A chapter-level event role identification method guided by knowledge graph information
By constructing text knowledge graphs and combining graph convolution networks and Transformer coding ends, the problem of long-sequence and long-distance dependence is solved, and the accuracy and deep understanding of event role recognition are improved.
Patent Information
- Application Number
- CN202210570151.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-24
- Publication Date
- 2025-04-11
- Estimated Expiration
- 2042-05-24
AI Technical Summary
When the existing event role recognition model crosses sentence limitations, it is difficult to effectively capture long-distance text features in long sequences, resulting in inaccurate and in-depth semantic understanding, and ignores the correlation between text key information and its information.
By constructing a text knowledge graph, using a graph convolution network to extract text knowledge graph features, and integrating them into text feature representation, combining Transformer encoding end and Bi-LSTM network to achieve the fusion of text information of different granularity, and finally identifying event role information through the CRF layer.
It enhances the understanding of text semantics, improves the accuracy and depth of event role recognition, and realizes effective processing of long-sequence and long-distance dependence.
Smart Images
Figure CN114880434B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a chapter-level event role recognition method based on knowledge graph information guidance, and belongs to the technical field of natural language processing. Background Art
[0002] Chinese news text event role recognition aims to use sequence annotation patterns to recognize event role information in news texts. The event role information of a document is an expression of the main information of the document, which helps to quickly understand the text content. Most of the existing event role extraction models focus on the sentence level, and document-level event role extraction is different from extracting events from independent sentences. The document-level event extraction model has special requirements for sequence feature extraction. Crossing sentence limitations and capturing long-distance text feature extraction in long sequences, the loss of text information is still the biggest difficulty in the document event role recognition task. In the current feature extraction process, the model only focuses on the semantic information of the context, but ignores the influence of the key information of the text and the correlation between its information on the semantic information of the text, resulting in the model's semantic understanding of the text is still not accurate and in-depth enough. The present invention is of great significance in solving the problem of text information loss in the process of long sequence feature extraction by constructing a text knowledge graph. Summary of the invention
[0003] The present invention provides a chapter-level event role recognition method based on knowledge graph information guidance, which is used to solve the long-sequence and long-distance dependency problem in the current chapter-level event role recognition.
[0004] The technical solution of the present invention is: a chapter-level event role identification method based on knowledge graph information guidance, and the specific steps of the chapter-level event role identification method based on knowledge graph information guidance are as follows:
[0005] Step 1: Extract triple information from text sentences and cluster triple elements based on entities, high-frequency words, keywords and other information to build associations between triple elements and form knowledge graph information of the document;
[0006] Step 2: Vectorize the text using two different methods: pre-trained Bert and pre-trained word vectors as the input of the encoder;
[0007] Step 3: Encode the document through the Transformer encoding end;
[0008] Step 4: Encode the text sentence level through the Bi-LSTM network;
[0009] Step 5: Use the graph convolution module to jointly learn the vector representation of the text and the knowledge graph information to obtain the vector representation of the fused knowledge graph information;
[0010] Step 6. Finally, the fusion of text information of different granularities is realized through the gating fusion mechanism in the fusion reasoning module, and then the CRF layer is used to perform event role information recognition reasoning to complete the recognition of event role information.
[0011] As a preferred embodiment of the present invention, the specific steps of Step 1 are:
[0012] Step 1.1, use the Stanford tool to extract the triple information of the sentences in the text, and build interdependent relationships between the elements in each triple;
[0013] Step 1.2: Clustering is performed based on the elements of the sentence triples by judging whether the elements of the sentence triples are high-frequency words, keywords, entity information, etc. of the document. The elements of the same type are built into relationships with each other, and the graph structure within the triples is combined to build a global graph structure of the text, thereby forming a knowledge graph of the document's semantic information;
[0014] As a preferred embodiment of the present invention, the specific steps of Step 2 are:
[0015] Step 2.1: First, convert the text into a vector representation according to the pre-trained word vector table, and obtain the character embedding X of the text. c :
[0016] X c =Embedding(T c ) (1)
[0017] Among them, T c ={c1,c2,c3……c n} represents each character of the text. Denotes a character vector representing a set.
[0018] After the text data is segmented by the Bert vocabulary, the input sequence D = {d1, d2, d3...d m}, the final output hidden state vector representation X is obtained through the Bert pre-training model b :
[0019] X b = Bert(D) (2)
[0020] in, A collection of Bert representations representing text.
[0021] Step 2.2, use a cross attention network between the pre-trained word vector text representation and the pre-trained model Bert text representation. The character-level vector representation of the text is obtained from the pre-trained vector table And through the Bert model, we get In order to enrich the semantic information of text character embedding representation, X b The representation of X c Then we get the jth character representation x′ in the text j :
[0022]
[0023]
[0024]
[0025] Among them, X′ e ={x′1,x′2,x′3……x′ n} represents the text representation set obtained by the representation module. Attention weight α j.i is the corresponding matching score m j.i Then based on the feature vector and The matching score is calculated by the bilinear product of
[0026] This preferred design is an important part of the present invention, which mainly provides a vector encoding process for the present invention, uses an encoder to vectorize the text, and on this basis, effectively integrates the text representation of the traditional pre-trained word vector and the text representation of the pre-trained Bert through a cross-attention machine network.
[0027] As a preferred solution of the present invention, the specific steps of Step 3 are: extracting features from the text through the Transform encoding end. Embedding the text representation X′ obtained by the representation fusion module e ={x′1,x′2,x′3……x′ n}, when X′ e After entering the Transformer encoding end, it first undergoes the calculation of the self-attention mechanism, so that the model can learn the correlation between the semantics within the text, and then obtain The calculation process is shown in Equation 6-Equation 7:
[0028] Query = Value = Key = X' e (6)
[0029]
[0030] Obtained through the self-attention mechanism It realizes the dynamic fusion of related information within the text. The first layer output H1 of the Transformer encoder is obtained through the feedforward network, and then the final output H of the Transformer encoder is obtained through multiple layers of encoding layers. t :
[0031]
[0032] h t =Attention(H t-1 ,H t-1 ,H t-1 ), t>1 (9)
[0033] H t =FNN(h t ), t>1 (10)
[0034] As a preferred solution of the present invention, the specific steps of Step 4 are: extracting features from the sentence-level text of the text through the BI-LSTM encoder. Obtaining the sentence-level text vector representation of the text from the pre-trained vector table By extracting text features at the sentence level, the model can learn fine-grained semantic information through sentence-level context encoding, and then obtain sentence-level text representation.
[0035]
[0036] in, The vector representation of the first sentence of the text, Sentence-level representation of each token in a paragraph.
[0037] The Step 5 includes:
[0038] Feature extraction of text knowledge graph information through graph convolutional network;
[0039] The cross-attention network is used to achieve a deep fusion of the features of knowledge graph information and text features.
[0040] As a preferred embodiment of the present invention, the specific steps of Step 5 are:
[0041] Step 5.1. Based on the knowledge graph, we create nodes with the characters of the elements in the triples. Edges are added between the characters of the elements within the triples, and edges are added between the characters of the elements of the cross-sentence triples built together by clustering, thus forming a graph ζ=(γ,ε), where γ=v1,v2,v3……v n (|γ|=n) and ε are the node set and edge set respectively. i It represents the character v in the triple iNodes, each edge (v i ,v j )∈ε is the character v i and the character v j In order to enable the information flow between nodes, we equate the relationship between nodes to the forward flow, reverse flow and self-circulation flow of information between nodes, so the final type label is K(v i ,v j ):
[0042]
[0043] Step 5.2: Text feature representation H t As the input of the graph convolutional network, to initialize the node representation of the first layer of the network In the kth layer of the syntactic graph convolutional network module, the graph convolution vector of node v∈γ can be calculated by the following method
[0044]
[0045] Where k(u,v) represents the adjacency matrix constructed by the type label of edge (u,v); and are the weight matrix and bias of a certain type label k(u,v); N(v) is the neighborhood set of v, including v (due to self-loop); f is a nonlinear activation function.
[0046] Step 5.3: Obtain the text hidden state vector from the Transform encoding end Then, through the cross attention network, the knowledge graph feature representation obtained by the graph convolution module is integrated into the text representation to obtain a new text representation document-level representation H′={h′1,h′2…h′ n}.
[0047] This preferred design can effectively aggregate key information in the text. The graph convolution module is designed to capture the correlation between key information in the text. The relationship between the elements of the text triplet can enhance the dependency between information across sentences, thereby improving the model's deep understanding of text semantics.
[0048] As a preferred embodiment of the present invention, the specific steps of Step 6 are:
[0049] Step 6.1: Gating flux through gating mechanism Control sentence level and document-level representation (n: the i-th representation of the n-th sentence in the vector of the document-level representation) Fusion of text representations at two granularities:
[0050]
[0051]
[0052] where ⊙ represents the element-wise product.
[0053] Step 6.2: Joint modeling through conditional random fields to mark the decision. After passing through a linear layer, we get P′, where P′ i,j represents the score of the label j of the i-th character. Let the input sequence be X = {x1, x2, x3…, x m ,}, the label sequence is Y = {y1,y2,y3…,y m}, the score of label prediction is:
[0054]
[0055]
[0056] Among them, Q is the transfer score matrix, Q i,j represents the score of transferring from label i to label j. Apply the Softmax function to the scores of all possible label sequences to obtain the probability P(x / y) of label sequence y given the input x. During model training, the logarithmic probability of the positive label sequence is maximized. During decoding, the model predicts the output sequence with the highest score and uses it as the event role recognition sequence.
[0057] This preferred solution designs a model for the knowledge graph constructed from text, which helps to better understand the text semantic information and improve the performance of identifying event roles. The beneficial effects of the present invention are:
[0058] 1. The chapter-level event role recognition method based on knowledge graph information guidance of the present invention uses the knowledge graph information constructed by text triples, extracts text knowledge graph features through graph convolutional networks and integrates them into the feature representation of text, thereby enhancing the semantic understanding of text;
[0059] 2. The chapter-level event role recognition method based on knowledge graph information guidance of the present invention realizes effective fusion of text information of different granularities through a selective fusion network, thereby achieving effective utilization of text semantics. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 This is a flow chart of the chapter-level event role identification method based on knowledge graph information guidance proposed by the present invention;
[0061] Figure 2 This is a model diagram of the chapter-level event role identification method based on knowledge graph information guidance proposed by the present invention;
[0062] Figure 3 The knowledge graph construction of the chapter-level event role identification method based on knowledge graph information guidance proposed in the present invention. DETAILED DESCRIPTION
[0063] Example 1: Figure 1-Figure 3 As shown, the chapter-level event role identification method based on knowledge graph information guidance of the present invention comprises:
[0064] Step 1: Extract triple information from text sentences and cluster triple elements based on entities, high-frequency words, keywords and other information to build the association relationship between triple elements and form the knowledge graph information of the document.
[0065] Step 2: Vectorize the text using two different methods: pre-trained Bert and pre-trained word vectors as the input of the encoder.
[0066] Step 3: Encode the document through the Transformer encoding end.
[0067] Step 4: Encode the text at the sentence level through the Bi-LSTM network.
[0068] Step 5: Use the graph convolution module to jointly learn the vector representation of the text and the knowledge graph information to obtain the vector representation of the fused knowledge graph information.
[0069] Step 6. Finally, the fusion of text information of different granularities is realized through the gating fusion mechanism in the fusion reasoning module, and then the CRF layer is used to perform event role information recognition reasoning to complete the recognition of event role information.
[0070] As a preferred embodiment of the present invention, the specific steps of Step 1 are:
[0071] Example:
[0072] [S1] On May 6, the second batch of 28 tons of chicken products from a certain city base under a certain group was shipped to a certain country.
[0073] [S2] On March 17, the base in a certain city exported 24 tons of chicken to a certain country for the first time. This is the group's active integration into the green belt and road and its successful entry into the chicken markets of ten countries in a certain alliance.
[0074] [S3] This year, we plan to export 5,000 tons of chicken products to a certain country, and we are expected to enter the market of ten countries in a certain alliance next year or the year after!
[0075] …
[0076] Step 1.1, use the Stanford tool to extract the triple information of the sentences in the text, and build interdependent relationships between the elements in each triple;
[0077] Step 1.2: Clustering is performed based on the elements of the sentence triples by judging whether the elements of the sentence triples are high-frequency words, keywords, entity information, etc. of the document. The elements of the same type are built into relationships with each other, and the graph structure within the triples is combined to build a global graph structure of the text, thereby forming a knowledge graph of the document's semantic information;
[0078] As a preferred embodiment of the present invention, the specific steps of Step 2 are:
[0079] Step 2.1: First, convert the text into a vector representation according to the pre-trained word vector table, and obtain the character embedding X of the text. c :
[0080] X c =Embedding(T c ) (1)
[0081] Among them, T c ={c1,c2,c3……c n} represents each character of the text. Denotes a character vector representing a set.
[0082] After the text data is segmented by the Bert vocabulary, the input sequence D = {d1, d2, d3...d m}, the final output hidden state vector representation X is obtained through the Bert pre-training model b :
[0083] X b = Bert(D) (2)
[0084] in, A collection of Bert representations representing text.
[0085] Step 2.2, use a cross attention network between the pre-trained word vector text representation and the pre-trained model Bert text representation. The character-level vector representation of the text is obtained from the pre-trained vector table And through the Bert model, we get In order to enrich the semantic information of text character embedding representation, X b The representation of X c Then we get the jth character representation x′ in the text j :
[0086]
[0087]
[0088]
[0089] Among them, X′ e ={x′1,x′2,x′3……x′ n} represents the text representation set obtained by the representation module. Attention weight α j.i is the corresponding matching score m j.i Then based on the feature vector and The matching score is calculated by the bilinear product of
[0090] This preferred design is an important part of the present invention, which mainly provides a vector encoding process for the present invention, uses an encoder to vectorize the text, and on this basis, effectively integrates the text representation of the traditional pre-trained word vector and the text representation of the pre-trained Bert through a cross-attention machine network.
[0091] As a preferred solution of the present invention, the specific steps of Step 3 are: extracting features from the text through the Transform encoding end. Embedding the text representation X′ obtained by the representation fusion module e ={x′1,x′2,x′3……x′ n}, when X′ e After entering the Transformer encoding end, it first undergoes the calculation of the self-attention mechanism, so that the model can learn the correlation between the semantics within the text, and then obtain The calculation process is shown in Equation 6-Equation 7:
[0092] Query = Value = Key = X e (6)
[0093]
[0094] Obtained through the self-attention mechanism It realizes the dynamic fusion of related information within the text. The first layer output H1 of the Transformer encoder is obtained through the feedforward network, and then the final output H of the Transformer encoder is obtained through multiple layers of encoding layers. t :
[0095]
[0096] h t =Attention(Ht-1 ,H t-1 ,H t-1 ), t>1 (9)
[0097] H t =FNN(h t ), t>1 (10)
[0098] As a preferred solution of the present invention, the specific steps of Step 4 are: extracting features from the sentence-level text of the text through the BI-LSTM encoder. Obtaining the sentence-level text vector representation of the text from the pre-trained vector table By extracting text features at the sentence level, the model can learn fine-grained semantic information through sentence-level context encoding, and then obtain sentence-level text representation.
[0099]
[0100] in, The vector representation of the first sentence of the text, Sentence-level representation of each token in a paragraph.
[0101] As a preferred embodiment of the present invention, the specific steps of Step 5 are:
[0102] Step 5.1. Based on the knowledge graph, we create nodes with the characters of the elements in the triples. Edges are added between the characters of the elements within the triples, and edges are added between the characters of the elements of the cross-sentence triples built together by clustering, thus forming a graph ζ=(γ,ε), where γ=v1,v2,v3……v n (|γ|=n) and ε are the node set and edge set respectively. i It represents the character v in the triple i Nodes, each edge (v i ,v j )∈ε is the character v i and the character v j In order to enable the information flow between nodes, we equate the relationship between nodes to the forward flow, reverse flow and self-circulation flow of information between nodes, so the final type label is K(v i ,v j ):
[0103]
[0104] Step 5.2: Text feature representation H t As the input of the graph convolutional network, to initialize the node representation of the first layer of the network In the kth layer of the syntactic graph convolutional network module, the graph convolution vector of node v∈γ can be calculated by the following method
[0105]
[0106] Where k(u,v) represents the adjacency matrix constructed by the type label of edge (u,v); and are the weight matrix and bias of a certain type label k(u,v); N(v) is the neighborhood set of v, including v (due to self-loop); f is a nonlinear activation function.
[0107] Step 5.3: Obtain the text hidden state vector from the Transform encoding end Then, through the cross attention network, the knowledge graph feature representation obtained by the graph convolution module is integrated into the text representation to obtain a new text representation document-level representation H′={h′1,h′2…h′ n}.
[0108] This preferred design can effectively aggregate key information in the text. The graph convolution module is designed to capture the correlation between key information in the text. The dependency between information across sentences can be enhanced through the relationship between text triple elements.
[0109] As a preferred embodiment of the present invention, the specific steps of Step 6 are:
[0110] Step 6.1: Gating flux through gating mechanism Control sentence level and document-level representation (n: the i-th representation of the n-th sentence in the vector of the document-level representation) Fusion of text representations at two granularities:
[0111]
[0112]
[0113] where ⊙ represents the element-wise product.
[0114] Step 6.2: Joint modeling through conditional random fields to mark the decision. After passing through a linear layer, we get P′, where P′ i,j represents the score of the label j of the i-th character. Let the input sequence be X = {x1, x2, x3…, x m ,}, the label sequence is Y = {y1,y2,y3…,y m}, the score of label prediction is:
[0115]
[0116]
[0117] Among them, Q is the transfer score matrix, Q i,j represents the score of transferring from label i to label j. Apply the Softmax function to the scores of all possible label sequences to obtain the probability P(x / y) of label sequence y given the input x. During model training, the log probability of the positive label sequence is maximized. During decoding, the model predicts the output sequence with the highest score and uses it as the event role recognition sequence.
[0118] Step 7. Parameter comparison tests and ablation experiments were carried out in combination with the model in this paper to prove the rationality and efficiency of the model settings. The model was compared with the existing models to prove that this method has a good effect in event role recognition.
[0119] The experiment evaluates the experimental results from two aspects: "Head Noun Match" and "Exact Match"; "Head Noun Match" compares whether the extracted content is within the range of the label content. "Exact Match" compares whether the extracted content is completely consistent with the label content. The results of this paper are reported as the precision (P), recall (R), and F-measure (F-1) scores of the macro average of all event roles.
[0120] Precision (P): The proportion of correctly predicted events in the total predicted events.
[0121] Recall (R): The proportion of correctly predicted events to real events.
[0122] F-1:
[0123] In order to verify the impact of the number of graph convolution layers on model performance, this chapter set up 1 layer, 2 layers, 3 layers and 4 layers for experiments, and finally found the optimal number of layers for knowledge graph information feature extraction. The experimental results are shown in Table 1:
[0124] Table 1 Experimental results of graph convolution layers
[0125]
[0126] The experiment of graph convolution layers is shown in Table 1. As the number of convolution layers increases, the performance of the model gradually improves. When the number of graph convolution layers reaches 3, the model performance is optimal. Comparing the results of 1-layer and 2-layer graph convolution, it can be analyzed that when the number of graph convolution modules is shallow, the knowledge graph information is not fully utilized, and the model's understanding of the text is not deep enough. When the number of layers is 4, the model performance decreases. Since the model layer is too deep at this time, there is a tendency to overfit the text features, which affects the performance of the model in the test set. Therefore, in subsequent experiments, the model uses 3-layer graph convolution.
[0127] To verify the effectiveness of the proposed method for event role recognition, four baseline methods were selected for testing. Table 2 shows the comparative experimental results of the document-level event extraction model, and Table 3 shows the results of each event role. The experiments show that the F1 value of the model in this chapter exceeds that of other baseline models.
[0128] GLACIER: This method consists of a sentence event classifier and a set of role filling identifiers. Finally, event information is extracted by multiplying the normalized sentence and phrase probabilities.
[0129] TIER: This method divides documents into three stages: narrative document classification, event sentence recognition, and noun phrase analysis, through which event role information is determined.
[0130] Cohesion Extraction: This method adopts a bottom-up approach, first actively identifying candidate role filler words in the document, and then using a sentence classifier to refine the candidate set to determine the location of event information.
[0131] Multi-Granularity Reader: This method dynamically aggregates information obtained from representations of different granularities (such as sentences and paragraphs) to enable the model to extract event roles by using semantic information from multi-level document contexts.
[0132] Table 2 Comparative experimental results
[0133]
[0134] (1) Compared with the traditional end-to-end model, the performance of our model is significantly better than that of the existing model in terms of "Head NounMatch" and "Exact Match" with the assistance of knowledge graph information.
[0135] (2) Compared with the long text input models such as "GLACIER", "TIER" and "Cohesion Extract", "Multi-Granularity Reader" uses sentence information dynamically to enable the model to pay attention to a wider context and shows better results. This paper aggregates multi-granular information (sentences, paragraphs and knowledge graph structures) to enable the neural network model to have a deeper understanding of the text, and the model shows better performance. Paying attention to local context and broader context plays an important role in the task of event detection.
[0136] (3) By comparing the R values, the values of the proposed model in "Head Noun Match" and "Exact Match" are significantly higher than those of other models, indicating that with the assistance of graph information, the proposed model is very accurate in locating positive samples, which ultimately greatly improves the overall performance of the model. All of these show that with the assistance of knowledge graph information, the model can have a deeper understanding of the text.
[0137] To verify the effectiveness of the structured knowledge graph information and Transformer structured encoding in this model, as shown in Table 3 (the table description "(-)" indicates that the network structure is not used):
[0138] Table 3 Ablation experiment results
[0139]
[0140] The results of the ablation experiment are shown in Table 3. The P, R, and F1 values of the model in this paper all exceed those of other baseline models. When the model does not use the Transform encoding structure and does not incorporate knowledge graph information, the P, R, and F1 of the model all decrease. Therefore, it can be proved that the Transform encoding structure and the integration of knowledge graph information can better assist the model in understanding the semantics of the text, which is very important for improving the performance of the model and is also an indispensable part of the model. At the same time, compared with the Transform encoding structure, since the traditional Bi-LSTM encoder does not perform deep feature extraction on the text, the model performance also decreases, indicating that the Transform encoding structure is also helpful for model improvement. When the model does not use the graph convolution module to understand the knowledge graph information, the P, R, and F1 of the model are all decreasing. Therefore, it can be proved that the model can effectively extract and integrate the deep features of the sentence with the help of knowledge graph information, which can better assist the model in accurately identifying event information.
[0141] The specific implementation modes 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 implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.
Claims
1. A chapter-level event role identification method based on knowledge graph information guidance, characterized by: The specific steps of the chapter-level event role identification method based on knowledge graph information guidance are as follows: Step 1: Extract triple information from text sentences and cluster triple elements based on entity, high-frequency word, and keyword information to build the association relationship between triple elements and form the knowledge graph information of the document; Step 2: Vectorize the text using two different methods: pre-trained Bert and pre-trained word vectors as the input of the encoder; Step 3: Encode the document through the Transformer encoding end; Step 4: Encode the text sentence level through the Bi-LSTM network; Step 5: Use the graph convolution module to jointly learn the vector representation of the text and the knowledge graph information to obtain the vector representation of the fused knowledge graph information; Step 6. Finally, the fusion of text information of different granularities is realized through the gating fusion mechanism in the fusion reasoning module, and then the CRF layer is used to perform event role information recognition reasoning to complete the recognition of event role information; The specific steps of Step 5 are as follows: Step 5.
1. On the basis of the constructed knowledge graph, nodes are created with the characters of the elements in the triples; edges are added between the characters of the elements within the triples, and edges are added between the characters of the elements of the cross-sentence triples constructed together by clustering, thus forming a graph ζ=(γ,ε), where γ=v1,v2,v3……v n (|γ|=n) and ε are the node set and edge set respectively; each v i It represents the character v in the triple i Nodes, each edge (v i ,v j )∈ε is the character v i and the character v j In order to enable the information flow between nodes, the relationship between nodes is equivalent to the forward flow, reverse flow and self-circulation flow of information between nodes, so the final type label is K(v i ,v j ): Step 5.2: Text feature representation H t As the input of the graph convolutional network, to initialize the node representation of the first layer of the network In the kth layer of the syntactic graph convolutional network module, the graph convolution vector of node v∈γ is calculated by the following method Where k(u,v) represents the adjacency matrix constructed by the type label of edge (u,v); and are the weight matrix and bias of a certain pointing type label k(u,v); N(v) is the neighborhood set of v, including v; f is a nonlinear activation function; Step 5.3: Obtain the text hidden state vector from the Transform encoding end Then, through the cross attention network, the knowledge graph feature representation obtained by the graph convolution module is integrated into the text representation to obtain a new text representation, the document-level representation H ′ ={h ′ 1,h ′ 2…h ′ n }; The specific steps of Step 6 are: Step 6.1, through the gating fusion network, the sentence-level feature information of the text is gated and fused with the text feature information, and the model has a more comprehensive understanding of the text semantics through text information of different granularities; Step 6.2: Use the CRF network to identify and annotate the text to achieve effective identification of event role information.
2. The method for identifying chapter-level event roles based on knowledge graph information guidance according to claim 1 is characterized by: The specific steps of Step 1 are: Step 1.1, use the Stanford tool to extract the triple information of the sentences in the text, and build interdependent relationships between the elements in each triple; Step 1.2: Extract high-frequency words, keywords, and entity words from the text, and determine whether the triple elements of different sentences belong to a certain type. Cluster the triple elements and build the association between cross-sentence triples to form a text knowledge graph.
3. The method for identifying chapter-level event roles based on knowledge graph information guidance according to claim 1 is characterized in that: The specific steps of Step 2 are: Step 2.1, first use the pre-trained word vector and pre-trained Bert model to vectorize the text respectively; Step 2.2, then use the cross attention network to jointly learn the obtained pre-trained word vector representation and the vector representation of the pre-trained Bert model to obtain the text vector representation that integrates the pre-trained Bert model.
4. The method for identifying chapter-level event roles based on knowledge graph information guidance according to claim 1 is characterized in that: In the Step 3, the feature encoding of the text vector is performed using the Transformer encoding end to obtain the feature representation of the text.
5. The method for identifying chapter-level event roles based on knowledge graph information guidance according to claim 1 is characterized in that: In the Step 4, a BI-LSTM encoder is used to perform feature encoding on each sentence in the text to obtain a sentence-level feature representation of the text.
6. The method for identifying chapter-level event roles based on knowledge graph information guidance according to claim 1 is characterized by: The Step 5 includes: Feature extraction of text knowledge graph information through graph convolutional network; The cross-attention network is used to achieve a deep fusion of the features of knowledge graph information and text features.