An archive intelligent abstract generation system and method based on natural language processing technology

Abstract

The application relates to the technical field of intelligent abstract generation, in particular to an archive intelligent abstract generation system and method based on a natural language processing technology; the system comprises the following: a multi-layer semantic generation core unit which carries out semantic hierarchical segmentation analysis on an archive original text, constructs a multi-layer semantic graph, and constructs an abstract generation model to generate abstract content; a knowledge graph fusion engine unit which is connected to extract terms in the archive original text, constructs an entity mapping relationship, and sets expression weights of semantic graph nodes in the abstract generation model according to a semantic association confidence score; a paging index cache retrieval unit which carries out segment division processing on each node in the multi-layer semantic graph, and constructs an abstract content segment index structure containing a node set; and a version tracing transaction management unit which records versions of generation and modification operations of the abstract content. The application constructs an archive intelligent abstract generation system which is fused with a multi-layer semantic graph construction and a knowledge graph association.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent summary generation, and particularly relates to an archive intelligent summary generation system and method based on natural language processing technology. BACKGROUND

[0002] With the increasing of unstructured text data such as government archives, enterprise files and academic data, archive intelligent summary technology, as an important means to improve information reading efficiency and support literature management automation, is gradually becoming an important research direction in the field of natural language processing. At present, the automatic summary method based on deep learning has been widely applied in news, medical treatment, law and other scenes, mainly relying on a sequence encoder at the text level to generate a fixed-length summary content. However, the traditional method often has difficulty in fine modeling of the complex semantic structure in the original archive, lacks the recognition and expression control ability of key knowledge units such as terms, entities and indicators, and thus the generated summary has certain deficiencies in accuracy, structural clarity and traceability.

[0003] The existing technology still has the following two key problems in actual application: firstly, the traditional summary generation method is mostly based on single plane semantic modeling or shallow representation, and lacks the modeling ability of the hierarchical semantic structure existing in the archive text, cannot effectively identify and retain the contact between the multi-level and multi-path semantic units, and thus the summary content lacks semantic integrity and structural logic; secondly, the current summary system generally ignores the deep fusion with the external knowledge graph, lacks the quantitative expression and dynamic control mechanism of the semantic contact strength between the archive terms and the external entities, and thus the generated summary does not pay enough attention to the key terms, affects the professionalism and explainability of the content, and cannot support the version management and on-demand query of the summary content and other operations. SUMMARY

[0004] The application aims to provide an archive intelligent summary generation system and method based on natural language processing technology to solve the two key problems proposed in the background technology.

[0005] To achieve the above-mentioned purpose, the application provides an archive intelligent summary generation system based on natural language processing technology, which comprises:

[0006] The multi-layer semantic generation core unit is used for performing semantic hierarchical segmentation analysis on the archive original text to construct a multi-layer semantic graph, scoring the structural relationship nodes in the multi-layer semantic graph, constructing an summary generation model based on the multi-layer semantic graph and the node scores, and generating summary content in combination with the structural path relationship of each node.

[0007] The knowledge graph fusion engine unit is configured to associate terms in the extracted archive original text, construct an entity mapping relationship between the terms and an external knowledge graph, calculate a semantic association confidence score based on semantic correlation between the terms and entity nodes in the external knowledge graph, and set a semantic graph node expression weight in a summary generation model according to the semantic association confidence score, and meanwhile, construct an extended reference matrix integrated into the summary generation model.

[0008] The paging index cache retrieval unit is configured to perform fragment division processing on each node in the multi-layer semantic graph, construct a summary content fragment index structure containing a node set, and filter out matching summary content based on a user input summary query condition using a matching page block filtering operation.

[0009] The version tracing transaction management unit is configured to record versions of the generation and modification operations of the summary content, and construct a multi-summary version index linked list structure to trace and manage the summary content.

[0010] The multi-layer semantic generation core unit includes a semantic graph construction module.

[0011] The semantic graph construction module is configured to perform semantic hierarchical segmentation analysis on the archive original text, and construct a multi-layer semantic graph based on inter-sentence logical connection structure, theme word clustering boundaries and paragraph pragmatic relationships.

[0012] The semantic hierarchical segmentation analysis includes syntax dependency analysis, semantic role labeling and context hierarchical division, divides the archive original text into a plurality of semantic units according to grammatical structure relationships, identifies subject-predicate-object structures and context dependencies in each semantic unit, and constructs a semantic nested hierarchical text structure graph.

[0013] The semantic nested hierarchical text structure graph is used to construct a multi-layer semantic graph, and the specific method is as follows:

[0014] The semantic layers are constructed based on the path depth of syntax dependency analysis, the semantic units in the same path depth are classified into the same semantic layer, the intra-layer edges are established based on semantic dependency relationships between predicate argument structures, and the cross-layer edges are established based on event pre- and post-context relationships.

[0015] The nodes of the multi-layer semantic graph are semantic units, the intra-layer edges of the multi-layer semantic graph are syntax dependency relationships between the same semantic layer nodes, and the cross-layer edges of the multi-layer semantic graph represent event context progressive relationships between different semantic layers.

[0016] Preferably, the multi-layer semantic generation core unit includes a summary modeling control module.

[0017] The abstract modeling control module constructs an abstract generation model based on the structural relationship of each node in the multi-layer semantic graph and the structural relationship node score;

[0018] The structural relationship of each node in the multi-layer semantic graph includes the syntactic dependency relationship between the same layer semantic units and the event context progression relationship between the cross-layer semantic units. The structural relationship node score is weighted and calculated based on the structural depth hierarchical factor, the syntactic connection centrality factor and the context aggregation density factor of each node in the semantic graph;

[0019] The structural depth hierarchical factor represents the semantic level of the node in the multi-layer semantic graph. The syntactic connection centrality factor represents the number of in-out edges of the node in the multi-layer semantic graph. The context aggregation density factor represents the density of co-occurring nodes in the context sliding window.

[0020] Preferably, the abstract modeling control module comprises:

[0021] The attention mechanism of the limited window is used to limit the local window range of the attention mechanism calculation based on the path distance between each node in the multi-layer semantic graph when constructing the abstract generation model. The guided semantic output control structure is used to preferentially generate the semantic unit content with high semantic weight based on the path centrality index and the structural relationship node of each node in the multi-layer semantic graph during the abstract generation process.

[0022] Preferably, the knowledge graph fusion engine unit comprises an entity mapping construction module;

[0023] The entity mapping construction module is used to identify terms from the original text of the archives, and based on the structural features between the terms and the entity nodes in the external knowledge graph, to construct the mapping path between the nodes in the multi-layer semantic graph and the entity nodes in the external knowledge graph. Specifically, the following steps are included:

[0024] S21.1, extract the term nodes from the original text of the archives, and calculate the context semantic vector of each term node; extract the attribute description information of the entity nodes in the external knowledge graph, and construct the corresponding entity node semantic vector;

[0025] S21.2, based on the similarity between the context semantic vector of the term node and the entity node semantic vector, filter the candidate node pairs that meet the preset semantic similarity threshold to form a candidate mapping set;

[0026] S21.3, respectively construct the context path subgraph of the candidate term node in the multi-layer semantic graph and the structural neighborhood subgraph of the candidate entity node in the external knowledge graph;

[0027] S21.4, respectively, on the context path subgraph and the structure neighborhood subgraph Structure embedding coding is carried out to obtain path structure feature representation, and based on the structural similarity between the path structure feature representation, combined with the semantic similarity, the joint confidence score of the mapping path is calculated;

[0028] S21.5, the joint confidence score of the node pair exceeding the preset score threshold is determined as the matching node, and the path mapping structure is constructed to form a node mapping path set, and finally the mapping path between the nodes in the multi-layer semantic graph and the entity nodes in the external knowledge graph is obtained.

[0029] Preferably, the knowledge graph fusion engine unit further comprises an association weight regulation module;

[0030] The association weight regulation module calculates the semantic association confidence score based on the semantic correlation between the term and the entity node in the external knowledge graph, and sets the semantic graph node expression weight in the abstract generation model according to the semantic association confidence score, which specifically comprises the following steps:

[0031] S22.1, the entity definition, attribute label and adjacency relationship of the entity node in the external knowledge graph are extracted to construct the multi-dimensional vector embedding representation of the entity node;

[0032] S22.2, the semantic correlation score between the context semantic vector of the term node and the entity node semantic vector is calculated using the bidirectional attention mechanism method, and the semantic association confidence score is calculated by combining the structural position index of the term node in the multi-layer semantic graph;

[0033] S22.3, based on the semantic association confidence score, the semantic graph node expression weight of the multi-layer semantic graph node corresponding to the term node in the abstract generation model is allocated.

[0034] Preferably, the extended reference matrix is a structured semantic reference matrix between the term node and the entity node in the external knowledge graph, which is used to describe the external knowledge graph entity path relationship referenced by the node in the multi-layer semantic graph in the semantic hierarchical structure;

[0035] The specific method for constructing the extended reference matrix is: extracting the term node set in the multi-layer semantic graph and the candidate entity node set in the external knowledge graph; based on the entity mapping path and the semantic association confidence score, the joint weight value of each pair of nodes in the term node set and the candidate entity node set is calculated, and the joint weight value greater than the preset reference confidence threshold is taken to construct the extended reference matrix;

[0036] The joint weight value greater than the preset reference confidence threshold indicates whether the term node references the candidate entity node, and the size of the joint weight value represents the reference strength, that is, the greater the joint weight value, the greater the reference strength.

[0037] Preferably, the page index cache retrieval unit comprises a semantic segment division module, a segment index construction module and a conditional matching retrieval module.

[0038] The semantic segment division module divides each node in the multi-layer semantic graph into a segment cluster based on the node structure path relationship and semantic clustering characteristics, to generate a plurality of abstract segment subgraphs.

[0039] The segment index construction module is used to encode each abstract segment subgraph into segment metadata containing node index numbers and path start and end markers, and to construct a page index table structure for user on-demand positioning and calling.

[0040] The conditional matching retrieval module constructs a query semantic vector based on the user input abstract query condition, and performs similarity scoring matching with each abstract segment subgraph in the page index table structure, to select a target abstract segment that meets the matching threshold.

[0041] Preferably, in the version tracing transaction management unit, the multi-abstract version index linked list structure is a unidirectional or bidirectional chain structure composed of a plurality of abstract version nodes, used for sequentially recording, traceable management and version content difference tracking of historical abstract versions after the user operates the abstract content.

[0042] The specific method for constructing the multi-abstract version index linked list structure is as follows: after each operation and editing of the abstract content, an abstract version node is constructed, which includes abstract compressed content, a timestamp, a unique identifier and original semantic segment mapping records; if the current is the first time to generate abstract content by the system, the first constructed abstract version node is taken as the start node of the linked list; otherwise, the current abstract version node is linked to the previous abstract version node, and a bidirectional index pointer is set; between each pair of continuous abstract version nodes, an abstract difference description vector is constructed using a word-level difference identification technology method, and is attached to the index structure between the pair of linked list nodes, to finally obtain the multi-abstract version index linked list structure.

[0043] On the other hand, the present application provides an archive intelligent abstract generation method based on natural language processing technology, which is used for the archive intelligent abstract generation system based on natural language processing technology described above, and comprises the following steps:

[0044] S10.1, performing semantic hierarchical segmentation analysis on the archive original text to construct a multi-layer semantic graph, using structure relationship node scoring on each node in the multi-layer semantic graph, constructing an abstract generation model based on the multi-layer semantic graph and the node scoring, and generating abstract content in combination with the structure path relationship of each node;

[0045] S10.2, extract the terms in the archive original text, construct the entity mapping relationship between the terms and the external knowledge graph, calculate the semantic correlation confidence score based on the semantic correlation between the terms and the entity nodes in the external knowledge graph, and set the semantic graph node expression weight in the summary generation model according to the semantic correlation confidence score, and construct an extended reference matrix integrated into the summary generation model;

[0046] S10.3, segmenting each node in the multi-layer semantic graph, constructing a summary content segment index structure containing a node set, and filtering out matching summary content based on user input summary query conditions using a matching page block filtering operation;

[0047] S10.4, record the version of the summary content generation and modification operation, and construct a multi-summary version index linked list structure to manage the summary content.

[0048] Compared with the prior art, the above technical scheme of the present application has the following beneficial technical effects:

[0049] 1. In the present application, the summary generation model is constructed based on the multi-layer semantic graph and the structure relationship node score, which can realize the hierarchical analysis and structure path control of the semantic units in the archive original text, and improve the generation quality of the summary content in terms of semantic structure integrity, information focus and expression controllability.

[0050] 2. In the present application, the entity mapping relationship between the terms and the external knowledge graph is constructed, and the semantic graph node expression weight is set based on the semantic correlation confidence score, which realizes the importance regulation of the key node expression in the summary generation model, and further enhances the context semantic coverage ability between the archive terms and the graph entities by constructing the extended reference matrix. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A principle block diagram of an embodiment of the present application is provided.

[0052] Reference signs: 1, multi-layer semantic generation core unit; 11, semantic graph construction module; 12, summary modeling control module; 2, knowledge graph fusion engine unit; 21, entity mapping construction module; 22, correlation weight regulation module; 3, paging index cache retrieval unit; 31, semantic segment division module; 32, segment index construction module; 33, condition matching retrieval module; 4, version trace transaction management unit. DETAILED DESCRIPTION

[0053] Embodiment one, as shown in the figure, provides an archive intelligent summary generation system based on natural language processing technology, which comprises: Figure 1

[0054] ​The multi-layer semantic generation core unit 1 is used for segmenting and analyzing the original text of the archives according to semantic levels, constructing a multi-layer semantic graph, scoring nodes in the multi-layer semantic graph according to structural relationships, constructing an abstract generation model based on the multi-layer semantic graph and the node scores, and generating abstract content in combination with the structural path relationships of the nodes;

[0055] The multi-layer semantic generation core unit 1 comprises a semantic graph construction module 11.

[0056] The semantic graph construction module 11 is used for segmenting and analyzing the original text of the archives according to semantic levels, constructing a multi-layer semantic graph based on inter-sentence logical connection structures, subject word clustering boundaries, and paragraph pragmatic relationships.

[0057] The semantic level segmentation and analysis comprises syntax dependency analysis, semantic role labeling, and context level segmentation, the original text of the archives is divided into a plurality of semantic units according to the grammatical structure relationship, and the subject-predicate-object structure and context dependency in each semantic unit are identified, so as to construct a text structure graph with semantic nested levels.

[0058] The text structure graph with semantic nested levels is used for constructing the multi-layer semantic graph, and the specific method is as follows:

[0059] The semantic layers are constructed according to the path depth of the syntax dependency analysis, the semantic units in the same path depth are classified into the same semantic layer, the intra-layer edges are established based on the semantic dependency relationship between the predicate argument structures, and the cross-layer edges are established based on the context relationship before and after the event.

[0060] The nodes of the multi-layer semantic graph are the semantic units, the intra-layer edges of the multi-layer semantic graph are the syntax dependency relationships between the nodes in the same semantic layer, and the cross-layer edges of the multi-layer semantic graph represent the event context progressive relationship of different semantic layers.

[0061] In this embodiment, the path depth of the syntax dependency analysis refers to the shortest path length required for the semantic unit to start from the syntax root node and reach the node through the dependency relationship edge in the syntax dependency tree, the path depth is represented by the minimum edge number in the graph structure, and is used for reflecting the subordinate level of the node in the overall syntax structure.

[0062] In this embodiment, the multi-layer semantic generation core unit 1 comprises an abstract modeling control module 12.

[0063] The abstract modeling control module 12 constructs an abstract generation model based on the structural relationships of the nodes in the multi-layer semantic graph and the node scores of the structural relationships.

[0064] Among them, the structural relationships of each node in the multi-layer semantic graph include the syntactic dependency relationship between semantic units at the same layer and the event context progression relationship between semantic units across layers; the structural relationship node score is calculated by weighting the structural depth hierarchical factor, syntactic connectivity centrality factor and context aggregation density factor of each node in the semantic graph.

[0065] Among them, the structural depth hierarchical factor represents the semantic level of a node in a multi-layer semantic graph; the syntactic connectivity centrality factor represents the number of inbound and outbound edges of a node in a multi-layer semantic graph; and the context aggregation density factor represents the density of co-occurring nodes of a node in a context sliding window.

[0066] In this embodiment, the syntactic dependency relationship between semantic units at the same level is based on the main component connection structure of dependency grammar, which is used to represent the grammatical control and modification direction between the subject, predicate, object and modifier in the same semantic layer, thereby reflecting the primary and secondary structure and logical connection relationship between semantic units.

[0067] The event context progression relationship between cross-level semantic units is based on the sequential orientation structure of event trigger words and context causal words. It is used to characterize the contextual presupposition and causal groundwork of upper-level events on lower-level semantic units, and to reflect the pragmatic evolution path of the semantic succession chain in the summary generation.

[0068] In this embodiment, the basic idea of ​​the structural depth hierarchical factor is: the closer a node is to the top layer, the stronger its control over the entire semantic organization structure and the more representative its generation. The hierarchical weight is represented by the reverse normalized value of the layer number, as defined below:

[0069] Suppose that the multi-layer semantic graph has L layers and node v q The layer it is located on is l q Then its structural depth stratification factor is:

[0070]

[0071] Among them, f str (v q ) is node v q The structural depth hierarchical factor, and f str (v q )∈[0,1], the closer to the first layer, the larger its value;

[0072] The syntactic connectivity factor measures the "connectivity activity" of a node in a structural connectivity graph, representing the strength of its syntactic association with other semantic units, and is defined as follows:

[0073] Let node v q The set of outgoing edges is O(v q The set of incoming edges is I(v) q If ), then its syntactic connectivity centrality factor is:

[0074]

[0075] Among them, f syn (v q ) is node v q Syntactic connectivity factor; v j Let V be the j-th node; V is the set of all nodes; the larger the syntactic connection centrality factor, the more syntactic connection hub the node is, and the higher the representativeness of the summary.

[0076] The context aggregation density factor measures the semantic density of a node within its context neighborhood, and is defined as follows:

[0077] For node v q Set the context window W q This contains the set of all nodes semantically related to it within the sliding time window / structural window, denoted as N. c (v q If ), then its context aggregation density factor is:

[0078]

[0079] Among them, f ctx (v q ) is node v q Context aggregation density factor;

[0080] Finally, node v q The structural relationship node scores are as follows:

[0081] S str (v q )=α1f str (v q )+α2f syn (v q )+α3f ctx (v q );

[0082] Among them, S str (v q ) is node v q The structural relationship node score; α1 is the structural depth hierarchical factor weight; α2 is the syntactic connection centrality factor weight; α3 is the context aggregation density factor weight.

[0083] In this embodiment, the summary modeling control module 12 includes a window-limiting attention mechanism and a guided semantic output control structure;

[0084] The attention mechanism of the defined window is used to limit the local window range of the attention mechanism calculation based on the path distance between nodes in the multi-layer semantic graph when constructing the summary generation model, and only the attention distribution calculation is performed in the node set that meets the preset path length threshold, so as to improve the semantic focusing ability between nodes and suppress the cross-section semantic interference; the guided semantic output control structure is used to generate the semantic unit content with high semantic weight based on the path centrality index and the structural relationship node of each node in the multi-layer semantic graph during the summary generation process.

[0085] In the embodiment, the window refers to the path perception domain size set in the process of constructing the summary generation model to limit the range of the attention mechanism, which is used to limit the range of the node set that can be focused on by any node in the multi-layer semantic graph when calculating attention. The window is not a fixed position sliding window in the traditional text sequence, but a graph path window based on the multi-layer semantic graph structure. The path distance between nodes in the multi-layer semantic graph refers to the shortest structural path length between two nodes in the semantic graph, which is usually measured by the number of hops in the shortest path in the graph, that is, the number of edges that need to be crossed from a node to another node. The local window range refers to the node set in the multi-layer semantic graph with all path distances not exceeding a preset threshold with a certain center node as a reference point.

[0086] The path centrality index of each node in the multi-layer semantic graph is used to measure the importance of the node as a node in the semantic information transmission path in the graph. The path centrality index is calculated as the frequency of the node appearing in all shortest paths, specifically as follows:

[0087]

[0088] Wherein, C pc (v q ) is the path centrality index of node v q ; represents the number of all shortest paths from node v q1 to node v q2 and passing through node v q ; represents the number of all shortest paths from node v q1 to node v q2 .

[0089] The knowledge graph fusion engine unit 2 is used to associate the terms in the extracted archive original text, construct the entity mapping relationship between the terms and the external knowledge graph, calculate the semantic association confidence score based on the semantic correlation between the terms and the entity nodes in the external knowledge graph, and set the semantic graph node expression weight in the summary generation model according to the semantic association confidence score, and construct an extended reference matrix integrated into the summary generation model;

[0090] In the embodiment, the knowledge graph fusion engine unit 2 comprises an entity mapping construction module 21;

[0091] The entity mapping construction module 21 is configured to identify terms from the original text of the archives, and construct a mapping path between a node in the multi-layer semantic graph and an entity node in the external knowledge graph based on structural features between the terms and the entity nodes in the external knowledge graph, specifically comprising the following steps:

[0092] S21.1, extract term nodes from the original text of the archives, and calculate a context semantic vector of each term node; extract attribute description information of entity nodes in the external knowledge graph, and construct a corresponding entity node semantic vector;

[0093] S21.2, based on the similarity between the context semantic vector of the term node and the entity node semantic vector, filter a candidate node pair that satisfies a preset semantic similarity threshold, and construct a candidate mapping set;

[0094] S21.3, respectively construct a context path subgraph of the candidate term node in the multi-layer semantic graph and a structural neighborhood subgraph of the candidate entity node in the external knowledge graph;

[0095] S21.4, respectively perform structural embedding coding on the context path subgraph and the structural neighborhood subgraph to obtain path structure feature representations, and calculate a joint confidence score of the mapping path based on the structural similarity between the path structure feature representations and the semantic similarity;

[0096] S21.5, determine a node pair with a joint confidence score exceeding a preset score threshold as a matched node, construct a path mapping structure thereof, form a node mapping path set, and finally obtain a mapping path between a node in the multi-layer semantic graph and an entity node in the external knowledge graph.

[0097] In the embodiment, the external knowledge graph refers to a collection of data structures that express terms, entities and semantic relationships between them in an external field in a structured graph form, which is usually derived from existing authoritative databases or automatically extracted through a graph construction tool, and includes entity nodes such as persons, place names, organizations, etc., and attribute edges such as belong to, contain, involve, etc.

[0098] The attribute description information of the entity node in the external knowledge graph refers to the descriptive text data, category label, upper and lower entity information and associated edge weight features carried by each entity node in the graph, which are used to represent the semantic boundary and context features of the node and serve as the basis for constructing the entity semantic representation vector;

[0099] The preset semantic similarity threshold refers to a judgment lower limit for judging whether the archive term node and the knowledge graph entity node satisfy the initial matching relationship. The threshold is an experience set or a similarity score boundary automatically selected through cross-validation, representing the minimum acceptance standard of the cosine similarity between semantic vectors.

[0100] The context path subgraph refers to a local path structure graph formed by extending outward from the archive term node to a set of adjacent nodes of a fixed step or semantic level in the multi-layer semantic graph, used to capture the structural nesting, co-occurrence link and semantic propagation characteristics of the term in the original context.

[0101] The structural neighborhood subgraph refers to a local subgraph structure formed by extending outward from the target entity node based on the connection relationship between nodes in the knowledge graph, including the direct associated nodes, semantic attribute edges and their adjacent relationships of the entity, usually controlled within two steps, used to construct the structural context representation of the graph entity.

[0102] In the embodiments S21.4 and S21.5, the joint confidence score of the mapping path is calculated, and the path mapping structure is constructed, and the specific method is as follows:

[0103] The context path subgraph of the term node in the multi-layer semantic graph is extracted, the structural adjacent relationship set in which the term node is located in the original archive text is constructed, and a semantic weight is assigned to each adjacent node in the set. The semantic weight is calculated according to the position depth and path frequency of the node in the path:

[0104]

[0105] Wherein, is the semantic weight of node v q ; is the path frequency of node v q ; is the position depth of node v q ; and λ is the path depth penalty factor.

[0106] The structural neighborhood subgraph of the target entity in the knowledge graph is extracted, the structural attribute context set of the entity is constructed, and an attribute description vector is calculated for each adjacent entity in the set. The semantic embedding matrix is generated using the word vector representation method.

[0107] The semantic correlation matrix R between the term context subgraph and the entity structural neighborhood subgraph is calculated. The joint confidence score is calculated comprehensively according to the maximum matching value and the average value of the matching path in the correlation matrix R. The joint confidence score is compared with the preset semantic similarity threshold. When the joint confidence score is greater than or equal to the preset semantic similarity threshold, it is determined that there is an effective mapping relationship between the term node and the entity node, and is recorded as a mapping path.

[0108] In the embodiment, the knowledge graph fusion engine unit 2 further comprises an association weight regulation module 22;

[0109] The association weight regulation module 22 calculates a semantic association confidence score based on semantic correlation between the term and the entity node in the external knowledge graph, and sets the semantic graph node expression weight in the summary generation model according to the semantic association confidence score, and specifically includes the following steps:

[0110] S22.1, extract the entity definition, attribute label and adjacency relationship of the entity node in the external knowledge graph, and construct the multi-dimensional vector embedding representation of the entity node;

[0111] S22.2, calculate the semantic correlation score between the context semantic vector of the term node and the semantic vector of the entity node using a bidirectional attention mechanism method, and calculate the semantic association confidence score combined with the structural position index of the term node in the multi-layer semantic graph;

[0112] S22.3, based on the semantic association confidence score, assign the semantic graph node expression weight of the multi-layer semantic graph node corresponding to the term node in the summary generation model.

[0113] In the embodiment, in the entity node in the external knowledge graph, the entity definition refers to the text description information of the standardized archive term represented by the graph node, which is the main semantic expression of the entity node; the attribute label refers to the attribute set associated with the entity node, such as category, source, timestamp, legal classification number, etc.; the adjacency relationship represents the semantic or logical connection between the entity node and other entity nodes;

[0114] The multi-dimensional vector embedding of the entity node adopts a combined feature learning mechanism based on the Transformer encoder to extract cross-domain entity semantics and perform multi-dimensional vector embedding.

[0115] In the embodiment, the extended reference matrix is a structured semantic reference matrix between the term node and the entity node in the external knowledge graph, which is used to describe the external knowledge graph entity path relationship referred to by the node in the semantic hierarchical structure in the multi-layer semantic graph;

[0116] The specific method for constructing the extended reference matrix is: extracting a set of term nodes in the multi-layer semantic graph and a set of candidate entity nodes in the external knowledge graph; calculating the joint weight value of each pair of nodes in the set of term nodes and the set of candidate entity nodes based on the entity mapping path and the semantic association confidence score, and constructing the extended reference matrix by taking the joint weight value greater than the preset reference confidence threshold;

[0117] The joint weight value greater than the preset reference confidence threshold value indicates whether the term node references the candidate entity node, and the size of the joint weight value represents the reference strength, and the greater the joint weight value, the greater the reference strength.

[0118] The paging index cache retrieval unit 3 is configured to perform fragment division processing on each node in the multi-layer semantic graph, construct an abstract content fragment index structure containing a node set, and filter out matching abstract content based on a user input abstract query condition using a matching page block filtering operation.

[0119] In this embodiment, the paging index cache retrieval unit 3 includes a semantic fragment division module 31, a fragment index construction module 32, and a condition matching retrieval module 33.

[0120] The semantic fragment division module 31 performs fragment clustering division on each node in the multi-layer semantic graph based on the node structure path relationship and semantic clustering features, and generates a plurality of abstract fragment subgraphs.

[0121] The fragment index construction module 32 is configured to encode each abstract fragment subgraph into fragment metadata containing node index numbers and path start and end markers, and construct a paging index table structure for user on-demand positioning and calling.

[0122] The condition matching retrieval module 33 constructs a query semantic vector based on a user input abstract query condition, and performs similarity scoring matching with each abstract fragment subgraph in the paging index table structure, and selects a target abstract fragment that satisfies the matching threshold.

[0123] The version tracing transaction management unit 4 is configured to record versions of the generation and modification operations of the abstract content, and construct a multi-abstract version index linked list structure to manage the abstract content.

[0124] In this embodiment, the multi-abstract version index linked list structure in the version tracing transaction management unit 4 is a unidirectional or bidirectional chain structure composed of a plurality of abstract version nodes, which is used to sequentially record, traceably manage, and track version content differences of historical abstract versions after the user performs operations on the abstract content.

[0125] The specific method for constructing the multi-abstract version index linked list structure is as follows: after each operation and editing of the abstract content, an abstract version node is constructed, the abstract version node comprising abstract compressed content, a timestamp, a unique identifier, and an original semantic fragment mapping record; if the current is the first generation of abstract content by the system, the first constructed abstract version node is taken as the start node of the linked list; otherwise, the current abstract version node is linked to the previous abstract version node, and a bidirectional index pointer is set; between each pair of continuous abstract version nodes, an abstract difference description vector is constructed using a word-level difference identification technology method, and is attached to the index structure between the pair of linked list nodes, and finally the multi-abstract version index linked list structure is obtained.

[0126] In the embodiment, the multi-abstract version index linked list structure is further provided with a version rollback interface for supporting a user to query a target version in the linked list according to a version number or a timestamp, and performing an abstract content rollback or difference comparison operation.

[0127] In embodiment two, the present application proposes an archive intelligent abstract generation method based on natural language processing technology, which is used for the archive intelligent abstract generation system based on natural language processing technology in the above-mentioned embodiment one, and comprises the following steps:

[0128] S10.1, performing semantic hierarchical segmentation analysis on the archive original text to construct a multi-layer semantic graph, using structure relationship node scoring on each node in the multi-layer semantic graph, constructing an abstract generation model based on the multi-layer semantic graph and the node scoring, and generating abstract content in combination with the structure path relationship of each node;

[0129] S10.2, correlating and extracting terms in the archive original text, constructing an entity mapping relationship between the terms and an external knowledge graph, calculating a semantic correlation confidence score based on the semantic correlation between the entity nodes in the terms and the external knowledge graph, setting an expression weight of a semantic graph node in the abstract generation model according to the semantic correlation confidence score, and simultaneously constructing an extended reference matrix integrated into the abstract generation model;

[0130] S10.3, performing fragment division processing on each node in the multi-layer semantic graph, constructing an abstract content fragment index structure comprising a node set, and using a matching page block filtering operation to filter out matching abstract content based on a user input abstract query condition;

[0131] S10.4, recording versions of the generation and modification operations of the abstract content, and constructing a multi-abstract version index linked list structure to trace manage the abstract content.

[0132] The embodiments of the present application are described in detail above in combination with the drawings, but the present application is not limited thereto, and various changes can be made within the knowledge range possessed by those skilled in the art without departing from the purpose of the present application.

Claims

1. An archive intelligent summary generation system based on natural language processing technology, characterized by, Comprise: A multi-layer semantic generation core unit (1) for segmenting and parsing the original text of the archives to construct a multi-layer semantic graph, scoring the nodes in the multi-layer semantic graph based on the structure relationship, constructing an abstract generation model based on the multi-layer semantic graph and node scoring, and generating abstract content combined with the structure path relationship of each node; A knowledge graph fusion engine unit (2) for associating and extracting terms in the original text of the archives, constructing the entity mapping relationship between the terms and the external knowledge graph, calculating the semantic association confidence score based on the semantic correlation between the terms and the entity nodes in the external knowledge graph, and setting the semantic graph node expression weight in the abstract generation model according to the semantic association confidence score, and constructing an extended reference matrix integrated into the abstract generation model; The knowledge graph fusion engine unit (2) comprises an entity mapping construction module (21) for identifying terms from the original text of the archives, constructing a mapping path between nodes in the multi-layer semantic graph and entity nodes in the external knowledge graph based on the structural features between the terms and the entity nodes in the external knowledge graph, and specifically comprising the following steps: S21.1, extract term nodes from the original text of the archives, and calculate the context semantic vector of each term node; extract the attribute description information of the entity nodes in the external knowledge graph, and construct the corresponding entity node semantic vector; S21.2, based on the similarity between the context semantic vector of the term node and the entity node semantic vector, filter the candidate node pairs that meet the preset semantic similarity threshold to form a candidate mapping set; S21.3, respectively construct the context path subgraph of the candidate term node in the multi-layer semantic graph and the structural neighborhood subgraph of the candidate entity node in the external knowledge graph; S21.4, respectively structure embedding coding of the context path subgraph and the structural neighborhood subgraph, get the path structure feature representation, calculate the joint confidence score of the mapping path based on the structural similarity between the path structure feature representation and the semantic similarity; S21.5, determine the node pair with joint confidence score exceeding the preset score threshold as the matching node, construct its path mapping structure, form the node mapping path set, and finally get the mapping path between the nodes in the multi-layer semantic graph and the entity nodes in the external knowledge graph; The knowledge graph fusion engine unit (2) further comprises an association weight regulation module (22) for calculating the semantic association confidence score based on the semantic correlation between the terms and the entity nodes in the external knowledge graph, and setting the semantic graph node expression weight in the abstract generation model according to the semantic association confidence score, specifically comprising the following steps: S22.1, extract the entity definition, attribute label and adjacency relationship of the entity nodes in the external knowledge graph, and construct the multi-dimensional vector embedding representation of the entity nodes; S22.2, calculate the semantic correlation score between the context semantic vector of the term node and the entity node semantic vector using the bidirectional attention mechanism method, and combine the structural position index of the term node in the multi-layer semantic graph to calculate the semantic association confidence score; S22.3, based on the semantic association confidence score, assign the semantic graph node expression weight of the multi-layer semantic graph node corresponding to the term node in the summary generation model; The paging index cache retrieval unit (3) is used for segmenting the nodes in the multi-layer semantic graph, constructing a summary content fragment index structure containing a node set, and filtering out matching summary content using a matching page block filtering operation based on a user input summary query condition; The version tracking transaction management unit (4) is used for version recording of the generation and modification operations of the summary content, and constructs a multi-summary version index linked list structure to track and manage the summary content. 2.The natural language processing technique based archive intelligent summary generation system of claim 1, wherein, The multi-layer semantic generation core unit (1) includes a semantic graph construction module (11); The semantic graph construction module (11) is used for semantic hierarchical segmentation analysis of the original archive text, and constructs a multi-layer semantic graph based on inter-sentence logical connection structure, theme word clustering boundary and paragraph pragmatic relationship; Wherein, the semantic hierarchical segmentation analysis includes syntax dependency analysis, semantic role labeling and context hierarchical segmentation, the original archive text is divided into a plurality of semantic units according to the grammatical structure relationship, and the subject-predicate-object structure and context dependence in each semantic unit are identified, and a semantic nested hierarchical text structure graph is constructed; Wherein, the semantic nested hierarchical text structure graph is used to construct a multi-layer semantic graph, and the specific method is as follows: The semantic layers are constructed based on the path depth of syntax dependency analysis, the semantic units in the same path depth are classified into the same semantic layer, and the intra-layer edges are established based on the semantic dependence relationship between the predicate argument structure, and the cross-layer edges are established based on the event pre-post context relationship; Wherein, the nodes of the multi-layer semantic graph are semantic units, the intra-layer edges of the multi-layer semantic graph are the syntax dependency relationships between the same semantic layer nodes, and the cross-layer edges of the multi-layer semantic graph represent the event context progressive relationship of different semantic layers. 3.The archive intelligent summarization generation system based on natural language processing technology of claim 2, wherein, The multi-layer semantic generation core unit (1) includes an abstract modeling control module (12); The abstract modeling control module (12) constructs an abstract generation model based on the structural relationship of each node in the multi-layer semantic graph and the structural relationship node score; Wherein, the structural relationship of each node in the multi-layer semantic graph includes the syntax dependency relationship between the same layer semantic units and the event context progressive relationship between the cross-layer semantic units; the structural relationship node score is weighted calculated based on the structural depth layer factor, the syntax connection centrality factor and the context aggregation density factor of each node in the semantic graph; Wherein, the structural depth layer factor represents the semantic level of the node in the multi-layer semantic graph; the syntax connection centrality factor represents the number of in-out edges of the node in the multi-layer semantic graph; the context aggregation density factor represents the density of co-occurring nodes in the context sliding window. 4.The archive intelligent summarization generation system based on natural language processing technology of claim 3, wherein, The abstract modeling control module (12) includes; The attention mechanism of the defined window is used to define the local window range of the attention mechanism calculation based on the path distance between nodes in the multi-layer semantic graph when constructing the summary generation model; and the guided semantic output control structure is used to preferentially generate the semantic unit content with high semantic weight based on the path centrality index and the structural relationship node of each node in the multi-layer semantic graph during the summary generation process. 5.The natural language processing technique based archive intelligent summary generation system of claim 4, wherein, The extended reference matrix is a structured semantic reference matrix between the term nodes and the external knowledge graph entity nodes, and is used to describe the path relationship of the external knowledge graph entity referenced by the nodes in the multi-layer semantic graph in the semantic hierarchical structure. The specific method for constructing the extended reference matrix is as follows: extracting a set of term nodes in the multi-layer semantic graph and a set of candidate entity nodes in the external knowledge graph; calculating the joint weight value of each pair of nodes in the set of term nodes and the set of candidate entity nodes based on the entity mapping path and the semantic association confidence score; and constructing the extended reference matrix by taking the joint weight values greater than the preset reference confidence threshold. The joint weight value greater than the preset reference confidence threshold indicates whether the term node references the candidate entity node, and the size of the joint weight value represents the reference strength, that is, the greater the joint weight value, the greater the reference strength. 6.The natural language processing technique based archive intelligent summary generation system of claim 5, wherein, The paging index cache retrieval unit (3) comprises a semantic segment division module (31), a segment index construction module (32), and a conditional matching retrieval module (33). The semantic segment division module (31) performs segment clustering division on each node in the multi-layer semantic graph based on the node structure path relationship and the semantic clustering feature, and generates a plurality of summary segment subgraphs. The segment index construction module (32) is used to encode each summary segment subgraph into segment metadata containing node index numbers and path start and end markers, and to construct a paging index table structure for user on-demand positioning and calling. The conditional matching retrieval module (33) constructs a query semantic vector based on the user input summary query condition, and performs similarity scoring matching with each summary segment subgraph in the paging index table structure to select a target summary segment that meets the matching threshold. 7.The natural language processing technique based archive intelligent summarization generation system of claim 6, wherein, In the version tracing transaction management unit (4), the multi-summary version index linked list structure is a unidirectional or bidirectional chain structure composed of a plurality of summary version nodes, and is used to sequentially record, traceably manage, and version content difference track the historical summary versions after the user operates the summary content. The specific method for constructing the multi-summary version index linked list structure is as follows: after each operation and editing of the summary content, a summary version node is constructed, which comprises summary compression content, a timestamp, a unique identifier, and original semantic segment mapping records; if the current is the first time that the system generates the summary content, the first constructed summary version node is taken as the starting node of the linked list; Otherwise, the current summary version node is linked to the previous summary version node, and a bidirectional index pointer is set; between each pair of continuous summary version nodes, a summary difference description vector is constructed using a word-level difference identification technology method, and is attached to the index structure between the pair of linked list nodes, and finally the multi-summary version index linked list structure is obtained.

8. An intelligent archive summary generation method based on natural language processing technology, used for an intelligent archive summary generation system based on natural language processing technology as claimed in any one of claims 1-7, characterized in that: The method comprises the following steps: S10.1, the original text of the archive is segmented and parsed to construct a multi-layer semantic graph, the nodes in the multi-layer semantic graph are scored based on the structure relationship, an abstract generation model is constructed based on the multi-layer semantic graph and the node score, and the abstract content is generated combined with the structure path relationship of each node; S10.2, the terms in the original text of the archive are extracted and associated, the entity mapping relationship between the terms and the external knowledge graph is constructed, the semantic correlation confidence score is calculated based on the semantic correlation between the entity nodes in the terms and the external knowledge graph, and the semantic graph node expression weight in the abstract generation model is set according to the semantic correlation confidence score, and an extended reference matrix is constructed and integrated into the abstract generation model; S10.3, each node in the multi-layer semantic graph is divided into segments, an abstract content segment index structure containing a node set is constructed, and matching page block filtering operation is used to filter out matching abstract content based on user input abstract query conditions; S10.4, record the version of the generation and modification operation of the abstract content, and construct a multi-abstract version index linked list structure to trace and manage the abstract content.

Images