Contract text classification method and equipment based on heterogeneous graph convolutional neural network
Through the method based on heterogeneous graph convolution neural network, the problem of poor classification of contract texts in the existing technology is solved, and stronger representation learning and feature fusion are achieved, which significantly improves the accuracy of contract text classification.
Patent Information
- Application Number
- CN202411264555.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2025-01-14
AI Technical Summary
The existing neural network text classification methods cannot effectively utilize and exert the potential of contract text context sentences, resulting in poor classification results and affecting the use of contract functions by intelligent ERP systems.
Using a method based on heterogeneous graph convolution neural network, the heterogeneous graphs are constructed by constructing contract text data sets, obtaining word embedding vectors, defining heterogeneous graph nodes and node relationships, and the heterogeneous graph heterogeneous graphs are constructed, and the graph convolution network is trained to achieve Stronger representation learning and feature fusion.
It significantly improves the accuracy and accuracy of contract text classification, makes full use of the feature fusion capability of graph convolutional neural network, and improves the classification effect of contract text.
Smart Images
Figure CN119311887A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of natural language processing technology, and in particular to a contract text classification method and device based on a heterogeneous graph convolutional neural network. Background Art
[0002] With the deepening of global economic integration and the booming digital transformation of enterprises, many companies have established financial shared-use centers (FSCs), platforms dedicated to processing shared financial data. Contract texts are a crucial text resource within these centers. Automated text classification within these diverse contract documents is crucial to enabling business personnel to quickly identify underlying contract attributes and make relevant decisions. However, traditional manual classification methods require actual reading of the relevant contract texts and possess relevant business judgment skills. Furthermore, the length of the text cannot be controlled, resulting in labor-intensive and time-consuming tasks and a high risk of misidentification, which can cause losses to both the company and its customers.
[0003] Contracts are a special type of text, characterized by their legality, equality, consistency, and binding nature. Common contract types include standard contracts, custom contracts, legal contracts, sales contracts, employment contracts, lease contracts, cooperation agreements, confidentiality agreements, loan agreements, and authorization agreements. Many other types of contracts exist, depending on the industry, field, and specific needs. Early work primarily used support vector machines (SVMs) for text classification. Later, neural networks, such as long short-term memory (LSTM) neural networks and attention-based neural networks, gained application due to their flexible structure and automatic feature recognition capabilities. However, these networks failed to fully leverage the potential of contextual sentences, resulting in less than satisfactory contract classification results.
[0004] In summary, existing neural network text classification methods fail to fully utilize and unleash the potential of contextual sentences in contract texts, and their effectiveness falls short of enterprise-level applications. If this issue remains unresolved, it will impact the entire intelligent ERP system's ability to utilize contract functions and hinder the preservation and use of contract texts. Summary of the Invention
[0005] The embodiments of the present invention provide a contract text classification method and device based on a heterogeneous graph convolutional neural network, which is used to solve the following technical problem: the existing neural network text classification method cannot well utilize and give full play to the potential of the context sentences of the contract text, and the effect fails to reach the level of enterprise application.
[0006] The embodiment of the present invention adopts the following technical solutions:
[0007] In one aspect, an embodiment of the present invention provides a contract text classification method based on a heterogeneous graph convolutional neural network, the method comprising: collecting contract text data of various types and constructing a contract text dataset;
[0008] Based on long short-term memory networks and graph convolutional networks, an intelligent classification model is built and initialized;
[0009] Based on the contract text dataset, a word embedding vector is obtained, and the intelligent classification model is preliminarily trained using the word embedding vector to obtain a hidden layer state vector;
[0010] Defining heterogeneous graph nodes and node relationships, and initializing the heterogeneous graph nodes using the hidden layer state vector;
[0011] Constructing a contract text heterogeneous graph based on the contract text dataset, the heterogeneous graph nodes and node relationships;
[0012] Further training the intelligent classification model through the contract text heterogeneous graph to obtain a trained intelligent classification model;
[0013] The contract text to be classified is input into the trained intelligent classification model to obtain the corresponding contract type.
[0014] In a feasible implementation, various types of contract text data are collected to construct a contract text dataset, specifically including:
[0015] Collect various types of contract documents on the Internet and extract the contract text data;
[0016] Dividing the contract text data into sentences to obtain a plurality of sentences;
[0017] An aspect set is constructed for each sentence, and a classification label and a predicted label are generated for each aspect in the aspect set.
[0018] In a feasible implementation, constructing an aspect set for each sentence and generating a classification label and a prediction label for each aspect in the aspect set specifically includes:
[0019] Split the current sentence into corresponding word sequences S = {w1,w2,…w n}; where n is the total number of words in the current sentence, w n is the nth word in the current sentence;
[0020] Take m words in the word sequence S in order to form an aspect of the sentence T = {w r+1 ,w r+2 ,…w r+m}; where r means starting to extract from the rth word;
[0021] Generate classification labels for the aspect T and the predicted label y of the classification model; where, p is the total number of contract categories, class p is the pth contract category;
[0022] All aspects that can be extracted from the current sentence are constructed as the aspect set.
[0023] In a feasible implementation, an intelligent classification model is constructed and initialized based on a long short-term memory network and a graph convolutional network, specifically including:
[0024] Construct a word embedding layer and initialize it with GloVe vectors;
[0025] Based on the long short-term memory network, a Bi-LSTM layer is constructed;
[0026] Based on the graph convolutional network and heterogeneous graph structure, a heterogeneous GCN network layer is constructed;
[0027] The word embedding layer, the Bi-LSTM layer, the heterogeneous GCN network layer, the pooling layer, and the fully connected layer are sequentially connected to obtain the intelligent classification model;
[0028] The weight of each layer in the intelligent classification model is initialized by a uniform distribution method, and initial parameters are set for each layer.
[0029] In a feasible implementation, obtaining a word embedding vector based on the contract text dataset specifically includes:
[0030] Map each word in the contract text dataset to a low-dimensional space to obtain a low-dimensional vector matrix containing all words Where R is a set of real numbers, W is the number of words in the low-dimensional vector matrix, and d w is the vector length of a certain dimension of a word;
[0031] Find the low-dimensional vector of each word in the word sequence S in the low-dimensional vector matrix to obtain the word embedding vector X={x1,x2,…,x n},in Where n is the total number of words in the word sequence S, x n is the low-dimensional vector of the nth word in the word sequence S.
[0032] In a feasible implementation, the intelligent classification model is preliminarily trained using the word embedding vector to obtain a hidden layer state vector, specifically including:
[0033] The word embedding vector is input into the Bi-LSTM layer in the intelligent classification model, the Bi-LSTM layer is trained, and the Get the corresponding hidden layer state vector H; where, n is the length of the word embedding vector, d h is the dimension of the hidden state vector; {x1,x2,…,x n} is the word embedding vector, and are all intermediate vectors, h n is the hidden state vector of the nth word in the word sequence S.
[0034] In a feasible implementation, defining heterogeneous graph nodes and node relationships, and initializing the heterogeneous graph nodes using the hidden layer state vector, specifically includes:
[0035] Sentence nodes, aspect nodes, aspect vocabulary nodes, and context vocabulary nodes are defined to constitute the heterogeneous graph nodes; wherein the sentence nodes represent information about the entire sentence; the aspect nodes represent information about a specific aspect; the aspect vocabulary nodes represent aspect vocabulary in the sentence; and the context vocabulary nodes represent common vocabulary in the sentence, i.e., vocabulary that is not considered an aspect vocabulary;
[0036] Defining node relationships between nodes of different heterogeneous graphs; wherein the node relationships include at least: a subordinate relationship, an information extraction relationship, and a mutual indication relationship;
[0037] Extracting a hidden layer state aspect vector representing aspect T from the hidden layer state vector;
[0038] The sentence node is initialized by the hidden layer state vector; and the aspect node is initialized by the hidden layer state aspect vector.
[0039] In a feasible implementation, constructing a contract text heterogeneous graph based on the contract text dataset, the heterogeneous graph nodes, and the node relationships specifically includes:
[0040] The representation vectors of sentence nodes and aspect nodes are added to the hidden layer state vector to obtain the heterogeneous graph representation vector matrix H * ={h1,h2,…,h n ,s,t}; where H={h1,h2,…,h n} is the hidden state vector, s is the representation vector of the sentence node, and t is the representation vector of the aspect node;
[0041] Based on the defined heterogeneous graph nodes and node relationships, the contract text dataset is constructed as a contract text heterogeneous graph G = (V, E, A, H * ); where V = (|V| = N) is the set of nodes in the heterogeneous graph, E is the set of edges in the heterogeneous graph; A is the adjacency matrix, which is used to represent the relationship between adjacent nodes in the heterogeneous graph.
[0042] In a feasible implementation, further training the intelligent classification model using the contract text heterogeneous graph to obtain a trained intelligent classification model specifically includes:
[0043] The heterogeneous graph representation vector matrix H in the heterogeneous graph G of the contract text * , and the adjacency matrix A are input into the GCN network layer of the intelligent classification model to perform GCN network layer training, so that the GCN network layer can learn the representation of each node through the syntactic path and obtain the neighbor relationship information between the nodes;
[0044] At this point, the training of the intelligent classification model is completed.
[0045] On the other hand, an embodiment of the present invention further provides a contract text classification device based on a heterogeneous graph convolutional neural network, the device comprising:
[0046] at least one processor; and,
[0047] a memory communicatively connected to the at least one processor; wherein,
[0048] The memory stores instructions that can be executed by the at least one processor, so that the at least one processor can execute the contract text classification method based on heterogeneous graph convolutional neural network.
[0049] Compared with the existing technology, the contract text classification method and device based on heterogeneous graph convolutional neural network provided by the embodiment of the present invention have the following beneficial effects:
[0050] Based on the existing heterogeneous graphs, the present invention proposes a new convolutional neural network for heterogeneous graphs, which innovatively defines the representation of sentences and aspects as different entity nodes, enabling them to establish relationships with other nodes, exchange information, and achieve stronger representation learning. By integrating heterogeneous graphs with multiple relationships, significant improvements have been made in the classification accuracy in the field of contract text classification. The present invention mainly obtains the feature representation of the input text by giving a context, and then uses a powerful heterogeneous graph convolutional neural network to develop the potential of contextual sentences, thereby making full use of the feature fusion ability of the graph convolutional neural network, improving the classification accuracy of the contract text, and accurately obtaining the classification expression for a certain aspect. The intelligent classification model provided by the present invention has great advantages in contract text classification, and the heterogeneous graph plays a great role in achieving good performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments described in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:
[0052] Figure 1 A flow chart of a contract text classification method based on a heterogeneous graph convolutional neural network provided by an embodiment of the present invention;
[0053] Figure 2 A schematic diagram of the structure of an intelligent classification model provided by an embodiment of the present invention;
[0054] Figure 3 A heterogeneous graph node relationship dependency graph provided by an embodiment of the present invention;
[0055] Figure 4 A schematic diagram of the structure of a contract text classification device based on a heterogeneous graph convolutional neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0056] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.
[0057] The embodiment of the present invention provides a contract text classification method based on heterogeneous graph convolutional neural network, such as Figure 1 As shown, the contract text classification method based on heterogeneous graph convolutional neural network specifically includes steps S101-S106:
[0058] S101. Collect various types of contract text data and construct a contract text dataset.
[0059] Specifically, various types of contract documents are first collected on the Internet, and the contract text data is extracted. The contract text data is then divided into sentences to obtain a number of sentences.
[0060] Furthermore, an aspect set is constructed for each sentence, and a classification label and a predicted label are generated for each aspect in the aspect set, specifically including:
[0061] Split the current sentence into corresponding word sequences S = {w1,w2,…w n}; where n is the total number of words in the current sentence, w n is the nth word in the current sentence. Then m words are sequentially taken out from the word sequence S to form an aspect of the sentence T = {w r+1 ,w r+2 ,…w r+m}; where r means starting extraction from the rth word.
[0062] Furthermore, a classification label is generated for aspect T and the predicted label y of the classification model; where, p is the total number of contract categories, class p is the pth contract category. Finally, all aspects that can be extracted from the current sentence are constructed into an aspect set.
[0063] S102. Based on the long short-term memory network and graph convolutional network, build an intelligent classification model and initialize it.
[0064] Specifically, we construct a word embedding layer and initialize it with GloVe vectors. We then build a Bi-LSTM layer based on a long short-term memory network. We also build a heterogeneous GCN network layer based on a graph convolutional network and heterogeneous graph structures.
[0065] Furthermore, the word embedding layer, Bi-LSTM layer, heterogeneous GCN network layer, pooling layer, and fully connected layer are sequentially connected to obtain an intelligent classification model. The weights of each layer in the intelligent classification model are then initialized using the uniform distribution method, and initial parameters are set for each layer.
[0066] As a feasible implementation method, Figure 2 A schematic diagram of an intelligent classification model structure provided by an embodiment of the present invention is shown in FIG. Figure 2As shown in the figure, the intelligent classification model consists of an input layer, a word embedding layer, a Bi-LSTM layer, a heterogeneous GCN layer, a pooling layer, a fully connected layer, and an output layer.
[0067] In one embodiment, the word embedding layer is initialized with a 300-dimensional GloVe vector. All weights of each layer in the intelligent classification model are initialized with a uniform distribution U(0.01, 0.01), and all biases are set to zero. The hidden dimension of the Bi-LSTM layer is set to 50, and the hidden dimension of the heterogeneous GCN layer is set to 100. The dropout rate of the word embedding layer is set to 0.7, while the dropout rate of the Bi-LSTM layer and the heterogeneous GCN layer is set to 0.1. The model batch size is 32, and the L2 coefficient is 10 -4 .
[0068] S103. Based on the contract text dataset, obtain word embedding vectors, and perform preliminary training on the intelligent classification model using the word embedding vectors to obtain a hidden state vector.
[0069] Specifically, each word in the contract text dataset is mapped to a low-dimensional space to obtain a low-dimensional vector matrix containing all words Among them, R is a set of real numbers, W is the number of words in the low-dimensional vector matrix, and d w is the vector length of a certain dimension of a word.
[0070] Furthermore, the low-dimensional vector of each word in the word sequence S is searched in the low-dimensional vector matrix to obtain the word embedding vector X = {x1, x2, ..., x n},in Among them, n is the total number of words in the word sequence S, x n is the low-dimensional vector of the nth word in the word sequence S.
[0071] Furthermore, the word embedding vector is input into the Bi-LSTM layer in the intelligent classification model, the Bi-LSTM layer is trained, and the Get the corresponding hidden layer state vector H; where, n is the length of the word embedding vector, d h is the dimension of the hidden state vector; {x1,x2,…,x n} is the word embedding vector, and are all intermediate vectors, h n is the hidden state vector of the nth word in the word sequence S.
[0072] S104: Define heterogeneous graph nodes and node relationships, and initialize the heterogeneous graph nodes through hidden layer state vectors.
[0073] Specifically, sentence nodes, aspect nodes, aspect vocabulary nodes and context vocabulary nodes are defined to constitute heterogeneous graph nodes; among them, sentence nodes represent the information of the entire sentence; aspect nodes represent the information of a specific aspect; aspect vocabulary nodes represent aspect vocabulary in the sentence; context vocabulary nodes represent ordinary vocabulary in the sentence, that is, vocabulary that is not an aspect vocabulary.
[0074] Furthermore, node relationships between nodes of different heterogeneous graphs are defined; wherein the node relationships include at least: subordinate relationships, information extraction relationships, and mutual indication relationships.
[0075] As a feasible implementation method, the present invention innovatively defines the representation of sentences and aspects as different entity nodes, so that they can establish relationships with other nodes and exchange information, thus achieving stronger representation learning. The representation of sentence nodes and aspect nodes are represented by s and t respectively, where
[0076] As a feasible implementation method, Figure 3 A heterogeneous graph node relationship dependency diagram provided by an embodiment of the present invention, such as Figure 3 As shown in the figure, the heterogeneous graph nodes in the present invention include sentence nodes, aspect nodes, aspect vocabulary nodes, and context vocabulary nodes. The lines between each two nodes represent different node relationships. Among them, the dotted line between the sentence node and the context vocabulary node indicates that these context vocabulary belong to the same sentence and are extracted using the TF-IDF algorithm; the dotted line between the sentence node and the aspect vocabulary node indicates that these aspect vocabulary also belong to the sentence; the dotted line between the aspect node and the aspect vocabulary node indicates that these aspect vocabulary are extracted from the aspect using a window mechanism. When experimenting with the window mechanism, the most appropriate window size is selected based on the length distribution of each data; the solid line between the sentence node and the aspect node indicates that there is a mutual indicative relationship between the two.
[0077] Furthermore, in the hidden layer state vector H = {h1,h2,…,h n}, extract the hidden state aspect vector H representing aspect T t ={h r+1 ,h r+2 ,…,h r+m}; where r and m have the same meaning as the parameters in the representation vector of aspect T.
[0078] Furthermore, the sentence node s is initialized through the hidden state vector, s represents the sentence node, which is the average value of the hidden state vectors H of all words in the entire sentence. The aspect node t is initialized through the hidden state aspect vector, t represents the aspect node, which is the hidden state aspect vector H of all words in the entire sentence. t The average value of .
[0079] S105. Construct a contract text heterogeneous graph based on the contract text dataset, heterogeneous graph nodes, and node relationships.
[0080] Specifically, the representation vectors of sentence nodes and aspect nodes are added to the hidden layer state vector to obtain the heterogeneous graph representation vector matrix H * ={h1,h2,…,h n ,s,t}; where H={h1,h2,…,h n} is the hidden state vector, s is the representation vector of the sentence node, t is the representation vector of the aspect node, and
[0081] Furthermore, based on the defined heterogeneous graph nodes and node relationships, the contract text dataset is constructed as a contract text heterogeneous graph G = (V, E, A, H * ); where V = (|V| = N) is the set of nodes in the heterogeneous graph, E is the set of edges in the heterogeneous graph; A is the adjacency matrix, which is used to represent the relationship between adjacent nodes in the heterogeneous graph.
[0082] S106: Further train the intelligent classification model using the heterogeneous graph of the contract text to obtain a trained intelligent classification model. Input the contract text to be classified into the trained intelligent classification model to obtain the corresponding contract type.
[0083] Specifically, the heterogeneous graph representation vector matrix H in the heterogeneous graph G of the contract text is * , and the adjacency matrix A are input into the GCN network layer of the intelligent classification model, and the GCN network layer is trained so that the GCN network layer can learn the representation of each node through the syntactic path and obtain the neighbor relationship information between the nodes.
[0084] At this point, the training of the intelligent classification model is completed.
[0085] As a specific implementation method, the constructed heterogeneous graph is input into the GCN network, and the convolution operation is used to effectively encode local information based on the relationship in the graph. The input of GCN is the representation vector matrix H of the graph * and the adjacency matrix A. Adjacency here refers to the adjacent positions of nodes in the heterogeneous graph, rather than the sequential positions of words in the sentence. Therefore, GCN can effectively learn the representation of each node along the syntactic path, which is the difference between GCN and CNN. By increasing the number of GCN layers, information can be obtained from further neighbors.
[0086] Furthermore, the contract text to be classified is input into the trained intelligent classification model to output the corresponding contract type.
[0087] In addition, the embodiment of the present invention also provides a contract text classification device based on heterogeneous graph convolutional neural network, such as Figure 4 As shown in the figure, the contract text classification device based on heterogeneous graph convolutional neural network specifically includes:
[0088] at least one processor; and a memory communicatively connected to the at least one processor; wherein,
[0089] The memory stores instructions executable by at least one processor, so as to enable the at least one processor to perform:
[0090] Collect various types of contract text data and build a contract text dataset;
[0091] Based on long short-term memory networks and graph convolutional networks, an intelligent classification model is built and initialized;
[0092] Based on the contract text dataset, a word embedding vector is obtained, and the intelligent classification model is preliminarily trained using the word embedding vector to obtain a hidden layer state vector;
[0093] Defining heterogeneous graph nodes and node relationships, and initializing the heterogeneous graph nodes using the hidden layer state vector;
[0094] Constructing a contract text heterogeneous graph based on the contract text dataset, the heterogeneous graph nodes and node relationships;
[0095] Further training the intelligent classification model through the contract text heterogeneous graph to obtain a trained intelligent classification model;
[0096] The contract text to be classified is input into the trained intelligent classification model to obtain the corresponding contract type.
[0097] The various embodiments of the present invention are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer storage medium embodiments are generally similar to the method embodiments, so their descriptions are simplified. For relevant details, refer to the descriptions of the method embodiments.
[0098] The above description of specific embodiments of the present invention is provided. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0099] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations may be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention are intended to be included within the scope of protection of the present invention.
Claims
1. A contract text classification method based on heterogeneous graph convolutional neural network, characterized in that: The method comprises: Collect various types of contract text data and build a contract text dataset; Based on long short-term memory network and graph convolution network, build intelligent classification model and initialize it; Based on the contract text data set, a word embedding vector is obtained, and the intelligent classification model is preliminarily trained by the word embedding vector to obtain a hidden layer state vector; Defining heterogeneous graph nodes and node relationships, and initializing the heterogeneous graph nodes through the hidden layer state vector; Constructing a heterogeneous graph of contract text based on the contract text data set, the heterogeneous graph nodes and node relationships; Further training the intelligent classification model through the contract text heterogeneous graph to obtain a trained intelligent classification model; The contract text to be classified is input into the trained intelligent classification model to obtain the corresponding contract type.
2. According to claim 1, a contract text classification method based on heterogeneous graph convolutional neural network is characterized in that: Collect various types of contract text data and build a contract text dataset, including: Collect various types of contract documents on the Internet and extract the contract text data; Dividing the contract text data into sentences to obtain a plurality of sentences; An aspect set is constructed for each sentence, and a classification label and a predicted label are generated for each aspect in the aspect set.
3. According to claim 2, a contract text classification method based on heterogeneous graph convolutional neural network is characterized in that: Construct an aspect set for each sentence, and generate a classification label and a predicted label for each aspect in the aspect set, specifically including: Split the current sentence into corresponding word sequences S = {w1,w2,…w n }; where n is the total number of words in the current sentence, w n is the nth word in the current sentence; Take m words in the word sequence S in order to form an aspect of the sentence T = {w r+1 ,w r+2 ,…w r+m }; where r means starting to extract from the rth word; Generate classification labels for the aspect T and the predicted label y of the classification model; where p is the total number of contract categories, class p is the pth contract category; All aspects that can be extracted from the current sentence are constructed as the aspect set.
4. According to claim 1, a contract text classification method based on heterogeneous graph convolutional neural network is characterized in that: Based on the long short-term memory network and graph convolutional network, an intelligent classification model is built and initialized, including: Construct a word embedding layer and initialize the word embedding layer with GloVe vectors; Based on the long short-term memory network, a Bi-LSTM layer is constructed; Based on the graph convolutional network and the heterogeneous graph structure, a heterogeneous GCN network layer is constructed; The word embedding layer, the Bi-LSTM layer, the heterogeneous GCN network layer, the pooling layer, and the fully connected layer are sequentially connected to obtain the intelligent classification model; The weight of each layer in the intelligent classification model is initialized by a uniform distribution method, and initial parameters are set for each layer.
5. According to claim 1, a contract text classification method based on heterogeneous graph convolutional neural network is characterized in that: Based on the contract text dataset, word embedding vectors are obtained, specifically including: Map each word in the contract text dataset to a low-dimensional space to obtain a low-dimensional vector matrix containing all words Where R is a real number set, W is the number of words in the low-dimensional vector matrix, and d w is the vector length of a certain dimension of a word; Find the low-dimensional vector of each word in the word sequence S in the low-dimensional vector matrix to obtain the word embedding vector X = {x1, x2, ..., x n },in Where n is the total number of words in the word sequence S, x n is the low-dimensional vector of the nth word in the word sequence S.
6. According to claim 1, a contract text classification method based on heterogeneous graph convolutional neural network is characterized in that: The intelligent classification model is preliminarily trained by the word embedding vector to obtain a hidden layer state vector, which specifically includes: The obtained word embedding vector is input into the Bi-LSTM layer in the intelligent classification model, the Bi-LSTM layer is trained, and the Get the corresponding hidden layer state vector H; where, n is the length of the word embedding vector, d h is the dimension of the hidden layer state vector; {x1,x2,…,x n } is the word embedding vector, and are all intermediate vectors, h n is the hidden state vector of the nth word in the word sequence S.
7. According to claim 1, a contract text classification method based on heterogeneous graph convolutional neural network is characterized in that: Defining heterogeneous graph nodes and node relationships, and initializing the heterogeneous graph nodes through the hidden layer state vector, specifically includes: Sentence nodes, aspect nodes, aspect vocabulary nodes and context vocabulary nodes are defined to constitute the heterogeneous graph nodes; wherein the sentence node represents the information of the entire sentence; the aspect node represents the information of a specific aspect; the aspect vocabulary node represents the aspect vocabulary in the sentence; the context vocabulary node represents the common vocabulary in the sentence, that is, the vocabulary that is not used as aspect vocabulary; Defining node relationships between nodes of different heterogeneous graphs; wherein the node relationships at least include: a subordinate relationship, an information extraction relationship, and a mutual indication relationship; Extracting a hidden layer state aspect vector representing aspect T from the hidden layer state vector; The sentence node is initialized by the hidden layer state vector; and the aspect node is initialized by the hidden layer state aspect vector.
8. According to claim 1, a contract text classification method based on heterogeneous graph convolutional neural network is characterized in that: Based on the contract text data set, the heterogeneous graph nodes and node relationships, constructing a contract text heterogeneous graph, specifically including: The representation vectors of sentence nodes and aspect nodes are added to the hidden layer state vector to obtain the heterogeneous graph representation vector matrix H * ={h1,h2,…,h n ,s,t}; where H={h1,h2,…,h n } is the hidden state vector, s is the representation vector of the sentence node, and t is the representation vector of the aspect node; Based on the defined heterogeneous graph nodes and node relationships, the contract text dataset is constructed as a contract text heterogeneous graph G = (V, E, A, H * ), where V = (|V| = N) is the set of nodes in the heterogeneous graph, E is the set of edges in the heterogeneous graph, and A is the adjacency matrix, which is used to represent the relationship between adjacent nodes in the heterogeneous graph.
9. The method for classifying contract texts based on heterogeneous graph convolutional neural networks according to claim 8 is characterized in that: The intelligent classification model is further trained through the contract text heterogeneous graph to obtain a trained intelligent classification model, which specifically includes: The heterogeneous graph representation vector matrix H in the heterogeneous graph G of the contract text * , and the adjacency matrix A are input into the GCN network layer of the intelligent classification model, and the GCN network layer is trained so that the GCN network layer can learn the representation of each node through the syntactic path and obtain the neighbor relationship information between the nodes; At this point, the training of the intelligent classification model is completed.
10. A contract text classification device based on heterogeneous graph convolutional neural network, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor so that the at least one processor can execute a contract text classification method based on a heterogeneous graph convolutional neural network according to any one of claims 1-9.