Knowledge base construction method and system based on search enhancement and knowledge graph

Abstract

The application provides a knowledge base construction method and system based on retrieval enhancement and knowledge graph. First, the preprocessing stage converts the text data into structured data set containing entities, attributes and relationships, realizing the conversion from unstructured or semi-structured text data to machine interpretable data. Then, through block encapsulation, an atomic knowledge unit is formed, which becomes the smallest knowledge carrier that can be associated and operated. Then, on the one hand, by constructing a composite index based on vectorization technology, efficient semantic retrieval association between atomic knowledge units is established; on the other hand, by constructing a knowledge graph, a deep logical relationship association between atomic knowledge units is established, forming a reasoning semantic network. Finally, through system integration, an enhanced knowledge base with efficient semantic retrieval and deep relationship reasoning capability is formed, thereby solving the technical problems of low data structuring degree and insufficient data association in the prior art.

Description

A method and system for knowledge base construction based on retrieval enhancement generation and knowledge graph. Technical Field

[0001] This application relates to the field of knowledge base technology, and in particular to a method and system for constructing a knowledge base based on retrieval enhancement generation and knowledge graph. Background Technology

[0002] A meeting knowledge base is a system that centrally stores and manages all meeting content. It is used to organize and save scattered meeting minutes, discussion points, and decisions so that team users can find and review them at any time, avoid losing important information, and improve collaboration efficiency.

[0003] In existing technical solutions, conference audio data is transcribed into text and stored in an unstructured or semi-structured format in the conference knowledge base. This results in a single data storage format and a lack of in-depth processing for knowledge retrieval. When a user queries, the system typically relies on keywords to traverse the entire text. Because the text is directly transcribed from audio, knowledge remains in the form of raw text paragraphs, making it impossible for the system to understand the semantic relationship between the query intent and the text content. Users can only search through literal keyword matching, which leads to decreased retrieval efficiency with increasing text length, high time complexity, and a tendency to miss semantically related but wordedly different information.

[0004] Furthermore, the meeting knowledge base lacks the ability to establish connections between various data types, including meeting audio, transcribed text, decision minutes, and action items. When a user finds a specific piece of information, if they want to further trace its background context, related decision chains, relevant action items, or different speakers' views on the same topic, the system cannot provide this information directly. Users must manually sift through large amounts of original text to sort through it, resulting in high costs and low efficiency for knowledge tracing and deep association queries. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of the prior art and provide a knowledge base construction method and system based on retrieval enhancement generation and knowledge graph, so that the knowledge base has efficient semantic retrieval and deep relational reasoning capabilities.

[0006] The technical solution adopted in this application is as follows:

[0007] In the first aspect, a method for constructing a knowledge base based on retrieval enhancement generation and knowledge graph is provided, the method comprising:

[0008] The text data transcribed from audio data is preprocessed to obtain a structured dataset; wherein the structured dataset includes entities extracted from the text data, attributes of the entities, semantic relationships between the entities, and source metadata corresponding to the text data;

[0009] The structured dataset is divided into blocks based on a preset block division strategy, and each block is encapsulated as an atomic knowledge unit; wherein each atomic knowledge unit includes at least the knowledge content consisting of the entity, the attributes of the entity and the semantic relationship between the entities, the knowledge type, the source metadata and the list of associated knowledge unit identifiers;

[0010] The knowledge content of each atomic knowledge unit is vectorized based on a preset vector model to generate corresponding vector data; and a composite index structure for retrieval is constructed based on the vector data and the source metadata of the atomic knowledge unit based on a preset index construction strategy.

[0011] Based on the entities and semantic relationships between the entities contained in the knowledge content of the atomic knowledge units, a knowledge graph is constructed with entities as nodes and semantic relationships as edges.

[0012] Based on a preset storage engine and a preset hierarchical strategy, the composite index structure and the knowledge graph are integrated to construct an enhanced knowledge base.

[0013] Furthermore, the preprocessing of the text data transcribed from the audio data to obtain a structured dataset includes:

[0014] The text data is parsed based on a preset enhanced parsing model to extract text elements and source metadata;

[0015] During the parsing process, the validity of the text elements and the system resource consumption of the parsing process are verified simultaneously.

[0016] If all verifications pass, the text elements will be corrected at multiple levels based on a preset error correction mechanism.

[0017] The text elements after multi-level error correction are subjected to knowledge enhancement processing to extract the entities, the attributes of the entities, and the semantic relationships between the entities, and a structured dataset is generated by combining the source metadata.

[0018] Furthermore, the multi-level error correction of the text elements based on the preset error correction mechanism includes:

[0019] At least two rounds of error correction processing are performed on the text elements sequentially, wherein:

[0020] The first round of error correction is used to correct character and syntactic errors in the text elements;

[0021] The second round of error correction is used to standardize the terminology and format of the text elements based on an industry terminology database.

[0022] Furthermore, the step of performing knowledge enhancement processing on the text elements after multi-level error correction to extract the entities, the attributes of the entities, and the semantic relationships between the entities, and combining this with the source metadata to generate a structured dataset, includes:

[0023] Based on a preset matching template and a preset language model, the entity, the entity's attributes, and the semantic relationships between the entities are extracted from the text elements.

[0024] The entity is associated with and matched with a general knowledge graph and a private knowledge base based on entity linking technology to supplement the attribute information of the entity;

[0025] The entity, the entity's attributes, and the semantic relationships between the entities are verified based on the preset verification rules and the preset language model.

[0026] The verified entities, their attributes, semantic relationships between entities, and source metadata are integrated to generate a structured dataset that conforms to a preset format.

[0027] Furthermore, the preset segmentation strategy is a dual-dimensional segmentation strategy of semantics and time series.

[0028] The step of dividing the structured dataset into blocks based on a preset block-division strategy and encapsulating each block into an atomic knowledge unit includes:

[0029] Based on the semantic and time-series dual-dimensional segmentation strategy, the structured dataset is divided into multiple text blocks; wherein each text block contains a portion of the entities segmented from the structured dataset, the attributes of the entities, and the semantic relationships between the entities;

[0030] Generate multi-level descriptive tags for each text block, representing its contextual hierarchy, core content, and summary information;

[0031] Each text block, its corresponding multi-level description tag, and the source metadata corresponding to the text block inherited from the structured dataset are encapsulated into an atomic knowledge unit, and the atomic knowledge unit is assigned a knowledge type and a list of associated knowledge unit identifiers.

[0032] Furthermore, the vector data includes at least a dense vector for capturing global semantic similarity, a sparse vector for representing keyword features, and an interactive vector for fine-grained matching.

[0033] Furthermore, the construction of a composite index structure for retrieval based on the preset index construction strategy, based on the vector data and the source metadata of the atomic knowledge units, includes:

[0034] Construct a vector index based on the vector data;

[0035] A metadata index for attribute filtering is constructed based on the source metadata;

[0036] A keyword index for keyword matching is constructed based on the knowledge content of the atomic knowledge units.

[0037] The vector index, the metadata index, and the keyword index are integrated to form the composite index structure.

[0038] Furthermore, the construction of an enhanced knowledge base based on a preset storage engine and a preset hierarchical strategy, integrating the composite index structure and the knowledge graph, includes:

[0039] Construct a layered storage architecture consisting of a basic storage layer, a domain storage layer, and a cache layer;

[0040] The atomic knowledge units, the composite index structure, and the knowledge graph are stored in the basic storage layer;

[0041] Based on the knowledge type of the atomic knowledge unit and / or the source metadata, it is logically categorized and mapped to different domain storage layers;

[0042] The frequently accessed atomic knowledge units and their associated vector data are loaded into the cache layer.

[0043] Furthermore, the method also includes:

[0044] Dynamic updates to the enhanced knowledge base are triggered by listening to preprocessing completion events or responding to user feedback.

[0045] The dynamic update includes at least: synchronously updating the atomic knowledge units involved, as well as the associated composite index structure and knowledge graph.

[0046] In the second aspect, a knowledge base construction system based on retrieval enhancement generation and knowledge graph is provided, the system comprising:

[0047] A preprocessing unit is used to preprocess text data transcribed from audio data to obtain a structured dataset; wherein the structured dataset includes entities extracted from the text data, attributes of the entities, semantic relationships between the entities, and source metadata corresponding to the text data;

[0048] The block encapsulation unit is used to divide the structured dataset into blocks based on a preset block strategy and encapsulate each block into an atomic knowledge unit; wherein each atomic knowledge unit includes at least the knowledge content consisting of the entity, the attributes of the entity and the semantic relationship between the entities, the knowledge type, the source metadata and the list of associated knowledge unit identifiers;

[0049] The vector index construction unit is used to vectorize the knowledge content of each atomic knowledge unit based on a preset vector model to generate corresponding vector data; and to construct a composite index structure for retrieval based on the vector data and the source metadata of the atomic knowledge units based on a preset index construction strategy.

[0050] A knowledge graph construction unit is used to construct a knowledge graph with entities as nodes and semantic relationships as edges based on the entities contained in the knowledge content of the atomic knowledge units and the semantic relationships between the entities.

[0051] The knowledge base construction unit is used to integrate the composite index structure and the knowledge graph based on a preset storage engine and a preset layering strategy to build an enhanced knowledge base.

[0052] In a third aspect, a computer device is provided, including a memory and a processor; the memory stores a computer program that can run on the processor, and when the computer program is executed by the processor, it performs the method provided in the first aspect.

[0053] In the fourth aspect, a computer-readable storage medium is provided that stores a computer program, which, when executed, implements the method provided in the first aspect.

[0054] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0055] This application relates to the field of knowledge base technology, and provides a method and system for constructing a knowledge base based on retrieval enhancement generation and knowledge graph. The method includes: preprocessing text data transcribed from audio data to obtain a structured dataset; wherein the structured dataset includes entities extracted from the text data, entity attributes, semantic relationships between entities, and source metadata corresponding to the text data; dividing the structured dataset into blocks based on a preset block-division strategy, and encapsulating each block into an atomic knowledge unit; wherein each atomic knowledge unit includes at least knowledge content consisting of entities, entity attributes, and semantic relationships between entities, knowledge type, source metadata, and a list of associated knowledge unit identifiers; vectorizing the knowledge content of each atomic knowledge unit based on a preset vector model to generate corresponding vector data; constructing a composite index structure for retrieval based on a preset index construction strategy, based on the vector data and the source metadata of the atomic knowledge units; constructing a knowledge graph with entities as nodes and semantic relationships as edges based on the entities and semantic relationships contained in the knowledge content of the atomic knowledge units; and integrating the composite index structure and the knowledge graph based on a preset storage engine and a preset hierarchical strategy to construct an enhanced knowledge base. This solution constructs a knowledge base with deep semantic association capabilities through hierarchical technical means. First, the preprocessing stage transforms text data into a structured dataset containing entities, attributes, and relationships, realizing the transformation from unstructured or semi-structured text data to machine-parseable data, laying the data foundation for all subsequent operations. Next, atomic knowledge units are formed through block encapsulation, giving knowledge an independent identity and a clear type, making it the smallest knowledge carrier that can be associated and manipulated. Based on this, this solution establishes efficient semantic retrieval associations between atomic knowledge units by constructing a composite index based on vectorization technology; and it establishes deep logical relationship associations between atomic knowledge units by constructing a knowledge graph, forming a reasonable semantic network. Finally, through system integration, an enhanced knowledge base with efficient semantic retrieval and deep relational reasoning capabilities is formed, thus solving the technical problems of low data structuring and insufficient data association in existing technologies. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 is a general flowchart of the knowledge base construction method based on retrieval enhancement generation and knowledge graph provided in the embodiments of this application;

[0058] Figure 2 is a flowchart of text data preprocessing provided in an embodiment of this application;

[0059] Figure 3 is a flowchart of knowledge enhancement processing for text elements provided in an embodiment of this application;

[0060] Figure 4 is a flowchart of block encapsulation of structured datasets provided in an embodiment of this application;

[0061] Figure 5 is a flowchart of constructing a knowledge base provided in an embodiment of this application;

[0062] Figure 6 is a flowchart of the knowledge base update provided in an embodiment of this application;

[0063] Figure 7 is a schematic diagram of the composition of the knowledge base construction system based on retrieval enhancement generation and knowledge graph provided in an embodiment of this application;

[0064] Figure 8 is a schematic diagram of the computer device provided in an embodiment of this application. Detailed Implementation

[0065] The technical solutions in the embodiments of this application will be clearly and completely described below. 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 are within the scope of protection of this application.

[0066] As described in the background section, in existing technical solutions, conference audio data is transcribed into text and stored in a conference knowledge base in an unstructured or semi-structured format. This results in a single data storage format and a lack of in-depth processing for knowledge retrieval. When a user queries, the system typically relies on keywords to traverse the entire text. Because the text is directly transcribed from audio, knowledge always exists in the form of raw text paragraphs, making it impossible for the system to understand the semantic relationship between the query intent and the text content. Users can only search through literal keyword matching, which reduces retrieval efficiency as text length increases, resulting in high time complexity and the potential to miss semantically related but wordedly different information.

[0067] Furthermore, the meeting knowledge base lacks the ability to establish connections between various data types, including meeting audio, transcribed text, decision minutes, and action items. When a user finds a specific piece of information, if they want to further trace its background context, related decision chains, relevant action items, or different speakers' views on the same topic, the system cannot provide this information directly. Users must manually sift through large amounts of original text to sort through it, resulting in high costs and low efficiency for knowledge tracing and deep association queries.

[0068] Based on this, this application provides a knowledge base construction method and system based on retrieval enhancement generation and knowledge graph, aiming to solve the technical problems of low data structuring degree and insufficient data correlation in the prior art.

[0069] The embodiments of this application are described in detail below with reference to the accompanying drawings.

[0070] Example 1

[0071] This embodiment provides a method for constructing a knowledge base based on retrieval enhancement generation and knowledge graph. Referring to Figure 1, the method includes:

[0072] S1: Preprocess the text data transcribed from audio data to obtain a structured dataset; the structured dataset includes entities extracted from the text data, entity attributes, semantic relationships between entities, and source metadata corresponding to the text data;

[0073] S2: Divide the structured dataset into blocks based on a preset block division strategy, and encapsulate each block into an atomic knowledge unit; wherein each atomic knowledge unit includes at least the knowledge content consisting of entities, entity attributes and semantic relationships between entities, knowledge type, source metadata and a list of associated knowledge unit identifiers;

[0074] S3: Based on the preset vector model, the knowledge content of each atomic knowledge unit is vectorized to generate corresponding vector data; and based on the preset index construction strategy, a composite index structure for retrieval is constructed based on the vector data and the source metadata of the atomic knowledge units.

[0075] S4: Based on the entities contained in the knowledge content of atomic knowledge units and the semantic relationships between entities, construct a knowledge graph with entities as nodes and semantic relationships as edges;

[0076] S5: Based on a preset storage engine and preset hierarchical strategy, it integrates a composite index structure and knowledge graph to build an enhanced knowledge base.

[0077] For ease of understanding, the following description will primarily use a meeting scenario as an example.

[0078] In step S1, the purpose of preprocessing is to refine the text data transcribed from audio data, extract core information, and provide a reliable data foundation for subsequent knowledge base construction. Preprocessing can convert unstructured or semi-structured text data into a structured dataset, which provides the sole operational basis for data association. In a structured dataset, entities are the core objects identified from the text data; their attributes describe the characteristics of the object; semantic relationships define the logical connections between different entities; and source metadata records the location and context of this information in the original data fragments. For example, preprocessing can match and parse text data using predefined regular expressions, keyword dictionaries, and sentence templates. It can also automatically identify entity boundaries and entity types in text data by training machine learning models using labeled corpora, further determining relationships between entities through classification models. Alternatively, it can directly utilize large-scale pre-trained language models, fine-tuning them to complete structured tasks such as entity recognition, attribute extraction, and relationship classification. Specific preprocessing methods can be selected or combined based on the data characteristics and resource conditions of the specific scenario to achieve the goal of data structuring.

[0079] In text data from a meeting setting, the specific composition of a structured dataset can be described as follows:

[0080] Entities are the core objects extracted from meeting discussions, such as specific project names, responsible persons (e.g., Zhang San), action items (e.g., submitting project proposals), and key values ​​(e.g., budget of 1 million yuan or deadline of 2025-12-30). Entity attributes describe the characteristics of these entities; for example, the attribute of the responsible person, Zhang San, can be supplemented by indicating that he belongs to the technical department, and the attribute of the key value of 1 million yuan clearly indicates that it falls under the category of project budget. Semantic relationships between entities are the logical links connecting different entities, such as [Zhang San] being responsible for [Project XX], [Proposal XX] being formulated based on [Market Analysis Report], or [Action Item A] containing [Subtask A1]. Source metadata is contextual data recording the origin of the above information, including at least the meeting's unique identifier (ID), the start timestamp of the information's appearance in the meeting, and the corresponding speaker's identifier, ensuring that all data can be traced back to the original discussion context.

[0081] In step S2, the purpose of segmentation and encapsulation is to transform continuous structured datasets into independent, manageable, and inherently interconnected standardized atomic knowledge units, thereby laying the object foundation for knowledge base retrieval and association. The preset segmentation strategy can adopt a dual-dimensional segmentation strategy based on semantics and time series. Specifically, although structured datasets contain entities, entity attributes, and semantic relationships between entities, they are still coherent data segments in form, making precise retrieval and independent management difficult. The preset segmentation strategy cuts the structured dataset into appropriately sized, semantically relatively complete knowledge fragments, allowing each fragment to focus on a core issue or decision point. Encapsulation, based on segmentation, not only transforms and reorganizes the knowledge content (entities, attributes, and relationships) obtained from segmentation into the core of the knowledge unit, but also gives it a complete identity by adding knowledge types, a list of associated identifiers, and inherited source metadata. Here, knowledge type refers to the classification identifier of the knowledge content carried by the atomic knowledge unit, such as the categories mentioned above like decision, action item, terminology, numerical value, or logical relationship; the list of associated identifiers refers to a list pointing to unique identifiers of other related atomic knowledge units. Thus, atomic knowledge units can be directly used as input for index and graph construction to support efficient retrieval and deep association.

[0082] It's important to understand that the knowledge content within atomic knowledge units (often represented in structured formats like RDF) is not isolated knowledge detached from its source. The original text data is the sole source of the knowledge content, which is a structured presentation of the semantic essence of the original text data. The two are linked through source metadata (such as meeting ID, timestamp, and speaker) within the atomic knowledge unit, ensuring a one-to-one correspondence and bidirectional traceability. This traceable link between text data and knowledge content aligns with the goal of transforming meeting data into searchable knowledge.

[0083] As an example, when segmenting text blocks, the base length and context overlap can be controlled by pre-configuring the block size (chunk_size) and block overlap (chunk_overlap) character length, such as chunk_size=800 and chunk_overlap=150. In a conference scenario, text blocks can be further divided based on the boundaries of "speech paragraphs" and "discussion topics".

[0084] In step S3, the purpose of vectorization and index construction is to transform the machine-readable structured knowledge content in atomic knowledge units into machine-matchable mathematical representations and index structures, thereby injecting semantic retrieval capabilities into the knowledge base.

[0085] As an example, the preset vector model can adopt the BGE-M3 model. The vectorization process involves applying the model to the knowledge content of each atomic knowledge unit. The model automatically encodes the knowledge content and simultaneously generates at least three different types of vector data (also known as vector representations): dense vectors for capturing global semantic similarity, sparse vectors for representing keyword features, and interactive vectors (ColBERT vectors) for fine-grained matching. These vectors together constitute the complete mathematical representation of the knowledge unit, providing a numerical foundation for the subsequent construction of a composite index structure that can be directly calculated and matched.

[0086] As an example, the preset index building strategy is a triple indexing strategy consisting of vector index, metadata index, and keyword index. Based on this preset index building strategy and the source metadata of vector data and atomic knowledge units, the specific steps for constructing a composite index structure for retrieval include: building a vector index based on the vector data; building a metadata index for attribute filtering based on the source metadata; building a keyword index for keyword matching based on the knowledge content of atomic knowledge units; and integrating the vector index, metadata index, and keyword index to form a composite index structure.

[0087] Among them, vector indexes can be constructed based on dense vectors, sparse vectors, and interactive vectors in vector data; attribute filtering in metadata indexes allows users to filter by conditions such as speaker and time, filtering out irrelevant conditions; and keyword indexes are adapted for BM25 algorithm retrieval. For example, when processing a complex query, the metadata index can be used to filter out the content of a specific speaker, and then the vector index can be used to perform semantic search within it, while the keyword index ensures that specific terms, names, and other information are not missed.

[0088] In step S4, the purpose of knowledge graph construction is to connect all atomic knowledge units, forming a global, reasonable semantic network, thereby injecting logical association and reasoning capabilities into the knowledge base. Entities (such as projects, personnel, and matters) within atomic knowledge units serve as nodes in the graph, and the semantic relationships (such as responsible, approved, and contained) between entities within atomic knowledge units serve as edges connecting the nodes. The knowledge graph construction process can utilize graph databases (such as Neo4j) to identify, merge, and link entities scattered across different atomic knowledge units, forming a unified, machine-traversable, and queryable network structure.

[0089] In step S5, the preset storage engine can be the ChromaDB engine, and the preset layering strategy is a three-level architecture partitioning strategy of "basic storage layer, domain storage layer and cache layer".

[0090] In summary, firstly, the preprocessing stage transforms text data into a structured dataset containing entities, attributes, and relationships, realizing the transformation from unstructured or semi-structured text data to machine-parseable data, laying the data foundation for all subsequent operations. Next, atomic knowledge units are formed through block encapsulation, giving knowledge an independent identity and a clear type, making it the smallest associative and manipulable knowledge carrier. Based on this, this solution establishes efficient semantic retrieval associations between atomic knowledge units by constructing a composite index based on vectorization technology; secondly, it establishes deep logical relationship associations between atomic knowledge units by constructing a knowledge graph, forming a reasonable semantic network. Finally, through system integration, an enhanced knowledge base with efficient semantic retrieval and deep relational reasoning capabilities is formed, thus solving the technical problems of low data structuring and insufficient data association in existing technologies.

[0091] Further, referring to Figure 2, S1: The text data transcribed from the audio data is preprocessed to obtain a structured dataset, including:

[0092] S10: Parse the text data based on the preset enhanced parsing model to extract text elements and source metadata;

[0093] S11: During the parsing process, the text validity of the text elements and the system resource consumption of the parsing process are verified simultaneously.

[0094] S12: If all verifications pass, multi-level error correction is performed on the text elements based on the preset error correction mechanism;

[0095] S13: Perform knowledge enhancement processing on the text elements after multi-level error correction to extract entities, entity attributes, semantic relationships between entities, and generate a structured dataset by combining source metadata.

[0096] The preset enhanced parsing model can use the MinerU enhanced parser. The text elements are the core text obtained after targeted parsing, and the source metadata includes at least a timestamp and speaker tag.

[0097] During the parsing process, to ensure the validity of the parsed text elements and minimize system load, it is necessary to verify the text validity and system resource usage. The first step is text element validity verification. This involves iterating through the text elements and calculating the ratio of valid text length to total text length. If this ratio meets or exceeds a preset value, the process proceeds to the next step. If not, the batch of text elements is marked for re-inspection, requiring manual verification to ensure accurate problem localization. This preset value can be any one of 80%, 85%, or 90%. The second step is system resource usage verification. This involves monitoring CPU usage, memory usage, and I / O read / write speeds during parsing to determine if core resource usage is below a preset value. This ensures that the parsing process does not affect the normal operation of other system modules. This preset value can be any one of 70%, 75%, or 80%. The third step is to verify the results and trigger the process. If both verifications are successful, the parsing is deemed qualified, and the parsed text elements and source metadata are synchronously output to the subsequent multi-level error correction process. If either verification fails, a three-way concurrent control and basic parser degradation scheme is triggered to ensure the processing stability of the preprocessing link.

[0098] When verification fails, the specific steps for initiating a three-way concurrency control and basic parser degradation scheme include: Regarding three-way concurrency control, firstly, the parsing task is split into meeting segments, with each concurrent link assigned an independent parsing process. Link 1 is responsible for parsing the first third of the meeting segment, Link 2 is responsible for parsing the middle third of the meeting segment, and Link 3 is responsible for parsing the last third of the meeting segment. Each link is configured with independent resource quotas, such as 25% CPU usage and 25% memory usage, to avoid resource contention between links. Secondly, a concurrent link coordination mechanism is set up to synchronize the parsing progress of each link through a message queue. When the parsing progress of a certain link lags behind other links by a preset time, the resource quota of that link is dynamically adjusted, such as increasing the CPU usage to 30%, to ensure that all three links can complete parsing synchronously. The message queue can be Kafka, and the preset time can be any value among 20s, 25s, and 30s. Finally, concurrent results are aggregated. After all three links have completed parsing, the aggregation module merges the parsing results, checking the continuity of speaker identifiers and timestamps in the source metadata to avoid missing segments or information errors. The merged results are then output to the verification logic for secondary validation. Regarding the basic parser degradation scheme, the degradation trigger conditions are first clearly defined. When system resource usage exceeds a preset value and continues for a preset time, or when the MinerU enhanced parser experiences an abnormal process crash or parsing error, degradation is immediately triggered, suspending the high-load parsing function of the MinerU enhanced parser. The preset value can be any one of 70%, 75%, or 80%, and the preset time can be any one of 8s, 10s, or 12s. Second, the basic parser is switched, enabling a preset lightweight basic parser. This lightweight basic parser is built based on a rule engine and is used to replace the MinerU enhanced parser in performing core parsing tasks. When the lightweight basic parser runs, it prioritizes retaining the core functions of text extraction, speaker identifier matching, and timestamp association, while temporarily disabling the initial semantic filtering function to ensure that the parsing task is not interrupted. Finally, a degradation recovery mechanism is introduced. By monitoring the system resource usage and the status of the MinerU enhanced parser in real time, the system automatically switches back to the enhanced parser when the system resource usage is less than a preset value and continues for a preset time, or when the MinerU enhanced parser completes its fault repair. The preset value can be 70%, and the preset time can be 15s or 20s. At the same time, the parsing results of the basic parser are compared with those of the MinerU enhanced parser to supplement and improve the semantic filtering information, ensuring that the parsing quality is not reduced and laying the foundation for subsequent multi-level error correction and knowledge enhancement processing.

[0099] Once all verification criteria are met, multi-level error correction is performed on the text elements based on a preset error correction mechanism. The specific steps include:

[0100] At least two rounds of error correction are performed on the text elements sequentially, where:

[0101] The first round of error correction is used to correct character and syntactic errors in text elements;

[0102] The second round of error correction is used to standardize the terminology and format of text elements based on an industry terminology database.

[0103] The pre-defined error correction mechanism can be built based on the GLM-4 model. The first round of error correction (basic error correction) is mainly used to correct errors in the transcribed text data, such as homophones, similar-looking characters, punctuation errors, and missing / excessive characters, while also marking suspected errors to provide direction for subsequent error correction. The second round of error correction (standardization error correction) mainly completes the standardization of industry terminology, such as aligning the medical field with pharmacopoeias and the financial field with regulatory standards, as well as the standardization of text format, such as unifying the format of time, responsible persons, and action items, while filtering out redundant content such as interjections, repetitive expressions, and invalid colloquialisms. After two rounds of error correction, the accuracy of text elements is improved.

[0104] As an example, three rounds of error correction can be performed sequentially on text elements to further improve their accuracy. The first and second rounds of error correction are consistent with the previous two rounds of error correction schemes. The third round of error correction (semantic cross-validation and error correction) is used to resolve minor errors and semantic inconsistencies remaining from the first two rounds. The specific steps include: First, semantic consistency verification: extract the entities of the current text elements, such as project name, person in charge, and key values, and cross-compare them with the relevant statements of other speakers in the same session and the meeting topic to correct semantic inconsistencies. For example, if the previous round standardized "project deadline December 2024" to "2024.12", the third round of verification will find that the meeting explicitly mentions "December 31, 2024", and this time will be completed. Second, suspected error review: call the RoBERTa fine-tuning model to perform contextual semantic reasoning on the suspected errors marked in the first round, combined with the meeting language. The first step involves determining a correction scheme from the lexicon. For example, the vague expression "this node" can be corrected to "the acceptance node of the XX project" based on the context. The second step is to integrate the statements of multiple speakers. For the same matter, the differences in the statements of different speakers are addressed, and the biased statements are corrected based on the decision-making tendency of the meeting. For example, if speaker A says "the plan needs to be modified" and speaker B adds "the plan needs to be modified in accordance with XX standard", the third round will integrate and correct it to "the plan needs to be modified in accordance with XX standard". The third step is to output the results. After the correction is completed, the accuracy of the text is calculated. If it reaches or exceeds a preset value, the text element is output to the knowledge enhancement processing stage. The preset value can be any one of 95%, 97%, or 99%. If it does not reach the standard, the uncorrected error is marked, triggering a light manual review. During the review, only the substandard segments can be checked to reduce manual costs.

[0105] As a further example, four rounds of error correction can be performed on text elements sequentially to further improve their accuracy. The first, second, and third rounds of error correction are consistent with the previous three-round error correction scheme. The fourth round of error correction (terminology ambiguity resolution and numerical precision verification) is mainly aimed at terminology-intensive meetings such as those in the medical and financial fields. The specific steps include: First, terminology ambiguity resolution. Through entity linking technology, ambiguous terms in text elements, such as "risk control," which may refer to "risk management" in a financial meeting and "access control" in an IT meeting, are linked to the enterprise's private terminology database and a general knowledge graph. The unique meaning is determined and the labeling is corrected based on the domain attributes of the meeting. Second, numerical precision verification. It should be clarified that the extracted numerical information is not meaningless numbers, but quantitative data that is strongly related to the core matters in the meeting scenario and has actual business significance. Specifically, it includes four categories: First, budget-related data, such as "project budget 1..." The data is categorized into four types: 1) Budget of 1 million yuan, hardware cost of 500,000 yuan; 2) Time-related data, such as "Deadline: December 31, 2024, Duration: 30 days"; 3) Progress-related data, such as "Project completion rate: 80%, Review pass rate: 90%"; 4) Specification / quantity-related data, such as "Deploy 3 servers, Submit 2 versions of the plan". The data is validated using regular expressions and the RoBERTa model to ensure the format and logic of these values. For example, "Budget of 1 million yuan, of which 5 million yuan is for hardware" has a clear logical contradiction and will trigger correction. Simultaneously, it is linked to meeting action items and decision-making matters to ensure the values ​​match the context. The third step is a second-round format calibration. This second calibration addresses potential formatting deviations from the first three rounds of correction, such as inconsistent capitalization of terms and numerical units, ensuring consistency with corporate document standards and meeting record standards.

[0106] As a further example, five rounds of error correction can be performed sequentially on text elements to further improve their accuracy. The first, second, third, and fourth rounds of error correction are consistent with the previous four-round error correction scheme. The specific steps of the fifth round of error correction (human-assisted error correction combined with model feedback iteration) include: First, light human intervention: low-confidence segments and semantically complex segments that still exist after the fourth round of error correction, such as segments with multiple speakers discussing each other or dense technical terms, are pushed to the human error correction interface. Humans do not need to modify word by word; the system will provide multiple candidate correction results, and humans only need to confirm the correction scheme. Second, model feedback iteration: the text segments and correction logic after human error correction are synchronously fed back to the GLM-4 and RoBERTa fine-tuning models to update the model's training samples and optimize the error correction rules. Third, final verification: after human error correction is completed, the system automatically performs a final semantic, format, and numerical verification of the entire text to ensure that there are no residual errors, in order to adapt to the text accuracy requirements of core and important meetings, such as project reviews and major decision-making meetings.

[0107] In summary, the multi-level error correction mechanism systematically purifies and elevates colloquial and noisy transcribed text data to high-precision structured data through a progressive processing chain involving basic character correction, terminology standardization, semantic consistency verification, ambiguity resolution and numerical verification, and manual-assisted iteration. The resulting structured data forms the data foundation for constructing atomic knowledge units, composite index structures, and knowledge graphs.

[0108] Further, referring to Figure 3, S13: Knowledge enhancement processing is applied to the text elements after multi-level error correction to extract entities, entity attributes, and semantic relationships between entities. This is combined with source metadata to generate a structured dataset, including:

[0109] S130: Extract entities, entity attributes, and semantic relationships between entities from text elements based on preset matching templates and preset language models;

[0110] S131: Based on entity linking technology, entities are associated and matched with general knowledge graphs and private knowledge bases to supplement the attribute information of entities;

[0111] S132: Verify entities, entity attributes, and semantic relationships between entities based on preset verification rules and preset language models;

[0112] S133: Integrate the validated entities, entity attributes, semantic relationships between entities, and source metadata to generate a structured dataset that conforms to a preset format.

[0113] The overall process of knowledge augmentation follows five stages: pre-processing, core information extraction, knowledge association and supplementation, quality verification, and output adaptation, forming a complete closed loop. First, pre-processing mainly involves loading dependent resources and completing data adaptation. Next, multi-dimensional core information extraction is performed, using a parallel approach of rule extraction and model extraction. Then, knowledge association and supplementation are achieved through entity linking technology combined with general knowledge graph adaptation. Following this is extraction quality verification, ensuring the accuracy and completeness of the extracted information. Finally, structured output is performed, adapting to the format of subsequent atomic knowledge unit encapsulation and knowledge graph construction.

[0114] First, regarding pre-processing, one aspect involves loading core dependency resources. These resources reuse existing resources from the original text to ensure consistency in terminology and technology, avoiding inconsistencies. The first type of resource is rule resources, which include a pre-defined regular expression template library specific to the meeting scenario, categorized by industry. For example, medical meetings include templates related to "treatment plans and number of cases," while financial meetings include templates related to "budget and rate of return." The second type of resource is model resources, loading the RoBERTa model, which has been fine-tuned in the preprocessing stage. This model is specifically tuned for meeting scenario corpora, primarily to optimize the extraction accuracy of domain terminology and numerical information. The third type of resource is knowledge resources, loading a general knowledge graph, a private enterprise library, and a meeting terminology library. The general knowledge graph uses a lightweight subset specifically adapted to the meeting scenario to avoid redundant data affecting processing efficiency, while the private enterprise library contains core information such as internal projects, responsible persons, and terminology. Second, data adaptation processing converts the standardized text elements after multi-level error correction into UTF-8 format that the model can parse. At the same time, it retains the speaker identifier and timestamp information marked in the parsing and verification process. Then, it splits the text into segments according to the meeting paragraphs and time order. The segment length will be adapted to the block size character of the subsequent atomic knowledge unit blocks to ensure that the information extracted later can be successfully associated with the core source metadata such as speaker and time, laying the foundation for subsequent steps.

[0115] Second, regarding the extraction of core information, a parallel approach is adopted, with rule-based extraction as the primary method and model-based extraction as a secondary method. Different types of core information are extracted separately to avoid the bias of a single extraction method and ensure that the extracted information is accurate and comprehensive.

[0116] The rule extraction method employs a preset matching template, which can be a regular expression template. This template is used to extract three types of structured information in batches to meet the core needs of meeting scenarios. The first type of structured information is basic core information. Through keyword matching combined with regular expression templates, the meeting theme is extracted. For example, for the "XX Project Q3 Progress Review Meeting," the matching rule is set to keywords in the meeting name (review, discussion, deployment, etc.) and core entities (project, product, etc.) to ensure accurate extraction of the meeting theme. The second type of structured information is decision-making information, extracting clearly defined decision-making matters from the meeting, such as approving the XX plan or determining the XX budget. The matching rule is set to decision-making verbs (approval, confirmation, review and approval, etc.) and core matters to ensure no decision-making information is omitted. The third type of structured information is action item information, extracting clearly defined action items from the meeting, including complete attributes (responsible person, deadline, action content). The matching rule is set to a combination of action verbs (promote, complete, submit, etc.), content, responsible person identifier, and time expression to ensure that the extracted action items can be directly associated with action item information types in subsequent atomic knowledge units without requiring additional format adjustments.

[0117] The model extraction method employs fine-tuned RoBERTa model inference to deeply mine unstructured or semi-structured information that rules cannot cover. This information specifically includes three categories: The first category is domain terminology, extracting industry-specific terms from the text that are not matched by rules, such as targeted therapy in medical conferences or microservice deployment in IT conferences. The model uses semantic understanding, combined with a conference terminology database, to filter out non-terminal words, improving accuracy. The second category is numerical information, specifically referring to quantitative data strongly related to core matters and possessing practical business significance in the conference scenario, rather than meaningless numbers. This is further subdivided into four categories: budget-related, such as a project budget of 1 million yuan and hardware costs of 500,000 yuan; time-related, such as a deadline of December 31, 2024, and a 30-day construction period; and progress-related, such as project completion rate of 80% and review... The pass rate is 90%. The fourth type is specifications or quantity, such as deploying 3 servers or submitting 2 versions of the plan. This type of numerical information will serve as the core attributes in the atomic knowledge unit and will also provide data support for numerical associations in the subsequent knowledge graph, such as project-budget-1 million. The third type of information is logical relationships. We extract the causal, conditional, and inclusion logical relationships between core entities. For example, the causal relationship of delaying the project launch due to insufficient budget, the inclusion relationship of action items including the submission of the plan, and the organization of review. We extract these relationships using the triple (subject-relationship-object) format, such as project launch-delay reason-insufficient budget. These extracted logical relationships will provide basic materials for the edge construction of the subsequent meeting-specific knowledge graph.

[0118] Third, regarding knowledge association and supplementation, entity linking technology is used to match all extracted core entities (projects, responsible persons, domain terms, action items, etc.) with general knowledge graphs and enterprise private databases respectively. The matching logic adopts dual judgment of entity name similarity and contextual semantic similarity. When the similarity reaches a preset value or above, it is judged as a successful match. After the matching is completed, the relevant attributes of the entity are supplemented, such as responsible person - department - technology department, domain term - definition - the core role of XX technology, to ensure that the entity information is more complete.

[0119] It's important to clarify that the relationship between the general knowledge graph discussed here and the knowledge graph construction discussed later is one of foundational supplementation and deep integration. Specifically, this relationship manifests in two ways: First, the entity attributes supplemented by the general knowledge graph are encapsulated as association attributes within atomic knowledge units, enriching their information dimensions. For example, the term "microservice" within an atomic knowledge unit might include the basic definition from the general knowledge graph, making the unit's information more complete. Second, the general semantic relationships provided by the general knowledge graph offer a foundational reference for cross-conference knowledge fusion and semantic association reasoning in subsequent conference-specific knowledge graphs. For instance, the project-responsible person association logic in the general knowledge graph can be reused in the entity associations of the knowledge graph, improving the efficiency and accuracy of knowledge graph construction. Furthermore, the general knowledge graph itself does not participate in the core construction of the knowledge graph, nor is it stored in the hierarchical knowledge base of this embodiment. It serves only as an external auxiliary resource for knowledge enhancement, providing support for knowledge association supplementation in the preprocessing stage.

[0120] Fourth, regarding quality verification, rule-based verification and model verification are employed to ensure output quality. Quality verification mainly includes three aspects: first, accuracy, verifying whether the extracted information and associated attributes are consistent with the meeting text, such as whether the responsible person's identifier matches the speaker and whether numerical information matches the original text, avoiding deviations; second, completeness, verifying whether all core entities (projects, responsible persons, key action items) have been extracted and associated without omissions; and third, relevance, verifying whether the association results conform to time series constraints and semantic logic, such as no cross-time erroneous associations and no logically contradictory associations. The implementation methods for quality verification are: first, rule-based verification, using preset verification rules to batch-check obvious errors, such as negative values, incorrect time formats, and contradictory associations, improving verification efficiency; second, model re-checking, calling the RoBERTa model to perform secondary reasoning on segments that failed verification, correcting extraction or association errors; and third, anomaly handling, information that failed verification is marked for re-checking and pushed to manual lightweight verification.

[0121] Fifth, regarding structured output, the extracted and supplemented information that has passed verification will be standardized and output in a unified format to form a structured dataset of "entity-attribute-relationship-source metadata". The specific format will be adapted to the needs of subsequent atomic knowledge unit encapsulation and knowledge graph construction. Specifically, each core entity corresponds to one structured record, including entity name, type, attributes (including attributes supplemented by the general knowledge graph), related entities and relationships, and source metadata (speaker identifier, timestamp, and corresponding text fragment). The output format of the structured dataset adopts JSON-LD, which is consistent with the encoding format of atomic knowledge units, ensuring that it can be directly called by subsequent intelligent segmentation and knowledge unit encapsulation stages.

[0122] When constructing associations, time-series information needs to be given special consideration, as it is one of the core constraints for association construction. Two key methods are employed to ensure the accuracy and temporal consistency of associations: First, during association matching, entities within the same time dimension are forcibly associated. For example, the extracted "project progress 80%" will only match the project entity corresponding to that time point (e.g., the end of Q3), avoiding erroneous associations across time dimensions (e.g., associating Q3 progress with the Q2 project), ensuring the accuracy of the association logic. Second, the association results will be synchronously bound with the timestamp information retained in the original text, binding and storing "time-entity-association relationship." For example, "end of Q3 2024 - XX project - progress 80% - association - project budget 1 million," this time-series information will be synchronously written into the source metadata of the atomic knowledge unit, providing core support for subsequent time associations in the knowledge graph (such as progress associations at different times for the same project) and cross-conference knowledge fusion (such as the progress connection between Q2 and Q3 meetings for the same project), ensuring the accuracy and temporal consistency of knowledge associations.

[0123] Furthermore, the preset segmentation strategy is a dual-dimensional segmentation strategy of semantics and time series; referring to Figure 4, S2: the structured dataset is segmented into blocks based on the preset segmentation strategy, and each block is encapsulated into an atomic knowledge unit, including:

[0124] S20: Based on a dual-dimensional segmentation strategy of semantics and time series, the structured dataset is divided into multiple text blocks; each text block contains a portion of entities from the structured dataset, the attributes of the entities, and the semantic relationships between the entities;

[0125] S21: Generate multi-level descriptive tags for each text block, representing its contextual hierarchy, core content, and summary information;

[0126] S22: Encapsulate each text block and its corresponding multi-level description label, as well as the source metadata corresponding to the text block inherited from the structured dataset, into an atomic knowledge unit, and assign the atomic knowledge unit a knowledge type and a list of associated knowledge unit identifiers.

[0127] In the semantic and time-series dual-dimensional segmentation strategy, the semantic dimension ensures that each segmented text block is relatively independent and complete in terms of content theme, while the time-series dimension ensures that the text blocks conform to the natural time sequence of the discussion and the speaking rounds.

[0128] As an example, a three-level tag is generated for each text block, including a main heading for the meeting stage, subheadings, and a core summary. The main heading identifies the macro-level stage of the text block within the overall meeting process, such as project initiation discussion, solution review, or summary and decision-making—this establishes the contextual level of the knowledge. Subheadings summarize the core topics discussed within the text block, such as disputes over budget allocation or data migration plans. The core summary extracts the key information from the text block, summarizing the core conclusions or key discussion points in one or two sentences. In this way, the three-level descriptive tags enable precise classification, categorization, and rapid browsing of atomic knowledge units, laying the foundation for refined knowledge management.

[0129] Furthermore, referring to Figure 5, S5: Based on a preset storage engine and a preset hierarchical strategy, an enhanced knowledge base is constructed by integrating a composite index structure and a knowledge graph, including:

[0130] S50: Construct a layered storage architecture consisting of a basic storage layer, a domain storage layer, and a cache layer;

[0131] S51: Store atomic knowledge units, composite index structures, and knowledge graphs in the basic storage layer;

[0132] S52: Based on the knowledge type and / or source metadata of atomic knowledge units, logically classify and map them to different domain storage layers;

[0133] S53: Load frequently accessed atomic knowledge units and their associated vector data into the cache layer.

[0134] The architecture employs a three-tiered structure: a basic storage layer, a domain storage layer, and a cache layer. The basic storage layer (which can utilize a distributed storage system such as HDFS) persistently stores all atomic knowledge units, vector data corresponding to composite index structures, and graph data of the knowledge graph, ensuring data reliability. The domain storage layer logically categorizes atomic knowledge units into different domains (such as healthcare and finance) based on their knowledge type (e.g., decision, action item) and / or the item identifier in their source metadata, and configures access control lists (ACLs) for each domain to achieve fine-grained permission management. The cache layer (which can utilize a Redis cluster) dynamically loads frequently accessed atomic knowledge units and their associated hot vector data to provide fast retrieval responses. The entire architecture supports incremental synchronization; once new meeting data is processed, it automatically triggers collaborative updates across all layers.

[0135] Furthermore, referring to Figure 6, the method also includes:

[0136] S6: Trigger dynamic updates to the enhanced knowledge base by listening to preprocessing completion events or responding to user feedback;

[0137] Dynamic updates include at least the synchronous updates of atomic knowledge units involved, as well as the associated composite index structure and knowledge graph.

[0138] The system employs a dual-drive mechanism to achieve dynamic updates to the knowledge base. Specifically, the automated update chain listens for preprocessing completion events via a message queue (such as Kafka). Once this event is captured, it automatically triggers the encapsulation, vectorization, and knowledge graph fusion operations of newly added structured data into atomic knowledge units, achieving incremental updates. The manual update chain provides an interactive interface to respond to user feedback on corrections, additions, or annotations. Upon receiving this feedback, the system performs targeted updates on the specified atomic knowledge units and their associated indexes and graph nodes. Both update chains ensure the synchronous correction of atomic knowledge units, composite indexes, and the knowledge graph, and record detailed version logs to support status traceability and data rollback.

[0139] As an example, the specific process for listening to and triggering this update step in the automated update chain is as follows:

[0140] The first step is to trigger a completion verification of the preprocessing chain. After the structured output step of the preprocessing stage is completed, the system automatically initiates a completion verification, focusing on three core aspects: First, all stages of the entire preprocessing chain (parsing verification, multi-level error correction, and knowledge enhancement processing) have been executed normally without interruption or errors; second, the output structured dataset format conforms to the JSON-LD standard, is compatible with the atomic knowledge unit encoding format, and contains complete information on entity-attribute-relationship-source metadata; third, the data quality meets the standards, i.e., the accuracy of the structured dataset exceeds the first preset value, the completeness rate of core entity extraction exceeds the second preset value, and there are no segments that fail verification and need to be re-examined. If any substandard items exist, the system will trigger a rework in the preprocessing stage until the standards are met before proceeding with subsequent steps.

[0141] The second step is to generate and push a preprocessing completion event identifier. After successful verification, the system automatically generates a unique preprocessing completion event identifier (including the meeting UUID, preprocessing completion timestamp, and dataset version number), and encapsulates the core event information (basic meeting information, preprocessing execution logs, and structured dataset storage path). This identifier is then pushed to the Kafka message queue via a preset interface to complete the event reporting.

[0142] The third step involves Kafka listening for and forwarding events. The Kafka message queue continuously listens for reports of preprocessed events. Once it captures the event identifier and core information, it immediately unpacks and verifies the message. After confirming that the event information is complete and accurate, it forwards the event to the trigger module of the automated update link according to the preset routing rules, ensuring accurate event delivery.

[0143] The fourth step triggers the entire automated knowledge base update process. After receiving the preprocessing completion event forwarded by Kafka, the trigger module synchronously starts the automated knowledge base update: First, it triggers the splitting and encapsulation of atomic knowledge units. The system reads the storage path of the structured dataset and, according to a semantic and time-series dual-dimensional segmentation strategy, splits the structured dataset into text blocks, generates three-level tags, and encapsulates them into standardized atomic knowledge units, completing the batch generation and encoding of atomic knowledge units (JSON-LD format). Next, it triggers multimodal vectorization and index update, calling the BGE-M3 model to generate multi-dimensional vectors (Dense vector, S...) for the newly generated atomic knowledge units. Parse vectors and ColBERT vectors (each vector has 1024 dimensions), and simultaneously update the composite index structure of "vector index + metadata index + keyword index" to ensure that new atomic knowledge units can be quickly retrieved. Then, trigger knowledge graph fusion. Based on the semantic associations of the newly generated atomic knowledge units, extract core entities and semantic relationships, and simultaneously update the knowledge graph specific to the meeting. If it is new meeting data from the same project / team, cross-meeting knowledge fusion algorithms, such as entity co-occurrence combined with semantic similarity matching algorithms, are used to realize the association and fusion of new atomic knowledge units with existing knowledge graphs, thereby improving the global knowledge network.

[0144] The fifth step is to verify and log the update results. After the knowledge base is updated, the system automatically verifies the update results, checking whether all newly generated atomic knowledge units are synchronized to the layered knowledge base (basic storage layer, domain storage layer, and cache layer), whether the vector index update is effective, and whether the knowledge graph fusion is accurate and free of logical contradictions. If the verification passes, a complete update log is recorded (including event identifier, update content, update time, and execution status), supporting rollback to subsequent versions. If the verification fails, the system automatically triggers a retry mechanism. If the retry fails after 3 attempts, it is marked as abnormal and pushed to the manual operation and maintenance end to ensure the reliability of the automated update chain.

[0145] In summary, by establishing a user feedback-driven model iteration mechanism, the quality of the knowledge base is continuously optimized, adapting to the personalized needs of conferences in different fields.

[0146] Furthermore, this embodiment also provides an implementation method for transcribing audio data into text data. As an example, to address issues such as multiple speakers, overlapping speech, and a mixture of technical terms and colloquialisms in meeting scenarios, the Whisper-medium model was customized and fine-tuned with optimized strategies. The core of this approach is the introduction of a dedicated speaker separation algorithm, speech overlap resolution strategy, and semantic correction mechanism to improve transcription accuracy and better adapt to complex meeting scenarios. The specific operation is as follows:

[0147] First, a speaker separation algorithm is introduced, employing the Speaker Diarization algorithm based on the Transformer architecture, combined with Mel spectrogram feature extraction technology. The conference audio is first segmented at the frame level (with preset frame length and frame shift), and then a clustering algorithm is used to distinguish the speech features of different speakers, achieving accurate separation of multiple speakers. At the same time, the separated speech is associated with the subsequent transcribed text, and a unique speaker identifier is automatically added to each transcribed content. This solves the semantic confusion problem caused by multiple speakers and lays the foundation for subsequent analysis of speaker-related information.

[0148] Second, a speech overlap resolution strategy is adopted. In meetings, it is common for 2-3 people to speak simultaneously. To address this speech overlap problem, a three-step approach is used: overlap detection, speech separation, and completion / repair. First, overlapping speech segments are accurately identified through energy threshold detection and phase difference analysis. Second, an improved Conv-TasNet model is used to separate the overlapping segments, extracting the independent speech signal of each speaker. Third, by combining contextual semantics and a meeting terminology database, the incomplete speech segments are completed and repaired, ensuring the integrity and coherence of the transcribed text and fundamentally solving the problems of transcription loss and semantic breaks caused by overlapping speeches.

[0149] Third, a semantic correction mechanism for matching terminology with spoken language is introduced. Considering the characteristics of dense terminology and chaotic spoken language in meeting scenarios, a dual semantic correction strategy is formulated: First, terminology adaptation correction. Based on meeting terminology databases of various industries (such as medical, financial, IT, etc.), the Whisper-medium model is fine-tuned to enhance the model's ability to identify domain-specific terms and reduce terminology transcription errors. At the same time, through a terminology mapping table, colloquial expressions are corrected to standardized terms. For example, the commonly used phrase "this plan" corresponds to the standardized "XX project implementation plan". Second, colloquial redundancy filtering and semantic correction. A BERT-based semantic filtering model is introduced to automatically identify and filter colloquial filler words such as "um," "that," and "then." At the same time, for colloquial phrasing and semantically ambiguous expressions, such as "it will be completed around next week," logical correction is performed, and specific information is added based on the meeting context, such as "it is expected to be completed next week X." This improves the standardization and semantic clarity of the transcribed text, provides higher-quality input data for subsequent preprocessing links, and ensures that the audio transcription results can be seamlessly connected and efficiently linked with the preprocessing stage.

[0150] Example 2

[0151] This second embodiment provides a knowledge base construction system based on retrieval enhancement generation and knowledge graph. Referring to Figure 7, the system includes:

[0152] A preprocessing unit is used to preprocess text data transcribed from audio data to obtain a structured dataset; wherein the structured dataset includes entities extracted from the text data, attributes of the entities, semantic relationships between the entities, and source metadata corresponding to the text data;

[0153] The block encapsulation unit is used to divide the structured dataset into blocks based on a preset block strategy and encapsulate each block into an atomic knowledge unit; wherein each atomic knowledge unit includes at least the knowledge content consisting of the entity, the attributes of the entity and the semantic relationship between the entities, the knowledge type, the source metadata and the list of associated knowledge unit identifiers;

[0154] The vector index construction unit is used to vectorize the knowledge content of each atomic knowledge unit based on a preset vector model to generate corresponding vector data; and to construct a composite index structure for retrieval based on the vector data and the source metadata of the atomic knowledge units based on a preset index construction strategy.

[0155] A knowledge graph construction unit is used to construct a knowledge graph with entities as nodes and semantic relationships as edges based on the entities contained in the knowledge content of the atomic knowledge units and the semantic relationships between the entities.

[0156] The knowledge base construction unit is used to integrate the composite index structure and the knowledge graph based on a preset storage engine and a preset layering strategy to build an enhanced knowledge base.

[0157] First, the preprocessing stage transforms text data into a structured dataset containing entities, attributes, and relationships, realizing the transformation from unstructured or semi-structured text data to machine-parseable data, laying the data foundation for all subsequent operations. Next, atomic knowledge units are formed through block encapsulation, giving knowledge an independent identity and a clear type, making it the smallest associative and manipulable knowledge carrier. Based on this, this solution establishes efficient semantic retrieval associations between atomic knowledge units by constructing a composite index based on vectorization technology; and it establishes deep logical relationship associations between atomic knowledge units by constructing a knowledge graph, forming a reasonable semantic network. Finally, through system integration, an enhanced knowledge base with efficient semantic retrieval and deep relational reasoning capabilities is formed, thus solving the technical problems of low data structuring and insufficient data association in existing technologies.

[0158] Example 3

[0159] This embodiment three provides a computer device, including a memory and a processor; 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 knowledge base construction method based on retrieval enhancement generation and knowledge graph provided in embodiment one above.

[0160] Referring to Figure 8, the computer device of this embodiment is illustrated by way of example. Specifically, it may 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 be communicatively connected via a communication bus 1530.

[0161] 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 in this application.

[0162] 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 computer device, and the basic input / output system 1522 for controlling the low-level operations of the computer 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. In summary, when implementing the technical solution provided in this application through software or firmware, the relevant program code is stored in the memory 1520 and is called and executed by the processor 1510.

[0163] 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.

[0164] 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.).

[0165] The communication bus 1530 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.

[0166] In addition, the device can also obtain information on specific claim conditions from the virtual resource object claim condition information database for condition judgment, and so on.

[0167] 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, communication bus 1530, 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 solution of this application, and does not necessarily include all the components shown in the figures.

[0168] Example 4

[0169] This fourth embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed, it implements the knowledge base construction method based on retrieval enhancement generation and knowledge graph provided in the first embodiment.

[0170] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, 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 of various embodiments or some parts of the embodiments of this application.

[0171] The foregoing has provided a detailed description of the knowledge base construction method and system based on retrieval enhancement generation and knowledge graph provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for constructing a knowledge base based on retrieval enhancement generation and knowledge graph, characterized in that, The method includes: preprocessing text data transcribed from audio data to obtain a structured dataset; dividing the structured dataset into blocks based on a preset block-division strategy, and encapsulating each block into an atomic knowledge unit; vectorizing the knowledge content of each atomic knowledge unit based on a preset vector model to generate corresponding vector data; and constructing a composite index structure for retrieval based on the vector data and the source metadata of the atomic knowledge units, based on a preset index construction strategy, including: constructing a vector index based on the vector data; constructing a metadata index for attribute filtering based on the source metadata; constructing a keyword index for keyword matching based on the knowledge content of the atomic knowledge units; and connecting the vector index, the metadata index, and the keyword index. The atomic knowledge units are integrated to form the composite index structure; based on the entities contained in the knowledge content of the atomic knowledge units and the semantic relationships between the entities, a knowledge graph is constructed with entities as nodes and semantic relationships as edges; based on a preset storage engine and a preset layering strategy, the composite index structure and the knowledge graph are integrated to construct an enhanced knowledge base, including: constructing a layered storage architecture consisting of a basic storage layer, a domain storage layer, and a cache layer; storing the atomic knowledge units, the composite index structure, and the knowledge graph in the basic storage layer; logically classifying and mapping the atomic knowledge units to different domain storage layers based on their knowledge types and / or source metadata; and loading frequently accessed atomic knowledge units and their associated vector data into the cache layer.

2. The knowledge base construction method based on retrieval enhancement generation and knowledge graph as described in claim 1, characterized in that, The preprocessing of text data transcribed from audio data to obtain a structured dataset includes: parsing the text data based on a preset enhanced parsing model to extract text elements and source metadata; simultaneously verifying the text validity of the text elements and the system resource consumption of the parsing process during the parsing process; if all verifications pass, performing multi-level error correction on the text elements based on a preset error correction mechanism; and performing knowledge enhancement processing on the text elements after multi-level error correction to extract the entities, the attributes of the entities, and the semantic relationships between the entities, and generating a structured dataset in combination with the source metadata.

3. The knowledge base construction method based on retrieval enhancement generation and knowledge graph as described in claim 2, characterized in that, The multi-level error correction based on the preset error correction mechanism for the text elements includes: performing at least two rounds of error correction processing on the text elements in sequence, wherein: the first round of error correction is used to correct character and syntactic errors in the text elements; the second round of error correction is used to perform terminology standardization and format normalization processing on the text elements based on an industry terminology database.

4. The knowledge base construction method based on retrieval enhancement generation and knowledge graph as described in claim 2, characterized in that, The step of performing knowledge augmentation processing on the text elements after multi-level error correction to extract the entities, the attributes of the entities, and the semantic relationships between the entities, and generating a structured dataset in conjunction with the source metadata, includes: extracting the entities, the attributes of the entities, and the semantic relationships between the entities from the text elements based on a preset matching template and a preset language model; associating and matching the entities with a general knowledge graph and a private knowledge base based on entity linking technology to supplement the attribute information of the entities; validating the entities, the attributes of the entities, and the semantic relationships between the entities based on preset validation rules and the preset language model; and integrating the validated entities, the attributes of the entities, the semantic relationships between the entities, and the source metadata to generate a structured dataset conforming to a preset format.

5. The knowledge base construction method based on retrieval enhancement generation and knowledge graph as described in claim 1, characterized in that, The preset segmentation strategy is a semantic and time-series dual-dimensional segmentation strategy. The step of segmenting the structured dataset based on the preset segmentation strategy and encapsulating each segment into an atomic knowledge unit includes: dividing the structured dataset into multiple text blocks based on the semantic and time-series dual-dimensional segmentation strategy; wherein each text block contains a portion of the entities segmented from the structured dataset, the attributes of the entities, and the semantic relationships between the entities; generating multi-level descriptive tags representing the contextual level, core content, and summary information of each text block; encapsulating each text block, its corresponding multi-level descriptive tags, and the source metadata inherited from the structured dataset corresponding to the text block into an atomic knowledge unit, and assigning the atomic knowledge unit a knowledge type and a list of associated knowledge unit identifiers.

6. The knowledge base construction method based on retrieval enhancement generation and knowledge graph as described in claim 1, characterized in that, The vector data includes at least a dense vector for capturing global semantic similarity, a sparse vector for characterizing keyword features, and an interactive vector for fine-grained matching.

7. The knowledge base construction method based on retrieval enhancement generation and knowledge graph as described in claim 1, characterized in that, The method further includes: triggering dynamic updates to the enhanced knowledge base by listening to preprocessing completion events or responding to user feedback; wherein the dynamic updates include at least: synchronously updating the atomic knowledge units involved, as well as the associated composite index structure and knowledge graph.

8. A knowledge base construction system based on retrieval enhancement generation and knowledge graph, characterized in that, The system includes: a preprocessing unit for preprocessing text data transcribed from audio data to obtain a structured dataset; a block encapsulation unit for dividing the structured dataset into blocks based on a preset block strategy and encapsulating each block into an atomic knowledge unit; a vector index construction unit for vectorizing the knowledge content of each atomic knowledge unit based on a preset vector model to generate corresponding vector data; and, based on a preset index construction strategy, constructing a composite index structure for retrieval based on the vector data and the source metadata of the atomic knowledge units, including: constructing a vector index based on the vector data; constructing a metadata index for attribute filtering based on the source metadata; constructing a keyword index for keyword matching based on the knowledge content of the atomic knowledge units; and connecting the vector index, the metadata index, and the keyword index. The keyword index is integrated to form the composite index structure; the knowledge graph construction unit is used to construct a knowledge graph with entities as nodes and semantic relationships as edges based on the entities contained in the knowledge content of the atomic knowledge units and the semantic relationships between the entities; the knowledge base construction unit is used to integrate the composite index structure and the knowledge graph based on a preset storage engine and a preset layering strategy to construct an enhanced knowledge base, including: constructing a layered storage architecture consisting of a basic storage layer, a domain storage layer and a cache layer; storing the atomic knowledge units, the composite index structure and the knowledge graph in the basic storage layer; logically classifying and mapping the atomic knowledge units to different domain storage layers based on their knowledge types and / or source metadata; and loading frequently accessed atomic knowledge units and their associated vector data into the cache layer.

Images