Corporate training dialogue coaching method and system

Abstract

The application provides a kind of enterprise training dialogue accompanying training method and system, it is related to enterprise training technical field, this method includes: receiving the scene description information of user input;Scene description information is parsed based on natural language processing technology, and accompanying training configuration information is generated in combination with knowledge retrieval;According to accompanying training configuration information and the accompanying training mode selected by user, start the dialogue accompanying training process with student;After dialogue accompanying training, the key event that matches the scoring dimension in accompanying training configuration information is identified from dialogue text, and the corresponding original dialogue segment is bound as evidence chain, so as to generate the interpretable accompanying training report for evaluating student performance in combination with target achievement degree;Accompanying training report is synchronized to management end workstation, for training progress tracking and management decision-making.

Description

Technical Field

[0001] This invention relates to the field of corporate training technology, and in particular to a method and system for corporate training dialogue practice. Background Technology

[0002] Current corporate training primarily relies on human instructors, standardized courseware, or traditional online exams. While these methods can impart basic knowledge and processes, the training effectiveness often depends on the instructor's skill level, lacks personalization and practical simulations, and makes it difficult to effectively assess trainees' adaptability and comprehensive abilities.

[0003] In recent years, some companies have begun to try introducing AI-powered dialogue practice tools, but most existing technologies adopt a fixed script model or rule-based question-and-answer robots. Fixed script practice requires administrators to manually fill in scene, role, and dialogue information, which is complex and inefficient. Furthermore, during the dialogue, it can only compare the preset dialogue line by line, making it unable to flexibly cope with complex and uncertain real-world scenarios.

[0004] Furthermore, existing dialogue training technologies generally use the accuracy of surface-level rhetoric as an evaluation standard, lacking intelligent judgment on whether trainees have truly achieved the training objectives, and making it difficult to comprehensively measure trainees' adaptability and task completion in real-world situations. At the same time, limited support for multiple languages, dialects, and knowledge across different job roles results in insufficient training coverage and practicality. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for corporate training dialogue practice, so as to alleviate the problems of complex setup, rigid dialogue, and unscientific evaluation in existing corporate training dialogue practice technologies.

[0006] Firstly, the present invention provides a method for corporate training dialogue practice, comprising: Receive scene description information input by the user; The scenario description information is parsed using natural language processing technology and combined with knowledge retrieval to generate coaching configuration information, which includes coaching scenario, dialogue role, candidate dialogue and scoring dimensions. Based on the coaching configuration information and the coaching mode selected by the user, the dialogue coaching process with the student is initiated; among them, the coaching modes include a free dialogue mode based on goals and round constraints, or a fixed script mode based on sentence-by-sentence comparison of standard dialogues. After the dialogue practice ends, key events that match the scoring dimensions in the practice configuration information are identified from the dialogue text, and the corresponding original dialogue fragments are bound as evidence chains. In this way, an interpretable practice report that evaluates the student's performance is generated by combining the goal achievement. The training reports are synchronized to the management console for tracking training progress and making management decisions.

[0007] In an optional implementation, the scene description information is parsed using natural language processing technology, and combined with knowledge retrieval to generate coaching configuration information, including: The scene description information is segmented, dependency parsing and semantic parsing are performed to obtain preliminary structured information, which includes scene keywords, role candidates and target intent; Based on the target intent, the system calls the pre-set enterprise knowledge base and extracts the script template by searching relevant business processes, frequently asked questions and answers and business cases. Based on the scene keywords, the scene description information is refined using a large language model to generate a complete training scenario; Based on the role candidates, the dialogue roles are identified from the scene description information through named entity recognition and few-shot classification algorithms; Based on the dialogue template, candidate dialogues for the dialogue characters are generated by using retrieval-enhanced generation technology, combined with knowledge base content and generation models. Based on the target intent and the pre-set template rules, the training objectives are broken down into multi-level scoring dimensions, and scoring points are defined for each scoring dimension. The scoring dimensions are arranged from high to low as follows: competency dimension, key task, specific action, and standard script.

[0008] In an optional implementation, the execution process of the free dialogue mode includes: Preset dialogue goals and maximum number of dialogue rounds; During the dialogue, a soft finite state machine model is used to track the dialogue state. The state of the soft finite state machine model includes the current round, the goal completion state, the filling of the slot, and the yaw risk indicator. The system judges the speaker's intention in real time. If the confidence level between the speaker's intention and the dialogue goal is lower than the preset first threshold, it is judged as a deviation and a fallback strategy is triggered to guide the dialogue back to the main path. If the dialogue objective is achieved before reaching the maximum number of dialogue rounds, the training session is considered successful; if the dialogue objective is not achieved even after reaching the maximum number of dialogue rounds, the training session is considered unsuccessful.

[0009] In an optional implementation, the execution process of the fixed script mode includes: Load the preset standard dialogue sequence; Receive trainees' responses line by line; Calculate the semantic vector similarity between the student's response and the corresponding standard response. If the similarity exceeds the preset second threshold, it is considered a match; otherwise, it is marked as a deviation. The overall accuracy of students' answers is calculated based on the sentence-by-sentence matching results.

[0010] In an optional implementation, key events matching the scoring dimensions in the coaching configuration information are identified from the dialogue text, and the corresponding original dialogue fragments are bound as a chain of evidence. This, combined with the goal achievement, generates an interpretable coaching report evaluating the trainee's performance, including: The event extraction model is invoked to segment and analyze the dialogue text, identifying multiple key events; Establish a mapping relationship between key events and scoring dimensions; For each scoring dimension, key events are matched with original dialogue fragments to form a chain of evidence. Based on the chain of evidence, a weighted scoring model is used to calculate the scores for each scoring dimension and the weighted total score. A training report is generated based on the chain of evidence, scores for each scoring dimension, weighted total score, and the degree of goal achievement corresponding to the dialogue text.

[0011] In optional implementations, the corporate training dialogue practice method also includes: During the dialogue practice, a language recognition device determines the language type of the student's input speech in real time. Based on the identified target language type, the system dynamically switches to the corresponding speech recognition model for speech-to-text conversion, and uses shallow fusion technology to inject the enterprise's proprietary terminology hot word list into the speech recognition decoding process. To generate speech output by binding specific voices or digital human figures to dialogue characters.

[0012] In an optional implementation, the management console performs the following functions: The system aggregates and displays training reports from multiple trainees, including goal achievement rate, average score, and dimensional distribution statistics. Based on the rules engine, reminder notifications are automatically triggered for students who have not completed their tutoring tasks for a long time or have consistently received low scores; It provides a modular interface for training plan development, task approval, and trainee progress tracking.

[0013] Secondly, the present invention provides an enterprise training dialogue practice system, comprising: The information receiving module is used to receive scene description information input by the user; The information parsing module is used to parse the scene description information based on natural language processing technology and generate coaching configuration information by combining knowledge retrieval. The coaching configuration information includes coaching scene, dialogue role, candidate dialogue and scoring dimensions. The dialogue practice module is used to initiate the dialogue practice process with the student based on the practice configuration information and the practice mode selected by the user. The practice modes include a free dialogue mode based on goals and round constraints, or a fixed script mode based on sentence-by-sentence comparison of standard dialogue. The coaching evaluation module is used to identify key events from the dialogue text that match the scoring dimensions in the coaching configuration information after the dialogue coaching is completed, and bind the corresponding original dialogue fragments as evidence chains, thereby generating an interpretable coaching report that evaluates the trainee's performance. The report synchronization module is used to synchronize training reports to the management workbench for training progress tracking and management decision-making.

[0014] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the enterprise training dialogue practice method of any of the foregoing embodiments.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when run by a processor, executes the enterprise training dialogue practice method described in any of the foregoing embodiments.

[0016] The enterprise training dialogue practice method and system provided by this invention includes the following steps: receiving scenario description information input by a user; parsing the scenario description information based on natural language processing technology and generating practice configuration information by combining knowledge retrieval, the practice configuration information including practice scenario, dialogue roles, candidate dialogues, and scoring dimensions; initiating a dialogue practice process with the trainee according to the practice configuration information and the practice mode selected by the user; wherein, the practice mode includes a free dialogue mode based on goal and round constraints, or a fixed script mode based on sentence-by-sentence comparison of standard dialogues; after the dialogue practice ends, identifying key events from the dialogue text that match the scoring dimensions in the practice configuration information, and binding the corresponding original dialogue fragments as evidence chains, thereby generating an interpretable practice report evaluating the trainee's performance based on goal achievement; and synchronizing the practice report to the management console for training progress tracking and management decision-making. This automatically constructs practice scenarios, roles, and dialogues, supports flexible and free dialogue, and performs intelligent evaluation based on goal achievement and evidence chains, improving the efficiency and effectiveness of training. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a corporate training dialogue practice method provided in an embodiment of the present invention; Figure 2This is a schematic diagram of the structure of a corporate training dialogue coaching system provided in an embodiment of the present invention; Figure 3 A system architecture diagram provided for an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a method and system for corporate training dialogue practice, which can automatically construct practice scenarios, roles and scripts, support flexible and free dialogue, and perform intelligent evaluation based on goal achievement, thereby improving the efficiency and effectiveness of training.

[0021] The embodiments of this invention aim to solve the problems of complex setup, rigid dialogue, and unscientific evaluation in existing enterprise training coaching systems. The goal is: It provides a low-barrier, high-efficiency method for configuring training partners, reducing the difficulty for administrators to set up. It provides a goal-oriented dialogue mode that can dynamically simulate complex situations, enhancing realism and flexibility; It provides an intelligent evaluation mechanism based on goal achievement and evidence chain, making the scoring more scientific and interpretable; It provides a multilingual, voice-binding-enabled voice interaction method to enhance the system's coverage and immersive experience; It provides a management-side insight and reminder mechanism to achieve closed-loop management of the entire training process.

[0022] To achieve the above objectives, this invention proposes a solution for one-click creation and goal achievement assessment of dialogue practice for enterprise training, which mainly includes the following parts: 1. Create a coaching setup with one click.

[0023] The user enters a description of the scenario; The system automatically generates scenarios, roles, and candidate dialogues by combining natural language parsing and knowledge retrieval with a generative model; Multi-level scoring dimensions (ability → task → action → communication) are generated based on templated rules.

[0024] 2. Dual-mode dialogue coaching.

[0025] Free Dialogue Mode: Employs objective and round constraints, combined with a Soft Finite State Machine (Soft-FSM), supports yaw and return to the main path, and is used to examine adaptability; Fixed script mode: Load standard scripts, compare trainees' answers sentence by sentence, and use semantic similarity to determine consistency, which is suitable for standardized training.

[0026] 3. Intelligent assessment and evidence chain generation.

[0027] Use event extraction models to identify key events in a dialogue; Match events with preset scoring points to form a target-dimension-event mapping; By binding the original dialogue fragments as a chain of evidence, a weighted model is used to generate interpretable scoring results.

[0028] 4. Multi-voice recognition and timbre binding.

[0029] The system automatically identifies the input language (such as Mandarin, Cantonese, or English) through a language recognizer and calls the corresponding ASR (Automatic Speech Recognition) model. By integrating a list of enterprise terminology hot words through Shallow Fusion technology, the recognition accuracy in professional scenarios can be improved. It supports binding characters with specific voices or digital human figures to enhance immersion.

[0030] 5. Training reports and management workbench.

[0031] The system automatically generates a training report, which includes the goal completion status, dimension scores, evidence chain fragments, and improvement suggestions; The report is synchronized to the management console, supporting task approval, trainee progress tracking, training plans, and data insights; Based on the rule engine and anomaly detection module, it automatically reminds students who have not completed tasks for a long time or have low scores, thus achieving closed-loop management.

[0032] The embodiments of the present invention have the following beneficial effects: Significantly lowers the barrier to entry: automates the entire process from natural language processing to coaching setup; Enhance the realism of training: The free dialogue mode combined with the soft FSM closely resembles real business scenarios; To ensure the scientific rigor of the assessment: By linking event extraction with the chain of evidence, the scoring results are verifiable and interpretable; Expanding system applicability: Supports multiple languages, multiple voices, and enterprise-specific terminology; Achieve closed-loop training management: The management platform has reminder and data insight functions to ensure training effectiveness.

[0033] To facilitate understanding of this embodiment, a detailed description of a corporate training dialogue coaching method disclosed in this embodiment of the invention will be provided first.

[0034] This invention provides a method for enterprise training dialogue practice, which can be executed by an electronic device with data processing capabilities. See also... Figure 1 The diagram shows a flowchart of a corporate training dialogue practice method, which mainly includes the following steps S110 to S150: Step S110: Receive scene description information input by the user.

[0035] Users can enter a short description in the coaching management interface, such as "simulating customer inquiries about after-sales issues".

[0036] Step S120: The scene description information is parsed based on natural language processing technology, and the training configuration information is generated by combining knowledge retrieval. The training configuration information includes the training scene, dialogue role, candidate dialogue and scoring dimension.

[0037] The training scenario, dialogue characters, and target intent can be extracted using LLM (Large Language Model), and candidate dialogues can be generated by calling the knowledge base RAG (Retrieval-Augmented Generation), as well as the scoring dimensions can be generated in a templated manner.

[0038] In some possible embodiments, step S120 above may include: performing word segmentation, dependency parsing, and semantic analysis on the scene description information to obtain preliminary structured information, which includes scene keywords, role candidates, and target intent; calling a preset enterprise knowledge base based on the target intent, and extracting dialogue templates by retrieving relevant business processes, frequently asked questions, and business cases; refining the scene description information using a large language model based on the scene keywords to generate a complete training scenario; identifying dialogue roles from the scene description information based on the role candidates using named entity recognition and few-shot classification algorithms; generating candidate dialogues for the dialogue roles based on the dialogue templates using retrieval enhancement generation technology, combined with knowledge base content and generation models; and decomposing the training objective into multi-level scoring dimensions based on the target intent and preset template rules, and defining scoring points for each scoring dimension; wherein, the scoring dimensions are arranged from high to low as ability dimensions, key tasks, specific actions, and standard dialogues.

[0039] The aforementioned capability dimensions are the top-level, macro-level qualities or skills. In customer after-sales consultation scenarios, capability dimensions can include information acquisition ability, problem-solving ability, and customer reassurance ability. Key tasks are the core work steps or milestones that must be completed to achieve the next higher-level capability dimension. For example, the key tasks for information acquisition ability could be clarifying the customer's problem and obtaining necessary information (such as product model, fault details, etc.). Specific actions are the concrete behaviors demonstrated in completing key tasks. For example, corresponding to the key task of "clarifying the customer's problem," specific actions could include asking clarifying questions and restating and confirming the customer's problem. Standard scripts are the bottom-level, concrete language expressions that achieve the actions. For example, corresponding to the action of "asking clarifying questions," standard scripts could include "Could you please describe in detail what problem the phone is experiencing?" or "When you say 'cannot turn on,' do you mean it's completely unresponsive, or does the screen light up briefly and then go out?" The above scoring points are evaluation criteria applied to key tasks to measure their completion quality.

[0040] In specific implementation, taking the user input "simulating customer inquiries about after-sales issues" as an example, the input and parsing process begins as follows: 1. The system calls the natural language processing module to perform word segmentation, dependency parsing, and semantic analysis on the input text, obtaining preliminary structured information, including: 1) Scenario keywords: after-sales, customer inquiry; 2) Role candidates: customer, customer service; 3) Target intent: problem answering, after-sales processing. 2. Based on the semantic parsing results, the system calls the knowledge base to retrieve relevant SOPs (Standard Operating Procedures), FAQs (Frequently Asked Questions), and business cases, extracting referable script templates.

[0041] Then, a one-click setup for training is implemented: 1. Scenario Refinement: The Large Language Model (LLM) is used to rewrite user input, generating a more complete scenario description, such as "At the after-sales service center, a customer inquires about repair solutions due to a mobile phone malfunction." 2. Role Recognition: Named Entity Recognition (NER) and Few-shot classification algorithms are used to identify two typical dialogue roles in the scenario: customer and customer service representative. 3. Script Generation: The system retrieves knowledge base content using RAG and then generates candidate scripts, such as standard greetings from customer service representatives, problem clarification scripts, and solution expressions. 4. Scoring Dimension Generation: The system breaks down training objectives into competency dimensions, such as "information acquisition," "problem solving," and "customer reassurance," and defines key tasks and scoring points for each dimension.

[0042] Step S130: Based on the training configuration information and the training mode selected by the user, start the dialogue training process with the student; wherein, the training mode includes a free dialogue mode based on goal and round constraints, or a fixed script mode based on standard dialogue word by word comparison.

[0043] This embodiment provides two coaching modes for users to choose from flexibly. The coaching configuration information and the user-selected coaching mode can be loaded to initiate the coaching dialogue process with the student. Coaching modes include free dialogue mode and fixed script mode. Users can select either "free dialogue mode" or "fixed script mode" in the coaching management interface. If the user selects free dialogue mode, the system requires the user to set the dialogue objective (such as "clarifying customer needs and providing solutions") and the maximum number of rounds (such as 8 rounds). If the user selects fixed script mode, the system will load the preset SOP (Standard Operating Procedure) flow corresponding to the coaching configuration information, such as "greeting → understanding the problem → providing solutions → closing remarks".

[0044] During dialogue training, trainees need to log in to the trainee client and enter the practice dialogue interface. In free mode: the AI ​​plays the role of the customer, designing dynamic questions and feedback based on the dialogue objectives. Each round, the system judges whether the trainee has completed the target sub-task (e.g., whether they asked clarifying questions, whether they provided solutions). In fixed mode: the AI ​​plays the customer dialogue sentence by sentence, and the trainee answers sentence by sentence according to the SOP. The system compares each sentence with the standard script for consistency.

[0045] In some possible embodiments, the execution process of the above-mentioned free dialogue mode may include: setting a preset dialogue goal and a maximum number of dialogue rounds; during the dialogue, a soft finite state machine model is used to track the dialogue state, the state of the soft finite state machine model includes the current round, goal completion status, slot filling, and yaw risk flag; real-time judgment of the student's speaking intention, if the confidence of the speaking intention and the dialogue goal is lower than a preset first threshold, it is judged as yaw and a fallback strategy is triggered to guide the dialogue back to the main path; if the dialogue goal is completed before reaching the maximum number of dialogue rounds, the training is judged as successful; if the dialogue goal is not completed even after reaching the maximum number of rounds, the training is judged as failed. The dialogue goal may include multiple goals.

[0046] In some possible embodiments, the execution process of the above fixed script mode may include: loading a preset standard dialogue sequence; receiving the student's response dialogue line by line; calculating the semantic vector similarity between the student's response dialogue and the corresponding standard dialogue dialogue, and determining a match when the similarity exceeds a preset second threshold, otherwise marking it as a deviation; and calculating the overall accuracy of the student's response based on the line-by-line matching results.

[0047] The first and second thresholds mentioned above can be set according to actual needs, and no restrictions are imposed here.

[0048] Optionally, this embodiment also supports multiple languages, multiple voices, and enterprise-specific terminology. Based on this, the above-mentioned enterprise training dialogue practice method further includes: during the dialogue practice, using a language recognizer to determine the language type of the student's input speech in real time; dynamically switching to the corresponding speech recognition model to convert speech to text based on the identified target language type, and using shallow fusion technology to inject a list of enterprise-specific terminology hot words into the speech recognition decoding process; and binding specific voices or digital human images to the dialogue characters for speech synthesis output.

[0049] Step S140: After the dialogue training ends, identify key events from the dialogue text that match the scoring dimensions in the training configuration information, and bind the corresponding original dialogue fragments as evidence chains, thereby generating an interpretable training report that evaluates the trainee's performance in combination with the goal achievement.

[0050] This embodiment can also perform intelligent evaluation of trainee performance based on the evidence chain and goal achievement, where goal achievement refers to the completion of the dialogue goal. In some possible embodiments, step S140 above may include: calling an event extraction model to perform segmented analysis on the dialogue text and identify multiple key events; establishing a mapping relationship between key events and scoring dimensions; binding the original dialogue fragments to the key events matched for each scoring dimension to form an evidence chain; calculating the score and weighted total score for each scoring dimension using a weighted scoring model based on the evidence chain; and generating a coaching report based on the evidence chain, the scores of each scoring dimension, the weighted total score, and the goal achievement corresponding to the dialogue text.

[0051] In practice, after the dialogue ends, the system invokes an event extraction model to segment and analyze the dialogue text, identifying events such as "obtaining customer needs," "proposing solutions," and "confirming transactions." The system matches the extracted events with scoring points, establishing a target-dimensional-event mapping. The system binds each scoring point to an original dialogue fragment, forming a chain of evidence. The final scoring uses a weighted model: score for a scoring dimension = number of completed events under that scoring dimension / total number of events under that scoring dimension; weighted total score = Σ(weight of each scoring dimension × score of the corresponding scoring dimension) - penalty points for prohibited phrases. Penalties for prohibited phrases (such as "shifting responsibility") are deducted. The training report generated by the system can include: the completion status of each target, the scores and weighted total score of each evaluation dimension, corresponding dialogue fragments as evidence, and system-generated improvement suggestions (such as "further inquiry into customer details is needed").

[0052] Step S150: Synchronize the training report to the management console for training progress tracking and management decision-making.

[0053] The training reports can be automatically synchronized to the management console, where administrators can view training progress, remind trainees to redo their work, and generate annual training statistics.

[0054] In some possible embodiments, the aforementioned management console can perform the following functions: aggregate and display training reports for multiple trainees, including target achievement rate, average score, and dimensional distribution statistics; automatically trigger reminder notifications for trainees who have not completed training tasks for a long time or have consistently received low scores, based on a rule engine; and provide a componentized interface for training plan creation, task approval, and trainee progress tracking.

[0055] The enterprise training dialogue practice method provided in this invention can automatically construct practice scenarios, roles, and scripts, support flexible and free dialogue, and perform intelligent evaluation based on goal achievement and evidence chain, thereby improving the efficiency and effectiveness of training.

[0056] Corresponding to the above-described enterprise training dialogue practice method, this invention also provides an enterprise training dialogue practice system. See [link to related document]. Figure 2 The diagram shown illustrates the structure of a corporate training dialogue practice system. The system includes: The information receiving module 201 is used to receive scene description information input by the user; The information parsing module 202 is used to parse the scene description information based on natural language processing technology and generate coaching configuration information by combining knowledge retrieval. The coaching configuration information includes coaching scene, dialogue role, candidate dialogue and scoring dimension. The dialogue practice module 203 is used to initiate the dialogue practice process with the student based on the practice configuration information and the practice mode selected by the user. The practice modes include a free dialogue mode based on goals and round constraints, or a fixed script mode based on sentence-by-sentence comparison of standard dialogue. The training evaluation module 204 is used to identify key events in the dialogue text that match the scoring dimensions in the training configuration information after the dialogue training is completed, and bind the corresponding original dialogue fragments as evidence chains to generate an interpretable training report that evaluates the trainee's performance. The report synchronization module 205 is used to synchronize training reports to the management workbench for training progress tracking and management decision-making.

[0057] The enterprise training dialogue practice system provided in this embodiment of the invention can automatically construct practice scenarios, roles and scripts, support flexible and free dialogue, and perform intelligent evaluation based on goal achievement and evidence chain, thereby improving the efficiency and effectiveness of training.

[0058] Furthermore, the aforementioned information parsing module 202 is specifically used for: performing word segmentation, dependency parsing, and semantic parsing on the scene description information to obtain preliminary structured information, which includes scene keywords, role candidates, and target intent; based on the target intent, calling a preset enterprise knowledge base, and extracting dialogue templates by retrieving relevant business processes, frequently asked questions, and business cases; refining the scene description information using a large language model based on the scene keywords to generate a complete training scenario; identifying dialogue roles from the scene description information based on the role candidates using named entity recognition and few-shot classification algorithms; generating candidate dialogues for the dialogue roles based on the dialogue templates using retrieval enhancement generation technology, combined with knowledge base content and generation models; and decomposing the training objective into multi-level scoring dimensions based on the target intent and preset template rules, and defining scoring points for each scoring dimension; wherein, the levels of the scoring dimensions from high to low are: capability dimension, key tasks, specific actions, and standard dialogues.

[0059] Furthermore, the execution process of the aforementioned free dialogue mode includes: setting a preset dialogue goal and a maximum number of dialogue rounds; during the dialogue, a soft finite state machine model is used to track the dialogue state, and the state of the soft finite state machine model includes the current round, goal completion status, slot filling, and yaw risk indicator; real-time judgment of the trainee's speaking intention, if the confidence of the speaking intention and the dialogue goal is lower than a preset first threshold, it is judged as yaw and a fallback strategy is triggered to guide the dialogue back to the main path; if the dialogue goal is completed before reaching the maximum number of dialogue rounds, the training is judged as successful; if the dialogue goal is not completed even after reaching the maximum number of rounds, the training is judged as failed.

[0060] Furthermore, the execution process of the above-mentioned fixed script mode includes: loading a preset standard dialogue sequence; receiving the students' response dialogues sentence by sentence; calculating the semantic vector similarity between the students' response dialogues and the corresponding standard dialogues, and determining a match when the similarity exceeds a preset second threshold, otherwise marking it as a deviation; and calculating the overall accuracy of the students' answers based on the sentence-by-sentence matching results.

[0061] Furthermore, the aforementioned training evaluation module 204 is specifically used for: calling the event extraction model to perform segmented analysis of the dialogue text and identify multiple key events; establishing a mapping relationship between key events and scoring dimensions; binding the original dialogue fragments to the key events matched for each scoring dimension to form an evidence chain; calculating the score and weighted total score for each scoring dimension using a weighted scoring model based on the evidence chain; and generating a training report based on the evidence chain, the scores of each scoring dimension, the weighted total score, and the target achievement degree corresponding to the dialogue text.

[0062] Furthermore, the aforementioned dialogue practice module 203 is also used to: determine the language type of the student's input speech in real time through a language recognizer during the dialogue practice process; dynamically switch to the corresponding speech recognition model to convert speech to text based on the identified target language type, and inject the enterprise's exclusive terminology hot word list into the speech recognition decoding process using shallow fusion technology; and bind a specific timbre or digital human image to the dialogue character for speech synthesis output.

[0063] Furthermore, the aforementioned management console implements the following functions: aggregates and displays training reports from multiple trainees, including target achievement rate, average score, and dimensional distribution statistics; automatically triggers reminder notifications for trainees who have not completed training tasks for a long time or have consistently received low scores, based on a rules engine; and provides a componentized interface for training plan creation, task approval, and trainee progress tracking.

[0064] The corporate training dialogue coaching system provided in this embodiment has the same implementation principle and technical effects as the aforementioned corporate training dialogue coaching method embodiment. For the sake of brevity, any parts not mentioned in the corporate training dialogue coaching system embodiment can be referred to the corresponding content in the aforementioned corporate training dialogue coaching method embodiment.

[0065] See Figure 3 The system architecture diagram shown in this embodiment of the invention includes a user layer 301, a service layer 302, an algorithm layer 303, a data layer 304, and a management layer 305. The student and administrator ends connect to the service layer, which calls the algorithm layer to complete dialogue practice and assessment. The data layer stores knowledge and logs, and the management layer provides a workbench and insight functions. These five layers are progressively invoked in the main workflow of the enterprise training dialogue practice system. The main workflow may include: user inputs a sentence → one-click service parsing → mode selection (free / fixed) → dialogue simulation → intelligent assessment → report output → management workbench.

[0066] The key aspects of the embodiments of the present invention are described below.

[0067] I. Generation of sparring partner configuration information: Input a one-sentence description D; extract {Scene S, Role Candidate Rc, Target Gc} using LLM; generate the dialogue H by calling the knowledge base RAG; generate the scoring dimension M using templates; output P={S,R,H,M}.

[0068] II. Comparison Algorithm under Fixed Script Mode: 1. The system loads the standard dialogue sequence. H = h 1, h 2,···, h n ; 2. Students answer sentence by sentence.U = u 1, u 2,···, u m ; 3. The system uses the following sentence vector similarity calculation formula to calculate the similarity between the standard script and the student's response: ; When similarity sim If the value is ≥0.8, it is considered a match; otherwise, it is marked as a deviation. 4. The system records the matching results sentence by sentence and calculates the overall accuracy.

[0069] III. Goal and Round Control in Free Dialogue Mode: 1. During configuration, the user sets the objective G={information gathering, requirements clarification, solution provision}, with a maximum number of rounds N=10; 2. In actual dialogue, the system updates goal_state in real time: if "information collection" is completed, then goal_state[1]=true; if the number of rounds > 10 and not all goals are completed, then it is judged as failure.

[0070] 3. Yaw Handling: If the confidence level of the intention is less than the threshold θ=0.6, the system triggers a fallback strategy, such as "customer repeats the same question", to guide the trainee back to the topic.

[0071] A soft FSM is introduced to support yaw-return, ensuring the objective is achieved within a limited number of rounds. The implementation method can be: State representation: S_t={turn,goal_state,slot,risk_flag}; Transfer rule: Input u_t → Classify intent_t → Update goal_state → If turn > max_turn, the goal has not been achieved → Determine failure; A fallback strategy is triggered when the plane yaws.

[0072] IV. Generation of Evidence Chains for Intelligent Assessment: 1. Extract events from the dialogue text T to obtain the event set E = {e1: asking about needs, e2: explaining the solution, e3: confirming the result}.

[0073] 2. Match the event with the rating points M = {information gathering, problem solving, customer reassurance}.

[0074] 3. Generate a chain of evidence, for example: Scoring point "Information gathering" ← Event e1 ← Dialogue excerpt [Round 2 student speech].

[0075] 4. The final evaluation report states that "the trainee raised clarification questions in the second round, thus meeting the information gathering objective."

[0076] V. Multilingual and Voice Support: 1. When a student selects Cantonese as the input language in the practice dialogue interface, the system calls LID (Language Identification) to determine that the input language is Cantonese.

[0077] 2. The system switches to the Cantonese ASR model and uses Shallow Fusion to inject corporate terms (such as "maintenance order" and "inspection code") into WFST (Weighted Finite State Transducer) to improve the recognition rate.

[0078] 3. The system assigns a specific Cantonese accent to the "customer" role and corresponds it to the digital human image to ensure multimodal consistency.

[0079] VI. Task Insights from the Management Console: 1. Administrators can log in to the management console and view the following on the workbench: To be done: exam paper grading, training approval, student reminders; Training plan: Annual goals and current progress; Student progress: The percentage of students who have completed the task and those who have not.

[0080] 2. The system automatically triggers reminders based on a rules engine, such as "If a student does not complete the tutoring within three days → push a reminder to the administrator and the student".

[0081] 3. The system supports anomaly detection: it identifies students who score low three times in a row and automatically marks them as high-risk.

[0082] The following section introduces the overall architecture, data flow and core processes, and innovative design of the aforementioned corporate training dialogue and coaching system.

[0083] 1. Overall architecture.

[0084] The system architecture of this invention is divided into five layers: user layer, service layer, algorithm layer, data layer, and management layer. It can be deployed in the cloud or locally and supports access from PC (Personal Computer), mobile devices, and VR (Virtual Reality) training devices.

[0085] 1) User layer: student end (to provide training), administrator end (to configure and supervise training).

[0086] 2) Service layer: one-click construction service, dialogue simulation service (supports free dialogue mode and fixed script mode), intelligent evaluation service, multilingual and voice service.

[0087] 3) Algorithm layer: Scene parsing and character generation, free dialogue control (goal-oriented + round constraint + soft FSM), event extraction and evidence chain generation, multilingual recognition and speech synthesis (ASR + TTS (Text To Speech)).

[0088] 4) Data layer: Enterprise knowledge base (SOP, FAQ, case studies), training log and scoring data warehouse, hot word list and tone library.

[0089] 5) Management level: Componentized workbench, report aggregation and data insights, training programs and reminder engine.

[0090] 2. Data flow and core processes.

[0091] 1) Input: The user submits a scenario description D; 2) Analysis: One-click construction of service processing D, generating {Scenario S, Role R, Script Set H, Scoring Dimension M}; 3) Mode selection: Free dialogue mode (goal + round) or fixed script mode (sentence-by-sentence comparison); 4) Dialogue execution: AI acts as a customer to interact with students, updating the goal_state in real time; 5) Evaluation: Identify key events in the dialogue through event extraction, align them with scoring points M, and generate a chain of evidence and a report; 6) Output: The report is sent to the management end for trainees to review and for administrators to gain insights.

[0092] 3. Innovative design.

[0093] 1) Automated one-click build: Combining LLM, knowledge base and rule templates to achieve low-threshold configuration; 2) Free-flowing dialogue soft FSM: Allows yaw and determines the main regression path based on confidence level; 3) Goal-oriented assessment: Emphasizing the completion rate of goals and rounds, rather than comparing sentence by sentence; 4) Evidence chain scoring: Binding event fragments to form an interpretable chain of "target-dimension-event-evidence"; 5) Enhanced multi-speech hot word recognition: Utilizes Shallow Fusion to enhance the recognition of industry terminology; 6) Intelligent reminders: Based on the rule engine and time triggers, a closed-loop training management is formed.

[0094] The relevant algorithms involved in the embodiments of the present invention are described below.

[0095] 1. One-click algorithm construction.

[0096] 1) Input: Scene description D; 2) Analysis: ; Scene refinement: Structured processing using Prompt templates; Role recognition: NER+Few-shot classification to extract identities in conversations; 3) Script generation: ; 4) Generation of rating dimensions: M = Capability → Key Tasks → Actions → Communication Skills; Output: Training partner configuration P={S,R,H,M}.

[0097] 2. Free-flowing dialogue control (goal-oriented + soft FSM).

[0098] 1) Technical solution: Introduce a soft FSM to support yaw-return and ensure that the goal is achieved within a limited number of rounds.

[0099] 2) Implementation method: State representation: .

[0100] 3) Transfer rules: enter u t →Classification intent t ; Update target status; If turn > max_turn and the target is not completed, the decision is considered a failure. When yaw occurs, a fallback strategy is triggered, and the path returns to the main path after confidence is restored.

[0101] 4) Innovation: Avoids the rigidity of traditional FSM and supports yaw-regression mechanism.

[0102] 3. Comparison of fixed script patterns.

[0103] By comparing similarity scores, sentence-by-sentence matching and accuracy are calculated.

[0104] Input: Standard script H and student response U; similarity calculation formula: .

[0105] Output: Sentence-by-sentence comparison results and overall accuracy.

[0106] 4. Intelligent evaluation (event extraction + evidence chain).

[0107] Event extraction and evidence chain binding enable interpretability of the "target-dimension-event-evidence" framework.

[0108] The implementation method is as follows: input dialogue text T; event extraction. ;match Chain of evidence Scoring formula: ;in, w i Indicates the first i The weights of each rating dimension, # completed_events i Indicates the first i Number of completed events under each rating dimension, # total_events i Indicates the first i Total number of events under each rating dimension penalty This indicates that points will be deducted for using inappropriate language.

[0109] 5. Multilingual and timbre binding.

[0110] It involves speech recognition (LID), dynamic switching of ASR, hot word enhancement, and timbre binding.

[0111] The implementation method is as follows: LID determination language; calling the ASR_CN (Mandarin), ASR_HK (Cantonese), or ASR_EN (English) model; Shallow Fusion fusion of enterprise hot word list; establishing role IDs. Sound ID mapping.

[0112] 6. Management insights.

[0113] It provides a modular workbench, report aggregation, and alert engine.

[0114] The implementation method is as follows: input the training report set {R1,…,Rn}; aggregate and calculate the target achievement rate, average number of rounds, and dimensional distribution; identify students with consistently low scores through anomaly detection; trigger reminders through the rule engine; and output charts, reports, and task reminders.

[0115] In summary, the key features of this invention include: one-click construction of integrated semantic parsing, knowledge retrieval, and templated dimension generation; free dialogue using soft FSM, supporting yaw regression; intelligent evaluation using event extraction + evidence chain binding to ensure interpretability; multi-speech recognition combined with hot word enhancement to solve industry terminology issues; and the introduction of a rule engine and anomaly detection on the management end to form closed-loop management.

[0116] The innovative aspects of this invention include: 1. One-click innovation: integrating semantic parsing, knowledge retrieval, and templated dimension generation.

[0117] 2. Dialogue control innovation: Soft FSM mechanism, supports yaw regression, and avoids infinite loops.

[0118] 3. Intelligent assessment innovation: Event extraction + evidence chain binding enhances objectivity and interpretability.

[0119] 4. Speech Recognition Innovation: Multilingual recognition combined with hot word enhancement solves the problems of industry terminology and dialects.

[0120] 5. Management innovation: Rule engine + anomaly detection to achieve closed-loop training management.

[0121] The embodiments of the present invention can lower the threshold for setting up training scenarios, improve the flexibility and realism of dialogue training, and make the evaluation process more scientific, objective and interpretable.

[0122] like Figure 4 As shown, an electronic device 400 provided in this embodiment of the invention includes: a processor 401, a memory 402 and a bus. The memory 402 stores a computer program that can run on the processor 401. When the electronic device 400 is running, the processor 401 and the memory 402 communicate through the bus, and the processor 401 executes the computer program to realize the above-mentioned enterprise training dialogue practice method.

[0123] Specifically, the memory 402 and processor 401 mentioned above can be general-purpose memory and processor, without any specific limitations here.

[0124] This invention also provides a computer-readable storage medium storing a computer program. When a processor runs the computer program, it executes the enterprise training dialogue practice method described in the preceding method embodiments. The computer-readable storage medium includes various media capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), RAM, magnetic disk, or optical disk.

[0125] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0126] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0127] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0128] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface, and the indirect coupling or communication connection between modules may be electrical, mechanical, or other forms.

[0129] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0130] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dialogue practice in corporate training, characterized in that, include: Receive scene description information input by the user; The scenario description information is parsed using natural language processing technology and combined with knowledge retrieval to generate coaching configuration information, which includes coaching scenario, dialogue role, candidate dialogue, and scoring dimensions. Based on the training configuration information and the training mode selected by the user, the dialogue training process with the student is initiated; wherein, the training mode includes a free dialogue mode based on goal and round constraints, or a fixed script mode based on sentence-by-sentence comparison of standard dialogue. After the dialogue training ends, key events that match the scoring dimensions in the training configuration information are identified from the dialogue text, and the corresponding original dialogue fragments are bound as evidence chains, thereby generating an interpretable training report that evaluates the trainee's performance in combination with the goal achievement. The training reports will be synchronized to the management console for training progress tracking and management decision-making. The process of parsing the scene description information using natural language processing technology and generating training configuration information by combining knowledge retrieval includes: performing word segmentation, dependency parsing, and semantic analysis on the scene description information to obtain preliminary structured information, which includes scene keywords, role candidates, and target intent; based on the target intent, calling a preset enterprise knowledge base to extract dialogue templates by retrieving relevant business processes, frequently asked questions, and business cases; refining the scene description information using a large language model based on the scene keywords to generate a complete training scenario; identifying dialogue roles from the scene description information using named entity recognition and few-shot classification algorithms based on the role candidates; generating candidate dialogues for the dialogue roles based on the dialogue templates using retrieval-enhanced generation technology, combined with knowledge base content and a generation model; and decomposing the training objective into multi-level scoring dimensions based on the target intent and preset template rules, and defining scoring points for each scoring dimension; wherein the levels of the scoring dimensions, from high to low, are: capability dimension, key tasks, specific actions, and standard dialogue.

2. The enterprise training dialogue practice method according to claim 1, characterized in that, The execution process of the free dialogue mode includes: Preset dialogue goals and maximum number of dialogue rounds; During the dialogue, a soft finite state machine model is used to track the dialogue state. The state of the soft finite state machine model includes the current round, the goal completion state, the filling of the slot, and the yaw risk indicator. The system judges the speaker's intention in real time. If the confidence level between the speaker's intention and the dialogue target is lower than a preset first threshold, it is determined to be a deviation and a fallback strategy is triggered to guide the dialogue back to the main path. If the dialogue objective is achieved before reaching the maximum number of dialogue rounds, the training session is considered successful; if the dialogue objective is not achieved even after reaching the maximum number of dialogue rounds, the training session is considered unsuccessful.

3. The corporate training dialogue practice method according to claim 1, characterized in that, The execution process of the fixed script pattern includes: Load the preset standard dialogue sequence; Receive trainees' responses line by line; Calculate the semantic vector similarity between the student's response and the corresponding standard response. If the semantic vector similarity exceeds a preset second threshold, it is determined to be a match; otherwise, it is marked as a deviation. The overall accuracy of students' answers is calculated based on the sentence-by-sentence matching results.

4. The enterprise training dialogue practice method according to claim 1, characterized in that, The process of identifying key events from the dialogue text that match the scoring dimensions in the coaching configuration information, and binding the corresponding original dialogue fragments as a chain of evidence, thereby generating an interpretable coaching report that evaluates the trainee's performance based on the goal achievement, includes: The event extraction model was invoked to segment and analyze the dialogue text, identifying multiple key events; Establish a mapping relationship between the key events and the scoring dimensions; For each of the scoring dimensions, key events are matched with original dialogue fragments to form a chain of evidence. Based on the chain of evidence, a weighted scoring model is used to calculate the scores and weighted total scores for each scoring dimension. The training report is generated based on the chain of evidence, the scores of each of the scoring dimensions, the weighted total score, and the goal achievement degree corresponding to the dialogue text.

5. The corporate training dialogue practice method according to claim 1, characterized in that, The aforementioned corporate training dialogue practice methods also include: During the dialogue practice, a language recognition device determines the language type of the student's input speech in real time. Based on the identified target language type, the system dynamically switches to the corresponding speech recognition model for speech-to-text conversion, and uses shallow fusion technology to inject the enterprise's proprietary terminology hot word list into the speech recognition decoding process. To generate speech output by binding specific voices or digital human figures to dialogue characters.

6. The corporate training dialogue practice method according to claim 1, characterized in that, The management console performs the following functions: The system aggregates and displays training reports from multiple trainees, including goal achievement rate, average score, and dimensional distribution statistics. Based on the rules engine, reminder notifications are automatically triggered for students who have not completed their tutoring tasks for a long time or have consistently received low scores; It provides a modular interface for training plan development, task approval, and trainee progress tracking.

7. A corporate training dialogue practice system, characterized in that, include: The information receiving module is used to receive scene description information input by the user; The information parsing module is used to parse the scene description information based on natural language processing technology and generate coaching configuration information in combination with knowledge retrieval. The coaching configuration information includes coaching scene, dialogue role, candidate dialogue and scoring dimension. The dialogue practice module is used to initiate a dialogue practice process with the student based on the practice configuration information and the practice mode selected by the user; wherein, the practice mode includes a free dialogue mode based on goal and round constraints, or a fixed script mode based on sentence-by-sentence comparison of standard dialogue. The coaching evaluation module is used to identify key events from the dialogue text that match the scoring dimensions in the coaching configuration information after the dialogue coaching is completed, and bind the corresponding original dialogue fragments as evidence chains, thereby generating an interpretable coaching report that evaluates the student's performance. The report synchronization module is used to synchronize the training reports to the management console for training progress tracking and management decision-making. The information parsing module is specifically used for: performing word segmentation, dependency parsing, and semantic parsing on the scenario description information to obtain preliminary structured information, which includes scenario keywords, role candidates, and target intent; based on the target intent, calling a preset enterprise knowledge base to extract dialogue templates by retrieving relevant business processes, frequently asked questions, and business cases; refining the scenario description information using a large language model based on the scenario keywords to generate a complete training scenario; identifying dialogue roles from the scenario description information using named entity recognition and few-shot classification algorithms based on the role candidates; generating candidate dialogues for the dialogue roles based on the dialogue templates using retrieval enhancement generation technology, combined with knowledge base content and generation models; and decomposing the training objective into multi-level scoring dimensions based on the target intent and preset template rules, and defining scoring points for each scoring dimension; wherein, the levels of the scoring dimensions, from high to low, are: capability dimension, key tasks, specific actions, and standard dialogue.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the enterprise training dialogue practice method as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by the processor, performs the enterprise training dialogue practice method as described in any one of claims 1-6.

Images