Human-computer interaction intention recognition method and system based on user portrait and storage medium

Abstract

The invention provides a human-computer interaction intention recognition method and system based on a user portrait and a storage medium, and relates to the technical field of human-computer interaction. The method comprises the steps that keystroke behavior characteristics of a user in the input process are collected, and basic parameter vectors containing semantic purity, backspace frequency and pause variance are extracted; calculating a semantic entropy increase trajectory index reflecting the character logic stability and dynamic pressure feedback stiffness reflecting the instantaneous emotion pressure of the user; in combination with the geometric distance between the current input vector and the historical preference centroid of the user portrait, reconstructing and generating collaborative intention confidence; and according to the confidence value, self-adaptive switching is carried out in three types of shunting paths including direct retrieval execution, semantic induced clarification and manual cooperation takeover. Through deep coupling of behaviors and semantics, the problem that traditional intention recognition is low in accuracy in a logic discrete scene is solved, and the method is suitable for improvement of an intelligent customer service system.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction technology, and in particular to a method, system, and storage medium for recognizing human-computer interaction intent based on user profiles. Background Technology

[0002] With the popularization of Natural Language Processing (NLP) technology, intelligent customer service and interactive dialogue systems have been widely used in e-commerce, finance, government affairs, and other fields. Traditional intent recognition solutions mainly rely on deep learning models to extract semantic features from user-input text sequences and map them into a pre-defined intent label space.

[0003] However, in real-world human-computer interaction scenarios, relying solely on text semantics for intent recognition has significant limitations: The disconnect between semantics and interactive behavior: Existing technologies often focus only on "what the user said," while ignoring the "interactive behavior characteristics" of the user during the input process (such as keystroke rhythm, backspace modification frequency, etc.). When users are emotionally anxious or their expression is logically confused, the semantics of the text can be deceptive, causing NLP engines to output high-confidence incorrect recognition results and triggering invalid interactions.

[0004] User profiles are underutilized and static: While some systems have introduced user profiles, most treat them as static labels, lacking a dynamic mechanism to deeply couple users' historical preferences with their current real-time interactive behaviors. This prevents the system from personalizing confidence levels based on different users' input habits (such as habitual pauses or frequent corrections).

[0005] The identification decision-making process lacks closed-loop feedback capability: Existing triage decisions (such as automatic response or referral to human intervention) are usually based on isolated probability threshold judgments, lacking the ability to spontaneously adjust the weights of the underlying identification algorithm based on the interaction results. When the system consistently misclassifies within a certain profile group, it cannot achieve parameter-level self-evolution and alignment.

[0006] Therefore, how to construct an intent recognition optimization scheme that can integrate text semantics, interactive behavior and user profiles, and has adaptive learning capabilities has become a key technical problem that urgently needs to be solved in the field of human-computer interaction. Summary of the Invention

[0007] The purpose of this invention is to provide a method, system, and storage medium for recognizing human-computer interaction intent based on user profiles, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention proposes a human-computer interaction intent recognition method based on user profiles, comprising the following steps: S1. Acquire the user's original text sequence in real time through the text interaction interface, and collect the user's keystroke behavior features during the input process within a predetermined input behavior collection window. Extract the basic parameter vector containing semantic purity, backspace frequency, and pause variance. ; S2, Based on the aforementioned basic parameter vector Parallel computation is performed on the semantic entropy increase trajectory index A, which reflects the stability of text logic, and the dynamic pressure feedback stiffness B, which reflects the instantaneous emotional pressure of the user. A state alignment algorithm is then used to map A and B to a unified interactive state sequence space. S3. Based on the semantic entropy increase trajectory index A and the dynamic pressure feedback stiffness B, and combined with the geometric distance between the current input vector and the centroid of the user profile's historical preferences, the collaborative intent confidence C is generated through nonlinear transformation. S4. Based on the value and trend of the confidence level C of collaborative intent, adaptively switch between the three diversion paths: direct retrieval execution, semantic induced clarification, and manual collaborative takeover. S5. Based on the intent recognition results output under different traffic paths, obtain user satisfaction and dynamically update the calculated weight coefficients of A and B according to user satisfaction.

[0009] As a preferred embodiment, the extraction includes a basic parameter vector comprising semantic purity, backspace frequency, and pause variance. Specifically, it includes: Identify and calculate the proportion of non-business-related modal particles, emoticons, and repetitive / redundant characters in the total number of characters in the current input text, and extract the semantic purity of the text. ; The average backspace frequency of users in a single round of dialogue interaction is collected through the front-end interaction interface. Variance of input pauses Collect thermal parameters of user keystroke behavior; Constructing the basic parameter vector .

[0010] As a preferred embodiment, the semantic entropy increase trajectory index A is used to quantify the degree of logical breakdown during the user's text organization process, and the calculation formula is as follows: In the formula, These are preset semantic weight coefficients; The formula for calculating the dynamic pressure feedback stiffness B is as follows: In the formula, Enter the total duration for the current round of text input; A correction constant to prevent overflow during division by zero; These are preset behavior weighting coefficients.

[0011] As a preferred embodiment, the process of mapping A and B to a unified interaction state sequence space using a state alignment algorithm specifically includes: A hidden Markov model state alignment algorithm is employed, using the sampling clock of the dynamic pressure feedback stiffness B as a reference, to interpolate and complete the semantic entropy increase trajectory exponent A, thereby constructing a state flow in the continuous time domain. .

[0012] As a preferred approach, step S3 involves reconstructing the collaborative intent confidence C through nonlinear transformation, specifically including: In a high-dimensional semantic vector space, calculate the Mahalanobis distance between the current input vector and the centroid of user profile preferences. ; The formula for calculating the confidence level C of collaborative intent is as follows: In the formula, is the original intent confidence score output by the NLP engine; k is the logical behavior mismatch adjustment coefficient. The probability distribution of the original intent is smoothed by using the confidence level C of the collaborative intent. When the product of A and B exceeds a preset threshold, the probability advantage of the Top-1 intent is forcibly reduced.

[0013] As a preferred option, in step S4, there are three traffic diversion paths: direct retrieval execution, semantic-guided clarification, and manual collaborative takeover. The adaptive switching logic is as follows: Direct retrieval execution, that is: when When the intent is clear, the business knowledge base is retrieved directly based on the recognition results, and a standard response is pushed out. Semantic induced clarification, that is: when When the intent is ambiguous, the conflict slot with the highest weight in the current text is extracted to generate a disambiguation prompt containing candidate options; Human-assisted takeover, that is: when If the instantaneous jump rate of B is detected to exceed the preset threshold, the interaction is determined to be in an abnormal state, the automatic recognition engine is automatically suspended and a manual collaboration channel is established. in, , This is the preset path switching threshold.

[0014] As a preferred option, in step S5, user satisfaction S is defined as a binary signal: The calculated weight coefficients of A and B are dynamically updated based on user satisfaction S, as follows: When S=1, then 1-S=0, and the weight remains unchanged; When S=0, then 1-S=1, and the weight increases by one step. .

[0015] As a preferred embodiment, the input behavior acquisition window is limited to a closed-loop time interval from when the user activates the input box control to when the command is triggered.

[0016] Furthermore, this invention also proposes a human-computer interaction intent recognition optimization system based on user profiles, the system comprising: The input behavior awareness unit is used to acquire the user's original text sequence in real time. Within a predetermined input behavior acquisition window, it collects the user's keystroke behavior features during the input process and extracts a basic parameter vector containing semantic purity, backspace frequency, and pause variance. ; Semantic geometry mapping unit, used in the basic parameter vector Based on this, we perform parallel computation of the semantic entropy increase trajectory index A, which reflects the stability of text logic, and the dynamic pressure feedback stiffness B, which reflects the instantaneous emotional pressure of users, and use a state alignment algorithm to map A and B to a unified interactive state sequence space. The feature parallel computing unit is used to reconstruct the collaborative intent confidence C through nonlinear transformation based on the output of the semantic geometric mapping unit, combined with the geometric distance between the current input vector and the centroid of the user profile's historical preferences. The intent disambiguation routing module is used to adaptively switch between three types of routing paths—direct retrieval execution, semantic inducement clarification, and manual collaborative takeover—based on the magnitude and trend of the collaborative intent confidence C. The feedback update unit is used to obtain user satisfaction and dynamically update the calculated weight coefficients of A and B based on user satisfaction.

[0017] Furthermore, the present invention also proposes a computer storage medium that stores one or more programs, which, when executed by an electronic device including multiple applications, cause the electronic device to perform the above-described human-computer interaction intent recognition method based on user profiles.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention not only analyzes the surface meaning of the text, but also collects behavioral details such as semantic purity, backspace frequency, and pause variance to explore the user's emotional state and logical stability during the input process, which significantly improves the accuracy of intent classification in complex environments.

[0019] (2) By introducing the spatial geometric distance between the current interaction features and the user profile’s historical preferences, the system achieves differentiated compatibility for users with different input habits. For different users’ hesitation, correction or fragmented expression, the system can perform personalized confidence compensation based on the profile background, avoiding the problem of false triggering or rejection caused by fixed thresholds in traditional solutions, and greatly improving the smoothness of the interaction process.

[0020] (3) This invention establishes a closed-loop online optimization of weights based on user satisfaction feedback, enabling the system to automatically adjust the sensitivity of semantic and behavioral features according to business results. While ensuring that user demands are responded to in a high-quality manner, it achieves the best balance between system automation efficiency and manual costs, reducing the difficulty of long-term operation and maintenance. Attached Figure Description

[0021] Figure 1 The flowchart is a human-computer interaction intent recognition method provided in an embodiment of the present invention.

[0022] Figure 2 This is a diagram illustrating the architecture of a human-computer interaction intent recognition system provided in an embodiment of the present invention. Detailed Implementation

[0023] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0024] The application of intelligent customer service systems is becoming increasingly widespread.

[0025] When dealing with highly sensitive issues such as "payment inquiries" or "long delays in logistics," users are often in a state of anxiety. At this time, users are often unable to express their needs in a standard logical structure, exhibiting symptoms such as speaking very quickly, frequently correcting themselves (self-correction), using a large number of colloquial interjections, and rapidly jumping between different sub-requests (such as "checking progress" and "complaining about personnel").

[0026] User anxiety leads to semantic fragmentation (improper sentence segmentation) and logical loops (repeatedly conveying the same information). This endogenous perturbation makes it difficult for standard NLP sequence labeling models to extract effective slots.

[0027] Redundant historical information in multi-turn dialogues can interfere with the current intent. When a user suddenly overturns their assumption from round N-2 in round N, the system is prone to falling into a "logical dead loop".

[0028] Sudden adjustments to the backend knowledge base or business policies (such as a surge in inquiries caused by a sudden large-scale power outage) can cause user interaction patterns to deviate from the historically preset "normal profile" in a short period of time.

[0029] Furthermore, user profiles are typically based on historical purchase / credit records (slow variables), but the user's identification intent at this moment is dominated by instantaneous emotions (fast variables). The adjustment weight of the "slow variable profile" to the "fast variable intent" is difficult to quantify dynamically, causing the model to still push redundant solutions according to the normal logic when the user's emotions erupt, leading to interaction crashes.

[0030] Deconstructing ambiguous intents requires collecting features from more modalities / dimensions, but as user emotional disturbances intensify, the entropy (noise) of the input signal increases exponentially. How to extract a deterministic core intent in an environment of escalating semantic noise remains a current technological bottleneck.

[0031] Therefore, this embodiment discloses a method for recognizing human-computer interaction intent based on user profiles, see [link to relevant documentation]. Figure 1 As shown. This process can build a complete human-computer interaction intent recognition system, see... Figure 2 As shown, the system consists of an input behavior perception unit, a semantic geometry mapping unit, a feature parallel computing unit, an intent disambiguation routing module, and a feedback update unit. These units work collaboratively to ensure that intent recognition dynamically evolves based on individual user interaction characteristics. This system can be used to improve ultra-fast response intelligent customer service systems in the financial or e-commerce after-sales fields. The implementation process of this invention will be described in detail below, combining the technical mechanism with specific scenarios.

[0032] S1. Data Acquisition and Basic Parameter Vector Construction The system monitors the user's input process through the front-end text interaction interface. In this embodiment, the input behavior collection window is limited to the closed-loop time interval from when the user clicks or activates the input box control to when the user clicks the "send" button or triggers a send command (such as pressing the Enter key).

[0033] Within this window, the system performs the following operations in real time: Semantic purity extraction: The system parses the input text, identifies and removes modal particles (such as "um," "that"), emoticons, and redundant characters resulting from repeated input. Semantic purity The calculation is shown in the following formula (1).

[0034] in The number of valid characters for the business. This represents the total number of input characters.

[0035] Behavioral thermal parameter acquisition: Synchronously record user actions within the window period and obtain the average backspace frequency. (Number of backspaces per second) and variance of input pauses (Reflecting the stability of the input rhythm), see formulas (2) and (3) below respectively.

[0036] in This refers to the number of times the backspace key is triggered within a single round of interaction. The input duration for this round. The time interval between two consecutive keystrokes. This represents the average time interval.

[0037] Vector encapsulation: Encapsulate the above parameters into a basic parameter vector. .

[0038] S2, Parallel computation of sub-features and state alignment Get Then, the system starts two computational dimensions in parallel: Dimension 1, Semantic Entropy Increase Trajectory Index A, is used to quantify the degree of logical collapse in the process of users organizing text. The calculation model is as shown in Equation (4).

[0039] in, These are preset semantic weight coefficients. When rise and When the value decreases (e.g., when users frequently correct errors and use colloquial language), the A value increases non-linearly, indicating poor stability of the representation logic.

[0040] Dimension 2, Dynamic Pressure Feedback Stiffness B, reflects the instantaneous emotional pressure of the user during input. The calculation model is shown in Equation (5).

[0041] in, To input the total duration, To correct the constant, This is a preset behavioral weighting coefficient. If the input rhythm fluctuates drastically, the B value will surge, indicating that the user is in a state of anxiety or hesitation.

[0042] The system employs a Hidden Markov Model (HMM) state alignment algorithm. Since the sampling frequency of behavioral feature B is typically higher than that of semantic logical feature A, the system uses the sampling clock of B as a reference to interpolate and complete A, ultimately constructing a cooperative state flow in the continuous time domain. .

[0043] S3. Reconstruction and generation of confidence in collaborative intent This step aims to deeply couple current behavioral characteristics with the user profile context.

[0044] Distance calculation: In a high-dimensional semantic vector space, the current input feature vector is... (i.e., the corresponding basic parameter vector) ) and the centroid of user profile historical preferences (This represents the user's past steady-state interaction habits) are compared, and their Mahalanobis distance is calculated. See equation (6).

[0045] in, Let covariance be the distribution of user historical interaction features. Its inverse matrix; This is the transpose operation of a matrix.

[0046] User profile, historical preferences, centroid It is the arithmetic mean vector of the user's historical steady-state characteristics in k-dimensional space, as shown in equation (7): Where N is the preset sliding window size (e.g., the size of the last 50 successful interactions) to ensure that the centroid can slowly shift with subtle changes in the user's long-term habits.

[0047] To calculate the Mahalanobis distance, we also need to calculate the covariance matrix of these historical sample points. , used to describe the correlation between various user input behavior indicators, see equation (8).

[0048] This matrix records the user's unique "input habit distribution." For example, if the user is accustomed to thinking while typing, their historical data will show... The dimension will have a large variance, and this characteristic will be recorded in middle.

[0049] Reconstruct the confidence level C of the collaborative intent according to formula (9).

[0050] in, is the confidence level of the original intent output by the NLP engine; k is the adjustment coefficient for logical behavior mismatch.

[0051] Entropy smoothing: when When an item exceeds a preset threshold (i.e., a severe mismatch between logic and pressure), the system forcibly reduces the probability advantage of the Top-1 intent and prevents the system from blindly providing recognition results through entropy smoothing.

[0052] S4, Path Diversion Decision Logic The system implements a three-way adaptive load splitting strategy based on the value of C: Direct retrieval execution: If The system determines that the identification result is highly clear, directly retrieves the business database and pushes a standard response (e.g., directly process the business).

[0053] Semantic induced clarification: If The system determines that there is ambiguity, extracts the conflict slot with the highest weight (e.g., time or location is unknown), and generates a disambiguation prompt containing candidate options.

[0054] Human-assisted takeover: If If a transient rate of change exceeding the threshold is detected in the B value, the interaction is determined to be in an abnormal state (such as extreme anger or input interruption caused by the incident), the engine is automatically suspended and a manual collaboration channel is established.

[0055] S5. Dynamic feedback update of weight parameters After the interaction is completed, the system obtains the user satisfaction S: S=1: The user clicks to rate the user as satisfied or the task is successfully completed (e.g., the order is successfully submitted).

[0056] S=0 indicates user dissatisfaction, task interruption, or forced transfer to human intervention midway.

[0057] The system uses the satisfaction level S to adjust the weighting coefficients according to equation (10): If S=1, then 1-S=0, and the weight remains unchanged; if S=0, then 1-S=1, and the weight increases by one step. By increasing the weights, the system will be more "vigilant" when faced with similar discrete behaviors in the future, resulting in a lower confidence level C, thus prioritizing the triggering of safe diversion paths (such as clarification or transfer to human intervention).

[0058] The technical solution disclosed in this embodiment shows improvements over traditional NLP solutions in four key quantitative indicators: intent recognition accuracy, interaction efficiency, and self-evolution performance. See Table 1.

[0059] Table 1 Comparison of Traditional NLP Solutions and the Solution of This Invention Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for recognizing human-computer interaction intent based on user profiles, characterized in that, Includes the following steps: S1. Acquire the user's original text sequence in real time through the text interaction interface, and collect the user's keystroke behavior features during the input process within a predetermined input behavior collection window. Extract the basic parameter vector containing semantic purity, backspace frequency, and pause variance. ; S2, Based on the aforementioned basic parameter vector Parallel computation is performed on the semantic entropy increase trajectory index A, which reflects the stability of text logic, and the dynamic pressure feedback stiffness B, which reflects the instantaneous emotional pressure of the user. A state alignment algorithm is then used to map A and B to a unified interactive state sequence space. S3. Based on the semantic entropy increase trajectory index A and the dynamic pressure feedback stiffness B, and combined with the geometric distance between the current input vector and the centroid of the user profile's historical preferences, the collaborative intent confidence C is generated through nonlinear transformation. S4. Based on the value and trend of the confidence level C of collaborative intent, adaptively switch between the three diversion paths: direct retrieval execution, semantic induced clarification, and manual collaborative takeover. S5. Based on the intent recognition results output under different traffic paths, obtain user satisfaction and dynamically update the calculated weight coefficients of A and B according to user satisfaction.

2. The human-computer interaction intent recognition method based on user profiles according to claim 1, characterized in that, The extraction process includes a basic parameter vector comprising semantic purity, backspace frequency, and pause variance. Specifically, it includes: Identify and calculate the proportion of non-business-related modal particles, emoticons, and repetitive / redundant characters in the total number of characters in the current input text, and extract the semantic purity of the text. ; The average backspace frequency of users in a single round of dialogue interaction is collected through the front-end interaction interface. Variance of input pauses Collect thermal parameters of user keystroke behavior; Constructing the basic parameter vector .

3. The human-computer interaction intent recognition method based on user profiles according to claim 2, characterized in that, The semantic entropy increase trajectory index A is used to quantify the degree of logical breakdown in the user's text organization process, and the calculation formula is as follows: In the formula, These are preset semantic weight coefficients; The formula for calculating the dynamic pressure feedback stiffness B is as follows: In the formula, Enter the total duration for the current round of text input; A correction constant to prevent overflow during division by zero; These are preset behavior weighting coefficients.

4. The human-computer interaction intent recognition method based on user profiles according to claim 1 or 3, characterized in that, The process of mapping A and B to a unified interaction state sequence space using a state alignment algorithm specifically includes: A hidden Markov model state alignment algorithm is employed, using the sampling clock of the dynamic pressure feedback stiffness B as a reference, to interpolate and complete the semantic entropy increase trajectory exponent A, thereby constructing a state flow in the continuous time domain. .

5. The human-computer interaction intent recognition method based on user profiles according to claim 1, characterized in that, Step S3 involves reconstructing the collaborative intent confidence C through nonlinear transformation, specifically including: In a high-dimensional semantic vector space, calculate the Mahalanobis distance between the current input vector and the centroid of user profile preferences. ; The formula for calculating the confidence level C of collaborative intent is as follows: In the formula, is the original intent confidence score output by the NLP engine; k is the logical behavior mismatch adjustment coefficient. The probability distribution of the original intent is smoothed by using the confidence level C of the collaborative intent. When the product of A and B exceeds a preset threshold, the probability advantage of the Top-1 intent is forcibly reduced.

6. The human-computer interaction intent recognition method based on user profiles according to claim 1 or 5, characterized in that, In step S4, there are three traffic diversion paths: direct retrieval execution, semantic-guided clarification, and manual collaborative takeover. The adaptive switching logic is as follows: Direct retrieval execution, that is: when When the intent is clear, the business knowledge base is retrieved directly based on the recognition results, and a standard response is pushed out. Semantic induced clarification, that is: when When the intent is ambiguous, the conflict slot with the highest weight in the current text is extracted to generate a disambiguation prompt containing candidate options; Human-assisted takeover, that is: when If the instantaneous jump rate of B is detected to exceed the preset threshold, the interaction is determined to be in an abnormal state, the automatic recognition engine is automatically suspended and a manual collaboration channel is established. in, , This is the preset path switching threshold.

7. The human-computer interaction intent recognition method based on user profiles according to claim 1, characterized in that, In step S5, user satisfaction S is defined as a binary signal: The calculated weight coefficients of A and B are dynamically updated based on user satisfaction S, as follows: When S=1, then 1-S=0, and the weight remains unchanged; When S=0, then 1-S=1, and the weight increases by one step. .

8. The human-computer interaction intent recognition method based on user profiles according to claim 1, characterized in that, The input behavior acquisition window is defined as a closed-loop time interval from when the user activates the input box control to when the command is triggered.

9. A human-computer interaction intent recognition and optimization system based on user profiles, characterized in that, include: The input behavior awareness unit is used to acquire the user's original text sequence in real time. Within a predetermined input behavior acquisition window, it collects the user's keystroke behavior features during the input process and extracts a basic parameter vector containing semantic purity, backspace frequency, and pause variance. ; Semantic geometry mapping unit, used in the basic parameter vector Based on this, we perform parallel computation of the semantic entropy increase trajectory index A, which reflects the stability of text logic, and the dynamic pressure feedback stiffness B, which reflects the instantaneous emotional pressure of users, and use a state alignment algorithm to map A and B to a unified interactive state sequence space. The feature parallel computing unit is used to reconstruct the collaborative intent confidence C through nonlinear transformation based on the output of the semantic geometric mapping unit, combined with the geometric distance between the current input vector and the centroid of the user profile's historical preferences. The intent disambiguation routing module is used to adaptively switch between three types of routing paths—direct retrieval execution, semantic inducement clarification, and manual collaborative takeover—based on the magnitude and trend of the collaborative intent confidence C. The feedback update unit is used to obtain user satisfaction and dynamically update the calculated weight coefficients of A and B based on user satisfaction.

10. A computer storage medium storing one or more programs, characterized in that: When the one or more programs are executed by an electronic device including multiple applications, the electronic device performs the human-computer interaction intent recognition method based on user profiles as described in any one of claims 1 to 8.

Images