Industrial economic operation data retrieval method and system based on semantic large model

Abstract

This invention provides a method and system for retrieving industrial economic operation data based on a semantic large-scale model. First, it acquires a set of event records describing economic activities in natural language text, with each record carrying source system and time information. Then, it uses a pre-trained semantic large-scale model to extract entity semantic units and quantify the semantic association strength between units. Next, it fuses all semantic units and association strengths across event records, constructing an industrial economic operation semantic association network with semantic units as nodes and association strength as edge weights. Upon receiving a query request in natural language form, it parses the query intent using the same semantic large-scale model to obtain the query semantic unit group, and performs semantic alignment and association path extraction within the semantic association network, backtracking and matching relevant economic operation event records as the retrieval results. This invention achieves deep semantic understanding and precise association retrieval of massive amounts of industrial economic operation data.

Description

Technical Field

[0001] This application relates to the field of text processing, and in particular to a method and system for retrieving industrial economic operation data based on a semantic big data model. Background Technology

[0002] With the widespread deployment of industrial information systems, industrial economic operation data retrieval technology has emerged. This refers to the technology of finding data records related to specific economic indicators or business events from unstructured event description texts generated in industrial production, distribution, sales, and other economic operation processes. Currently, common retrieval technologies typically rely on keyword full-text search engines. Users input keyword combinations, and the system performs literal matching scans on the record set, filtering texts containing the keywords and outputting them in order of frequency. However, this literal matching-based approach can only identify shallow string consistency and cannot capture synonymous expressions, contextual semantic dependencies, and implicit relationships across event records. This leads to a large number of missed semantically relevant but literally mismatched records or falsely detected literally matched but semantically irrelevant noisy records. The comprehensiveness and accuracy of the retrieval cannot meet the needs of deep semantic understanding. Summary of the Invention

[0003] This invention provides a method and system for retrieving industrial economic operation data based on a semantic large model.

[0004] In a first aspect, embodiments of the present invention provide a method for retrieving industrial economic operation data based on a semantic large model, the method comprising: The original data set of industrial economic operation records is obtained. The original data set of industrial economic operation records contains multiple event description texts in the form of natural language text describing the process of economic operation activities. Each event description text corresponds to an independent economic operation event record. The economic operation event record carries the event source system identifier and event occurrence time context description information. Using a pre-trained semantic big model, semantic units are extracted from the original record dataset of industrial economic operations. This yields a set of semantic units corresponding to each economic operation event record and a parameter for the strength of association between semantic units. The set of semantic units consists of semantic entity terms extracted from the event description text of the economic operation event record. The parameter for the strength of association between semantic units represents a measure of the closeness of semantic association between different semantic entity terms within the economic operation event record. The semantic association network of industrial economic operation is formed based on the set of semantic units and the correlation strength parameter between semantic units. The semantic association network of industrial economic operation includes a set of semantic unit nodes with all semantic entity terms extracted from all economic operation event records as semantic unit nodes, and a set of semantic association edges with the correlation strength parameter between semantic units as edge weights. The semantic unit nodes are connected by semantic association edges to form a network association structure. Receive target retrieval requests, which contain a text describing the query intent in natural language, targeting changes in target economic indicators or related business events. Based on the semantic association network of industrial economic operation, the query semantic unit group is obtained by parsing the query intent of the target retrieval request through the semantic big model. Then, semantic alignment and location processing and association path extraction are performed in the semantic association network of industrial economic operation according to the query semantic unit group, and a set of economic operation data retrieval results consisting of matching economic operation event records is generated.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory for storing computer-executable instructions or computer programs; and a processor for executing the computer-executable instructions or computer programs stored in the memory to implement the above-mentioned method for retrieving industrial economic operation data based on a semantic big model.

[0006] The embodiments of this application have the following beneficial effects: The industrial economic operation data retrieval method based on a semantic big data model provided in this invention obtains event description text describing economic operation activities in natural language, extracts the semantic unit set and the correlation strength parameter between semantic units for each economic operation event record using a semantic big data model, and automatically parses out semantic entity terms related to the economic field and their semantic correlation strength measure. Based on the semantic entity terms and correlation strength parameters extracted from all event records, an industrial economic operation semantic correlation network is constructed, which condenses semantic knowledge scattered from different sources and time points into a mesh correlation structure with semantic unit nodes as basic units and correlation strength parameters as edge weights, comprehensively depicting economic operation. The multidimensional relationships between concepts avoid the problem of lost semantic connections in traditional keyword indexing. It receives target retrieval requests with query intent described in natural language, obtains query semantic unit groups through semantic big data model parsing, performs semantic alignment and location and extracts association paths of query semantic unit groups in semantic association network, accurately locates semantically matching economic operation event records based on the true semantics of query intent, and extracts deeply related search result sets based on semantic association paths. The entire retrieval is driven by deep semantic understanding, which can capture the substantial semantic alignment between event records and query intent across literal expression differences, thereby significantly improving the comprehensiveness and accuracy of industrial economic operation data retrieval. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the architecture of the application scenario provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the computer system provided in the embodiments of this application; Figure 3This is a flowchart illustrating the industrial economic operation data retrieval method based on a semantic big model provided in this application embodiment. Detailed Implementation

[0008] See Figure 1 , Figure 1 This is a schematic diagram of the application scenario provided in this application embodiment. The industrial information system 400 is connected to the computer system 200 through a network 300. The network 300 can be a wide area network (WAN), a local area network (LAN), or a combination of both, using wireless or wired links to achieve data transmission. In some embodiments, the computer system 200 can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0009] The following describes the computer system that implements the industrial economic operation data retrieval method based on a semantic large model provided in the embodiments of this application. See also Figure 2 , Figure 2 This is a schematic diagram of the structure of the computer system provided in the embodiments of this application. Figure 2 The computer system shown includes at least one processor 210, memory 250, at least one network interface 220, and an external interface 230. The various components in the computer system 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 240.

[0010] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0011] External interface 230 may include, for example, one or more speakers and / or one or more visual displays. External interface 230 may also include one or more input devices 232, such as a keyboard, mouse, microphone, touch screen display, camera, etc.

[0012] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0013] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this application embodiment is intended to include any suitable type of memory.

[0014] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0015] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; The network communication module 252 is used to reach the industrial information system via one or more (wired or wireless) network interfaces 220, such as Bluetooth, WiFi, and Universal Serial Bus (USB). Presentation module 253 is configured to enable the display of information (e.g., external interface for operating peripheral devices and displaying content and information) via one or more output devices 231 (e.g., display screen, speaker, etc.) associated with external interface 230; The input processing module 254 is used to detect and translate one or more user inputs or interactions from one or more input devices 232.

[0016] Based on the above description of the application scenarios and computer systems provided in the embodiments of this application, the following describes the industrial economic operation data retrieval method based on a semantic large model provided in the embodiments of this application. In actual implementation, the industrial economic operation data retrieval method based on a semantic large model provided in the embodiments of this application can be implemented by a computer system. See also Figure 3 , Figure 3 This is a flowchart illustrating the industrial economic operation data retrieval method based on a semantic large model provided in this application embodiment. Next, we will combine... Figure 3 The steps shown are explained.

[0017] Step S100: Obtain the set of original records of industrial economic operation. The set of original records of industrial economic operation contains multiple event description texts in the form of natural language text describing the process of economic operation activities. Each event description text corresponds to an independent economic operation event record. The economic operation event record carries the event source system identifier and the event occurrence time context description information.

[0018] In this embodiment, the industrial economic operation raw record data set aggregates massive economic operation activity records reported from different industrial information systems. These industrial information systems each undertake the operational monitoring and management functions of a specific link in the enterprise's production and operation chain. During continuous operation, they constantly generate descriptive text materials in natural language text form, such as operation logs, status reports, anomaly notifications, and statistical summaries. These materials collectively constitute the raw material of the industrial economic operation raw record data set. Each event description text is essentially a written record of an economic operation activity process that has occurred under specific spatiotemporal conditions and has relatively complete semantic boundaries. Its content typically involves specific business facts such as real-time changes in economic indicators, switching of production equipment operating status, material flow at supply chain nodes, phased statistics of energy consumption, judgment of quality inspection results, increases and decreases in warehouse inventory, and confirmation of market transaction execution.

[0019] Each event description text corresponds to an independent economic operation event record. In addition to storing the event description text as the main data field, the economic operation event record also forcibly binds two auxiliary metadata fields: an event source system identifier and an event occurrence time context description. The event source system identifier is used to accurately trace the origin of the event description text. The specific encoding rules follow the system coding and naming conventions uniformly managed by the Enterprise Information Asset Metadata Registry. This convention adopts a hierarchical namespace structure, concatenating information from four levels—enterprise identifier, plant area identifier, system type identifier, and system instance number—with predetermined separators into a globally unique system identifier string. The event occurrence time context description is used to accurately mark the time position of the economic operation event. It is composed of the event start timestamp and the event end timestamp, both timestamps are stored in Coordinated Universal Time (UTC) standard format with a time precision of seconds. For events involving only instantaneous state changes without duration, both timestamps are filled with the same time value. Through the binding of the event source system identifier and the event occurrence time context description, subsequent steps in constructing semantic association networks and tracing the source of retrieval results can always maintain a complete data source chain and time dimension constraint reference.

[0020] Step S200: Use a pre-trained semantic large model to extract semantic units from the original data set of industrial economic operation records, and obtain the semantic unit set corresponding to each economic operation event record and the correlation strength parameter between semantic units. The semantic unit set consists of semantic entity terms extracted from the event description text of the economic operation event record. The correlation strength parameter between semantic units represents the measure of the degree of semantic association between different semantic entity terms within the economic operation event record.

[0021] The pre-trained semantic large-scale model is the core computational carrier for performing semantic unit extraction and semantic relation quantification. Its internal structure, based on functional division, can be divided into five collaborative functional components: an input embedding transformation component, a multi-layer context encoding network, a semantic entity boundary recognition processing component, an entity category labeling processing component, and a semantic relation inference channel. The input embedding transformation component transforms the event description text string from a discrete character sequence into a sequence of real-number vectors, forming a numerical semantic representation that can be directly consumed by the subsequent encoding network. The multi-layer context encoding network, based on the language prior knowledge obtained through large-scale pre-training, performs progressive contextual semantic fusion on the event description text, ensuring that each semantic symbol in the text receives a deep semantic representation constrained by the entire sentence context. The semantic entity boundary recognition processing component and the entity category labeling processing component respectively handle the boundary localization and type determination tasks of entity semantic units, working together to accurately extract the set of entity semantic units from the event description text. The semantic relation inference channel further performs joint inference on the semantic dependency direction and dependency tightness between any two entity semantic units, outputting quantified parameters of the association strength between semantic units. Before being used for industrial economic operation data retrieval tasks, this semantic big data model underwent self-supervised pre-training on a large-scale corpus of multi-source industrial economic texts, including industrial news, technical standards, economic operation reports, supply chain documents, and equipment operation and maintenance records. The pre-training tasks included self-supervised objectives such as masked language modeling and next-sentence prediction, ensuring that the model parameters internalized lexical knowledge, syntactic patterns, entity co-occurrence rules, and event logical association patterns within the industrial economic domain. The complete processing flow of the semantic big data model for each economic operation event record is detailed in steps S210 to S280.

[0022] Step S210: Input each economic operation event record in the original data set of industrial economic operation into the input embedding transformation component of the semantic big model in sequence, perform semantic symbolization mapping on the event description text of the economic operation event record, and generate an initial semantic symbol sequence.

[0023] When an economic event record is fed into the semantic big data model, the input embedding transformation component first preprocesses and vectorizes the event description text field. Internally, the input embedding transformation component maintains a sub-word-level lexicon built specifically for the industrial economic field. This lexicon covers frequently occurring vocabulary units in industrial economic texts, such as industrial production terms, economic indicator names, abbreviations of quantity units, organizational abbreviations, geographical location names, and product model specifications. It also retains a set of basic single-character entries to cover out-of-vocabulary words. After the event description text string is fed into the lexicon matcher, it searches the lexicon segment by segment from left to right according to the maximum matching principle, segmenting the event description text into a sequence of entries. Each entry is then mapped to a corresponding lexicon index integer, resulting in an integer-type semantic symbol index sequence.

[0024] Next, the input embedding transformation component calls a trainable word embedding matrix, whose row capacity equals the vocabulary size and column dimension equals the hidden layer representation dimension of the semantic large model. Using each integer index in the semantic symbol index sequence as a row number, the corresponding row vector is read row by row from the word embedding matrix, outputting an initial vector sequence of the same length as the semantic symbol sequence. Each semantic symbol thus obtains a trainable, dense real-valued vector representation that has been optimized and tuned using an industrial-scale corpus during pre-training.

[0025] To inject positional information of semantic symbols into the initial vector sequence, the input embedding transformation component then appends a positional encoding vector corresponding to the position index to the initial vector at each position. The positional encoding vectors are generated by alternately evaluating sine and cosine functions at each vector dimension index, with the input variables being the product of the position index and the frequency parameter. The characteristic of this encoding method is that relative positional relationships can be derived through linear transformations between positional encoding vectors, enabling the model to capture the order and distance relationships between semantic symbols in the sequence without relying on explicit positional indices.

[0026] Furthermore, the input embedding transformation component uniformly appends a trainable segment label vector to the vectors at all positions in the sequence. This segment label vector is learned during the pre-training phase to be a type label identifying whether the input text belongs to the semantic segment of the event description text. After the triple transformation of word embedding mapping, position encoding superposition, and segment label superposition, the input embedding transformation component generates an initial semantic symbol sequence. The vector at each position in the sequence carries simultaneously the lexical semantic information of the corresponding semantic symbol, the position information of the symbol in the event description text, and the text type attribution information.

[0027] Step S220: Perform multi-layer self-attention context encoding on the initial semantic symbol sequence. In each encoding layer, fuse the contextual association information between the semantic symbols within the initial semantic symbol sequence to obtain a deep semantic understanding encoding sequence. Each encoding position in the deep semantic understanding encoding sequence stores the semantic representation data of the corresponding semantic symbol in the context of the complete event description text.

[0028] The initial sequence of semantic symbols is passed to a multi-layer context encoding network of a large semantic model. This network consists of multiple identical encoding layers stacked end-to-end, with the output sequence of each layer directly serving as the input sequence for the next. The number of encoding layers determines the depth of semantic abstraction of the text; the more layers, the more refined the long-range semantic dependencies the model can capture. Each encoding layer integrates two core computational modules: a multi-head self-attention mechanism and a feedforward fully connected transformation mechanism. Residual connections and layer normalization operations are applied between the input and output of each module to stabilize gradient propagation and training convergence of the deep network.

[0029] In the multi-head self-attention module, each position vector of the input sequence is first transformed by three linear transformations through the query projection matrix, key projection matrix, and value projection matrix, respectively, to obtain three role vectors: query representation, key representation, and value representation. The multi-head self-attention mechanism divides the dimensions of the three role vectors into multiple subspaces along the feature direction, with each subspace corresponding to an attention head. For each attention head, the query representation and key representation within that head are inner-producted, and the result forms a square matrix. The element in the i-th row and j-th column of this matrix represents the contribution strength of the semantic information of position j to understanding position i from the perspective of the current position i. This square matrix is ​​then processed by a normalized exponential function within each row to transform it into a probability distribution with a sum of 1. This probability distribution is then used as a weighting coefficient to weight and aggregate the value representations, resulting in the output representation of the attention head that incorporates global context information. The output representations of all attention heads are reconstructed along the feature dimension and mapped back to the hidden layer dimension of the encoding layer through an output projection matrix, forming the output of the multi-head self-attention module.

[0030] The output is element-wise added to its input to form a residual connection result. Then, layer normalization is performed independently on each position vector of the residual connection result. The layer normalization operation first calculates the mean and standard deviation of the vector along the feature dimension, then subtracts the mean from each component of the vector and divides it by the standard deviation for standardization scaling, and finally performs an affine transformation on the standardized result through a trainable scaling and translation parameter to obtain the intermediate representation sequence after self-attention fusion and normalization.

[0031] The intermediate sequence then enters the feedforward fully connected module, which consists of two cascaded linear projection transformations. The first linear projection transformation expands the hidden layer dimension to a pre-defined intermediate multiple and applies a non-linear activation function to each element to perform a non-linear feature transformation. The non-linear activation function used is the Gaussian error linear unit function, which maintains linear growth in the positive interval and performs soft suppression in the negative interval, effectively alleviating the gradient vanishing problem in deep networks. The second linear projection transformation then compresses the dimension back to the original hidden layer dimension, restoring the same shape as the input. The output of the feedforward fully connected module also undergoes residual connection and layer normalization processing to obtain the final output sequence of the current encoding layer.

[0032] The aforementioned multi-head self-attention and feedforward fully connected computation process is repeated in each encoding layer. Each time the initial semantic symbol sequence passes through an encoding layer, each position vector in the sequence integrates contextual information from semantic symbols at more distant locations within its range, and gradually performs feature transfer in the vector space towards a more abstract semantic level. When the sequence penetrates the final encoding layer, each encoded position vector in the resulting deep semantic understanding encoding sequence has fully absorbed the contextual information across the entire event description text, forming the most accurate semantic representation of the semantic symbol at the current position under the constraints of the complete event description text context.

[0033] Step S230: The semantic entity boundary recognition processing component of the semantic big model performs boundary scanning on the deep semantic understanding encoding sequence to determine the start position marker sequence and end position marker sequence of the entity semantic unit in the event description text. Based on the start position marker sequence and end position marker sequence, the corresponding text fragments are extracted from the event description text to form a set of entity semantic unit fragments.

[0034] The semantic entity boundary recognition processing component is directly deployed at each encoding position of the deep semantic understanding encoding sequence. This component contains two parallel linear classifiers: a start boundary classifier and a termination boundary classifier. The start boundary classifier consists of a fully connected linear layer with parameters of a weight matrix and a bias vector. Its function is to reduce the dimensionality of the hidden layer dimension vector at each position in the deep semantic understanding encoding sequence to a real scalar. This real scalar is then fed into a logistic function for numerical compression, outputting a start boundary probability value between 0 and 1. The start boundary probability value essentially measures the confidence that the encoding position is the starting boundary of a certain entity semantic unit's text span. When this probability value crosses a pre-set start boundary determination threshold, the position is marked as a start position. The termination boundary classifier has the exact same structural design as the start boundary classifier, but its fully connected linear layer uses independent weight matrix and bias vector parameters. It also outputs a termination boundary probability value for each encoding position vector. When the probability value exceeds a pre-set termination boundary determination threshold, the position is marked as a termination position. The decision thresholds used by the two classifiers are calibrated during the model training phase using boundary recognition performance metrics on the validation set.

[0035] After obtaining the start and end position marker sequences, the semantic entity boundary recognition processing component performs a boundary pairing step. Boundary pairing follows a nearest-distance greedy matching principle: it traverses the start and end position markers from left to right in the event description text, pairing any unpaired start position marker with its nearest unpaired end position marker to its right. Each successfully paired start and end position marker defines a character range for an entity semantic unit in the event description text; this range corresponds to a continuous substring. The substring extracted from the original event description text based on this character range is an entity semantic unit fragment, which fully preserves the literal representation of the entity semantic unit in the original text. All successfully paired entity semantic unit fragments are merged into an entity semantic unit fragment set; isolated start or end position markers that fail to pair are considered false boundary detections and are discarded.

[0036] Step S240: Call the entity category labeling processing component of the semantic big model to assign a category to each entity semantic unit fragment in the entity semantic unit fragment set, assign an industrial economic operation domain entity type label to each entity semantic unit fragment, and take the entity semantic unit fragment carrying the industrial economic operation domain entity type label as the entity semantic unit to obtain the entity semantic unit set of the economic operation event record.

[0037] For each entity semantic unit fragment already located in the entity semantic unit fragment set, the entity category labeling processing component needs to further determine its semantic type classification. First, the component indexes all encoded position vectors corresponding to the start and end positions of the entity semantic unit fragment from the deep semantic understanding encoding sequence. These encoded position vectors constitute the feature matrix of the fragment in the context semantic space. This feature matrix is ​​then averaged along the sequence length, with the element values ​​in each feature dimension calculated as an arithmetic mean, resulting in an aggregated vector with a fixed dimension consistent with the hidden layer dimension of the large semantic model. This averaging operation can equally weight and compress the semantic features of each semantic symbol within the entity semantic unit fragment, resulting in a compact fragment-level overall semantic representation.

[0038] The entity category labeling component feeds the aggregated vector into a fully connected linear classifier for category determination. The input dimension of this classifier is the same as the dimension of the aggregated vector, and the output dimension equals the preset total number of entity types in the industrial economic operation domain. The weight matrix within the classifier linearly maps the aggregated vector to an entity category score vector. The value of each dimension in the entity category score vector reflects the original discrimination score of the entity semantic unit fragment belonging to the corresponding entity type. The score vector is then passed to a normalized exponential function layer. This layer performs an exponential transformation on the score vector and normalizes it item by item according to the sum of the exponents of all categories, outputting a probability distribution vector where all components are positive and the sum is 1. The entity type label string corresponding to the dimension index with the highest probability value in the distribution vector is the category labeling result of that entity semantic unit fragment.

[0039] The entity type labeling system used in this embodiment is hierarchically organized according to the business entity ontology of the industrial economic field. It comprehensively covers the basic entity categories involved in industrial economic operation activities, such as subject entities, object entities, attribute entities, activity entities, spatiotemporal entities, and measurement entities. Each basic category is further subdivided into multiple subcategories. The entity category labeling processing component merges and encapsulates the entity semantic unit fragment, along with its determined industrial economic operation field entity type label and its span position information in the event description text, into a complete entity semantic unit. This entity semantic unit is then entered into the entity semantic unit set corresponding to the economic operation event record.

[0040] Step S250: According to the local text window where entity semantic units appear together in the event description text, pair and combine the entity semantic units in the entity semantic unit set to generate an entity semantic unit pairing sample set. Each entity semantic unit pairing sample in the entity semantic unit pairing sample set contains a first entity semantic unit and a second entity semantic unit.

[0041] The pairing and combination process of entity semantic units is based on the linear co-occurrence relationship of entity semantic units in the event description text. Each entity semantic unit in the entity semantic unit set is used as an anchor entity semantic unit. The center position of the span of this anchor entity semantic unit in the event description text is taken as the origin of the window, extending forward and backward by a predetermined window coverage span. The unit of measurement for this window coverage span is the number of entity semantic units. For each other entity semantic unit within the window coverage area, excluding the anchor entity semantic unit, a pairing sample is formed with the anchor entity semantic unit. In this pairing sample, the anchor entity semantic unit is identified as the first entity semantic unit, and the other entity semantic units selected within the window are identified as the second entity semantic unit. This process is repeated for all entity semantic units in the entity semantic unit set. The union of all generated entity semantic unit pairing samples constitutes the entity semantic unit pairing sample set of the economic operation event record.

[0042] Step S260: Input each entity semantic unit pairing sample in the entity semantic unit pairing sample set into the semantic relationship reasoning channel of the semantic big model, perform joint reasoning on the semantic dependency direction and dependency tightness between the first entity semantic unit and the second entity semantic unit in the entity semantic unit pairing sample, and output the semantic association direction label and the corresponding semantic unit association strength parameter component corresponding to the entity semantic unit pairing sample.

[0043] The semantic relationship inference channel is a reasoning network specifically designed within the semantic large model to determine the direction and quantify the closeness of semantic relationships between entity semantic units. This inference channel consists of four functional layers connected sequentially in the data flow order: a sequence concatenation processing layer, a direction-sensitive positional encoding injection layer, a bidirectional semantic influence analysis structure, and a global association strength measurement structure. Each functional layer undertakes a relatively independent computational step in the inference process. The semantic relationship inference channel completes the full inference of the paired samples of the current entity semantic unit through steps S261 to S268.

[0044] Step S261: From the deep semantic understanding encoding sequence, extract the encoding fragment of the first entity semantic unit according to the start position marker sequence and the end position marker sequence corresponding to the first entity semantic unit in the entity semantic unit pairing sample, and extract the encoding fragment of the second entity semantic unit according to the start position marker sequence and the end position marker sequence corresponding to the second entity semantic unit.

[0045] In the two-dimensional matrix of the deep semantic understanding encoding sequence, the start and end position markers of each entity semantic unit, as annotated in step S230, are essentially two delimiters on the row index interval of this matrix. For the first entity semantic unit in the entity semantic unit pairing sample, the semantic relation inference channel performs a row slicing operation according to its start and end position marker indices, extracting the complete encoding vector of each row within the row index interval in its original order and assembling it into a sub-matrix. This sub-matrix is ​​the encoding fragment of the first entity semantic unit. The number of rows in the matrix is ​​equal to the number of semantic symbols contained in the first entity semantic unit, and the number of columns is the dimension of the hidden layer of the semantic large model. For the second entity semantic unit, the semantic relation inference channel uses the same row slicing operation, extracting the second entity semantic unit encoding fragment with the same column dimension and the same number of rows as the number of semantic symbols contained in it, according to its own start and end position marker indices. The two encoding fragments maintain the semantic constraints of the complete context of the event description text within their respective local intervals, without losing contextual information due to fragmented extraction.

[0046] Step S262: Input the first entity semantic unit encoding fragment and the second entity semantic unit encoding fragment into the sequence splicing processing layer of the semantic relationship reasoning channel, and perform front-to-back splicing operation along the sequence length dimension to generate a paired joint encoding sequence. The paired joint encoding sequence maintains the sequential splicing order of the first entity semantic unit encoding fragment and the second entity semantic unit encoding fragment in the sequence structure.

[0047] The sequence concatenation layer contains no trainable parameters and only performs tensor shape-level recombination operations. This layer receives two two-dimensional matrices with different numbers of rows: a first entity semantic unit encoding fragment and a second entity semantic unit encoding fragment. It performs a vertical concatenation operation directly along the row dimension, placing all row vectors of the first entity semantic unit encoding fragment on top and all row vectors of the second entity semantic unit encoding fragment immediately following it on the bottom. This generates a paired joint encoding sequence with the number of rows equal to the sum of the number of rows in the two input sub-matrices, while maintaining the column number while preserving the hidden layer dimension of the semantic model. The top rows of the concatenated sequence completely correspond to the first entity semantic units, and the bottom rows completely correspond to the second entity semantic units—a perturbation-free front-to-back partitioning arrangement.

[0048] Step S263: Perform direction-sensitive position coding injection on the paired joint coding sequence, add a first entity discrimination identifier to each position element in the paired joint coding sequence that comes from the first entity semantic unit coding segment, and add a second entity discrimination identifier to each position element that comes from the second entity semantic unit coding segment, to obtain a direction-discriminative joint coding sequence.

[0049] The direction-sensitive positional encoding injection layer maintains two trainable parameter vectors with the same dimension as the hidden layers of the semantic large model, denoted as the first entity discrimination vector and the second entity discrimination vector, respectively. These two vectors are optimized during model training to produce discriminative feature offsets in the representation space for vectors from different entities. The direction-sensitive positional encoding injection layer iterates through the paired joint encoding sequence row by row according to the row index: for all row vectors whose row index falls within the interval corresponding to the first entity semantic unit encoding segment, the first entity discrimination vector is element-wise superimposed onto each component of that row vector; for all row vectors whose row index falls within the interval corresponding to the second entity semantic unit encoding segment, the second entity discrimination vector is element-wise superimposed onto each component of that row vector. After the superposition operation, the original paired joint encoding sequence is renamed the direction-discriminative joint encoding sequence. In this sequence, each row vector in the first half is appended with a feature imprint from the first entity source, and each row vector in the second half is appended with a feature imprint from the second entity source. This allows the bidirectional semantic influence analysis structure to distinguish between the two types of input sources and calculate the semantic influence intensity in different directions.

[0050] Step S264: Perform positive semantic influence reasoning on the directional discriminative joint coding sequence through the bidirectional semantic influence analysis structure of the semantic relationship reasoning channel to obtain the positive semantic influence strength value of the first entity semantic unit on the second entity semantic unit, and perform reverse semantic influence reasoning on the directional discriminative joint coding sequence to obtain the reverse semantic influence strength value of the second entity semantic unit on the first entity semantic unit.

[0051] The bidirectional semantic influence analysis structure consists of two branches: a forward semantic influence inference branch and a reverse semantic influence inference branch. These two branches share a completely identical internal computational architecture but each maintains its own independent set of network parameters. The input to both branches is the directional discriminative joint coding sequence output from step S263.

[0052] Within the positive semantic influence inference branch, a self-attention pooling operation is first performed. The execution process of self-attention pooling is as follows: A trainable pooling query vector is maintained, with the same dimension as each row vector in the direction-discriminative joint encoding sequence. This pooling query vector is used as the query, and all row vectors in the direction-discriminative joint encoding sequence are used as both keys and values. The inner product similarity scalar between the pooling query vector and each row vector in the direction-discriminative joint encoding sequence is calculated. This set of similarity scalars is compressed by a normalized exponential function layer and transformed into a set of attention weight coefficients that sum to 1. These attention weight coefficients are then used to perform a weighted summation of all row vectors in the direction-discriminative joint encoding sequence, resulting in a pooled representation vector that integrates information from all positional vectors in the sequence. The dimension of this pooled representation vector remains the same as the hidden layer dimension. The pooled representation vector is then input into a scoring network, which consists of two cascaded fully connected layers. The first fully connected layer maps the pooled representation vector to an intermediate dimension, applying non-linear activation to each feature component of the intermediate dimension for feature transformation. The second fully connected layer then maps the intermediate dimension to a scalar output of dimension 1, which is the positive semantic influence strength value. The positive semantic influence strength value semantically measures the degree of semantic influence exerted by the first entity semantic unit as the agent of semantic influence on the second entity semantic unit as the receiver of semantic influence in the paired sample of entity semantic units.

[0053] The internal computation process of the reverse semantic influence inference branch is completely equivalent to that of the forward branch: A separate, trainable pooling query vector is used to perform self-attention pooling on all row vectors of the directional discriminative joint encoding sequence, resulting in a reverse pooling representation vector. This reverse pooling representation vector is then processed by an independent scoring network to output a scalar of dimension 1, which represents the reverse semantic influence strength value. Its semantic meaning is the degree to which the second entity semantic unit exerts a semantic influence on the first entity semantic unit. The two branches operate independently during computation, extracting semantic influence measures in opposite directions from the same input sequence.

[0054] Step S265: Based on the positive semantic influence strength value and the negative semantic influence strength value, the direction is determined by comparison. When the positive semantic influence strength value is greater than the negative semantic influence strength value, the semantic association direction is marked as the first entity semantic unit pointing to the second entity semantic unit. Otherwise, the semantic association direction is marked as the second entity semantic unit pointing to the first entity semantic unit.

[0055] The direction comparison determination is performed by a numerical comparator. The numerical comparator takes the positive semantic influence strength value and the negative semantic influence strength value as two scalar inputs and performs a greater than comparison operation. The result of the comparison operation directly determines the generated content of the semantic association direction tag: if the positive semantic influence strength value is greater than the negative semantic influence strength value, it means that in the pairing sample of entity semantic units, the semantic dominance of the first entity semantic unit over the second entity semantic unit is stronger than that of the negative one. Therefore, the natural direction of the semantic dependency relationship is from the first entity semantic unit to the second entity semantic unit. At this time, the semantic association direction tag is assigned the direction identifier from the first entity semantic unit to the second entity semantic unit. If the positive semantic influence strength value is less than or equal to the negative semantic influence strength value, it means that the natural direction of the semantic dependency relationship is the opposite direction or the negative one is dominant. At this time, the semantic association direction tag is assigned the direction identifier from the second entity semantic unit to the first entity semantic unit.

[0056] Step S266: Input the directional discriminative joint coding sequence into the global association strength measurement structure of the semantic relationship reasoning channel, and perform multi-head semantic interaction projection on the directional discriminative joint coding sequence through multiple sets of learnable semantic interaction transformation matrices to generate association strength feature signals under multiple semantic interaction subspaces.

[0057] The global association strength measurement structure and the direction determination branch are executed in parallel. Their task is to measure the overall tightness of the semantic association between two entity semantic units in the paired sample of the entity semantic unit, rather than the directionality. This structure adopts a multi-head projection strategy, allowing the position vectors in the direction-discriminative joint encoding sequence to perform semantic interaction modeling in multiple different projection subspaces, so as to capture different aspects of association strength from multiple perspectives. The specific execution process is gradually unfolded through steps S2661 to S2667.

[0058] Step S2661: Copy the orientation-discriminative joint coding sequence into multiple copies of the parallel input sequence, the number of copies being the same as the number of semantic interaction projection heads preset in the global association strength measurement structure.

[0059] When the directional discriminative joint encoded sequence enters the global association strength metric structure, it is first broadcast and copied into several identical replica sequences. The number of replica sequences is equal to the number of predefined semantic interaction projectors in the global association strength metric structure. Each semantic interaction projector is associated with a completely independent set of projection parameters. In subsequent steps, each projector performs independent and non-interfering projection and interaction calculations on its assigned replica sequence, allowing different projectors to observe the association patterns in the paired samples of the entity semantic unit from different projection perspectives.

[0060] Step S2662: Apply a set of independent first linear projection transformation parameters and second linear projection transformation parameters to each copy sequence, and project the copy sequence to the query semantic space and the key semantic space simultaneously to obtain the query projection vector sequence and the key projection vector sequence corresponding to the semantic interaction projection head.

[0061] Within a semantic interactive projector, the first linear projection transformation parameter is a trainable weight matrix. The number of columns in this matrix corresponds to the subspace dimension assigned to the semantic interactive projector, and the number of rows corresponds to the hidden layer dimension of the semantic large model. This weight matrix performs matrix multiplication on each row vector in the replica sequence, linearly mapping each row vector from the original hidden layer dimension to a lower-dimensional query semantic space. The mapping result is a sequence of query projection vectors of equal length but with reduced dimensionality. The second linear projection transformation parameter is another trainable weight matrix with the same shape as the first linear projection transformation parameter, but its parameter values ​​are trained independently. This matrix maps each row vector in the replica sequence to the key semantic space, resulting in a sequence of key projection vectors. The dimensions of the query semantic space and the key semantic space are set to integer values ​​obtained by dividing the hidden layer dimension of the semantic large model by the number of semantic interactive projectors.

[0062] Step S2663: Apply a set of independent third linear projection transformation parameters to each copy sequence to project the copy sequence onto the value semantic space, thereby obtaining the value projection vector sequence corresponding to the semantic interaction projection head.

[0063] Within the same semantic interactive projection head, the third linear projection transformation parameter is also a trainable weight matrix. This matrix maps each row vector of the replica sequence to the value semantic space, which has the same dimension as the query semantic space and the key semantic space. The resulting sequence of value projection vectors is the value representation used for subsequent weighted aggregation under this semantic interactive projection head.

[0064] Step S2664: Within each semantic interaction projection head, a semantic interaction weight matrix is ​​generated by querying the projection vector sequence and the key projection vector sequence. The semantic interaction weight distribution data of the semantic interaction projection head is generated by the transpose interaction operation of the query projection vector sequence and the key projection vector sequence.

[0065] Within each semantic interaction projector, the query projection vector sequence is organized into a matrix with the number of rows equal to the length of the replica sequence and the number of columns equal to the subspace dimension. The key projection vector sequence is organized into a matrix of the same shape. Multiplying the query projection vector sequence matrix by the transpose of the key projection vector sequence matrix yields a square matrix with both the number of rows and columns equal to the length of the replica sequence. This square matrix represents the original semantic interaction weight matrix of the semantic interaction projector on the current replica sequence. Each element in the square matrix records the original attention interaction score of one row position vector to another row position vector in the replica sequence from the perspective of the subspace represented by the projector. A normalized exponential function operation is independently performed on each row of this square matrix, converting the scores of each element in a row into probability values ​​that sum to 1, thus obtaining the semantic interaction weight distribution data of the semantic interaction projector.

[0066] Step S2665: Apply the semantic interaction weight distribution data to the value projection vector sequence, perform weighted aggregation, and generate a context-aware representation sequence that integrates cross-positional semantic interaction information under the semantic interaction projection head.

[0067] Within each semantic interaction projection head, the semantic interaction weight distribution data is applied to the value projection vector sequence via left multiplication. The output of the matrix multiplication is a new matrix with the number of rows equal to the length of the copy sequence and the number of columns equal to the subspace dimension. This new matrix is ​​the context-aware representation sequence. Each row vector in this sequence is a weighted linear combination of all row vectors in the value projection vector sequence, with the weighting coefficients given by the corresponding row of the semantic interaction weight distribution data. This weighted aggregation essentially achieves complete information exchange between the position vectors within the direction-discriminative joint coding sequence in the projection subspace. The final representation obtained for each position not only contains its own projection information in the subspace but also incorporates the projection information from all other positions in the sequence according to attention weights.

[0068] Step S2666: Perform global average aggregation on the context-aware representation sequence, calculate the mean of the representation elements along the sequence length dimension, and obtain the association strength feature signal corresponding to the semantic interaction projection head.

[0069] Within each semantic interaction projection head, an arithmetic mean compression is performed on the context-aware representation sequence along the sequence length dimension. This involves averaging the matrix element-wise along the row direction, resulting in a vector with a dimension equal to the subspace dimension. This vector represents the association strength feature signal corresponding to that semantic interaction projection head. The global averaging and aggregation operation uniformly compresses information from all positions across the entire sequence into a fixed-length feature vector. This vector concisely expresses the overall semantic association state across positions between the first and second entity semantic units within the semantic interaction subspace.

[0070] Step S2667: Concatenate the association strength feature signals corresponding to all semantic interaction projection heads along the head dimension to generate a multi-dimensional association strength concatenated signal. Perform linear mapping compression on the multi-dimensional association strength concatenated signal to compress the dimension of the multi-dimensional association strength concatenated signal to a single dimension, and generate the association strength parameter components between semantic units of entity semantic unit pairing samples.

[0071] Each semantic interaction projection head independently outputs a subspace-dimensional association strength feature signal vector. These vectors are concatenated end-to-end according to the projection head's arrangement. The total dimension of the concatenated signal equals the sum of the subspace dimensions of all projection heads, i.e., the hidden layer dimension of the semantic large model, thus obtaining a multi-dimensional association strength concatenated signal. This concatenated signal integrates the association strength feature expressions of the paired samples of the entity semantic unit, which are modeled separately in multiple different semantic interaction subspaces. The multi-dimensional association strength concatenated signal is then fed into a linear mapping compression layer. The parameter of this layer is a fully connected linear transformation matrix with 1 row and the number of columns equal to the dimension of the concatenated signal. This matrix compresses and maps the concatenated signal into a one-dimensional real scalar value, which is the association strength parameter component between the semantic units corresponding to the paired samples of the entity semantic unit. The magnitude of this parameter component quantifies the overall semantic tightness of the two entity semantic units in the semantic context of the economic operation event record; a larger value indicates a tighter semantic connection.

[0072] Step S267: Fusion and compression of association strength feature signals under multiple semantic interaction subspaces to generate semantic unit association strength parameter components of entity semantic unit pairing samples.

[0073] In step S266, each semantic interaction projection head independently outputs an association strength feature signal vector, the dimension of which is equal to the dimension of the subspace allocated to each projection head. Different projection heads capture different semantic interaction patterns within their respective subspaces; some focus on local dependencies, while others focus on global co-occurrence patterns. Therefore, the association strength information carried by each vector is complementary. In step S267, the association strength feature signal vectors output by all semantic interaction projection heads are first concatenated along the head dimension to obtain a multi-dimensional association strength concatenated signal with a total dimension equal to the dimension of the hidden layer of the semantic large model. This concatenated signal gathers multi-angle feature expressions of association strength from all projection subspaces. Subsequently, the multi-dimensional association strength concatenated signal is input into a fully connected linear transformation layer. The parameter matrix of this fully connected linear transformation layer compresses and maps the concatenated signal from a high dimension to a single-dimensional real scalar. This scalar represents the association strength parameter component between semantic units of the paired samples of the entity semantic unit. The parameters of the fully connected linear transformation layer are learned through training. Their role is to adaptively assign fusion weights to the feature components of different projection heads during compression, preserving information beneficial to the determination of association strength while suppressing redundancy or noise. The larger the value of the association strength parameter component in the final output semantic unit, the stronger the semantic association between the two entity semantic units within the context of that economic event record.

[0074] Step S268: Pair and bind the semantic association direction marker and the semantic unit association strength parameter component to obtain the semantic relationship reasoning result of the paired sample of the entity semantic unit.

[0075] The final data encapsulation layer of the semantic relation reasoning channel combines and encapsulates the semantic association direction marker output in step S265 and the semantic unit association strength parameter components output in step S267 according to a mutually bound data structure. In this embodiment, the binding method adopts a key-value pair structure, where the key field stores the semantic association direction marker and the value field stores the numerical value of the semantic unit association strength parameter components. The combined key-value pair is output to the outside as the complete semantic relation reasoning result of the paired sample of the entity semantic unit.

[0076] Step S270: Collect the semantic association direction labels and semantic association strength parameter components corresponding to all entity semantic unit pairing samples in the entity semantic unit pairing sample set, and construct the semantic unit association strength parameter set with entity semantic unit pairing samples as index units.

[0077] Each entity semantic unit pairing sample in the entity semantic unit pairing sample set is registered one by one in a dictionary structure with the entity semantic unit pairing sample identifier as the index key. This dictionary structure is the set of association strength parameters between semantic units. Each entry in it fully records the directional semantic association relationship between a pair of entity semantic units in the economic operation event record and its association strength quantification value.

[0078] Step S280: Combine and bind the entity semantic unit set and the set of semantic unit association strength parameters to serve as the semantic unit set and semantic unit association strength parameters corresponding to the economic operation event record.

[0079] For the current economic operation event record, the set of entity semantic units generated in step S240 and the set of semantic unit association strength parameters generated in step S270 are packaged together into an output data packet. This output data packet is the final processing result of the economic operation event record in step S200 and is used in the next stage for the construction of the industrial economic operation semantic association network.

[0080] Step S300: Form an industrial economic operation semantic association network based on the semantic unit set and the semantic unit association strength parameter. The industrial economic operation semantic association network includes a semantic unit node set with all semantic entity terms extracted from all economic operation event records as semantic unit nodes, and a semantic association edge set with the semantic unit association strength parameter as edge weight. The semantic unit nodes are connected by semantic association edges to form a network association structure.

[0081] The task of constructing the semantic association network for industrial economic operation is to perform global disambiguation and fusion across records for all independently generated entity semantic units and the correlation strength parameters between semantic units, forming a unified graph structure representing the entity concepts and their semantic relationships within the entire semantic space of industrial economic operation. The necessity of cross-record fusion lies in the fact that entity semantic units describing the same entity or concept in different event description texts may be completely identical or highly similar in literal terms. However, in step S200, they are independently extracted and assigned independent local identifiers. If they are not merged at the global level, it will lead to a large number of duplicate synonym nodes and redundant association edges in the semantic association network, weakening the network's semantic consistency and retrieval utility.

[0082] Specifically, step S300 completes global fusion and network construction through the systematic process of steps S310 to S370.

[0083] Step S310: Initialize an empty semantic unit node registration pool and an empty semantic association edge registration pool; traverse the semantic unit set corresponding to each economic operation event record, take each entity semantic unit in the semantic unit set as a candidate semantic unit node, and extract the entity semantic unit text surface form and industrial economic operation domain entity type mark of the candidate semantic unit node.

[0084] The semantic unit node registration pool is the main storage container for all registered semantic unit nodes globally, and it is created as an empty data structure before traversal begins. The semantic association edge registration pool is the main storage container for all registered semantic association edges globally, and it is also created as an empty data structure before traversal begins. Subsequently, the semantic unit set corresponding to each economic operation event record is read one by one, and a candidate registration process is performed on each entity semantic unit in the set. The first step of candidate registration is to extract the entity semantic unit text surface form and the industrial economic operation domain entity type mark of the entity semantic unit. The entity semantic unit text surface form is the original string extracted from the original event description text of the entity semantic unit, and the industrial economic operation domain entity type mark is the entity type mark string determined in step S240. These two attributes together constitute the key comparison basis for determining whether the candidate entity semantic unit refers to the same entity as the globally registered nodes.

[0085] Step S320: For each candidate semantic unit node, search the semantic unit node registration pool to see if there is a registered semantic unit node that is identical to the candidate semantic unit node in both the entity semantic unit text surface form and the entity type marker in the industrial economic operation field. If not, register the candidate semantic unit node as a new semantic unit node and assign it a unique semantic unit node identifier. If it exists, obtain the unique semantic unit node identifier of the registered semantic unit node and establish a coreference mapping relationship between the candidate semantic unit node and the registered semantic unit node.

[0086] Step S320 is the core step in global coreference resolution. For each candidate semantic unit node, its identity is compared with all registered semantic unit nodes already existing in the semantic unit node registration pool. The matching criterion for comparison is a two-field exact match, that is, the standardized result of the entity semantic unit text surface form and the entity type label in the industrial economic operation field must be completely consistent for it to be determined that they refer to the same entity. This matching strategy avoids the possible erroneous merging of entity types by only judging the similarity of text surface form. The specific execution of step S320 is further refined by steps S321 to S327.

[0087] Step S321: Construct the retrieval index structure of the semantic unit node registration pool. The retrieval index structure uses standardized strings in the text surface form of entity semantic units as the main retrieval key and entity type tags in the industrial economic operation field as auxiliary retrieval keys.

[0088] The retrieval index structure is a two-level nested index organization. The first-level index uses the standardized entity semantic unit text surface form string as the key, aggregating all registered semantic unit nodes in the semantic unit node registration pool that share the same standardized text surface form into the same index bucket. Within each first-level index bucket, the entity type tag string is used as the second-level index key for further subdivision, establishing a fast mapping from the entity type tag to the registered semantic unit nodes under that type. This two-level index structure allows for rapid narrowing of the search scope first through the text surface form and then precise positioning through the type tag when performing two-field exact match retrieval, without needing to traverse all registered nodes.

[0089] Step S322: Receive the candidate semantic unit nodes to be retrieved, obtain the entity semantic unit text surface form of the candidate semantic unit nodes, perform case-consistent conversion and whitespace character standardization on the entity semantic unit text surface form to obtain the standardized entity semantic unit text surface form.

[0090] Case unification converts all uppercase letters in the text surface form of entity semantic units to their corresponding lowercase letters, eliminating literal matching failures caused by inconsistencies in letter case due to different input systems. Whitespace normalization scans the text surface form string, removing all whitespace characters at the beginning and end of the string, and merging and replacing multiple consecutive whitespace characters in the middle of the string with a standard space character. The string after these two steps is the standardized entity semantic unit text surface form, which retains the substantive semantic content of the text surface form and only eliminates irrelevant differences at the format presentation level.

[0091] Step S323: Obtain the industrial economic operation domain entity type tag of the candidate semantic unit node, and form a retrieval key tuple by combining the standardized entity semantic unit text surface form and the industrial economic operation domain entity type tag.

[0092] The first element in the retrieval key tuple is the text surface form of the standardized entity semantic unit, and the second element is the entity type label in the field of industrial economic operation. This tuple is the complete query condition for performing an exact match retrieval in the retrieval index structure constructed in step S321.

[0093] Step S324: Perform an exact match search in the search index structure using the search key tuple to determine whether there exists a registered semantic unit node that simultaneously possesses a standardized entity semantic unit text surface form and an entity type marker in the industrial economic operation field that are completely consistent with the search key tuple.

[0094] Using the first element of the search key tuple as the lookup key, the system enters the first-level index and locates the corresponding text surface form index bucket. Within this bucket, using the second element of the search key tuple as the lookup key, it enters the second-level index and attempts to retrieve the registered semantic unit node record corresponding to this type of tag. If at least one record is successfully found in the second-level index, it is determined that a matching registered semantic unit node exists in the semantic unit node registration pool. If the first-level index does not find a match or the second-level index returns null, it is determined that no matching registered semantic unit node exists.

[0095] Step S325: When the exact match retrieval returns an empty set, it is determined that there is no registered semantic unit node in the semantic unit node registration pool that matches the candidate semantic unit node. A new unique identifier for the candidate semantic unit node is generated. The candidate semantic unit node is associated with the new unique identifier and stored in the semantic unit node registration pool. At the same time, the retrieval index structure is updated to include the newly registered node.

[0096] The new semantic unit node's unique identifier is generated using an auto-incrementing integer allocation strategy: the semantic unit node registration pool maintains a globally monotonically increasing integer counter. Each time a new identifier needs to be allocated, the current value of the counter is taken as the new identifier, and then the counter value is incremented by 1. The newly registered semantic unit node's storage record in the semantic unit node registration pool includes the semantic unit node's unique identifier, the standardized entity semantic unit text surface form, the entity type marker in the industrial economic operation domain, and an initially empty economic operation event source list field. Simultaneously with storage in the registration pool, the new node's reference information is synchronously written to the corresponding bucket position in the two-layer retrieval index structure, enabling subsequent identical candidate nodes to be retrieved.

[0097] Step S326: When the exact match retrieval returns a non-empty set, determine that there is a registered semantic unit node in the semantic unit node registration pool that matches the candidate semantic unit node, and extract the unique identifier of the existing semantic unit node from the returned registered semantic unit node record.

[0098] The value of the unique identifier field of the semantic unit node is directly read from the registered semantic unit node record returned by the second-level index. This value is the unified node identifier that the candidate semantic unit node should be mapped to in the global scope.

[0099] Step S327: Construct a coreference mapping record from the original entity semantic unit of the candidate semantic unit node in its economic operation event record to the unique identifier of the existing semantic unit node, and add the coreference mapping record to the coreference mapping set.

[0100] The coreference mapping record is stored as a key-value pair. The key is the original local identifier of the candidate semantic unit node in its respective economic operation event record, and the value is the unique identifier of the existing semantic unit node obtained in step S326. This key-value pair is added to a global coreference mapping set, which provides a globally unified identifier conversion basis when replacing the node identifier of the association strength parameter between semantic units in the subsequent step S330.

[0101] Step S330: Based on the common reference mapping relationship, replace the entity semantic units involved in the semantic unit association strength parameter within each economic operation event record with the corresponding semantic unit node unique identifier, and generate internal semantic association edge candidates represented by the semantic unit node unique identifier. Each internal semantic association edge candidate includes the starting semantic unit node unique identifier, the ending semantic unit node unique identifier, the semantic association direction marker, and the semantic unit association strength parameter component.

[0102] For each association strength record in the set of semantic unit association strength parameters for each economic operation event record, its index key is an entity semantic unit pairing sample. The first and second entity semantic units within this pairing sample are represented by local identifiers within the event record. The unique identifier of the global semantic unit node corresponding to the local identifier of the first entity semantic unit is searched in the coreference mapping set to obtain the unique identifier of the starting semantic unit node; the unique identifier of the global semantic unit node corresponding to the local identifier of the second entity semantic unit is searched to obtain the unique identifier of the ending semantic unit node. Simultaneously, the semantic association direction marker and the semantic unit association strength parameter components are extracted from the semantic relationship inference results of this association strength record. The four fields—the unique identifier of the starting semantic unit node, the unique identifier of the ending semantic unit node, the semantic association direction marker, and the semantic unit association strength parameter components—are merged into a candidate internal semantic association edge. All the quadruplets that have completed the replacement together constitute the candidate list of internal semantic association edges for this economic operation event record.

[0103] Step S340: Merge the candidate internal semantic association edge with the registered semantic association edges in the semantic association edge registration pool that are identified by the same starting semantic unit node unique identifier and the same ending semantic unit node unique identifier. If there are no registered semantic association edges identified by the same node pair in the semantic association edge registration pool, the candidate internal semantic association edge is registered as a new semantic association edge in the semantic association edge registration pool. If there are, the semantic unit association strength parameter component of the candidate internal semantic association edge is fused with the edge weight of the registered semantic association edge through a preset strength fusion function, and the edge weight of the registered semantic association edge is updated.

[0104] In the semantic association edge registration pool, each registered semantic association edge uses a tuple of a starting semantic unit node's unique identifier and a terminating semantic unit node's unique identifier as the unique key to distinguish different semantic association edges. This tuple is the node pair identifier. When internal semantic association edge candidates from different economic operation event records have the exact same node pair identifier, these candidate edges repeatedly reflect the semantic association relationship between the same pair of semantic unit nodes in different event contexts. It is necessary to fuse and merge the association strength information they carry across records to form a comprehensive evaluation of the strength of the semantic association relationship.

[0105] Specifically, the merging determination and fusion update process in step S340 is elaborated in detail through steps S341 to S347.

[0106] Step S341: Extract the unique identifiers of the starting semantic unit node and the ending semantic unit node of the candidate internal semantic association edge to form a node pair identifier tuple.

[0107] Step S342: Using the node pair identifier tuple as the query condition, traverse and search the semantic association edge registration pool to obtain the set of registered semantic association edges where both the unique identifier of the starting semantic unit node and the unique identifier of the ending semantic unit node match the node pair identifier tuple.

[0108] The semantic edge registration pool internally maintains a hash map index with node pair identifier tuples as keys. The node pair identifier tuples of candidate semantic edges are input into this hash map, which returns a list of registered semantic edges associated with that node pair identifier in constant time complexity. If the node pair identifier is not registered in the hash map, an empty list is returned.

[0109] Step S343: Detect whether the set of registered semantic association edges contains at least one registered semantic association edge.

[0110] The list returned in step S342 is checked for the number of elements. If the number of elements is zero, it is determined that there are no registered semantic association edges with the same node pair identifier; if the number of elements is greater than zero, it is determined that there are.

[0111] Step S344: When the set of registered semantic association edges is empty, the semantic association direction label and the semantic unit association strength parameter component of the candidate internal semantic association edges are used as the initial edge weight attribute value of the semantic association edge to generate a new registered semantic association edge, and the new registered semantic association edge is written into the semantic association edge registration pool.

[0112] The newly generated registered semantic association edge data record contains the following fields: the unique identifiers of the starting and ending semantic unit nodes are taken from the corresponding fields of the internal semantic association edge candidates; the semantic association direction marker is taken from the internal semantic association edge candidates; the edge weight attribute value field is assigned the current value of the association strength parameter component between semantic units of the internal semantic association edge candidates; the edge weight last update time field is assigned the current system timestamp; and the historical cumulative update count field is assigned the initial value. This new record is written to the semantic association edge registration pool, and simultaneously, the hash mapping index within the semantic association edge registration pool is updated synchronously, registering the reference of the new record under the corresponding node pair identifier tuple key.

[0113] Step S345: When the set of registered semantic association edges is not empty, select the registered semantic association edge whose edge weight was last updated earliest from the set of registered semantic association edges as the semantic association edge to be fused.

[0114] For all registered semantically related edge records in the set of registered semantically related edges, compare the value of the most recent update time field of their edge weights one by one, and select the record with the smallest timestamp value as the semantically related edge to be fused. The basis for selecting the edge with the earliest update time for fusion is that this edge has not been updated for the longest time, and its current edge weight value most needs to be corrected by the new observation samples.

[0115] Step S346: Obtain the current edge weight attribute value of the semantic association edge to be fused, input the current edge weight attribute value and the association strength parameter component between semantic units of the candidate semantic association edge into the preset strength fusion function, the strength fusion function calculates the weighted harmonic average value of the current edge weight attribute value and the association strength parameter component between semantic units, and outputs the fused edge weight attribute value.

[0116] The strength fusion function uses a weighted harmonic mean as the fusion operator. The calculation path of the weighted harmonic mean is as follows: take the reciprocal of the current edge weight attribute value, multiply this reciprocal by the historical cumulative update count of the semantically related edge to be fused, and obtain the first weighting term; take the reciprocal of the semantic unit association strength parameter component of the candidate semantically related edge, and use 1 as its weight, to obtain the second weighting term. Summing the first and second weighting terms and dividing by the total weight of the update count plus 1, we obtain the reciprocal value of the weighted harmonic mean. Taking the reciprocal of this reciprocal value again yields the fused edge weight attribute value. The characteristic of the weighted harmonic mean as a fusion operator is that it has a strong damping effect on abnormally low values ​​close to zero, and can effectively avoid the drastic reduction of the long-term accumulated edge weight estimate due to extremely low association strength noise extracted from individual event records.

[0117] Step S347: The edge weight update operation of the semantically related edge to be merged is completed by overwriting the edge weight attribute value of the merged edge weight attribute value with the merged edge weight attribute value.

[0118] Replace the edge weight attribute value field of the semantic association edge to be merged with the merged edge weight attribute value calculated in step S346, update the edge weight most recent update time field of the semantic association edge to be merged to the current system timestamp, and increase the value of the historical cumulative update count field of the semantic association edge to be merged by 1.

[0119] Step S350: Traverse each semantic association edge in the semantic association edge registration pool, extract the unique identifier of the starting semantic unit node and the unique identifier of the ending semantic unit node of the semantic association edge, and establish a semantic association edge connection with edge weight between the corresponding semantic unit nodes.

[0120] After merging and aggregating all internal semantic association edge candidates, each registered semantic association edge stored in the semantic association edge registration pool represents a semantic association relationship between a pair of semantic unit nodes in the semantic space of industrial economic operation, achieved through cross-record fusion. Each semantic association edge record in the registration pool is traversed. The starting semantic unit node data object is located in the semantic unit node registration pool based on its unique identifier, and the ending semantic unit node data object is located based on its unique identifier. A directed connection edge data structure instance is created between the two node data objects. The direction of this connection edge is consistent with the semantic association direction marker stored in the semantic association edge record, and the weight attribute of the connection edge is assigned the edge weight attribute value of the semantic association edge record.

[0121] Step S360: Collect all semantic unit nodes in the semantic unit node registration pool to form a semantic unit node set; collect all semantic association edges in the semantic association edge registration pool to form a semantic association edge set.

[0122] The semantic unit node set is a list of nodes formed by transforming all registered semantic unit node records stored in the semantic unit node registration pool into a graph node data structure. The semantic association edge set is a list of edges formed by transforming all registered semantic association edge records stored in the semantic association edge registration pool into a directed weighted edge data structure.

[0123] Step S370: Generate an industrial economic operation semantic association network based on the set of semantic unit nodes and the set of semantic association edges. The industrial economic operation semantic association network uses semantic unit nodes as semantic unit nodes and semantic association edges as semantic association edges to obtain a network association structure.

[0124] Using all nodes in the semantic unit node set as the vertex set of the graph structure, and all edges in the semantic association edge set as the directed weighted edge set of the graph structure, a complete directed weighted graph data structure is constructed in memory. The internal storage form of the graph adopts an adjacency list representation. Each vertex's adjacency list entry stores the target vertex index of all adjacent edges originating from that vertex, as well as the edge weight and direction label information of each adjacent edge. This directed weighted graph is the internal data form of the semantic association network of industrial economic operation. At the global level, it weaves the entity semantic units and their semantic associations contained in the massive industrial economic operation event records into a semantic network that can be efficiently retrieved and traversed.

[0125] Step S400: Receive a target retrieval request, which contains a text description of the query intent in natural language, describing changes in target economic indicators or related business events.

[0126] The target retrieval request is the input carrier for users to initiate semantic retrieval of industrial economic operation data. Its receiving process relies on an externally open request receiving interface. Various sub-functional modules attached to this interface work together to complete preprocessing tasks such as interception, verification, cleaning, language adaptation, and intent type labeling of the request message. Finally, it is assembled into a standardized structured retrieval request data unit and sent to the retrieval engine.

[0127] In an optional embodiment, the complete implementation process of step S400 is carried out from step S410 to step S460.

[0128] Step S410: Establish a request receiving interface for industrial economic operation retrieval services. The request receiving interface continuously listens for retrieval request messages from industrial economic operation management terminal devices. When the request receiving interface intercepts a retrieval request message, it extracts the payload field content of the retrieval request message to obtain the original retrieval request string.

[0129] The request receiving interface is constructed according to the standard pattern of the network service framework. It binds a listening socket to a specified network port and waits for connection requests and data packets from the industrial economic operation management terminal equipment in an asynchronous event loop or blocking reception manner. When a complete packet is successfully received, the request receiving interface first performs transport layer protocol parsing and encoding format identification on the packet, and then extracts the user's search input content carried in string form from the packet's payload field. This content is the original search request string.

[0130] Step S420: Perform text validity validation on the original search request string. Check whether the original search request string contains non-whitespace printable characters. If it contains non-whitespace printable characters, the original search request string is determined to be a valid request text. Otherwise, the search request message is rejected and an empty request prompt is returned.

[0131] The text validity verification module iterates through the original search request string character by character, determining whether each character is a printable character and not a whitespace character based on its category attribute in the Unicode character set. If at least one character meeting the criteria is found before reaching the end of the string, it indicates that the original search request string contains substantial text content, the verification passes, and the string is marked as valid request text and proceeds to the next processing stage. If no matching characters are found after traversing the entire string, the verification fails, the request receiving interface directly returns a predefined error message indicating an empty request to the requesting terminal, and the current search request processing flow terminates.

[0132] Step S430: Normalize the original search request string that is determined to be a valid request text by unifying the format of full-width and half-width characters, removing leading and trailing whitespace characters, and merging consecutive whitespace characters to obtain normalized search request text.

[0133] Character normalization is a multi-stage string cleaning and standardization pipeline, consisting of multiple sub-steps from S431 to S437 arranged sequentially. Each sub-step applies progressively increasing normalization transformations to the input string. Specifically: Step S431: Construct a mapping conversion table from full-width characters to half-width characters. The mapping conversion table contains mapping entries from full-width letter characters, full-width number characters, and full-width punctuation characters to their corresponding half-width characters.

[0134] The mapping conversion table is a one-way lookup dictionary, compiled based on the encoding correspondence between full-width characters and their corresponding half-width characters as defined in the Unicode standard. The dictionary's key fields contain the code points of uppercase and lowercase Latin letters, Arabic numerals, and common punctuation marks in full-width form, while the value fields contain the code points of half-width characters that are semantically equivalent to the full-width characters in the key fields but occupy half the display width.

[0135] Step S432: Scan each character in the original search request string. When a scanned character is located in the mapping conversion table, replace the character with the corresponding half-width character through the mapping conversion table until all characters have been scanned, and obtain the half-width unified search request string.

[0136] Starting from the first character of the original search request string and ending with the last character, each character is read sequentially, and its Unicode code point is used as the key to look up the mapping conversion table dictionary. If the key exists in the dictionary, the character is replaced with the corresponding half-width character in the dictionary; if the key does not exist in the dictionary, the original character is left unchanged. After all characters have been processed, the new string obtained is the half-width unified search request string.

[0137] Step S433: Remove leading and trailing whitespace characters from the half-width unified search request string. Locate the first non-whitespace character in the half-width unified search request string as the content start position and the last non-whitespace character as the content end position. Extract the substring between the content start position and the content end position as the whitespace-free search request string.

[0138] Starting from the first character of the unified whitespace search request string and proceeding towards the end, each character is checked for whitespace attributes. When the first character that does not meet the whitespace condition is encountered, its offset index in the string is recorded as the start position of the content. Simultaneously, starting from the last character of the unified whitespace search request string and proceeding towards the beginning, the first character that does not meet the whitespace condition is encountered, and its offset index is recorded as the end position of the content. Using the start position of the content as the starting point and the end position of the content as the ending position, a substring is extracted from the original string. The resulting substring is the whitespace-removed search request string.

[0139] Step S434: Compress the whitespace-free search request string by sequentially scanning the characters in the whitespace-free search request string. When a sequence of consecutive whitespace characters is encountered, replace the consecutive whitespace character sequence with a single standard space character to generate a whitespace-free search request string.

[0140] During the sequential scanning of the whitespace-removed search request string, a character output buffer and an internal whitespace status flag are maintained. When the currently scanned character is a non-whitespace character, it is directly appended to the end of the output buffer, and the whitespace status flag is set to false. When the currently scanned character is a whitespace character and the whitespace status flag is currently false, a standard space character is appended to the end of the output buffer, and the whitespace status flag is set to true. When the currently scanned character is a whitespace character and the whitespace status flag is currently true, no append operation is performed, and the scanning of the next character continues. After the scan is completed, the character sequence in the output buffer is the whitespace-free search request string.

[0141] Step S435: Detect whether there are non-printable control characters in the search request string without consecutive whitespace. If so, remove the non-printable control characters from the search request string without consecutive whitespace and generate a clean search request string.

[0142] For each character in the search request string without consecutive whitespace, determine its Unicode category to see if it belongs to the control character category or the invisible formatting character category. Keep all characters that do not belong to the above categories in their original order, discard all characters that belong to the above categories, and reassemble the kept characters into a clean search request string in sequence.

[0143] Step S436: Use a preset spell correction dictionary to perform lexical matching correction on the clean search request string, identify the word elements in the clean search request string that do not completely match the standard word forms in the spell correction dictionary, replace the word elements with the corresponding standard word forms in the spell correction dictionary, and obtain the spell-corrected search request string.

[0144] The spelling correction dictionary pre-includes standard and normative word forms of common terms and search keywords in the field of industrial and economic operations, and records the common abbreviations, acronyms, and spelling variations of each standard word form in actual user input. The clean search request string is segmented to obtain a word sequence. For each word in the sequence, its edit distance to each standard word form and its variations in the spelling correction dictionary is calculated. The dictionary entry with the smallest edit distance within the allowed threshold range is searched. If the minimum edit distance is zero, the word form is already a standard word form and remains unchanged. If the edit distance is greater than zero and within the threshold range, the word form is replaced with the standard word form associated with that dictionary entry. If the minimum edit distance exceeds the threshold range, the original word form is retained. After all word forms have been processed, they are concatenated in their original order to restore the spell-corrected search request string.

[0145] Step S437: Normalize the sentence boundaries of the spell-corrected retrieval request string by appending a standard sentence end identifier to the end of the spell-corrected retrieval request string to obtain the normalized retrieval request text.

[0146] After checking the spelling correction, the last character of the search request string is checked. If the last character is already a standard sentence end identifier, the original string is returned as the result; otherwise, a standard sentence end identifier character is directly appended to the end of the original string to obtain the normalized search request text.

[0147] Step S440: Identify the language attribute of the normalized retrieval request text. When the language attribute indicates that it is the natural language used by the major industrial economic participants, the normalized retrieval request text is directly used as the query intent description text. Otherwise, the normalized retrieval request text is converted into the natural language used by the major industrial economic participants through the embedded language conversion processing unit before being used as the query intent description text.

[0148] Language attribute identification is performed by a built-in language identification model. This model extracts character-level byte-pair encoding statistical features from the normalized search request text and inputs these features into a fully connected classifier to predict the language encoding of the normalized search request text. If the predicted language encoding matches the natural language encoding used by the main industrial economic participants in the system's preset list, the normalized search request text is directly passed down as the query intent description text. If the predicted language encoding does not match the preset encoding, the embedded language conversion processing unit is activated. This unit uses a sequence-to-sequence machine translation structure to convert the normalized search request text into a language that is semantically consistent with the original text and is used as the query intent description text.

[0149] Step S450: Determine the conceptual category of the target economic operation indicator or the event type category of the related business event referred to by the query intent description text, and generate a search intent type tag.

[0150] The search intent type label is a conceptual, structured classification label that categorizes the search needs expressed in the query intent description text. This label will serve as an important constraint feature for subsequent semantic alignment and search result ranking. For example, step S450 is implemented through steps S451 to S456.

[0151] Step S451: Input the query intent description text into the hierarchical intent understanding processing component of the semantic big model. Semantically encode the query intent description text through the basic semantic encoding layer of the hierarchical intent understanding processing component to generate a query intent semantic encoding sequence.

[0152] The hierarchical intent understanding processing component is a set of modules in the semantic big model specifically designed for understanding query intent. Its basic semantic encoding layer shares the main structure with the multi-layer context encoding network used in step S220. After the query intent description text undergoes input embedding transformation and self-attention encoding and feedforward fully connected encoding across multiple encoding layers, a query intent semantic encoding sequence is formed at the output of the last encoding layer. The vector at each position in the sequence is a deep semantic representation of the semantic symbol at that position in the corresponding query intent description text under the constraints of the complete query context.

[0153] Step S452: The intent trigger word recognition layer of the hierarchical intent understanding processing component performs trigger word scanning on the semantic encoding sequence of the query intent, identifies the economic operation indicator mention fragments pointing to specific economic operation indicators or event mention fragments pointing to related business events in the query intent description text, and outputs a set of intent trigger word fragments.

[0154] The intent trigger word recognition layer performs trigger word boundary discrimination at each encoding position of the query intent semantic encoding sequence. The discrimination method is the same as that of entity semantic unit boundary recognition in step S230, that is, the start position score and the end position score are generated by the start boundary linear classifier and the end boundary linear classifier respectively and compared with their respective thresholds. However, the two classifiers of the intent trigger word recognition layer are fine-tuned and trained on the labeled data of industrial economic operation indicator mentions and business event mentions, so they are specifically inclined to detect mention fragments pointing to economic operation indicators and business events. All successfully detected mention fragments, together with their trigger type labels that are identified as economic operation indicator mentions or event mentions, constitute the intent trigger word fragment set.

[0155] Step S453: Perform concept normalization on each intent trigger word fragment in the intent trigger word fragment set, and map the intent trigger word fragment to the standard concept node in the predefined industrial economic operation concept classification system to obtain the standard concept node identifier.

[0156] The Industrial Economic Operation Concept Classification System is a pre-constructed hierarchical concept classification tree or graph covering the core concepts of various fields of industrial economic operation. Each standard concept node in the system represents a standardized industrial economic concept or event type, and the concept nodes form a hierarchical structure through parent-child relationship edges. For each intent-triggered word fragment in the intent-triggered word fragment set, the concept normalization operation transforms its textual surface form into a dense vector, and then calculates its similarity with the pre-computed vector representations of all standard concept nodes in the Industrial Economic Operation Concept Classification System. The similarity is measured using cosine similarity, and the standard concept node corresponding to the maximum cosine similarity is selected as the normalization mapping target to obtain the unique standard concept node identifier of that standard concept node in the concept classification system.

[0157] Step S454: Based on the hierarchical position and parent-child concept relationship of the standard concept node identifier in the industrial economic operation concept classification system, determine the category of economic operation indicator concept or business event type corresponding to the standard concept node identifier.

[0158] In the data structure of the concept classification system, the position of a concept node in the system can be directly located through the standard concept node identifier, and the complete concept hierarchy path can be traced upwards to the root node. This concept hierarchy path completely describes the concept category or event type category to which the query intent belongs in the industrial economic operation concept classification system, and each ancestor node on the path constitutes a higher-level semantic classification constraint for the intent.

[0159] Step S455: When the intent trigger word fragment set contains multiple intent trigger word fragments, obtain the economic operation indicator concept category or business event type category corresponding to each intent trigger word fragment to obtain a candidate intent type set, and select the dominant intent type from the candidate intent type set through a preset voting priority rule.

[0160] Specifically, the frequency of occurrence of each economic operation indicator concept category or business event type category in the candidate intent type set is counted, and the category with the highest frequency is selected as the dominant intent type. When multiple categories have the same highest frequency, the one whose corresponding intent trigger word fragment appears first in the query intent description text is selected as the dominant intent type.

[0161] Step S456: Encode the dominant intent type into a fixed-format retrieval intent type tag string, which includes the concept classification system version number and a standard concept node path description.

[0162] The encoding of the search intent type tag string adopts a prefix-delimiter-suffix concatenation format. The prefix is ​​the current version number string of the industrial economic operation concept classification system, and the suffix is ​​the standard concept node path description. The path description is formed by sequentially connecting the identifiers of each level node along the path from the root node of the concept classification system to the standard concept node corresponding to the dominant intent type, according to the path delimiter.

[0163] Step S460: Assemble the query intent description text, the search intent type label, and the request receipt timestamp into a target search request. The target search request stores the query intent description text, the search intent type label, and the request receipt timestamp in the form of structured data units.

[0164] The request receipt timestamp is read and temporarily stored by the system clock when the request receipt interface intercepts the retrieval request message in step S410. The three fields—query intent description text, retrieval intent type label, and request receipt timestamp—are encapsulated into a structured data unit according to the field name-field value mapping structure. This data unit is the target retrieval request and serves as the input parameter for the retrieval execution phase in step S500.

[0165] Step S500: Based on the semantic association network of industrial economic operation, the target retrieval request is parsed and processed to obtain the query semantic unit group through the semantic big model. Then, semantic alignment and positioning processing and association path extraction are performed in the semantic association network of industrial economic operation according to the query semantic unit group to generate a set of economic operation data retrieval results consisting of matching economic operation event records.

[0166] For example, the complete implementation process of step S500 is carried out from step S510 to step S570.

[0167] Step S510: Extract semantic units from the query intent description text in the target retrieval request based on the semantic big model to obtain the query semantic unit set and the correlation strength parameters between query semantic units. Each query semantic unit in the query semantic unit set shares a unified semantic entity term expression form and an entity type labeling system in the industrial economic operation semantic association network with the semantic unit nodes in the industrial economic operation domain.

[0168] The semantic unit extraction and reuse step S200 is a pipeline for extracting complete semantic units from the event description text. The query intent description text is sequentially transformed into an initial semantic symbol sequence by an input embedding transformation component, then passed through a multi-layer context encoding network to obtain a deep semantic understanding encoding sequence, and finally processed by a semantic entity boundary recognition processing component to extract a set of entity semantic unit fragments. Each fragment is then assigned an entity type label from the industrial economic operation domain by an entity category labeling processing component, resulting in a set of query semantic units. Finally, the set of query semantic units is obtained through entity semantic unit pairing and semantic relationship inference channels. Throughout the extraction process, the entity type labeling system used is completely consistent with that used when constructing the semantic association network, ensuring full alignment of the query semantic units on the query side and the semantic unit nodes on the network side in terms of type system.

[0169] Step S520: For each query semantic unit in the query semantic unit set, perform node matching retrieval in the semantic unit node set of the industrial economic operation semantic association network, find matching semantic unit nodes that have the same entity semantic unit text surface form and the same industrial economic operation domain entity type label as the query semantic unit, establish an alignment mapping relationship from the query semantic unit to the matching semantic unit node, and if there is no matching semantic unit node, mark the query semantic unit as an unaligned semantic unit.

[0170] In one embodiment, the node matching retrieval process is specifically implemented through the following steps S521 to S527.

[0171] Step S521: Extract the entity semantic unit text surface form of the query semantic unit, and standardize the entity semantic unit text surface form in accordance with the standardization processing rules used in the semantic unit node registration pool of the industrial economic operation semantic association network to obtain the standardized query text surface form.

[0172] The standardization processing rules are consistent with those used in step S322 when constructing the semantic unit node registration pool. That is, the same operations of uniform conversion of uppercase and lowercase letters, removal of leading and trailing whitespace characters, and merging of consecutive whitespace characters are performed on the query side to ensure that the standardized text surface form on the query side is completely aligned with the standardized text surface form used by the network side registration nodes in terms of format processing.

[0173] Step S522: Extract the entity type marker of the industrial economic operation field of the query semantic unit and combine it with the standardized query text surface form to form the query semantic unit retrieval key.

[0174] The query semantic unit retrieval key adopts the same structure as the retrieval key tuple in step S323. The first field is the standardized query text surface form, and the second field is the entity type marker in the field of industrial economic operation.

[0175] Step S523: Construct a hash index structure for the set of semantic unit nodes of the semantic association network of industrial economic operation. The hash index structure uses the joint hash value of the standardized entity semantic unit text surface form and the entity type mark in the industrial economic operation field as the index slot.

[0176] When constructing this hash index structure, the standardized entity semantic unit text surface form of each semantic unit node in the semantic unit node set is concatenated with its industrial economic operation domain entity type label using a fixed separator to form a joint string. A hash function is applied to this joint string to calculate the hash value. The hash value is then moduloed by the total number of slots in the hash index structure to obtain the storage slot number of the semantic unit node in the hash index structure. All semantic unit nodes are assigned to the candidate semantic unit node linked list of the corresponding slot according to this rule.

[0177] Step S524: Calculate the joint hash value by querying the semantic unit retrieval key, locate the candidate hash slot in the hash index structure by the joint hash value, and obtain the list of candidate semantic unit nodes with the same joint hash value from the candidate hash slot.

[0178] After concatenating the two fields of the query semantic unit retrieval key using the same concatenation rules as in step S523, the hash value is calculated using the same hash function and modulo the total number of slots to obtain the hit slot number in the hash index structure. All candidate semantic unit nodes are retrieved from the candidate semantic unit node linked list corresponding to that slot to form a candidate semantic unit node list.

[0179] Step S525: For each candidate semantic unit node in the candidate semantic unit node list, perform a precise comparison: compare whether the standardized entity semantic unit text surface form of the candidate semantic unit node is consistent with the standardized query text surface form character by character, and compare whether the industrial economic operation field entity type label of the candidate semantic unit node is consistent with the industrial economic operation field entity type label of the query semantic unit category by category.

[0180] The precise comparison consists of two independent string equality checks. The first check compares the standardized entity semantic unit text surface form string of the candidate semantic unit node with the standardized query text surface form string of the query semantic unit character-by-character. The check is true only if the two strings are of equal length and all character code points are identical. The second check compares the industrial economic operation domain entity type marker string of the candidate semantic unit node with the industrial economic operation domain entity type marker string of the query semantic unit character-by-character. The check is true only if they are completely identical.

[0181] Step S526: When both are exactly matched, the candidate semantic unit node is determined as the matching semantic unit node, the alignment mapping relationship between the query semantic unit and the matching semantic unit node is recorded, and the retrieval process of the query semantic unit is terminated.

[0182] The alignment mapping is recorded in a mapping entry, which stores the correspondence between the local identifier of the query semantic unit and the unique identifier of the global semantic unit node of the matching semantic unit node.

[0183] Step S527: When no candidate semantic unit node is found that matches the other two after traversing the candidate semantic unit node list, mark the query semantic unit as an unaligned semantic unit, and temporarily store the standardized query text surface form and industrial economic operation domain entity type mark of the unaligned semantic unit in the unaligned semantic unit buffer.

[0184] The unaligned semantic unit buffer is used to record all query semantic units that failed to match the corresponding semantic unit node in the semantic association network, for subsequent result feedback or manual maintenance reference.

[0185] Step S530: Based on the alignment mapping relationship, map the starting query semantic unit and the ending query semantic unit in the correlation strength parameter between query semantic units to the corresponding starting matching semantic unit node and ending matching semantic unit node in the industrial economic operation semantic correlation network, respectively, to obtain the query semantic correlation edge template.

[0186] For each association strength record in the query semantic unit association strength parameter, the local identifiers of the first and second entity semantic units are extracted. The unique identifiers of the global semantic unit nodes corresponding to these two local identifiers are then searched in the alignment mapping relationship to obtain the unique identifiers of the starting and ending matching semantic unit nodes. Simultaneously, the semantic association direction marker and the semantic unit association strength parameter components are read from the association strength record. The unique identifiers of the starting and ending matching semantic unit nodes, the semantic association direction marker, and the semantic unit association strength parameter components are combined into a query semantic association edge template. If the starting or ending query semantic unit involved in a certain association strength record is marked as unaligned, no query semantic association edge template is generated for that record.

[0187] Step S540: By querying the semantic association edge template, perform association edge matching retrieval in the semantic association edge set of the industrial economic operation semantic association network, and filter out matching semantic association edges that are completely consistent with the query semantic association edge template at the starting matching semantic unit node and the ending matching semantic unit node and have the same semantic association direction label, thus forming a matching semantic association edge set.

[0188] For each query semantic association edge template, retrieve registered semantic association edges from the semantic association edge set of the industrial economic operation semantic association network that simultaneously satisfy the following three constraints: the edge's starting semantic unit node unique identifier equals the template's starting matching semantic unit node unique identifier; the edge's ending semantic unit node unique identifier equals the template's ending matching semantic unit node unique identifier; and the edge's semantic association direction label equals the template's semantic association direction label. All edges that meet the conditions are included in the matching semantic association edge set.

[0189] Step S550: Using each matching semantic association edge in the set of matching semantic association edges as the starting edge of the path, perform adjacency expansion traversal along the semantic association edges of the industrial economic operation semantic association network. During the traversal, the edge weights are weighted according to the correlation strength parameter between query semantic units, and the expansion path with priority given to the correlation weight value is selected to obtain a semantic association path covering multiple matching semantic unit nodes.

[0190] The adjacency expansion traversal employs a priority queue-based heuristic graph search strategy. Initially, each matching semantically related edge in the set of matching semantically related edges is treated as an independent single-step path and placed into the priority queue. Each single-step path records the node sequence and cumulative conformity value. The initial value of the cumulative conformity value is set to the reciprocal or inverse of the absolute value of the deviation between the edge weight of that matching edge and the component of the correlation strength parameter between the corresponding query semantic unit. The priority queue is arranged in descending order of cumulative conformity value.

[0191] In each iteration of the search, the path with the highest cumulative conformity value is popped from the priority queue. Starting from the last node of this path, all adjacent nodes connected to the last node through semantic association edges in the semantic association network of industrial economic operation are traversed. For each adjacent node, the deviation metric between the edge weight of the connecting edge and the corresponding semantic unit association strength parameter component is calculated. The deviation metric is accumulated to the cumulative conformity value of the current path according to a certain influence coefficient. If the adjacent node has not yet appeared in the node sequence of this path, a new path is generated and pushed into the priority queue. The search terminates when the priority queue is empty, the cumulative conformity value of the dequeued path is lower than a preset truncation threshold, or the number of expansion steps reaches the preset maximum path length. From the paths retained in the priority queue at the end of the search, the batch of paths with the highest cumulative conformity value is selected as semantic association paths.

[0192] Step S560: Extract all semantic unit nodes on the semantic association path, trace back the economic operation event records associated with each semantic unit node, obtain the event description text and event source system identifier of each economic operation event record, and obtain a set of candidate economic event records.

[0193] Each semantic unit node on the semantic association path is bound to an economic operation event source list in the semantic unit node registration pool. This list records which entity semantic units from which economic operation event records were merged by the semantic unit node during the construction phase through coreference mapping.

[0194] Specifically, step S560 completes the backtracking from the path node to the event step by step through steps S561 to S567.

[0195] Step S561: Perform node traversal on the semantic association path, collect the unique identifiers of all semantic unit nodes appearing on the semantic association path, and form a set of unique identifiers of path nodes after deduplication.

[0196] Step S562: Traverse each unique identifier of a path node in the set of unique identifiers of path nodes, access the semantic unit node registration pool of the semantic association network of industrial economic operation, find the semantic unit node registration information corresponding to the unique identifier of the path node, and extract the list of economic operation event sources recorded by the semantic unit node in the semantic unit node registration pool. The list of economic operation event sources contains the event record identifier of the economic operation event record from which the semantic unit node was extracted.

[0197] Step S563: Merge the economic operation event source lists corresponding to the unique identifiers of all path nodes, remove duplicate event record identifiers, and obtain a set of candidate event record identifiers.

[0198] Step S564: For each event record identifier in the candidate event record identifier set, retrieve the complete data unit of the economic operation event record corresponding to the event record identifier from the original storage. The complete data unit includes the event description text, the event source system identifier, and the event occurrence time context description information.

[0199] Step S565: Weight the source credibility of economic operation event records by using the event source system identifier, assign a higher source credibility weight to economic operation event records from the preset core monitoring system, and assign a lower source credibility weight to economic operation event records from the external reporting system.

[0200] Source credibility weighting is performed based on a pre-configured system credibility weight table. This table categorizes the system codes of each industrial information system into different credibility levels according to objective indicators such as the degree of automation in data acquisition, direct connection to physical equipment, and the number of relay levels in the data link. Core monitoring systems belong to the highest credibility level and are assigned the highest weight, while peripheral reporting systems belong to lower credibility levels and are assigned lower weights. The source system identifier carried in the economic operation event record is used to perform code matching in the system credibility weight table; once a match is successful, the corresponding source credibility weight value is read.

[0201] Step S566: Combine the source credibility weight and the semantic coverage of the event description text of the economic operation event record with the query semantic unit group to calculate the final retrieval relevance score for each economic operation event record.

[0202] When determining semantic coverage, the total number of query semantic units that have successfully established alignment mapping relationships in the query semantic unit set can be counted as the total number of query semantic units. Then, the number of entity semantic units in the event description text of the current economic operation event record that are semantically matched with these aligned query semantic units through semantic alignment can be counted as the number of covered semantic units. The ratio of the number of covered semantic units to the total number of query semantic units is used as the semantic coverage. The final retrieval relevance score can be obtained by multiplying the semantic coverage, the source credibility weight, and a correlation path conformity factor from the semantic association path search stage.

[0203] Step S567: Sort the economic operation event records from high to low according to the final retrieval relevance score to obtain a candidate economic event record set.

[0204] When sorting, the records of economic operation events are arranged in descending order according to the final retrieval relevance score. When multiple economic event records have the same score, they are further sorted in a secondary order according to the event start timestamp in the context description information of their respective event occurrence time.

[0205] Step S570: Sort and organize the economic operation event records in the candidate economic event record set according to the overall semantic coverage with the query semantic unit group to generate an economic operation data retrieval result set.

[0206] The economic operation data retrieval results are organized in the form of a structured array or list. Each array element corresponds to an economic operation event record. The entry includes the original text of the event description, the system identifier of the event source, the contextual description information of the event occurrence time, and the final retrieval relevance score of the record. After the entries are sorted in descending order of the score, the entire result set is returned to the initiator of the target retrieval request, completing the full semantic retrieval of industrial economic operation data.

[0207] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A method for retrieving industrial economic operation data based on a semantic large model, characterized in that, The method includes: Obtain a set of original records of industrial economic operation data. The set of original records of industrial economic operation data contains multiple event description texts in the form of natural language text describing the process of economic operation activities. Each event description text corresponds to an independent economic operation event record. The economic operation event record carries the event source system identifier and event occurrence time context description information. Using a pre-trained semantic big model, semantic units are extracted from the original data set of industrial economic operation records to obtain the semantic unit set corresponding to each economic operation event record and the correlation strength parameter between semantic units. The semantic unit set is composed of semantic entity terms extracted from the event description text of the economic operation event record, and the correlation strength parameter between semantic units represents the measure of the degree of semantic association between different semantic entity terms within the economic operation event record. An industrial economic operation semantic association network is formed based on the semantic unit set and the correlation strength parameters between the semantic units. This network includes a semantic unit node set consisting of all semantic entity terms extracted from all economic operation event records as semantic unit nodes, and a semantic association edge set consisting of the correlation strength parameters between semantic units as edge weights. The semantic unit nodes are connected by these semantic association edges to form a mesh-like association structure. Specifically, this includes: initializing an empty semantic unit node registration pool and an empty semantic association edge registration pool; traversing the semantic unit set corresponding to each economic operation event record, and extracting candidate semantic unit nodes from each entity semantic unit set as candidate semantic unit nodes. The entity semantic unit text surface form and industrial economic operation domain entity type mark are defined. For each candidate semantic unit node, the semantic unit node registration pool is searched to see if there is a registered semantic unit node with the same entity semantic unit text surface form and industrial economic operation domain entity type mark as the candidate semantic unit node. If not, the candidate semantic unit node is registered as a new semantic unit node and assigned a unique semantic unit node identifier. If it exists, the unique semantic unit node identifier of the registered semantic unit node is obtained, and a coreference mapping relationship is established from the candidate semantic unit node to the registered semantic unit node. Based on the coreference mapping relationship, the semantic units within each economic operation event record are strongly associated. The entity semantic units involved in the degree parameter are replaced with the corresponding semantic unit node unique identifiers, generating internal semantic association edge candidates represented by semantic unit node unique identifiers. Each internal semantic association edge candidate includes a starting semantic unit node unique identifier, an ending semantic unit node unique identifier, a semantic association direction marker, and a semantic unit association strength parameter component. The internal semantic association edge candidates are merged and judged with the existing registered semantic association edges in the semantic association edge registration pool that are identified by the same starting semantic unit node unique identifier and the same ending semantic unit node unique identifier. If there are no registered semantic association edges identified by the same node pair in the semantic association edge registration pool, the internal semantic association edge candidates are used as new semantic association edge candidates. Semantic association edges are registered in the semantic association edge registration pool. If they exist, the association strength parameter components between semantic units of the candidate semantic association edges are fused with the edge weights of the registered semantic association edges using a preset strength fusion function, and the edge weights of the registered semantic association edges are updated. Each semantic association edge in the semantic association edge registration pool is traversed, and the unique identifiers of the starting and ending semantic unit nodes of the semantic association edge are extracted. Semantic association edge connections with the edge weights are established between the corresponding semantic unit nodes. All semantic unit nodes in the semantic unit node registration pool are collected to form a semantic unit node set, and all semantic association edges in the semantic association edge registration pool are collected to form a semantic association edge set.An industrial economic operation semantic association network is generated based on the set of semantic unit nodes and the set of semantic association edges. The industrial economic operation semantic association network uses semantic unit nodes as semantic unit nodes and semantic association edges as semantic association edges to obtain the network association structure. Receive a target retrieval request, which includes a query intent description text in natural language text form, describing changes in target economic operation indicators or related business events; Based on the aforementioned industrial economic operation semantic association network, the target retrieval request is parsed using the aforementioned semantic big model to obtain a query semantic unit group. Then, based on the query semantic unit group, semantic alignment and location processing and association path extraction are performed in the industrial economic operation semantic association network to generate a set of economic operation data retrieval results consisting of matched economic operation event records.

2. The method according to claim 1, characterized in that, The process involves using a pre-trained semantic large model to extract semantic units from the original industrial economic operation record dataset, obtaining a set of semantic units corresponding to each economic operation event record and parameters of the correlation strength between semantic units, including: Each economic operation event record in the original data set of industrial economic operation is sequentially input into the input embedding and transformation component of the semantic big model, and the event description text of the economic operation event record is semantically symbolized and mapped to generate an initial semantic symbol sequence. The initial semantic symbol sequence is subjected to multi-layer self-attention context encoding. In each encoding layer, the contextual association information between each semantic symbol within the initial semantic symbol sequence is fused to obtain a deep semantic understanding encoding sequence. Each encoding position in the deep semantic understanding encoding sequence stores the semantic representation data of the corresponding semantic symbol in the context of the complete event description text. The semantic entity boundary recognition processing component of the semantic big model performs boundary scanning on the deep semantic understanding encoding sequence to determine the start position marker sequence and end position marker sequence of the entity semantic unit in the event description text. Based on the start position marker sequence and end position marker sequence, the corresponding text fragments are extracted from the event description text to form an entity semantic unit fragment set. The entity category labeling processing component of the semantic big model is invoked to assign a category to each entity semantic unit fragment in the entity semantic unit fragment set, and an industrial economic operation domain entity type label is assigned to each entity semantic unit fragment. The entity semantic unit fragment carrying the industrial economic operation domain entity type label is used as the entity semantic unit to obtain the entity semantic unit set of the economic operation event record. Based on the local text window where entity semantic units co-occur in the event description text, the entity semantic units in the entity semantic unit set are paired and combined to generate an entity semantic unit pairing sample set. Each entity semantic unit pairing sample in the entity semantic unit pairing sample set contains a first entity semantic unit and a second entity semantic unit. Each entity semantic unit pairing sample in the entity semantic unit pairing sample set is input into the semantic relationship reasoning channel of the semantic big model. Joint reasoning is performed on the semantic dependency direction and dependency tightness between the first entity semantic unit and the second entity semantic unit in the entity semantic unit pairing sample. The semantic association direction label and the corresponding semantic unit association strength parameter component are output for the entity semantic unit pairing sample. Collect the semantic association direction markers and semantic association strength parameter components corresponding to all entity semantic unit pairing samples in the entity semantic unit pairing sample set, and construct a semantic unit association strength parameter set with entity semantic unit pairing samples as index units; The set of entity semantic units and the set of association strength parameters between semantic units are combined and bound together to form the set of semantic units and the association strength parameters between semantic units corresponding to the economic operation event record.

3. The method according to claim 2, characterized in that, The step involves inputting each entity semantic unit pairing sample from the entity semantic unit pairing sample set into the semantic relationship inference channel of the semantic big model, jointly inferring the semantic dependency direction and dependency tightness between the first entity semantic unit and the second entity semantic unit in the entity semantic unit pairing sample, and outputting the semantic association direction label and the corresponding semantic unit association strength parameter components corresponding to the entity semantic unit pairing sample, including: From the deep semantic understanding encoding sequence, the first entity semantic unit encoding fragment is extracted based on the start position marker sequence and the end position marker sequence corresponding to the first entity semantic unit in the entity semantic unit pairing sample; and the second entity semantic unit encoding fragment is extracted based on the start position marker sequence and the end position marker sequence corresponding to the second entity semantic unit. The first entity semantic unit encoding fragment and the second entity semantic unit encoding fragment are input into the sequence splicing processing layer of the semantic relationship reasoning channel, and splicing operation is performed along the sequence length dimension to generate a paired joint encoding sequence. Direction-sensitive positional coding injection is performed on the paired joint coding sequence. A first entity discrimination identifier is added to each positional element in the paired joint coding sequence that comes from the first entity semantic unit coding segment, and a second entity discrimination identifier is added to each positional element that comes from the second entity semantic unit coding segment, to obtain a direction-discriminative joint coding sequence. By performing positive semantic influence reasoning on the directional distinguishable joint coding sequence through the bidirectional semantic influence analysis structure of the semantic relationship reasoning channel, the positive semantic influence strength value of the first entity semantic unit on the second entity semantic unit is obtained, and by performing reverse semantic influence reasoning on the directional distinguishable joint coding sequence, the reverse semantic influence strength value of the second entity semantic unit on the first entity semantic unit is obtained. The direction is determined by comparing the positive semantic influence strength value and the negative semantic influence strength value. When the positive semantic influence strength value is greater than the negative semantic influence strength value, the semantic association direction is marked as the first entity semantic unit pointing to the second entity semantic unit. Otherwise, the semantic association direction is marked as the second entity semantic unit pointing to the first entity semantic unit. The directional discriminative joint coding sequence is input into the global association strength measurement structure of the semantic relationship reasoning channel. Multiple sets of learnable semantic interaction transformation matrices are used to perform multi-head semantic interaction projection on the directional discriminative joint coding sequence to generate association strength feature signals in multiple semantic interaction subspaces. The association strength feature signals under the multiple semantic interaction subspaces are fused and compressed to generate the semantic unit association strength parameter components of the entity semantic unit pairing samples; The semantic association direction marker and the association strength parameter component between the semantic units are paired and bound together for output, so as to obtain the semantic relationship reasoning result of the paired sample of the entity semantic unit.

4. The method according to claim 3, characterized in that, The step involves inputting the directional discriminative joint coding sequence into the global association strength measurement structure of the semantic relationship inference channel, and performing multi-head semantic interaction projection on the directional discriminative joint coding sequence through multiple sets of learnable semantic interaction transformation matrices to generate association strength feature signals in multiple semantic interaction subspaces, including: The direction-discriminative joint coding sequence is copied into multiple copies of the sequence, which are input in parallel. The number of copies is the same as the number of semantic interaction projection heads preset in the global association strength measurement structure. For each copy sequence, apply a set of independent first linear projection transformation parameters and second linear projection transformation parameters to project the copy sequence onto the query semantic space and the key semantic space simultaneously, to obtain the query projection vector sequence and the key projection vector sequence corresponding to the semantic interaction projection head; Apply a set of independent third linear projection transformation parameters to each copy sequence to project the copy sequence onto the value semantic space to obtain the value projection vector sequence corresponding to the semantic interaction projection head; Within each semantic interaction projection head, a semantic interaction weight matrix is ​​generated using the query projection vector sequence and the key projection vector sequence. Semantic interaction weight distribution data for that semantic interaction projection head is generated through the transpose interaction operation of the query projection vector sequence and the key projection vector sequence. The semantic interaction weight distribution data is applied to the value projection vector sequence and weighted aggregation is performed to generate a context-aware representation sequence that integrates cross-position semantic interaction information under the semantic interaction projection head. The context-aware representation sequence is globally averaged and aggregated, and the mean of the representation elements is calculated along the sequence length dimension to obtain the association strength feature signal corresponding to the semantic interaction projection head. All the association strength feature signals corresponding to the semantic interaction projection heads are spliced ​​along the head dimension to generate a multi-dimensional association strength spliced ​​signal. The multi-dimensional association strength spliced ​​signal is then linearly mapped and compressed to reduce its dimension to a single dimension, thereby generating the semantic unit association strength parameter components of the entity semantic unit pairing samples.

5. The method according to claim 1, characterized in that, Also includes: Construct a retrieval index structure for the semantic unit node registration pool. The retrieval index structure uses standardized strings in the text surface form of entity semantic units as the main retrieval key and entity type tags in the field of industrial economic operation as auxiliary retrieval keys. Receive candidate semantic unit nodes to be retrieved, obtain the entity semantic unit text surface form of the candidate semantic unit nodes, perform case-consistent conversion and whitespace character standardization on the entity semantic unit text surface form to obtain a standardized entity semantic unit text surface form. Obtain the industrial economic operation domain entity type tag of the candidate semantic unit node, and form a retrieval key tuple by combining the standardized entity semantic unit text surface form and the industrial economic operation domain entity type tag. The search index structure is used to perform precise matching search by the search key tuple to determine whether there is a registered semantic unit node that has both a standardized entity semantic unit text surface form and an entity type mark in the field of industrial economic operation that are completely consistent with the search key tuple. When an exact match retrieval returns an empty set, it is determined that there is no registered semantic unit node in the semantic unit node registration pool that matches the candidate semantic unit node. A new unique identifier for the candidate semantic unit node is generated, and the candidate semantic unit node is associated with the new unique identifier and stored in the semantic unit node registration pool. At the same time, the retrieval index structure is updated to include the new registered node. When the exact match retrieval returns a non-empty set, it is determined that there is a registered semantic unit node in the semantic unit node registration pool that matches the candidate semantic unit node, and the unique identifier of the existing semantic unit node is extracted from the returned registered semantic unit node record. Construct a coreference mapping record from the original entity semantic unit of the candidate semantic unit node in its respective economic operation event record to the unique identifier of the existing semantic unit node, and add the coreference mapping record to the coreference mapping set.

6. The method according to claim 1, characterized in that, Also includes: Extract the unique identifiers of the starting semantic unit node and the ending semantic unit node of the candidate internal semantic association edge to form a node pair identifier tuple. Using the node pair identifier tuple as the query condition, the semantic association edge registration pool is traversed and retrieved to obtain the set of registered semantic association edges in which the unique identifiers of the starting semantic unit node and the ending semantic unit node match the node pair identifier tuple. Detect whether the set of registered semantic association edges contains at least one registered semantic association edge; When the set of registered semantic association edges is empty, the semantic association direction label and the semantic unit association strength parameter component of the candidate internal semantic association edges are used as the initial edge weight attribute value of the semantic association edge to generate a new registered semantic association edge, and the new registered semantic association edge is written into the semantic association edge registration pool. When the set of registered semantic association edges is not empty, the registered semantic association edge with the earliest last updated edge weight is selected from the set of registered semantic association edges as the semantic association edge to be fused. Obtain the current edge weight attribute value of the semantic association edge to be fused, input the current edge weight attribute value and the semantic unit association strength parameter component of the internal semantic association edge candidate into a preset strength fusion function, the strength fusion function calculates the weighted harmonic average value of the current edge weight attribute value and the semantic unit association strength parameter component, and outputs the fused edge weight attribute value. The edge weight update operation of the semantically related edge to be merged is completed by overwriting the edge weight attribute value of the merged edge weight attribute value with the merged edge weight attribute value.

7. The method according to claim 1, characterized in that, The receiving of a target retrieval request, wherein the target retrieval request includes a query intent description text expressed in natural language text form, targeting changes in target economic operation indicators or related business events, including: A request receiving interface for industrial economic operation retrieval services is established. The request receiving interface continuously listens for retrieval request messages from industrial economic operation management terminal devices. When the request receiving interface intercepts a retrieval request message, it extracts the payload field content of the retrieval request message to obtain the original retrieval request string. The original search request string is validated for text validity. The original search request string is checked to see if it contains non-whitespace printable characters. If it contains non-whitespace printable characters, the original search request string is determined to be a valid request text. Otherwise, the search request message is rejected and an empty request prompt is returned. The original search request string that is determined to be a valid request text is normalized by unifying the format of full-width and half-width characters, removing leading and trailing whitespace characters, and merging consecutive whitespace characters to obtain the normalized search request text. The language attribute of the normalized retrieval request text is identified. When the language attribute indicates that it is the natural language used by the major industrial economic participants, the normalized retrieval request text is directly used as the query intent description text. Otherwise, the normalized retrieval request text is converted into the natural language used by the major industrial economic participants through the embedded language conversion processing unit before being used as the query intent description text. Determine the conceptual category of the target economic operation indicator or the event type category of the related business event referred to by the query intent description text, and generate a search intent type tag. The query intent description text, the search intent type tag, and the request receipt timestamp are assembled into a target search request, and the target search request stores the query intent description text, the search intent type tag, and the request receipt timestamp in the form of structured data units.

8. The method according to claim 7, characterized in that, The process of determining the conceptual category of the target economic operation indicator or the event type category of the related business event referred to by the query intent description text, and generating search intent type tags, includes: The query intent description text is input into the hierarchical intent understanding and processing component of the semantic big model. The basic semantic encoding layer of the hierarchical intent understanding and processing component performs semantic encoding on the query intent description text to generate a query intent semantic encoding sequence. The intent trigger word recognition layer of the hierarchical intent understanding processing component scans the semantic encoding sequence of the query intent by identifying the economic operation indicator mention fragments or event mention fragments that point to specific economic operation indicators or related business events in the query intent description text, and outputs a set of intent trigger word fragments. Each intent trigger word fragment in the intent trigger word fragment set is conceptually normalized, and the intent trigger word fragment is mapped to a standard concept node in a predefined industrial economic operation concept classification system to obtain a standard concept node identifier. Based on the hierarchical position and parent-child concept relationship of the standard concept node identifier in the industrial economic operation concept classification system, determine the economic operation indicator concept category or business event type category corresponding to the standard concept node identifier. When the intent trigger word fragment set contains multiple intent trigger word fragments, the economic operation indicator concept category or business event type category corresponding to each intent trigger word fragment is obtained to obtain a candidate intent type set. The dominant intent type is selected from the candidate intent type set through a preset voting priority rule. The dominant intent type is encoded into a fixed-format search intent type tag string, which includes the concept classification system version number and the standard concept node path description.

9. A computer system, characterized in that, include: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the industrial economic operation data retrieval method based on a semantic large model as described in any one of claims 1 to 8.

Images