Recurrent neural network multi-tree fusion method in biomedical event extraction
Through the multi-tree fusion method, the syntactic analysis trees generated by multiple parsers are heterogeneously fused, and combined with branch pruning and node distillation, the problem of missing dependencies in a single syntactic analysis tree in biomedical event extraction is solved, achieving more efficient and accurate event extraction.
Patent Information
- Application Number
- CN202510866654.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-26
- Publication Date
- 2025-10-14
AI Technical Summary
In existing technologies, a single syntactic parse tree cannot accurately parse complex biomedical events, resulting in missing or inaccurate dependency relationships. Especially when dealing with texts with complex structures and many professional terms, existing research multimodal methods have failed to effectively solve the problem of heterogeneous semantic fusion.
Multiple parsers are used to generate multiple syntactic analysis trees, which are then merged through the union merging method. Branch distillation and branch pruning are performed from the bottom up to form a multi-branch tree structure. Duplicate nodes are processed and the final merged multi-branch tree is output.
The accuracy and efficiency of biomedical event extraction have been improved, especially the semantic representation of complex events has been significantly improved, and the overall model performance has been improved.
Smart Images
Figure CN120781827A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of biomedical information extraction, and particularly relates to a recursive neural network multi-tree fusion method in biomedical event extraction. BACKGROUND
[0002] The research content of biomedical event extraction mainly includes trigger word recognition and event argument extraction. Among them, the trigger word is the core of the biomedical event, which is usually a verb or a noun phrase, and can clearly indicate the occurrence of a biomedical event. The event argument is the participant or attribute of the event, such as protein, gene, etc. The event type is classified according to different characteristics of the biomedical event, such as simple event, binding event and regulation event, etc.
[0003] The nodes in the recursive neural network recursively process the input information according to the connection order, and each node performs the same operation, that is, combines the information of its child nodes to generate its own representation. The recursive neural network generates the representation of the parent node by recursively combining the information of the child nodes, so as to realize the coding of the entire tree structure. In the training process, the recursive neural network calculates the output of each node through the forward propagation algorithm, and calculates the gradient using the back propagation algorithm to update the network parameters. Different built-in algorithms of different syntax parsers result in that a single syntax analysis tree cannot parse all dependency relationships. For example, the rule-based method may not be able to handle some sentence structures that do not conform to the rules, and the statistical or deep learning-based method may not be able to accurately identify all syntactic dependencies in the sentence, especially in the case of complex structure and difficult-to-understand professional terms in biomedical event extraction tasks, resulting in the lack of some edges in the analysis tree or the inaccurate labeling of the edges. The existing research usually adopts a multi-modal method for single modal. For the phenomenon of missing dependency relationships of single syntax parsing tree, other modal forms of structure cannot be found, but heterogeneous semantic fusion can be supplemented through semantic association of other structures. The existing research focuses more on the research on merging binary tree into forest structure, but less on the research on multi-ary tree structure syntax parsing tree. SUMMARY
[0004] The application provides a recursive neural network multi-tree fusion method in biomedical event extraction, to solve the technical problem of missing dependency relationships of traditional syntax parsing tree in the prior art.
[0005] In order to achieve the above purpose, the technical scheme adopted by the application is as follows: A recursive neural network multi-tree fusion method in biomedical event extraction, comprising the following steps: Generating multiple syntax analysis trees from text content by using different parsers; create a new multi-tree, set a root node for the merged tree; The multi-syntactic analysis trees are searched, and a parallel set-merging method is used, for the same nodes in the multi-syntactic analysis trees, the corresponding sub-trees are merged, and for the nodes existing in only one syntactic analysis tree, the nodes and the sub-trees are directly added to the new multi-tree; For the preliminary merged multi-tree, from bottom to top, a branch distillation and branch pruning method is used to process the repeated nodes, and the final merged multi-tree is output.
[0006] The creating of the new multi-tree and the setting of the root node for the merged tree are specifically as follows: a new multi-tree T is created, and the following formula is used for merging:
[0007] In the formula, is an aggregation function, is any attribute of the multi-tree T, , are multi-syntactic analysis trees generated by different parsers, is the contribution of the current node.
[0008] The parallel set-merging method is used when any two syntactic analysis trees A and B are merged, and the same nodes exist: The node r exists in the B tree, the node r exists in the A tree, and the node r is retained after merging: ; The parallel set-merging method is used when any two syntactic analysis trees A and B are merged, and the nodes exist in only one syntactic analysis tree: The node n does not exist in the B tree, the node n exists in the A tree, and the node n is retained after merging: ; The node m exists in the B tree, the node m does not exist in the A tree, and the node m is retained after merging: 。
[0009] For the preliminary merged multi-tree, from bottom to top, for the repeated sub-nodes, the parent nodes are traced back, the branches formed by the sub-nodes and the parent nodes are extracted, branch matching is performed, if the branch matching is successful, branch pruning is performed, and if the matching fails, node distillation is performed.
[0010] The branch pruning, specifically, after the sub-tree merging, if the same parent node has repeated or matchable sub-node branches, the parent node is traced back, the branch formed by the parent node and the sub-node is extracted for matching, when the matching is successful, the sub-node with higher depth value and the connection between the sub-node and the parent node are directly deleted, and only the branch with lower depth value is reserved.
[0011] The node distillation, specifically, after the sub-tree merging, if the same parent node has repeated or matchable sub-node branches, the parent node is traced back, the branch formed by the parent node and the sub-node is extracted for matching, when the matching fails, for the nodes with equal depth values, the node importance ranking is calculated by using a self-attention algorithm, the nodes with higher rankings are reserved, and the nodes with lower rankings are deleted; if a node has a depth significantly higher than other nodes and has no matchable branch, the node is directly deleted.
[0012] The method of adopting and merging, specifically, all nodes of two or more syntax analysis trees are merged together, after the repeated elements are removed, a set is obtained, all nodes and edges of the multiple sub-trees are merged into a new tree structure, and the original connection relationship is maintained.
[0013] The traversal search on the multiple syntax analysis trees adopts a depth-first traversal method, starts from a root node, searches as deeply as possible along a path, until a target node is reached or the search cannot continue, then backtracks and tries other paths.
[0014] The traversal search on the multiple syntax analysis trees adopts a breadth-first traversal method, starts from a root node, first visits all adjacent nodes, then starts from the adjacent nodes, visits the unvisited adjacent nodes of the adjacent nodes, and so on.
[0015] Compared with the prior art, the present application has the following beneficial effects: The application provides a biomedical event extraction recursive neural network multi-tree fusion method, different parsers are used to generate diversified tree structures, information complementation is performed through structural isomerism fusion, and a heterogeneous syntax parsing tree merging method is provided. BRIEF DESCRIPTION OF DRAWINGS
[0016] Figure 1 FIG. 1 is a biomedical event extraction flowchart based on the recursive neural network multi-tree fusion method provided in the application; Figure 2 FIG. 3 is a sub-tree set merging and post-processing process schematic diagram of path three in the embodiment of the application; Figure 3 FIG. 4 is a sub-tree set merging and post-processing process schematic diagram of path six in the embodiment of the application; Figure 4 FIG. 6 is a comparison result of the isomerism fusion method of the application; Figure 5 FIG. 7 is a fusion path traversal comparison schematic diagram of the application; Figure 6 FIG. 8 is a flowchart of a biomedical event extraction recursive neural network multi-tree fusion method. DETAILED DESCRIPTION
[0017] For further understanding of the present application, the following will make a detailed description of the present application in combination with the accompanying drawings and specific embodiments. It should be understood that the embodiments are only to explain the present application and not to limit it.
[0018] The embodiments of the present application will be described in detail below in combination with the accompanying drawings.
[0019] Embodiment 1 This embodiment proposes a recursive neural network multi-tree fusion method in biomedical event extraction, as shown in Figure 6 The implementation method includes the following steps, that is: Different parsers are used to generate multiple syntax analysis trees from the text content; A new multi-tree is created, and a root node is set for the merged tree; The multiple syntax analysis trees are searched, and the parallel set merging method is used. For the same nodes in multiple syntax analysis trees, the corresponding sub-trees are merged. For the nodes only existing in one syntax analysis tree, the node and its sub-tree are directly added to the new multi-tree; For the preliminary merged multi-tree, from bottom to top, the branch distillation and branch pruning methods are used to process the repeated nodes, and the final merged multi-tree is output.
[0020] Based on the above-mentioned recursive neural network multi-tree fusion method in biomedical event extraction, biomedical event extraction is carried out, and the biomedical event extraction framework is as shown in Figure 1 The input layer, the word embedding layer, the semantic update layer and the output layer are included. In the field of biomedical event extraction, due to the diversity of data sources (such as literature, electronic medical records, experimental reports, etc.), the data structure is heterogeneous, which leads to differences in structure and semantics of the sub-trees generated by text analysis. Therefore, different parsers are used for syntax analysis to generate different syntax analysis trees. In order to achieve accurate biomedical event extraction, the present application provides multiple multi-tree fusion paths.
[0021] First, a batch of biomedical literature data is obtained, and the text in the literature is preprocessed by natural language processing technology. The syntax analysis tool and the semantic role labeling algorithm are used to analyze each literature text into a syntax analysis tree. The semantic update layer adopts a recursive neural network based on heterogeneous fusion, which is used to merge multiple syntax analysis trees and post-process the repeated redundant nodes. The tree structure after merging and post-processing is used as the input of the semantic update layer. After the semantic update layer, the fused tree structure is traversed, and biomedical events are extracted according to the information of the nodes and edges in the tree. Finally, the biomedical event extraction result is output through the output layer.
[0022] Embodiment 2 Based on the recursive neural network multi-tree fusion method in biomedical event extraction proposed in embodiment 1, this embodiment takes two sub-trees T1 and T2 as an example, adopts the depth-first traversal or breadth-first traversal method to traverse the two sub-trees, and adopts the parallel set union to merge the two sub-trees. The parallel set union refers to merging all nodes of two or more sub-trees together to obtain a set after removing duplicate elements. This way merges all nodes and edges of multiple sub-trees into a new tree structure while maintaining the original connection relationship. For the two sub-trees T1 and T2, a new multi-tree production A→BC is initialized, B is taken as the root of T1, C is taken as the root of T2, and a new tree is merged, the root node is A, and the child nodes are B and C. For any attribute of tree T , , , ,…, , the formula is:
[0023] wherein g is an aggregation function, which can adopt addition or maximum value, is the contribution of the current node.
[0024] The parallel set union method adopted in the application, when merging any two sub-trees A and B, for multiple syntax analysis trees that have the same node, the corresponding sub-trees are merged, and for the node that only exists in one syntax analysis tree, the node and its sub-tree are directly added to the new multi-tree; the specific merging of the same node is as follows: B tree has node r, A tree has node r, and node r is retained after merging: ; B tree has no node n, A tree has node n, and node n is retained after merging: ; B tree has node m, A tree has no node m, and node m is retained after merging: 。
[0025] For the preliminary merged multi-ary tree, branch distillation and branch pruning methods are used from bottom to top to process repeated nodes. For repeated child nodes, backtracking to the parent node, extracting the branch formed by the child node and the parent node, and entering branch matching. If branch matching is successful, branch pruning is performed, and the child node with a higher depth value and its connection with the parent node are directly deleted, and only the branch with a lower depth value is retained. If the matching fails, node distillation is performed. For nodes with equal depth values, the self-attention algorithm is used to calculate the node importance ranking, and the nodes with high ranking are retained and the nodes with low ranking are deleted. If a node is significantly higher than other nodes in depth and has no matching branch, the node is directly deleted.
[0026] In further preferred embodiments of the application, eight different multi-tree fusion paths are designed: Path one: BFS-M{prun} (breadth-first merge-subtree pruning) The breadth-first traversal method is used for priority merging. Taking the root node as the starting point, the child trees are merged layer by layer based on the above merging rules according to the breadth-first search strategy. First, the child trees at the same level are merged to form a preliminary tree structure. In the merging process, the node attributes of the newly generated tree structure are assigned according to the production rule, and the attributes of the new tree are calculated through the attribute aggregation formula. The preliminary merged tree structure is traversed from bottom to top. For repeated sub-tree branches, the node pruning method is used to delete the sub-tree branch with a higher depth value, and the sub-tree branch with a lower depth value is retained to obtain the final fused tree structure.
[0027] Path two: BFS-M{distillation} (breadth-first merge-node distillation) Similarly, the breadth-first traversal method is used for priority merging. The initial child trees are merged layer by layer to construct a preliminary tree structure. The preliminary merged tree structure is traversed from bottom to top. For unmatched nodes, the node distillation method is used to calculate the ranking of these nodes using the self-attention algorithm. According to the ranking, the nodes with high importance are retained and the nodes with low ranking are deleted to obtain the fused tree structure.
[0028] Path three: BFS-M{prun|distillation} (breadth-first merge-subtree pruning|node distillation) As shown in Figure 2 Similarly, the breadth-first traversal method is used for priority merging. The initial child trees are merged layer by layer to construct a preliminary tree structure. The preliminary merged tree structure is traversed from bottom to top. In the merged tree structure, there are both repeated sub-tree branches describing the same biomedical relationship and nodes with attribute conflicts. By simultaneously performing pruning and distillation operations, the tree structure is optimized to obtain the final tree structure.
[0029] Path four: BFS-M{distillation | prun} (breadth-first merge → node distillation | subtree pruning) The breadth-first traversal method is used for priority merging, and the initial subtree is merged layer by layer to construct a preliminary tree structure. The preliminary merged tree structure is traversed from bottom to top, and the node distillation operation is performed on the mismatched nodes in the preliminary merged tree structure to solve the node attribute conflict and importance screening problem. After node distillation, the tree structure is traversed again to check and delete duplicate subtree branches, perform node pruning operation, and obtain the final fusion tree structure.
[0030] Path five: DFS-M{prun} (depth-first merge → subtree pruning) Starting from the root node, according to the depth-first traversal method, the subtree is merged as deeply as possible along one branch until the branch is merged, and then the other branches are processed in reverse, forming a preliminary tree structure. The preliminary merged tree structure is traversed from bottom to top, and the node pruning method is used to delete the duplicate branches with high depth values to obtain the final tree structure.
[0031] Path six: DFS-M{distillation} (depth-first merge → node distillation) As shown in Figure 3 Starting from the root node, according to the depth-first traversal method, a preliminary tree structure is constructed. The preliminary merged tree structure is traversed from bottom to top, and the node distillation method is used to process the deep mismatched nodes to optimize the tree structure and obtain the final tree structure.
[0032] Path seven: DFS-M{prun | distillation} (depth-first merge → subtree pruning | node pruning) Starting from the root node, according to the depth-first traversal method, a preliminary tree structure is constructed. The preliminary merged tree structure is traversed from bottom to top, and the node pruning operation is performed on the duplicate subtree branches that appear in the deep layer, and the node distillation operation is performed on the mismatched nodes to comprehensively optimize the tree structure and obtain the final tree structure.
[0033] Path eight: BFS-M{distillation | prun} (breadth-first merge → node distillation | subtree pruning) The breadth-first traversal method is used for priority merging, and a preliminary tree structure is constructed. The preliminary merged tree structure is traversed from bottom to top, and the node distillation operation is performed on the mismatched nodes to screen out important nodes. Then, the tree structure is traversed to delete duplicate subtree branches, perform node pruning operation, and obtain the final fused tree structure.
[0034] Based on the above eight different multi-tree fusion paths,Figure 2 and Figure 3 In the sub-tree merging and post-processing process shown in FIG. 6, the input sentence is "Thalidomide inhibited the formation of capillary tubes.", which is parsed by the Stanford parser and the spaCy parser to generate two syntax analysis trees, respectively. The sub-tree merging method is used to obtain a merged tree, and the sub-tree pruning and node distillation are performed on the merged tree, respectively, to finally generate a simplified merged tree as the input of the semantic update layer. It is worth noting that not every merged result will produce eight paths. According to the actual merged result, the algorithm will switch to the corresponding path. In the example, the algorithm switches to path three and path six according to the requirements. After processing the biomedical literature data through the above eight different multi-tree fusion paths, the biomedical event extraction task is tested. Path one and path five can quickly eliminate repeated structures and improve the efficiency of event extraction when processing similar structured data; path two and path six can effectively filter key nodes and improve the accuracy of event extraction when processing highly heterogeneous data. Users can flexibly choose the appropriate fusion path according to the characteristics and fusion requirements of the actual data, so as to realize more efficient and accurate biomedical event extraction and provide powerful data support for biomedical research and application.
[0035] In the biomedical event extraction task, to explore the effectiveness of different sub-tree heterogeneous merging rules, post-processing rules and fusion paths, the present embodiment is based on the BioNLP'09 dataset, and experiments are conducted to compare different sub-tree heterogeneous merging rules. The specific process is as follows: Sub-tree (stanford): only using Stanford syntax parsing tree; Sub-tree (malt): syntax parsing tree generated based on MaltParser; Sub-tree (spaCy): syntax parsing tree constructed using spaCy tool.
[0036] BFS-M{sub-tree (stanford)+sub-tree (malt)}: breadth-first merging of sub-trees {sub-tree (stanford)+sub-tree (malt)}; merging the sub-trees generated by Stanford and MaltParser through breadth-first strategy; BFS-M{sub-tree (spaCy)+sub-tree (malt)}: breadth-first merging of sub-trees {sub-tree (spaCy)+sub-tree (malt)}; breadth-first merging of spaCy and MaltParser sub-trees; BFS-M {Subtree(stanford) + Subtree(malt) + Subtree(spaCy)}: Breadth-First Merging of Subtrees {Subtree(stanford) + Subtree(malt) + Subtree(spaCy)}; Breadth-First Merging of Subtrees generated by three tools; DFS-M {Subtree(stanford) + Subtree(malt)}: Depth-First Merging of Subtrees {Subtree(stanford) + Subtree(malt)}; Depth-First Merging of Stanford and MaltParser Subtrees; DFS-M {Subtree(spaCy) + Subtree(malt)}: Depth-First Merging of Subtrees {Subtree(spaCy) + Subtree(malt)}; Depth-First Merging of spaCy and MaltParser Subtrees; DFS-M {Subtree(stanford) + Subtree(malt) + Subtree(spaCy)}: Depth-First Merging of Subtrees {Subtree(stanford) + Subtree(malt) + Subtree(spaCy)}; Depth-First Merging of Subtrees generated by three tools.
[0037] By comparing the different subtree isomerization merging rules described above, the experimental results are shown in Table 1 as follows: Figure 4 The experimental results show that: whether breadth-first or depth-first, the structure after subtree merging is better than the result of single syntactic parsing tree. For the subtree merging process, the effect of merging three subtrees is better than merging two. The more the number of merged subtrees, the richer the information carried by the isomerization structure, and the more complete the tree structure after merging. Especially for complex biomedical events with nested structures, the node representation semantic update effect is better after structure interaction fusion. The breadth-first traversal strategy preferentially accesses the neighbor nodes, because the merging process has strong compatibility for subtree merging, so the breadth-first effect is better.
[0038] In further preferred embodiments of the application, based on the BFS-M {Subtree(stanford) + Subtree(malt)} merging scheme, the subtree post-processing link is compared. For the subtree deduplication process, branch pruning is performed on the basis of merging. For repeated redundant subnodes, node distillation is performed. The priority of branch pruning (prun) and node distillation (distillation) can be manually set.
[0039] The experimental results in the BioNLP'09 dataset are shown in Table 1. It can be seen that the performance of the combined results after sub-tree post-processing is obviously improved. Branch pruning and node distillation improve the performance of the model. The reason is that sub-tree post-processing simplifies the combined tree model. After only node distillation, the effect is similar to branch pruning. However, when branch pruning and node distillation are used at the same time, the sub-tree structure is further simplified, and the overall performance of the model is generally improved. For the priority selection of branch pruning and node distillation, first perform sub-tree pruning, and then perform node distillation, which is better for binding and regulating complex events. The reason is that this process has a repeated branch judgment, which can ensure the integrity of the sub-tree structure as much as possible. However, if node distillation is performed first, and then the repeated node removal is performed at the same time, the related edges are also removed, which will cause the destruction of the tree structure.
[0040] Table 1
[0041] In a further preferred embodiment of the application, based on the BioNLP'09 dataset, the following eight heterogeneous fusion paths are compared using the "sub-tree (stanford) + sub-tree (malt) + sub-tree (spaCy)" merging scheme: Path one: BFS-M {prun}: sub-tree pruning after breadth-first merging; Path two: BFS-M {distillation}: node distillation after breadth-first merging; Path three: BFS-M {prun|distillation}: pruning first and then distillation after breadth-first merging; Path four: BFS-M {distillation|prun}: distillation first and then pruning after breadth-first merging; Path five: DFS-M {prun}: sub-tree pruning after depth-first merging; Path six: DFS-M {distillation}: node distillation after depth-first merging; Path seven: DFS-M {prun|distillation}: pruning first and then distillation after depth-first merging; Path eight: DFS-M {distillation|prun}: distillation first and then pruning after depth-first merging.
[0042] The experimental comparison is as follows: Figure 5As shown, the experimental results show that the fusion path using breadth-first traversal is better than the fusion path using DFS depth-first traversal, and the compatibility advantage of BFS in sub-tree set merging is verified again. The reason is that the merging process has strong compatibility for sub-tree set merging. Path three and path seven first perform sub-tree pruning, and then perform node distillation, which can guarantee the integrity of the sub-tree structure as much as possible, and perform best in processing complex biomedical events. This order ensures the integrity of the tree structure by preferentially preserving the complete branch structure, thereby improving the accuracy of complex event semantic analysis. Path two and path six rely only on a single post-processing operation, which can easily cause the integrity of the tree structure to be broken during the merging process, resulting in lower event extraction performance than the combined strategy path, and all effects are less than paths three / seven.
[0043] In addition, it should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that those skilled in the art can understand. The above is only to illustrate the technical idea of the present application, and cannot limit the protection scope of the present application. Any modification made on the basis of the technical solutions of the present application falls within the protection scope of the claims of the present application.
Claims
1. A recursive neural network multi-tree fusion method for biomedical event extraction, characterized in that: The following steps are involved: Use different parsers to generate multiple syntactic analysis trees from the text content; Create a new multitree and set the root node for the merged tree; Traverse and search multiple parse trees, using the union-merge method. For nodes that exist in multiple parse trees, merge their corresponding subtrees. For nodes that only exist in one parse tree, directly add the node and its subtree to the new multi-tree. For the multi-branch tree after the initial merger, the node distillation and branch pruning methods are used from bottom to top to process the duplicate nodes and output the final merged multi-branch tree.
2. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 1, characterized in that: The method of creating a new multi-branch tree and setting the root node for the merged tree is as follows: creating a new multi-branch tree T and merging it using the following formula: Where, is an aggregate function, is any attribute of the multitree T, 、 Multiple syntactic analysis trees generated by different parsers, The contribution of the current node.
3. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 1, characterized in that: In the union merging method, when any two parse trees A and B are merged, there are the same nodes: There is a node r in the B-tree and a node r in the A-tree. After merging, the node r is retained: 。 4. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 1, characterized in that: In the union merging method, when any two parse trees A and B are merged, there is a situation where a node exists in only one parse tree: There is no node n in the B-tree, but there is node n in the A-tree. After merging, node n is retained: ； There is a node m in the B-tree, but not in the A-tree. After merging, the node m is retained: 。 5. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 1, characterized in that: The multi-branch tree after the initial merger is traversed from bottom to top. For repeated child nodes, it traces back to the parent node, extracts the branches formed by the child nodes and the parent node, and enters branch matching. If the branch matching is successful, branch pruning is performed. If the matching fails, node distillation is performed.
6. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 5, characterized in that: The branch pruning is specifically as follows: after the subtrees are merged, if duplicate or matchable child node branches are found under the same parent node, the parent node is traced back to extract the branches consisting of the parent node and the child nodes for matching. If the match is successful, the child node with the higher depth value and its connection with the parent node are directly deleted, and only the branch with the lower depth value is retained.
7. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 5, characterized in that: The branch distillation is specifically as follows: after the subtrees are merged, if repeated or matching child node branches are found under the same parent node, the parent node is traced back to extract the branches consisting of the parent node and the child nodes for matching. When the matching fails, for nodes with equal depth values, the self-attention algorithm is used to calculate the node importance ranking, retaining the nodes with higher rankings and deleting the nodes with lower rankings; if the depth of a node is significantly higher than that of other nodes and there is no matching branch, the node is directly deleted.
8. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 1, characterized in that: The method of adopting the union merging method is specifically as follows: merging all nodes of two or more syntactic analysis trees together, removing duplicate elements to obtain a set, and merging all nodes and edges of multiple subtrees into a new tree structure while maintaining the original connection relationship.
9. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 1, characterized in that: The traversal search of the multiple syntactic analysis trees adopts a depth-first traversal method, starting from the root node, searching as deeply as possible along a path until the target node is reached or it is impossible to continue, and then backtracking and trying other paths.
10. The recursive neural network multi-tree fusion method for biomedical event extraction according to claim 1, characterized in that: The traversal search of the multiple syntactic analysis trees adopts a breadth-first traversal method, starting from the root node, first visiting all adjacent nodes, and then starting from these adjacent nodes, visiting their unvisited adjacent nodes, and so on.