A Method and System for Constructing Knowledge Networks and Analyzing Structural Deviations in Damaged Ancient Books

Abstract

This application provides a method and system for constructing knowledge networks and analyzing structural deviations in damaged ancient books, relating to the field of ancient book digitization technology. The method includes: acquiring at least one complete reference text and at least one real damaged text; performing the same unified text processing mapping on both and independently generating a complete reference network and a real damaged network respectively; using the complete reference network as a benchmark, performing degradation simulation using candidate degradation operators to generate a simulated degradation network; then, using the complete reference network as a common benchmark, calculating the structural deviation vectors of the real damaged network and the simulated degradation network respectively; obtaining the comprehensive fitting loss through the weighted distance between the two; and ranking and outputting the candidate degradation operators accordingly. This method can transform the knowledge network structural deviations caused by text damage into calculable, comparable, and verifiable quantitative results, facilitating the assessment of the impact of damage and the analysis of candidate degradation mechanisms.

Description

Technical Field

[0001] This application relates to the field of ancient book digitization technology, specifically to a method and system for constructing knowledge networks and analyzing structural deviations for damaged ancient books. Background Technology

[0002] Ancient books, manuscripts, bamboo and silk scrolls, and printed documents may suffer from insect damage, mold, breakage, loss of volumes, war damage, copying omissions, and later reconstruction during their long transmission. Traditional document collation usually focuses on missing characters, fragmented sentences, and the recovery of lost chapters, while digital processing mainly focuses on image enhancement, optical character recognition, handwritten text recognition, layout analysis, text completion, and fragment splicing. These technologies can restore or identify local content, but generally do not answer a higher-level question: what structural changes occur between the knowledge network constructed based on existing texts and the relatively complete reference network when some figures, places, and their narrative connections are no longer visible?

[0003] Existing schemes for constructing historical knowledge graphs and literary social networks typically include steps such as text preprocessing, entity recognition, name disambiguation, relation extraction or co-occurrence graph construction, graph database storage, and network analysis. For complete modern texts, subject-verb-object triples can be obtained through syntactic analysis or relation extraction models. For damaged ancient texts, missing characters, broken sentences, incomplete grammar, and variant characters increase the probability of parsing failure; therefore, local co-occurrence relations are often used to reduce dependence on complete syntactic structures. However, existing co-occurrence graph construction schemes usually only generate a network and perform centrality or community analysis, without treating text incompleteness as a simulable and measurable variable.

[0004] Specifically, the closest existing solutions can be roughly divided into four categories: The first category is the ancient book named entity recognition and knowledge graph construction solution, which usually performs text segmentation, entity recognition, alias alignment, co-occurrence statistics and stores it in a graph database. Its output is a knowledge graph, which cannot directly evaluate the structural deviation caused by the incomplete underlying text. The second category is the ancient book image restoration, text recognition and content completion solution, which mainly optimizes the character-level recognition rate or text restoration quality, but lacks the step of further mapping the preserved text into an entity relationship network and calculating the topological deviation. The third category is the complex network robustness and attack simulation solution, which usually takes a structured network as input, deletes hubs and edges by degree or betweenness and observes connectivity indicators, does not retain the positional evidence of relations in the original text, and does not include ancient book entity extraction and cross-text alignment. The fourth category is the incomplete knowledge graph or missing network comparison solution, which usually calculates node cover, edge cover or graph distance, but generally does not construct text position-driven degradation operators for damaged ancient books, nor does it build graphs independently for complete reference text and real incomplete text through the same processing flow.

[0005] The aforementioned approaches suffer from several common problems: Firstly, because they typically generate only a single network, they cannot distinguish between existing observational differences in real text materials and simulated effects caused by artificial degradation algorithms. Secondly, due to the lack of a scale-matched baseline, differences between candidate mechanisms may primarily stem from the number of deletions rather than structural selection rules. Thirdly, because co-occurrence anchor points are not recorded, missing sections or fragments can only be roughly approximated as random edge deletions. Fourthly, because the entity extraction process is tied to specific tools, the network node and edge sets change after model replacement, making it difficult to verify the stability of the conclusions. Fifthly, due to the single evaluation metric, two networks with similar numbers of nodes may be completely different in their connectivity structure and central entities. These shortcomings make it difficult for existing technologies to generate repeatable and interpretable assessment results of residual impact. Summary of the Invention

[0006] The purpose of this application is to provide a method, system, electronic device and storage medium for knowledge network construction and structural deviation analysis of damaged ancient books, which can transform the network structural deviation caused by text incompleteness into calculable, comparable and verifiable quantitative results.

[0007] In a first aspect, embodiments of this application provide a method for constructing knowledge networks and analyzing structural deviations in damaged ancient books, including: Obtain at least one complete reference text and at least one actual incomplete text; The same unified text processing mapping is performed on the complete reference text and the real incomplete text to generate complete reference networks and real incomplete networks independently, respectively. The unified text processing mapping includes text preprocessing, candidate entity generation and context determination, name normalization and alias normalization, entity mention location recording, and weighted knowledge network construction based on local co-occurrence, and records the original text location anchor point corresponding to each co-occurrence relationship. Using the complete reference network as a benchmark, one or more candidate degradation operators are used to simulate the degradation of the complete reference network to generate a simulated degradation network; the candidate degradation operators are used to simulate the impact of text damage on the knowledge network structure. Using the complete reference network as a common benchmark, the first deviation vector of the real incomplete network relative to the complete reference network and the second deviation vector of the simulated degenerate network relative to the complete reference network are calculated respectively; the deviation vectors contain multiple structural deviation indices. Calculate the weighted distance between the second deviation vector and the first deviation vector, and use it as the overall fitting loss; Based on the magnitude of the comprehensive fitting loss, the candidate degradation operators are sorted, and the sorting results are output.

[0008] Secondly, embodiments of this application provide a knowledge network construction and structural deviation analysis system for damaged ancient books, including: The text acquisition module is used to acquire at least one complete reference text and at least one real incomplete text. The knowledge network construction module is used to perform the same unified text processing mapping on the complete reference text and the real incomplete text, and to independently generate the complete reference network and the real incomplete network respectively; the unified text processing mapping includes text preprocessing, candidate entity generation and context determination, name normalization and alias normalization, entity mention position recording and weighted knowledge network construction steps based on local co-occurrence, and records the original text position anchor point corresponding to each co-occurrence relationship; The degradation simulation module is used to simulate the degradation of the complete reference network using one or more candidate degradation operators as a benchmark, thereby generating a simulated degradation network; the candidate degradation operators are used to simulate the impact of text damage on the knowledge network structure. The structural deviation assessment module is used to calculate, with the complete reference network as a common benchmark, a first deviation vector of the real incomplete network relative to the complete reference network, and a second deviation vector of the simulated degraded network relative to the complete reference network; the deviation vector includes multiple structural deviation indices. The fitting and result output module is used to calculate the weighted distance between the second deviation vector and the first deviation vector as the comprehensive fitting loss; according to the magnitude of the comprehensive fitting loss, the candidate degradation operators are sorted and the sorting results are output.

[0009] Thirdly, this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in the first aspect.

[0010] Fourthly, this application provides a computer-readable storage medium having computer-readable instructions stored thereon, which can be executed by a processor to implement the method described in the first aspect.

[0011] Compared to existing technologies, the method for knowledge network construction and structural deviation analysis of damaged ancient books provided in this application obtains at least one complete reference text and at least one real damaged text. The same unified text processing mapping is performed on both, and a complete reference network and a real damaged network are generated independently respectively. Using the complete reference network as a benchmark, a simulated degradation network is generated by employing candidate degradation operators. Then, using the complete reference network as a common benchmark, the structural deviation vectors of the real damaged network and the simulated degradation network are calculated separately. The comprehensive fitting loss is obtained through the weighted distance between the two, and the candidate degradation operators are ranked and output accordingly. This method can transform the knowledge network structural deviation caused by text damage into a calculable, comparable, and verifiable quantitative result, facilitating the assessment of the impact of damage and the analysis of candidate degradation mechanisms. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This document illustrates a flowchart of a method for constructing a knowledge network and analyzing structural deviations for damaged ancient books, provided in an embodiment of this application. Figure 2 This illustration shows a schematic diagram of the overall processing framework of the complete reference network, the real incomplete network, and the simulated degenerate network provided in the embodiments of this application; Figure 3 This illustration shows a schematic diagram of the entity processing and knowledge network construction process provided in an embodiment of this application; Figure 4 This illustration shows a schematic diagram of a knowledge network construction and structural deviation analysis system for damaged ancient books provided in an embodiment of this application. Figure 5 A schematic diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0014] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0015] Furthermore, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order. Additionally, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to those processes, methods, products, or apparatuses.

[0016] To facilitate understanding, the main symbols and data objects used below will be explained first: : Full reference text; "ref" means reference.

[0017] : Real incomplete text; "frag" stands for fragmentary.

[0018] : Unified text processing mapping; η is the set of parameters for preprocessing, entity extraction, normalization, and graph construction.

[0019] : Fully referenced from the internet.

[0020] : A real, incomplete network.

[0021] The simulated degenerate network obtained under the k-th candidate mechanism, parameter Θ, and random seed s.

[0022] V, E, and w represent the set of nodes, the set of edges, and the edge weight function of the network, respectively.

[0023] u, v: Two different normalized entity nodes.

[0024] : The set of starting positions of entity u in the entire text.

[0025] The position of the first character mentioned in entity u.

[0026] W: Width of the local co-occurrence window.

[0027] : Indicator function, returns 1 if the condition is true, otherwise returns 0.

[0028] : The set of anchor points for the co-occurrence of entities u and v.

[0029] : The k-th class of degenerate operators or combinations of operators.

[0030] : These are the total parameter vector and the parameter sub-vector of the k-th mechanism, respectively.

[0031] : These represent a single random seed and a set of random seeds, respectively.

[0032] These represent node retention rate, edge retention rate, and relative total edge weight, respectively.

[0033] Jensen–Shannon divergence of degree distribution.

[0034] : These represent the proportion of nodes in the largest connected subgraph and the proportion of isolated nodes, respectively.

[0035] : The change in connectivity ratio of the network to be compared relative to the reference network.

[0036] PageRank Top-K Entity Set Overlap Rate

[0037] : Mapping of the m-th structural deviation index.

[0038] : The non-negative weight of the m-th indicator.

[0039] : The combined fitting loss corresponding to parameter Θ and random seed s.

[0040] M: The set of indexes included in the overall fit loss.

[0041] N: Monte Carlo repetition count.

[0042] n m : These represent the target number of nodes and the target number of edges for scale matching, respectively.

[0043] ε, α, and γ: respectively, to prevent the denominator from being zero, the overall loss intensity, and the degree of dependence index.

[0044] The technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0045] Please refer to Figure 1 , Figure 1 This is a flowchart of a method for constructing a knowledge network and analyzing structural deviations for damaged ancient books, provided in an embodiment of this application, including the following steps: S101. Obtain at least one complete reference text and at least one real incomplete text.

[0046] Receive at least one relatively complete reference text and at least one real incomplete text The two types of texts should have comparable themes, events, people, or geographical scope, but are not required to be complete or incomplete versions of the same physical manuscript.

[0047] The system identifies each text record with a text identifier, version source, character encoding, chapter boundaries, page number or volume number, whether it is damaged, a description of the location of damage, and corpus authorization information. For multiple reference texts, they can be mapped separately and then merged, or they can be spliced ​​together into a reference corpus while keeping the chapter boundaries from being crossed.

[0048] The complete reference network and the real incomplete network are generated independently using the following formula: ; in, For complete reference network; For a real, incomplete network; For complete reference text; It is a real, incomplete text; The semicolon indicates a unified text processing mapping; the semicolon separates the input data from the configuration parameters; η is a set of unified configuration parameters, including cleaning rules, text block length, overlap length, candidate entity generator set, context judgment rules, alias mapping table, stop table, window width, weighting method, and minimum edge weight threshold; the equal sign indicates that the network is generated by the corresponding text through a unified mapping.

[0049] In one specific implementation, the complete reference text is selected from the complete biographical text related to a certain historical event in the *Old Book of Tang*, while the authentic incomplete text is selected from the corresponding damaged manuscript in the Dunhuang Bianwen (narrative poetry). The system assigns text identifiers T_ref_001 and T_frag_001 to the two texts respectively, recording the version source, character encoding (UTF-8), chapter boundaries, and descriptions of the damaged locations (the damaged manuscript marks 15 missing characters).

[0050] S102. Perform the same unified text processing mapping on the complete reference text and the real incomplete text to generate the complete reference network and the real incomplete network independently.

[0051] For ease of understanding, Figure 2 A schematic diagram of the overall processing framework is shown, comprising a complete reference network, a real incomplete network, and a simulated degenerate network. Figure 3 A schematic diagram of the entity processing and knowledge network construction process is shown.

[0052] The unified text processing mapping includes the steps of text preprocessing, candidate entity generation and context determination, name normalization and alias unification, entity mention position recording, and weighted knowledge network construction based on local co-occurrence, and records the original text position anchor corresponding to each co-occurrence relation.

[0053] S21: Text preprocessing The same character encoding unification, invisible character removal, layout noise elimination, optional variant character conversion, punctuation standardization and text boundary marking are performed on the above two types of texts. A "original position - processed position" mapping table is established for any deletion or replacement operation, so that entity mentions and co-occurrence relations can be traced back to the original text.

[0054] The text can be partitioned by a fixed number of characters, sentences / segments, or layout regions; overlapping areas are set for adjacent blocks (for example, partitioning every 400 characters with an overlap of 80 characters between adjacent blocks) to prevent entities from being truncated across blocks. Each text block at least stores the block number, start and end character positions, affiliated chapter, and original text hash value.

[0055] In a specific implementation, the biographies text of *Old Book of Tang* and the Dunhuang bianwen text are respectively subjected to UTF-8 encoding unification, whitespace character removal, and variant character conversion (e.g., converting "於" to "于"), and are partitioned by 400 characters with an overlap of 80 characters between adjacent blocks. The system establishes a mapping table between the original position and the processed position for each character, recording the original offset and the processed offset of each character.

[0056] S22: Candidate entity generation and context determination Candidate entity generation is not limited to a specific model. The system receives rule or dictionary matching results, statistical sequence labeling results, pre-trained named entity recognition results, large language model extraction results, external knowledge base recall results, or any combination thereof through a unified interface. Each candidate at least includes a surface string, candidate type, start and end positions, source model, confidence, and local context. Candidates generated from multiple sources are deduplicated or merged according to the text identifier, character interval, string, and candidate type: ; wherein, is the candidate entity set of text T; T is any input text; Q is the total number of connected candidate generators; q is the serial number of the candidate generator, with a value range from 1 to Q; represents the q-th candidate generator; represents the candidate set output by the generator for text T; ∪ represents the set union operation, which is used to merge candidates from various sources; the subscript q=1 and superscript Q represent the traversal range of the union; the parentheses represent function inputs.

[0057] For each candidate entity, extract local context of a predetermined length (e.g., 80 characters before and after). Determine whether to retain the entity and its type using a rule-based classifier, statistical classifier, neural network, large language model, or manual review interface. Identifiable entity types such as people, places, organizations, times, events, works, official titles, or other domains are defined in the configuration file. For candidates containing generic titles, pronouns, common nouns, abstract concepts, formatting remnants, or no valid context, set a deletion flag. The output of this step is an entity mention record with stable location coordinates and type labels.

[0058] In one specific implementation, two candidate generators are simultaneously integrated: an entity recognition statistical model and a large language model. The entity recognition statistical model identifies candidate entities such as "Taizong," "Li Shimin," "Chang'an," and "capital," while the large language model identifies entities such as "Tang Taizong" and "Xijing." After merging and deduplicating by character range, the system extracts 80 characters of context before and after each candidate. A rule classifier retains candidates of person and location types, while invalid candidates such as generic titles of official positions (e.g., "Duke" and "Marquis") and pronouns (e.g., "of" and "its") are deleted.

[0059] S23: Name normalization, aliasing and entity alignment Name normalization refers to mapping variant characters, abbreviations, courtesy names, pen names, posthumous titles, place name variants, and different word segmentation results to standardized names. Normalization can be achieved through a combination of precise alias tables, fuzzy string matching, contextual similarity, external knowledge base candidates, and manual verification. To prevent erroneous merging, each mapping saves the criteria type, confidence level, source, and whether it was manually verified. For candidates with the same name but different people, different entity identifiers can be generated based on era, location, co-occurrence neighborhood, or text context. ; in, This represents a name normalization function; c The surface string representing the candidate entity; x Auxiliary features representing candidate entities, such as local context, type, and location; This represents a collection of alias dictionaries, authoritative knowledge bases, and rule bases; u The normalized identifier represents a unique entity; commas separate function inputs; the equals sign indicates that the candidate is mapped to an entity after normalization. .

[0060] In one specific implementation, the alias mapping table maps "Taizong", "Tang Taizong", and "Emperor Wen" to the standard name "Li Shimin"; and maps "Chang'an", "Xijing", and "Capital" to the standard name "Chang'an". For candidates with the same name but different people (such as two different "Li Mi"), they are distinguished by era information (early Tang and mid-Tang) and different entity identifiers are generated.

[0061] S24: Entity mention location record Record each canonical entity u The set of all starting positions mentioned throughout the text: ; in, Representing a canonical entity u The set of all starting positions of all mentions; u To regulate entities; Representing entities u No. i The position of the first character mentioned; i To mention the serial number; For entities u Total number of mentions in the text; curly braces indicate a set; This indicates that the middle position is omitted; subscripts 1, 2 and... Indicates a range of positional sequences. Positions can be represented by character offsets, page numbers-line numbers-character numbers, page coordinates, or timestamps.

[0062] The system synchronously records textual location evidence supporting the co-occurrence relationship when establishing it. For each edge (u,v), an anchor point set is established. Each anchor point can contain the reference positions of u and v, the start and end positions of the window, the chapter identifier, and the hash of the original text segment. This anchor point enables continuous interval erasure to directly delete relational evidence located within the missing interval, rather than randomly deleting edges indiscriminately.

[0063] In one specific implementation, the canonical entity "Li Shimin" appears 25 times in the entire text, and the starting positions of all 25 mentions are recorded, such as the 150th character, the 380th character, the 520th character, etc. The canonical entity "Chang'an" appears 18 times in the entire text. For the edge (Li Shimin, Chang'an), the system records the anchor point corresponding to each co-occurrence. For example, the first anchor point includes mention position u 150, mention position v 158, window start and end positions [100, 200], chapter identifier "Volume 1", and original text fragment hash.

[0064] S25: Construction of a weighted knowledge network based on local co-occurrence For two distinct canonical entities u and v, local co-occurrence is defined as the distance between the starting positions of any two mentions not exceeding the window width W within the same chapter or allowed context. An undirected weighted network is preferred, but directed networks can also be constructed based on text order, syntactic direction, or event roles. The system excludes entity self-loops and allows setting a minimum edge weight threshold. :

[0065] in, Let be the edge weight between entity u and entity v; u and v are two different entities; and , i and j are the number of times u and v are mentioned, respectively; i and j are the mention numbers of u and v, respectively. and These represent the starting positions mentioned; the minus sign indicates the difference between the two positions; the vertical bar |·| indicates the absolute value; W is the width of the co-occurrence window; ≤ indicates that the position distance is not greater than W; This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Accumulate all mentions of u and v respectively; u≠v indicates exclusion of self-loops.

[0066] ; Where G is the generated weighted knowledge network; V is the set of canonical entity nodes; E is the set of edges; w is the edge weight function; (u,v) is the edge connecting entities u and v; vertical lines It means "satisfied"; This represents the minimum edge weight threshold; ≥ indicates that the edge weight reaches the threshold; the semicolon is used to separate multiple constraints. If a directed relation is used, then... and They can exist separately; if an undirected relation is used, they are considered as the same edge.

[0067] This step can use pair weighting, window counting, segment counting, distance decay weighting, or semantic confidence weighting. Regardless of the weighting method used, the complete reference text and the actual incomplete text must use the same configuration.

[0068] Obtained independently through the aforementioned steps and In the technical implementation, an "independent mapping flag" should be set to prohibit the use of... Directly initialized to The nodes or edges are deleted only after a copy is created. This constraint ensures that the real incomplete network only reflects the entities and relationships observable in the real incomplete text.

[0069] In one specific implementation, the window width W is set to 50 characters, pair weighting is used, self-loops are excluded, and the minimum edge weight threshold is used. =2. Independent graph construction yielded... (120 nodes, 280 edges, total edge weight 520) and (68 nodes, 96 edges, total edge weight 185). The system has verified that the independent graph construction flag is in effect. Not from Obtained through deletion and modification.

[0070] S103. Using the complete reference network as a benchmark, perform degradation simulation on the complete reference network using one or more candidate degradation operators to generate a simulated degradation network.

[0071] The candidate degradation operator is used to simulate the impact of text damage on the knowledge network structure.

[0072] Candidate degeneracy operators only accept Its location anchor point, entity type, and centrality, and other additional attributes, output : ; in, The simulated degradation network generated for the k-th candidate degradation mechanism under parameters Θ and random seed s; For the k-th class of degenerate operators or combinations of operators; For a complete reference network; A is the set of co-occurrence anchor points for all edges; Θ is the parameter vector; s is the random seed; superscript Identify the simulation conditions.

[0073] Candidate degenerate operators can be executed individually or in combination according to a preset order. The system saves the operator name, operation object, input parameters, deleted or merged records, random seed, and output network hash for each execution.

[0074] The following eight basic operators are preferably configured, and each operator can be executed individually or in combination in a preset order: (1) Continuous interval erasure operator Select one or more consecutive intervals on the text position axis. Delete anchor points falling into the interval Evidence of co-occurrence. If the edge If some anchor points are retained, the edge weight is updated according to the remaining number of anchor points or the remaining proportion; if all anchor points are deleted, the edge is deleted. This operator is applicable to scroll fragments, bamboo slips, missing pages, and partially unidentifiable areas.

[0075] in, The edge weights after interval erasure; For the original edge weight; For the edge B is the set of all co-occurring anchor points; B is the text location range or set of ranges to be erased. Represents the difference between sets; This indicates the number of anchor points that did not fall within the erasure zone; The original number of anchor points is represented by the number of anchor points; the fraction represents the percentage of anchor points retained; the multiplication sign indicates the total number of anchor points. This indicates scaling by the retention ratio. If If the value is 0, then the edge should not be included in this operator or should be deleted directly.

[0076] In one specific implementation, a continuous interval B = [800, 1500] is set, representing the deletion of all text evidence between the 800th and 1500th characters. For the edge (Li Shimin, Chang'an), its anchor point set is {150, 380, 520, 790, 820, 1100, 1350, 1600}, where 820, 1100, and 1350 fall into the erasure interval and are deleted, leaving 5 anchor points. The edge weight is updated to 5 / 8 = 0.625 times the original weight.

[0077] (2) Independent random edge deletion operator ; in, For the edge Retained random variables; It indicates that the probability distribution follows a certain probability distribution; Bernoulli represents the Bernoulli distribution; Let this be the probability of edge deletion; Preserve the probability for the edge; The set of edges after deletion; It is the original edge set; Indicates belonging to; This indicates that the condition is met; This indicates that the edge is retained. This operator does not consider node degree, type, or community location, and can be used as a relation loss model without structural preferences.

[0078] In one specific implementation, the edge deletion probability is set. =0.50, for Each of the 280 edges is independently sampled using Bernoulli, and approximately 140 edges are deleted. The system records the set of deleted edges and the set of retained edges.

[0079] (3) Location Neighborhood Deletion Operator Select one or more target nodes from the nodes of type "Location". l The neighborhood is obtained according to the condition that the graph distance is no greater than the radius r. This function deletes nodes, edges, or only relationships related to the target location within the neighborhood. The target location can be randomly selected, or it can be selected based on the frequency of place names, spatial range, or manually specified. ; in, Location node The r-order neighborhood; The target location node; r is a non-negative integer network distance threshold; Let V be any node; V is the set of nodes; For network G arrive The shortest path distance; ≤ indicates the distance does not exceed r; set brackets and This represents the set of nodes that satisfy the conditions.

[0080] In one specific implementation, the target node "Chang'an" is selected, the radius r=2 is set, and all nodes and their associated edges that are no more than 2 hops away from "Chang'an" are deleted, resulting in the deletion of 23 nodes and 67 edges.

[0081] (4) Community center node deletion operator First, a community set is obtained using Louvain, Leiden, label propagation, spectral clustering, or other community discovery algorithms. Then, the central nodes in the target community are selected and deleted based on betweenness centrality, degree centrality, PageRank, or combined score.

[0082] In one specific implementation, the Louvain algorithm is used to... The nodes were divided into 3 communities, and the two nodes with the highest betweenness centrality were removed from each community, for a total of 6 nodes.

[0083] (5) Degree-dependent edge node loss operator

[0084] in, For nodes The deletion probability; min indicates taking the smaller of the two values ​​in parentheses; 1 is the upper bound of the probability; α is the overall churn intensity; For nodes The degree of the node is ε; ε is a positive number to prevent the denominator from being zero when the degree is 0; γ is the degree-dependent exponent; the exponent operation indicates that the higher the degree of the node, the lower the probability of deletion is usually; parentheses are used to limit the denominator and the input of the min function. By adjusting γ, the priority of edge nodes can be controlled. When γ=0, it degenerates into a probability deletion that is independent of degree.

[0085] In one specific implementation, α=0.5, γ=1.2, and ε=1. For an edge node with a degree of 1, the deletion probability is min(1, 0.5 / 2^1.2)=0.218; for a center node with a degree of 10, the deletion probability is min(1, 0.5 / 11^1.2)=0.027. The deletion probability of edge nodes is much higher than that of center nodes.

[0086] (6) Node merging operator Select the source node from low-degree nodes, nodes with similar names, or nodes with similar semantics. Select the target node from high centrality nodes or canonical entities. ,Will Adjacent edges and edge weights are transferred , and delete .like If an edge already exists between two neighbors, the edge weights are either added together or merged according to preset rules. ; in, The source node to be merged; is the target node for receiving the relationship; x is any node adjacent to the source node or the target node; The boundary rights after merging; The original edge weight of the target node is used; if it does not exist, it is set to 0. ρ is the edge weight between the source node and x; ρ is the edge weight transfer coefficient, which can be 0 to 1 or set according to business rules; the plus sign indicates merging the relationship strength of the two parts.

[0087] In one specific implementation, the source node "Li Er" (degree=3) and the target node "Li Shimin" (degree=28) are selected, and ρ=1 is set. The three edges of "Li Er" are transferred to "Li Shimin". If "Li Shimin" already has an edge weight of 5 with "Wang", and "Li Er" has an edge weight of 1 with "Wang", then the edge weight after merging is 6.

[0088] (7) PageRank hierarchical filtering operator Nodes are divided into a central layer and an outer layer based on their weighted PageRank values, and deletion probabilities are configured for edges inside the central layer, edges crossing layers, and edges inside the outer layer. , and Preferred satisfaction This makes it easier to lose external relationships. The stratification threshold can be determined by the top q% of PageRank rankings.

[0089] In one specific implementation, the stratification threshold is set to the top 20% of PageRank. =0.05, =0.20, =0.40. That is, the probability of deleting edges inside the core layer is 5%, cross-layer edges are 20%, and outer layer internal edges are 40%.

[0090] (8) Local perturbation operator of the central node Select one or more central nodes with the highest degree, betweenness, or PageRank, and delete a certain proportion of their associated edges, or change their local edge weights. The target node can be specified by the user or selected automatically.

[0091] In one specific implementation, the top 3 nodes in PageRank ("Li Shimin", "Changsun Wuji", "Wei Zheng") are automatically selected, and 20% of their associated edges are deleted.

[0092] S104. Using the complete reference network as a common benchmark, calculate the first deviation vector of the real incomplete network relative to the complete reference network, and the second deviation vector of the simulated degraded network relative to the complete reference network; the deviation vector includes multiple structural deviation indices.

[0093] For any real incomplete network or any simulated degenerate network, we will collectively refer to it as the network to be compared. =( , , All deviations are expressed in terms of... As a common benchmark.

[0094] The structural deviation index includes at least one or more of the following indicators: ; in, For real incomplete networks or simulated degenerate networks; The retention rate is used as a reference node. Use the edge retention rate as a reference. For relative total edge weight; and for The set of nodes and the set of edges; and For reference, the set of network nodes and the set of edges; Indicates intersection; Indicates the size of the set; e represents an edge; ∈ indicates belonging to; and Let e ​​represent the edge weight of edge e in the network to be compared and the reference network, respectively; This represents the summation of all edges in the set; the fraction bar represents the ratio. If the denominator is 0, the system outputs an undefined flag and skips the index.

[0095] ; in, denoted as Jensen–Shannon divergence; P is the normalized distribution of the reference network; Q is the normalized distribution of the network to be compared; M is the average distribution of P and Q. Kullback–Leibler divergence; symbol The parameter divider represents the two parameters distributed in the divergence function; 1 / 2 represents the weighting coefficients of the divergences in the two directions; the plus sign indicates summation; P+Q represents dimension-wise addition, and dividing by 2 indicates averaging. To handle the zero probability term, a smoothing constant can be added and the function renormalized.

[0096] ; in, The proportion of nodes in the largest connected subgraph of network G; Let G be the set of nodes of the largest connected subgraph in G; The proportion of isolated nodes; This is the set of isolated nodes with a degree of 0. Let G be the set of all nodes; Indicates the number of nodes. For directed networks, weakly connected components or strongly connected components can be configured.

[0097] ; Where Δ represents the change of the network to be compared relative to the reference network; For the maximum connected subgraph scale change; The proportion of isolated nodes changes; the minus sign indicates that... The difference is taken as the baseline; the meanings of the other symbols are the same as in the above formula.

[0098] ; in, The overlap rate of the PageRank Top-K sets; Let K be the set of the top K nodes in network G, sorted by PageRank value; K is a preset rank cutoff value, such as 50; K This is the actual usable cutoff value; min indicates taking the minimum value among K, the number of reference nodes, and the number of nodes to be compared. Let K be the set of nodes in the networks to be compared; let K be the intersection of the two networks that both networks share a common entry into the top K; and let K be the denominator. Used for normalization when there are fewer than K nodes.

[0099] Optional supplementary metrics include Jaccard similarity between node and edge sets, normalized Laplacian spectral distance, modularity change, mean absolute rank drift, maximum rank drift, Spearman correlation coefficient, and Kendall correlation coefficient. These supplementary metrics can be used for reporting purposes or incorporated into the fitting loss based on application requirements.

[0100] Based on the aforementioned structural deviation indices, the first deviation vector of the real incomplete network relative to the complete reference network can be obtained. The second deviation vector of the simulated degenerate network relative to the complete reference network .

[0101] ; in, This represents the actual observation bias vector; For parameters The simulated deviation vector corresponding to the random seed s; Φ is the structural deviation index mapping, whose output can sequentially include , , , , , and .

[0102] S105. Calculate the weighted distance between the second deviation vector and the first deviation vector as the comprehensive fitting loss.

[0103] The overall fitting loss is calculated using the following formula: ; in, Let be the parameter vector of the candidate degeneracy operator. Use a random seed. M To include the set of indexes in the fit, For the first m The non-negative weight of each indicator For the first m Item structure deviation index function, For the complete reference network, This refers to the actual incomplete network. The simulated degradation network is defined as follows: the smaller the comprehensive fitting loss, the closer the degradation mode of the candidate degradation operator is to the actual residual mode.

[0104] The following is a weight vector of a specific embodiment: ; in, , , , , , and These are the weights corresponding to node retention rate, edge retention rate, relative total edge weight, degree distribution divergence, change in the proportion of the most connected subgraph, change in the proportion of isolated nodes, and PageRank Top-50 overlap rate, respectively. The weights can be fixed in advance or determined by the validation set, expert rules, or multi-objective optimization.

[0105] In one specific implementation, in order to eliminate randomness, the average overall fitting loss is calculated for multiple random seed simulations under the same parameter configuration according to the following formula:

[0106] in, The number of times is random. For the first i A random seed is generated, and the standard deviation or confidence interval of the comprehensive fitting loss is output simultaneously.

[0107] In one specific implementation, N=30 is set, and the "continuous interval erasure" operator is repeatedly simulated using 30 random seeds (s=1,2,...,30) to obtain 30 loss values. The average loss is then calculated. =0.18, standard deviation σ=0.03. The system also outputs a 95% confidence interval of [0.12, 0.24].

[0108] S106. Sort the candidate degradation operators according to the magnitude of the comprehensive fitting loss and output the sorting results.

[0109] The system generates candidate mechanisms in ascending order of average fitting loss and simultaneously outputs the difference relative to the scale-matched random null model: ; Where, Θ Configure the parameters that minimize the average fitting loss in the candidate parameter space; This indicates that the parameters that minimize the objective function are returned; Θ is any candidate parameter vector; Ω is a predefined candidate parameter space; ∈ indicates that Θ belongs to Ω; This represents the average loss. It only represents the best structure fit result under the given corpus, processing flow, indicators and parameter space, and does not directly prove the causal mechanism of real historical damage.

[0110] To prevent the structural fit from being misinterpreted as historical fact, the output results are divided into three levels of evidence: The first layer is and The differences in observed structures include network size statistics and the values ​​of various indicators of the true deviation vector; The second layer is the structural fit of the candidate mechanism to the observed differences, including the parameters of each candidate mechanism, the random seed for each simulation, the mean and fluctuation range of the simulation index, the overall fit loss and ranking. The third layer is a historical interpretation formed by combining version origins, fragment form, writing materials, physical testing, and collation evidence.

[0111] The system automatically completes only the first two layers by default; the third layer requires manual or multimodal evidence verification.

[0112] The results report should include at least the following: input text and version, entity extraction configuration, normalization rules, network size statistics, real observation bias, parameters for each candidate mechanism, random seed for each simulation, mean and fluctuation range of simulation indicators, overall fitting loss, candidate mechanism ranking, and anomaly logs.

[0113] In a preferred embodiment, before generating the simulated degradation network in step S103, a scale-matching baseline construction step is further included: The target number of nodes is determined based on the number of shared nodes between the complete reference network and the real incomplete network, and the target number of edges is determined based on the number of shared edges between the complete reference network and the real incomplete network. The deletion or merging operations of the candidate degeneracy operators are controlled so that the number of nodes and the number of edges of the generated simulated degenerate network approach the target number of nodes and the target number of edges, respectively.

[0114] Specifically, the number of target nodes and the number of target edges are calculated using the following formulas: ; Where, n m is the target number of nodes. Let the target number of edges be , and These are the sets of nodes for the complete reference network and the real incomplete network, respectively. and These are the sets of edges from the two networks, respectively. Represents the intersection of sets; Indicates the number of elements in the set; asterisk This indicates that the scale matches the target rather than multiplication.

[0115] Size matching stochastic null model from Randomly select n 1 node and generate an induced subgraph; when the number of edges in the induced subgraph exceeds m Randomly delete redundant edges; when the number of edges is less than m In this case, you can choose not to add edges and record the approximate matching state, or recover candidate edges from the original reference network that will not introduce new nodes.

[0116] Under the constraints of the same number of target nodes and target edges, edge-priority loss baselines and hub-priority loss baselines can be established separately, so that the three have the same number of target nodes and target edges, and are used to compare the fitting effect with the candidate degradation operator.

[0117] In one specific implementation, There are 120 nodes and 280 edges. There are 68 nodes and 96 edges, with 55 shared nodes and 70 shared edges. Therefore, the target number of nodes, n, is set. = 55 ， Target number of edges m =70. The system controls the deletion intensity of each degradation operator, making the number of nodes in the simulated network approach 55 and the number of edges approach 70. The edge-first churn baseline is set from... In the middle, prioritize deleting nodes with the lowest degree until 55 remain, then delete weak relationships until the number of edges reaches 70. Prioritize deleting nodes with the highest degree from the hub and baseline, and record approximate matches when the number of edges is below 70.

[0118] In a preferred embodiment, for each candidate mechanism and parameter configuration, random seeds are sequentially read from the random seed set S, and the degradation operator is repeatedly executed to obtain a family of simulated degenerate networks. Operation logs and metric results are saved separately for each repetition. N can be set to 10, 30, 100, or greater depending on the network size and time requirements. ; in, This is the set of simulated degenerate networks corresponding to parameter Θ; To use the i-th random seed The generated simulated network; to There are N random seeds; N is the number of repetitions.

[0119] In one specific implementation, N=30 is set, and the combined (interval erasure + random edge deletion) operator is repeatedly simulated using 30 random seeds. The number of edges and nodes deleted in each simulation is slightly different. The average loss of the 30 results is 0.12, and the standard deviation is 0.02, indicating that the operator is stable.

[0120] The method for constructing knowledge networks and analyzing structural deviations for damaged ancient books provided in this application has the following beneficial effects: First, this invention does not perform static analysis on a single existing text network, but establishes a verifiable comparative closed loop through a complete reference network, a real incomplete network, and a simulated degradation network. This can separate the differences in real observations from the effects of artificial simulation, overcoming the shortcomings of existing schemes that only generate a single network and cannot distinguish between differences in text observations and simulated degradation effects.

[0121] Second, this invention associates co-occurring edges with anchor points of the original text location. Continuous interval erasure can delete specific relationship evidence and simultaneously adjust edge weights, which is closer to continuous damage forms such as missing pages, missing scroll segments, and scattered bamboo slips than simply randomly deleting edges.

[0122] Third, this invention uses the number of shared nodes and the number of shared edges to construct a scale matching baseline, so that the evaluation of candidate mechanisms is no longer dominated by the number of deletions, which is helpful in judging whether the structure selection rules provide additional explanatory power.

[0123] Fourth, this invention reduces randomness by using multiple random seeds for repeated simulation and complete log output, enabling the reproduction of experiments and the location of specific node, edge, and text evidence changes.

[0124] Fifth, this invention clearly distinguishes between structural fitting and historical causal interpretation, and can be combined with textual criticism, collation, fragment images, and material testing evidence to reduce the risk of over-interpretation of automated analysis results. It can be directly extended to damaged archives, inscriptions, bamboo slips, local chronicles, genealogies, manuscripts, and other incomplete cultural heritage texts, demonstrating strong engineering versatility.

[0125] In the above embodiments, a method for constructing and analyzing structural deviations in knowledge networks for damaged ancient books is provided. Correspondingly, this application also provides a system for constructing and analyzing structural deviations in knowledge networks for damaged ancient books. The system for constructing and analyzing structural deviations in knowledge networks for damaged ancient books provided in this application can implement the above-described method for constructing and analyzing structural deviations in knowledge networks for damaged ancient books. This system can be implemented through software, hardware, or a combination of both. For example, the system can include integrated or separate functional modules or units to perform the corresponding steps in the above methods.

[0126] Please refer to Figure 4 As shown, this application provides a knowledge network construction and structural deviation analysis system 10 for damaged ancient books, including: The text acquisition module 101 is used to acquire at least one complete reference text and at least one real incomplete text. The knowledge network construction module 102 is used to perform the same unified text processing mapping on the complete reference text and the real incomplete text, and to independently generate the complete reference network and the real incomplete network respectively; the unified text processing mapping includes text preprocessing, candidate entity generation and context determination, name normalization and alias normalization, entity mention position recording and weighted knowledge network construction steps based on local co-occurrence, and records the original text position anchor point corresponding to each co-occurrence relationship; The degradation simulation module 103 is used to simulate the degradation of the complete reference network using one or more candidate degradation operators as a benchmark, thereby generating a simulated degradation network; the candidate degradation operators are used to simulate the impact of text damage on the knowledge network structure. The structural deviation evaluation module 104 is used to calculate, with the complete reference network as a common benchmark, a first deviation vector of the real incomplete network relative to the complete reference network and a second deviation vector of the simulated degraded network relative to the complete reference network; the deviation vector includes multiple structural deviation indices. The fitting and result output module 105 is used to calculate the weighted distance between the second deviation vector and the first deviation vector as the comprehensive fitting loss; according to the magnitude of the comprehensive fitting loss, the candidate degradation operators are sorted and the sorting result is output.

[0127] In one possible implementation, the degradation simulation module 103 is further configured to perform a scale-matching baseline construction step before generating the simulated degradation network: The target number of nodes is determined based on the number of shared nodes between the complete reference network and the real incomplete network, and the target number of edges is determined based on the number of shared edges between the complete reference network and the real incomplete network. The deletion or merging operations of the candidate degeneracy operators are controlled so that the number of nodes and the number of edges of the generated simulated degenerate network approach the target number of nodes and the target number of edges, respectively.

[0128] The knowledge network construction and structural deviation analysis system for damaged ancient books provided in this application embodiment and the knowledge network construction and structural deviation analysis method for damaged ancient books provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0129] This application also provides an electronic device corresponding to the method provided in the foregoing embodiments. The electronic device can be an electronic device for a client, such as a mobile phone, laptop, tablet computer, desktop computer, etc., to execute the above-mentioned method for knowledge network construction and structural deviation analysis of damaged ancient books.

[0130] Please refer to Figure 5 This illustrates a schematic diagram of an electronic device provided by some embodiments of this application. For example... Figure 5 As shown, the electronic device 20 includes: a processor 200, a memory 201, a bus 202, and a communication interface 203. The processor 200, the communication interface 203, and the memory 201 are connected via the bus 202. The memory 201 stores a computer program that can run on the processor 200. When the processor 200 runs the computer program, it executes the knowledge network construction and structural deviation analysis method for damaged ancient books provided in any of the foregoing embodiments of this application.

[0131] The memory 201 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 203 (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0132] Bus 202 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. Memory 201 is used to store programs. After receiving an execution instruction, processor 200 executes the program. The knowledge network construction and structural deviation analysis method for damaged ancient books disclosed in any of the foregoing embodiments of this application can be applied to processor 200, or implemented by processor 200.

[0133] The processor 200 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 200 or by instructions in software form. The processor 200 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 201. The processor 200 reads the information in memory 201 and, in conjunction with its hardware, completes the steps of the above method.

[0134] The electronic device provided in this application embodiment and the knowledge network construction and structural deviation analysis method for damaged ancient books provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0135] This application also provides a computer-readable storage medium corresponding to the knowledge network construction and structural deviation analysis method for damaged ancient books provided in the foregoing embodiments, wherein a computer program (i.e., program product) is stored thereon. When the computer program is run by a processor, it executes the knowledge network construction and structural deviation analysis method for damaged ancient books provided in any of the foregoing embodiments.

[0136] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0137] The computer-readable storage medium provided in the above embodiments of this application and the knowledge network construction and structural deviation analysis method for damaged ancient books provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application.

Claims

1. A method for constructing knowledge networks and analyzing structural deviations in damaged ancient books, characterized in that, include: Obtain at least one complete reference text and at least one actual incomplete text; The same unified text processing mapping is performed on the complete reference text and the real incomplete text to generate complete reference networks and real incomplete networks independently, respectively. The unified text processing mapping includes text preprocessing, candidate entity generation and context determination, name normalization and alias normalization, entity mention location recording, and weighted knowledge network construction based on local co-occurrence, and records the original text location anchor point corresponding to each co-occurrence relationship. Using the complete reference network as a benchmark, one or more candidate degradation operators are used to simulate the degradation of the complete reference network to generate a simulated degradation network; the candidate degradation operators are used to simulate the impact of text damage on the knowledge network structure. Using the complete reference network as a common benchmark, the first deviation vector of the real incomplete network relative to the complete reference network and the second deviation vector of the simulated degenerate network relative to the complete reference network are calculated respectively; the deviation vectors contain multiple structural deviation indices. Calculate the weighted distance between the second deviation vector and the first deviation vector, and use it as the overall fitting loss; Based on the magnitude of the comprehensive fitting loss, the candidate degradation operators are sorted, and the sorting results are output.

2. The method according to claim 1, characterized in that, The candidate degradation operators include at least one or more of the following basic operators, each of which can be executed individually or in combination in a preset order: Continuous interval erasure operator: Select one or more continuous intervals on the text position axis, delete all co-occurrence evidence where the co-occurrence anchor point falls within the interval, and update the edge weight of the corresponding edge proportionally according to the number of remaining anchor points; Independent random edge deletion operator: Each edge in the complete reference network is independently sampled using a preset edge deletion probability, and the edge whose sampled value is deleted is the edge to be deleted. Location Neighborhood Deletion Operator: Select one or more target nodes from nodes of type "location", determine the neighborhood according to the condition that the network distance is not greater than a preset radius, and delete the nodes, edges or only the relationships related to the target location in the neighborhood; Community center node deletion operator: The complete reference network is divided into communities using a community discovery algorithm, and a center node is selected and deleted in each community based on the centrality index; Degree-dependent edge node churn operator: Calculates the deletion probability based on the node's degree, where the deletion probability is inversely proportional to the node's degree; Node merge operator: Select a source node from low-degree nodes or nodes with similar names, select a target node from high-centrality nodes or canonical entities, transfer the adjacent edges and edge weights of the source node to the target node, and delete the source node; PageRank hierarchical filtering operator: Divides nodes into a central layer and an outer layer based on weighted PageRank values, and configures different deletion probabilities for edges inside the central layer, edges across layers, and edges inside the outer layer. Local perturbation operator for central nodes: Select one or more nodes with the highest centrality, delete their associated edges at a preset ratio, or modify their local edge weights.

3. The method according to claim 2, characterized in that, In the degree-dependent edge node loss operator, the node The probability of deletion is calculated using the following formula: ； in, For nodes The probability of deletion, For the overall loss intensity, For nodes The degree, To prevent positive numbers with a denominator of zero, The degree of dependence index.

4. The method according to claim 2, characterized in that, In the node merge operator, the source node is merged. The edge weight is transferred to the target node. When merging, the edge weights are calculated as follows: ； in, The source node to be merged, For the target node of the receiving relationship, Any node adjacent to the source node or the target node. For the merged border rights, The original edge weights of the target node, For the source node and The boundary rights between them This represents the edge weight transfer coefficient.

5. The method according to claim 1, characterized in that, Before generating the simulated degenerate network, a scale-matched baseline construction step is also included: The target number of nodes is determined based on the number of shared nodes between the complete reference network and the real incomplete network, and the target number of edges is determined based on the number of shared edges between the complete reference network and the real incomplete network. The deletion or merging operations of the candidate degeneracy operators are controlled so that the number of nodes and the number of edges of the generated simulated degenerate network approach the target number of nodes and the target number of edges, respectively.

6. The method according to claim 1, characterized in that, The structural deviation index includes at least one or more of the following indicators: Node retention rate is used to measure the proportion of nodes shared by the network under comparison and the complete reference network to the total number of nodes in the complete reference network. Edge retention rate is used to measure the proportion of edges shared by the network under comparison and the complete reference network to the total number of edges in the complete reference network. Relative total edge weight is used to measure the ratio of the total edge weight of the network to be compared to the total edge weight of the complete reference network; The Jensen-Shannon divergence of the degree distribution is used to measure the difference in degree distribution between the network to be compared and the complete reference network; The change in the proportion of the largest connected subgraph is used to measure the amount of change in the proportion of the largest connected subgraph nodes of the network to be compared relative to the complete reference network. The change in the proportion of isolated nodes is used to measure the amount of change in the proportion of isolated nodes in the network under comparison relative to the complete reference network. PageRank Top-K set overlap rate is used to measure the degree of overlap between the network being compared and the set of the top K core nodes in the complete reference network.

7. The method according to claim 1, characterized in that, The overall fitting loss is calculated according to the following formula: ； in, Let be the parameter vector of the candidate degeneracy operator. Use a random seed. M To include the set of indexes in the fit, For the first m The non-negative weight of each indicator For the first m Item structure deviation index function, For the complete reference network, This refers to the actual incomplete network. The simulated degradation network is defined as follows: the smaller the comprehensive fitting loss, the closer the degradation mode of the candidate degradation operator is to the actual residual mode.

8. The method according to claim 7, characterized in that, For multiple random seed simulations under the same parameter configuration, the average overall fitting loss is calculated according to the following formula: ； in, The number of times is random. For the first i A random seed is generated, and the standard deviation or confidence interval of the comprehensive fitting loss is output simultaneously.

9. A system for constructing knowledge networks and analyzing structural deviations in damaged ancient books, characterized in that, include: The text acquisition module is used to acquire at least one complete reference text and at least one real incomplete text. The knowledge network construction module is used to perform the same unified text processing mapping on the complete reference text and the real incomplete text, and to independently generate the complete reference network and the real incomplete network respectively; the unified text processing mapping includes text preprocessing, candidate entity generation and context determination, name normalization and alias normalization, entity mention position recording and weighted knowledge network construction steps based on local co-occurrence, and records the original text position anchor point corresponding to each co-occurrence relationship; The degradation simulation module is used to simulate the degradation of the complete reference network using one or more candidate degradation operators as a benchmark, thereby generating a simulated degradation network; the candidate degradation operators are used to simulate the impact of text damage on the knowledge network structure. The structural deviation assessment module is used to calculate, with the complete reference network as a common benchmark, a first deviation vector of the real incomplete network relative to the complete reference network, and a second deviation vector of the simulated degraded network relative to the complete reference network; the deviation vector includes multiple structural deviation indices. The fitting and result output module is used to calculate the weighted distance between the second deviation vector and the first deviation vector as the comprehensive fitting loss; according to the magnitude of the comprehensive fitting loss, the candidate degradation operators are sorted and the sorting results are output.

10. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method as claimed in any one of claims 1 to 8.

Images