Information extraction method based on topological graph non-independent sampling and context association learning

Abstract

The invention provides an information extraction method based on topological graph non-independent sampling and context correlation learning, and relates to the technical field of biomedicine, and the method comprises the steps: carrying out the data extraction based on a biomedical event extraction task, and obtaining original sample data; obtaining a low-frequency trigger word pair set and a trigger word candidate set based on the trigger word type of the original sample data; the method comprises the following steps: extracting context representations of a low-frequency trigger word pair set and a trigger word candidate set by utilizing context association learning, obtaining a corresponding low-frequency trigger word pair representation vector and a trigger word candidate representation vector through average pooling, calculating a transition probability, obtaining an enhanced training set through topological graph non-independent sampling, and training a hybrid neural network to obtain a low-frequency trigger word pair representation vector and a trigger word candidate representation vector. And obtaining a trained hybrid neural network, and analyzing the biomedical event extraction task to obtain an information extraction result. According to the method, the problem that the existing trigger word recognition technology has limitation in the aspects of long-tail distribution modeling and context semantic modeling is solved.

Description

Technical Field

[0001] This specification relates to the field of biomedical technology, and in particular to an information extraction method based on topological graph non-independent sampling and contextual association learning. Background Technology

[0002] With the explosive growth in the number of biomedical documents, extracting structured event information from unstructured text has become a key task in bioinformatics research and clinical decision support. Biomedical event extraction aims to identify events and their elements related to biological processes such as diseases, genes, and proteins, with trigger word detection playing a fundamental and crucial role.

[0003] Trigger words are typically central words that identify the occurrence of an event, such as "activate," "expression," and "phosphorylation." Identifying these words not only helps in the structural modeling of events but also has a decisive impact on subsequent role identification and event classification. Therefore, high-quality trigger word identification results directly determine the effectiveness of the entire event extraction system. In existing mainstream methods, trigger word detection is usually modeled as a sequence labeling task, where each input sentence is divided into a sequence of tokens, and the model needs to predict the role of each token in the event trigger words. The commonly used labeling system is the BIO (Begin-Inside-Outside) format, and its modeling objective can be formalized as follows: Each of them Indicates the first in the sentence Each token has a corresponding output tag sequence. ,in, Indicates the beginning of the trigger word. Indicates the interior of the trigger word, This represents a non-trigger word. The model's goal is to learn a mapping function. This makes the output label sequence... It can accurately reflect the position and type of all trigger words in a sentence.

[0004] However, current biomedical trigger word extraction problems face the challenge of severe long-tail distribution and difficulty in learning low-frequency trigger words. Biomedical event trigger words exhibit a typical power-law distribution, meaning that a very small number of high-frequency trigger words occupy the majority of training instances, while trigger word types in most original sample data appear only a few dozen times or even less. This imbalanced distribution causes the model to be biased towards learning high-frequency categories during training, thus ignoring or misjudging semantically important but sparsely occurring low-frequency event types. To quantify this distributional bias, we can define low-frequency trigger word pairs as... .in Indicates the trigger word type of the original sample data. Indicate its context, This indicates the number of times the two elements co-occur in the training set. The preset threshold is used. Most important trigger words (such as "localization", "catabolism", etc.) belong to this type of low-frequency pair, and the model lacks enough samples to learn its discrimination boundary.

[0005] Secondly, biomedical trigger words have complex semantic dependencies, making them difficult to model using static windows and standard attention mechanisms. Biomedical sentences often contain multiple entities and modifying structures, and the occurrence and semantic discrimination of trigger words are highly dependent on the context structure. For example, "expression" is an event trigger word in "gene expression increased," but not in "expression of interest." If the model uses a fixed window or standard self-attention mechanism, it may fail to perceive distant dependencies or syntactic features under different sentence structures, leading to misjudgments or missed detections.

[0006] Meanwhile, traditional loss functions struggle to focus on difficult samples, and the training process is frequency-driven. Standard cross-entropy loss assigns equal weights to all samples, which can lead to the model ignoring low-frequency or difficult-to-learn categories during training when data distribution is severely uneven. Although some studies have attempted to use class reweighting mechanisms, dynamic adjustment strategies for difficult samples are still lacking. Therefore, methods such as Focal Loss have emerged in recent years. Their core idea is to reduce the weight of easily classified samples and increase attention to difficult samples; it is defined as… ,in For position The predicted probability of getting the correct label. For category weights, >0 is the focus factor.

[0007] Despite significant progress in trigger word detection regarding model structure and pre-trained language models, key challenges remain in low-frequency category enhancement, context-dependent modeling, and loss function design. These issues are particularly acute in biomedical texts due to their more complex semantic structures and sparser labeling systems. Summary of the Invention

[0008] To address the aforementioned shortcomings in existing technologies, this invention provides an information extraction method based on non-independent sampling of topological graphs and contextual association learning, which solves the limitations of existing trigger word recognition technologies in long-tail distribution modeling and contextual semantic modeling.

[0009] To achieve the aforementioned objectives, the present invention employs the following technical solution: an information extraction method based on topological graph non-independent sampling and contextual association learning, comprising: S1: Based on the biomedical event extraction task, data extraction is performed to obtain raw sample data; S2: Based on the trigger word type of the original sample data, the set of low-frequency trigger word pairs and the candidate set of trigger words are obtained through calculation; S3: Utilize context association learning to extract the context representations of the low-frequency trigger word pair set and the trigger word candidate set, and obtain the corresponding low-frequency trigger word pair representation vector and trigger word candidate representation vector through average pooling; S4: Calculate the transition probability using the low-frequency trigger word pair representation vector and the candidate trigger word representation vector, and obtain the enhanced training set through non-independent sampling of the topological graph; S5: Use the augmented training set to train the hybrid neural network to obtain a trained hybrid neural network; S6: Utilize the trained hybrid neural network to analyze the biomedical event extraction task, obtain information extraction results, and complete the information extraction.

[0010] Further, S4 includes: The transition probability is obtained by using the low-frequency trigger word pair representation vector and the trigger word candidate representation vector. Based on the transition probability, seed samples are obtained by performing Markov chain sampling on the graph through non-independent sampling of the topological graph. Based on seed samples, new sentences and labels are generated using a large language model to obtain an augmented dataset; By filtering the augmented dataset, an augmented training set is obtained.

[0011] Furthermore, the expression for the transition probability is: ; in, This indicates that the node represents the node. Transition to node The transition probability, This represents the source node in the trigger word topology graph. This represents a directed edge from the source node to the target node. This represents the target node in the trigger word topology graph. Indicates the control ratio. This represents a similarity function based on trigger word type. Represents a node The corresponding trigger word type, Represents a node The corresponding trigger word type, The representation is a similarity function based on context representation. Represents the node The corresponding context feature vector, Represents the node The corresponding context feature vector.

[0012] Furthermore, the hybrid neural network includes: The encoding layer is used to encode the augmented training set to obtain the context representation of each token; The dynamic context association module is used to process the context representation of each token using a self-attention mechanism to obtain a context-enhanced representation with discriminative capabilities. The Transformer layer is used to process the context-enhanced representation to obtain a globally context-enhanced token representation; The Conditional Random Field layer is used to process the globally context-enhanced token representation using label scores and label transitions to obtain sequence labels; The prediction layer is used for analysis based on sequence labels to obtain information extraction results.

[0013] Furthermore, the expression for the context-enhanced representation is: ; ; ; in, This represents the attention weight of the i-th target word to the j-th context word under trigger word type t. This represents the transpose of the attention weight vector. This represents the context representation vector of the i-th word in the input sequence. Let represent the context representation vector of the j-th word in the input sequence. , and Represents the learnable projection matrix. The type embedding vector represents the trigger word type t of the original sample data, k represents the total number of trigger word types, and n represents the number of words in the current sequence.

[0014] Furthermore, the expression for the sequence tag is: ; in, This represents the conditional probability of obtaining the label sequence Y given the input sequence X. Representation and Label The transpose of the corresponding learnable weight vector, This represents the feature vector at position i. This represents the transition score function between adjacent labels. Indicates the i-th A label in one location, This represents the label at the i-th position. This represents the normalization factor obtained by summing over all possible label sequences. This represents the input text or word sequence. This represents the label sequence corresponding to the input sequence.

[0015] Furthermore, the expression for the loss function of the hybrid neural network is: ; ; ; ; in, Represents the total loss function. This represents the negative log-likelihood loss of a conditional random field layer. Indicates the weighting coefficient. Denotes the auxiliary loss function. Represents the category weight coefficient. This represents the probability that position i is predicted to be label c. Indicates the focus adjustment parameter, This represents the true label sequence corresponding to a given input sequence. Indicates the true label, Let represent the indicator function, where i represents the position index in the sequence, c represents the label category index, and n represents the total length of the labels in the sequence. Represents the set of all possible label sequences. express A candidate label sequence in the, This represents the sequence scoring function computed for a given feature representation U and label sequence. This represents the set of feature representations of the entire input sequence obtained by a hybrid neural network. Represents the set of real labels. This represents the emission score vector at the i-th position. Indicates from label y i-1 Move to label The transfer score.

[0016] The beneficial effects of this invention are as follows: This invention provides an information extraction method based on topological graph non-independent sampling and context association learning. Based on the low-frequency label non-independent sampling mechanism of the topological graph, a label co-occurrence graph is constructed based on the co-occurrence frequency and semantic distance of trigger words in the training data. Using this as a priori, a large-scale language model is guided to generate new samples that satisfy specific label combinations, thereby significantly expanding the semantic distribution space of low-frequency trigger events in the training set. The constructed dynamic context modeling module realizes the modeling of context dependencies under different trigger types by fusing trigger word type, entity position and context representation. Unlike the unified attention mechanism of traditional Transformers, the DCA module dynamically adjusts the contextual focus region based on different labels, effectively identifying long-distance dependencies and ambiguous expressions, and significantly improving the model's ability to distinguish trigger words in complex semantic structures and its generalization ability. During the decoding stage, a Conditional Random Field (CRF) is introduced to model the transition relationships between BIO labels, enhancing the consistency of the output sequence. Simultaneously, a Focal Loss mechanism is used to assign higher weights to low-confidence or hard-to-classify samples, thereby strengthening the model's ability to learn low-frequency labels. By introducing a label combination sampling and entity consistency check mechanism based on a topology graph, the large model is guided to generate only within the legal label subspace, avoiding invalid combinations and semantic drift. Furthermore, biomedical semantic style constraints are imposed on the generated text (such as preserving event verbs and normalizing entity representations) to ensure that the generated results maintain a high degree of consistency with the original data in style, semantics, and structure. Attached Figure Description

[0017] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart illustrating an information extraction method based on non-independent sampling of topological graphs and contextual association learning, according to some embodiments of this specification. Figure 2 This is an exemplary schematic diagram illustrating an information extraction method based on non-independent sampling of topological graphs and contextual association learning, according to some embodiments of this specification; Figure 3 This is an exemplary schematic diagram of the data-enhanced tag frequency distribution shown in some embodiments of this specification. Detailed Implementation

[0018] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0019] Example Figure 1 This is an exemplary flowchart illustrating an information extraction method based on non-independent sampling of topological graphs and contextual association learning, according to some embodiments of this specification. Figure 1 and Figure 2 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0020] S1: Based on the biomedical event extraction task, data extraction is performed to obtain raw sample data.

[0021] The biomedical event extraction task includes trigger word detection, a key subtask within the overall biomedical event extraction process. Its goal is to identify words or phrases representing specific biomedical events within text. The results serve as the foundation for subsequent tasks such as event role recognition and event classification.

[0022] Data extraction, specifically biomedical event extraction, is the task of transforming unstructured biomedical text into structured event information. It is widely used in knowledge graph construction, literature mining, and clinical decision support. Typical events include protein phosphorylation, gene expression, and apoptosis.

[0023] The original training set is ,in Original sentence, This indicates a BIO sequence label.

[0024] S2: Based on the trigger word type of the original sample data, the set of low-frequency trigger word pairs and the candidate set of trigger words are obtained through calculation.

[0025] In some embodiments, the processor may define low-frequency trigger word-context pairs. This yields a set of low-frequency trigger word pairs. Trigger word candidate set A subset of samples containing at least one low-frequency trigger word pair; wherein, Indicates the trigger word type. Indicates context, This indicates the frequency of their occurrence in the training set. The threshold value is set.

[0026] S3: Utilize contextual association learning to extract the contextual representations of the low-frequency trigger word pair set and the trigger word candidate set. Through average pooling, obtain the corresponding low-frequency trigger word pair representation vector and trigger word candidate representation vector.

[0027] In some embodiments, the processor may utilize BioBERT to extract a context representation: The trigger word phrases are subjected to average pooling to obtain their representation vectors. , .

[0028] S4: Calculate the transition probability using the low-frequency trigger word pair representation vector and the trigger word candidate representation vector, and obtain the enhanced training set through non-independent sampling of the topological graph.

[0029] The topology graph is a semantic transition graph of trigger words, a directed graph structure used to define the semantic transition relationships between trigger words. Nodes in the graph represent labeled trigger words and their context representations, while edges represent transition probabilities under both type consistency and context similarity metrics. This structure serves as the basis for the MCMC sampling distribution and is used for selecting high-quality seed instances.

[0030] In some embodiments, the processor can use the low-frequency trigger word pair representation vector and the trigger word candidate representation vector to calculate the transition probability; based on the transition probability, perform Markov chain sampling on the graph through non-independent sampling of the topological graph to obtain seed samples; based on the seed samples, use a large language model to generate new sentences and labels to obtain an augmented dataset; and obtain an augmented training set by filtering the augmented dataset.

[0031] Topology graph non-independent sampling, also known as semantically guided Markov sampling, is a non-independent sampling mechanism based on trigger word semantic graphs to mine diverse contexts of low-frequency trigger words. This method combines the trigger word type similarity and context vector similarity of the original sample data to construct a topology transition graph, and then runs MCMC sampling on this graph to select semantically coherent seed instances, guiding the large language model to generate enhanced samples.

[0032] In some embodiments, the processor can use the low-frequency trigger word pair representation vector and the trigger word candidate representation vector to calculate the transition probability; wherein, In the output data, nodes are represented by trigger words, and edges are similar edges with transition probabilities (used for MCMC sampling).

[0033] In some embodiments, the transition probability is expressed as: ; in, This indicates that the node represents the node. Transition to node The transition probability, This represents the source node in the trigger word topology graph. This represents a directed edge from the source node to the target node. This represents the target node in the trigger word topology graph. Indicates the control ratio. This represents a similarity function based on trigger word type. Represents a node The corresponding trigger word type, Represents a node The corresponding trigger word type, The representation is a similarity function based on context representation. Represents the node The corresponding context feature vector, Represents the node The corresponding context feature vector.

[0034] In some embodiments, such as Figure 3 As shown, the processor can perform Markov chain sampling on the graph (following...). Each round of sampling selects one seed sample. Call the large language model to generate new sentences and generate tags Enhanced sample composition The generated sentences need to meet the following criteria: label consistency (the presence of BIO labels of type t), BIO validity, semantic similarity (cosine similarity ≥ 0.55), and non-repetition (Jaccard similarity ≤ 0.8). A prompt is then constructed, and the selected augmented dataset is used. The enhanced training set is obtained as .

[0035] S5: Use the augmented training set to train the hybrid neural network to obtain a trained hybrid neural network.

[0036] Hybrid neural networks are neural network models used for information extraction.

[0037] In some embodiments, the hybrid neural network includes: an encoding layer for encoding the augmented training set to obtain a context representation for each token; a dynamic context association module for processing the context representation for each token using a self-attention mechanism to obtain a context-enhanced representation with discriminative capabilities; a Transformer layer for processing the context-enhanced representation to obtain a globally context-enhanced token representation; a Conditional Random Field layer for processing the globally context-enhanced token representation using label scores and label transitions to obtain sequence labels; and a prediction layer for analyzing based on the sequence labels to obtain information extraction results.

[0038] Enter sentence , can be represented as: The output data is the context representation of each token. .

[0039] The encoding process includes a BIO structure repair and consistency screening mechanism, a post-processing mechanism specifically designed to ensure the quality of data generated from large models. This method includes dictionary-driven BIO annotation generation, structure validity repair, trigger word type consistency verification of original sample data, and semantic similarity and length diversity screening, ensuring enhanced data usability and semantic consistency.

[0040] BIO (Boundary Intent) tagging is a commonly used sequence tagging format for identifying the boundaries and categories of entities or trigger words. Here, B indicates the beginning of a segment, I indicates the interior, and O indicates that it does not belong to any tagging object.

[0041] The dynamic context association module is a context modeling module based on gating attention mechanism, designed to capture the multi-granular dynamic dependencies between trigger words and their context. This module dynamically adjusts the effective receptive field through token-level attention and gating mechanisms, effectively enhancing the semantic discrimination ability in trigger word recognition, especially performing well in contexts with complex syntactic structures and dense biological entities.

[0042] In some embodiments, the processor can represent the input sentence. Use a self-attention mechanism to capture the relationships between tokens. A learnable projection matrix is ​​used to incorporate prior information at the label level, mapping the input vector to a unified attention space. Used to map semantic associations after non-linear activation to scalar attention scores. This is used to enhance the non-linearity of the model's representation. After computing the context, a gating mechanism is used for fusion, and the output data is a context-enhanced representation with discriminative capabilities. .

[0043] In some embodiments, the expression for context-enhanced representation is: ; ; ; in, This represents the attention weight of the i-th target word to the j-th context word under trigger word type t. This represents the transpose of the attention weight vector. This represents the context representation vector of the i-th word in the input sequence. Let represent the context representation vector of the j-th word in the input sequence. , and Represents the learnable projection matrix. The type embedding vector represents the trigger word type t of the original sample data, k represents the total number of trigger word types, and n represents the number of words in the current sequence.

[0044] In some embodiments, the processor can use a Transformer neural network to process the input data. After processing, a token representation with global context enhancement is obtained. .

[0045] The Transformer neural network can be represented as: The configuration is as follows: 2 layers, 8 heads, FFN dimension 3072, Dropout=0.1.

[0046] In some embodiments, the expression for the sequence tag is: ; in, This represents the conditional probability of obtaining the label sequence Y given the input sequence X. Representation and Label The transpose of the corresponding learnable weight vector, This represents the feature vector at position i. This represents the transition score function between adjacent labels. Indicates the i-th A label in one location, This represents the label at the i-th position. This represents the normalization factor obtained by summing over all possible label sequences. This represents the input text or word sequence. This represents the label sequence corresponding to the input sequence.

[0047] The loss function of the hybrid neural network is a combined hard-example perceptual loss function, which integrates two mechanisms: Conditional Random Field (CRF) and Weighted Focal Loss. The former models the transition dependency of the label sequence, while the latter focuses on learning difficult samples (such as low-frequency trigger words). By linearly combining these mechanisms, a balance is achieved between structural consistency and long-tail robustness, thereby improving the overall detection accuracy and generalization ability.

[0048] In some embodiments, the processor can define a tag transition matrix for the CRF. ,in Indicates from the label Transferred to The score. Calculate the score for the entire label sequence. The overall path score; the entire CRF layer is optimized by maximizing the conditional probability of the true label path relative to all possible paths.

[0049] It can effectively guide the model to learn the globally optimal label path, and at the same time, the Viterbi algorithm can efficiently decode the path with the maximum score during the inference stage, ensuring the structural legality and label consistency of the output sequence.

[0050] In some embodiments, the expression for the loss function of a hybrid neural network is: ; ; ; ; in, Represents the total loss function. This represents the negative log-likelihood loss of a conditional random field layer. Indicates the weighting coefficient. Denotes the auxiliary loss function. Represents the category weight coefficient. This represents the probability that position i is predicted to be label c. Indicates the focus adjustment parameter, This represents the true label sequence corresponding to a given input sequence. Indicates the true label, Let represent the indicator function, where i represents the position index in the sequence, c represents the label category index, and n represents the total length of the labels in the sequence. Represents the set of all possible label sequences. express A candidate label sequence in the, This represents the sequence scoring function computed for a given feature representation U and label sequence. This represents the set of feature representations of the entire input sequence obtained by a hybrid neural network. Represents the set of real labels. This represents the emission score vector at the i-th position. Indicates from label y i-1 Move to label The transfer score.

[0051] Conditional Random Fields (CRFs) are discriminative graphical models used for sequence labeling. They can model transition dependencies between labels and are often used as decoding layers in tasks such as named entity recognition and event detection.

[0052] Focal loss is a loss function used to address class imbalance problems. By adjusting the loss weights for easy and difficult samples, it effectively improves the model's ability to identify minority or difficult samples.

[0053] S6: Utilize the trained hybrid neural network to analyze the biomedical event extraction task, obtain information extraction results, and complete the information extraction.

[0054] The information extraction result is the BIO tag prediction result for each token. .

[0055] Based on the co-occurrence frequency and semantic similarity among trigger word types in the training set, a tag co-occurrence topology is constructed. A non-independent Markov chain sampling strategy is introduced to dynamically sample representative and diverse tag combinations. Compared with traditional independent sampling methods, this mechanism can more effectively cover the low-frequency tag space, guiding subsequent generative models to focus on the co-occurrence dependencies and semantic consistency among long-tail trigger word types, significantly improving the targeting and effectiveness of data augmentation. Based on the sampled tag combinations, this invention uses a pre-trained large language model as a generator to generate high-quality sentences that meet the requirements of the target trigger word combination under structured prompts. To address the semantic ambiguity of trigger words in different contexts, this invention introduces a context dynamic modeling module based on a gating mechanism. This module integrates trigger word type, entity position, and associated semantics in the context representation, and dynamically controls the attention weights of different context fragments to the target word. This module effectively models the changing correlation between event trigger words and their syntactic and semantic dependent contexts. In the decoding stage, a combined optimization strategy of Conditional Random Field (CRF) and Focal Loss is adopted. By explicitly modeling the transition structure relationship between BIO tags, the syntactic validity of the output sequence is guaranteed. At the same time, Focal Loss is used to strengthen the training focus on hard-to-classify samples (such as low-frequency trigger words), thereby improving the robustness and accuracy of the model in low-sample scenarios.

[0056] In some embodiments of this specification, an information extraction method based on topological graph non-independent sampling and context association learning is provided. The low-frequency label non-independent sampling mechanism based on the topological graph constructs a label co-occurrence graph based on the co-occurrence frequency and semantic distance of trigger words in the training data. Using this as a priori, a large language model is guided to generate new samples that satisfy specific label combinations, thereby significantly expanding the semantic distribution space of low-frequency trigger events in the training set. The constructed dynamic context modeling module realizes context dependency modeling under different trigger types by fusing trigger word type, entity location, and context representation. Unlike the unified attention mechanism of traditional Transformers, the DCA module dynamically adjusts the contextual focus region based on different labels, effectively identifying long-distance dependencies and ambiguous expressions, and significantly improving the model's ability to distinguish trigger words in complex semantic structures and its generalization ability. During the decoding stage, a Conditional Random Field (CRF) is introduced to model the transition relationships between BIO labels, enhancing the consistency of the output sequence. Simultaneously, a Focal Loss mechanism is used to assign higher weights to low-confidence or hard-to-classify samples, thereby strengthening the model's ability to learn low-frequency labels. By introducing a label combination sampling and entity consistency check mechanism based on a topology graph, the large model is guided to generate only within the legal label subspace, avoiding invalid combinations and semantic drift. Furthermore, biomedical semantic style constraints are imposed on the generated text (such as preserving event verbs and normalizing entity representations) to ensure that the generated results maintain a high degree of consistency with the original data in style, semantics, and structure.

Claims

1. An information extraction method based on topological graph non-independent sampling and contextual association learning, characterized in that, include: S1: Based on the biomedical event extraction task, data extraction is performed to obtain raw sample data; S2: Based on the trigger word type of the original sample data, the set of low-frequency trigger word pairs and the candidate set of trigger words are obtained through calculation; S3: Utilize context association learning to extract the context representations of the low-frequency trigger word pair set and the trigger word candidate set, and obtain the corresponding low-frequency trigger word pair representation vector and trigger word candidate representation vector through average pooling; S4: Calculate the transition probability using the low-frequency trigger word pair representation vector and the candidate trigger word representation vector, and obtain the enhanced training set through non-independent sampling of the topological graph; S5: Use the augmented training set to train the hybrid neural network to obtain a trained hybrid neural network; S6: Utilize the trained hybrid neural network to analyze the biomedical event extraction task, obtain information extraction results, and complete the information extraction.

2. The information extraction method based on topological graph non-independent sampling and contextual association learning according to claim 1, characterized in that, S4 includes: The transition probability is obtained by using the low-frequency trigger word pair representation vector and the trigger word candidate representation vector. Based on the transition probability, seed samples are obtained by performing Markov chain sampling on the graph through non-independent sampling of the topological graph. Based on seed samples, new sentences and labels are generated using a large language model to obtain an augmented dataset; By filtering the augmented dataset, an augmented training set is obtained.

3. The information extraction method based on topological graph non-independent sampling and contextual association learning according to claim 1, characterized in that, The expression for the transition probability is: ； in, This indicates that the node represents the node. Transition to node The transition probability, This represents the source node in the trigger word topology graph. This represents a directed edge from the source node to the target node. This represents the target node in the trigger word topology graph. Indicates the control ratio. This represents a similarity function based on trigger word type. Represents a node The corresponding trigger word type, Represents a node The corresponding trigger word type, The representation is a similarity function based on context representation. Represents the node The corresponding context feature vector, Represents the node The corresponding context feature vector.

4. The information extraction method based on topological graph non-independent sampling and context association learning according to claim 1, characterized in that, The hybrid neural network includes: The encoding layer is used to encode the augmented training set to obtain the context representation of each token; The dynamic context association module is used to process the context representation of each token using a self-attention mechanism to obtain a context-enhanced representation with discriminative capabilities. The Transformer layer is used to process the context-enhanced representation to obtain a globally context-enhanced token representation; The Conditional Random Field layer is used to process the globally context-enhanced token representation using label scores and label transitions to obtain sequence labels; The prediction layer is used for analysis based on sequence labels to obtain information extraction results.

5. The information extraction method based on topological graph non-independent sampling and context association learning according to claim 4, characterized in that, The expression for the context-enhanced representation is: ； ； ； in, This represents the attention weight of the i-th target word to the j-th context word under trigger word type t. This represents the transpose of the attention weight vector. This represents the context representation vector of the i-th word in the input sequence. Let represent the context representation vector of the j-th word in the input sequence. , and Represents the learnable projection matrix. The type embedding vector represents the trigger word type t of the original sample data, k represents the total number of trigger word types, and n represents the number of words in the current sequence.

6. The information extraction method based on topological graph non-independent sampling and context association learning according to claim 4, characterized in that, The expression for the sequence tag is: ； in, This represents the conditional probability of obtaining the label sequence Y given the input sequence X. Representation and Label The transpose of the corresponding learnable weight vector, This represents the feature vector at position i. This represents the transition score function between adjacent labels. Indicates the i-th A label in one location, This represents the label at the i-th position. This represents the normalization factor obtained by summing over all possible label sequences. This represents the input text or word sequence. This represents the label sequence corresponding to the input sequence.

7. The information extraction method based on topological graph non-independent sampling and contextual association learning according to claim 4, characterized in that, The expression for the loss function of the hybrid neural network is: ； ； ； ； in, Represents the total loss function. This represents the negative log-likelihood loss of a conditional random field layer. Indicates the weighting coefficient. Denotes the auxiliary loss function. Represents the category weight coefficient. This represents the probability that position i is predicted to be label c. Indicates the focus adjustment parameter, This represents the true label sequence corresponding to a given input sequence. Indicates the true label, Let represent the indicator function, where i represents the position index in the sequence, c represents the label category index, and n represents the total length of the labels in the sequence. Represents the set of all possible label sequences. express A candidate label sequence in the, This represents the sequence scoring function computed for a given feature representation U and label sequence. This represents the set of feature representations of the entire input sequence obtained by a hybrid neural network. Represents the set of real labels. This represents the emission score vector at the i-th position. Indicates from label y i-1 Move to label The transfer score.

Images