Automatic file analysis method based on semantic evolution graph

Abstract

The invention discloses an automatic file analysis method based on a semantic evolution graph. The method comprises the following steps: generating a policy document standardized corpus; obtaining a policy document structured corpus unit set; constructing an initial semantic map snapshot containing a version stamp; establishing a time slice semantic evolution graph sequence; obtaining a multi-time sequence fusion vector; generating a multi-time sequence optimization word embedding vector matrix; performing time continuous regularization on the multi-time-sequence optimization word embedding vector matrix, outputting time continuous word embedding vectors, and mapping the time continuous word embedding vectors to nodes and edges of the time slice semantic evolution graph sequence to form a global semantic evolution graph fusion representation sequence; and taking the global semantic evolution graph fusion representation sequence as feature input to obtain a policy document analysis result set. According to the method, the timeliness, the accuracy and the interpretability of automatic analysis of the policy document are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of file parsing, and in particular to a file automatic parsing method based on semantic evolution graph. BACKGROUND

[0002] With the continuous expansion and dynamic adjustment of the policy and regulation system, the number and complexity of policy documents continue to grow, and the demand for policy interpretation, regulatory comparison, and risk judgment is increasing. In the prior art, policy document automatic parsing mainly relies on natural language processing and static knowledge graph technology, usually using subject-object-action triple extraction, policy clause classification, and regulation relationship annotation methods. However, the existing methods are generally based on fixed structures and limited samples for supervised learning, and have obvious limitations in the actual scenarios of frequent policy terminology iteration, large clause expression changes, and departmental coordination.

[0003] On the one hand, the existing static knowledge graph cannot effectively express the continuous evolution and version replacement of policy concepts, entities, and their relationships over time, lacks time sequence modeling of the whole process, and makes it difficult to track and quantify the evolution path, internal connection, and transition logic of key concepts between different versions of policies. On the other hand, traditional word embedding migration or semantic modeling methods have difficulty in supporting high-robustness analysis under weak or no annotation conditions when facing long-tail terms in the policy text, new terms, and homonym expressions due to sparse annotation and uneven sample distribution. SUMMARY

[0004] One object of the present application is to provide a file automatic parsing method based on semantic evolution graph, which greatly improves the timeliness, accuracy, and interpretability of policy document automatic parsing.

[0005] According to the file automatic parsing method based on semantic evolution graph, the method comprises the following steps:

[0006] Collecting multiple source policy document original texts, establishing a policy document original text set, performing unified coding, format standardization, structure annotation, and release time annotation on the policy document original text set, and generating a policy document standardized corpus;

[0007] Based on the policy document standardized corpus, preprocessing is performed to obtain a policy document structured corpus unit set;

[0008] Extracting entities and their semantic relationships from the policy document structured corpus unit set to construct an initial semantic graph snapshot containing version stamps;

[0009] Dividing the initial semantic graph snapshot into multiple time slices according to the release time, and establishing a time slice semantic evolution graph sequence;

[0010] The improved word embedding model is called to perform vectorization representation on the structured corpus unit set of the policy document, to obtain a multi-time fusion vector;

[0011] Based on the anchor words in the domain, anchor word alignment is performed on the multi-time fusion vector, to obtain an aligned word embedding vector and perform synonym aggregation and polysemy separation, to generate a multi-time optimized word embedding vector matrix;

[0012] Time-continuous regularization is performed on the multi-time optimized word embedding vector matrix, to output a time-continuous word embedding vector, and map it to the nodes and edges of the time-slice semantic evolution graph sequence, to form a global semantic evolution graph fusion representation sequence;

[0013] The global semantic evolution graph fusion representation sequence is taken as a feature input, to obtain a policy document analysis result set.

[0014] Optionally, the preprocessing of the policy document standardized corpus includes:

[0015] Each policy document original text in the policy document standardized corpus is subjected to clause segmentation processing, to obtain a clause unit set;

[0016] Each clause unit in the clause unit set is subjected to title and body structure recognition, to divide each clause unit into a title part and a body part;

[0017] Cross-document reference link parsing is performed on each clause body part, to extract all reference fragments in the body part, and map the reference fragments to the pointed standardized clause numbers;

[0018] Each policy term in the title part and the body part is subjected to term normalization processing, to standardize the policy terms with the same meaning but different expression forms, to obtain a standardized term set;

[0019] The structured corpus unit set is constructed by integrating the clause segmentation result, the title and body structure recognition result, the clause reference mapping set, and the standardized term set.

[0020] Optionally, the construction of the initial semantic graph snapshot containing the version stamp includes:

[0021] For each structured corpus unit in the structured corpus unit set, a policy theme entity, a policy tool entity, and a target object entity are extracted, to form a policy core semantic entity set;

[0022] For each policy core semantic entity set, semantic relationships among the policy theme entity, the policy tool entity, and the target object entity are extracted, and combined into a semantic relationship triple set;

[0023] For each structured corpus unit, the release time of the original text of the policy document corresponding to the structured corpus unit is taken as a version stamp, the version stamp is bound with the semantic relation triple set to obtain a semantic relation triple set with a version stamp attribute;

[0024] The semantic relation triple set with the version stamp attribute is summarized to construct an initial semantic graph snapshot, and a policy core semantic entity node and a semantic relation edge with a version stamp attribute are provided.

[0025] Optionally, the initial semantic graph snapshot is divided into multiple time slices according to the release time, including:

[0026] All policy core semantic entity nodes with a version stamp attribute in the initial semantic graph snapshot are divided into multiple time slices in a self-adaptive sliding window manner;

[0027] According to the time slices, the initial semantic graph snapshot is split into multiple time slice semantic graphs Each time slice semantic graph includes a policy core semantic entity node set and a semantic relation edge set, and for two adjacent time slice semantic graphs, a policy automatic analysis semantic evolution tensor Δ m is calculated.

[0028]

[0029] Among them, are respectively the indication vectors of the newly added node set, the modified node set and the abolished node set under the policy automatic analysis theme, and α + , α ° , α - are the corresponding weight coefficients under the automatic analysis policy.

[0030] The time slice semantic graphs are sequentially spliced according to the time sequence and the policy automatic analysis semantic evolution tensor to form a time slice semantic evolution graph sequence.

[0031] Optionally, the construction of the multi-time sequence fusion vector includes:

[0032] For each standardized policy term in the structured corpus unit set, an improved word embedding model is called to perform vectorization representation to obtain a policy document initial word embedding vector matrix.

[0033] Each word embedding vector in the policy document initial word embedding vector matrix is injected with a time identification vector according to the time slice where it is located, the corresponding time slice vector is spliced to form a word embedding vector with time sequence information, and is bound with the time slice semantic graph node one by one to form a multi-time sequence fusion vector for the policy document automatic analysis theme.

[0034] Optionally, the improved word embedding model construction comprises:

[0035] Anchors in the anchor set are obtained by summing the square of the distance between the general pre-training word embedding vector and the domain word embedding vector, and the anchor alignment loss value of each anchor in the anchor set under the general pre-training word embedding vector and the domain word embedding vector is obtained.

[0036] The local semantic contrast loss value of each policy term and its positive semantic sample and negative semantic sample in the local semantic contrast sample set is obtained by taking the logarithm of the ratio of the similarity of the policy term and the positive semantic sample vector to the similarity of the policy term and the negative semantic sample vector, and summing them up.

[0037] For each standardized policy term corresponding to the word embedding vector in the continuous time slice, the distance between the word embedding vectors of the same policy term in the continuous time slice is squared to obtain the time continuous regularization loss value.

[0038] The total loss value is obtained by weighted summing the anchor alignment loss value, the local semantic contrast loss value and the time continuous regularization loss value, and the total loss value is used for the optimization goal of the improved word embedding model in the automatic parsing of policy documents.

[0039] Optionally, the aligned word embedding vector and the near-synonym aggregation and the heteronym separation comprise:

[0040] For the multi-time sequence fusion vector, the stable anchor word in the automatic parsing scene of the policy document is selected, the word embedding vector and the anchor word in the time slice are aligned to obtain the aligned word embedding vector.

[0041] For the standardized policy term pair in the aligned word embedding vector in the time slice, the near-synonym pair and the heteronym pair are constructed, and the local contrast learning mechanism is adopted to maximize the vector distance of the heteronym pair and minimize the vector distance of the near-synonym pair, so as to realize the near-synonym aggregation and the heteronym separation, and generate the optimized word embedding vector.

[0042] The optimized word embedding vector in each time slice is summarized to form a multi-time sequence optimized word embedding vector matrix.

[0043] Optionally, the formation of the global semantic evolution graph fusion representation sequence comprises:

[0044] The vector smoothness measure value between the optimized word embedding vectors corresponding to each standardized policy term in the adjacent time slices in the time slice semantic evolution graph sequence is calculated.

[0045] On the basis of minimizing the vector smoothness measure value, the time continuous word embedding vector set is output.

[0046] mapping each time continuous word embedding vector in the set of time continuous word embedding vectors to an entity node of the time slice semantic evolution graph atlas sequence;

[0047] performing embedding synthesis operation on the time continuous word embedding vectors of the two entity nodes connected by the same policy relationship edge in the set of time continuous word embedding vectors for each time slice to form a time continuous embedding representation at the edge level;

[0048] fusing the node-level time continuous word embedding vector under each time slice with the edge-level time continuous embedding representation to form a time slice semantic evolution graph fusion vector representation, and sequentially splicing the semantic evolution graph fusion vector representations under all time slices to form a global semantic evolution graph fusion representation sequence.

[0049] Optionally, the set of policy document analysis results is constructed, including:

[0050] taking the global semantic evolution graph fusion representation sequence as an input feature, constructing a policy clause classification model according to a policy field classification standard, classifying each clause unit in the set of structured corpus units, and outputting a set of clause classification results;

[0051] taking the global semantic evolution graph fusion representation sequence as an input feature, constructing a policy relationship prediction model, performing relationship prediction on the clause pairs in the set of structured corpus units, and outputting a set of relationship prediction results;

[0052] based on the global semantic evolution graph fusion representation sequence, performing dynamic comparison on the embedding vectors and relationship vectors of the same standardized policy term, clause or entity under different time slices, and outputting a set of cross-version difference comparison results;

[0053] summarizing the set of clause classification results, the set of relationship prediction results and the set of cross-version difference comparison results to form a set of policy document analysis results.

[0054] The beneficial effects of the present application are:

[0055] The present application realizes fine modeling and multi-time sequence automatic analysis of policy document semantic evolution, breaks through the key bottlenecks of the prior art in time sequence modeling, semantic drift quantification, long-tail term differentiation and weakly labeled analysis, and greatly improves the timeliness, accuracy and interpretability of policy document automatic analysis. BRIEF DESCRIPTION OF DRAWINGS

[0056] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0057] Figure 1A flow chart of a file automatic parsing method based on semantic evolution graph is provided. DETAILED DESCRIPTION

[0058] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams which show the basic structure of the application in a schematic manner only, and thus only show the components relevant to the application.

[0059] REFERENCE Figure 1 A file automatic parsing method based on semantic evolution graph, comprising:

[0060] Collecting multi-source policy document original texts, establishing a policy document original text set, performing unified coding, format standardization, structure annotation and release time annotation on the policy document original text set, and generating a policy document standardized corpus;

[0061] Based on the policy document standardized corpus, preprocessing is performed to obtain a policy document structured corpus unit set;

[0062] In this embodiment, the preprocessing of the policy document standardized corpus includes:

[0063] Each policy document original text in the policy document standardized corpus is subjected to clause segmentation processing to obtain a clause unit set. Each clause unit represents the smallest clause granularity text unit formed after the policy document original text is structured and divided. The clause unit set is obtained by dividing each policy document original text according to the structured identification information and language connection mode;

[0064] Each clause unit in the clause unit set is subjected to title and body structure recognition to distinguish each clause unit into a title part and a body part. The title part represents the title text of the clause unit, and the body part represents the body text of the clause unit. The title part and the body part jointly constitute the content boundary of the complete clause unit;

[0065] Cross-document reference link parsing is performed for each clause body part to extract all reference segments in the body part, and the reference segments are mapped to the standardized clause numbers pointed to. A clause reference mapping set is used to record each reference in the clause body part and its corresponding standardized clause number. The clause reference mapping set supports standardized reference pointing across documents and across levels;

[0066] Each policy term in the title part and the body part is subjected to term normalization processing. Policy terms with the same meaning but different expressions are standardized and normalized to obtain a standardized term set;

[0067] The clause segmentation result, the title and the text structure recognition result, the clause reference mapping set and the standardized term set are integrated to construct a structured corpus unit set, each structured corpus unit comprising a title part, a text part, a reference mapping set of clauses and a standardized term set related to the clauses.

[0068] The structured corpus unit set of the policy document is extracted to extract entities and semantic relationships, and an initial semantic graph snapshot containing a version stamp is constructed.

[0069] In the embodiment, the construction of the initial semantic graph snapshot containing the version stamp comprises:

[0070] For each structured corpus unit in the structured corpus unit set, policy theme entities, policy tool entities and target object entities are extracted to form a policy core semantic entity set, which is used to represent all core semantic entity types contained in the structured corpus unit.

[0071] For each policy core semantic entity set, semantic relationships between policy theme entities, policy tool entities and target object entities are extracted and combined into a semantic relationship triple set, which is used to represent the correspondence between all entities and relationships in the structured corpus unit.

[0072] For each structured corpus unit, the publication time of the original text of the policy document corresponding to the structured corpus unit is taken as the version stamp, which is used to mark the time sequence information of the policy document to which the structured corpus unit belongs. The version stamp is bound with the semantic relationship triple set to obtain a semantic relationship triple set with a version stamp attribute, which is used to record the time sequence state of entities and relationships.

[0073] The semantic relationship triple set with the version stamp attribute is summarized to construct an initial semantic graph snapshot, and policy core semantic entity nodes and semantic relationship edges with the version stamp attribute are provided. The initial semantic graph snapshot comprises all policy core semantic entity nodes with the version stamp attribute and semantic relationship edges with the version stamp attribute, and is used to uniformly express the global structure of policy core semantic entities and their semantic relationships at the same time sequence.

[0074] The initial semantic graph snapshot is divided into multiple time slices according to the publication time, and a time slice semantic evolution graph sequence is established.

[0075] In the embodiment, the initial semantic graph snapshot is divided into multiple time slices according to the publication time, comprising:

[0076] The policy core semantic entity nodes with version stamp attributes in the initial semantic graph snapshot are divided into multiple time slices in a self-adaptive sliding window manner;

[0077] The time slice width is determined by multiplying the adaptive coefficient by the ratio of the number of policy clause units involved in the automatic parsing scenario and revised in the time range to the total number of policy document structured corpus units and taking the integer greater than or equal to the product, and each time slice width is used to represent the degree of centralized release of policy clauses in the automatic parsing scenario of the policy document in the corresponding time range.

[0078]

[0079] wherein λ is a time slice adaptive coefficient, represents the number of policy clause units involved in the automatic parsing scenario and revised in the time interval q m all C

[0080] The initial semantic graph snapshot is divided into multiple time slice semantic graphs according to the time slices. Each time slice semantic graph includes a set of policy core semantic entity nodes and a set of semantic relationship edges, and for two adjacent time slice semantic graphs, a policy automatic parsing semantic evolution tensor Δ m is calculated, which is used to represent the change characteristics of policy core semantic entities and relationships between two adjacent time slices in the automatic parsing scenario of the policy, and is obtained by concatenating the indication vectors of the set of added nodes, the set of modified nodes and the set of abolished nodes in the automatic parsing scenario of the policy multiplied by the corresponding weight coefficients in the automatic parsing strategy.

[0081]

[0082] wherein are the indication vectors of the set of added nodes, the set of modified nodes and the set of abolished nodes in the automatic parsing scenario of the policy, respectively, and α + , α ° , α - are the corresponding weight coefficients in the automatic parsing strategy.

[0083] The time slice semantic graphs are sequentially concatenated with the policy automatic parsing semantic evolution tensor to form a time slice semantic evolution graph sequence, which is used to record the global time sequence evolution trajectory of the policy core semantic entity nodes and the semantic relationship edges with the release time in the automatic parsing scenario of the policy document.

[0084] ​The improved word embedding model is called to perform vectorization representation on the structured corpus unit set of the policy document, to obtain a multi-time fusion vector;

[0085] In this embodiment, the construction of the multi-time fusion vector includes:

[0086] For each standardized policy term in the structured corpus unit set, the improved word embedding model is called to perform vectorization representation, to obtain a policy document initial word embedding vector matrix, which is used to represent the initial semantic features of all standardized policy terms in the policy document automatic parsing scenario, wherein each policy document initial word embedding vector is obtained by performing vectorization mapping on the corresponding standardized policy term by the improved word embedding model;

[0087] Each word embedding vector in the policy document initial word embedding vector matrix is injected with a time identifier vector according to the time slice where it is located, and a corresponding time sequence identifier vector is spliced according to the corresponding time slice vector, to form a word embedding vector with time sequence information, which is used to reflect the change characteristics of the policy term with the release time in the policy document automatic parsing scenario. Each word embedding vector with time sequence information is further indexed and bound one by one with the time slice semantic graph nodes, to form a multi-time fusion vector for the policy document automatic parsing theme.

[0088] In this embodiment, the construction of the improved word embedding model includes:

[0089] The distance square sum between the general pre-training word embedding vector and the domain word embedding vector is obtained, to obtain an anchor word alignment loss value of each anchor word in the anchor word set under the general pre-training word embedding vector and the domain word embedding vector, which is used to measure the vector consistency of the anchor word in different word embedding spaces in the policy document automatic parsing theme;

[0090] The local semantic contrast loss value of each policy term and its positive semantic sample and negative semantic sample in the local semantic contrast sample set is obtained by taking the logarithm of the ratio of the policy term and the positive semantic sample vector similarity and the policy term and the negative semantic sample vector similarity, and summing them up, which is used to measure the semantic distinguishability of the policy term in the policy document automatic parsing theme;

[0091] For the word embedding vector of each standardized policy term in the continuous multiple time slices, the distance square sum of the word embedding vectors of the same policy term in the continuous time slices is obtained, to obtain a time continuous regularization loss value, which is used to constrain the vector smoothness of the same policy term in different time slices in the policy document automatic parsing theme;

[0092] The total loss value is obtained by weighted sum of the anchor word alignment loss value, the local semantic contrast loss value and the time continuity regularization loss value, and the total loss value is used to improve the optimization goal of the word embedding model under the automatic parsing of policy documents.

[0093] Based on the anchor words in the domain, the anchor word alignment is performed on the multi-time sequence fusion vector to obtain the aligned word embedding vector and perform the synonym aggregation and the polysemy separation to generate a multi-time sequence optimized word embedding vector matrix.

[0094] In the embodiment, the aligned word embedding vector and the synonym aggregation and the polysemy separation include:

[0095] For the multi-time sequence fusion vector, stable anchor words in the automatic parsing of policy documents are selected, the word embedding vector under the time slice is aligned with the anchor words to obtain the aligned word embedding vector, the aligned word embedding vector is used to measure the difference between the word embedding vectors of the same anchor word under different time slices in the automatic parsing of policy documents, and the aligned word embedding vector is calculated by minimizing the sum of squares of distances between the word embedding vectors of the same anchor word under all time slices and the mean of all time slice embedding vectors of the anchor word. The word embedding vector of the anchor word under the time slice is used to represent the word semantic feature of the anchor word under the corresponding time slice, and the mean of the all time slice embedding vectors of the anchor word is used to represent the overall word semantic center under all time slices.

[0096] For the time slice, the standardized policy term pairs in the aligned word embedding vector are constructed, the local contrast learning mechanism is adopted, the vector distance of the polysemy pairs is maximized and the vector distance of the synonym pairs is minimized to realize the synonym aggregation and the polysemy separation, and the optimized word embedding vector is generated.

[0097] The optimized word embedding vector is used to measure the semantic gathering and scattering relationship of the policy terms in the automatic parsing of policy documents, the optimized word embedding vector is obtained by maximizing the distance between the word embedding vectors of the polysemy pairs under the same time slice and minimizing the distance between the word embedding vectors of the synonym pairs under the same time slice, the synonym pairs are used to represent the standardized policy term pairs expressing the same or similar policy meanings in the automatic parsing of policy documents, the polysemy pairs are used to represent the standardized policy term pairs expressing different policy meanings in the automatic parsing of policy documents, and the distance between the word embedding vectors is used to measure the semantic similarity or difference between the policy terms.

[0098] The optimized word embedding vectors under each time slice are summarized to form a multi-time sequence optimized word embedding vector matrix, the multi-time sequence optimized word embedding vector matrix is used to support the downstream tasks of policy clause classification, policy relationship prediction and cross-version difference comparison in the automatic parsing of policy documents, and the multi-time sequence optimized word embedding vector matrix is obtained by aggregating the optimized word embedding vectors under all time slices.

[0099] performing time-continuous regularization on the multi-time-series optimized word embedding vector matrix, outputting time-continuous word embedding vectors, and mapping the time-continuous word embedding vectors to nodes and edges of the time-slice semantic evolution graph sequence to form a global semantic evolution graph fusion representation sequence;

[0100] In the embodiment, the formation of the global semantic evolution graph fusion representation sequence includes:

[0101] calculating vector smoothness metric values between optimized word embedding vectors corresponding to each standardized policy term in adjacent time slices in the time-slice semantic evolution graph sequence;

[0102] The vector smoothness metric value is used to measure the degree of continuity of the semantic change of the same standardized policy term in adjacent time slices. The vector smoothness metric value is obtained by accumulating the squares of the vector distances of all standardized policy terms and all adjacent time slices, by calculating the square of the vector distance between the optimized word embedding vector of the same standardized policy term in the current time slice and the optimized word embedding vector in the previous time slice;

[0103] On the basis of minimizing the vector smoothness metric value, a set of time-continuous word embedding vectors is outputted;

[0104] The set of time-continuous word embedding vectors is used to represent the word embedding vector of each standardized policy term in each time slice after time-continuous regularization. The set of time-continuous word embedding vectors is used to ensure the continuity and stability of the semantic change trajectory of the same standardized policy term in the time dimension in the automatic parsing scene of policy documents;

[0105] Each time-continuous word embedding vector in the set of time-continuous word embedding vectors is mapped to an entity node of the time-slice semantic evolution graph sequence;

[0106] The entity node mapping operation is realized by a node-level mapping function. The node-level mapping function performs one-to-one mapping between the time-continuous word embedding vector in each time slice and the policy core semantic entity node corresponding thereto. The node-level mapping operation is used to represent the semantic features of each policy core semantic entity node in the corresponding time slice;

[0107] For each time slice, the time-continuous word embedding vectors of the two entity nodes connected by the same policy relationship edge in the set of time-continuous word embedding vectors are subjected to embedding synthesis operation to form a time-continuous embedding representation at the edge level;

[0108] The embedding synthesis operation includes vector splicing, vector averaging or vector weighting. The time-continuous embedding representation at the edge level is used to express the semantic relationship dynamic characteristics of the same policy relationship edge in the corresponding time slice;

[0109] The node-level time continuous word embedding vector and the edge-level time continuous embedding representation of each time slice are fused to form a time slice semantic evolution graph fusion vector representation, and the semantic evolution graph fusion vector representations of all time slices are sequentially spliced to form a global semantic evolution graph fusion representation sequence.

[0110] The global semantic evolution graph fusion representation sequence is used to comprehensively express the dynamic evolution path of the policy terms and the concept relationship thereof on the time axis in the automatic parsing scene of the policy document.

[0111] The global semantic evolution graph fusion representation sequence is used to comprehensively express the dynamic evolution path of the policy terms and the concept relationship thereof on the time axis in the automatic parsing scene of the policy document.

[0112] In the embodiment, the policy document parsing result set is constructed, including:

[0113] The global semantic evolution graph fusion representation sequence is used to comprehensively express the dynamic evolution path of the policy terms and the concept relationship thereof on the time axis in the automatic parsing scene of the policy document.

[0114] The clause classification rule includes dividing the clause unit into different policy categories according to the semantic features and node relationships of the standardized policy terms in the global semantic evolution graph fusion representation sequence, and the clause classification result set is used to represent the policy category attribute of each clause unit.

[0115] The global semantic evolution graph fusion representation sequence is used to comprehensively express the dynamic evolution path of the policy terms and the concept relationship thereof on the time axis in the automatic parsing scene of the policy document.

[0116] The relationship prediction rule includes predicting the policy semantic relationship type and the time sequence feature between the clause pair based on the node and edge fusion vectors of the standardized policy terms in the global semantic evolution graph fusion representation sequence and the historical policy evolution rule, and the relationship prediction result set is used to record the policy relationship type between the clause pair.

[0117] Based on the global semantic evolution graph fusion representation sequence, the embedding vectors and relationship vectors of the same standardized policy term, clause or entity in different time slices are dynamically compared, and a cross-version difference comparison result set is output.

[0118] The cross-version difference comparison rule includes determining the semantic change type of the standardized policy term, clause or entity according to the node-level and edge-level fusion vector changes thereof in different time slices, including consistency, deviation, differentiation or abolition, and the cross-version difference comparison result set is used to express the semantic evolution features of the policy document between different versions.

[0119] The aggregated clause classification result set, the relationship prediction result set, and the cross-version difference comparison result set form a policy document analysis result set, and the output format of the policy document analysis result set is as follows:

[0120] Policy clause classification result:

[0121] Category n: classification rule or threshold, including standardized policy term A coverage ≤ X% or relationship degree ≤ Y;

[0122] Policy relationship prediction result:

[0123] Relationship type m: prediction rule or condition, including standardized policy term A and standardized policy term B co-occurrence and semantic correlation ≥ Z;

[0124] Cross-version difference comparison result:

[0125] Consistency: including the vector distance of the same standardized policy term under each time slice ≤ threshold;

[0126] Offset: including the vector distance of the same standardized policy term under adjacent time slices within the threshold interval;

[0127] Differentiation: including the appearance of a split in the standardized policy term, resulting in multiple new nodes;

[0128] Abolition: including the disappearance of a standardized policy term or clause node.

[0129] Embodiment 1: A policy research institution in a certain region is entrusted to automatically analyze and compare multiple versions of a large number of policy documents in the field of digital infrastructure and new energy development in the past five years, aiming to accurately track the evolution of core policy terms, analyze changes in similar policy clauses between different periods and different publication levels, and provide the latest policy trends and potential compliance risk warnings for decision-makers and enterprises. The institution uses the conventional method of static knowledge graph combined with fine-tuning of mainstream language models and conducts parallel experiments and effect comparisons with traditional solutions on the same policy domain dataset.

[0130] The research team collected more than 1,000 digital and new energy-related policy documents issued by multiple policy publishing units in the past five years, covering guidelines, rules, draft proposals, and implementation plans. The total number of original documents exceeds 35 million words. All documents are numbered, formatted, and structured as shown in Table 1:

[0131] Table 1 Structure of Document Display

[0132]

[0133] All original policy documents are standardized in the implementation of the present application, including unified text encoding, automatic clause segmentation, title and body separation, and reference normalization. The "Digital Infrastructure Promotion Action Plan" in Example 1 is segmented into several clauses, each clause automatically establishes an index of references to related clauses in other policy documents, and all references to "network facilities" and "penetration rate" are unified into standard terms. After sorting, the following Table 2 structure is formed:

[0134] Table 2 structured clause unit

[0135]

[0136] Subsequently, the entity-relation joint extraction model of the present application is used to automatically identify policy topics ("new energy demonstration zone construction", "digital governance"), policy tools ("financial subsidies", "technical standards"), target objects ("local management agencies", "related enterprises") core semantic entities and their associated relationships, while each data is labeled with a release time point as a version stamp. The policy clause "promote the construction of emerging energy demonstration zones" in Example 1 is extracted as:

[0137] Policy theme entity: "emerging energy demonstration zone";

[0138] Policy tool entity: "demonstration";

[0139] Target object entity: "relevant areas";

[0140] Version stamp: "year-month-day";

[0141] The system uses an adaptive sliding window to dynamically generate time slices according to the number of policy documents and the concentration of revisions each year. In Example 1, the number of policy clauses released in a certain year increased sharply, and the system automatically set it as an independent time slice, while some years were combined. The difference between the semantic graph nodes and edges in each time slice is analyzed using an evolution tensor to accurately record the entire process of adding, adjusting, and abolishing "subsidy measures" and "project approval" concepts.

[0142] In the word embedding migration link, the basic word vectors of standardized policy terms are obtained through a general large model, and then combined with high-frequency stable anchor words ("digital platform" "subsidy amount") for anchor word alignment under each time slice, maintaining the consistency of policy basic semantics. The system can automatically identify that "information service station" and "digital service center" have different expressions but similar meanings in two years, and aggregate them in the semantic space; while "pilot demonstration zone" and "innovation demonstration zone" are effectively distinguished due to different policy orientations. If a new term such as "distributed new energy consumption" appears in a certain year, the system can accurately include the new term in the overall vector space through adaptive adjustment with only a few samples, effectively distinguishing its differences from existing terms.

[0143] By time continuous regularization, the system automatically tracks the semantic evolution track of the important term of "energy subsidy standard" in different annual documents, and it is found in embodiment 1 that the semantic center of a certain term gradually changes from "financial support" to "market support" in the past three years, which provides direct quantitative data for policy analysis and industry response. The multi-time sequence embedding vectors of all standardized terms correspond to the nodes of the evolution graph, and finally the global fusion vector is obtained, which can be directly used for clause classification, relationship prediction and policy evolution trend analysis intelligent tasks.

[0144] Taking the "energy subsidy policy" scene as an example, more than 300 relevant policy documents of different levels of units in five years are selected, and the main terms related to "subsidy standard", "demonstration project" and "approval mechanism" are extracted. The present method and the traditional scheme are used to model and train data samples (part of the clause content) as shown in Table 3:

[0145] Table 3 Training data sample

[0146] Clause number Title Text content Standard term B2-1 Subsidy standard New projects are subsidized X yuan per unit of output Subsidy standard B2-2 Exit mechanism New projects after a certain year will no longer enjoy subsidies Subsidy exit B2-3 Demonstration project Support the construction of "new energy demonstration projects" Demonstration project B2-4 Innovation pilot Allow "innovation pilot" projects to have priority access Innovation pilot ... ... ... ...

[0147] The comparative experiment strictly uses the same training set and test set allocation: training set: 220 policy documents, covering energy subsidies and equipment management topics, a total of 1620 human work annotation samples; test set: 80 new documents not seen before, a total of 520 policy clauses, spanning five years and multiple release levels.

[0148] The main evaluation indicators and result data (part of the sample) are shown in Table 4:

[0149] Table 4 Comparison experiment evaluation results

[0150]

[0151]

[0152] In the "subsidy standard adjustment" clause tracking, the present method can accurately identify the historical evolution of a certain term from "X yuan / unit" to "Y yuan / unit", and timely detect the first appearance of "innovation pilot" related clauses and subsequent expansion impact. The traditional method is easy to confuse "innovation pilot" with "demonstration project", and the tracking of semantic changes in different years is not accurate. The recognition ability of the newly introduced terms is higher than 80% in the present method under the condition of only a small amount of annotation, while the traditional method is less than 60%.

[0153] In the cross-version policy difference comparison, the present method can output the following structured comparison results:

[0154] Policy clause classification results:

[0155] Subsidy policy: standard term "subsidy standard" <= certain threshold or clause contains "no longer subsidized";

[0156] Demonstration project: clause contains "demonstration project" or semantic correlation with "innovation pilot" is higher than a certain threshold;

[0157] Policy relationship prediction results:

[0158] "Innovation pilot" and "approval mechanism" clauses exist in parallel or reference relationship, and the prediction weight is higher than 0.9;

[0159] Cross-version difference comparison results:

[0160] Consistency: as "subsidy exit" has very small semantic vector distance between adjacent annual files;

[0161] Offset: "innovation pilot" is gradually expanded year by year after its first appearance, and the semantic vector change increases;

[0162] Finally, the policy research team pointed out in the feedback that the method of the present invention has a significant improvement in efficiency, accuracy and intelligence in clause fine-grained identification, cross-year tracking, few-sample new term learning and automatic structuring compared with traditional solutions, effectively meeting the demand for automatic analysis and decision support in high complexity policy environment.

[0163] In the present application, by extracting entities and their semantic relationships from multi-source policy document original texts, an initial semantic graph snapshot containing version stamp attributes is formed, and a self-adaptive sliding window segmentation mechanism is used to dynamically generate multi-resolution time slices according to the centralized distribution of policy clause units in the automatic analysis scene, and a time-sequenced policy automatic analysis semantic evolution tensor is constructed by combining node evolution indicator vectors and weights, realizing the dynamic tracking and quantitative expression of policy core semantic entities and their relationships under different time slices. In the aspect of word embedding migration optimization, the improved word embedding model is called for the standardized policy terms in the structured corpus unit set to obtain the initial word embedding vector, and it is spliced with the time slice time sequence identifier to form a multi-time sequence fusion word embedding vector. The multi-time sequence word embedding vector is aligned with the high-frequency stable anchor word in the field, and the semantic consistency is maintained by minimizing the semantic distance of the same anchor word in different time slices. Further, within each time slice, a local contrast learning mechanism is used to construct synonym pairs and heteronym pairs, maximize the heteronym distance and minimize the synonym distance to generate high-discrimination optimized word embedding vectors, and perform time continuous regularization across time slices to output continuous and stable semantic evolution vectors, which fully reflect the semantic evolution trajectory of the same policy concept at different time nodes. The time-continuous word embedding vector is mapped and synthesized with the time slice semantic evolution graph nodes and relationships one by one to obtain a global semantic evolution graph fusion representation sequence, which can realize multi-granularity automatic classification, complex relationship intelligent prediction and cross-version difference dynamic comparison of policy clauses. The output policy document analysis result set can be structured to express different clause categories, relationship types and their evolution and differentiation characteristics on the time axis, effectively supporting policy comparison, risk warning and policy impact tracing in multiple scenarios. The present application realizes the fine modeling and multi-time sequence automatic analysis of policy document semantic evolution, breaks through the key bottlenecks of existing technologies in time sequence modeling, semantic drift quantification, long-tail term differentiation and weakly labeled analysis, and greatly improves the timeliness, accuracy and interpretability of policy document automatic analysis.

[0164] The above is only the preferred embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for automatic file parsing based on semantic evolution graph, characterized in that: include: Collect original texts of policy documents from multiple sources, establish a collection of original texts of policy documents, perform unified coding, format standardization, structure annotation and release time annotation on the original text collection of policy documents, and generate a standardized corpus of policy documents; Preprocessing is performed based on the standardized corpus of policy documents to obtain a set of structured corpus units of policy documents; Extracting entities and their semantic relationships from the set of structured corpus units of the policy document, and constructing an initial semantic graph snapshot including a version stamp; Dividing the initial semantic graph snapshot into multiple time slices according to the release time, and establishing a time slice semantic evolution graph sequence; The improved word embedding model is used to vectorize the set of structured corpus units in policy documents to obtain a multi-time series fusion vector. Based on the anchor words in the domain, anchor word alignment is performed on the multi-time series fusion vector to obtain the aligned word embedding vector and perform synonym aggregation and synonym separation to generate a multi-time series optimized word embedding vector matrix. Performing time-continuous regularization on the multi-time-series optimized word embedding vector matrix, outputting a time-continuous word embedding vector, and mapping it to the nodes and edges of the time-sliced ​​semantic evolution graph sequence to form a global semantic evolution graph fusion representation sequence; The global semantic evolution graph fusion representation sequence is used as feature input to obtain a set of policy document parsing results.

2. The method for automatic file parsing based on semantic evolution graph according to claim 1, characterized in that: The preprocessing of the standardized policy document corpus includes: Perform clause segmentation on the original text of each policy document in the standardized corpus of policy documents to obtain a set of clause units; Identifying the title and body structure of each clause unit in the clause unit set, and dividing each clause unit into a title part and a body part; Perform cross-document reference link parsing for the body of each clause, extract all reference fragments in the body, and map the reference fragments to the standardized clause numbers they point to; Performing term normalization processing on each policy term in the title part and the body part, normalizing policy terms with the same meaning but different expression forms to obtain a standardized term set; The structured corpus unit set is constructed by integrating the clause segmentation results, title and text structure recognition results, clause reference mapping set and standardized term set.

3. The method for automatic file parsing based on semantic evolution graph according to claim 2, characterized in that: The construction of the initial semantic graph snapshot including the version stamp includes: For each structured corpus unit in the structured corpus unit set, extracting a policy subject entity, a policy tool entity, and a target object entity to form a policy core semantic entity set; For each of the policy core semantic entity sets, extract the semantic relationships between the policy subject entity, the policy tool entity, and the target object entity, and combine them into a set of semantic relationship triples; For each structured corpus unit, the release time of the original text of the policy document corresponding to the structured corpus unit is used as a version stamp, and the version stamp is bound to the set of semantic relationship triples to obtain a set of semantic relationship triples with version stamp attributes; The set of semantic relationship triples with version stamp attributes is aggregated to construct an initial semantic graph snapshot, and policy core semantic entity nodes and semantic relationship edges with version stamp attributes are provided.

4. The method for automatic file parsing based on semantic evolution graph according to claim 3, characterized in that: The initial semantic graph snapshot is divided into multiple time slices according to the release time, including: Divide all policy core semantic entity nodes with version stamp attributes in the initial semantic graph snapshot into multiple time slices according to an adaptive sliding window method; Split the initial semantic graph snapshot into multiple time-sliced ​​semantic graphs based on time slices Each time slice semantic graph includes a set of policy core semantic entity nodes and a set of semantic relationship edges. For two adjacent time slice semantic graphs, the policy automatic parsing semantic evolution tensor Δ is calculated. m ; in, They are the indicator vectors of newly added node sets, modified node sets and abolished node sets under the policy automatic parsing theme, α + , α ° , α - is the corresponding weight coefficient under the automatic parsing strategy; The time-sliced ​​semantic graph is sequentially spliced ​​with the policy automatic parsing semantic evolution tensor in time series to form a time-sliced ​​semantic evolution graph sequence.

5. The method for automatic file parsing based on semantic evolution graph according to claim 4, characterized in that: The construction of the multi-time series fusion vector includes: For each standardized policy term in the set of structured corpus units, calling the improved word embedding model to perform vector representation to obtain an initial word embedding vector matrix of the policy document; A time identification vector is injected into each word embedding vector in the initial word embedding vector matrix of the policy document according to the time slice in which it is located, and the corresponding time series identification vector is spliced ​​according to the corresponding time slice vector to form a word embedding vector with time series information, and a one-to-one corresponding index binding is performed with the time slice semantic graph node to form a multi-time series fusion vector for automatic parsing of policy documents.

6. The method for automatic file parsing based on semantic evolution graph according to claim 5, characterized in that: The improved word embedding model construction includes: The anchor word alignment loss value of each anchor word in the anchor word set under the general pre-trained word embedding vector and the domain word embedding vector is obtained by summing the squares of the distances between the general pre-trained word embedding vector and the domain word embedding vector; The local semantic comparison loss value of each policy term and its positive semantic sample and reverse semantic sample in the local semantic comparison sample set is obtained by taking the logarithm of the ratio of the similarity between the policy term and the positive semantic sample vector and the similarity between the policy term and the reverse semantic sample vector and summing them up. For each standardized policy term’s corresponding word embedding vector in multiple consecutive time slices, the distance between the word embedding vectors of the same policy term in consecutive time slices is squared and summed to obtain the time-continuous regularization loss value; The total loss value is obtained by weighted summing the anchor word alignment loss value, the local semantic contrast loss value and the time-continuous regularization loss value. The total loss value is used to improve the optimization goal of the word embedding model under the topic of automatic parsing of policy documents.

7. The method for automatic file parsing based on semantic evolution graph according to claim 5, characterized in that: The alignment of word embedding vectors and performing synonym aggregation and synonym separation include: For the multi-time series fusion vector, a stable anchor word in the policy document automatic parsing scenario is selected, and the word embedding vector under the time slice is aligned with the anchor word to obtain an aligned word embedding vector; Aligning the standardized policy term pairs in the word embedding vectors for the time slices, constructing synonym pairs and synonym pairs, and using a local contrastive learning mechanism to achieve synonym aggregation and synonym separation by maximizing the vector distance of synonym pairs and minimizing the vector distance of synonym pairs, thereby generating an optimized word embedding vector; The optimized word embedding vectors under each time slice are aggregated to form a multi-time series optimized word embedding vector matrix.

8. The method for automatic file parsing based on semantic evolution graph according to claim 7, characterized in that: The formation of the global semantic evolution graph fusion representation sequence includes: Calculate the vector smoothness metric between the optimized word embedding vectors corresponding to each standardized policy term in the time slice semantic evolution graph sequence at adjacent time slices; Based on minimizing the vector smoothness metric, output a set of time-continuous word embedding vectors; Map each time-continuous word embedding vector in the time-continuous word embedding vector set to an entity node of the time-sliced ​​semantic evolution graph sequence; For each time slice, the time-continuous word embedding vectors of the two entity nodes connected to the same policy relationship edge in the time-continuous word embedding vector set are embedded and synthesized to form an edge-level time-continuous embedding representation; The node-level time-continuous word embedding vector and the edge-level time-continuous embedding representation under each time slice are fused to form a time-slice semantic evolution graph fusion vector representation. The semantic evolution graph fusion vector representations under all time slices are sequentially spliced ​​to form a global semantic evolution graph fusion representation sequence.

9. The method for automatic file parsing based on semantic evolution graph according to claim 8, characterized in that: The policy document parsing result set is constructed, including: Using the global semantic evolution graph fusion representation sequence as input features, a policy clause classification model is constructed according to the policy field classification standard, each clause unit in the structured corpus unit set is classified, and a clause classification result set is output; Using the global semantic evolution graph fusion representation sequence as input features, constructing a policy relationship prediction model, performing relationship prediction on the clause pairs in the structured corpus unit set, and outputting a relationship prediction result set; Based on the fusion representation sequence of the global semantic evolution graph, the embedding vectors and relationship vectors of the same standardized policy terms, clauses or entities at different time slices are dynamically compared, and a set of cross-version difference comparison results is output; The clause classification result set, the relationship prediction result set and the cross-version difference comparison result set are aggregated to form a policy document parsing result set.

Images