Knowledge graph-based industrial industry development report labeling method and system

Abstract

The application provides an industrial industry development report annotation method and system based on a knowledge graph, semantic topology analysis is performed on initial text data sets of an industrial industry development report to be annotated to obtain context association paths and hierarchical distribution descriptions of industrial entity nodes in report paragraph units; a pre-constructed industrial knowledge graph is called to perform cyclic path mapping to generate knowledge path annotation records of the industrial entity nodes; the industrial development context trend descriptions and the spatiotemporal interweaving distribution descriptions of the knowledge units in the report text are determined; report annotation instructions containing knowledge unit positioning sequences are generated based on the industrial development context trend descriptions and the spatiotemporal interweaving distribution descriptions, and are sent to a target report processing terminal to start annotation synchronization operations. The application significantly improves the accuracy and annotation efficiency of industrial report knowledge unit positioning.

Description

Technical Field

[0001] This application relates to the field of text processing, and in particular to a method and system for annotating industrial development reports based on knowledge graphs. Background Technology

[0002] With the advancement of industrial digitalization, structured parsing of development reports containing industry descriptions and assigning knowledge labels to industry entities within the report can aid in industry trend analysis and knowledge management. Currently, most common report annotation methods involve acquiring the report text, performing entity recognition, and assigning corresponding category labels to the identified entities based on preset rule templates or keyword matching methods to complete the annotation of the report content. However, this annotation method only focuses on the identification and classification of the entities themselves, and the annotation results are independent of each other, making it difficult to reflect the evolution of the industry theme in the report and the inheritance and derivation relationships between industry knowledge. When faced with lengthy, cross-domain industrial development reports, the resulting annotation results are often fragmented and unable to trace the development trajectory, thus limiting the usability of the annotation information in industry analysis. Summary of the Invention

[0003] This invention provides a method and system for annotating industrial development reports based on knowledge graphs.

[0004] In a first aspect, embodiments of the present invention provide a method for annotating industrial development reports based on knowledge graphs, including: Obtain the initial text data set corresponding to the industrial development report to be annotated. The initial text data set contains multiple report paragraph units with time sequence identifiers that are recorded continuously. The report paragraph units are divided according to the natural paragraphs of the report and maintain the original record order in time sequence. Semantic topology parsing is performed on the initial text data set to obtain the contextual association path of the industry entity nodes in the report paragraph unit and the hierarchical distribution description of the knowledge structure in which the industry entity nodes are located. The contextual association path represents the semantic flow direction relationship and theme continuity relationship between industry entity nodes in adjacent paragraphs, and the hierarchical distribution description represents the hierarchical boundary integrity and hierarchical subordinate relationship of industry entity nodes in the knowledge structure. The pre-built industry knowledge graph is invoked to perform cyclic path mapping on the context association path and hierarchical distribution description, and generate knowledge path annotation records for industry entity nodes in the report paragraph unit. The knowledge path annotation records indicate the association path branches and node positioning identifiers of industry entity nodes in the knowledge graph. Based on the knowledge path annotation records, determine the industrial development path description corresponding to the industrial development report to be annotated and the spatiotemporal interweaving distribution description of knowledge units in the report text. The spatiotemporal interweaving distribution description includes the evolution path of the coverage area and the trend of distribution density change of knowledge units in the report paragraphs. Based on the description of the development trend and the spatiotemporal interweaving distribution, a report annotation instruction containing the knowledge unit positioning sequence is generated, and the report annotation instruction is sent to the target report processing terminal to start the annotation synchronization operation.

[0005] Secondly, embodiments of the present invention provide a computer system including a memory and a processor. The memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above method.

[0006] The present invention provides a knowledge graph-based method for annotating industrial development reports. After segmenting the report text into paragraphs, it extracts the semantic transmission trajectory of industrial entity nodes between adjacent paragraphs and their hierarchical distribution in the knowledge structure through semantic topological parsing. This ensures that the report annotation is based on both local contextual flow relationships and global knowledge hierarchy relationships. Based on this, the industrial knowledge graph is invoked to perform cyclical path mapping on the above two types of information. This establishes a correspondence between industrial entity nodes in the report and the associated paths and node locations in the knowledge graph. The generated knowledge path annotation records express the dual position of the entity in the report context and the knowledge system. Furthermore, based on the knowledge path annotation records, the industrial development trajectory corresponding to the entire report is reconstructed, and the evolution path and distribution density change trend of knowledge units in the report paragraphs are constructed, forming a spatiotemporally interwoven distribution description. This unifies the macro-characteristics of the industrial development stage with the dynamic distribution of micro-knowledge units. The final generated report annotation instructions carry the location sequence of knowledge units in the report, enabling the annotation operation to be automatically synchronized along the report's timeline. The resulting annotation results are faithful to the stage theme of industrial development and retain the knowledge hierarchy topology information, providing consistent technical support for the structured analysis of industrial development reports and the mining of industrial knowledge associations. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the specification, serve to explain the technical solutions of the present invention.

[0008] Figure 1 This is a schematic diagram illustrating the implementation process of a knowledge graph-based industrial development report annotation method provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the composition structure of a report annotation device provided in an embodiment of the present invention.

[0010] Figure 3 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] This invention provides a method for labeling industrial development reports based on knowledge graphs, which can be executed by a computer system processor.

[0013] Figure 1 This is a schematic diagram illustrating the implementation process of a knowledge graph-based industrial development report annotation method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Step S100: Obtain the initial text data set corresponding to the industrial development report to be annotated. The initial text data set contains multiple report paragraph units with time sequence identifiers that are continuously recorded. The report paragraph units are divided according to the natural paragraphs of the report and maintain the original record order in time sequence.

[0014] In this embodiment, the industrial development report to be annotated exists in the report management system as an electronic document. This electronic document consists of a series of natural paragraphs, and the order in which the paragraphs are written indicates the chronological order of industrial events, policy promulgation, technological breakthroughs, or market changes. Step S100 uses a file parsing adapter to read the electronic document as a binary stream and decodes the binary stream into a plain text character sequence according to the document encoding format. During the decoding process, the file parsing adapter synchronously monitors paragraph markers in the document. Paragraph markers include, for example, carriage return / line feed characters, paragraph indentation spaces, and combinations of line breaks appearing more than twice consecutively. Once a paragraph marker is detected, the file parsing adapter extracts the character sequence from the end of the previous paragraph marker to the beginning of the current paragraph marker into an independent report paragraph unit.

[0015] Each extracted report paragraph unit is encapsulated into a data structure containing three members: a paragraph number, a paragraph text string, and a time sequence identifier. The paragraph number is an integer starting from 1 and incrementing sequentially according to the order in which the report paragraph units appear in the document. The time sequence identifier is generated by assigning the time sequence identifier of the first report paragraph unit to a baseline start timestamp, which can be the internal time value of the report's publication date or the start date of the time interval covered by the report. The time sequence identifier of each subsequent report paragraph unit is increased by a fixed unit time increment based on the time sequence identifier of the previous report paragraph unit. This unit time increment is equal to the overall time span of the report divided by the total number of report paragraph units, thus creating an evenly spaced time sequence of time sequence identifiers for consecutive report paragraph units. This processing method explicitly transforms the originally implicit temporal relationships in the report into quantifiable temporal coordinates.

[0016] The initial text data set is stored as an ordered array, with array indices corresponding one-to-one with paragraph numbers. Each array element is the complete text string of that paragraph along with its associated time sequence identifier. During the generation process, the original text content was not deleted, rewritten, or rearranged in any way, thus fully preserving the original logical order between paragraphs and the sentence and word layout within each paragraph.

[0017] Step S200: Perform semantic topological parsing on the initial text data set to obtain the contextual association path of the industry entity nodes in the report paragraph unit and the hierarchical distribution description of the knowledge structure in which the industry entity nodes are located. The contextual association path represents the semantic flow direction relationship and topic continuity relationship between industry entity nodes in adjacent paragraphs, and the hierarchical distribution description represents the hierarchical boundary integrity and hierarchical subordinate relationship of industry entity nodes in the knowledge structure.

[0018] In an optional embodiment, step S200 may specifically include the following steps S210 to S260: Step S210: Decompose the initial text data set into paragraph structures. Based on the natural paragraph separators and semantic coherence constraints in the report, divide the initial text data set into multiple text paragraph slices with sequential relationships, and ensure that each text paragraph slice maintains a one-to-one correspondence with the paragraph sequence identifiers in the original report.

[0019] Step S210 requires checking the semantic coherence of each natural paragraph break to determine whether to merge or split the paragraph boundaries. While natural paragraph break markers provide an initial segmentation framework, in actual industry reports, a complete discussion of an industry event is often divided into two natural paragraphs. For example, the first paragraph describes the background of the event, and the second paragraph immediately follows by describing the specific impact of the event. The two paragraphs constitute the same event sentence group at the semantic level. If the segmentation is mechanically done according to natural paragraph breaks, events that should belong to the same context will be split into two text paragraph slices, resulting in information breaks during subsequent entity association capture.

[0020] When semantic coherence is constrained, for two adjacent report paragraph units, several words from the end of the preceding paragraph and several words from the beginning of the following paragraph are extracted. These words are then input into a pre-trained masking language model. This masking language model, based on a deep converter architecture, can predict the most likely word for the masked position based on the context. Specifically, the end of the preceding paragraph and the beginning of the following paragraph are concatenated into a temporary text fragment. The perplexity of the model is calculated when the end words are masked and when the complete text is retained. The difference between these two perplexity values ​​serves as a quantification of semantic cohesion. A smaller perplexity difference indicates a smoother and more natural connection between the end and the beginning, and higher semantic cohesion; a larger difference indicates lower semantic cohesion.

[0021] Furthermore, the cross-paragraph continuity of syntactic dependency chains is also incorporated into the semantic coherence constraints. The syntactic dependency parser constructs syntactic dependency trees for each of two adjacent report paragraph units, detecting whether there are referential chains originating from the noun phrase at the end of the previous paragraph that extend across the paragraph boundary to the next paragraph. For example, if the last sentence of the previous paragraph ends with "this technical route," and the subject of the first sentence of the next paragraph is omitted but the predicate immediately follows "this technical route," then cross-paragraph referencing is determined to exist. The ratio of the number of cross-paragraph syntactic dependency chains to the total number of syntactic chains in the two paragraphs is counted as the syntactic cross-paragraph ratio.

[0022] When the semantic cohesion value is higher than the preset high cohesion threshold and the syntactic segment crossing ratio also exceeds the preset ratio threshold, it indicates that the two report paragraph units are closely connected at both the semantic and syntactic levels. In step S210, these two report paragraph units are merged into a single text paragraph slice. Conversely, when the semantic cohesion value is lower than the preset low cohesion threshold and the syntactic segment crossing ratio is also lower than the preset ratio threshold, the two report paragraph units are maintained as independent text paragraph slices. In a report, the merging operation may occur between some adjacent paragraphs, resulting in a slightly smaller total number of text paragraph slices than the initial total number of report paragraph units. However, since each text paragraph slice no longer contains semantic breakpoints, the accuracy of subsequent entity recognition and association capture will be improved. The paragraph number and time sequence identifier of each text paragraph slice are taken from the former paragraph number and time sequence identifier of the two merged report paragraph units when the merging occurs.

[0023] Step S220: Perform entity recognition and association capture on the text paragraph slices. By recognizing the semantic tags carried by industry entity nodes in adjacent text paragraph slices, capture the semantic transmission trajectory of industry entity nodes in cross-paragraph context, and generate a context association path that includes a consistent description of semantic transmission direction and a description of topic continuity.

[0024] The entity recognition process is implemented based on a sequence labeling architecture. Its underlying encoder is a multi-layer bidirectional transformer network (MTN), which computes the latent state representation of each word in the input text segment within the overall context. This MTN consists of alternating stacked self-attention and feedforward sublayers, each followed by residual connections and layer normalization operations. In the self-attention sublayer, the query matrix, key matrix, and value matrix are all obtained by transforming the input of the current layer through three independent linear projection matrices. The multi-head mechanism of self-attention performs dot-product scaling attention computation on the query, key, and value in multiple parallel representation subspaces, allowing each word to simultaneously attend to other words at different positions within the text segment, capturing long-distance semantic dependencies. The effect of multi-layer stacking is that self-attention in shallow layers tends to capture collocational relationships between local phrases and neighboring words, while self-attention in deeper layers tends to capture discourse-level semantic coherence across sentences.

[0025] Following the top layer of the multilayer bidirectional converter network, a Conditional Random Field (CRF) decoding layer is connected. This decoding layer receives the hidden state vector of each term as the basis for the emission probability and constructs a global transition probability matrix for the entire label sequence on top of it. The element in the i-th row and j-th column of the transition probability matrix represents the probability score of transitioning from label i to label j. During the training phase, the CRF decoding layer maximizes the log-likelihood of the correct label sequence. During the decoding phase, it searches for the path with the highest total score among all possible label paths using the Viterbi algorithm as the final predicted label sequence. The label type set is subdivided into two main categories: boundary labels for industry entity nodes and type labels. Boundary labels indicate whether a term is an entity start term, an entity internal term, an entity end term, or a non-entity term. Type labels, based on the boundary labels, indicate the semantic label of the entity node, distinguishing whether the entity node belongs to an industry semantic category such as raw materials, components, equipment, system integration, technical indicators, or policies and regulations.

[0026] Entity association capture retrieves the identified industry entity node lists from two adjacent text segment slices. The entity list of the preceding text segment is denoted as the source entity list, and the entity list of the following segment is denoted as the target entity list. The capture process does not simply assume that every node in the source entity list corresponds to a node in the target entity list. Instead, it first constructs a cross-segment semantic transit candidate graph. The nodes in the graph represent all industry entity nodes in the source and target entity lists, and the edges represent potential semantic transit relationships. The edge establishment rule is as follows: For any node in the source entity list, calculate its pairwise semantic distance with all nodes in the target entity list. The semantic distance is based on the cosine dissimilarity between the average pooling vectors of the corresponding segments in the word sequence output by the multilayer bidirectional converter network for each node. If the cosine dissimilarity is less than a preset upper limit for semantic connection distance, a candidate transit edge is established between the two nodes. Simultaneously, the semantic label compatibility matrix is ​​queried. If the semantic label type of the source node and the semantic label type of the target node are marked as incompatible in the matrix, no candidate transit edge is established even if the semantic distance between the two nodes is very close. For example, the semantic tag "raw materials" may be compatible with the semantic tag "manufacturing equipment" because upstream raw materials often correspond to midstream manufacturing equipment; however, the direct jump between the semantic tag "raw materials" and the semantic tag "consumer end products" often lacks intermediate links in industrial logic, resulting in low compatibility.

[0027] After candidate transit edges are established, for nodes where multiple candidate edges intersect, a graph matching algorithm is used for disambiguation to remove spurious correspondences with conflicting directions or low transit strength. The entity node pairs that remain are the entity node pairs that have undergone semantic transit. For each entity node pair, subsequent steps will calculate descriptions of directional consistency and topic continuity, integrating them into a contextual association path.

[0028] In an optional embodiment, step S220 may specifically include the following steps S221 to S226: Step S221: Identify the semantic change region between adjacent text paragraph slices, extract the entity node pairs that have undergone semantic transmission by comparing the differences between the entity node sets in the preceding and following paragraphs, and generate a flow distribution description of the entity node pairs in the semantic space. The flow distribution description indicates the direction path and intensity contour of semantic transmission from the source paragraph to the target paragraph.

[0029] The comparison of the source entity set and the target entity set is performed simultaneously through set operations to complete three statistical metrics: the number of semantically continuing nodes, the number of semantically fading nodes, and the number of semantically newly generated nodes. These three metrics collectively characterize the semantic change region at the transition point between adjacent text paragraph slices. Locations with drastic semantic changes are typically characterized by a significant increase in both the number of semantically fading and semantically newly generated nodes, while the number of semantically continuing nodes decreases sharply, indicating a topic shift or industry theme inflection point in the report. Locations with moderate semantic changes are dominated by the number of semantically continuing nodes, with very few semantically fading and semantically newly generated nodes.

[0030] The directional path described in the flow distribution is not a simple arc, but a multi-dimensional directional vector obtained by projecting and decomposing several predefined industry semantic benchmark directions in the semantic embedding space. The industry semantic benchmark directions are determined by the types of relationships with clear directions in the industry knowledge graph. For example, the "supply" relationship can learn a directional representation in the semantic embedding space, pointing to a semantic shift tendency from supply to demand; the "technology iteration" relationship corresponds to a directional representation from older generation technology nodes to newer generation technology nodes. Projecting the semantic transfer vector onto these benchmark directions yields a set of projection weights, which, when combined, constitute the multi-dimensional depiction of the directional path. The intensity profile is obtained by weighted harmonic adjustment of the magnitude of the semantic transfer vector with the semantic saliency of the source and target nodes within their respective segments. Semantic saliency was obtained in step S221 by fusing word frequency-inverse paragraph frequency, syntactic depth, and modifier information. The intensity profile is the magnitude multiplied by the geometric mean of the source node saliency and the target node saliency. This ensures that even if the magnitude of the transmission vector is high, if the saliency of the source node or the target node is extremely low, the final intensity profile will not be excessively amplified, thus avoiding interference from occasional rare nodes on the transmission analysis.

[0031] Step S222: Determine the semantic migration direction offset angle and semantic migration intensity change of the industry entity node between adjacent text segment slices based on the flow distribution description. The semantic migration direction offset angle represents the angle by which the semantic transmission trajectory deviates from the original entity semantic center, and the semantic migration intensity change represents the magnitude of change in semantic salience during the transmission process.

[0032] The determination of the original entity semantic center relies not only on the semantic embedding vector of the industry entity node in its first occurrence paragraph, but also on the relative position of the node's corresponding node and its direct parent concept node in the industry knowledge graph. Specifically, each node in the industry knowledge graph is associated with a pre-trained graph embedding vector. The difference between the node's graph embedding vector and the graph embedding vector of its direct parent concept node is used as an offset calibration vector. The semantic embedding vector of the first occurrence paragraph is superimposed with this offset calibration vector, and the result is used as the calibrated original entity semantic center. The reason for introducing hierarchical offset calibration from the industry knowledge graph is that the semantic embedding of entity nodes in plain text paragraphs may be shifted due to temporary modifications by the local context of the paragraph, but their structural positioning in the industry knowledge system is relatively stable. By superimposing the offset calibration vector, the accidental semantic drift caused by the local context of the paragraph can be corrected, making the calculation of the semantic migration direction offset angle more robust.

[0033] The semantic transfer direction offset angle is calculated in high-dimensional space using the vector angle calculation method: take the original entity semantic center vector after calibration, take the direction vector of the semantic transfer vector in step S221, and calculate the angle between the two. The high-dimensional vector angle is obtained by dividing the inner product of the two vectors by the product of the magnitudes of the two vectors and then performing an inverse cosine operation.

[0034] While the change in semantic migration intensity is defined as the difference in semantic saliency between the source and target nodes, it is necessary to further evaluate the relationship between this difference and the overall semantic density change trend at paragraph transition locations. If the overall semantic density at paragraph transition locations shows a general upward trend—for example, when a report transitions from a macro overview to a technical detail paragraph, the semantic saliency of most entity nodes systematically increases—then the change in semantic migration intensity of a particular industry entity node at this location should be reduced by the overall average semantic saliency change at that paragraph transition location to isolate the interference of global trends and obtain the net change attributed solely to the semantic migration of that entity node itself.

[0035] Step S223: Construct a semantic trajectory continuity analysis process for industrial entity nodes based on the semantic migration direction offset angle and semantic migration intensity change. The semantic trajectory continuity analysis process verifies the smoothness of the semantic transmission trajectory based on the offset angle threshold and intensity change tolerance, and outputs a trajectory continuity evaluation indicator.

[0036] The offset angle threshold and intensity variation tolerance are not fixed constants, but dynamically adapted based on the semantic tag type of the industry entity node and the node's depth in the industry knowledge graph. For industry entity nodes with semantic tags of abstract policy or macro strategy, the angle of their cross-paragraph semantic discussion is often more divergent, and the allowed offset angle threshold is relatively lenient. For industry entity nodes with semantic tags of specific technical parameters or specific product models, the concentration of their cross-paragraph semantic discussion is higher, and the offset angle threshold is relatively strict. Similarly, the intensity variation tolerance for specific product nodes located deeper in the industry knowledge graph is stricter than that for industry category nodes located shallower, because once a specific product is mentioned between paragraphs, its semantic salience should fluctuate less; otherwise, it may indicate a jump in the report's logical argument or abnormal data citation.

[0037] The dynamic parameters of the offset angle threshold and intensity change tolerance are stored in a parameter table that is jointly indexed by semantic tag type and hierarchical depth. This parameter table is obtained through large-scale statistical analysis of historical industry report corpora. The statistical method is as follows: for entity node samples manually labeled as semantic trajectory continuous and trajectory broken in historical reports, the distribution of their semantic migration direction offset angle and semantic migration intensity change are statistically analyzed respectively, and the boundary value of the overlapping area of ​​the two types of sample distributions is taken as the recommended threshold value under the corresponding conditions.

[0038] Step S224: Perform continuity calculation on the semantic transmission trajectory of industry entity nodes in continuous text segment slices through the semantic trajectory continuity analysis process to generate a semantic transmission direction consistency description. The semantic transmission direction consistency description reflects whether the semantic flow direction of industry entity nodes remains consistent between multiple segments and whether the fluctuation range is within the allowable range.

[0039] In step S224, the semantic trajectory continuity analysis process performs an aggregation operation on the trajectory continuity evaluation indication sequence output in step S223. However, the aggregation operation is not simply a count of "continuous points," but rather a distinction between the continuity quality levels within each "continuous point." Even in continuous paragraph transitions, there are differences in the magnitude of the offset angle and intensity change: "continuous points" with offset angles close to the center of the allowable offset angle interval and intensity changes close to zero are marked as "high-quality continuous points"; those with offset angles close to the edge of the allowable interval but not yet exceeding the limit and intensity changes close to the upper limit of the allowable interval are marked as "low-quality continuous points." An additional dimension is added to the semantic transmission direction consistency description to record the proportion of high-quality continuous points. The higher the proportion of high-quality continuous points, the more coherent and stable the semantic flow direction; a high proportion of low-quality continuous points indicates that while the flow is coherent, the fluctuation amplitude is close to a critical state, suggesting that subsequent paragraphs may experience directional breaks. The proportion of continuous segments is calculated by dividing the total number of continuous points by the total number of transition positions. The maximum continuous segment span only counts the number of segments spanned by continuous segments composed of high-quality continuous points. The directional offset angle smoothing value is only calculated for the offset angle at high-quality continuous points.

[0040] Step S225: Perform theme periodicity pattern recognition on the semantic transmission trajectory of industrial entity nodes, transform the semantic transmission trajectory of each entity node in the frequency domain according to the paragraph span, extract the theme fluctuation cycle that recurs within the preset paragraph span interval, and calculate the amplitude stability of the theme fluctuation cycle as a description of theme continuity.

[0041] Frequency domain transformation processes the semantic saliency sequence of industry entity nodes in each text segment. Before feeding the semantic saliency sequence into the Fast Fourier Transform (FFT), it needs to be pre-processed to detrend, because semantic saliency may show a macro trend of overall increase or decrease as the report progresses from beginning to end. For example, in an industry development report, the later parts involve more technical details, and the average semantic saliency of entity nodes will generally increase. Detrending processing estimates the low-frequency trend components in the sequence and subtracts them from the original sequence using a locally weighted regression smoothing method, while the residual sequence retains the mid-to-high-frequency fluctuation components. Performing an FFT on the residual sequence yields a spectrum that more cleanly reflects the periodic fluctuations caused by the switching of topics between paragraphs, rather than overall trend drift.

[0042] In the frequency domain, the frequency range corresponding to the preset paragraph span interval is the result of mapping the total number of paragraphs in the actual report onto the frequency axis. Multiple span intervals are set, namely short-cycle intervals, medium-cycle intervals, and long-cycle intervals. For each interval, the frequency component with the most significant amplitude is extracted. This approach is because thematic fluctuations of an industry entity node may simultaneously exhibit periodicity across multiple time scales: short-cycle fluctuations might correspond to a writing style where the focus of the description changes every few paragraphs; medium-cycle fluctuations might correspond to a quarterly time rhythm for the report's progress; and long-cycle fluctuations might correspond to the macro-narrative arc of the entire report, from problem identification and analysis to conclusion. Thematic continuity descriptions are expanded from a single amplitude stability index to a triple including short-cycle amplitude stability, medium-cycle amplitude stability, and long-cycle amplitude stability. Each element is calculated from the amplitude variation coefficient of its respective cycle interval, comprehensively reflecting the thematic continuity performance of the industry entity node at different time scales.

[0043] Step S226: Combine and encapsulate the semantic transmission direction consistency description and the topic continuity description to obtain the context association path. The context association path uses entity nodes as indexes and paragraph order as axes to record the semantic transmission direction trajectory and topic continuity intensity map of each industry entity node in the full text of the report.

[0044] The combined encapsulation output format is a mapping table with industry entity node identifiers as keys. Each entry in the mapping table points to a two-dimensional matrix unfolded along the paragraph axis. The first dimension of the matrix is ​​the paragraph number, and the second dimension traverses the descriptive fields of the entity node in that paragraph, including whether the entity node exists in that paragraph, semantic tag, semantic saliency value, coherence assessment result with the previous transition position, offset angle, high-quality continuous point identifier, number of theme fluctuation cycles, and amplitude stability values ​​at different time scales. This two-dimensional matrix can also serve as input data for the visualization module, generating a semantic transmission direction trajectory map of entity nodes and a theme continuation intensity heatmap. The horizontal axis of the trajectory map is the paragraph timeline, and the vertical axis is the offset angle, with arrow segments connecting adjacent paragraph positions. The intensity of the arrow color indicates the absolute value of the intensity change. The horizontal axis of the heatmap is the paragraph timeline, and the vertical axis is the period at different time scales. The intensity of the color blocks indicates the magnitude of the theme fluctuation amplitude at the location near the paragraph at that scale.

[0045] Step S230: Perform knowledge hierarchy segmentation on the text segment slices. Based on the hierarchical relationship of industry entity nodes in the knowledge system, perform hierarchical classification of entity nodes within each text segment slice to obtain a hierarchical distribution description that includes the boundary integrity description of the segmentation hierarchy and the degree of subordinate association between the hierarchical levels.

[0046] In an optional embodiment, step S230 may specifically include the following steps S231 to S236: Step S231: Perform semantic boundary enhancement on the text paragraph slices. By extracting contextual modifiers around entity nodes in each paragraph slice, the semantic outline boundary information of entity nodes is strengthened, resulting in paragraph slices with enhanced semantic boundaries.

[0047] The extraction of contextual modifiers is not limited to a single sentence. When an industry entity node is located at the end of a sentence or in a clause connection position, its contextual modifiers may cross syntactic boundaries and appear in adjacent clauses. When constructing the syntactic dependency tree, the dependency parser also includes cross-sentence referential chains and parallel structures within the dependency arc's coverage. Specifically, for each industry entity node in a text segment, the dependency parser not only collects modifiers directly connected by arcs of attributive-head, adverbial-head, and noun-modification relationships, but also traces back along the referential chains. If a pronoun is found to refer back to the entity node and is modified by an adjective or determiner, the latent state vectors of these modifiers are included in the contextual modifier set.

[0048] Attention-weighted calculations are based on the cosine similarity between modifiers and industry entity nodes in the semantic space. However, to ensure that the reinforcement effect does not overemphasize the influence of modifiers with extreme similarity, a temperature parameter is introduced for scaling and sharpness control. When the temperature parameter is less than 1, similarity differences are amplified, and the weight distribution is more concentrated on the few modifiers with the highest similarity, suitable for situations where there are many contextual modifiers for entity nodes but their quality varies. When the temperature parameter is greater than 1, similarity differences are compressed, and the weight distribution is more even, suitable for situations where there are relatively few contextual modifiers. The adaptive selection of the temperature parameter depends on the average number of modifiers for entity nodes in the current text segment.

[0049] Step S232: Based on the semantic contour boundary information in the paragraph slices after semantic boundary enhancement, perform knowledge region growth segmentation on each text paragraph slice. Using the superior concept of the entity node as the seed point, grow and aggregate the subordinate entities outward according to the semantic inclusion relationship to obtain multiple non-overlapping knowledge sub-region units. Each knowledge sub-region unit represents an independent knowledge level group.

[0050] The seed point selection strategy of the knowledge region growing segmentation algorithm is as follows: For a text segment, the direct superior concepts of all industry entity nodes in the industry knowledge system are taken. These direct superior concepts are sorted in descending order of frequency of occurrence. The direct superior concept node with the highest frequency that is not covered by other seed points at higher levels is selected as the first seed point. After selecting the first seed point, the direct superior concepts that are already covered by the industry entity nodes associated with the seed point are excluded. Among the direct superior concepts of the remaining uncovered industry entity nodes, the one with the highest frequency is selected as the second seed point, and so on, until all direct superior concepts of all industry entity nodes in the text segment have been assigned, or the number of seed points reaches a preset upper limit.

[0051] During the region growth process, the semantic inclusion strength is not solely based on cosine similarity and path hop count. An additional dimension is added: external co-occurrence statistics between two nodes in the industry knowledge system, based on high-frequency co-occurrence in the same policy provisions, technical standards, or market research reports. These external co-occurrence statistics are pre-extracted from a large-scale industry text corpus and stored as edge weights. These weighted sums, along with cosine similarity and path hop count, constitute the semantic inclusion strength. When calculating path hop count-related terms, the reciprocal of the path hop count is not directly used as a monotonically decaying factor. Instead, exponential decay is employed, with the decay rate factor adjusted according to the granularity of different branches of the industry knowledge system: in finer-grained technical branches, the semantic distance between adjacent higher and lower-level nodes is smaller, resulting in a correspondingly smaller decay rate factor; in coarse-grained branches, the semantic gap between adjacent levels is larger, leading to a larger decay rate factor. This adaptive decay ensures a relatively balanced growth and aggregation effect across different knowledge branches, preventing lower-level entities from struggling to cross the inclusion threshold due to overly dense hierarchical divisions in a particular branch, and also preventing irrelevant nodes from being easily absorbed into the same knowledge sub-region unit due to sparse hierarchical divisions in a particular branch.

[0052] Step S233: Extract semantic concentration distribution information within each knowledge sub-region unit. By statistically analyzing the number of entity nodes within the region and the semantic correlation between them, generate a semantic concentration distribution description at the region level. The semantic concentration distribution description reflects the degree of aggregation of knowledge topics within the knowledge sub-region unit.

[0053] The normalization benchmark for the semantic concentration distribution description is not limited to the maximum value of each knowledge sub-region unit within the current text paragraph slice, but also references the statistical distribution of the regional semantic concentration of the same superordinate concept node in the historical reports of similar paragraph slices within the industry knowledge system. Assuming the mean and standard deviation of the regional semantic concentration of the current superordinate concept node in the historical corpus are known, the semantic concentration distribution description, in addition to providing the normalized value, also includes a percentile ranking of that value within the historical distribution of the same superordinate concept. If the semantic concentration distribution description of a certain knowledge sub-region unit is in an extremely high percentile in the historical distribution, it indicates that the paragraph's exposition of the superordinate concept is exceptionally dense and concentrated, possibly representing a focused thematic discussion; if it is in an extremely low percentile, it indicates that the paragraph's mention of the superordinate concept is scattered and does not form a thematic cluster.

[0054] Step S234: Calculate the deviation of the knowledge sub-region unit in the knowledge dimension based on the semantic concentration distribution description, compare the semantic concentration in the region with the preset hierarchical semantic benchmark concentration, and obtain the dispersion description that measures the clarity of the hierarchical boundary. The larger the dispersion description, the more blurred the boundary between the knowledge sub-region unit and the adjacent level.

[0055] The hierarchical semantic baseline concentration index table is indexed not only by hierarchical depth but also by the industry sector category of the parent concept node. This is because, at the same hierarchical depth, the natural semantic clustering density of entity nodes in different industry sectors may vary significantly in the report text. For example, at an industry hierarchical depth of 2, the number of entity nodes under the parent concept node of "electronic information" is usually large, resulting in a relatively high semantic concentration baseline; while the number of entity nodes under the parent concept node of "rare metal materials" may be smaller, resulting in a relatively low semantic concentration baseline. Using a uniform hierarchical depth baseline concentration across the entire industry would lead to a systematic overestimation or underestimation of the deviation in certain industry sectors.

[0056] The calculation of knowledge dimension deviation, based on the absolute value of relative deviation, further considers the sign of the deviation direction. When the semantic concentration of a knowledge sub-region unit is higher than the hierarchical semantic baseline concentration, the deviation is positive, meaning that entity nodes within that region are excessively densely packed in the same hierarchical space, potentially indicating that nodes that should be split into finer-grained levels are forcibly merged. When the semantic concentration is significantly lower than the baseline concentration, the deviation is negative, meaning that information within that region is sparse, possibly due to segmentation strategies that have assigned nodes belonging to that level to other hierarchical groups. The sign information of the deviation is preserved and passed to the discreteness description, enabling the discreteness description to reflect not only the degree of boundary ambiguity but also the skewness trend of the ambiguity direction.

[0057] Step S235: Determine the boundary integrity description of the knowledge sub-region unit based on the discreteness description. The boundary integrity description is configured to be inversely correlated with the discreteness description, and the boundary integrity description decreases when the discreteness description increases.

[0058] The specific form of the inverse mapping function is a monotonically decreasing nonlinear mapping. This nonlinearity causes the boundary integrity description to change gradually when the discreteness description is at a low level, and then accelerates its decline as the discreteness description crosses the medium range and enters a higher level. The rationale for this nonlinear mapping is that when the hierarchical boundaries are relatively clear, small fluctuations in the boundaries have little impact on boundary integrity; however, when the boundaries become blurred to a certain extent, further blurring will drastically increase the damage to the effectiveness of hierarchical segmentation. When the granularity of the higher-level concepts in the industry knowledge system is inconsistent, the numerical value of the boundary integrity description can be calibrated by multiplying it with the granularity factor of the higher-level concept nodes. This allows coarse-grained nodes, due to their naturally large hierarchical span, to have a relatively high tolerance for boundary integrity; while fine-grained nodes, due to the need for more precise distinctions between levels, have more stringent requirements for boundary integrity.

[0059] Step S236: Analyze the relative positional relationships between different knowledge sub-region units, and generate a hierarchical subordinate association degree descriptor containing a distance distribution descriptor and a direction angle description. The distance distribution descriptor indicates the interval between different knowledge sub-region units in the semantic space, and the direction angle description indicates the subordinate direction deviation between levels.

[0060] The construction of the hierarchical dependency descriptor does not simply involve calculating the Euclidean distance and orientation angle between each pair of knowledge sub-region units; instead, it involves constructing a hierarchical relationship matrix. The rows and columns of this matrix correspond to different knowledge sub-region units, and the matrix elements store the scalar values ​​of the joint encoding of the distance and orientation angle between the unit in that row and the unit in that column. The joint encoding method involves standardizing the distance distribution descriptor and the orientation angle descriptor according to their respective distribution scales, and then summing them by weight according to preset relative importance weights.

[0061] In an industry knowledge system, the direction of hierarchical subordination is not only from lower-level nodes to upper-level nodes, but also involves the semantic convergence degree of the lower-level nodes towards the upper-level nodes. When the parent concept nodes corresponding to two knowledge sub-regions have a direct parent-child relationship in the industry knowledge system, the angle of orientation is calculated with reference to the directed path direction of these two parent concept nodes in the industry knowledge system. When two knowledge sub-regions are separated by multiple hops in the knowledge system, the angle of orientation is referenced to the weighted average direction of the direction vectors at each step of the shortest path. In the weighted average, the closer the number of hops to the current level, the higher the weight is given, because the subordination direction of directly adjacent levels is more critical for judging the subordination degree between the current paragraph segmentation levels, while the influence of the far-end path direction on local subordination is relatively weakened.

[0062] Step S240: The context association path and hierarchical distribution description input semantic alignment processing are bridged. By constructing a unified semantic mapping framework across the description space, the semantic transmission trajectory of the context association path and the hierarchical subordinate information of the hierarchical distribution description are aligned to the same semantic representation plane, resulting in a set of text semantic coupling descriptions with a unified representation form.

[0063] The unified semantic mapping framework employs two encoding towers to process heterogeneous structured descriptive data. The encoding tower for context-related paths, when processing paragraph sequences of variable temporal length, uses its internal self-attention pooling layer to perform attention-weighted pooling of features at each paragraph position within the path description structure along the paragraph time axis. The attention weights are calculated based on a combination of the semantic saliency value of the entity node at each paragraph position and the coherence evaluation result, ensuring that paragraphs occupying more semantically important positions on the full-text time axis contribute more to the final embedding vector. The encoding tower for hierarchical distribution description processes the spatial distribution information of knowledge sub-region units. Its self-attention pooling layer performs cross-regional weighted aggregation of the boundary integrity descriptions and subordinate association descriptors of multiple knowledge sub-region units within the same text paragraph slice. The weights are determined based on the semantic concentration distribution description of each knowledge sub-region unit within the paragraph.

[0064] After the two encoding towers complete their respective embedding vector outputs, the supervision mechanism for cross-description space alignment is not only effective during the training phase. During the inference application phase, the unified semantic mapping framework also includes a cross-tower embedding consistency verification module. This module calculates the cosine similarity in real time between the context-related path embedding vectors and hierarchical distribution description embedding vectors belonging to the same industry entity node. If the similarity is lower than a preset consistency threshold, it indicates a semantic conflict or inconsistency between the two types of descriptions for that entity node, triggering a conflict flag. This flag is passed along with the textual semantic coupling description to subsequent steps S250 and S300 for reference during saliency evaluation and cyclic path mapping, thereby reducing the weight contribution of nodes with conflicting descriptions during fusion.

[0065] Step S250: Based on the saliency measurement results of different description dimensions in the text semantic coupling description set, extract the path saliency description corresponding to the context association path and the hierarchical saliency description corresponding to the hierarchical distribution description respectively. The path saliency description reflects the contribution of the semantic transmission trajectory between adjacent paragraphs, and the hierarchical saliency description reflects the distinguishing power of hierarchical division in the knowledge structure.

[0066] The fully connected neural network in the saliency evaluator is structured with three hidden layers. The first hidden layer maps the concatenated vector to the intermediate representation dimension, using a modified linear unit with leakage as the activation function. The second hidden layer splits the intermediate representation into two parallel branches, corresponding to the contextual path dimension and the hierarchical distribution description dimension, respectively, with each branch having multiple layers of perceptual processing. The third hidden layer compresses the outputs of the two branches into two scalars. The attention weighting module, which introduces competitive normalization, uses flexible maximum normalization to transform the two scalars. The normalization process introduces a learnable temperature parameter, which determines the sharpness of the saliency allocation: the lower the temperature, the more extreme the saliency allocation, with the dominant saliency dimension receiving almost all the weights and the weaker dimension receiving weights close to zero; the higher the temperature, the more balanced the allocation. The temperature parameter is learned end-to-end by a fully connected neural network. Different industry entity nodes may learn different temperature values. For fluctuating nodes where semantic transmission trajectories are clearly dominant, the learned temperature is relatively small. For structural nodes where hierarchical relationships are clearly dominant, the learned temperature is also relatively small, but it may be reflected in the dominance of scalar in another dimension. For nodes where the two types of information are balanced and complementary, the learned temperature is relatively large, maintaining the dual-path emphasis of information.

[0067] Step S260: Based on the path saliency description and the hierarchical saliency description, adjust the fusion weight configuration of context-related paths and hierarchical distribution descriptions in the cyclic path mapping process to obtain the adjusted fusion weight configuration. The fusion weight configuration is used to coordinate the influence of semantic transmission trajectories and hierarchical subordination relationships in the process of industry knowledge graph mapping.

[0068] Adjusting the weighting of data integration is not simply a matter of comparing global averages, but rather a fine-grained adjustment based on different development stages within the report. Industry development reports typically have a clear phased structure, such as an introductory environmental scan, a middle section of in-depth analysis, and a concluding summary and outlook. The relative importance of path-specific and hierarchical descriptions may change across different stages: in the environmental scan phase, the report mainly lists the dynamics of various industry sectors, and the cross-segment semantic transmission of entity nodes is relatively active, so the contribution of path-specific descriptions may increase at this stage; in the in-depth analysis phase, the report focuses on the vertical analysis of specific technologies or markets, and the differentiating effect of hierarchical distribution may become dominant at this stage.

[0069] When step S260 is executed, it first uses the industrial development path description, which will be fully formed after step S400, to perform a pre-segmentation, roughly dividing the full-text report paragraphs into several segments according to development stages. Within each segment, the average value of the segment with significant path description and the average value of the segment with significant hierarchical description are calculated, and a segment-level fusion weight configuration is dynamically generated for that segment based on its respective segment average value. The segment-level fusion weight configurations of each segment constitute a piecewise function that varies along the report's time axis, rather than a single configuration pair that is constant throughout the entire text. When the segmented fusion weight configuration is passed to step S300, the path fusion layer will query the corresponding fusion weight configuration segment value based on the segment where the entity node is located when processing industrial entity nodes at different paragraph positions.

[0070] Step S300: Call the pre-built industry knowledge graph to perform cyclic path mapping on the context association path and hierarchical distribution description, and generate knowledge path annotation records for industry entity nodes in the report paragraph unit. The knowledge path annotation records indicate the association path branches and node positioning identifiers of industry entity nodes in the knowledge graph.

[0071] The node embedding vectors in the industry knowledge graph are not statically fixed, but support lightweight dynamic fine-tuning based on the input context. Specifically, the pre-trained graph embedding vector of each node is stored as a basic vector. When a node is queried, its basic vector is subjected to context bias correction based on a gating mechanism according to the fused path description vector from the input. The corrected node vector better reflects the semantic characteristics of the current report segment without destroying the original graph topological constraints. The gating bias is calculated by a lightweight gating unit. The input is the concatenation of the fused path description vector and the basic node embedding vector, and the output is a bias vector and a gating coefficient vector. The gating coefficient controls the degree of injection of the bias vector, preventing the context bias from overly covering the basic topological features.

[0072] In an optional embodiment, step S300 may specifically include the following steps S310 to S350: Step S310: Input the context-related path and hierarchical distribution description into the path fusion layer of the industry knowledge graph, configure the association modeling of the semantic transmission trajectory description of the context-related path and the hierarchical subordinate information of the hierarchical distribution description according to the fusion weight, and generate a fusion path description vector with semantic topological consistency constraints.

[0073] In the residual multilayer perceptron architecture, the first fully connected layer projects the input features into a high-dimensional latent space, the second fully connected layer performs a nonlinear transformation in this latent space, and the third fully connected layer compresses the high-dimensional representation back to the same dimension as the node embedding vector. The residual connection directly sums the input of the first fully connected layer to the output of the third fully connected layer, ensuring that the feature transformation unit retains the path of the linear identity mapping even when learning complex nonlinear mappings, which contributes to gradient propagation and training stability.

[0074] After the gated fusion unit performs element-wise weighted mixing of the semantic propagation trajectory vector and the hierarchical subordinate vector, it further adjusts the results through semantic topological consistency constraints. The method for maximizing the lower bound of mutual information is as follows: a discriminator network is constructed, which alternately receives positive and negative sample pairs. Positive sample pairs consist of the current fusion path description vector and the embedding vector of the correct corresponding node in the industry knowledge graph for that industry entity node. Negative sample pairs consist of the current fusion path description vector and the embedding vectors of other randomly selected nodes. The discriminator network outputs a score representing the probability that the input pair belongs to a positive sample. The estimation of the lower bound of mutual information is based on the contrastive loss function output by the discriminator. Optimizing this contrastive loss function indirectly maximizes the lower bound of mutual information. The semantic topological consistency constraint is superimposed on the main objective of the gated fusion unit as an auxiliary objective. Its gradient is propagated back to the gated fusion unit and the two feature transformation units through backpropagation, adjusting their parameters to ensure that the produced fusion path description vector aligns as closely as possible with the correct nodes and avoids heterogeneous nodes in the representation space of the industry knowledge graph.

[0075] Step S320: Through multi-scale perception processing of the industry knowledge graph, the fusion path description vector is extracted in a coordinated manner with local detail semantics and global context information. The fusion path description vector is split into multiple representation subspaces along the depth direction. The direct and indirect association patterns of industry entity nodes in the knowledge graph are extracted at different granularities to obtain a multi-scale path perception description set.

[0076] In an optional embodiment, step S320 may specifically include the following steps S321 to S326: Step S321: Input the fused path description vector into the path segmentation layer of the multi-scale perception processing, and uniformly divide the representation channels of the fused path description vector according to the preset path segmentation ratio to generate a dual-path parallel semantic stream containing local representation branches and global representation branches. The local representation branches focus on extracting neighborhood association details of entity nodes, while the global representation branches focus on extracting the macro-context of entity nodes in the full report.

[0077] The path segmentation ratio is selected based on the total number of representation channels in the fused path description vector and the average node degree of the industry knowledge graph. When the industry knowledge graph is dense and the average node degree is high, the neighborhood information of entity nodes is rich and complex, and more representation channels should be allocated to local representation branches to provide sufficient representation capacity to encode first-order and multi-order neighborhood structures. Conversely, when the graph is sparse, the relative importance of global contextual information increases, and the channel quota of global representation branches increases. The optimal setting of the path segmentation ratio parameter is automatically searched during the model training phase through structure search or gradient-based channel importance measurement.

[0078] Step S322: In the local representation branch, multiple context ranges with different expansion amplitudes are used to extract the local representation branch and perform multi-level semantic receptive field expansion. Local detail semantic description fragments at different spatial scales are extracted from the local representation branch. The channel capacity of the local detail semantic description fragments is aligned with the channel capacity of the global representation branch.

[0079] The specific structure of the dilated path convolutional kernel is a one-dimensional dilated convolutional kernel that slides along the representation channel direction. Different dilation magnitudes are achieved by inserting gaps between the weight elements of the convolutional kernel. For example, a dilation magnitude of 2 means inserting one gap between adjacent weight elements. In this case, the actual physical span of the convolutional kernel on the representation channel is expanded, and the receptive field is multiplied. Multiple parallel dilated path convolutional paths are each set with an increasing dilation magnitude factor, forming a pyramid structure with a progressively expanding receptive field. At the same level, it synchronously captures structural patterns ranging from direct association details of one-hop neighbors of entity nodes to two-hop or even multi-hop community-scale patterns.

[0080] The local detail semantic description fragments output by each dilation path convolutional path maintain the same channel capacity as the global representation branch. This consistency is achieved by setting the number of output channels of the dilation path convolutional layer to be equal to the number of channels of the global representation branch, ensuring that the subsequent semantic fusion process is executed under dimension matching.

[0081] Step S323: In the global representation branch, global statistical information is condensed. The global representation branch is compressed into a condensed context description with global summary information through cross-channel mean aggregation operation. The condensed context description is then reshaped in terms of spatial dimension to obtain a global context semantic description that matches the size of the local detail semantic description fragment.

[0082] The cross-channel mean aggregation operation is performed independently along the representation channel dimension: for each channel in the global representation branch, all element values ​​within that channel are summed and divided by the total number of elements in that channel to obtain a scalar value. The scalar values ​​of all channels are arranged in channel order to form a one-dimensional condensed context description. This condensed context description captures the average activation level of the fusion path description vector across all channels of the global representation branch. The average activation level reflects the overall response strength and multi-channel response pattern of the entity node within the macro-context of the full report.

[0083] The linear projection layer maps the one-dimensional condensed context description to a two-dimensional representation with the same spatial length and number of channels as the local detail semantic description fragment. The linear projection layer contains a fully connected weight matrix; the input dimension of the matrix is ​​the length of the condensed context description, and the output dimension is the product of the spatial length and the number of channels of the local detail semantic description fragment. After matrix multiplication and shape transformation, the global context semantic description obtains a size structure that perfectly matches the local detail semantic description fragment.

[0084] Step S324: Input the local detail semantic description fragment and the global context semantic description into the semantic fusion processing process. Generate the fusion allocation ratio of each channel based on the channel attention mechanism. Perform cross-scale information synthesis on the local detail semantic description fragment and the global context semantic description through the fusion allocation ratio to generate a multi-level fusion semantic description map. Each level of the multi-level fusion semantic description map corresponds to a local and global fusion result at a receptive field scale.

[0085] In the bottleneck structure, the first fully connected layer performs dimensionality reduction, compressing the number of channel descriptors to a reduced ratio of the original number of channels. This reduction ratio is typically set so that the number of channels at the bottleneck is significantly lower than the input and output channels, forcing the network to learn the most fundamental inter-channel dependencies. The second fully connected layer performs dimensionality increase, restoring the compressed channel descriptors to the original number of channels. The activation function between the two layers introduces a non-linear expression, enabling the bottleneck structure to learn not only linear correlations between channels but also capture higher-order non-linear interactions. The output of the bottleneck structure, after being processed by a normalized exponential function, yields a fusion allocation ratio vector of length equal to the number of channels, with each element ranging from 0 to 1.

[0086] During the element-weighted summation process, for a given channel, let the fusion allocation ratio for that channel be p. Then, the fused channel value is the element value of the local detail semantic description fragment in that channel multiplied by p, plus the element value of the global context semantic description in the same channel multiplied by (1-p). The weighted summation operation is performed element-wise at each spatial location and in each channel, ultimately yielding a fused semantic description map of the same size as the input local detail semantic description fragment.

[0087] Step S325: Perform channel normalization on the multi-level fusion semantic description mapping, adjust the fusion mappings of different scales to a unified numerical range, and splice the normalized fusion mappings along the channel dimension to generate a spliced ​​semantic description set with a unified representation dimension.

[0088] Layer normalization is performed independently on the fused semantic description map at each scale. For each map, the mean and variance estimates of all its elements are calculated, with the estimation range covering the set of elements across all channels and spatial locations of that map. Subtracting the mean from each element of the map and then dividing by the square root of the sum of the variance and a small constant achieves numerical translation and scaling. After layer normalization, maps at different scales exhibit comparable statistical properties in their numerical distributions.

[0089] The concatenation operation sequentially concatenates the three-level normalized fused semantic description maps along the channel dimension. Assuming that each of the three-level mappings has C channels, the total number of channels in the concatenated semantic description set is 3×C. The spatial dimension length has been kept the same in the three-level mappings, so the concatenated set remains aligned in the spatial dimension.

[0090] Step S326: Perform channel semantic reorganization on the spliced ​​semantic description set, calculate the mutual information between the representation vectors of each channel, cluster and merge the channels based on the mutual information, and reconstruct the semantic expression of each channel according to the cluster center vector to obtain the reorganized multi-scale path-aware description set.

[0091] When calculating mutual information, each pair of channel representation vectors forms a sample sequence based on the element values ​​of that channel at all spatial locations. The joint probability distribution and individual marginal probability distributions of the two sample sequences are estimated using a histogram. Then, the relative entropy between the product of the joint distribution and the marginal distributions is calculated according to the definition of mutual information. The greater the mutual information, the more information the two channel representation vectors share, and the greater the likelihood of redundancy or the same semantic dimension being captured holistically.

[0092] Hierarchical clustering algorithms, based on the mutual information matrix, iteratively merge the two channel clusters with the highest mutual information into a new cluster. The mutual information between the new cluster and the remaining clusters is updated using the average inter-cluster mutual information. The clustering process terminates when the mutual information between the two most recent clusters falls below a preset clustering stopping threshold. The clustering stopping threshold is set based on the compression ratio between the desired final number of channels and the original number of channels.

[0093] For each cluster, the element-wise average of the channel representation vectors within the cluster in the spatial dimension is taken as the cluster center vector. This cluster center vector is used as an independent channel after recombination, replacing all channels in the original cluster. The recombined multi-scale path-aware description set is more compact in the channel dimension. Each recombined channel is a condensed summary of multiple highly overlapping channels, resulting in a more refined expressive power.

[0094] Step S330: The semantic dependency model of the multi-scale path-aware description set is processed by the semantic attention mechanism of the industry knowledge graph to generate an attention distribution description map on the semantic dimension. The attention distribution description map indicates the distribution of the importance of descriptions at different scales in the knowledge path mapping.

[0095] In an optional embodiment, step S330 may specifically include the following steps S331 to S336: Step S331: Input the multi-scale path-aware description set into the query construction layer processed by the semantic attention mechanism, and generate a corresponding query description vector for each scale path-aware description in the multi-scale path-aware description set. The query description vector is used to represent the focus of attention of the path-aware description at that scale in the knowledge path mapping.

[0096] In the query construction layer, the size of the linear mapping parameter matrix corresponding to each scale's path-aware description is the number of channels in that scale's path-aware description multiplied by a preset query vector dimension. Matrix multiplication flattens the spatial dimension of the path-aware description to one dimension and multiplies it with the parameter matrix to obtain the query description vector. The dimension of the query description vector output at each scale is unified to the query vector dimension, unaffected by differences in the number of channels at the original scale. This ensures that subsequent dot product operations between the query and the key are performed under the premise of consistent dimension.

[0097] Step S332: Simultaneously input the multi-scale path-aware description set into the key-value construction layer processed by the semantic attention mechanism, perform linear mapping on the path-aware description at each scale, and generate key description vector and value description vector. The key description vector is used to calculate the matching degree with the query description vector, and the value description vector is used to carry the original semantic content of the path-aware description at that scale.

[0098] In the key-value construction layer, the dimensions of the key mapping parameter matrix and the value mapping parameter matrix are set to be consistent with the query mapping parameter matrix, ensuring that the dimensions of the key description vectors and the value description vectors are equal. The key matrix stacks key description vectors of all scales row-wise, and the value matrix stacks value description vectors of all scales row-wise. This dimensional alignment between the key and query matrices facilitates the subsequent computation of the dot product between each query-key pair.

[0099] Step S333: Calculate the semantic fit between each query description vector and all key description vectors. By performing a dot product operation on the query description vector and the key description vector and adjusting the scaling, the original attention distribution description map is generated. The original attention distribution description map represents the relevance strength of path-aware descriptions at each scale from the current query perspective.

[0100] The dot product operation is performed as follows: Take a query description vector and multiply it by the transpose of the key matrix to obtain a row vector. Each element of this row vector is the dot product of the query description vector and the corresponding column key description vector. This operation is repeated for all query description vectors. The matrix formed by stacking all the row vectors is the original attention distribution description map. The scaling factor used for scaling is the square root of the key vector dimension. This square root is derived from the variance stabilization analysis of the dot product operation under the random variable assumption, which avoids excessively large dot product values ​​of high-dimensional vectors entering the activation function saturation region.

[0101] Step S334: Normalize the original attention distribution description map by converting the correlation strengths of each element in the original attention distribution description map into a fixed total distribution ratio, thus obtaining a normalized attention distribution description map. The normalized attention distribution description map represents the attention distribution across all scales.

[0102] Normalization applies a normalization exponential function independently to each row, mapping all elements of that row to a probability distribution of positive numbers that sums to 1. After normalization, each element represents the proportion of attention allocated to the path-aware description of the corresponding scale in that column from the perspective of the query in that row, intuitively showing the relative attention level of each scale under the current query.

[0103] Step S335: Perform smoothing adjustment on the normalized attention distribution description map to suppress abnormally sharp peaks in the attention distribution map, so that the attention allocation has continuity between adjacent scales, and obtain a smoothed attention distribution description map.

[0104] The smoothing adjustment operates by applying a one-dimensional smoothing filter to the scale dimension of the normalized attention distribution descriptor. The filter's kernel function has a symmetrically decreasing weight distribution, with the largest weight at the center and decreasing towards the sides. The window width of the smoothing filter controls the smoothing strength; a wider window results in a stronger smoothing effect and a more gradual flattening of the attention distribution. After smoothing adjustment, the attention distribution is ensured to be perceived as continuous across scale gradients, preventing the model from over-focusing on a single scale and neglecting complementary information at other scales.

[0105] Step S336: Extract statistical information from the smoothed attention distribution description map, including the scale identifier corresponding to the attention distribution peak and the description of the dispersion of the attention distribution, and output the smoothed attention distribution description map and statistical information.

[0106] The scale identifier of the attention distribution peak is determined by finding the column index corresponding to the element with the largest attention allocation value in each row, which maps back to the corresponding scale level. The dispersion of the attention distribution is described by the ratio of the information entropy of the attention allocation vector in that row to the entropy of a uniform distribution in the same dimension. The closer the ratio is to 1, the more uniform and dispersed the attention distribution; the smaller the ratio, the more concentrated the attention distribution is at a few scales. Statistical information provides interpretable references for subsequent path mapping, allowing annotators to retrospectively view which receptive field scale an entity node primarily relied on in the knowledge path mapping.

[0107] Step S340: Assign and aggregate the multi-scale path-aware description set according to the attention distribution description map, and synthesize the multi-scale path-aware descriptions proportionally according to the allocation ratio corresponding to each scale in the attention distribution description map to obtain the semantically enhanced path vector that strengthens semantic relevance.

[0108] During the allocation and aggregation process, for each scale in the multi-scale path-aware description set, the allocation ratio of its corresponding column in the smoothed attention distribution description map is obtained. This allocation ratio is a vector corresponding to the query point. First, the allocation ratios of all query points need to be pooled and summarized, typically using average pooling or query point saliency-weighted pooling, merging them into a unified scale-level weight scalar. The path-aware descriptions at that scale are then weighted by scalar multiplication using this scale-level weight. Finally, all weighted scale-level path-aware descriptions are summed element-wise to obtain the semantically enhanced path vector. The semantically enhanced path vector inherits the local and global fusion information from multiple scales and strengthens the scale feature representation highly correlated with knowledge path mapping through an attention mechanism.

[0109] Step S350: Call the path classification output layer of the industry knowledge graph to perform knowledge path mapping on the semantically enhanced path vector. By evaluating the matching degree between the semantically enhanced path vector and the pre-stored path template in the knowledge graph, a knowledge path annotation record containing a knowledge path association strength distribution description is generated. The knowledge path association strength distribution description is used to represent the association strength distribution corresponding to the development trends of different industries.

[0110] Each path template in the path classification output layer's internal path template library records not only the node sequence but also the type sequence of relationships along each edge. This type sequence is crucial for distinguishing seemingly similar but logically different industrial development paths. For example, both paths lead from the "lithium battery" node to the "new energy vehicle" node, but the "supply" relationship represents the upstream and downstream direction of the industrial chain, while the "technology support" relationship represents the direction of collaborative technological evolution. These two paths have drastically different industrial implications. The path template's embedding vector is obtained by interleaving and fusing the embedding vectors of the node sequence and the relationship type sequence, then passing it through a sequence encoder. The sequence encoder employs a unidirectional long short-term memory network or a converter encoder structure, capable of capturing the sequential order of path nodes and the progressive logic of relationships.

[0111] During matching evaluation, cosine similarity is calculated between the semantically enhanced path vector and the embedding vector of each path template to obtain an initial matching score corresponding to each path template. Then, all initial matching scores are input into a ranking network. The ranking network imposes global constraints on the initial matching scores based on the prior industry logic compatibility between path templates; for example, two contradictory development paths should not simultaneously have high matching scores. The ranking network outputs calibrated matching scores and forms a knowledge path association strength distribution description. This distribution description indicates the development trends of multiple possible industry development paths and their respective normalized association strengths. The knowledge path association strength distribution descriptions of all industry entity nodes in each report paragraph unit are summarized into knowledge path annotation records.

[0112] Step S400: Based on the knowledge path annotation records, determine the industrial development trajectory description corresponding to the industrial development report to be annotated, as well as the spatiotemporal interweaving distribution description of knowledge units in the report text. The spatiotemporal interweaving distribution description includes the evolution path of the coverage area of ​​knowledge units in the report paragraphs and the trend of distribution density changes.

[0113] After completing the path mapping of each industry entity node in the industry knowledge graph in step S300, step S400 performs two levels of aggregation and derivation analysis: First, it aggregates vertically along the report's timeline, connecting industry entity nodes distributed across different paragraphs but belonging to the same industry development path branch to construct a complete description of the industry development trajectory. Second, it tracks and models the occurrence position and spatiotemporal migration of each knowledge unit in each paragraph of the report, generating a spatiotemporal interweaving distribution description reflecting the evolutionary patterns of knowledge units in the report text. A knowledge unit is a more granular conceptual unit than an industry entity node. A knowledge unit refers to an indivisible information particle in the report text that carries complete industry semantics. At the instance level, it can be a specific instance of an industry entity node appearing in a particular context, or a composite semantic configuration formed by multiple co-occurring industry entity nodes.

[0114] In an optional embodiment, step S400 may specifically include the following steps S410 to S460: Step S410: Parse the knowledge path association strength distribution description in the knowledge path annotation record, detect the path branches whose association strength exceeds the preset strength value in the knowledge path association strength distribution description and their corresponding timestamp information, and the timestamp information corresponds to the time sequence identifier of the report paragraph unit.

[0115] The knowledge path association strength distribution is described as a probability distribution. For each industry entity node at each report paragraph unit position, the distribution value represents the confidence strength of the node belonging to different industry development path branches at that position. Step S410 first performs threshold screening on this distribution, and path branches exceeding the preset strength value are marked as valid activated branches. The timestamp information is directly taken from the time sequence identifier assigned to the report paragraph unit in step S100. Because the same industry entity node may be validly activated on the same path branch in multiple paragraphs, step S410 also needs to continuously merge these repeatedly activated timestamp information. If the same path branch in adjacent paragraphs is continuously activated, these timestamps are merged into a continuous time activation interval to reduce fragmentation during subsequent development stage segmentation.

[0116] Step S420: Based on the timestamp information corresponding to the detected path branches, and combined with the continuous change trend of the correlation strength of the path branches in the preceding and following report paragraph units, determine the start and end time points of each development stage in the industrial development context, and calculate the time span metric between the start and end time points as a description of the duration of the corresponding development stage.

[0117] The segmentation of development stages is not only based on the presence or absence of path branches, but more importantly, on observing the continuous changing trend of the association strength of these branches. Step S420 obtains the association strength time series for each path branch at each paragraph position in the entire text. A change point detection algorithm is applied to this series to detect time points where the association strength significantly increases or decreases. The change point detection algorithm is based on cumulative sums and statistics. By continuously accumulating the deviation between the sequence value and the sequence mean, once the cumulative deviation exceeds the upper or lower limit of a preset decision interval, that position is marked as a change point. An increasing change point is considered the starting time point of the development stage, and a decreasing change point is considered the ending time point of the development stage. The development stage time span metric is equal to the absolute value of the paragraph time series value at the ending time point minus the paragraph time series value at the starting time point, expressed in paragraph time increments.

[0118] Step S430: Extract the set of industry entity nodes in the corresponding paragraph unit of the path branch, analyze the hierarchical position of the set of industry entity nodes in the knowledge graph, determine the combination relationship of core industry nodes and peripheral industry nodes involved in each development stage in the industrial development path, and generate a description of the direction of industrial development.

[0119] After extracting the set of industry entity nodes, each node in the set is queried for its hierarchical depth and network centrality index in the industry knowledge graph. Network centrality indexes include degree centrality, proximity centrality, and betweenness centrality, which respectively measure the number of direct connections to the node in the graph, the reciprocal of the average shortest path length to all other nodes, and the frequency of being on the shortest path. Nodes with shallow hierarchical depth and high levels of all three centrality indices are identified as core industry nodes. Core industry nodes dominate this development stage and are the axis of the industry narrative. Nodes with deeper hierarchical depth or lower centrality indices are identified as peripheral industry nodes, which play a supporting or subordinate role in this development stage. The description of the industry development trajectory is framed by a stage sequence, with each stage listing its core industry node list, peripheral industry node list, and a topological summary of the knowledge graph relationships between nodes.

[0120] Step S440: For each knowledge unit, based on the location information of its industry entity node in the knowledge path annotation record at different time periods, construct the initial paragraph position of the knowledge unit in the report text and the displacement vector in subsequent paragraphs to obtain the spatiotemporal movement trajectory of the knowledge unit.

[0121] In an optional embodiment, step S440 may specifically include the following steps S441 to S446: Step S441: Extract the location identifiers of the industry entity nodes associated with the target knowledge unit in each report paragraph unit from the knowledge path annotation record to obtain a location sequence arranged in chronological order. The location identifiers include the paragraph number and the node index in that paragraph.

[0122] The paragraph number is the paragraph number of the report paragraph unit, and the node index is the position number of the industry entity node in the paragraph according to its order of appearance. If an industry entity node appears multiple times in the same paragraph, generating multiple node indices, the median of these node indices is taken as the composite node index within that paragraph to represent the average position of the node in that paragraph.

[0123] Step S442: Based on the positioning sequence, determine the initial appearance paragraph position of the target knowledge unit, and take the initial appearance paragraph position as the trajectory starting point. Extract the paragraph number and node index from the positioning identifier corresponding to the trajectory starting point as the starting point coordinates.

[0124] The initial paragraph position is the earliest position in the location sequence in chronological order, and the location identifier content is not empty. The starting coordinates are represented by two-dimensional coordinates, with the first coordinate being the paragraph number and the second coordinate being the node index.

[0125] Step S443: For each subsequent position point in the positioning sequence, excluding the starting point coordinates, calculate the displacement vector of that position point relative to the previous position point. The displacement vector includes the direction indicator of the crossing between paragraphs and the change in the node index. The direction indicator indicates the migration pattern of the knowledge unit between paragraphs: moving forward, referencing backward, or remaining unchanged.

[0126] The rules for determining the direction indicator are as follows: If the paragraph number of the current position is greater than the paragraph number of the previous position, the direction indicator is marked as "move forward"; if the paragraph number of the current position is less than the paragraph number of the previous position, the direction indicator is marked as "backreference"; if the paragraph numbers are equal but the node index changes, the direction indicator is marked as "drift within paragraph"; if neither the paragraph number nor the node index changes, the direction indicator is marked as "unchanged". The change in the node index is the difference between the node index of the current position and the node index of the previous position.

[0127] Step S444: Generate the spatiotemporal movement trajectory of the knowledge unit based on the displacement vector. The spatiotemporal movement trajectory is constructed with the starting point coordinates as the origin, the paragraph number as the horizontal axis, and the relative position of the node index within the paragraph as the vertical axis, to construct a trajectory curve describing the movement path of the knowledge unit in the report text.

[0128] The spatiotemporal movement trajectory is constructed by plotting each location point as a scatter plot in a two-dimensional coordinate system with the segment number as the horizontal axis and the node index as the vertical axis, in chronological order. Then, adjacent location points are connected sequentially by straight line segments in chronological order to form a polyline trajectory. If the direction between two adjacent points is marked as "backward reference", the connecting line is drawn as a dashed line to distinguish it from the solid line connecting "forward movement", reflecting an abnormal migration of backward reference in the trajectory.

[0129] Step S445: Smooth the spatiotemporal movement trajectory by connecting displacement vectors to construct a continuous trajectory and removing jagged edges caused by fluctuations in the node sequence within the paragraph, thus obtaining the smoothed spatiotemporal movement trajectory of the knowledge unit.

[0130] The trajectory smoothing method employs a locally weighted regression approach. At each trajectory point, a window is drawn centered on that point along the horizontal axis. A weighted polynomial fit is then applied to the original data points within this window, and the fitted value replaces the original point's ordinate value. The width of the window is adaptively set proportionally to the total number of report paragraphs to accommodate differences in data density across reports of varying lengths. After smoothing, small oscillations in node indexes caused by referencing the same entity node at different locations within the same paragraph are effectively eliminated, while preserving long-range migration trends across paragraphs.

[0131] Step S446: Associate the smoothed spatiotemporal movement trajectory of the knowledge unit with the hierarchical information of the industry entity nodes associated with the knowledge unit in the knowledge graph, so that the spatiotemporal movement trajectory carries semantic information of hierarchical shifting up or down, and generates a spatiotemporal movement trajectory of the knowledge unit with hierarchical labeling.

[0132] The association of hierarchical information is achieved by querying the hierarchical depth of the industry entity node associated with that knowledge unit in the knowledge graph for each time point on the smoothed trajectory, and binding the hierarchical depth value to the attribute of that trajectory point. The hierarchical depth of adjacent time points is compared point by point. If the hierarchical depth value of a later point is less than that of an earlier point, it is marked as "hierarchical upward movement," indicating that the knowledge unit is generalizing towards a higher-level concept; if the hierarchical depth value of a later point is greater than that of an earlier point, it is marked as "hierarchical downward movement," indicating that the knowledge unit is refining towards a lower-level concept; if the hierarchical depth value remains unchanged, it is marked as "hierarchical translation." The spatiotemporal movement trajectory of knowledge units with hierarchical markings provides a three-dimensional semantic description of the horizontal axis position, vertical axis position, and hierarchical migration direction for each trajectory point.

[0133] Step S450: Perform density change analysis on the spatiotemporal movement trajectory of knowledge units, statistically analyze the frequency changes of knowledge units appearing in consecutive paragraphs and the increasing or decreasing trend of the number of paragraphs covered, and generate a spatiotemporal intertwined distribution description of knowledge units in the report text. The spatiotemporal intertwined distribution description includes a description of the distribution density changes of knowledge units and the directional trend of expansion or contraction of the coverage area.

[0134] In an optional embodiment, step S450 may specifically include the following steps S451 to S456: Step S451: Obtain the spatiotemporal movement trajectory of the knowledge unit with hierarchical tags, and the existence tag of the knowledge unit in each report paragraph unit.

[0135] The existence flag is a sequence of Boolean values ​​with a length equal to the total number of report paragraph units in the entire text. A value of true at the i-th position in the sequence indicates that the knowledge unit is located and appears in the i-th report paragraph unit, while a value of false indicates that it does not appear.

[0136] Step S452: Scan along the paragraph sequence in a sliding window manner, count the cumulative number of occurrences of the target knowledge unit within each sliding window, and obtain the occurrence frequency sequence. The width of the sliding window is adaptively set according to the total number of reported paragraphs.

[0137] The width of the sliding window is set to a preset percentage of the total number of report paragraphs, rounded down, and the minimum window width is not less than the minimum window size to accommodate extremely short reports. The window slides along the paragraph sequence in steps of one paragraph each. The number of paragraphs with a true presence marker within each sliding window is counted and recorded as the frequency of occurrence of that window. The frequency of occurrence of all windows is arranged in the order of the paragraphs whose centers are located in the windows to form the frequency sequence.

[0138] Step S453: Calculate the change in frequency between adjacent windows based on the frequency sequence to obtain a frequency change description. The frequency change description reflects the density increase or decrease trend of knowledge units at different reporting positions. A positive change in frequency indicates an increase in density, while a negative change indicates a decrease in density.

[0139] The change in frequency between adjacent windows is calculated by subtracting the frequency of the previous window from the frequency of the subsequent window. A sign analysis is performed on the frequency change sequence: consecutive positive values ​​indicate that the knowledge unit has entered a period of increasing density, while consecutive negative values ​​indicate a period of decreasing density.

[0140] Step S454: Combining the spatiotemporal movement trajectory of knowledge units with hierarchical markings, determine the direction of hierarchical movement of knowledge units during density changes. When density increases, identify whether the hierarchy is cohesive upwards or generalized downwards. When density decreases, identify whether the hierarchy is abstracted upwards or subdivided downwards. Generate a correlation description between density changes and hierarchical linkage.

[0141] The logic for the linkage analysis between density and hierarchy is as follows: In the density-increasing interval, observe the comparison of the number of hierarchical upward movement markers and hierarchical downward movement markers among the trajectory points marked with hierarchical markers within that interval. If hierarchical upward movement markers are in the majority, it indicates that while density is increasing, the hierarchy is converging upwards, and knowledge units are converging from multiple scattered subordinate concepts to a higher-level core concept; if hierarchical downward movement markers are in the majority, it indicates that while density is increasing, the hierarchy is generalizing downwards, and higher-level concepts are being refined and supported by a large number of subordinate instances. The same marker statistical comparison is performed in the density-decreasing interval. If hierarchical upward movement is in the majority, it indicates that density is decreasing along with hierarchical upward abstraction, and knowledge units are regressing from specific instances to upstream concepts; if hierarchical downward movement is in the majority, it indicates that density is decreasing along with hierarchical downward subdivision, and knowledge units no longer appear as holistic concepts but are scattered fragmented in detailed paragraphs. The correlation description records these four linkage patterns corresponding to each density change interval.

[0142] Step S455: Calculate the total number of paragraphs covered by the knowledge unit and its distribution span throughout the report, starting from the first paragraph and ending at the last paragraph, to determine the trend of changes in the coverage of the knowledge unit, including a description of the rate of expansion or contraction of coverage.

[0143] The total number of covered paragraphs is obtained by counting the positions that are true in the existence marker sequence. The distribution span is the span of paragraph indices between the first and last true position in the existence marker sequence. The rate of expansion of the coverage trend is defined as the ratio of the distribution span to the total number of paragraph slices in the whole text. Furthermore, the ratio of the total number of covered paragraphs in the first half and the second half of the report can be calculated. If the ratio is greater than 1, it indicates that the knowledge units are concentrated in the first half of the report, and if it is less than 1, it indicates that they are concentrated in the second half of the report.

[0144] Step S456: Combine the frequency change description, density change and hierarchical linkage description, and coverage change trend into a spatiotemporal intertwined distribution description of knowledge units in the report text. The spatiotemporal intertwined distribution description provides a complete picture of the intertwined evolution of knowledge units in time and space dimensions.

[0145] The final output format of the spatiotemporal interwoven distribution description is a multidimensional dictionary. The key is the knowledge unit identifier, and the value is a record structure containing fields such as the frequency sequence of occurrence, continuous density increase or decrease intervals and their corresponding hierarchical linkage patterns, the total number of covered paragraphs, the distribution span, and the half-area coverage ratio. This dictionary can be directly queried and called by downstream report annotation instruction generation steps and report visualization modules.

[0146] Step S460: Establish a stage-related mapping between the description of the development trend and the spatiotemporal intertwined distribution description, associate and bind each development stage with the spatiotemporal movement trajectory segment of the knowledge unit within the corresponding time period, and write the binding relationship into the description of the development trend and the spatiotemporal intertwined distribution description respectively.

[0147] Stage association mapping is achieved through time alignment: the start and end timestamps of each development stage in the description of the industrial development trajectory are aligned with the time axis of the spatiotemporal movement trajectory of the knowledge unit under the same time coordinate reference system. Spatiotemporal movement trajectory segments within the same start and end time interval are assigned to the corresponding trajectory segments of that development stage. Once the binding relationship is established, each stage entry in the description of the industrial development trajectory adds a reference field, pointing to the list of active knowledge units and the spatiotemporal interweaving distribution description index within that stage; each knowledge unit entry in the spatiotemporal interweaving distribution description adds a stage affiliation field, indicating which segments of the knowledge unit's spatiotemporal movement trajectory belong to which development stage.

[0148] Step S500: Based on the description of the development trend and the spatiotemporal interweaving distribution, generate a report annotation instruction containing the knowledge unit positioning sequence, and send the report annotation instruction to the target report processing terminal to start the annotation synchronization operation.

[0149] The report annotation instructions are the machine-readable instruction set that is finally sent to the report processing terminal to perform the annotation work. Each instruction in the instruction set precisely specifies the annotation content, annotation position, annotation intensity, and annotation stage classification information, so that the target report processing terminal does not need to perform semantic understanding or judgment. It only needs to insert annotation labels at the corresponding positions in the report and apply the corresponding visual or structured effects according to the coordinate parameters in the instructions.

[0150] In an optional embodiment, step S500 may specifically include the following steps S510 to S560: Step S510: Analyze the description of the development trend, extract the development stage sequence, each stage includes a set of core industry nodes and a set of peripheral industry nodes, as well as the start and end timestamps of the stage, and construct the temporal skeleton sequence of the development stage.

[0151] The temporal skeleton sequence is a linear list in which elements are arranged chronologically according to development stages. Each element records the name or index of that development stage, the start timestamp, the end timestamp, a list of core industry nodes, and a list of peripheral industry nodes. The criteria for distinguishing between core and peripheral industry nodes follow the judgment results from step S430.

[0152] Step S520: Based on the temporal skeleton sequence of the industrial development stage, query the active knowledge unit set in each development stage from the spatiotemporal intertwined distribution description, and sort the queried knowledge unit set according to its occurrence time and distribution density in the stage to generate the knowledge unit positioning sequence in the stage.

[0153] The query condition for an active set of knowledge units is that the spatiotemporal movement trajectory of the knowledge unit intersects with the start and end timestamps of the current development stage. The longer the intersection, the higher the activity level of the knowledge unit within this stage. During sorting, a coarse sort is first performed based on the order of the first appearance of each knowledge unit within the stage. For knowledge units with the same first appearance time, a fine sort is performed based on the mean of their frequency sequence within the stage, from high to low, so that knowledge units with higher distribution density are ranked higher in the sequence.

[0154] Step S530: For each knowledge unit in the knowledge unit location sequence, combine the coverage change trend and hierarchical linkage description in the spatiotemporal intertwined distribution description to mark the specific location coordinates and hierarchical depth of the knowledge unit in the report paragraph. The location coordinates are defined by the paragraph number and the offset within the paragraph.

[0155] The specific location coordinates are determined by retrieving the trajectory points of the knowledge unit's spatiotemporal movement trajectory within the target paragraph. When calculating the offset within a paragraph, the node index of the knowledge unit within that paragraph is obtained. This node index is then divided by the total number of industry entity nodes within the paragraph, normalized to a relative position between 0 and 1, which represents the offset within the paragraph. The hierarchical depth label is directly obtained from the spatiotemporal movement trajectory of the knowledge unit with hierarchical labels, showing the point's hierarchical depth value at that time.

[0156] Step S540: Based on the knowledge unit location sequence labeled with location coordinates and hierarchical depth markers, construct a spatiotemporal layout diagram of knowledge units for report annotation. Normalize the report paragraph timeline and knowledge hierarchical depth to the same numerical range. Arrange them in two dimensions with the normalized report paragraph timeline as the horizontal axis and the normalized knowledge hierarchical depth as the vertical axis. Draw each knowledge unit as a layout node according to its normalized location coordinates, while retaining its original report paragraph location coordinates and hierarchical depth markers.

[0157] The report paragraph timeline normalization maps paragraph numbers to the interval 0 to 1, with the first paragraph mapped to 0 and the last paragraph to 1. Knowledge level depth normalization takes the global minimum and maximum values ​​of the level depth values ​​of all knowledge units obtained in step S530, and linearly maps the level depth of each knowledge unit to the interval 0 to 1. Each knowledge unit occupies a node in the layout diagram. The horizontal axis of the node represents its normalized paragraph position, and the vertical axis represents its normalized level depth. The data entity of the node in the diagram still retains its original report paragraph position coordinates and original level depth value.

[0158] Step S550: By using the set of core industry nodes in the description of the development path of the industry, the spatiotemporal layout diagram of knowledge units is semantically anchored and corrected. Knowledge units that deviate from the semantic range of the core industry nodes of the stage are finely adjusted in position so that the layout of knowledge units is aligned with the stage theme of the development path of the industry.

[0159] In an optional embodiment, step S550 may specifically include the following steps S551 to S556: Step S551: Extract the set of core industry nodes for each stage of development from the description of the development path. The set of core industry nodes for each stage is the representative node of the theme of that stage and has the highest semantic centrality.

[0160] In addition to the comprehensive score of the three network centrality indicators in step S430, the selection of the core industry node set for each stage also includes a ranking filter based on the path association strength value of the node in the current stage of the industry development trend description. Nodes with higher rankings and higher comprehensive network centrality scores are selected.

[0161] Step S552: For each knowledge unit in the spatiotemporal layout diagram of knowledge units, calculate the semantic distance between it and the set of core industry nodes of the development stage to which it belongs. The semantic distance is comprehensively evaluated based on the shortest path length and relation weight between nodes in the industry knowledge graph.

[0162] The semantic distance is calculated as follows: In the industry knowledge graph, the industry entity nodes associated with the knowledge unit are taken as starting nodes, and each core node in the core industry node set of the stage is taken as the target node. A bidirectional breadth-first search is used to obtain the shortest paths from all starting nodes to all target nodes. The number of hops for each shortest path and the sum of the inverses of the relation weights of each edge on the path are recorded. The value with the smallest total weight distance among all shortest paths is taken as the semantic distance between the knowledge unit and the core industry node set of the stage.

[0163] Step S553: ​​Preset a semantic distance tolerance threshold, and mark knowledge units whose semantic distance exceeds the semantic distance tolerance threshold as deviating units. The semantic distance tolerance threshold is dynamically determined based on the overall dispersion of knowledge units within the development stage.

[0164] The quantification method for the overall dispersion of knowledge units within a development stage is to calculate the median and median absolute deviation of the semantic distance between all knowledge units within that stage and the core industry node set of that stage. The semantic distance tolerance threshold is set as the median plus a multiple of the median absolute deviation. This multiple is a preset coefficient that can be adjusted according to the report type to control the labeling sensitivity of deviation units.

[0165] Step S554: For the deviated unit, based on the semantic attraction relationship between the deviated unit and the core industry node set of the stage, make a fine-tuning of its position along the normalized horizontal and vertical axes in the normalized knowledge unit spatiotemporal layout diagram, and move it to the normalized coordinate area corresponding to the core industry node with the closest semantics. The adjustment range is proportional to the mapping value of semantic distance in the normalized space. After fine-tuning, the adjustment amount is simultaneously mapped back to the original coordinate system, and the original report paragraph position coordinates and hierarchical depth markers of the knowledge unit are updated.

[0166] The direction vector for position fine-tuning points from the current normalized coordinates of the offset unit to the normalized coordinates of the knowledge unit associated with the core industry node that is semantically closest to it in the layout diagram. The adjustment is performed along this direction vector, and the movement distance is proportional to the portion of the semantic distance that exceeds the tolerance threshold. After fine-tuning, the coordinate changes in the normalized space are transformed back to the original paragraph number coordinates and the original level depth value according to the inverse function of the normalization mapping, thereby updating the original position coordinates and level depth markers recorded in step S530 for that knowledge unit.

[0167] Step S555: After position fine-tuning, recalculate the spatial overlap between the knowledge unit and surrounding knowledge units in the normalized spatiotemporal layout diagram of the knowledge unit. If an overlap exceeding the preset overlap tolerance limit occurs, adjust the position diffusion of the surrounding knowledge units and map the adjustment result back to the original coordinate system to update the original coordinate information of the affected knowledge units until the spatial distribution of knowledge unit nodes in the layout diagram meets the preset separation conditions.

[0168] The degree of spatial overlap is measured by whether the Euclidean distance between two knowledge unit nodes in the normalized layout diagram is less than the sum of their radii. If overlap occurs, the repulsive force model from the force-directed layout algorithm is used to apply a repulsive displacement along the line connecting the overlapping knowledge unit pairs. The displacement is proportional to the overlap depth. After the repulsive displacement calculation is completed, the moved knowledge units are simultaneously mapped back to their original coordinate system for coordinate updates. Overlap detection and repulsive force adjustment are iteratively performed until the overlap depth between all node pairs is within a preset tolerance limit, achieving the spatial distribution separation condition.

[0169] Step S556: Output the corrected spatiotemporal layout diagram of knowledge units, wherein each knowledge unit node retains the normalized layout coordinates and the original report paragraph position coordinates and hierarchical depth markers, and transmit the corrected spatiotemporal layout diagram of knowledge units to the report annotation instruction generation step.

[0170] In the corrected spatiotemporal layout diagram of knowledge units, each node has a position representation under two coordinate systems. The normalized layout coordinates are used for visualization and global spatial relationship analysis, while the original coordinates and hierarchical depth markers are used for precise positioning and labeling of the downlink report processing terminal.

[0171] Step S560: Based on the corrected spatiotemporal layout diagram of knowledge units, generate a report annotation instruction containing the knowledge unit positioning sequence. Each annotation entry in the report annotation instruction is associated with a knowledge unit and its corresponding paragraph position coordinates and hierarchical depth marker in the report, and carries a stage affiliation tag, which is used to drive the target report processing terminal to perform annotation synchronization operation.

[0172] In an optional embodiment, step S560 may specifically include the following steps S561 to S566: Step S561: Traverse all layout nodes in the corrected knowledge unit spatiotemporal layout diagram, extract the knowledge unit identifier, the original paragraph position coordinates updated after correction, the original level depth marker, and the stage affiliation label corresponding to each layout node, and form a label entry.

[0173] The traversal order follows the spatial order of the layout nodes in the layout diagram from left to right and from top to bottom, so that the generated annotation entries have a spatially continuous presentation logic.

[0174] Step S562: Sort the annotation entries according to the order of the original paragraph position coordinates after correction and update to obtain a knowledge unit positioning sequence that is consistent with the paragraph order in the report. Each element in the knowledge unit positioning sequence indicates an annotation operation.

[0175] The sorting key is first the paragraph number, with smaller paragraph numbers listed first; paragraphs with the same number are sorted by their intra-paragraph offset from smallest to largest, ensuring that the annotation entries strictly follow the reading order of the report text.

[0176] Step S563: For each annotation operation in the knowledge unit positioning sequence, assign an annotation intensity level according to the original hierarchical depth mark. The annotation intensity level is inversely proportional to the hierarchical depth, and the deeper the hierarchical depth, the higher the annotation intensity.

[0177] The annotation intensity levels are categorized in descending order of the hierarchy depth. Superordinate concept nodes at shallower hierarchy depths have lower annotation intensity and may be annotated with light highlights or light backgrounds. Conversely, specific nodes at deeper hierarchy depths have higher annotation intensity and are annotated with bright highlights, bolding, or enclosed annotations. The intensity level mapping table predefines the correspondence between hierarchy depth ranges and annotation intensity levels, mapping hierarchy depth values ​​to several preset discrete intensity levels.

[0178] Step S564: Group the knowledge unit location sequence with assigned annotation intensity levels according to the stage affiliation label, and generate a grouped annotation instruction sequence based on the development stage. Each group of annotation instruction sequences corresponds to a development stage segment interval in the report.

[0179] Each group of annotation instruction sequences is encapsulated independently after grouping, and a group separator is set between each group. The separator records the start timestamp of the development stage corresponding to each group. The target report processing terminal can locate the correct report paragraph interval based on the start timestamp and start executing the annotation of that group.

[0180] Step S565: Add stage summary information to each set of annotation instruction sequences. The stage summary information includes the names of the core industry nodes, time range, and an overview of the number of knowledge units in that development stage, to obtain a complete report annotation instruction.

[0181] The phase summary information is embedded in the header of each group of annotation instruction sequences in the form of a text summary, serving as the metadata block of that group. The names of the core industry nodes in the metadata block are given in comma-separated list format, the time range is indicated by the order of start time and end time, and the knowledge unit quantity overview gives the total number of annotation entries in the group and the distribution of entries according to the annotation intensity level.

[0182] Step S566: Send the complete report annotation instruction to the target report processing terminal through the communication interface, so as to drive the target report processing terminal to perform the knowledge unit annotation synchronization operation at the corresponding report paragraph position according to the knowledge unit positioning sequence, annotation intensity level and stage affiliation label in the annotation instruction.

[0183] The communication interface employs a secure network transmission protocol. Report annotation instructions are serialized into structured data format and then compressed and encrypted before being sent. Upon receiving the annotation instructions, the target report processing terminal parses the data, reads each group sequentially, locates the corresponding report paragraph interval based on the group summary information, and then reads the annotation operations one by one according to the knowledge unit location sequence. Knowledge unit annotation anchor points are placed at the offset within the paragraph, a visual style corresponding to the annotation intensity level is applied, and a stage affixing label is added. The annotation synchronization operation is differentially overlaid and stored with the original report version, forming a final annotation report where the annotation layer is separated from the original content. This allows the annotation results to be displayed by stage and switched by intensity level, completing the entire intelligent annotation process for industrial development reports.

[0184] Based on the foregoing embodiments, this invention provides a report annotation device. The units and modules included in the device can be implemented by a processor in a computer device; of course, they can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0185] Figure 2 This is a schematic diagram of the composition structure of a report annotation device provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the report labeling device 200 includes: The text acquisition module 210 is used to acquire the initial text data set corresponding to the industrial development report to be labeled. The initial text data set contains multiple report paragraph units with time sequence identifiers that are recorded continuously. The report paragraph units are divided according to the natural paragraphs of the report and maintain the original record order in time sequence. The semantic parsing module 220 is used to perform semantic topological parsing on the initial text data set to obtain the contextual association path of the industry entity nodes in the report paragraph unit and the hierarchical distribution description of the knowledge structure in which the industry entity nodes are located. The contextual association path represents the semantic flow direction relationship and topic continuity relationship between industry entity nodes in adjacent paragraphs, and the hierarchical distribution description represents the hierarchical boundary integrity and hierarchical subordinate relationship of industry entity nodes in the knowledge structure. The path annotation module 230 is used to call the pre-built industry knowledge graph to perform cyclic path mapping on the context association path and hierarchical distribution description, and generate knowledge path annotation records of industry entity nodes in the report paragraph unit. The knowledge path annotation records indicate the association path branches and node positioning identifiers of industry entity nodes in the knowledge graph. The trend determination module 240 is used to determine the industrial development path description and the spatiotemporal interweaving distribution description of knowledge units in the report text, based on the knowledge path annotation records. The spatiotemporal interweaving distribution description includes the evolution path of the coverage area and the trend of distribution density change of knowledge units in the report paragraphs. The annotation synchronization module 250 is used to generate report annotation instructions containing knowledge unit positioning sequences based on the description of the development trend and the spatiotemporal interweaving distribution, and send the report annotation instructions to the target report processing terminal to start the annotation synchronization operation.

[0186] The descriptions of the apparatus embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. In some embodiments, the functions or modules included in the apparatus provided by the present invention can be used to perform the methods described in the method embodiments above. For technical details not disclosed in the apparatus embodiments of the present invention, please refer to the descriptions of the method embodiments of the present invention for understanding.

[0187] It should be noted that, in the embodiments of the present invention, if the above-described method for labeling industrial development reports based on knowledge graphs is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to related technologies, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, computer system, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of the present invention are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0188] Figure 3 A hardware entity diagram of a computer system provided as an embodiment of the present invention, such as... Figure 3 As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0189] The memory 1002 stores computer programs that can run on the processor. The memory 1002 is configured to store instructions and applications that can be executed by the processor 1001. It can also cache data to be processed or already processed by the processor 1001 and various modules in the system 1000 (e.g., image data, audio data, voice communication data, and video communication data). It can be implemented by flash memory or random access memory (RAM).

[0190] When processor 1001 executes a program, it implements the steps of the knowledge graph-based industrial development report annotation method described above. Processor 1001 typically controls the overall operation of computer system 1000.

[0191] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for annotating industrial development reports based on knowledge graphs, characterized in that, The method includes: Obtain the initial text data set corresponding to the industrial development report to be annotated. The initial text data set contains multiple report paragraph units with time sequence identifiers that are recorded continuously. The report paragraph units are divided according to the natural paragraphs of the report and maintain the original record order in time. Semantic topology parsing is performed on the initial text data set to obtain the context association path of the industry entity node in the report paragraph unit and the hierarchical distribution description of the knowledge structure in which the industry entity node is located. The context association path represents the semantic flow direction relationship and topic continuity relationship between industry entity nodes in adjacent paragraphs, and the hierarchical distribution description represents the hierarchical boundary integrity and hierarchical subordinate relationship of industry entity nodes in the knowledge structure. The process involves: 1) Calling a pre-built industry knowledge graph to perform cyclic path mapping on the context association path and the hierarchical distribution description, generating knowledge path annotation records for industry entity nodes in the report paragraph unit. These knowledge path annotation records indicate the association path branches and node location identifiers of the industry entity nodes in the knowledge graph. Specifically, this includes: inputting the context association path and the hierarchical distribution description into the path fusion layer of the industry knowledge graph; configuring association modeling for the semantic transmission trajectory description of the context association path and the hierarchical subordinate information of the hierarchical distribution description based on the fusion weight, generating a fusion path description vector with semantic topological consistency constraints; 2) using the multi-scale perception processing of the industry knowledge graph to collaboratively extract local detail semantics and global contextual information from the fusion path description vector, obtaining a multi-scale path-aware description set; and 3) using the semantic attention mechanism of the industry knowledge graph. The process involves semantic dependency modeling of the multi-scale path-aware description set to generate an attention distribution description map on the semantic dimension. This map indicates the distribution of importance of descriptions at different scales in knowledge path mapping. Based on the attention distribution description map, the multi-scale path-aware description set is allocated and aggregated. The multi-scale path-aware descriptions are proportionally synthesized according to the allocation ratio corresponding to each scale in the attention distribution description map, resulting in semantically enhanced path vectors with strengthened semantic relevance. The path classification output layer of the industry knowledge graph is then invoked to perform knowledge path mapping on the semantically enhanced path vectors. By evaluating the matching degree between the semantically enhanced path vectors and pre-stored path templates in the knowledge graph, knowledge path annotation records containing a knowledge path relevance strength distribution description are generated. This knowledge path relevance strength distribution description represents the relevance strength distribution corresponding to different industrial development trends. Based on the knowledge path annotation record, determine the industrial development path description and the spatiotemporal interweaving distribution description of the knowledge units in the report text corresponding to the industrial development report to be annotated. The spatiotemporal interweaving distribution description includes the evolution path of the coverage area and the trend of distribution density change of the knowledge units in the report paragraphs. Based on the description of the industrial development trend and the description of the spatiotemporal intertwined distribution, a report annotation instruction containing a knowledge unit positioning sequence is generated, and the report annotation instruction is sent to the target report processing terminal to start the annotation synchronization operation.

2. The method according to claim 1, characterized in that, The semantic topological parsing of the initial text data set yields the contextual association paths of industry entity nodes in the report paragraph units and the hierarchical distribution description of the knowledge structure in which the industry entity nodes reside, including: The initial text data set is decomposed into paragraph structure. Based on the natural paragraph separators and semantic coherence constraints in the report, the initial text data set is divided into multiple text paragraph slices with sequential relationships, and each text paragraph slice maintains a one-to-one correspondence with the paragraph sequence identifiers in the original report. Entity recognition and association capture are performed on the text segment slices. By recognizing the semantic tags carried by industry entity nodes in adjacent text segment slices, the semantic transmission trajectory of industry entity nodes in cross-segment context is captured, and a context association path containing a consistent description of semantic transmission direction and a description of topic continuity is generated. The text segment slices are divided into knowledge levels. Based on the hierarchical relationship of industry entity nodes in the knowledge system, the entity nodes within each text segment slice are classified into hierarchical categories to obtain a hierarchical distribution description that includes the boundary integrity description of the segmentation level and the degree of subordinate association between levels. The context association path and the hierarchical distribution description are bridged in the semantic alignment process. By constructing a unified semantic mapping framework across the description space, the semantic transmission trajectory of the context association path and the hierarchical subordinate information of the hierarchical distribution description are aligned to the same semantic representation plane, resulting in a set of text semantic coupling descriptions with a unified representation form. Based on the saliency measurement results of different description dimensions in the text semantic coupling description set, the path saliency description corresponding to the context association path and the hierarchical saliency description corresponding to the hierarchical distribution description are extracted respectively. Based on the path saliency description and the hierarchical saliency description, the fusion weight configuration of the context association path and the hierarchical distribution description in the cyclic path mapping process is adjusted to obtain the adjusted fusion weight configuration. The fusion weight configuration is used to coordinate the influence of semantic transmission trajectory and hierarchical subordination relationship in the process of industry knowledge graph mapping.

3. The method according to claim 2, characterized in that, The entity recognition and association capture of the text segment slices involves identifying the semantic tags carried by industry entity nodes in adjacent text segment slices, capturing the semantic transmission trajectory of industry entity nodes in cross-segment contexts, and generating a context association path that includes a description of semantic transmission direction consistency and a description of topic continuity. Identify semantic change regions between adjacent text paragraph slices, extract entity node pairs that undergo semantic transmission by comparing the differences in entity node sets in the preceding and following paragraphs, and generate a flow distribution description of the entity node pairs in the semantic space. The flow distribution description indicates the direction path and intensity profile of semantic transmission from the source paragraph to the target paragraph. Based on the flow distribution description, determine the semantic migration direction offset angle and semantic migration intensity change of the industry entity node between adjacent text segment slices; The semantic trajectory continuity analysis process of the industrial entity node is constructed based on the semantic migration direction offset angle and the semantic migration intensity change. The semantic trajectory continuity analysis process verifies the smoothness of the semantic transmission trajectory according to the offset angle threshold and intensity change tolerance, and outputs the trajectory continuity evaluation indicator. The semantic trajectory continuity analysis process is used to perform continuity calculations on the semantic transmission trajectory of the industry entity nodes in continuous text segment slices, and generate a consistent description of the semantic transmission direction. The semantic transmission trajectory of the industrial entity nodes is subjected to theme periodicity pattern recognition. The semantic transmission trajectory of each entity node is transformed in the frequency domain according to the paragraph span. The theme fluctuation cycle that recurs within the preset paragraph span interval is extracted, and the amplitude stability of the theme fluctuation cycle is calculated as the theme continuity description. The semantic transmission direction consistency description and the topic continuity description are combined and encapsulated to obtain the context association path. The context association path uses entity nodes as indexes and paragraph order as axes to record the semantic transmission direction trajectory and topic continuity intensity map of each industry entity node in the full text of the report.

4. The method according to claim 2, characterized in that, The process of performing knowledge-level segmentation on the text segment pieces involves classifying the entity nodes within each text segment based on their hierarchical relationships within the knowledge system. This yields a hierarchical distribution description that includes a description of the boundary integrity of the segmentation levels and a description of the degree of subordination between levels, including: The text segment slices are semantically enhanced by extracting contextual modifiers around entity nodes in each segment to strengthen the semantic outline boundary information of entity nodes, thus obtaining semantically enhanced segment slices. Based on the semantic contour boundary information in the paragraph slice after semantic boundary enhancement, each text paragraph slice is divided into knowledge regions by growth. The superior concept of the entity node is used as the seed point, and the subordinate entities are aggregated outward according to the semantic inclusion relationship to obtain multiple non-overlapping knowledge sub-region units. Each knowledge sub-region unit represents an independent knowledge level group. Extract semantic concentration distribution information within each knowledge sub-region unit, and generate a semantic concentration distribution description by region by statistically analyzing the number of entity nodes within the region and the semantic correlation between them. The deviation of the knowledge sub-region unit in the knowledge dimension is calculated based on the semantic concentration distribution description. The semantic concentration in the region is compared with the preset hierarchical semantic benchmark concentration to obtain a discreteness description that measures the clarity of the hierarchical boundary. The boundary integrity description of the knowledge sub-region unit is determined based on the discreteness description. The boundary integrity description is configured to be inversely correlated with the discreteness description, and the boundary integrity description decreases when the discreteness description increases. The relative positional relationships between different knowledge sub-region units are analyzed to generate a hierarchical subordinate association degree descriptor that includes a distance distribution descriptor and a direction angle description. The distance distribution descriptor indicates the interval between different knowledge sub-region units in the semantic space, and the direction angle description indicates the subordinate direction deviation between levels.

5. The method according to claim 1, characterized in that, The multi-scale perceptual processing of the industry knowledge graph is used to collaboratively extract local detail semantics and global contextual information from the fused path description vector, resulting in a multi-scale path-aware description set, including: The fused path description vector is input into the path segmentation layer of the multi-scale perception processing. The representation channels of the fused path description vector are evenly divided according to the preset path segmentation ratio to generate a dual-path parallel semantic stream containing local representation branches and global representation branches. In the local representation branch, multiple context ranges with different expansion amplitudes are used to extract the local representation branch and perform multi-level semantic receptive field expansion. Local detail semantic description fragments at different spatial scales are extracted from the local representation branch, and the channel capacity of the local detail semantic description fragments is aligned with the channel capacity of the global representation branch. In the global representation branch, global statistical information is condensed. The global representation branch is compressed into a condensed context description with global summary information through cross-channel mean aggregation operation. The condensed context description is then reshaped in terms of spatial dimensions to obtain a global context semantic description that matches the size of the local detail semantic description fragment. The local detail semantic description fragment and the global context semantic description are input into the semantic fusion processing process. Based on the channel attention mechanism, the fusion allocation ratio of each channel is generated. The local detail semantic description fragment and the global context semantic description are synthesized across scales through the fusion allocation ratio to generate a multi-level fusion semantic description map. Each level of the multi-level fusion semantic description map corresponds to a local and global fusion result at a receptive field scale. Channel normalization is performed on the multi-level fusion semantic description mapping to adjust the fusion mappings of different scales to a unified numerical range, and the normalized fusion mappings are spliced ​​along the channel dimension to generate a spliced ​​semantic description set with a unified representation dimension. The concatenated semantic description set is subjected to channel semantic recombination. The mutual information between the representation vectors of each channel is calculated. Based on the mutual information, the channels are clustered and merged. The semantic expression of each channel is reconstructed according to the cluster center vector to obtain the recombined multi-scale path-aware description set.

6. The method according to claim 1, characterized in that, The step of processing the multi-scale path-aware description set through the semantic attention mechanism of the industry knowledge graph to perform semantic dependency modeling and generate an attention distribution description graph in the semantic dimension includes: The multi-scale path-aware description set is input into the query construction layer processed by the semantic attention mechanism to generate a corresponding query description vector for each scale path-aware description in the multi-scale path-aware description set. The multi-scale path-aware description set is simultaneously input into the key-value construction layer processed by the semantic attention mechanism, and a linear mapping is performed on each scale path-aware description to generate a key description vector and a value description vector. Calculate the semantic fit between each query description vector and all key description vectors. Generate the original attention distribution description map by performing a dot product operation on the query description vector and the key description vector and adjusting the scaling. The original attention distribution description map represents the relevance strength of path perception description at each scale under the current query perspective. The original attention distribution description map is normalized by converting the correlation intensity of each element in the original attention distribution description map into a fixed total distribution ratio, thus obtaining a normalized attention distribution description map. The normalized attention distribution description map is smoothed to obtain a smoothed attention distribution description map; Extract statistical information from the smoothed attention distribution description map, including the scale identifier corresponding to the attention distribution peak and the description of the dispersion of the attention distribution, and output the smoothed attention distribution description map and the statistical information.

7. The method according to claim 1, characterized in that, The process of determining the industrial development trajectory description and the spatiotemporal distribution description of knowledge units in the report text corresponding to the industrial development report to be annotated based on the knowledge path annotation records includes: The knowledge path association strength distribution description in the knowledge path annotation record is parsed, and the path branches whose association strength exceeds the preset strength value in the knowledge path association strength distribution description and their corresponding timestamp information are detected. The timestamp information corresponds to the time sequence identifier of the report paragraph unit. Based on the timestamp information corresponding to the detected path branches, and combined with the continuous change trend of the correlation strength of the path branches in the preceding and following report paragraph units, the start and end times of each development stage in the industrial development context are determined, and the time span between the start and end times is calculated as a description of the duration of the corresponding development stage. Extract the set of industry entity nodes in the corresponding paragraph unit of the path branch, analyze the hierarchical position of the set of industry entity nodes in the knowledge graph, determine the combination relationship of core industry nodes and peripheral industry nodes involved in each development stage in the industrial development path, and generate a description of the direction of the industrial development path. For each knowledge unit, based on the location information of its industry entity node in the knowledge path annotation record at different time periods, the initial paragraph position and displacement vector of the knowledge unit in the report text are constructed to obtain the spatiotemporal movement trajectory of the knowledge unit. The spatiotemporal movement trajectory of the knowledge unit is analyzed for density change. The frequency change of the knowledge unit in continuous paragraphs and the trend of increase or decrease in the number of paragraphs covered are statistically analyzed. A spatiotemporal intertwined distribution description of the knowledge unit in the report text is generated. The spatiotemporal intertwined distribution description includes a description of the distribution density change of the knowledge unit and the directional trend of expansion or contraction of the coverage area. Establish a stage-related mapping between the description of the development trend and the description of the spatiotemporal intertwined distribution, associate and bind each development stage with the spatiotemporal movement trajectory segment of the knowledge unit within the corresponding time period, and write the binding relationship into the description of the development trend and the description of the spatiotemporal intertwined distribution respectively.

8. The method according to claim 7, characterized in that, For each knowledge unit, based on the position information of its respective industry entity node in the knowledge path annotation record at different time periods, the initial paragraph position and displacement vector of the knowledge unit in the report text are constructed to obtain the spatiotemporal movement trajectory of the knowledge unit, including: The location identifiers of the industry entity nodes associated with the target knowledge unit in each report paragraph unit are extracted from the knowledge path annotation records to obtain a location sequence arranged in chronological order. The location identifiers include the paragraph number and the node index in that paragraph. Based on the positioning sequence, the initial appearance paragraph position of the target knowledge unit is determined, and the initial appearance paragraph position is taken as the trajectory starting point. The paragraph number and node index in the positioning identifier corresponding to the trajectory starting point are extracted as the starting point coordinates. For each subsequent position point in the positioning sequence, excluding the starting coordinates, calculate the displacement vector of that position point relative to the previous position point. The displacement vector includes the direction indicator of the crossing between paragraphs and the change in the node index. The direction indicator indicates the migration pattern of the knowledge unit between paragraphs: moving forward, referencing backward, or remaining unchanged. The spatiotemporal movement trajectory of the knowledge unit is generated based on the displacement vector. The spatiotemporal movement trajectory is constructed with the starting point coordinates as the origin, the paragraph number as the horizontal axis, and the relative position of the node index within the paragraph as the vertical axis, to construct a trajectory curve describing the movement path of the knowledge unit in the report text. The spatiotemporal movement trajectory is smoothed by connecting displacement vectors to construct a continuous trajectory, resulting in the smoothed spatiotemporal movement trajectory of the knowledge unit. The smoothed spatiotemporal movement trajectory of the knowledge unit is associated with the hierarchical information of the industry entity nodes associated with the knowledge unit in the knowledge graph to generate a spatiotemporal movement trajectory of the knowledge unit with hierarchical markings.

9. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images