CN114896978B - Entity recognition method, system and storage medium based on multi-graph collaborative semantic network - Google Patents

Abstract

The invention discloses an entity identification method, an entity identification system and a storage medium based on a multi-graph collaborative semantic network, which can be widely applied to the technical field of entity identification. The method comprises the steps of extracting dictionary feature data and part-of-speech feature data in original data containing a plurality of Chinese sentences to form feature data, constructing a multi-graph matrix according to the original data and the feature data, merging the original data and the feature data to obtain merged data, merging the merged data and the multi-graph matrix to perform collaborative training to obtain a feature merging result, decoding the feature merging result to obtain a Chinese named entity recognition result of the original data, and recognizing Chinese entities in the Chinese sentences by considering the dictionary feature data and the part-of-speech feature data, so that the accuracy of the entity recognition result can be effectively improved.

Description

Entity recognition method, system and storage medium based on multi-graph collaborative semantic network

Technical Field

The present invention relates to the field of entity recognition technology, and in particular to an entity recognition method, system and storage medium based on a multi-graph collaborative semantic network.

Background Art

In the related technologies, named entity recognition is not only one of the most important directions of natural language processing, but also a preprocessing step for downstream tasks such as relationship extraction, information retrieval, and knowledge graphs. Named entity recognition aims to extract entities such as names of people, places, and organizations from unstructured texts. Traditional Chinese named entity recognition is mainly divided into character-level methods and word-level methods. Character-level methods cannot fully utilize the potential sequence information between words in a sentence, resulting in weak generalization of the model; word-level methods over-rely on word-segmentation tools, and word-segmentation errors will continue throughout the task process, and error propagation will lead to a decrease in recognition performance. With the emergence of deep learning, powerful tools have been provided for named entity recognition to improve the recognition ability of the model. However, named entity recognition based on deep learning methods overly relies on enhancement methods such as dictionaries, ignoring the interference of dictionary enhancement methods on the model. Over-reliance on dictionaries often easily leads to the phenomenon of identifying incorrect entities.

Summary of the invention

The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes an entity recognition method, system and storage medium based on a multi-graph collaborative semantic network, which can effectively improve the accuracy of entity recognition results.

On the one hand, an embodiment of the present invention provides an entity recognition method based on a multi-graph collaborative semantic network, comprising the following steps:

Acquire original data, wherein the original data includes a plurality of Chinese sentences;

Extracting feature data of the original data, wherein the feature data includes dictionary feature data and part-of-speech feature data;

Constructing a multi-image matrix according to the original data and the feature data;

Merging the original data and the characteristic data to obtain merged data;

The combined data and the multi-image matrix are fused and then collaboratively trained to obtain a feature fusion result;

The feature fusion result is decoded to obtain the Chinese named entity recognition result of the original data.

In some embodiments, extracting feature data of the original data includes:

Using a dictionary tool to match word feature data on the original data to obtain dictionary feature data;

The original data is parsed by a part-of-speech parsing tool to obtain part-of-speech feature data.

In some embodiments, after extracting the feature data of the original data, the method further comprises the following steps:

Inputting the dictionary feature data into an embedding layer to obtain a dictionary feature vector;

Inputting the part-of-speech feature data into an embedding layer to obtain a part-of-speech feature vector;

The original data is input into the embedding layer to obtain a context information vector.

In some embodiments, constructing a multi-image matrix according to the original data and the feature data includes:

fusing the dictionary feature vector and the context information vector to generate a boundary map;

The part-of-speech feature vector and the context information vector are fused to generate a relationship graph and an inclusion graph respectively.

In some embodiments, the relationship graph includes dependencies between individual items and dependencies between words.

In some embodiments, the step of fusing the combined data with the multi-image matrix for collaborative training includes:

Inputting the fusion result of the first merged data and the boundary graph into a first graph attention network for graph convolution processing, wherein the first merged data includes merged data of the dictionary feature vector and the context information vector;

Inputting the fusion result of the second merged data and the relationship graph into a second graph attention network for graph convolution processing, wherein the second merged data includes merged data of the part-of-speech feature vector and the context information vector;

The fusion result of the second merged data and the inclusion graph is input into a third graph attention network for graph convolution processing.

In some embodiments, the step of fusing the combined data with the multi-image matrix for collaborative training further includes:

Inputting the boundary graph, the relationship graph, the inclusion graph and the context information of the original data into a fusion layer for parameter learning;

Collaborative training via multi-image fusion.

On the other hand, an embodiment of the present invention provides an entity recognition system based on a multi-graph collaborative semantic network, comprising:

An acquisition module, used for acquiring original data, wherein the original data includes a plurality of Chinese sentences;

An extraction module, used to extract feature data of the original data, wherein the feature data includes dictionary feature data and part-of-speech feature data;

A construction module, used for constructing a multi-image matrix according to the original data and the feature data;

A merging module, used for merging the original data and the feature data to obtain merged data;

A collaborative training module, used for fusing the combined data with the multi-image matrix and performing collaborative training to obtain a feature fusion result;

A decoding module is used to decode the feature fusion result to obtain the Chinese named entity recognition result of the original data.

On the other hand, an embodiment of the present invention provides an entity recognition system based on a multi-graph collaborative semantic network, comprising:

at least one memory for storing a program;

At least one processor is used to load the program to execute the entity recognition method based on multi-graph collaborative semantic network.

On the other hand, an embodiment of the present invention provides a storage medium, in which a computer-executable program is stored. When the computer-executable program is executed by a processor, it is used to implement the entity recognition method based on a multi-graph collaborative semantic network.

The embodiment of the present invention provides an entity recognition method based on a multi-graph collaborative semantic network, which has the following beneficial effects:

This embodiment extracts dictionary feature data and part-of-speech feature data from original data containing several Chinese sentences to form feature data, and constructs a multi-graph matrix based on the original data and the feature data, and merges the original data and the feature data to obtain merged data, and then fuses the merged data and the multi-graph matrix for collaborative training to obtain a feature fusion result, and then decodes the feature fusion result to obtain a Chinese named entity recognition result of the original data. This embodiment identifies Chinese entities in Chinese sentences by considering dictionary feature data and part-of-speech feature data, thereby effectively improving the accuracy of entity recognition results.

Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be further described below with reference to the accompanying drawings and embodiments, wherein:

FIG1 is a flow chart of an entity recognition method based on a multi-graph collaborative semantic network according to an embodiment of the present invention;

FIG2 is a process diagram of an entity recognition method based on a multi-graph collaborative semantic network according to an embodiment of the present invention.

DETAILED DESCRIPTION

Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

In the description of the present invention, "several" means more than one, "many" means more than two, "greater than", "less than", "exceed", etc. are understood to exclude the number itself, and "above", "below", "within", etc. are understood to include the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

In the description of the present invention, unless otherwise clearly defined, words such as "setting" should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above words in the present invention in combination with the specific content of the technical solution.

In the description of the present invention, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

At present, named entity recognition based on deep learning methods over-relies on enhancement methods such as dictionaries, ignoring the interference of dictionary enhancement methods on the model. Over-reliance on dictionaries often easily leads to the phenomenon of identifying wrong entities. For example: for the sentence "Nanjing Yangtze River Bridge", the dictionary can match "Yangtze River Bridge" to help the model identify the place name entity "Yangtze River Bridge". But at the same time, the dictionary will also match the words "Nanjing City", "Nanjing", "Mayor" and "Jiang Bridge". The matching word "Nanjing" will make the model mistakenly think that "Nanjing" is a real entity in this sentence, which has always led to the subsequent recognition of two wrong entities "Mayor" and "Jiang Bridge". Therefore, the meaning of this sentence will become that the name of the mayor of Nanjing is Jiang Bridge. This is the problem caused by over-reliance on dictionary enhancement methods. In addition, most methods ignore the most important point in Chinese sentences, the part-of-speech dependency information of Chinese. Compared with other languages, Chinese sentences have richer expressive power. For example: "Guangdong Student Union", for the machine, it is difficult to associate the two words "Guang" or "Da" with the words "Student Union", but in fact "Guangdong Student Union" is a whole. Most current methods cannot solve this problem.

Based on this, referring to FIG1 , an embodiment of the present invention provides an entity recognition method based on a multi-graph collaborative semantic network, and the embodiment of the method can be applied to a processor of a cloud server or a Chinese named entity recognition platform. Specifically, as shown in FIG1 , the embodiment of the method includes but is not limited to the following steps:

Step 110: obtaining original data; wherein the original data includes a plurality of Chinese sentences;

Step 120: extracting feature data of the original data; wherein the feature data includes dictionary feature data and part-of-speech feature data;

Step 130, constructing a multi-image matrix according to the original data and the feature data;

Step 140: Merge the original data and the feature data to obtain merged data;

Step 150, fusing the merged data and the multi-image matrix and performing collaborative training to obtain a feature fusion result;

Step 160: decode the feature fusion result to obtain the Chinese named entity recognition result of the original data.

In the embodiment of the present application, the original data is a set of Chinese sentences, that is, the present embodiment performs Chinese named entity recognition on sentences that need Chinese named entity recognition. Specifically, assuming that the original data input of the character-based training model is s={c ₁ ,c ₂ ,...,c _n }, where c _i is the i-th character in a sentence, each character c _i can be represented by looking up an embedded vector, as shown in formula (1):

Among them, e ^c represents the character embedding vector table.

In an embodiment of the present application, for extracting the feature data of the original data, obtaining the dictionary feature data and the part-of-speech feature data, the original data can be matched with the word feature data by using a dictionary tool to obtain the dictionary feature data; and the original data can be parsed with the part-of-speech parsing tool to obtain the part-of-speech feature data. In this embodiment, since vocabulary knowledge helps to enhance the representation of characters to improve the performance of the NER model, this embodiment represents the entries matched by the dictionary matching tool for the characters of each sentence in the original data as l={l ₁ ,l ₂ ,...,l _m }. This embodiment uses the part-of-speech and relational dependency of Chinese vocabulary, and cuts and separates the sentences in the original data through the part-of-speech parsing tool, and obtains the cut vocabulary represented as w={w ₁ ,w ₂ ,...,w _k }.

In an embodiment of the present application, for constructing a multi-graph matrix according to the original data and the feature data, a boundary matrix graph can be generated by fusing the dictionary feature data and the original data; and a relationship graph and an inclusion graph are generated by fusing the part-of-speech feature data and the original data. Specifically, the present embodiment constructs three word-character interaction graphs, the vertices and edges of the three graphs are different, and an adjacency matrix is introduced to represent the graph structure, and the values in the adjacency matrix represent whether there is a relationship between the vertices in the graph. The present embodiment takes the original data s _i = (I, love, Beijing, Forbidden City, Museum, Museum) as an example, and performs dictionary matching on each character in the original data to make full use of potential characters, and matches to obtain a dictionary feature data set l _i = {Beijing, Forbidden City, Museum}, which is used to construct a boundary graph B-graph, and fuses the dictionary feature data and the original data to obtain a boundary graph, and the vertex set V is the merged result of the dictionary feature data and the original data, and the data is stored in the adjacency matrix ^AB . If the matched word has multiple characters, the position of the word in the matrix and the first and last characters in the word is recorded as 1. Taking _si as an example, the dependency analyzer is used to perform part-of-speech analysis on the original data, and the original data set is cut into a set of characters and words d _i = {I, love, Beijing, Forbidden City, Museum}, which is used to construct the relationship graph R-graph and the inclusion graph C-graph. The word dependency information set is defined as _fi = {SBV, HED, ATT, ATT, VOB}, and is represented by a set index index _i = {2, 0, 5, 5, 2}. Among them, regarding the relationship graph R-graph, the dependency relationship between individual characters and the dependency relationship between words are considered at the same time, and the adjacency matrix ^AR and ^AC are used to store data. For the adjacency matrix ^AR , if the cut word contains multiple characters, the position of the word and all the characters contained in the word in the matrix is recorded as 1, and the position of each character and the remaining characters in the word is recorded as 1 in turn; if the cut word is related to another word, its position is recorded as 1. For the adjacency matrix ^AC , the positions of the cut word and the first and last characters in the word are recorded as 1, and the positions of the cut character and its predecessor and successor characters are recorded as 1.

In an embodiment of the present application, after extracting the feature data of the original data, the dictionary feature data can be input into an embedding layer to obtain a dictionary feature vector; the part-of-speech feature data can be input into an embedding layer to obtain a part-of-speech feature vector; and the original data can be input into an embedding layer to obtain a context information vector. After obtaining the feature data in vector form and the context information of the original data, the multi-graph matrix is constructed according to the original data and the feature data, and a boundary graph can be generated by fusing the dictionary feature vector and the context information vector, and a relationship graph and an inclusion graph can be generated by fusing the part-of-speech feature vector and the context information vector, respectively.

Specifically, as shown in FIG2 , the dictionary feature data set is defined as l = {l ₁ ,l ₂ ,...,l _m }, which is sent to the pre-trained embedding layer embedding to obtain a low-dimensional dense vector of the dictionary feature data, namely, the dictionary feature vector Where e ^w is a vocabulary embedding lookup table. Similarly, the part-of-speech feature data set w = {w ₁ ,w ₂ ,...,w _k } and the original data s = {c ₁ ,c ₂ ,..., _cn } are respectively sent to the embedding layer to obtain the part-of-speech feature vector x ^w = e ^w (w) and the vector representation of the original data x ^c = e ^c (c), where e ^c is the character embedding vector table. The original data is processed using the LSTM model. The specific process is described in formulas (2), (3) and (4):

The above data vector of the original data vector And the following data vector Perform XOR operation to obtain the context information vector H = {h ₁ ,h ₂ ,...,h _n } of the original data. Combine the dictionary feature vector with the original data vector to obtain the boundary graph B-graph, whose vertex set is The part-of-speech feature vector is combined with the original data vector to obtain the relationship graph R-graph and the inclusion graph C-graph, both of which share a set of vertex sets.

In an embodiment of the present application, as shown in FIG2 , the collaborative training after fusing the merged data and the multi-image matrix includes but is not limited to the following steps:

Inputting the fusion result of the first merged data and the boundary graph B-graph into the first graph attention network GAT1 for graph convolution processing, wherein the first merged data includes the merged data of the dictionary feature vector and the context information vector;

Inputting the fusion result of the second merged data and the relationship graph R-graph into the second graph attention network GAT2 for graph convolution processing, wherein the second merged data includes the merged data of the part-of-speech feature vector and the context information vector;

The fusion result of the second merged data and the inclusion graph C-graph is input into the third graph attention network GAT3 for graph convolution processing.

In this embodiment, the graph attention network GAT is used to model the three interaction graphs. Specifically, in the M-layer GAT of this embodiment, the input representation of the j-th layer consists of a set of node features, NF _J ={f ₁ ,f ₂ ,...,f _n }. In addition, this embodiment also introduces the adjacency matrix A, and satisfies _fi ∈ ^RF , A∈R ^N*N where F represents the dimension of the j-th layer features and N represents the number of nodes. The output representation of the j-th layer is a set of new node features NF(j+1) ={f1',f2',...,fN'} that are different from other nodes. Each GAT has K independent and different attention heads, which are specifically represented as shown in formulas (5) and (6):

Among them, || represents the connection operation; σ represents the nonlinear activation function; _Ni represents the neighboring nodes of node i in the graph, represents the attention parameter, W ^k ∈ ^RF' , a∈R ^2F' represents a single-layer feedforward neural network. KF' represents the dimension of the j-th layer output _fi ', F' represents the dimension of the final output feature, and the result retains the average value. Specifically, the average value is calculated as shown in formula (7):

Specifically, this embodiment constructs three independent graph attention networks GAT1, GAT2 and GAT3 to model three interaction graphs. The boundary graph B-graph is used to capture the boundary entity information of vocabulary knowledge, and its vertex set is The output node feature representation of the GAT1 model is shown in formula (8):

G ₁ =GAT1(X _B ,A ^B ) Formula (8)

Where G ₁ ∈ ^RF'* ( ^N+M) , n represents the number of characters in the sentence, and m represents the number of matched words. For the relationship graph R-graph and the inclusion graph C-graph, they share the same set of vertices. The output node feature representations of the GAT2 and GAT3 models are G ₂ and G ₃ , respectively, as shown in formula (9) and formula (10):

G ₂ =GAT2(X, ^AR ) Formula (9)

G ₃ =GAT3(X, ^AC ) Formula (10)

Among them, G _y ∈ ^RF'*(n+k) , y ∈ (2,3), n represents the number of characters contained in the sentence, and k represents the number of words obtained by part-of-speech parsing. For the output node features of the graph attention network, _Gi , i∈ (1,2,3), only the character representations of the first n columns of the matrix, _Qi = _Gi [:,0:n], i∈ (1,2,3), are retained for use by the decoding layer.

In an embodiment of the present application, the collaborative training after fusing the merged data and the multi-graph matrix also includes: inputting the context information of the boundary graph, the relationship graph, the inclusion graph and the original data into the fusion layer for parameter learning; and then performing collaborative training through multi-graph fusion to obtain feature fusion results. Specifically, the three character-word interaction graphs constructed in this embodiment have different functions, and the three interaction graphs are fused by the fusion layer to fully utilize their respective advantages. At the same time, considering the auxiliary role of the context information of the original data on Chinese NER, this embodiment inputs it together with the retained features output by the three graph attention networks into the fusion module R, where R is shown in formula (11):

R＝W ₁ H+W ₂ Q ₁ +W ₃ Q ₂ +W ₄ Q _3Formula (11)

Among them, W _y is the training matrix, y∈(1,2,3,4), and finally a collaboration matrix R is obtained from the fusion layer, which integrates the different functions of these graphs.

In an embodiment of the present application, the prediction result can be obtained by using the CRF decoding feature fusion result. Specifically, the standard conditional random field discriminant model CRF is used to capture the dependency between continuous labels. For any input data s = {c ₁ , c ₂ , ..., c _n }, the multi-image fusion is collaboratively trained to obtain the feature fusion result R = (r ₁ , r ₂ , ..., r _n ) as the input data of the CRF decoding layer. The Chinese named entity recognition result corresponding to the input data is obtained by decoding the CRF decoding layer. Specifically, given the predicted label sequence Y = (y ₁ , y ₂ , ..., y _n ), the calculation formula of the label sequence Y is shown in formula (12):

Where _y'i represents any tag sequence, and All are training parameters. This example uses the first-order Viterbi algorithm to find the label sequence with the highest score on the character-based input representation. Given a manually labeled training data The optimized model is obtained by using the L2 regularized sentence log-likelihood loss as shown in formula (13).

Among them, λ represents the L2 regularization parameter and Θ represents the training parameter set.

The embodiment of the present invention provides an entity recognition system based on a multi-graph collaborative semantic network, comprising:

An acquisition module, used for acquiring original data, wherein the original data includes a plurality of Chinese sentences;

An extraction module, used to extract feature data of the original data, wherein the feature data includes dictionary feature data and part-of-speech feature data;

A construction module, used for constructing a multi-image matrix according to the original data and the feature data;

A merging module, used for merging the original data and the feature data to obtain merged data;

A collaborative training module, used for fusing the combined data with the multi-image matrix and performing collaborative training to obtain a feature fusion result;

A decoding module is used to decode the feature fusion result to obtain the Chinese named entity recognition result of the original data.

The contents of the method embodiments of the present invention are all applicable to the system embodiments. The functions specifically implemented by the system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

The embodiment of the present invention provides an entity recognition system based on a multi-graph collaborative semantic network, comprising:

at least one memory for storing a program;

At least one processor is used to load the program to execute the entity recognition method based on multi-graph collaborative semantic network shown in FIG1 .

The contents of the method embodiments of the present invention are all applicable to the system embodiments. The functions specifically implemented by the system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

An embodiment of the present invention provides a storage medium storing a computer-executable program, wherein the computer-executable program is used to implement the entity recognition method based on a multi-graph collaborative semantic network shown in FIG1 when executed by a processor.

The embodiment of the present invention also provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the entity recognition method based on the multi-graph collaborative semantic network shown in FIG1.

The embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in the relevant technical field without departing from the purpose of the present invention. In addition, the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

Claims (7)

1. An entity recognition method based on a multi-graph collaborative semantic network, characterized by comprising the following steps: Acquire original data, wherein the original data includes a plurality of Chinese sentences; Extracting feature data of the original data, wherein the feature data includes dictionary feature data and part-of-speech feature data; Constructing a multi-image matrix according to the original data and the feature data; Merging the original data and the characteristic data to obtain merged data; The combined data and the multi-image matrix are fused and then collaboratively trained to obtain a feature fusion result; Decoding the feature fusion result to obtain the Chinese named entity recognition result of the original data; Wherein, after extracting the characteristic data of the original data, the method further comprises the following steps: Inputting the dictionary feature data into an embedding layer to obtain a dictionary feature vector; Inputting the part-of-speech feature data into an embedding layer to obtain a part-of-speech feature vector; Input the original data into the embedding layer to obtain a context information vector; The constructing a multi-image matrix according to the original data and the feature data comprises: fusing the dictionary feature vector and the context information vector to generate a boundary map; Fusion of the part-of-speech feature vector and the context information vector to generate a relationship graph and an inclusion graph respectively; The step of fusing the combined data with the multi-image matrix for collaborative training includes: Inputting the fusion result of the first merged data and the boundary graph into a first graph attention network for graph convolution processing, wherein the first merged data includes merged data of the dictionary feature vector and the context information vector; Inputting the fusion result of the second merged data and the relationship graph into a second graph attention network for graph convolution processing, wherein the second merged data includes merged data of the part-of-speech feature vector and the context information vector; The fusion result of the second merged data and the inclusion graph is input into a third graph attention network for graph convolution processing. 2. The entity recognition method based on multi-graph collaborative semantic network according to claim 1, characterized in that the step of extracting feature data of the original data comprises: Using a dictionary tool to match word feature data on the original data to obtain dictionary feature data; The original data is parsed by a part-of-speech parsing tool to obtain part-of-speech feature data. 3. According to claim 1, an entity recognition method based on a multi-graph collaborative semantic network is characterized in that the relationship graph includes dependency relationships between single times and dependency relationships between words. 4. The entity recognition method based on multi-graph collaborative semantic network according to claim 1, characterized in that the step of fusing the merged data and the multi-graph matrix for collaborative training further comprises: Inputting the boundary graph, the relationship graph, the inclusion graph and the context information of the original data into a fusion layer for parameter learning; Collaborative training via multi-image fusion. 5. An entity recognition system based on multi-graph collaborative semantic network, characterized by comprising: An acquisition module, used for acquiring original data, wherein the original data includes a plurality of Chinese sentences; An extraction module, used to extract feature data of the original data, wherein the feature data includes dictionary feature data and part-of-speech feature data; A construction module, used for constructing a multi-image matrix according to the original data and the feature data; A merging module, used for merging the original data and the feature data to obtain merged data; A collaborative training module, used for fusing the combined data with the multi-image matrix and performing collaborative training to obtain a feature fusion result; A decoding module, used to decode the feature fusion result to obtain the Chinese named entity recognition result of the original data; Wherein, after extracting the feature data of the original data, the extraction module is further used to perform the following steps: Inputting the dictionary feature data into an embedding layer to obtain a dictionary feature vector; Inputting the part-of-speech feature data into an embedding layer to obtain a part-of-speech feature vector; Input the original data into the embedding layer to obtain a context information vector; The constructing a multi-image matrix according to the original data and the feature data comprises: fusing the dictionary feature vector and the context information vector to generate a boundary map; Fusion of the part-of-speech feature vector and the context information vector to generate a relationship graph and an inclusion graph respectively; The step of fusing the combined data with the multi-image matrix for collaborative training includes: Inputting the fusion result of the first merged data and the boundary graph into a first graph attention network for graph convolution processing, wherein the first merged data includes merged data of the dictionary feature vector and the context information vector; Inputting the fusion result of the second merged data and the relationship graph into a second graph attention network for graph convolution processing, wherein the second merged data includes merged data of the part-of-speech feature vector and the context information vector; The fusion result of the second merged data and the inclusion graph is input into a third graph attention network for graph convolution processing. 6. An entity recognition system based on multi-graph collaborative semantic network, characterized by comprising: at least one memory for storing a program; At least one processor is used to load the program to execute the entity recognition method based on a multi-graph collaborative semantic network as described in any one of claims 1-4. 7. A storage medium, characterized in that a computer executable program is stored therein, and when the computer executable program is executed by a processor, it is used to implement the entity recognition method based on a multi-graph collaborative semantic network as described in any one of claims 1-4.