Document-level event relation extraction method based on hybrid expert model and contrastive learning

Abstract

The present application belongs to the technical field of natural language processing and information extraction, and particularly relates to a document-level event relation extraction method based on a hybrid expert model and contrastive learning. In view of the problems of feature confusion and non-relation class bias in the unified modeling of multiple types of events in the existing method, the present application first encodes the input document, obtains the context representation of the event mention, and constructs event pair representation and interaction features. Secondly, the event pair is assigned to multiple expert models through a dynamic routing mechanism, each expert model models from the perspectives of semantic features, structural features and bilinear interaction features, and the output results are weighted and fused to obtain the initial relation prediction. Further, a contrastive learning discriminator is introduced to model whether there is an effective relation between the event pair. Finally, in the reasoning stage, the non-relation class is dynamically adjusted by combining the prediction calibration strategy and the adaptive negative class penalty mechanism.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and information extraction technology, specifically relating to a document-level event relationship extraction method based on hybrid expert models and contrastive learning. Background Technology

[0002] Document-level event relation extraction aims to identify semantic relationships between event mentions in unstructured text, including coreference, temporal relationships, causal relationships, and sub-event relationships. This task plays an important role in downstream applications such as information extraction, event graph construction, and complex reasoning.

[0003] Existing methods mostly model single-relation subtasks, relying on the semantic features of specific relations, making it difficult to extend to unified multi-relation modeling scenarios. Furthermore, document-level modeling typically requires capturing long-distance dependencies across sentences. While existing methods introduce graph structures or logical reasoning mechanisms, they generally employ a static modeling paradigm with shared parameters, easily leading to feature confusion between different fine-grained relations. Simultaneously, event relations exhibit an extremely sparse distribution, with no-relation categories dominating, making the model prone to severe class bias during training, thus affecting the ability to identify effective relations. Therefore, how to effectively model multi-type event relations within a unified framework and alleviate the bias problem caused by severe class imbalance has become a key challenge in document-level event relation extraction. Summary of the Invention

[0004] To address the problems of feature confusion and unrelated category bias in the unified modeling of multi-type event relationships in existing methods, this invention provides a document-level event relationship extraction method based on hybrid expert models and contrastive learning.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A document-level event relation extraction method based on a hybrid expert model and contrastive learning includes the following steps:

[0007] Step 1: Encode the input document to obtain a single event representation, construct multiple interaction features of event pairs, and extract the structural features of event pairs.

[0008] Step 2: Based on the interaction characteristics of the event pairs, calculate the expert weights through the router and select the corresponding experts using the Top-K strategy;

[0009] Step 3: Input the interaction features and structural features of the event pairs into multiple expert modules. Each expert models the relationship based on semantic features, structural features, and bilinear interaction features, and outputs the relationship prediction score.

[0010] Step 4: Perform weighted aggregation on the output results of the selected experts to obtain the initial relationship prediction results;

[0011] Step 5: Introduce a contrastive learning discriminator to model whether there is a valid relationship between event pairs;

[0012] Step 6: Combine the output of the contrastive learning discriminator to perform prediction calibration, introduce an adaptive negative class penalty mechanism to adjust the unrelated categories, and output the final event relationship prediction result.

[0013] Furthermore, step 1 specifically includes:

[0014] Step 1.1, process the input document Encode to obtain word-level hidden representations Regarding the incident mentioned and Average pooling is performed on the lexical hidden representations within each span to obtain single-event representations. and ;

[0015] Step 1.2, for any event pair Construct various interaction features for event pairs, including simple splicing features. Differential feature representation and element-wise interaction features , respectively represented as:

[0016] ;

[0017] ;

[0018] ;

[0019] in, This represents the element-wise multiplication operation;

[0020] Step 1.3: Extract the structural features of event pairs It originates from the structural information of event pairs, including discrete features such as sentence distance between events, relative position and event type, and is represented as a vector through embedding mapping.

[0021] Furthermore, step 2 specifically includes:

[0022] Step 2.1, given the element-wise interaction features For each type of relation The router calculates the initial route weight distribution through two layers of linear transformation and activation function. The calculation formula is:

[0023] ;

[0024] in, ReLU represents the activation function. and For relation types The routing parameter matrix;

[0025] Step 2.2, using a hard routing mechanism, set... For the initial route with the highest weight A set of expert indexes, only activate the set. The experts in the pool are identified, and their weights are renormalized to obtain the final activated expert weights. :

[0026] ;

[0027] in, Representing relation type Next The original routing weights of each expert, This indicates the activation of the expert index set. Represents a set Expert index in; This represents the normalized expert routing weight.

[0028] Furthermore, step 3 specifically includes:

[0029] Step 3.1, for the activated first The expert group comprises three functionally differentiated branches: the semantic branch, the structural branch, and the bilinear branch.

[0030] Utilizing simple splicing features Entering the semantic branch, fine-grained contextual information is encoded to obtain semantic features. :

[0031] ;

[0032] in, and These represent a multilayer perceptron with residual connections and layer normalization, respectively.

[0033] Structural features of event pairs Input structural branches to enhance the structure awareness capability in relation discrimination and obtain structural features. :

[0034] ;

[0035] in, It is a learnable scalar parameter used to adjust the contribution of structural branches;

[0036] Represented by a single event and Input a bilinear branch to characterize the multiplicative interaction and obtain bilinear features. :

[0037] ;

[0038] in, and The weight matrix that maps event representations to a low-rank interaction space;

[0039] Step 3.2, semantic features Structural features and bilinear features The concatenated data is fed into a fusion network to obtain a fused representation. :

[0040] ;

[0041] The fusion representation is mapped to a specific relation type. Preliminary prediction score :

[0042] ;

[0043] in, A mapping matrix for a specific task.

[0044] Furthermore, step 4 specifically includes:

[0045] For each relation type The multi-expert rating system is based on the obtained activated expert weights. Preliminary prediction scores of activated experts Perform weighted aggregation to generate event pairs Initial relationship prediction results :

[0046] .

[0047] Furthermore, step 5 specifically includes:

[0048] Step 5.1, for each relation type Two binary classifiers are introduced respectively. The binary classifier distinguishes between valid and unrelational categories. Difference feature representation based on event pairs Perform coarse-grained discrimination and represent the difference features. Extract discriminative hidden representations from the input multilayer perceptron with residual connections. The calculation formula is:

[0049] ;

[0050] Among them, ResMLP is a multilayer perceptron with residual connections;

[0051] Step 5.2, the discriminative hidden representation Input the contrast head network to generate contrastive embeddings for optimizing the embedding space distribution. :

[0052] ;

[0053] Among them, ConsHead is the comparison head network;

[0054] Step 5.3, the discriminative hidden representation The input discriminant head network, through linear mapping and activation function, outputs the probability confidence that a valid relationship exists between the event pairs. :

[0055] ;

[0056] in, and These represent the learnable classification weights and biases corresponding to the discriminator. This is the activation function.

[0057] Furthermore, step 6 specifically includes:

[0058] Step 6.1: Calculate the bias term based on the probability confidence level output by the discriminator. The calculation formula is:

[0059] ;

[0060] in, To control the hyperparameters of intervention intensity, A smoothing constant is used to prevent logarithmic operations from overflowing. This is an adaptive scaling factor;

[0061] Step 6.2, set the bias term Injected into the initial relation prediction results In the biased intervention, set For specific relationship category labels, the relationship prediction score after bias intervention. The calculation is as follows:

[0062] ;

[0063] in, Representing relation type The total number of all fine-grained categories, NONE indicates no relation;

[0064] Step 6.3, based on the post-intervention score Calculate the predicted probability distribution Obtain the maximum prediction probability and their corresponding prediction categories When the predicted category is unrelated and the confidence level is low, i.e. and When this occurs, the adaptive negative class penalty mechanism is triggered, and the penalty term is calculated. :

[0065] ;

[0066] in, The preset trigger threshold, Penalty scaling factor;

[0067] Step 6.4, the penalty item This is applied to the scores of unrelated categories to obtain the final calibrated scores for unrelated categories. :

[0068] ;

[0069] Step 6.5: After the relationship score is corrected by both bias intervention and adaptive negative class penalty, it is normalized by Softmax to obtain the final event relationship probability distribution. The category label with the highest probability is selected as the final event relationship prediction result for that event and output.

[0070] Compared with the prior art, the present invention has the following advantages:

[0071] This invention introduces a hybrid expert model to achieve dynamic decomposition and modeling of different relationship patterns, effectively alleviating the feature confusion problem between multiple types of relationships. At the same time, by combining a contrastive learning discriminator and an adaptive negative class penalty mechanism in the inference stage, it significantly alleviates the prediction bias problem caused by class imbalance, improves the model's ability to identify sparse and effective relationships, and thus enhances the overall performance and robustness of document-level event relationship extraction. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the overall model architecture of the present invention. Detailed Implementation

[0073] The document-level event relationship extraction method based on hybrid expert models and contrastive learning described in this invention is implemented through a computer program. The specific implementation methods of the technical solution proposed in this invention will be detailed below according to the workflow.

[0074] like Figure 1 As shown in the figure, a document-level event relationship extraction method based on a hybrid expert model and contrastive learning in this embodiment includes the following steps:

[0075] Step 1 involves encoding the input document to obtain a single-event representation, constructing multiple interaction features of event pairs, and extracting structural features of the event pairs. Specifically, this includes:

[0076] Step 1.1, process the input document Encode to obtain word-level hidden representations Regarding the incident mentioned and Average pooling is performed on the lexical hidden representations within each span to obtain single-event representations. and ;

[0077] Step 1.2, for any event pair Construct various interaction features for event pairs, including simple splicing features. Differential feature representation and element-wise interaction features The simple concatenation features are used for basic semantic modeling, the difference features are used by the contrastive learning discriminator, and the element-wise interaction features are used to mix the expert module's routing decisions and expert scores. These are formed by concatenating the above components by calculating the interaction terms between single event representations, encompassing individual event representations, element-wise products, and absolute differences. The various interaction features are represented as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] in, This represents the element-wise multiplication operation;

[0082] Step 1.3, Structural Features of Event Pairs The structural information derived from event pairs, including discrete features such as sentence distance, relative position, and event type between events, is mapped to vector representations through embedding to construct the structural features of event pairs. .

[0083] Step 2: Based on the interaction characteristics of the event pairs, calculate the expert weights using a router, and select the corresponding expert using a Top-K strategy. Specifically, this includes:

[0084] Step 2.1, given the element-wise interaction features For each type of relation The router calculates the initial route weight distribution through two layers of linear transformation and activation function. The calculation formula is:

[0085] ;

[0086] in, ReLU represents the activation function. and For relation types The routing parameter matrix;

[0087] Step 2.2, to encourage expert functional differentiation, a hard routing mechanism is adopted, setting... For the initial route with the highest weight A set of expert indices; the model only activates the set. The experts in the pool are identified, and their weights are renormalized to obtain the final activated expert weights. :

[0088] ;

[0089] in, Representing relation type Next The original routing weights of each expert, This indicates that the expert index set is activated. Represents a set Expert index in the database; This represents the normalized expert routing weight.

[0090] Step 3 involves inputting the interaction and structural features of the event pairs into multiple expert modules. Each expert models the relationships based on semantic features, structural features, and bilinear interaction features, and outputs a relationship prediction score. Specifically, this includes:

[0091] Step 3.1, for the activated first The expert group comprises three functionally differentiated branches: the semantic branch, the structural branch, and the bilinear branch.

[0092] First, utilize simple splicing features Entering the semantic branch, fine-grained contextual information is encoded to obtain semantic features. :

[0093] ;

[0094] in, and These represent a multilayer perceptron with residual connections and layer normalization, respectively.

[0095] Secondly, the structural features of event pairs Input structural branches to enhance the structure awareness capability in relation discrimination and obtain structural features. :

[0096] ;

[0097] in, It is a learnable scalar parameter used to adjust the contribution of structural branches;

[0098] Finally, use a single event representation. and Input a bilinear branch to characterize the multiplicative interaction and obtain bilinear features. :

[0099] ;

[0100] in, and The weight matrix that maps event representations to a low-rank interaction space;

[0101] Step 3.2, semantic features Structural features and bilinear characteristics The concatenated data is fed into a fusion network to obtain a fused representation. :

[0102] ;

[0103] The fusion representation is mapped to a specific relation type. Preliminary prediction score :

[0104] ;

[0105] in, A mapping matrix for a specific task.

[0106] Step 4 involves weighted aggregation of the selected experts' outputs to obtain the initial relationship prediction results. This includes:

[0107] For each relation type The multi-expert rating system uses the activated expert weights obtained in step 2.2. Preliminary prediction scores of activated experts Perform weighted aggregation to generate event pairs Initial relationship prediction results :

[0108] .

[0109] Step 5: Introduce a contrastive learning discriminator to model whether there is a valid relationship between event pairs. This specifically includes:

[0110] Step 5.1, to distinguish between valid relationships and non-relationship (NONE) categories, for each relationship type... (Except for the coreference resolution task), a specific binary classifier is introduced for each task. Unlike fine-grained hybrid expert scorers, the binary classifier... Difference feature representation based on event pairs Perform coarse-grained discrimination and represent the differential features from step 1.2. Extract discriminative hidden representations from the input multilayer perceptron (ResMLP) with residual connections. The calculation formula is:

[0111] ;

[0112] Among them, ResMLP is a multilayer perceptron with residual connections;

[0113] Step 5.2, the discriminative hidden representation Inputting a contrastive head network generates contrastive embeddings for optimizing the embedding space distribution. :

[0114] ;

[0115] Among them, ConsHead is the comparison head network;

[0116] Contrastive embeddings are optimized during the training phase using supervised contrastive loss, which encourages event pairs with real relationships to cluster in the representation space and separate them from irrelevant event pairs.

[0117] Step 5.3, the discriminative hidden representation The input is a discriminator head network, which, through linear mapping and activation functions, outputs the probability confidence level that a valid relationship (i.e., a non-NONE relationship) exists between the event pairs. :

[0118] ;

[0119] in, and These represent the learnable classification weights and biases corresponding to the discriminator. This is the activation function.

[0120] Step 6: In the re-inference stage, the output of the contrastive learning discriminator is combined for prediction calibration, and an adaptive negative class penalty mechanism is introduced to adjust unrelated categories, outputting the final event relationship prediction result. Specifically, this includes:

[0121] Step 6.1: Calculate the bias term based on the probability confidence level output by the discriminator. The calculation formula is:

[0122] ;

[0123] in, To control the hyperparameters of intervention intensity, A smoothing constant is used to prevent logarithmic operations from overflowing. This is an adaptive scaling factor;

[0124] Step 6.2, set the bias term Injected into the initial relation prediction results obtained in step 4 In the biased intervention, set For specific relationship category labels, the relationship prediction score after bias intervention. The calculation is as follows:

[0125] ;

[0126] in, Representing relation type The total number of all fine-grained categories, NONE indicates no relation;

[0127] Step 6.3, based on the post-intervention score Calculate the predicted probability distribution Obtain the maximum prediction probability and their corresponding prediction categories When the predicted category is unrelated and the confidence level is low, i.e. and When this occurs, the adaptive negative class penalty mechanism is triggered, and the penalty term is calculated. :

[0128] ;

[0129] in, The preset trigger threshold, Penalty scaling factor;

[0130] Step 6.4, the penalty item This is applied to the scores of unrelated categories to obtain the final calibrated scores for unrelated categories. :

[0131] ;

[0132] Step 6.5: After the relationship score is corrected by both bias intervention and adaptive negative class penalty, it is normalized by Softmax to obtain the final event relationship probability distribution. The category label with the highest probability is selected as the final event relationship prediction result for that event and output.

[0133] Example 2

[0134] To verify the effectiveness of this method, experiments were conducted on the MAVEN-ERE, EventStoryLine, and HiEve datasets. Specific dataset information is shown in Table 1.

[0135] Table 1. Dataset Statistics

[0136] Number of documents 4480 258 100 Number of events 103193 4732 2734 Number of time-series relations 1216217 8111 - Quantity of causal relationships 57992 4584 - Number of sub-event relationships 15841 - 3648 training set 2913 233 60 Validation set 710 25 20 test set 857 - 20

[0137] Report micro-averaged Precision, Recall, and F1 scores for temporal, causal, and sub-event relationships. For event coreference resolution tasks, the average F1 score of four evaluation strategies is used as the final result: MUC, B... 3 , CEAFe and BLANC.

[0138] The experimental results are shown in Tables 2 and 3. The proposed method demonstrates significant performance advantages across multiple datasets and relation types. On the MAVEN-ERE dataset, the method achieves leading or near-optimal F1 scores in four relation types: coreference, temporal, causal, and sub-event. The overall performance reaches 55.3, showing a stable improvement over existing representative methods (such as RoBERTa-base, ProtoERE, and LLMERE), especially in complex semantic modeling tasks involving causal and sub-event relations, demonstrating the model's powerful ability in fine-grained relation discrimination. Furthermore, compared to few-shot methods based on large language models (such as Qwen3 and DeepSeek), the proposed method significantly outperforms other methods across all metrics, indicating stronger stability and generalization ability in supervised learning scenarios.

[0139] Table 2 Experimental results on MAVEN-ERE

[0140] ERGO 89.3 51.2 28 21.8 47.6 RoBERTa-base 91.8 55.5 31.5 24.8 50.9 ProtoERE 89.8 53.8 31.8 27.9 50.8 TacoERE - - 34.1 30.6 - MAQInstruct 91.3 53.8 32.5 25.2 - GraphERE - 54.7 31.1 27.3 - Logie - 55 - 30.3 - LLMERE 90.9 54.7 36 28.2 52.5 Qwen3 66.3 21 7.9 6.7 25.5 DeepSeek 65.1 24.6 7.5 5.7 25.7 Ours (joint) 91.6 59 36.7 33.8 55.3

[0141] Table 3 Experimental results of EventStoryLine and HiEve

[0142] Qwen3(10-shot) 14.8 1.6 2.9 9.8 8.8 9.3 DeepSeek (10-shot) 5.0 8.3 6.3 7.1 14.8 9.6 Hierarchical 30.1 10.2 13.1 21.4 17.3 16.7 RoBERTa-base 30.6 11.1 14.8 13.7 22.8 17.1 SIEF 32.4 11.3 14.8 21.8 17.4 18.6 SCS-EERE 32.7 10.9 15.1 20.6 19.7 19.2 TacoERE 32.9 12.3 16.4 22.6 19.5 20.8 Ours 27.5 17.9 20.9 20.2 22.9 21.3

[0143] Furthermore, in the EventStoryLine causal relationship extraction task, the F1 score of this invention reaches 20.9, significantly outperforming existing methods; in the HiEve sub-event relationship extraction task, the F1 score reaches 21.3, also achieving optimal performance. Notably, this invention shows a particularly significant improvement in recall, indicating its ability to more effectively identify sparse and difficult-to-capture valid relationships, validating the effectiveness of the proposed contrastive learning discriminator and adaptive negative class penalty mechanism in mitigating class imbalance.

[0144] Experimental results across different datasets and relation types demonstrate that this invention achieves dynamic decoupling modeling of multiple relation patterns through a hybrid expert model, significantly reducing feature interference between different relations. Simultaneously, by combining contrastive learning and inference stage calibration strategies, it effectively improves the identification ability of long-tail and complex relations, thereby improving the overall accuracy, recall, and F1 score of document-level event relation extraction tasks. Therefore, this invention not only exhibits superior performance on a single dataset but also demonstrates good generalization and stability in cross-task and cross-dataset scenarios, possessing strong practical application value and promotional significance.

[0145] The foregoing has shown and described the main features and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0146] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A document-level event relationship extraction method based on a hybrid expert model and contrastive learning, characterized in that, Includes the following steps: Step 1: Encode the input document to obtain a single event representation, construct multiple interaction features of event pairs, and extract the structural features of event pairs. Step 2: Based on the interaction characteristics of the event pairs, calculate the expert weights through the router and select the corresponding experts using the Top-K strategy; Step 3: Input the interaction features and structural features of the event pairs into multiple expert modules. Each expert models the relationship based on semantic features, structural features, and bilinear interaction features, and outputs the relationship prediction score. Step 4: Perform weighted aggregation on the output results of the selected experts to obtain the initial relationship prediction results; Step 5: Introduce a contrastive learning discriminator to model whether there is a valid relationship between event pairs; Step 6: Combine the output of the contrastive learning discriminator to perform prediction calibration, introduce an adaptive negative class penalty mechanism to adjust the unrelated categories, and output the final event relationship prediction result.

2. The document-level event relationship extraction method based on hybrid expert models and contrastive learning according to claim 1, characterized in that, Step 1 specifically involves: Step 1.1, process the input document Encode to obtain word-level hidden representations Regarding the incident mentioned and Average pooling is performed on the lexical hidden representations within each span to obtain single-event representations. and ; Step 1.2, for any event pair Construct various interaction features for event pairs, including simple splicing features. Differential feature representation and element-wise interaction features , respectively represented as: ； ； ； in, This represents the element-wise multiplication operation; Step 1.3: Extract the structural features of event pairs. It originates from the structural information of event pairs, including discrete features such as sentence distance between events, relative position and event type, and is represented as a vector through embedding mapping.

3. The document-level event relationship extraction method based on hybrid expert models and contrastive learning according to claim 2, characterized in that, Step 2 specifically involves: Step 2.1, given the element-wise interaction features For each type of relation The router calculates the initial route weight distribution through two layers of linear transformation and activation function. The calculation formula is: ； in, ReLU represents the activation function. and For relation types The routing parameter matrix; Step 2.2, using a hard routing mechanism, set... For the initial route with the highest weight A set of expert indexes, only activate the set. The experts in the pool are identified, and their weights are renormalized to obtain the final activated expert weights. : ； in, Representing relation type Next The original routing weights of each expert, This indicates the activation of the expert index set. Represents a set Expert index in; This represents the normalized expert routing weight.

4. The document-level event relationship extraction method based on hybrid expert models and contrastive learning according to claim 3, characterized in that, Step 3 specifically involves: Step 3.1, for the activated first The expert group comprises three functionally differentiated branches: the semantic branch, the structural branch, and the bilinear branch. Utilizing simple splicing features Entering the semantic branch, fine-grained contextual information is encoded to obtain semantic features. : ； in, and These represent a multilayer perceptron with residual connections and layer normalization, respectively. Structural features of event pairs Input structural branches to enhance the structure awareness in relation discrimination and obtain structural features. : ； in, It is a learnable scalar parameter used to adjust the contribution of structural branches; Represented by a single event and Input a bilinear branch to characterize the multiplicative interaction and obtain bilinear features. : ； in, and The weight matrix that maps event representations to a low-rank interaction space; Step 3.2, semantic features Structural features and bilinear characteristics The concatenated data is fed into a fusion network to obtain a fused representation. : ； The fusion representation is mapped to a specific relation type. Preliminary prediction score : ； in, A mapping matrix for a specific task.

5. The document-level event relationship extraction method based on hybrid expert models and contrastive learning according to claim 3, characterized in that, Step 4 specifically involves: For each relation type The multi-expert rating system is based on the obtained activated expert weights. Preliminary prediction scores of activated experts Perform weighted aggregation to generate event pairs Initial relationship prediction results : 。 6. The document-level event relationship extraction method based on hybrid expert models and contrastive learning according to claim 5, characterized in that, Step 5 specifically involves: Step 5.1, for each relation type Two-class discriminants are introduced respectively. The binary classifier distinguishes between valid and unrelational categories. Difference feature representation based on event pairs Perform coarse-grained discrimination and represent the differential features. Extract discriminative hidden representations from the input multilayer perceptron with residual connections. The calculation formula is: ； Among them, ResMLP is a multilayer perceptron with residual connections; Step 5.2, the discriminative hidden representation Input the contrast head network to generate contrastive embeddings for optimizing the embedding space distribution. : ； Among them, ConsHead is the comparison head network; Step 5.3, the discriminative hidden representation The input discriminant head network, through linear mapping and activation function, outputs the probability confidence that a valid relationship exists between the event pairs. : ； in, and These represent the learnable classification weights and biases corresponding to the discriminator. This is the activation function.

7. The document-level event relationship extraction method based on hybrid expert models and contrastive learning according to claim 6, characterized in that, Step 6 specifically involves: Step 6.1: Calculate the bias term based on the probability confidence level output by the discriminator. The calculation formula is: ； in, To control the hyperparameters of intervention intensity, A smoothing constant is used to prevent logarithmic operations from overflowing. This is an adaptive scaling factor; Step 6.2, set the bias term Injected into the initial relation prediction results In the biased intervention, set For specific relationship category labels, the relationship prediction score after bias intervention. The calculation is as follows: ； in, Representing relation type The total number of all fine-grained categories; NONE indicates no relation. Step 6.3, based on the post-intervention score Calculate the predicted probability distribution Obtain the maximum prediction probability and their corresponding prediction categories When the predicted category is unrelated and the confidence level is low, i.e. and When this occurs, the adaptive negative class penalty mechanism is triggered, and the penalty term is calculated. : ； in, The preset trigger threshold, Penalty scaling factor; Step 6.4, the penalty item This is applied to the scores of unrelated categories to obtain the final calibrated scores for unrelated categories. : ； Step 6.5: After the relationship score is corrected by both bias intervention and adaptive negative class penalty, it is normalized by Softmax to obtain the final event relationship probability distribution. The category label with the highest probability is selected as the final event relationship prediction result for that event and output.

Images