediction results can both inherit the historical behavior evolution law and reflect the current user's long-term interaction habits, thereby improving the consistency between subsequent behavior predictions and the user's actual interaction behavior.

[0048] Extract the joint behavior state node corresponding to the next interaction round in the predicted behavior evolution path to obtain the predicted behavior state node; parse the predicted time-frequency state unit and the predicted domain-frequency state unit corresponding to the predicted behavior state node; determine the predicted time-frequency state unit as the target time-frequency state of the next interaction round, and determine the predicted domain-frequency state unit as the target domain-frequency state of the next interaction round.

[0049] The predicted behavior state node refers to the joint behavior state node corresponding to the next interaction round in the predicted behavior evolution path. It contains both time-frequency state unit and domain-frequency state unit information. In specific implementation, the first joint behavior state node after the current user interaction state is located according to the time sequence of the predicted behavior evolution path, and it is used as the predicted behavior state node.

[0050] The predicted behavior state nodes are parsed, and the time-frequency state units and domain-frequency state units contained therein are extracted respectively. The predicted time-frequency state units are used to reflect the interaction rhythm and temporal continuity that the user may have in the next interaction round, such as predicting whether the user will still maintain a short inquiry interval and a high inquiry time density. The predicted domain-frequency state units are used to reflect the coverage of product attribute domains that the user may focus on in the next interaction round, such as predicting whether the user will continue to consult in depth about the current product attributes, or further expand to new product attribute domains such as price, delivery, and after-sales service.

[0051] After parsing, the predicted time-frequency state unit is directly determined as the target time-frequency state for the next interaction round, and the predicted domain-frequency state unit is determined as the target domain-frequency state for the next interaction round. The target time-frequency state and the target domain-frequency state are not the actual interaction behaviors that the user has already generated, but rather the target behavior states predicted based on the historical behavior evolution patterns and user interaction habits. They are used to guide the digital human to organize information that matches the user's next interaction needs in advance when generating response content. This allows the digital human's response to not only respond to the current query, but also take into account the possible behavioral changes that the user may produce in the next interaction round, thereby improving the continuity and adaptability between the response content and the user's continuous interaction process.

[0052] Step S2: Interaction constraint analysis; For all current online users, obtain the online user set; Extract user interaction feedback features from the online user set to obtain the user interaction feedback feature set; Construct single-round joint response constraints for the digital human based on the user interaction feedback feature set; The online user set refers to the set of all users currently interacting with the digital human. Each online user corresponds to a user interaction feedback feature. The user interaction feedback feature is a comprehensive behavioral description result obtained from the previous steps. It is composed of the current user interaction state, interaction habit pattern, target time-frequency state, and target domain-frequency state, and is used to comprehensively reflect the user's current interaction needs and the development trend of the next interaction round.

[0053] In practice, the system first counts all users currently in an interactive state in real time and establishes an online user set based on the user interaction number. Then, it extracts the user interaction feedback features corresponding to each online user and organizes them according to a unified data structure to form a user interaction feedback feature set. Among them, each user interaction feedback feature is stored in association according to the user interaction number, so that the feedback features of different users remain independent of each other, while providing a unified data input for subsequent analysis of common relationships between different users.

[0054] The specific steps for constructing single-round joint response constraints for digital humans based on user interaction feedback feature sets include: Based on the user interaction feedback feature set, output the corresponding state feedback set, habit feedback set, time-frequency feedback set, and domain-frequency feedback set; perform correlation analysis on each state feedback set, habit feedback set, time-frequency feedback set, and domain-frequency feedback set, calculate the correlation influence relationship between different feedback sets, and generate joint feedback relationship nodes for each online user; construct a user feedback correlation network according to the joint feedback relationship nodes between each online user. The status feedback set refers to the data set consisting of the current user interaction status of all online users; the habit feedback set refers to the data set consisting of the interaction habit patterns of all online users; the time-frequency feedback set refers to the data set consisting of the target time-frequency status of all online users; and the domain-frequency feedback set refers to the data set consisting of the target domain-frequency status of all online users.

[0055] In practice, the four types of feedback information are first extracted from the user interaction feedback feature set to form corresponding data sets. Correlation analysis is then performed on the four types of feedback sets. Specifically, each online user is taken as the analysis object, and the similarity between their current user interaction state, interaction habit pattern, target time-frequency state, and target domain-frequency state and other online users is compared. The correlation and influence relationship between each type of feedback information is then calculated.

[0056] The correlation influence relationship describes the degree of common change among different feedback features. It can be implemented using a weighted similarity calculation method, which involves calculating the similarity of the four types of feedback features separately, and then performing a comprehensive calculation based on preset weights to obtain the correlation influence value between any two online users. Each online user and its corresponding correlation influence relationship are organized into a joint feedback relationship node. Each joint feedback relationship node records not only the user's own feedback features but also the degree of correlation between the user and other online users. Using the joint feedback relationship nodes as network nodes and the correlation influence value as the connection relationship between nodes, a user feedback correlation network is constructed. This makes the entire set of online users a data network that reflects the correlation relationships between user feedback, providing a foundation for subsequent target user group segmentation.

[0057] Based on the user feedback association network, the current online user set is clustered to obtain several target user groups; the joint feedback features of the target user groups are output; and the corresponding response guidance targets are output based on the joint feedback features of the groups. The target user group refers to a set of online users whose interaction feedback characteristics are highly similar; the group joint feedback characteristics are the overall characteristics obtained by statistically analyzing the feedback characteristics of all users within the same target user group, used to describe the common interaction characteristics of the group; the response guidance goal is the response direction that the digital human should focus on in the current interaction round.

[0058] In practice, the nodes of the joint feedback relationship in the user feedback association network are used as clustering objects, and feature clustering is performed based on the magnitude of the association influence value between nodes. During the clustering process, online users with larger association influence values ​​are prioritized to be assigned to the same target user group, and the node merging process is repeated until all joint feedback relationship nodes have been grouped.

[0059] Statistical analysis is performed on the current user interaction status, interaction habit patterns, target time-frequency status, and target domain-frequency status of all online users in each target user group. The frequency of occurrence of each type of feedback feature in the current group is calculated, and the feedback feature with the highest frequency or the largest proportion is selected as the group joint feedback feature of the target user group.

[0060] Based on the analysis of the shared feedback characteristics of the current group of users, the system identifies their common interaction needs and generates corresponding response guidance goals. For example, if a target user group primarily engages in fast-paced interactions and focuses on product prices and delivery information, the response guidance goal would be to quickly provide price and delivery details. Conversely, if another target user group focuses on product performance and maintains a slower interaction pace, the response guidance goal would be to provide a detailed description of product performance. This ensures that the digital human's responses better align with the overall needs of the current user group.

[0061] Based on the response guidance objectives corresponding to each target user group, query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints are constructed. The constraint association priority is calculated based on the group joint feedback characteristics corresponding to each target user group, and a corresponding joint response constraint set is generated. Conflict detection is performed based on each joint response constraint set to obtain the single-round joint response constraints of the digital human.

[0062] Among them, the query response constraint is used to specify the user query content that the digital human should respond to first; the time-frequency guidance constraint is used to specify the adaptation requirements between the digital human's response content and the user's predicted interaction rhythm; the domain-frequency guidance constraint is used to specify the product attribute domain that the digital human's response content should cover; and the interaction habit adaptation constraint is used to specify the matching requirements between the digital human's response method and the user's long-term interaction habits.

[0063] In practice, the four types of response constraints are first established based on the response guidance objectives of each target user group, and then organized into a joint response constraint set for the corresponding target user group. The importance of the four types of response constraints is calculated based on the joint feedback characteristics of the group. Specifically, the importance can be quantified based on the proportion of each feedback characteristic in the target user group. The higher the proportion, the higher the importance of the corresponding response constraint, and the constraint association priority is generated accordingly.

[0064] In practice, for the query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints in the target user group, the occurrence ratio of the corresponding feedback features in the target user group is statistically analyzed, and the consistency degree of each feedback feature within the group is calculated. Among them, the occurrence ratio is used to represent the proportion of online users with the same feedback features to the total number of users in the target user group, and the consistency degree is used to represent whether the feedback changes among online users with the same feedback features are consistent, which can be calculated based on the similarity between the corresponding feedback features of each online user.

[0065] When the same feedback feature not only has a high occurrence rate but also a high degree of consistency, it indicates that the feedback feature can represent most of the interaction needs of the target user group. The occurrence rate and the degree of consistency are weighted and fused according to preset weights to obtain the constraint association value of the corresponding response constraint. The constraint association values ​​are sorted from largest to smallest to obtain the constraint association priority of each response constraint. The larger the constraint association value, the stronger the representativeness of the corresponding response constraint in the current target user group, and the higher its constraint association priority. It is retained first in the subsequent joint response constraint construction process.

[0066] The specific steps for conflict detection based on each joint response constraint set include: When there are constraint conflicts between different joint response constraint sets, constraint coordination rules are constructed based on the degree of change in user interaction state, the degree of change in target time-frequency state, and the degree of change in target domain-frequency state. Based on the constraint coordination rules, the joint response constraint sets with conflicts are coordinated and corrected to obtain the single-round joint response constraints of the digital human.

[0067] It should be noted that constraint conflict refers to different joint response constraint sets proposing inconsistent response requirements for the same type of user inquiry. For example, one joint response constraint set requires the digital human to prioritize detailed description of product performance, while another joint response constraint set requires priority description of product price; or one joint response constraint set requires a concise and quick response, while another joint response constraint set requires a complete and detailed response. In such cases, a constraint conflict is considered to exist.

[0068] To reasonably coordinate different response requirements, we first calculate the degree of change in user interaction status, the degree of change in target time-frequency status, and the degree of change in target domain-frequency status for each target user group.

[0069] Among them, the degree of change in user interaction state reflects the magnitude of change in the current interaction state of the target user group relative to the previous interaction round. It can be calculated by comparing whether the interaction evolution stages corresponding to two consecutive rounds of interaction have changed and the number of changes. The more changes, the higher the degree of change. The degree of change in target time-frequency state reflects the magnitude of change in the predicted interaction rhythm of the target user group relative to the current interaction rhythm. It can be calculated by comparing the change ratio of the query time interval and query time density between the target time-frequency state unit and the current time-frequency state unit. The larger the change ratio, the more obvious the change in the user's interaction rhythm in the next round. The degree of change in target domain-frequency state reflects the degree to which the content of user attention has expanded or contracted relative to the current interaction round. It can be calculated by comparing the difference between the coverage ratio of the corresponding product attribute domains of the target domain-frequency state unit and the current domain-frequency state unit. The larger the change in the coverage ratio, the more obvious the change in the content of user attention. The three levels of change are normalized to a uniform numerical range and then weighted and fused according to preset weights to obtain the comprehensive change value for the corresponding target user group. Constraint coordination rules are established based on the magnitude of the comprehensive change value for each target user group. When a joint response constraint set with a larger comprehensive change value conflicts with a joint response constraint set with a smaller comprehensive change value, the joint response constraint with the larger comprehensive change value is retained first. When the comprehensive change values ​​are similar, the joint response constraint with the larger degree of change in user interaction state is retained first. If the degree of change in user interaction state is still the same, the degree of change in target time-frequency state and target domain-frequency state are compared sequentially, and the joint response constraint with the larger degree of change is selected as the priority coordination object, thus forming a unified constraint coordination rule.

[0070] Step S3: Digital Human Interaction Optimization; Construct an interaction reasoning model based on single-round joint response constraints, and generate corresponding digital human interaction content based on the interaction reasoning model; Obtain new user interaction data generated by online users in response to the digital human interaction content, provide feedback correction to the single-round joint response constraints and interaction reasoning model, and complete the real-time update of digital human interaction.

[0071] The specific steps for constructing an interactive reasoning model based on single-round joint response constraints and generating corresponding digital human interactive content based on the interactive reasoning model include: The single-round joint response constraints of the digital human are analyzed to obtain response constraint parameters. The response constraint parameters are organized according to query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints. The correlation between the response constraint parameters is established to construct an interactive reasoning model. The interactive content of digital humans includes not only the output text content, but also the adjustable facial expressions, tone of voice, and text expression methods of digital humans during the interaction process; facial expressions are used to reflect the way digital humans display their emotions, tone of voice is used to reflect the strength, pace, and tone of the digital human's response, and text expression methods are used to reflect the level of detail, structure, and wording of the response content.

[0072] Interactive reasoning models refer to response generation models built based on single-round joint response constraints. They are used to perform correlation analysis between user query data and response constraint parameters, and generate corresponding digital human interactive content.

[0073] In practice, the user query data and corresponding single-round joint response constraints are first obtained in the current interaction round. Based on the query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints in the single-round joint response constraints, the text content, expressions, and tone of the digital human are uniformly controlled. The user query data is input into the interaction inference model, which outputs the corresponding digital human interaction content and synchronously outputs the generated text, expressions, and tone to the online user to complete the digital human's response in the current interaction round.

[0074] The response constraint parameters are a set of parameters obtained by structured parsing of the single-round joint response constraints. They are used to describe the response requirements that the digital human should meet in the current interaction round. In specific implementation, the single-round joint response constraints are first parsed, and the control information corresponding to the query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints are extracted respectively.

[0075] Among them, the query response constraint is used to determine the user query content that the digital human needs to answer first; the time-frequency guidance constraint is used to determine the rhythm, tone, and duration of the digital human's response; the domain-frequency guidance constraint is used to determine the product attribute domain or business information scope that the digital human's response needs to cover; and the interaction habit adaptation constraint is used to determine the level of detail, text style, expression display method, and tone of voice of the digital human's response.

[0076] The above response constraint parameters are organized according to a unified data structure, and the relationships between the parameters are established. The relationships are used to describe the synergistic effect between different response constraint parameters. For example, when the frequency-guided constraint requires a quick response, the text expression should be appropriately concise; when the interaction habit adaptation constraint requires a detailed explanation, the text content should be appropriately expanded; when the domain frequency-guided constraint requires coverage of multiple attribute domains, the response content should be organized according to the attribute domains.

[0077] After completing the parameter organization, the response constraint parameters and their relationships are configured into the preset interactive reasoning model to form a target interactive reasoning model suitable for the current interaction round, enabling the model to control the three interactive targets of text content, facial expression and tone of voice simultaneously.

[0078] Input the user query data from all user interaction data in the current interaction round into the interaction inference model, perform inference analysis on the user query data according to the response constraint parameters, generate the corresponding digital human interaction content, and output the digital human interaction content to online users to complete the digital human response in the current interaction round.

[0079] In practice, the user query data in the current interaction round is first extracted and input into the target interaction reasoning model. After receiving the user query data, the interaction reasoning model first performs semantic parsing on the user query content to identify the product attribute domain, consultation intent, and query focus that the user is currently interested in. The parsing results are then used for reasoning analysis in conjunction with the response constraint parameters.

[0080] Specifically, the core questions that need to be answered first are determined based on the query response constraints; the response rhythm, tone speed, and response length are determined based on the time-frequency guidance constraints; the attribute domains that need to be covered are determined based on the domain-frequency guidance constraints; and the text expression method, emoticon display method, and tone expression method are determined based on the interaction habit adaptation constraints.

[0081] After completing the above analysis, the interaction reasoning model generates corresponding digital human interaction content, including text response content, corresponding facial expression parameters, and corresponding tone parameters. The generated text content, facial expression parameters, and tone parameters are synchronously output to online users, enabling the digital human to complete the response of the current interaction round with the corresponding text, facial expression, and tone, thereby improving the matching degree between the digital human's response content and the user's actual interaction needs.

[0082] New user interaction data refers to subsequent inquiry data generated by online users after the digital human completes its current round of responses. This includes new user inquiry content, corresponding user interaction timestamps, and user interaction numbers. In practice, the system first acquires new user interaction data after online users receive the digital human's interaction content and then performs correlation analysis with the digital human's interaction content from the current round.

[0083] The time-frequency state unit, domain-frequency state unit, and current user interaction state corresponding to the new user interaction data are reconstructed and compared with the target time-frequency state and target domain-frequency state predicted in the current round to obtain the state deviation result; the correspondence between the user's new inquiry content and the digital human's response content is analyzed to determine whether the digital human's response meets the user's current needs, and whether the digital human's text, expression, and tone match the user's interaction habits.

[0084] Based on the state deviation results and matching results, the single-round joint response constraints are corrected, such as adjusting the query response priority, time-frequency guidance method, domain-frequency coverage, and interaction habit adaptation method; the corrected response constraint parameters are then reconfigured into the interaction inference model, updating the interaction inference model; after the update is completed, the digital human uses the updated single-round joint response constraints and interaction inference model in subsequent interaction rounds to generate new digital human interaction content, thereby realizing real-time updates and continuous optimization of the digital human interaction process.

[0085] In this embodiment, all calculations, analyses, predictions, and simulations are performed on a framework combining rule analysis and statistical learning. The collected user interaction data undergoes preprocessing, including anomaly removal, missing data completion, and time-consistent processing. Specifically, missing user interaction timestamps are filled using adjacent timestamp interpolation; duplicate queries are deduplicated based on user interaction numbers and time order; and abnormal time data is corrected according to the principle of time continuity. After preprocessing, the user interaction data is sorted by time and divided into consecutive interaction rounds according to preset rules, providing a unified data foundation for subsequent analysis.

[0086] In the time-frequency state analysis, the user interaction timestamps in consecutive interaction rounds are used as the analysis object. The query time interval between two adjacent user queries is calculated, and the number of queries per unit time is further calculated to obtain the query time density. Interaction rhythm features are extracted through statistical analysis methods to form time-frequency state units. In the domain-frequency state analysis, semantic parsing of user query content is performed. Word segmentation, keyword extraction, and text vector matching methods can be used to identify the product attribute domains that the user is currently interested in. Based on the identification results, the types of product attribute domains involved in the current interaction round are counted, and the coverage ratio of product attribute domains is calculated to form domain-frequency state units. The time-frequency state units and domain-frequency state units are associated and combined in chronological order to form joint behavioral state nodes, which further constitute an interaction behavior state sequence.

[0087] For behavioral evolution analysis, this embodiment uses a combination of sliding window analysis and cluster analysis. The current behavioral analysis window is composed of several consecutive joint behavioral state nodes. The state change trend between the joint behavioral state nodes within the window is extracted to obtain behavioral evolution features. Cluster analysis algorithms are used to classify the behavioral evolution features. Specifically, mature algorithms such as hierarchical clustering, mean clustering, or density clustering can be used to divide behavioral features with similar evolution trends into the same interactive evolution stage.

[0088] When it is necessary to predict the user's behavior in the next round of interaction, a historical path matching algorithm is used to compare the similarity between the current interaction behavior state sequence and the historical behavior evolution path one by one. Euclidean distance, cosine similarity or dynamic time warping methods can be used to calculate the similarity between paths. The historical behavior evolution path with the highest similarity is selected as the target behavior evolution path, and the path is corrected by combining the user's interaction habit pattern to obtain the predicted behavior evolution path, thereby predicting the target time-frequency state and the target domain-frequency state.

[0089] For user feedback correlation analysis, this embodiment first constructs a status feedback set, a habit feedback set, a time-frequency feedback set, and a domain-frequency feedback set, respectively. Then, it uses a feature correlation analysis method to calculate the correlation degree between each feedback set. In specific implementation, the similarity between each feedback feature can be calculated separately, and weighted fusion can be performed according to preset weights to obtain feedback correlation values. Then, a user feedback correlation network is established based on the feedback correlation values. On the basis of the user feedback correlation network, a clustering analysis algorithm is used to complete the division of target user groups, and query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints are established based on the joint feedback features of the groups. An interaction reasoning model is established through a rule-based reasoning method. The user query content, response constraint parameters, and response control rules are input into the reasoning model. The reasoning model sequentially completes the organization of response content, adjustment of response rhythm, adjustment of product attribute domain coverage, and control of digital human expression, tone, and text style, and finally generates digital human interaction content. When new user interaction data is generated, feedback correction is completed based on the deviation between the predicted results and the actual results to realize the real-time update of the digital human interaction model.

[0090] To address the issue of data unit unification in this embodiment, a unified data unit mapping relationship is first established. Specifically, time is uniformly represented by seconds; query time intervals are uniformly represented by seconds; query time density is uniformly represented by the number of queries per unit time; product attribute domain coverage ratios are uniformly represented by dimensionless ratio values, ranging from zero to one; user interaction states, interaction habit patterns, time-frequency states, and domain-frequency states are all represented by standardized feature vectors. When different data sources use different units, the data type corresponding to each data point is first identified, and then unit conversion is performed according to the corresponding data unit mapping relationship. For example, time data is uniformly converted to seconds, minutes to sixty seconds, and milliseconds to 0.001 seconds; ratios are uniformly converted to decimals between zero and one; text features are uniformly converted to fixed-dimensional numerical feature vectors; and category labels are uniformly converted to continuous numbering or one-hot encoding. After unit unification, minimum-maximum standardization or mean-variance standardization methods are used to map all values ​​to the same numerical range, ensuring that data from different sources can participate in subsequent calculations and improving the consistency and stability of the calculation results.

[0091] In this embodiment, all preset models are determined by using historical data to establish initial models, trial runs to correct model parameters, and long-term adaptive updates. Historical interaction data generated during continuous operation is collected, with 70% used as model building data and the remaining 30% as model validation data. Based on historical interaction behavior, the initial establishment of the interaction behavior state model, interaction behavior clustering model, behavior evolution path model, and interaction reasoning model is completed. Trial runs are conducted in a real-world operating environment, and the parameters in the model are corrected by comparing the deviation between the model prediction results and the actual user interaction results. When the model prediction error remains stable over multiple consecutive operating cycles, the current model parameters are determined as the formal operating parameters. During long-term operation, the latest user interaction data is collected again according to the preset update cycle, and the model parameters are recalculated using the latest data, allowing the model to continuously adapt to changes in user interaction behavior.

[0092] The preset interaction round division rules in this embodiment can be determined based on the distribution of consecutive user query time intervals according to historical interaction data. The consecutive query time intervals of all historical users are statistically analyzed, the time interval distribution curve is calculated, and the end time of most users' consecutive queries is selected as the basis for dividing the interaction rounds. If the time interval between two consecutive queries exceeds this time, it is divided into a new interaction round; otherwise, it remains in the same interaction round. Subsequently, the time interval distribution is periodically re-statistically analyzed based on actual operation, and the division rules are dynamically revised.

[0093] For the preset number of nodes in the behavior analysis window, the distribution of historical behavior evolution path lengths can be statistically analyzed, the recognition accuracy of interaction states corresponding to different window lengths can be calculated, and the window length with the highest recognition accuracy and moderate computational complexity can be selected as the preset number of nodes. During long-term operation, the window length can be automatically adjusted according to changes in recognition accuracy, so that the window can adapt to different business scenarios.

[0094] For the preset matching thresholds involved in the matching process during the interaction evolution stage, offline testing can be conducted using historical interaction behavior data. First, the matching degree between a large number of historical behavior samples is calculated, and then the state recognition accuracy corresponding to different matching thresholds is tested respectively. The matching threshold with the highest overall accuracy and the lowest false recognition rate is selected as the initial value. Subsequently, it is continuously fine-tuned based on the actual recognition results to keep the matching threshold stable.

[0095] For the preset weights in the user feedback correlation analysis, the contribution analysis method can be used to determine them based on historical user feedback data. The contribution of status feedback, habitual feedback, time-frequency feedback, and domain-frequency feedback to the improvement of the digital human's response effect can be calculated separately, and then the corresponding weights can be allocated according to the contribution ratio. During the operation, the response satisfaction and interaction completion rate can be continuously counted, and the weights of each item can be adjusted appropriately according to the latest statistical results so that the weights always reflect the importance of each feedback feature.

[0096] The preset weights in the constraint association priority calculation can be determined based on the degree of influence of different response constraints on the final response quality; the impact of each response constraint on the user's follow-up question rate, interaction completion rate, and response satisfaction during historical interactions is statistically analyzed, and the more significant the impact, the higher the corresponding weight; fine-tuning is carried out based on actual operating results during the trial operation phase, and automatic updates are made based on the latest statistical data during long-term operation.

[0097] For the thresholds for degree of change, conflict determination, and priority determination parameters in the coordination rules of constraint conflict detection, the following methods can be used to determine them: initial values ​​based on historical statistics, corrections during on-site trial operation, and adaptive updates during long-term operation. First, the normal fluctuation range of various degrees of change is statistically analyzed based on historical interaction data, and the upper limit of normal fluctuation is used as the initial threshold. Then, the false positive rate and false negative rate are recorded during actual operation, and the threshold size is gradually adjusted according to the statistical results. After entering the stable operation stage, the latest operation data is re-statistically analyzed according to a fixed update cycle, new statistical characteristics are calculated using a sliding time window, and the corresponding thresholds are automatically updated to continuously adapt to changes in interaction behavior under different user groups and different business scenarios.

[0098] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Parts not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A digital human interaction method based on time-frequency-domain frequency analysis for model inference, characterized in that, Includes the following steps: Step S1: Real-time data acquisition; Real-time acquisition of user interaction data from online users; The user interaction data package contains the user interaction number, user query data, and user interaction timestamp; For online users: Obtain corresponding historical interaction data based on user interaction ID; Generate corresponding interaction behavior state sequences based on historical interaction data and user interaction data; Output the current user interaction state and interaction habit pattern based on the interaction behavior state sequence; Predict the target time-frequency state and target domain-frequency state of the online user in the next interaction round based on the current user interaction state; Combine the current user interaction state, interaction habit pattern, target time-frequency state, and target domain-frequency state to obtain user interaction feedback features; Step S2: Interaction constraint analysis; For all current online users, obtain the online user set; Extract user interaction feedback features from the online user set to obtain the user interaction feedback feature set; Constructing single-round joint response constraints for digital humans based on user interaction feedback feature sets; Step S3: Digital Human Interaction Optimization; Construct an interaction reasoning model based on single-round joint response constraints, and generate corresponding digital human interaction content based on the interaction reasoning model; Acquire new user interaction data generated by online users in response to digital human interaction content, provide feedback and correction to single-round joint response constraints and interaction reasoning models, and complete real-time updates of digital human interaction.

2. The digital human interaction method based on time-frequency-domain frequency analysis for model inference as described in claim 1, characterized in that, The specific steps for generating corresponding interaction behavior state sequences based on historical interaction data and user interaction data include: Extract user interaction timestamps from historical interaction data and user interaction data to sort all user query data; divide the sorted user query data according to the preset interaction round division rules to obtain several consecutive interaction rounds; For any consecutive interaction rounds, extract all user query data within the consecutive interaction rounds to form the corresponding target query set; for each target query set, calculate the query time interval and query time density based on the user interaction timestamp; perform feature extraction based on the query time interval and query time density to obtain the corresponding time-frequency state unit; identify the attribute domain of the target query set to obtain the corresponding product attribute domain; calculate the coverage ratio of the product attribute domain to the preset product attribute domain space to obtain the corresponding domain-frequency state unit; According to the time sequence of consecutive interaction rounds, the time-frequency state unit and the domain-frequency state unit are associated and combined to obtain the joint behavior state node; all joint behavior state nodes are combined to obtain the interaction behavior state sequence.

3. The digital human interaction method based on time-frequency-domain frequency analysis for model construction and reasoning as described in claim 2, characterized in that, The specific steps for outputting the current user interaction state and interaction habit pattern based on the interaction behavior state sequence include: Evolutionary features are extracted based on the sequence of interactive behavior states to obtain behavioral evolution features; feature clustering is performed on the behavioral evolution features to obtain different interactive evolution stages; According to the arrangement order of the joint behavior state nodes in the interaction behavior state sequence, a preset number of joint behavior state nodes are extracted as the current behavior analysis window; feature matching is performed on the interaction state relationship of the joint behavior state nodes in the current behavior analysis window to obtain the current interaction evolution stage; the current interaction evolution stage is determined as the current user interaction state. In continuous interaction rounds, user inquiry data after receiving a digital human response is set as user feedback data, and the corresponding digital human response data is extracted. The user feedback data and the corresponding digital human response data are then combined to obtain user feedback data pairs. Feature extraction is performed on the user feedback data pairs to obtain the interaction habit patterns.

4. The digital human interaction method based on time-frequency-domain frequency analysis for model construction and reasoning as described in claim 3, characterized in that, The specific steps for predicting the target time-frequency state and target domain-frequency state of online users in the next interaction round based on the current user interaction state include: Match the historical behavior evolution path corresponding to the current user interaction state in the interaction behavior state sequence to obtain the target behavior evolution path; combine the interaction habit pattern to make user-personalized corrections to the target behavior evolution path to obtain the corresponding predicted behavior evolution path. Extract the joint behavior state node corresponding to the next interaction round in the predicted behavior evolution path to obtain the predicted behavior state node; parse the predicted time-frequency state unit and the predicted domain-frequency state unit corresponding to the predicted behavior state node; determine the predicted time-frequency state unit as the target time-frequency state of the next interaction round, and determine the predicted domain-frequency state unit as the target domain-frequency state of the next interaction round.

5. The digital human interaction method based on time-frequency-domain frequency analysis for model inference as described in claim 4, characterized in that, The specific steps for constructing single-round joint response constraints for digital humans based on user interaction feedback feature sets include: Based on the user interaction feedback feature set, output the corresponding state feedback set, habit feedback set, time-frequency feedback set, and domain-frequency feedback set; perform correlation analysis on each state feedback set, habit feedback set, time-frequency feedback set, and domain-frequency feedback set, calculate the correlation influence relationship between different feedback sets, and generate joint feedback relationship nodes for each online user; construct a user feedback correlation network according to the joint feedback relationship nodes between each online user. Based on the user feedback association network, the current online user set is clustered to obtain several target user groups; the joint feedback features of the target user groups are output; and the corresponding response guidance targets are output based on the joint feedback features of the groups. Based on the response guidance objectives corresponding to each target user group, query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints are constructed. The constraint association priority is calculated based on the group joint feedback characteristics corresponding to each target user group, and a corresponding joint response constraint set is generated. Conflict detection is performed based on each joint response constraint set to obtain the single-round joint response constraints of the digital human.

6. The digital human interaction method based on time-frequency-domain frequency analysis for model construction and reasoning as described in claim 5, characterized in that, The specific steps for constructing an interactive reasoning model based on single-round joint response constraints and generating corresponding digital human interactive content based on the interactive reasoning model include: The single-round joint response constraints of the digital human are analyzed to obtain response constraint parameters. The response constraint parameters are organized according to query response constraints, time-frequency guidance constraints, domain-frequency guidance constraints, and interaction habit adaptation constraints. The correlation between the response constraint parameters is established to construct an interactive reasoning model. Input the user query data from all user interaction data in the current interaction round into the interaction inference model, perform inference analysis on the user query data according to the response constraint parameters, generate the corresponding digital human interaction content, and output the digital human interaction content to online users to complete the digital human response in the current interaction round.

7. The digital human interaction method based on time-frequency-domain frequency analysis for model inference as described in claim 6, characterized in that, The specific steps for conflict detection based on each joint response constraint set include: When there are constraint conflicts between different joint response constraint sets, constraint coordination rules are constructed based on the degree of change in user interaction state, the degree of change in target time-frequency state, and the degree of change in target domain-frequency state. Based on the constraint coordination rules, the joint response constraint sets with conflicts are coordinated and corrected to obtain the single-round joint response constraints of the digital human.

Images