Automatic full-process intelligent recruitment system based on competency assessment

Abstract

The invention relates to an automatic full-process intelligent recruitment system based on competency evaluation, and belongs to the technical field of intelligent interview. The system comprises a data acquisition module, a question generation module, an interactive execution module, a dynamic questioning module, a dimension switching module, an anomaly monitoring module, a dual-camera verification module, a score report module and a context maintenance module. Through cooperative work of the modules, the system can realize full-process automation from candidate data collection to final scoring report generation. The data acquisition module is responsible for collecting various information of candidates and providing basic data for subsequent interview links. Automation, intelligentization and standardization of the recruitment process are realized, and the recruitment efficiency and quality are effectively improved. By generating interview questions and dynamic questions in real time, the system can more accurately inspect ability performance of candidates in different competency dimensions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent interviews, and particularly relates to an automatic full-process intelligent recruitment system based on competency evaluation. BACKGROUND

[0002] Traditional recruitment system: the resume analysis technology based on NLP has realized keyword matching and basic information extraction, but lacks deep semantic analysis of work experience and project experience; in terms of structured interviews, existing systems mostly adopt a preset question bank mode, and question generation relies on a standardized question bank prepared by humans, lacking the ability to dynamically adjust to the individual experiences of candidates; enterprises generally define competency dimensions through job descriptions, but the evaluation process relies on the subjective judgment of interviewers, lacking fine-grained quantitative indicators; video interview platforms in automated interview processes have supported asynchronous questioning and voice recognition, but the follow-up mechanism mostly uses fixed logic jumps, and cannot generate targeted questions in real time according to the answer content; some systems (such as the "Recruitment Efficiency Assistant" of Zhaopin) use large models to optimize resume screening and interview question generation, but are limited to single-round questioning and have not realized cross-round context association; advanced systems (such as the "Multi-Face AI" of Liepin) have attempted to generate follow-up questions based on resume content, but the follow-up logic relies on rule engines, lacking dynamic adjustment to the competency evaluation state. SUMMARY

[0003] To solve the above problems in the prior art, the application provides an automatic full-process intelligent recruitment system based on competency evaluation; The object of the application can be achieved by the following technical solutions: An automatic full-process intelligent recruitment system based on competency evaluation, comprising: a data acquisition interface for acquiring resume information data of a candidate and a preset post competency framework; a large language model agent core engine for loading and running a large language model, the configuration including a question generation module, an interaction and dynamic follow-up module, a dimension switching module, an abnormality monitoring and analysis module, a scoring and report generation module, a context perception and maintenance module; The question generation module is used to start the process, the interaction and dynamic follow-up module processes the answers of the candidate and triggers multiple follow-up questions; the abnormality monitoring and analysis module runs in real time and in parallel during the interaction process and analyzes the answer content; the dimension switching module is used to determine whether the current dimension is ended, and if so, triggers switching to the next dimension and reactivates the question generation module; the scoring and report generation module performs scoring and summary at the end of the process or at the end of a dimension; the context perception and maintenance module is used to build, store and update the dynamic graph structure of the entire interview dialogue, providing decision basis and input information for other functional modules; The double-camera verification subsystem is configured to capture real-time dynamic images of the upper body posture and facial feature information of the candidate through double-camera shooting, calculate a pixel occlusion rate of an effective facial recognition area to generate an occlusion event, and count a frequency of the candidate leaving a center area of a picture within a unit time window to generate a leaving event; and encode the occlusion event and the leaving event into a structured message.

[0004] The master coordination unit is configured to receive the abnormality monitoring event from the large language model intelligent agent core engine and the occlusion / leaving event from the double-camera verification subsystem, and send an instruction to the large language model intelligent agent core engine based on a face reference behavior or a leaving frequency exceeding a standard behavior to activate a stress test type template in a dynamic follow-up question mode.

[0005] Specifically, the question generation module includes three generation modes, including a general reasoning type, a chain reasoning type, and an adversarial type. The general reasoning type is configured to generate a standardized question sequence for a general post based on a basic evaluation dimension in a preset post competency framework through a prompt engineering technology of a large language model. The chain reasoning type is configured to decompose the project experience of the candidate into multiple sub-tasks through a thinking chain decomposition technology, and generate a follow-up question group with a logical progressive relationship for each sub-task. The adversarial type is configured to generate open scenario simulation questions and challenging assumption conditions based on a post complexity threshold and a stress test parameter, for evaluating the emergency decision-making ability of the candidate.

[0006] Specifically, the interaction and dynamic follow-up question module provides the generated questions to the candidate, receives the answer content of the candidate, and autonomously decides and dynamically generates follow-up questions based on real-time understanding of the current answer content, historical dialogue context, and current competency evaluation state. The mode uses a context understanding model to analyze semantic association, and adjusts follow-up behavior in three dimensions of follow-up depth, question aggressiveness, and topic switching rate according to a preset follow-up rule library (including a detail clarification type, a contradiction detection type, and a stress test type template) and a current evaluation state (including competency score, answer completeness, and semantic coherence index).

[0007] Specifically, the abnormality monitoring and analysis module analyzes the answer content of the candidate in real time during the interaction, detects text structure abnormalities through text feature extraction, and calculates the similarity between the answer content and historical high-frequency templates to identify face reference behaviors by associating a face reference knowledge base. If text structure abnormalities or face reference behavior abnormalities are detected, an abnormality event is triggered.

[0008] Specifically, the scoring and report generation module calculates the deviation degree of the answer content from a preset standard answer through a semantic distance model according to the answer performance of the candidate in each competency dimension to generate a dimension score, and generates a comprehensive score and a report containing the score and the evaluation result by calling a preset weight template according to the type of the post to weight and sum the dimension scores.

[0009] Specifically, the context perception and maintenance module dynamically stores and maintains the entire interview dialogue history in a graph structure, models the candidate as a central node, and associates and times the question and answer; it drives the question generation, dynamic follow-up, dimension switching and scoring decision-making process throughout.

[0010] Specifically, the master coordination unit has a resource allocation strategy table, dynamically adjusts computing resources during the interview process, and preferentially allocates GPU resources to the reasoning task of the agent when the follow-up question is activated, and binds a dedicated video processing unit to the real-time analysis task of the dual camera.

[0011] Specifically, the master coordination unit broadcasts system state machine change events through the publish-subscribe mode to drive system mode switching.

[0012] Specifically, in the context perception and maintenance module of the large language model agent core engine, the graph structure includes a semantic correlation weight calculation unit configured to: vectorize the question and answer content through a word vector model to generate semantic correlation weight values of question nodes and answer nodes; calculate a conversation coherence decay coefficient according to the time stamp difference, the decay coefficient being negatively correlated with the interval time between adjacent questions and answers; dynamically weight and fuse the semantic correlation weight values and the decay coefficient to generate an edge attribute weight matrix; the edge attribute weight matrix is used for question path optimization decision-making in the dynamic follow-up mode.

[0013] Specifically, the cross-round dynamic adjustment of the large language model agent core engine includes: After each round of question and answer, the newly generated question, answer and time stamp are encapsulated as a triple node and inserted into the graph structure, and the edge attribute weight matrix is dynamically updated according to the semantic correlation weight calculation unit; When generating follow-up questions, the content, depth and method of the follow-up questions are adjusted autonomously in combination with the current competence evaluation state and the historical question and answer context information stored in the graph structure; The edge attribute weight matrix in the graph structure is continuously scanned, and when the semantic correlation weight is detected to be lower than a preset threshold, cross-dimension follow-up intervention is autonomously decided to guide the dialogue back to the core evaluation point and switch dimensions.

[0014] The beneficial effects of the present application are: The automatic full-process intelligent recruitment system of the application has many significant advantages. In terms of recruitment efficiency, the system can generate interview questions in real time according to the candidate resume information and the preset competence framework through a large language model, avoiding the complexity of relying on manual preparation of a standardized question bank in traditional recruitment, and greatly shortening the interview preparation time. Moreover, in the interview process, the dynamic follow-up module can generate follow-up questions in real time according to the candidate's answers, making the interview process more compact and efficient and reducing unnecessary time waste.

[0015] In terms of evaluation accuracy, the system ensures accurate judgment of the candidate's competence through a multi-dimensional comprehensive evaluation mechanism. The three generation modes of the question generation module examine the candidate from different angles. The general reasoning type generates a standardized question sequence for general posts, which can cover the basic evaluation dimensions of the post; the chain reasoning type conducts logical progressive follow-up questions for the candidate's project experience, and deeply understands the project execution ability; the confrontation type evaluates the candidate's emergency decision-making ability through open scenario simulation and challenging assumptions. The scoring report module calculates the deviation degree of the candidate's answers from the standard answers through a semantic distance model, and then combines the preset weight template to perform weighted summation, obtaining an objective and accurate comprehensive score.

[0016] In terms of interview fairness, the system greatly reduces the interference of human factors. The abnormality monitoring module can identify abnormal behaviors in the candidate's answer content, such as text structure abnormalities or face reference behavior abnormalities, and timely report them to prevent the candidate's improper behavior from affecting the interview results. The dual-camera verification module monitors the candidate's face and body posture in real time, and timely marks and feeds back situations such as face covering and frequent leaving of the screen, ensuring the standardization and fairness of the interview process.

[0017] In addition, the cross-round dynamic adjustment function of the system enhances the flexibility and adaptability of the interview. The dynamic updating of the graph structure, the context perception adjustment and the cross-round intervention enable the system to dynamically adjust the questions and follow-up questions according to the historical question and answer content and the current competence evaluation state, ensuring a comprehensive and in-depth understanding of the candidate's ability and quality throughout the interview process, and selecting the most suitable talent for the post for the enterprise. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to facilitate the understanding of those skilled in the art, the application will be further described below with reference to the accompanying drawings.

[0019] Fig. 1 The structure diagram of an automatic full-process intelligent recruitment system based on competence evaluation of the application; Fig. 2 The timing processing structure schematic diagram of an automatic full-process intelligent recruitment system based on competence evaluation in the application. DETAILED DESCRIPTION

[0020] To further clarify the technical means and effects taken by the present application to achieve the predetermined inventive purpose, the specific embodiments, structures, features and effects thereof according to the present application are described in detail below in conjunction with the drawings and preferred embodiments.

[0021] Please refer to Figs. 1-2 An automatic full-process intelligent recruitment system based on competency evaluation, comprising: A data acquisition interface is configured to acquire resume information data of a candidate and a preset post competency framework. A large language model agent core engine is configured to load and run a large language model, and the configuration includes a question generation module, an interaction and dynamic follow-up question module, a dimension switching module, an abnormality monitoring and analysis module, a scoring and report generation module, and a context perception and maintenance module. The question generation module is configured to start the process, the interaction and dynamic follow-up question module is configured to process the candidate's answers and trigger multiple follow-up questions, the abnormality monitoring and analysis module is configured to run in real time and in parallel during the interaction process and analyze the answer content, the dimension switching module is configured to determine whether the current dimension is ended, and if so, trigger switching to the next dimension and reactivate the question generation module, the scoring and report generation module is configured to score and summarize at the end of the process or at the end of the dimension, and the context perception and maintenance module is configured to build, store and update the dynamic graph structure of the entire interview dialogue, and provide decision basis and input information for other functional modules. A dual-camera verification subsystem is configured to capture real-time dynamic images of the candidate's upper body posture and facial feature information through dual-camera positions, calculate the pixel occlusion rate of the effective facial recognition area to generate an occlusion event, and count the frequency of the candidate's departure from the center area of the screen within a unit time window to generate a departure event, and encode the occlusion event and the departure event into a structured message.

[0022] A master control coordination unit is configured to receive abnormality monitoring events from the large language model agent core engine and occlusion / departure events from the dual-camera verification subsystem, and based on the face reference behavior or the departure frequency exceeding the standard behavior, send instructions to the large language model agent core engine to activate the stress test type template in the dynamic follow-up question mode.

[0023] Specifically, the question generation module includes three generation modes, including general reasoning type, chain reasoning type, and confrontation type; the general reasoning type is configured to generate a standardized question sequence for a general post based on a basic evaluation dimension in a preset post competence framework through a prompt engineering technology of a large language model; the chain reasoning type is configured to decompose a candidate's project experience into multiple sub-tasks, and generate a group of follow-up questions with a logical progressive relationship for each sub-task; and the confrontation type is configured to generate open scenario simulation questions and challenging assumption conditions based on a post complexity threshold and a stress test parameter, for evaluating the emergency decision-making ability of the candidate.

[0024] Specifically, the interaction and dynamic follow-up question module provides the generated questions to the candidate, receives the answer content of the candidate, and autonomously decides and dynamically generates follow-up questions based on real-time understanding of the current answer content, historical dialogue context, and current competence evaluation state; the mode utilizes a context understanding model to analyze semantic association, and adjusts follow-up behavior in three dimensions of follow-up depth, question aggressiveness, and topic switching rate according to a preset follow-up rule library (including detail clarification type, contradiction detection type, and stress test type templates) and the current evaluation state (including competence score, answer completeness, and semantic coherence index).

[0025] Specifically, the abnormality monitoring and analysis module analyzes the answer content of the candidate in real time during the interaction, performs text structure abnormality detection through text feature extraction, and calculates the similarity between the answer content and historical high-frequency templates to identify face reference behavior by associating a face experience knowledge base; if text structure abnormality or face reference behavior abnormality is detected, an abnormal event is triggered.

[0026] Specifically, the scoring and report generation module calculates the deviation degree of the answer content from a preset standard answer through a semantic distance model according to the answer performance of the candidate in each competence dimension, to generate a dimension score; and according to the type of the post, a preset weight template is called to weight and sum the dimension scores, to generate a comprehensive score and a report containing the score and the evaluation result.

[0027] Specifically, the context perception and maintenance module dynamically stores and maintains the entire interview dialogue history in a graph structure, models the candidate as a center node, and associates and times the question and answer; and drives the question generation, dynamic follow-up, dimension switching, and scoring decision processes.

[0028] Specifically, the main control coordination unit has a resource allocation strategy table built-in, dynamically adjusts computing resources during the interview, and preferentially allocates GPU resources to the reasoning task of the agent when the agent activates follow-up questions, and binds a dedicated video processing unit to the real-time analysis task of the dual camera.

[0029] Specifically, the master coordination unit broadcasts system state machine change events through a publish-subscribe mode to drive system mode switching.

[0030] Specifically, in the context perception and maintenance module of the large language model agent core engine, the graph structure includes a semantic correlation weight calculation unit configured to: vectorize the question and answer content through a word vector model to generate semantic correlation weight values of question nodes and answer nodes; calculate a conversation coherence decay coefficient according to a timestamp difference, the decay coefficient being negatively correlated with adjacent question and answer interval time; and dynamically weight and fuse the semantic correlation weight values and the decay coefficient to generate an edge attribute weight matrix; the edge attribute weight matrix is used for optimization decision-making of a follow-up question mode.

[0031] Specifically, the cross-round dynamic adjustment of the large language model agent core engine includes: After each round of question and answer, the newly generated question, answer and timestamp are encapsulated as a triple node and inserted into the graph structure, and the edge attribute weight matrix is dynamically updated according to the semantic correlation weight calculation unit; When generating a follow-up question, the content, depth and method of the follow-up question are autonomously adjusted in combination with the current competence evaluation state and historical question and answer context information stored in the graph structure; The edge attribute weight matrix in the graph structure is continuously scanned, and when the semantic correlation weight is detected to be lower than a preset threshold, cross-dimension follow-up question intervention is autonomously decided to guide the dialogue back to the core evaluation point and switch dimensions.

[0032] In this embodiment, a bidirectional Transformer architecture based on BERT is used to analyze entities (such as technical terms and project roles), intents (clarification, verification, and stress testing), and semantic correlations in the dialogue history in real time. Through a self-attention mechanism, key information points in the candidate's answer are captured, for example, when the candidate describes "leading database optimization of a project", the model automatically identifies the behavior subjectivity of "leading" and the technical dimension of "database optimization", and maps them to the preset "technical leadership" competence framework. Combined with the global text representation capability of the GPT series model, a multi-dimensional dialogue graph containing timelines, causal relationships, and technical difficulties is constructed to provide semantic support for subsequent follow-up questions.

[0033] The rule base contains three types of templates, which are dynamically loaded using the Drools rule engine: When the completeness of the answer is less than 60%, trigger, such as "Please specify the proportion of SQL optimization work you undertook in the project?" ; Detect statement contradictions through dependency syntax analysis, such as "The resume mentions that project A was completed in 3 months, but the previous answer mentioned a development cycle of half a year. Please explain the time difference" ; Based on the post complexity threshold, automatically generate challenge scenarios, such as "If the current optimization scheme leads to system compatibility problems, how will the responsible person make decisions?"

[0034] The abnormality monitoring module constructs a multi-level analysis pipeline: Vocabulary layer: Calculate pronoun density (I / We ratio), passive voice frequency (trigger words such as "by" and "through"); Syntactic layer: Detect abnormal sentence structures such as more than 3 layers of nested conditional clauses through Stanford Parser; Semantic layer: Use Text-CNN to extract n-gram features and detect homogenization expressions (same technical term repetition rate > 70%); Dynamic matching of face-to-face knowledge base, store historical high-frequency face-to-face answers through template library, and vectorize encoding through Doc2Vec; In real time, if the candidate's answer has a cosine similarity > 0.85 with the standard answer of "Redis cache penetration solution", and the syntactic tree matching degree > 90%, it is determined that it is a template answer.

[0035] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any equivalent embodiments with equivalent changes are equivalent. Any simple modification, equivalent change and modification of the above embodiments made in accordance with the technical essence of the present application are still within the scope of the technical solution of the present application.

Claims

1. An automatic full-process intelligent recruitment system based on competency assessment, characterized by: include: Data acquisition interface: used to obtain candidate resume information data and preset job competency framework; Large language model agent core engine: used to load and run large language models. Configuration includes question generation module, interaction and dynamic questioning module, dimension switching module, anomaly monitoring and analysis module, scoring and report generation module, and context awareness and maintenance module. The question generation module is used to start the process, and the interaction and dynamic questioning module processes the candidate's answers and triggers multiple follow-up questions. The anomaly monitoring and analysis module runs in real time and in parallel during the interaction process to analyze the answer content. The dimension switching module is used to determine whether the current dimension has ended. If it has ended, it will trigger a switch to the next dimension and reactivate the question generation module. The scoring and report generation module summarizes the scores at the end of the process or the end of the dimension. The context perception and maintenance module is used to build, store and update the dynamic graph structure of the entire interview dialogue, providing decision-making basis and input information for other functional modules. Dual-camera verification subsystem: used to capture the candidate's upper body posture dynamic images and facial feature information in real time through dual cameras; Calculate the pixel occlusion rate of the effective facial recognition area to generate an occlusion event, and count the frequency of candidates leaving the center area of ​​the screen within a unit time window to generate a departure event; Encode occlusion events and departure events into structured messages; Main control coordination unit: receiving abnormal monitoring events from the large language model intelligent agent core engine and occlusion / exit events from the dual-camera verification subsystem; Based on the reference behavior of the interview experience or the behavior of exceeding the frequency of leaving the interview, instructions are sent to the core engine of the large language model intelligent agent to activate the stress test template in the dynamic questioning mode.

2. The system according to claim 1, wherein: The question generation module includes three generation modes, including general reasoning, chain reasoning, and adversarial. The general reasoning mode is configured to generate a standardized question sequence for common positions based on the basic evaluation dimensions in the preset position competency framework through the prompt engineering technology of the large language model; the chain reasoning mode is configured as a thinking chain decomposition technology, which breaks down the candidate's project experience into multi-level subtasks, and generates a group of follow-up questions with a logical progressive relationship for each subtask; the adversarial mode is configured to combine the position complexity threshold and stress test parameters to generate open scenario simulation questions and challenging assumptions for evaluating the candidate's emergency decision-making ability.

3. The system according to claim 1, wherein: The interactive and dynamic follow-up questioning module provides generated questions to candidates, receives their answers, and independently decides and dynamically generates follow-up questions based on real-time understanding of the current answer content, historical conversation context, and current competency assessment status. This mode uses a context understanding model to parse semantic associations and regulates follow-up behavior in three dimensions: depth of questioning, aggressiveness of questioning, and topic switching rate, based on a preset follow-up questioning rule library (including detail clarification, contradiction detection, and stress testing templates) and the current assessment status (including competency scoring, answer completeness, and semantic coherence indicators).

4. The system according to claim 1, wherein: During the interaction process, the anomaly monitoring and analysis module analyzes the candidate's answer content in real time, detects text structure anomalies through text feature extraction, and calculates the similarity between the answer content and historical high-frequency templates in conjunction with the interview knowledge base to identify interview reference behavior; If an abnormal text structure or abnormal face reference behavior is detected, an abnormal event will be triggered.

5. The system according to claim 1, wherein: The scoring and report generation module calculates the deviation between the candidate's answer performance in each competency dimension and the preset standard answer through a semantic distance model to generate a dimension score; calls a preset weight template according to the position type, performs weighted summation on the dimension scores, and generates a comprehensive score and a report containing the score and evaluation results.

6. The system according to claim 1, wherein: The context perception and maintenance module dynamically stores and maintains the entire interview conversation history in a graph structure, uses the candidate as the central node to model the question-answer association and time sequence; and runs through and drives the question generation, dynamic questioning, dimension switching, and scoring decision-making process.

7. The system according to claim 1, wherein: The main control coordination unit has a built-in resource allocation strategy table, which dynamically adjusts computing resources during the interview process, prioritizes allocating GPU resources to its reasoning tasks when the intelligent agent activates follow-up questions, and binds the dual-camera real-time analysis tasks to a dedicated video processing unit.

8. The system according to claim 1, wherein: The master control coordination unit broadcasts system state machine change events in a publish-subscribe mode to drive the system working mode switching.

9. The system according to claim 1, wherein: In the context perception and maintenance module of the large language model intelligent agent core engine, the graph structure includes a semantic relevance weight calculation unit, which is configured as follows: vectorizing the question and answer content through a word vector model to generate a semantic relevance weight value between the question node and the answer node; calculating the conversation coherence attenuation coefficient based on the timestamp difference, and the attenuation coefficient is negatively correlated with the time interval between adjacent questions and answers; dynamically weighted fusion of the semantic relevance weight value and the attenuation coefficient to generate an edge attribute weight matrix; the edge attribute weight matrix is ​​used for question path optimization decision-making in the dynamic questioning mode.

10. The system according to claim 1, wherein: The specific implementation method of the cross-round dynamic adjustment of the core engine of the large language model agent includes: After each round of question-answering, the newly generated question, answer, and timestamp are encapsulated as a triple node and inserted into the graph structure, and the edge attribute weight matrix is ​​dynamically updated based on the semantic relevance weight calculation unit; When generating follow-up questions, the system automatically adjusts the content, depth, and method of the questions based on the current competency assessment status and the historical question and answer context information stored in the graph structure. Continuously scan the edge attribute weight matrix in the graph structure. When it detects that the semantic relevance weight is lower than the preset threshold, autonomous decision-making triggers cross-dimensional questioning intervention, guiding the conversation back to the core evaluation point and switching dimensions.

Images