Method, system and device for full-process automatic interview assistance based on large language model

Abstract

This invention belongs to the field of artificial intelligence technology. The embodiments of this invention provide a fully automated interview assistance method, system, and device based on a large language model. First, it acquires recruitment requirements, initial interview questions, and multimodal data of job seekers from the recruitment end to construct a multimodal dynamic knowledge base. This multimodal dynamic knowledge base includes content dynamic mechanisms, structure dynamic mechanisms, and iterative dynamic mechanisms, used to update the knowledge base content and sub-question sequences, update the structure of the knowledge graph in the multimodal dynamic knowledge base, and optimize sub-question generation and retrieval strategies, respectively. Then, based on the multimodal dynamic knowledge base, it determines the job seeker's structured question set, and subsequently determines the job seeker's interview assistance strategy. This invention, through three update mechanisms, makes knowledge application more closely aligned with actual interview scenarios, overcoming the shortcomings of traditional static and rigid knowledge bases.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method, system and device for full-process automatic interview assistance based on a large language model. Background Technology

[0002] With breakthroughs in large language models in the field of natural language processing, enterprises are increasingly demanding intelligent and automated interview assistance tools for recruitment. Modern recruitment involves job requirements from the recruiting side, multimodal job seeker data (resumes, videos, portfolios, etc.), and complex reasoning and evaluation. It requires efficient handling of the entire process, including information matching, dynamic interaction, and multi-dimensional evaluation. Traditional manual or semi-automated methods are no longer sufficient to meet the needs of large-scale and precise recruitment.

[0003] Current interview assistance technologies suffer from static knowledge bases, relying on fixed question banks and templates. They cannot integrate new industry skills, changes in job requirements, and dynamic data of job seekers in real time, making it difficult to uncover implicit abilities from resumes and videos. Summary of the Invention

[0004] In view of this, the present invention provides a fully automated interview assistance method, system and device based on a large language model, which aims to solve the problem of existing interview assistance technologies relying on fixed question banks and templates.

[0005] The first aspect of this invention provides a fully automated interview assistance method based on a large language model, comprising:

[0006] Obtain multimodal data on recruitment needs, initial interview questions, and job seekers from the recruitment platform;

[0007] Based on recruitment needs, initial interview questions, and multimodal data, a multimodal dynamic knowledge base is constructed. This multimodal dynamic knowledge base incorporates three mechanisms: a content dynamic mechanism, a structure dynamic mechanism, and an iterative dynamic mechanism. The content dynamic mechanism updates the knowledge base content and sub-question sequences; the structure dynamic mechanism updates the structure of the knowledge graph within the multimodal dynamic knowledge base; and the iterative dynamic mechanism optimizes sub-question generation and retrieval strategies.

[0008] Based on a multimodal dynamic knowledge base, determine the structured question set for job seekers;

[0009] Based on a set of structured questions, determine interview support strategies for job seekers.

[0010] In one possible implementation, a multimodal dynamic knowledge base is constructed based on recruitment needs, initial interview questions, and multimodal data, including:

[0011] Obtain the initial knowledge graph;

[0012] The question bank is expanded based on recruitment needs, initial interview questions, and incremental web crawling technology. The content of SAO triples and sub-problem sequences are updated based on the expanded question bank and content dynamic mechanism. Among them, SAO triples are subject-behavior-object triples.

[0013] Update the entities and relationships of the SAO triples based on the updated SAO triple content, the job seeker's multimodal data, and the structural dynamics mechanism;

[0014] Based on the updated sub-question sequence, job seekers' multimodal data, and iterative dynamic mechanisms, we infer sub-question generation and retrieval strategies.

[0015] In one possible implementation, the content of the SAO triples and the sequence of sub-problems are updated based on the expanded question bank and a dynamic content mechanism, including:

[0016] New skill requirements, question types, and scenarios are extracted from the expanded question bank, and new SAO triples are generated and added to the knowledge base.

[0017] Through the dynamic retrieval function of the content dynamic mechanism, the new SAO triples are associated and matched with the content of existing SAO triples in the initial knowledge graph, and entities with inconsistent descriptions are corrected.

[0018] Based on the newly added SAO triples and the core entities of recruitment needs, new sub-problems are generated and integrated into the sub-problem sequence.

[0019] In one possible implementation, the entities and relations of the SAO triples are updated based on the updated content of the SAO triples, the job seeker's multimodal data, and structural dynamics mechanisms, including:

[0020] Extract new SAO triples from job seekers' multimodal data and identify newly added entities within them;

[0021] The entity alignment function of the structural dynamic mechanism is used to semantically match the new entity with the existing entity in the knowledge graph.

[0022] Based on the updated SAO triples, the relationships between entities are supplemented, and conflicting relationships are corrected;

[0023] An entity tree construction algorithm is used to adjust the hierarchical structure of the knowledge graph, setting the core skill entity as the root node and the associated project experience and tool usage as child nodes, thereby optimizing the association paths between entities.

[0024] In one possible implementation, a sub-question generation and retrieval strategy is inferred based on the updated sub-question sequence, job seeker multimodal data, and an iterative dynamic mechanism, including:

[0025] We analyze the interview response performance of job seekers in multimodal data, and use natural language processing technology to calculate the relevance of the answers to sub-questions, the completeness of the answers, and the accuracy of professional terminology, and mark inefficient sub-questions.

[0026] The reinforcement learning module based on the iterative dynamic mechanism uses the above analysis results as a reward signal to fine-tune the parameters of the sub-problem generation model and optimize the generation weights of the sub-problems.

[0027] The retrieval strategy is dynamically adjusted based on the strength of the associations between entities in the knowledge graph.

[0028] It receives feedback from the recruitment end regarding the interview results, iteratively corrects the logical structure of the sub-question sequence based on the feedback, supplements sub-questions for corresponding dimensions, and updates the retrieval benchmark through an entity tree segment filtering mechanism.

[0029] In one possible implementation, a structured set of questions for job seekers is determined based on a multimodal dynamic knowledge base, including:

[0030] By using a query-aware entity tree segment filtering mechanism, the core entities and semantic relationships of job requirements are extracted from the knowledge graph.

[0031] By using complex interview questions and paragraphs associated with entity trees as prompts, semantically independent sets of sub-questions are generated.

[0032] Semantic speech activity detection technology is used to support real-time interruption and contextual continuation, ensuring that the sequence of sub-questions conforms to the logic of natural dialogue and forms a structured question set.

[0033] In one possible implementation, interview support strategies for job seekers are determined based on a set of structured questions, including:

[0034] It adopts a multi-stream full-duplex architecture to achieve natural dialogue, interacts with job seekers in real time based on a structured question set, and dynamically retrieves knowledge base content to answer questions;

[0035] We use a multimodal thinking chain to evaluate job seekers' interaction data and output an assessment report with logical chains.

[0036] In one possible implementation, the method also includes:

[0037] Push the assessment report to the recruitment end and receive detailed instructions; push improvement suggestions based on SAO differences to candidates;

[0038] Obtain recruitment decisions and feed them back to the iterative dynamic mechanism.

[0039] A second aspect of the present invention provides a fully automated interview assistance system based on a large language model, comprising:

[0040] The acquisition module is used to acquire recruitment requirements, initial interview questions, and multimodal data of job seekers from the recruitment platform.

[0041] A construction module is used to build a multimodal dynamic knowledge base based on the recruitment requirements, the initial interview questions, and the multimodal data. The multimodal dynamic knowledge base includes a content dynamic mechanism, a structure dynamic mechanism, and an iterative dynamic mechanism. The content dynamic mechanism is used to update the knowledge base content and sub-question sequences. The structure dynamic mechanism is used to update the structure of the knowledge graph in the multimodal dynamic knowledge base. The iterative dynamic mechanism is used to optimize the sub-question generation and retrieval strategies.

[0042] The retrieval module is used to determine the job seeker's structured question set based on the multimodal dynamic knowledge base;

[0043] The determination module is used to determine the interview support strategies for job seekers based on the structured question set.

[0044] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the full-process automatic interview assistance method based on a large language model as described in the first aspect above.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the full-process automatic interview assistance method based on a large language model as described in the first aspect above.

[0046] The present invention provides a fully automated interview assistance method, system, and device based on a large language model. First, it acquires recruitment requirements, initial interview questions, and multimodal data of job seekers from the recruitment end. Then, based on the recruitment requirements, initial interview questions, and multimodal data, it constructs a multimodal dynamic knowledge base. This multimodal dynamic knowledge base includes a content dynamic mechanism, a structure dynamic mechanism, and an iterative dynamic mechanism. The content dynamic mechanism updates the knowledge base content and sub-question sequences; the structure dynamic mechanism updates the structure of the knowledge graph in the multimodal dynamic knowledge base; and the iterative dynamic mechanism optimizes sub-question generation and retrieval strategies. Next, based on the multimodal dynamic knowledge base, it determines the job seeker's structured question set. Finally, based on the structured question set, it determines the job seeker's interview assistance strategy. This invention ensures the freshness of knowledge by updating knowledge content and sub-question sequences in real time through the content dynamic mechanism, optimizes the entities and relationships of the knowledge graph through the structure dynamic mechanism to ensure a reasonable structure, and further strengthens the synergy between the two through the iterative dynamic mechanism's adaptation strategy. This makes knowledge application more closely aligned with actual interview scenarios, overcoming the shortcomings of traditional static and rigid knowledge bases. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating the implementation of the fully automated interview assistance method based on a large language model provided in this embodiment of the invention. Detailed Implementation

[0049] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0050] Figure 1 This is a flowchart illustrating the implementation of the fully automated interview assistance method based on a large language model provided in this embodiment of the invention. Figure 1 As shown, in some embodiments, the end-to-end automated interview assistance method based on a large language model includes:

[0051] S110: Obtain recruitment requirements, initial interview questions, and multimodal data of job seekers from the recruitment end.

[0052] S120: Based on recruitment needs, initial interview questions, and multimodal data, a multimodal dynamic knowledge base is constructed. This multimodal dynamic knowledge base includes a content dynamic mechanism, a structure dynamic mechanism, and an iterative dynamic mechanism. The content dynamic mechanism is used to update the knowledge base content and sub-question sequences; the structure dynamic mechanism is used to update the structure of the knowledge graph in the multimodal dynamic knowledge base; and the iterative dynamic mechanism is used to optimize sub-question generation and retrieval strategies.

[0053] S130, Based on the multimodal dynamic knowledge base, determine the job seeker's structured question set;

[0054] S140, Based on the structured question set, determine the interview support strategies for job seekers.

[0055] In this embodiment of the invention, structured and unstructured recruitment requirements are received from the recruitment end. The structured requirements include job title, job description, required skills, and priority conditions, while the unstructured requirements cover text descriptions such as work mode and team collaboration requirements. The core entities (such as skills, experience, and tools) and related relationships (such as "job-requirements-skills") are extracted through natural language processing technology.

[0056] Collect initial interview questions provided by recruiters, categorize and store them according to assessment dimensions (technical skills, behavioral qualities, project experience, etc.), including open-ended questions (such as "Describe the most complex technical problem you have solved"), closed-ended questions (such as "Do you have mastered the Spark framework?"), and scenario-based questions (such as "How to handle data skew issues"), and label the assessment objectives for each question (such as "Evaluate problem-solving ability").

[0057] We collect multimodal data from job seekers, specifically text, video, and portfolio data. All data is stored after being standardized in format and desensitized for privacy, providing raw data support for subsequent SAO triple extraction, implicit ability mining, and knowledge graph construction.

[0058] In some embodiments, a multimodal dynamic knowledge base is constructed based on recruitment needs, initial interview questions, and multimodal data, including: obtaining an initial knowledge graph; expanding the question bank based on recruitment needs, initial interview questions, and incremental crawling technology, and updating the content and sub-question sequence of SAO triples based on the expanded question bank and content dynamic mechanism; wherein, the SAO triple is a subject-behavior-object triple; updating the entities and relations of the SAO triples based on the updated content of the SAO triples, the job seeker's multimodal data, and the structural dynamic mechanism; and reasoning sub-question generation and retrieval strategies based on the updated sub-question sequence, the job seeker's multimodal data, and the iterative dynamic mechanism.

[0059] In this embodiment of the invention, the process of constructing a multimodal dynamic knowledge base can be broken down into four progressive steps:

[0060] 1. Initialization of basic knowledge graph: First, obtain an initial knowledge graph containing basic information such as industry-wide skills, job standards, and common interview dimensions, which will serve as the underlying framework of the knowledge base.

[0061] 2. Dynamic Content Updates: Combining recruitment needs and initial interview questions, new questions are crawled from industry platforms, enterprise question banks, and other channels using incremental crawling technology to expand the question bank. Then, using a dynamic content mechanism, "subject-behavior-object" (SAO) triples are extracted from the new questions and added to the knowledge base. At the same time, new sub-questions are generated by associating them with the core entities of recruitment needs, and integrated to form an updated sequence of sub-questions.

[0062] 3. Structural Dynamic Optimization: Based on the SAO triple content updated in the previous step, combined with job seeker multimodal data, new entities are identified through a structural dynamic mechanism, and semantic matching and alignment are performed with existing entities in the knowledge graph; at the same time, relationships between entities are supplemented, conflict information is corrected, and the hierarchical structure of the knowledge graph is optimized.

[0063] 4. Strategy Iteration and Adaptation: Based on the updated sub-question sequence and the job seeker's response performance, the strategy is continuously optimized through an iterative dynamic mechanism: on the one hand, the sub-question generation logic is adjusted, and on the other hand, the knowledge retrieval strategy is optimized to make the sub-question generation and knowledge retrieval more in line with the actual interview scenario.

[0064] In some embodiments, the content of SAO triples and the sub-problem sequence are updated according to the expanded question bank and the content dynamic mechanism, including: extracting new skill requirements, question type scenarios, and other information from the expanded question bank, parsing and generating new SAO triples, and adding them to the knowledge base; using the dynamic retrieval function of the content dynamic mechanism, associating and matching the new SAO triples with the content of existing SAO triples in the initial knowledge graph, and correcting entities with inconsistent expressions; and generating new sub-problems based on the newly added SAO triples and the core entities of recruitment requirements, and integrating them into the sub-problem sequence.

[0065] In this embodiment of the invention, the process of updating the SAO triplet content and sub-question sequence according to the expanded question bank and content dynamic mechanism specifically includes the following three steps:

[0066] First, extract new information and generate SAO triples.

[0067] From the question bank expanded using incremental web crawling technology, key information is extracted using natural language processing techniques (such as entity recognition and relation extraction):

[0068] Identify new skill requirements (such as "distributed system design" and "A / B testing implementation capabilities"), question types and scenarios (such as "database optimization in high-concurrency scenarios" and "cross-departmental collaboration conflict resolution"), and assessment dimensions (such as "algorithm complexity analysis" and "project risk control").

[0069] This information is parsed into SAO triple structures. For example, "Redis-solve-cache penetration problem" is extracted from "How to use Redis to solve cache penetration problem?" and "user growth project-leader-job seeker" is extracted from "Describe the user growth project you led". These newly generated SAO triples are then added to the knowledge base to enrich the basic content of the knowledge graph.

[0070] Then, perform association matching and entity correction.

[0071] Through the dynamic retrieval function of the content dynamic mechanism, semantic association and consistency verification are performed between the new SAO triples and the existing SAO triples in the initial knowledge graph:

[0072] Based on the semantic similarity calculation of the pre-trained language model, the entities in the new triples (such as "cache penetration problem") are matched with the existing entities in the knowledge graph (such as "data cache anomaly") to identify synonyms or near-synonyms.

[0073] Normalize and correct entities with inconsistent descriptions (e.g., unify "Redis cache optimization" and "distributed cache tuning" to the core entity "caching technology optimization") to ensure consistency of entity terminology in the knowledge graph and avoid ambiguity and redundancy.

[0074] Finally, new subproblems are generated and the sequences are integrated.

[0075] Based on the newly added SAO triples and the core entities of recruitment requirements (such as "big data processing ability" and "team management experience" in job requirements), targeted sub-questions are generated through a large language model:

[0076] For skill-based SAO triples (such as "Hadoop - processing - massive log data"), generate practical questions (such as "When using Hadoop to process TB-level log data, how do you optimize the efficiency of MapReduce tasks?").

[0077] For scenario-based SAO triples (such as "cross-team project - coordination - resource conflict"), generate behavioral description questions (such as "Please give an example of how you coordinate resource conflicts in a cross-team project?").

[0078] The newly generated sub-problems are categorized according to the assessment dimensions (technical skills, soft skills, project experience, etc.), and are subjected to redundancy checks (such as merging duplicate or highly similar questions) and logical sorting (such as adjusting the order according to the progressive relationship of "basic skills → advanced skills → comprehensive scenarios"), ultimately forming an updated sub-problem sequence that ensures comprehensive coverage, logical coherence, and alignment with recruitment needs.

[0079] In some embodiments, the entities and relationships of SAO triples are updated based on the updated SAO triple content, job seekers' multimodal data, and structural dynamics mechanisms. This includes: extracting new SAO triples from job seekers' multimodal data and identifying newly added entities within them; semantically matching the newly added entities with existing entities in the knowledge graph using the entity alignment function of the structural dynamics mechanism; supplementing the relationships between entities based on the updated SAO triple content and correcting conflicting relationships; and adjusting the hierarchical structure of the knowledge graph using an entity tree construction algorithm, setting core skill entities as root nodes and associated project experience and tool usage as child nodes, thereby optimizing the association paths between entities.

[0080] In some embodiments of the present invention, the process of updating the SAO triplet content and sub-question sequence based on the expanded question bank and content dynamic mechanism is achieved through systematic information extraction, association verification, and question generation, as follows:

[0081] For the question bank expanded from industry databases, enterprise recruitment platforms, and other channels through incremental web crawling technology, natural language understanding technology is used for deep analysis. This accurately identifies key elements in the questions, such as new skill requirements (e.g., "microservice architecture design" and "LLM model fine-tuning ability"), question types and scenarios (e.g., "service disaster recovery solutions during e-commerce promotions" and "analysis scenarios of declining user retention rates"), and assessment objectives (e.g., "evaluating system design thinking" and "testing data sensitivity"). The extracted information is then structured according to a "subject-behavior-object" structure. These newly generated triples are then used as foundational content to supplement the knowledge base, enriching the skill and scenario dimensions of the knowledge graph.

[0082] Leveraging the real-time retrieval and matching capabilities of the dynamic content mechanism, semantic alignment is performed between the new SAO triples and existing content in the initial knowledge graph. Specifically, based on pre-trained language models (such as BERT), the semantic similarity between entities in the new triples (e.g., "container status monitoring") and existing entities in the knowledge graph (e.g., "service operation status monitoring" and "containerized deployment and maintenance") is calculated to establish an association mapping. Entities with different expressions but consistent core meanings are normalized (e.g., "LLM fine-tuning" and "large language model parameter tuning" are uniformly associated with the core entity "large language model fine-tuning technology"). Conflicting entities (e.g., "must be proficient in Python" in the test question and "Python is a basic requirement" in the initial graph) are marked, and expressions matching the recruitment requirements are prioritized for retention. This ensures the consistency and accuracy of entity terminology in the knowledge graph and avoids ambiguity in subsequent retrieval and reasoning.

[0083] By combining the newly added SAO triples and the core entities of recruitment requirements (such as "core skills: distributed systems" and "preferred conditions: experience in AI project implementation" in the job description), targeted sub-questions are generated through a large language model and systematically integrated. For skill-based SAO triples (such as "Kubernetes-Orchestration-Container Cluster"), generate technical detail questions (such as "When using Kubernetes to orchestrate container clusters, how do you resolve the conflict between node affinity and resource scheduling?"); for scenario-based SAO triples (such as "AI Project-Implementation-User Growth Scenario"), generate in-depth case study questions (such as "Please explain how you balanced model performance and engineering implementation costs in the user growth scenario of the AI ​​project you led?"); the newly generated sub-questions and the original sequence undergo multi-dimensional processing—removing duplicate questions through text similarity algorithms (such as "How to optimize database performance?" and "What are the methods for database performance tuning?"), and sorting them logically according to "basic ability → professional depth → comprehensive application" (such as first asking "Have you used Kubernetes?", then asking "How to resolve Kubernetes network plugin conflicts?"), ultimately forming a comprehensive, clearly structured, and recruitment-relevant sub-question sequence, providing precise question support for subsequent structured interviews.

[0084] In some embodiments, based on the updated sub-question sequence, job seekers' multimodal data, and an iterative dynamic mechanism, a sub-question generation and retrieval strategy is inferred, including: analyzing the interview answer performance in the job seekers' multimodal data; calculating the relevance of answers to sub-questions, answer completeness, and accuracy of professional terminology using natural language processing technology; and marking inefficient sub-questions; using a reinforcement learning module based on the iterative dynamic mechanism as a reward signal to fine-tune the parameters of the sub-question generation model and optimize the generation weight of sub-questions; dynamically adjusting the retrieval strategy by combining the association strength of entities in the knowledge graph; receiving feedback from the recruiter on the interview results; iteratively correcting the logical structure of the sub-question sequence based on the feedback; supplementing sub-questions in corresponding dimensions; and updating the retrieval benchmark through an entity tree segment filtering mechanism.

[0085] In this embodiment of the invention, a comprehensive evaluation is conducted on job seekers' responses to interviews based on a sequence of sub-questions, using multimodal data parsing and natural language processing techniques. Specifically, a deep structured semantic model is used to calculate the semantic matching degree between the answer and the corresponding sub-question. For example, if the sub-question is "How to optimize the response latency of a distributed system?", and the answer focuses on "single-machine performance tuning," the relevance is low. Through text summarization and keyword coverage analysis, it is determined whether the answer fully covers the core assessment points of the sub-question (e.g., when the sub-question examines "project full-process management," it is checked whether the answer includes requirements analysis, implementation, risk control, etc.). The consistency of technical terms (such as "CAP theorem" and "MVCC mechanism") appearing in the answer with standard terms in the knowledge graph is compared to identify terminology misuse or incorrect expressions. Sub-questions with a relevance score below a threshold (e.g., 0.6), a completeness score of less than 50%, or a low terminology accuracy rate are marked as inefficient questions (e.g., recurring "basic skills description" questions, or questions with low relevance to the core requirements of the position).

[0086] The reinforcement learning module of the iterative dynamic mechanism transforms the above analysis results into reward and punishment signals to optimize the parameters and weights of the sub-problem generation model: positive rewards are given to sub-problems with high relevance (e.g., ≥0.8) and high completeness (e.g., ≥80%), and their generation weights are increased proportionally (e.g., from 1.0 to 1.5); negative rewards are given to inefficient sub-problems, and their generation weights are reduced (e.g., from 1.0 to 0.3); the generation strategy of the large language model is adjusted through reinforcement learning algorithms. For example, for sub-problems of the "distributed system design" category with high-quality responses, the generation weight of advanced questions in the same domain (e.g., "How to design a distributed transaction solution?") is increased; for sub-problems of the "cross-team collaboration" category with weak responses, the scenario-based description of sub-problems in this dimension is strengthened (e.g., asking questions in conjunction with specific conflict cases).

[0087] Based on the strength of associations between entities in the knowledge graph (calculated through entity vector distance using models such as TransE), the priority and scope of knowledge retrieval are optimized in real time. For example, the association strength between "distributed system" and "microservice architecture" is 0.9 (strong association), while the association strength between "distributed system" and "UI design" is 0.2 (weak association). When a sub-question focuses on "distributed system," entity knowledge with an association strength ≥ 0.7 in the knowledge graph (such as "service registration and discovery" and "load balancing algorithm") is retrieved first to support the generation of follow-up questions. For sub-questions marked as inefficient, the retrieval scope is expanded to entities with an association strength of 0.5-0.7 (such as supplementing knowledge on "coordination between distributed system and cloud computing") to help adjust the question wording. During the job seeker's answer, the retrieval benchmark is dynamically updated based on the currently mentioned entity (such as "K8s container orchestration"), prioritizing the retrieval of related knowledge of that entity (such as "container resource limitations" and "pod scheduling strategy") to ensure the relevance of follow-up questions.

[0088] Based on feedback from recruiters regarding interview results (e.g., "insufficient depth of technical questions" or "lack of assessment of resilience"), the sub-question sequence and retrieval benchmark are updated using an entity tree segment filtering mechanism. For feedback regarding "insufficient depth of technical questions," deeper related questions from the "core skills" entity in the knowledge graph are added to the sub-question sequence (e.g., deepening "Do you know Java?" to "How to solve Java memory leak problems?"). For "missing assessment of resilience," sub-questions based on the "behavioral scenarios" entity are added (e.g., "Describe your coping strategies when a project is delayed"). The segment weights of the knowledge graph are readjusted using the entity tree algorithm; for example, the priority of entity segments related to "resilience," such as "crisis management cases" and "team collaboration conflicts," is increased to ensure that relevant knowledge is prioritized for question generation during retrieval. The revised sub-question sequence and retrieval strategy are applied to the next round of interviews, continuously collecting response data and recruiter feedback to form an iterative closed loop of "analysis-optimization-verification," gradually adapting sub-question generation and retrieval strategies to recruitment needs and job seeker characteristics.

[0089] In some embodiments, a structured question set for job seekers is determined based on a multimodal dynamic knowledge base, including: extracting core entities and semantic relationships of job requirements from a knowledge graph through a query-aware entity tree paragraph filtering mechanism; generating a semantically independent sub-question set by using complex interview questions and entity tree-related paragraphs as prompts; and using semantic speech activity detection technology to support real-time interruption and contextual continuation to ensure that the sub-question sequence conforms to natural dialogue logic, thus forming a structured question set.

[0090] In this embodiment of the invention, a query-aware entity tree segment filtering mechanism is adopted based on the knowledge graph of a multimodal dynamic knowledge base to accurately locate core entities and their relationships from job requirements. Using the job title (e.g., "Senior Data Analyst") as the root node, core entities (e.g., "skills," "tools," "project experience," "ability") are extracted from the SAO triples of the recruitment requirements using entity recognition technology as first-level child nodes. Second-level child nodes are then constructed based on the hierarchical relationships between entities (e.g., "skills" is subordinate to "advanced SQL query," "data visualization," and "machine learning modeling"), forming a hierarchical entity tree. For each node in the entity tree, the associated segment information in the knowledge graph is called (e.g., "advanced SQL query" is associated with "window function application scenarios" and "complex query optimization cases"), and an attention mechanism is used to filter segments strongly related to the job requirements. Simultaneously, based on the association strength of entities in the knowledge graph (e.g., the association degree between "data visualization tools" and "Tableau" and "Power BI" is ≥0.8), implicit semantic relationships are mined (e.g., the causal relationship between "skill mastery" and "project application": "Mastering Tableau" → "Supporting the construction of e-commerce data dashboards").

[0091] Based on an entity tree structure and related paragraphs, a large language model is used to break down complex interview questions into a set of semantically independent and logically coherent sub-questions. For open-ended complex questions (such as "Please describe your data analysis work experience in detail"), corresponding sub-questions are generated using the first-level child nodes of the entity tree as the decomposition dimension. For example, based on the "Skills" node, "What are your three most proficient data analysis skills?" is generated; based on the "Project Experience" node, "What is the largest data analysis project you have participated in?" is generated; based on the "Tools" node, "What are the data analysis tools you commonly use and their application scenarios?" is generated. Related paragraphs of the entity tree (such as "Tableau's interactive chart design guidelines") are added as prompts to the input of the large language model to constrain the professionalism and contextualization of the sub-questions. For example, for the "Data Visualization" entity, combined with paragraph information, "When using Tableau to create sales trend charts, how do you optimize the readability of the charts to suit non-technical personnel?" is generated. At the same time, semantic similarity calculation (such as based on cosine similarity) ensures that the sub-questions are semantically independent (similarity ≤ 0.3) and avoids repeated questioning.

[0092] Semantic Voice Activity Detection (SemanticVAD) technology is introduced to adapt the sub-question sequence to natural dialogue, ensuring that the structured question set maintains logical coherence during interaction. Voice Endpoint Detection (VAD) identifies interruption signals from job seekers (e.g., "Wait a minute, I'd like to add some clarification on this question"), while a semantic understanding model determines the interruption intent (question, supplementary information, or request for pause). If it's a supplementary answer, the system pauses the current sub-question push and caches the question sequence; if it's a question (e.g., "Does this skill refer to basic applications or advanced abilities?"), it calls the definition paragraph of the corresponding entity in the knowledge graph (e.g., "Advanced SQL abilities include window functions, stored procedures, etc.") to generate a response and then continues the original question sequence. Based on the semantic vector of the dialogue history (encoded using a BERT model), the semantic relevance between the question and the preceding dialogue is calculated after the interruption ends. Sub-questions with a relevance ≥ 0.6 are prioritized (e.g., if the preceding discussion was "e-commerce data analysis project," the question is followed by "How did you handle missing values ​​in user behavior data in this project?") to avoid topic gaps. The resulting structured question set not only covers the core assessment dimensions of job requirements but also adapts to the dynamic interactive scenarios of natural dialogue, achieving a balance between standardization and flexibility.

[0093] In some embodiments, an interview support strategy for job seekers is determined based on a structured question set, including: using a multi-stream full-duplex architecture to achieve natural dialogue, interacting with job seekers in real time based on the structured question set, dynamically retrieving knowledge base content to answer questions, and using a multimodal thinking chain to evaluate the job seeker's interaction data and output an evaluation report with logical chains.

[0094] In this embodiment of the invention, the multi-stream full-duplex architecture processes voice streams, text streams, and knowledge streams in parallel to achieve real-time interaction and accurate responses during the interview process. The specific mechanism is as follows:

[0095] The architecture includes a semantic speech activity detection module and a parallel processing channel:

[0096] When the system pushes sub-questions according to a structured question set (such as "How do you handle missing values ​​using Python?"), the voice stream channel captures the job seeker's voice input in real time and converts it into a text stream simultaneously; the text stream channel performs real-time semantic parsing through a pre-trained language model (such as BERT) to determine whether the job seeker is answering, interrupting (such as "I would like to explain the project background first"), or asking new questions (such as "What scenario does 'missing value' refer to here?").

[0097] If an interruption is detected, the system immediately pauses the current question push and caches the sequence of sub-questions to be asked through a multi-stream collaborative mechanism, prioritizing the processing of job seekers' input: for interruptions involving supplementary explanations, a temporary buffer channel is opened to record the content and associate it with the corresponding sub-question; for interruptions involving questions, dynamic retrieval of the knowledge flow channel is triggered.

[0098] In this embodiment of the invention, the knowledge flow channel and the multimodal dynamic knowledge base are linked in real time, supporting accurate responses in the following scenarios:

[0099] Answering job seekers' questions: When a job seeker inquires about job details (such as "Does this job involve real-time data processing?"), the system retrieves the SAO triples of the recruitment requirements from the knowledge graph (such as "Job - Responsibility - Real-time data pipeline construction"), and outputs a concise response (such as "Yes, this position requires participation in the daily maintenance of the real-time data processing system"), combined with the natural language generation model.

[0100] Enhance the relevance of questions: During the job seeker's answers (such as mentioning "using Excel to process data"), the system searches the knowledge graph in real time for related entities of "Excel" and job requirements (such as "job requirements - skills - Python data processing"), and generates guiding follow-up questions (such as "Have you tried using Python to replace Excel to improve processing efficiency?"), focusing the dialogue on the core requirements of the job.

[0101] Supplementing contextual information: If a job seeker's answer deviates from the sub-question (e.g., asking about "project duration" but discussing "team composition"), the system retrieves relevant paragraphs from the knowledge base for that sub-question (e.g., "core elements of project duration assessment") and generates subtle guidance (e.g., "Can you explain your role in the context of project duration?") to ensure that the interaction revolves around the assessment dimensions of the structured question set.

[0102] In this embodiment of the invention, the multimodal thinking chain integrates interactive data such as text, voice, and video to construct interpretable evaluation logic. The multimodal interactive data of job seekers is then analyzed in layers:

[0103] Text layer: Calculate the semantic relevance of the answer to the sub-question (e.g., the relevance of "answering Python data processing methods" to the sub-question "how to handle missing values") using a deep structured semantic model (DSSM), extract professional terms (e.g., "Pandas" and "interpolation") and compare them with skill entities in the knowledge graph to calculate terminology accuracy;

[0104] Speech layer: Analyzes speech features (such as speech rate, intonation, and pause frequency) through emotion recognition models to infer the fluency of communication (e.g., a higher score is given if the speech rate is stable and the pauses are reasonable).

[0105] Video layer: If video interaction is included, body language features (such as the frequency of eye contact and the naturalness of gestures) are extracted using computer vision technology as auxiliary indicators for soft skills assessment (such as for scoring the "communication and expression ability" dimension).

[0106] A hierarchical reasoning chain is constructed based on multi-dimensional features, and the final output is an evaluation report with logical chains:

[0107] For each sub-question, an independent reasoning chain is generated. For example: "Professional skills matching score: 75 points → Reason 1: The correlation between 'using Pandas to handle missing values' in the answer and 'job requirements - Python data processing' in the knowledge graph is 0.85 (strong match); Reason 2: 'Real-time data processing experience' in the job requirements was not mentioned (deduct 25 points)"; "Communication fluency score: 80 points → Reason 1: The speech recognition shows that there are no long pauses in the answer (0 pauses of ≥3 seconds); Reason 2: The text is logically coherent (the accuracy of the use of inter-sentence connectors is 85%)."

[0108] Individual indicators are aggregated according to the assessment dimensions (such as "technical ability", "soft skills" and "project experience"), and the total score is calculated by a weighted algorithm (the weights are based on the importance of the core entities in the recruitment requirements). For example: "Total score for technical ability: 78 points → derived by weighting sub-items such as 'Python skills (80 points)' and 'data processing experience (75 points)', with the weights referring to the association strength of the 'core skills' entity in the knowledge graph."

[0109] Output a report with logical chains. The report includes: ① a table showing the correspondence between sub-questions and answers; ② a hierarchical reasoning chain for each score (linked to knowledge graph entity IDs and matching degree data); ③ a marker for items to be verified (e.g., "The resume mentions 'proficient in SQL' but the answer does not mention it; it is recommended to confirm in a second interview"). The final report presents both quantitative scores and explains the scoring basis through a traceable logical chain, providing decision support for recruiters.

[0110] Through the above strategies, the interview assistance process can not only simulate the flexibility of natural conversation, but also ensure the objectivity and interpretability of the results through multimodal assessment, achieving a synergistic effect of "natural interaction, accurate assessment, and transparent decision-making".

[0111] In some embodiments, the method further includes: pushing the assessment report to the recruitment end and receiving refined instructions; pushing improvement suggestions based on SAO differences to candidates; obtaining recruitment decisions and feeding the recruitment decisions back to the iterative dynamic mechanism.

[0112] In some embodiments, to further approximate the natural communication rhythm in real interview scenarios, the interaction mechanism is refined and upgraded based on the original multi-stream full-duplex architecture:

[0113] The dynamic response logic is optimized by using streaming output technology to push model-generated content in segments in real time, avoiding conversation interruptions caused by waiting for complete sentences to be generated. Simultaneously, the response speed is automatically adjusted based on the conversation topic—the pace is appropriately accelerated in technical question answering to match the efficiency of professional communication, while appropriate pauses are maintained in behavioral description questions to simulate the thinking pauses of human interviewers, enhancing the naturalness of the conversation.

[0114] Enhance the collaborative interpretation capabilities of multimodal signals by real-time correlation between facial micro-expressions captured in the video stream (such as frowning and nodding) and intonation changes in the audio stream (such as slowing down speech and emphasizing tone) and text semantics. For example, when the system detects frequent pauses or hesitation in a job seeker's answers, it automatically extends the waiting time or generates guiding prompts (such as "You can start by describing the difficulties at the beginning of the project") to avoid interrupting the train of thought; if the job seeker's tone is firm and the logic is clear, it smoothly moves on to the next question, maintaining a smooth dialogue.

[0115] Upgraded contextual awareness capabilities enable intelligent handling of "real-time interruptions" by constructing a semantic association network of dialogue history. When job seekers add new information to their answers (such as "the tool I just mentioned actually has another application in subsequent projects"), the system can automatically associate this content with the sub-questions currently being discussed and dynamically adjust the direction of subsequent follow-up questions to ensure that the added information is effectively incorporated into the evaluation while maintaining the coherence of the overall dialogue logic.

[0116] In some embodiments, to enhance the knowledge base's adaptability to complex interview scenarios, the dynamic update and knowledge representation logic are further refined based on the original three dynamic mechanisms:

[0117] The content dynamic mechanism adds a "core knowledge anchoring" module. When expanding the question bank through incremental crawling, it simultaneously anchors historical core knowledge (such as essential job skills and industry-standard knowledge). Newly added SAO triples need to undergo semantic consistency verification with the anchored knowledge. If a conflict occurs (such as a significant deviation between the description of a skill in a new question and the industry-standard definition), a secondary verification is initiated—prioritizing based on the core entities of the recruitment requirements to ensure the stability of core knowledge is not affected by new content, while retaining valuable emerging knowledge (such as new industry terminology).

[0118] The structural dynamic mechanism optimizes the hierarchical adaptability of the knowledge graph, employing a layered iterative strategy to adjust entity association paths. Using core skill entities as the root node, the weights of their child nodes (such as tool usage and project experience) are dynamically adjusted based on high-frequency information in the job seeker's multimodal data. If a certain type of project experience frequently appears in recent interviews and is strongly correlated with core skills (e.g., the mention rate of "distributed computing projects" has significantly increased in the "big data processing" position), the association priority of that child node is automatically increased, shortening the retrieval path and ensuring efficient access to high-frequency information.

[0119] Expanding the contextual dimension of knowledge representation, we introduce the "scenario-behavior-outcome" association unit based on the SAO triple. For example, the characterization of "team collaboration ability" not only includes the basic relationship of "job seeker-coordinator-cross-departmental team", but also associates it with specific scenarios (such as "project delays"), behavioral details (such as "organizing daily synchronization meetings" and "prioritizing tasks"), and actual results (such as "delivering 3 days ahead of schedule"). This allows the knowledge graph to more accurately support in-depth assessment of complex abilities, rather than just remaining at the description of a single behavior.

[0120] In some embodiments, to improve the relevance of the problem and the objectivity of the evaluation results, the generation logic and evaluation system are refined:

[0121] The generation of structured question sets incorporates a "competency-dimensional hierarchical" strategy. Based on the hierarchical relationship of core entities in job requirements, a progressive question sequence of "basic knowledge - professional application - comprehensive innovation" is constructed. For example, regarding "data analysis ability," basic questions (such as "What data cleaning methods do you commonly use?") are used to verify knowledge reserves, followed by application-oriented questions (such as "How to use these methods to process unstructured data?") to examine practical experience. Finally, comprehensive questions (such as "How would you verify data conclusions when they conflict with business intuition?") assess in-depth thinking ability. Simultaneously, information gain analysis is used to filter for discriminatory questions, avoiding repetitive or inefficient questioning (e.g., reducing the proportion of basic conceptual questions for experienced job seekers).

[0122] The multimodal thinking chain assessment model adds a "traceable reasoning node" mechanism, breaking down the assessment dimensions into clear logical chains. Taking the assessment of "technical ability" as an example, the chain includes: ① the degree of matching between professional terminology and standard terminology in the knowledge graph; ② the relevance of the answer content to the core points of the question; ③ the scenario adaptability of the solution (such as "whether the mentioned algorithm is applicable to the business scenario described in the problem description"). Each node corresponds to specific interview performance evidence (such as the matching record between the "decision tree algorithm" mentioned by the job seeker and the "classification task" in the job requirements), making the assessment conclusion traceable layer by layer and avoiding ambiguous judgments.

[0123] To strengthen the fairness verification process, a semantic filtering module is used during question generation to eliminate potentially biased statements (such as avoiding descriptions of personal background unrelated to ability). In the evaluation phase, a "feature balancing" strategy is introduced to balance objective indicators (such as accuracy of technical terminology) and subjective performance (such as fluency of communication) in multimodal data, preventing a single dimension from excessively influencing the results. Simultaneously, the evaluation model is regularly subjected to "bias detection"—by analyzing the distribution of evaluation results across different groups, if an unreasonable correlation is found between a certain feature (such as expression style) and the score, the model weights are dynamically adjusted to ensure the objectivity of the evaluation.

[0124] In some embodiments, a fully automated interview assistance system based on a large language model includes:

[0125] The acquisition module is used to acquire recruitment requirements, initial interview questions, and multimodal data of job seekers from the recruitment platform.

[0126] A construction module is used to build a multimodal dynamic knowledge base based on the recruitment requirements, the initial interview questions, and the multimodal data. The multimodal dynamic knowledge base includes a content dynamic mechanism, a structure dynamic mechanism, and an iterative dynamic mechanism. The content dynamic mechanism is used to update the knowledge base content and sub-question sequences. The structure dynamic mechanism is used to update the structure of the knowledge graph in the multimodal dynamic knowledge base. The iterative dynamic mechanism is used to optimize the sub-question generation and retrieval strategies.

[0127] The retrieval module is used to determine the job seeker's structured question set based on the multimodal dynamic knowledge base;

[0128] The determination module is used to determine the interview support strategies for job seekers based on the structured question set.

[0129] The fully automated interview assistance system based on a large language model provided in this embodiment can be used to execute the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again in this embodiment.

[0130] In some embodiments, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the embodiments of the full-process automated interview assistance methods based on large language models described above.

[0131] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0132] The memory can be an internal storage unit of the control device, such as the hard drive or RAM. It can also be an external storage device, such as a plug-in hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units. The memory is used to store computer programs and other programs and data required by the terminal. It can also be used to temporarily store data that has been output or will be output.

[0133] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0134] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A fully automated interview assistance method based on a large language model, characterized in that, include: Obtain multimodal data on recruitment needs, initial interview questions, and job seekers from the recruitment platform; Based on the recruitment requirements, the initial interview questions, and the multimodal data, a multimodal dynamic knowledge base is constructed. This multimodal dynamic knowledge base includes a content dynamic mechanism, a structure dynamic mechanism, and an iterative dynamic mechanism. The content dynamic mechanism is used to update the knowledge base content and sub-question sequences. The structure dynamic mechanism is used to update the structure of the knowledge graph in the multimodal dynamic knowledge base. The iterative dynamic mechanism is used to optimize the sub-question generation and retrieval strategies. Based on the recruitment requirements, the initial interview questions, and the multimodal data, a multimodal dynamic knowledge base is constructed, including: Obtain the initial knowledge graph; The question bank is expanded based on the recruitment requirements, the initial interview questions, and incremental web crawling technology. New skill requirements, question types, and scenarios are extracted from the expanded question bank, and new SAO triples are generated and added to the knowledge base. The SAO triples are subject-behavior-object triples. Through the dynamic retrieval function of the content dynamic mechanism, the new SAO triples are associated and matched with the content of existing SAO triples in the initial knowledge graph, and entities with inconsistent descriptions are corrected. Based on the new SAO triples and the core entities of recruitment needs, new sub-problems are generated and integrated into the sub-problem sequence; Extract new SAO triples from job seekers' multimodal data and identify newly added entities within them; The entity alignment function of the structural dynamic mechanism is used to semantically match the new entity with the existing entity in the knowledge graph. Based on the new SAO triple content, the relationships between entities are supplemented and conflicting relationships are corrected; The knowledge graph hierarchy is adjusted by using an entity tree construction algorithm, with core skill entities set as root nodes and related project experience and tool usage set as child nodes, thus optimizing the association paths between entities. We analyzed the interview responses of job seekers in multimodal data and used natural language processing technology to determine the analysis results, including: the relevance of the answers to sub-questions, the completeness of the answers, and the accuracy of professional terminology. The reinforcement learning module based on the iterative dynamic mechanism uses the above analysis results as a reward signal to fine-tune the parameters of the sub-problem generation model and optimize the generation weights of the sub-problems. The retrieval strategy is dynamically adjusted based on the strength of the associations between entities in the knowledge graph. Receive feedback from the recruitment end on the interview results, iteratively revise the logical structure of the sub-question sequence based on the feedback, supplement sub-questions of the corresponding dimensions, and update the retrieval benchmark through the entity tree segment filtering mechanism; Based on the multimodal dynamic knowledge base, determine the job seeker's structured question set; Based on the structured question set, determine the interview support strategies for job seekers.

2. The fully automated interview assistance method based on a large language model according to claim 1, characterized in that, Based on the aforementioned multimodal dynamic knowledge base, a set of structured questions for job seekers is determined, including: By using a query-aware entity tree segment filtering mechanism, the core entities and semantic relationships of job requirements are extracted from the knowledge graph. By using complex interview questions and paragraphs associated with entity trees as prompts, semantically independent sets of sub-questions are generated. Semantic speech activity detection technology is used to support real-time interruption and contextual continuation, ensuring that the sequence of sub-questions conforms to the logic of natural dialogue and forms a structured question set.

3. The fully automated interview assistance method based on a large language model according to claim 1, characterized in that, Based on the structured question set, determine the job seeker's interview support strategies, including: It adopts a multi-stream full-duplex architecture to achieve natural dialogue, interacts with job seekers in real time based on a structured question set, and dynamically retrieves knowledge base content to answer questions; We use a multimodal thinking chain to evaluate job seekers' interaction data and output an assessment report with logical chains.

4. The fully automated interview assistance method based on a large language model according to claim 3, characterized in that, The method further includes: Push the assessment report to the recruitment end and receive detailed instructions; push improvement suggestions based on SAO differences to candidates; Obtain recruitment decisions and feed them back to the iterative dynamic mechanism.

5. A fully automated interview assistance system based on a large language model, characterized in that, include: The acquisition module is used to acquire recruitment requirements, initial interview questions, and multimodal data of job seekers from the recruitment platform. A construction module is used to build a multimodal dynamic knowledge base based on the recruitment requirements, the initial interview questions, and the multimodal data. The multimodal dynamic knowledge base includes a content dynamic mechanism, a structure dynamic mechanism, and an iterative dynamic mechanism. The content dynamic mechanism is used to update the knowledge base content and sub-question sequences. The structure dynamic mechanism is used to update the structure of the knowledge graph in the multimodal dynamic knowledge base. The iterative dynamic mechanism is used to optimize the sub-question generation and retrieval strategies. The retrieval module is used to determine the job seeker's structured question set based on the multimodal dynamic knowledge base; The determination module is used to determine the interview support strategies for job seekers based on the structured question set; Build modules, specifically used for: Obtain the initial knowledge graph; The question bank is expanded based on the recruitment requirements, the initial interview questions, and incremental web crawling technology. New skill requirements, question types, and scenarios are extracted from the expanded question bank, and new SAO triples are generated and added to the knowledge base. The SAO triples are subject-behavior-object triples. Through the dynamic retrieval function of the content dynamic mechanism, the new SAO triples are associated and matched with the content of existing SAO triples in the initial knowledge graph, and entities with inconsistent descriptions are corrected. Based on the new SAO triples and the core entities of recruitment needs, new sub-problems are generated and integrated into the sub-problem sequence; Extract new SAO triples from job seekers' multimodal data and identify newly added entities within them; The entity alignment function of the structural dynamic mechanism is used to semantically match the new entity with the existing entity in the knowledge graph. Based on the new SAO triple content, the relationships between entities are supplemented and conflicting relationships are corrected; The knowledge graph hierarchy is adjusted by using an entity tree construction algorithm, with core skill entities set as root nodes and related project experience and tool usage set as child nodes, thus optimizing the association paths between entities. We analyzed the interview responses of job seekers in multimodal data and used natural language processing technology to determine the analysis results, including: the relevance of the answers to sub-questions, the completeness of the answers, and the accuracy of professional terminology. The reinforcement learning module based on the iterative dynamic mechanism uses the above analysis results as a reward signal to fine-tune the parameters of the sub-problem generation model and optimize the generation weights of the sub-problems. The retrieval strategy is dynamically adjusted based on the strength of the associations between entities in the knowledge graph. The system receives feedback from the recruitment end regarding the interview results, iteratively modifies the logical structure of the sub-question sequence based on the feedback, supplements sub-questions for corresponding dimensions, and updates the retrieval benchmark through an entity tree segment filtering mechanism.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the full-process automatic interview assistance method based on a large language model as described in any one of claims 1 to 4.

Images