A language structure learning method and system for entity relationship extraction

Abstract

The invention relates to the technical field of language structure learning, in particular to a language structure learning method and system for entity relationship extraction, which extracts relation mentioning sentences, entity pairs and semantic relationship categories between entity pairs from a data set; combines entity feature sequences with entity tag sentences to obtain text sequences, which are input into a neural network model; maps words in the text sequences into vectors to obtain an abstract representation matrix; constructs an enhanced feature association adjacency matrix as a variable parameter for training the neural network model; performs graph convolution operations on the abstract representation matrix and the enhanced feature association adjacency matrix, and performs residual links to obtain text language structure features; performs full connection operations on the text language structure features, and outputs semantic relationship classification results through activation functions and linear classification. By constructing a feature association adjacency matrix, the matrix is ​​continuously updated in the hidden layer of the neural network, so as to achieve the purpose of learning language structure relations and improve the performance of entity relationship extraction.

Description

Technical Field

[0001] The present invention relates to language structure learning technology, and in particular to a language structure learning method and system oriented to entity relationship extraction. Background Art

[0002] With the rapid development of information technology, a large amount of data is presented in digital form, including various data types such as video, audio, image and text. The structured processing and storage of text data is particularly critical because it can effectively represent these data types. Therefore, information extraction technology came into being.

[0003] Information extraction technology is widely used in knowledge graph construction, search engine optimization, sentiment analysis and other fields. Its core task is to extract structured information from unstructured or semi-structured natural language text, covering named entity recognition, relationship extraction, event extraction, etc. As the core component of information extraction, relationship extraction plays a fundamental and key role in natural language processing technology.

[0004] The goal of relation extraction is to identify the semantic relationship between named entity pairs in text, and to provide theoretical and technical support for tasks such as question-answering systems, opinion mining, and semantic analysis. Named entity pairs in text refer to proper nouns such as names of people, places, and organizations. For example, in the sentence "But according to estimates by the Organization for Security and Cooperation in Europe, at least 1,000 people are still in prison", the relation extraction model needs to identify the semantic relationship of the geographical location LOCATION between the named entities "1,000 people" and "prison" based on the known content. The key to identifying semantic relationships lies in the text representation of named entity pairs in text. Entity pairs are usually annotated in the data set and exist in the form of continuous characters. The relation extraction model uses the contextual content of entity pairs in the text to learn the language structure to identify the semantic relationship between them. The research on relation extraction technology not only provides theoretical support for the development of natural language processing technology, but also has important research and application value for tasks such as semantic role labeling, paragraph understanding, and machine translation.

[0005] Existing natural language processing technology has the following technical problems:

[0006] (1) Traditional relation extraction methods rely on artificial rules or syntax trees to capture the language structure of text;

[0007] (2) The fixed feature association adjacency matrix used by traditional neural networks may cause the entity features to not match the connection relationship between the fixed feature association adjacency matrix, resulting in the model being unable to accurately predict the semantic relationship category between real entity pairs;

[0008] (3) Semi-supervised learning and unsupervised learning face the problem of semantic drift, which causes the model’s relationship extraction performance to deteriorate when faced with new, unseen language structures. Summary of the invention

[0009] In order to solve the above technical problems, the present invention provides a language structure learning method and system for entity relationship extraction, improves the existing natural language processing technology, overcomes the feature sparsity problem, improves the relationship extraction performance, and effectively utilizes the language structure information.

[0010] To achieve the above object, the technical solution adopted by the present invention is: firstly, a language structure learning method for entity relationship extraction is provided, the steps comprising:

[0011] S1 Extract relation instance

[0012] Extract relation instances from the dataset;

[0013] S2 feature sequence extraction

[0014] Extract entity features to obtain entity feature sequences, use the natural language processing library to combine the entity feature sequences with entity tag sentences, and obtain the text sequence S_ as the input of the neural network model;

[0015] S3 Text Sequence Abstract Representation

[0016] Construct a word vector lookup table for the word embedding layer of the neural network model, map the words in the text sequence S_ into vectors through the word vector lookup table, and obtain an abstract representation matrix M;

[0017] S4 constructs feature correlation adjacency matrix

[0018] The feature association adjacency matrix A is constructed by using random numbers within a set interval, and combined with the unit matrix to obtain the enhanced feature association adjacency matrix Ã, which is used as a variable parameter for neural network model training. The semantic relationship category is used as the parameter adjustment target to update the enhanced feature association adjacency matrix Ã.

[0019] S5 Learning language structure features

[0020] The connection between the abstract representation matrix M and the enhanced feature association adjacency matrix à is established through the convolution operation of two graph convolution layers, and then the text language structure feature C is extracted from the graph convolution operation; wherein the abstract representation matrix M constitutes the node Node of the graph structure G, and the value in the enhanced feature association adjacency matrix à represents the association relationship between the nodes Node in the graph structure G, that is, the edge Edge, and the graph structure G is expressed as: G=<Node,Edge> ;

[0021] S6 semantic relationship classification

[0022] A full connection operation is performed on the text language structure feature C, the global feature is adjusted, and the semantic relationship classification result is output through the activation function SoftMax and linear classification operation.

[0023] Preferably, the step S1 of extracting relationship instances includes: using the xml.etree.ElementTree library in the Python standard library to extract required data from the XML file in the data set, parsing the XML file through an efficient programming interface API, obtaining relationship instances including relationship mention statements, entity pairs, and semantic relationship categories between entity pairs, and storing them in json or txt files.

[0024] Preferably, the data set in step S1 includes: ACE2005, SemEval-2010-task 8, Re-TACRED, and Chinese literary text corpus.

[0025] Preferably, the entity features in step S2 are extracted from a data set or through a third-party tool; the entity features include: entity type ET, entity subtype EST, entity head noun EHN, entity part-of-speech tag EPOS, and entity structure ES.

[0026] Preferably, the operation process of the text sequence S_ in step S2 is as follows:

[0027] Suppose the relation mention statement is S, and the two entities are and , expressed as: , , ,in, , Indicates a relation mentioning two entities in a sentence;

[0028] entity The characteristic sequence of is expressed as: ;

[0029] entity The characteristic sequence of is expressed as: ;

[0030] entity and The common feature sequence is ;

[0031] The entity feature sequence is expressed as: ,entity and The marking features are represented as and , combine it with the relation mention statement S to obtain the entity tag statement , expressed as: ,

[0032] The entity feature sequence and the entity tag sentence Combined, we get the text sequence S_, expressed as: ,in, and Represented as the boundary of a solid.

[0033] Preferably, the initialization method of the word vector lookup table in step S3 includes: random initialization using a certain probability distribution within a certain range; loading word vector initialization from a pre-trained file; loading a pre-trained language model initialization.

[0034] Preferably, the calculation process of the text language structure feature C in step S5 is as follows:

[0035] The abstract representation matrix M and the enhanced feature association adjacency matrix à are subjected to a first graph convolution operation, which is expressed as: GraphConv(M, Ã);

[0036] The output of the first graph convolution operation is connected to the abstract representation matrix M for the first time by residual connection, which is expressed as: Residual(GraphConv(M,Ã),M);

[0037] The first residual connection result and the enhanced feature associated adjacency matrix à are subjected to a second graph convolution operation, which is expressed as: GraphConv(Residual(GraphConv(M, Ã),M), Ã);

[0038] The output of the second graph convolution operation is connected to the output of the first graph convolution operation for a second residual connection to obtain the text language structure feature C, which is expressed as: C=Residual(GraphConv(Residual(GraphConv(M,Ã),M),Ã),GraphConv(M,Ã)),

[0039] Among them, GraphConv represents the graph convolution operation, and Residual represents the residual connection.

[0040] The present invention also provides a language structure learning system for entity relationship extraction, comprising:

[0041] Relation instance extraction module: used to extract relationship instances from the data set;

[0042] Feature sequence extraction module: used to extract entity features, obtain entity feature sequences, use the natural language processing library to combine the entity feature sequences with entity tag sentences, and obtain the text sequence S_ as the input of the neural network model;

[0043] Text sequence abstract representation module: used to construct a word vector lookup table of the word embedding layer of the neural network model, and map the words in the text sequence S_ into vectors through the word vector lookup table to obtain an abstract representation matrix M;

[0044] Feature association adjacency matrix construction module: used to construct the feature association adjacency matrix A using random numbers within a set interval, and combine it with the unit matrix to obtain the enhanced feature association adjacency matrix Ã, which is used as a variable parameter for neural network model training. The semantic relationship category is used as the parameter adjustment target to update the enhanced feature association adjacency matrix Ã;

[0045] Language structure feature learning module: used to combine the abstract representation matrix M with the enhanced feature association adjacency matrix à through two graph convolution operations, establish the connection between the abstract representation features, and obtain the output of the graph convolution layer, i.e., the text language structure feature C, through two residual connections;

[0046] Semantic relationship classification module: used to perform a full connection operation on the text language structure feature C, adjust the global features, and output the semantic relationship classification result through the activation function SoftMax and linear classification operation.

[0047] The above technical solution has the following advantages or beneficial effects:

[0048] (1) Based on the use of relationship instances, the present invention extracts entity features and introduces neural network technology to give full play to the ability of neural network to automatically extract text language structure features in a hierarchical manner. A graph convolution layer is constructed in the hidden layer of the neural network model, and residual connections are used to avoid semantic loss.

[0049] (2) During the training process of the neural network model targeting entity semantic relations, the enhanced feature association adjacency matrix à is continuously learned and adjusted to adapt to the semantic relationship categories between entity pairs, which effectively alleviates the feature sparsity problem and improves the performance of relationship extraction.

[0050] (3) By extracting the entity features, the entity feature sequence is obtained and compared with the entity tag sentence Combined with the input of the neural network model, a feature association adjacency matrix A is constructed in the hidden layer of the neural network model, and combined with the unit matrix to obtain the enhanced feature association adjacency matrix Ã, so that it participates in graph convolution and residual connection operations to obtain text language structure feature information, thereby enhancing the model's ability to recognize the influence of semantic relationships of words other than entities in the text.

[0051] The above summary is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features of the present invention will be readily apparent by reference to the accompanying drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the multiple drawings represent the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings only depict some embodiments disclosed in the present invention and should not be regarded as limiting the scope of the present invention.

[0053] Figure 1 A flow chart of a language structure learning method for entity relationship extraction provided by the present invention;

[0054] Figure 2 The overall network structure diagram of a language structure learning system for entity relationship extraction provided by the present invention;

[0055] Figure 3 A flowchart of the steps of an embodiment provided by the present invention;

[0056] Figure 4 A schematic diagram of the functional modules of a language structure learning system for entity relationship extraction provided by the present invention. DETAILED DESCRIPTION

[0057] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0058] The present invention first provides a language structure learning method for entity relationship extraction, the process is as follows Figure 1 As shown, the following steps are included:

[0059] S1 Extract relation instance

[0060] The xml.etree.ElementTree library in the Python standard library is used to extract the required data from the XML files in the dataset. The XML files are parsed through an efficient programming interface API to obtain relationship instances including relationship mention sentences, entity pairs, and semantic relationship categories between entity pairs, which are stored in json or txt files. The datasets include: ACE2005, SemEval-2010-task 8, Re-TACRED, and Chinese literature text corps.

[0061] S2 feature sequence extraction

[0062] Extract entity features from the dataset or through third-party tools, including entity type ET, entity subtype EST, entity head noun EHN, entity part-of-speech tag EPOS, entity structure ES, and obtain entity feature sequences. Use the natural language processing library to combine the entity feature sequences with entity tag sentences. Combine and obtain the text sequence S_ as the input of the neural network model;

[0063] The operation process of the text sequence S_ is as follows:

[0064] Suppose the relation mention statement is S, and the two entities are and , expressed as: , , ,in, , Indicates a relation mentioning two entities in a sentence;

[0065] entity The characteristic sequence of is expressed as: ;

[0066] entity The characteristic sequence of is expressed as: ;

[0067] entity and The common feature sequence is ;

[0068] The entity feature sequence is expressed as: ,

[0069] entity and The marking features are represented as and , combine it with the relation mention statement S to obtain the entity tag statement , expressed as: ,

[0070] The entity feature sequence and the entity tag sentence Combined, we get the text sequence S_, expressed as: ,in, and Represented as the boundary of an entity, the entity and The mark feature can be expressed in many ways, including: "*", "#", " <e1> ”、" <e2>”.

[0071] S3 Text Sequence Abstract Representation

[0072] Construct a word vector lookup table for the word embedding layer of the neural network model, map the words in the text sequence S_ into vectors through the word vector lookup table, and obtain an abstract representation matrix M;

[0073] The word vector lookup table is initialized by: using a certain probability distribution to initialize randomly within a certain range; loading a word vector from a pre-trained file for initialization; and loading a pre-trained language model for initialization.

[0074] S4 constructs feature correlation adjacency matrix

[0075] A feature association adjacency matrix A with a shape of n×n and a size based on the abstract representation matrix M is constructed using random numbers within a set interval (-0.35, 0.35). Since each node has a strong correlation with itself, in order to enhance the language information of each node, the feature association adjacency matrix A is combined with the unit matrix I to obtain an enhanced feature association adjacency matrix Ã=A+I, so that the nodes are self-associated;

[0076] The enhanced feature association adjacency matrix à is used as a variable parameter for training the neural network model, and the semantic relationship category is used as the parameter adjustment target to update the enhanced feature association adjacency matrix Ã, so as to achieve the purpose of learning the text language structure of the neural network model.

[0077] S5 Learning language structure features

[0078] The connection between the abstract representation matrix M and the enhanced feature association adjacency matrix à is established through the convolution operation of two graph convolution layers, and two residual connections are performed. In the process of layer-by-layer mapping, a complex function is fitted, and the weights required for each local mapping are learned, so as to extract the text language structure feature C from the graph convolution operation; wherein the abstract representation matrix M constitutes the nodes Node of the graph structure G, and the values ​​in the enhanced feature association adjacency matrix à represent the association relationship between the nodes Node in the graph structure G, that is, the edge Edge, and the graph structure G is expressed as: G=<Node,Edge> .

[0079] The graph convolution layer is defined as:

[0080] The value of the feature association adjacency matrix A is randomly initialized; in the forward feedback process, the potential relationship between the input abstract representation features is calculated as D 1 / 2 ÃD 1 / 2 X, where is a diagonal matrix of node association values, H (0) Initialized to H (0) =X, the output of the first layer of graph convolution is represented by H (l) ;

[0081] Multiple layers stacked together are represented as: H (l+1) =σ(D 1 / 2 Ã （l） D 1 / 2 H (l) W (1) )

[0082] Among them, W (1) represents a trainable parameter used to encode the discriminability of the learned text language structure representation, and σ represents a nonlinear activation function, including the ReLU activation function.

[0083] The calculation process of the text language structure feature C is as follows:

[0084] The abstract representation matrix M and the enhanced feature association adjacency matrix à are subjected to a first graph convolution operation, which is expressed as: GraphConv(M, Ã);

[0085] The output of the first graph convolution operation is connected to the abstract representation matrix M for the first residual connection, expressed as: Residual(GraphConv(M,Ã),M);

[0086] The first residual connection result and the enhanced feature associated adjacency matrix à are subjected to a second graph convolution operation, which is expressed as: GraphConv(Residual(GraphConv(M, Ã),M), Ã);

[0087] The output of the second graph convolution operation is connected to the output of the first graph convolution operation for a second residual connection to obtain the text language structure feature C, which is expressed as: C=Residual(GraphConv(Residual(GraphConv(M,Ã),M),Ã),GraphConv(M,Ã))

[0088] Among them, GraphConv represents the graph convolution operation, and Residual represents the residual connection.

[0089] S6 semantic relationship classification

[0090] A full connection operation is performed on the text language structure feature C, the global feature is adjusted, and the semantic relationship classification result is output through the activation function SoftMax and linear classification operation.

[0091] The present invention also provides a language structure learning system for entity relationship extraction, such as Figure 2 , Figure 4 As shown, including:

[0092] Extracting relation instances module: used to extract relation instances from the data set;

[0093] Feature sequence extraction module: used to extract entity features, obtain entity feature sequences, use the natural language processing library to combine the entity feature sequences with entity tag sentences, and obtain the text sequence S_ as the input of the neural network model;

[0094] Text sequence abstract representation module: used to construct a word vector lookup table of the word embedding layer of the neural network model, and map the words in the text sequence S_ into vectors through the word vector lookup table to obtain an abstract representation matrix M;

[0095] Constructing feature association adjacency matrix module: It is used to construct feature association adjacency matrix A using random numbers within a set interval, and combine it with the unit matrix to obtain enhanced feature association adjacency matrix Ã, which is used as a variable parameter for neural network model training. The semantic relationship category is used as the parameter adjustment target to update the enhanced feature association adjacency matrix Ã.

[0096] Learning language structure feature module: used to establish the connection between the abstract representation matrix M and the enhanced feature association adjacency matrix à through the convolution operation of two graph convolution layers, and then extract the text language structure feature C from the graph convolution operation;

[0097] Semantic relationship classification module: used to perform a full connection operation on the text language structure feature C, adjust the global features, and output the semantic relationship classification result through the activation function SoftMax and linear classification operation.

[0098] According to the above method and system, this embodiment provides an example, taking the ACE2005 data set as an example, extracting relationship instances and learning the language structure relationship method, such as Figure 3 As shown, the steps are as follows:

[0099] S1 extracts a relation instance from the ACE2005 dataset

[0100] The relationship instances include: a relationship mention sentence "But according to estimates by the Organization for Security and Cooperation in Europe, at least 1,000 people are still in prison"; entity pairs "1,000 people", "prison"; and a semantic relationship category "geographic location relationship LOCATION" between entity pairs.

[0101] S2 feature sequence extraction

[0102] Extract entity features. The feature sequences of entities "1000 people" and "prison" are: [PER, Group, person, NN, NN, <*>,< / *> ] and [FAC, Building-Grounds, Prison, NN, NN, <#>,< / #> ];

[0103] The common feature sequence of "1000 people" and "prison" is [e1before e2];

[0104] The marking features of the entities "1000 people" and "prison" are represented by "<*>",< / *> " and "<#>,< / #> ”;

[0105] The entity tagged sentence is "However, the Organization for Security and Cooperation in Europe estimates that there are at least <*> 1,000< / *> Imprisoned in prison< / #> inside";

[0106] The entity feature sequence is combined with the entity tag sentence to obtain the text sequence S_ as "e1before e2, PER, Group, People, NN, NN, FAC, Building-Grounds, Prison, NN, NN, but according to the Organization for Security and Cooperation in Europe, there are at least <*> 1000 people< / *> Imprisoned in prison< / #> inside".

[0107] S3 text sequence abstract representation

[0108] In this embodiment, the text sequence S_ is vectorized and mapped through the Bert-base pre-trained language model, and the unregistered words in the sequence are vectorized by random initialization to obtain the abstract representation matrix M of the text sequence S_.

[0109] S4 builds the feature association matrix A

[0110] Random numbers within the set interval (-0.35, 0.35) are used to construct a feature association adjacency matrix A with a shape of n×n and a size based on the abstract representation matrix M. The matrix is ​​combined with the unit matrix I to obtain an enhanced feature association adjacency matrix Ã=A+I, which is used as a variable parameter for neural network model training. The semantic relationship category is used as the adjustment target of the parameter to update the enhanced feature association adjacency matrix Ã, so as to achieve the purpose of learning the text language structure of the neural network model.

[0111] S5 Learning language structure features

[0112] The abstract representation matrix M is subjected to a first graph convolution operation with the enhanced feature associated adjacency matrix Ã, the output of the first graph convolution is subjected to a first residual connection with the abstract representation matrix M, the first residual connection result is subjected to a second graph convolution operation with the enhanced feature associated adjacency matrix Ã, the output of the second graph convolution is subjected to a second residual connection with the output of the first graph convolution, and the text language structure feature C is obtained: C=Residual(GraphConv(Residual(GraphConv(M, Ã),M),Ã), GraphConv(M, Ã)).

[0113] S6 semantic relationship classification

[0114] The obtained text language structure feature C is fully connected, the global features are adjusted, and the semantic relationship classification results are output through the activation function SoftMax and linear classification operations.

[0115] The above technical solution has the following advantages or beneficial effects: The present invention extracts entity features based on the use of relationship instances, introduces neural network technology, gives full play to the ability of neural networks to automatically extract text language structure features in layers, constructs graph convolution layers in the hidden layers of the neural network model, and uses residual connections to avoid semantic loss. In the process of training the neural network model with entity semantic relationships as the target, the enhanced feature association adjacency matrix à is continuously learned and adjusted to adapt to the semantic relationship categories between entity pairs, effectively alleviating the feature sparsity problem and improving the performance of relationship extraction. By extracting the entity features, the entity feature sequence is obtained, which is compared with the entity labeling statement Combined with the input of the neural network model, a feature association adjacency matrix A is constructed in the hidden layer of the neural network model, and combined with the unit matrix to obtain the enhanced feature association adjacency matrix Ã, so that it participates in graph convolution and residual connection operations to obtain text language structure feature information, thereby enhancing the model's ability to recognize the influence of semantic relationships of words other than entities in the text.

[0116] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of various changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims. < / e1>

Claims

1. A language structure learning method for entity relationship extraction, characterized in that the steps include: S1 extracts the relationship instance: Extract relation instances from the dataset; S2 feature sequence extraction: Extract entity features to obtain entity feature sequences, use the natural language processing library to combine the entity feature sequences with entity tag sentences, and obtain the text sequence S_ as the input of the neural network model; S3 text sequence abstract representation: Construct a word vector lookup table for the word embedding layer of the neural network model, map the words in the text sequence S_ into vectors through the word vector lookup table, and obtain an abstract representation matrix M; S4 constructs the feature association adjacency matrix: The feature association adjacency matrix A is constructed by using random numbers within a set interval, and combined with the unit matrix to obtain the enhanced feature association adjacency matrix Ã, which is used as a variable parameter for neural network model training. The semantic relationship category is used as the parameter adjustment target to update the enhanced feature association adjacency matrix Ã. S5 learning language structure features: The connection between the abstract representation matrix M and the enhanced feature association adjacency matrix à is established through the convolution operation of two graph convolution layers, and then the text language structure feature C is extracted from the graph convolution operation; wherein the abstract representation matrix M constitutes the node Node of the graph structure G, and the value in the enhanced feature association adjacency matrix à represents the association relationship between the nodes Node in the graph structure G, that is, the edge Edge, and the graph structure G is expressed as: G=<Node,Edge> ; The calculation process of the text language structure feature C is as follows: The abstract representation matrix M and the enhanced feature association adjacency matrix à are subjected to the first graph convolution operation, which is expressed as: GraphConv(M, Ã); The output of the first graph convolution operation is connected to the abstract representation matrix M for the first residual connection, which is expressed as: Residual(GraphConv(M,Ã),M); The result of the first residual connection is associated with the enhanced feature adjacency matrix à for the second graph convolution operation, which is expressed as: GraphConv(Residual(GraphConv(M,Ã),M),Ã); The output of the second graph convolution operation is connected to the output of the first graph convolution operation for a second residual connection to obtain the text language structure feature C, which is expressed as: C=Residual(GraphConv(Residual(GraphConv(M,Ã),M),Ã),GraphConv(M,Ã)), Among them, GraphConv represents the graph convolution operation, and Residual represents the residual connection; S6 semantic relationship classification: A full connection operation is performed on the text language structure feature C, the global feature is adjusted, and the semantic relationship classification result is output through the activation function SoftMax and linear classification operation.

2. A language structure learning method for entity relationship extraction as claimed in claim 1, characterized in that: The step S1 of extracting relationship instances includes: using the xml.etree.ElementTree library in the Python standard library to extract required data from the XML file in the data set, parsing the XML file through an efficient programming interface API, obtaining relationship instances including relationship mention statements, entity pairs, and semantic relationship categories between entity pairs, and storing them in json or txt files.

3. A language structure learning method for entity relationship extraction according to claim 1 or 2, characterized in that: The data sets in step S1 include: ACE2005, SemEval-2010-task 8, Re-TACRED, and Chinese literary text corpus.

4. The method for learning a language structure for entity relationship extraction according to claim 1, characterized in that: In step S2, the entity features are extracted from the data set or through a third-party tool; the entity features include: entity type ET, entity subtype EST, entity head noun EHN, entity part-of-speech tag EPOS, and entity structure ES.

5. The method for learning a language structure for entity relationship extraction according to claim 4, characterized in that: The operation process of the text sequence S_ in step S2 is as follows: Suppose the relation mention statement is S, and the two entities are and , expressed as: , , ,in, , Indicates a relation mentioning two entities in a sentence; entity The characteristic sequence of is expressed as: ； entity The characteristic sequence of is expressed as: ； entity and The common feature sequence is ; The entity feature sequence is expressed as: ,entity and The marking features are represented as and , combine it with the relation mention statement S to obtain the entity tag statement , expressed as: , the entity feature sequence and the entity tag sentence Combined, we get the text sequence S_, expressed as: ,in, and Represented as the boundary of a solid.

6. The language structure learning method for entity relationship extraction according to claim 1, characterized in that: The initialization method of the word vector lookup table in step S3 includes: random initialization using a certain probability distribution within a certain range; loading word vector initialization from a pre-trained file; loading a pre-trained language model initialization.

7. A language structure learning system for entity relationship extraction, characterized in that: The system is used to execute the language structure learning method for entity relationship extraction as claimed in claim 1, comprising: Relation instance extraction module: used to extract relationship instances from the data set; Feature sequence extraction module: used to extract entity features, obtain entity feature sequences, use the natural language processing library to combine the entity feature sequences with entity tag sentences, and obtain the text sequence S_ as the input of the neural network model; Text sequence abstract representation module: used to construct a word vector lookup table of the word embedding layer of the neural network model, and map the words in the text sequence S_ into vectors through the word vector lookup table to obtain an abstract representation matrix M; Feature association adjacency matrix construction module: used to construct the feature association adjacency matrix A using random numbers within a set interval, and combine it with the unit matrix to obtain the enhanced feature association adjacency matrix Ã, which is used as a variable parameter for neural network model training. The semantic relationship category is used as the parameter adjustment target to update the enhanced feature association adjacency matrix Ã; Language structure feature learning module: used to combine the abstract representation matrix M with the enhanced feature association adjacency matrix à through two graph convolution operations, establish the connection between the abstract representation features, and obtain the output of the graph convolution layer, i.e., the text language structure feature C, through two residual connections; Semantic relationship classification module: used to perform a full connection operation on the text language structure feature C, adjust the global features, and output the semantic relationship classification result through the activation function SoftMax and linear classification operation.

Images