Equipment knowledge field document-level relation extraction method based on heterogeneous graph evidence guidance

Abstract

The invention discloses an equipment knowledge field document-level relation extraction method based on heterogeneous graph evidence guidance, and relates to the technical field of deep learning. According to the equipment knowledge field document-level relation extraction method based on heterogeneous graph evidence guidance, through a teacher-student framework based on an improved ATLOP model, effective evidence is automatically retrieved from remote supervision data, dependence on a large amount of manual annotation data is reduced, the data annotation cost is reduced, the expandability of the model is improved, and the data annotation efficiency is improved. By constructing a weighted heterogeneous graph and utilizing a contribution degree distinguishing rule, the model can focus on sentences with high contribution degree to relation prediction, noise brought by irrelevant sentences is removed, feature enhancement is performed on the weighted heterogeneous graph in combination with a graph convolutional neural network, nodes can better integrate adjacent node information, and the correlation prediction accuracy is improved. Meanwhile, the relation prediction performance of the entity is further improved through multiple embedding modes, and it is ensured that the model can capture key information more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and specifically to a method for document-level relation extraction in the equipment knowledge field guided by heterogeneous graph evidence. Background Art

[0002] With the rapid development of information technology, a large amount of semi-structured and unstructured data in the equipment knowledge field has emerged. These data are often scattered and redundant. In view of the massive complex text data in the equipment knowledge field, in order to better obtain valuable information from it and make full use of it, it is usually necessary to convert it into structured equipment domain knowledge graph data. In response to this demand, information extraction technology for text has emerged, and relation extraction is one of the core technologies of information extraction. In the actual application scenarios of the equipment knowledge field, a large number of relation facts are often expressed by multiple sentences, that is, they are contained in the long text in the complex context of the equipment knowledge field. According to statistics on Wikipedia data, a certain amount of relation facts need to be inferred from multiple sentences. In view of these situations, the sentence-level relation extraction method will no longer be applicable. Therefore, document-level relation extraction for the equipment knowledge field is a technical difficulty that needs to be focused on overcoming currently. Document-level relation extraction in the equipment knowledge field can better handle complex long text extraction tasks such as cross-sentence, cross-paragraph, and cross-document in the equipment knowledge field, providing a new solution idea for the relation extraction of long text in the complex context of the equipment knowledge field, and having important application value and significance for constructing the equipment knowledge graph.

[0003] Currently, there are still certain limitations in document-level relation extraction in the equipment knowledge field based on heterogeneous graphs. On the one hand, there is the problem of sentence noise. When inferring relations in multiple sentences, the importance of sentences in the document to entity pairs is not equal. Some sentences irrelevant to relation prediction will introduce noise and affect the performance of the model. The graph convolutional neural network depends on the information transfer of neighbor nodes when updating nodes. If the neighbors contain irrelevant sentence nodes, it may cause these irrelevant information to interfere with node update, thus affecting the accuracy of the model. On the other hand, there is the problem of limited quantity of manually labeled evidence. In the document-level relation extraction dataset, the quantity of manually labeled evidence is limited, and the method of obtaining evidence annotation at low cost has not been fully explored. Although it is possible to automatically collect silver training data for document-level relation extraction through remote supervision, it is still very important to find silver evidence data for document-level relation extraction in the document. In view of the deficiencies of the prior art, the present invention provides a method for document-level relation extraction in the equipment knowledge field guided by heterogeneous graph evidence to solve the above problems. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides a method for document-level relation extraction in the field of equipment knowledge guided by heterogeneous graph evidence. Through a teacher-student framework based on an improved ATLOP model, valid evidence is automatically retrieved from remotely supervised data, reducing the dependence on a large amount of manually annotated data, lowering the data annotation cost, and improving the scalability of the model. By constructing a weighted heterogeneous graph and using contribution degree discrimination rules, the model can focus on sentences with high contribution to relation prediction, removing the noise brought by irrelevant sentences. Combining with a graph convolutional neural network to enhance the features of the weighted heterogeneous graph enables nodes to better integrate adjacent node information. At the same time, through various embedding methods, the performance of entity pairs in relation prediction is further improved, ensuring that the model can more accurately capture key information.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for document-level relation extraction in the field of equipment knowledge guided by heterogeneous graph evidence, which adopts a heterogeneous graph evidence guidance strategy. The specific steps include:

[0006] Step S1, encoding and embedding the document through a BERT pre-trained language model to construct an unweighted heterogeneous graph containing mention nodes, entity nodes, and sentence nodes;

[0007] Step S2, using a teacher model constructed based on ATLOP trained on manually annotated data to automatically retrieve silver evidence from remotely supervised data, and training a student model on the remotely supervised data to obtain valid evidence data;

[0008] Step S3, using contribution degree discrimination rules to construct the unweighted heterogeneous graph into a weighted heterogeneous graph based on the valid evidence data;

[0009] Step S4, using a graph convolutional neural network for feature enhancement to enable nodes to contain weighted adjacent node information;

[0010] Step S5, constructing entity pair embeddings based on multiple embeddings and making predictions on the basis of an entity pair classifier.

[0011] Preferably, through the relationship reasoning between mention nodes, entity nodes, and sentence nodes in the unweighted heterogeneous graph, combined with the text features generated by BERT, the accuracy and robustness of document-level relation extraction are improved.

[0012] Preferably, the teacher model is trained on manually annotated data and applied to remotely supervised data to extract silver evidence. The student model is trained using remotely supervised data, and automatic evidence retrieval is performed under the supervision of silver evidence to obtain valid evidence data to improve the contribution degree discrimination of the weighted heterogeneous graph.

[0013] Preferably, the teacher model based on the improved ATLOP directly guides the model to focus on the evidence, supervises the calculation of the entity's specific local context embedding, which is generated by weighted summation of all token embeddings based on the attention mechanism of the BERT encoder, and its training objective is to assign different weights to the evidence, with higher or lower weights.

[0014] Preferably, the supervision of the student model consists of the binary cross-entropy loss of document-level relation extraction and the self-training loss of automatic evidence retrieval. The KL divergence loss is used to train the automatic evidence retrieval of the student model, and the KL divergence loss formula is as follows:

[0015]

[0016] Preferably, the loss of all entity pairs is calculated to prevent the loss of important instances, and the overall loss is balanced by hyperparameters. The hyperparameter formula is as follows:

[0017]

[0018] Preferably, after training on remotely supervised data, the student model is further fine-tuned using manually annotated data, and the supervision signal improves its knowledge of document-level relation extraction and automatic evidence retrieval.

[0019] Preferably, for automatic evidence retrieval, static thresholding is applied, and sentences with importance higher than a predefined threshold are selected as evidence to obtain the evidence sentences for entity pairs (e h , e t )

[0020]

[0021] Preferably, the contribution degree discrimination module assigns contribution weights to each edge according to specific rules; the types of edges are divided into the following four types:

[0022] First, M-E: independent of the sentence and without self-loop edges;

[0023] Second, M-M: independent of the sentence but with self-loop edges;

[0024] Third, M-S and E-S: related to the sentence and without self-loop edges;

[0025] Fourth, S-S: related to the sentence and with self-loop edges;

[0026] Self-loop edges are introduced to retain the self-information of the nodes. An edge pointing to itself is added to each node to avoid the loss of node information.

[0027] Preferably, the entity pair classifier predicts the relationship based on the final entity pair embedding, and the main components of the entity pair embedding are as follows:

[0028] The head entity embedding e h and the tail entity embedding e t : capturing the semantic expression of the context information of the head and tail entities in the heterogeneous graph;

[0029] The absolute value of the difference between the head and tail entity vectors |e h -e t |: quantifying the feature differences between entities and reflecting the relationship strength and directionality existing between entities;

[0030] The element-wise product of the head and tail entity vectors e h ⊙e t : capturing the interaction patterns between the head and tail entity embeddings and providing multi-dimensional semantic information for relationship modeling;

[0031] Evidence sentence information Prioritize the sentence information closely related to the entity pair relationship to ensure that the model can more accurately capture the key information supporting the relationship prediction.

[0032]

[0033] The probability function is as follows.

[0034] P(r|e h ,e t ) = Sigmoid(W2σ(W1(e h ,e t ) + b1) + b2).

[0035] The present invention discloses a method for document-level relationship extraction in the field of equipment knowledge guided by heterogeneous graph evidence, and its beneficial effects are as follows:

[0036] 1. The method for document-level relationship extraction in the field of equipment knowledge guided by heterogeneous graph evidence constructs a teacher-student framework through an improved ATLOP model, and uses the teacher model to automatically retrieve silver evidence from remotely supervised data, and further trains the student model to obtain effective evidence data. This method not only reduces the dependence on a large amount of manually labeled data, but also expands the training data through automatic evidence retrieval, significantly improves the generalization ability and adaptability of the model, reduces the data annotation cost, and improves the scalability of the relationship extraction task.

[0037] 2. The method for document-level relation extraction in the equipment knowledge domain guided by heterogeneous graph evidence constructs a weighted heterogeneous graph and assigns weights to different types of edges using the contribution degree discrimination rule, enabling entity nodes to pay more attention to sentences with high contribution to relation prediction, thus effectively removing the noise brought by irrelevant sentences. In addition, the weighted heterogeneous graph is feature-enhanced through a graph convolutional neural network, further improving the node's ability to integrate information of adjacent nodes and enhancing the model's reasoning ability for complex relations. These improvements significantly enhance the accuracy and robustness of document-level relation extraction, enabling it to better handle the long-text relation extraction task in the complex context of the equipment knowledge domain.

[0038] 3. The method for document-level relation extraction in the equipment knowledge domain guided by heterogeneous graph evidence guides the student model to focus on evidence sentences through the teacher model and dynamically adjusts the threshold of evidence retrieval using the improved ATLOP model to ensure that the model can accurately identify sentences closely related to relation prediction. In addition, the weighted heterogeneous graph is feature-enhanced through a graph convolutional neural network. The model can not only integrate the information of adjacent nodes but also further improve the performance of entities in relation prediction through various embedding methods, such as head and tail entity embeddings, evidence sentence information, etc. This optimized evidence retrieval and feature enhancement mechanism ensures that the model can more accurately capture the key information supporting relation prediction, thus significantly improving the overall performance of relation extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0040] Figure 1 is a flowchart of a method for document-level relation extraction in the equipment knowledge domain guided by heterogeneous graph evidence according to an embodiment of the present disclosure.

[0041] Figure 2 is a model architecture diagram of a method for document-level relation extraction in the equipment knowledge domain guided by heterogeneous graph evidence according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described below. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention fall within the protection scope of the present invention.

[0043] To better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0044] Example 1:

[0045] The embodiment of the present invention discloses a method for document-level relation extraction in the field of equipment knowledge guided by heterogeneous graph evidence. Figure 1 It is a flowchart of a method for document-level relation extraction in the field of equipment knowledge guided by heterogeneous graph evidence according to an embodiment of the present disclosure.

[0046] The present disclosure provides a method for document-level relation extraction in the field of equipment knowledge guided by heterogeneous graph evidence, which includes: encoding and embedding the document through a BERT pre-trained language model to construct an unweighted heterogeneous graph containing mention nodes, entity nodes, and sentence nodes; using a teacher model constructed based on ATLOP trained on manually annotated data to automatically retrieve silver evidence from remotely supervised data, training a student model on the remotely supervised data, using the silver evidence to supervise the automatic evidence retrieval to obtain effective evidence data; using a contribution degree discrimination rule to construct the unweighted heterogeneous graph into a weighted heterogeneous graph based on the effective evidence data; using a graph convolutional neural network for feature enhancement so that the nodes contain information of weighted adjacent nodes; constructing entity pair embeddings based on multiple embeddings and making predictions based on an entity pair classifier.

[0047] Figure 2 It is a model architecture diagram of a method for document-level relation extraction in the field of equipment knowledge guided by heterogeneous graph evidence according to an embodiment of the present disclosure.

[0048] In the present disclosure, the student model trained by the improved ATLOP is used to obtain effective evidence information, and the contribution degrees of different edge types of the unweighted heterogeneous graph are distinguished using the evidence information, improving the performance of document-level relation extraction. First, after encoding and embedding the text, an unweighted heterogeneous graph is constructed based on the association relationships among mentions, entities, and sentences; then, the improved ATLOP teacher model is trained using manually annotated data and applied to remotely supervised data to obtain silver evidence. The silver evidence is used to supervise the student model also trained on remotely supervised data for automatic evidence retrieval, and the student model is fine-tuned using manually annotated data to improve the adaptability of the model. Then, the extracted effective evidence information is used to assign specific contribution degrees to different edge types in the unweighted heterogeneous graph according to the contribution degree discrimination rule, so that the nodes pay more attention to the sentence information with high contribution degrees. And a graph convolutional neural network is performed to aggregate the neighbor node information with different weights. Finally, entity pair embeddings are constructed based on multiple embeddings including evidence sentence embeddings, and predictions are made based on an entity pair classifier.

[0049] Improve the accuracy and robustness of document-level relation extraction through relation reasoning among mentioned nodes, entity nodes, and sentence nodes in a heterogeneous graph without weights, combined with the text features generated by BERT.

[0050] Train a teacher model on manually annotated data and apply it to remotely supervised data to extract silver evidence. Use the remotely supervised data to train a student model, and perform automatic evidence retrieval under the supervision of silver evidence to obtain effective evidence data to improve the contribution degree differentiation of the heterogeneous graph with weights.

[0051] The teacher model based on the improved ATLOP directly guides the model to focus on evidence, supervises the calculation of the local context embedding of specific entity pairs. The local context embedding is generated by weighted summation of all token embeddings based on the attention mechanism of the BERT encoder, and its training objective is to assign different high and low weights to evidence.

[0052] The supervision of the student model consists of the binary cross-entropy loss of document-level relation extraction and the self-training loss of automatic evidence retrieval. Use the KL divergence loss to train the automatic evidence retrieval of the student model. The KL divergence loss formula is as follows:

[0053]

[0054] Prevent the loss of important instances by calculating the loss of all entity pairs, and balance the overall loss through hyperparameters. The hyperparameter formula is as follows:

[0055]

[0056] After training on remotely supervised data, further fine-tune the student model using manually annotated data, and the supervision signal improves its knowledge about document-level relation extraction and automatic evidence retrieval.

[0057] For automatic evidence retrieval, apply static thresholding, select sentences with importance higher than a predefined threshold as evidence, and obtain the evidence sentences for entity pairs (e h , e t )

[0058]

[0059] The contribution degree differentiation module assigns contribution weights to each edge through specific rules; the edge types are divided into the following four types:

[0060] First, M-E: Irrelevant to sentences and without self-loop edges;

[0061] Second, M-M: Irrelevant to sentences but with self-loop edges;

[0062] Third, M-S and E-S: Relevant to sentences and without self-loop edges;

[0063] Fourth, S-S: Related to the sentence and has a self-loop edge;

[0064] Introduce a self-loop edge to retain the self-owned information of the node, add an edge pointing to itself for each node, and avoid the loss of node information.

[0065] The entity pair classifier predicts the relationship based on the final entity pair embedding. The main components of the entity pair embedding are as follows:

[0066] The head entity embedding e h and the tail entity embedding e t : Capture the semantic expressions of the context information of the head and tail entities in the heterogeneous graph;

[0067] The absolute value of the difference between the head and tail entity vectors |e h -e t |: Quantify the feature differences between entities, and reflect the relationship strength and directionality existing between entities;

[0068] The element-wise product of the head and tail entity vectors e h ⊙e t : Capture the interaction patterns between the head and tail entity embeddings, and provide multi-dimensional semantic information for relationship modeling;

[0069] Evidence sentence information Give priority to the sentence information closely related to the entity pair relationship to ensure that the model can more accurately capture the key information supporting the relationship prediction.

[0070]

[0071] The probability function is as follows.

[0072] P(r|e h , e t ) = Sigmoid(W2σ(W1(e h , e t ) + b1) + b2).

[0073] The following uses specific examples to illustrate the method for document-level relationship extraction in the equipment knowledge field based on heterogeneous graph evidence guidance of the present disclosure. Due to the lack of a publicly annotated high-quality document-level dataset for the equipment field, a self-constructed document-level dataset for the equipment knowledge field is used, and the publicly available dataset DocRED is used to verify the generalization performance of the model. The original data of this dataset comes from foreign open-source military websites such as WIKI, AMERICANSPECIALOPS, and MilitaryFactory. The detailed statistical information of the dataset is shown in Table 1.

[0074] Table 1 Detailed statistical information of the DLDED dataset

[0075]

[0076]

[0077] The hyperparameter settings of the model are shown in Table 2.

[0078] Table 2 Experimental parameters

[0079] Parameter Name Parameter Value Embedding Dimension 512 Hidden Size of R-GCN 512 Epoch 30 Batch_size 4 Learning rate 2e-4 lr_decay 0.05 Optimizer AdamW Warmup ratio 0.08

[0080] The present invention compares the performance of each model based on the values of evaluation metrics. The evaluation metrics used are IgnF1, F1, and EviF1 values as the main evaluation metrics to test the performance and effectiveness of the document-level relation extraction model on two datasets. Among them, F1 and EviF1 are used as auxiliary evaluation metrics to measure the comprehensive performance of relation prediction and evidence extraction respectively, while the calculation of IgnF1 takes into account the overall consistency performance of the model for both relation prediction and evidence extraction tasks.

[0081] To evaluate the proposed model, the present invention selects some competitive entity alignment models for comparison. They mainly include: SSAN, ATLOP, E2GRE, DocuNet, EIDER, and SAIA. The following is a brief introduction to these models:

[0082] SSAN is a model based on the multi-head attention mechanism, which uses shared subspace representation to model relation extraction. By comprehensively representing the context information of entity pairs, this model improves the ability to capture long-distance dependency relationships and performs well in multiple document-level relation extraction tasks.

[0083] ATLOP adopts an adaptive threshold mechanism, which improves the accuracy of relation prediction by dynamically adjusting each entity pair. The model also introduces local context embedding and global context embedding to enhance the ability to identify entity relationships in complex documents.

[0084] E2GRE is a model focusing on global relation modeling of entity pairs. By explicitly introducing the contribution of global information and multi-granularity context to relations, it combines the document-level relation extraction task with the evidence sentence selection task, significantly improving the model's ability to handle complex relations.

[0085] DocuNet uses graph neural networks to model sentences and entities in a document and captures the complex interaction relationships between entities and sentences by constructing a document-level heterogeneous graph. This model shows excellent performance in relation extraction tasks, especially in dealing with cross-sentence relations.

[0086] EIDER introduces an attention-based multi-layer nested structure, which captures the potential dependencies between entities by hierarchically modeling the relational context information in the document. The model shows high robustness on multiple public datasets.

[0087] SAIA adopts an adaptive interactive attention mechanism, which enhances the model's predictive ability for complex relationships in long documents by dynamically learning the dependencies between entity pairs. Its innovation lies in the refined modeling of information interaction at different granularities.

[0088] Experimental results show that the performance of the model of the present invention is almost better than all comparison methods on the self-built DLDED dataset and the DocRED dataset.

[0089] Table 3 Performance comparison of the models

[0090]

[0091] It can be seen that the performance of the model of the present invention is better than that of other models. This is because the model obtains more effective evidence information through the student model constructed based on ATLOP, effectively solving the problem of insufficient manually annotated evidence data. And based on the obtained effective evidence information, different types of edges on the unweighted heterogeneous graph are assigned corresponding contribution degrees according to the contribution degree discrimination rule, enabling the final entity pair classification to focus on a small number of sentences, removing the sentence noise problem, and thus improving the overall performance of the model.

[0092] It should be noted that in the present invention, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the process, method, article or device comprising the element.

[0093] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A document-level relationship extraction method for equipment knowledge domain based on heterogeneous graph evidence, characterized in that: This method adopts a heterogeneous graph evidence-guided strategy, and the specific steps include: Step S1, encoding and embedding the document through the BERT pre-trained language model to construct an unweighted heterogeneous graph containing mention nodes, entity nodes, and sentence nodes; Step S2, using the teacher model based on ATLOP trained on the manually annotated data, automatically retrieve the silver evidence from the remote supervision data, and train the student model on the remote supervision data to obtain the valid evidence data; Step S3, using the contribution differentiation rule to construct the unweighted heterogeneous graph into a weighted heterogeneous graph based on the valid evidence data; Step S4, using graph convolutional neural network to perform feature enhancement so that the node contains the authorized adjacent node information; In step S5, an entity pair embedding is constructed based on multiple embeddings, and prediction is performed based on the entity pair classifier.

2. According to claim 1, the method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance is characterized in that: By reasoning about the relationships among mention nodes, entity nodes, and sentence nodes in unweighted heterogeneous graphs and combining them with text features generated by BERT, the accuracy and robustness of document-level relationship extraction are improved.

3. The method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance according to claim 1 is characterized in that: Based on the improved ATLOP model, the teacher model is trained on manually annotated data and applied to remote supervision data to extract silver evidence. The student model is trained using remote supervision data, and automatic evidence retrieval is performed under the supervision of silver evidence to obtain effective evidence data to improve the contribution distinction of weighted heterogeneous graphs.

4. The method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance according to claim 3 is characterized in that: The teacher model based on the improved ATLOP directly guides the model to pay attention to the evidence and supervises the entity to calculate the specific local context embedding. The local context embedding is based on the attention mechanism of the BERT encoder and is generated by weighted summation of all tag embeddings. Its training goal is to give different weights to the evidence.

5. The method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance according to claim 3 is characterized in that: The supervision of the student model consists of the binary cross entropy loss for document-level relation extraction and the self-training loss for automatic evidence retrieval. The KL divergence loss is used to train the automatic evidence retrieval of the student model. The KL divergence loss formula is as follows:

6. The method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance according to claim 5 is characterized in that: The loss of important instances is prevented by calculating the loss of all entity pairs, and the overall loss is balanced by the hyperparameters. The hyperparameter formula is as follows:

7. The method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance according to claim 6 is characterized in that: After training on distant supervision data, the student model is further fine-tuned using human annotated data, and the supervisory signal refines its knowledge about document-level relation extraction and automatic evidence retrieval.

8. The method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance according to claim 7 is characterized in that: For automatic evidence retrieval, static thresholding is applied to select sentences with importance higher than a predefined threshold as evidence to obtain entity pairs (e h ,e t ) Evidence Sentence 9. The method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance according to claim 2 is characterized in that: The contribution differentiation module assigns contribution weights to each edge through specific rules, and divides the edge types into the following four types: First, ME: It is independent of the sentence and has no self-loop edges; Second, MM: irrelevant to the sentence but has self-loop edges; Third, MS and ES: related to the sentence and without self-loop edges; Fourth, SS: related to the sentence and has self-loop edges; Self-loop edges are introduced to retain the node's own information, and an edge pointing to each node is added to avoid node information loss.

10. The method for extracting document-level relations in equipment knowledge domain based on heterogeneous graph evidence guidance according to claim 2 is characterized in that: The entity pair classifier predicts the relationship based on the final entity pair embedding. The main components of the entity pair embedding are as follows: Header entity embedding h and tail entity embedding t : Capture the semantic expression of contextual information of head and tail entities in heterogeneous graphs; The absolute value of the difference between the head and tail entity vectors |e h -e t |: Quantify the feature differences between entities and reflect the strength and direction of the relationship between entities; Element-wise product of head and tail entity vectors: captures the interaction pattern between head and tail entity embeddings, providing multi-dimensional semantic information for relationship modeling; Evidence sentence information Prioritize sentence information that is closely related to entity pairs, ensuring that the model can more accurately capture key information that supports relationship prediction; The probability function is as follows: P(r|e h ,and t )=Sigmoid(W2σ(W1(e h ,and t )+b1)+b2)。

Images