AI Agent-based verbal skill quality inspection method and device, electronic equipment and storage medium

Abstract

The invention relates to the field of power system supply and demand interaction, and discloses an AI Agent-based verbal skill quality inspection method and device, electronic equipment and a storage medium, and the method comprises the following steps: S1, obtaining real-time dialogue data, and mapping the real-time dialogue data to a node on a preset dialogue strategy map; s2, when the conversation state advances to a key decision node; s3, portraying based on the key decision node and the role ability of the current executor; s4, actually selecting a verbal strategy at the key decision node for an executor; and S5, generating a quality inspection result based on the opportunity cost score. According to the method, the user dialogue data are mapped to the preset nodes in real time by adopting a dynamic state tracking technology based on the dialogue strategy graph. Compared with a static data processing scheme in the prior art, the problem that a traditional method cannot respond to user requirements in time is solved, and the flexibility and adaptability of dialogues are improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to AI Agent-based speech quality inspection methods, devices, electronic devices, and storage media. Background Technology

[0002] In existing technologies, many dialogue management systems rely on traditional static data processing methods. This approach typically processes user dialogue information in batches and lacks real-time response capabilities. This results in the system's inability to adapt to changes in user needs in a timely manner, impacting the overall dialogue experience. For example, a user might instantly change their intention during a conversation; if the system cannot react quickly, it may miss a conversion opportunity. From the user's perspective, this sluggish response is clearly a significant drawback.

[0003] Traditional dialogue quality assessments often rely on single metrics, such as success rate or repetition rate. Such dependence on these metrics alone can easily lead to misleading results. Furthermore, a single metric cannot provide a comprehensive analysis of dialogue quality or fully reflect the complex interactions between users and the system. This results in many systems adopting untargeted or even counterproductive measures when optimizing dialogue strategies. Such systems are like the blind men and the elephant, lacking a true understanding of the situation, ultimately leading to very limited and unscientific strategy choices.

[0004] In terms of strategy optimization, existing technologies often lack flexibility, relying primarily on fixed solutions. Many systems simply adopt a single strategy without the ability to dynamically adjust based on real-time feedback. This approach prevents the system from responding promptly to changing user needs. User experience has always been crucial to the success of dialogue systems; systems that cannot respond flexibly severely impact user satisfaction. Summary of the Invention

[0005] The purpose of this invention is to provide a method, device, electronic device and storage medium for dialogue quality inspection based on AI Agent, which solves the problems of untimely response to user needs and insufficient dialogue quality caused by the lack of real-time dynamic tracking, comprehensive evaluation and flexible strategy adjustment in existing dialogue systems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI Agent-based speech quality inspection method, comprising the following steps: S1. Acquire real-time dialogue data and map the real-time dialogue data to nodes on a preset dialogue strategy graph in order to track the transitions of the dialogue state in the dialogue strategy graph. S2. When the dialogue state progresses to a critical decision node, calculate the dialogue entropy of the critical decision node to quantify the uncertainty of the current dialogue state. S3. Based on the role and capability profiles of key decision-making nodes and the current executor, screen candidate dialogue strategies starting from key decision-making nodes; S4. Calculate the opportunity cost score of the actual verbal strategy selected by the executor at key decision nodes, based on the other candidate verbal strategies that were screened and not selected, as well as the dialogue entropy. S5. Generate quality inspection results based on opportunity cost scores.

[0007] Preferably, the step of mapping real-time dialogue data to nodes in a preset dialogue strategy graph in step S1 specifically includes: Speech recognition processing is performed on real-time dialogue data to obtain the dialogue text; Natural language processing is performed on the dialogue text to identify the rhetorical strategies and intentions contained therein; Match the verbal strategy intent with each strategy edge in the dialogue strategy graph that originates from the node where the current dialogue state is located, in order to determine a matching strategy edge. The next node to which the dialogue state will transition is determined based on the matched strategy edge.

[0008] Furthermore, by implementing speech recognition processing, the real-time dialogue content is converted into text format, ensuring accurate information capture. Subsequently, natural language understanding processing is performed on the text to identify potential discourse strategy intentions. After parsing, the identified intentions are matched with the policy edges of the current state in the dialogue policy graph, ensuring that the selected policy edges reflect the actual progress of the dialogue.

[0009] Preferably, the step of calculating the dialogue entropy of the key decision node in step S2 includes: Based on real-time data from the current dialogue, semantic uncertainty score, emotional conflict score, and topic drift score are obtained respectively; the three scores are weighted and summed according to preset weights to obtain the dialogue entropy.

[0010] Furthermore, real-time data of the current dialogue is first collected to calculate the dialogue entropy. This process includes obtaining semantic uncertainty scores, sentiment conflict scores, and topic drift scores by analyzing the dialogue content. The semantic uncertainty score reflects the ambiguity of the intended expression in the dialogue, the sentiment conflict score measures the emotional disharmony present in the dialogue, and the topic drift score judges the stability of the topic. These three scores are then weighted and summed according to preset weights to obtain the overall dialogue entropy.

[0011] Preferably, the step of obtaining the semantic uncertainty score specifically includes: A pre-defined natural language understanding model is used to process real-time dialogue data to obtain intent classification results and corresponding confidence scores. The confidence score is then reverse-normalized to obtain the semantic uncertainty score.

[0012] Furthermore, a pre-defined natural language understanding model is used to process the real-time dialogue in depth. This model can not only classify intents but also attach a confidence score for each classification result, which helps to assess the level of understanding of the dialogue. Next, the semantic uncertainty score is obtained by inverse normalization of the confidence scores.

[0013] Preferably, the step of screening candidate dialogue strategies starting from key decision nodes in step S3 specifically includes: Determine the strategy type corresponding to the candidate verbal strategies starting from key decision nodes; Based on the strategy type, obtain the corresponding ability score from the current executor's role ability profile; The ability score is compared with a preset ability threshold, and candidate speech strategies with an ability score not lower than the ability threshold are selected to pass the screening, so as to form the final set of feasible strategies.

[0014] Furthermore, the system first clarifies the specific type of candidate dialogue strategies and then obtains a significant score based on the current executor's role and ability profile. By comparing this score with preset ability thresholds, the system can filter out candidate dialogue strategies that meet the criteria.

[0015] Preferably, the step of calculating the opportunity cost score of the actually selected verbal strategy in step S4 specifically includes: for the other candidate verbal strategies that have been screened but not selected, calculating their respective potential values ​​based on their preset historical success rates and emotional impact coefficients; The potential value of other candidate verbal strategies is summed to obtain a total potential value. The dialogue entropy is used as a risk discount factor to reduce the total potential value, thereby generating the final opportunity cost score.

[0016] Furthermore, the system first analyzes the unselected candidate dialogue strategies, calculating the potential value of each strategy based on its historical success rate and emotional impact coefficient. The sum of these potential values ​​allows the system to arrive at a total value. Based on this, dialogue entropy is introduced as a risk discount factor to evaluate the actual value of the unselected strategies.

[0017] Preferably, the step of generating the quality inspection result based on the opportunity cost score in step S5 specifically includes: Based on the complete transition path of real-time dialogue data in the dialogue policy graph, a visualized policy path diagram is generated; Based on the dialogue entropy of each key decision node calculated during the dialogue process, a dialogue entropy evolution curve is generated. Key decision nodes with opportunity cost scores higher than a preset threshold are selected, and a comparative analysis report is generated for the selected key decision nodes, including the selected dialogue strategies and other candidate dialogue strategies that were not selected after the selection.

[0018] Furthermore, a complete policy transition path diagram is first generated based on real-time dialogue data, visually displaying the state changes during the dialogue process. Simultaneously, based on the dialogue entropy calculated at each key decision node, a corresponding evolution curve is generated, allowing users to clearly see the dynamic changes in dialogue quality.

[0019] AIAgent-based speech quality inspection device, including: The data mapping module is used to map real-time dialogue data to nodes in the dialogue policy graph; The dialogue entropy calculation module is used to calculate the dialogue entropy of the target node. The strategy filtering module is used to filter the set of feasible strategy edges based on the character's ability profile; The opportunity cost analysis module is used to calculate the opportunity cost score of the actual chosen strategy; The report generation module is used to generate quality inspection results.

[0020] An electronic device includes a processor and a memory, the memory storing a computer program, the processor executing the program to implement a method.

[0021] A storage medium storing computer instructions that, when executed by a processor, implement a method.

[0022] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention employs dynamic state tracking technology based on dialogue strategy graphs to map user dialogue data to preset nodes in real time. This achieves accurate tracking and analysis of dialogue states. Compared to existing static data processing solutions, it solves the problem that traditional methods cannot respond to user needs in a timely manner, thus improving the flexibility and adaptability of dialogue.

[0023] 2. This invention utilizes a multidimensional uncertainty assessment model to calculate dialogue entropy, quantifying the uncertainty of the current dialogue state. This achieves a comprehensive analysis of dialogue quality from multiple dimensions. Compared to the single-index evaluation method in existing technologies, it overcomes the lack of in-depth analysis in traditional methods, making dialogue strategy selection more scientific and reasonable.

[0024] 3. This invention utilizes data mining and machine learning to optimize dialogue strategies, achieving the technical effect of real-time adjustment and improvement of dialogue quality. Compared with the use of a single strategy in existing technologies, this solution overcomes the limitations of a single strategy selection, making the system more flexible and capable of quickly adjusting dialogue strategies based on user feedback, thereby improving user satisfaction. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the dialogue strategy graph of the present invention; Figure 3 This is a schematic diagram of the dialogue entropy calculation model of the present invention; Figure 4 This is a diagram of the feedback database structure of the present invention; Figure 5 This is a structural diagram of the device of the present invention; Figure 6 This is a schematic diagram of the computer device structure of the present invention.

[0026] Among them, 40 is computer equipment; 41 is processor; 42 is memory; and 43 is storage medium. Detailed Implementation

[0027] The following is in conjunction with the appendix Figure 1 -Appendix Figure 4 The present invention will be further described in detail below.

[0028] This invention provides a dialogue quality inspection method based on AI Agent, comprising the following steps: S1. Acquire real-time dialogue data and map the real-time dialogue data to nodes on a preset dialogue strategy graph in order to track the transitions of the dialogue state in the dialogue strategy graph. Specifically, in step S1 of this embodiment, real-time dialogue data is effectively mapped to a preset dialogue strategy graph, thereby realizing real-time tracking and analysis of the dialogue state.

[0029] First, a dialogue strategy graph is constructed. This graph uses a directed graph model and consists of the following elements: Node (S): Each node represents a dialogue state or context.

[0030] Edge (E): An edge connects nodes and marks a possible dialogue transition path. Each edge corresponds to a specific strategy, representing the action that the executor may take at the corresponding node.

[0031] In the implementation process, the audio input received from the user is first processed through an audio processing module. This module uses speech recognition technology, which is based on a deep learning model and can accurately convert speech information into structured dialogue text.

[0032] Next, the obtained dialogue text will be input into the Natural Language Understanding (NLU) module. The main function of this module is to parse customer intent and extract key information. The NLU module employs a range of natural language processing algorithms, including but not limited to intent recognition, sentiment analysis, and entity extraction. For example, the system can use a classifier model to analyze the input text: "I am interested in this product," to identify the intent for further exploration.

[0033] Once the intent in the dialogue is identified, the system performs an edge matching process. Specifically, the system matches the current intent with policy edges in the dialogue policy graph. This matching process utilizes a similarity calculation method, typically employing the cosine similarity formula as follows: Similarity In the formula, A and B represent the vector representations of the current intent graph and the policy edge, respectively; A·B represents the dot product of vectors A and B; and ||A|| and ||B|| represent the magnitudes of vectors A and B, respectively.

[0034] This formula provides a quantitative method to assess the similarity between different intentions and strategy edges, and ultimately select the edge that best fits the current dialogue state.

[0035] Next, state transitions are performed based on the matching results. By selecting a strategy edge that matches the current dialogue state, the node states in the graph can be effectively changed. For example, if a user expresses purchase interest, the transition can move from the needs exploration node to the product recommendation node. This dynamic change ensures the real-time nature and smoothness of the dialogue.

[0036] The entire dialogue path will be tracked, including visited nodes and their corresponding policy edges, to provide detailed data support for subsequent quality inspection and analysis. Recording the dialogue path not only reflects the executor's response strategies but also provides important references for future dialogue optimization.

[0037] In this way, real-time dialogue data is effectively mapped to nodes in the dialogue policy graph, thereby enabling accurate tracking of the dialogue state.

[0038] S2. When the dialogue state progresses to a critical decision node, calculate the dialogue entropy of the critical decision node to quantify the uncertainty of the current dialogue state. Specifically, the goal of step S2 in this embodiment is to calculate the dialogue entropy, which serves as a quantification tool for the uncertainty of the dialogue state. The calculation of dialogue entropy provides an information basis for subsequent strategy selection.

[0039] First, a multidimensional uncertainty assessment model is established, comprising three main components: semantic uncertainty score, emotional conflict score, and topic drift score. Each component provides information from different dimensions for assessing the uncertainty of a conversation.

[0040] In the implementation process, semantic information is first extracted from real-time dialogue data to calculate semantic uncertainty scores. This process is completed by the semantic analysis module. This module uses a trained Natural Language Understanding (NLU) model to process the user's real-time dialogue content, generate intent classification results and related confidence scores, and sets the intent category as C. i The confidence score is denoted as P(C). i ), indicating that the current input belongs to intent category C. i The estimated probability.

[0041] Inverse normalization is performed using the following formula, so that the semantic uncertainty score is defined as: U sem =1-P(C max ): In the formula, U sem P(C) represents semantic uncertainty. max C is the probability of the class with the highest confidence level. max It is the category with the highest confidence level.

[0042] This formula effectively reflects the degree of uncertainty in intent determination. The lower the confidence score, the more ambiguous the user's intent, and the higher the score.

[0043] Calculate the emotional conflict score. This score measures the degree of emotional antagonism between the two parties in the conversation. The sentiment analysis module extracts the emotional polarity of each party, denoted as E. user and E agent E user and E agent The range is from -1 (strongly negative) to +1 (strongly positive). The formula for calculating the emotional conflict score is as follows; U emo =|E user -E agent |: In the formula, U emo E represents the score for emotional conflict. user E represents the user's emotional state. agent It indicates the emotional state of an agent or system.

[0044] By calculating |E user -E agent This allows us to obtain the absolute difference in the emotional states of the two individuals, and thus measure the degree of emotional conflict.

[0045] This formula can effectively reflect the emotional consistency between users and agents; the larger the value, the more obvious the emotional conflict.

[0046] Then, the topic drift score is calculated. Topic drift assesses the degree of deviation between the dialogue content and the core issue of the current decision node. The topic identification module first defines the core issue T for each node. node And calculate the current dialogue content D and T using a text matching algorithm. node The semantic similarity between them. For example: U top =1-similarity(D,T) node ): In the formula, U top The similarity score represents the degree of relevance between topics. node () indicates the content D and the node topic T. node The similarity between them.

[0047] Finally, based on the above three scores, the dialogue entropy H is calculated. Weights w are set. sem w emo and w top These are weighted coefficients for semantic uncertainty, emotional conflict, and topic drift, respectively. The formula for calculating dialogue entropy is: H = w sem ·U sem +w emo ·U emo +w top ·U top : In the formula, H represents the final overall score, and w sem w emo and w top There are three scores, w sem The weighting coefficient w represents the semantic uncertainty. emo w represents the weighting coefficient of emotional conflict. top U represents the weighting coefficient for topic conflict. sem U emo and U top These are scores for semantic uncertainty, sentiment conflict, and topic conflict, which respectively measure the differences between semantics, sentiment, and topic.

[0048] In this formula, the weights of each item can be adjusted according to historical data and business needs, providing flexibility and adaptability to the model.

[0049] Through the above process, the uncertainty of the dialogue state is comprehensively assessed from multiple dimensions, effectively quantifying the dialogue entropy at the current key decision-making node. This step achieves a scientific analysis of the dialogue state, thus providing sufficient basis for subsequent strategy selection.

[0050] S3. Based on the role and capability profiles of key decision-making nodes and the current executor, screen candidate dialogue strategies starting from key decision-making nodes; Specifically, the purpose of step S3 in this embodiment is to start from key decision nodes, screen possible candidate verbal strategies based on the executor's capability profile and strategy type, in order to determine a feasible set of strategies.

[0051] Identify the key decision-making nodes in the current dialogue. These nodes represent the state in the dialogue process where the executor needs to make an important decision. Based on the information flow, the system will identify the specific category of this key node based on the dialogue entropy, intent, and real-time data calculated in the previous step, such as handling price objections or confirming information.

[0052] Extract all candidate dialogue strategies connected to the current key decision node from the dialogue strategy graph. These strategies may correspond to different types, including emotion guidance, information delivery, etc. This process is accomplished by traversing the out-degree edges of the current node in the graph to obtain the strategy information of all connected edges.

[0053] Based on the strategy type of the candidate dialogue strategies, we need to obtain relevant ability scores from the current executor's role ability profile. The ability profile is a quantitative model representing the executor's mastery of different strategy types. Let the strategy type be T. j The corresponding ability score is P. j By understanding the correlation between each strategy type and ability score, we can effectively understand the executor's potential in implementing each strategy.

[0054] Perform strategy screening. Set a preset capability threshold C. threshold The ability score is compared to this threshold. If P j ≥C threshold If so, the candidate speech strategy is considered to have passed the screening and is added to the final set of feasible strategies. The specific filtering criteria formula is as follows; Through screening: P j ≥C threshold : In the formula, P j C represents the score or probability value of strategy j. threshold It is the threshold for strategy selection; only when the strategy's score or probability value P is met... j When this threshold is reached or exceeded.

[0055] This condition ensures that only strategies with ability scores that reach or exceed the threshold are considered, thus avoiding the selection of unsuitable strategies.

[0056] This selection process will ultimately result in a set of feasible strategies that includes all candidate strategies suitable for the current executor's capabilities and the current state of the conversation. This set will provide the executor with clear action guidelines, helping to optimize their subsequent conversational performance and decision-making.

[0057] During this process, the capability profile used can be dynamically updated through historical data, real-time feedback, and training models to ensure that it always reflects the executor's true capability level. Furthermore, the system continuously learns and optimizes the execution effects of various strategies, thereby achieving more efficient and personalized dialogue strategy recommendations.

[0058] S4. Calculate the opportunity cost score of the actual verbal strategy selected by the executor at key decision nodes, based on the other candidate verbal strategies that were screened and not selected, as well as the dialogue entropy. Specifically, in step S4 of this embodiment, dialogue behavior is executed according to the selected candidate dialogue strategy, and the execution effect is monitored and evaluated in real time to optimize customer experience and dialogue quality.

[0059] From the selected set of candidate dialogue strategies, select a high-performance strategy S. k The strategy selection process employs a decision-making algorithm based on effectiveness scoring, which considers multiple factors, including historical performance and user preferences. The mathematical expression of this process is as follows: S selected =argm k ax(E k ): In the formula, S selected E represents the final chosen strategy. k argmax represents the effectiveness score of each candidate policy k. k (E k () indicates the selection score E k The biggest strategy.

[0060] The selection criteria aim to ensure that the chosen strategy has demonstrated a high success rate and user acceptance in past scenarios.

[0061] Next, the selected strategy S selected The content will be transformed into executable dialogue and embedded into the dialogue system. This process is achieved through a Natural Language Generation (NLG) module, which generates corresponding dialogue text based on the policy content.

[0062] Real-time monitoring of user feedback is used to assess conversation quality. This monitoring includes recording users' immediate reactions to strategy implementation, their emotional state, and conversation duration. The emotional state of user feedback is assessed using an emotional scoring model. This model calculates an emotional score E based on the user's response text. user E is defined as the degree to which a user expresses emotion in a conversation. user =f(D user ): In the formula, D user The user inputs content, and f is a sentiment analysis function that can extract potential emotional features from the text.

[0063] To address potential negative emotional feedback and maintain flexibility, conversation strategies or topics can be adjusted in real time. For example, when a user's emotional score falls below a set threshold C... threshold When this happens, the system will proactively switch to a preset emotional soothing strategy to improve user experience satisfaction.

[0064] At the end of the dialogue, a comprehensive evaluation of the overall effectiveness is conducted to optimize future dialogue choices. Evaluation criteria include the effectiveness of the dialogue, user satisfaction, and the duration of effective communication. The overall score is calculated using the following formula: Q = w1·E dialog +w2·E user +w3·D: In the formula, Q represents the overall score of the consumer experience, and E... dialog The score represents the fluency and effectiveness of the dialogue, D represents the duration of the dialogue, and w1, w2 and w3 represent the weighting coefficients of the indicators.

[0065] This comprehensive analysis provides data support for subsequent dialogue optimization, ensuring that the system continuously improves dialogue quality.

[0066] S5. Generate quality inspection results based on opportunity cost scores.

[0067] Specifically, in step S5 of this embodiment, the dialogue strategy is dynamically adjusted and optimized based on existing feedback data and dialogue execution results to improve the quality of subsequent dialogues and user satisfaction.

[0068] Collect and store user feedback data from step S4. This data includes, but is not limited to, real-time user sentiment score E. user Dialogue effectiveness score E dialog And a performance score Q calculated based on various factors. All this feedback information will be integrated into a feedback database D. feedback This database has a hierarchical structure and can store detailed parameters for each dialogue interaction, including timestamps, selected strategies, and user sentiment fluctuations.

[0069] Data mining and machine learning algorithms are used to conduct in-depth analysis of stored historical feedback data to identify key factors influencing dialogue outcomes. This process employs methods such as regression analysis or cluster analysis to reveal the potential relationships between user responses and dialogue effectiveness. To build a predictive model, a strategy type is assumed to be named S. j Its utility prediction function can be expressed as: E predict (S j )=β0+β1·E dialog +β2·E user +β3·D: In the formula, E predict (S j ) represents strategy S j The predictive utility, where β0 represents the intercept term of the model, and β1 represents the dialogue fluency and effectiveness score E. dialog The regression coefficient, β2, represents the user satisfaction score E. user The regression coefficients, β3 represents the regression coefficient of the dialogue duration D, and E dialog E indicates whether the conversation is natural, coherent, and purposeful. user represents the user satisfaction rating, and D represents the duration of the conversation.

[0070] During model training, the system uses historical data for comparison and employs cross-validation to verify the model's accuracy, ensuring its universality across different dialogue scenarios. Mean squared error (MSE) will be used as a key metric to evaluate model performance, quantifying the difference between predicted and true values. In the formula, MSE represents the squared mean of the error between the predicted result and the actual value, n represents the sample size, and E actual,i In the i-th sample, E predict,i This indicates that in the i-th sample...

[0071] Based on the predictions generated by the model, the system will automatically generate strategy optimization suggestions. These suggestions may include improvements, replacements, or combinations of existing script strategies. For example, if strategy S... k Predictive utility E predict (S k (Below the set threshold T) threshold If the system automatically suggests reducing the frequency of using this strategy, or proposes adjustments to the content to better meet user needs.

[0072] By combining the feedback monitoring module with the strategy selection module, a closed-loop feedback mechanism is formed. Specifically, when new feedback data is collected, optimization suggestions are reflected in the dialogue strategy graph in real time. This leads to a dynamic adjustment of the weights used in strategy usage, ensuring that future dialogues tend to favor strategies that perform well.

[0073] The above process ensures that feedback after each conversation is used efficiently, providing data support for future strategy optimization. Through continuous learning and refinement, the system will enhance the flexibility and adaptability of its conversation strategies, improve user satisfaction, and optimize overall conversation quality.

[0074] The AI ​​Agent-based speech quality inspection device described below can be referred to in correspondence with the AI ​​Agent-based speech quality inspection method described above.

[0075] Please see the appendix Figure 5 The present invention also provides an AI Agent-based speech quality inspection device, comprising: The data mapping module is used to map real-time dialogue data to nodes in the dialogue policy graph; The dialogue entropy calculation module is used to calculate the dialogue entropy of the target node. The strategy filtering module is used to filter the set of feasible strategy edges based on the character's ability profile; The opportunity cost analysis module is used to calculate the opportunity cost score of the actual chosen strategy; The report generation module is used to generate quality inspection results.

[0076] The device in this embodiment can be used to execute the above method embodiments, and its principle and technical effects are similar, so they will not be described again here.

[0077] Please see the appendix Figure 6 The present invention also provides a computer device 40, including a processor 41 and a memory 42, wherein the memory 42 stores a computer program executable by the processor, and the computer program performs the method described above when executed by the processor.

[0078] The present invention also provides a storage medium storing a computer program, which is executed by a processor to perform the method described above.

[0079] The storage medium 43 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dialogue quality inspection method based on AI Agent, characterized in that, Includes the following steps: S1. Acquire real-time dialogue data and map the real-time dialogue data to nodes on a preset dialogue strategy graph to track the transitions of the dialogue state in the dialogue strategy graph. S2. When the dialogue state progresses to a critical decision node, calculate the dialogue entropy of the critical decision node to quantify the uncertainty of the current dialogue state. S3. Based on the role and capability profiles of key decision-making nodes and the current executor, screen candidate dialogue strategies starting from key decision-making nodes; S4. Calculate the opportunity cost score of the actual verbal strategy selected by the executor at key decision nodes, based on the other candidate verbal strategies that were screened and not selected, as well as the dialogue entropy. S5. Generate quality inspection results based on opportunity cost scores.

2. The AI ​​Agent-based speech quality inspection method according to claim 1, characterized in that, The step of mapping real-time dialogue data to nodes in a preset dialogue strategy graph in step S1 specifically includes: Speech recognition processing is performed on real-time dialogue data to obtain the dialogue text; Natural language processing is performed on the dialogue text to identify the rhetorical strategies and intentions contained therein; Match the verbal strategy intent with each strategy edge in the dialogue strategy graph that originates from the node where the current dialogue state is located, in order to determine a matching strategy edge. The next node to which the dialogue state will transition is determined based on the matched strategy edge.

3. The AI ​​Agent-based speech quality inspection method according to claim 1, characterized in that, The step of calculating the dialogue entropy of the key decision node in step S2 includes: Based on real-time data from the current dialogue, semantic uncertainty score, sentiment conflict score, and topic drift score are obtained respectively. The three scores are weighted and summed according to preset weights to obtain the dialogue entropy.

4. The AI ​​Agent-based speech quality inspection method according to claim 3, characterized in that, The steps for obtaining the semantic uncertainty score specifically include: A pre-defined natural language understanding model is used to process real-time dialogue data to obtain intent classification results and corresponding confidence scores. The confidence score is then reverse-normalized to obtain the semantic uncertainty score.

5. The AI ​​Agent-based speech quality inspection method according to claim 1, characterized in that, The step of screening candidate speech strategies starting from key decision nodes in step S3 specifically includes: Determine the strategy type corresponding to the candidate verbal strategies starting from key decision nodes; Based on the strategy type, obtain the corresponding ability score from the current executor's role ability profile; The ability score is compared with a preset ability threshold, and candidate speech strategies with an ability score not lower than the ability threshold are selected to pass the screening, so as to form the final set of feasible strategies.

6. The AI ​​Agent-based dialogue quality inspection method according to claim 5, characterized in that, The step of calculating the opportunity cost score of the actual selected rhetorical strategy in step S4 specifically includes: For the other candidate verbal strategies that were screened but not selected, their potential value was calculated based on their preset historical success rate and emotional impact coefficient. The potential value of other candidate verbal strategies is summed to obtain a total potential value. The dialogue entropy is used as a risk discount factor to reduce the total potential value, thereby generating the final opportunity cost score.

7. The AI ​​Agent-based speech quality inspection method according to claim 1, characterized in that, The step of generating quality inspection results based on opportunity cost scores in step S5 specifically includes: Based on the complete transition path of real-time dialogue data in the dialogue policy graph, a visualized policy path diagram is generated; Based on the dialogue entropy of each key decision node calculated during the dialogue process, a dialogue entropy evolution curve is generated. Key decision nodes with opportunity cost scores higher than a preset threshold are selected, and a comparative analysis report is generated for the selected key decision nodes, including the selected dialogue strategies and other candidate dialogue strategies that were not selected after the selection.

8. An AI Agent-based speech quality inspection device, applied to the AI ​​Agent-based speech quality inspection method according to any one of claims 1-7, characterized in that, include; The data mapping module is used to map real-time dialogue data to nodes in the dialogue policy graph; The dialogue entropy calculation module is used to calculate the dialogue entropy of the target node. The strategy filtering module is used to filter the set of feasible strategy edges based on the character's ability profile; The opportunity cost analysis module is used to calculate the opportunity cost score of the actual chosen strategy; The report generation module is used to generate quality inspection results.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the program to implement the method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The system stores computer instructions that, when executed by a processor, implement the method as described in any one of claims 1-7.

Images