Text data structurization processing method and system based on semantic recognition

Abstract

The application discloses a text data structured processing method and system based on semantic recognition, and relates to the technical field of natural language processing.The method comprises the following steps: introducing a predetermined recognition and extraction strategy to process standardized original text data, obtaining text key information and analyzing to obtain a first relationship triple, establishing a triple set, and then performing graph embedding analysis to obtain a target entity matrix, adjusting initial semantic features by combining a domain dictionary library after clustering processing to obtain target semantic features, and finally taking the structured result of the target semantic features as the structured result of the original text data.The application solves the technical problem that the traditional text data processing method cannot effectively utilize semantic recognition technology, cannot accurately capture deep semantic associations of the text, and cannot extract key information completely and accurately, and achieves the technical effect of semantic-level structured processing of text data and improvement of the completeness of key information extraction and the accuracy of structured results.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and particularly relates to a text data structured processing method and system based on semantic recognition. BACKGROUND

[0002] In the process of text data processing, the demand for semantic information extraction of unstructured text is increasing. The existing technology mostly uses rule-based or word frequency statistical methods to process text data. These methods are difficult to deeply understand the connotation of the text when facing complex semantic texts. Due to the insufficient use of semantic recognition technology, the traditional method cannot accurately capture the semantic association in the text, resulting in incomplete extraction of key information, insufficient accuracy of structured processing results, and difficulty in meeting the demand for deep analysis and application of text data. SUMMARY

[0003] The present application provides a text data structured processing method and system based on semantic recognition, which is used to solve the technical problem that the traditional text data processing method cannot accurately capture the deep semantic association of the text due to the ineffective use of semantic recognition technology, resulting in incomplete extraction of key information and inaccurate structured processing results.

[0004] In a first aspect, the present application provides a text data structured processing method based on semantic recognition, which comprises: introducing a predetermined recognition and extraction strategy to recognize and extract the original text data after standardization processing, to obtain text key information; analyzing the text key information to obtain a first relationship triple of a first entity, and establishing a triple set; calling a predetermined graph embedding strategy to perform graph embedding analysis on the triple set, to obtain a target entity matrix; performing vectorization clustering processing on the target entity matrix, to obtain a target clustering result, wherein the target clustering result comprises a plurality of clustering clusters; analyzing the plurality of clustering clusters to obtain initial semantic features, and adjusting the initial semantic features in combination with a domain dictionary library, to obtain target semantic features; and taking a target structured result of the target semantic features as a structured result of the original text data.

[0005] In a second aspect of the present application, a text data structured processing system based on semantic recognition is provided, comprising: a text key information acquisition module, configured to introduce a predetermined recognition and extraction strategy to perform recognition and extraction analysis on the original text data after standardization processing, to obtain text key information; a triple set construction module, configured to analyze the text key information to obtain a first relationship triple of a first entity, and construct a triple set; a target entity matrix acquisition module, configured to call a predetermined graph embedding strategy to perform graph embedding analysis on the triple set, to obtain a target entity matrix; a target clustering result acquisition module, configured to perform vectorization clustering processing on the target entity matrix, to obtain a target clustering result, wherein the target clustering result comprises a plurality of clustering clusters; a target semantic feature acquisition module, configured to analyze the plurality of clustering clusters to obtain initial semantic features, and adjust the initial semantic features in combination with a domain dictionary library, to obtain target semantic features; and a structured result acquisition module, configured to take a target structured result of the target semantic features as a structured result of the original text data.

[0006] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0007] In the present application, the semantic recognition technology is introduced to perform recognition and extraction analysis on the original text after standardization processing to obtain key information, to construct a relationship triple set and perform graph embedding analysis, to perform vectorization clustering and fuse and adjust semantic features with a domain dictionary library, to optimize a structured result in combination with a data enhancement strategy, so as to accurately realize semantic-level structured processing of text data, to make the semantic information extraction and processing result of unstructured text more accurate and reliable, and to achieve the technical effects of semantic-level structured processing of text data and improvement of key information extraction integrity and structured result accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0009] Figure 1 is a flowchart of a text data structured processing method based on semantic recognition provided by the embodiments of the present application.

[0010] Figure 2 is a structural diagram of a text data structured processing system based on semantic recognition provided by the embodiments of the present application.

[0011] The reference signs are explained as follows: a text key information acquisition module 1, a triple set construction module 2, a target entity matrix acquisition module 3, a target clustering result acquisition module 4, a target semantic feature acquisition module 5, and a structured result acquisition module 6. DETAILED DESCRIPTION

[0012] The application provides a text data structured processing method and system based on semantic recognition, which is used to solve the technical problem that the traditional text data processing method cannot effectively utilize semantic recognition technology, cannot accurately capture deep semantic association of text, and cannot extract key information completely and cannot obtain accurate structured processing result.

[0013] The technical solutions in the embodiments of the application will be clearly and completely described in the specification combined with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the application.

[0014] It should be noted that the terms "first", "second", and the like in the specification and the above drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or modules not clearly listed or inherent to the process, method, product or device.

[0015] Embodiment one, as shown in the figure, a text data structured processing method based on semantic recognition, wherein the method comprises: Figure 1

[0016] Step A100: introducing a predetermined recognition and extraction strategy to recognize and extract the original text data after standardization processing, to obtain text key information.

[0017] In the embodiments of the application, the predetermined recognition and extraction strategy refers to a combined strategy including an entity recognition plan and a relationship extraction plan.

[0018] Specifically, the entity recognition plan and the relationship extraction plan in the predetermined recognition and extraction strategy are used to respectively perform entity recognition processing and relationship extraction processing on the original text data after standardization processing, and the obtained entity set and relationship set jointly constitute the text key information, and the specific steps are described in detail in A110-A130.​

[0019] Step A200: analyzing the text key information to obtain a first relationship triple of a first entity, and assembling a triple set.

[0020] In the embodiment of the application, the first entity is any one of the entities in the entity set extracted from the original text data through the entity recognition plan. The first relationship triple is a triple structure composed of two entities and the relationship between them.

[0021] Optionally, first, two related entities are extracted from the obtained entity set, such as artificial intelligence and machine learning, and then the semantic relationship between them is matched from the relationship set, such as containing and association, so as to form a triple structure in the form of (entity 1, relationship, entity 2). For example, if the entity set contains system and user, and the relationship set contains the relationship of serving, then the first relationship triple of the first entity (system, serves, user) can be constructed.

[0022] This process is usually implemented by means of knowledge graph construction technology. First, the extracted entity is aligned with the external knowledge base (such as Wikipedia and general domain knowledge graph) through entity linking technology, for example, when the text contains Ruifeng Technology, the unique organization entity in the knowledge base is determined through entity linking to avoid the ambiguity problem of the same name entity; at the same time, the relationship disambiguation algorithm is used to calibrate the semantics of the preliminary extracted relationship, such as analyzing the context of the word connection in the text to determine whether it represents a technical interface connection or a cooperative relationship connection, so as to improve the accuracy of the relationship expression in the triple. In general domain text processing, when using the Neo4j graph database tool to store and manage triples, the entity set is used as a node in the graph database, each node contains entity name, type and other attributes, the relationship set is used as a directed edge between nodes, and the edge contains relationship type, weight and other attributes, for example, the triple structure of system (node) - serves (edge) - user (node) is constructed, the Cypher query language of Neo4j is used to support efficient semantic query, such as quickly searching all entities that have a research and development relationship with an artificial intelligence system, and using the graph traversal ability of the graph database to realize semantic reasoning, such as deducing the indirect association between entities through the relationship chain.

[0023] Further, by traversing the Cartesian product of the entity set and the relationship set, the entity-relationship combination that meets the semantic logic is filtered out, and a large number of triples are gradually generated to form a triple set. For example, in the processing scene of scientific and technological literature, about 12,000 valid triples can be constructed from the text key information containing 2000 entities and 500 relationships to form a structured knowledge unit set. This set not only retains the core semantic elements in the text, but also reveals the deep semantic network between entities through the association relationship of triples.

[0024] The conversion from unstructured text key information to structured triple set is realized through semantic association construction of entity set and relation set, triple validity verification and batch generation strategy, which provides a standardized knowledge representation form for subsequent graph embedding analysis, semantic clustering and other processing, and significantly improves the semantic structuring efficiency and application value of text data.

[0025] Step A300: calling a predetermined graph embedding strategy to perform graph embedding analysis on the triple set to obtain a target entity matrix.

[0026] In an embodiment of the present application, by calling a predetermined graph embedding strategy, the triple set containing multiple relations of any entity is extracted from the triple set, the target undirected graph is formed with the entity as the vertex and the relation as the edge, and then the target entity matrix is obtained by analysis. The specific steps are A310-A

[0027] 330 are described in detail.

[0028] Step A400: performing vectorization clustering processing on the target entity matrix to obtain a target clustering result, wherein the target clustering result includes multiple clustering clusters.

[0029] Specifically, first, the high-dimensional matrix data is converted into a vector representation suitable for clustering. Since the eigenvectors of the target entity matrix (Laplacian matrix) contain semantic association information in the graph structure, a spectral clustering algorithm is used to process it. Specifically, by performing eigenvalue decomposition on the target entity matrix, the eigenvectors corresponding to the first d smallest eigenvalues (d is a pre-set dimension reduction dimension, such as d=128) are extracted, and these eigenvectors are concatenated by columns to form a low-dimensional feature matrix. For example, when processing a target entity matrix containing 1000 entities, after eigenvalue decomposition, the first 128 eigenvectors are selected, the matrix dimension can be reduced from 1000x1000 to 1000x128, effectively reducing the data dimension while retaining more than 85% of the graph structure semantic information.

[0030] Subsequently, the low-dimensional feature matrix is standardized to eliminate the dimensional differences of different feature dimensions by Z-Score standardization. For example, the elements of each eigenvector are subtracted by the mean and divided by the standard deviation, so that the processed data has a mean of 0 and a variance of 1, avoiding the influence of extreme values on the clustering algorithm.

[0031] Then, the K-means clustering algorithm is used to cluster the standardized feature matrix. The optimal number of clusters k is determined by the elbow method (Elbow Method), for example, in the processing of scientific and technological field text, k usually takes a value of 5-10. In specific operation, k cluster centers are randomly initialized, the Euclidean distance of each vector to each center is calculated and assigned to the nearest cluster, and the cluster centers are iteratively updated until convergence.

[0032] Finally, the obtained target clustering result contains multiple clustering clusters, and the entities in each cluster have high similarity in semantic relationship. For example, when processing enterprise technical documents, clustering clusters such as artificial intelligence technology, data processing method, system architecture design can be formed, and the entities (such as machine learning algorithm, neural network model) in each cluster are closely connected through edge relationship (such as belongs to, is associated with).

[0033] By performing feature decomposition dimension reduction, standardization processing and K-means clustering on the target entity matrix, the effect of converting the text semantic structure into a semantic similar cluster is achieved, so that entities with similar semantic relationships are gathered together, providing hierarchical semantic units for subsequent semantic feature extraction and structured result generation, and significantly improving the semantic organization and interpretability of the text data.

[0034] Step A500: analyzing the plurality of clustering clusters to obtain initial semantic features, and adjusting the initial semantic features in combination with a domain dictionary library to obtain target semantic features.

[0035] In the embodiments of the present application, the initial semantic features are general semantic descriptions extracted from the clustering clusters, which are obtained by analyzing the common semantic features of the entities and relationships in the clustering clusters. The domain dictionary library is a pre-constructed specific domain term set, which stores standardized domain entity names, professional interpretations and semantic association relationships.

[0036] Specifically, a first clustering cluster is extracted from the plurality of clustering clusters, and a first general semantic feature is obtained by analyzing the first clustering cluster as an initial semantic feature. Then, the entities in the clustering cluster are traversed in the domain dictionary library, and a first domain semantic feature is obtained according to the traversal result. Then, the initial semantic feature is adjusted based on the domain semantic feature to obtain a target semantic feature, and the specific steps are described in detail in A510-A550.

[0037] Step A600: taking the target structured result of the target semantic feature as the structured result of the original text data.

[0038] In the embodiments of the present application, the target structured result is a standardized data structure formed by disassembling, mapping and integrating the target semantic feature, which carries the core semantic information of the original text in a machine-readable format.

[0039] Specifically, first, the target semantic feature is deeply disassembled by means of natural language processing technology. The core entities are located by a named entity recognition (NER) model, for example, extracting entities such as digital currency, distributed smart contract and storage technology from distributed smart contract storage technology applied to digital currency. This process uses a pre-training model based on BERT.

[0040] The BERT-based pre-training model is built on a bidirectional Transformer encoder architecture, containing 12 layers (Base version) or 24 layers (Large version) of network, with 768-dimensional (Base version) or 1024-dimensional (Large version) hidden units and 12 / 16 attention heads in each layer, capturing the context dependency of the text through self-attention mechanism. The model training is based on large-scale unlabeled text, which is first segmented into WordPiece subwords, generating input sequences containing mask markers [MASK], sentence separators [SEP]. The training process optimizes the model parameters through two major tasks: one is the Masked Language Model (MLM), which randomly selects 15% of the tokens to replace [MASK] (80%), random words (10%), or original words (10%), requiring the model to predict the original token; the second is the Next Sentence Prediction (NSP), which determines whether the input two sentences are continuous text, with 50% positive examples and 50% randomly selected negative examples from the corpus. The training uses the Adam optimizer with a learning rate of 5e-5, distributed training on 16 / 64 TPU, iterating for 1 million steps, about 40 epochs, finally enabling the model to have deep semantic understanding, with an average score of 80.5% on the 11 NLP tasks of the GLUE benchmark without fine-tuning, and significantly improving the performance of downstream tasks such as entity recognition and relation extraction, providing a strong context representation foundation for precise disassembly of target semantic features.

[0041] At the same time, the dependency syntax analysis is used to determine the relationship type between entities, such as applying to the association direction between distributed smart contracts and digital currencies, with an F1 value of 88% for relation extraction, ensuring that the entity set and relationship network are accurately stripped from the semantic features.

[0042] Subsequently, the information mapping is performed according to the predefined structured template of the text domain. The template is designed in a standardized format containing entity type, association relationship, domain label, semantic summary, etc. fields, such as the JSON structure commonly used in the financial technology field: {entity set: [digital currency, distributed smart contract, storage technology], relationship network: [{subject: distributed smart contract, relationship: applied to, object: digital currency}], domain attribute: financial technology, semantic summary: distributed smart contract storage technology applied to digital currency}. The template mapping algorithm is used to automatically fill the extracted entities and relationships into the corresponding fields, where the domain attribute directly inherits the annotation results from the domain dictionary library, such as financial technology.

[0043] Then, for complex features involving multi-level semantic association, such as blockchain -> underlying technology -> distributed storage -> support -> smart contract, a graph database technology is used to construct an entity relationship graph. Taking Neo4j as an example, each entity is taken as a node and attributes are attached, such as blockchain labeling technology architecture type, the relationship between entities is taken as an edge and the weight is set, such as the support relationship weight 0.85, to form a visual semantic network.

[0044] Finally, the processed template data or graph structure is output as the structured result of the original text, supporting multiple format exports such as JSON, CSV, XML, etc., and can be directly connected to downstream tasks such as knowledge graph construction and intelligent question answering, significantly enhancing the computability and interactivity of text data in intelligent systems.

[0045] Through semantic feature disassembly, domain template mapping, and graph structure modeling, the effect of converting unstructured text semantics into machine-readable standardized data structures is achieved, making the core information of the original text present in a clear hierarchical and logically rigorous manner, providing an efficient infrastructure for deep analysis and cross-system application of data, and significantly improving the practicality and technical value of text data structured processing.

[0046] Further, the step A100 in the method provided by the embodiment of the application comprises:

[0047] A110: performing entity recognition processing on the original text data according to an entity recognition plan in the predetermined recognition and extraction strategy to obtain an entity set.

[0048] A120: performing relationship extraction processing on the original text data according to a relationship extraction plan in the predetermined recognition and extraction strategy to obtain a relationship set.

[0049] A130: the entity set and the relationship set constitute the text key information.

[0050] In the embodiment of the application, the entity recognition plan is a strategy in the predetermined recognition and extraction strategy for performing entity recognition processing on the standardized processed original text data to obtain an entity set. The relationship extraction plan is a strategy in the predetermined recognition and extraction strategy for performing relationship extraction processing on the original text data to obtain a relationship set.

[0051] Specifically, in the text data structured processing, first, the original text is standardized processed, and through word segmentation, part-of-speech tagging, and removing stop words, etc., the unstructured text is converted into a standardized processable format, laying a foundation for subsequent recognition and extraction. In the text processing scenario, standardization processing can improve the format uniformity of text data, significantly reducing the noise interference of subsequent model processing.

[0052] Next, the original text data is identified to obtain a candidate entity set by using a preset classifier, including a CRF classifier, an HMM classifier and a conditional random field, and the candidate entity set is verified and analyzed based on a weighted fusion strategy to obtain an entity set, which is specifically described in A111-A113.

[0053] Then, when performing relation extraction processing on the original text data, first, based on the identified entity set, all possible entity pairs in the text after standardization processing are located. For example, if the entity set contains two entities of a company and a product, the system will search for the paragraphs or sentences in which the two entities appear simultaneously in the text to form a candidate entity pair set. This step relies on entity position labeling technology to ensure accurate matching of entity pairs by marking the start and end positions of each entity in the text.

[0054] Next, the context of the candidate entity pair is extracted using the core algorithm in the relation extraction plan. Deep learning models are used to learn features such as word sequences and syntactic structures around the entity pair. For example, a bidirectional long short-term memory network (BiLSTM) is used to capture semantic dependencies in the text, and an attention mechanism is used to focus on key words for relation judgment. In experiments on a certain public data set (such as the NYT relation extraction data set), the feature extraction accuracy of this type of model for common entity relationships (such as ownership and production) can reach more than 85%.

[0055] Then, the classifier is used to determine the relationship type of the extracted features. The classifier can use a support vector machine (SVM) or a multi-layer perceptron (MLP) model, which is pre-trained on a labeled relationship data set. For example, after training on a data set containing 100,000 labeled samples, the model's F1 score for identifying types such as cooperation and competition relationships can reach 90%. The system assigns one or more relationship types to each candidate entity pair to form a preliminary relationship set.

[0056] After that, a rule verification mechanism is introduced to optimize the preliminary relationship set. For example, if there are contradictory relationship descriptions for the same entity pair in the same paragraph, the rule is to use the relationship that appears more frequently, or to verify the reasonableness of the relationship through a domain dictionary library, such as a company and a product being more likely to have a production relationship than an employment relationship. In practical applications, this rule verification can improve the accuracy of relation extraction by about 5-8%.

[0057] Finally, after the above steps, the relationship set obtained after processing forms the key information of the text together with the entity set. For example, when processing an enterprise news text, the entity set may contain certain city Rui Fu Technology Co., Ltd. and artificial intelligence system, and the relationship set may contain R&D-artificial intelligence system and release-2025 June, etc. The combination of the two can fully describe the core information in the text.

[0058] By positioning entity pairs, extracting context features, classifying relationship types, and rule checking, and combining the advantages of existing deep learning models and rule methods, the effect of accurately extracting the relationship between entities in the text is achieved, so that the key information of the text can be fully and structuredly presented, and reliable relationship data support is provided for subsequent triple construction and semantic analysis.

[0059] Further, the step A110 in the method provided by the embodiment of the application comprises:

[0060] A111: obtaining a candidate entity set of the original text data by a preset classifier.

[0061] A112: verifying and analyzing the candidate entity set based on a weighted fusion strategy to obtain the entity set.

[0062] A113: wherein the preset classifier comprises a CRF classifier, an HMM classifier, and a conditional random field.

[0063] In the embodiment of the application, the preset classifier refers to a classifier used to identify the candidate entity set of the original text data.

[0064] Optionally, when performing entity recognition processing on the original text data, first, the standardized text data is input into the preset classifier. These preset classifiers include a CRF classifier, an HMM classifier, and a conditional random field, each of which analyzes the text based on different algorithm logic.

[0065] When constructing the CRF classifier, first, a training corpus containing entity annotations is prepared, for example, 100,000 text sentences with annotated entity types (such as organization name, product name, etc.) are selected. Then, feature engineering is performed on the corpus, and features such as words, context words before and after the word, part-of-speech labels, and text position information are extracted, and these features are converted into feature functions recognizable by the CRF model. Then, the L-BFGS optimization algorithm is used to train the model, and the model parameters are adjusted through iteration to make the entity annotation accuracy of the model on the training set reach more than 85%, and finally a CRF classifier with sequence feature capturing capability is obtained, which can accurately identify the entity boundary in the text based on the context.

[0066] When constructing the HMM classifier, first define the set of hidden states, such as setting the state categories of organization entity, product entity, location entity, etc., and the observation sequence is the word sequence in the text. Then, the initial state probability is calculated from the annotated corpus, for example, the probability of starting with an organization entity in all sentences is 0.35; the state transition probability is calculated, such as the probability of an organization entity followed by a product entity is 0.28; the emission probability is obtained by maximum likelihood estimation, i.e. the probability of a certain word belonging to a specific entity type, such as the probability of Ruisi belonging to the organization entity is 0.92. To avoid zero probability problem, Laplace smoothing is performed on the probability parameters. The final trained HMM classifier can achieve a recall rate of 80% on the common entity types in the public test set, and can predict the entity types in the text based on the probability model.

[0067] When constructing the conditional random field classifier, the linear chain conditional random field is used as the basic framework. First, a multi-dimensional feature vector is constructed, including word-level features (such as the current word, the left and right two words), syntactic features (such as dependency syntactic relations), and domain dictionary features (such as whether it appears in the preset product word table). Then, annotated cross-domain text data is used as the training set, such as 50,000 samples containing enterprise news and technical documents. The gradient descent algorithm is used to optimize the potential function parameters of the model, so that the entity recognition F1 score of the model on the training set reaches 88%. This classifier can combine the context and domain knowledge of the text to jointly predict the entity boundary and type, and has good robustness in handling long-distance dependent entity recognition tasks.

[0068] After the above classifiers process the standardized text data, the corresponding candidate entity sets will be output. There may be some repeated or different entity recognition results in these candidate entity sets.

[0069] Subsequently, the above candidate entity sets are verified and analyzed based on a weighted fusion strategy. This strategy assigns different weight coefficients to each preset classifier based on its performance in historical similar text recognition tasks. Assuming that in the past experiments, the CRF classifier has an accuracy of 85% in recognizing entities in a certain type of text, the HMM classifier has an accuracy of 78%, and the conditional random field has an accuracy of 82%, then the weights can be set as 0.4, 0.3, and 0.3 respectively. Then, for each entity in each candidate entity set, the recognition results and corresponding weights in different classifiers are weighted to obtain a comprehensive score. When the comprehensive score of an entity exceeds a preset threshold, it is included in the final entity set.

[0070] By using multiple preset classifiers to identify candidate entity sets and combining the weighted fusion strategy to verify the results, the accuracy and robustness of entity recognition are improved, providing a more reliable entity data basis for subsequent extraction of key information from text and overall structured processing.

[0071] Further, the step A300 in the method provided by the embodiments of the present application comprises:

[0072] A310: extracting any triple in the set of triples, and the any triple comprising a plurality of relation triples of any entity.

[0073] A320: forming a target undirected graph according to the predetermined graph embedding strategy, taking the any entity as a vertex and the plurality of relation triples as an edge.

[0074] A330: analyzing to obtain the target entity matrix of the target undirected graph.

[0075] Specifically, first, a plurality of relation triples containing any entity are extracted from the set. For example, when processing a text in the field of science and technology, if an entity is an artificial intelligence system, the corresponding relation triples of the entity can include (artificial intelligence system, applied to, financial field), (artificial intelligence system, dependent on, machine learning algorithm), (artificial intelligence system, developer, technology company), etc. This step filters out all the relation descriptions centered on the entity by traversing the set of triples, forms a local relation network, and ensures the richness of the subsequent graph structure construction.

[0076] Then, according to the predetermined graph embedding strategy, the extracted entity is taken as a vertex, and the corresponding plurality of relation triples are taken as edges connecting the vertices, to construct a target undirected graph. Taking the artificial intelligence system as an example, the artificial intelligence system is taken as a central vertex, related entities such as the financial field, the machine learning algorithm, and the technology company are taken as adjacent vertices, and each relation type (such as applied to, dependent on, and developer) is converted into an attribute of an edge, to realize storage and visualization through the node-relation model of the Neo4j graph database, that is, each entity corresponds to a node in the graph database, with an entity type attribute, each relation corresponds to a directed edge, and the edge attribute records the relation type. Since the undirected graph is constructed, the direction of the edge does not affect the vertex connection relationship, that is, the vertices at both ends of the edge are adjacent to each other.

[0077] After the target undirected graph is formed, a target entity matrix is further analyzed and obtained, and a target degree matrix and a target adjacency matrix of the target undirected graph are sequentially obtained, and the difference between the two is recorded as the target entity matrix. The specific steps are described in detail in A331-A332.

[0078] By extracting the entity relation triples to construct an undirected graph and converting it into a matrix representation, the effect of converting the semantic information of unstructured text into a calculable mathematical model is achieved, a standardized data structure is provided for subsequent clustering analysis and semantic feature extraction based on matrix operation, and the efficiency and accuracy of text semantic mining are significantly improved.

[0079] Further, the method provided in the embodiments of the present application further comprises the following step A330:

[0080] A331: sequentially obtaining a target degree matrix and a target adjacency matrix of the target undirected graph.

[0081] A332: taking the difference between the target degree matrix and the target adjacency matrix as the target entity matrix.

[0082] Specifically, first, the target adjacency matrix of the graph is obtained. Taking an undirected graph containing N vertices as an example, the adjacency matrix is an N*N matrix. If there is a connection edge between vertex i and vertex j, the element values of the i-th row and the j-th column and the j-th row and the i-th column in the matrix are 1, and the undirected graph edge is symmetrical, otherwise 0. For example, an undirected graph containing three vertices of artificial intelligence system, machine learning algorithm and financial field, if the artificial intelligence system has a connection with the other two vertices, the corresponding position in the adjacency matrix is 1, and the rest is 0. The matrix is generated by traversing the connection relationship of all edges in the graph. In actual application, the Cypher query of Neo4j graph database can efficiently extract the adjacency matrix. For a graph with 100,000 vertices, the adjacency matrix extraction time is about 200 milliseconds.

[0083] Then the target degree matrix is obtained. The degree matrix is a diagonal matrix, and the main diagonal elements represent the degree of the corresponding vertex, that is, the number of edges connected to the vertex, and the rest of the elements are 0. Continue to take the above three-vertex graph as an example, if the artificial intelligence system connects 2 edges, the machine learning algorithm and the financial field each connect 1 edge, then the diagonal elements of the degree matrix are 2, 1 and 1 respectively. The calculation of the degree matrix can be obtained by counting the number of non-zero elements in each row (or column) of the adjacency matrix. In graph theory algorithm, the time complexity of this step is O(N+E), where N is the number of vertices and E is the number of edges. For a medium-sized graph (N=1000, E=5000), the generation time of the degree matrix is less than 50 milliseconds.

[0084] After obtaining the target degree matrix and the target adjacency matrix, the difference between the degree matrix and the adjacency matrix is obtained, which is the target entity matrix, that is, the Laplacian matrix of the graph. Taking the three-vertex graph as an example, after the degree matrix is subtracted from the adjacency matrix, the non-diagonal elements are -1 (corresponding to the position of the edge), and the diagonal elements are the degrees of the vertices. The matrix plays an important role in graph embedding technology, for example, in spectral clustering algorithm, the eigenvectors of the Laplacian matrix can be used to represent the vertex embedding of the graph.

[0085] By sequentially constructing the adjacency matrix and the degree matrix of the target undirected graph, and calculating the difference between the two to obtain the Laplacian matrix, the effect of converting the semantic information of the graph structure into a calculable matrix form is achieved, which provides a mathematical basis for subsequent vectorization clustering and semantic feature extraction based on matrix operation, and effectively improves the accuracy and efficiency of the structured processing of text data.

[0086] Further, the step A500 in the method provided by the embodiments of the present application comprises:

[0087] A510: extracting a first clustering cluster in the plurality of clustering clusters.

[0088] A520: analyzing the first clustering cluster to obtain a first general semantic feature, and taking the first general semantic feature as the initial semantic feature.

[0089] A530: traversing entities in the first clustering cluster in the domain dictionary library to obtain a traversal result.

[0090] A540: analyzing the first domain semantic feature of the first clustering cluster according to the traversal result.

[0091] A550: adjusting the initial semantic feature based on the first domain semantic feature to obtain the target semantic feature.

[0092] In the embodiments of the present application, the domain dictionary library is a pre-constructed structured knowledge base, which stores standardized terms in a specific domain (such as science and technology, finance, etc.) and corresponding explanations, domain classification and association.

[0093] Specifically, first, a first clustering cluster is extracted from the target clustering result. The cluster is usually selected from the cluster with the most entities or the most closely related semantic associations, for example, in the processing of science and technology texts, the clustering cluster containing high-frequency entities such as artificial intelligence and machine learning is preferentially extracted.

[0094] Next, the first clustering cluster is analyzed for general semantic features. Through word frequency statistics and co-occurrence relationship analysis, the common semantics of entities and relationships in the cluster are extracted. For example, if the cluster contains entities such as neural network models, deep learning frameworks, convolutional calculations, and relationships such as belonging to and being applied to, the artificial intelligence algorithm technology can be generalized as the first general semantic feature. This process usually uses the TF-IDF algorithm to calculate the term weight, and the specific process is as follows:

[0095] In the process of calculating the term weight by the TF-IDF algorithm, firstly, the occurrence frequency (TF) of each term in the single document in the cluster is counted, and the normalization processing (such as dividing the total number of words in the document) is performed to eliminate the influence of the difference in document length. Then the inverse document frequency (IDF) is calculated, which reflects the rarity of the term in the entire data set by dividing the total number of documents in all clusters by the number of documents containing the term and taking the natural logarithm. The TF-IDF value of a certain term is the product of its TF and IDF. The higher the value, the higher the importance of the term in the current cluster and the lower the frequency of its appearance in other clusters, and the more suitable it is as a key term to distinguish the semantics of the cluster. For example, in the processing of the 20Newsgroups data set, the TF-IDF values of terms such as algorithm and program are significantly higher than those in other clusters. By selecting the top 20% of the terms with the highest TF-IDF values, the semantic aggregation can accurately extract general semantic features related to computer technology, providing a quantitative weight basis and scientific screening standard for the initial semantic features.

[0096] Subsequently, the entities in the first cluster are iteratively queried in the domain dictionary library. The domain dictionary library stores standardized domain terms and explanations, such as the explanation of the Transformer model in the science and technology domain dictionary library being a sequence modeling architecture based on self-attention mechanism. The iteration process is performed by combining exact matching and fuzzy matching, for example, when the entity CNN exists, the dictionary library can return the complete explanation and domain classification of convolutional neural network.

[0097] According to the obtained iteration results, the first domain semantic features of the first cluster are analyzed. For example, if the entity BERT in the cluster corresponds to a pre-trained language model in the dictionary library, which is used for natural language understanding tasks, the domain semantic features can be refined to the pre-trained model technology in the natural language processing domain. By aggregating the domain explanations of all entities, a domain-specific semantic description of the cluster is formed. In the domain text test, this method can improve the domain relevance of semantic features by 15%-20%.

[0098] Finally, the general explanation of the entity is obtained, and the domain explanation in the domain dictionary library is fused at the first level. Then the first general semantic features and the first domain semantic features are fused at the second level. Finally, the target semantic features are obtained by integrating the two-level fusion results. The specific steps are described in detail in A551-A555.

[0099] By extracting the core cluster, analyzing the general semantic features, combining the domain dictionary library iteration, and aggregating the domain features, the general semantic features and the domain professional knowledge are fused, which makes the target semantic features not only retain the common semantic framework of the text, but also have accurate domain-specific explanations, providing key support for the accuracy and professionalism of the subsequent text structured results.

[0100] Further, the method provided in the embodiments of the present application further comprises the following step A550:

[0101] A551: obtaining a general interpretation of the entity.

[0102] A552: obtaining a domain interpretation corresponding to the entity in the traversal result.

[0103] A553: performing first-level fusion on the general interpretation and the domain interpretation to obtain a first-level fusion result.

[0104] A554: performing second-level fusion on the first general semantic feature and the first domain semantic feature to obtain a second-level fusion result.

[0105] A555: performing fusion on the first-level fusion result and the second-level fusion result to obtain the target semantic feature.

[0106] In the embodiments of the present application, the general interpretation is a non-domain-specific explanation obtained from a public general knowledge base, which describes the basic concept of the entity in simple and easy-to-understand language, and does not involve professional terms or deep application scenarios of a specific industry. The domain interpretation is a professional explanation stored in a pre-constructed domain dictionary library, which defines the entity for a specific domain (such as technology, finance), and contains domain-specific terms, technical principles or application scenarios.

[0107] In one embodiment, the general interpretation and the domain interpretation of the entity are obtained first. The general interpretation can be obtained from a public knowledge base such as Wikipedia or Baidu Encyclopedia, for example, the general interpretation of the entity blockchain is a distributed database technology that ensures data cannot be tampered with through cryptographic methods; the domain interpretation is extracted from a pre-constructed domain dictionary library, such as the interpretation of blockchain in the financial domain dictionary library, which is a distributed ledger technology applied to digital currency and smart contracts, with high security and transparency.

[0108] Then the two types of paraphrase are fused at the first level. Adopting the word vector weighted fusion strategy, first convert the general paraphrase and domain paraphrase into multi-dimensional vector representation to construct the word vector space through the Word2Vec model. The model is based on the Skip-Gram architecture and uses the negative sampling technique, with 5 negative samples. Train on the mixed data set of the domain corpus (such as 10 million financial texts) and the general corpus (such as the full text of Wikipedia). Set the sliding window size to 8, the vector dimension to 300, and the number of iterations to 15, so that the generated word vector can capture both general semantic and domain semantic features. After training, input the segmented general paraphrase and domain paraphrase into the model to get the corresponding multi-dimensional vector representation. Then calculate the cosine similarity of the two. Assuming that the similarity of the general and domain paraphrase vectors of the blockchain is 0.72, and then dynamically allocating weights according to the similarity, when the similarity is greater than or equal to 0.7, the weight of the general paraphrase is 0.4 and the weight of the domain paraphrase is 0.6. When fused, the vector elements at the corresponding positions are linearly combined according to the weights to obtain the first fusion result vector.

[0109] Then the second level fusion is performed to combine the first general semantic feature and the first domain semantic feature. The general semantic feature is generated by clustering the word frequency statistics, such as the general feature of the blockchain technology cluster being the distributed data storage method. The domain semantic feature is extracted based on the keyword aggregation of the dictionary library paraphrase, such as digital currency application and smart contract technology. The TextRank algorithm is used to sort the keywords of the two types of features, and the specific process is as follows:

[0110] Step a: Tokenize the first general semantic feature and the first domain semantic feature and construct a vocabulary, for example, the general feature distributed data storage method is tokenized as [distributed, data, storage, method], and the domain feature digital currency application, smart contract technology is tokenized as [digital, currency, application, smart, contract, technology], and after merging and removing duplicates, an initial vocabulary containing multiple words is formed. Then build a word co-occurrence graph, with nodes representing words in the vocabulary. If two words co-occur in the same feature phrase (window size set to 2), an undirected edge is established between the nodes and the weight is assigned, such as distributed and storage co-occurring 2 times, with an edge weight of 2.

[0111] Step b: iteratively calculate the PageRank value of each node, and initially set the PageRank value of all nodes to 1. In each iteration, the PageRank value of node i is updated to PR(i) = (1-d) + d x ∑(PR(j) / In(j)), where d is the damping coefficient (set to 0.85), j is the node pointing to i, and In(j) is the out-degree of node j. After 10 iterations, the PageRank values of each word are obtained, such as distributed 0.12, smart contract 0.15, and digital currency 0.14. After sorting in descending order of value, the top 5 keywords are selected, such as distributed, smart contract, digital currency, storage, and technology. According to the source, the domain attribute is labeled, such as digital currency labeled as financial domain.

[0112] Step c: perform keyword fusion, and preferentially retain domain attribute words such as digital currency and smart contract. General words (such as distributed and storage) are adjusted in terms of domain relevance, such as distributed storage adjusted to distributed ledger storage, and domain-related verbs are supplemented, such as applied to, and finally integrated into distributed smart contract storage technology applied to digital currency.

[0113] Finally, the primary fusion result and the secondary fusion result are deeply integrated. A feature fusion neural network is constructed, including 1 hidden layer with 128 nodes. The vector representation of the primary fusion and the keyword vector of the secondary fusion are input as input, and the vector representation of the target semantic feature is output.

[0114] Through hierarchical entity interpretation, word vector weighted fusion, keyword aggregation, and neural network deep integration, the general semantic and domain professional knowledge are deeply combined, the target semantic feature maintains the basic semantic framework of the text, and has domain-specific accurate expression, which provides core support for the professionalism and accuracy of the text structured result.

[0115] Further, the method provided in the embodiment of the application comprises the following steps A700:

[0116] A710: processing the original text data according to a predetermined data enhancement strategy to obtain enhanced text data.

[0117] A720: optimizing the target structured result according to the enhanced text data.

[0118] A730: wherein the predetermined data enhancement strategy comprises synonym replacement and sentence reorganization.

[0119] Optionally, before structuring the original text data, the data quality is improved through a predetermined data enhancement strategy. First, synonym replacement is implemented, and the WordNet synonym dictionary in the NLTK library (Natural Language Toolkit) or the BERT-based (Bidirectional Transformer Encoder Representation) synonym generation model is used to replace 10%-20% of the non-core words in the text. Core words are selected by TF-IDF weight, and words with a weight greater than or equal to 0.3 are not replaced. For example, the technology in artificial intelligence technology is replaced by technology, and the cosine similarity of the word vectors before and after replacement is calculated through the Word2Vec model (threshold set to 0.7) to ensure semantic consistency.

[0120] Next, sentence restructuring is performed, and based on the dependency syntax analysis results of the spaCy library, the syntax structure of long sentences is adjusted. Specifically, the order of parallel clauses is exchanged, such as he developed the algorithm and applied it to the system, which is restructured to he applied the algorithm to the system and completed the development; active and passive sentences are converted, such as the team trained the model, which is changed to the model was trained by the team, and the surface structure is changed while keeping the propositional logic unchanged. When restructuring, a sentence length threshold is set, and sentences longer than 20 words are restructured to avoid semantic deviation caused by excessive conversion.

[0121] After obtaining the enhanced text data, the entire process from constructing the target undirected graph to generating the target semantic features is re-executed: the triples in the enhanced text are reconstructed into a graph structure, and the difference between the degree matrix and the adjacency matrix is calculated to obtain the target entity matrix. After spectral clustering and domain semantic fusion, a new structured result is generated. By comparing the structured results of the original and enhanced data, the edit distance and semantic similarity (such as BERTScore) are used as double indicators for optimization: when the BERTScore of the structured result of the enhanced data and the original result is greater than or equal to 0.9, the original result is retained; if it is lower than the threshold, the intersection of entities and the union of relationships of the two are taken to generate the optimized result after fusion.

[0122] By synonym replacement to improve vocabulary diversity and sentence restructuring to enrich text structure, combined with the whole process of enhanced data re-computation and result fusion strategy, the robustness of structured processing is strengthened, the final structured result can still maintain high accuracy when facing text expression variation, effectively solving the problem of insufficient structured generalization ability caused by single expression of original text, and providing a more reliable technical solution for data processing in complex semantic scenarios.

[0123] Further, the step A700 in the method provided by the embodiment of the application comprises:

[0124] A740: displaying the target structured result through a user interface.

[0125] In one embodiment, when the target structured result is displayed, a knowledge graph visualization interface is first constructed, entities in the target structured result are converted into nodes, and relationships are converted into edges. Each node is set to a different color according to the field to which it belongs, and the size is adjusted according to the importance. The thickness of the edge represents the relationship strength. The user can view the detailed information of the entity by hovering the mouse, support node dragging and relationship path highlighting functions, and when processing a document containing 500 entities, the rendering speed can reach 60 frames per second.

[0126] At the same time, a table view component is developed to convert data such as entity sets and relationship networks in the structured result into a sortable and filterable table. Modern interface design is adopted, and the table row height is set to 48 pixels to ensure content readability, and the table header is fixedly displayed. Virtual scrolling technology is integrated, and when processing large data sets of more than 10,000 records, memory occupancy is reduced by 70%, scrolling frame rate is maintained at more than 50 frames per second, and the table supports multi-field filtering, and filtering response time is controlled within 200 milliseconds.

[0127] To support deep semantic analysis, a semantic association analysis module is developed. Based on the target structured result, the semantic similarity between entities is calculated, and when the similarity reaches a certain threshold, an association prompt is established. In the interface, the entity association strength is displayed in the form of a heat map, and the user can view the first and second level association networks by clicking any entity.

[0128] Finally, mobile adaptation is realized, and the responsive design principle is adopted. When the screen width is less than 768 pixels, it automatically switches to a vertical layout. The touch interaction experience is optimized, and the target area is expanded to 44x44 pixels, supporting operations such as zooming in on the knowledge graph, sliding to filter table data, etc. Through background thread processing of complex computing tasks, smooth experience is ensured on mobile devices, and the interface response time on mainstream mobile devices is averaged at 120 milliseconds.

[0129] Through the steps of building a knowledge graph visualization, a responsive table view, a semantic association analysis module, and mobile adaptation, the abstract structured result is converted into an intuitive and easy-to-use visualization interface, enabling users to efficiently explore the semantic network of text data, significantly improving the efficiency of understanding complex semantic information and decision support capabilities, and providing a friendly interactive interface for deep application of text data.

[0130] In summary, the text data structured processing method based on semantic recognition provided by the embodiments of the present application has the following technical effects:

[0131] This application extracts the first cluster from multiple clusters, analyzes this cluster to obtain initial semantic features, and then traverses the domain dictionary to obtain domain semantic features for the entities within it. The initial semantic features are adjusted to obtain target semantic features, which are then obtained through entity interpretation fusion, semantic feature fusion, and other steps to obtain the final target semantic features. Finally, the original text data is enhanced, the structured results are optimized, and a user interface is constructed for display. This achieves high-precision semantic recognition and structured processing of the original text data, making the structured results more accurate, intuitive, and practical. It achieves the technical effect of semantic-level structured processing of text data, improving the completeness of key information extraction and the accuracy of structured results.

[0132] Example 2, as Figure 2 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides a text data structuring processing system based on semantic recognition, the system comprising:

[0133] The text key information acquisition module 1 is used to introduce a predetermined recognition and extraction strategy to identify, extract and analyze the standardized original text data to obtain text key information.

[0134] The triple set construction module 2 is used to analyze the key information of the text to obtain the first relation triples of the first entity and construct the triple set.

[0135] The target entity matrix acquisition module 3 is used to invoke a predetermined graph embedding strategy to perform graph embedding analysis on the triple set to obtain the target entity matrix.

[0136] The target clustering result acquisition module 4 is used to perform vectorized clustering processing on the target entity matrix to obtain the target clustering result, wherein the target clustering result includes multiple clusters.

[0137] The target semantic feature acquisition module 5 is used to analyze the multiple clusters to obtain initial semantic features, and adjust the initial semantic features in combination with the domain dictionary to obtain target semantic features.

[0138] The structured result acquisition module 6 is used to take the target structured result of the target semantic feature as the structured result of the original text data.

[0139] Furthermore, the text key information acquisition module 1 is used to perform the following steps:

[0140] According to the entity recognition plan in the predetermined recognition and extraction strategy, the original text data is subjected to entity recognition processing to obtain an entity set; according to the relationship extraction plan in the predetermined recognition and extraction strategy, the original text data is subjected to relationship extraction processing to obtain a relationship set; the entity set and the relationship set form the text key information.

[0141] Further, the text key information acquisition module 1 is configured to perform the following steps:

[0142] A candidate entity set of the original text data is obtained through a preset classifier; the candidate entity set is subjected to verification analysis based on a weighted fusion strategy to obtain the entity set; wherein the preset classifier includes a CRF classifier, an HMM classifier and a conditional random field.

[0143] Further, the target entity matrix acquisition module 3 is configured to perform the following steps:

[0144] Any triple in the triple set is extracted, and the any triple includes a plurality of relationship triples of any entity; according to the predetermined graph embedding strategy, the any entity is taken as a vertex and the plurality of relationship triples are taken as edges to form a target undirected graph; the target entity matrix of the target undirected graph is obtained by analysis.

[0145] Further, the target entity matrix acquisition module 3 is configured to perform the following steps:

[0146] A target degree matrix and a target adjacency matrix of the target undirected graph are sequentially obtained; the difference between the target degree matrix and the target adjacency matrix is denoted as the target entity matrix.

[0147] Further, the target semantic feature acquisition module 5 is configured to perform the following steps:

[0148] A first clustering cluster in the plurality of clustering clusters is extracted; a first general semantic feature of the first clustering cluster is obtained by analysis, and the first general semantic feature is taken as the initial semantic feature; entities in the first clustering cluster are traversed in the domain dictionary library to obtain a traversal result; a first domain semantic feature of the first clustering cluster is obtained by analysis according to the traversal result; the initial semantic feature is adjusted based on the first domain semantic feature to obtain the target semantic feature.

[0149] Further, the target semantic feature acquisition module 5 is configured to perform the following steps:

[0150] The general interpretation of the entity is obtained, the domain interpretation corresponding to the entity in the traversal result is obtained, first-level fusion is performed on the general interpretation and the domain interpretation to obtain a first-level fusion result, second-level fusion is performed on the first general semantic feature and the first domain semantic feature to obtain a second-level fusion result, and the first-level fusion result and the second-level fusion result are fused to obtain the target semantic feature.

[0151] Further, the structured result obtaining module 6 is configured to perform the following steps:

[0152] The original text data is processed according to a predetermined data enhancement strategy to obtain enhanced text data, and the target structured result is optimized according to the enhanced text data; wherein the predetermined data enhancement strategy includes synonym replacement and sentence reorganization.

[0153] Further, the structured result obtaining module 6 is configured to perform the following steps:

[0154] The target structured result is displayed through a user interface.

[0155] The text data structured processing system based on semantic recognition provided by the embodiment of the application can execute the text data structured processing method based on semantic recognition provided by any embodiment of the application, has the corresponding function modules and beneficial effects of the execution method.

[0156] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server, and each unit and module included is only divided according to the functional logic, but is not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific name of each functional unit is only for easy mutual differentiation, and does not limit the protection scope of the present application.

[0157] The above specific embodiments do not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent replacement and improvement made within the spirit and principles of the present application should be included in the protection scope of the present application. In some cases, the actions or steps described in the present application can be executed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

Claims

1. A method for text data structure processing based on semantic recognition, characterized in that, The method comprises the following steps: introducing a predetermined recognition and extraction strategy to analyze the original text data after standardization to obtain text key information; analyzing the text key information to obtain a first relationship triple of a first entity and establishing a triple set; calling a predetermined graph embedding strategy to analyze the triple set to obtain a target entity matrix; performing vectorization clustering on the target entity matrix to obtain a target clustering result, wherein the target clustering result comprises a plurality of clustering clusters; analyzing the plurality of clustering clusters to obtain initial semantic features, and adjusting the initial semantic features in combination with a domain dictionary library to obtain target semantic features; taking the target structured result of the target semantic features as the structured result of the original text data; analyzing the plurality of clustering clusters to obtain initial semantic features, and adjusting the initial semantic features in combination with a domain dictionary library to obtain target semantic features, comprising: extracting a first clustering cluster from the plurality of clustering clusters; analyzing the first clustering cluster to obtain a first general semantic feature, and taking the first general semantic feature as the initial semantic feature; traversing the entities in the first clustering cluster in the domain dictionary library to obtain a traversal result; analyzing the first clustering cluster based on the traversal result to obtain a first domain semantic feature; adjusting the initial semantic features based on the first domain semantic features to obtain the target semantic features; adjusting the initial semantic features based on the first domain semantic features to obtain the target semantic features, comprising: obtaining a general explanation of the entity; obtaining a domain explanation corresponding to the entity in the traversal result; performing first-level fusion on the general explanation and the domain explanation to obtain a first-level fusion result; performing second-level fusion on the first general semantic feature and the first domain semantic feature to obtain a second-level fusion result; fusing the first-level fusion result and the second-level fusion result to obtain the target semantic features.

2. The method of claim 1, wherein the method further comprises: introducing a predetermined recognition and extraction strategy to analyze the original text data after standardization to obtain text key information, comprising: performing entity recognition processing on the original text data according to an entity recognition plan in the predetermined recognition and extraction strategy to obtain an entity set; performing relationship extraction processing on the original text data according to a relationship extraction plan in the predetermined recognition and extraction strategy to obtain a relationship set; the entity set and the relationship set constitute the text key information.

3. The method of claim 2, wherein the semantic recognition-based text data structuring process is characterized by, performing entity recognition processing on the original text data according to an entity recognition plan in the predetermined recognition and extraction strategy to obtain an entity set, comprising: identifying a candidate entity set of the original text data through a preset classifier; performing verification analysis on the candidate entity set based on a weighted fusion strategy to obtain the entity set; wherein the preset classifier comprises a CRF classifier, an HMM classifier and a conditional random field.

4. The method of claim 1, wherein the method further comprises: calling a predetermined graph embedding strategy to analyze the triple set to obtain a target entity matrix, comprising: extracting any triple from the triple set, and the any triple comprises a plurality of relationship triples of any entity; According to the predetermined graph embedding strategy, the arbitrary entity is taken as a vertex, and the plurality of relationship triples are taken as edges to form a target undirected graph; The target entity matrix of the target undirected graph is analyzed and obtained.

5. The method of claim 4, wherein the semantic recognition-based text data structuring process is characterized by, The target entity matrix of the target undirected graph is analyzed and obtained, including: A target degree matrix and a target adjacency matrix of the target undirected graph are sequentially obtained; The difference between the target degree matrix and the target adjacency matrix is taken as the target entity matrix.

6. The method of claim 1, wherein the method further comprises: Before taking the target structured result of the target semantic feature as the structured result of the original text data, the method further includes: According to a predetermined data enhancement strategy, the original text data is processed to obtain enhanced text data; The target structured result is optimized according to the enhanced text data; The predetermined data enhancement strategy includes synonym replacement and sentence reorganization.

7. The method of claim 1, wherein the method further comprises: The target structured result is displayed through a user interface.

8. A text data structuring system based on semantic recognition, characterized by, The system for implementing the text data structured processing method based on semantic recognition according to any one of claims 1-7, the system comprises: A text key information acquisition module is configured to introduce a predetermined recognition and extraction strategy to recognize and extract the standardized original text data to obtain text key information; A triple set construction module is configured to analyze the text key information to obtain first relationship triples of a first entity and construct a triple set; A target entity matrix acquisition module is configured to call a predetermined graph embedding strategy to perform graph embedding analysis on the triple set to obtain a target entity matrix; A target clustering result acquisition module is configured to perform vectorization clustering processing on the target entity matrix to obtain a target clustering result, wherein the target clustering result comprises a plurality of clustering clusters; A target semantic feature acquisition module is configured to analyze the plurality of clustering clusters to obtain initial semantic features and adjust the initial semantic features in combination with a domain dictionary library to obtain target semantic features; A structured result acquisition module is configured to take a target structured result of the target semantic features as a structured result of the original text data.

Images