Dynamic intention recognition method and system, medium and product

Abstract

The invention discloses a dynamic intention recognition method and system, a medium and a product, and relates to the technical field of data processing. The method comprises the following steps: firstly, obtaining a dialogue text and multi-dimensional behavior data of a user in a specified time window; the system identifies a plurality of candidate intentions based on the data, and calculates initial probability distribution of the candidate intentions by using intention distribution of similar user groups in community historical data as a prior probability. The hypothetical intent is screened by analyzing the probability diversity, and the digital human then generates tentative interaction content and collects user feedback. And the system updates the probability distribution based on the consistency evaluation value of the feedback data and the assumed intention, and determines the candidate intention with the highest probability as a final recognition result when the probability convergence degree exceeds a threshold value. By implementing the technical scheme provided by the invention, the accuracy of user intention recognition can be improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a dynamic intent recognition method, system, medium, and product. Background Technology

[0002] With the rapid development of artificial intelligence technology, digital humans, as a new type of human-computer interaction interface, are widely used in various online communities. Digital humans interact with users through natural language understanding technology, providing services such as information consultation and emotional support. In these application scenarios, accurately understanding the user's true intentions is key to providing precise services.

[0003] Currently, intent recognition primarily employs rule-based matching or machine learning methods. Rule-based matching predefines a series of intent templates and keywords, identifying intent by matching user input against these templates. Machine learning-based methods, on the other hand, establish a mapping between input text and intent categories through training on large amounts of labeled data. However, in practical applications, it has been found that user intent often changes dynamically throughout the interaction process and may be expressed in multiple different ways. Relying solely on the content of a single moment in the dialogue for intent recognition is prone to misunderstanding, leading to a mismatch between the digital human's response and the user's needs. This is especially true in community scenarios, where user intents are often interconnected, posing a significant challenge to accurate recognition.

[0004] Therefore, how to improve the accuracy of intent recognition, reduce comprehension bias, and enhance the quality of digital human services in a community environment is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This application provides a dynamic intent recognition method, system, medium, and product that can improve the accuracy of user intent recognition.

[0006] The first aspect of this application provides a dynamic intent recognition method, comprising: Acquire the user's dialogue text sequence and multi-dimensional behavioral data with the digital human within a preset time window in the project. The multi-dimensional behavioral data includes browsing trajectory data, content creation data, and social interaction data. Based on the dialogue text sequence and the multidimensional behavioral data, multiple candidate intentions are determined. The intention distribution of related user groups is extracted from the historical data of the project as a prior probability, and the probability distribution of the candidate intentions is calculated based on the prior probability. Calculate the probability difference between each candidate intent in the initial probability distribution, and filter the hypothetical intents among each candidate intent based on the probability difference. The digital human generates probing interactive content based on the hypothetical intention and sends it to the user, and collects feedback data generated by the user after receiving the probing interactive content within a preset observation time. Calculate the consistency evaluation value between the feedback data and the hypothetical intent, update the probability distribution based on the consistency evaluation value, and calculate the probability convergence of each candidate intent based on the updated probability distribution; When the probability convergence is greater than a preset threshold, the candidate intent with the highest current probability is output as the user's intent recognition result.

[0007] By employing the aforementioned technical solution, multiple candidate intentions are identified based on dialogue text sequences and multidimensional behavioral data. The intention distribution of related user groups is extracted from historical data as prior probabilities, enabling a more comprehensive capture of users' potential intention characteristics. Hypothetical intentions are filtered by calculating the probability differences between candidate intentions, and targeted trial interaction content is generated to effectively verify the user's true intentions. Furthermore, a consistency evaluation value is calculated based on user feedback data to the trial interaction content, and the probability distribution is dynamically updated until the probability convergence reaches a preset threshold, at which point the recognition result is output. This achieves continuous tracking and accurate recognition of user intentions, avoiding comprehension biases that may result from relying solely on dialogue content at a single moment. This solution, through a closed-loop mechanism of multidimensional data analysis, interaction verification, and dynamic probability updates, improves the accuracy of digital human intention recognition in a project environment.

[0008] Optionally, the dialogue text sequence is converted into a dialogue semantic vector based on a pre-trained semantic encoding model. The similarity between the dialogue semantic vector and a preset intent template vector is calculated, and intents with similarity exceeding a preset text threshold are selected as candidate dialogue intents. Behavioral feature sequences are extracted based on the multidimensional behavioral data, and the behavioral feature sequences are matched with the behavioral patterns corresponding to each intent in the project. Intents with matching scores exceeding a preset behavioral threshold are selected as candidate behavioral intents. The candidate dialogue intents and the candidate behavioral intents are merged and duplicates are removed to obtain multiple candidate intents.

[0009] Optionally, from the historical data of the project, a group of related users with the same interest tags and active periods as the user are selected. The trigger frequency of each intent of the related user group in the corresponding dialogue scenario and behavior pattern of the user is counted, and the trigger frequency is converted into the prior probability value of each candidate intent. For each candidate intent, historical successful interaction cases between the digital human and the user under the candidate intent are extracted from the project. The semantic distance between the current dialogue text sequence of the user and the dialogue text in the historical successful interaction cases, as well as the pattern similarity between the current multidimensional behavior data of the user and the behavior data in the historical successful interaction cases, are calculated. The semantic distance and the pattern similarity are weighted and calculated to obtain the observation confidence coefficient. The prior probability value of each candidate intent is multiplied by the corresponding observation confidence coefficient to obtain the adjusted probability value, and the adjusted probability value is normalized to obtain the probability distribution of the candidate intent.

[0010] Optionally, an intent verification decision tree is constructed for each of the hypothetical intents, where the root node of the intent verification decision tree is the corresponding hypothetical intent, and each branch node is a key feature dimension of the corresponding hypothetical intent. The key feature dimensions include the target object, time range, constraints, and expected result. The feature dimension with the largest information gain is selected from the intent verification decision tree as the current trial point. Comparative trial content is generated for the current trial point, and the comparative trial content includes differentiated options for different hypothetical intents. Each differentiated option embeds the feature value of the corresponding hypothetical intent at the current trial point. The comparative trial content is sent to the user through the digital human, and an option identifier is assigned to each differentiated option.

[0011] Optionally, for each key feature dimension in the intent verification decision tree, obtain the set of values ​​for each hypothetical intent on the key feature dimension, calculate the entropy value of the set of values ​​as the initial uncertainty of the corresponding key feature dimension; divide the set of values ​​into multiple subsets according to the hypothetical intent category, calculate the entropy value of each subset, and sum the entropy values ​​of each subset according to the proportion of hypothetical intents contained in the corresponding subset to obtain the expected entropy of the key feature dimension; calculate the difference between the initial uncertainty and the expected entropy as the information gain of the corresponding key feature dimension; select the key feature dimension with the largest information gain as the current trial point.

[0012] Optionally, the system collects the user's explicit selection operations on each differentiated option in the comparative trial content, and determines the feature value of the user's selection based on the option identifier corresponding to the explicit selection operation; simultaneously monitors the implicit feedback signal of the user when no explicit selection operation is performed, the implicit feedback signal including the visual dwell time on the differentiated option, mouse hover trajectory, and facial micro-movements; performs multimodal fusion on the implicit feedback signal to calculate the user's implicit preference for each differentiated option; obtains the user's behavioral sequence of bypassing the comparative trial content in the project within the preset observation time, matches the behavioral sequence with the behavioral patterns of each hypothetical intention to obtain a behavior matching vector; and combines the implicit preference and the behavior matching vector to form the feedback data.

[0013] Optionally, the entropy value of the updated probability distribution is calculated, the entropy value reflecting the concentration of the updated probability distribution; the entropy reduction between the entropy value of the probability distribution before the update and the entropy value of the updated probability distribution is calculated; and the ratio of the entropy reduction to the entropy value of the probability distribution before the update is used as the probability convergence.

[0014] In a second aspect, embodiments of this application provide a dynamic intent recognition system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, which includes computer instructions, and the one or more processors call the computer instructions to cause the dynamic intent recognition system to perform the method described in the first aspect and any possible implementation thereof.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a dynamic intent recognition system, cause the dynamic intent recognition system to perform the method described in the first aspect and any possible implementation thereof.

[0016] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a dynamic intent recognition system, cause the dynamic intent recognition system to perform the method described in the first aspect and any possible implementation thereof.

[0017] By employing the aforementioned technical solution, multiple candidate intentions are identified based on dialogue text sequences and multidimensional behavioral data. The intention distribution of related user groups is extracted from historical data as prior probabilities, enabling a more comprehensive capture of users' potential intention characteristics. Hypothetical intentions are filtered by calculating the probability differences between candidate intentions, and targeted trial interaction content is generated to effectively verify the user's true intentions. Furthermore, a consistency evaluation value is calculated based on user feedback data to the trial interaction content, and the probability distribution is dynamically updated until the probability convergence reaches a preset threshold, at which point the recognition result is output. This achieves continuous tracking and accurate recognition of user intentions, avoiding comprehension biases that may result from relying solely on dialogue content at a single moment. This solution, through a closed-loop mechanism of multidimensional data analysis, interaction verification, and dynamic probability updates, improves the accuracy of digital human intention recognition in a project environment. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a dynamic intent recognition method disclosed in an embodiment of this application; Figure 2 This is another flowchart illustrating a dynamic intent recognition method disclosed in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a system provided in an embodiment of this application.

[0019] Explanation of reference numerals in the attached drawings: 301, Central Processing Unit; 302, Read-Only Memory; 303, Random Access Memory; 304, Bus; 305, Input / Output Interface; 306, Input Section; 307, Output Section; 308, Storage Section; 309, Communication Section; 310, Driver; 311, Removable Media. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0021] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0022] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0023] This application provides a dynamic intent recognition method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating a dynamic intent recognition method provided in an embodiment of this application. The method is applied to a system, which refers to a hardware and software integrated platform capable of executing a dynamic intent recognition program. The system can execute a dynamic intent recognition program, and the method includes steps 101 to 106, as follows: Step 101: Obtain the text sequence of conversations between the user and the digital human within a preset time window in the project, as well as multi-dimensional behavioral data, including browsing trajectory data, content creation data, and social interaction data.

[0024] The project is an online community platform for user-digital human interaction, providing various interactive functions and services. The preset time window refers to a fixed observation period set by the system, such as the last 30 minutes or the last 24 hours. The dialogue text sequence is a collection of texts formed in chronological order of the back-and-forth conversations between the user and the digital human. Multidimensional behavioral data includes browsing trajectory data, recording the user's page access paths, dwell time, and click behavior within the community; content creation data, including original content such as articles, comments, and images posted by users and their attribute information; and social interaction data, recording social behaviors such as following, liking, and forwarding between users and other community members or digital humans.

[0025] Specifically, data acquisition is achieved through the community platform's data collection module. The system receives the user ID and a preset time window parameter, extracts all dialogue records between the specified user and the digital human within that time window from the community platform's dialogue log database, and constructs a dialogue text sequence sorted by timestamp. Simultaneously, the system extracts multi-dimensional behavioral data from the user behavior log database. The data collection module acquires browsing history through a page access logger, recording visited page URLs, entry times, exit times, page dwell times, and click coordinate sequences within the page; it extracts user-created content from the content management system, including attributes such as content type, topic, publication time, word count, and citation sources; and it extracts interactive behaviors from the social relationship graph, recording features such as the interaction object, interaction type, interaction time, and interaction intensity. After data collection is complete, the system converts all data into a unified structured format for storage. For example, when a user discusses machine learning algorithms with a digital human within a 30-minute time window, the system records the complete conversation, along with data on the user's behavior while viewing relevant technical documents (5 minutes of page viewing time, clicking on 3 articles consecutively), posting technical comments (200 words, citing 2 references), and interacting with 3 professional users in technical discussions (2 comments, 3 likes). This data collectively constitutes a complete record of the user's behavior within that time window, providing a data foundation for subsequent intent recognition.

[0026] Step 102: Determine multiple candidate intents based on dialogue text sequences and multidimensional behavioral data, extract the intent distribution of related user groups from the project's historical data as prior probabilities, and calculate the probability distribution of candidate intents based on the prior probabilities.

[0027] Candidate intents refer to the set of possible intents initially identified from user behavior. Associated user groups refer to the set of historical users with similar characteristics or behavioral patterns to the current user. Intent distribution represents the statistical frequency of different intents. Prior probability is an initial probability estimate based on historical data. Probability distribution describes the relative likelihood of each candidate intent.

[0028] Specifically, the candidate intent set is first determined through two parallel sub-processes. The first sub-process uses a pre-trained BERT semantic encoding model to convert the dialogue text sequence into a 768-dimensional semantic vector. It calculates the cosine similarity between this vector and each vector in a pre-defined intent template vector library (containing standard vector representations of 100 common intents), adding intents with a similarity greater than 0.8 to the candidate set. The second sub-process converts the user's multi-dimensional behavioral data into a feature sequence, including features such as page dwell time ratio, content creation frequency, and social interaction intensity. This sequence is matched against a predefined intent behavior pattern library, adding intents with a matching score greater than 0.7 to the candidate set. The results of the two sub-processes are merged and deduplicated to obtain the candidate intent set. The system then filters related user groups from historical community data. The filtering criteria include: having the same interest tags as the target user (e.g., "technical research," "life sharing"), an overlap of more than 60% in active time periods, and similar behavioral patterns within the past 30 days. The frequency of these related users triggering various intents in similar scenarios is statistically analyzed, and the prior probability of each intent is obtained by dividing the frequency by the total number of samples. The system extracts successful interaction cases (historical records of the digital human correctly understanding and meeting user needs) for each candidate intent from historical data. For the current user's dialogue text sequence, it calculates the semantic distance (using Euclidean distance of vectors encoded with BERT) between it and each successful case; for the current user's behavioral data, it calculates the behavioral pattern similarity with successful cases (using the DTW algorithm to calculate the matching degree of temporal features). The semantic distance and behavioral similarity are weighted at a 6:4 ratio to obtain the observed confidence coefficient. The final probability value of each candidate intent is equal to the product of its prior probability and the observed confidence coefficient, which is then normalized to obtain the probability distribution. For example, if the prior probability of a candidate intent is 0.3 and the observed confidence coefficient is 0.8, its adjusted probability is 0.24, and after normalization, it may obtain a probability value of 0.35. In this way, the system comprehensively utilizes text semantics, behavioral patterns, and historical data to calculate the probability distribution of each candidate intent, providing a foundation for subsequent intent recognition.

[0029] In one possible implementation, multiple candidate intents are determined based on the dialogue text sequence and multidimensional behavioral data, specifically including steps 1021-1023, as follows: Step 1021: Convert the dialogue text sequence into dialogue semantic vectors based on the pre-trained semantic encoding model, calculate the similarity between the dialogue semantic vectors and the preset intent template vectors, and take the intents with similarity exceeding the preset text threshold as dialogue candidate intents.

[0030] Semantic encoding models are deep learning models that convert natural language text into numerical vectors, such as BERT and GPT. Dialogue semantic vectors are high-dimensional numerical vector representations of text after semantic encoding. Intent template vectors are vector representations of various predefined standard intents. Similarity is a numerical metric that measures the degree of similarity between two vectors. A preset text threshold is a standard value used to determine whether the similarity is high enough. Dialogue candidate intents are a set of possible intents initially identified through text analysis.

[0031] Specifically, firstly, the dialogue text sequence is concatenated chronologically, using [SEP] tags to separate different rounds of dialogue, forming a complete text input. The input text undergoes word segmentation and encoding preprocessing, including adding [CLS] and [SEP] tags, converting words to vocabulary IDs, generating positional and segmented codes. The preprocessed codes are input into the BERT model, extracting the hidden state at the last layer's [CLS] position as a 768-dimensional dialogue semantic vector. The system maintains an intent template vector library containing 100 predefined standard vector representations of intents, obtained through training on a large amount of labeled data. For the current dialogue semantic vector, its cosine similarity is calculated with each vector in the intent template vector library. The cosine similarity formula is the dot product of the two vectors divided by their respective magnitudes, with values ​​ranging from -1 to 1. A preset text threshold of 0.8 is set, and intents with similarity exceeding this threshold are considered candidate dialogue intents. For example, a user's dialogue sequence with a digital human is "I want to learn machine learning, any suggestions? [SEP] I want to learn systematically from the basics [SEP] I'm mainly interested in deep learning." After semantic encoding to obtain a vector, the similarity with the "learning consultation" intent template is 0.85, and the similarity with the "technical discussion" intent template is 0.82. Both intents exceed the threshold, so they are included in the dialogue candidate intent set. In this way, the system achieves an initial mapping from text to intent.

[0032] Step 1022: Extract behavioral feature sequences from multidimensional behavioral data, match the behavioral feature sequences with the behavioral patterns corresponding to each intent in the project, and select intents with a matching degree exceeding a preset behavioral threshold as candidate behavioral intents.

[0033] A behavioral feature sequence is a set of time-ordered feature values ​​extracted from multidimensional behavioral data, encompassing user browsing, creation, and interaction behaviors. A behavioral pattern refers to the typical combination of behavioral features exhibited by a user under a specific intent and its temporal variation. Matching degree indicates the degree of fit between the behavioral feature sequence and a predefined behavioral pattern. A preset behavioral threshold is the standard value for determining whether the matching degree meets the requirements. Candidate behavioral intents are the set of possible intents identified through behavioral analysis.

[0034] Specifically, the system first constructs a standardized sequence of behavioral features. For browsing trajectory data, it extracts the page type (converted to one-hot encoding), dwell time (statistically counted and normalized by minutes), and click frequency (number of clicks per minute) for each time point. For content creation data, it extracts the creation type (converted to category encoding), content length (normalized word count), and creation frequency (number of creations per hour). For social interaction data, it extracts the number of interacting objects, the composition of interaction types (converted to proportional distribution), and interaction intensity (weighted according to preset weights). These features are organized into a feature sequence in chronological order, and a sliding time window (10 minutes window size, 1 minute step) is used for feature aggregation to obtain a unified-dimensional sequence of behavioral features. The system pre-establishes an intent-behavior pattern library in the project, with each intent corresponding to a set of standard behavioral patterns. The Dynamic Time Warping (DTW) algorithm is used to calculate the matching degree between the current behavioral feature sequence and each standard behavioral pattern. The DTW algorithm uses dynamic programming to find the optimal alignment between two time sequences and calculates the aligned distance as the matching degree index. A preset behavioral threshold of 0.7 is set, and intents with a matching degree exceeding this threshold are included in the candidate intent set. For example, when a user views technical documents, they exhibit a behavioral sequence of "continuously browsing multiple related articles - taking notes - posting comments - discussing with others." This sequence has a match score of 0.85 with the standard behavioral pattern of "deep learning" intent and a match score of 0.78 with "technical research" intent. Both of these intents are identified as candidate behavioral intents. Through the extraction and matching of behavioral features, the system captures the intent information contained in the user's behavior.

[0035] Step 1023: Merge dialogue candidate intents and behavior candidate intents and remove duplicates to obtain multiple candidate intents.

[0036] Dialogue candidate intents are a preliminary set of intents obtained through dialogue text analysis. Each intent includes attributes such as intent ID, intent type, and confidence level. Behavioral candidate intents are a preliminary set of intents obtained through behavioral analysis, possessing the same attribute structure. Duplicates refer to intents with the same intent ID in both sets. Candidate intents are the final set of intents after merging and deduplication.

[0037] Specifically, an empty target set is first constructed to store the final candidate intents. For each intent in the dialogue candidate intent set, the system checks whether its intent ID already exists in the target set. If it does not exist, the intent is directly added to the target set; if it already exists, the intent with the higher confidence score is retained. In practice, the system stores the two candidate intent sets in ordered arrays, with each array element containing fields such as intent ID, intent type, and confidence score. Two pointers are used to traverse the two arrays, comparing intent IDs for merging. When encountering the same intent ID, the confidence scores of the two intents are compared, and the intent with the higher confidence score is retained in the target set. For example, dialogue analysis yields candidate intents {"Learning Consultation": 0.85, "Technical Discussion": 0.82}, and behavior analysis yields candidate intents {"Technical Discussion": 0.78, "Resource Search": 0.75}. During merging, it is found that "Technical Discussion" appears in both sets, and the confidence score obtained from dialogue analysis (0.82) is higher than that from behavior analysis (0.78), therefore, the intent with a confidence score of 0.82, "Technical Discussion", is retained. The final candidate intent set is {"Learning Consultation": 0.85, "Technical Discussion": 0.82, "Resource Search": 0.75}. Through this merging and deduplication operation, the system obtains a complete candidate intent set that integrates the results of dialogue and behavioral analysis.

[0038] In one possible implementation, the intent distribution of associated user groups is extracted from the project's historical data as a prior probability, and the probability distribution of candidate intents is calculated based on the prior probability. Specifically, this includes steps 1024-1026, as follows: Step 1024: Filter the historical data of the project to find related user groups that have the same interest tags and active time periods as the user. Calculate the trigger frequency of each intent of the related user groups in the corresponding dialogue scenarios and behavior patterns of the user, and convert the trigger frequency into the prior probability value of each candidate intent.

[0039] Interest tags identify a user's primary interests within the community, such as "technical research" or "lifestyle sharing." Active time periods refer to the distribution of time periods during which users are most active in the community. Associated user groups are historical user sets with similar characteristics to the target user. Dialogue scenarios are the specific contexts in which a user interacts with the digital avatar. Behavioral patterns are combinations of behavioral characteristics exhibited by users in the community. Trigger frequency is the number of times various intentions occur in historical data. Prior probability values ​​are initial probability estimates based on historical statistics.

[0040] Specifically, the system extracts the target user's interest tag set and active time period distribution from user profile data. Interest tags are multi-tag, with each user having multiple tags; active time periods are divided into 48 half-hour periods within a 24-hour period, and the activity level for each period is recorded. An interest tag overlap rate threshold of 0.6 and an active time period similarity threshold of 0.7 are set. The system calculates the interest tag overlap rate (number of common tags / total number of tags) and the cosine similarity of the active time period distribution between all historical users and the target user. Users who simultaneously meet both thresholds are selected as the associated user group. Next, the system extracts the target user's current dialogue scenario features (such as dialogue topic, sentiment tendency, and interaction frequency) and behavioral pattern features (such as page visit sequence, content creation type, and social interaction method). In the historical data of the associated user group, interaction records with similar scenario features and behavioral patterns are retrieved, and the trigger frequency of each intent is counted. For example, in the "technology learning" scenario, the intent "learning consultation" appears 800 times, the intent "technology discussion" appears 600 times, and the intent "resource search" appears 400 times, totaling 1800 intent triggers in the associated user group. Dividing the trigger frequency by the total number of triggers yields the prior probability values ​​for each intent: {"Learning Consultation": 0.44, "Technical Discussion": 0.33, "Resource Search": 0.23}. Through this statistical analysis based on associated user groups, the system obtains the prior probability distribution of each candidate intent.

[0041] Step 1025: For each candidate intent, extract historical successful interaction cases between the digital human and the user under the candidate intent from the project, calculate the semantic distance between the user's current dialogue text sequence and the dialogue text in the historical successful interaction cases, and the pattern similarity between the user's current multidimensional behavioral data and the behavioral data in the historical successful interaction cases.

[0042] Historical successful interaction cases refer to the historical record of the digital human accurately understanding and meeting user needs. Semantic distance represents the distance between two pieces of text in the semantic space; the smaller the value, the more semantically similar they are. Pattern similarity measures the degree of matching between two sets of behavioral data; the larger the value, the more similar the behavioral patterns. Candidate intent is the set of possible intents identified through preceding steps.

[0043] Specifically, the system first extracts historical successful interaction cases from the project's interaction log database. For each candidate intent, it queries historical records marked "successful interaction" to extract interaction cases containing complete dialogue text and behavioral data. The criteria for successful interaction include: user confirmation feedback of "satisfaction," interaction goal completion rate exceeding 90%, and subsequent user behavior conforming to intent expectations. For the current user's dialogue text sequence and the dialogue text in historical cases, the BERT model is used for semantic encoding to obtain corresponding semantic vectors. The Euclidean distance between the two semantic vectors is calculated as the semantic distance metric. The specific calculation formula is: d=sqrt(sum((v1-v2)^2)), where v1 and v2 are the semantic vectors of the current dialogue and historical cases, respectively. For calculating the pattern similarity of behavioral data, the system first standardizes the multi-dimensional behavioral data of the current user and historical cases, converting them into feature vectors of the same dimension. Behavioral features include: browsing behavior features (page access sequence, dwell time distribution, click patterns), creation behavior features (content type distribution, creation frequency, citation patterns), and social behavior features (distribution of interaction objects, proportion of interaction types, interaction intensity). The Dynamic Time Warping (DTW) algorithm is used to calculate the matching degree between two sets of behavioral sequences, thus obtaining behavioral pattern similarity. For example, the semantic distance between the user's current dialogue text "I want to learn machine learning algorithms in depth, please recommend a learning path" and a historical case is 0.15, indicating that the semantics are very close; the current user's behavioral sequence of "continuously browsing tutorials - taking notes - asking questions and discussing" has a behavioral pattern similarity of 0.82 with the historical case, indicating that the behavioral pattern is highly matched. By calculating semantic distance and pattern similarity, the system quantifies the degree of matching between the current interaction and historical successful cases.

[0044] Step 1026: Calculate the weighted semantic distance and pattern similarity to obtain the observation confidence coefficient; multiply the prior probability value of each candidate intent by the corresponding observation confidence coefficient to obtain the adjusted probability value, and normalize each adjusted probability value to obtain the probability distribution of the candidate intent.

[0045] The observation confidence coefficient is a comprehensive index calculated by weighting semantic distance and pattern similarity, reflecting the overall matching degree between the current observation and historical cases. The adjusted probability value is the probability value after adjusting the prior probability using the observation confidence coefficient. Normalization ensures that the sum of all probability values ​​is 1, resulting in a standard probability distribution. The candidate intent is the set of possible intents identified in the preceding steps.

[0046] Specifically, the semantic distance is first converted into a semantic similarity score. Since a smaller semantic distance indicates greater similarity, a conversion is needed: semantic similarity = 1 / (1 + semantic distance), mapping the result to the [0, 1] interval. Pattern similarity is already within the [0, 1] interval and does not require conversion. The semantic similarity weight is set to 0.6, and the pattern similarity weight to 0.4. The observation confidence coefficient is calculated by weighted summation: confidence coefficient = 0.6 × semantic similarity + 0.4 × pattern similarity. For each candidate intent, its prior probability value is multiplied by the corresponding observation confidence coefficient to obtain the adjusted probability value. Finally, all adjusted probability values ​​are normalized: each adjusted probability value is divided by the sum of all adjusted probability values. For example, for three intents in the candidate intent set, the calculation process is as follows: Intent A has a semantic distance of 0.15, a converted semantic similarity of 0.87, a pattern similarity of 0.82, and an observation confidence coefficient of 0.85; Intent B has a semantic similarity of 0.75, a pattern similarity of 0.78, and an observation confidence coefficient of 0.76; Intent C has a semantic similarity of 0.65, a pattern similarity of 0.70, and an observation confidence coefficient of 0.67. The prior probabilities of these intents are 0.44, 0.33, and 0.23, respectively. Multiplying these probabilities by the observation confidence coefficients yields adjusted probability values ​​of 0.374, 0.251, and 0.154. After normalization, the final probability distributions are obtained: 0.48 for Intent A, 0.32 for Intent B, and 0.20 for Intent C. Through this calculation process, the system obtains a standardized probability distribution that comprehensively considers prior probabilities and observed data.

[0047] Step 103: Calculate the probability difference between each candidate intent in the initial probability distribution, and filter the hypothetical intents among each candidate intents based on the probability difference.

[0048] The initial probability distribution is a set of probability values ​​for each candidate intent. The probability dissimilarity is a numerical metric that measures the difference in probability values ​​between different candidate intents. Assume that intents are a subset of candidate intents that require further verification. The probability value reflects the likelihood that each intent will be judged as the user's true intent.

[0049] Specifically, the probability difference matrix between each candidate intent is first calculated. For any two candidate intents i and j, their probability difference D(i,j) is calculated using the following formula: D(i,j) = |P(i) - P(j)| / max(P(i), P(j)), where P(i) and P(j) represent the probability values ​​of intents i and j, respectively. This calculation method standardizes the probability difference to the interval [0,1]. For n candidate intents, an n×n symmetric matrix is ​​generated, where each element represents the probability difference between the corresponding two intents. Next, the system sets a probability difference threshold (e.g., 0.3) and identifies intent pairs with a difference less than this threshold. The intents in these pairs are considered as competing hypothetical intents. For example, a candidate intent set obtained in a certain identification might be: {"Learning Consultation": 0.35, "Technical Discussion": 0.32, "Resource Search": 0.20, "Question Answering": 0.13}. The calculated probability difference between "learning consultation" and "technical discussion" is 0.09, which is less than the threshold of 0.3. Therefore, both of these intentions are included in the hypothetical intention set. "Resource search" and "question answering" have probability differences greater than 0.3 compared to other intentions and are not included in the hypothetical intention set. The final hypothetical intention set is {"learning consultation", "technical discussion"}. Through this probability difference-based screening mechanism, the system identifies a set of intentions with similar probabilities that require further verification.

[0050] Step 104: Generate probing interactive content based on the assumed intent using a digital human and send it to the user. Collect feedback data generated by the user after receiving the probing interactive content within a preset observation period.

[0051] Assume that the intent is a set of intents selected from candidate intents that require further verification. The probing interaction content is dialogue content generated by a digital human to verify a specific intent. The preset observation time is a fixed period during which the system collects user feedback data. Feedback data includes both direct user responses to the probing content and indirect behavioral expressions.

[0052] Specifically, firstly, an intent validation decision tree is constructed for each hypothetical intent. The root node of the decision tree is the hypothetical intent, and the branch nodes are the key feature dimensions of that intent (such as target object, time range, constraints, expected results, etc.). Based on information gain calculation, the most discriminative feature dimension is selected as the current trial point. For the selected trial point, comparative trial content containing differentiated options is generated. For example, for the two hypothetical intents "learning consultation" and "technical discussion," the system selects "learning goal" as the trial point and generates the trial content: "Do you want to systematically learn the knowledge in this field, or do you want to solve a specific technical problem?", where the first option corresponds to the "learning consultation" intent, and the second option corresponds to the "technical discussion" intent. The digital human sends the trial content to the user and simultaneously activates the feedback data collection module. Within a preset 30-second observation period, the system collects multi-dimensional feedback data: explicit selection operations: recording the specific options selected by the user; implicit feedback signals: including the user's visual dwell time on each option, mouse hover trajectory, facial micro-movements, etc.; interactive behaviors: page jumps, content viewing, social interactions, etc., performed by the user during the observation period. In the example above, the system might collect the following feedback data: the user selected the "System Learning" option, spent 80% of their time on this option, and subsequently visited and bookmarked the "Machine Learning Introduction Tutorial" page. This feedback data will be used for subsequent intent verification analysis. Through this proactive probing and multi-dimensional feedback collection method, the system obtains the key data needed to verify the intent of the hypothesis.

[0053] In one possible implementation, feedback data generated by the user within a preset observation period after receiving the exploratory interaction content is collected, specifically including steps 1041-1044, as follows: Step 1041: Collect the user's explicit selection operations on each differentiated option in the comparative trial content, and determine the feature value selected by the user based on the option identifier corresponding to the explicit selection operation.

[0054] Comparative heuristics consist of structured questions with multiple differentiated options. Differentiated options are choices designed to address different hypothetical intentions. Explicit selection is the user's direct action of explicitly choosing an option. Option identifiers are unique identifiers assigned to each option by the system. Feature values ​​are the specific values ​​of the intention features corresponding to each option.

[0055] Specifically, firstly, each differentiated option in the trial content is assigned a unique option identifier, constructed using the format "intent ID_feature dimension_value". For example, for the trial question "Do you want to systematically learn knowledge in this field, or do you want to solve a specific technical problem?", the two options are assigned the identifiers "LEARNING_GOAL_SYSTEMATIC" and "DISCUSSION_GOAL_SPECIFIC" respectively. The system embeds the option identifiers in the user interface, but they are not visible to the user. When the user makes a selection, the system records detailed information about the selection event, including: event timestamp, option identifier, selection method (e.g., click, touch, voice), selection duration, coordinate trajectory during the selection process, etc. For multiple selections, the selection sequence number is also recorded. The system parses the identifier of the selected option and extracts its feature value information. For example, when the user selects the "systematically learn knowledge in this field" option, the system records the identifier "LEARNING_GOAL_SYSTEMATIC", parses it to obtain the feature dimension as "learning goal", and the feature value as "systematically learn". If a user repeatedly switches between multiple options before finally making a choice, the system records the complete sequence of selection changes, but extracts feature values ​​based on the final selected option. In this way, the system accurately records the user's explicit selection behavior and maps it to specific values ​​in the intent feature space. These feature values ​​serve as important evidence for intent verification.

[0056] Step 1042: Simultaneously monitor the implicit feedback signals of the user when no explicit selection is made. Implicit feedback signals include the duration of visual dwell on differentiated options, mouse hover trajectory, and facial micro-movements.

[0057] Implicit feedback signals are non-active behavioral characteristics that users naturally generate during interaction. Visual dwell time refers to the cumulative time a user gazes at each option. Mouse hover trajectory records the movement path and stopping position of the mouse pointer on the interface. Facial micro-movements are subtle changes in a user's facial expressions, including eyebrow movements, changes in the corners of the mouth, etc.

[0058] Specifically, for monitoring visual dwell time, the system uses a camera to capture the user's eye movement data at a sampling frequency of 60Hz. An eye-tracking algorithm calculates the coordinates of the user's gaze on the screen in real time to determine if the gaze falls within each option area. The system sets a rectangular bounding box for each option area; timing begins when the gaze falls within the bounding box and stops when it leaves, accumulating the visual dwell time for each option. For collecting mouse hover trajectories, the system records the mouse pointer's coordinates every 10 milliseconds, forming a coordinate point sequence. The coordinate sequence is segmented to extract features such as mouse movement speed, acceleration, and directional changes. When the mouse dwells in an option area for more than 100 milliseconds, it is recorded as a valid hover. Simultaneously, subtle jitter patterns during hovering are recorded to reflect the user's degree of hesitation. For recognizing facial micro-movements, the system uses a deep learning model to analyze the user's facial video stream, extracting the motion trajectories of 68 facial key points. Special attention is paid to changes in the eyebrow area (reflecting confusion or understanding), the eye area (reflecting focus), and the mouth area (reflecting emotional inclination). For example, the system detected that when a user was reading the "System Learning" option, their visual dwell time was 2.3 seconds, the mouse hovered in a circular motion over the area for 0.8 seconds, and their eyebrows slightly raised to indicate interest. When reading the "Solve Specific Problems" option, the visual dwell time was 1.1 seconds, the mouse moved quickly across the screen, and the facial expression remained neutral. Through this multi-dimensional collection of implicit feedback signals, the system obtained information about the user's potential preferences for each option.

[0059] Step 1043: Perform multimodal fusion on the implicit feedback signal and calculate the user's implicit preference for each differentiated option.

[0060] Implicit feedback signals include data such as visual dwell time, mouse movements, and facial micro-expressions. Implicit preference is an indicator of user inclination towards each option, obtained through multimodal signal analysis. Differentiated options are choices designed for different intentions within the trial content.

[0061] Specifically, the process begins with feature extraction and standardization of various feedback signals. For visual dwell data, the relative dwell time ratio (dwell time of a single option / total dwell time) and gaze frequency (gaze frequency of a single option / total gaze frequency) for each option are calculated. For mouse trajectory data, features such as hover duration ratio, trajectory smoothness (path length / straight-line distance), and hover frequency are extracted. For facial expression micro-movement data, the intensity of positive emotions (e.g., smiling, raised eyebrows) and the intensity of negative emotions (e.g., frowning, downturned corners of the mouth) are calculated. All feature values ​​are normalized to the [0, 1] interval. Then, a weighted fusion method is used to calculate the implicit preference, specifically: Preference = w1 × visual feature score + w2 × mouse feature score + w3 × facial expression feature score. The weights w1 = 0.4, w2 = 0.35, and w3 = 0.25, set based on feature reliability. Visual feature score = 0.6 × dwell time ratio + 0.4 × gaze frequency. Mouse feature score = 0.5 × hover ratio + 0.3 × trajectory smoothness + 0.2 × hover frequency. Facial expression feature score = 0.7 × positive intensity - 0.3 × negative intensity. For example, for the "system learning" option, users exhibit a high visual dwell ratio (0.85), smooth mouse hover (0.75), and positive facial expression response (0.80), resulting in an implicit preference score of 0.81; while for the "solving specific problems" option, the indicators are 0.35, 0.40, and 0.30 respectively, resulting in an implicit preference score of 0.35. Through this multimodal feature fusion calculation, the system obtains a quantitative implicit preference evaluation result.

[0062] Step 1044: Obtain the sequence of user behaviors performed in the project while bypassing the comparative probing content within a preset observation period, match the sequence of behaviors with the behavior patterns of each hypothetical intention to obtain a behavior matching vector; combine the implicit preference and the behavior matching vector to form feedback data.

[0063] A behavioral sequence is a time-ordered record of a user's actions within the community. A behavioral pattern is a combination of typical behavioral characteristics exhibited by users under various intentions. A behavior matching vector records the degree of matching between the behavioral sequence and the behavioral pattern for each intention. Implicit preference is a user tendency indicator obtained through multimodal signal analysis. Feedback data is a comprehensive evaluation data integrating implicit preferences and behavioral matching.

[0064] Specifically, the system first records all user behavior within a preset observation period (usually 2 minutes). It employs an event listening mechanism to capture every action a user performs on the community platform, including: page access events (URL, entry time, exit time), content operation events (browsing, saving, sharing), and social interaction events (commenting, liking, following). These events are then sorted by timestamp to form a behavior sequence. Next, the system performs behavior pattern matching, using a sequence matching algorithm to calculate the similarity between the behavior sequence and the standard behavior patterns of each hypothetical intent. For each behavior event, features such as type, object, and parameters are extracted and matched against standard features in the intent behavior pattern library. An edit distance algorithm is used to calculate sequence similarity, considering the influence of event order and time intervals. The matching scores of each intent are combined to form a behavior matching vector. For example, a user's behavior sequence after receiving probing content might be: visiting the "Machine Learning Tutorial" page (30 seconds) → saving the tutorial → visiting the "Learning Roadmap Planning" page (45 seconds) → following 3 AI experts. The system matches this sequence with the standard pattern of the "learning consultation" intent (viewing tutorials → saving learning materials → following domain experts), obtaining a match score of 0.85; and with the standard pattern of the "technical discussion" intent, the match score is 0.35. Finally, the implicit preference vector [0.81, 0.35] and the behavioral matching vector [0.85, 0.35] are combined into a feedback data matrix, with each row corresponding to a hypothetical intent, and the columns representing the implicit preference and behavioral matching scores, respectively. In this way, the system obtains feedback data that comprehensively reflects the user's intent tendencies.

[0065] Step 105: Calculate the consistency evaluation value between the feedback data and the hypothetical intent, update the probability distribution based on the consistency evaluation value, and calculate the probability convergence of each candidate intent based on the updated probability distribution.

[0066] Feedback data is an evaluation matrix containing implicit preference scores and behavioral matching vectors. The consistency evaluation value measures how well the feedback data supports the hypothesized intention. The probability distribution represents the relative likelihood of each candidate intention. The probability convergence reflects the stability of the probability distribution. The hypothesized intention is the set of potential intentions that need to be verified.

[0067] Specifically, the consistency evaluation value between the feedback data and the hypothetical intent is first calculated. For each hypothetical intent, the cosine similarity between its feedback data vector (containing two components: implicit preference and behavioral matching degree) and the preset ideal feedback pattern vector is calculated. The ideal feedback pattern vector is set to [1, 1], representing a perfect match. The formula for calculating the consistency evaluation value is: consistency = (p×w1 + m×w2) / sqrt((w1^2 + w2^2)×(p^2 + m^2)), where p is the implicit preference degree, m is the behavioral matching degree, and w1 and w2 are the weights of the two components (both set to 1). Then, the probability distribution is updated based on the consistency evaluation value. The Bayesian update formula is used: P(intent|feedback) = P(feedback|intent)×P(intent) / P(feedback). Where P(feedback|intent) is determined by the consistency evaluation value, P(intent) is the current probability distribution, and P(feedback) is the normalization factor. For example, for the intent "learning consultation," the feedback data is [0.81, 0.85], resulting in a consistency evaluation value of 0.83; the current probability is 0.48, and the probability increases to 0.65 after the update. For the intent "technical discussion," the feedback data is [0.35, 0.35], the consistency evaluation value is 0.35, and the probability decreases from 0.32 to 0.20. Finally, the probability convergence is calculated using the entropy reduction as the metric. The entropy value of the probability distribution before and after the update is calculated as: H = -∑P(i)×log(P(i)), where P(i) is the probability of each intent. The probability convergence is equal to the entropy reduction divided by the entropy value before the update. In this example, the entropy value before the update is 0.95, and the entropy value after the update is 0.65, resulting in a convergence of 0.32. In this way, the system achieves dynamic updating of the probability distribution and convergence evaluation based on feedback data.

[0068] In one possible implementation, the probability convergence of each candidate intent is calculated based on the updated probability distribution, specifically including steps 1051-1052, as follows: Step 1051: Calculate the entropy value of the updated probability distribution. The entropy value reflects the degree of concentration of the updated probability distribution.

[0069] A probability distribution is the set of numerical values ​​representing the likelihood that each candidate intent will be judged as the user's true intent; the sum of all probability values ​​is 1. Entropy is a mathematical indicator describing the uncertainty of a probability distribution; a smaller entropy value indicates a more concentrated distribution, while a larger entropy value indicates a more dispersed distribution. The degree of concentration reflects the dispersion of the probability distribution among the candidate intents.

[0070] Specifically, the updated probability distribution vector is first obtained, which contains the probability values ​​of all candidate intentions. The entropy value is calculated using the information entropy formula: H = -∑(P(i) × log2(P(i))), where P(i) represents the probability value of the i-th candidate intention, and log2 represents the logarithm to the base 2. During the calculation, when a probability value is 0, the corresponding term is defined as 0 to avoid invalid values ​​in the logarithmic operation. Vectorized operations are used in the system implementation to improve computational efficiency. For example, for a probability distribution containing three candidate intentions [0.65, 0.20, 0.15], the calculation process is: H = -(0.65 × log2(0.65) + 0.20 × log2(0.20) + 0.15 × log2(0.15)) = -(0.65 × (-0.62) + 0.20 × (-2.32) + 0.15 × (-2.74)) = 1.23. In contrast, if the probability distribution is more concentrated, such as [0.90, 0.05, 0.05], the calculated entropy value is 0.47, indicating a more deterministic distribution; if the probability distribution is more uniform, such as [0.34, 0.33, 0.33], the calculated entropy value is 1.58, indicating a more uncertain distribution. By calculating the entropy value, the system obtains a quantified index of the concentration of the probability distribution. This index will be used in subsequent convergence calculations to help determine the credibility of the intent recognition results.

[0071] Step 1052: Calculate the entropy reduction between the entropy value of the probability distribution before the update and the entropy value of the probability distribution after the update; use the ratio of the entropy reduction to the entropy value of the probability distribution before the update as the probability convergence.

[0072] The probability distribution before the update is a set of probability values ​​based on historical data and initial analysis. The probability distribution after the update is a set of probability values ​​adjusted with feedback data. The entropy reduction is the difference between the entropy values ​​before and after the update, reflecting the degree of improvement in the determinism of the distribution. The probability convergence is the ratio of the entropy reduction to the initial entropy value, indicating the convergence of the intent recognition process.

[0073] Specifically, first, obtain the probability distribution vectors before and after the update. Let the probability distribution before the update be P1, and the probability distribution after the update be P2. Calculate the entropy values ​​H1 and H2 of the two distributions respectively, using the formula: H = -∑(P(i) × log2(P(i))). For example, for the probability distribution before the update, P1 = [0.48, 0.32, 0.20], we calculate H1 = -(0.48 × log2(0.48) + 0.32 × log2(0.32) + 0.20 × log2(0.20)) = 1.48; for the probability distribution after the update, P2 = [0.65, 0.20, 0.15], we calculate H2 = -(0.65 × log2(0.65) + 0.20 × log2(0.20) + 0.15 × log2(0.15)) = 1.23. The entropy reduction ΔH is calculated as ΔH = H1 - H2 = 1.48 - 1.23 = 0.25. Finally, the probability convergence C is calculated as C = ΔH / H1 = 0.25 / 1.48 = 0.169. The convergence value ranges from [0, 1], with a larger value indicating a more concentrated distribution. The system sets a convergence threshold of 0.15; when the convergence exceeds this threshold, the intent recognition result is considered sufficiently reliable. In this example, the convergence of 0.169 exceeds the threshold, indicating that the probability distribution has reached a relatively stable state, and the recognition result can be output. Through this convergence calculation based on entropy changes, the system establishes a quantitative standard for judging whether the intent recognition process is complete.

[0074] Step 106: When the probability convergence is greater than the preset threshold, output the candidate intent with the highest probability as the user's intent recognition result.

[0075] Probabilistic convergence is the ratio of entropy reduction to the initial entropy value, reflecting the stability of the probability distribution. The preset threshold is the standard value for judging whether the convergence meets the requirements. Candidate intents are the set of possible intents identified by the system. The intent recognition result is the user's true intent finally determined by the system. The candidate intent with the highest current probability is the intent with the highest probability value in the probability distribution.

[0076] Specifically, the system first checks whether the current probability convergence meets the output conditions. The system reads a preset convergence threshold (set to 0.15) and compares the currently calculated probability convergence with this threshold. When the convergence is greater than the threshold, the system enters the result output stage. It searches for the maximum probability value and its corresponding intent in the probability distribution vector. A vector maximum value search algorithm is used, and the index position of the maximum value is recorded. The intent identifier corresponding to this index is converted into complete intent description information, including intent type, intent parameters, and confidence level. For example, if the current probability distribution is [0.65, 0.20, 0.15], and the convergence is 0.169, which is greater than the threshold of 0.15, the system finds that the maximum probability of 0.65 corresponds to the intent "learning consultation," whose complete description is: {intent type: "learning consultation", learning goal: "systematic learning", learning domain: "machine learning", confidence level: 0.65}. The system outputs this intent description as the final recognition result to guide the digital human's subsequent responses. Through this judgment mechanism based on convergence and maximum probability, the system ensures that the output intent recognition results have sufficient reliability.

[0077] In the above embodiments, a basic intent recognition framework was implemented through semantic encoding and intent matching. To further improve the accuracy of intent recognition and reduce the impact of user expression on the recognition results, this application also provides a dynamic probing method based on decision trees. This method constructs an intent verification decision tree, analyzes the information gain of feature dimensions, and performs comparative probing interactions, enabling the system to more accurately handle intent recognition needs in complex scenarios. The following section combines... Figure 2 Another dynamic intent recognition method in the embodiments of this application is described below: Please see Figure 2 This is a flowchart illustrating a dynamic intent recognition method in an embodiment of this application.

[0078] Step 201: Construct an intent verification decision tree for each hypothetical intent. The root node of the intent verification decision tree is the corresponding hypothetical intent, and each branch node is the key feature dimension of the corresponding hypothetical intent. The key feature dimensions include the target object, time range, constraints, and expected results.

[0079] An intent validation decision tree is a tree-like logical structure used to validate a specific intent. The root node is the hypothetical intent to be validated. Branch nodes are the constituent elements of the intent, representing different feature dimensions. Key feature dimensions include: target object (the entity on which the intent acts), time range (the time span of intent execution), constraints (the limiting conditions for intent realization), and expected result (the expected state of intent completion).

[0080] Specifically, firstly, the system creates a decision tree data structure for each hypothetical intent. The node definition uses a tree structure: each node includes attributes such as node type (root node / branch node / leaf node), feature identifier, value range, and a list of child nodes. Taking the "learning consultation" intent as an example, the construction process is as follows: a root node is created, identified as "LEARNING_CONSULTATION". Four first-level branch nodes are created under the root node, corresponding to four key feature dimensions. The target object branch includes two sub-features: learning topic (e.g., programming language, algorithm theory, framework tools) and difficulty level (beginner, intermediate, expert). The time range branch includes two sub-features: overall cycle (short-term, medium-term, long-term) and learning pace (centralized, decentralized). The constraint branch includes two sub-features: time investment (weekly class hours) and prerequisite knowledge (necessary foundation, relevant experience). The expected result branch includes two sub-features: mastery level (understanding, application, proficiency) and practical goal (project development, research innovation). For each sub-feature, the system presets its possible value set as the basis for generating subsequent trial content. Another example is the decision tree for the "Technical Discussion" intent: the root node is identified as "TECHNICAL_DISCUSSION," and four branch nodes are created. The target object branch focuses on specific technical issues and application scenarios; the time frame branch considers the urgency of the issue and the depth of discussion; the constraints branch includes technology stack limitations and resource constraints; and the expected outcome branch includes the solution type and verification criteria. Through this hierarchical feature decomposition, the system establishes a complete intent verification logic framework.

[0081] Step 202: Select the feature dimension with the largest information gain from the intent verification decision tree as the current test point; generate comparative test content for the current test point. The comparative test content includes differentiated options for different hypothetical intents. Each differentiated option embeds the feature value of the corresponding hypothetical intent at the current test point.

[0082] Information gain is a metric that measures how much a feature contributes to distinguishing intent. The current heuristic is the feature dimension selected to generate the heuristic content. The contrastive heuristic content is a structured question with multiple options. Differentiated options are choices designed for different intent feature values. Feature values ​​are the specific values ​​that a feature dimension takes.

[0083] Specifically, the information gain of each feature dimension is calculated first. For each feature dimension in the intent verification decision tree, its information gain value is calculated. The calculation formula is: Gain(S, A) = Entropy(S) - ∑((|Sv| / |S|) × Entropy(Sv)), where S is the current intent set, A is the feature dimension, and Sv is the subset corresponding to the feature value v. For each feature dimension, the following calculations are performed: 1) Calculate the entropy value Entropy(S) of the current intent set; 2) Statistically analyze the distribution of intents under each value of the feature and calculate the conditional entropy; 3) Subtract the two to obtain the information gain. For example, for the intent set {"Learning Consultation": 0.48, "Technical Discussion": 0.32, "Resource Search": 0.20}, the information gain of the "Target Object" feature is calculated: the initial entropy value is 1.48, the feature values ​​include "Knowledge System", "Specific Problem" and "Resource Content", the intent distribution under each value is statistically analyzed, the conditional entropy is calculated to be 0.75, and the information gain is 0.73. Similarly, calculate the information gain of other feature dimensions and select the feature corresponding to the maximum value as the current trial point.

[0084] Then, comparative trial content is generated based on the selected trial point. Option templates are designed, embedding feature values ​​of different intentions into the templates to generate differentiated options. For example, when "target object" is selected as the trial point, the trial question is generated: "What type of help do you want?", containing three options: "The system understands the knowledge system of a certain field" (corresponding to the learning consultation intention); "Solve the specific technical problem currently encountered" (corresponding to the technical discussion intention); "Find learning resources on a specific topic" (corresponding to the resource search intention). Each option is assigned a unique identifier in the format "Intent ID_Feature Dimension_Feature Value", such as "LEARNING_OBJECT_KNOWLEDGE". These identifiers are used internally by the system to track user selections but are not visible to the user. In this way, the system generates trial content with clear differentiation.

[0085] In one possible implementation, the feature dimension with the largest information gain is selected from the intent verification decision tree as the current trial point, specifically including steps 2021-2023, as follows: Step 2021: For each key feature dimension in the intent verification decision tree, obtain the set of values ​​for each hypothetical intent on the key feature dimension, and calculate the entropy value of the set of values ​​as the initial uncertainty of the corresponding key feature dimension.

[0086] Key feature dimensions are the constituent elements of intent, including the target object, time frame, constraints, and expected outcome. The set of values ​​represents the specific possible combinations of values ​​for the feature dimension. Initial uncertainty is the degree of disorder in the distribution of feature values, expressed as entropy. Entropy is a mathematical metric for measuring information uncertainty.

[0087] Specifically, first, each key feature dimension in the intent verification decision tree is traversed, collecting the set of values ​​for each hypothetical intent in that dimension. For each value, its frequency of occurrence in all hypothetical intents is calculated to form a probability distribution. For example, for the "target object" dimension, the set of hypothetical intents includes three intents: {"learning consultation", "technical discussion", and "resource search"}. The target object value for the "learning consultation" intent is "knowledge system", the value for the "technical discussion" intent is "specific problem", and the value for the "resource search" intent is "resource content". The statistical probability distribution of the values ​​is: P(knowledge system) = 0.33, P(specific problem) = 0.33, P(resource content) = 0.33. The initial uncertainty of this dimension is calculated using the information entropy formula: H = -∑(P(i) × log2(P(i))), where P(i) is the probability of the i-th value. For the above probability distribution, the calculation process is: H = -(0.33×log2(0.33) + 0.33×log2(0.33) + 0.33×log2(0.33)) = 1.58. Similarly, the initial uncertainty of other feature dimensions is calculated. For the "time range" dimension, the value set is {"short-term", "medium-term", "long-term"}, assuming a probability distribution of [0.5, 0.3, 0.2], and the calculated entropy value is 1.49. For the "constraints" dimension, the value set is {"basic requirements", "resource limitations", "time investment"}, assuming a probability distribution of [0.4, 0.4, 0.2], and the calculated entropy value is 1.52. For the "expected results" dimension, the value set is {"knowledge acquisition", "problem solving", "resource acquisition"}, assuming a probability distribution of [0.35, 0.35, 0.3], and the calculated entropy value is 1.57. In this way, the system obtains the initial uncertainty quantification values ​​for each feature dimension, laying the foundation for subsequent information gain calculation.

[0088] Step 2022: Divide the set of values ​​into multiple subsets according to the hypothetical intent category, calculate the entropy value of each subset, and sum the entropy values ​​of each subset according to the proportion of hypothetical intents contained in the corresponding subset to obtain the expected entropy of the key feature dimension.

[0089] Assume that intent categories are the basic classifications of intents. Subsets are groups of values ​​grouped according to intent categories. Subset entropy is the uncertainty of the value distribution within a single group. Expected entropy is the weighted average of the entropy values ​​of all subsets, reflecting the uncertainty of the feature dimension given the intent category.

[0090] Specifically, the set of values ​​is grouped according to intent categories. A mapping table is created to record which intent category each value belongs to. For example, for the "target object" dimension, its complete set of values ​​is {"knowledge system", "specific problem", "resource content", "tool usage", "experience sharing", "dataset"}. This is divided into three subsets according to intent categories: learning consultation {"knowledge system", "experience sharing"}, technical discussion {"specific problem", "tool usage"}, and resource search {"resource content", "dataset"}. Then, the entropy value of each subset is calculated. For the learning consultation subset, the probability distribution of values ​​is [0.5, 0.5], and the entropy value H1 is calculated as -(0.5×log2(0.5)+0.5×log2(0.5))=1.0. For the technical discussion subset, the probability distribution of values ​​is [0.5, 0.5], and the entropy value H2=1.0. For the resource search subset, with a probability distribution of [0.5, 0.5], the entropy value H3 = 1.0 is calculated. The weights of each subset are calculated, representing the proportion of hypothetical intentions in the subset relative to the total. Assuming there are currently 6 hypothetical intentions: 2 for learning consultation, 2 for technical discussion, and 2 for resource search, the weights are w1 = 0.33, w2 = 0.33, and w3 = 0.33, respectively. Finally, the expected entropy is calculated: E = w1 × H1 + w2 × H2 + w3 × H3 = 0.33 × 1.0 + 0.33 × 1.0 + 0.33 × 1.0 = 1.0. By calculating the expected entropy, the system obtains the average uncertainty of the feature dimension under the known intention category. This value will be used to calculate the information gain later. The same calculation process is performed for other feature dimensions. For example, for the "time range" dimension, subsets are divided according to intent categories and the expected entropy is calculated; for the "constraints" dimension, the same subset division and expected entropy calculation are performed; for the "expected results" dimension, the same subset entropy calculation and weighted summation are performed.

[0091] Step 2023: Calculate the difference between the initial uncertainty and the expected entropy, and use it as the information gain of the corresponding key feature dimension; select the key feature dimension with the largest information gain as the current trial point.

[0092] Initial uncertainty is the overall entropy of the feature dimension's value distribution. Expected entropy is the conditional entropy of the feature dimension under a known intent category. Information gain is the difference between initial uncertainty and expected entropy, representing the feature dimension's contribution to intent differentiation. The current probe point is the feature dimension with the largest information gain, used to generate probe content.

[0093] Specifically, for each key feature dimension, the initial uncertainty and expected entropy are calculated, and information gain is calculated. Taking four feature dimensions as an example: the initial uncertainty of the "target object" dimension is H0=1.58, the expected entropy is E=1.0, and the information gain is G1=H0-E=0.58; the initial uncertainty of the "time range" dimension is H0=1.49, the expected entropy is E=1.1, and the information gain is G2=0.39; the initial uncertainty of the "constraints" dimension is H0=1.52, the expected entropy is E=1.2, and the information gain is G3=0.32; and the initial uncertainty of the "expected result" dimension is H0=1.57, the expected entropy is E=1.15, and the information gain is G4=0.42. An information gain ranking table is created, recording each dimension and its information gain value: [{Dimension: "Target Object", Gain: 0.58}, {Dimension: "Expected Result", Gain: 0.42}, {Dimension: "Time Range", Gain: 0.39}, {Dimension: "Constraints", Gain: 0.32}]. The dimension with the highest information gain is selected as the current probe point; in this example, it is the "Target Object" dimension (information gain 0.58). The system encapsulates the selected probe point information into structured data: {Probe Point Dimension: "Target Object", Information Gain: 0.58, Value Set: ["Knowledge System", "Specific Problem", "Resource Content", "Tool Usage", "Experience Sharing", "Dataset"], Dimension Description: "Core Object Type of User Needs"}. This probe point information will be used to generate comparative probe content subsequently. Through this information gain-based selection mechanism, the system finds the most distinctive feature dimension, improving the efficiency of the probe process.

[0094] Step 203: Send comparative trial content to the user via digital human and assign an option identifier to each differentiated option.

[0095] Comparative heuristics consist of structured questions with multiple options. Differentiated options are choices designed for different intentions. Option identifiers are unique codes assigned to each option by the system, used to track and analyze user choices.

[0096] Specifically, the probe content is first encapsulated in a structured manner. A message object is created, containing the following fields: message type (set as "probe_question"), message ID (a unique identifier generated using a timestamp and a random number), probe point information (feature dimension name and description), and an option list (each option includes display text and an identifier). An identifier is generated for each differentiated option, using a multi-segment encoding rule: the first segment is the intent type code (e.g., "LEARNING", "TECHNICAL", "RESOURCE"), the second segment is the feature dimension code (e.g., "OBJECT", "TIME", "CONDITION"), and the third segment is the feature value code (e.g., "KNOWLEDGE", "PROBLEM", "CONTENT"), connected by underscores. For example, for the probing question "What type of help do you want?", the system generates the following structure: {Message ID: "probe_20251203_143022_8374", Message Type: "probe_question", Probing Point: {Dimension: "Target Object", Description: "Core Object Type of User Needs"}, Option List: [{Text: "The system understands the knowledge system of a certain field", Identifier: "LEARNING_OBJECT_KNOWLEDGE"}, {Text: "Solve the specific technical problem encountered", Identifier: "TECHNICAL_OBJECT_PROBLEM"}, {Text: "Find learning resources on a specific topic", Identifier: "RESOURCE_OBJECT_CONTENT"}]}. Then, it calls the digital human's dialogue interface to send the message, while simultaneously setting up the interactive response mechanism for the options in the user interface, including click event listening and selection state management. When the user selects an option, the system records the identifier of that option for subsequent intent verification analysis. Through this structured message passing and identifier management mechanism, the system achieves precise delivery and selection tracking of trial content.

[0097] The following describes a dynamic intent recognition system according to an embodiment of the present invention from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 3 This is a schematic diagram of the structure of a dynamic intent recognition system in an embodiment of this application.

[0098] It should be noted that, Figure 3 The structure of the dynamic intent recognition system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0099] like Figure 3As shown, a dynamic intent recognition system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on a program stored in Read-Only Memory (ROM) 302 or a program loaded from storage portion 308 into Random Access Memory (RAM) 303, such as performing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0100] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0101] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0102] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0103] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0104] Specifically, a dynamic intent recognition system according to this embodiment includes a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements a dynamic intent recognition method provided in the above embodiment.

[0105] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in a dynamic intent recognition system described in the above embodiments; or it may exist independently and not assembled into the dynamic intent recognition system. The storage medium carries one or more computer programs, which, when executed by a processor of the dynamic intent recognition system, enable the dynamic intent recognition system to implement the dynamic intent recognition method based on encrypted data transmission of the Internet of Things provided in the above embodiments.

Claims

1. A dynamic intent recognition method, characterized in that, The method includes: Acquire the user's dialogue text sequence and multi-dimensional behavioral data with the digital human within a preset time window in the project. The multi-dimensional behavioral data includes browsing trajectory data, content creation data, and social interaction data. Based on the dialogue text sequence and the multidimensional behavioral data, multiple candidate intentions are determined. The intention distribution of related user groups is extracted from the historical data of the project as a prior probability, and the probability distribution of the candidate intentions is calculated based on the prior probability. Calculate the probability difference between each candidate intent in the probability distribution, and filter the hypothetical intents among each candidate intent based on the probability difference. The digital human generates probing interactive content based on the hypothetical intention and sends it to the user, and collects feedback data generated by the user after receiving the probing interactive content within a preset observation time. Calculate the consistency evaluation value between the feedback data and the hypothetical intent, update the probability distribution based on the consistency evaluation value, and calculate the probability convergence of each candidate intent based on the updated probability distribution; When the probability convergence is greater than a preset threshold, the candidate intent with the highest current probability is output as the user's intent recognition result.

2. The method according to claim 1, characterized in that, The determination of multiple candidate intents based on the dialogue text sequence and the multidimensional behavioral data includes: The dialogue text sequence is converted into a dialogue semantic vector based on a pre-trained semantic encoding model. The similarity between the dialogue semantic vector and the preset intent template vector is calculated. Intents with similarity exceeding a preset text threshold are taken as candidate dialogue intents. Based on the multidimensional behavioral data, behavioral feature sequences are extracted, and the behavioral feature sequences are matched with the behavioral patterns corresponding to each intent in the project. Intents with a matching degree exceeding a preset behavioral threshold are selected as candidate behavioral intents. The dialogue candidate intents and the behavior candidate intents are merged and duplicates are removed to obtain multiple candidate intents.

3. The method according to claim 1, characterized in that, The step of extracting the intent distribution of associated user groups from the historical data of the project as a prior probability, and calculating the probability distribution of the candidate intent based on the prior probability, includes: Filter the historical data of the project to identify related user groups that have the same interest tags and active periods as the user. Calculate the trigger frequency of each intention of the related user groups in the corresponding dialogue scenarios and behavioral patterns of the user, and convert the trigger frequency into the prior probability value of each candidate intention. For each candidate intent, extract historical successful interaction cases between the digital human and the user under the candidate intent from the project, calculate the semantic distance between the user's current dialogue text sequence and the dialogue text in the historical successful interaction cases, and the pattern similarity between the user's current multidimensional behavioral data and the behavioral data in the historical successful interaction cases. The observation confidence coefficient is obtained by weighting the semantic distance and the pattern similarity. The prior probability value of each candidate intention is multiplied by the corresponding observation confidence coefficient to obtain the adjusted probability value, and each adjusted probability value is normalized to obtain the probability distribution of the candidate intention.

4. The method according to claim 1, characterized in that, The step of generating probing interactive content based on the hypothetical intent using the digital human and sending it to the user includes: For each of the hypothetical intentions, an intention verification decision tree is constructed. The root node of the intention verification decision tree is the corresponding hypothetical intention, and each branch node is a key feature dimension of the corresponding hypothetical intention. The key feature dimension includes the target object, time range, constraints, and expected result. Select the feature dimension with the largest information gain from the intent verification decision tree as the current probe point; A comparative trial content is generated for the current trial point. The comparative trial content includes differentiated options for different hypothetical intentions. Each differentiated option embeds the feature value of the corresponding hypothetical intention at the current trial point. The comparative probing content is sent to the user through the digital human, and an option identifier is assigned to each of the differentiated options.

5. The method according to claim 4, characterized in that, The step of selecting the feature dimension with the largest information gain from the intent verification decision tree as the current probe point includes: For each key feature dimension in the intent verification decision tree, obtain the set of values ​​for each hypothetical intent on the key feature dimension, and calculate the entropy value of the set of values ​​as the initial uncertainty of the corresponding key feature dimension; The set of values ​​is divided into multiple subsets according to the hypothetical intent category. The entropy value of each subset is calculated. The entropy values ​​of each subset are weighted and summed according to the proportion of hypothetical intents contained in the corresponding subset to obtain the expected entropy of the key feature dimension. The difference between the initial uncertainty and the expected entropy is calculated as the information gain of the corresponding key feature dimension; The key feature dimension with the largest information gain is selected as the current probe point.

6. The method according to claim 4, characterized in that, The collection of feedback data generated by the user within a preset observation period after receiving the exploratory interaction content includes: Collect the user's explicit selection operations for each differentiated option in the comparative trial content, and determine the feature value selected by the user based on the option identifier corresponding to the explicit selection operation; Simultaneously monitor the implicit feedback signals of the user when no explicit selection operation is performed. The implicit feedback signals include the duration of visual dwell on the differentiated options, mouse hover trajectory, and facial micro-movements. Multimodal fusion is performed on the implicit feedback signals to calculate the user's implicit preference for each of the differentiated options; Obtain the sequence of behaviors performed by the user in the project within the preset observation time, bypassing the comparative probing content; match the sequence of behaviors with the behavioral patterns of each hypothetical intention to obtain a behavior matching vector. The implicit preference and the behavior matching vector are combined to form the feedback data.

7. The method according to claim 1, characterized in that, The step of calculating the probability convergence of each candidate intent based on the updated probability distribution includes: Calculate the entropy value of the updated probability distribution, where the entropy value reflects the degree of concentration of the updated probability distribution; Calculate the entropy reduction between the entropy value of the probability distribution before the update and the entropy value of the probability distribution after the update; The ratio of the entropy reduction to the entropy value of the probability distribution before the update is taken as the probability convergence.

8. A dynamic intent recognition system, characterized in that, The dynamic intent recognition system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the dynamic intent recognition system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the dynamic intent recognition system, it causes the dynamic intent recognition system to perform the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the dynamic intent recognition system, the dynamic intent recognition system performs the method as described in any one of claims 1-7.

Images