A knowledge management reasoning system based on accurate positioning retrieval

Abstract

The application relates to the field of information processing and discloses a knowledge management reasoning system based on accurate positioning retrieval, which is centered on the semantic anchor point, performs bidirectional logical correlation detection in the starting direction and the ending direction of the text respectively, extracts a first logical coupling degree when detecting in the starting direction, divides the preceding sentence into a slicing range when the first logical coupling degree is greater than a preset correlation threshold, extracts a second logical coupling degree when detecting in the ending direction, divides the subsequent sentence into the slicing range when the second logical coupling degree is greater than the preset correlation threshold, stops the detection until an adjacent semantic anchor point is extracted or the second logical coupling degree is lower than a preset breaking threshold, and generates a variable-length logical slice based on the sentence boundary divided into the slicing range. The application dynamically identifies a logical boundary through a semantic anchor point, eliminates the semantic breaking problem caused by mechanical cutting, and greatly improves the quality of knowledge management.

Description

Technical Field

[0001] This invention relates to the field of information processing, and in particular to a knowledge management reasoning system based on precise location retrieval. Background Technology

[0002] With the rapid development of natural language processing and large-scale language models, knowledge management systems are playing an increasingly prominent role in processing complex unstructured data and assisting in decision-making and reasoning. The core of modern knowledge base construction lies in transforming massive amounts of documents into computable semantic vectors, thereby achieving efficient information retrieval and knowledge association. This technical architecture is widely used in scientific research, engineering, and production processes, and has significant technical value in improving the automation level and scientific rigor of decision-making in complex information systems.

[0003] Document slicing and vector indexing techniques are crucial for determining the quality of knowledge reasoning, aiming to effectively extract and map knowledge points through preprocessing of the original corpus. To ensure retrieval precision, the system needs to perform fine-grained boundary segmentation of documents to construct accurate logical units within complex knowledge structures. An ideal retrieval system should be able to accurately identify the semantic core of the text, ensuring that each extracted knowledge element is logically consistent and complete.

[0004] However, existing technologies generally employ fixed-length physical slicing mechanisms, which easily lead to the mechanical truncation of complete technical definitions or causal logic chains in different data blocks, resulting in a severe loss of semantic integrity. To alleviate the context fragmentation problem, traditional methods often rely on redundant slices with overlapping regions. This not only significantly increases the resource overhead of vector storage and computation but also introduces a large amount of irrelevant background noise, directly interfering with the accuracy of knowledge reasoning. Furthermore, existing retrieval models lack the ability to parse deep logical anchor points in documents, making it difficult to achieve atomic-level precise positioning of specific operational steps or technical parameters within a document. This results in output results often containing a large amount of irrelevant information, failing to meet the needs of high-precision knowledge services.

[0005] Therefore, a knowledge management reasoning system based on precise location retrieval is desired. Summary of the Invention

[0006] The purpose of this invention is to provide a knowledge management reasoning system based on precise location retrieval, which can effectively solve the problems of semantic integrity impairment, low retrieval signal-to-noise ratio, and ambiguous location accuracy in the aforementioned background technologies. A knowledge management reasoning system based on precise location retrieval includes: The semantic parsing device is configured to receive the original unstructured document, perform word segmentation and part-of-speech tagging, and construct a dependency tree with predicates as the core nodes. Based on the dependency strength values ​​between words and their connected nodes in the dependency tree, it filters the modifiers of preset parts of speech and extracts the logical backbone. By matching a preset logical connection lexicon, it marks specific words with logical guidance functions in the logical backbone and segments the sentences containing the specific words into precondition nodes and execution action nodes. An anchor point identification device, connected to the semantic parsing device, is configured to receive the logical backbone, calculate the comprehensive semantic weight of each word based on the part-of-speech weight score of the word, the degree center score of the word in the dependency relationship tree, and the domain relevance score obtained by querying the built-in technical field dictionary, mark the word with the comprehensive semantic weight greater than a preset threshold as a semantic anchor point; and when a causal logic chain containing the precondition node and the execution action node is extracted, set the execution action node in the causal logic chain as the core action anchor point; A dynamic slice generation device, connected to the anchor point recognition device, is configured to perform bidirectional logical correlation detection in both the start and end directions of the text, centered on the semantic anchor point. When detecting in the start direction, it extracts precondition features, constraint parameter features, environmental description features, or logical subject definition features from the preceding statement to calculate a first logical coupling degree with the semantic anchor point. If the first logical coupling degree is greater than a preset correlation threshold, the preceding statement is included in the slice range. When detecting in the end direction, it extracts supplementary explanation features, effect feedback features, anomaly handling features, or numerical constraint features from the following statement to calculate a second logical coupling degree with the semantic anchor point. If the second logical coupling degree is greater than the preset correlation threshold, the following statement is included in the slice range. Detection continues until an adjacent semantic anchor point is extracted or the second logical coupling degree is lower than a preset breakage threshold. Variable-length logical slices are generated based on the boundaries of the statements included in the slice range. The retrieval adaptation device is connected to the dynamic slice generation device and the preset index library. It is configured to perform word segmentation and semantic vectorization processing on the received query request string to extract the query vector, calculate the dimensional feature parameters and information entropy value of the query vector, calculate and extract the knowledge granularity level based on the dimensional feature parameters and the information entropy value, and extract the corresponding single variable-length logical slice or a composite knowledge unit composed of multiple related variable-length logical slices from the index library based on the knowledge granularity level and output it. The management and monitoring device is communicatively connected to the semantic parsing device, the anchor point recognition device, the dynamic slice generation device, and the retrieval adaptation device, respectively. It is configured to collect the operating status parameters and computing resource consumption data of each device. When the document parsing delay time of the semantic parsing device exceeds a predetermined delay threshold, it triggers a load balancing mechanism to schedule the central processing unit and graphics processing unit resources of each computing node.

[0007] Compared with the prior art, the present invention has the following beneficial effects: 1. Ensures the logical integrity of knowledge points: This invention dynamically identifies logical boundaries through semantic anchors, abandoning the traditional fixed-length physical slicing mode, ensuring that each extracted knowledge element is semantically self-consistent and coherent, fundamentally eliminating the semantic breakage problem caused by mechanical cutting, and greatly improving the quality of knowledge management.

[0008] 2. Significantly improved precision in knowledge retrieval: By adopting atomic-level precise positioning technology based on logical backbone, the system can directly map query requests to the most relevant logical units, avoiding interference from irrelevant and redundant information, making the retrieval results more accurate and efficient, and meeting the high-precision requirements for knowledge acquisition in high-precision engineering fields.

[0009] 3. Optimized computing resource utilization and reduced interference: The variable-length logical slices generated by this invention eliminate a large number of overlapping areas set up by conventional techniques to maintain context while ensuring information integrity. This significantly reduces the storage scale of vector indexes, reduces the computational load of the backend inference model when processing redundant contexts, and improves the system's response speed and stability.

[0010] 4. Enhanced intelligent adaptability of the system: Through the analysis of query complexity and dynamic control of knowledge granularity by the retrieval adaptation device, the system can provide knowledge services of different depths according to the needs of different users, realizing a smooth transition from macro-level document guidance to micro-level atomic knowledge positioning, and has extremely strong application flexibility.

[0011] 5. Possesses excellent generalization ability and industry application prospects: The logical analysis and anchor point recognition mechanism of this invention does not depend on specific document formats and can be widely applied to the automated processing of various complex texts such as scientific research reports, engineering procedures, and legal provisions, providing core technical support for building a high-performance industry professional knowledge base. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall technical architecture of a knowledge management reasoning system based on precise location retrieval proposed in this invention. Detailed Implementation

[0013] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0014] like Figure 1 A knowledge management reasoning system based on precise location retrieval includes a semantic parsing device, an anchor point recognition device, a dynamic slice generation device, a retrieval adaptation device, and a management and monitoring device. The semantic parsing device receives and preprocesses raw unstructured documents. Through syntactic dependency analysis and semantic topology construction, it identifies and extracts the logical backbone of the document, transforming the linear text stream into a structured representation with logical hierarchy. As the input processing hub of the entire system, the semantic parsing device's core function is to eliminate linguistic noise in the raw text and reorganize the scattered text sequence into a low-level skeleton reflecting the technical logic. The semantic parsing device is internally configured with a document format standardization unit, which is configured to uniformly process unstructured documents from different data sources. These data sources include, but are not limited to, portable document formats, word processing software formats, Hypertext Markup Language (HTML) formats, and plain text formats. The document format standardization unit, through a built-in encoding conversion plugin, converts various documents into a unified markup language format, such as Extensible Markup Language (XML), while preserving paragraph hierarchy, heading weights, and chart index information from the original document during the conversion process. The semantic parsing device further includes a syntactic dependency analysis unit, which employs a multi-threaded parallel processing mode to perform parallel syntax parsing on a large-scale document library. The syntactic dependency analysis unit invokes a built-in linguistic processing unit to perform in-depth word segmentation and part-of-speech tagging on the standardized text. During word segmentation, the linguistic processing unit identifies each basic semantic unit in the text by matching a pre-defined general dictionary with a domain-specific dictionary. In the part-of-speech tagging stage, the linguistic processing unit identifies the subject, predicate, object, and various adverbial, attributive, and complement components in the text.

[0015] After completing part-of-speech tagging, the semantic parsing device constructs a dependency tree through the syntactic dependency analysis unit. This dependency tree reflects the dominance and subordination relationships between words, with predicates typically serving as core nodes, and other components attached to corresponding branches based on their semantic connection strength with the predicate. The semantic parsing device analyzes the dependency strength between predicates and their arguments, performing semantic pruning to filter out non-substantive modifying components such as adjectives and adverbs, thereby identifying the logical backbone containing core engineering logic or technical parameters. Furthermore, the semantic parsing device is equipped with a logical relationship recognition unit, which identifies deep logical connections in the text, including but not limited to causal relationships, progressive relationships, parallel relationships, adversative relationships, hypothetical relationships, and purposive relationships. This logical relationship recognition unit matches a pre-defined logical connection lexicon, marking specific words with logical guidance functions, such as "if," "then," "because," "therefore," "and," and "but," thereby determining the start and end points of the logical chain. When a statement such as "If the pressure exceeds 10 MPa, the pressure relief valve will be automatically triggered" is detected, the logic relationship identification unit locks the complete logical structure of the causal chain through the logic connection unit, identifies "pressure exceeds 10 MPa" as a prerequisite, and identifies "automatically trigger the pressure relief valve" as an action to be performed.

[0016] The anchor point recognition device is connected to the semantic parsing device and is used to locate semantic nodes representing core knowledge points or technical actions based on the extracted logical backbone, and solidify them as semantic anchor points as the logical core for subsequent knowledge segmentation. The anchor point recognition device receives the logical backbone information output by the semantic parsing device and performs in-depth semantic weight evaluation on it. The anchor point recognition device integrates a weight calculation unit, which calculates the semantic weight of each word in the logical backbone and selects words with weight values ​​higher than a preset threshold as candidate anchor points. The semantic weight is determined by a combination of scores from multiple dimensions, specifically including the word's part-of-speech weight score, the word's degree centrality score in the semantic topology graph, and the domain relevance score. The part-of-speech weight score is allocated according to a preset priority table; for example, the weight scores of verbs and core nouns are set higher than those of adjectives and function words. The degree centrality score reflects the tightness of the connection of the word in the dependency tree, specifically the number of edges connected to the node; the more edges a node has, the higher its degree centrality score. The domain relevance score is obtained by querying a dynamically updated technical field dictionary built into the anchor point recognition device. This technical field dictionary covers core terms, abbreviations, process names, and parameter units for a specific industry. When a term is identified as a core technical parameter or key action verb in that industry, its domain relevance score is assigned a maximum value.

[0017] After calculating the comprehensive weight of candidate anchor points, the anchor point identification device identifies action verbs or key technical parameter nodes with top weight values ​​and clear technical meaning as formal semantic anchor points. In the context of engineering technical specifications, verbs such as "trigger," "start," "shut down," and "adjust," as well as parameter nouns such as "temperature," "pressure," "flow rate," and "voltage," are typically fixed as semantic anchor points. For causal logic chains, the anchor point identification device defines the core execution action in the chain as the core action anchor point and assigns it the highest level of retrieval priority. For example, in the logical backbone of "automatically triggering the pressure relief valve," "trigger" is identified as the core action anchor point, carrying the core technical value of this text.

[0018] The dynamic slice generation device is configured to perform bidirectional logical correlation detection centered on the semantic anchor point, moving towards the start and end directions of the text. It dynamically identifies the logical boundaries of knowledge elements, thereby generating logical slices of variable length, ensuring the self-consistency and completeness of knowledge expression within each slice. This dynamic slice generation device completely abandons the traditional physical slicing mode that truncates to a fixed number of characters. When performing forward logical correlation detection, the dynamic slice generation device starts from the selected semantic anchor point and traces back sentence by sentence to the preceding content. During the tracing process, the dynamic slice generation device calculates the logical coupling degree between each preceding statement and the current semantic anchor point. The determination of the logical coupling degree is based on whether the preceding statement contains the preconditions, constraint parameters, environmental descriptions, or logical subject definitions of the current anchor point. If the logical coupling degree between a preceding statement and the current anchor point is higher than a preset correlation threshold, the dynamic slice generation device includes that statement in the current slice and continues to probe upwards.

[0019] Simultaneously, when performing backward logical correlation probing, the dynamic slice generation device starts from the semantic anchor point and probes downwards for the semantic extensibility of subsequent statements. The device analyzes whether subsequent statements supplement the current anchor point, provide effect feedback, handle anomalies, or impose numerical constraints to determine whether to expand the slice boundary. If a subsequent statement further explains the action of the current anchor point, such as "Note: the pressure relief time should not exceed 5 seconds," the device uses a logical correlation determination algorithm to include this explanatory text in the current atomic knowledge slice. This backward probing process continues until a new independent semantic anchor point is encountered, or the logical coupling is detected to be below a preset breakage threshold. Through this bidirectional probing mechanism, the system can generate variable-length logical slices that reflect complete logical units.

[0020] The dynamic slice generation device also features numerical logic protection. When processing text containing numerical logic, if a preset physical quantity limit or time limit is identified, the dynamic slice generation device automatically treats the numerical value and its corresponding physical unit as part of the logical core, ensuring that the numerical value and its defined physical object are completely contained within the same logical slice, preventing the numerical value from being decoupled from its context. Furthermore, the generated variable-length logical slices undergo a secondary verification by the correlation verification unit before being stored in the index. If the length of a slice exceeds the predetermined physical length limit due to an excessively long logical chain, the correlation verification unit will search for secondary weak correlation points within the slice for secondary splitting. The selection of these secondary weak correlation points is based on the local minimum of semantic correlation to ensure that high retrieval efficiency and semantic concentration balance are maintained after splitting. The dynamic slice generation device also features logical conflict detection. When two adjacent slices are detected to overlap or contradict each other at their logical boundaries, the device calculates their global semantic similarity and performs a merging or forced splitting operation to maintain the uniqueness of the index.

[0021] The retrieval adaptation device receives user-input query requests and performs semantic dimensionality reduction analysis on the complexity of the request. It then dynamically matches and extracts atomic-level knowledge elements of corresponding granularity from a pre-defined index, achieving precise location and output of the target knowledge. The retrieval adaptation device internally includes a request parsing unit, a granularity control unit, a feedback correction unit, and a context reconstruction unit. The request parsing unit performs word segmentation, noise reduction, and semantic vectorization on the user-input query string. The granularity control unit determines the required knowledge granularity level based on the dimensional characteristics and information entropy of the query vector. For high-complexity query requests, such as a user asking "the complete operation procedure and precautions of a pressure relief system," the granularity control unit configures the retrieval adaptation device to retrieve composite knowledge units containing multiple logical slices from the index to provide a comprehensive knowledge service. When receiving low-complexity or parameter-specific query requests, such as "what is the trigger pressure of the pressure relief valve," the granularity control unit configures the retrieval adaptation device to extract only a single atomic-level knowledge element directly related to that parameter.

[0022] The retrieval adaptation device employs a hybrid retrieval strategy, combining literal keyword matching based on inverted indexes with deep semantic vector retrieval based on a multi-level vector space model. The device calculates the semantic distance between candidate slice vectors and query request vectors, and then reorders the retrieval results based on logical granularity weights. Furthermore, the device features context reconstruction capabilities. When extracted atomic-level knowledge elements, while precise, still require background information for logical expression, the device automatically retrieves adjacent logical slices from the original document to provide necessary background information to the user. In addition, the feedback correction unit dynamically adjusts the weight distribution of semantic anchors and the threshold for determining logical relevance based on user click behavior, dwell time, and explicit evaluations of the retrieval results. Through this closed-loop learning mechanism, the system achieves self-evolution of retrieval performance, increasingly aligning with the query habits of users in specific domains.

[0023] The management and monitoring device is used to monitor the operating status and computing resource consumption of each module in real time. The management and monitoring device establishes bidirectional communication connections with the semantic parsing device, anchor point recognition device, dynamic slice generation device, and retrieval adaptation device. When the semantic parsing device experiences parsing latency exceeding a predetermined threshold while processing extremely large documents, the management and monitoring device automatically triggers a load balancing mechanism, dynamically scheduling the CPU and GPU resources of computing nodes to improve the system's parallel processing capability. Simultaneously, the management and monitoring device is also responsible for maintaining the health of the index database, periodically cleaning up redundant indexes, and monitoring core technical indicators such as precision and response time of vector retrieval.

[0024] The index library employs a multi-level vector space model, storing logical slices hierarchically according to semantic topics, logical depth, and atomicity. Each logical slice is associated with detailed metadata in the index library, including the path information of its original document, its logical hierarchical position within the document, associated semantic anchor identifiers, and its industry category tag. This multi-dimensional storage architecture provides underlying support for rapid retrieval and multi-granularity positioning. Example 2

[0025] This embodiment, based on Embodiment 1, further illustrates the system's operational details and hardware implementation through specific engineering examples. In this embodiment, the knowledge management reasoning system based on precise location retrieval is deployed in a distributed computing environment to address the knowledge management needs of massive amounts of engineering drawings, technical specifications, and operation guides.

[0026] The semantic parsing device relies on a server cluster equipped with high-performance graphics processing units (GPUs) for its underlying hardware. When performing syntactic dependency analysis, the semantic parsing device leverages the computing power of the GPUs to accelerate tensor operations, significantly improving the speed of dependency tree construction. The linguistic processing unit pre-stores a professional ontology library in the engineering field, which defines the hierarchical topological relationships between engineering components, processes, and physical attributes. When processing a technical manual on a nuclear power plant's secondary loop depressurization system, the semantic parsing device uses a multi-threaded parallel processing mode to simultaneously parse the text descriptions, tabular data, and diagram annotations in the manual.

[0027] In this embodiment, the anchor point recognition device incorporates a statistical feature-based mechanism for identifying uncommon words. When uncommon terms or names of new devices appear outside the technical field dictionary in a document, the anchor point recognition device does not ignore them directly. Instead, it calculates the degree of aggregation and degree of freedom of the uncommon word in a specific context using contextual statistical features. If the word frequently appears in multiple predicate positions with similar logical structures, the anchor point recognition device identifies it as a potential new semantic anchor and assigns it a temporary weight. The dynamically updated technical field dictionary is periodically expanded based on these identification results, thereby ensuring the system's generalization ability for new technology documents.

[0028] The dynamic slice generation device employs a strategy combining a sliding window and an association probability model when performing logical association detection. The size of the sliding window is dynamically adjusted based on the attributes of the semantic anchors. For action-type anchors, the sliding window probes further backward to capture the complete action instruction sequence; for parameter-type anchors, the sliding window focuses more on probing forward to its defined physical boundaries and preconditions. The logical coupling degree calculated by the dynamic slice generation device depends not only on the lexical overlap but also on the cosine similarity of the semantic vectors. The calculation process for the cosine similarity is described as follows: extract the feature vectors of two sentences, calculate the dot product of these two feature vectors, and divide the dot product by the product of the magnitudes of the two feature vectors. The quotient is the cosine similarity value. If this value is greater than a preset association threshold, the logical association is determined to be valid.

[0029] In this embodiment, when processing a maintenance procedure containing tightly nested logic, the dynamic slice generation device identifies the statement "If abnormal flow fluctuations are detected when the reactor is in a low-pressure shutdown state and the residual heat removal system is in operation, the main pump power supply should be immediately cut off." The dynamic slice generation device uses "cut off" as the core action anchor point, detecting two parallel preconditions: "the reactor is in a low-pressure shutdown state" and "the residual heat removal system is in operation"; simultaneously, it identifies a trigger condition: "abnormal flow fluctuations are detected." The dynamic slice generation device integrates all the logically related content into a single atomic-level knowledge slice. Compared to traditional fixed 512-character slices, the variable-length slices generated by this system effectively avoid forcibly separating preconditions and execution actions, thus ensuring the self-consistency of the retrieved knowledge elements.

[0030] The retrieval adaptation device is connected to the front-end application and back-end storage array via a high-speed network interface at the hardware level. The built-in request parsing unit of the retrieval adaptation device can identify implicit logical requirements in natural language queries. When a user inputs "response measures when the system is abnormal," the request parsing unit decomposes the query request into two semantic dimensions: "abnormal scenario" and "processing action," through semantic dimensionality reduction analysis. The granularity control unit then matches all variable-length logical slices containing "processing action" anchors and whose preconditions involve "abnormality," based on the distribution of knowledge elements in the index. The retrieval adaptation device calculates the comprehensive similarity between each candidate slice and the query request by executing a hybrid retrieval strategy. The comprehensive similarity is calculated by multiplying the literal similarity score by a first preset weight, multiplying the semantic vector similarity score by a second preset weight, adding the two together, and finally weighting and correcting according to the logical granularity of the slice. The result is used as the ranking basis.

[0031] The retrieval adaptation device also integrates a context-based rapid reconstruction unit. For frequently queried atomic-level knowledge elements, the context reconstruction unit pre-loads their related logical slices into a high-speed random access memory. Once a user clicks to view detailed background, the system can output background information in milliseconds without rescanning the original document. The feedback correction unit collects user interaction trajectories, such as mouse hover positions, copied and pasted content, and corrected keywords after the query, and uses reinforcement learning algorithms to adjust internal logical judgment thresholds. For example, if a large number of users actively consult a subsequent paragraph after retrieving a certain atomic slice, the feedback correction unit will automatically increase the logical coupling evaluation between these two paragraphs, so that they are merged into a more logically complete slice in the subsequently generated index.

[0032] In this embodiment, the management and monitoring device acts as a system performance scheduler. Because semantic parsing and vector computation exhibit periodic fluctuations in memory and computing resource usage, the management and monitoring device monitors memory usage, CPU load percentage, and disk read / write bandwidth in real time. When the remaining memory of the current computing node falls below a predetermined safety threshold, the management and monitoring device automatically migrates new document parsing tasks to other physical servers with lower loads in the cluster and reprioritizes existing tasks. This resource scheduling mechanism ensures that even when processing terabyte-scale engineering document libraries, the system maintains stable index building efficiency and retrieval response time. Example 3

[0033] This embodiment focuses on describing a cloud-based integrated architecture based on this system, used for a large-scale knowledge reasoning platform across industries. In this architecture, the system provides knowledge retrieval and location capabilities externally through a microservice model.

[0034] The semantic parsing device, as a core node of the microservice cluster, receives document processing requests from clients in different industries via an API gateway. The document format standardization unit within the semantic parsing device supports the parsing of multilingual documents and incorporates dependency analysis models for different languages. When processing multilingual technical standards in multinational engineering projects, the semantic parsing device can extract the cross-language logical framework and utilize a multilingual pre-trained model to map terms from different languages ​​into a unified cross-language vector space.

[0035] In this embodiment, the anchor point identification device is equipped with an industry isolation mechanism. For different industries such as chemical, aerospace, and power, the anchor point identification device activates the corresponding industry knowledge graph to assist in the accurate positioning of anchor points. In the chemical industry mode, the anchor point identification device defines chemical reaction equations and valve position states in the process flow as high-priority anchor points. In the aerospace manufacturing mode, it focuses more on assembly tolerances, process steps, and quality inspection parameters. The weight calculation unit adjusts the ratio between the part-of-speech weight score and the domain relevance score according to the currently activated industry model to achieve deep adaptation to the vertical domain.

[0036] The dynamic slice generation device employs an incremental indexing strategy in a cloud environment. When the original document library undergoes partial updates or revisions, the dynamic slice generation device does not need to re-slice the entire document; instead, it performs bidirectional probing only on the changed paragraphs and the logical chains within their affected scope. The logical conflict detection function plays a crucial role at this time, accurately updating the affected vector indexes by comparing the semantic similarity and anchor point distribution between the old and new logical slices. This incremental processing method significantly reduces the maintenance costs and energy consumption of the cloud system.

[0037] The retrieval adaptation device supports multi-tenant personalized configuration within a cloud architecture. Different enterprise tenants can set different retrieval granularity preferences based on their own business scenarios. For example, research users tend to obtain long logic slices containing complete derivation processes, while on-site operation users prefer to obtain extremely concise instruction-type atomic slices. The granularity control unit dynamically adjusts the judgment threshold of logical coupling based on the tenant's preference configuration file. In high-precision industrial maintenance scenarios, the retrieval adaptation device, through its interface with augmented reality devices, pushes precisely located atomic-level knowledge elements to the operator's field of vision in real time. When the operator scans a specific equipment nameplate using the augmented reality device, the system automatically initiates a query request based on that equipment identifier. The retrieval adaptation device directly locates and returns the corresponding operation logic slice for that equipment, achieving a deep integration of knowledge and actual operational scenarios.

[0038] The management and monitoring device, integrated within a cloud architecture, undertakes more complex security auditing and multi-tenant isolation monitoring tasks. It monitors each tenant's usage of computing resources and allocates bandwidth and concurrency according to a pre-defined service level agreement. Simultaneously, through deep packet inspection of retrieval traffic, it identifies and intercepts abnormal crawling requests, protecting the intellectual property security of the underlying knowledge base. In the event of a system-level failure, the management and monitoring device can trigger cross-regional disaster recovery switching, seamlessly migrating the retrieval service to a remote cloud data center, ensuring the continuous high availability of the knowledge reasoning service.

[0039] Furthermore, the index library in this embodiment not only stores semantic vectors but also the topological relationships between logical slices. This graph-vector hybrid storage structure allows the system to perform complex logical path retrieval. For example, a user can query "all cascading protection logic related to pressure increase." The retrieval adapter first locates the anchor points related to "pressure increase" through vector retrieval, and then, based on the logical topological relationships stored in the index library, extracts all related trigger action slices and cascading protection slices, thereby forming a complete knowledge graph path. This multi-dimensional knowledge association capability enables the system to evolve from simple "information retrieval" to "knowledge navigation" with preliminary logical reasoning capabilities. Example 4

[0040] This embodiment describes a special application configuration of the system when processing high-density, multi-parameter engineering indicator documents. In some high-precision manufacturing fields, technical documents are filled with a large number of technical indicator values ​​and complex constraints, which places extremely high demands on the accuracy of knowledge location.

[0041] In this embodiment, the semantic parsing device enhances its ability to extract parameterized expressions. The logical relation recognition unit not only identifies textual logical connectors but also recognizes mathematical logical relation descriptions within the text. Although not expressed using algebraic formulas, the logical relation recognition unit can accurately identify descriptive logical quantifiers such as "not less than," "between the two," and "increasing linearly." These quantifiers are mapped to logical edges with specific directions, connecting parameter nodes and numerical nodes in the dependency tree.

[0042] For high-density parameter documents, the anchor point recognition device employs a composite anchor point recognition strategy. This device not only defines single verbs or parameter nouns as anchor points, but also defines composite units consisting of "parameters plus corresponding thresholds" as semantic anchor points. For example, in the context of "when the inlet temperature reaches 250 degrees Celsius and lasts for more than 300 seconds," the anchor point recognition device uses both "temperature reaches 250 degrees Celsius" and "lasts for 300 seconds" as a composite anchor point. This composite anchor point mechanism improves the stability of recognition under complex technical constraints.

[0043] When processing such documents, the dynamic slice generation device incorporates a logical integrity self-checking unit. After generating a logical slice, this unit simulates a logical integrity check on that slice. For example, if a slice mentions "adjusting the voltage to the rated value," but the slice does not contain a specific value or a description of its source, the self-checking unit determines that the slice's logic is incomplete. In this case, the dynamic slice generation device automatically triggers a second round of bidirectional probing, expanding the search range to find missing logical elements until the semantic information within the slice can be closed looped. This mechanism ensures that even extremely small atomic slices possess independent reference value.

[0044] The retrieval adaptation device has been deeply optimized for parameter queries. When the granularity control unit recognizes that the query request involves the location of a specific physical quantity, it automatically switches to "granularity mode." In this mode, the retrieval adaptation device uses high-dimensional semantic alignment technology to accurately match the parameter requirements in the query request with the numerical descriptions in the slices. Furthermore, the context reconstruction unit also has a multi-dimensional association display function. When a user queries the limit value of a specific parameter, the retrieval adaptation device not only returns the logical slice containing that value but also automatically associates and displays other limit value slices of that parameter under different operating conditions, providing the user with a comprehensive parameter comparison view.

[0045] In this embodiment, the management and monitoring device adds a real-time verification function for the model's localization accuracy. The device periodically extracts samples from the index and uses a preset benchmark set to evaluate the accuracy of semantic parsing and anchor point recognition. If a trend of declining localization accuracy is detected for a specific document format, the device automatically triggers parameter fine-tuning procedures for relevant devices or sends a reminder to the system administrator to re-annotate training samples, ensuring that the system's knowledge management quality remains at a high level.

[0046] This embodiment further illustrates the system's robustness in handling complex engineering logic. Through this layered, logically interlocked device configuration, the present invention effectively solves the core problems of semantic fragmentation, information redundancy, and low retrieval accuracy in the prior art when processing unstructured technical documents, providing a solid technical path for building an industrial-grade atomic knowledge location system.

[0047] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0048] At the software architecture level, all devices and units involved in this invention can be implemented using instruction sets stored in a computer-readable storage medium. When executed by a processor, these instruction sets can drive hardware devices to complete the aforementioned complex semantic analysis, anchor point location, and dynamic slicing logic. Simultaneously, the distributed storage and computing architecture adopted in this invention can achieve flexible horizontal scaling according to business scale. For those skilled in the art, this invention is not limited to the specific hardware environment or software version described above; any variations based on the core logic of this invention fall within the protection scope of this patent. In actual deployment, the communication protocol between modules can be selected in synchronous or asynchronous mode according to real-time requirements, and various encryption algorithms can be introduced during data transmission to ensure the security of knowledge assets. Through a systematic architectural design, this invention achieves end-to-end automation from text input to precise atomic knowledge output, greatly improving the level of intelligence in knowledge management.

Claims

1. A knowledge management reasoning system based on precise location retrieval, characterized in that, include: The semantic parsing device is configured to receive the original unstructured document, perform word segmentation and part-of-speech tagging, and construct a dependency tree with predicates as the core nodes. Based on the dependency strength values ​​between words and their connected nodes in the dependency tree, the device filters the modifiers of preset parts of speech and extracts the logical backbone. By matching a preset logical connection lexicon to mark specific words with logical guidance functions in the logical backbone, the statement containing the specific words is divided into precondition nodes and execution action nodes. An anchor point identification device, connected to the semantic parsing device, is configured to receive the logical backbone, calculate the comprehensive semantic weight of each word based on the part-of-speech weight score of the word, the degree center score of the word in the dependency relationship tree, and the domain relevance score obtained by querying the built-in technical field dictionary, mark the word with the comprehensive semantic weight greater than a preset threshold as a semantic anchor point; and when a causal logic chain containing the precondition node and the execution action node is extracted, set the execution action node in the causal logic chain as the core action anchor point; A dynamic slice generation device, connected to the anchor point recognition device, is configured to perform bidirectional logical correlation detection in both the start and end directions of the text, centered on the semantic anchor point. When detecting in the start direction, it extracts precondition features, constraint parameter features, environmental description features, or logical subject definition features from the preceding statement to calculate a first logical coupling degree with the semantic anchor point. If the first logical coupling degree is greater than a preset correlation threshold, the preceding statement is included in the slice range. When detecting in the end direction, it extracts supplementary explanation features, effect feedback features, anomaly handling features, or numerical constraint features from the following statement to calculate a second logical coupling degree with the semantic anchor point. If the second logical coupling degree is greater than the preset correlation threshold, the following statement is included in the slice range. Detection continues until an adjacent semantic anchor point is extracted or the second logical coupling degree is lower than a preset breakage threshold. Variable-length logical slices are generated based on the boundaries of the statements included in the slice range. The retrieval adaptation device is connected to the dynamic slice generation device and the preset index library. It is configured to perform word segmentation and semantic vectorization processing on the received query request string to extract the query vector, calculate the dimensional feature parameters and information entropy value of the query vector, calculate and extract the knowledge granularity level based on the dimensional feature parameters and the information entropy value, and extract the corresponding single variable-length logical slice or a composite knowledge unit composed of multiple related variable-length logical slices from the index library based on the knowledge granularity level and output it. The management and monitoring device is communicatively connected to the semantic parsing device, the anchor point recognition device, the dynamic slice generation device, and the retrieval adaptation device, respectively. It is configured to collect the operating status parameters and computing resource consumption data of each device. When the document parsing delay time of the semantic parsing device exceeds a predetermined delay threshold, it triggers a load balancing mechanism to schedule the central processing unit and graphics processing unit resources of each computing node.

2. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The semantic parsing device is internally equipped with a document format standardization unit; The document format standardization unit has a built-in encoding conversion plugin configured to convert the original unstructured documents from different data source formats into Extensible Markup Language (XML) format, and to extract and retain the paragraph level coordinates, heading weight values, and chart index data of the original unstructured documents in the converted XML format text.

3. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The dynamic slice generation device is internally equipped with a numerical logic protection module and an correlation verification unit; The numerical logic protection function module is configured to, when extracting physical quantity limits or time limits from the text, divide the extracted physical quantity limits or time limits, the corresponding physical units, and the limited physical object data into the same variable-length logic slice. The correlation verification unit is configured to calculate the semantic correlation value between each statement node within the variable-length logic slice when the physical length value of the generated variable-length logic slice exceeds a predetermined physical length limit, extract the coordinates of the local minimum point of the semantic correlation value as a secondary weak correlation point, and perform a secondary boundary splitting operation on the variable-length logic slice along the coordinates of the secondary weak correlation point.

4. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The anchor point recognition device is equipped with a rare word recognition mechanism based on statistical features; The anchor point recognition device is configured to extract the context text of the rare word when it is extracted and calculate the aggregation degree parameter and the degree of freedom parameter of the rare word in the current context. When the frequency of occurrence of the rare word in multiple predicate positions with the same syntactic dependency structure is found to be greater than a preset frequency threshold, the anchor point recognition device marks the rare word as a temporary semantic anchor point and assigns a temporary weight score, and writes the characters of the rare word into the update queue of the technical field dictionary.

5. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The dynamic slice generation device invokes a sliding window detection mechanism configured with an association probability model when performing the bidirectional logical correlation detection. The dynamic slice generation device is configured to dynamically adjust the detection direction and span parameters of the sliding window according to the attributes of the semantic anchor point: when the extracted center node is the core action anchor point, the detection depth value of the sliding window in the end direction is set to be greater than the detection depth value in the start direction; when the extracted center node is a parameter-type anchor point containing technical parameter terms, the detection depth value of the sliding window in the start direction is set to be greater than the detection depth value in the end direction. The calculation parameters for the first logical coupling degree and the second logical coupling degree are obtained by calculating the cosine similarity value. The algorithm for calculating the cosine similarity value is as follows: extract the feature vectors of the two statements to be compared, perform a dot product operation, and divide the result of the dot product operation by the product value of the magnitudes of the two feature vectors.

6. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The retrieval adaptation device performs extraction operations based on the knowledge granularity level through its internal granularity control unit: When the extracted information entropy value is greater than the high complexity determination threshold, the granularity control unit generates a composite knowledge unit extraction instruction to extract the variable-length logical slice containing multiple logical associations from the index library; when the extracted information entropy value is less than or equal to the high complexity determination threshold, the granularity control unit generates an extraction instruction to extract a single variable-length logical slice from the index library. The retrieval adaptation device performs a ranking calculation based on a hybrid retrieval strategy: it calculates the literal similarity score between the candidate variable-length logical slice vector and the query vector and multiplies it by a first preset weight parameter to obtain a first product; it calculates the semantic vector similarity score between the two under the multi-level vector space model and multiplies it by a second preset weight parameter to obtain a second product; it adds the first product and the second product together and then performs a weighted conversion with the logical granularity value of the extracted slice to obtain a re-ranked comprehensive similarity value; and it outputs the final result according to the descending sequence of the re-ranked comprehensive similarity value.

7. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The retrieval adaptation device also includes a context reconstruction unit and a feedback correction unit; The context reconstruction unit is configured to pre-extract the target variable-length logical slice whose query frequency is higher than a preset frequency threshold, and load the associated logical slices that are adjacent to each other in the original unstructured document into the random access memory, and read and output the associated logical slices from the random access memory when a background retrieval instruction is received. The feedback correction unit is configured to collect the mouse hover position coordinates returned by the user terminal, the copied and pasted paragraph text, and the correction keyword characters input in the secondary query as user interaction trajectory data. The user interaction trajectory data is imported into the reinforcement learning algorithm model. The distribution value of the comprehensive semantic weight of the semantic anchor in the anchor recognition device is updated according to the feedback parameters output by the reinforcement learning algorithm model. The preset association threshold and the preset breakage threshold recorded in the dynamic slice generation device are also modified.

8. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The dynamic slice generation device is equipped with an incremental slice generation mechanism and a logic conflict detection unit. The dynamic slice generation device is configured to, upon receiving document text with local revision identifiers, extract the changed paragraph coordinates and their associated logical chain regions, and perform the bidirectional logical association detection in incremental mode to generate updated logical slices for the changed paragraph coordinates and their associated logical chain regions. The logical conflict detection unit is configured to extract the feature vectors of the newly generated variable-length logical slice and the adjacent variable-length logical slice when the text overlaps at the boundary position, and calculate the global semantic similarity parameter. Based on the interval distribution of the global semantic similarity parameter, the variable-length logical slices that have text overlap are either merged or forced to be segmented along the text boundary coordinates.

9. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The preset index library adopts a graph-vector hybrid multi-dimensional storage architecture; The preset index library stores each variable-length logical slice in layers according to semantic topic, logical depth and atomicity, and binds a multi-dimensional metadata field to each variable-length logical slice in the data table. The multi-dimensional metadata field includes the network path characters of the original unstructured document to which it belongs, the logical level position coordinates within the document, the globally unique identifier of the associated semantic anchor point and the industry classification label code. The preset index library stores topology path graph edge data for recording the context association order between the variable-length logical slices. The retrieval adaptation device is configured to extract the connected trigger action slice node data and chain protection slice node data based on the topology path graph edge data after locking the variable-length logical slice where the target semantic anchor point is located, and concatenate the variable-length logical slice where the target semantic anchor point is located with the extracted node data to form a logical association path set.

10. The knowledge management reasoning system based on precise location retrieval according to claim 1, characterized in that, The logical relationship recognition unit in the semantic parsing device is configured with parameterized expression extraction rules. The logical relationship recognition unit extracts mathematical logical relationship description quantifiers in the text according to the parameterized expression extraction rules, converts the mathematical logical relationship description quantifiers into logical edge data with directional features, and attaches the logical edge data between the parameter name node and the parameter limit value node in the dependency relationship tree. The anchor point identification device is equipped with a composite anchor point generation strategy, which combines and binds the extracted parameter name node and the logically connected parameter limit node to confirm it as a composite semantic anchor point. The dynamic slice generation device is equipped with a logic integrity self-checking unit. The logic integrity self-checking unit is configured to increase the detection span parameter range of the sliding window and trigger a secondary bidirectional logic correlation detection operation when the generated variable-length logic slice contains numerical adjustment action nodes and lacks matching quantized numerical node data, until the matching quantized numerical node data is extracted within the redefined slice boundary range.

Images