Event prediction method based on pair event relation learning and multi-view evidence fusion

Abstract

The application relates to an event prediction method based on pair event relation learning and multi-view evidence fusion, and belongs to the technical field of natural language processing. The method comprises the following steps: pair event relation learning: a pre-training language model is trained by constructing positive and negative samples in a supervised learning framework; an event relation graph is constructed: the pre-training language model after pair event relation learning is used to deduce the dependency relation of the pair event, the probability output by the model is used as the weight of the edge, and thus the event relation graph is constructed; multi-view evidence learning: quantitative evidence of a candidate event is learned from a text semantic view and a graph structure view; credible evidence fusion: the uncertainty of each view is modeled by using a Dirichlet distribution, and multi-view evidence is dynamically fused by using Dempster-Shafer theory to generate a final prediction result. Under the condition of not depending on an external knowledge base, the application realizes a prediction accuracy of 64.62% on an NYT data set, and exceeds an existing baseline model.

Description

Technical Field

[0001] This invention relates to an event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion, applicable to downstream tasks such as narrative understanding, public opinion analysis, and question answering systems, and belongs to the field of natural language processing technology. Background Technology

[0002] Event prediction, based on a given event chain, understands the evolutionary logic of events by capturing the inherent relationships between them, thereby selecting the most reasonable subsequent events from a set of candidate events. Event prediction can provide strong support for downstream tasks such as public opinion analysis and question-answering systems. However, in scenarios where event information is sparse, how to infer the implicit relationships between events and accurately predict subsequent events remains a pressing technical challenge.

[0003] Currently, mainstream event prediction methods can be divided into the following three categories:

[0004] 1. Based on the principle of statistical co-occurrence: This method quantifies the correlation between events by calculating the co-occurrence frequency of event arguments and models event interactions based on event pairs or event graphs. However, this approach ignores deep semantic information, often introducing noise and false associations. Furthermore, its generalization ability is weak, only capturing high-frequency, common event relationship patterns, and its ability to identify long-tail, sparse event patterns is significantly insufficient.

[0005] 2. This type of method provides additional common-sense support for event prediction tasks by integrating event relationships contained in external knowledge bases, thereby improving the model's prediction performance to some extent. However, external knowledge bases typically encode static and generalized background knowledge, which is difficult to adapt to the dynamic evolution characteristics of specific narrative contexts. Therefore, when facing complex and ever-changing real-world scenarios, it is easy to introduce noise information that is inconsistent with the context, which in turn interferes with the model's accurate judgment of event logic.

[0006] 3. Methods based on pre-trained language models: These methods primarily utilize the deep semantic understanding capabilities of pre-trained language models, employing different fine-tuning strategies to explore the model's potential in understanding event relationships. However, these methods lack explicit integration of complex relationship graphs, which limits the model's ability to capture global narrative logic.

[0007] In summary, existing technologies have the following technical problems in event prediction tasks:

[0008] The complexity of constructing inter-event dependencies: Accurate characterization of inter-event dependencies requires a deep integration of semantic understanding and common-sense reasoning, and is difficult to obtain directly through shallow grammar or co-occurrence patterns.

[0009] Noise interference problem: Event dependency modeling is susceptible to noise interference, leading to systematic biases in prediction results. Traditional methods have inherent limitations in noise suppression, lacking effective robustness mechanisms and struggling to distinguish between true dependencies and random co-occurring noise, thus reducing the practical usability of the prediction model. Summary of the Invention

[0010] To address the aforementioned issues, this invention provides an event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion, which significantly improves the accuracy of event association.

[0011] The technical solution of this invention is: an event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion, comprising:

[0012] S1. Paired event relationship learning: A supervised learning framework is adopted to train a pre-trained language model by constructing positive and negative samples to distinguish between relevant and irrelevant event pairs;

[0013] S2. Constructing an event relationship graph: Using a pre-trained language model learned from pairwise event relationships, infer the dependencies between pairwise events, and use the probabilities output by the model as the weights of the edges to construct an event relationship graph;

[0014] S3, Multi-perspective evidence learning: Learn quantitative evidence for candidate events from both textual semantic and graph structure perspectives;

[0015] S4. Credible Evidence Fusion: The uncertainty of each perspective is modeled using the Dirichlet distribution, and the evidence from multiple perspectives is dynamically fused through the Dempster-Shafer theory to generate the final prediction result.

[0016] Further, S1 includes:

[0017] S1.1 Constructing positive and negative samples: Positive samples are defined as event pairs with semantic or temporal correlation, coming from the same event chain; negative samples are defined as unrelated event pairs, coming from different event chains.

[0018] S1.2 Constructing a structured input sequence: Constructing a structured input sequence for each event pair;

[0019] S1.3, Paired Event Relationship Inference and Model Optimization: Input the constructed sequence into the pre-trained language model, to... <s>Bit output vector As an aggregate representation of the entire input, then Input classifier; the process is described as follows:

[0020]

[0021] Where C represents the context event, and For the target event pairs that require relation inference. This represents the construction of a structured input sequence, with the classifier weight matrix being... The bias term is K is the number of categories, which are set as relevant / irrelevant, K=2. The loss function is the cross-entropy loss function commonly used in classification tasks, and the goal is to make the probability distribution... Fit the true labels.

[0022] Further, S2 includes:

[0023] Step 2.1, Paired Event Relationship Inference: Input event pairs into the pre-trained language model to obtain the probability distribution of relevance / irrelevance. =[ , In this probability distribution, the relevant probability is... The probability of being unrelated is , ;

[0024] Step 2.2, Event Relationship Graph Construction: Based on Relevant / Unrelated Probability Distributions =[ , Construct a weighted directed graph where the edge weights are determined by a probability distribution, described as follows:

[0025] ;

[0026] in, This represents a probability threshold; if the probability of association between two events exceeds a certain threshold... Then, connecting edges are established between them.

[0027] Further, S3 includes:

[0028] Step 3.1, Text Semantic Perspective Learning: Concatenate the event chain and candidate events into a sequence, and obtain event representations through a pre-trained language model; design an attention pooling module, which realizes the attention aggregation between candidate events and historical events; use the candidate event itself as the query, and the event chain as the key and value, and calculate the weighted aggregate representation of the event chain through the attention mechanism;

[0029]

[0030]

[0031] in, The weighted aggregation representation of the event chain. Candidate events The hidden state, The hidden state of candidate events K and V are the hidden states of the event chain. n is the length of the event chain;

[0032] Finally, the weighted aggregate representation of the event chain ah and the candidate events are concatenated and fed into the classifier. The scalar output of the classifier is used as the relevance score of the candidate events.

[0033]

[0034] in, This represents the hidden state of a candidate event. , , , , The dimension of the hidden state. This indicates vector concatenation;

[0035] Step 3.2, Graph Structure Perspective Learning: Based on the event relationship graph, a graph convolutional network is used to update the event representation and calculate the relevance score; the graph structure perspective adds a graph convolutional module between the pre-trained language model and the classifier.

[0036] The global event relationships are modeled using a message passing mechanism, and the event representations are updated accordingly; the learning process is represented as follows:

[0037]

[0038] in, To add self-loops to the adjacency matrix, , It is the identity matrix. for The degree matrix, These are the trainable parameters for this layer. This represents the characteristics of all nodes in this layer. Initial representation of the event .

[0039] Further, S4 includes:

[0040] Step 4.1 Output from a specific perspective These are parameters of a multinomial distribution, representing the amount of evidence for each candidate event, obtained through... Mapping evidence to Dirichlet concentration parameters The purpose of adding 1 is to ensure Nonnegativity;

[0041] Step 4.2: Derive the confidence level of each candidate event and the overall uncertainty of the candidate set from the Dirichlet distribution; the calculation formula is as follows:

[0042]

[0043] in, Let m be the confidence quality of view m for the k-th candidate event. The higher the confidence quality, the more evidence the model provides to support this candidate event. The number of candidate events, The overall quality of uncertainty is inversely proportional to the total amount of evidence. Also known as Dirichlet strength;

[0044] Step 4.3: Introduce a simplified Dempster-Shafer combination rule to integrate opinions from different sources, forming a joint opinion; define view opinions as a set of confidence quality and uncertainty quality, and text view opinions as... The diagram is as follows The fusion result Calculated using the following formula:

[0045]

[0046] in, and These represent the overall uncertainty of the text view and the graph view, respectively. and Text view and graph view respectively for the first The confidence level of each candidate event. It is a measure of conflict between viewpoints from different sources. and For the text view and graph view, calculate the confidence level of different candidate events.

[0047] Step 4.4: Obtain the fusion result Then, the Dirichlet concentration parameters are calculated in reverse and used for the final decision:

[0048]

[0049] By combining the rules, the concentration parameters of the integrated Dirichlet distribution were obtained. This allows us to calculate the final probability of each candidate event, i.e. .

[0050] The present invention also provides an event prediction system based on pairwise event relationship learning and multi-perspective evidence fusion, the system comprising: a module for executing the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion.

[0051] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the event prediction method based on pairwise event relationship learning and multi-view evidence fusion.

[0052] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion.

[0053] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion.

[0054] The beneficial effects of this invention are:

[0055] 1. By using probabilistic event relationship graphs, noise in traditional co-occurrence methods is reduced, improving the accuracy of event association;

[0056] 2. The multi-perspective evidence fusion mechanism enhances the model's ability to handle uncertainty and maintains stable performance in noisy environments;

[0057] 3. Without relying on external knowledge, it achieves an accuracy of 64.62% on the NYT dataset, which is better than existing baseline models. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0059] Figure 1 This is a flowchart of the present invention;

[0060] Figure 2 This is a flowchart of the present invention;

[0061] Figure 3 These are the results of the hyperparameter sensitivity experiment of this invention. Detailed Implementation

[0062] Example 1: As Figures 1-3 As shown, the event prediction method based on pairwise event relationship learning and multi-view evidence fusion includes:

[0063] S1. Paired event relationship learning: A supervised learning framework is adopted to train a pre-trained language model by constructing positive and negative samples to distinguish between relevant and irrelevant event pairs;

[0064] Further, S1 includes:

[0065] S1.1 Constructing positive and negative samples: Positive samples are defined as event pairs with semantic or temporal correlation, coming from the same event chain, because these events share the same narrative context; negative samples are defined as unrelated event pairs, coming from different event chains, because the probability of correlation between cross-chain events is low.

[0066] S1.2 Constructing a structured input sequence: Constructing a structured input sequence for each event pair: <s><Context Prompt><Context Events><Target Prompt>< >and< >< / s> .in, <contextprompt> It is a natural language phrase that guides the model to understand the context: "In the event chain context:";<Context Events> To arrange context events in chronological order, use [interchangeable] events.< / contextprompt> < / s> Separate;<Target Prompt> Used to separate and introduce target event pairs: "Now, determine the relation between:"; > and < > is the target event pair;

[0067] S1.3, Paired Event Relationship Inference and Model Optimization: Input the constructed sequence into the pre-trained language model, to... <s>Bit output vector As an aggregate representation of the entire input, then Input classifier; the process is described as follows:

[0068]

[0069] Where C represents the context event, and For the target event pairs that require relation inference. This represents the construction of a structured input sequence, with the classifier weight matrix being... The bias term is K is the number of categories, which are set as relevant / irrelevant, K=2. The loss function is the cross-entropy loss function commonly used in classification tasks, and the goal is to make the probability distribution... Fit the true labels.

[0070] In specific implementation, the present invention adopts the following parameter configuration:

[0071] RoBERTa-large was used as the pre-trained language model.

[0072] The maximum length of the context event sequence is set to 512 tokens;

[0073] The hidden layer dimension of the classifier is d=1024;

[0074] Set the batch size to 32;

[0075] The learning rate was set to 2e-5, and the AdamW optimizer was used.

[0076] S2. Constructing an event relationship graph: Using a pre-trained language model learned from pairwise event relationships, infer the dependencies between pairwise events, and use the probabilities output by the model as the weights of the edges to construct an event relationship graph;

[0077] Further, S2 includes:

[0078] Step 2.1, Paired Event Relationship Inference: Input event pairs into the pre-trained language model to obtain the probability distribution of relevance / irrelevance. =[ , In this probability distribution, the relevant probability is... The probability of being unrelated is , ;

[0079] Step 2.2, Event Relationship Graph Construction: Based on Relevant / Unrelated Probability Distributions =[ , Construct a weighted directed graph where the edge weights are determined by a probability distribution, described as follows:

[0080] ;

[0081] in, This represents a probability threshold; if the probability of association between two events exceeds a certain threshold... Then, connecting edges are established between them.

[0082] After completing the learning of pairwise event relationships, this invention uses the trained model to construct an event relationship graph. The specific implementation process includes:

[0083] Probability threshold setting: After grid search, the optimal probability threshold τ=0.5 was determined;

[0084] Graph construction algorithm: Adjacency list is used to store the graph structure, and sparse matrix is ​​used to store weight information;

[0085] Graph optimization: Prune the constructed graph by removing edges with weights below a threshold.

[0086] S3, Multi-perspective evidence learning: Learn quantitative evidence for candidate events from both textual semantic and graph structure perspectives;

[0087] Further, S3 includes:

[0088] Step 3.1, Text Semantic Perspective Learning: Concatenate the event chain and candidate events into a sequence, and obtain event representations through a pre-trained language model; design an attention pooling module, which realizes the attention aggregation between candidate events and historical events; use the candidate event itself as the query, and the event chain as the key and value, and calculate the weighted aggregate representation of the event chain through the attention mechanism;

[0089]

[0090]

[0091] in, The weighted aggregation representation of the event chain. Candidate events The hidden state, The hidden state of candidate events K and V are the hidden states of the event chain. n is the length of the event chain;

[0092] Finally, the weighted aggregate representation of the event chain ah and the candidate events are concatenated and fed into the classifier. The scalar output of the classifier is used as the relevance score of the candidate events.

[0093]

[0094] in, This represents the hidden state of a candidate event. , , , , The dimension of the hidden state. This indicates vector concatenation;

[0095] Step 3.2, Graph Structure Perspective Learning: Based on the event relationship graph, a graph convolutional network is used to update the event representation and calculate the relevance score; the graph structure perspective adds a graph convolutional module between the pre-trained language model and the classifier.

[0096] The global event relationships are modeled using a message passing mechanism, and the event representations are updated accordingly; the learning process is represented as follows:

[0097]

[0098] in, To add self-loops to the adjacency matrix, , It is the identity matrix. for The degree matrix, These are the trainable parameters for this layer. This represents the characteristics of all nodes in this layer. Initial representation of the event The graph convolutional network uses a 2-layer GCN structure with 512 hidden units per layer. It uses the ReLU activation function, adds a Dropout layer to prevent overfitting, and sets the dropout rate to 0.1.

[0099] S4. Credible Evidence Fusion: The uncertainty of each perspective is modeled using the Dirichlet distribution, and the evidence from multiple perspectives is dynamically fused through the Dempster-Shafer theory to generate the final prediction result.

[0100] Further, S4 includes:

[0101] Step 4.1 Output from a specific perspective These are parameters of a multinomial distribution, representing the amount of evidence for each candidate event, obtained through... Mapping evidence to Dirichlet concentration parameters The purpose of adding 1 is to ensure Nonnegativity;

[0102] Step 4.2: Derive the confidence level of each candidate event and the overall uncertainty of the candidate set from the Dirichlet distribution; the calculation formula is as follows:

[0103]

[0104] in, Let m be the confidence quality of view m for the k-th candidate event. The higher the confidence quality, the more evidence the model provides to support this candidate event. The number of candidate events, The overall quality of uncertainty is inversely proportional to the total amount of evidence. Also known as Dirichlet strength;

[0105] Step 4.3: Introduce a simplified Dempster-Shafer combination rule to integrate opinions from different sources into a joint opinion; define view opinions as a set of confidence quality and uncertainty quality, and text view opinions as... The diagram is as follows The fusion result Calculated using the following formula:

[0106]

[0107] in, and These represent the overall uncertainty of the text view and the graph view, respectively. and Text view and graph view respectively for the first The confidence level of each candidate event. It is a measure of conflict between viewpoints from different sources. and For the text view and graph view, calculate the confidence level of different candidate events.

[0108] Step 4.4: Obtain the fusion result Then, the Dirichlet concentration parameters are calculated in reverse and used for the final decision:

[0109]

[0110] By combining the rules, the concentration parameters of the integrated Dirichlet distribution were obtained. This allows us to calculate the final probability of each candidate event, i.e. .

[0111] This invention also provides an event prediction system based on pairwise event relationship learning and multi-perspective evidence fusion. The system includes a module for executing the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion. Specifically, it may include:

[0112] The pairwise event relationship learning module employs a supervised learning framework to train a pre-trained language model using positive and negative samples to distinguish between relevant and unrelated event pairs. It is used to construct an event relationship graph: leveraging the pre-trained language model after pairwise event relationship learning, it infers the dependencies between pairs of events, using the model's output probabilities as edge weights to construct the event relationship graph. This method helps reduce noise introduced by traditional co-occurrence statistical methods.

[0113] The evidence learning module is used to learn quantitative evidence for candidate events from both textual semantic and graph structure perspectives.

[0114] The credible evidence fusion module is used to model the uncertainty of each perspective using the Dirichlet distribution and dynamically fuse multi-perspective evidence through Dempster-Shafer theory to generate the final prediction result; it significantly improves the robustness of model decision-making.

[0115] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the event prediction method based on pairwise event relationship learning and multi-view evidence fusion.

[0116] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion.

[0117] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion.

[0118] In summary, this event prediction method is based on pairwise event relationship learning and multi-perspective evidence fusion. By learning pairwise event relationships, the model can perceive events and the connections between them, overcoming the noise interference problem in event relationships encountered by traditional methods. Simultaneously, the use of multi-perspective evidence fusion enables the model to make more reliable decisions even in the face of noise interference. Experimental results on public datasets demonstrate that the proposed method outperforms state-of-the-art baseline models without relying on external knowledge.

[0119] Furthermore, to verify the effectiveness of this event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion, the following comparative experiment was conducted in this embodiment:

[0120] The experiments were evaluated on the NYT dataset, which contains 160,331 samples, with 140,331 samples used as the training set, and 10,000 samples each for the development and test sets. Each sample consists of an event chain containing 8 events and a candidate list containing 5 candidate events. All events were extracted from the New York Times in the Gigaword corpus and represented as quadruples, i.e., (subject, verb, object, indirect object).

[0121] Table 1 shows the main experimental results.

[0122] Table 1 shows the performance on the NYT dataset, where our method outperforms all baseline models. From the experimental results, we can draw the following observations: Generally, models that utilize external knowledge outperform those that do not. For example, SGNN+Int+Senti outperforms SGNN by 3.58%, thanks to the external knowledge base enhancing the semantics of events and providing rich common-sense constraints on event relationships. Our method, however, achieves competitive results without relying on external knowledge, even outperforming state-of-the-art models that utilize external knowledge by 1%.

[0123] Furthermore, our method also outperforms strong baseline models that undergo extensive event-centric post-training, such as EventBERT. While EventBERT learns event knowledge through post-training on additional large corpora, it is costly to train, requiring 199.9M tokens with 500 GPU hours of computation. Our method achieves a 1% performance improvement over EventBERT, yet our proposed event-relationship-centric self-supervised learning method can be trained in just 20 GPU hours. This demonstrates that, compared to paths relying on expensive post-training or external knowledge injection, designing lightweight and accurate pre-training tasks can also effectively stimulate the event reasoning capabilities of language models.

[0124] Table 2 shows the results of the large-scale model experiment.

[0125] Table 2 presents the comparison with large models. Our method was tested on large generative language models such as ChatGPT1, Claude-3-5-sonnet, and Llama3.170b in a zero-shot setting. The experimental results are shown in Table 2. The accuracies of GPT-3.5 and Llama3.170B are only 28.36% and 39.32%, respectively, indicating significant difficulties in modeling event dependencies. Although Claude-3-5-sonnet and GPT-4o show some improvement over earlier models, they are still insufficient in handling subtle interactions between events. In contrast, smaller supervised learning models perform better, mainly due to the limitations of large models in effectively modeling complex event dependencies.

[0126] This invention also performs hyperparameter sensitivity analysis. Based on the pairwise event association probabilities output by the pre-trained language model, this invention constructs an event relationship graph. Specifically, when the association probability between two events is higher than a set probability threshold τ, an edge is added to the graph. To evaluate the impact of the probability threshold τ on the quality of event dependency representation, this invention compares the model performance under different τ values, and the results are as follows: Figure 3 As shown, the model achieves the best performance when the probability threshold τ is set to 0.5.

[0127] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.< / s>

Claims

1. An event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion, characterized in that: include: S1. Paired event relationship learning: A supervised learning framework is adopted to train a pre-trained language model by constructing positive and negative samples to distinguish between relevant and irrelevant event pairs; S2. Constructing an event relationship graph: Using a pre-trained language model learned from pairwise event relationships, infer the dependencies between pairwise events, and use the probabilities output by the model as the weights of the edges to construct an event relationship graph; S3, Multi-perspective evidence learning: Learn quantitative evidence for candidate events from both textual semantic and graph structure perspectives; S4. Credible Evidence Fusion: The uncertainty of each perspective is modeled using the Dirichlet distribution, and the evidence from multiple perspectives is dynamically fused through the Dempster-Shafer theory to generate the final prediction result.

2. The event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion according to claim 1, characterized in that: S1 includes: S1.1 Constructing positive and negative samples: Positive samples are defined as event pairs with semantic or temporal correlation, coming from the same event chain; negative samples are defined as unrelated event pairs, coming from different event chains. S1.2 Constructing a structured input sequence: Constructing a structured input sequence for each event pair; S1.3, Paired Event Relationship Inference and Model Optimization: Input the constructed sequence into the pre-trained language model, to... <s>Bit output vector As an aggregate representation of the entire input, then Input classifier; the process is described as follows:< / s> <s> ； Where C represents the context event, and For the target event pairs that require relation inference. This represents the construction of a structured input sequence, with the classifier weight matrix being... The bias term is K is the number of categories, which are set as relevant / irrelevant, K=2. The loss function is the cross-entropy loss function commonly used in classification tasks, and the goal is to make the probability distribution... Fit the true labels.

3. The event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion according to claim 1, characterized in that: S2 includes: Step 2.1, Paired Event Relationship Inference: Input event pairs into the pre-trained language model to obtain the probability distribution of relevance / irrelevance. =[ , In this probability distribution, the relevant probability is... The probability of being unrelated is , ; Step 2.2, Event Relationship Graph Construction: Based on Relevant / Unrelated Probability Distributions =[ , Construct a weighted directed graph where the edge weights are determined by a probability distribution, described as follows: ； in, This represents a probability threshold; if the probability of association between two events exceeds a certain threshold... Then, connecting edges are established between them.

4. The event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion according to claim 1, characterized in that: S3 includes: Step 3.1, Text Semantic Perspective Learning: Concatenate the event chain and candidate events into a sequence, and obtain event representations through a pre-trained language model; design an attention pooling module, which realizes the attention aggregation between candidate events and historical events; use the candidate event itself as the query, and the event chain as the key and value, and calculate the weighted aggregate representation of the event chain through the attention mechanism; ； ； in, The weighted aggregation representation of the event chain. Candidate events The hidden state, The hidden state of candidate events K and V are the hidden states of the event chain. n is the length of the event chain; Finally, the weighted aggregate representation of the event chain ah and the candidate events are concatenated and fed into the classifier. The scalar output of the classifier is used as the relevance score of the candidate events. ； in, This represents the hidden state of a candidate event. , , , , The dimension of the hidden state. This indicates vector concatenation; Step 3.2, Graph Structure Perspective Learning: Based on the event relationship graph, a graph convolutional network is used to update the event representation and calculate the relevance score; the graph structure perspective adds a graph convolutional module between the pre-trained language model and the classifier. The global event relationships are modeled using a message passing mechanism, and the event representations are updated accordingly; the learning process is represented as follows: ； in, To add self-loops to the adjacency matrix, , It is the identity matrix. for The degree matrix, These are the trainable parameters for this layer. This represents the characteristics of all nodes in this layer. Initial representation of the event .

5. The event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion according to claim 1, characterized in that: S4 includes: Step 4.1 Output from a specific perspective These are parameters of a multinomial distribution, representing the amount of evidence for each candidate event, obtained through... Mapping evidence to Dirichlet concentration parameters The purpose of adding 1 is to ensure Nonnegativity; Step 4.2: Derive the confidence level of each candidate event and the overall uncertainty of the candidate set from the Dirichlet distribution; the calculation formula is as follows: ； in, Let m be the confidence quality of view m for the k-th candidate event. The higher the confidence quality, the more evidence the model provides to support this candidate event. The number of candidate events, The overall quality of uncertainty is inversely proportional to the total amount of evidence. Also known as Dirichlet strength; Step 4.3: Introduce a simplified Dempster-Shafer combination rule to integrate opinions from different sources, forming a joint opinion; define view opinions as a set of confidence quality and uncertainty quality, and text view opinions as... The diagram is as follows The fusion result Calculated using the following formula: ； in, and These represent the overall uncertainty of the text view and the graph view, respectively. and Text view and graph view respectively for the first The confidence level of each candidate event. It is a measure of conflict between viewpoints from different sources. and For the text view and graph view, calculate the confidence level of different candidate events. Step 4.4: Obtain the fusion result Then, the Dirichlet concentration parameters are calculated in reverse and used for the final decision: ； By combining the rules, the concentration parameters of the integrated Dirichlet distribution were obtained. This allows us to calculate the final probability of each candidate event, i.e. .

6. An event prediction system based on pairwise event relationship learning and multi-perspective evidence fusion, characterized in that, The system includes a module for performing the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion as described in any one of claims 1 to 5.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the event prediction method based on pairwise event relationship learning and multi-perspective evidence fusion as described in any one of claims 1 to 5. < / s>

Images