Knowledge distillation-based human brain glioma gene entity relationship prediction method and equipment

Abstract

The invention discloses a knowledge distillation-based human brain glioma gene entity relationship prediction method and equipment, and the method comprises the specific steps: SA01, training a high-precision teacher model through an artificial data set, carrying out the prediction on a remote supervision data set through the teacher model, and generating a soft label; sA02, performing preliminary learning in combination with a soft label and a student model, and constructing a bank evidence set by adopting a bank evidence set construction step; sA03, finely selecting evidence sentences in the bank evidence set by utilizing a reinforcement learning technology; sA04, combining the attention obtained from the selected evidence sentence with the attention of the soft label to create a new and enhanced attention mechanism, and then combining the attention mechanism with the co-occurrence relation characteristics obtained in the co-occurrence relation characteristic construction step to predict the final relation between the entities; sA05, the self-trained student model is subjected to fine adjustment through manual annotation data, evidence sentence information can be fully utilized, the relation between target relations can be considered, and therefore more accurate prediction can be made.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method and device for predicting human brain glioma gene entity relationships based on knowledge distillation. Background Art

[0002] Gliomas are highly malignant intracranial tumors, and their treatment has always faced great challenges and difficulties, and the prognosis of patients is often poor. The occurrence, development, and resistance to radiotherapy and chemotherapy of gliomas are closely related to the complex gene regulatory network of tumors. In addition, given the complexity of glioma treatment and the diversity of gene regulatory networks, the development of efficient and accurate gene relationship prediction models is of great significance for guiding medical research and drug development. With the rapid development of high-throughput sequencing technology, researchers can obtain massive amounts of bioinformatics data, and then analyze the interaction relationship between genes through systems biology methods, and publish a large number of academic papers. Therefore, considering the use of natural language processing technology to mine document relationships has become a research hotspot in this field.

[0003] At present, the contextual information contained in documents is becoming more and more rich and complex, resulting in an increasing diversity of relational facts. The relationships between entities are often no longer single, but multiple. Compared with the traditional single-relation extraction problem, multi-relation extraction can provide richer and deeper knowledge, especially in knowledge graph construction and text analysis.

[0004] Traditional document-level relation extraction methods usually predict the relationship between entities by semantically modeling the global document, but this global modeling method makes it difficult for the model to capture local information. In addition, this method is easily disturbed by sentences that are irrelevant to the target relationship, thereby reducing the accuracy of relation extraction. To address this problem, recent document-level relation extraction methods have introduced the concept of evidence sentences, which are the most important key information related to the target relationship. Unlike traditional global document modeling, these methods focus more on the evidence sentences in the document. However, in actual documents, these sentences are often scattered throughout the document, rather than often appearing in a few consecutive sentences. Therefore, how to effectively select evidence sentences is also a key issue facing current document-level relation extraction methods.

[0005] Furthermore, when using evidence sentences to solve multiple relation extraction problems, document relation extraction methods mainly rely on the semantic information of evidence sentences, while ignoring the connections between target relations. This is because, especially when the target relations are vague and ambiguous, the semantic information of evidence sentences may not be able to identify the subtle differences between certain relations. Therefore, an important problem faced by most relation extraction methods is that they may only focus on predicting some major relations while ignoring other important relations, resulting in incomplete relation predictions. Especially in practical applications, such as knowledge graphs, text analysis, etc., incomplete relations will not provide comprehensive information, which will greatly reduce task performance. Therefore, we believe that there is an urgent need for a new extraction method that can fully utilize the information of evidence sentences and consider the connections between target relations to make more accurate predictions. Summary of the invention

[0006] The technical problem solved by the present invention is to provide a human brain glioma gene entity relationship prediction method based on knowledge distillation, which can fully utilize evidence sentence information and consider the connection between target relationships, thereby making more accurate predictions.

[0007] The technical solution adopted by the present invention to solve the technical problem is: a method for predicting the entity relationship of human brain glioma genes based on knowledge distillation, characterized in that it includes the following steps:

[0008] Data preprocessing step: Process, extract and encode a large amount of text data to obtain the relationship between entity pairs;

[0009] Silver evidence set construction step: construct the silver evidence set through the heuristic rules of continuous path, coreference path and bridge path respectively. Evidence sentence selection step: identify and select the evidence sentences that are most critical to the entity relationship judgment;

[0010] Steps for constructing co-occurrence relationship features: Use the SOTA model to obtain the co-occurrence frequency of each relationship in different document-level relationship data sets, and obtain the co-occurrence relationship features;

[0011] Knowledge distillation prediction step: It includes a teacher model trained with a manually annotated dataset and a student model that is self-trained with soft labels and fine-tuned with a manually annotated dataset. The specific steps are:

[0012] SA01: Train a high-precision teacher model with an artificial dataset, use the teacher model to make predictions on a remote supervision dataset, and generate and save soft labels for the remote supervision dataset;

[0013] SA02: Combining the distant supervision dataset and soft labels, the student model conducts preliminary learning and builds a silver evidence set;

[0014] SA03: Use reinforcement learning technology to finely select evidence sentences in the silver evidence set;

[0015] SA04: The attention obtained from the selected evidence sentences is combined with the attention of the soft labels to create a new, enhanced attention mechanism, which is then combined with the co-occurrence relationship features obtained in the co-occurrence relationship feature construction step to predict the final relationship between entities;

[0016] SA05: Fine-tune the self-trained student model using manually annotated data.

[0017] Furthermore, in the data preprocessing step, a large amount of text data is processed, extracted and encoded to obtain the relationship between entity pairs, specifically:

[0018] S110: Embedding extraction of entity mentions: For each entity e in the document s , collect the embeddings of all its textual mentions m These embeddings are obtained from the start position tokens of entity mentions and are represented as The m represents the synonyms mentioned in the text;

[0019] S120: Embedded information encoding: All mentioned embeddings are integrated using the Log-Sum-Exp technique to form a comprehensive entity embedding. The specific formula is as follows:

[0020]

[0021] S130: Calculation of attention mechanism between entity pairs: For the head entity e h and tail entity e t , calculate the multi-head attention vector q between them (h，t) , which is realized by Hadamard product and normalization operation. The specific formula is:

[0022]

[0023] S140: Using context aggregation technology, the system aggregates the importance of n sentences in document D respectively. The specific formula is:

[0024] c (h，t) =H T q (h，t) ;

[0025] Where H represents the global context embedding of the document, c (h，t) It is the weighted average of all token embeddings, reflecting the contextual relationship of entity pairs in the document;

[0026] S150: Sentence-level correlation analysis: Calculate the contribution of each sentence to the entity pair relationship. The specific formula is:

[0027] Among them, q (h，t) is the multi-head attention vector, w n Represents the beginning of each sentence, w n′ Represents the end of each sentence.

[0028] Furthermore, the silver evidence set construction step is to construct the silver evidence set for the continuous path, the co-referential path and the bridge path respectively, specifically:

[0029] S210: Continuous paths: When the head entity and the tail entity can be found in consecutive sentences, the length of these continuous paths is limited to 3 sentences, that is, |ji≤2|, where i represents the first sentence selected in the text and j represents the last sentence selected in the text;

[0030] S220: Co-reference path: When the same entity has different mention representations, for well-defined co-reference mentions, sentences where different mentions of the same entity pair appear simultaneously are used as evidence sentences. For some datasets that do not have co-reference partitioning, the pre-trained Hoi model is used for partitioning.

[0031] S230: Bridging path: when there is an indirect association between entities, the intermediate entity that plays the role of association is set as a bridging entity, the maximum number of bridging entities is set to two, and the maximum number of bridging sentences is set to three;

[0032] S240: Other paths: When the sentence where the head entity and the tail entity first appear is neither in the three consecutive sentences nor meets the conditions of the coreference path and the bridging path, the sentence containing the first appearance of the head entity and the tail entity will be used as the evidence sentence.

[0033] Furthermore, in the evidence sentence selection step, the most critical evidence sentences for the entity relationship judgment are identified and selected, specifically:

[0034] S310: Integrate the embedding vectors of the head entity and the tail entity, and introduce the contextual dependency and sentence information between entity pairs to define the system state;

[0035] S320: using the Monte Carlo method to calculate the probability of each sentence being an evidence sentence according to a predetermined decision strategy, and based on the probabilities and the silver evidence set constructed in the silver evidence set construction step, selecting the sentence with the highest probability in the silver evidence set;

[0036] S330: Using the REINFORCE algorithm to optimize the decision strategy in step S320.

[0037] Furthermore, in the step of constructing the co-occurrence relationship feature, the SOTA model is used to obtain the co-occurrence frequency of each relationship in different document-level relationship data sets, and the co-occurrence relationship feature is obtained, specifically:

[0038] Using prior knowledge and the SOTA model in the field of document relationship extraction, the co-occurrence frequency of each relationship is obtained in different document-level relationship data sets. The co-occurrence matrix P is built based on the co-occurrence frequency, and the threshold τ is used to filter the noise edge. The specific formula is:

[0039]

[0040] Where Q∈R n*n It is a co-occurrence matrix after binary processing, R represents the relationship, n represents the number of relationships, i and j both represent an entity. In the process of constructing graph nodes through the matrix, re-weighting is used to update the node features in GCN:

[0041]

[0042] When the parameter p is close to 1, the weight of the node's own features is small, and the relationship with the features of other related nodes is emphasized. When p is close to 0, the weight of the node's own features is large, emphasizing the node's own features. Finally, different GAT networks are used to transform and integrate the relationship features to obtain a new co-occurrence relationship feature G, which is specifically expressed as follows:

[0043] F rel =f1,f2...f n ,n=97;

[0044] H rel =GAT(F rel ,Q' ij );

[0045] G=H rel *W;

[0046] Among them, Frel represents a relationship matrix after all relationships are summarized, and n=97 is because there are 97 types of relationships in training.

[0047] Among them, f i Represents the embedding vector of the i-th relation, similar to word embedding. W is a weight matrix used to transform the aggregated features. H rel represents the relational feature representation after aggregation, r represents the number of relations, d Q Dimensions that characterize each relationship.

[0048] The present invention also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the human brain glioma gene entity relationship prediction method based on knowledge distillation are implemented.

[0049] The present invention also discloses a computer device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; wherein:

[0050] The memory is used to store computer programs;

[0051] The processor is used to execute the steps of the above-mentioned method for predicting human brain glioma gene entity relationships based on knowledge distillation by running the program stored in the memory.

[0052] The beneficial effect of the present invention is that the method can reduce the model size and improve the computing efficiency while ensuring the reasoning effect by transferring the knowledge of the large model to the small model. Through the entity multi-relationship prediction method based on knowledge distillation, it is possible to make full use of the existing bioinformatics data and gene regulatory network knowledge, and accurately predict the regulatory relationship between glioma-related gene entities by constructing an efficient prediction model. This not only helps to deeply understand the pathogenesis of glioma, but also provides a scientific basis for the development of therapeutic drugs for specific gene targets. In the future, with the continuous advancement of technology and the continuous accumulation of data, the entity multi-relationship prediction method based on knowledge distillation will show broad application prospects in the field of gene therapy and precision medicine for glioma. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Schematic diagram of the structure of the human glioma gene entity relationship prediction model based on knowledge distillation.

[0054] Figure 2 Schematic diagram of model structure construction for the silver evidence set of human glioma gene entity relationship prediction model based on knowledge distillation.

[0055] Figure 3 Schematic diagram of the model structure for constructing the co-occurrence relationship features of the human glioma gene entity relationship prediction model based on knowledge distillation. DETAILED DESCRIPTION

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0057] like Figure 1As shown, the embodiment of the present application discloses a method for predicting human brain glioma gene entity relationships based on knowledge distillation, which is characterized by comprising the following steps:

[0058] Data preprocessing step: Process, extract and encode a large amount of text data to obtain the relationship between entity pairs;

[0059] Silver evidence set construction steps: construct the silver evidence set through the heuristic rules of continuous path, coreference path and bridge path respectively;

[0060] Evidence sentence selection step: Identify and select the evidence sentences that are most critical to the entity relationship judgment;

[0061] Steps for constructing co-occurrence relationship features: Use the SOTA model to obtain the co-occurrence frequency of each relationship in different document-level relationship data sets, and obtain the co-occurrence relationship features;

[0062] Knowledge distillation prediction step: It includes a teacher model trained with a manually annotated dataset and a student model that is self-trained with soft labels and fine-tuned with a manually annotated dataset. The specific steps are:

[0063] SA01: Train a high-precision teacher model with an artificial dataset, use the teacher model to make predictions on a remote supervision dataset, and generate and save soft labels for the remote supervision dataset;

[0064] SA02: Combined with the distant supervision dataset and soft labels, the student model performs preliminary learning and uses the silver evidence set construction step to build a silver evidence set;

[0065] SA03: Use reinforcement learning technology to finely select evidence sentences in the silver evidence set;

[0066] SA04: The attention obtained from the selected evidence sentences is combined with the attention of the soft labels to create a new, enhanced attention mechanism, which is then combined with the co-occurrence relationship features obtained in the co-occurrence relationship feature construction step to predict the final relationship between entities;

[0067] SA05: Fine-tune the self-trained student model using manually annotated data.

[0068] In this method, first, we use a carefully annotated artificial dataset to train a high-precision teacher model and use its predicted output as soft labels to guide the self-training of the student model. Then, we combine the remote supervision data and these soft labels to perform preliminary learning on the student model. In addition, we use heuristic rules to construct silver evidence sets and use reinforcement learning techniques to perform fine-grained selection and screening of silver evidence sets in the remote supervision dataset. In both the teacher model and the student model, we integrate co-occurrence relationship features to build models to enhance the model's prediction ability for different entity relationships.

[0069] As the final stage of model training, we fine-tune the self-trained student model using manually annotated data. This is to fully utilize the knowledge embedded in the teacher model, improve the student model's ability to handle complex data, and address potential differences caused by the diversity of data sources.

[0070] Specifically, in the data preprocessing step, a large amount of text data is processed, extracted and encoded to obtain the relationship between entity pairs, specifically:

[0071] S110: Embedding extraction of entity mentions: For each entity e in the document s , collect the embeddings of all its textual mentions m These embeddings are derived from the start position marker of the entity mention (e.g., "*") and are represented as The m represents the synonyms mentioned in the text, that is, different words but referring to the same thing;

[0072] S120: Embedded information encoding: All mentioned embeddings are integrated using the Log-Sum-Exp technique to form a comprehensive entity embedding. The specific formula is as follows:

[0073]

[0074] S130: Calculation of attention mechanism between entity pairs: For the head entity e h and tail entity e t , calculate the multi-head attention vector q between them (h，t) , which is realized by Hadamard product and normalization operation. The specific formula is:

[0075]

[0076] It should be explained that the above “.” refers to the Hadamard product operation.

[0077] S140: Using context aggregation technology, the system aggregates the importance of n sentences in document D respectively. The specific formula is:

[0078] c (h,t) =H T q (h，t) ;

[0079] Where H represents the global context embedding of the document, c (h，t) It is the weighted average of all token embeddings, reflecting the contextual relationship of entity pairs in the document;

[0080] S150: Sentence-level correlation analysis: Calculate the contribution of each sentence to the entity pair relationship. The specific formula is:

[0081] Among them, q (h，t) is the multi-head attention vector, w n Represents the beginning of each sentence, w n′ Represents the end of each sentence.

[0082] This formula represents the calculation of the importance of each sentence in the text

[0083] The advantage of this preprocessing method is that it can automatically process large amounts of text data, accurately extract and encode the relationships between entities, and provide high-quality input for subsequent entity relationship extraction tasks. This method improves the efficiency and accuracy of entity relationship recognition tasks and is a technological innovation of great value in the field of natural language processing.

[0084] In this embodiment, the silver evidence set construction step is: constructing the silver evidence set for the continuous path, the co-referencing path and the bridging path respectively, such as Figure 2 As shown, specifically,

[0085] S210: Continuous paths: When the head entity and the tail entity can be found in consecutive sentences, the length of these continuous paths is limited to 3 sentences, that is, |ji≤2|, where i represents the first sentence selected in the text and j represents the last sentence selected in the text;

[0086] The above steps are specifically explained as follows: When people write documents, they tend to subconsciously arrange related parts together continuously, so most inter-sentence relations can be found in consecutive sentences. Based on this observation, we set an assumption that no more than 3 consecutive sentences are sufficient to support the judgment of the relationship when making inferences. Therefore, we limit the length of these continuous paths to at most 3 sentences, that is, |ji≤2|. It should be noted that this definition can also naturally apply to the case where the head entity and the tail entity are in the same sentence, so we regard this case as a special type of continuous path. There may be multiple continuous paths between the same pair of entities because they can be mentioned multiple times in the text.

[0087] S220: Co-reference path: When the same entity has different mention representations, for well-defined co-reference mentions, sentences where different mentions of the same entity pair appear simultaneously are used as evidence sentences. For some datasets that do not have co-reference partitioning, the pre-trained Hoi model is used for partitioning.

[0088] S230: Bridging path: when there is an indirect association between entities, the intermediate entity that plays the role of association is set as a bridging entity, the maximum number of bridging entities is set to two, and the maximum number of bridging sentences is set to three;

[0089] For example, to express the relationship between entity A and entity C, it is necessary to present it through sentences between entity A and entity B, and sentences between entity B and entity C. In this case, we call entity B a bridge entity. Usually, we limit the maximum number of bridge entities to two, so the maximum number of bridge sentences is three, that is, (A, B), (B, C), (C, D).

[0090] S240: Other paths: When the sentence where the head entity and the tail entity first appear is neither in the three consecutive sentences nor meets the conditions of the coreference path and the bridging path, the sentence containing the first appearance of the head entity and the tail entity will be used as the evidence sentence.

[0091] In this embodiment, in the evidence sentence selection step, the most critical evidence sentences for entity relationship judgment are identified and selected, specifically:

[0092] S310: Integrate the embedding vectors of the head entity and the tail entity, and introduce the contextual dependency and sentence information between entity pairs to define the system state. The design of this state ensures that each selected evidence sentence will affect the selection of subsequent sentences.

[0093] S320: Using the Monte Carlo method to calculate the probability of each sentence being an evidence sentence according to a predetermined decision strategy, based on these probabilities and the silver evidence set constructed in the silver evidence set construction step, select the sentence with the highest probability in the silver evidence set. In this process, the system does not immediately obtain a reward, but generates a reward based on the effect of the selected evidence sentence at the end of the processing cycle of each entity pair. ;

[0094] S330: The decision strategy in step S320 is optimized using the REINFORCE algorithm. The strategy is defined by the Q values ​​output by the neural network, and these Q values ​​are adjusted by the policy gradient method to maximize the expected cumulative reward. The parameters included in the strategy are updated according to the cumulative reward through each iteration, so as to continuously optimize the strategy to improve the accuracy of entity relationship inference.

[0095] Through this method, the system can efficiently identify and select the most critical evidence sentences for entity relationship judgment, supporting in-depth analysis and information extraction in complex text data. The development of this technology provides a new tool in the field of natural language processing, especially for application scenarios that require automatic extraction and verification of information from large amounts of text.

[0096] In this embodiment, Figure 3 As shown, in the co-occurrence relationship feature construction step, the SOTA model is used to obtain the co-occurrence frequency of each relationship in different document-level relationship data sets, and the co-occurrence relationship features are obtained, specifically:

[0097] Using prior knowledge and the SOTA model in the field of document relationship extraction, the co-occurrence frequency of each relationship is obtained in different document-level relationship data sets. The co-occurrence matrix P is built based on the co-occurrence frequency, and the threshold τ is used to filter the noise edge. The specific formula is:

[0098]

[0099] Where Q∈R n*n It is a co-occurrence matrix after binary processing. R represents the relationship, n represents the number of relationships, and i and j both represent an entity. In the process of constructing graph nodes through the matrix, reweighting is used to update the node features in GCN:

[0100]

[0101] It should be explained that GCN stands for Graph Convolutional Networks (GCN), which is a common term in deep learning, graph convolutional network.

[0102] When the parameter p is close to 1 (where p is a trainable parameter that is constantly updated during the training process), the weight of the node's own features is small and almost not considered, and the relationship with the features of other related nodes is emphasized. When p is close to 0, the weight of the node's own features is large, emphasizing the node's own features and ignoring the information of neighboring nodes. Finally, different GAT networks are used to transform and integrate the relationship features to obtain a new representation G:, which is specifically expressed as follows:

[0103] F rel =f1,f2...f n ,n=97;

[0104] H rel =GAT(F rel ,Q i ' j );

[0105] G=H rel *W;

[0106] Among them, Frel represents a relationship matrix after all relationships are summarized, and n=97 is because there are 97 types of relationships in training.

[0107] Among them, f i Represents the embedding vector of the i-th relation, similar to word embedding. W is a weight matrix used to transform the aggregated features. H rel represents the relational feature representation after aggregation, r represents the number of relations, d Q Dimensions that characterize each relationship.

[0108] The present invention also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the human brain glioma gene entity relationship prediction method based on knowledge distillation are implemented.

[0109] In addition, the computer-readable storage medium of this embodiment may adopt any combination of one or more computer-readable storage media, wherein the computer-readable storage medium includes electrical, optical, electromagnetic, infrared or semiconductor systems, devices or components, or any combination thereof.

[0110] The present invention also discloses a computer device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; wherein:

[0111] The memory is used to store computer programs;

[0112] The processor is used to execute the steps of the above-mentioned method for predicting human brain glioma gene entity relationships based on knowledge distillation by running the program stored in the memory.

[0113] As an embodiment of the present invention, the communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0114] As an implementation mode of the present invention, the communication interface is used for communication between the above-mentioned terminal and other devices.

[0115] As an embodiment of the present invention, the memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0116] As an embodiment of the present invention, the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0117] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting entity relationships of human glioma genes based on knowledge distillation, characterized in that: The steps include: Data preprocessing step: Process, extract and encode a large amount of text data to obtain the relationship between entity pairs; Silver evidence set construction steps: construct the silver evidence set through the heuristic rules of continuous path, coreference path and bridge path respectively; Evidence sentence selection step: Identify and select the evidence sentences that are most critical to the entity relationship judgment; Steps for constructing co-occurrence relationship features: Use the SOTA model to obtain the co-occurrence frequency of each relationship in different document-level relationship data sets, and obtain the co-occurrence relationship features; Knowledge distillation prediction step: It includes a teacher model trained with a manually annotated dataset and a student model that is self-trained with soft labels and fine-tuned with a manually annotated dataset. The specific steps are: SA01: Train a high-precision teacher model with an artificial dataset, use the teacher model to make predictions on a remote supervision dataset, and generate and save soft labels for the remote supervision dataset; SA02: Combined with the remote supervision dataset and soft labels, the student model conducts preliminary learning and builds a silver evidence set; SA03: Use reinforcement learning technology to finely select evidence sentences in the silver evidence set; SA04: The attention obtained from the selected evidence sentences is combined with the attention of the soft labels to create a new, enhanced attention mechanism, which is then combined with the co-occurrence relationship features obtained in the co-occurrence relationship feature construction step to predict the final relationship between entities; SA05: Fine-tune the self-trained student model using manually annotated data.

2. The method for predicting entity relationships of human brain glioma genes based on knowledge distillation according to claim 1, characterized in that: In the data preprocessing step, a large amount of text data is processed, extracted and encoded to obtain the relationship between entity pairs, specifically: S110: Embedding extraction of entity mentions: For each entity e in the document s , collect the embeddings of all its textual mentions m These embeddings are obtained from the start position tokens of entity mentions and are represented as The m represents the synonyms mentioned in the text; S120: Embedded information encoding: All mentioned embeddings are integrated using the Log-Sum-Exp technique to form a comprehensive entity embedding. The specific formula is as follows: S130: Calculation of attention mechanism between entity pairs: For the head entity e h and tail entity e t , calculate the multi-head attention vector q between them (h，t) , which is realized by Hadamard product and normalization operation. The specific formula is: S140: Using context aggregation technology, the system aggregates the importance of n sentences in document D respectively. The specific formula is: c (h，t) =H T q (h，t) ; Where H represents the global context embedding of the document, c (h，t) It is the weighted average of all token embeddings, reflecting the contextual relationship between entity pairs in the document. S150: Sentence-level correlation analysis: Calculate the contribution of each sentence to the entity pair relationship. The specific formula is: Among them, q (h，t) is the multi-head attention vector, w n Represents the beginning of each sentence, w n′ Represents the end of each sentence.

3. The method for predicting human brain glioma gene entity relationships based on knowledge distillation according to claim 1, characterized in that: The silver evidence set construction step: constructing the silver evidence set for the continuous path, the coreference path and the bridge path respectively, specifically: S210: Continuous paths: When the head entity and the tail entity can be found in consecutive sentences, the length of these continuous paths is limited to 3 sentences, that is, |Ji≤2|, where i represents the first sentence selected in the text and j represents the last sentence selected in the text; S220: Co-reference path: When the same entity has different mention representations, for well-defined co-reference mentions, sentences where different mentions of the same entity pair appear simultaneously are used as evidence sentences. For some datasets that do not have co-reference partitioning, the pre-trained Hoi model is used for partitioning. S230: Bridging path: when there is an indirect association between entities, the intermediate entity that plays the role of association is set as a bridging entity, the maximum number of bridging entities is set to two, and the maximum number of bridging sentences is set to three; S240: Other paths: When the sentence where the head entity and the tail entity first appear is neither in the three consecutive sentences nor meets the conditions of the coreference path and the bridging path, the sentence containing the first appearance of the head entity and the tail entity will be taken as the evidence sentence.

4. The method for predicting entity relationships of human brain glioma genes based on knowledge distillation according to claim 1, characterized in that: In the evidence sentence selection step, identify and select the most critical evidence sentences for entity relationship judgment, specifically: S310: Integrate the embedding vectors of the head entity and the tail entity, and introduce the contextual dependency and sentence information between entity pairs to define the system state; S320: using the Monte Carlo method to calculate the probability of each sentence being an evidence sentence according to a predetermined decision strategy, and based on these probabilities and the silver evidence set constructed in the silver evidence set construction step, selecting the sentence with the highest probability in the silver evidence set; S330: Using the REINFORCE algorithm to optimize the decision strategy in step S320.

5. The method for predicting entity relationships of human brain glioma genes based on knowledge distillation according to claim 1, characterized in that: In the co-occurrence relationship feature construction step, the SOTA model is used to obtain the co-occurrence frequency of each relationship in different document-level relationship data sets, and the co-occurrence relationship features are obtained, specifically: Using prior knowledge and the SOTA model in the field of document relationship extraction, we obtain the co-occurrence frequency of each relationship in different document-level relationship data sets, build the co-occurrence matrix P through the co-occurrence frequency, and use the threshold T to filter the noise edge. The specific formula is: Where Q∈R n*n It is a co-occurrence matrix after binary processing. R represents the relationship, n represents the number of relationships, and i and j both represent an entity. In the process of constructing graph nodes through the matrix, reweighting is used to update the node features in GCN: When the parameter p is close to 1, the weight of the node's own features is small, and the relationship with the features of other related nodes is emphasized. When p is close to 0, the weight of the node's own features is large, emphasizing the node's own features. Finally, different GAT networks are used to transform and integrate the relationship features to obtain a new co-occurrence relationship feature G, which is specifically expressed as follows: F rel =f1,f2…f n ,n=97#F rel =f1,f2…f n ,n=97; H rel =GAT(F rel ,Q′ ij ); G=H rel *W; Among them, Frel represents a relationship matrix after all relationships are summarized, and n=97 is because there are 97 types of relationships in training. Among them, f i Represents the embedding vector of the i-th relation, similar to word embedding. W is a weight matrix used to transform the aggregated features. H rel represents the relational feature representation after aggregation, r represents the number of relations, d Q Dimensions that characterize each relationship.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for predicting human brain glioma gene entity relationships based on knowledge distillation according to any one of claims 1 to 5 are implemented.

7. A computer device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; wherein: The memory is used to store computer programs; The processor is used to execute the steps of the human brain glioma gene entity relationship prediction method based on knowledge distillation according to any one of claims 1 to 5 by running the program stored in the memory.

Images