Personalized prompt word generation method and system combining context awareness and interaction feedback

Abstract

The application discloses a personalized prompt word generation method and system combining context awareness and interactive feedback, relates to the technical field of data processing, and comprises the following steps: setting a user interaction dynamic sequence, performing intention association analysis, configuring a prompt word adjustment vector, combining a user preference database and a scene demand knowledge base, and determining intention recognition deviation; according to the prompt word adjustment vector and the intention recognition deviation, a prompt word optimization instruction set is configured; based on the prompt word optimization instruction set, the underlying configuration parameters of a double-channel intention prediction model are dynamically adjusted, and a personalized prompt word candidate set is generated; at the same time, based on different interactive scene types, the personalized prompt word candidate set is synchronously matched and called, and according to the intention matching accuracy and the user feedback response rate, dynamic feedback adjustment is performed in combination with a preset utility target. The application solves the technical problem that the personalized prompt word generation adaptation effect is poor in the prior art, and achieves the technical effect of improving the accuracy of personalized prompt words.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a personalized prompt word generation method and system combining context awareness and interactive feedback. BACKGROUND

[0002] In the existing personalized prompt word generation process, it is difficult to comprehensively integrate and analyze the user input data, the running state of the device and the environmental context parameters. These factors have important influence on the adaptation and generation of prompt words. Current technical methods often fail to fully consider these multi-dimensional data, resulting in that the generated prompt words cannot accurately reflect the real needs of users or adapt to different interactive scenarios, so that the personalization and precision of prompt words are limited, affecting the user's interactive experience. SUMMARY

[0003] The present application provides a personalized prompt word generation method and system combining context awareness and interactive feedback, which is used to solve the technical problem that in the existing personalized prompt word generation process, various factors such as user input data, device state and environmental context cannot be fully considered, resulting in poor prompt word adaptation effect.

[0004] In view of the above problems, the present application provides a personalized prompt word generation method and system combining context awareness and interactive feedback.

[0005] In a first aspect of the present application, a personalized prompt word generation method combining context awareness and interactive feedback is provided, which comprises:

[0006] Collecting user input data, device running state data and environmental context parameters, setting a user interactive dynamic sequence; performing intent association analysis according to the user interactive dynamic sequence, configuring a prompt word adjustment vector, and determining an intent recognition deviation in combination with a user preference database and a scene demand knowledge base; configuring a prompt word optimization instruction set according to the prompt word adjustment vector and the intent recognition deviation; dynamically adjusting the bottom configuration parameters of a double-channel intent prediction model based on the prompt word optimization instruction set to generate a personalized prompt word candidate set; at the same time, based on different interactive scene types, the personalized prompt word candidate set is synchronously matched and called, and dynamic feedback adjustment is performed in combination with a preset utility target according to the intent matching accuracy and the user feedback response rate.

[0007] In a second aspect of the present application, a personalized prompt word generation system combining context awareness and interactive feedback is provided, which comprises:

[0008] The data collection module is used for collecting user input data, equipment running state data and environmental context parameters, and setting a user interaction dynamic sequence; the intention correlation analysis module is used for performing intention correlation analysis according to the user interaction dynamic sequence, configuring a prompt word adjustment vector, and determining intention recognition bias in combination with a user preference database and a scene demand knowledge base; the instruction set configuration module is used for configuring a prompt word optimization instruction set according to the prompt word adjustment vector and the intention recognition bias; the prompt word generation module is used for dynamically adjusting underlying configuration parameters of a double-channel intention prediction model based on the prompt word optimization instruction set, and generating a personalized prompt word candidate set; and the feedback adjustment module is used for simultaneously performing synchronous matching calling on the personalized prompt word candidate set based on different interactive scene types, and performing dynamic feedback adjustment in combination with a preset utility target according to intention matching accuracy and user feedback response rate.

[0009] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0010] The present application collects user input data, equipment running state data and environmental context parameters, sets a user interaction dynamic sequence, performs intention correlation analysis according to the user interaction dynamic sequence, configures a prompt word adjustment vector, and determines intention recognition bias in combination with a user preference database and a scene demand knowledge base, configures a prompt word optimization instruction set according to the prompt word adjustment vector and the intention recognition bias, dynamically adjusts underlying configuration parameters of a double-channel intention prediction model based on the prompt word optimization instruction set, generates a personalized prompt word candidate set, simultaneously performs synchronous matching calling on the personalized prompt word candidate set based on different interactive scene types, and performs dynamic feedback adjustment in combination with a preset utility target according to intention matching accuracy and user feedback response rate. The present application solves the technical problem that in the prior art, multiple factors such as user input data, equipment state and environmental context cannot be fully considered in the personalized prompt word generation process, resulting in poor prompt word adaptation effect, and through collecting user input data, equipment state and environmental parameters, and performing intention recognition and dynamic adjustment in combination with a user preference database and a scene demand knowledge base, the generation and matching process of prompt words is optimized, and the technical effect of improving the accuracy of personalized prompt words is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0012] Figure 1A schematic diagram of the personalized prompt word generation method combining context awareness and interactive feedback provided in the embodiments of this application;

[0013] Figure 2 This is a schematic diagram of the structure of a personalized prompt word generation system that combines context awareness and interactive feedback, provided in an embodiment of this application.

[0014] Figure labeling: Data acquisition module 11, intent association analysis module 12, instruction set configuration module 13, prompt word generation module 14, feedback adjustment module 15. Detailed Implementation

[0015] This application provides a personalized prompt word generation method and system that combines context awareness and interactive feedback. It addresses the technical problem in existing technologies where the generation of personalized prompt words fails to adequately consider various factors such as user input data, device status, and environmental context, leading to poor prompt word adaptation. By collecting user input data, device status, and environmental parameters, and combining this with a user preference database and a scenario requirement knowledge base for intent recognition and dynamic adjustment, the application optimizes the prompt word generation and matching process, thereby improving the accuracy of personalized prompt words.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0018] Example 1, as Figure 1 As shown, this application provides a method for generating personalized prompt words that combines context awareness and interactive feedback, the method comprising:

[0019] Step S100: Collect user input data, device operating status data, and environmental context parameters, and set a dynamic sequence of user interaction.

[0020] In this embodiment, user input data, device operating status data, and environmental parameters are collected synchronously first. User input data includes images of questions uploaded by students and voice prompts, control commands and authorization codes set by parents, and teaching highlights and batch exercises annotated by teachers. Device operating status data covers screen orientation, battery level, and network connectivity. Environmental parameters are acquired through terminal sensors, including time information from parents and the status of external devices connected to teachers. All these multi-source data streams are timestamped and fused in a structured manner to ultimately construct a dynamic sequence of user interactions.

[0021] Step S200: Perform intent association analysis based on the dynamic sequence of user interaction, configure prompt word adjustment vectors, and determine intent recognition bias by combining user preference database and scenario requirement knowledge base.

[0022] In this embodiment, intent association analysis is performed based on the dynamic sequence of user interactions. This process is driven by a dual-channel intent prediction model, which performs deep analysis of the dynamic sequence of user interactions. Its explicit intent channel directly processes the structured input information in the sequence through a Transformer architecture, identifying the explicit needs clearly expressed by the user; simultaneously, the latent intent channel, based on a graph neural network architecture, mines the latent needs reflected in implicit information such as device operating status data and environmental context parameters within the sequence. Based on this multi-dimensional analysis result, a prompt word adjustment vector is configured. This prompt word adjustment vector is a multi-dimensional feature set that quantifies various adjustment parameters.

[0023] Subsequently, personalized historical behavior data stored in the user preference database and standardized scenario requirement parameters defined in the scenario requirement knowledge base were used to cross-validate the preliminary intent recognition results output by the dual-channel intent prediction model. By comparing the preliminary intent recognition results output by the model, historical preference features in the user preference database, and standard scenario requirement parameters in the scenario requirement knowledge base, the intent recognition bias was finally determined. The intent recognition bias refers to the quantitative difference between the preliminary intent recognition results output by the dual-channel intent prediction model and the historical preference features in the user preference database and the standard scenario requirement parameters in the scenario requirement knowledge base.

[0024] Step S300: Configure the prompt word optimization instruction set based on the prompt word adjustment vector and the intent recognition deviation.

[0025] In this embodiment, when configuring the prompt word optimization instruction set based on the prompt word adjustment vector and intent recognition deviation, the parameters of the prompt word adjustment vector are first parsed, and the quantified adjustment requirements in the prompt word adjustment vector are converted into specific operation instructions. For example, when the prompt word adjustment vector includes a parameter to increase parsing detail, it is converted into an operation instruction to increase the decomposition level of the problem-solving steps.

[0026] Then, intent recognition bias is introduced as a key constraint, and an optimization environment is constructed using a reinforcement learning framework. The state space of this environment consists of the parameter configuration of the current prompting strategy, the action space is defined as the adjustment operations on prompt word complexity, semantic depth, and interaction method, and the reward function comprehensively calculates the balance between intent matching accuracy and user cognitive load. During the optimization process, firstly, multiple rounds of strategy exploration are conducted in the simulated environment based on a proximal policy optimization algorithm, continuously adjusting the semantic encoding dimension in the explicit demand response scheme; then, the graph neural network attention mechanism in the latent demand mining strategy is optimized; and simultaneously, the hierarchical threshold in the prompt complexity adaptation parameters is calibrated. Through these adjustments, the final output prompting strategy can reduce intent recognition bias to within the preset dynamic optimization threshold range. After multiple rounds of iterative optimization, the optimal parameter configuration is encapsulated to form a prompt word optimization instruction set. The prompt word optimization instruction set includes explicit demand response schemes, latent demand mining strategies, and prompt complexity adaptation parameters. Specifically, the explicit demand response scheme specifies how to respond to the user's explicitly expressed needs, the latent demand mining strategy defines the method for discovering deep needs, and the prompt complexity adaptation parameters control the information density and presentation format of the final output.

[0027] Step S400: Based on the prompt word optimization instruction set, dynamically adjust the underlying configuration parameters of the dual-channel intent prediction model to generate a personalized prompt word candidate set.

[0028] In this embodiment, when dynamically adjusting the underlying configuration parameters of the dual-channel intent prediction model based on the prompt word optimization instruction set, the semantic encoding dimension is adjusted according to the explicit demand response scheme, the potential demand mining depth of the graph neural network is optimized according to the potential demand mining strategy, and the output complexity hierarchical threshold of the model is calibrated according to the prompt complexity adaptation parameters. After the parameter adjustment is completed, the optimized dual-channel intent prediction model is deployed in a lightweight simulation interaction environment for inference verification. When the simulation results meet the preset utility target, the model generates a personalized prompt word candidate set containing multiple candidate prompt words based on the current interaction context. Each prompt word is evaluated from multiple dimensions such as intent matching degree, user preference conformity, and scene adaptability, forming an optimized result set that can be directly invoked.

[0029] Furthermore, in the method provided in the application embodiments, the underlying configuration parameters of the dual-channel intent prediction model are dynamically adjusted based on the prompt word optimization instruction set, and the method further includes:

[0030] Based on the prompt word optimization instruction set, which includes explicit demand response schemes, potential demand mining strategies, and prompt complexity adaptation parameters, the semantic encoding dimension of the explicit channel and the potential demand mining depth of the potential channel are dynamically adjusted according to the intent matching coverage. Based on the user cognitive load value, the prompt word complexity layering threshold and semantic expression conciseness parameters are optimized. When the interaction process simplification rate is detected to be lower than the efficiency setting value, the asynchronous intent enhancement network layer is activated.

[0031] In this embodiment, a dynamic parameter adjustment process for the dual-channel intent prediction model is initiated based on a prompt word optimization instruction set that includes explicit demand response schemes, latent demand mining strategies, and prompt complexity adaptation parameters. During the intent matching coverage monitoring and response phase, model performance is evaluated by continuously tracking user interaction behavior. Specifically, this includes recording user adoption, correction, and ignore / regenerate behaviors related to generated prompt words, and quantifying these behaviors into numerical scores. Direct adoption scores 1 point, edited use scores 0.5 points, and ignore scores 0 points. The average of these scores is then calculated within a set time window to obtain the intent matching coverage. When the intent matching coverage is detected to be below a preset threshold of 85%, the semantic encoding dimension of the explicit channel is adjusted according to the explicit demand response scheme. Semantic understanding is enhanced by increasing the number of heads in the self-attention mechanism of the Transformer architecture. Simultaneously, the latent demand mining depth of the latent channel is adjusted according to the latent demand mining strategy, strengthening the detection capability of the user's implicit intent by increasing the number of graph convolutional layers in the graph neural network.

[0032] During the user cognitive load management phase, the user cognitive load value is assessed by monitoring efficiency parameters such as user operation response latency, frequency of interactive misoperations, and task completion path. When the user cognitive load value exceeds the safety threshold of 0.75, the complexity stratification threshold of the prompt words is automatically increased according to the prompt complexity adaptation parameter, and the semantic expression conciseness parameter is adjusted simultaneously to reduce the user's cognitive burden by simplifying sentence structure and controlling information density.

[0033] During the interaction flow monitoring phase, the interaction flow simplification rate is obtained. This process first defines a standard task flow and sets its optimal number of steps as a benchmark. Then, the actual number of steps required to complete the task during user interaction is recorded. Finally, the interaction flow simplification rate is obtained by calculating the ratio of the actual number of steps to the optimal number of steps. When the interaction flow simplification rate is detected to be consistently below the efficiency setpoint of 70%, it indicates that the current interaction efficiency is not meeting expectations, and the asynchronous intent enhancement network layer is immediately activated. This network layer performs in-depth analysis of the user's interaction history through parallel-running Long Short-Term Memory (LSTM) networks and temporal convolutional networks. Its output is integrated with the main output of the dual-channel intent prediction model through a feature fusion gating mechanism to improve the accuracy of intent recognition and interaction efficiency in complex scenarios.

[0034] Furthermore, the method provided in the application embodiments also includes:

[0035] During the execution of the prompt optimization path, a prompt configuration item dependency graph is constructed. Using the prompt configuration item dependency graph, when an intent parsing conflict is detected, the intent reasoning constraints are updated, and the optimized path after the conflict is resolved is marked as a priority prompt strategy snapshot and associated with similar user profile files.

[0036] In this embodiment, during the execution of the prompt optimization path, a dependency graph of prompt configuration items is constructed using graph neural network technology. This graph adopts a directed acyclic graph structure, where nodes represent specific configuration items such as explicit demand response schemes, potential demand mining strategies, and prompt complexity adaptation parameters, and edges represent parameter dependencies and execution order constraints between configuration items. When an intent parsing conflict is detected through real-time monitoring of user interaction behavior data, i.e., when there is a semantic logical contradiction between the explicit channel output and the potential channel output of the dual-channel intent prediction model, a conflict resolution mechanism based on the dependency graph is initiated. This involves first locating the key configuration node causing the conflict using a graph topology sorting algorithm, and then updating the intent reasoning constraints using a constraint propagation algorithm. These constraints explicitly define the value range and interaction relationships of each configuration parameter in the form of first-order logical predicates.

[0037] During conflict resolution, a case-based reasoning approach is employed to mark the optimized complete parameter configuration path as a priority suggestion strategy snapshot. This snapshot contains validated configuration item combinations and their effectiveness evaluation data in actual interactions, including quantitative metrics such as intent matching accuracy. Finally, a user profile classification mechanism based on the K-means clustering algorithm is used to associate the priority suggestion strategy snapshot with user profile files that share similar characteristics.

[0038] Furthermore, in the method provided in the application embodiment, before dynamically adjusting the underlying configuration parameters of the dual-channel intent prediction model based on the prompt word optimization instruction set, it further includes:

[0039] Before dynamically adjusting the underlying configuration parameters of the dual-channel intent prediction model, record the state snapshot and utility difference data of the dual-channel intent prediction model to locate the key configuration items that cause the decrease in the satisfaction score of the prompt words; based on the key configuration items, generate compensation adjustment instructions and insert optimization path breakpoints.

[0040] In this embodiment, before dynamically adjusting the underlying configuration parameters of the dual-channel intent prediction model, a model state recording and diagnostic process is first executed. This process records a snapshot of the dual-channel intent prediction model's state using serialization technology. This snapshot saves the complete configuration state of the underlying configuration parameters at the current moment, including the semantic encoding dimensions of the explicit channels (such as the number of self-attention heads and the hidden layer dimension configuration of the feedforward network), the potential demand mining depth of the potential channels (such as the number of graph convolutional layers and the adjacency matrix connectivity configuration), and the cue complexity adaptation parameters (such as the complexity layering threshold setting and the semantic expression conciseness parameter configuration).

[0041] Simultaneously, time series analysis was used to compare and analyze recent user interaction data to calculate utility difference data. Specifically, a regression model was established between changes in configuration parameters and prompt satisfaction scores. By analyzing the correlation between the changing trends of each configuration parameter and the prompt satisfaction scores, utility difference data was obtained. This utility difference data reflects the degree of influence of changes in each configuration parameter on prompt satisfaction scores in the form of relative changes. For example, when the semantic encoding dimension of the explicit channel increases from a low value to a high value, the prompt satisfaction score shows a downward trend, thus recording the utility difference data for that configuration item as having a negative impact.

[0042] The satisfaction score for the prompt words is obtained using a user behavior analysis method. By statistically analyzing users' adoption, modification, and ignore / regenerate behaviors in relation to the generated prompt words, higher scores are assigned to direct adoption, medium scores to edited use, and lower scores to ignore. Then, the weighted average of these behavior scores is calculated within a set time window to obtain a quantitative satisfaction score for the prompt words.

[0043] Based on recorded state snapshots and utility difference data, a root cause analysis method using a decision tree algorithm is employed to locate key configuration items that lead to a decline in cue word satisfaction scores. By constructing a feature importance ranking of configuration parameters and satisfaction scores, specific configuration parameters with the greatest impact on score decline are identified as key configuration items. For example, when the analysis reveals that changes in the semantic encoding dimension of the explicit channel contribute significantly more to the decline in satisfaction scores than other parameters, that parameter is marked as a key configuration item.

[0044] Subsequently, compensation adjustment instructions are generated based on these key configuration items. These instructions include parameter correction schemes and adjustment ranges for the key configuration items. For example, for the key configuration item of explicit channel semantic encoding dimension, a compensation adjustment instruction is generated to reduce its value from a high level to a moderate level, and the corresponding adjustment range is set. Finally, an interruption handling mechanism is used to insert these compensation adjustment instructions into the optimization path interruption point, ensuring that these compensation adjustments are executed first in subsequent parameter adjustment processes, thereby effectively restoring and improving the satisfaction score of the prompt words.

[0045] Furthermore, the method provided in the application embodiments also includes:

[0046] In multi-device collaborative interaction scenarios, the interaction modal differences of each device are analyzed; based on the interaction modal differences of each device, intent parsing tasks are dynamically allocated and the prompt word parameter configuration between each device is optimized synchronously; when the prompt word response time difference between each device exceeds the time difference threshold, the intent parsing load migration mechanism is enabled.

[0047] In this embodiment, in a multi-device collaborative interaction scenario, the interaction modal differences of each device are first analyzed by identifying device type and interaction method. Specifically, this includes identifying device types such as tablets, smartphones, or smartwatches; detecting supported input methods (text input, voice reception, or image capture); output methods (screen display or voice broadcast); and processing capabilities (computing resources and network latency). Based on the analyzed interaction modal differences, a load balancing algorithm is used to dynamically allocate intent parsing tasks. Computationally intensive latent demand mining tasks are assigned to devices with strong processing capabilities, while real-time explicit intent parsing tasks are assigned to devices with low latency. Simultaneously, a synchronization engine is configured to optimize the prompt word parameter configurations between devices in real time, including core parameters such as semantic encoding dimension, latent demand mining depth, and prompt complexity adaptation parameters.

[0048] When the time difference between prompt word responses across devices exceeds a preset threshold, the intent parsing load balancing mechanism is immediately activated via the time synchronization protocol. This mechanism first identifies device nodes with response delays using a task status monitor, then uses a task migration algorithm to redistribute unfinished intent parsing tasks to other available devices. Simultaneously, a configuration synchronization engine maintains consistency in prompt word parameter configurations across devices. During load balancing, task allocation status and device load records are continuously updated to ensure a stable interactive experience and high intent parsing efficiency in a multi-device collaborative environment.

[0049] Furthermore, the method provided in the application embodiments also includes:

[0050] Referring to the dependency graph of the configuration items, a configuration prompt influence factor graph is generated. This graph is used to quantify the weight of each configuration item on intent matching accuracy and user experience. When a user inputs a modification request, the graph predicts the trend of prompt utility changes and generates optimization suggestion alerts. Redundant configuration items are identified through a graph backtracking mechanism. Based on the redundant configuration items, potential interference factors are mined using the prompt influence factor graph, and the prompt word optimization instruction set is iteratively corrected until the prompt word utility test passes.

[0051] In this embodiment, a prompt influence factor map is first configured by referring to the dependency graph of the prompt configuration items. This prompt influence factor map is constructed using a multiple linear regression analysis method based on historical interaction data. During the construction of the prompt influence factor map, configuration item parameter values, intent matching accuracy data, and user satisfaction score data are collected from historical interaction data. Intent matching accuracy data is calculated by comparing the output of the dual-channel intent prediction model with the actual user adoption behavior. User satisfaction score data is calculated by statistically analyzing the weighted average of users' direct adoption, edited use, and ignored behaviors of prompt words. The regression coefficients of each configuration item parameter are solved using the least squares method, and then the regression coefficients are standardized to obtain the weight values ​​of each configuration item on intent matching accuracy and user experience, thus completing the configuration of the prompt influence factor map.

[0052] When a user inputs a modification request, a time series prediction algorithm is used based on the prompt influence factor map to predict the trend of prompt utility changes. This prediction process first identifies configuration items affected by the user's modification request, then obtains the weight values ​​of these configuration items in the prompt influence factor map, and quantifies the magnitude of the user's modification request. By analyzing the product relationship between the configuration item weight values ​​and the modification magnitude, the expected impact on intent matching accuracy and user satisfaction rating is calculated. A time series prediction model is established to predict the trajectory of prompt utility changes over a future period based on historical utility data trends. Simultaneously, by comparing the predicted prompt utility change trend with the preset utility target, when a deviation from the preset target is detected, an optimization suggestion alarm prompt containing specific parameter adjustment suggestions is generated.

[0053] Subsequently, a graph backtracking mechanism was used to identify redundant configuration items. This mechanism employs a depth-first search algorithm to traverse the dependency graph of cue configuration items, searching for those whose impact on intent matching accuracy and user experience weight is below a preset threshold. Based on the identified redundant configuration items, the Apriori association rule mining algorithm was used to uncover potential demand interference factors, using the cue influence factor graph as a foundation. These potential demand interference factors were obtained by statistically analyzing the co-occurrence relationship between redundant configuration items and negative user feedback behaviors in historical interaction data.

[0054] Finally, based on the identified potential demand interference factors, the prompt word optimization instruction set is iteratively revised. This revision is achieved by adjusting the parameter settings of explicit demand response schemes, the execution logic of the potential demand mining strategy, and the threshold configuration of prompt complexity adaptation parameters within the prompt word optimization instruction set. This iterative revision process continues until the prompt word effectiveness verification is passed in an online testing environment. The standard for passing the prompt word effectiveness verification is that the intent matching accuracy data and user satisfaction score data generated by the revised prompt words in the testing environment both reach the preset threshold requirements.

[0055] Step S500: Simultaneously, based on different interaction scenario types, the personalized prompt word candidate set is synchronously matched and invoked, and dynamic feedback adjustment is performed based on the intent matching accuracy, user feedback response rate, and preset utility target.

[0056] In this embodiment, personalized prompt word candidate sets are synchronously matched and invoked based on different interaction scenario types. The interaction scenario types are identified using an attention-based neural network scene classifier. This classifier analyzes device operating status data and environmental context parameters in the dynamic sequence of user interactions to output scenario type identifiers, including learning scenarios, tutoring scenarios, and lesson preparation scenarios. During the matching process, a semantic similarity-based retrieval algorithm is used to match the current interaction scenario type with the scenario tags in the personalized prompt word candidate set, generating a scenario-appropriate prompt word subset.

[0057] The matching results are then evaluated based on intent matching accuracy, which is calculated by comparing the consistency between the actual prompts used and the user's true needs. Specifically, the accuracy calculation formula is to divide the number of correctly matched prompts by the total number of prompts invoked. Simultaneously, user feedback response rate is monitored. This metric is obtained by statistically analyzing the ratio of the number of direct user interactions with prompts (including adoption, modification, and ignoring) to the total number of prompts displayed.

[0058] Next, based on real-time data of intent matching accuracy and user feedback response rate, dynamic feedback adjustment is performed in conjunction with a preset utility target. This adjustment process employs a PID control algorithm, using intent matching accuracy and user feedback response rate as input variables, comparing them with the preset utility target value, and outputting adjustment parameters for the personalized prompt word candidate set through proportional, integral, and derivative calculations. These adjustment parameters include the candidate set update frequency, matching threshold, and priority weight, etc. Based on these parameters, the composition structure and calling strategy of the personalized prompt word candidate set are optimized in real time to complete closed-loop adjustment.

[0059] Furthermore, the method provided in the application embodiments also includes:

[0060] Collect prompt word call frequency data and corresponding user satisfaction score snapshots from historical interaction tasks; based on the prompt word call frequency data and corresponding user satisfaction score snapshots, extract semantic association features and demand mining delay indicators corresponding to parallel processing tasks in the parallel inference path; optimize the decision nodes of the dual-channel intent prediction model based on semantic association features and demand mining delay indicators, so that the utility improvement ratio of the generated prompt word optimization path meets the dynamic optimization threshold.

[0061] In this embodiment, the frequency data of prompt words in historical interaction tasks and the corresponding user satisfaction rating snapshots are first collected. That is, by using a sliding time window statistical method, the number of times each prompt word is called is recorded at a fixed time period to form the call frequency data. At the same time, the five-star satisfaction rating after each use is collected through the user interaction interface to form the user satisfaction rating snapshot.

[0062] Subsequently, based on the frequency data of prompt words and the corresponding user satisfaction rating snapshots, semantic association features and demand mining latency indicators are extracted for parallel processing tasks in the parallel inference path. In this process, based on the frequency data of prompt words and the user satisfaction rating snapshots, word vector embedding technology is used to convert the prompt words in the parallel inference path into numerical vectors. Semantic association features are obtained by calculating the cosine similarity between these vectors, and demand mining latency indicators are obtained by measuring the time interval from when a user submits a question to when the corresponding prompt word is generated.

[0063] Finally, based on semantic association features and demand mining delay indicators, the gradient descent algorithm is used to adjust the parameter weights in the decision nodes of the dual-channel intent prediction model. Through iterative optimization, the weight coefficients of semantic association features and the tolerance values ​​of demand mining delay indicators are optimized, ultimately achieving a dynamic optimization threshold for the utility improvement ratio of the generated prompt word optimization path. The utility improvement ratio is calculated as the ratio of the increase in user satisfaction scores before and after optimization to the growth rate of prompt word call frequency. This process completes the optimization of the decision nodes of the dual-channel intent prediction model.

[0064] Furthermore, the method provided in the application embodiments also includes:

[0065] Configure cue utility monitoring metrics, including intent matching coverage, user cognitive load, and interaction flow simplification rate; obtain semantic consistency deviation during multimodal input synchronous parsing, and use the instruction backlog depth and priority inversion number of the intent prediction queue as auxiliary metrics to measure demand response efficiency.

[0066] In this embodiment, the cue utility monitoring metrics are first configured, including intent matching coverage, user cognitive load, and interaction flow simplification rate. Intent matching coverage is calculated by statistically analyzing user adoption, modification, and ignoring of cue words. Specifically, a weighted average algorithm is used, assigning 1 point for direct adoption, 0.5 points for editing and using, and 0 points for ignoring. The average score of these behaviors is calculated within a set time window to obtain the intent matching coverage. User cognitive load is obtained through multi-dimensional behavioral feature analysis, including measuring user operation response latency, statistically analyzing the frequency of interaction errors, and recording the number of task completion steps. These parameters are input into a support vector machine-based evaluation model to obtain a quantified user cognitive load value. This support vector machine evaluation model uses operation response latency, error frequency, and task step count from historical interaction data as training input data, and user cognitive load values ​​labeled by technical experts as training output labels for training. The interaction flow simplification rate is calculated by comparing the actual number of user operation steps with a preset optimal number of steps, and the ratio of the actual number of steps to the optimal number of steps is used as the interaction flow simplification rate.

[0067] Next, the semantic consistency deviation during synchronous parsing of multimodal inputs is obtained. This deviation is calculated by comparing the semantic representation vectors generated after processing text, image, and speech inputs through their respective parsing channels. The cosine similarity algorithm is used to calculate the difference between the semantic vectors of different modalities as the semantic consistency deviation. Simultaneously, the instruction backlog depth and priority inversion count in the intent prediction queue are used as auxiliary indicators to measure demand response efficiency. The instruction backlog depth is obtained by counting the number of pending instructions in the intent prediction queue, and the priority inversion count is obtained by recording the number of times high-priority instructions are blocked by low-priority instructions. When the instruction backlog depth continuously increases and the priority inversion count occurs frequently, it indicates that demand response efficiency is declining. These auxiliary indicators, together with the prompt utility monitoring indicators, constitute a complete performance evaluation system, providing comprehensive data support for optimizing prompt generation strategies.

[0068] Furthermore, in the method provided in the application embodiment, based on the prompt word optimization instruction set, the underlying configuration parameters of the dual-channel intent prediction model are dynamically adjusted to generate a personalized prompt word candidate set, which further includes:

[0069] A lightweight simulation interaction environment is embedded to simulate the reasoning behavior of a dynamically adjusted dual-channel intent prediction model. When the simulation results meet the preset utility target, the prompt word configuration changes are batch-retained in the personalized prompt word candidate set. Utility deviation data is introduced and combined with the prompt influence factor map to make targeted adjustments to the prompt word optimization instruction set.

[0070] In this embodiment, a lightweight simulated interaction environment is embedded. This environment is built based on containerization technology and generates simulated input by loading historical interaction data, executing the inference process of a dynamically adjusted dual-channel intent prediction model. During the simulation, cue utility monitoring indicators such as intent matching coverage, user cognitive load, and interaction flow simplification rate are calculated in real time. When all indicators continuously reach the preset utility target threshold, the verified cue word configuration changes are batch-stored into a personalized cue word candidate set.

[0071] Simultaneously, utility deviation data, obtained by comparing the difference in cue utility monitoring indicators between the simulated and production environments, is introduced. Combined with a cue influence factor graph recording the weights of each configuration item, targeted adjustments are made to the cue word optimization instruction set. This adjustment process, based on the magnitude and direction of change of the utility deviation data and referring to the weight ratios of configuration items in the cue influence factor graph, specifically modifies the parameter configurations of explicit demand response schemes, the execution logic of potential demand mining strategies, and the threshold settings of cue complexity adaptation parameters within the cue word optimization instruction set. Ultimately, this targeted adjustment of the cue word optimization instruction set is completed, generating a new, optimized version of the cue word optimization instruction set.

[0072] In summary, the embodiments of this application have at least the following technical effects:

[0073] This application collects user input data, device operating status data, and environmental context parameters to set a dynamic sequence of user interactions. Based on this sequence, it performs intent association analysis, configures prompt word adjustment vectors, and determines intent recognition deviations by combining user preference databases and scenario requirement knowledge bases. Based on the prompt word adjustment vectors and intent recognition deviations, it configures a prompt word optimization instruction set. Based on this optimization instruction set, it dynamically adjusts the underlying configuration parameters of the dual-channel intent prediction model to generate a personalized prompt word candidate set. Simultaneously, based on different interaction scenario types, it synchronously matches and calls the personalized prompt word candidate set, and dynamically adjusts the feedback based on intent matching accuracy, user feedback response rate, and a preset utility target. This invention solves the technical problem in existing technologies where the personalized prompt word generation process cannot fully consider multiple factors such as user input data, device status, and environmental context, resulting in poor prompt word adaptation. By collecting user input data, device status, and environmental parameters, and combining user preference databases and scenario requirement knowledge bases for intent recognition and dynamic adjustment, it optimizes the prompt word generation and matching process, thereby improving the accuracy of personalized prompt words.

[0074] Example 2, based on the same inventive concept as the personalized prompt word generation method combining context awareness and interactive feedback in the aforementioned examples, such as... Figure 2As shown, this application provides a personalized prompt word generation system that combines context awareness and interactive feedback. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0075] The data acquisition module 11 is used to collect user input data, device operating status data, and environmental context parameters, and set a dynamic sequence of user interaction. The intent association analysis module 12 is used to perform intent association analysis based on the dynamic sequence of user interaction, configure prompt word adjustment vectors, and determine intent recognition deviations by combining user preference databases and scenario requirement knowledge bases. The instruction set configuration module 13 is used to configure an optimized prompt word instruction set based on the prompt word adjustment vectors and the intent recognition deviations. The prompt word generation module 14 is used to dynamically adjust the underlying configuration parameters of the dual-channel intent prediction model based on the optimized prompt word instruction set, and generate a personalized prompt word candidate set. The feedback adjustment module 15 is used to simultaneously match and call the personalized prompt word candidate set based on different interaction scenario types, and perform dynamic feedback adjustment based on intent matching accuracy, user feedback response rate, and preset utility targets.

[0076] Furthermore, the system is also used to implement the following functions:

[0077] Collect prompt word call frequency data and corresponding user satisfaction score snapshots from historical interaction tasks; based on the prompt word call frequency data and corresponding user satisfaction score snapshots, extract semantic association features and demand mining delay indicators corresponding to parallel processing tasks in the parallel inference path; optimize the decision nodes of the dual-channel intent prediction model based on semantic association features and demand mining delay indicators, so that the utility improvement ratio of the generated prompt word optimization path meets the dynamic optimization threshold.

[0078] Furthermore, the system is also used to implement the following functions:

[0079] Configure cue utility monitoring metrics, including intent matching coverage, user cognitive load, and interaction flow simplification rate; obtain semantic consistency deviation during multimodal input synchronous parsing, and use the instruction backlog depth and priority inversion number of the intent prediction queue as auxiliary metrics to measure demand response efficiency.

[0080] Furthermore, the system is also used to implement the following functions:

[0081] Based on the prompt word optimization instruction set, which includes explicit demand response schemes, potential demand mining strategies, and prompt complexity adaptation parameters, the semantic encoding dimension of the explicit channel and the potential demand mining depth of the potential channel are dynamically adjusted according to the intent matching coverage. Based on the user cognitive load value, the prompt word complexity layering threshold and semantic expression conciseness parameters are optimized. When the interaction process simplification rate is detected to be lower than the efficiency setting value, the asynchronous intent enhancement network layer is activated.

[0082] Furthermore, the system is also used to implement the following functions:

[0083] During the execution of the prompt optimization path, a prompt configuration item dependency graph is constructed. Using the prompt configuration item dependency graph, when an intent parsing conflict is detected, the intent reasoning constraints are updated, and the optimized path after the conflict is resolved is marked as a priority prompt strategy snapshot and associated with similar user profile files.

[0084] Furthermore, the system is also used to implement the following functions:

[0085] Before dynamically adjusting the underlying configuration parameters of the dual-channel intent prediction model, record the state snapshot and utility difference data of the dual-channel intent prediction model to locate the key configuration items that cause the decrease in the satisfaction score of the prompt words; based on the key configuration items, generate compensation adjustment instructions and insert optimization path breakpoints.

[0086] Furthermore, the system is also used to implement the following functions:

[0087] In multi-device collaborative interaction scenarios, the interaction modal differences of each device are analyzed; based on the interaction modal differences of each device, intent parsing tasks are dynamically allocated and the prompt word parameter configuration between each device is optimized synchronously; when the prompt word response time difference between each device exceeds the time difference threshold, the intent parsing load migration mechanism is enabled.

[0088] Furthermore, the system is also used to implement the following functions:

[0089] Referring to the dependency graph of the configuration items, a configuration prompt influence factor graph is generated. This graph is used to quantify the weight of each configuration item on intent matching accuracy and user experience. When a user inputs a modification request, the graph predicts the trend of prompt utility changes and generates optimization suggestion alerts. Redundant configuration items are identified through a graph backtracking mechanism. Based on the redundant configuration items, potential interference factors are mined using the prompt influence factor graph, and the prompt word optimization instruction set is iteratively corrected until the prompt word utility test passes.

[0090] Furthermore, the system is also used to implement the following functions:

[0091] A lightweight simulation interaction environment is embedded to simulate the reasoning behavior of a dynamically adjusted dual-channel intent prediction model. When the simulation results meet the preset utility target, the prompt word configuration changes are batch-retained in the personalized prompt word candidate set. Utility deviation data is introduced and combined with the prompt influence factor map to make targeted adjustments to the prompt word optimization instruction set.

[0092] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A personalized prompt word generation method combining context awareness and interactive feedback, characterized in that, The method includes: Collect user input data, device operating status data, and environmental context parameters, and set dynamic sequences for user interactions; Based on the dynamic sequence of user interactions, intent association analysis is performed, prompt word adjustment vectors are configured, and intent recognition bias is determined by combining user preference database and scenario requirement knowledge base. Based on the prompt words, adjust the vector and the intent recognition deviation, and configure a prompt word optimization instruction set; Based on the aforementioned prompt word optimization instruction set, the underlying configuration parameters of the dual-channel intent prediction model are dynamically adjusted to generate a personalized prompt word candidate set; Meanwhile, based on different interaction scenario types, the personalized prompt word candidate set is synchronously matched and invoked, and dynamic feedback adjustment is performed based on intent matching accuracy, user feedback response rate, and preset utility goals. The method of dynamically adjusting the underlying configuration parameters of the dual-channel intent prediction model based on the prompt word optimization instruction set also includes: Based on the prompt word optimization instruction set, which includes explicit demand response schemes, potential demand mining strategies and prompt complexity adaptation parameters, the semantic encoding dimension of the explicit channel and the potential demand mining depth of the potential channel are dynamically adjusted according to the intent matching coverage. Based on the user's cognitive load value, the threshold for the complexity of prompt words and the parameters for the conciseness of semantic expression are optimized. When the simplification rate of the interaction process is detected to be lower than the efficiency setting value, the asynchronous intent enhancement network layer is activated. This includes: During the execution of the prompt optimization path, a dependency graph of prompt configuration items is constructed; Using the aforementioned prompt configuration item dependency graph, when an intent parsing conflict is detected, the intent reasoning constraints are updated, and the optimized path after the conflict is resolved is marked as a priority prompt strategy snapshot and associated with similar user profile files.

2. The personalized prompt word generation method combining context awareness and interactive feedback as described in claim 1, characterized in that, Collect data on the frequency of prompt words in historical interaction tasks and corresponding snapshots of user satisfaction scores; Based on the frequency data of the prompt words and the corresponding user satisfaction rating snapshots, semantic association features and demand mining delay indicators are extracted from the parallel processing tasks in the parallel inference path. The decision nodes of the dual-channel intent prediction model are optimized based on semantic association features and demand mining delay indicators, so that the utility improvement ratio of the generated prompt words to optimize the path meets the dynamic optimization threshold.

3. The personalized prompt word generation method combining context awareness and interactive feedback as described in claim 2, characterized in that, Configure cue utility monitoring metrics, including intent matching coverage, user cognitive load, and interaction flow simplification rate; We obtain the semantic consistency deviation during synchronous parsing of multimodal inputs, and use the instruction backlog depth and priority inversion number of the intent prediction queue as auxiliary indicators to measure demand response efficiency.

4. The personalized prompt word generation method combining context awareness and interactive feedback as described in claim 1, characterized in that, Based on the aforementioned prompt word optimization instruction set, the underlying configuration parameters of the dual-channel intent prediction model are dynamically adjusted. Prior to this, the method also includes: Before dynamically adjusting the underlying configuration parameters of the dual-channel intent prediction model, record the state snapshot and utility difference data of the dual-channel intent prediction model to identify the key configuration items that cause the decline in the satisfaction score of the prompt words. Based on the key configuration items, compensation adjustment instructions are generated and optimized path breakpoints are inserted.

5. The personalized prompt word generation method combining context awareness and interactive feedback as described in claim 4, characterized in that, The method includes: In multi-device collaborative interaction scenarios, the differences in interaction modalities among various devices are analyzed; Based on the differences in interaction modalities among various devices, intent parsing tasks are dynamically allocated and prompt word parameter configurations are simultaneously optimized across various devices. When the time difference between the prompts and responses of different devices exceeds the time difference threshold, the intent resolution load migration mechanism is activated.

6. The personalized prompt word generation method combining context awareness and interactive feedback as described in claim 5, characterized in that, The method includes: Referring to the dependency graph of the configuration items, configure the prompt influence factor graph, which is used to quantify the weight of each configuration item on the intent matching accuracy and user experience; When a user inputs a modification request, the system predicts the trend of utility changes based on the graph and generates optimization suggestions and alarms, and identifies redundant configuration items through the graph backtracking mechanism. Based on the redundant configuration items, potential demand interference factors are mined using the prompt influence factor map, and the prompt word optimization instruction set is iteratively corrected until the prompt word utility test is passed.

7. The personalized prompt word generation method combining context awareness and interactive feedback as described in claim 6, characterized in that, Based on the aforementioned prompt word optimization instruction set, the underlying configuration parameters of the dual-channel intent prediction model are dynamically adjusted to generate a personalized prompt word candidate set. The method includes: A lightweight simulation interaction environment is embedded to simulate the reasoning behavior of a dynamically adjusted dual-channel intent prediction model. When the simulation results meet the preset utility target, the prompt word configuration changes are batch-retained into the personalized prompt word candidate set. By introducing utility bias data and combining it with the aforementioned prompt influence factor map, the prompt word optimization instruction set is adjusted in a targeted manner.

8. A personalized prompt word generation system combining context awareness and interactive feedback, characterized in that, The system is used to execute the personalized prompt word generation method combining context awareness and interactive feedback as described in any one of claims 1-7, the system comprising: The data acquisition module is used to collect user input data, device operating status data, and environmental context parameters, and to set dynamic sequences of user interactions; The intent association analysis module is used to perform intent association analysis based on the dynamic sequence of user interactions, configure prompt word adjustment vectors, and determine intent recognition deviation by combining user preference database and scenario requirement knowledge base; The instruction set configuration module is used to configure an instruction set optimized by the prompt words based on the prompt word adjustment vector and the intent recognition deviation. The prompt word generation module is used to dynamically adjust the underlying configuration parameters of the dual-channel intent prediction model based on the prompt word optimization instruction set, and generate a personalized prompt word candidate set. The feedback adjustment module is used to simultaneously match and call the personalized prompt word candidate set based on different interaction scenario types, and to dynamically adjust the feedback based on the intent matching accuracy, user feedback response rate, and preset utility goals.

Images