A non-heritage knowledge graph construction method and system

Abstract

The application discloses a non-heritage knowledge graph construction method and system, relates to the technical field of multi-source text entity alignment, and provides the following scheme, which comprises the following steps: obtaining multi-source non-heritage text materials, extracting observation values of mentions of skills, patterns or processes on each characteristic position, recording no observation on the characteristic position for the mentions without extracted observation values, merging the mentions into variant nodes according to the containing and non-intersecting relationship between the mention sets corresponding to each observation value, and obtaining initial variant node division and undirected pedigree topology. According to the application, the identity determination between records is changed from the pair similarity of mention pairs to the explanation event count composed of the sum of edge events and observation events, and the variant node division, the undirected pedigree topology and the normalized state are taken as joint unknown quantities for common minimization, so that the problems of component attribution loss and difficulty in giving consideration to excessive merging and pedigree fragmentation after the characteristic differences are aggregated into a single similarity in the prior art entity alignment are solved.

Description

Technical Field

[0001] This invention relates to the field of multi-source text entity alignment technology, and more specifically, this application relates to a method and system for constructing an intangible cultural heritage knowledge graph. Background Technology

[0002] The digital organization of intangible cultural heritage knowledge in traditional fine arts usually relies on multiple sources of text, such as local chronicles, application texts for the list of representative intangible cultural heritage projects, museum collection records, and oral records of inheritors. These materials describe the names of techniques, the composition of patterns, materials and tools, the order of procedures, and regional schools. The same craft object may be described in multiple ways due to differences in the recording, local terminology, or the level of detail in the text. It may also form real branches due to continuous changes in the practice of apprenticeship. Therefore, the construction of intangible cultural heritage knowledge maps needs to distinguish between differences in records and changes in inheritance while merging duplicate records.

[0003] Existing knowledge graph construction methods generally first identify objects and attributes from text, then aggregate the word similarity, attribute overlap, contextual proximity or vector distance between two records into pairwise similarity, and complete entity merging with the help of threshold, clustering or global consistency constraints. The implicit premise of this type of method is that the differences in features between records can be aggregated into a single similarity without loss of judgment information, and the described object remains static during the data formation period, and whether two records point to the same object depends on the two records themselves.

[0004] However, in the living transmission of traditional art and intangible cultural heritage materials, the basis for distinguishing between recording errors and transmission changes lies in which records jointly hold the feature bits to which the differences belong and their values. However, the above-mentioned component attribution and joint holding relationships are no longer retained when aggregated into pairwise similarity. Adjusting the threshold, using hierarchical clustering or low-dimensional vector representation still affects the aggregated values. Introducing the age, lineage, or school level is limited by data gaps and circular dependencies. Thus, in scenarios where there are continuous and gradual changes along the transmission chain and multiple sources of variant texts coexist, relaxing the scale can easily cause multiple branches to be merged into a single graph node and lose the branch level. Tightening the scale can easily cause variant text records of the same object to be split into isolated nodes and lose the genealogical relationship between adjacent variants. Moreover, the two types of errors increase or decrease with the adjustment of the judgment scale. In addition, when the same craft element appears repeatedly in branches without direct lineage relationships, it is also easy to be mistakenly merged or discarded, making it difficult to form a corresponding recording structure in the knowledge graph. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper provides a method and system for constructing an intangible cultural heritage knowledge graph. This technical solution resolves the issues raised in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] Firstly, this application provides a method for constructing an intangible cultural heritage knowledge graph, the method comprising:

[0008] Obtain multi-source intangible cultural heritage text data, extract the observation values ​​of the mentions of techniques, patterns or crafts at each feature position, and record the mentions at the feature position as unobserved if no observation value is extracted.

[0009] Based on the inclusion and disjoint relationships between the mention sets corresponding to each observation value, the mentions are merged into variant nodes to obtain the initial variant node partition and undirected spectrum topology. The mode observation value of the mentions in the variant node is taken as the initial normalized state for each feature bit. If there is no observation, the normalized state is recorded as undefined.

[0010] Using variant node partitioning, undirected spectral topology, and normalized state as joint unknowns, the normalized state at both ends of each edge of the undirected spectral topology is compared with the feature bits. Feature bits that are not undefined and inconsistent are counted as edge events. The observed values ​​mentioned are compared with the normalized state of the variant node to which they belong. Observed and inconsistent feature bits are counted as observed events. The number of edge events and the number of observed events are summed to obtain the number of explained events.

[0011] Starting from the initial value of the joint unknown, state reassignment, mention reassignment, and topology reconnection are performed alternately. The state reassignment is to change the normalized state from the observed values ​​of each feature bit. The mention reassignment will reassign the mention to other or newly created variant nodes. The topology reconnection changes the access position of the variant node. Each operation only retains those that cause the count of explained events to decrease until the count no longer decreases in one round of traversal, thus obtaining the converged variant node partition, undirected spectral topology, and normalized state.

[0012] Statistically identify the connected components formed by variant nodes with the same non-undefined normalized state at the same feature position in the converged undirected spectral topology. When there are two or more connected components, establish cross-branch shared edges between variant nodes corresponding to different connected components.

[0013] By using the variant nodes and their normalized states in the converged variant node partition as graph nodes and node attributes, and the edges of the converged undirected spectral topology and cross-branch shared edges as graph edges, an intangible cultural heritage knowledge graph is constructed.

[0014] Secondly, this application provides a system for constructing an intangible cultural heritage knowledge graph, used to implement the aforementioned method for constructing an intangible cultural heritage knowledge graph, including:

[0015] The text observation module is used to acquire multi-source intangible cultural heritage text data, extract the observation values ​​of the mentions of techniques, patterns or crafts at each feature position, and record the mentions at the feature position as unobserved if no observation value is extracted.

[0016] The initial spectrum module is used to merge the mentions into variant nodes based on the inclusion and disjoint relationships between the mention sets corresponding to each observation value, thereby obtaining the initial variant node partition and undirected spectrum topology. The mode observation value of the mentions in the variant node is used as the initial normalized state for each feature bit. If there are no observations, the normalized state is marked as undefined.

[0017] The event counting module is used to compare the normalized states at both ends of each edge of the undirected spectral topology with variant node partitioning, undirected spectral topology and normalized state as joint unknowns. It counts non-undefined and inconsistent feature bits as edge events, and compares each mentioned observation value with the normalized state of its variant node. Observed and inconsistent feature bits are counted as observation events. The number of edge events and the number of observation events are summed to obtain the explanatory event count.

[0018] The alternating convergence module is used to alternately execute state reassignment, mention reassignment, and topology reconnection starting from the initial value of the joint unknown. The state reassignment is to change the normalized state from the observed values ​​of each feature bit. The mention reassignment will reassign the mention to other or newly created variant nodes. The topology reconnection changes the access position of the variant node. Each operation only retains those that cause the count of explained events to decrease until the count no longer decreases in one round of traversal, thus obtaining the converged variant node partition, undirected spectral topology, and normalized state.

[0019] The cross-branch sharing module is used to count the connected components formed by variant nodes with the same non-undefined normalized state on the same feature bit in the converged undirected spectral topology. When there are more than two connected components, cross-branch sharing edges are established between variant nodes corresponding to different connected components.

[0020] The graph generation module is used to construct an intangible cultural heritage knowledge graph by taking the variant nodes and their normalized states in the converged variant node partition as graph nodes and node attributes, and taking the edges of the converged undirected spectral topology and cross-branch shared edges as graph edges.

[0021] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described method for constructing an intangible cultural heritage knowledge graph.

[0022] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for constructing an intangible cultural heritage knowledge graph.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] This application solves the problems of component attribution loss and difficulty in balancing excessive merging and spectrum fragmentation after existing entity alignment aggregates various feature differences into a single similarity by changing the determination of identity between records from pairwise similarity of mention pairs to explanatory event count consisting of the sum of edge events and observed events, and by minimizing variant node partitioning, undirected spectrum topology and normalized state as joint unknowns. This allows multi-source variant texts that change continuously along the inheritance chain to retain branch relationships while merging duplicate records without constructing pairwise similarity matrices or setting a merging threshold with similarity as the independent variable.

[0025] This application solves the problem that endangered missing feature bits and rare stable feature bits are difficult to distinguish under a single observation and that the decision boundary depends on manually set hyperparameters by changing the contribution of each feature bit to the count from a fixed weight to the side event count coefficient and the observation event count coefficient generated by the joint ordering of observation density and cross-source recurrence rate. The two types of coefficients are written back with the independent occurrence degree and source concentration after the spectral convergence. This makes the decision scale driven by the residual of the spectral structure itself, rather than depending on the free weight coefficient.

[0026] This application solves the problem that existing methods often mistakenly merge or discard recurring process elements in branches without direct mentorship relationships, and lack corresponding structures to record them in the graph, by establishing cross-branch shared edges between variant nodes with the same feature bit value but not connected to each other. This enables such cross-branch repetitive relationships to be explicitly recorded as searchable and traceable graph edges. Attached Figure Description

[0027] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0028] Figure 1 This is a flowchart of a method for constructing an intangible cultural heritage knowledge graph proposed in this invention;

[0029] Figure 2 This is a structural block diagram of an intangible cultural heritage knowledge graph construction system proposed in this invention. Detailed Implementation

[0030] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0031] To facilitate understanding, the following example uses the compilation of multi-source intangible cultural heritage texts on a local paper-cutting technique as a consistent illustration. Assume that, focusing on the same category of paper-cutting, data was collected from four sources: local chronicles, application texts for the representative intangible cultural heritage list, museum catalogs, and oral accounts of inheritors. Several mentions of techniques, patterns, or processes were extracted. The same craft may appear in different versions across different sources due to variations in recording methods, local terminology, or level of detail. Furthermore, continuous modifications during apprenticeship practice may lead to actual branches. This consistent example demonstrates how each step is implemented sequentially on the same batch of data.

[0032] Before proceeding with each step, let's define two concepts that run throughout the text for direct reference in subsequent paragraphs: Unobserved is defined at the level of <mention, feature bit>, meaning that no observed value was extracted for a certain mention at a certain feature bit. This could be because the original text did not describe the attribute, or because the extraction component failed to locate it. Undefined is defined at the level of <variant node, feature bit>, meaning that all mentions within a variant node are unobserved at that feature bit, and therefore no value can be obtained. The two are at different levels and are not interchangeable: Unobserved is the missing detection of a single mention, while Undefined is the entire node having no value at that bit. Throughout the text, Unobserved does not participate in any consistency comparisons, and neither edge events nor observation events are counted. Undefined acts as a placeholder that does not generate conflicts during comparison, meaning that no edge events are generated when the undefined end is involved.

[0033] This embodiment is applicable to the digital organization of knowledge of intangible cultural heritage in the field of traditional fine arts, which is passed down in a living manner. Its typical application environment is: to construct an intangible cultural heritage knowledge graph that supports intelligent retrieval and inheritance analysis from multiple heterogeneous texts such as local chronicles, application texts for the representative list of intangible cultural heritage projects, museum collection records, and oral records of inheritors. The common feature of this type of work is that multiple records of the same craft object contain both recording errors and real inheritance changes along the apprenticeship chain. The two need to be distinguished in the same graph rather than being erased.

[0034] The initial input comes from the information extraction stage of the above multi-source text: the named entity recognition and relation extraction component (which can use a sequence labeling model in conjunction with rule post-processing) locates descriptive fragments such as technique names, pattern composition, materials and tools, process sequence, and regional schools from each text, registers them one by one as a mention, and fills in the observation value according to the pre-agreed feature position template; the feature position template is a set of structured fields, such as "main pattern motif", "border composition", "number of folds", "starting point of the cut", "paper material", "first step of the process", etc., and each field corresponds to a feature position; when a mention is explicitly described in the text at a certain feature position, its observation value is recorded, and when it is not described, the position is left blank.

[0035] Reference Figure 1 As shown, this application proposes a method for constructing an intangible cultural heritage knowledge graph:

[0036] Obtain multi-source intangible cultural heritage text data, extract the observation values ​​of the mentions of techniques, patterns or crafts at each feature position, and record the mentions at the feature position as unobserved if no observation value is extracted.

[0037] It should be noted that the observed values ​​come from the normalized output of the mention description fragment by the information extraction component: for the same feature position, a controlled value table is first established (e.g., the values ​​under "main pattern motif" can be "floral pattern", "double happiness character", "fish pattern", etc.), and then the extracted free text is mapped to the controlled values ​​through synonym merging. When the mapping is uncertain, the original text is retained and merged into the controlled value table for manual review; the mapped controlled value is the observed value of the mention at that feature position.

[0038] Based on the inclusion and disjoint relationships between the mention sets corresponding to each observation value, the mentions are merged into variant nodes to obtain the initial variant node partition and undirected spectrum topology. The mode observation value of the mentions in the variant node is taken as the initial normalized state for each feature bit. If there is no observation, the normalized state is recorded as undefined.

[0039] It should be noted that a variant node is a set of references formed by merging the reference sets corresponding to each observation value in this step based on the inclusion and disjoint relationships. Its meaning is entirely determined by this merging process. Semantically, it corresponds to the same process variant, but its determination does not depend on any prior definition of the process variant. Variant node partitioning means all references are assigned to each variant node. Undirected phylogenetic topology is an undirected graph with variant nodes as vertices and direct adjacency relationships between variant nodes as edges. It is used to carry possible inheritance relationships between adjacent variants without specifying a direction. Normalized state is the representative value assigned to a variant node at a certain feature bit, which serves as the value for that node to participate in external comparisons at that bit.

[0040] The initial normalized state is obtained by statistical analysis of each feature bit: for a variant node, all observed values ​​mentioned in that feature bit are collected, and the observation value with the highest frequency is taken as the normalized state of that bit. Unobserved values ​​are not included in the frequency statistics. If all mentions in that bit are unobserved, it is recorded as undefined. The strict distinction between unobserved mentions at the mention level and undefined mentions at the node level is to avoid treating the omission of a single mention as the absence of a value for the entire node at that bit, thus incorrectly causing edge events to be ignored and the differences between adjacent branches to be smoothed out.

[0041] Additionally, when taking the mode for each feature bit, if the most frequent observation value is not unique, the state of that bit is not fixed by taking the first occurrence or randomly selecting one, but the normalized state is directly recorded as undefined and left to be redefined by subsequent states, so as to ensure that no human decision independent of the count of the explained events is introduced during the initialization phase.

[0042] Using variant node partitioning, undirected spectral topology, and normalized state as joint unknowns, the normalized state at both ends of each edge of the undirected spectral topology is compared with the feature bits. Feature bits that are not undefined and inconsistent are counted as edge events. The observed values ​​mentioned are compared with the normalized state of the variant node to which they belong. Observed and inconsistent feature bits are counted as observed events. The number of edge events and the number of observed events are summed to obtain the number of explained events.

[0043] It should be noted that the process of calculating the explanatory event count is as follows: For each edge of the undirected spectral topology, the normalized states of the two variant nodes at each feature position are compared. An edge event is counted only if both ends are not undefined at that position and their values ​​are different (if either end is undefined, no edge event is generated at that position). For each mention, its observed value is compared with the normalized state of the variant node at that position. An observation event is counted only if the mention has been observed at that position and its value is different from the normalized state. The total number of edge events is directly added to the total number of observation events, and the resulting non-negative integer is the explanatory event count. The variant node partitioning, the undirected spectral topology, and the normalized state are adjusted together in subsequent steps, rather than fixing one of them first and then calculating the others.

[0044] It should be further explained that although edge events and observed events are counted at the lineage edge and reference levels respectively, they are both necessary explanations for a single inconsistency. Both are dimensionless frequency measurements with non-negative integer ranges, thus they are homogeneous counts and can be directly added: each observed inconsistency is either explained as a transmission change along the lineage edge and counted as an edge event, or explained as a recording bias and counted as an observed event. The sum of these two is the total number of hypotheses needed to explain all observations. Because they are in the same count and compete with each other—that is, the division tends to merge, so the observed events increase, and the division tends to separate, so the edge events increase—the observed events increase. The rise, excessive merging and phylogenetic fragmentation can be simultaneously constrained by the same criterion, without the need to set an adjustable relative weight for each of these two types of events; thus, the distinction between recording error and inheritance change is not given by any pre-defined boundary, but is endogenously given by the attribution method that minimizes the number of explanatory events, that is, if a difference can be carried by a few phylogenetic edges, it is classified as inheritance change, and if it can only exist as a deviation from the normalized state of its variant node, it is classified as recording error; based on this, the embodiment provides an explanation that requires the fewest assumptions under a given observation, and its feasibility and beneficial effects are not based on the premise of restoring historical truth.

[0045] Starting from the initial value of the joint unknown, state reassignment, mention reassignment, and topology reconnection are performed alternately. Each operation retains only those that decrease the count of explained events until the count no longer decreases in one round of traversal, resulting in a converged variant node partition, undirected spectral topology, and normalized state.

[0046] It should be noted that state reassignment is the process of changing the normalized state of a certain feature bit of a variant node under the premise of fixed partitioning and topology; mention reassignment is the process of reassigning a mention from the current variant node to an adjacent or newly created variant node; topology reconnection is the process of removing a variant node along with its subordinate branches from the current connection position and connecting it to another location; alternating execution will inevitably terminate: the event count is interpreted as a non-negative integer, and each accepted operation strictly decreases it. Strictly decreasing non-negative integers cannot be performed indefinitely, so the count stops decreasing after one round of traversal (trying the three types of operations in turn), and the process terminates.

[0047] Additionally, it should be noted that the joint minimization of the three types of operations is computationally difficult to obtain the global optimum precisely in polynomial time. Therefore, this embodiment adopts an alternating descent approximation solution, and the result is referred to as "converged" rather than "optimal". The feasibility and beneficial effects of this embodiment do not depend on obtaining the global optimum, but on changing the judgment object from the similarity of mention pairs to the spectrum interpretation cost, and the replacement of the basic term type of the objective function. In order to reduce the risk of getting trapped in bad local solutions, a multi-starting point strategy can be adopted in the implementation: the aforementioned mode initialization is the main starting point, and another random selection of a tie feature bit is used to form a faulty starting point. After each starting point descents independently and alternately, the result with the smallest number of interpretation events is taken as the convergent result.

[0048] Statistically identify the connected components formed by variant nodes with the same non-undefined normalized state at the same feature position in the converged undirected spectral topology. When there are two or more connected components, establish cross-branch shared edges between variant nodes corresponding to different connected components.

[0049] It should be noted that a cross-branch shared edge is a newly created edge between two variant nodes that have the same value at the same feature bit but are not connected in the convergent spectral topology. It is used to explicitly record the relationship of the same process element recurring in branches without a direct mentor-mentee relationship in the graph. The process of obtaining it is as follows: fix a certain feature bit, take out all variant nodes whose normalized state is a certain non-undefined value, and examine their connectivity only along the existing spectral edges of the convergent undirected spectral topology. If these nodes fall in two or more unconnected connected components, then cross-branch shared edges are established in pairs between variant nodes belonging to different connected components. If they all fall in the same connected component, then this value does not generate a cross-branch shared edge. Cross-branch shared edges are the output results that need to be retained in this scheme, so no penalty term is set for their quantity.

[0050] The variant nodes and their normalized states in the converged variant node partition are used as graph nodes and node attributes, and the edges and cross-branch shared edges of the converged undirected spectral topology are used as graph edges to construct an intangible cultural heritage knowledge graph.

[0051] It should be noted that the final output intangible cultural heritage knowledge graph is carried by a graph data structure, and its fields are composed of: a set of nodes, each node corresponding to a variant node, carrying node attributes, namely the list of identifiers mentioned by the node, the normalization status of each feature position (including those marked as undefined), and the set of data sources involved in the node; and a set of edges containing two types of edges, namely lineage edges (from convergent undirected lineage topology, labeled as "lineage" type) and cross-branch shared edges (labeled as "cross-branch shared" type and with its causal feature position attached).

[0052] For example, an output record can be: Node V7 {mention = [m12, m19, m23], main motif = floral pattern, number of folds = four folds, starting point of the cut = undefined, source = {local gazetteer, collection record}; edge e {type = lineage, endpoint = V7-V3}; edge g {type = cross-branch sharing, endpoint = V7-V15, causal feature = main motif: floral pattern}.

[0053] Through the above technical solution, this embodiment uses the sum of edge events and observation events for each feature bit as the unique criterion, and organizes the differences between records by combining the minimization of variant node partitioning, undirected spectral topology, and normalized state. This ensures that the consistency and inconsistency between records are implemented at specific feature bits and specific spectral edges rather than being compressed into a single value. Thus, while merging duplicate records, the differences and levels between branches are preserved, taking into account both the merging of duplicate records and the differentiation of real inheritance branches.

[0054] In an optional embodiment, mentions are merged into variant nodes based on the inclusion and disjointness relationships between the mention sets corresponding to each observation value, specifically including:

[0055] Each observation value is used to collect mentions with that observation value to obtain the support set for that observation value, and all mentions are combined to form the total support set;

[0056] It should be noted that the process of obtaining the support set is as follows: traverse all mentions with <feature bit, observed value> as the key, and include mentions that take the observed value at that feature bit into the same set. This set is the support set for this <feature bit, observed value>. Observers not at that feature bit are not included in any support set under that feature bit. The support set is a custom intermediate object in this step, which carries the component attribution information with the set relationship of "which mentions share a certain value", replacing pairwise similarity. The total support set is a set containing all mentions. Its function is to ensure that after merging it into the subsequent maximal hierarchical subfamily, any mention will be covered by at least one selected support set, thus ensuring that each mention can find a merging unit, and the initial undirected spectral topology is necessarily connected and does not produce isolated mentions.

[0057] Compare the inclusion, disjointness and intersection relationships of each support set, select a maximal hierarchical subfamily from the support set where any two support sets are only included or disjoint, and merge the total support set into the maximal hierarchical subfamily.

[0058] It should be noted that "layered" here refers to a family of sets satisfying the following condition: any two support sets either contain each other or are disjoint. The selection process is as follows: first, determine the relationship between each pair of support sets as either containment, disjointness, or intersection (containment means one set is a subset of the other, disjointness means the intersection is empty, and intersection means the intersection is non-empty and the sets do not contain each other); then, exclude the support sets corresponding to the observation values ​​that intersect with any other support set from the layered structure. No other support sets can be added to the remaining support sets without violating the layered property. The maxima of his support set constitute a maximal hierarchical subfamily; the support sets and observation values ​​of intersecting observation values ​​excluded from the hierarchical structure are all retained as is, without deletion or weight reduction, and participate in the counting of subsequent edge events and observation events as usual; because intersection only indicates that the value cannot coexist with other values ​​in the same spectral branch (lateral borrowing, parallel evolution, and shared motifs will all produce intersection), not noise. If it is removed or weighted at this stage, the relationship of the same element appearing repeatedly in different branches will not be identified as a cross-branch shared edge after convergence;

[0059] It should be noted that when the maximal subfamily satisfying the layered property is not unique, one is selected in a deterministic order: first, the one with the largest number of support sets is selected; if they are still tied, the one with the largest number of covered mentions is selected; if they are still tied, one is selected arbitrarily. The starting point only affects the search path of alternating descent, and the final decision is still made by comparing the descent of the explanatory event counts. Therefore, the juxtaposition does not affect feasibility.

[0060] The variant nodes are formed by merging the support sets containing the least number of mentions in the maximal hierarchical subfamily and connecting the variant nodes according to the direct inclusion relationship between the merged units to obtain the initial variant node partitioning and undirected spectral topology.

[0061] It should be noted that a merging unit refers to the finest-grained set of support selected for a particular mention. The process of obtaining it is as follows: for each mention, all support sets containing the mention are sifted out from the maximal subfamily, and the one with the fewest mentions is taken as the merging unit of the mention. The mentions contained in the same merging unit constitute a variant node, and all variant nodes constitute the initial variant node partition. The criterion for connecting variant nodes is the direct inclusion relationship between merging units: if merging unit U properly includes merging unit W and there is no third merging unit X between them such that U properly includes X and X properly includes W, then U and W are said to be directly included, and an undirected edge is connected between the corresponding two variant nodes, thus obtaining the initial undirected spectral topology. Since the total support set has been merged into the maximal subfamily and contains all mentions, its corresponding variant node must be located at the top of the inclusion relationship, ensuring topological connectivity.

[0062] Through the above technical solution, this embodiment uses the inclusion and disjoint relationships between support sets, and completes the initial merging and connection of the reference to the variant node through the maximal layered subfamily and the least covered merging unit. The intersection values ​​are retained as is, and only they are excluded from the layered structure. This allows the repeated appearance of the same element in different branches to be retained until it is subsequently identified as a cross-branch shared edge, providing a starting point for the initial partition that is both connected and does not lose the repetitive relationship.

[0063] In an optional embodiment, state resetting, mention reassignment, and topology reaccess are performed alternately, specifically including:

[0064] When resetting the execution state, fix the variant node partition and the undirected spectral topology, try to obtain the observed values ​​that have appeared at each feature bit, re-count the edge events of the associated edges of the variant node and the observation events mentioned therein, and when the minimum value of the sum of the two is less than the current value, replace the normalized state of the variant node at that feature bit with the corresponding observation value.

[0065] It should be noted that after fixing the partition and topology, each feature bit is independent of the others. The exact solution for the separable feature bits of the state re-determining involves two scans: Since the undirected spectral topology is not yet oriented at this point, before scanning, an arbitrary variant node is selected as the traversal reference point, and a traversal tree is generated using breadth-first or depth-first search. Based on this, the relative order of the leaves and the root is determined (the selection of the reference point only determines the scanning order and does not affect the obtained normalized state, because this subproblem has a unique minimum solution under the given partition and topology); the first scan proceeds from the leaves to the root, and each variant node collects the candidate state set uploaded by its child nodes and the state set mentioned by itself. The observed values ​​form a set of candidate states for the node at that feature position. In the second pass from root to leaf, for each variant node in the candidate state set, the value that minimizes the sum of (the number of edge events of the associated edges + the number of observed events mentioned within the node) at that feature position is selected as the normalized state. Undefined values ​​do not generate edge events for selection as agreed in the previous text. This subproblem can be solved precisely in time with a polynomial relationship to the scale, so state re-determination does not produce an "increment" and all optimal states can be obtained at once for all feature positions. The range of trials is limited to the observed values ​​that have appeared at that feature position to avoid introducing values ​​that have never appeared in the corpus.

[0066] When performing mention reassignment, mentions are tried to be assigned to adjacent variant nodes and newly created variant nodes respectively. The observed values ​​of the mention are used to form the normalized state of the newly created variant node. The observed events of the mention and the events of the removed and assigned variant nodes and their associated edges are recounted to obtain the increment of each candidate reassignment.

[0067] It should be noted that the candidate reassignment increment refers to the difference in the number of explained events before and after a trial assignment. A negative value indicates a decrease in the count and is an acceptable direction. Its local computability comes from the decomposition of events according to <mention × feature bit> and <feature bit × spectrum edge>: a reassignment only needs to recalculate the number of observed events of the mention itself, the normalized state of the two variant nodes that were removed and accepted due to changes in membership, and the number of edge events on the edges associated with these two nodes, without affecting other nodes. The target nodes of the trial assignment include the adjacent variant nodes of the current node in the undirected spectrum topology, and a newly created variant node with the observed values ​​of the mention itself as the normalized state. When a mention is unobserved on all feature bits, the normalized state of its newly created variant node is undefined, producing neither observed events nor edge events. Such a creation does not change the count but will meaninglessly increase empty nodes. Therefore, in practice, the trial assignment of newly created nodes is not performed for such unobserved mentions. This is a limitation on the search range and not a separate judgment set outside the number of explained events.

[0068] When performing topology reconnection, the variant node and its subordinate branches are pruned from the current position and the remaining variant nodes are tried to be connected. The edge events of the pruning position and the trial connection position are re-counted to obtain the increment of each candidate reconnection.

[0069] It should be noted that the process of obtaining the candidate re-access increment is as follows: the variant node to be moved, along with its entire sub-branch under its jurisdiction on the undirected spectral topology, is cut off from the current access edge, and the number of edge events resolved by the original associated edge at the cut position is recorded; then the access end of the sub-branch is tried to be connected to each of the remaining variant nodes one by one, and the number of edge events added by the new associated edge at each tried access position is recorded; the (number of newly added edge events - number of resolved edge events) is used as the candidate re-access increment; since both pruning and access only change a few local spectral edges, the recalculation of edge events is also limited to these edges.

[0070] When there are candidates for reassignment increment or candidate reaccess increment that are less than zero, retain the result corresponding to the smallest value, terminate the alternating execution, and shrink and merge adjacent variant nodes with the same set of defined feature bits and the same corresponding normalization state.

[0071] It should be noted that the order and termination of multipath operations are as follows: within one round of traversal, state reassignment, mention reassignment, and topology reconnection are performed sequentially; among all candidate increments generated by mention reassignment and topology reconnection, if there is one less than zero, only the result corresponding to the smallest value (the largest decrease) is adopted and the next round is entered; otherwise, this type of operation remains unchanged; when none of the three types of operations can make the count continue to decrease after one round of traversal, the process terminates; after termination, a contraction and merging operation is performed: adjacent variant nodes with the same set of defined feature bits and whose normalized states on these feature bits are consistent are merged into one;

[0072] It should be noted that the shrinking and merging is performed after minimization, without changing the explanatory event count or entering the cost function. Therefore, it must not be rewritten as a node number penalty term or a minimum description length regularization term, otherwise it will introduce a free weight coefficient, which contradicts the principle of this scheme that the determination does not depend on the parameter value. In addition, "the same set of defined feature bits" is a necessary condition, that is, feature bits that are defined on one side but not on the other side are not considered to be consistent in state, so as to avoid merging a large number of redundant nodes with undefined feature bits into adjacent nodes.

[0073] Through the above technical solution, this embodiment alternately decreases the local incremental comparison of two scans for accurate solution of state re-determination, mention reassignment and topology re-access, and after convergence, it uses shrinkage and merging to sort out redundant nodes without entering the cost function, so that each adjustment only needs to recalculate a few events directly related to it to evaluate the merits. While ensuring computational feasibility, it gradually reduces the number of explanatory events and sorts out variant node partitions with less redundancy.

[0074] In an optional embodiment, before comparing the normalized states at both ends of each edge of the undirected spectral system topology, a process for synthesizing composite feature bits is further included, specifically including:

[0075] First, it is necessary to explain its necessity and the problem it addresses: the independent counting of each feature bit implies the assumption that the changes of each feature bit are independent of each other. However, in traditional art, needlework and texture, materials and color sequence, process sequence and tools and finished product form are often coupled in pairs. A real inheritance drift will cause inconsistencies on several coupled feature bits at the same time and be counted as multiple side events. This deviation has a fixed direction, that is, the side events are systematically overestimated. Minimizing it is to suppress the overestimated side events and tend to over-merge, which will cause adjacent variants that should be distinguished to be merged into a single node. The conventional approach is to weight the relevant features or use principal components or embedding to reduce dimensionality, but both of these are aggregations that will destroy the component identities again. This solution does not aggregate, but changes the counting unit so that a drift is counted as only one side event.

[0076] In the initial variant node partitioning and undirected spectral topology, the number of times the same variant node is mentioned as being jointly observed and the number of times the normalized state changes synchronously at both ends of the same side are counted for each feature pair. Feature pairs whose two counts are not less than their median among all feature pairs are denoted as candidate coupled feature pairs.

[0077] It should be noted that candidate coupled feature pairs refer to two feature bits that have been screened by both co-occurrence and synchronous change and determined to be potentially coupled. The process of calculating the number of both items is as follows: For any feature pair, firstly, traverse all variant nodes and count the sum of the number of times the same reference in the same variant node is observed in both feature bits (i.e., neither of the two items is unobserved), which is taken as the number of co-observations; secondly, traverse each edge of the undirected spectrum topology and count the sum of the number of times the normalized state of both ends of the edge "changes simultaneously" in both feature bits (both ends are not undefined in the two items and take different values), which is taken as the number of synchronous changes; then, calculate the median of these two items in all feature pairs as a threshold. Only when the number of both items of a feature pair is not lower than their respective thresholds is it recorded as a candidate coupled feature pair.

[0078] The co-occurrence and synchronous change counts are all anchored to the initial variant node partitioning and undirected spectral topology for statistics, without using the convergence results that are still being solved, to avoid the circular dependency of "constructing counting units with the variant structure to be solved, and then using the counting units to inversely solve the variant structure". The median is used as the threshold here because the median is the midpoint of the distribution of the two counts over all feature pairs. It is determined by the data itself rather than given by humans and has no room for adjustment. Taking the midpoint of the distribution ensures that no more than half of the candidates for coupling are judged, avoiding the inclusion of weakly coupled feature pairs into composite feature pairs, which would lead to overly coarse counting units.

[0079] This step does not construct counting units based on convergence results, which is not the same as writing back the counting coefficients with convergence structure later: the composite feature bits change the definition of the counting unit of which is counted as an event. If it is constructed with the structure to be solved, the object to be minimized will drift with the solution. However, writing back the counting coefficients does not change the definition of the event, but only changes the counting scale of each feature bit, and is equipped with loop detection as the termination criterion.

[0080] From the candidate coupled feature bit pairs, select the maximal feature bit set that does not share feature bits with each other, and where any two members in the group constitute a candidate coupled feature bit pair, to obtain the feature bit group, where each feature bit is a constituent feature bit.

[0081] It should be noted that the process of obtaining the feature bit group is as follows: construct a relational graph with feature bits as vertices and candidate coupled feature bit pairs as edges. Select a set of vertices that are connected to any two vertices (i.e., each pair is a candidate coupled feature bit pair) and cannot be expanded further as a feature bit group. Each feature bit in the group is called a constituent feature bit. Different feature bit groups are required not to share constituent feature bits. It is necessary to require that each pair in the group is a candidate coupled feature bit pair. Pairs that meet the conditions cannot be merged into a connected set of vertices through transit. Otherwise, when A and B are coupled, B and C are coupled, but A and C are not coupled, the three will be mistakenly merged into the same composite feature bit. The requirement that feature bit groups do not share constituent feature bits is because if the same original feature bit falls into multiple composite feature bits, its event will be counted repeatedly, distorting the interpretation of event count.

[0082] The normalized state of each component feature bit is used as the normalized state of the composite feature bit, and the observed value of the composite feature bit is used as the observed value mentioned on the composite feature bit. When both the normalized state and the observed value of the composite feature bit are composed of all the component feature bits, the composite feature bit replaces its individual component feature bits in the event counting.

[0083] It should be noted that a composite feature bit is a new counting unit composed of all the component feature bits of a feature bit group. Its normalized state is taken as an ordered combination of the normalized values ​​of each component feature bit in the group, and its observed value is taken as an ordered combination of the observed values ​​of each component feature bit in the group. Only when a variant node is not undefined on all component feature bits, or a mention has been observed on all component feature bits, thus completing the value combination, is the composite feature bit used to replace its individual component feature bits in the counting of edge events and observation events. That is, in this case, the composite bit is counted as one event, while each component bit is no longer counted as an event separately, so that a drift across several coupling bits is counted as only one edge event. If any component feature bit is missing or the value combination is incomplete, the composite bit will not take effect on the node or mention, and will still be counted independently as each component feature bit.

[0084] Statistically count the connected components at each component feature bit and establish cross-branch shared edges;

[0085] It should be noted that the statistics of connected components and the establishment of cross-branch shared edges are not performed on the composite feature bits, but are always performed separately on each component feature bit. This is because the state of the composite feature bit is a combination of values. If the two branches are different in any component, the composite state will be different. The set of variant nodes holding the same composite state will become fragmented and degenerate into a single variant node. As a result, the number of connected components formed on the convergent undirected spectral topology will not reach more than two. As a result, a large number of cross-branch shared edges will be missed (for example, two branches that share the same pattern motif but have different color orders will have different composite states due to the different color order components. They will not be considered as having the same value and their connectivity will not be examined together. These two branches will not be identified as sharing the motif). Therefore, the composite feature bit is only used for event counting to correct the overestimation of the directionality of edge events. The identification of cross-branch sharing is still completed independently on each component feature bit.

[0086] Through the above technical solution, this embodiment uses the co-occurrence and synchronous change of the initial structure to screen out candidate coupled feature bit pairs, aggregates the feature bits that are paired and do not share each other into composite feature bits, and replaces the constituent feature bits with composite feature bits of value combination to participate in the counting, so that the real drift across several coupled feature bits is counted only once, avoiding the excessive merging caused by the repeated counting of edge events on coupled feature bits.

[0087] In an optional embodiment, the sum of the number of edge events and the number of observed events is used to obtain the explanatory event count, specifically including:

[0088] The observation density of a feature position is obtained by calculating the proportion of the observed mentions to the total number of mentions.

[0089] It should be noted that observation density refers to the prevalence of a certain feature bit in the text description. The process of obtaining it is: divide the number of observed mentions at that feature bit by the total number of mentions, with a value range of (0,1]. The data source label is taken from the source field retained in the information extraction process (each mention is recorded with its own title and source category, i.e., local chronicles, list application texts, collection records, and oral records of inheritors).

[0090] The multi-source intangible cultural heritage text data are divided according to their sources to obtain each source. Among the variant nodes that include mentions from at least two sources and have been observed at the feature position, the proportion of variant nodes with consistent observation values ​​from different sources is counted to obtain the cross-source recurrence rate of the feature position.

[0091] It should be noted that cross-source recurrence rate refers to the degree to which the value of a certain feature bit is independently reproduced across different sources. It is used to distinguish between "rare and stable" and "endangered and missing". The calculation process is as follows: First, the data is divided into various data sources according to the source field; then, for a certain feature bit, variant nodes that "simultaneously contain mentions from at least two different data sources and have been observed at that position" are selected as the denominator set; the number of variant nodes in the denominator set that "these mentions from different data sources have the same observed value at that position" is counted as the numerator, and the numerator divided by the denominator is the cross-source recurrence rate of that feature bit; if there are no variant nodes that satisfy the denominator condition for a certain feature bit (no one is observed by more than two sources), then no cross-source recurrence rate can be obtained for that position.

[0092] The feature bits that yield cross-source recurrence rate are denoted as sortable feature bits. The sortable feature bits are sorted according to observation density from low to high and cross-source recurrence rate from high to low. The two order bits of each sortable feature bit are summed and sorted again from small to large. The proportion of the resulting reverse order position to the total number of sortable feature bits is taken as the edge event count coefficient and observation event count coefficient of that feature bit.

[0093] It should be noted that the calculation process for the two types of count coefficients is a ranking conversion rather than a weight setting: First, all sortable feature bits are sorted once in ascending order of observation density to obtain their respective density order (the rarer the feature, the higher the order), and then sorted once in descending order of cross-source recurrence rate to obtain their respective recurrence order (the more stable the feature, the higher the order); Second, for each sortable feature bit, the two orders are added together to obtain a combined order sum, and then sorted once in ascending order of the combined order sum to obtain a comprehensive ranking; Third, the reverse order ranking corresponding to this comprehensive ranking (i.e., the ranking obtained by adding one to the total number of sortable feature bits and then subtracting the forward order ranking) is divided by the total number of sortable feature bits. The resulting ratio is used as both the edge event count coefficient and the observation event count coefficient for that feature bit. The domain falls within (0,1]; the two statistics must be jointly ordered and cannot each be determined separately as a coefficient, because low observation density alone cannot distinguish between rare and stable and endangered and missing, and must be separated by cross-source recurrence rate; the two count coefficients are deterministically derived from the ordinal relationship between observation density and cross-source recurrence rate, and their values ​​do not change with any adjustable parameters, so there is no need for calibration or training for specific corpora during implementation; ordinalization makes the coefficients independent of the numerical scale of the two statistics, thus eliminating the need to select the scale and weight ratio in advance when numerically weighting statistics with different dimensions; the two ordinal values ​​are summed with equal status because after ordinalization, the two are both positional quantities and there is no scale difference, so there is no need or way to set relative weights for them;

[0094] It should be noted that if there are tied positions when the two ordinal positions are summed and then sorted again, the average position should be taken. If the number of sortable feature positions is too small and the sum of the two ordinal positions is insufficient to produce a distinction, the counting coefficient can be determined directly by the joint order relationship of the two statistics.

[0095] When the cross-source reproducibility rate of a feature bit cannot be obtained, the median of the percentage of the reversed position of each sortable feature bit is taken as its two-class counting coefficient, and the feature bit is marked as a feature bit to be corrected.

[0096] It should be noted that the characteristic positions to be corrected refer to those that are temporarily included in the counting with neutral coefficients due to the lack of cross-source reproducibility evidence, and will be corrected after the phylogenetic analysis is completed. The two types of counting coefficients are taken as the median of the proportion of the inverted position of each sortable characteristic position. That is, the median is the neutral value under the ordinal position proportion system. It is necessary to avoid taking the coefficient as one, because taking it as one will pull the weight of these endangered characteristic positions to the maximum, which is contrary to the design intention. The median is also a statistical quantity determined by the data itself rather than a value given by humans. Since the two types of counting coefficients are taken from the proportion of the inverted position and distributed according to position on (0,1], the median is the median point of the distribution. Taking it will ensure that these characteristic positions are neither adjusted up nor down. The actual correction is postponed until after the phylogenetic analysis is completed, and will be made according to the structural residuals by the counting coefficient structural residual correction process described later.

[0097] Additionally, if the cross-source recurrence rate cannot be obtained for all feature bits, resulting in an empty set of sortable feature bits, then the two types of count coefficients for all feature bits are all set to one, i.e., degenerated into unweighted original explanatory event counts. After the spectrum is solved, the differential coefficients are uniformly assigned by the count coefficient structure residual correction process.

[0098] The edge events and observed events of each feature bit are accumulated according to the two types of counting coefficients to obtain the explanatory event count;

[0099] It should be noted that the accumulation process is as follows: each edge event is weighted and accumulated according to the edge event count coefficient of its feature position, and each observation event is weighted and accumulated according to the observation event count coefficient of its feature position. The sum of the two weighted sums is the explanatory event count. The two types of count coefficients only enter into the composition of the explanatory event count and do not appear in any "whether to merge" judgment statement. The judgment is always a decrease comparison of the explanatory event count.

[0100] Through the above technical solution, this embodiment generates two types of counting coefficients falling in (0,1) by joint ordinal conversion of observation density and cross-source recurrence rate, and obtains the explanatory event count by differentially accumulating the edge events and observed events. This allows feature bits with sufficient evidence and cross-source stability to contribute more to the judgment, while rare or endangered missing feature bits are treated differently. Moreover, the two types of counting coefficients are determined by the ordinal relationship of the data itself and do not change with any adjustable parameters. No corpus labeling or training is required during implementation.

[0101] In an optional embodiment, after obtaining the converged variant node partitioning, undirected spectral topology, and normalized state, a counting coefficient correction process is further included, specifically comprising:

[0102] For each feature bit, count the number of connected components formed by each non-undefined normalized state in the converged undirected spectrum topology, and take the maximum number as the independent occurrence degree of that feature bit.

[0103] It should be noted that the degree of independent occurrence refers to the degree to which the value of a certain feature bit has repeated independent origins in the spectrum. The process of calculating it is as follows: fix a certain feature bit, take each non-undefined normalized state value of the bit, count the number of connected components formed by the variant nodes holding the value in the topology of the convergent undirected spectrum, and take the maximum value of the number of connected components among all the values ​​of the bit as the degree of independent occurrence of the feature bit.

[0104] The number of observed events for each feature bit across all data sources is counted, and the proportion of the largest number to the total number of observed events for that feature bit is taken as the source concentration of that feature bit.

[0105] It should be noted that source concentration refers to the extent to which the inconsistency of a certain feature bit is concentrated from a single data source. The process of obtaining it is as follows: classify and count the observed events that occur at the feature bit according to the data source to which they are mentioned, take the maximum value of the number of observed events at a single source, divide it by the total number of observed events at the feature bit, and obtain the source concentration, with a value range of (0,1).

[0106] The feature bits are sorted from high to low based on their independent occurrence, and the proportion of their position to the total number of feature bits is taken as the updated edge event counting coefficient. The feature bits are also sorted from high to low based on their source concentration, and the proportion of their position to the total number of feature bits is taken as the updated observation event counting coefficient. The feature bits to be corrected are updated with the two types of counting coefficients based on their independent occurrence and source concentration, and the marking is removed.

[0107] It should be noted that the update direction corresponds to the two types of event sources: High independent occurrence indicates a low degree of stable continuation of the normalized state along the phylogenetic lineage, more like independent origins. Therefore, the effectiveness of its side events as evidence of inheritance drift should be reduced. The updated side event count coefficient is obtained by sorting by independent occurrence in descending order and taking the proportion of the ordinal position to the total number of feature positions (the earlier the coefficient, the smaller it is, i.e., reduced). High source concentration indicates that the inconsistency of this position mainly stems from differences in the caliber of a single source, more like recording noise. Therefore, the effectiveness of its observed events should be reduced. The updated observed event count coefficient is obtained by sorting by source concentration in descending order and taking the proportion of the ordinal position. Feature positions to be corrected can now participate in the above two sorting processes based on their independent occurrence and source concentration, thereby obtaining the updated two types of coefficients and removing the correction mark. The write-back here is based on the structural position of the residuals (number of connected components, source concentration), not the numerical value of the residuals.

[0108] The alternating execution is re-executed with the updated two types of counting coefficients to obtain new converged variant node partitions, undirected spectral topologies, and normalized states. The variant node partitions, undirected spectral topologies, and normalized states obtained in each round are recorded. When the result obtained in this round is the same as the corresponding result in any round recorded in previous rounds, the counting coefficient correction process is terminated.

[0109] It should be noted that after each round of updating the count coefficients, the scale of the explanatory event count changes accordingly, and the count values ​​of the two rounds cannot be directly compared. Therefore, the outer layer cannot use the fact that the count value no longer decreases as the termination condition. This embodiment uses a loop detection method, that is, it fully records the variant node partitioning, undirected spectrum topology, and normalized state triplet of each round (the normalized state must be recorded together, as it directly determines the degree of independent occurrence and the final node attributes; partitioning and topology being the same but the state being different are not considered the same result). The loop terminates when the triplet of the current round is exactly the same as any previous round. This criterion can cover both the stable convergence case and the case of back-and-forth between two structures, and does not introduce the iteration number parameter. The graph result after termination is selected as follows: among the rounds that constitute this loop, the round with the smallest explanatory event count is taken as the final output; if they are tied, the last round is taken.

[0110] Through the above technical solution, this embodiment uses the number of connected components on the spectral topology and the concentration of the sources of observed events as structural residuals, writes back two types of counting coefficients according to their structural positions, and terminates the outer layer with cyclic detection, so that the counting contribution of each feature bit can be automatically adjusted according to the structural evidence after the spectral solution is obtained, and the system can converge stably when the structure no longer changes, without introducing the iteration number parameter.

[0111] In an optional embodiment, after establishing cross-branch shared edges between variant nodes corresponding to different connected components, the process further includes phylogenetic orientation calibration and hierarchical structure generation, specifically including:

[0112] Obtain asymmetric evidence corresponding to the mention, wherein the asymmetric evidence includes at least one of the following: document date markers, data formation time order, outgroup mentions, ancestral state markers, and records of changes in normalized state complexity;

[0113] It should be noted that asymmetric evidence refers to evidence that can provide a chronological indication of the direction of transmission. Its sources and acquisition are as follows: document date markers are taken from the writing or recording date fields of each text; the order of data formation time is taken from the recording time in the collected metadata; external group mentions refer to mentions that are determined to be outside the genealogy of this category and can be used as directional references, which are taken from the category labeling in the extraction process; ancestral state markers are taken from the manual annotations of domain experts on certain values ​​of earlier forms; normalized state complexity change records are taken from the sorting results of the same feature value according to the predetermined complexity measures such as the complexity of the pattern and the number of processes; at least one of these is required to enter directional labeling, and multiple ones can be used together when they coexist.

[0114] Each variant node in the converged undirected spectrum topology is taken as a candidate root variant node. The number of edges that do not match the asymmetric evidence after orientation of each candidate root variant node is counted, and the candidate root variant node with the fewest edges is selected.

[0115] It should be noted that the candidate root variant node refers to the variant node that serves as the starting point of the lineage. The process of obtaining it is as follows: taking each variant node as a candidate root variant node in turn, temporarily orienting all edges on the undirected lineage topology outward from the candidate root, checking whether the direction matches the asymmetric evidence for each edge, counting the number of inconsistent edges, and selecting the candidate root variant node with the fewest inconsistent edges. The reason for using a fixed root instead of each edge orientation is that each edge orientation will fail at the point where the age of the endangered branch is missing, while this scheme only needs one piece of asymmetric evidence to determine the root position, and the direction of the entire tree topology is determined accordingly. The other edges do not need their own age evidence.

[0116] The determination of discrepancies in various types of evidence is as follows: the order of document date markers or data formation time is opposite to the direction of the oriented edge; the mention of the outgroup does not fall at the upstream of the oriented edge; the node where the ancestral state marker is located is not upstream of the oriented edge; or the normalized state complexity (the ordinal value given according to the predetermined complexity measure such as the complexity of the pattern and the number of processes) decreases along the oriented edge (it should increase with inheritance or at least not decrease). Any of these discrepancies will be counted as discrepancies for that edge.

[0117] When a candidate root variant node is unique, or when the same candidate root variant node is obtained after successively removing one asymmetric piece of evidence and re-screening, it is recorded as a root variant node; otherwise, the edge of the undirected spectral topology is recorded as a spectral edge with undetermined direction.

[0118] It should be noted that when the one with the fewest inconsistent edges is not unique, a perturbation test is performed. That is, after removing one piece of asymmetric evidence from the evidence set each time, the above root selection is repeated. If the same candidate root variant node is still obtained after each removal, it is recorded as a root variant node. Otherwise, it indicates that the directional basis is insufficient, and the edge of the undirected spectral topology is recorded as a spectral edge with undetermined direction and is not forcibly oriented. When there is only one type of available asymmetric evidence and it is impossible to remove it item by item, if the one with the fewest inconsistent edges under the unique evidence is still not unique, the edge is directly recorded as a spectral edge with undetermined direction. The object of the perturbation test is the evidence, not the topology: it should not be changed to remove an edge in the topology one by one, because the undirected spectral topology is a tree, and removing an edge breaks it into two disconnected parts. The root found again on the incomplete topology is no longer the root of the same problem, and it is meaningless to judge stability based on this.

[0119] After determining the root variant node, the edges of the undirected spectral topology are oriented as directed spectral edges from the root variant node outward. The level depth of each variant node is obtained according to the number of edges from each variant node to the root variant node. Variant nodes with the same level depth and the same parent variant node on the directed spectral edge are grouped into the same level node. Cross-level inclusion edges are established between two level nodes with adjacent level depths and parent-child relationships between the variant nodes they contain.

[0120] It should be noted that hierarchy depth refers to the generational distance of a variant node from the root in the genealogy. It is calculated by the number of edges traversed by the variant node along the directed genealogical edges to the root variant node. Hierarchical nodes are a group of variant nodes aggregated according to the same hierarchy depth and the same parent. They are calculated by first determining the parent variant node of each variant node according to the directed genealogical edges, and then grouping variant nodes with the same hierarchy depth and the same parent variant node into the same hierarchy node. Cross-hierarchical containment edges are containment edges established between adjacent hierarchy nodes due to the parent-child relationship between their members. The criterion is: when a member of an upper-level hierarchy node is the parent variant node of a member of a lower-level hierarchy node in two adjacent hierarchy nodes, a cross-hierarchical containment edge is established between these two hierarchy nodes. Semantically, hierarchy nodes correspond to the hierarchy of genre differentiation, and cross-hierarchical containment edges correspond to the coverage of a higher-level genre by a lower-level genre. Both are automatically generated by hierarchy depth and forking structure, without introducing any manually specified labels.

[0121] The final graph structure supports retrieval and inheritance analysis as follows: The output graph contains variant nodes and their normalized state attributes, directed lineage edges (or lineage edges with undetermined directions), cross-branch shared edges, hierarchical nodes, and cross-level contained edges. Based on this, one can trace the differentiation path and origin of a variant along the directed lineage edges, trace the cross-branch circulation of the same technological element between branches without direct mentorship along the cross-branch shared edges, and search step by step according to the school of thought along the cross-level contained edges, thereby supporting intelligent retrieval and inheritance analysis.

[0122] For example, a hierarchical output can be: hierarchical node T2 {hierarchical depth = 2, parent = T1, members = [V7, V9], semantics = a second-level sub-branch under a certain branch}; cross-hierarchical containing edge {endpoint = T1-T2, basis = V3 is the parent of V7}.

[0123] Through the above technical solution, this embodiment uses asymmetric evidence to determine root variant nodes through root determination and evidence removal perturbation tests. Then, it orients and automatically generates hierarchical nodes and cross-hierarchical inclusion edges according to hierarchical depth and bifurcation structure, so that the direction of the entire lineage can be determined by only one reliable root position. Even when the evidence is sparse or missing, it can still give a robust direction or clearly mark it as undetermined. And it automatically forms a hierarchical structure of corresponding school differentiation according to the bifurcation structure.

[0124] In summary, the intangible cultural heritage knowledge graph construction method proposed in this application, in a scenario where multiple sources of variant texts coexist and continuously change along the inheritance chain, uses the sum of edge events and observation events at each feature position as the sole criterion, and minimizes the differences in organization and recording by combining variant node division, undirected phylogenetic topology, and normalized state. This method can distinguish between recording errors and inheritance changes while merging duplicate records, and avoid multiple branches being merged into a single node or the same object being split into isolated nodes. It can also explicitly record the recurrence of craft elements in branches without direct apprenticeship relationships by sharing edges across branches. Optionally, it corrects the overestimation of edge events caused by coupling through composite feature positions, stabilizes the determination by ordinalization of count coefficients and structural residual correction, and provides the school hierarchy through root determination and hierarchical generation, thereby forming a structured intangible cultural heritage knowledge graph that can support intelligent retrieval and inheritance analysis.

[0125] See Figure 2 As shown, this solution proposes an intangible cultural heritage knowledge graph construction system to implement the aforementioned intangible cultural heritage knowledge graph construction method, including:

[0126] The text observation module is used to acquire multi-source intangible cultural heritage text data, extract the observation values ​​of the mentions of techniques, patterns or crafts at each feature position, and record the mentions at the feature position as unobserved if no observation value is extracted.

[0127] The initial spectrum module is used to merge the mentions into variant nodes based on the inclusion and disjoint relationships between the mention sets corresponding to each observation value, thereby obtaining the initial variant node partition and undirected spectrum topology. The mode observation value of the mentions in the variant node is used as the initial normalized state for each feature bit. If there are no observations, the normalized state is marked as undefined.

[0128] The event counting module is used to compare the normalized states at both ends of each edge of the undirected spectral topology with variant node partitioning, undirected spectral topology and normalized state as joint unknowns. It counts non-undefined and inconsistent feature bits as edge events, and compares each mentioned observation value with the normalized state of its variant node. Observed and inconsistent feature bits are counted as observation events. The number of edge events and the number of observation events are summed to obtain the explanatory event count.

[0129] The alternating convergence module is used to alternately execute state reassignment, mention reassignment, and topology reconnection starting from the initial value of the joint unknown. The state reassignment is to change the normalized state from the observed values ​​of each feature bit. The mention reassignment will reassign the mention to other or newly created variant nodes. The topology reconnection changes the access position of the variant node. Each operation only retains those that cause the count of explained events to decrease until the count no longer decreases in one round of traversal, thus obtaining the converged variant node partition, undirected spectral topology, and normalized state.

[0130] The cross-branch sharing module is used to count the connected components formed by variant nodes with the same non-undefined normalized state on the same feature bit in the converged undirected spectral topology. When there are more than two connected components, cross-branch sharing edges are established between variant nodes corresponding to different connected components.

[0131] The graph generation module is used to construct an intangible cultural heritage knowledge graph by taking the variant nodes and their normalized states in the converged variant node partition as graph nodes and node attributes, and taking the edges of the converged undirected spectral topology and cross-branch shared edges as graph edges.

[0132] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0133] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0134] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps described above.

[0135] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for constructing an intangible cultural heritage knowledge graph, characterized in that, The method includes: Obtain multi-source intangible cultural heritage text data, extract the observation values ​​of the mentions of techniques, patterns or crafts at each feature position, and record the mentions at the feature position as unobserved if no observation value is extracted. Based on the inclusion and disjoint relationships between the mention sets corresponding to each observation value, the mentions are merged into variant nodes to obtain the initial variant node partition and undirected spectrum topology. The mode observation value of the mentions in the variant node is taken as the initial normalized state for each feature bit. If there is no observation, the normalized state is recorded as undefined. Using variant node partitioning, undirected spectral topology, and normalized state as joint unknowns, the normalized state at both ends of each edge of the undirected spectral topology is compared with the feature bits. Feature bits that are not undefined and inconsistent are counted as edge events. The observed values ​​mentioned are compared with the normalized state of the variant node to which they belong. Observed and inconsistent feature bits are counted as observed events. The number of edge events and the number of observed events are summed to obtain the number of explained events. Starting from the initial value of the joint unknown, state reassignment, mention reassignment, and topology reconnection are performed alternately. The state reassignment is to change the normalized state from the observed values ​​of each feature bit. The mention reassignment will reassign the mention to other or newly created variant nodes. The topology reconnection changes the access position of the variant node. Each operation only retains those that cause the count of explained events to decrease until the count no longer decreases in one round of traversal, thus obtaining the converged variant node partition, undirected spectral topology, and normalized state. Statistically identify the connected components formed by variant nodes with the same non-undefined normalized state at the same feature position in the converged undirected spectral topology. When there are two or more connected components, establish cross-branch shared edges between variant nodes corresponding to different connected components. By using the variant nodes and their normalized states in the converged variant node partition as graph nodes and node attributes, and the edges of the converged undirected spectral topology and cross-branch shared edges as graph edges, an intangible cultural heritage knowledge graph is constructed.

2. The method according to claim 1, characterized in that, Based on the inclusion and disjoint relationships between the mention sets corresponding to each observation value, mentions are grouped into variant nodes, specifically including: Each observation value is used to collect mentions with that observation value to obtain the support set for that observation value, and all mentions are combined to form the total support set; Compare the inclusion, disjointness and intersection relationships of each support set, select a maximal hierarchical subfamily from the support set where any two support sets are only included or disjoint, and merge the total support set into the maximal hierarchical subfamily. The variant nodes are formed by merging the support sets containing the least number of mentions in the maximal hierarchical subfamily. The variant nodes are then connected according to the direct inclusion relationship between the merged units to obtain the initial variant node partitioning and undirected spectral topology.

3. The method according to claim 1, characterized in that, Alternating execution of state reassignment, mention reassignment, and topology reaccess includes: When resetting the execution state, fix the variant node partition and the undirected spectral topology, try to obtain the observed values ​​that have appeared at each feature bit, re-count the edge events of the associated edges of the variant node and the observation events mentioned therein, and when the minimum value of the sum of the two is less than the current value, replace the normalized state of the variant node at that feature bit with the corresponding observation value. When performing mention reassignment, mentions are tried to be assigned to adjacent variant nodes and newly created variant nodes respectively. The observed values ​​of the mention are used to form the normalized state of the newly created variant node. The observed events of the mention and the events of the removed and assigned variant nodes and their associated edges are recounted to obtain the increment of each candidate reassignment. When performing topology reconnection, the variant node and its subordinate branches are pruned from the current position and the remaining variant nodes are tried to be connected. The edge events of the pruning position and the trial connection position are re-counted to obtain the increment of each candidate reconnection. When there are candidates with a reassignment increment or candidate reaccess increment that are less than zero, retain the result corresponding to the smallest value. After terminating the alternation execution, shrink and merge adjacent variant nodes with the same set of defined feature bits and the same corresponding normalization state.

4. The method according to claim 1, characterized in that, Before comparing the normalized states at both ends of each edge of the undirected spectral system topology, the process also includes the synthesis of composite feature bits, specifically including: In the initial variant node partitioning and undirected spectral topology, the number of times the same variant node is mentioned as being jointly observed and the number of times the normalized state changes synchronously at both ends of the same side are counted for each feature pair. Feature pairs whose two counts are not less than their median among all feature pairs are denoted as candidate coupled feature pairs. Select the maximal feature bit set from the candidate coupled feature bit pairs that do not share feature bits with each other and any two members in the group constitute the candidate coupled feature bit pair, and obtain the feature bit group, where each feature bit is a component feature bit; The normalized state of each component feature bit is used as the normalized state of the composite feature bit, and the observed value is used as the observed value mentioned on the composite feature bit. When both the normalized state and the observed value of the composite feature bit are composed of all the component feature bits, the composite feature bit replaces each of its component feature bits in the event counting. Connectivity components are counted at each component feature bit, and cross-branch shared edges are established.

5. The method according to claim 1, characterized in that, Summing the number of edge events with the number of observed events yields the count of explained events, which includes: The observation density of a feature position is obtained by calculating the proportion of the observed mentions to the total number of mentions. The multi-source intangible cultural heritage text data are divided according to their sources to obtain each source. Among the variant nodes that include mentions from at least two sources and have been observed at the feature position, the proportion of variant nodes with consistent observation values ​​from different sources is counted to obtain the cross-source recurrence rate of the feature position. The feature bits that yield cross-source recurrence rate are denoted as sortable feature bits. The sortable feature bits are sorted according to observation density from low to high and cross-source recurrence rate from high to low. The two order bits of each sortable feature bit are summed and sorted again from small to large. The proportion of the resulting reverse order position to the total number of sortable feature bits is taken as the edge event count coefficient and observation event count coefficient of that feature bit. When the cross-source reproducibility rate of a feature bit cannot be obtained, the median of the percentage of the reversed position of each sortable feature bit is taken as its two-class counting coefficient, and the feature bit is marked as a feature bit to be corrected. The edge events and observed events of each feature bit are accumulated according to the two types of counting coefficients to obtain the explanatory event count.

6. The method according to claim 5, characterized in that, After obtaining the converged variant node partitioning, undirected spectral topology, and normalized state, the process also includes a count coefficient correction procedure, specifically including: For each feature bit, count the number of connected components formed by each non-undefined normalized state in the converged undirected spectrum topology, and take the maximum number as the independent occurrence degree of that feature bit. The number of observed events for each feature bit across all data sources is counted, and the proportion of the largest number to the total number of observed events for that feature bit is taken as the source concentration of that feature bit. The feature bits are sorted from high to low based on their independent occurrence, and the proportion of their position to the total number of feature bits is taken as the updated edge event counting coefficient. The feature bits are also sorted from high to low based on their source concentration, and the proportion of their position to the total number of feature bits is taken as the updated observation event counting coefficient. The feature bits to be corrected are updated with the two types of counting coefficients based on their independent occurrence and source concentration, and the marking is removed. The alternating execution is re-executed with the updated two types of counting coefficients to obtain new converged variant node partitions, undirected spectral topology and normalized state, and the variant node partitions, undirected spectral topology and normalized state obtained in each round are recorded. The counting coefficient correction process terminates when the result obtained in this round is the same as the corresponding result in any round in previous records.

7. The method according to claim 1, characterized in that, After establishing cross-branch shared edges between variant nodes corresponding to different connected components, the process also includes phylogenetic orientation calibration and hierarchical structure generation, specifically including: Obtain asymmetric evidence corresponding to the mention, wherein the asymmetric evidence includes at least one of the following: document date markers, data formation time order, outgroup mentions, ancestral state markers, and records of changes in normalized state complexity; Each variant node in the converged undirected spectrum topology is taken as a candidate root variant node. The number of edges that do not match the asymmetric evidence after orientation of each candidate root variant node is counted, and the candidate root variant node with the fewest edges is selected. When a candidate root variant node is unique, or when the same candidate root variant node is obtained after successively removing one asymmetric piece of evidence and re-screening, it is recorded as a root variant node; otherwise, the edge of the undirected spectral topology is recorded as a spectral edge with undetermined direction. After determining the root variant node, the edges of the undirected spectral topology are oriented as directed spectral edges from the root variant node outwards. The level depth of each variant node is obtained according to the number of edges from each variant node to the root variant node. Variant nodes with the same level depth and the same parent variant node on the directed spectral edge are grouped into the same level node. Cross-level inclusion edges are established between two level nodes with adjacent level depths and parent-child relationships between the contained variant nodes.

8. A system for constructing a knowledge graph of intangible cultural heritage, characterized in that, A method for constructing an intangible cultural heritage knowledge graph as described in any one of claims 1-7, comprising: The text observation module is used to acquire multi-source intangible cultural heritage text data, extract the observation values ​​of the mentions of techniques, patterns or crafts at each feature position, and record the mentions at the feature position as unobserved if no observation value is extracted. The initial spectrum module is used to merge the mentions into variant nodes based on the inclusion and disjoint relationships between the mention sets corresponding to each observation value, thereby obtaining the initial variant node partition and undirected spectrum topology. The mode observation value of the mentions in the variant node is used as the initial normalized state for each feature bit. If there are no observations, the normalized state is marked as undefined. The event counting module is used to compare the normalized states at both ends of each edge of the undirected spectral topology with variant node partitioning, undirected spectral topology and normalized state as joint unknowns. It counts non-undefined and inconsistent feature bits as edge events, and compares each mentioned observation value with the normalized state of its variant node. Observed and inconsistent feature bits are counted as observation events. The number of edge events and the number of observation events are summed to obtain the explanatory event count. The alternating convergence module is used to alternately execute state reassignment, mention reassignment, and topology reconnection starting from the initial value of the joint unknown. The state reassignment is to change the normalized state from the observed values ​​of each feature bit. The mention reassignment will reassign the mention to other or newly created variant nodes. The topology reconnection changes the access position of the variant node. Each operation only retains those that cause the count of explained events to decrease until the count no longer decreases in one round of traversal, thus obtaining the converged variant node partition, undirected spectral topology, and normalized state. The cross-branch sharing module is used to count the connected components formed by variant nodes with the same non-undefined normalized state on the same feature bit in the converged undirected spectral topology. When there are more than two connected components, cross-branch sharing edges are established between variant nodes corresponding to different connected components. The graph generation module is used to construct an intangible cultural heritage knowledge graph by taking the variant nodes and their normalized states in the converged variant node partition as graph nodes and node attributes, and taking the edges of the converged undirected spectral topology and cross-branch shared edges as graph edges.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images