A key entity extraction method for digitized archival texts

Abstract

The application relates to the technical fields of electric digital data processing and knowledge graph fusion, and discloses a key entity extraction method for digital archive texts, which comprises the following steps: acquiring multiple-source digital archive texts and generating an initial sequence boundary probability matrix; constructing a local topological network and comparing the local topological network with a global prior network to identify a topological conflict subgraph and determine a conflict probability matrix; triggering boundary probability consistency verification to calculate vector similarity, and generating a final de-conflict boundary probability rewriting matrix when a node time span constraint and a step reduction rule are satisfied; and overwriting the corresponding initial sequence boundary probability matrix by using the final de-conflict boundary probability rewriting matrix to adjust the cutting boundary probability. The application constructs a bidirectional dependence closed loop of a graph topology and a boundary probability, effectively suppresses cascade amplification of input noise and suppresses boundary probability iterative divergence, and improves the entity extraction accuracy.

Description

Technical Field

[0001] This invention relates to the fields of electronic digital data processing and knowledge graph fusion technology, and in particular to a method for extracting key entities from digitized archival texts. Background Technology

[0002] Currently, in the process of converting digitized archival text into structured graph networks, a serial processing method is usually adopted. The processor first calculates the initial boundary probability matrix of the character sequence through sequence labeling rules, and determines the segmentation position of the entity text accordingly. Then, the candidate segments obtained from the segmentation are sent into the graph network to establish the association between multi-source data and store relevant knowledge. This unidirectional processing method, which gradually enters the graph network construction from character segmentation, usually assumes that the previous segmentation result has reached a local optimum, and also assumes that the local boundary division is not affected by the subsequent global network topology constraints.

[0003] For text streams obtained from paper archives through image recognition, when the original archives have missing text due to weathering, misformatted layout, or excessive noise, the pre-sequence annotation process lacks back feedback from graph topological relationships. The processor is prone to probability distribution shifts when determining character boundaries, causing the generated initial boundary probability matrix to deviate from the actual text structure. Since the pre-level character segmentation and the subsequent graph network construction are separate, small local segmentation deviations continue to propagate to the subsequent graph network and gradually amplify as relationships between character nodes are established. This causes the topological connections in the graph network to become discrete, ultimately resulting in structural differences after merging multi-source data exceeding the preset range. Existing improvements to address this error amplification problem typically involve adjusting some parameters of the pre-level sequence model or increasing the number of network layers to improve boundary prediction accuracy. However, these methods consume more computational resources, are prone to state divergence when data relationships are sparse, and cannot eliminate the separation between character segmentation and graph network fusion.

[0004] Therefore, how to enable the previous character segmentation result to be dynamically adjusted according to the topological constraints of the subsequent graph, suppress the cascading amplification of errors caused by noise, and ensure that the boundary probability adjustment process remains stable and convergent under high interference conditions has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems in the background art, the present invention provides the following technical solution: a method for extracting key entities from digitized archival text, comprising the following steps:

[0006] Step S1: Obtain multi-source digitized archival text, segment the character stream of candidate segments in the multi-source digitized archival text, and generate multiple initial sequence boundary probability matrices corresponding to different text sources and representing the segmentation boundary probability of each character node;

[0007] Step S2: Based on the boundary probability matrices of each initial sequence, construct a local topology network containing the associated edges between character nodes, compare the local topology network with the stored global prior network, identify the topological conflict subgraphs that do not satisfy the topological constraints of the global graph, and determine the boundary probability matrices of the initial sequences with topological conflicts as the conflict probability matrices.

[0008] Step S3: Based on the topological conflict subgraph, trigger the boundary probability consistency check, calculate the vector similarity between the conflict probability matrices, and when the node temporal span constraint and step size reduction rule are satisfied, suppress the divergence tendency of the boundary probability of the conflict probability matrix in the iteration process, and fuse to generate the final conflict-free boundary probability rewrite matrix.

[0009] Step S4: Overwrite the corresponding initial sequence boundary probability matrix with the final conflict-free boundary probability rewrite matrix, adjust the segmentation boundary probability of each character node, generate the corrected entity extraction sequence, and output the target entity extraction boundary of the multi-source digitized archival text.

[0010] Preferably, step S2 includes the following sub-steps: step S21, calculating the temporal span of the character nodes at both ends of each associated edge in the local topology network in the character stream; step S22, when a temporal span of more than 18 nodes is detected and the corresponding structural confidence is less than 0.65, the topological weight of the associated edge is reset to 0 to filter out long-distance interfering network edges and identify topological conflict subgraphs that do not satisfy the global graph topology constraints.

[0011] Preferably, in step S3, triggering the boundary probability consistency check includes the following sub-steps: Step S31, when the same character node generates multiple mutually exclusive conflict probability matrices in a multi-source heterogeneous text stream, the vector similarity between the conflict probability matrices is calculated using the following formula: ,in, The vector similarity between the conflict probability matrices. Let S32 be the directional angle of the conflict probability matrix in the feature space; Step S32: When the calculated vector similarity is less than 0.35, extract the node degree centrality value of each local topology network in the global prior network to determine the arbitration priority of the conflict probability matrix.

[0012] Preferably, in step S3, the step size reduction rule includes the following sub-steps: Step S33, real-time monitoring of the monotonic variation of the probability correction value in the final conflict-free boundary probability rewriting matrix; Step S34, when the calculated monotonic variation is continuously less than 0.05, the adjustment step size of the probability correction value is reduced by 50% to suppress the iterative divergence of the boundary probability and complete the probability convergence within 5 iterations.

[0013] Preferably, after step S32, triggering the boundary probability consistency check further includes the following sub-steps: step S35, assigning normalized weights to multiple mutually exclusive conflict probability matrices based on the obtained node degree centrality values; step S36, using the normalized weights to perform linear weighted fusion of the conflict probability matrices to generate a unique final conflict-free boundary probability rewriting matrix.

[0014] Preferably, step S1 includes the following sub-steps: step S11, inputting the multi-source digitized archive text into a pre-stored sequence labeling model to obtain the temporal feature vector of each character node; step S12, calculating the segmentation probability between adjacent character nodes based on the temporal feature vector, and constructing a multidimensional probability distribution based on the segmentation probability to generate an initial sequence boundary probability matrix.

[0015] Preferably, in step S2, the construction of the global prior network includes the following sub-steps: step S23, obtaining the historical archive association graph; step S24, calculating the node co-occurrence frequency in the historical archive association graph, filtering out association edges whose node co-occurrence frequency is lower than a preset stability threshold, and storing the remaining topological association network as the global prior network.

[0016] Preferably, in step S34, completing probability convergence within 5 iterations includes the following sub-steps: Step S341, setting up a loop counter, and incrementing the count value of the loop counter by 1 after each adjustment step reduction; Step S342, when the count value of the loop counter reaches 5 and the boundary probability has not converged, terminating the probability overwrite loop and locking the probability correction value of the current control cycle.

[0017] Preferably, the noun objects included in the modified entity extraction sequence are derived from the original structure of the archive, and the target entity extraction boundary corresponds to the person name, place name, organization name, historical event name, and archive file number.

[0018] Preferably, after outputting the target entity extraction boundary, the following steps are also included: Step S5, based on the target entity extraction boundary, extract key structured key-value pairs from the multi-source digital archive text, and write the key structured key-value pairs into the digital archive knowledge graph in the form of triples to update the graph topology of the digital archive knowledge graph.

[0019] The beneficial effects of this invention are:

[0020] 1. By comparing the local topology network with the global prior network and correcting the initial sequence boundary probability matrix based on the identified topological conflict subgraphs, the character segmentation of the preceding level can be adjusted using the topological constraints of the subsequent level graph. This reduces the continuous propagation of local segmentation deviations to the subsequent level graph network, suppresses the cascading amplification of errors caused by input noise, improves the accuracy of target entity extraction boundaries, and enhances the structural consistency of multi-source archival data.

[0021] 2. By dynamically adjusting the step size of the probability correction value through the step size reduction rule and restricting the probability overwrite loop, the fluctuation of the boundary probability during repeated correction can be reduced, and the iteration process can be prevented from being difficult to converge due to continuous oscillation. This method can keep the boundary probability correction process stable without increasing the processor's computational burden, and obtain usable correction results within a limited number of iterations.

[0022] 3. By identifying and filtering out long-distance interfering network edges with low confidence through node temporal span constraints, the impact of invalid associations on local topology networks can be reduced, and the identification of long-distance erroneous dependencies on interfering topological conflict subgraphs can be avoided. This allows the correction of boundary probabilities to focus on character nodes with more explicit associations, improves the reliability of local topological relationship judgment, and reduces the possibility of the continuous spread of erroneous associations in sparse graph networks. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0024] Figure 1 This is a flowchart of the probability correction process for the boundary of text entities in digital archives according to the present invention.

[0025] Figure 2 This is a schematic diagram of the entity boundary probability de-conflict control system of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] A method for extracting key entities from digitized archival text includes the following steps:

[0028] Step S1: Obtain multi-source digitized archival text, segment the character stream of candidate segments in the multi-source digitized archival text, and generate multiple initial sequence boundary probability matrices corresponding to different text sources and representing the segmentation boundary probability of each character node;

[0029] Step S2: Based on the boundary probability matrices of each initial sequence, construct a local topology network containing the associated edges between character nodes, compare the local topology network with the stored global prior network, identify the topological conflict subgraphs that do not satisfy the topological constraints of the global graph, and determine the boundary probability matrices of the initial sequences with topological conflicts as the conflict probability matrices.

[0030] Step S3: Based on the topological conflict subgraph, trigger the boundary probability consistency check, calculate the vector similarity between the conflict probability matrices, and when the node temporal span constraint and step size reduction rule are satisfied, suppress the divergence tendency of the boundary probability of the conflict probability matrix in the iteration process, and fuse to generate the final conflict-free boundary probability rewrite matrix.

[0031] Step S4: Overwrite the corresponding initial sequence boundary probability matrix with the final conflict-free boundary probability rewrite matrix, adjust the segmentation boundary probability of each character node, generate the corrected entity extraction sequence, and output the target entity extraction boundary of the multi-source digitized archival text.

[0032] Preferably, step S2 includes the following sub-steps: step S21, calculating the temporal span of the character nodes at both ends of each associated edge in the local topology network in the character stream; step S22, when a temporal span of more than 18 nodes is detected and the corresponding structural confidence is less than 0.65, the topological weight of the associated edge is reset to 0 to filter out long-distance interfering network edges and identify topological conflict subgraphs that do not satisfy the global graph topology constraints.

[0033] Preferably, in step S3, triggering the boundary probability consistency check includes the following sub-steps: Step S31, when the same character node generates multiple mutually exclusive conflict probability matrices in a multi-source heterogeneous text stream, the vector similarity between the conflict probability matrices is calculated using the following formula: ,in, The vector similarity between the conflict probability matrices. Let S32 be the directional angle of the conflict probability matrix in the feature space; Step S32: When the calculated vector similarity is less than 0.35, extract the node degree centrality value of each local topology network in the global prior network to determine the arbitration priority of the conflict probability matrix.

[0034] Preferably, in step S3, the step size reduction rule includes the following sub-steps: Step S33, real-time monitoring of the monotonic variation of the probability correction value in the final conflict-free boundary probability rewriting matrix; Step S34, when the calculated monotonic variation is continuously less than 0.05, the adjustment step size of the probability correction value is reduced by 50% to suppress the iterative divergence of the boundary probability and complete the probability convergence within 5 iterations.

[0035] Preferably, after step S32, triggering the boundary probability consistency check further includes the following sub-steps: step S35, assigning normalized weights to multiple mutually exclusive conflict probability matrices based on the obtained node degree centrality values; step S36, using the normalized weights to perform linear weighted fusion of the conflict probability matrices to generate a unique final conflict-free boundary probability rewriting matrix.

[0036] Preferably, step S1 includes the following sub-steps: step S11, inputting the multi-source digitized archive text into a pre-stored sequence labeling model to obtain the temporal feature vector of each character node; step S12, calculating the segmentation probability between adjacent character nodes based on the temporal feature vector, and constructing a multidimensional probability distribution based on the segmentation probability to generate an initial sequence boundary probability matrix.

[0037] Preferably, in step S2, the construction of the global prior network includes the following sub-steps: step S23, obtaining the historical archive association graph; step S24, calculating the node co-occurrence frequency in the historical archive association graph, filtering out association edges whose node co-occurrence frequency is lower than a preset stability threshold, and storing the remaining topological association network as the global prior network.

[0038] Preferably, in step S34, completing probability convergence within 5 iterations includes the following sub-steps: Step S341, setting up a loop counter, and incrementing the count value of the loop counter by 1 after each adjustment step reduction; Step S342, when the count value of the loop counter reaches 5 and the boundary probability has not converged, terminating the probability overwrite loop and locking the probability correction value of the current control cycle.

[0039] Preferably, the noun objects included in the modified entity extraction sequence are derived from the original structure of the archive, and the target entity extraction boundary corresponds to the person name, place name, organization name, historical event name, and archive file number.

[0040] Preferably, after outputting the target entity extraction boundary, the following steps are also included: Step S5, based on the target entity extraction boundary, extract key structured key-value pairs from the multi-source digital archive text, and write the key structured key-value pairs into the digital archive knowledge graph in the form of triples to update the graph topology of the digital archive knowledge graph.

[0041] Example 1:

[0042] The electronic digital data processing system continuously receives multi-source digitized archival texts containing ink smears, layout errors, or character recognition noise. The processor acquires candidate segments from each text source, segments the character stream of the candidate segments using a pre-stored sequence labeling model, and obtains the temporal feature vector of each character node. The processor calculates the segmentation probability based on the temporal feature vectors of adjacent character nodes, and constructs a multidimensional probability distribution according to the arrangement order of character nodes in the candidate segments. This generates multiple initial sequence boundary probability matrices corresponding to different text sources and representing the segmentation boundary probability of each character node. When constructing a local topology network, character nodes are used as network nodes. Association edges are established based on the order relationship of character nodes in the candidate segments and the segmentation probability between adjacent character nodes, and the corresponding association results are recorded as the topological weights of the association edges.

[0043] The global prior network is constructed based on the historical archive association graph. The processor counts the co-occurrence frequency between nodes in the historical archive association graph and determines the node co-occurrence frequency based on the number of archive texts involved in the count. After filtering out association edges whose node co-occurrence frequency is lower than a preset stability threshold, the topological association network formed by the remaining nodes and association edges is stored as the global prior network. The node relationships and association edge relationships retained in this network constitute the global graph topological constraints. The processor compares the nodes and association edges in each local topological network with the global prior network one by one, and classifies the association edges that do not satisfy the global graph topological constraints and their connected character nodes into the topological conflict subgraph. Then, based on the text source and matrix position corresponding to the character nodes in the topological conflict subgraph, the processor further classifies them into the topological conflict subgraph. The processor backtracks the corresponding initial sequence boundary probability matrix and determines the initial sequence boundary probability matrix with topological conflicts as the conflict probability matrix. During the associated edge detection process, the processor calculates the temporal span of the character nodes at both ends of each associated edge in the local topological network in the character stream according to the node temporal span constraint. When constructing the local topological network, the confidence result output by the sequence labeling model for the association relationship between the two end character nodes is recorded as the structural confidence of the corresponding associated edge. When the temporal span exceeds 18 nodes and the structural confidence is lower than 0.65, the topological weight of the associated edge is reset to 0, long-distance interfering network edges are filtered out from the local topological network, and the topological conflict subgraph is determined according to the filtered nodes and the relationship of the associated edges.

[0044] When the same character node forms mutually exclusive segmentation results in multiple conflict probability matrices corresponding to multi-source heterogeneous text streams, the system triggers boundary probability consistency verification. The processor extracts the corresponding boundary probabilities from each conflict probability matrix according to the same character node order and forms a vector. The vector similarity between the conflict probability matrices is calculated using the following formula: ,in, The vector similarity between the conflict probability matrices. Let be the directional angle of the conflict probability matrix in the feature space. When the vector similarity is less than 0.35, the processor counts the number of valid associated edges connected to the conflicting character nodes in the global prior network to obtain the node degree centrality value corresponding to each local topology network. The larger the node degree centrality value, the higher the arbitration priority of the corresponding conflict probability matrix. The processor compares each node degree centrality value with the sum of the degree centrality values ​​of all nodes to be arbitrated to obtain the normalized weights corresponding to the multiple mutually exclusive conflict probability matrices. Then, the normalized weights are used to perform linear weighted fusion of the conflict probability matrices to generate a unique final conflict-free boundary probability rewriting matrix.

[0045] During the fusion process, the processor executes a step-size reduction rule, using the difference between the probability values ​​at the same matrix position of the final conflict-free boundary probability rewritten matrix and the corresponding conflict probability matrix as the probability correction value, and determining the monotonic variation based on the absolute value of the difference between the probability correction values ​​obtained from two adjacent iterations. When the monotonic variation is consistently less than 0.05, the adjustment step size of the probability correction value is reduced by 50%, and the boundary probability is updated again according to the reduced adjustment step size. The processor sets up a loop counter, which increments by 1 each time the adjustment step size is reduced. When the boundary probability remains stable in subsequent iterations, probability convergence is determined to be complete, and the iteration is stopped, limiting the probability convergence process to within 5 iterations. The current control cycle is the processing cycle for boundary probability overwriting of this batch of multi-source digitized archival texts. When the loop counter reaches 5 and the boundary probability has not yet converged, the probability overwriting loop is terminated, and the probability correction value of the current control cycle is locked.

[0046] The processor overwrites the corresponding initial sequence boundary probability matrix using the final conflict-resolving boundary probability rewrite matrix, adjusts the segmentation boundary probability of each character node, and performs boundary determination according to the arrangement order of character nodes in the character stream. Positions satisfying the segmentation conditions are written into the corrected entity extraction sequence. The processor records the start and end positions of each continuous character interval, forming the target entity extraction boundary of the multi-source digitized archival text. The noun objects contained in the corrected entity extraction sequence are taken from the original archival text structure, and the target entity extraction boundary corresponds to personal names, place names, organization names, historical event names, and archival file numbers. The output target entity extraction... After identifying the boundaries, the processor extracts the corresponding entity text according to each target entity, using the entity type as the key and the entity text as the value to form key-value pairs. Then, based on the relationships between entities in the archival text, the processor combines the corresponding entities and their relationships into subject, predicate, and object triples, and writes them into the digital archival knowledge graph, updating the nodes, edges, and graph topology in the digital archival knowledge graph. Through the above processing, the topological constraints in the global prior network can adjust the initial sequence boundary probability matrix in reverse, reducing the spread of local segmentation bias in the local topological network and suppressing the divergence of boundary probabilities during the iteration process.

[0047] Example 2:

[0048] The offline archive automated processing test bench uses a general-purpose digital processor and a high-speed random access memory. The general-purpose digital processor processes 200 character streams per second. The test data is taken from heterogeneous text streams in the digitized archives and superimposed with Gaussian-distributed character segmentation drift noise to simulate the interference caused by ink annihilation and typesetting errors on character boundary recognition. The experiment sets up three independent test channels: the first channel uses the complete key entity extraction method as the test group using the method of this invention; the second channel removes the reverse feedback overwriting process between the local topology network and the global prior network as the control group; the third channel changes the truncation threshold of the node temporal span constraint as the out-of-range control group, and sets up two test branches with 25 nodes and 10 nodes respectively. Each test channel inputs test data according to three levels of noise: low noise, medium noise, and high noise. Irregular random fluctuations are added to the initial sequence boundary probability matrix, and the initial value of the segmentation boundary probability of each character node is uniformly distributed in the range of 0.45 to 0.55.

[0049] In the test channel employing the method of this invention, the sequence labeling model segments the character stream in the heterogeneous text stream, outputting multiple initial sequence boundary probability matrices. A general-purpose digital processor (DPC) then constructs a local topology network containing character nodes and their associated edges. During testing, the truncation threshold for node temporal span constraints is progressively adjusted to balance the preservation of long-distance semantic associations with the filtering of long-distance interfering network edges. As the candidate segment length increases, the truncation threshold is adjusted towards a decreasing value, and the working threshold is determined to be 18 nodes. When the temporal span of the character nodes at both ends of an associated edge exceeds this working threshold, and... When the structural confidence is below 0.65, the topological weight of the corresponding associated edge is reset to 0. The topological conflict subgraph is determined based on the remaining nodes and associated edges, and the initial sequence boundary probability matrix with topological conflict is determined as the conflict probability matrix. When the same character node forms multiple mutually exclusive conflict probability matrices in a multi-source heterogeneous text stream, the system triggers boundary probability consistency verification. After extracting the boundary probabilities in each conflict probability matrix according to the arrangement order of the same character node, the vector similarity between the conflict probability matrices is calculated. The vector similarity obtained in this test is 0.282, which is lower than 0.35.

[0050] Based on this, the general-purpose digital processor extracts the node degree centrality values ​​of each local topology network in the global prior network, determines the arbitration priority of the conflict probability matrix, and assigns normalized weights according to the node degree centrality values. The conflict probability matrix is ​​then linearly weighted and fused using the normalized weights to generate a unique final conflict-free boundary probability rewriting matrix. During the iterative update process, the monotonic variation of the probability correction value is monitored in real time. When the monotonic variation is consistently less than 0.05, the adjustment step size of the probability correction value is reduced by 50%. Each time the adjustment step size is reduced, the count value of the loop counter is incremented by 1. When the count value reaches 5 and the boundary probability has not yet converged, the probability rewriting loop is terminated, and the probability correction value of the current control cycle is locked.

[0051] At low noise levels, the target entity extraction boundary accuracy rates of the test group, control group, and low-threshold out-of-range control group using the method of this invention are 96.4%, 88.2%, and 85.1%, respectively. At this level, the initial sequence boundary probability matrix of each test channel still retains relatively clear character boundary features, and the impact of the reverse feedback overwriting process on the results is relatively limited. However, after removing this process or filtering out long-distance associated edges in advance, the target entity extraction boundary accuracy rate decreases. At medium noise levels, the target entity extraction boundary accuracy rates of the three corresponding test channels are 94.1%, 72.5%, and 70.3%, respectively. As the character segmentation drift noise increases, the local segmentation deviation in the control group continues to propagate to the local topology network, while the low-threshold out-of-range control group filters out some real long sentence associations, and the gap between the two and the test group using the method of this invention widens accordingly. At high noise levels, ink annihilation and typesetting errors cause the local segmentation deviation in the control group to accumulate continuously along the unidirectional processing process, and its target entity extraction boundary accuracy rate drops to 54.3%.

[0052] The test group using the method of this invention adjusted the initial sequence boundary probability matrix in reverse by using the topological relationships in the global prior network, and the accuracy of the target entity extraction boundary was 91.6%. In the out-of-range control group, the high-threshold branch retained more long-distance interfering network edges, and the conflict associations in the local topological network were not filtered out in time, resulting in an accuracy of 61.2% for the target entity extraction boundary. The low-threshold branch truncated some real long sentence associations in advance, resulting in an accuracy of 65.8% for the target entity extraction boundary. The test results under various noise levels and different truncation thresholds show that by using topological conflict subgraph identification, conflict probability matrix weighted fusion, and boundary probability feedback overwriting, the spread of local segmentation deviations in the local topological network can be limited, and the boundary probabilities can be kept converged during the iteration process.

[0053] Example 3:

[0054] Large-scale text layout errors and local color-changing occlusions can cause discontinuous character boundaries in multi-source digitized archival texts. Local segmentation deviations output by the sequence labeling model, after entering the local topological network during data processing, can easily lead to mismatched associated edges and cause the heterogeneity of the graph topology structure after merging multi-source heterogeneous texts to deviate from the preset convergence boundary. This embodiment uses parameter calibration, node temporal span constraints, and step size reduction rules to adjust the segmentation boundary probability of character nodes. Before the system enters the online processing cycle, the general-purpose digital processor performs gradient testing on the target sample set containing manually labeled boundaries in an offline environment to determine the monotonic variation threshold and the calibration benchmark for adjusting the step size. The offline archive automated processing test bench extracts the sample heterogeneous text stream, injects character segmentation drift noise conforming to a Gaussian distribution into the feature space, calls the pre-stored sequence labeling model to calculate the segmentation probability between adjacent character nodes, calculates the vector similarity between conflict probability matrices, and monitors the iterative convergence process of the normalized weights, thereby determining the feedback adjustment interval used to suppress the divergence of boundary probabilities.

[0055] After receiving multi-source digitized archival texts online, the general-purpose digital processor allocates specific hardware register addressing space to carry the input data stream, calls the sequence labeling model to segment the character stream of candidate segments, and obtains the temporal feature vector of each character node. Based on the temporal feature vector, it calculates the segmentation probability between adjacent character nodes, and constructs a multidimensional probability distribution based on the segmentation probability to generate multiple initial sequence boundary probability matrices corresponding to different text sources and representing the segmentation boundary probability of each character node. Subsequently, based on each initial sequence boundary probability matrix, it constructs a local topology network containing the associated edges between character nodes, compares it one by one with the global prior network stored in high-speed random access memory, identifies topological conflict subgraphs that do not satisfy the topological constraints of the global graph, and determines the initial sequence boundary probability matrix with topological conflict as the conflict probability matrix.

[0056] During the topology conflict identification process, the general-purpose digital processor calculates the temporal span of the character nodes at both ends of each associated edge in the local topology network within the character stream. When the temporal span exceeds 18 nodes and the corresponding structural confidence is below 0.65, the topology weight of the associated edge is reset to 0, long-distance interfering network edges are filtered out, and the topology conflict subgraph is determined based on the remaining nodes and associated edges. When the same character node generates multiple mutually exclusive conflict probability matrices in a multi-source heterogeneous text stream, the system triggers a boundary probability consistency check and calculates the vector similarity between the conflict probability matrices. The vector similarity obtained in this process is 0.282, which is less than 0.35. Based on this, the general-purpose digital processor extracts the node degree centrality value of each local topology network in the global prior network, determines the arbitration priority of the conflict probability matrix, and assigns normalized weights according to the multiple conflict probability matrices with mutually exclusive node degree centrality values.

[0057] After linearly weighting and fusing the conflict probability matrix using normalized weights, a unique final conflict-free boundary probability rewriting matrix is ​​generated. During the boundary probability iteration process, the control operator monitors the monotonic variation of the probability correction value in the final conflict-free boundary probability rewriting matrix in real time. When the monotonic variation is consistently less than 0.05, the adjustment step size of the probability correction value is reduced by 50% by modifying the working state of the increment register. Each time the adjustment step size is reduced, the count value of the loop counter is incremented by 1. When the boundary probability converges before the count value reaches 5, the current iteration stops. When the count value of the loop counter reaches 5 and the boundary probability has not yet converged, the probability rewriting loop is terminated, and the probability correction value of the current control cycle is locked.

[0058] The general-purpose digital processor uses the final conflict-resolving boundary probability rewriting matrix to overwrite the corresponding initial sequence boundary probability matrix, adjusts the segmentation boundary probabilities of each character node, generates a corrected entity extraction sequence, and outputs the target entity extraction boundary of the multi-source digitized archival text. The noun objects in the corrected entity extraction sequence originate from the original archival text structure, and the target entity extraction boundary corresponds to person names, place names, organization names, historical event names, and archival file numbers. After completing the boundary output, the system extracts key structured key-value pairs from the multi-source digitized archival text based on the target entity extraction boundary, writes them into the digitized archival knowledge graph in the form of triples, and updates the nodes, associated edges, and graph topology in the digitized archival knowledge graph. Thus, the topological constraints in the global prior network can adjust the initial sequence boundary probability matrix in reverse, limit the topological conflicts caused by long-distance interfering network edges, and keep the boundary probability stable during the feedback adjustment process.

[0059] Example 4:

[0060] This embodiment combines Figures 1 to 2 This document describes a method for extracting key entities from digitized archival texts, such as... Figure 1As shown, in step S1, multi-source digitized archival text is acquired, and the character stream of candidate segments in the multi-source digitized archival text is segmented to generate multiple initial sequence boundary probability matrices corresponding to different text sources and representing the segmentation boundary probability of each character node; in step S2, a local topology network containing the associated edges between character nodes is constructed based on each initial sequence boundary probability matrix, and the local topology network is compared with the stored global prior network to identify topological conflict subgraphs that do not satisfy the topological constraints of the global graph, and the initial sequence boundary probability matrices with topological conflicts are determined as conflict probability matrices; in step S3, boundary probability consistency verification is triggered based on the topological conflict subgraphs to calculate the vector similarity between conflict probability matrices, and the divergence tendency of the boundary probability of the conflict probability matrix in the iteration process is suppressed when the node temporal span constraint and step size reduction rule are satisfied, so as to fuse and generate the final conflict-free boundary probability rewriting matrix; in step S4, the final conflict-free boundary probability rewriting matrix overwrites the corresponding initial sequence boundary probability matrix to adjust the segmentation boundary probability of each character node and generate the corrected entity extraction sequence, thereby outputting the target entity extraction boundary of the multi-source digitized archival text.

[0061] like Figure 2 As shown, the memory management unit is connected to the node for opening a static alignment buffer and the node for storing the global prior network, respectively. The control operator is connected to the node for generating the collision-free boundary probability rewrite matrix, the node for executing the step size reduction rule, and the node for real-time monitoring of monotonic changes, respectively. The general-purpose digital processor is connected to the node for acquiring multi-source digitized archive text, the node for triggering boundary probability consistency verification, the node for generating the collision-free boundary probability rewrite matrix, the node for adjusting the segmentation boundary probability, and the node for outputting the target entity extraction boundary; Figure 2In the control system topology shown, the node for acquiring multi-source digitized archive text is connected to the node for generating the initial sequence boundary probability matrix, and the node for generating the initial sequence boundary probability matrix is ​​connected to the node for constructing the local topology network. The node for storing the global prior network points upward to the node for opening a static alignment buffer, and the node for identifying the topology conflict subgraph points upward to the node for storing the global prior network. Simultaneously, the node for identifying the topology conflict subgraph connects to the node for constructing the local topology network to the right and downwards to the edge node for filtering long-distance interference. The node for calculating the vector similarity between conflict probability matrices connects upwards to the edge node for filtering long-distance interference, and the node for triggering boundary probability consistency verification connects upwards to the node for calculating the conflict. The vector similarity node between probability matrices connects upwards to the node that generates the conflict-free boundary probability rewrite matrix and triggers the boundary probability consistency check node. The step size reduction rule execution node points upwards to the node that generates the conflict-free boundary probability rewrite matrix, and the real-time monitoring monotonic change node points upwards to the node that executes the step size reduction rule. The segmentation boundary probability adjustment node points to the node that generates the conflict-free boundary probability rewrite matrix to the upper left and connects downwards to the node that generates the corrected entity extraction sequence. The node that generates the corrected entity extraction sequence connects to the node that outputs the target entity extraction boundary to the upper right, and the node that outputs the target entity extraction boundary connects downwards to the graph topology node that updates the digitized archive knowledge graph.

[0062] Example 5:

[0063] When multi-source digitized archival text data streams are accessed in parallel, the global prior network has not yet been written into the high-speed random access memory. The memory management unit first allocates an independent static alignment buffer. The general-purpose digital processor reads the preset hardware initialization parameter file, uses the feature dictionary index of the original structure of the digitized archival text as the addressing reference, allocates contiguous physical memory blocks for the graph nodes in the global prior network, and establishes a one-dimensional reference mapping table to bind each physical memory block to the logical address of the node degree centrality value. Before the data stream is transmitted, the values ​​of the increment register and the loop counter are cleared to enable the static alignment buffer to have the storage conditions to carry the global prior network and boundary probability feedback data. After the storage space initialization is completed, the general-purpose digital processor obtains the historical archival association graph, counts the number of times each node co-occurs in the historical archival text, and determines the node co-occurrence frequency based on the archival texts involved in the statistics. Association edges with node co-occurrence frequencies lower than the preset stability threshold are filtered out. The remaining nodes and association edges constitute a topological association network and are written into the corresponding physical memory blocks according to the one-dimensional reference mapping table, and stored in the high-speed random access memory as the global prior network. The degree centrality value of each node is synchronously written into the corresponding logical address for subsequent boundary probability consistency verification.

[0064] Before the electronic digital data processing system goes online, the general-purpose electronic digital processor sends a single synchronization pulse signal to the static alignment buffer via the data bus to detect the initial read latency during the unidirectional pipeline processing. When the initial read latency is greater than 15 milliseconds, the control operator adjusts the bus bandwidth allocation and shortens the refresh period of the initial sequence boundary probability matrix in the dynamic storage area to 10 microseconds. Subsequently, hardware interrupt sampling is performed continuously until the read latency is less than 10 milliseconds for five consecutive measurements. Then, the physical refresh frequency of the dynamic storage area is locked. After the read latency and refresh period calibration are completed, the initial confidence parameters of the target entity boundary are extracted. Set to 0.50 as the initial baseline when the system processes multi-source digitized archival texts online.

[0065] Example 6:

[0066] After the newly formed multi-source digital archival text completes the target entity extraction boundary output, the general-purpose digital processor extracts the corresponding character ranges according to the extraction boundaries of each target entity, and determines the noun objects in the corrected entity extraction sequence by combining the original archival text structure. Each noun object corresponds to a person's name, place name, organization name, historical event name, or archive file number. The processor organizes its text content and corresponding category in the multi-source digital archival text into key structured key-value pairs. Based on the existing entity relationships in the original archival text structure, the processor combines the key structured key-value pairs, writing the related noun objects and their relationships into subject, predicate, and object triples. For noun objects already existing in the digital archival knowledge graph, the processor writes the associated edges according to the graph nodes corresponding to the noun objects. For noun objects that have not yet formed graph nodes, the processor writes the corresponding nodes and associated edges, thereby updating the graph topology of the digital archival knowledge graph. After completing the graph topology update, the processor updates the digital archival knowledge graph. The spectrum, as a historical archive association graph, counts the co-occurrence frequency of nodes based on the common occurrence of noun objects in the historical archive text. Edges with a co-occurrence frequency below a preset stability threshold are not written into the network used for subsequent processing. The remaining nodes and edges form a topological association network and are stored as a global prior network. When the next batch of multi-source digitized archive texts enters the processing flow, the processor constructs a local topological network based on the initial sequence boundary probability matrix and compares it with the updated global prior network one by one. Character nodes and edges that do not satisfy the global graph topological constraints are assigned to a topological conflict subgraph, and the corresponding initial sequence boundary probability matrix is ​​determined as a conflict probability matrix. After completing boundary probability consistency verification and linear weighted fusion, the processor uses the final de-conflict boundary probability rewriting matrix to adjust the segmentation boundary probability of character nodes. The resulting target entity extraction boundary continues to be used for key structured key-value pair extraction and digitized archive knowledge graph updates, enabling the entity relationships formed by newly added archive texts to enter the subsequent global prior network.

[0067] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for extracting key entities from digitized archival text, characterized in that, Includes the following steps: Step S1: Obtain multi-source digitized archival text, segment the character stream of candidate segments in the multi-source digitized archival text, and generate multiple initial sequence boundary probability matrices corresponding to different text sources and representing the segmentation boundary probability of each character node; Step S2: Based on the boundary probability matrices of each initial sequence, construct a local topology network containing the associated edges between character nodes, compare the local topology network with the stored global prior network, identify the topological conflict subgraphs that do not satisfy the topological constraints of the global graph, and determine the boundary probability matrices of the initial sequences with topological conflicts as the conflict probability matrices. Step S3: Based on the topological conflict subgraph, trigger the boundary probability consistency check, calculate the vector similarity between the conflict probability matrices, and when the node temporal span constraint and step size reduction rule are satisfied, suppress the divergence tendency of the boundary probability of the conflict probability matrix in the iteration process, and fuse to generate the final conflict-free boundary probability rewrite matrix. Step S4: Overwrite the corresponding initial sequence boundary probability matrix with the final conflict-free boundary probability rewrite matrix, adjust the segmentation boundary probability of each character node, generate the corrected entity extraction sequence, and output the target entity extraction boundary of the multi-source digitized archival text.

2. The method for extracting key entities from digitized archival text according to claim 1, characterized in that, Step S2 includes the following sub-steps: Step S21, calculate the temporal span of the character nodes at both ends of each associated edge in the local topology network in the character stream; Step S22, when a temporal span of more than 18 nodes is detected and the corresponding structure confidence is less than 0.65, reset the topological weight of the associated edge to 0 to filter out long-distance interfering network edges and identify the topological conflict subgraph that does not satisfy the global graph topology constraints.

3. The method for extracting key entities from digitized archival text according to claim 1, characterized in that, In step S3, triggering the boundary probability consistency check includes the following sub-steps: Step S31, when the same character node generates multiple mutually exclusive conflict probability matrices in a multi-source heterogeneous text stream, the vector similarity between the conflict probability matrices is calculated using the following formula: ,in, The vector similarity between the conflict probability matrices. Let S32 be the directional angle of the conflict probability matrix in the feature space; Step S32: When the calculated vector similarity is less than 0.35, extract the node degree centrality value of each local topology network in the global prior network to determine the arbitration priority of the conflict probability matrix.

4. The method for extracting key entities from digitized archival text according to claim 1, characterized in that, In step S3, the step size reduction rule includes the following sub-steps: Step S33, monitor the monotonic variation of the probability correction value in the final conflict-free boundary probability rewriting matrix in real time; Step S34, when the calculated monotonic variation is continuously less than 0.05, reduce the adjustment step size of the probability correction value by 50% to suppress the iterative divergence of the boundary probability and complete the probability convergence within 5 iterations.

5. The method for extracting key entities from digitized archival text according to claim 3, characterized in that, After step S32, triggering the boundary probability consistency check also includes the following sub-steps: Step S35, assigning normalized weights to multiple mutually exclusive conflict probability matrices based on the obtained node degree centrality values; Step S36, using the normalized weights to perform linear weighted fusion of the conflict probability matrices to generate a unique final conflict-free boundary probability rewriting matrix.

6. The method for extracting key entities from digitized archival text according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11, input the multi-source digitized archive text into the pre-stored sequence labeling model to obtain the temporal feature vector of each character node; Step S12, calculate the segmentation probability between adjacent character nodes based on the temporal feature vector, and construct a multi-dimensional probability distribution based on the segmentation probability to generate the initial sequence boundary probability matrix.

7. The method for extracting key entities from digitized archival text according to claim 1, characterized in that, In step S2, the construction of the global prior network includes the following sub-steps: Step S23, obtain the historical archive association graph; Step S24, calculate the node co-occurrence frequency in the historical archive association graph, filter out the association edges whose node co-occurrence frequency is lower than the preset stability threshold, and store the remaining topological association network as the global prior network.

8. The method for extracting key entities from digitized archival text according to claim 4, characterized in that, In step S34, completing probability convergence within 5 iterations includes the following sub-steps: Step S341, set up a loop counter, and increment the count value of the loop counter by 1 after each adjustment step reduction; Step S342, when the count value of the loop counter reaches 5 and the boundary probability has not converged, terminate the probability overwrite loop and lock the probability correction value of the current control cycle.

9. A method for extracting key entities from digitized archival text according to claim 1, characterized in that, The noun objects included in the revised entity extraction sequence are derived from the original structure of the archive, and the target entity extraction boundaries correspond to personal names, place names, organization names, historical event names, and archive file numbers.

10. A method for extracting key entities from digitized archival text according to claim 1, characterized in that, After outputting the target entity extraction boundary, the following steps are also included: Step S5, based on the target entity extraction boundary, extract key structured key-value pairs from the multi-source digital archive text, and write the key structured key-value pairs into the digital archive knowledge graph in the form of triples to update the graph topology of the digital archive knowledge graph.

Images