A Meeting Information Question Answering Method and System Based on Retrieval Enhancement Generation and Knowledge Graph

Abstract

This paper describes a method, system, computer device, and computer-readable storage medium for question answering in conference settings based on retrieval enhancement generation and knowledge graphs. The embodiments of this application significantly improve the accuracy and domain adaptability of question answering in conference scenarios by constructing a full-link technical framework encompassing domain-adaptive intent understanding, hybrid index retrieval, knowledge graph reasoning enhancement, reordering context optimization, and structured answer generation. This framework enables deep, cross-modal knowledge association and intelligent reasoning, while optimizing answer generation quality and user experience.

Description

Technical Field

[0001] This invention relates to the field of intelligent retrieval technology, and in particular to a method, system, computer device, and computer-readable storage medium for meeting information question answering based on retrieval enhancement generation and knowledge graph. Background Technology

[0002] In recent years, the combination of retrieval-enhanced generative techniques and knowledge graphs has provided a new paradigm for the development of intelligent question-answering systems. These general-purpose systems are typically pre-trained and built based on large-scale public corpora, capable of handling a wide range of open-ended questions, and demonstrating a certain level of information understanding and generation capabilities in general domains.

[0003] However, when such general frameworks are directly applied to the highly specialized, structured, and scenario-specific field of meeting information management and question-and-answer, their inherent limitations become apparent, mainly in the following aspects:

[0004] First, question-answering engines suffer from a severe lack of domain adaptability. Existing general-purpose knowledge base question-answering systems primarily train their core models (such as intent classification, semantic understanding, and text generation models) on publicly available, written general corpora, failing to deeply optimize for the unique language patterns and knowledge characteristics of meeting scenarios. Meeting discussions are generally characterized by frequent colloquial expressions (such as omissions, pronouns, and catchphrases), a high density of domain-specific terminology (such as specific project codes, internal process abbreviations, and technical jargon), and complex logical connections and strong contextual dependencies arising from multi-speaker interactions. General-purpose systems struggle to understand these features, directly resulting in low intent recognition accuracy in meeting scenarios and insufficient recall of answers to implicit information, making it difficult to meet users' needs for accurate and complete meeting information retrieval.

[0005] Secondly, there is a lack of deep-level knowledge association and tracing capabilities. Meeting knowledge is typically not isolated but distributed in a multimodal and fragmented manner across multiple data sources, such as audio recordings, automatically transcribed text, structured decision minutes, and to-do lists, with rich semantic connections between these data. Most general question-answering systems lack the modeling and reasoning capabilities for such complex connections, failing to effectively establish cross-modal and cross-document semantic relationship networks. Therefore, the system often only provides isolated answer fragments. To verify answers or obtain more comprehensive background information, users must leave the question-answering interface and manually trace back to the original meeting recordings, text records, or related documents, creating a secondary operational cost of "question-answer first, then tracing back," severely impacting the efficiency and experience of information acquisition.

[0006] In conclusion, how to construct an intelligent question-answering method that can accurately understand the intentions of conversational and domain-specific questions, deeply associate and reason about multimodal meeting knowledge, and ultimately provide accurate, complete, and efficiently traceable answers, has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a meeting information question-answering method, system, computer device, and computer-readable storage medium based on retrieval-enhanced generation and knowledge graphs to solve one or more technical problems in the prior art.

[0008] To solve one or more of the above-mentioned technical problems, the technical solution adopted in this application is:

[0009] Firstly, this application provides a meeting information question-answering method based on retrieval enhancement generation and knowledge graph, the method comprising:

[0010] The system receives user query information, uses a pre-trained classification model to obtain the intent classification of the user query information, and uses semantic role labeling to extract the core entities and constraints in the user query. Based on the intent classification and the core entities and constraints, a structured query vector is generated.

[0011] Based on the structured query vector, a hybrid retrieval is performed using a pre-constructed composite index structure to obtain a set of candidate knowledge units. The composite index structure includes a vector index constructed based on multi-dimensional vectors of standardized atomic knowledge units. The hybrid retrieval includes at least vector retrieval based on the vector index, keyword retrieval based on keyword matching, and fusion of vector retrieval results and keyword retrieval results with knowledge graph reasoning.

[0012] A reordering model is used to perform a secondary sorting of the knowledge units in the candidate knowledge unit set, select the top-ranked candidate knowledge units, and construct a context structure based on semantic logic and temporal order.

[0013] A preliminary answer is generated based on a preset contextualized prompt word template and the context structure. The preliminary answer is then post-processed to obtain a structured answer with source information. The post-processing includes formatting, filtering redundant information, and adding citation sources.

[0014] This application embodiment significantly improves the accuracy and domain adaptability of question answering in meeting scenarios by constructing a full-link technical framework that includes domain-adaptive intent understanding, hybrid retrieval with composite indexes, enhanced knowledge graph reasoning, reordering context optimization, and structured answer generation. It achieves deep, cross-modal knowledge association and intelligent reasoning, and optimizes the quality of answer generation and user experience.

[0015] In some embodiments, receiving user query information, using a pre-trained classification model to obtain the intent classification of the user query information, and using semantic role labeling to extract core entities and constraints in the user query, and generating a structured query vector based on the intent classification and the core entities and constraints, includes:

[0016] Receive user questions, including meeting-related information;

[0017] A pre-trained classification model is invoked to perform intent recognition on the user's query information to obtain intent classification, wherein the intent classification includes at least one of query action items, confirmation decisions, explanation terms, and tracing back discussions;

[0018] Semantic role labeling is performed on the user question information to identify and extract the core entities and constraints in the user question information. The core entities include at least one of the following: attendees, project name, document name, and meeting topic. The constraints include at least one of the following: time range, priority status, and task status.

[0019] The intent classification, the core entity, and the constraints are structurally represented and vectorized to generate a structured query vector.

[0020] In this embodiment, by constructing a meeting-specific understanding and representation mechanism, accurate intent capture and information structuring are achieved at the beginning of the question-and-answer process, laying a solid foundation for subsequent efficient and accurate retrieval and generation.

[0021] In some embodiments, the step of performing a hybrid retrieval based on the structured query vector using a pre-constructed composite index structure to obtain a set of candidate knowledge units includes:

[0022] By calling a pre-built vector index, the semantic similarity between the structured query vector and the multi-dimensional vectors of each standardized atomic knowledge unit stored in the vector index is calculated to obtain a first subset of candidate knowledge units. By calling a pre-built inverted index, keyword matching is performed based on the core entities and constraints contained in the structured query vector to obtain a second subset of candidate knowledge units.

[0023] The first subset of candidate knowledge units and the second subset of candidate knowledge units are fused together to obtain a set of candidate knowledge units.

[0024] This application embodiment effectively solves the limitations of a single retrieval mode in dealing with the complexity of conference knowledge by using a multi-strategy parallel and fusion mechanism, and achieves a significant improvement in the three dimensions of retrieval results: accuracy, recall, and knowledge completeness.

[0025] In some embodiments, the step of fusing the first subset of candidate knowledge units and the second subset of candidate knowledge units to obtain a set of candidate knowledge units includes:

[0026] The first subset of candidate knowledge units and the second subset of candidate knowledge units are deduplicated to obtain a set of fused knowledge units;

[0027] Calculate the overall score for each knowledge unit in the fused knowledge unit set;

[0028] The knowledge units in the fused knowledge unit set are sorted according to the comprehensive score, and the top-ranked knowledge units are selected to form a preliminary candidate knowledge unit set.

[0029] The knowledge units in the preliminary candidate knowledge unit set are supplemented and verified to obtain the candidate knowledge unit set.

[0030] In this embodiment of the application, by systematically integrating and optimizing the results of different retrieval strategies, the overall quality and credibility of the candidate knowledge unit set are significantly improved, laying a solid foundation for the subsequent generation of high-quality answers.

[0031] In some embodiments, calculating the comprehensive score of each knowledge unit in the fused knowledge unit set includes:

[0032] If a knowledge unit appears only in the first subset of candidate knowledge units, its comprehensive score is the semantic similarity score of the knowledge unit in vector retrieval multiplied by the first weight coefficient.

[0033] If a knowledge unit appears only in the second subset of candidate knowledge units, its overall score is the matching score of that knowledge unit in keyword retrieval multiplied by the second weighting coefficient.

[0034] If a knowledge unit appears in both the first subset of candidate knowledge units and the second subset of candidate knowledge units, its comprehensive score is the sum of the semantic similarity score multiplied by the first weight coefficient and the matching score multiplied by the second weight coefficient.

[0035] In this embodiment of the application, by introducing a weighting mechanism and a differentiated scoring strategy into the comprehensive score calculation rules, intelligent and refined evaluation and integration of retrieval results are achieved, effectively improving the scientific nature of the candidate knowledge unit ranking and the relevance of the final answer.

[0036] In some embodiments, the supplementary verification process for the knowledge units in the preliminary candidate knowledge unit set to obtain the candidate knowledge unit set includes:

[0037] Verify that the metadata of each knowledge unit is consistent with the constraints in the user's question information, and remove inconsistent knowledge units; and / or,

[0038] Verify that the confidence level of each knowledge unit is not lower than a preset threshold, and remove knowledge units with a confidence level lower than the preset threshold.

[0039] In this embodiment of the application, by implementing strict quality and consistency control, the candidate knowledge unit set is effectively purified, and the data reliability of the input downstream answer generation process is significantly improved, thereby ensuring the accuracy and credibility of the final answer.

[0040] In some embodiments, the step of performing a hybrid retrieval based on the structured query vector using a pre-constructed composite index structure to obtain a candidate knowledge unit set further includes:

[0041] The preliminary set of candidate knowledge units is input into a pre-constructed knowledge graph;

[0042] In the knowledge graph, starting from the knowledge units in the preliminary candidate knowledge unit set, reasoning and expansion are performed along a preset semantic relationship path to obtain extended knowledge units associated with the knowledge units in the preliminary candidate knowledge unit set.

[0043] The extended knowledge units are added to the preliminary candidate knowledge unit set to form the candidate knowledge unit set.

[0044] This application embodiment significantly enhances the completeness of the candidate knowledge unit set and the depth of the answer by mining deep semantic relationships between knowledge, effectively solving the limitations of traditional retrieval methods in dealing with complex and implicit information needs.

[0045] In some embodiments, the step of using a reordering model to perform a secondary sorting of the knowledge units in the candidate knowledge unit set, selecting the top-ranked candidate knowledge units, and constructing a context based on semantic logic and temporal order includes:

[0046] A re-ranking model is used to perform a secondary relevance score on all knowledge units in the candidate knowledge unit set. During the scoring process, the re-ranking model dynamically enhances the feature weights related to the time constraints and speaker relevance in the user's question information.

[0047] Based on the secondary scoring results, all knowledge units are sorted in descending order, and the top N knowledge units are selected as candidate knowledge units, where N is a preset positive integer.

[0048] In this embodiment, by performing refined scoring and intelligent organization of candidate knowledge units, the information quality and structure of the input generation model are significantly optimized, providing a key guarantee for generating high-quality and highly relevant answers.

[0049] In some embodiments, constructing the context based on semantic logic and temporal order includes:

[0050] Extract the metadata of the candidate knowledge units, wherein the metadata includes at least a meeting unique identifier, a timestamp, and a confidence level;

[0051] Construct a semantic logic sequence based on the semantic correlation between the candidate knowledge units;

[0052] Based on the semantic logic sequence, each candidate knowledge unit is arranged in chronological order according to its timestamp to form a context.

[0053] The context and the corresponding metadata are encapsulated together into a context structure.

[0054] In this embodiment of the application, by organizing the selected knowledge units in both logical and temporal dimensions, a clear and complete context is generated, providing a solid foundation for generating high-quality and traceable answers.

[0055] In some embodiments, generating a preliminary answer based on a preset contextualized prompt word template and the context structure includes:

[0056] A pre-trained large language model is invoked to generate an initial answer based on meeting knowledge constraints and a preset scenario-based prompt word template; wherein, the scenario-based prompt word template includes at least professional role settings, context injection rules and structured output format requirements.

[0057] This application embodiment effectively guides a large language model to generate preliminary answers that are both professional and accurate, as well as formatted correctly, by combining domain knowledge constraints with structured prompting engineering depth, thus significantly improving the quality and usability of question-and-answer results.

[0058] In some embodiments, the post-processing of the preliminary answer to obtain a structured answer with source information includes:

[0059] The preliminary answer is formatted and redundant information is filtered out to form a refined answer text;

[0060] Based on the metadata of the candidate knowledge units used to construct the context, source citations are added to the key information in the refined answer text to generate a structured answer with source tracing information;

[0061] The cited source includes at least the meeting unique identifier, timestamp, and confidence level of the corresponding knowledge unit, and supports direct location of the original knowledge unit content through the cited source.

[0062] In this embodiment, by refining the generated answer and fusing it with in-depth source information, the initial answer is transformed into a credible and verifiable structured result that can be directly delivered to users, thus achieving the final leap in the practical value of the question-and-answer system.

[0063] In some embodiments, the method further includes:

[0064] Based on the confidence level of the candidate knowledge units and the consistency between the preliminary answer and the context structure, the overall confidence level of the structured answer is calculated using a weighted fusion algorithm; and / or,

[0065] Based on the context and semantic associations in the knowledge base, several extended questions related to the user's question information are generated, and the extended questions are output together with the structured answer.

[0066] In this embodiment of the application, by introducing answer confidence assessment and intelligent question answering extension, the intelligence level and interactive experience of the system are significantly improved.

[0067] In some embodiments, the method further includes a multi-turn dialogue support step, including:

[0068] Cache the dialogue history context through a dialogue state tracking mechanism;

[0069] In subsequent user questions, the system automatically links entity information from previous questions to resolve cross-question references and continue intent.

[0070] In this embodiment of the application, by maintaining a coherent dialogue state and context, the system's understanding depth and response coherence in real and complex interaction scenarios are significantly improved, achieving a leap from single-turn question answering to intelligent continuous dialogue capabilities.

[0071] In a second aspect, this application also provides a conference information question-answering system based on retrieval-enhanced generation and knowledge graph, the system being used to implement the conference information question-answering method based on retrieval-enhanced generation and knowledge graph as described in any of the first aspects, the system comprising:

[0072] The vector generation module is configured to receive user query information, use a pre-trained classification model to obtain the intent classification of the user query information, use semantic role labeling to extract the core entities and constraints in the user query, and generate a structured query vector based on the intent classification and the core entities and constraints.

[0073] The hybrid retrieval and reasoning module is configured to perform hybrid retrieval based on the structured query vector using a pre-built composite index structure to obtain a set of candidate knowledge units. The composite index structure includes a vector index constructed based on multi-dimensional vectors of standardized atomic knowledge units. The hybrid retrieval includes at least vector retrieval based on the vector index, keyword retrieval based on keyword matching, and fusion of vector retrieval results and keyword retrieval results with knowledge graph reasoning.

[0074] The context building module is configured to use a reordering model to perform a secondary sorting of the knowledge units in the candidate knowledge unit set, select the top-ranked candidate knowledge units, and build a context structure based on semantic logic and temporal order.

[0075] The answer generation and source tracing module is configured to generate a preliminary answer based on a preset scenario-based prompt word template and the context structure, and to perform post-processing on the preliminary answer to obtain a structured answer with source tracing information. The post-processing includes formatting, filtering redundant information, and adding citation sources.

[0076] In a third aspect, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and when the computer program is executed by the processor, implements the meeting information question answering method based on retrieval enhancement generation and knowledge graph as described in any of the first aspects.

[0077] In a fourth aspect, this application also provides a computer-readable storage medium storing a computer program that, when executed, implements the meeting information question-answering method based on retrieval enhancement generation and knowledge graph as described in any of the first aspects.

[0078] It is understood that the features and aspects of this application described above with respect to various aspects of this application are equally applicable to embodiments of this application according to other aspects of this application, and can be suitably combined with these embodiments, and not only in the specific combinations described herein. Attached Figure Description

[0079] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein:

[0080] Figure 1 The flowcharts of some embodiments of the present application provide a meeting information question answering method based on retrieval enhancement generation and knowledge graph;

[0081] Figure 2 The diagram shows a schematic representation of the structure of a meeting information question-answering system based on retrieval enhancement generation and knowledge graph provided in some embodiments of this application;

[0082] Figure 3 A schematic diagram of the computer device is shown. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0084] As described in the background section, the general framework used in existing intelligent question-answering systems suffers from technical problems such as insufficient domain adaptability of the question-answering engine and lack of deep knowledge association and tracing capabilities when applied to highly specialized, structured, and scenario-specific fields such as conference information management and question answering.

[0085] Based on this, this application presents a novel question-answering method for meeting information based on retrieval enhancement generation and knowledge graph. This method significantly improves the accuracy and domain adaptability of question answering in meeting scenarios by constructing a full-link technical framework that includes domain-adaptive intent understanding, hybrid index retrieval, knowledge graph reasoning enhancement, reordering context optimization, and structured answer generation. It achieves deep, cross-modal knowledge association and intelligent reasoning, and optimizes the quality of answer generation and user experience.

[0086] The present application will be described in detail below through specific embodiments.

[0087] Example 1

[0088] Figure 1 This is a flowchart of a meeting information question-answering method based on retrieval enhancement generation and knowledge graph provided in some embodiments of this application, with reference to... Figure 1 As shown, the method includes the following steps:

[0089] S100: Receive user query information, use a pre-trained classification model to obtain the intent classification of the user query information, and use semantic role labeling to extract the core entities and constraints in the user query, and generate a structured query vector based on the intent classification and the core entities and constraints.

[0090] Specifically, the system first receives user-submitted questions related to topics such as meetings. Then, a classification model specifically tuned for meeting scenarios is invoked to identify the intent of the question and categorize it into a predefined intent category, such as "querying action items," "confirming decisions," "explaining terminology," or "tracing back to discussion." This clarifies the fundamental purpose of the user's question. Simultaneously, Semantic Role Labeling (SRL) technology can be used to perform deep semantic analysis on the user's question information, identifying and extracting key information elements, including but not limited to core entities (such as participants, project names, and meeting topics) and constraints (such as timeframes, task status, priorities, and other limiting information). Finally, the identified intent categories, core entities, and constraints are integrated and encoded to generate a unified structured query vector. This vector not only contains the literal information of the original question but also integrates its deep semantic structure and query intent, providing a precise and computable query blueprint for subsequent knowledge retrieval, fundamentally laying the foundation for accurate question-and-answer processing.

[0091] By combining intent classification achieved through a pre-trained classification model with semantic role labeling (SRL), not only is user intent (such as querying action items or tracing discussions) understood with high precision, but unstructured natural language queries are also accurately deconstructed into structured query vectors containing core entities and constraints. This process transforms vague user needs into precise retrieval instructions that machines can execute, laying a solid foundation for subsequent efficient and accurate retrieval and fundamentally avoiding the ambiguity and information loss problems of traditional keyword retrieval.

[0092] S200: Based on the structured query vector, a hybrid retrieval is performed using a pre-constructed composite index structure to obtain a set of candidate knowledge units. The composite index structure includes a vector index constructed based on multi-dimensional vectors of standardized atomic knowledge units. The hybrid retrieval includes at least vector retrieval based on the vector index, keyword retrieval based on keyword matching, and fusion of vector retrieval results and keyword retrieval results with knowledge graph reasoning.

[0093] It should be noted that this step aims to efficiently and comprehensively retrieve the most relevant information fragments to user questions from a large-scale conference knowledge base. Its core lies in utilizing a composite index structure and executing a hybrid retrieval process that integrates multiple retrieval strategies and knowledge reasoning. First, a composite index structure can be pre-constructed. The core foundation of this composite index structure is the Standardized Atomic Knowledge Unit (AKU)—an independent knowledge carrier formed by standardizing and structurally encapsulating the original conference content (such as audio clips or text paragraphs). Multi-dimensional semantic vectors generated for each AKU constitute a vector index, used to support deep semantic matching. The hybrid retrieval process unfolds in parallel on this composite index structure using vector retrieval and keyword retrieval. Subsequently, the results of these two retrieval paths are fused and subjected to knowledge graph reasoning. Ultimately, through this multi-level, hybrid retrieval mechanism of "vector retrieval + keyword retrieval + knowledge reasoning," the limitations of single retrieval methods can be overcome, resulting in a set of candidate knowledge units that is both accurate and comprehensive, possessing both explicit matching and implicit associations. This lays a rich and reliable material foundation for the subsequent generation of accurate answers.

[0094] As can be understood, an Atomic Knowledge Unit (AKU) is a data structure and knowledge representation unit, serving as the basic unit for a system to process, organize, and manage raw, unstructured meeting content. Specifically, an AKU is the smallest, complete, independent, and identifiable knowledge fragment extracted and encapsulated by the system from raw meeting materials (such as audio streams, transcribed text, and meeting minutes documents) through automated or semi-automated processes.

[0095] It should be noted that the hybrid retrieval in this embodiment is not a simple superposition of parallel retrieval results. Instead, it ensures semantic similarity through vector retrieval, guarantees key entity matching through keyword retrieval, and innovatively introduces knowledge graph reasoning to complete and logically verify the preliminary results. For example, when querying the "follow-up actions" of a project, this method can not only find directly related task records, but also associate relevant responsible persons, prior discussions, and dependent documents through the knowledge graph, achieving a leap from "point-like answers" to "networked knowledge," greatly improving the completeness and depth of information association.

[0096] S300: The knowledge units in the candidate knowledge unit set are sorted a second time using a reordering model, and the top-ranked candidate knowledge units are selected. A context structure is then constructed based on semantic logic and temporal order.

[0097] This step aims to refine and optimize the candidate knowledge unit set obtained from the initial retrieval, which may still contain noise or have uneven relevance levels, in order to construct the contextual information most suitable for generating high-quality answers. Specifically, firstly, a re-ranking model is used to perform a secondary relevance score on all knowledge units in the candidate knowledge unit set. This model is typically based on a more complex interactive architecture (such as BGE Reranker v2-m3), capable of measuring the semantic matching degree between the query and each knowledge unit more meticulously than the calculation in the initial retrieval stage. During this process, the model pays special attention to and enhances the feature weights related to key constraints in the user's question, such as time constraints and speaker relevance, ensuring that the ranking results better reflect the actual needs of the meeting scenario. Then, all candidate knowledge units are sorted in descending order based on the secondary scoring results, and the top-ranked units (e.g., Top-5 or Top-10) are selected. These units are considered the most relevant and reliable core knowledge fragments to the user's question across multiple dimensions, including semantics, time, and role. Finally, a contextual structure is constructed based on these candidate knowledge units according to semantic logic and temporal order.

[0098] S400: Generate a preliminary answer based on a preset scenario-based prompt word template and the context structure, and perform post-processing on the preliminary answer to obtain a structured answer with source information. The post-processing includes formatting, filtering redundant information, and adding citation sources.

[0099] This step is the final output of the question-and-answer process. Its goal is to transform the optimized and organized contextual information into a professional and reliable final answer that users can directly use. In practice, a pre-trained large language model can be invoked, and a preliminary answer can be generated based on a preset scenario-based prompt template. This template is an instruction framework specifically designed for meeting scenarios (especially in industries such as healthcare). Subsequently, a series of post-processing operations are performed on this preliminary answer, ultimately outputting a structured answer with complete traceability information. This structured answer not only provides clear and accurate conclusions but also ensures that every piece of information is traceable to a specific meeting source, thus meeting the stringent requirements for reliability, auditability, and high readability in professional scenarios while providing intelligent question-and-answer services.

[0100] By employing a reordering model to perform a refined secondary sorting of massive candidate knowledge units and constructing a contextual structure based on semantic logic and temporal order, the information input to the generative model is ensured to be the most relevant, coherent, and chronologically logical. Building upon this, pre-defined scenario-based prompts guide the generative model (such as a large language model) to produce preliminary answers that better align with conference domain norms and user intent. Finally, through a post-processing workflow including formatting, redundancy filtering, and adding source citations, the final product delivered to the user is a reliable answer with a clear structure, concise information, and every key information point traceable to the original conference knowledge source. This effectively solves the common "illusion" problem in generative models and greatly enhances the credibility and usability of the answers.

[0101] In some embodiments, a hybrid model architecture based on BERT-LSTM can be employed, with domain-specific fine-tuning to achieve high-precision intent classification. Specifically, this model fully leverages the powerful capabilities of BERT (Bidirectional Encoder Representations from Transformers) in deep semantic representation and contextual understanding, pre-encoding the input user questions to effectively capture key semantic information and subtle intent differences. To further enhance the model's ability to capture and judge intent classification sequence features, an LSTM (Long Short-Term Memory) layer can be connected after the BERT output layer. LSTM can further perform time-series modeling and information integration on the context-related features extracted by BERT, and is particularly suitable for analyzing possible dependencies and coherent logic of intent expression in user questions, thereby improving the stability and accuracy of classification. In practical applications, the hybrid model can be supervisedly fine-tuned using a large amount of labeled meeting domain question-and-answer data to specifically adapt it to distinguishing multiple core meeting intents, including querying action items, confirming decisions, explaining terminology, and tracing back discussions. Experiments have shown that the finely tuned BERT-LSTM model performs excellently in intent classification tasks in meeting scenarios, with an accuracy rate that can stably reach over 94%, significantly higher than traditional classification methods or single models, providing a reliable foundation for intent understanding for subsequent accurate retrieval and answer generation.

[0102] In some embodiments, receiving user query information, using a pre-trained classification model to obtain the intent classification of the user query information, and using semantic role labeling to extract core entities and constraints in the user query, and generating a structured query vector based on the intent classification and the core entities and constraints, includes:

[0103] Receive user questions, including meeting-related information;

[0104] A pre-trained classification model is invoked to perform intent recognition on the user's query information to obtain intent classification, wherein the intent classification includes at least one of query action items, confirmation decisions, explanation terms, and tracing back discussions;

[0105] Semantic role labeling is performed on the user question information to identify and extract the core entities and constraints in the user question information. The core entities include at least one of the following: attendees, project name, document name, and meeting topic. The constraints include at least one of the following: time range, priority status, and task status.

[0106] The intent classification, the core entity, and the constraints are structurally represented and vectorized to generate a structured query vector.

[0107] In a question-answering method for conference information based on retrieval enhancement generation and knowledge graphs, deep analysis and structured representation of user questions are crucial for ensuring the accuracy and efficiency of subsequent steps. Specifically, the method first receives natural language user questions, typically related to the conference itself, such as tasks, decisions, terminology, or historical discussions. Next, a classification model (such as a fine-tuned BERT-LSTM hybrid model) pre-trained on a large dataset of conference-related data is invoked to identify the intent of the question. This model can accurately categorize user questions into predefined intent categories, such as "querying action items," "confirming decisions," "explaining terminology," or "tracing back to discussions," thus accurately grasping the user's fundamental needs. To simultaneously capture specific details from user questions, the method can perform semantic role labeling analysis in parallel. This process deeply analyzes sentence structure, accurately identifies and extracts core entities that are the focus of the question (such as specific attendees, project names, document names, or conference topics), as well as constraints that limit the scope of the question (such as specific timeframes, priority states, or task states). Finally, the identified intent classifications, extracted core entities, and constraints are integrated and transformed into a standard format that can be seamlessly processed by machines: a structured query vector. This vector not only encapsulates the semantics of the user's intent but also precisely carries its entity and conditional information. It provides a highly structured and information-rich unified query input for subsequent hybrid retrieval in composite indexes, laying the foundation for the transition from fuzzy natural language to precise knowledge retrieval.

[0108] In some specific embodiments, the process of performing a hybrid retrieval based on the structured query vector using a pre-constructed composite index structure to obtain a set of candidate knowledge units includes:

[0109] By calling a pre-built vector index, the semantic similarity between the structured query vector and the multi-dimensional vectors of each standardized atomic knowledge unit stored in the vector index is calculated to obtain a first subset of candidate knowledge units. By calling a pre-built inverted index, keyword matching is performed based on the core entities and constraints contained in the structured query vector to obtain a second subset of candidate knowledge units.

[0110] The first subset of candidate knowledge units and the second subset of candidate knowledge units are fused together to obtain a set of candidate knowledge units.

[0111] Based on the structured query vector generated by the preceding steps, an efficient, multi-path hybrid retrieval process is initiated to comprehensively retrieve relevant information from massive amounts of conference knowledge. This process relies on a pre-built composite index structure, the core of which consists of a vector index and an inverted index working together. First, the vector index is invoked. This vector index stores multi-dimensional semantic vectors of all standardized atomic knowledge units (such as individual decision points, action items, terminology explanations, etc.). By calculating the semantic similarity between the structured query vector and these unit vectors (e.g., using cosine similarity), the system can find the content that best matches the user's question intent at the semantic level, thereby obtaining a first subset of candidate knowledge units. This method excels at capturing semantic relevance, even if the question does not directly mention certain keywords. Simultaneously, the inverted index is invoked in parallel. This inverted index is built based on keywords (such as entity names, status words, etc.) in the knowledge units. The system directly uses the core entities extracted from the structured query vector (such as "Zhang San", "Project Alpha") and constraints (such as "last week", "high priority") as keywords for precise matching, quickly locking down knowledge units containing these specific elements, thereby obtaining a second subset of candidate knowledge units. This method ensures accurate recall of key facts. Finally, the two subsets are merged. This process includes not only a simple union operation but may also involve weighted sorting and deduplication based on similarity scores and matching degrees, ultimately forming a complete and high-quality set of candidate knowledge units. This hybrid strategy combining "semantic similarity retrieval" and "keyword exact matching" balances the depth of understanding with the breadth of retrieval, significantly improving the comprehensiveness and accuracy of information retrieval and laying a rich foundation of knowledge materials for subsequent answer generation.

[0112] In some specific embodiments, the step of fusing the first subset of candidate knowledge units and the second subset of candidate knowledge units to obtain a set of candidate knowledge units includes:

[0113] The first subset of candidate knowledge units and the second subset of candidate knowledge units are deduplicated to obtain a set of fused knowledge units;

[0114] Calculate the overall score for each knowledge unit in the fused knowledge unit set;

[0115] The knowledge units in the fused knowledge unit set are sorted according to the comprehensive score, and the top-ranked knowledge units are selected to form a preliminary candidate knowledge unit set.

[0116] The knowledge units in the preliminary candidate knowledge unit set are supplemented and verified to obtain the candidate knowledge unit set.

[0117] After completing semantic retrieval based on vector indexes and keyword retrieval based on inverted indexes, the first and second subsets of candidate knowledge units obtained are deeply fused to construct a high-quality, highly relevant set of candidate knowledge units. This fusion process specifically includes the following steps:

[0118] First, deduplication is performed on the two subsets. Since different search paths may return the same knowledge units, duplicates can be merged using methods such as unique identifier comparison to form a unified fused knowledge unit set, which serves as the basis for subsequent processing. Next, a comprehensive score is calculated for each knowledge unit in the fused set. This score is a multi-dimensional evaluation metric that integrates not only the semantic similarity score in vector retrieval and the matching score in keyword retrieval, but may also incorporate other weighting factors, such as the authority, timeliness, or correlation strength between the knowledge unit and other nodes in the knowledge graph. This calculation process aims to quantify the overall relevance of each unit to the user's original query. Then, based on the calculated comprehensive score, all units in the fused knowledge unit set are sorted in descending order. After sorting, a certain number of the top-ranked knowledge units are selected to form a preliminary candidate knowledge unit set. This step effectively filters out low-relevance noise and focuses on core candidate information. Finally, to ensure the completeness and logical consistency of the information, the preliminary set undergoes supplementary verification. Through this step, an optimized and enhanced candidate knowledge unit set is finally formed for subsequent re-ranking and answer generation. This rigorous fusion and verification process significantly improves the accuracy, completeness, and reliability of search results.

[0119] In some embodiments, calculating the comprehensive score of each knowledge unit in the fused knowledge unit set includes:

[0120] If a knowledge unit appears only in the first subset of candidate knowledge units, its comprehensive score is the semantic similarity score of the knowledge unit in vector retrieval multiplied by the first weight coefficient.

[0121] If a knowledge unit appears only in the second subset of candidate knowledge units, its overall score is the matching score of that knowledge unit in keyword retrieval multiplied by the second weighting coefficient.

[0122] If a knowledge unit appears in both the first subset of candidate knowledge units and the second subset of candidate knowledge units, its comprehensive score is the sum of the semantic similarity score multiplied by the first weight coefficient and the matching score multiplied by the second weight coefficient.

[0123] In the fusion processing of mixed search results, to fairly and quantitatively evaluate the relevance of each knowledge unit to the query, a unified comprehensive score can be calculated for each unit in the fused knowledge unit set. This scoring strategy differentiates based on how well the knowledge unit is recalled by different search paths, with the specific rules as follows:

[0124] 1. Single Source: Semantic Priority

[0125] If a knowledge unit appears only in a subset of the first candidate knowledge units (i.e., retrieved solely through semantic similarity retrieval via vector indexing), it indicates a high degree of deep semantic relevance to the user's question, but may not contain explicit keyword matching. In this case, its overall score is directly derived by multiplying the semantic similarity score obtained by that unit in vector retrieval by a preset first weighting coefficient. This coefficient is used to calibrate the contribution of semantic similarity to the overall score.

[0126] 2. Single Source: Keyword Matching

[0127] If a knowledge unit appears only in the subset of the second candidate knowledge units (i.e., recalled only through keyword matching using the inverted index), it indicates that it precisely contains the core entity or constraint in the query, but may have slight deviations in the semantic vector space. In this case, its overall score is obtained by multiplying the matching score obtained by that unit in keyword retrieval by a preset second weighting coefficient. This coefficient is used to adjust the weight of precise keyword matching.

[0128] 3. Dual-path recall: Enhanced synergy

[0129] If a knowledge unit appears in both candidate subsets, it proves that it highly matches the query intent at the semantic level and contains key elements at the literal level, making it the most ideal candidate result. In this case, the system will perform a weighted sum of its two scores, resulting in a comprehensive score of (semantic similarity score × first weight coefficient) + (matching score × second weight coefficient). This processing method allows knowledge units jointly verified by both paths to obtain a higher comprehensive score, naturally occupying a higher position in the ranking.

[0130] Through the aforementioned rule-based and configurable scoring mechanism, the system can organically integrate the results of semantic understanding and keyword matching, transforming evidence from different retrieval paths into a comparable quantitative indicator. This provides a clear and reasonable basis for subsequent ranking and screening, thereby ensuring the quality and relevance of the final candidate set.

[0131] In some embodiments, the supplementary verification process for the knowledge units in the preliminary candidate knowledge unit set to obtain the candidate knowledge unit set includes:

[0132] Verify that the metadata of each knowledge unit is consistent with the constraints in the user's question information, and remove inconsistent knowledge units; and / or,

[0133] Verify that the confidence level of each knowledge unit is not lower than a preset threshold, and remove knowledge units with a confidence level lower than the preset threshold.

[0134] After forming a preliminary set of candidate knowledge units, supplementary verification processing can be performed to ensure the high accuracy and reliability of the knowledge materials ultimately used to generate answers. This processing mainly includes one or more of the following control steps:

[0135] 1. Consistency check between metadata and constraints

[0136] Each knowledge unit in the initial candidate knowledge units is individually verified to ensure that its metadata (such as creation time, status label, associated responsible person, etc.) completely matches the constraints extracted through semantic role labeling in the user's question (such as "completed last week", "high priority", "pending approval status"). For example, if the user explicitly asks "action items determined last week", but a knowledge unit's timestamp is two months ago, that unit will be considered inconsistent. The system will automatically eliminate all knowledge units that do not meet the constraints, thereby ensuring that the answer is strictly limited to the scope defined by the user and avoiding providing misleading information that is outdated or inconsistent in status.

[0137] 2. Confidence threshold verification

[0138] The system assesses the confidence level of each knowledge unit. This confidence level may be derived from its source (e.g., comparing authoritative conference minutes with automatic summaries), historical citations, or relevance in the current search. The system presets a confidence threshold, retaining only highly reliable knowledge units with a confidence level at or above (i.e., greater than or equal to) this threshold. Any knowledge unit with a confidence level below the preset threshold, even if its content is relevant, will be removed due to questionable credibility.

[0139] By performing one or more of the above checks, candidate information that is inconsistent in terms of timeliness, status, or credibility can be effectively filtered out, resulting in a purified and enhanced final set of candidate knowledge units. This step significantly improves the quality of knowledge input to the answer generation stage and is a crucial quality checkpoint to ensure the accuracy, reliability, and practicality of the final answer.

[0140] In some embodiments, the step of performing a hybrid retrieval based on the structured query vector using a pre-constructed composite index structure to obtain a candidate knowledge unit set further includes:

[0141] The preliminary set of candidate knowledge units is input into a pre-constructed knowledge graph;

[0142] In the knowledge graph, starting from the knowledge units in the preliminary candidate knowledge unit set, reasoning and expansion are performed along a preset semantic relationship path to obtain extended knowledge units associated with the knowledge units in the preliminary candidate knowledge unit set.

[0143] The extended knowledge units are added to the preliminary candidate knowledge unit set to form the candidate knowledge unit set.

[0144] After obtaining a preliminary set of candidate knowledge units through hybrid retrieval and fusion ranking, this method introduces a knowledge graph reasoning extension step to further enhance the association depth and information completeness of the answers. This step aims to overcome the limitations of simple text matching and utilize the rich semantic relationships inherent in knowledge for intelligent association discovery.

[0145] Specifically, the initial set of candidate knowledge units is used as input and located within a pre-constructed knowledge graph for the meeting domain. In this graph, knowledge units (such as decisions, tasks, personnel, and documents) are nodes, and their semantic relationships (such as "belonging to," "causing," "referring to," and "responsible for execution") are edges. Then, starting with each knowledge unit node in the initial candidate set, traversal and reasoning are performed along one or more pre-defined semantic relationship paths in the graph. For example, when the candidate set contains a "project review decision" node, its corresponding "follow-up action items" can be automatically found along the "derived" relationship, related "background documents" along the "refer to" relationship, or "key participants" along the "involved" relationship. Through this process, extended knowledge units that are not directly recalled in the hybrid retrieval but are closely related to the core candidate content can be discovered and acquired. Finally, these extended knowledge units discovered through graph reasoning are added to the initial candidate set, forming a final set of candidate knowledge units with more comprehensive information and clearer connections. This process achieves a leap from "results from retrieval and matching" to "results from inference and association," ensuring that the subsequently generated answers not only address the user's direct questions but also proactively provide relevant context and derivative information, significantly enhancing the depth and practicality of the answers.

[0146] The following uses a project review meeting scenario in the medical industry as an example to illustrate the hybrid search in this application embodiment.

[0147] Suppose a user asks, "In this cardiovascular disease treatment plan review meeting, which medical equipment budgets were involved in the action items that Zhang San was responsible for?", the complete process of vector retrieval, keyword retrieval, and result fusion is shown in the following example:

[0148] 1. Question preprocessing and query vector generation: By fine-tuning the BERT-LSTM model, the user intent is identified as "query action item + associated budget + limited speaker". The core entities "Zhang San", "cardiovascular disease diagnosis and treatment plan review meeting", "action item" and "medical equipment budget" are extracted to generate a structured query vector.

[0149] 2. Vector Retrieval Execution: Retrieve all AKU vectors for the conference stored in the vector index. Calculate the semantic similarity between the query vector and each AKU vector using the BGE-M3 model. Since vector retrieval focuses on semantic association matching, it will prioritize matching AKUs that contain the semantic association of "Zhang San + Action Item + Budget". Assuming the output is the Top-10 candidate AKUs, it includes 3 core relevant AKUs (corresponding to "Action Items Proposed by Zhang San: Compiling the Medical Equipment Procurement List and Budget", "Action Item Division of Labor: Zhang San is responsible for verifying the medical equipment budget", and "Medical Equipment Budget Details: Zhang San's Submitted Budget Table for Cardiovascular Diagnosis and Treatment Equipment"), and the remaining 7 are candidates with lower semantic relevance (such as budget discussions involving other speakers or non-medical equipment action items).

[0150] 3. Keyword Search Execution: The BM25 algorithm is used, with "Zhang San", "medical equipment budget", and "action items" as core keywords. The algorithm matches the AKU text content associated with the vector index and the keyword index. Keyword search focuses on literal matching and will prioritize matching AKUs that contain all core keywords. Assuming the output of Top-10 candidate AKUs, there are 2 core related AKUs (one of which overlaps with the 3 core AKUs output by the vector search, and one new one, "Zhang San reports on the progress of medical equipment budget and action items"), and the remaining 8 are irrelevant AKUs containing a single keyword (such as only mentioning Zhang San without mentioning the budget, or only mentioning the medical equipment budget without mentioning Zhang San).

[0151] 4. Result Fusion and Knowledge Graph Reasoning: The top-10 candidate AKUs from vector retrieval and keyword retrieval are fused and deduplicated to obtain a total of 19 candidate AKUs (removing 1 overlapping item). Then, knowledge graph path reasoning is performed using the Neo4j APOC reasoning library to mine hidden related knowledge (e.g., by using the semantic relationship of "Zhang San - responsible for - action item" and "action item - associated - medical equipment budget" in the knowledge graph, one AKU containing "action item associated with budget approval" is added and matched). Finally, a set of candidate AKUs covering the core needs of user questions is formed to ensure the comprehensiveness of the retrieval.

[0152] 5. Subsequent processing: The merged candidate AKUs are passed to the re-sorting stage for secondary filtering and sorting, providing a high-quality context for answer generation.

[0153] In some embodiments, the step of using a reordering model to perform a secondary sorting of the knowledge units in the candidate knowledge unit set, selecting the top-ranked candidate knowledge units, and constructing a context based on semantic logic and temporal order includes:

[0154] A re-ranking model is used to perform a secondary relevance score on all knowledge units in the candidate knowledge unit set. During the scoring process, the re-ranking model dynamically enhances the feature weights related to the time constraints and speaker relevance in the user's question information.

[0155] Based on the secondary scoring results, all knowledge units are sorted in descending order, and the top N knowledge units are selected as candidate knowledge units, where N is a preset positive integer.

[0156] After obtaining a rich set of candidate knowledge units through hybrid retrieval and knowledge graph expansion, the system performs a refined reordering step to select the best among them and construct a logically coherent context. This step aims to go beyond preliminary relevance calculation and dynamically evaluate the direct contribution of each candidate knowledge unit to generating the final answer through a more complex model.

[0157] Specifically, a specially trained re-ranking model (such as a cross-encoder based on BERT architecture) is used to perform a secondary relevance scoring on each knowledge unit in the candidate knowledge unit set. Unlike the first-round retrieval scoring, this re-ranking model can analyze the deep semantic interaction between the user's original question and individual knowledge units at a finer granular level. Crucially, this model possesses the ability to dynamically adjust feature weights. It pays particular attention to the temporal constraints (e.g., "last week's meeting," "Q1's goal") and speaker information (e.g., "Zhang San's opinion") inherent in the user's question, and significantly enhances the weights of features related to these constraints and speaker relevance during the scoring process. This allows knowledge units that perfectly match in time or are directly related to a specific speaker to receive higher re-ranking scores, even if they are similar in other aspects.

[0158] After completing the secondary scoring of all knowledge units, they are sorted in descending order based on the new scores. Then, the top N knowledge units (where N is a preset positive integer, such as 5 or 10) are selected as candidate knowledge units for the final answer generation stage. This selection mechanism, based on dynamic weight reordering, ensures that the materials ultimately used to construct the context are not only relevant but also highly aligned with the key constraints in the user's question, laying a solid foundation for generating accurate and relevant answers.

[0159] In some embodiments, constructing the context based on semantic logic and temporal order includes:

[0160] Extract the metadata of the candidate knowledge units, wherein the metadata includes at least a meeting unique identifier, a timestamp, and a confidence level;

[0161] Construct a semantic logic sequence based on the semantic correlation between the candidate knowledge units;

[0162] Based on the semantic logic sequence, each candidate knowledge unit is arranged in chronological order according to its timestamp to form a context.

[0163] The context and the corresponding metadata are encapsulated together into a context structure.

[0164] After refining and ranking the candidate knowledge units, these discrete units need to be organized into coherent and ordered information blocks to provide high-quality input for the generative model. This process is achieved by constructing a context structure, with the following specific steps:

[0165] First, the system extracts metadata from each candidate knowledge unit to prepare for subsequent logical organization and answer tracing. This metadata contains key information describing the attributes of the knowledge unit, including at least:

[0166] Meeting Unique Identifier: Used to trace the specific meeting to which the knowledge belongs;

[0167] Timestamp: Records the precise point in time when this knowledge unit was generated or associated;

[0168] Confidence level: Characterizes the reliability of the information in this unit.

[0169] Secondly, based on the semantic relevance between candidate knowledge units, a semantic logical sequence is constructed. This step is not a simple accumulation, but rather an analysis of the inherent connections between units, such as causal relationships (e.g., a decision leads to subsequent actions), explanatory relationships (e.g., a term definition explains the concepts in the discussion), or parallel supporting relationships, and they are initially arranged in the most logically coherent way that best reflects the essence of the problem.

[0170] Then, based on the established semantic logical sequence, the system further arranges the candidate knowledge units chronologically according to the timestamps in the metadata of each unit. For example, all units are finally arranged according to the order in which the meetings occurred or the timeline of event evolution. This step ensures that the context truly reflects the historical context and evolution of the meeting discussions or project progress.

[0171] Finally, the context (i.e., ordered knowledge unit content) formed after semantic and temporal sorting is encapsulated together with the metadata (identifier, timestamp, confidence level) corresponding to each unit to form a complete, machine-readable context structure. This structure not only contains the substantive content required to generate the answer, but also includes key supporting information to ensure the accuracy, timeliness, and traceability of the answer, providing optimal raw materials for generating a clearly structured and credibly sourced answer.

[0172] In some embodiments, generating a preliminary answer based on a preset contextualized prompt word template and the context structure includes:

[0173] A pre-trained large language model is invoked to generate an initial answer based on meeting knowledge constraints and a preset scenario-based prompt word template; wherein, the scenario-based prompt word template includes at least professional role settings, context injection rules and structured output format requirements.

[0174] After constructing the context structure, the initial answer generation stage begins. The core of this stage is to invoke a pre-trained large language model and carefully structure the input to ensure that the generated answers highly match the needs of the meeting scenario.

[0175] It should be noted that in this embodiment, a set of scenario-based prompt word templates is pre-designed to precisely control the generation behavior of the large language model. This template is not a simple instruction, but a composite framework containing multiple constraints and guidance, which explicitly defines key elements such as professional role setting, context injection rules, and structured output format requirements. Professional role setting means that the template assigns a professional identity to the model, such as "meeting secretary" or "project manager." This setting aims to guide the model to organize language and analyze information from the perspective, tone, and professionalism of a specific role, ensuring that the generated answer style is consistent with the professional tone of the meeting scenario. Context injection rules refer to the template defining how to "feed" the previously constructed context structure, containing ordered knowledge units and their metadata, to the large language model in the clearest and most effective way. This ensures that when generating answers, the model can fully understand and utilize all relevant historical discussions, decisions, action items, and other specific information, guaranteeing the accuracy and richness of the answers. Structured output format requirements mean that the template explicitly requires the model to organize the answer content according to a specific format. For example, answers must clearly distinguish between sections such as "Background Review," "Core Decisions," and "Follow-up Tasks," or must present action items, their responsible parties, and deadlines in list format. This directly ensures the final output answer is well-organized, readable, and readily applicable.

[0176] In some embodiments, the post-processing of the preliminary answer to obtain a structured answer with source information includes:

[0177] The preliminary answer is formatted and redundant information is filtered out to form a refined answer text;

[0178] Based on the metadata of the candidate knowledge units used to construct the context, source citations are added to the key information in the refined answer text to generate a structured answer with source tracing information;

[0179] The cited source includes at least the meeting unique identifier, timestamp, and confidence level of the corresponding knowledge unit, and supports direct location of the original knowledge unit content through the cited source.

[0180] After generating the initial answer, a crucial post-processing workflow can be executed to ensure that the final answer delivered to the user is highly usable, credible, and traceable. This workflow aims to transform the raw text generated by the model into a professional, concise, and structured answer with complete traceability information.

[0181] In practice, the initial answer is first formatted and redundant information filtered. Formatting ensures the answer conforms to preset display standards (such as heading levels, list indentation, etc.) to improve readability. Redundant information filtering removes repetitive statements, irrelevant comments, or ambiguous transitional sentences from the generated text through rules or lightweight models, resulting in a concise answer text that highlights core information and is precisely worded. Secondly, the source information addition step is performed. This process relies on the metadata of candidate knowledge units extracted and encapsulated during the construction of the context structure. The refined answer text is parsed to identify key information points (such as a specific decision, an action item, or a data conclusion) and mapped back to the original knowledge unit that provided the information. For each key information point, the system automatically adds a citation source. The citation source must at least include the meeting's unique identifier, timestamp, and confidence level of the corresponding knowledge unit. Furthermore, this citation mechanism is designed to support direct location, meaning that users or the system can directly jump to and view the complete content of the original knowledge unit that provided the information point (such as excerpts from meeting minutes or transcripts) through the source information (e.g., clicking the citation tag).

[0182] Through the above post-processing, the final structured answer with traceable information is not only concise and clearly formatted, but also possesses extremely high transparency and credibility. Users can not only obtain the conclusion, but also verify the source of information and its reliability with one click. This effectively solves the "black box" concerns commonly found in generative AI applications, and significantly enhances the system's practical value and trustworthiness in professional conference scenarios.

[0183] In some embodiments, the method further includes:

[0184] Based on the confidence level of the candidate knowledge units and the consistency between the preliminary answer and the context structure, the overall confidence level of the structured answer is calculated using a weighted fusion algorithm; and / or,

[0185] Based on the context and semantic associations in the knowledge base, several extended questions related to the user's question information are generated, and the extended questions are output together with the structured answer.

[0186] After generating and refining structured answers with traceability information, this method also includes one or more of the following enhancements to further improve the overall value of the answers and the user experience:

[0187] 1. Calculate the overall confidence level

[0188] To quantify the reliability of the final answer, its overall confidence level can be calculated based on two core dimensions. The first dimension is the confidence level of the candidate knowledge units that form the basis of the answer, reflecting the reliability of the source information. The second dimension is the consistency between the initial answer and the context in which it was constructed, assessing whether the generation process faithfully reflects the provided evidence. These two scores are combined using a pre-defined weighted fusion algorithm (e.g., assigning different weights to the confidence levels of the original knowledge units and the consistency of generation, then performing a comprehensive calculation), resulting in a single, interpretable, structured overall confidence level for the answer. This confidence level can serve as an important indicator, presented along with the answer in the form of a rating or visualization (such as a "high confidence / medium confidence" label), helping users quickly assess the reliability level of the answer.

[0189] 2. Generate related extended questions

[0190] To encourage users to explore deeper and think more comprehensively, the system automatically generates several related extended questions by leveraging the context used to construct the answer and its broader semantic connections within the overall knowledge base. For example, if the original question is "What decisions were made regarding Project A last week?", the system might analyze related information to generate questions such as "What is the current progress of the action project corresponding to this decision?" or "What discussions have been held regarding the technical risks related to Project A?" These extended questions are not randomly generated but are logically derived based on entity relationships, causal chains, or topic similarity. Ultimately, the system outputs these intelligently generated extended questions along with the main structured answer to the user. This not only helps users gain a more comprehensive knowledge map but also guides deeper conversations, enabling more proactive and intelligent knowledge services.

[0191] In some embodiments, the method further includes a multi-turn dialogue support step, including:

[0192] Cache the dialogue history context through a dialogue state tracking mechanism;

[0193] In subsequent user questions, the system automatically links entity information from previous questions to resolve cross-question references and continue intent.

[0194] To provide a continuous and natural interactive experience, this method also integrates efficient multi-turn dialogue support capabilities, ensuring that the system can understand and handle complex information needs across multiple question-and-answer sessions. For example, a dialogue state tracking mechanism can be used to maintain and update the contextual information of the dialogue. Specifically, a high-performance in-memory database such as Redis is used to cache the historical context of the dialogue, including core entities mentioned by the user, explicit intents, and fragments of answers already provided by the system. This establishes a dynamic and traceable dialogue memory. When the user asks subsequent questions, the system can automatically link the entity information from previous questions, intelligently resolving cross-turn referencing and continuing intent. For example, a user might initially ask, "Please describe the current progress of Project Alpha," and after receiving the answer, immediately ask, "Who is responsible for it?" Through dialogue state tracking, the system can accurately understand that "it" refers to "Project Alpha" mentioned earlier, thus eliminating the need for the user to repeat this information and seamlessly connecting the intents of the two rounds of question-and-answer. This allows the dialogue to unfold and deepen naturally around a core topic without repeating all the background information each time.

[0195] In some embodiments, the method further includes a method for constructing a meeting knowledge base based on retrieval enhancement. The meeting knowledge base stores all original meeting recording transcripts, meeting minutes, and project reports. When performing knowledge graph reasoning, information extraction (named entity recognition, relation extraction) can be performed on the documents in the meeting knowledge base to extract entities such as people, projects, tasks, and decisions, as well as the relationships between them (such as "propose", "oppose", "belong to", "cause"), and construct an interconnected semantic network.

[0196] It should be noted that the method for constructing a conference knowledge base based on retrieval enhancement provided in this application revolves around an architecture design of "layered storage + dynamic updates + multi-dimensional indexing," closely integrating the characteristics of RAG technology to systematically transform unstructured raw conference data into knowledge assets that facilitate accurate retrieval and intelligent reasoning. Its core construction process is as follows:

[0197] Intelligent chunking and Atomic Knowledge Unit (AKU) encapsulation are implemented. Standardized conference texts, after ASR transcription and multi-level preprocessing, are intelligently chunked using a dual-dimensional strategy of "semantic coherence" and "time series." Each text chunk is encapsulated as a standardized AKU containing complete metadata (conference UUID, timestamp, speaker, confidence level, etc.) and structured knowledge content (RDF format). The essence of this AKU design lies in achieving a strong binding and bidirectional traceability between knowledge content and the original text: knowledge content is a structured extraction of the core semantics of the original text, while the original text is the sole source and verification basis of the knowledge content. The two are precisely correlated and updated in tandem through the source metadata in the AKU, fundamentally solving the problem of knowledge traceability.

[0198] Multimodal vectorization and composite index construction are performed. Advanced models such as BGE-M3 are used to generate vector representations of three different dimensions (e.g., 1024 dimensions) for each AKU: Dense, Sparse, and ColBERT, to capture semantic similarity, keyword features, and fine-grained matching information, respectively. Based on this, a composite index structure is constructed, including vector indexes, metadata indexes, and keyword indexes, providing efficient and multi-dimensional data support for subsequent hybrid retrieval.

[0199] Establish a hierarchical knowledge base and persistent storage system. Design a three-tier storage architecture of "base library - domain library - cache library" to ensure data reliability, implement domain-specific access control, and improve the performance of high-frequency access, respectively. Through the combination of technologies such as ChromaDB, HDFS, and Redis cluster, ensure efficient data access and persistence, and support incremental synchronization, enabling new meeting knowledge to be quickly added to the database (latency ≤30 seconds).

[0200] Enhanced knowledge graph fusion: Based on core entities and semantic relationships extracted from AKU, a conference-specific knowledge graph is constructed and stored in graph databases such as Neo4j. Through cross-conference knowledge fusion algorithms, relevant knowledge scattered across different conferences is connected into a global knowledge network, significantly improving the relevance and completeness of knowledge.

[0201] A dual-drive dynamic update mechanism is designed. By listening for preprocessing completion events (i.e., events where the conference audio completes ASR optimization and the entire "parsing-correction-enhancement-structuring" process, outputting structured data that meets quality requirements), a complete pipeline of AKU encapsulation, vector updates, and knowledge graph fusion is automatically triggered. Simultaneously, a manual update pipeline is provided, allowing users to correct and supplement knowledge. All updates synchronize knowledge content with the original text and record version logs, thereby ensuring the continuous timeliness, accuracy, and consistency of the knowledge base.

[0202] This method, through the aforementioned end-to-end automated construction process, not only efficiently organizes massive amounts of meeting knowledge, but also ensures, through innovative strong binding design, that knowledge is searchable, verifiable, and associative throughout the entire process from generation and storage to application, laying a solid data foundation for high-quality meeting question answering based on RAG and knowledge graphs.

[0203] To achieve structured modeling and semantic association of data, and to address the issues of low structuring and poor processing efficiency of existing meeting data, some embodiments of the method also include a meeting text preprocessing method. This method, centered on a four-level chain of "parsing-error correction-enhancement-structuring," aims to deeply refine and structure the transcribed text after scenario-based optimization using the ASR model, providing a high-quality data foundation for subsequent knowledge base construction and question-and-answer applications. Its implementation steps are as follows:

[0204] Level 1: Transcription Text Parsing and Verification. An enhanced parser is used to perform targeted parsing on the optimized transcribed text (including timestamps, speaker information, etc.), and three layers of verification logic are executed simultaneously (text validity, system resource usage, and result judgment). It is equipped with 3-way concurrency control and a basic parser degradation scheme to ensure stability and efficiency under large-scale processing and to ensure complete alignment between the parsed results and the original information.

[0205] Level Two: Multi-level Error Correction and Standardization. A scalable error correction mechanism with two to five levels based on a large language model is constructed. In some embodiments, two-level error correction (basic error correction, terminology standardization, and redundancy filtering) is used to improve text accuracy to over 98%. To meet higher precision requirements, it can be flexibly extended to 3-5 rounds of progressive error correction. By adding semantic cross-validation, terminology ambiguity resolution, numerical precision calibration, and even manual feedback, residual errors, semantic contradictions, and industry terminology standardization issues are systematically resolved. Ultimately, the text accuracy can reach over 99.5%, meeting the precision requirements of core and important meetings.

[0206] Level 3: Knowledge Augmentation Preprocessing. Its goal is to extract core semantic information from standardized text and supplement knowledge associations, providing direct input for the subsequent construction of a structured knowledge base. This stage employs a strategy of "parallel rule extraction and model extraction": on the one hand, regular expression templates are used to accurately extract structured information such as meeting topics, decision items, and action items; on the other hand, a fine-tuned RoBERTa model is invoked to deeply extract domain terminology, key values, and logical relationships between entities. Furthermore, entity linking technology is used to associate the extracted entities with lightweight quantum sets of general knowledge graphs and enterprise private libraries, supplementing entity attributes and strictly binding time-series information to ensure the accuracy and temporal consistency of knowledge associations.

[0207] Level 4: Structured Output and Quality Closed Loop. The extracted and correlated information is standardized and encapsulated in a unified format to form a structured dataset (JSON-LD format) containing "entity-attribute-relationship-metadata". Before output, rigorous rule and model dual quality checks are performed to ensure the accuracy, completeness, and relevance of the information. This dataset can be directly used as the core input for subsequent construction of Atomic Knowledge Units (AKUs) and conference-specific knowledge graphs (CKGs), achieving seamless integration between preprocessing and knowledge base construction.

[0208] In summary, this preprocessing method, through its modular, configurable, and strongly validated process design, achieves automated transformation from raw transcribed text to high-quality structured knowledge, forming a complete technical closed loop covering "information extraction → error correction and purification → knowledge enhancement → structure encapsulation," providing a solid and reliable data foundation for upper-layer intelligent question-and-answer applications.

[0209] Example 2

[0210] Corresponding to Embodiment 1 above, this application also provides a conference information question-answering system based on retrieval-enhanced generation and knowledge graph. This system is used to implement the conference information question-answering method based on retrieval-enhanced generation and knowledge graph as described in any one of Embodiment 1. In this embodiment, content that is the same as or similar to that in Embodiment 1 above can be referred to the above description and will not be repeated hereafter.

[0211] Reference Figure 2 As shown, the system includes:

[0212] The vector generation module is configured to receive user query information, use a pre-trained classification model to obtain the intent classification of the user query information, use semantic role labeling to extract the core entities and constraints in the user query, and generate a structured query vector based on the intent classification and the core entities and constraints.

[0213] The hybrid retrieval and reasoning module is configured to perform hybrid retrieval based on the structured query vector using a pre-built composite index structure to obtain a set of candidate knowledge units. The composite index structure includes a vector index constructed based on multi-dimensional vectors of standardized atomic knowledge units. The hybrid retrieval includes at least vector retrieval based on the vector index, keyword retrieval based on keyword matching, and fusion of vector retrieval results and keyword retrieval results with knowledge graph reasoning.

[0214] The context building module is configured to use a reordering model to perform a secondary sorting of the knowledge units in the candidate knowledge unit set, select the top-ranked candidate knowledge units, and build a context structure based on semantic logic and temporal order.

[0215] The answer generation and source tracing module is configured to generate a preliminary answer based on a preset scenario-based prompt word template and the context structure, and to perform post-processing on the preliminary answer to obtain a structured answer with source tracing information. The post-processing includes formatting, filtering redundant information, and adding citation sources.

[0216] In some implementations, the meeting information question-and-answer system based on retrieval enhancement generation and knowledge graph can also implement other steps corresponding to the method described in Embodiment 1. For details, please refer to the detailed description in Embodiment 1 or 2, which will not be repeated here.

[0217] Example 3

[0218] Corresponding to Embodiment 1 above, this application also provides a computer device, including: a processor and a memory, wherein the memory stores a computer program that can run on the processor, and when the computer program is executed by the processor, it executes the meeting information question answering method based on retrieval enhancement generation and knowledge graph provided in any of the above embodiments.

[0219] in, Figure 3 An exemplary computer device 1500 is shown, which may specifically include a processor 1510, a video display adapter 1511, a disk drive 1512, an input / output interface 1513, a network interface 1514, and a memory 1520. The processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, and memory 1520 can communicate with each other via a communication bus 1530.

[0220] The processor 1510 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solution provided by the present invention.

[0221] The memory 1520 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1520 can store the operating system 1521 for controlling the operation of the electronic device, and the basic input / output system (BIOS) 1522 for controlling the low-level operations of the electronic device. Additionally, it can store a web browser 1523, a data storage management system 1524, and a device identification information processing system 1525, etc. The aforementioned device identification information processing system 1525 can be the application program that specifically implements the aforementioned steps in this embodiment of the invention. In summary, when the technical solution provided by this invention is implemented through software or firmware, the relevant program code is stored in the memory 1520 and is called and executed by the processor 1510.

[0222] Input / output interface 1513 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0223] Network interface 1514 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0224] The bus includes a pathway for transmitting information between various components of the device, such as processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, and memory 1520.

[0225] In addition, the electronic device can also obtain information on specific claim conditions from the virtual resource object claim condition information database for condition judgment, and so on.

[0226] It should be noted that although the above-described device only shows the processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, memory 1520, bus, etc., in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the present invention, and not necessarily all the components shown in the figures.

[0227] Example 4

[0228] Corresponding to Embodiment 1 above, this application also provides a computer-readable storage medium. In this embodiment, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description and will not be repeated hereafter.

[0229] The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the meeting information question-answering method based on retrieval enhancement generation and knowledge graph as described above.

[0230] In some implementations of this application, when the computer program is executed by the processor, it can also implement the steps corresponding to the method described in Embodiment 1. Please refer to the detailed description in Embodiment 1, which will not be repeated here.

[0231] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0232] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0233] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0234] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0235] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0236] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0237] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A meeting information question-answering method based on retrieval enhancement generation and knowledge graph, characterized in that, The method includes: The system receives user query information, uses a pre-trained classification model to obtain the intent classification of the user query information, and uses semantic role labeling to extract the core entities and constraints in the user query. Based on the intent classification and the core entities and constraints, a structured query vector is generated. Based on the structured query vector, a hybrid retrieval is performed using a pre-constructed composite index structure to obtain a set of candidate knowledge units. The composite index structure includes a vector index constructed based on multi-dimensional vectors of standardized atomic knowledge units. The hybrid retrieval includes at least vector retrieval based on the vector index, keyword retrieval based on keyword matching, and fusion of vector retrieval results and keyword retrieval results with knowledge graph reasoning. A reordering model is used to perform a secondary sorting of the knowledge units in the candidate knowledge unit set, select the top-ranked candidate knowledge units, and construct a context structure based on semantic logic and temporal order. A preliminary answer is generated based on a preset scenario-based prompt word template and the context structure. The preliminary answer is then post-processed to obtain a structured answer with source information. The post-processing includes formatting, filtering redundant information, and adding citation sources. Specifically, the process of performing a hybrid retrieval based on the structured query vector using a pre-constructed composite index structure to obtain a set of candidate knowledge units includes: By calling a pre-built vector index, the semantic similarity between the structured query vector and the multi-dimensional vectors of each standardized atomic knowledge unit stored in the vector index is calculated to obtain a first subset of candidate knowledge units. Then, by calling a pre-built inverted index, keyword matching is performed based on the core entities and constraints contained in the structured query vector to obtain a second subset of candidate knowledge units. The first subset of candidate knowledge units and the second subset of candidate knowledge units are fused to obtain a set of candidate knowledge units.

2. The meeting information question-answering method based on retrieval enhancement generation and knowledge graph as described in claim 1, characterized in that, The process of receiving user query information involves using a pre-trained classification model to obtain the intent classification of the user query information and using semantic role labeling to extract the core entities and constraints in the user query. A structured query vector is then generated based on the intent classification and the core entities and constraints, including: Receive user questions, including meeting-related information; A pre-trained classification model is invoked to perform intent recognition on the user's query information to obtain intent classification, wherein the intent classification includes at least one of query action items, confirmation decisions, explanation terms, and tracing back discussions; Semantic role labeling is performed on the user question information to identify and extract the core entities and constraints in the user question information. The core entities include at least one of the following: attendees, project name, document name, and meeting topic. The constraints include at least one of the following: time range, priority status, and task status. The intent classification, the core entity, and the constraints are structurally represented and vectorized to generate a structured query vector.

3. The meeting information question answering method based on retrieval enhancement generation and knowledge graph as described in claim 1, characterized in that, The step of fusing the first subset of candidate knowledge units and the second subset of candidate knowledge units to obtain a set of candidate knowledge units includes: The first subset of candidate knowledge units and the second subset of candidate knowledge units are deduplicated to obtain a set of fused knowledge units; Calculate the overall score for each knowledge unit in the fused knowledge unit set; The knowledge units in the fused knowledge unit set are sorted according to the comprehensive score, and the top-ranked knowledge units are selected to form a preliminary candidate knowledge unit set. The knowledge units in the preliminary candidate knowledge unit set are supplemented and verified to obtain the candidate knowledge unit set.

4. The meeting information question answering method based on retrieval enhancement generation and knowledge graph as described in claim 3, characterized in that, The calculation of the comprehensive score of each knowledge unit in the fused knowledge unit set includes: if a knowledge unit appears only in the first candidate knowledge unit subset, its comprehensive score is the semantic similarity score of the knowledge unit in vector retrieval multiplied by a first weight coefficient; if a knowledge unit appears only in the second candidate knowledge unit subset, its comprehensive score is the matching score of the knowledge unit in keyword retrieval multiplied by a second weight coefficient; if a knowledge unit appears in both the first candidate knowledge unit subset and the second candidate knowledge unit subset, its comprehensive score is the sum of the semantic similarity score multiplied by the first weight coefficient and the matching score multiplied by the second weight coefficient. And / or, the supplementary verification process for the knowledge units in the preliminary candidate knowledge unit set to obtain the candidate knowledge unit set includes: verifying whether the metadata of each knowledge unit is consistent with the constraints in the user question information, and removing inconsistent knowledge units; and / or, verifying whether the confidence level of each knowledge unit is not lower than a preset threshold, and removing knowledge units with a confidence level lower than the preset threshold. And / or, the step of performing hybrid retrieval based on the structured query vector using a pre-constructed composite index structure to obtain a candidate knowledge unit set further includes: inputting the preliminary candidate knowledge unit set into a pre-constructed knowledge graph; in the knowledge graph, taking the knowledge units in the preliminary candidate knowledge unit set as starting points, performing reasoning and expansion along a preset semantic relationship path to obtain extended knowledge units associated with the knowledge units in the preliminary candidate knowledge unit set; and adding the extended knowledge units to the preliminary candidate knowledge unit set to form the candidate knowledge unit set.

5. The meeting information question-answering method based on retrieval enhancement generation and knowledge graph as described in claim 1 or 2, characterized in that, The process involves employing a reordering model to perform a secondary sorting of the knowledge units in the candidate knowledge unit set, selecting the top-ranked candidate knowledge units, and constructing a context based on semantic logic and temporal order, including: A re-ranking model is used to perform a secondary relevance score on all knowledge units in the candidate knowledge unit set. During the scoring process, the re-ranking model dynamically enhances the feature weights related to the time constraints and speaker relevance in the user's question information. Based on the secondary scoring results, all knowledge units are sorted in descending order, and the top N knowledge units are selected as candidate knowledge units, where N is a preset positive integer.

6. The meeting information question answering method based on retrieval enhancement generation and knowledge graph according to claim 5, characterized in that, The construction of context based on semantic logic and temporal order includes: Extract the metadata of the candidate knowledge units, wherein the metadata includes at least a meeting unique identifier, a timestamp, and a confidence level; Construct a semantic logic sequence based on the semantic correlation between the candidate knowledge units; Based on the semantic logic sequence, each candidate knowledge unit is arranged in chronological order according to its timestamp to form a context. The context and the corresponding metadata are encapsulated together into a context structure.

7. The meeting information question-answering method based on retrieval enhancement generation and knowledge graph as described in claim 1 or 2, characterized in that, The process of generating a preliminary answer based on a preset contextualized prompt word template and the contextual structure includes: A pre-trained large language model is invoked to generate an initial answer based on meeting knowledge constraints and a preset scenario-based prompt word template; wherein, the scenario-based prompt word template includes at least professional role settings, context injection rules and structured output format requirements.

8. The meeting information question answering method based on retrieval enhancement generation and knowledge graph as described in claim 7, characterized in that, The post-processing of the preliminary answer to obtain a structured answer with source information includes: The preliminary answer is formatted and redundant information is filtered out to form a refined answer text; Based on the metadata of the candidate knowledge units used to construct the context, source citations are added to the key information in the refined answer text to generate a structured answer with source tracing information; The cited source includes at least the meeting unique identifier, timestamp, and confidence level of the corresponding knowledge unit, and supports direct location of the original knowledge unit content through the cited source.

9. The meeting information question answering method based on retrieval enhancement generation and knowledge graph as described in claim 8, characterized in that, The method further includes: Based on the confidence level of the candidate knowledge units and the consistency between the preliminary answer and the context structure, the overall confidence level of the structured answer is calculated using a weighted fusion algorithm; and / or, Based on the context and semantic associations in the knowledge base, several extended questions related to the user's question information are generated, and the extended questions are output together with the structured answer.

10. The meeting information question answering method based on retrieval enhancement generation and knowledge graph according to claim 1 or 2, characterized in that, The method also includes multi-turn dialogue support steps, including: Cache the dialogue history context through a dialogue state tracking mechanism; In subsequent user questions, the system automatically links entity information from previous questions to resolve cross-question references and continue intent.

11. A meeting information question-answering system based on retrieval-enhanced generation and knowledge graph, characterized in that, The system is used to implement the meeting information question answering method based on retrieval enhancement generation and knowledge graph as described in any one of claims 1 to 10, the system comprising: The vector generation module is configured to receive user query information, use a pre-trained classification model to obtain the intent classification of the user query information, use semantic role labeling to extract the core entities and constraints in the user query, and generate a structured query vector based on the intent classification and the core entities and constraints. The hybrid retrieval and reasoning module is configured to perform hybrid retrieval based on the structured query vector using a pre-built composite index structure to obtain a set of candidate knowledge units. The composite index structure includes a vector index constructed based on multi-dimensional vectors of standardized atomic knowledge units. The hybrid retrieval includes at least vector retrieval based on the vector index, keyword retrieval based on keyword matching, and fusion of vector retrieval results and keyword retrieval results with knowledge graph reasoning. The context building module is configured to use a reordering model to perform a secondary sorting of the knowledge units in the candidate knowledge unit set, select the top-ranked candidate knowledge units, and build a context structure based on semantic logic and temporal order. The answer generation and source tracing module is configured to generate a preliminary answer based on a preset scenario-based prompt word template and the context structure, and to perform post-processing on the preliminary answer to obtain a structured answer with source tracing information. The post-processing includes formatting, filtering redundant information, and adding citation sources. Specifically, the process of performing a hybrid retrieval based on the structured query vector using a pre-constructed composite index structure to obtain a set of candidate knowledge units includes: By calling a pre-built vector index, the semantic similarity between the structured query vector and the multi-dimensional vectors of each standardized atomic knowledge unit stored in the vector index is calculated to obtain a first subset of candidate knowledge units. Then, by calling a pre-built inverted index, keyword matching is performed based on the core entities and constraints contained in the structured query vector to obtain a second subset of candidate knowledge units. The first subset of candidate knowledge units and the second subset of candidate knowledge units are fused to obtain a set of candidate knowledge units.

Images